core.c revision 383efcd00053ec40023010ce5034bd702e7ab373
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 77#include <asm/switch_to.h> 78#include <asm/tlb.h> 79#include <asm/irq_regs.h> 80#include <asm/mutex.h> 81#ifdef CONFIG_PARAVIRT 82#include <asm/paravirt.h> 83#endif 84 85#include "sched.h" 86#include "../workqueue_internal.h" 87#include "../smpboot.h" 88 89#define CREATE_TRACE_POINTS 90#include <trace/events/sched.h> 91 92void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 93{ 94 unsigned long delta; 95 ktime_t soft, hard, now; 96 97 for (;;) { 98 if (hrtimer_active(period_timer)) 99 break; 100 101 now = hrtimer_cb_get_time(period_timer); 102 hrtimer_forward(period_timer, now, period); 103 104 soft = hrtimer_get_softexpires(period_timer); 105 hard = hrtimer_get_expires(period_timer); 106 delta = ktime_to_ns(ktime_sub(hard, soft)); 107 __hrtimer_start_range_ns(period_timer, soft, delta, 108 HRTIMER_MODE_ABS_PINNED, 0); 109 } 110} 111 112DEFINE_MUTEX(sched_domains_mutex); 113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115static void update_rq_clock_task(struct rq *rq, s64 delta); 116 117void update_rq_clock(struct rq *rq) 118{ 119 s64 delta; 120 121 if (rq->skip_clock_update > 0) 122 return; 123 124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 125 rq->clock += delta; 126 update_rq_clock_task(rq, delta); 127} 128 129/* 130 * Debugging: various feature bits 131 */ 132 133#define SCHED_FEAT(name, enabled) \ 134 (1UL << __SCHED_FEAT_##name) * enabled | 135 136const_debug unsigned int sysctl_sched_features = 137#include "features.h" 138 0; 139 140#undef SCHED_FEAT 141 142#ifdef CONFIG_SCHED_DEBUG 143#define SCHED_FEAT(name, enabled) \ 144 #name , 145 146static const char * const sched_feat_names[] = { 147#include "features.h" 148}; 149 150#undef SCHED_FEAT 151 152static int sched_feat_show(struct seq_file *m, void *v) 153{ 154 int i; 155 156 for (i = 0; i < __SCHED_FEAT_NR; i++) { 157 if (!(sysctl_sched_features & (1UL << i))) 158 seq_puts(m, "NO_"); 159 seq_printf(m, "%s ", sched_feat_names[i]); 160 } 161 seq_puts(m, "\n"); 162 163 return 0; 164} 165 166#ifdef HAVE_JUMP_LABEL 167 168#define jump_label_key__true STATIC_KEY_INIT_TRUE 169#define jump_label_key__false STATIC_KEY_INIT_FALSE 170 171#define SCHED_FEAT(name, enabled) \ 172 jump_label_key__##enabled , 173 174struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 175#include "features.h" 176}; 177 178#undef SCHED_FEAT 179 180static void sched_feat_disable(int i) 181{ 182 if (static_key_enabled(&sched_feat_keys[i])) 183 static_key_slow_dec(&sched_feat_keys[i]); 184} 185 186static void sched_feat_enable(int i) 187{ 188 if (!static_key_enabled(&sched_feat_keys[i])) 189 static_key_slow_inc(&sched_feat_keys[i]); 190} 191#else 192static void sched_feat_disable(int i) { }; 193static void sched_feat_enable(int i) { }; 194#endif /* HAVE_JUMP_LABEL */ 195 196static int sched_feat_set(char *cmp) 197{ 198 int i; 199 int neg = 0; 200 201 if (strncmp(cmp, "NO_", 3) == 0) { 202 neg = 1; 203 cmp += 3; 204 } 205 206 for (i = 0; i < __SCHED_FEAT_NR; i++) { 207 if (strcmp(cmp, sched_feat_names[i]) == 0) { 208 if (neg) { 209 sysctl_sched_features &= ~(1UL << i); 210 sched_feat_disable(i); 211 } else { 212 sysctl_sched_features |= (1UL << i); 213 sched_feat_enable(i); 214 } 215 break; 216 } 217 } 218 219 return i; 220} 221 222static ssize_t 223sched_feat_write(struct file *filp, const char __user *ubuf, 224 size_t cnt, loff_t *ppos) 225{ 226 char buf[64]; 227 char *cmp; 228 int i; 229 230 if (cnt > 63) 231 cnt = 63; 232 233 if (copy_from_user(&buf, ubuf, cnt)) 234 return -EFAULT; 235 236 buf[cnt] = 0; 237 cmp = strstrip(buf); 238 239 i = sched_feat_set(cmp); 240 if (i == __SCHED_FEAT_NR) 241 return -EINVAL; 242 243 *ppos += cnt; 244 245 return cnt; 246} 247 248static int sched_feat_open(struct inode *inode, struct file *filp) 249{ 250 return single_open(filp, sched_feat_show, NULL); 251} 252 253static const struct file_operations sched_feat_fops = { 254 .open = sched_feat_open, 255 .write = sched_feat_write, 256 .read = seq_read, 257 .llseek = seq_lseek, 258 .release = single_release, 259}; 260 261static __init int sched_init_debug(void) 262{ 263 debugfs_create_file("sched_features", 0644, NULL, NULL, 264 &sched_feat_fops); 265 266 return 0; 267} 268late_initcall(sched_init_debug); 269#endif /* CONFIG_SCHED_DEBUG */ 270 271/* 272 * Number of tasks to iterate in a single balance run. 273 * Limited because this is done with IRQs disabled. 274 */ 275const_debug unsigned int sysctl_sched_nr_migrate = 32; 276 277/* 278 * period over which we average the RT time consumption, measured 279 * in ms. 280 * 281 * default: 1s 282 */ 283const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 284 285/* 286 * period over which we measure -rt task cpu usage in us. 287 * default: 1s 288 */ 289unsigned int sysctl_sched_rt_period = 1000000; 290 291__read_mostly int scheduler_running; 292 293/* 294 * part of the period that we allow rt tasks to run in us. 295 * default: 0.95s 296 */ 297int sysctl_sched_rt_runtime = 950000; 298 299 300 301/* 302 * __task_rq_lock - lock the rq @p resides on. 303 */ 304static inline struct rq *__task_rq_lock(struct task_struct *p) 305 __acquires(rq->lock) 306{ 307 struct rq *rq; 308 309 lockdep_assert_held(&p->pi_lock); 310 311 for (;;) { 312 rq = task_rq(p); 313 raw_spin_lock(&rq->lock); 314 if (likely(rq == task_rq(p))) 315 return rq; 316 raw_spin_unlock(&rq->lock); 317 } 318} 319 320/* 321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 322 */ 323static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 324 __acquires(p->pi_lock) 325 __acquires(rq->lock) 326{ 327 struct rq *rq; 328 329 for (;;) { 330 raw_spin_lock_irqsave(&p->pi_lock, *flags); 331 rq = task_rq(p); 332 raw_spin_lock(&rq->lock); 333 if (likely(rq == task_rq(p))) 334 return rq; 335 raw_spin_unlock(&rq->lock); 336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 337 } 338} 339 340static void __task_rq_unlock(struct rq *rq) 341 __releases(rq->lock) 342{ 343 raw_spin_unlock(&rq->lock); 344} 345 346static inline void 347task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 348 __releases(rq->lock) 349 __releases(p->pi_lock) 350{ 351 raw_spin_unlock(&rq->lock); 352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 353} 354 355/* 356 * this_rq_lock - lock this runqueue and disable interrupts. 357 */ 358static struct rq *this_rq_lock(void) 359 __acquires(rq->lock) 360{ 361 struct rq *rq; 362 363 local_irq_disable(); 364 rq = this_rq(); 365 raw_spin_lock(&rq->lock); 366 367 return rq; 368} 369 370#ifdef CONFIG_SCHED_HRTICK 371/* 372 * Use HR-timers to deliver accurate preemption points. 373 * 374 * Its all a bit involved since we cannot program an hrt while holding the 375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a 376 * reschedule event. 377 * 378 * When we get rescheduled we reprogram the hrtick_timer outside of the 379 * rq->lock. 380 */ 381 382static void hrtick_clear(struct rq *rq) 383{ 384 if (hrtimer_active(&rq->hrtick_timer)) 385 hrtimer_cancel(&rq->hrtick_timer); 386} 387 388/* 389 * High-resolution timer tick. 390 * Runs from hardirq context with interrupts disabled. 391 */ 392static enum hrtimer_restart hrtick(struct hrtimer *timer) 393{ 394 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 395 396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 397 398 raw_spin_lock(&rq->lock); 399 update_rq_clock(rq); 400 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 401 raw_spin_unlock(&rq->lock); 402 403 return HRTIMER_NORESTART; 404} 405 406#ifdef CONFIG_SMP 407/* 408 * called from hardirq (IPI) context 409 */ 410static void __hrtick_start(void *arg) 411{ 412 struct rq *rq = arg; 413 414 raw_spin_lock(&rq->lock); 415 hrtimer_restart(&rq->hrtick_timer); 416 rq->hrtick_csd_pending = 0; 417 raw_spin_unlock(&rq->lock); 418} 419 420/* 421 * Called to set the hrtick timer state. 422 * 423 * called with rq->lock held and irqs disabled 424 */ 425void hrtick_start(struct rq *rq, u64 delay) 426{ 427 struct hrtimer *timer = &rq->hrtick_timer; 428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 429 430 hrtimer_set_expires(timer, time); 431 432 if (rq == this_rq()) { 433 hrtimer_restart(timer); 434 } else if (!rq->hrtick_csd_pending) { 435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 436 rq->hrtick_csd_pending = 1; 437 } 438} 439 440static int 441hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 442{ 443 int cpu = (int)(long)hcpu; 444 445 switch (action) { 446 case CPU_UP_CANCELED: 447 case CPU_UP_CANCELED_FROZEN: 448 case CPU_DOWN_PREPARE: 449 case CPU_DOWN_PREPARE_FROZEN: 450 case CPU_DEAD: 451 case CPU_DEAD_FROZEN: 452 hrtick_clear(cpu_rq(cpu)); 453 return NOTIFY_OK; 454 } 455 456 return NOTIFY_DONE; 457} 458 459static __init void init_hrtick(void) 460{ 461 hotcpu_notifier(hotplug_hrtick, 0); 462} 463#else 464/* 465 * Called to set the hrtick timer state. 466 * 467 * called with rq->lock held and irqs disabled 468 */ 469void hrtick_start(struct rq *rq, u64 delay) 470{ 471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 472 HRTIMER_MODE_REL_PINNED, 0); 473} 474 475static inline void init_hrtick(void) 476{ 477} 478#endif /* CONFIG_SMP */ 479 480static void init_rq_hrtick(struct rq *rq) 481{ 482#ifdef CONFIG_SMP 483 rq->hrtick_csd_pending = 0; 484 485 rq->hrtick_csd.flags = 0; 486 rq->hrtick_csd.func = __hrtick_start; 487 rq->hrtick_csd.info = rq; 488#endif 489 490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 491 rq->hrtick_timer.function = hrtick; 492} 493#else /* CONFIG_SCHED_HRTICK */ 494static inline void hrtick_clear(struct rq *rq) 495{ 496} 497 498static inline void init_rq_hrtick(struct rq *rq) 499{ 500} 501 502static inline void init_hrtick(void) 503{ 504} 505#endif /* CONFIG_SCHED_HRTICK */ 506 507/* 508 * resched_task - mark a task 'to be rescheduled now'. 509 * 510 * On UP this means the setting of the need_resched flag, on SMP it 511 * might also involve a cross-CPU call to trigger the scheduler on 512 * the target CPU. 513 */ 514#ifdef CONFIG_SMP 515 516#ifndef tsk_is_polling 517#define tsk_is_polling(t) 0 518#endif 519 520void resched_task(struct task_struct *p) 521{ 522 int cpu; 523 524 assert_raw_spin_locked(&task_rq(p)->lock); 525 526 if (test_tsk_need_resched(p)) 527 return; 528 529 set_tsk_need_resched(p); 530 531 cpu = task_cpu(p); 532 if (cpu == smp_processor_id()) 533 return; 534 535 /* NEED_RESCHED must be visible before we test polling */ 536 smp_mb(); 537 if (!tsk_is_polling(p)) 538 smp_send_reschedule(cpu); 539} 540 541void resched_cpu(int cpu) 542{ 543 struct rq *rq = cpu_rq(cpu); 544 unsigned long flags; 545 546 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 547 return; 548 resched_task(cpu_curr(cpu)); 549 raw_spin_unlock_irqrestore(&rq->lock, flags); 550} 551 552#ifdef CONFIG_NO_HZ 553/* 554 * In the semi idle case, use the nearest busy cpu for migrating timers 555 * from an idle cpu. This is good for power-savings. 556 * 557 * We don't do similar optimization for completely idle system, as 558 * selecting an idle cpu will add more delays to the timers than intended 559 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 560 */ 561int get_nohz_timer_target(void) 562{ 563 int cpu = smp_processor_id(); 564 int i; 565 struct sched_domain *sd; 566 567 rcu_read_lock(); 568 for_each_domain(cpu, sd) { 569 for_each_cpu(i, sched_domain_span(sd)) { 570 if (!idle_cpu(i)) { 571 cpu = i; 572 goto unlock; 573 } 574 } 575 } 576unlock: 577 rcu_read_unlock(); 578 return cpu; 579} 580/* 581 * When add_timer_on() enqueues a timer into the timer wheel of an 582 * idle CPU then this timer might expire before the next timer event 583 * which is scheduled to wake up that CPU. In case of a completely 584 * idle system the next event might even be infinite time into the 585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 586 * leaves the inner idle loop so the newly added timer is taken into 587 * account when the CPU goes back to idle and evaluates the timer 588 * wheel for the next timer event. 589 */ 590void wake_up_idle_cpu(int cpu) 591{ 592 struct rq *rq = cpu_rq(cpu); 593 594 if (cpu == smp_processor_id()) 595 return; 596 597 /* 598 * This is safe, as this function is called with the timer 599 * wheel base lock of (cpu) held. When the CPU is on the way 600 * to idle and has not yet set rq->curr to idle then it will 601 * be serialized on the timer wheel base lock and take the new 602 * timer into account automatically. 603 */ 604 if (rq->curr != rq->idle) 605 return; 606 607 /* 608 * We can set TIF_RESCHED on the idle task of the other CPU 609 * lockless. The worst case is that the other CPU runs the 610 * idle task through an additional NOOP schedule() 611 */ 612 set_tsk_need_resched(rq->idle); 613 614 /* NEED_RESCHED must be visible before we test polling */ 615 smp_mb(); 616 if (!tsk_is_polling(rq->idle)) 617 smp_send_reschedule(cpu); 618} 619 620static inline bool got_nohz_idle_kick(void) 621{ 622 int cpu = smp_processor_id(); 623 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 624} 625 626#else /* CONFIG_NO_HZ */ 627 628static inline bool got_nohz_idle_kick(void) 629{ 630 return false; 631} 632 633#endif /* CONFIG_NO_HZ */ 634 635void sched_avg_update(struct rq *rq) 636{ 637 s64 period = sched_avg_period(); 638 639 while ((s64)(rq->clock - rq->age_stamp) > period) { 640 /* 641 * Inline assembly required to prevent the compiler 642 * optimising this loop into a divmod call. 643 * See __iter_div_u64_rem() for another example of this. 644 */ 645 asm("" : "+rm" (rq->age_stamp)); 646 rq->age_stamp += period; 647 rq->rt_avg /= 2; 648 } 649} 650 651#else /* !CONFIG_SMP */ 652void resched_task(struct task_struct *p) 653{ 654 assert_raw_spin_locked(&task_rq(p)->lock); 655 set_tsk_need_resched(p); 656} 657#endif /* CONFIG_SMP */ 658 659#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 660 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 661/* 662 * Iterate task_group tree rooted at *from, calling @down when first entering a 663 * node and @up when leaving it for the final time. 664 * 665 * Caller must hold rcu_lock or sufficient equivalent. 666 */ 667int walk_tg_tree_from(struct task_group *from, 668 tg_visitor down, tg_visitor up, void *data) 669{ 670 struct task_group *parent, *child; 671 int ret; 672 673 parent = from; 674 675down: 676 ret = (*down)(parent, data); 677 if (ret) 678 goto out; 679 list_for_each_entry_rcu(child, &parent->children, siblings) { 680 parent = child; 681 goto down; 682 683up: 684 continue; 685 } 686 ret = (*up)(parent, data); 687 if (ret || parent == from) 688 goto out; 689 690 child = parent; 691 parent = parent->parent; 692 if (parent) 693 goto up; 694out: 695 return ret; 696} 697 698int tg_nop(struct task_group *tg, void *data) 699{ 700 return 0; 701} 702#endif 703 704static void set_load_weight(struct task_struct *p) 705{ 706 int prio = p->static_prio - MAX_RT_PRIO; 707 struct load_weight *load = &p->se.load; 708 709 /* 710 * SCHED_IDLE tasks get minimal weight: 711 */ 712 if (p->policy == SCHED_IDLE) { 713 load->weight = scale_load(WEIGHT_IDLEPRIO); 714 load->inv_weight = WMULT_IDLEPRIO; 715 return; 716 } 717 718 load->weight = scale_load(prio_to_weight[prio]); 719 load->inv_weight = prio_to_wmult[prio]; 720} 721 722static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 723{ 724 update_rq_clock(rq); 725 sched_info_queued(p); 726 p->sched_class->enqueue_task(rq, p, flags); 727} 728 729static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 730{ 731 update_rq_clock(rq); 732 sched_info_dequeued(p); 733 p->sched_class->dequeue_task(rq, p, flags); 734} 735 736void activate_task(struct rq *rq, struct task_struct *p, int flags) 737{ 738 if (task_contributes_to_load(p)) 739 rq->nr_uninterruptible--; 740 741 enqueue_task(rq, p, flags); 742} 743 744void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 745{ 746 if (task_contributes_to_load(p)) 747 rq->nr_uninterruptible++; 748 749 dequeue_task(rq, p, flags); 750} 751 752static void update_rq_clock_task(struct rq *rq, s64 delta) 753{ 754/* 755 * In theory, the compile should just see 0 here, and optimize out the call 756 * to sched_rt_avg_update. But I don't trust it... 757 */ 758#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 759 s64 steal = 0, irq_delta = 0; 760#endif 761#ifdef CONFIG_IRQ_TIME_ACCOUNTING 762 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 763 764 /* 765 * Since irq_time is only updated on {soft,}irq_exit, we might run into 766 * this case when a previous update_rq_clock() happened inside a 767 * {soft,}irq region. 768 * 769 * When this happens, we stop ->clock_task and only update the 770 * prev_irq_time stamp to account for the part that fit, so that a next 771 * update will consume the rest. This ensures ->clock_task is 772 * monotonic. 773 * 774 * It does however cause some slight miss-attribution of {soft,}irq 775 * time, a more accurate solution would be to update the irq_time using 776 * the current rq->clock timestamp, except that would require using 777 * atomic ops. 778 */ 779 if (irq_delta > delta) 780 irq_delta = delta; 781 782 rq->prev_irq_time += irq_delta; 783 delta -= irq_delta; 784#endif 785#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 786 if (static_key_false((¶virt_steal_rq_enabled))) { 787 u64 st; 788 789 steal = paravirt_steal_clock(cpu_of(rq)); 790 steal -= rq->prev_steal_time_rq; 791 792 if (unlikely(steal > delta)) 793 steal = delta; 794 795 st = steal_ticks(steal); 796 steal = st * TICK_NSEC; 797 798 rq->prev_steal_time_rq += steal; 799 800 delta -= steal; 801 } 802#endif 803 804 rq->clock_task += delta; 805 806#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 807 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 808 sched_rt_avg_update(rq, irq_delta + steal); 809#endif 810} 811 812void sched_set_stop_task(int cpu, struct task_struct *stop) 813{ 814 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 815 struct task_struct *old_stop = cpu_rq(cpu)->stop; 816 817 if (stop) { 818 /* 819 * Make it appear like a SCHED_FIFO task, its something 820 * userspace knows about and won't get confused about. 821 * 822 * Also, it will make PI more or less work without too 823 * much confusion -- but then, stop work should not 824 * rely on PI working anyway. 825 */ 826 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 827 828 stop->sched_class = &stop_sched_class; 829 } 830 831 cpu_rq(cpu)->stop = stop; 832 833 if (old_stop) { 834 /* 835 * Reset it back to a normal scheduling class so that 836 * it can die in pieces. 837 */ 838 old_stop->sched_class = &rt_sched_class; 839 } 840} 841 842/* 843 * __normal_prio - return the priority that is based on the static prio 844 */ 845static inline int __normal_prio(struct task_struct *p) 846{ 847 return p->static_prio; 848} 849 850/* 851 * Calculate the expected normal priority: i.e. priority 852 * without taking RT-inheritance into account. Might be 853 * boosted by interactivity modifiers. Changes upon fork, 854 * setprio syscalls, and whenever the interactivity 855 * estimator recalculates. 856 */ 857static inline int normal_prio(struct task_struct *p) 858{ 859 int prio; 860 861 if (task_has_rt_policy(p)) 862 prio = MAX_RT_PRIO-1 - p->rt_priority; 863 else 864 prio = __normal_prio(p); 865 return prio; 866} 867 868/* 869 * Calculate the current priority, i.e. the priority 870 * taken into account by the scheduler. This value might 871 * be boosted by RT tasks, or might be boosted by 872 * interactivity modifiers. Will be RT if the task got 873 * RT-boosted. If not then it returns p->normal_prio. 874 */ 875static int effective_prio(struct task_struct *p) 876{ 877 p->normal_prio = normal_prio(p); 878 /* 879 * If we are RT tasks or we were boosted to RT priority, 880 * keep the priority unchanged. Otherwise, update priority 881 * to the normal priority: 882 */ 883 if (!rt_prio(p->prio)) 884 return p->normal_prio; 885 return p->prio; 886} 887 888/** 889 * task_curr - is this task currently executing on a CPU? 890 * @p: the task in question. 891 */ 892inline int task_curr(const struct task_struct *p) 893{ 894 return cpu_curr(task_cpu(p)) == p; 895} 896 897static inline void check_class_changed(struct rq *rq, struct task_struct *p, 898 const struct sched_class *prev_class, 899 int oldprio) 900{ 901 if (prev_class != p->sched_class) { 902 if (prev_class->switched_from) 903 prev_class->switched_from(rq, p); 904 p->sched_class->switched_to(rq, p); 905 } else if (oldprio != p->prio) 906 p->sched_class->prio_changed(rq, p, oldprio); 907} 908 909void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 910{ 911 const struct sched_class *class; 912 913 if (p->sched_class == rq->curr->sched_class) { 914 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 915 } else { 916 for_each_class(class) { 917 if (class == rq->curr->sched_class) 918 break; 919 if (class == p->sched_class) { 920 resched_task(rq->curr); 921 break; 922 } 923 } 924 } 925 926 /* 927 * A queue event has occurred, and we're going to schedule. In 928 * this case, we can save a useless back to back clock update. 929 */ 930 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 931 rq->skip_clock_update = 1; 932} 933 934static ATOMIC_NOTIFIER_HEAD(task_migration_notifier); 935 936void register_task_migration_notifier(struct notifier_block *n) 937{ 938 atomic_notifier_chain_register(&task_migration_notifier, n); 939} 940 941#ifdef CONFIG_SMP 942void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 943{ 944#ifdef CONFIG_SCHED_DEBUG 945 /* 946 * We should never call set_task_cpu() on a blocked task, 947 * ttwu() will sort out the placement. 948 */ 949 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 950 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 951 952#ifdef CONFIG_LOCKDEP 953 /* 954 * The caller should hold either p->pi_lock or rq->lock, when changing 955 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 956 * 957 * sched_move_task() holds both and thus holding either pins the cgroup, 958 * see task_group(). 959 * 960 * Furthermore, all task_rq users should acquire both locks, see 961 * task_rq_lock(). 962 */ 963 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 964 lockdep_is_held(&task_rq(p)->lock))); 965#endif 966#endif 967 968 trace_sched_migrate_task(p, new_cpu); 969 970 if (task_cpu(p) != new_cpu) { 971 struct task_migration_notifier tmn; 972 973 if (p->sched_class->migrate_task_rq) 974 p->sched_class->migrate_task_rq(p, new_cpu); 975 p->se.nr_migrations++; 976 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 977 978 tmn.task = p; 979 tmn.from_cpu = task_cpu(p); 980 tmn.to_cpu = new_cpu; 981 982 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn); 983 } 984 985 __set_task_cpu(p, new_cpu); 986} 987 988struct migration_arg { 989 struct task_struct *task; 990 int dest_cpu; 991}; 992 993static int migration_cpu_stop(void *data); 994 995/* 996 * wait_task_inactive - wait for a thread to unschedule. 997 * 998 * If @match_state is nonzero, it's the @p->state value just checked and 999 * not expected to change. If it changes, i.e. @p might have woken up, 1000 * then return zero. When we succeed in waiting for @p to be off its CPU, 1001 * we return a positive number (its total switch count). If a second call 1002 * a short while later returns the same number, the caller can be sure that 1003 * @p has remained unscheduled the whole time. 1004 * 1005 * The caller must ensure that the task *will* unschedule sometime soon, 1006 * else this function might spin for a *long* time. This function can't 1007 * be called with interrupts off, or it may introduce deadlock with 1008 * smp_call_function() if an IPI is sent by the same process we are 1009 * waiting to become inactive. 1010 */ 1011unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1012{ 1013 unsigned long flags; 1014 int running, on_rq; 1015 unsigned long ncsw; 1016 struct rq *rq; 1017 1018 for (;;) { 1019 /* 1020 * We do the initial early heuristics without holding 1021 * any task-queue locks at all. We'll only try to get 1022 * the runqueue lock when things look like they will 1023 * work out! 1024 */ 1025 rq = task_rq(p); 1026 1027 /* 1028 * If the task is actively running on another CPU 1029 * still, just relax and busy-wait without holding 1030 * any locks. 1031 * 1032 * NOTE! Since we don't hold any locks, it's not 1033 * even sure that "rq" stays as the right runqueue! 1034 * But we don't care, since "task_running()" will 1035 * return false if the runqueue has changed and p 1036 * is actually now running somewhere else! 1037 */ 1038 while (task_running(rq, p)) { 1039 if (match_state && unlikely(p->state != match_state)) 1040 return 0; 1041 cpu_relax(); 1042 } 1043 1044 /* 1045 * Ok, time to look more closely! We need the rq 1046 * lock now, to be *sure*. If we're wrong, we'll 1047 * just go back and repeat. 1048 */ 1049 rq = task_rq_lock(p, &flags); 1050 trace_sched_wait_task(p); 1051 running = task_running(rq, p); 1052 on_rq = p->on_rq; 1053 ncsw = 0; 1054 if (!match_state || p->state == match_state) 1055 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1056 task_rq_unlock(rq, p, &flags); 1057 1058 /* 1059 * If it changed from the expected state, bail out now. 1060 */ 1061 if (unlikely(!ncsw)) 1062 break; 1063 1064 /* 1065 * Was it really running after all now that we 1066 * checked with the proper locks actually held? 1067 * 1068 * Oops. Go back and try again.. 1069 */ 1070 if (unlikely(running)) { 1071 cpu_relax(); 1072 continue; 1073 } 1074 1075 /* 1076 * It's not enough that it's not actively running, 1077 * it must be off the runqueue _entirely_, and not 1078 * preempted! 1079 * 1080 * So if it was still runnable (but just not actively 1081 * running right now), it's preempted, and we should 1082 * yield - it could be a while. 1083 */ 1084 if (unlikely(on_rq)) { 1085 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1086 1087 set_current_state(TASK_UNINTERRUPTIBLE); 1088 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1089 continue; 1090 } 1091 1092 /* 1093 * Ahh, all good. It wasn't running, and it wasn't 1094 * runnable, which means that it will never become 1095 * running in the future either. We're all done! 1096 */ 1097 break; 1098 } 1099 1100 return ncsw; 1101} 1102 1103/*** 1104 * kick_process - kick a running thread to enter/exit the kernel 1105 * @p: the to-be-kicked thread 1106 * 1107 * Cause a process which is running on another CPU to enter 1108 * kernel-mode, without any delay. (to get signals handled.) 1109 * 1110 * NOTE: this function doesn't have to take the runqueue lock, 1111 * because all it wants to ensure is that the remote task enters 1112 * the kernel. If the IPI races and the task has been migrated 1113 * to another CPU then no harm is done and the purpose has been 1114 * achieved as well. 1115 */ 1116void kick_process(struct task_struct *p) 1117{ 1118 int cpu; 1119 1120 preempt_disable(); 1121 cpu = task_cpu(p); 1122 if ((cpu != smp_processor_id()) && task_curr(p)) 1123 smp_send_reschedule(cpu); 1124 preempt_enable(); 1125} 1126EXPORT_SYMBOL_GPL(kick_process); 1127#endif /* CONFIG_SMP */ 1128 1129#ifdef CONFIG_SMP 1130/* 1131 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1132 */ 1133static int select_fallback_rq(int cpu, struct task_struct *p) 1134{ 1135 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); 1136 enum { cpuset, possible, fail } state = cpuset; 1137 int dest_cpu; 1138 1139 /* Look for allowed, online CPU in same node. */ 1140 for_each_cpu(dest_cpu, nodemask) { 1141 if (!cpu_online(dest_cpu)) 1142 continue; 1143 if (!cpu_active(dest_cpu)) 1144 continue; 1145 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1146 return dest_cpu; 1147 } 1148 1149 for (;;) { 1150 /* Any allowed, online CPU? */ 1151 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1152 if (!cpu_online(dest_cpu)) 1153 continue; 1154 if (!cpu_active(dest_cpu)) 1155 continue; 1156 goto out; 1157 } 1158 1159 switch (state) { 1160 case cpuset: 1161 /* No more Mr. Nice Guy. */ 1162 cpuset_cpus_allowed_fallback(p); 1163 state = possible; 1164 break; 1165 1166 case possible: 1167 do_set_cpus_allowed(p, cpu_possible_mask); 1168 state = fail; 1169 break; 1170 1171 case fail: 1172 BUG(); 1173 break; 1174 } 1175 } 1176 1177out: 1178 if (state != cpuset) { 1179 /* 1180 * Don't tell them about moving exiting tasks or 1181 * kernel threads (both mm NULL), since they never 1182 * leave kernel. 1183 */ 1184 if (p->mm && printk_ratelimit()) { 1185 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1186 task_pid_nr(p), p->comm, cpu); 1187 } 1188 } 1189 1190 return dest_cpu; 1191} 1192 1193/* 1194 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1195 */ 1196static inline 1197int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1198{ 1199 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1200 1201 /* 1202 * In order not to call set_task_cpu() on a blocking task we need 1203 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1204 * cpu. 1205 * 1206 * Since this is common to all placement strategies, this lives here. 1207 * 1208 * [ this allows ->select_task() to simply return task_cpu(p) and 1209 * not worry about this generic constraint ] 1210 */ 1211 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1212 !cpu_online(cpu))) 1213 cpu = select_fallback_rq(task_cpu(p), p); 1214 1215 return cpu; 1216} 1217 1218static void update_avg(u64 *avg, u64 sample) 1219{ 1220 s64 diff = sample - *avg; 1221 *avg += diff >> 3; 1222} 1223#endif 1224 1225static void 1226ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1227{ 1228#ifdef CONFIG_SCHEDSTATS 1229 struct rq *rq = this_rq(); 1230 1231#ifdef CONFIG_SMP 1232 int this_cpu = smp_processor_id(); 1233 1234 if (cpu == this_cpu) { 1235 schedstat_inc(rq, ttwu_local); 1236 schedstat_inc(p, se.statistics.nr_wakeups_local); 1237 } else { 1238 struct sched_domain *sd; 1239 1240 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1241 rcu_read_lock(); 1242 for_each_domain(this_cpu, sd) { 1243 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1244 schedstat_inc(sd, ttwu_wake_remote); 1245 break; 1246 } 1247 } 1248 rcu_read_unlock(); 1249 } 1250 1251 if (wake_flags & WF_MIGRATED) 1252 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1253 1254#endif /* CONFIG_SMP */ 1255 1256 schedstat_inc(rq, ttwu_count); 1257 schedstat_inc(p, se.statistics.nr_wakeups); 1258 1259 if (wake_flags & WF_SYNC) 1260 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1261 1262#endif /* CONFIG_SCHEDSTATS */ 1263} 1264 1265static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1266{ 1267 activate_task(rq, p, en_flags); 1268 p->on_rq = 1; 1269 1270 /* if a worker is waking up, notify workqueue */ 1271 if (p->flags & PF_WQ_WORKER) 1272 wq_worker_waking_up(p, cpu_of(rq)); 1273} 1274 1275/* 1276 * Mark the task runnable and perform wakeup-preemption. 1277 */ 1278static void 1279ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1280{ 1281 trace_sched_wakeup(p, true); 1282 check_preempt_curr(rq, p, wake_flags); 1283 1284 p->state = TASK_RUNNING; 1285#ifdef CONFIG_SMP 1286 if (p->sched_class->task_woken) 1287 p->sched_class->task_woken(rq, p); 1288 1289 if (rq->idle_stamp) { 1290 u64 delta = rq->clock - rq->idle_stamp; 1291 u64 max = 2*sysctl_sched_migration_cost; 1292 1293 if (delta > max) 1294 rq->avg_idle = max; 1295 else 1296 update_avg(&rq->avg_idle, delta); 1297 rq->idle_stamp = 0; 1298 } 1299#endif 1300} 1301 1302static void 1303ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1304{ 1305#ifdef CONFIG_SMP 1306 if (p->sched_contributes_to_load) 1307 rq->nr_uninterruptible--; 1308#endif 1309 1310 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1311 ttwu_do_wakeup(rq, p, wake_flags); 1312} 1313 1314/* 1315 * Called in case the task @p isn't fully descheduled from its runqueue, 1316 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1317 * since all we need to do is flip p->state to TASK_RUNNING, since 1318 * the task is still ->on_rq. 1319 */ 1320static int ttwu_remote(struct task_struct *p, int wake_flags) 1321{ 1322 struct rq *rq; 1323 int ret = 0; 1324 1325 rq = __task_rq_lock(p); 1326 if (p->on_rq) { 1327 ttwu_do_wakeup(rq, p, wake_flags); 1328 ret = 1; 1329 } 1330 __task_rq_unlock(rq); 1331 1332 return ret; 1333} 1334 1335#ifdef CONFIG_SMP 1336static void sched_ttwu_pending(void) 1337{ 1338 struct rq *rq = this_rq(); 1339 struct llist_node *llist = llist_del_all(&rq->wake_list); 1340 struct task_struct *p; 1341 1342 raw_spin_lock(&rq->lock); 1343 1344 while (llist) { 1345 p = llist_entry(llist, struct task_struct, wake_entry); 1346 llist = llist_next(llist); 1347 ttwu_do_activate(rq, p, 0); 1348 } 1349 1350 raw_spin_unlock(&rq->lock); 1351} 1352 1353void scheduler_ipi(void) 1354{ 1355 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1356 return; 1357 1358 /* 1359 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1360 * traditionally all their work was done from the interrupt return 1361 * path. Now that we actually do some work, we need to make sure 1362 * we do call them. 1363 * 1364 * Some archs already do call them, luckily irq_enter/exit nest 1365 * properly. 1366 * 1367 * Arguably we should visit all archs and update all handlers, 1368 * however a fair share of IPIs are still resched only so this would 1369 * somewhat pessimize the simple resched case. 1370 */ 1371 irq_enter(); 1372 sched_ttwu_pending(); 1373 1374 /* 1375 * Check if someone kicked us for doing the nohz idle load balance. 1376 */ 1377 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1378 this_rq()->idle_balance = 1; 1379 raise_softirq_irqoff(SCHED_SOFTIRQ); 1380 } 1381 irq_exit(); 1382} 1383 1384static void ttwu_queue_remote(struct task_struct *p, int cpu) 1385{ 1386 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1387 smp_send_reschedule(cpu); 1388} 1389 1390bool cpus_share_cache(int this_cpu, int that_cpu) 1391{ 1392 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1393} 1394#endif /* CONFIG_SMP */ 1395 1396static void ttwu_queue(struct task_struct *p, int cpu) 1397{ 1398 struct rq *rq = cpu_rq(cpu); 1399 1400#if defined(CONFIG_SMP) 1401 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1402 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1403 ttwu_queue_remote(p, cpu); 1404 return; 1405 } 1406#endif 1407 1408 raw_spin_lock(&rq->lock); 1409 ttwu_do_activate(rq, p, 0); 1410 raw_spin_unlock(&rq->lock); 1411} 1412 1413/** 1414 * try_to_wake_up - wake up a thread 1415 * @p: the thread to be awakened 1416 * @state: the mask of task states that can be woken 1417 * @wake_flags: wake modifier flags (WF_*) 1418 * 1419 * Put it on the run-queue if it's not already there. The "current" 1420 * thread is always on the run-queue (except when the actual 1421 * re-schedule is in progress), and as such you're allowed to do 1422 * the simpler "current->state = TASK_RUNNING" to mark yourself 1423 * runnable without the overhead of this. 1424 * 1425 * Returns %true if @p was woken up, %false if it was already running 1426 * or @state didn't match @p's state. 1427 */ 1428static int 1429try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1430{ 1431 unsigned long flags; 1432 int cpu, success = 0; 1433 1434 smp_wmb(); 1435 raw_spin_lock_irqsave(&p->pi_lock, flags); 1436 if (!(p->state & state)) 1437 goto out; 1438 1439 success = 1; /* we're going to change ->state */ 1440 cpu = task_cpu(p); 1441 1442 if (p->on_rq && ttwu_remote(p, wake_flags)) 1443 goto stat; 1444 1445#ifdef CONFIG_SMP 1446 /* 1447 * If the owning (remote) cpu is still in the middle of schedule() with 1448 * this task as prev, wait until its done referencing the task. 1449 */ 1450 while (p->on_cpu) 1451 cpu_relax(); 1452 /* 1453 * Pairs with the smp_wmb() in finish_lock_switch(). 1454 */ 1455 smp_rmb(); 1456 1457 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1458 p->state = TASK_WAKING; 1459 1460 if (p->sched_class->task_waking) 1461 p->sched_class->task_waking(p); 1462 1463 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1464 if (task_cpu(p) != cpu) { 1465 wake_flags |= WF_MIGRATED; 1466 set_task_cpu(p, cpu); 1467 } 1468#endif /* CONFIG_SMP */ 1469 1470 ttwu_queue(p, cpu); 1471stat: 1472 ttwu_stat(p, cpu, wake_flags); 1473out: 1474 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1475 1476 return success; 1477} 1478 1479/** 1480 * try_to_wake_up_local - try to wake up a local task with rq lock held 1481 * @p: the thread to be awakened 1482 * 1483 * Put @p on the run-queue if it's not already there. The caller must 1484 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1485 * the current task. 1486 */ 1487static void try_to_wake_up_local(struct task_struct *p) 1488{ 1489 struct rq *rq = task_rq(p); 1490 1491 if (WARN_ON_ONCE(rq != this_rq()) || 1492 WARN_ON_ONCE(p == current)) 1493 return; 1494 1495 lockdep_assert_held(&rq->lock); 1496 1497 if (!raw_spin_trylock(&p->pi_lock)) { 1498 raw_spin_unlock(&rq->lock); 1499 raw_spin_lock(&p->pi_lock); 1500 raw_spin_lock(&rq->lock); 1501 } 1502 1503 if (!(p->state & TASK_NORMAL)) 1504 goto out; 1505 1506 if (!p->on_rq) 1507 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1508 1509 ttwu_do_wakeup(rq, p, 0); 1510 ttwu_stat(p, smp_processor_id(), 0); 1511out: 1512 raw_spin_unlock(&p->pi_lock); 1513} 1514 1515/** 1516 * wake_up_process - Wake up a specific process 1517 * @p: The process to be woken up. 1518 * 1519 * Attempt to wake up the nominated process and move it to the set of runnable 1520 * processes. Returns 1 if the process was woken up, 0 if it was already 1521 * running. 1522 * 1523 * It may be assumed that this function implies a write memory barrier before 1524 * changing the task state if and only if any tasks are woken up. 1525 */ 1526int wake_up_process(struct task_struct *p) 1527{ 1528 WARN_ON(task_is_stopped_or_traced(p)); 1529 return try_to_wake_up(p, TASK_NORMAL, 0); 1530} 1531EXPORT_SYMBOL(wake_up_process); 1532 1533int wake_up_state(struct task_struct *p, unsigned int state) 1534{ 1535 return try_to_wake_up(p, state, 0); 1536} 1537 1538/* 1539 * Perform scheduler related setup for a newly forked process p. 1540 * p is forked by current. 1541 * 1542 * __sched_fork() is basic setup used by init_idle() too: 1543 */ 1544static void __sched_fork(struct task_struct *p) 1545{ 1546 p->on_rq = 0; 1547 1548 p->se.on_rq = 0; 1549 p->se.exec_start = 0; 1550 p->se.sum_exec_runtime = 0; 1551 p->se.prev_sum_exec_runtime = 0; 1552 p->se.nr_migrations = 0; 1553 p->se.vruntime = 0; 1554 INIT_LIST_HEAD(&p->se.group_node); 1555 1556/* 1557 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be 1558 * removed when useful for applications beyond shares distribution (e.g. 1559 * load-balance). 1560 */ 1561#if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED) 1562 p->se.avg.runnable_avg_period = 0; 1563 p->se.avg.runnable_avg_sum = 0; 1564#endif 1565#ifdef CONFIG_SCHEDSTATS 1566 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1567#endif 1568 1569 INIT_LIST_HEAD(&p->rt.run_list); 1570 1571#ifdef CONFIG_PREEMPT_NOTIFIERS 1572 INIT_HLIST_HEAD(&p->preempt_notifiers); 1573#endif 1574 1575#ifdef CONFIG_NUMA_BALANCING 1576 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1577 p->mm->numa_next_scan = jiffies; 1578 p->mm->numa_next_reset = jiffies; 1579 p->mm->numa_scan_seq = 0; 1580 } 1581 1582 p->node_stamp = 0ULL; 1583 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1584 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1585 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1586 p->numa_work.next = &p->numa_work; 1587#endif /* CONFIG_NUMA_BALANCING */ 1588} 1589 1590#ifdef CONFIG_NUMA_BALANCING 1591#ifdef CONFIG_SCHED_DEBUG 1592void set_numabalancing_state(bool enabled) 1593{ 1594 if (enabled) 1595 sched_feat_set("NUMA"); 1596 else 1597 sched_feat_set("NO_NUMA"); 1598} 1599#else 1600__read_mostly bool numabalancing_enabled; 1601 1602void set_numabalancing_state(bool enabled) 1603{ 1604 numabalancing_enabled = enabled; 1605} 1606#endif /* CONFIG_SCHED_DEBUG */ 1607#endif /* CONFIG_NUMA_BALANCING */ 1608 1609/* 1610 * fork()/clone()-time setup: 1611 */ 1612void sched_fork(struct task_struct *p) 1613{ 1614 unsigned long flags; 1615 int cpu = get_cpu(); 1616 1617 __sched_fork(p); 1618 /* 1619 * We mark the process as running here. This guarantees that 1620 * nobody will actually run it, and a signal or other external 1621 * event cannot wake it up and insert it on the runqueue either. 1622 */ 1623 p->state = TASK_RUNNING; 1624 1625 /* 1626 * Make sure we do not leak PI boosting priority to the child. 1627 */ 1628 p->prio = current->normal_prio; 1629 1630 /* 1631 * Revert to default priority/policy on fork if requested. 1632 */ 1633 if (unlikely(p->sched_reset_on_fork)) { 1634 if (task_has_rt_policy(p)) { 1635 p->policy = SCHED_NORMAL; 1636 p->static_prio = NICE_TO_PRIO(0); 1637 p->rt_priority = 0; 1638 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1639 p->static_prio = NICE_TO_PRIO(0); 1640 1641 p->prio = p->normal_prio = __normal_prio(p); 1642 set_load_weight(p); 1643 1644 /* 1645 * We don't need the reset flag anymore after the fork. It has 1646 * fulfilled its duty: 1647 */ 1648 p->sched_reset_on_fork = 0; 1649 } 1650 1651 if (!rt_prio(p->prio)) 1652 p->sched_class = &fair_sched_class; 1653 1654 if (p->sched_class->task_fork) 1655 p->sched_class->task_fork(p); 1656 1657 /* 1658 * The child is not yet in the pid-hash so no cgroup attach races, 1659 * and the cgroup is pinned to this child due to cgroup_fork() 1660 * is ran before sched_fork(). 1661 * 1662 * Silence PROVE_RCU. 1663 */ 1664 raw_spin_lock_irqsave(&p->pi_lock, flags); 1665 set_task_cpu(p, cpu); 1666 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1667 1668#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1669 if (likely(sched_info_on())) 1670 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1671#endif 1672#if defined(CONFIG_SMP) 1673 p->on_cpu = 0; 1674#endif 1675#ifdef CONFIG_PREEMPT_COUNT 1676 /* Want to start with kernel preemption disabled. */ 1677 task_thread_info(p)->preempt_count = 1; 1678#endif 1679#ifdef CONFIG_SMP 1680 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1681#endif 1682 1683 put_cpu(); 1684} 1685 1686/* 1687 * wake_up_new_task - wake up a newly created task for the first time. 1688 * 1689 * This function will do some initial scheduler statistics housekeeping 1690 * that must be done for every newly created context, then puts the task 1691 * on the runqueue and wakes it. 1692 */ 1693void wake_up_new_task(struct task_struct *p) 1694{ 1695 unsigned long flags; 1696 struct rq *rq; 1697 1698 raw_spin_lock_irqsave(&p->pi_lock, flags); 1699#ifdef CONFIG_SMP 1700 /* 1701 * Fork balancing, do it here and not earlier because: 1702 * - cpus_allowed can change in the fork path 1703 * - any previously selected cpu might disappear through hotplug 1704 */ 1705 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1706#endif 1707 1708 rq = __task_rq_lock(p); 1709 activate_task(rq, p, 0); 1710 p->on_rq = 1; 1711 trace_sched_wakeup_new(p, true); 1712 check_preempt_curr(rq, p, WF_FORK); 1713#ifdef CONFIG_SMP 1714 if (p->sched_class->task_woken) 1715 p->sched_class->task_woken(rq, p); 1716#endif 1717 task_rq_unlock(rq, p, &flags); 1718} 1719 1720#ifdef CONFIG_PREEMPT_NOTIFIERS 1721 1722/** 1723 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1724 * @notifier: notifier struct to register 1725 */ 1726void preempt_notifier_register(struct preempt_notifier *notifier) 1727{ 1728 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1729} 1730EXPORT_SYMBOL_GPL(preempt_notifier_register); 1731 1732/** 1733 * preempt_notifier_unregister - no longer interested in preemption notifications 1734 * @notifier: notifier struct to unregister 1735 * 1736 * This is safe to call from within a preemption notifier. 1737 */ 1738void preempt_notifier_unregister(struct preempt_notifier *notifier) 1739{ 1740 hlist_del(¬ifier->link); 1741} 1742EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1743 1744static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1745{ 1746 struct preempt_notifier *notifier; 1747 struct hlist_node *node; 1748 1749 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1750 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1751} 1752 1753static void 1754fire_sched_out_preempt_notifiers(struct task_struct *curr, 1755 struct task_struct *next) 1756{ 1757 struct preempt_notifier *notifier; 1758 struct hlist_node *node; 1759 1760 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1761 notifier->ops->sched_out(notifier, next); 1762} 1763 1764#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1765 1766static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1767{ 1768} 1769 1770static void 1771fire_sched_out_preempt_notifiers(struct task_struct *curr, 1772 struct task_struct *next) 1773{ 1774} 1775 1776#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1777 1778/** 1779 * prepare_task_switch - prepare to switch tasks 1780 * @rq: the runqueue preparing to switch 1781 * @prev: the current task that is being switched out 1782 * @next: the task we are going to switch to. 1783 * 1784 * This is called with the rq lock held and interrupts off. It must 1785 * be paired with a subsequent finish_task_switch after the context 1786 * switch. 1787 * 1788 * prepare_task_switch sets up locking and calls architecture specific 1789 * hooks. 1790 */ 1791static inline void 1792prepare_task_switch(struct rq *rq, struct task_struct *prev, 1793 struct task_struct *next) 1794{ 1795 trace_sched_switch(prev, next); 1796 sched_info_switch(prev, next); 1797 perf_event_task_sched_out(prev, next); 1798 fire_sched_out_preempt_notifiers(prev, next); 1799 prepare_lock_switch(rq, next); 1800 prepare_arch_switch(next); 1801} 1802 1803/** 1804 * finish_task_switch - clean up after a task-switch 1805 * @rq: runqueue associated with task-switch 1806 * @prev: the thread we just switched away from. 1807 * 1808 * finish_task_switch must be called after the context switch, paired 1809 * with a prepare_task_switch call before the context switch. 1810 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1811 * and do any other architecture-specific cleanup actions. 1812 * 1813 * Note that we may have delayed dropping an mm in context_switch(). If 1814 * so, we finish that here outside of the runqueue lock. (Doing it 1815 * with the lock held can cause deadlocks; see schedule() for 1816 * details.) 1817 */ 1818static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1819 __releases(rq->lock) 1820{ 1821 struct mm_struct *mm = rq->prev_mm; 1822 long prev_state; 1823 1824 rq->prev_mm = NULL; 1825 1826 /* 1827 * A task struct has one reference for the use as "current". 1828 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1829 * schedule one last time. The schedule call will never return, and 1830 * the scheduled task must drop that reference. 1831 * The test for TASK_DEAD must occur while the runqueue locks are 1832 * still held, otherwise prev could be scheduled on another cpu, die 1833 * there before we look at prev->state, and then the reference would 1834 * be dropped twice. 1835 * Manfred Spraul <manfred@colorfullife.com> 1836 */ 1837 prev_state = prev->state; 1838 vtime_task_switch(prev); 1839 finish_arch_switch(prev); 1840 perf_event_task_sched_in(prev, current); 1841 finish_lock_switch(rq, prev); 1842 finish_arch_post_lock_switch(); 1843 1844 fire_sched_in_preempt_notifiers(current); 1845 if (mm) 1846 mmdrop(mm); 1847 if (unlikely(prev_state == TASK_DEAD)) { 1848 /* 1849 * Remove function-return probe instances associated with this 1850 * task and put them back on the free list. 1851 */ 1852 kprobe_flush_task(prev); 1853 put_task_struct(prev); 1854 } 1855} 1856 1857#ifdef CONFIG_SMP 1858 1859/* assumes rq->lock is held */ 1860static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1861{ 1862 if (prev->sched_class->pre_schedule) 1863 prev->sched_class->pre_schedule(rq, prev); 1864} 1865 1866/* rq->lock is NOT held, but preemption is disabled */ 1867static inline void post_schedule(struct rq *rq) 1868{ 1869 if (rq->post_schedule) { 1870 unsigned long flags; 1871 1872 raw_spin_lock_irqsave(&rq->lock, flags); 1873 if (rq->curr->sched_class->post_schedule) 1874 rq->curr->sched_class->post_schedule(rq); 1875 raw_spin_unlock_irqrestore(&rq->lock, flags); 1876 1877 rq->post_schedule = 0; 1878 } 1879} 1880 1881#else 1882 1883static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1884{ 1885} 1886 1887static inline void post_schedule(struct rq *rq) 1888{ 1889} 1890 1891#endif 1892 1893/** 1894 * schedule_tail - first thing a freshly forked thread must call. 1895 * @prev: the thread we just switched away from. 1896 */ 1897asmlinkage void schedule_tail(struct task_struct *prev) 1898 __releases(rq->lock) 1899{ 1900 struct rq *rq = this_rq(); 1901 1902 finish_task_switch(rq, prev); 1903 1904 /* 1905 * FIXME: do we need to worry about rq being invalidated by the 1906 * task_switch? 1907 */ 1908 post_schedule(rq); 1909 1910#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1911 /* In this case, finish_task_switch does not reenable preemption */ 1912 preempt_enable(); 1913#endif 1914 if (current->set_child_tid) 1915 put_user(task_pid_vnr(current), current->set_child_tid); 1916} 1917 1918/* 1919 * context_switch - switch to the new MM and the new 1920 * thread's register state. 1921 */ 1922static inline void 1923context_switch(struct rq *rq, struct task_struct *prev, 1924 struct task_struct *next) 1925{ 1926 struct mm_struct *mm, *oldmm; 1927 1928 prepare_task_switch(rq, prev, next); 1929 1930 mm = next->mm; 1931 oldmm = prev->active_mm; 1932 /* 1933 * For paravirt, this is coupled with an exit in switch_to to 1934 * combine the page table reload and the switch backend into 1935 * one hypercall. 1936 */ 1937 arch_start_context_switch(prev); 1938 1939 if (!mm) { 1940 next->active_mm = oldmm; 1941 atomic_inc(&oldmm->mm_count); 1942 enter_lazy_tlb(oldmm, next); 1943 } else 1944 switch_mm(oldmm, mm, next); 1945 1946 if (!prev->mm) { 1947 prev->active_mm = NULL; 1948 rq->prev_mm = oldmm; 1949 } 1950 /* 1951 * Since the runqueue lock will be released by the next 1952 * task (which is an invalid locking op but in the case 1953 * of the scheduler it's an obvious special-case), so we 1954 * do an early lockdep release here: 1955 */ 1956#ifndef __ARCH_WANT_UNLOCKED_CTXSW 1957 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 1958#endif 1959 1960 context_tracking_task_switch(prev, next); 1961 /* Here we just switch the register state and the stack. */ 1962 switch_to(prev, next, prev); 1963 1964 barrier(); 1965 /* 1966 * this_rq must be evaluated again because prev may have moved 1967 * CPUs since it called schedule(), thus the 'rq' on its stack 1968 * frame will be invalid. 1969 */ 1970 finish_task_switch(this_rq(), prev); 1971} 1972 1973/* 1974 * nr_running and nr_context_switches: 1975 * 1976 * externally visible scheduler statistics: current number of runnable 1977 * threads, total number of context switches performed since bootup. 1978 */ 1979unsigned long nr_running(void) 1980{ 1981 unsigned long i, sum = 0; 1982 1983 for_each_online_cpu(i) 1984 sum += cpu_rq(i)->nr_running; 1985 1986 return sum; 1987} 1988 1989unsigned long long nr_context_switches(void) 1990{ 1991 int i; 1992 unsigned long long sum = 0; 1993 1994 for_each_possible_cpu(i) 1995 sum += cpu_rq(i)->nr_switches; 1996 1997 return sum; 1998} 1999 2000unsigned long nr_iowait(void) 2001{ 2002 unsigned long i, sum = 0; 2003 2004 for_each_possible_cpu(i) 2005 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2006 2007 return sum; 2008} 2009 2010unsigned long nr_iowait_cpu(int cpu) 2011{ 2012 struct rq *this = cpu_rq(cpu); 2013 return atomic_read(&this->nr_iowait); 2014} 2015 2016unsigned long this_cpu_load(void) 2017{ 2018 struct rq *this = this_rq(); 2019 return this->cpu_load[0]; 2020} 2021 2022 2023/* 2024 * Global load-average calculations 2025 * 2026 * We take a distributed and async approach to calculating the global load-avg 2027 * in order to minimize overhead. 2028 * 2029 * The global load average is an exponentially decaying average of nr_running + 2030 * nr_uninterruptible. 2031 * 2032 * Once every LOAD_FREQ: 2033 * 2034 * nr_active = 0; 2035 * for_each_possible_cpu(cpu) 2036 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; 2037 * 2038 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) 2039 * 2040 * Due to a number of reasons the above turns in the mess below: 2041 * 2042 * - for_each_possible_cpu() is prohibitively expensive on machines with 2043 * serious number of cpus, therefore we need to take a distributed approach 2044 * to calculating nr_active. 2045 * 2046 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 2047 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } 2048 * 2049 * So assuming nr_active := 0 when we start out -- true per definition, we 2050 * can simply take per-cpu deltas and fold those into a global accumulate 2051 * to obtain the same result. See calc_load_fold_active(). 2052 * 2053 * Furthermore, in order to avoid synchronizing all per-cpu delta folding 2054 * across the machine, we assume 10 ticks is sufficient time for every 2055 * cpu to have completed this task. 2056 * 2057 * This places an upper-bound on the IRQ-off latency of the machine. Then 2058 * again, being late doesn't loose the delta, just wrecks the sample. 2059 * 2060 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because 2061 * this would add another cross-cpu cacheline miss and atomic operation 2062 * to the wakeup path. Instead we increment on whatever cpu the task ran 2063 * when it went into uninterruptible state and decrement on whatever cpu 2064 * did the wakeup. This means that only the sum of nr_uninterruptible over 2065 * all cpus yields the correct result. 2066 * 2067 * This covers the NO_HZ=n code, for extra head-aches, see the comment below. 2068 */ 2069 2070/* Variables and functions for calc_load */ 2071static atomic_long_t calc_load_tasks; 2072static unsigned long calc_load_update; 2073unsigned long avenrun[3]; 2074EXPORT_SYMBOL(avenrun); /* should be removed */ 2075 2076/** 2077 * get_avenrun - get the load average array 2078 * @loads: pointer to dest load array 2079 * @offset: offset to add 2080 * @shift: shift count to shift the result left 2081 * 2082 * These values are estimates at best, so no need for locking. 2083 */ 2084void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2085{ 2086 loads[0] = (avenrun[0] + offset) << shift; 2087 loads[1] = (avenrun[1] + offset) << shift; 2088 loads[2] = (avenrun[2] + offset) << shift; 2089} 2090 2091static long calc_load_fold_active(struct rq *this_rq) 2092{ 2093 long nr_active, delta = 0; 2094 2095 nr_active = this_rq->nr_running; 2096 nr_active += (long) this_rq->nr_uninterruptible; 2097 2098 if (nr_active != this_rq->calc_load_active) { 2099 delta = nr_active - this_rq->calc_load_active; 2100 this_rq->calc_load_active = nr_active; 2101 } 2102 2103 return delta; 2104} 2105 2106/* 2107 * a1 = a0 * e + a * (1 - e) 2108 */ 2109static unsigned long 2110calc_load(unsigned long load, unsigned long exp, unsigned long active) 2111{ 2112 load *= exp; 2113 load += active * (FIXED_1 - exp); 2114 load += 1UL << (FSHIFT - 1); 2115 return load >> FSHIFT; 2116} 2117 2118#ifdef CONFIG_NO_HZ 2119/* 2120 * Handle NO_HZ for the global load-average. 2121 * 2122 * Since the above described distributed algorithm to compute the global 2123 * load-average relies on per-cpu sampling from the tick, it is affected by 2124 * NO_HZ. 2125 * 2126 * The basic idea is to fold the nr_active delta into a global idle-delta upon 2127 * entering NO_HZ state such that we can include this as an 'extra' cpu delta 2128 * when we read the global state. 2129 * 2130 * Obviously reality has to ruin such a delightfully simple scheme: 2131 * 2132 * - When we go NO_HZ idle during the window, we can negate our sample 2133 * contribution, causing under-accounting. 2134 * 2135 * We avoid this by keeping two idle-delta counters and flipping them 2136 * when the window starts, thus separating old and new NO_HZ load. 2137 * 2138 * The only trick is the slight shift in index flip for read vs write. 2139 * 2140 * 0s 5s 10s 15s 2141 * +10 +10 +10 +10 2142 * |-|-----------|-|-----------|-|-----------|-| 2143 * r:0 0 1 1 0 0 1 1 0 2144 * w:0 1 1 0 0 1 1 0 0 2145 * 2146 * This ensures we'll fold the old idle contribution in this window while 2147 * accumlating the new one. 2148 * 2149 * - When we wake up from NO_HZ idle during the window, we push up our 2150 * contribution, since we effectively move our sample point to a known 2151 * busy state. 2152 * 2153 * This is solved by pushing the window forward, and thus skipping the 2154 * sample, for this cpu (effectively using the idle-delta for this cpu which 2155 * was in effect at the time the window opened). This also solves the issue 2156 * of having to deal with a cpu having been in NOHZ idle for multiple 2157 * LOAD_FREQ intervals. 2158 * 2159 * When making the ILB scale, we should try to pull this in as well. 2160 */ 2161static atomic_long_t calc_load_idle[2]; 2162static int calc_load_idx; 2163 2164static inline int calc_load_write_idx(void) 2165{ 2166 int idx = calc_load_idx; 2167 2168 /* 2169 * See calc_global_nohz(), if we observe the new index, we also 2170 * need to observe the new update time. 2171 */ 2172 smp_rmb(); 2173 2174 /* 2175 * If the folding window started, make sure we start writing in the 2176 * next idle-delta. 2177 */ 2178 if (!time_before(jiffies, calc_load_update)) 2179 idx++; 2180 2181 return idx & 1; 2182} 2183 2184static inline int calc_load_read_idx(void) 2185{ 2186 return calc_load_idx & 1; 2187} 2188 2189void calc_load_enter_idle(void) 2190{ 2191 struct rq *this_rq = this_rq(); 2192 long delta; 2193 2194 /* 2195 * We're going into NOHZ mode, if there's any pending delta, fold it 2196 * into the pending idle delta. 2197 */ 2198 delta = calc_load_fold_active(this_rq); 2199 if (delta) { 2200 int idx = calc_load_write_idx(); 2201 atomic_long_add(delta, &calc_load_idle[idx]); 2202 } 2203} 2204 2205void calc_load_exit_idle(void) 2206{ 2207 struct rq *this_rq = this_rq(); 2208 2209 /* 2210 * If we're still before the sample window, we're done. 2211 */ 2212 if (time_before(jiffies, this_rq->calc_load_update)) 2213 return; 2214 2215 /* 2216 * We woke inside or after the sample window, this means we're already 2217 * accounted through the nohz accounting, so skip the entire deal and 2218 * sync up for the next window. 2219 */ 2220 this_rq->calc_load_update = calc_load_update; 2221 if (time_before(jiffies, this_rq->calc_load_update + 10)) 2222 this_rq->calc_load_update += LOAD_FREQ; 2223} 2224 2225static long calc_load_fold_idle(void) 2226{ 2227 int idx = calc_load_read_idx(); 2228 long delta = 0; 2229 2230 if (atomic_long_read(&calc_load_idle[idx])) 2231 delta = atomic_long_xchg(&calc_load_idle[idx], 0); 2232 2233 return delta; 2234} 2235 2236/** 2237 * fixed_power_int - compute: x^n, in O(log n) time 2238 * 2239 * @x: base of the power 2240 * @frac_bits: fractional bits of @x 2241 * @n: power to raise @x to. 2242 * 2243 * By exploiting the relation between the definition of the natural power 2244 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2245 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2246 * (where: n_i \elem {0, 1}, the binary vector representing n), 2247 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2248 * of course trivially computable in O(log_2 n), the length of our binary 2249 * vector. 2250 */ 2251static unsigned long 2252fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2253{ 2254 unsigned long result = 1UL << frac_bits; 2255 2256 if (n) for (;;) { 2257 if (n & 1) { 2258 result *= x; 2259 result += 1UL << (frac_bits - 1); 2260 result >>= frac_bits; 2261 } 2262 n >>= 1; 2263 if (!n) 2264 break; 2265 x *= x; 2266 x += 1UL << (frac_bits - 1); 2267 x >>= frac_bits; 2268 } 2269 2270 return result; 2271} 2272 2273/* 2274 * a1 = a0 * e + a * (1 - e) 2275 * 2276 * a2 = a1 * e + a * (1 - e) 2277 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2278 * = a0 * e^2 + a * (1 - e) * (1 + e) 2279 * 2280 * a3 = a2 * e + a * (1 - e) 2281 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2282 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2283 * 2284 * ... 2285 * 2286 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2287 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2288 * = a0 * e^n + a * (1 - e^n) 2289 * 2290 * [1] application of the geometric series: 2291 * 2292 * n 1 - x^(n+1) 2293 * S_n := \Sum x^i = ------------- 2294 * i=0 1 - x 2295 */ 2296static unsigned long 2297calc_load_n(unsigned long load, unsigned long exp, 2298 unsigned long active, unsigned int n) 2299{ 2300 2301 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2302} 2303 2304/* 2305 * NO_HZ can leave us missing all per-cpu ticks calling 2306 * calc_load_account_active(), but since an idle CPU folds its delta into 2307 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2308 * in the pending idle delta if our idle period crossed a load cycle boundary. 2309 * 2310 * Once we've updated the global active value, we need to apply the exponential 2311 * weights adjusted to the number of cycles missed. 2312 */ 2313static void calc_global_nohz(void) 2314{ 2315 long delta, active, n; 2316 2317 if (!time_before(jiffies, calc_load_update + 10)) { 2318 /* 2319 * Catch-up, fold however many we are behind still 2320 */ 2321 delta = jiffies - calc_load_update - 10; 2322 n = 1 + (delta / LOAD_FREQ); 2323 2324 active = atomic_long_read(&calc_load_tasks); 2325 active = active > 0 ? active * FIXED_1 : 0; 2326 2327 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2328 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2329 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2330 2331 calc_load_update += n * LOAD_FREQ; 2332 } 2333 2334 /* 2335 * Flip the idle index... 2336 * 2337 * Make sure we first write the new time then flip the index, so that 2338 * calc_load_write_idx() will see the new time when it reads the new 2339 * index, this avoids a double flip messing things up. 2340 */ 2341 smp_wmb(); 2342 calc_load_idx++; 2343} 2344#else /* !CONFIG_NO_HZ */ 2345 2346static inline long calc_load_fold_idle(void) { return 0; } 2347static inline void calc_global_nohz(void) { } 2348 2349#endif /* CONFIG_NO_HZ */ 2350 2351/* 2352 * calc_load - update the avenrun load estimates 10 ticks after the 2353 * CPUs have updated calc_load_tasks. 2354 */ 2355void calc_global_load(unsigned long ticks) 2356{ 2357 long active, delta; 2358 2359 if (time_before(jiffies, calc_load_update + 10)) 2360 return; 2361 2362 /* 2363 * Fold the 'old' idle-delta to include all NO_HZ cpus. 2364 */ 2365 delta = calc_load_fold_idle(); 2366 if (delta) 2367 atomic_long_add(delta, &calc_load_tasks); 2368 2369 active = atomic_long_read(&calc_load_tasks); 2370 active = active > 0 ? active * FIXED_1 : 0; 2371 2372 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2373 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2374 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2375 2376 calc_load_update += LOAD_FREQ; 2377 2378 /* 2379 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk. 2380 */ 2381 calc_global_nohz(); 2382} 2383 2384/* 2385 * Called from update_cpu_load() to periodically update this CPU's 2386 * active count. 2387 */ 2388static void calc_load_account_active(struct rq *this_rq) 2389{ 2390 long delta; 2391 2392 if (time_before(jiffies, this_rq->calc_load_update)) 2393 return; 2394 2395 delta = calc_load_fold_active(this_rq); 2396 if (delta) 2397 atomic_long_add(delta, &calc_load_tasks); 2398 2399 this_rq->calc_load_update += LOAD_FREQ; 2400} 2401 2402/* 2403 * End of global load-average stuff 2404 */ 2405 2406/* 2407 * The exact cpuload at various idx values, calculated at every tick would be 2408 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2409 * 2410 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2411 * on nth tick when cpu may be busy, then we have: 2412 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2413 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2414 * 2415 * decay_load_missed() below does efficient calculation of 2416 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2417 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2418 * 2419 * The calculation is approximated on a 128 point scale. 2420 * degrade_zero_ticks is the number of ticks after which load at any 2421 * particular idx is approximated to be zero. 2422 * degrade_factor is a precomputed table, a row for each load idx. 2423 * Each column corresponds to degradation factor for a power of two ticks, 2424 * based on 128 point scale. 2425 * Example: 2426 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2427 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2428 * 2429 * With this power of 2 load factors, we can degrade the load n times 2430 * by looking at 1 bits in n and doing as many mult/shift instead of 2431 * n mult/shifts needed by the exact degradation. 2432 */ 2433#define DEGRADE_SHIFT 7 2434static const unsigned char 2435 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2436static const unsigned char 2437 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2438 {0, 0, 0, 0, 0, 0, 0, 0}, 2439 {64, 32, 8, 0, 0, 0, 0, 0}, 2440 {96, 72, 40, 12, 1, 0, 0}, 2441 {112, 98, 75, 43, 15, 1, 0}, 2442 {120, 112, 98, 76, 45, 16, 2} }; 2443 2444/* 2445 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2446 * would be when CPU is idle and so we just decay the old load without 2447 * adding any new load. 2448 */ 2449static unsigned long 2450decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2451{ 2452 int j = 0; 2453 2454 if (!missed_updates) 2455 return load; 2456 2457 if (missed_updates >= degrade_zero_ticks[idx]) 2458 return 0; 2459 2460 if (idx == 1) 2461 return load >> missed_updates; 2462 2463 while (missed_updates) { 2464 if (missed_updates % 2) 2465 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2466 2467 missed_updates >>= 1; 2468 j++; 2469 } 2470 return load; 2471} 2472 2473/* 2474 * Update rq->cpu_load[] statistics. This function is usually called every 2475 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2476 * every tick. We fix it up based on jiffies. 2477 */ 2478static void __update_cpu_load(struct rq *this_rq, unsigned long this_load, 2479 unsigned long pending_updates) 2480{ 2481 int i, scale; 2482 2483 this_rq->nr_load_updates++; 2484 2485 /* Update our load: */ 2486 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2487 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2488 unsigned long old_load, new_load; 2489 2490 /* scale is effectively 1 << i now, and >> i divides by scale */ 2491 2492 old_load = this_rq->cpu_load[i]; 2493 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2494 new_load = this_load; 2495 /* 2496 * Round up the averaging division if load is increasing. This 2497 * prevents us from getting stuck on 9 if the load is 10, for 2498 * example. 2499 */ 2500 if (new_load > old_load) 2501 new_load += scale - 1; 2502 2503 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2504 } 2505 2506 sched_avg_update(this_rq); 2507} 2508 2509#ifdef CONFIG_NO_HZ 2510/* 2511 * There is no sane way to deal with nohz on smp when using jiffies because the 2512 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading 2513 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. 2514 * 2515 * Therefore we cannot use the delta approach from the regular tick since that 2516 * would seriously skew the load calculation. However we'll make do for those 2517 * updates happening while idle (nohz_idle_balance) or coming out of idle 2518 * (tick_nohz_idle_exit). 2519 * 2520 * This means we might still be one tick off for nohz periods. 2521 */ 2522 2523/* 2524 * Called from nohz_idle_balance() to update the load ratings before doing the 2525 * idle balance. 2526 */ 2527void update_idle_cpu_load(struct rq *this_rq) 2528{ 2529 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2530 unsigned long load = this_rq->load.weight; 2531 unsigned long pending_updates; 2532 2533 /* 2534 * bail if there's load or we're actually up-to-date. 2535 */ 2536 if (load || curr_jiffies == this_rq->last_load_update_tick) 2537 return; 2538 2539 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2540 this_rq->last_load_update_tick = curr_jiffies; 2541 2542 __update_cpu_load(this_rq, load, pending_updates); 2543} 2544 2545/* 2546 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed. 2547 */ 2548void update_cpu_load_nohz(void) 2549{ 2550 struct rq *this_rq = this_rq(); 2551 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2552 unsigned long pending_updates; 2553 2554 if (curr_jiffies == this_rq->last_load_update_tick) 2555 return; 2556 2557 raw_spin_lock(&this_rq->lock); 2558 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2559 if (pending_updates) { 2560 this_rq->last_load_update_tick = curr_jiffies; 2561 /* 2562 * We were idle, this means load 0, the current load might be 2563 * !0 due to remote wakeups and the sort. 2564 */ 2565 __update_cpu_load(this_rq, 0, pending_updates); 2566 } 2567 raw_spin_unlock(&this_rq->lock); 2568} 2569#endif /* CONFIG_NO_HZ */ 2570 2571/* 2572 * Called from scheduler_tick() 2573 */ 2574static void update_cpu_load_active(struct rq *this_rq) 2575{ 2576 /* 2577 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz(). 2578 */ 2579 this_rq->last_load_update_tick = jiffies; 2580 __update_cpu_load(this_rq, this_rq->load.weight, 1); 2581 2582 calc_load_account_active(this_rq); 2583} 2584 2585#ifdef CONFIG_SMP 2586 2587/* 2588 * sched_exec - execve() is a valuable balancing opportunity, because at 2589 * this point the task has the smallest effective memory and cache footprint. 2590 */ 2591void sched_exec(void) 2592{ 2593 struct task_struct *p = current; 2594 unsigned long flags; 2595 int dest_cpu; 2596 2597 raw_spin_lock_irqsave(&p->pi_lock, flags); 2598 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2599 if (dest_cpu == smp_processor_id()) 2600 goto unlock; 2601 2602 if (likely(cpu_active(dest_cpu))) { 2603 struct migration_arg arg = { p, dest_cpu }; 2604 2605 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2606 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2607 return; 2608 } 2609unlock: 2610 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2611} 2612 2613#endif 2614 2615DEFINE_PER_CPU(struct kernel_stat, kstat); 2616DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2617 2618EXPORT_PER_CPU_SYMBOL(kstat); 2619EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2620 2621/* 2622 * Return any ns on the sched_clock that have not yet been accounted in 2623 * @p in case that task is currently running. 2624 * 2625 * Called with task_rq_lock() held on @rq. 2626 */ 2627static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2628{ 2629 u64 ns = 0; 2630 2631 if (task_current(rq, p)) { 2632 update_rq_clock(rq); 2633 ns = rq->clock_task - p->se.exec_start; 2634 if ((s64)ns < 0) 2635 ns = 0; 2636 } 2637 2638 return ns; 2639} 2640 2641unsigned long long task_delta_exec(struct task_struct *p) 2642{ 2643 unsigned long flags; 2644 struct rq *rq; 2645 u64 ns = 0; 2646 2647 rq = task_rq_lock(p, &flags); 2648 ns = do_task_delta_exec(p, rq); 2649 task_rq_unlock(rq, p, &flags); 2650 2651 return ns; 2652} 2653 2654/* 2655 * Return accounted runtime for the task. 2656 * In case the task is currently running, return the runtime plus current's 2657 * pending runtime that have not been accounted yet. 2658 */ 2659unsigned long long task_sched_runtime(struct task_struct *p) 2660{ 2661 unsigned long flags; 2662 struct rq *rq; 2663 u64 ns = 0; 2664 2665 rq = task_rq_lock(p, &flags); 2666 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2667 task_rq_unlock(rq, p, &flags); 2668 2669 return ns; 2670} 2671 2672/* 2673 * This function gets called by the timer code, with HZ frequency. 2674 * We call it with interrupts disabled. 2675 */ 2676void scheduler_tick(void) 2677{ 2678 int cpu = smp_processor_id(); 2679 struct rq *rq = cpu_rq(cpu); 2680 struct task_struct *curr = rq->curr; 2681 2682 sched_clock_tick(); 2683 2684 raw_spin_lock(&rq->lock); 2685 update_rq_clock(rq); 2686 update_cpu_load_active(rq); 2687 curr->sched_class->task_tick(rq, curr, 0); 2688 raw_spin_unlock(&rq->lock); 2689 2690 perf_event_task_tick(); 2691 2692#ifdef CONFIG_SMP 2693 rq->idle_balance = idle_cpu(cpu); 2694 trigger_load_balance(rq, cpu); 2695#endif 2696} 2697 2698notrace unsigned long get_parent_ip(unsigned long addr) 2699{ 2700 if (in_lock_functions(addr)) { 2701 addr = CALLER_ADDR2; 2702 if (in_lock_functions(addr)) 2703 addr = CALLER_ADDR3; 2704 } 2705 return addr; 2706} 2707 2708#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2709 defined(CONFIG_PREEMPT_TRACER)) 2710 2711void __kprobes add_preempt_count(int val) 2712{ 2713#ifdef CONFIG_DEBUG_PREEMPT 2714 /* 2715 * Underflow? 2716 */ 2717 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2718 return; 2719#endif 2720 preempt_count() += val; 2721#ifdef CONFIG_DEBUG_PREEMPT 2722 /* 2723 * Spinlock count overflowing soon? 2724 */ 2725 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2726 PREEMPT_MASK - 10); 2727#endif 2728 if (preempt_count() == val) 2729 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2730} 2731EXPORT_SYMBOL(add_preempt_count); 2732 2733void __kprobes sub_preempt_count(int val) 2734{ 2735#ifdef CONFIG_DEBUG_PREEMPT 2736 /* 2737 * Underflow? 2738 */ 2739 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2740 return; 2741 /* 2742 * Is the spinlock portion underflowing? 2743 */ 2744 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2745 !(preempt_count() & PREEMPT_MASK))) 2746 return; 2747#endif 2748 2749 if (preempt_count() == val) 2750 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2751 preempt_count() -= val; 2752} 2753EXPORT_SYMBOL(sub_preempt_count); 2754 2755#endif 2756 2757/* 2758 * Print scheduling while atomic bug: 2759 */ 2760static noinline void __schedule_bug(struct task_struct *prev) 2761{ 2762 if (oops_in_progress) 2763 return; 2764 2765 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2766 prev->comm, prev->pid, preempt_count()); 2767 2768 debug_show_held_locks(prev); 2769 print_modules(); 2770 if (irqs_disabled()) 2771 print_irqtrace_events(prev); 2772 dump_stack(); 2773 add_taint(TAINT_WARN); 2774} 2775 2776/* 2777 * Various schedule()-time debugging checks and statistics: 2778 */ 2779static inline void schedule_debug(struct task_struct *prev) 2780{ 2781 /* 2782 * Test if we are atomic. Since do_exit() needs to call into 2783 * schedule() atomically, we ignore that path for now. 2784 * Otherwise, whine if we are scheduling when we should not be. 2785 */ 2786 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2787 __schedule_bug(prev); 2788 rcu_sleep_check(); 2789 2790 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2791 2792 schedstat_inc(this_rq(), sched_count); 2793} 2794 2795static void put_prev_task(struct rq *rq, struct task_struct *prev) 2796{ 2797 if (prev->on_rq || rq->skip_clock_update < 0) 2798 update_rq_clock(rq); 2799 prev->sched_class->put_prev_task(rq, prev); 2800} 2801 2802/* 2803 * Pick up the highest-prio task: 2804 */ 2805static inline struct task_struct * 2806pick_next_task(struct rq *rq) 2807{ 2808 const struct sched_class *class; 2809 struct task_struct *p; 2810 2811 /* 2812 * Optimization: we know that if all tasks are in 2813 * the fair class we can call that function directly: 2814 */ 2815 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2816 p = fair_sched_class.pick_next_task(rq); 2817 if (likely(p)) 2818 return p; 2819 } 2820 2821 for_each_class(class) { 2822 p = class->pick_next_task(rq); 2823 if (p) 2824 return p; 2825 } 2826 2827 BUG(); /* the idle class will always have a runnable task */ 2828} 2829 2830/* 2831 * __schedule() is the main scheduler function. 2832 * 2833 * The main means of driving the scheduler and thus entering this function are: 2834 * 2835 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2836 * 2837 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2838 * paths. For example, see arch/x86/entry_64.S. 2839 * 2840 * To drive preemption between tasks, the scheduler sets the flag in timer 2841 * interrupt handler scheduler_tick(). 2842 * 2843 * 3. Wakeups don't really cause entry into schedule(). They add a 2844 * task to the run-queue and that's it. 2845 * 2846 * Now, if the new task added to the run-queue preempts the current 2847 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2848 * called on the nearest possible occasion: 2849 * 2850 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2851 * 2852 * - in syscall or exception context, at the next outmost 2853 * preempt_enable(). (this might be as soon as the wake_up()'s 2854 * spin_unlock()!) 2855 * 2856 * - in IRQ context, return from interrupt-handler to 2857 * preemptible context 2858 * 2859 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2860 * then at the next: 2861 * 2862 * - cond_resched() call 2863 * - explicit schedule() call 2864 * - return from syscall or exception to user-space 2865 * - return from interrupt-handler to user-space 2866 */ 2867static void __sched __schedule(void) 2868{ 2869 struct task_struct *prev, *next; 2870 unsigned long *switch_count; 2871 struct rq *rq; 2872 int cpu; 2873 2874need_resched: 2875 preempt_disable(); 2876 cpu = smp_processor_id(); 2877 rq = cpu_rq(cpu); 2878 rcu_note_context_switch(cpu); 2879 prev = rq->curr; 2880 2881 schedule_debug(prev); 2882 2883 if (sched_feat(HRTICK)) 2884 hrtick_clear(rq); 2885 2886 raw_spin_lock_irq(&rq->lock); 2887 2888 switch_count = &prev->nivcsw; 2889 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2890 if (unlikely(signal_pending_state(prev->state, prev))) { 2891 prev->state = TASK_RUNNING; 2892 } else { 2893 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2894 prev->on_rq = 0; 2895 2896 /* 2897 * If a worker went to sleep, notify and ask workqueue 2898 * whether it wants to wake up a task to maintain 2899 * concurrency. 2900 */ 2901 if (prev->flags & PF_WQ_WORKER) { 2902 struct task_struct *to_wakeup; 2903 2904 to_wakeup = wq_worker_sleeping(prev, cpu); 2905 if (to_wakeup) 2906 try_to_wake_up_local(to_wakeup); 2907 } 2908 } 2909 switch_count = &prev->nvcsw; 2910 } 2911 2912 pre_schedule(rq, prev); 2913 2914 if (unlikely(!rq->nr_running)) 2915 idle_balance(cpu, rq); 2916 2917 put_prev_task(rq, prev); 2918 next = pick_next_task(rq); 2919 clear_tsk_need_resched(prev); 2920 rq->skip_clock_update = 0; 2921 2922 if (likely(prev != next)) { 2923 rq->nr_switches++; 2924 rq->curr = next; 2925 ++*switch_count; 2926 2927 context_switch(rq, prev, next); /* unlocks the rq */ 2928 /* 2929 * The context switch have flipped the stack from under us 2930 * and restored the local variables which were saved when 2931 * this task called schedule() in the past. prev == current 2932 * is still correct, but it can be moved to another cpu/rq. 2933 */ 2934 cpu = smp_processor_id(); 2935 rq = cpu_rq(cpu); 2936 } else 2937 raw_spin_unlock_irq(&rq->lock); 2938 2939 post_schedule(rq); 2940 2941 sched_preempt_enable_no_resched(); 2942 if (need_resched()) 2943 goto need_resched; 2944} 2945 2946static inline void sched_submit_work(struct task_struct *tsk) 2947{ 2948 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2949 return; 2950 /* 2951 * If we are going to sleep and we have plugged IO queued, 2952 * make sure to submit it to avoid deadlocks. 2953 */ 2954 if (blk_needs_flush_plug(tsk)) 2955 blk_schedule_flush_plug(tsk); 2956} 2957 2958asmlinkage void __sched schedule(void) 2959{ 2960 struct task_struct *tsk = current; 2961 2962 sched_submit_work(tsk); 2963 __schedule(); 2964} 2965EXPORT_SYMBOL(schedule); 2966 2967#ifdef CONFIG_CONTEXT_TRACKING 2968asmlinkage void __sched schedule_user(void) 2969{ 2970 /* 2971 * If we come here after a random call to set_need_resched(), 2972 * or we have been woken up remotely but the IPI has not yet arrived, 2973 * we haven't yet exited the RCU idle mode. Do it here manually until 2974 * we find a better solution. 2975 */ 2976 user_exit(); 2977 schedule(); 2978 user_enter(); 2979} 2980#endif 2981 2982/** 2983 * schedule_preempt_disabled - called with preemption disabled 2984 * 2985 * Returns with preemption disabled. Note: preempt_count must be 1 2986 */ 2987void __sched schedule_preempt_disabled(void) 2988{ 2989 sched_preempt_enable_no_resched(); 2990 schedule(); 2991 preempt_disable(); 2992} 2993 2994#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 2995 2996static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 2997{ 2998 if (lock->owner != owner) 2999 return false; 3000 3001 /* 3002 * Ensure we emit the owner->on_cpu, dereference _after_ checking 3003 * lock->owner still matches owner, if that fails, owner might 3004 * point to free()d memory, if it still matches, the rcu_read_lock() 3005 * ensures the memory stays valid. 3006 */ 3007 barrier(); 3008 3009 return owner->on_cpu; 3010} 3011 3012/* 3013 * Look out! "owner" is an entirely speculative pointer 3014 * access and not reliable. 3015 */ 3016int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 3017{ 3018 if (!sched_feat(OWNER_SPIN)) 3019 return 0; 3020 3021 rcu_read_lock(); 3022 while (owner_running(lock, owner)) { 3023 if (need_resched()) 3024 break; 3025 3026 arch_mutex_cpu_relax(); 3027 } 3028 rcu_read_unlock(); 3029 3030 /* 3031 * We break out the loop above on need_resched() and when the 3032 * owner changed, which is a sign for heavy contention. Return 3033 * success only when lock->owner is NULL. 3034 */ 3035 return lock->owner == NULL; 3036} 3037#endif 3038 3039#ifdef CONFIG_PREEMPT 3040/* 3041 * this is the entry point to schedule() from in-kernel preemption 3042 * off of preempt_enable. Kernel preemptions off return from interrupt 3043 * occur there and call schedule directly. 3044 */ 3045asmlinkage void __sched notrace preempt_schedule(void) 3046{ 3047 struct thread_info *ti = current_thread_info(); 3048 3049 /* 3050 * If there is a non-zero preempt_count or interrupts are disabled, 3051 * we do not want to preempt the current task. Just return.. 3052 */ 3053 if (likely(ti->preempt_count || irqs_disabled())) 3054 return; 3055 3056 do { 3057 add_preempt_count_notrace(PREEMPT_ACTIVE); 3058 __schedule(); 3059 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3060 3061 /* 3062 * Check again in case we missed a preemption opportunity 3063 * between schedule and now. 3064 */ 3065 barrier(); 3066 } while (need_resched()); 3067} 3068EXPORT_SYMBOL(preempt_schedule); 3069 3070/* 3071 * this is the entry point to schedule() from kernel preemption 3072 * off of irq context. 3073 * Note, that this is called and return with irqs disabled. This will 3074 * protect us against recursive calling from irq. 3075 */ 3076asmlinkage void __sched preempt_schedule_irq(void) 3077{ 3078 struct thread_info *ti = current_thread_info(); 3079 3080 /* Catch callers which need to be fixed */ 3081 BUG_ON(ti->preempt_count || !irqs_disabled()); 3082 3083 user_exit(); 3084 do { 3085 add_preempt_count(PREEMPT_ACTIVE); 3086 local_irq_enable(); 3087 __schedule(); 3088 local_irq_disable(); 3089 sub_preempt_count(PREEMPT_ACTIVE); 3090 3091 /* 3092 * Check again in case we missed a preemption opportunity 3093 * between schedule and now. 3094 */ 3095 barrier(); 3096 } while (need_resched()); 3097} 3098 3099#endif /* CONFIG_PREEMPT */ 3100 3101int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3102 void *key) 3103{ 3104 return try_to_wake_up(curr->private, mode, wake_flags); 3105} 3106EXPORT_SYMBOL(default_wake_function); 3107 3108/* 3109 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3110 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3111 * number) then we wake all the non-exclusive tasks and one exclusive task. 3112 * 3113 * There are circumstances in which we can try to wake a task which has already 3114 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3115 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3116 */ 3117static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3118 int nr_exclusive, int wake_flags, void *key) 3119{ 3120 wait_queue_t *curr, *next; 3121 3122 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3123 unsigned flags = curr->flags; 3124 3125 if (curr->func(curr, mode, wake_flags, key) && 3126 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3127 break; 3128 } 3129} 3130 3131/** 3132 * __wake_up - wake up threads blocked on a waitqueue. 3133 * @q: the waitqueue 3134 * @mode: which threads 3135 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3136 * @key: is directly passed to the wakeup function 3137 * 3138 * It may be assumed that this function implies a write memory barrier before 3139 * changing the task state if and only if any tasks are woken up. 3140 */ 3141void __wake_up(wait_queue_head_t *q, unsigned int mode, 3142 int nr_exclusive, void *key) 3143{ 3144 unsigned long flags; 3145 3146 spin_lock_irqsave(&q->lock, flags); 3147 __wake_up_common(q, mode, nr_exclusive, 0, key); 3148 spin_unlock_irqrestore(&q->lock, flags); 3149} 3150EXPORT_SYMBOL(__wake_up); 3151 3152/* 3153 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3154 */ 3155void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 3156{ 3157 __wake_up_common(q, mode, nr, 0, NULL); 3158} 3159EXPORT_SYMBOL_GPL(__wake_up_locked); 3160 3161void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3162{ 3163 __wake_up_common(q, mode, 1, 0, key); 3164} 3165EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3166 3167/** 3168 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3169 * @q: the waitqueue 3170 * @mode: which threads 3171 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3172 * @key: opaque value to be passed to wakeup targets 3173 * 3174 * The sync wakeup differs that the waker knows that it will schedule 3175 * away soon, so while the target thread will be woken up, it will not 3176 * be migrated to another CPU - ie. the two threads are 'synchronized' 3177 * with each other. This can prevent needless bouncing between CPUs. 3178 * 3179 * On UP it can prevent extra preemption. 3180 * 3181 * It may be assumed that this function implies a write memory barrier before 3182 * changing the task state if and only if any tasks are woken up. 3183 */ 3184void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3185 int nr_exclusive, void *key) 3186{ 3187 unsigned long flags; 3188 int wake_flags = WF_SYNC; 3189 3190 if (unlikely(!q)) 3191 return; 3192 3193 if (unlikely(!nr_exclusive)) 3194 wake_flags = 0; 3195 3196 spin_lock_irqsave(&q->lock, flags); 3197 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3198 spin_unlock_irqrestore(&q->lock, flags); 3199} 3200EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3201 3202/* 3203 * __wake_up_sync - see __wake_up_sync_key() 3204 */ 3205void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3206{ 3207 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3208} 3209EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3210 3211/** 3212 * complete: - signals a single thread waiting on this completion 3213 * @x: holds the state of this particular completion 3214 * 3215 * This will wake up a single thread waiting on this completion. Threads will be 3216 * awakened in the same order in which they were queued. 3217 * 3218 * See also complete_all(), wait_for_completion() and related routines. 3219 * 3220 * It may be assumed that this function implies a write memory barrier before 3221 * changing the task state if and only if any tasks are woken up. 3222 */ 3223void complete(struct completion *x) 3224{ 3225 unsigned long flags; 3226 3227 spin_lock_irqsave(&x->wait.lock, flags); 3228 x->done++; 3229 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3230 spin_unlock_irqrestore(&x->wait.lock, flags); 3231} 3232EXPORT_SYMBOL(complete); 3233 3234/** 3235 * complete_all: - signals all threads waiting on this completion 3236 * @x: holds the state of this particular completion 3237 * 3238 * This will wake up all threads waiting on this particular completion event. 3239 * 3240 * It may be assumed that this function implies a write memory barrier before 3241 * changing the task state if and only if any tasks are woken up. 3242 */ 3243void complete_all(struct completion *x) 3244{ 3245 unsigned long flags; 3246 3247 spin_lock_irqsave(&x->wait.lock, flags); 3248 x->done += UINT_MAX/2; 3249 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3250 spin_unlock_irqrestore(&x->wait.lock, flags); 3251} 3252EXPORT_SYMBOL(complete_all); 3253 3254static inline long __sched 3255do_wait_for_common(struct completion *x, long timeout, int state) 3256{ 3257 if (!x->done) { 3258 DECLARE_WAITQUEUE(wait, current); 3259 3260 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3261 do { 3262 if (signal_pending_state(state, current)) { 3263 timeout = -ERESTARTSYS; 3264 break; 3265 } 3266 __set_current_state(state); 3267 spin_unlock_irq(&x->wait.lock); 3268 timeout = schedule_timeout(timeout); 3269 spin_lock_irq(&x->wait.lock); 3270 } while (!x->done && timeout); 3271 __remove_wait_queue(&x->wait, &wait); 3272 if (!x->done) 3273 return timeout; 3274 } 3275 x->done--; 3276 return timeout ?: 1; 3277} 3278 3279static long __sched 3280wait_for_common(struct completion *x, long timeout, int state) 3281{ 3282 might_sleep(); 3283 3284 spin_lock_irq(&x->wait.lock); 3285 timeout = do_wait_for_common(x, timeout, state); 3286 spin_unlock_irq(&x->wait.lock); 3287 return timeout; 3288} 3289 3290/** 3291 * wait_for_completion: - waits for completion of a task 3292 * @x: holds the state of this particular completion 3293 * 3294 * This waits to be signaled for completion of a specific task. It is NOT 3295 * interruptible and there is no timeout. 3296 * 3297 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3298 * and interrupt capability. Also see complete(). 3299 */ 3300void __sched wait_for_completion(struct completion *x) 3301{ 3302 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3303} 3304EXPORT_SYMBOL(wait_for_completion); 3305 3306/** 3307 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3308 * @x: holds the state of this particular completion 3309 * @timeout: timeout value in jiffies 3310 * 3311 * This waits for either a completion of a specific task to be signaled or for a 3312 * specified timeout to expire. The timeout is in jiffies. It is not 3313 * interruptible. 3314 * 3315 * The return value is 0 if timed out, and positive (at least 1, or number of 3316 * jiffies left till timeout) if completed. 3317 */ 3318unsigned long __sched 3319wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3320{ 3321 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3322} 3323EXPORT_SYMBOL(wait_for_completion_timeout); 3324 3325/** 3326 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3327 * @x: holds the state of this particular completion 3328 * 3329 * This waits for completion of a specific task to be signaled. It is 3330 * interruptible. 3331 * 3332 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3333 */ 3334int __sched wait_for_completion_interruptible(struct completion *x) 3335{ 3336 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3337 if (t == -ERESTARTSYS) 3338 return t; 3339 return 0; 3340} 3341EXPORT_SYMBOL(wait_for_completion_interruptible); 3342 3343/** 3344 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3345 * @x: holds the state of this particular completion 3346 * @timeout: timeout value in jiffies 3347 * 3348 * This waits for either a completion of a specific task to be signaled or for a 3349 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3350 * 3351 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3352 * positive (at least 1, or number of jiffies left till timeout) if completed. 3353 */ 3354long __sched 3355wait_for_completion_interruptible_timeout(struct completion *x, 3356 unsigned long timeout) 3357{ 3358 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3359} 3360EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3361 3362/** 3363 * wait_for_completion_killable: - waits for completion of a task (killable) 3364 * @x: holds the state of this particular completion 3365 * 3366 * This waits to be signaled for completion of a specific task. It can be 3367 * interrupted by a kill signal. 3368 * 3369 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3370 */ 3371int __sched wait_for_completion_killable(struct completion *x) 3372{ 3373 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3374 if (t == -ERESTARTSYS) 3375 return t; 3376 return 0; 3377} 3378EXPORT_SYMBOL(wait_for_completion_killable); 3379 3380/** 3381 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3382 * @x: holds the state of this particular completion 3383 * @timeout: timeout value in jiffies 3384 * 3385 * This waits for either a completion of a specific task to be 3386 * signaled or for a specified timeout to expire. It can be 3387 * interrupted by a kill signal. The timeout is in jiffies. 3388 * 3389 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3390 * positive (at least 1, or number of jiffies left till timeout) if completed. 3391 */ 3392long __sched 3393wait_for_completion_killable_timeout(struct completion *x, 3394 unsigned long timeout) 3395{ 3396 return wait_for_common(x, timeout, TASK_KILLABLE); 3397} 3398EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3399 3400/** 3401 * try_wait_for_completion - try to decrement a completion without blocking 3402 * @x: completion structure 3403 * 3404 * Returns: 0 if a decrement cannot be done without blocking 3405 * 1 if a decrement succeeded. 3406 * 3407 * If a completion is being used as a counting completion, 3408 * attempt to decrement the counter without blocking. This 3409 * enables us to avoid waiting if the resource the completion 3410 * is protecting is not available. 3411 */ 3412bool try_wait_for_completion(struct completion *x) 3413{ 3414 unsigned long flags; 3415 int ret = 1; 3416 3417 spin_lock_irqsave(&x->wait.lock, flags); 3418 if (!x->done) 3419 ret = 0; 3420 else 3421 x->done--; 3422 spin_unlock_irqrestore(&x->wait.lock, flags); 3423 return ret; 3424} 3425EXPORT_SYMBOL(try_wait_for_completion); 3426 3427/** 3428 * completion_done - Test to see if a completion has any waiters 3429 * @x: completion structure 3430 * 3431 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3432 * 1 if there are no waiters. 3433 * 3434 */ 3435bool completion_done(struct completion *x) 3436{ 3437 unsigned long flags; 3438 int ret = 1; 3439 3440 spin_lock_irqsave(&x->wait.lock, flags); 3441 if (!x->done) 3442 ret = 0; 3443 spin_unlock_irqrestore(&x->wait.lock, flags); 3444 return ret; 3445} 3446EXPORT_SYMBOL(completion_done); 3447 3448static long __sched 3449sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3450{ 3451 unsigned long flags; 3452 wait_queue_t wait; 3453 3454 init_waitqueue_entry(&wait, current); 3455 3456 __set_current_state(state); 3457 3458 spin_lock_irqsave(&q->lock, flags); 3459 __add_wait_queue(q, &wait); 3460 spin_unlock(&q->lock); 3461 timeout = schedule_timeout(timeout); 3462 spin_lock_irq(&q->lock); 3463 __remove_wait_queue(q, &wait); 3464 spin_unlock_irqrestore(&q->lock, flags); 3465 3466 return timeout; 3467} 3468 3469void __sched interruptible_sleep_on(wait_queue_head_t *q) 3470{ 3471 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3472} 3473EXPORT_SYMBOL(interruptible_sleep_on); 3474 3475long __sched 3476interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3477{ 3478 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3479} 3480EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3481 3482void __sched sleep_on(wait_queue_head_t *q) 3483{ 3484 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3485} 3486EXPORT_SYMBOL(sleep_on); 3487 3488long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3489{ 3490 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3491} 3492EXPORT_SYMBOL(sleep_on_timeout); 3493 3494#ifdef CONFIG_RT_MUTEXES 3495 3496/* 3497 * rt_mutex_setprio - set the current priority of a task 3498 * @p: task 3499 * @prio: prio value (kernel-internal form) 3500 * 3501 * This function changes the 'effective' priority of a task. It does 3502 * not touch ->normal_prio like __setscheduler(). 3503 * 3504 * Used by the rt_mutex code to implement priority inheritance logic. 3505 */ 3506void rt_mutex_setprio(struct task_struct *p, int prio) 3507{ 3508 int oldprio, on_rq, running; 3509 struct rq *rq; 3510 const struct sched_class *prev_class; 3511 3512 BUG_ON(prio < 0 || prio > MAX_PRIO); 3513 3514 rq = __task_rq_lock(p); 3515 3516 /* 3517 * Idle task boosting is a nono in general. There is one 3518 * exception, when PREEMPT_RT and NOHZ is active: 3519 * 3520 * The idle task calls get_next_timer_interrupt() and holds 3521 * the timer wheel base->lock on the CPU and another CPU wants 3522 * to access the timer (probably to cancel it). We can safely 3523 * ignore the boosting request, as the idle CPU runs this code 3524 * with interrupts disabled and will complete the lock 3525 * protected section without being interrupted. So there is no 3526 * real need to boost. 3527 */ 3528 if (unlikely(p == rq->idle)) { 3529 WARN_ON(p != rq->curr); 3530 WARN_ON(p->pi_blocked_on); 3531 goto out_unlock; 3532 } 3533 3534 trace_sched_pi_setprio(p, prio); 3535 oldprio = p->prio; 3536 prev_class = p->sched_class; 3537 on_rq = p->on_rq; 3538 running = task_current(rq, p); 3539 if (on_rq) 3540 dequeue_task(rq, p, 0); 3541 if (running) 3542 p->sched_class->put_prev_task(rq, p); 3543 3544 if (rt_prio(prio)) 3545 p->sched_class = &rt_sched_class; 3546 else 3547 p->sched_class = &fair_sched_class; 3548 3549 p->prio = prio; 3550 3551 if (running) 3552 p->sched_class->set_curr_task(rq); 3553 if (on_rq) 3554 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3555 3556 check_class_changed(rq, p, prev_class, oldprio); 3557out_unlock: 3558 __task_rq_unlock(rq); 3559} 3560#endif 3561void set_user_nice(struct task_struct *p, long nice) 3562{ 3563 int old_prio, delta, on_rq; 3564 unsigned long flags; 3565 struct rq *rq; 3566 3567 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3568 return; 3569 /* 3570 * We have to be careful, if called from sys_setpriority(), 3571 * the task might be in the middle of scheduling on another CPU. 3572 */ 3573 rq = task_rq_lock(p, &flags); 3574 /* 3575 * The RT priorities are set via sched_setscheduler(), but we still 3576 * allow the 'normal' nice value to be set - but as expected 3577 * it wont have any effect on scheduling until the task is 3578 * SCHED_FIFO/SCHED_RR: 3579 */ 3580 if (task_has_rt_policy(p)) { 3581 p->static_prio = NICE_TO_PRIO(nice); 3582 goto out_unlock; 3583 } 3584 on_rq = p->on_rq; 3585 if (on_rq) 3586 dequeue_task(rq, p, 0); 3587 3588 p->static_prio = NICE_TO_PRIO(nice); 3589 set_load_weight(p); 3590 old_prio = p->prio; 3591 p->prio = effective_prio(p); 3592 delta = p->prio - old_prio; 3593 3594 if (on_rq) { 3595 enqueue_task(rq, p, 0); 3596 /* 3597 * If the task increased its priority or is running and 3598 * lowered its priority, then reschedule its CPU: 3599 */ 3600 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3601 resched_task(rq->curr); 3602 } 3603out_unlock: 3604 task_rq_unlock(rq, p, &flags); 3605} 3606EXPORT_SYMBOL(set_user_nice); 3607 3608/* 3609 * can_nice - check if a task can reduce its nice value 3610 * @p: task 3611 * @nice: nice value 3612 */ 3613int can_nice(const struct task_struct *p, const int nice) 3614{ 3615 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3616 int nice_rlim = 20 - nice; 3617 3618 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3619 capable(CAP_SYS_NICE)); 3620} 3621 3622#ifdef __ARCH_WANT_SYS_NICE 3623 3624/* 3625 * sys_nice - change the priority of the current process. 3626 * @increment: priority increment 3627 * 3628 * sys_setpriority is a more generic, but much slower function that 3629 * does similar things. 3630 */ 3631SYSCALL_DEFINE1(nice, int, increment) 3632{ 3633 long nice, retval; 3634 3635 /* 3636 * Setpriority might change our priority at the same moment. 3637 * We don't have to worry. Conceptually one call occurs first 3638 * and we have a single winner. 3639 */ 3640 if (increment < -40) 3641 increment = -40; 3642 if (increment > 40) 3643 increment = 40; 3644 3645 nice = TASK_NICE(current) + increment; 3646 if (nice < -20) 3647 nice = -20; 3648 if (nice > 19) 3649 nice = 19; 3650 3651 if (increment < 0 && !can_nice(current, nice)) 3652 return -EPERM; 3653 3654 retval = security_task_setnice(current, nice); 3655 if (retval) 3656 return retval; 3657 3658 set_user_nice(current, nice); 3659 return 0; 3660} 3661 3662#endif 3663 3664/** 3665 * task_prio - return the priority value of a given task. 3666 * @p: the task in question. 3667 * 3668 * This is the priority value as seen by users in /proc. 3669 * RT tasks are offset by -200. Normal tasks are centered 3670 * around 0, value goes from -16 to +15. 3671 */ 3672int task_prio(const struct task_struct *p) 3673{ 3674 return p->prio - MAX_RT_PRIO; 3675} 3676 3677/** 3678 * task_nice - return the nice value of a given task. 3679 * @p: the task in question. 3680 */ 3681int task_nice(const struct task_struct *p) 3682{ 3683 return TASK_NICE(p); 3684} 3685EXPORT_SYMBOL(task_nice); 3686 3687/** 3688 * idle_cpu - is a given cpu idle currently? 3689 * @cpu: the processor in question. 3690 */ 3691int idle_cpu(int cpu) 3692{ 3693 struct rq *rq = cpu_rq(cpu); 3694 3695 if (rq->curr != rq->idle) 3696 return 0; 3697 3698 if (rq->nr_running) 3699 return 0; 3700 3701#ifdef CONFIG_SMP 3702 if (!llist_empty(&rq->wake_list)) 3703 return 0; 3704#endif 3705 3706 return 1; 3707} 3708 3709/** 3710 * idle_task - return the idle task for a given cpu. 3711 * @cpu: the processor in question. 3712 */ 3713struct task_struct *idle_task(int cpu) 3714{ 3715 return cpu_rq(cpu)->idle; 3716} 3717 3718/** 3719 * find_process_by_pid - find a process with a matching PID value. 3720 * @pid: the pid in question. 3721 */ 3722static struct task_struct *find_process_by_pid(pid_t pid) 3723{ 3724 return pid ? find_task_by_vpid(pid) : current; 3725} 3726 3727/* Actually do priority change: must hold rq lock. */ 3728static void 3729__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3730{ 3731 p->policy = policy; 3732 p->rt_priority = prio; 3733 p->normal_prio = normal_prio(p); 3734 /* we are holding p->pi_lock already */ 3735 p->prio = rt_mutex_getprio(p); 3736 if (rt_prio(p->prio)) 3737 p->sched_class = &rt_sched_class; 3738 else 3739 p->sched_class = &fair_sched_class; 3740 set_load_weight(p); 3741} 3742 3743/* 3744 * check the target process has a UID that matches the current process's 3745 */ 3746static bool check_same_owner(struct task_struct *p) 3747{ 3748 const struct cred *cred = current_cred(), *pcred; 3749 bool match; 3750 3751 rcu_read_lock(); 3752 pcred = __task_cred(p); 3753 match = (uid_eq(cred->euid, pcred->euid) || 3754 uid_eq(cred->euid, pcred->uid)); 3755 rcu_read_unlock(); 3756 return match; 3757} 3758 3759static int __sched_setscheduler(struct task_struct *p, int policy, 3760 const struct sched_param *param, bool user) 3761{ 3762 int retval, oldprio, oldpolicy = -1, on_rq, running; 3763 unsigned long flags; 3764 const struct sched_class *prev_class; 3765 struct rq *rq; 3766 int reset_on_fork; 3767 3768 /* may grab non-irq protected spin_locks */ 3769 BUG_ON(in_interrupt()); 3770recheck: 3771 /* double check policy once rq lock held */ 3772 if (policy < 0) { 3773 reset_on_fork = p->sched_reset_on_fork; 3774 policy = oldpolicy = p->policy; 3775 } else { 3776 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3777 policy &= ~SCHED_RESET_ON_FORK; 3778 3779 if (policy != SCHED_FIFO && policy != SCHED_RR && 3780 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3781 policy != SCHED_IDLE) 3782 return -EINVAL; 3783 } 3784 3785 /* 3786 * Valid priorities for SCHED_FIFO and SCHED_RR are 3787 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3788 * SCHED_BATCH and SCHED_IDLE is 0. 3789 */ 3790 if (param->sched_priority < 0 || 3791 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3792 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3793 return -EINVAL; 3794 if (rt_policy(policy) != (param->sched_priority != 0)) 3795 return -EINVAL; 3796 3797 /* 3798 * Allow unprivileged RT tasks to decrease priority: 3799 */ 3800 if (user && !capable(CAP_SYS_NICE)) { 3801 if (rt_policy(policy)) { 3802 unsigned long rlim_rtprio = 3803 task_rlimit(p, RLIMIT_RTPRIO); 3804 3805 /* can't set/change the rt policy */ 3806 if (policy != p->policy && !rlim_rtprio) 3807 return -EPERM; 3808 3809 /* can't increase priority */ 3810 if (param->sched_priority > p->rt_priority && 3811 param->sched_priority > rlim_rtprio) 3812 return -EPERM; 3813 } 3814 3815 /* 3816 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3817 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3818 */ 3819 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3820 if (!can_nice(p, TASK_NICE(p))) 3821 return -EPERM; 3822 } 3823 3824 /* can't change other user's priorities */ 3825 if (!check_same_owner(p)) 3826 return -EPERM; 3827 3828 /* Normal users shall not reset the sched_reset_on_fork flag */ 3829 if (p->sched_reset_on_fork && !reset_on_fork) 3830 return -EPERM; 3831 } 3832 3833 if (user) { 3834 retval = security_task_setscheduler(p); 3835 if (retval) 3836 return retval; 3837 } 3838 3839 /* 3840 * make sure no PI-waiters arrive (or leave) while we are 3841 * changing the priority of the task: 3842 * 3843 * To be able to change p->policy safely, the appropriate 3844 * runqueue lock must be held. 3845 */ 3846 rq = task_rq_lock(p, &flags); 3847 3848 /* 3849 * Changing the policy of the stop threads its a very bad idea 3850 */ 3851 if (p == rq->stop) { 3852 task_rq_unlock(rq, p, &flags); 3853 return -EINVAL; 3854 } 3855 3856 /* 3857 * If not changing anything there's no need to proceed further: 3858 */ 3859 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3860 param->sched_priority == p->rt_priority))) { 3861 task_rq_unlock(rq, p, &flags); 3862 return 0; 3863 } 3864 3865#ifdef CONFIG_RT_GROUP_SCHED 3866 if (user) { 3867 /* 3868 * Do not allow realtime tasks into groups that have no runtime 3869 * assigned. 3870 */ 3871 if (rt_bandwidth_enabled() && rt_policy(policy) && 3872 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3873 !task_group_is_autogroup(task_group(p))) { 3874 task_rq_unlock(rq, p, &flags); 3875 return -EPERM; 3876 } 3877 } 3878#endif 3879 3880 /* recheck policy now with rq lock held */ 3881 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3882 policy = oldpolicy = -1; 3883 task_rq_unlock(rq, p, &flags); 3884 goto recheck; 3885 } 3886 on_rq = p->on_rq; 3887 running = task_current(rq, p); 3888 if (on_rq) 3889 dequeue_task(rq, p, 0); 3890 if (running) 3891 p->sched_class->put_prev_task(rq, p); 3892 3893 p->sched_reset_on_fork = reset_on_fork; 3894 3895 oldprio = p->prio; 3896 prev_class = p->sched_class; 3897 __setscheduler(rq, p, policy, param->sched_priority); 3898 3899 if (running) 3900 p->sched_class->set_curr_task(rq); 3901 if (on_rq) 3902 enqueue_task(rq, p, 0); 3903 3904 check_class_changed(rq, p, prev_class, oldprio); 3905 task_rq_unlock(rq, p, &flags); 3906 3907 rt_mutex_adjust_pi(p); 3908 3909 return 0; 3910} 3911 3912/** 3913 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3914 * @p: the task in question. 3915 * @policy: new policy. 3916 * @param: structure containing the new RT priority. 3917 * 3918 * NOTE that the task may be already dead. 3919 */ 3920int sched_setscheduler(struct task_struct *p, int policy, 3921 const struct sched_param *param) 3922{ 3923 return __sched_setscheduler(p, policy, param, true); 3924} 3925EXPORT_SYMBOL_GPL(sched_setscheduler); 3926 3927/** 3928 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3929 * @p: the task in question. 3930 * @policy: new policy. 3931 * @param: structure containing the new RT priority. 3932 * 3933 * Just like sched_setscheduler, only don't bother checking if the 3934 * current context has permission. For example, this is needed in 3935 * stop_machine(): we create temporary high priority worker threads, 3936 * but our caller might not have that capability. 3937 */ 3938int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3939 const struct sched_param *param) 3940{ 3941 return __sched_setscheduler(p, policy, param, false); 3942} 3943 3944static int 3945do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3946{ 3947 struct sched_param lparam; 3948 struct task_struct *p; 3949 int retval; 3950 3951 if (!param || pid < 0) 3952 return -EINVAL; 3953 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3954 return -EFAULT; 3955 3956 rcu_read_lock(); 3957 retval = -ESRCH; 3958 p = find_process_by_pid(pid); 3959 if (p != NULL) 3960 retval = sched_setscheduler(p, policy, &lparam); 3961 rcu_read_unlock(); 3962 3963 return retval; 3964} 3965 3966/** 3967 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3968 * @pid: the pid in question. 3969 * @policy: new policy. 3970 * @param: structure containing the new RT priority. 3971 */ 3972SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3973 struct sched_param __user *, param) 3974{ 3975 /* negative values for policy are not valid */ 3976 if (policy < 0) 3977 return -EINVAL; 3978 3979 return do_sched_setscheduler(pid, policy, param); 3980} 3981 3982/** 3983 * sys_sched_setparam - set/change the RT priority of a thread 3984 * @pid: the pid in question. 3985 * @param: structure containing the new RT priority. 3986 */ 3987SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3988{ 3989 return do_sched_setscheduler(pid, -1, param); 3990} 3991 3992/** 3993 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3994 * @pid: the pid in question. 3995 */ 3996SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3997{ 3998 struct task_struct *p; 3999 int retval; 4000 4001 if (pid < 0) 4002 return -EINVAL; 4003 4004 retval = -ESRCH; 4005 rcu_read_lock(); 4006 p = find_process_by_pid(pid); 4007 if (p) { 4008 retval = security_task_getscheduler(p); 4009 if (!retval) 4010 retval = p->policy 4011 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4012 } 4013 rcu_read_unlock(); 4014 return retval; 4015} 4016 4017/** 4018 * sys_sched_getparam - get the RT priority of a thread 4019 * @pid: the pid in question. 4020 * @param: structure containing the RT priority. 4021 */ 4022SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4023{ 4024 struct sched_param lp; 4025 struct task_struct *p; 4026 int retval; 4027 4028 if (!param || pid < 0) 4029 return -EINVAL; 4030 4031 rcu_read_lock(); 4032 p = find_process_by_pid(pid); 4033 retval = -ESRCH; 4034 if (!p) 4035 goto out_unlock; 4036 4037 retval = security_task_getscheduler(p); 4038 if (retval) 4039 goto out_unlock; 4040 4041 lp.sched_priority = p->rt_priority; 4042 rcu_read_unlock(); 4043 4044 /* 4045 * This one might sleep, we cannot do it with a spinlock held ... 4046 */ 4047 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4048 4049 return retval; 4050 4051out_unlock: 4052 rcu_read_unlock(); 4053 return retval; 4054} 4055 4056long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4057{ 4058 cpumask_var_t cpus_allowed, new_mask; 4059 struct task_struct *p; 4060 int retval; 4061 4062 get_online_cpus(); 4063 rcu_read_lock(); 4064 4065 p = find_process_by_pid(pid); 4066 if (!p) { 4067 rcu_read_unlock(); 4068 put_online_cpus(); 4069 return -ESRCH; 4070 } 4071 4072 /* Prevent p going away */ 4073 get_task_struct(p); 4074 rcu_read_unlock(); 4075 4076 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4077 retval = -ENOMEM; 4078 goto out_put_task; 4079 } 4080 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4081 retval = -ENOMEM; 4082 goto out_free_cpus_allowed; 4083 } 4084 retval = -EPERM; 4085 if (!check_same_owner(p)) { 4086 rcu_read_lock(); 4087 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 4088 rcu_read_unlock(); 4089 goto out_unlock; 4090 } 4091 rcu_read_unlock(); 4092 } 4093 4094 retval = security_task_setscheduler(p); 4095 if (retval) 4096 goto out_unlock; 4097 4098 cpuset_cpus_allowed(p, cpus_allowed); 4099 cpumask_and(new_mask, in_mask, cpus_allowed); 4100again: 4101 retval = set_cpus_allowed_ptr(p, new_mask); 4102 4103 if (!retval) { 4104 cpuset_cpus_allowed(p, cpus_allowed); 4105 if (!cpumask_subset(new_mask, cpus_allowed)) { 4106 /* 4107 * We must have raced with a concurrent cpuset 4108 * update. Just reset the cpus_allowed to the 4109 * cpuset's cpus_allowed 4110 */ 4111 cpumask_copy(new_mask, cpus_allowed); 4112 goto again; 4113 } 4114 } 4115out_unlock: 4116 free_cpumask_var(new_mask); 4117out_free_cpus_allowed: 4118 free_cpumask_var(cpus_allowed); 4119out_put_task: 4120 put_task_struct(p); 4121 put_online_cpus(); 4122 return retval; 4123} 4124 4125static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4126 struct cpumask *new_mask) 4127{ 4128 if (len < cpumask_size()) 4129 cpumask_clear(new_mask); 4130 else if (len > cpumask_size()) 4131 len = cpumask_size(); 4132 4133 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4134} 4135 4136/** 4137 * sys_sched_setaffinity - set the cpu affinity of a process 4138 * @pid: pid of the process 4139 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4140 * @user_mask_ptr: user-space pointer to the new cpu mask 4141 */ 4142SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4143 unsigned long __user *, user_mask_ptr) 4144{ 4145 cpumask_var_t new_mask; 4146 int retval; 4147 4148 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4149 return -ENOMEM; 4150 4151 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4152 if (retval == 0) 4153 retval = sched_setaffinity(pid, new_mask); 4154 free_cpumask_var(new_mask); 4155 return retval; 4156} 4157 4158long sched_getaffinity(pid_t pid, struct cpumask *mask) 4159{ 4160 struct task_struct *p; 4161 unsigned long flags; 4162 int retval; 4163 4164 get_online_cpus(); 4165 rcu_read_lock(); 4166 4167 retval = -ESRCH; 4168 p = find_process_by_pid(pid); 4169 if (!p) 4170 goto out_unlock; 4171 4172 retval = security_task_getscheduler(p); 4173 if (retval) 4174 goto out_unlock; 4175 4176 raw_spin_lock_irqsave(&p->pi_lock, flags); 4177 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4178 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4179 4180out_unlock: 4181 rcu_read_unlock(); 4182 put_online_cpus(); 4183 4184 return retval; 4185} 4186 4187/** 4188 * sys_sched_getaffinity - get the cpu affinity of a process 4189 * @pid: pid of the process 4190 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4191 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4192 */ 4193SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4194 unsigned long __user *, user_mask_ptr) 4195{ 4196 int ret; 4197 cpumask_var_t mask; 4198 4199 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4200 return -EINVAL; 4201 if (len & (sizeof(unsigned long)-1)) 4202 return -EINVAL; 4203 4204 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4205 return -ENOMEM; 4206 4207 ret = sched_getaffinity(pid, mask); 4208 if (ret == 0) { 4209 size_t retlen = min_t(size_t, len, cpumask_size()); 4210 4211 if (copy_to_user(user_mask_ptr, mask, retlen)) 4212 ret = -EFAULT; 4213 else 4214 ret = retlen; 4215 } 4216 free_cpumask_var(mask); 4217 4218 return ret; 4219} 4220 4221/** 4222 * sys_sched_yield - yield the current processor to other threads. 4223 * 4224 * This function yields the current CPU to other tasks. If there are no 4225 * other threads running on this CPU then this function will return. 4226 */ 4227SYSCALL_DEFINE0(sched_yield) 4228{ 4229 struct rq *rq = this_rq_lock(); 4230 4231 schedstat_inc(rq, yld_count); 4232 current->sched_class->yield_task(rq); 4233 4234 /* 4235 * Since we are going to call schedule() anyway, there's 4236 * no need to preempt or enable interrupts: 4237 */ 4238 __release(rq->lock); 4239 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4240 do_raw_spin_unlock(&rq->lock); 4241 sched_preempt_enable_no_resched(); 4242 4243 schedule(); 4244 4245 return 0; 4246} 4247 4248static inline int should_resched(void) 4249{ 4250 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4251} 4252 4253static void __cond_resched(void) 4254{ 4255 add_preempt_count(PREEMPT_ACTIVE); 4256 __schedule(); 4257 sub_preempt_count(PREEMPT_ACTIVE); 4258} 4259 4260int __sched _cond_resched(void) 4261{ 4262 if (should_resched()) { 4263 __cond_resched(); 4264 return 1; 4265 } 4266 return 0; 4267} 4268EXPORT_SYMBOL(_cond_resched); 4269 4270/* 4271 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4272 * call schedule, and on return reacquire the lock. 4273 * 4274 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4275 * operations here to prevent schedule() from being called twice (once via 4276 * spin_unlock(), once by hand). 4277 */ 4278int __cond_resched_lock(spinlock_t *lock) 4279{ 4280 int resched = should_resched(); 4281 int ret = 0; 4282 4283 lockdep_assert_held(lock); 4284 4285 if (spin_needbreak(lock) || resched) { 4286 spin_unlock(lock); 4287 if (resched) 4288 __cond_resched(); 4289 else 4290 cpu_relax(); 4291 ret = 1; 4292 spin_lock(lock); 4293 } 4294 return ret; 4295} 4296EXPORT_SYMBOL(__cond_resched_lock); 4297 4298int __sched __cond_resched_softirq(void) 4299{ 4300 BUG_ON(!in_softirq()); 4301 4302 if (should_resched()) { 4303 local_bh_enable(); 4304 __cond_resched(); 4305 local_bh_disable(); 4306 return 1; 4307 } 4308 return 0; 4309} 4310EXPORT_SYMBOL(__cond_resched_softirq); 4311 4312/** 4313 * yield - yield the current processor to other threads. 4314 * 4315 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4316 * 4317 * The scheduler is at all times free to pick the calling task as the most 4318 * eligible task to run, if removing the yield() call from your code breaks 4319 * it, its already broken. 4320 * 4321 * Typical broken usage is: 4322 * 4323 * while (!event) 4324 * yield(); 4325 * 4326 * where one assumes that yield() will let 'the other' process run that will 4327 * make event true. If the current task is a SCHED_FIFO task that will never 4328 * happen. Never use yield() as a progress guarantee!! 4329 * 4330 * If you want to use yield() to wait for something, use wait_event(). 4331 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4332 * If you still want to use yield(), do not! 4333 */ 4334void __sched yield(void) 4335{ 4336 set_current_state(TASK_RUNNING); 4337 sys_sched_yield(); 4338} 4339EXPORT_SYMBOL(yield); 4340 4341/** 4342 * yield_to - yield the current processor to another thread in 4343 * your thread group, or accelerate that thread toward the 4344 * processor it's on. 4345 * @p: target task 4346 * @preempt: whether task preemption is allowed or not 4347 * 4348 * It's the caller's job to ensure that the target task struct 4349 * can't go away on us before we can do any checks. 4350 * 4351 * Returns true if we indeed boosted the target task. 4352 */ 4353bool __sched yield_to(struct task_struct *p, bool preempt) 4354{ 4355 struct task_struct *curr = current; 4356 struct rq *rq, *p_rq; 4357 unsigned long flags; 4358 int yielded = 0; 4359 4360 local_irq_save(flags); 4361 rq = this_rq(); 4362 4363again: 4364 p_rq = task_rq(p); 4365 double_rq_lock(rq, p_rq); 4366 while (task_rq(p) != p_rq) { 4367 double_rq_unlock(rq, p_rq); 4368 goto again; 4369 } 4370 4371 if (!curr->sched_class->yield_to_task) 4372 goto out; 4373 4374 if (curr->sched_class != p->sched_class) 4375 goto out; 4376 4377 if (task_running(p_rq, p) || p->state) 4378 goto out; 4379 4380 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4381 if (yielded) { 4382 schedstat_inc(rq, yld_count); 4383 /* 4384 * Make p's CPU reschedule; pick_next_entity takes care of 4385 * fairness. 4386 */ 4387 if (preempt && rq != p_rq) 4388 resched_task(p_rq->curr); 4389 } 4390 4391out: 4392 double_rq_unlock(rq, p_rq); 4393 local_irq_restore(flags); 4394 4395 if (yielded) 4396 schedule(); 4397 4398 return yielded; 4399} 4400EXPORT_SYMBOL_GPL(yield_to); 4401 4402/* 4403 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4404 * that process accounting knows that this is a task in IO wait state. 4405 */ 4406void __sched io_schedule(void) 4407{ 4408 struct rq *rq = raw_rq(); 4409 4410 delayacct_blkio_start(); 4411 atomic_inc(&rq->nr_iowait); 4412 blk_flush_plug(current); 4413 current->in_iowait = 1; 4414 schedule(); 4415 current->in_iowait = 0; 4416 atomic_dec(&rq->nr_iowait); 4417 delayacct_blkio_end(); 4418} 4419EXPORT_SYMBOL(io_schedule); 4420 4421long __sched io_schedule_timeout(long timeout) 4422{ 4423 struct rq *rq = raw_rq(); 4424 long ret; 4425 4426 delayacct_blkio_start(); 4427 atomic_inc(&rq->nr_iowait); 4428 blk_flush_plug(current); 4429 current->in_iowait = 1; 4430 ret = schedule_timeout(timeout); 4431 current->in_iowait = 0; 4432 atomic_dec(&rq->nr_iowait); 4433 delayacct_blkio_end(); 4434 return ret; 4435} 4436 4437/** 4438 * sys_sched_get_priority_max - return maximum RT priority. 4439 * @policy: scheduling class. 4440 * 4441 * this syscall returns the maximum rt_priority that can be used 4442 * by a given scheduling class. 4443 */ 4444SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4445{ 4446 int ret = -EINVAL; 4447 4448 switch (policy) { 4449 case SCHED_FIFO: 4450 case SCHED_RR: 4451 ret = MAX_USER_RT_PRIO-1; 4452 break; 4453 case SCHED_NORMAL: 4454 case SCHED_BATCH: 4455 case SCHED_IDLE: 4456 ret = 0; 4457 break; 4458 } 4459 return ret; 4460} 4461 4462/** 4463 * sys_sched_get_priority_min - return minimum RT priority. 4464 * @policy: scheduling class. 4465 * 4466 * this syscall returns the minimum rt_priority that can be used 4467 * by a given scheduling class. 4468 */ 4469SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4470{ 4471 int ret = -EINVAL; 4472 4473 switch (policy) { 4474 case SCHED_FIFO: 4475 case SCHED_RR: 4476 ret = 1; 4477 break; 4478 case SCHED_NORMAL: 4479 case SCHED_BATCH: 4480 case SCHED_IDLE: 4481 ret = 0; 4482 } 4483 return ret; 4484} 4485 4486/** 4487 * sys_sched_rr_get_interval - return the default timeslice of a process. 4488 * @pid: pid of the process. 4489 * @interval: userspace pointer to the timeslice value. 4490 * 4491 * this syscall writes the default timeslice value of a given process 4492 * into the user-space timespec buffer. A value of '0' means infinity. 4493 */ 4494SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4495 struct timespec __user *, interval) 4496{ 4497 struct task_struct *p; 4498 unsigned int time_slice; 4499 unsigned long flags; 4500 struct rq *rq; 4501 int retval; 4502 struct timespec t; 4503 4504 if (pid < 0) 4505 return -EINVAL; 4506 4507 retval = -ESRCH; 4508 rcu_read_lock(); 4509 p = find_process_by_pid(pid); 4510 if (!p) 4511 goto out_unlock; 4512 4513 retval = security_task_getscheduler(p); 4514 if (retval) 4515 goto out_unlock; 4516 4517 rq = task_rq_lock(p, &flags); 4518 time_slice = p->sched_class->get_rr_interval(rq, p); 4519 task_rq_unlock(rq, p, &flags); 4520 4521 rcu_read_unlock(); 4522 jiffies_to_timespec(time_slice, &t); 4523 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4524 return retval; 4525 4526out_unlock: 4527 rcu_read_unlock(); 4528 return retval; 4529} 4530 4531static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4532 4533void sched_show_task(struct task_struct *p) 4534{ 4535 unsigned long free = 0; 4536 int ppid; 4537 unsigned state; 4538 4539 state = p->state ? __ffs(p->state) + 1 : 0; 4540 printk(KERN_INFO "%-15.15s %c", p->comm, 4541 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4542#if BITS_PER_LONG == 32 4543 if (state == TASK_RUNNING) 4544 printk(KERN_CONT " running "); 4545 else 4546 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4547#else 4548 if (state == TASK_RUNNING) 4549 printk(KERN_CONT " running task "); 4550 else 4551 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4552#endif 4553#ifdef CONFIG_DEBUG_STACK_USAGE 4554 free = stack_not_used(p); 4555#endif 4556 rcu_read_lock(); 4557 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4558 rcu_read_unlock(); 4559 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4560 task_pid_nr(p), ppid, 4561 (unsigned long)task_thread_info(p)->flags); 4562 4563 show_stack(p, NULL); 4564} 4565 4566void show_state_filter(unsigned long state_filter) 4567{ 4568 struct task_struct *g, *p; 4569 4570#if BITS_PER_LONG == 32 4571 printk(KERN_INFO 4572 " task PC stack pid father\n"); 4573#else 4574 printk(KERN_INFO 4575 " task PC stack pid father\n"); 4576#endif 4577 rcu_read_lock(); 4578 do_each_thread(g, p) { 4579 /* 4580 * reset the NMI-timeout, listing all files on a slow 4581 * console might take a lot of time: 4582 */ 4583 touch_nmi_watchdog(); 4584 if (!state_filter || (p->state & state_filter)) 4585 sched_show_task(p); 4586 } while_each_thread(g, p); 4587 4588 touch_all_softlockup_watchdogs(); 4589 4590#ifdef CONFIG_SCHED_DEBUG 4591 sysrq_sched_debug_show(); 4592#endif 4593 rcu_read_unlock(); 4594 /* 4595 * Only show locks if all tasks are dumped: 4596 */ 4597 if (!state_filter) 4598 debug_show_all_locks(); 4599} 4600 4601void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4602{ 4603 idle->sched_class = &idle_sched_class; 4604} 4605 4606/** 4607 * init_idle - set up an idle thread for a given CPU 4608 * @idle: task in question 4609 * @cpu: cpu the idle task belongs to 4610 * 4611 * NOTE: this function does not set the idle thread's NEED_RESCHED 4612 * flag, to make booting more robust. 4613 */ 4614void __cpuinit init_idle(struct task_struct *idle, int cpu) 4615{ 4616 struct rq *rq = cpu_rq(cpu); 4617 unsigned long flags; 4618 4619 raw_spin_lock_irqsave(&rq->lock, flags); 4620 4621 __sched_fork(idle); 4622 idle->state = TASK_RUNNING; 4623 idle->se.exec_start = sched_clock(); 4624 4625 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4626 /* 4627 * We're having a chicken and egg problem, even though we are 4628 * holding rq->lock, the cpu isn't yet set to this cpu so the 4629 * lockdep check in task_group() will fail. 4630 * 4631 * Similar case to sched_fork(). / Alternatively we could 4632 * use task_rq_lock() here and obtain the other rq->lock. 4633 * 4634 * Silence PROVE_RCU 4635 */ 4636 rcu_read_lock(); 4637 __set_task_cpu(idle, cpu); 4638 rcu_read_unlock(); 4639 4640 rq->curr = rq->idle = idle; 4641#if defined(CONFIG_SMP) 4642 idle->on_cpu = 1; 4643#endif 4644 raw_spin_unlock_irqrestore(&rq->lock, flags); 4645 4646 /* Set the preempt count _outside_ the spinlocks! */ 4647 task_thread_info(idle)->preempt_count = 0; 4648 4649 /* 4650 * The idle tasks have their own, simple scheduling class: 4651 */ 4652 idle->sched_class = &idle_sched_class; 4653 ftrace_graph_init_idle_task(idle, cpu); 4654 vtime_init_idle(idle); 4655#if defined(CONFIG_SMP) 4656 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4657#endif 4658} 4659 4660#ifdef CONFIG_SMP 4661void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4662{ 4663 if (p->sched_class && p->sched_class->set_cpus_allowed) 4664 p->sched_class->set_cpus_allowed(p, new_mask); 4665 4666 cpumask_copy(&p->cpus_allowed, new_mask); 4667 p->nr_cpus_allowed = cpumask_weight(new_mask); 4668} 4669 4670/* 4671 * This is how migration works: 4672 * 4673 * 1) we invoke migration_cpu_stop() on the target CPU using 4674 * stop_one_cpu(). 4675 * 2) stopper starts to run (implicitly forcing the migrated thread 4676 * off the CPU) 4677 * 3) it checks whether the migrated task is still in the wrong runqueue. 4678 * 4) if it's in the wrong runqueue then the migration thread removes 4679 * it and puts it into the right queue. 4680 * 5) stopper completes and stop_one_cpu() returns and the migration 4681 * is done. 4682 */ 4683 4684/* 4685 * Change a given task's CPU affinity. Migrate the thread to a 4686 * proper CPU and schedule it away if the CPU it's executing on 4687 * is removed from the allowed bitmask. 4688 * 4689 * NOTE: the caller must have a valid reference to the task, the 4690 * task must not exit() & deallocate itself prematurely. The 4691 * call is not atomic; no spinlocks may be held. 4692 */ 4693int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4694{ 4695 unsigned long flags; 4696 struct rq *rq; 4697 unsigned int dest_cpu; 4698 int ret = 0; 4699 4700 rq = task_rq_lock(p, &flags); 4701 4702 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4703 goto out; 4704 4705 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4706 ret = -EINVAL; 4707 goto out; 4708 } 4709 4710 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4711 ret = -EINVAL; 4712 goto out; 4713 } 4714 4715 do_set_cpus_allowed(p, new_mask); 4716 4717 /* Can the task run on the task's current CPU? If so, we're done */ 4718 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4719 goto out; 4720 4721 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4722 if (p->on_rq) { 4723 struct migration_arg arg = { p, dest_cpu }; 4724 /* Need help from migration thread: drop lock and wait. */ 4725 task_rq_unlock(rq, p, &flags); 4726 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4727 tlb_migrate_finish(p->mm); 4728 return 0; 4729 } 4730out: 4731 task_rq_unlock(rq, p, &flags); 4732 4733 return ret; 4734} 4735EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4736 4737/* 4738 * Move (not current) task off this cpu, onto dest cpu. We're doing 4739 * this because either it can't run here any more (set_cpus_allowed() 4740 * away from this CPU, or CPU going down), or because we're 4741 * attempting to rebalance this task on exec (sched_exec). 4742 * 4743 * So we race with normal scheduler movements, but that's OK, as long 4744 * as the task is no longer on this CPU. 4745 * 4746 * Returns non-zero if task was successfully migrated. 4747 */ 4748static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4749{ 4750 struct rq *rq_dest, *rq_src; 4751 int ret = 0; 4752 4753 if (unlikely(!cpu_active(dest_cpu))) 4754 return ret; 4755 4756 rq_src = cpu_rq(src_cpu); 4757 rq_dest = cpu_rq(dest_cpu); 4758 4759 raw_spin_lock(&p->pi_lock); 4760 double_rq_lock(rq_src, rq_dest); 4761 /* Already moved. */ 4762 if (task_cpu(p) != src_cpu) 4763 goto done; 4764 /* Affinity changed (again). */ 4765 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4766 goto fail; 4767 4768 /* 4769 * If we're not on a rq, the next wake-up will ensure we're 4770 * placed properly. 4771 */ 4772 if (p->on_rq) { 4773 dequeue_task(rq_src, p, 0); 4774 set_task_cpu(p, dest_cpu); 4775 enqueue_task(rq_dest, p, 0); 4776 check_preempt_curr(rq_dest, p, 0); 4777 } 4778done: 4779 ret = 1; 4780fail: 4781 double_rq_unlock(rq_src, rq_dest); 4782 raw_spin_unlock(&p->pi_lock); 4783 return ret; 4784} 4785 4786/* 4787 * migration_cpu_stop - this will be executed by a highprio stopper thread 4788 * and performs thread migration by bumping thread off CPU then 4789 * 'pushing' onto another runqueue. 4790 */ 4791static int migration_cpu_stop(void *data) 4792{ 4793 struct migration_arg *arg = data; 4794 4795 /* 4796 * The original target cpu might have gone down and we might 4797 * be on another cpu but it doesn't matter. 4798 */ 4799 local_irq_disable(); 4800 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4801 local_irq_enable(); 4802 return 0; 4803} 4804 4805#ifdef CONFIG_HOTPLUG_CPU 4806 4807/* 4808 * Ensures that the idle task is using init_mm right before its cpu goes 4809 * offline. 4810 */ 4811void idle_task_exit(void) 4812{ 4813 struct mm_struct *mm = current->active_mm; 4814 4815 BUG_ON(cpu_online(smp_processor_id())); 4816 4817 if (mm != &init_mm) 4818 switch_mm(mm, &init_mm, current); 4819 mmdrop(mm); 4820} 4821 4822/* 4823 * Since this CPU is going 'away' for a while, fold any nr_active delta 4824 * we might have. Assumes we're called after migrate_tasks() so that the 4825 * nr_active count is stable. 4826 * 4827 * Also see the comment "Global load-average calculations". 4828 */ 4829static void calc_load_migrate(struct rq *rq) 4830{ 4831 long delta = calc_load_fold_active(rq); 4832 if (delta) 4833 atomic_long_add(delta, &calc_load_tasks); 4834} 4835 4836/* 4837 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4838 * try_to_wake_up()->select_task_rq(). 4839 * 4840 * Called with rq->lock held even though we'er in stop_machine() and 4841 * there's no concurrency possible, we hold the required locks anyway 4842 * because of lock validation efforts. 4843 */ 4844static void migrate_tasks(unsigned int dead_cpu) 4845{ 4846 struct rq *rq = cpu_rq(dead_cpu); 4847 struct task_struct *next, *stop = rq->stop; 4848 int dest_cpu; 4849 4850 /* 4851 * Fudge the rq selection such that the below task selection loop 4852 * doesn't get stuck on the currently eligible stop task. 4853 * 4854 * We're currently inside stop_machine() and the rq is either stuck 4855 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4856 * either way we should never end up calling schedule() until we're 4857 * done here. 4858 */ 4859 rq->stop = NULL; 4860 4861 for ( ; ; ) { 4862 /* 4863 * There's this thread running, bail when that's the only 4864 * remaining thread. 4865 */ 4866 if (rq->nr_running == 1) 4867 break; 4868 4869 next = pick_next_task(rq); 4870 BUG_ON(!next); 4871 next->sched_class->put_prev_task(rq, next); 4872 4873 /* Find suitable destination for @next, with force if needed. */ 4874 dest_cpu = select_fallback_rq(dead_cpu, next); 4875 raw_spin_unlock(&rq->lock); 4876 4877 __migrate_task(next, dead_cpu, dest_cpu); 4878 4879 raw_spin_lock(&rq->lock); 4880 } 4881 4882 rq->stop = stop; 4883} 4884 4885#endif /* CONFIG_HOTPLUG_CPU */ 4886 4887#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4888 4889static struct ctl_table sd_ctl_dir[] = { 4890 { 4891 .procname = "sched_domain", 4892 .mode = 0555, 4893 }, 4894 {} 4895}; 4896 4897static struct ctl_table sd_ctl_root[] = { 4898 { 4899 .procname = "kernel", 4900 .mode = 0555, 4901 .child = sd_ctl_dir, 4902 }, 4903 {} 4904}; 4905 4906static struct ctl_table *sd_alloc_ctl_entry(int n) 4907{ 4908 struct ctl_table *entry = 4909 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4910 4911 return entry; 4912} 4913 4914static void sd_free_ctl_entry(struct ctl_table **tablep) 4915{ 4916 struct ctl_table *entry; 4917 4918 /* 4919 * In the intermediate directories, both the child directory and 4920 * procname are dynamically allocated and could fail but the mode 4921 * will always be set. In the lowest directory the names are 4922 * static strings and all have proc handlers. 4923 */ 4924 for (entry = *tablep; entry->mode; entry++) { 4925 if (entry->child) 4926 sd_free_ctl_entry(&entry->child); 4927 if (entry->proc_handler == NULL) 4928 kfree(entry->procname); 4929 } 4930 4931 kfree(*tablep); 4932 *tablep = NULL; 4933} 4934 4935static int min_load_idx = 0; 4936static int max_load_idx = CPU_LOAD_IDX_MAX; 4937 4938static void 4939set_table_entry(struct ctl_table *entry, 4940 const char *procname, void *data, int maxlen, 4941 umode_t mode, proc_handler *proc_handler, 4942 bool load_idx) 4943{ 4944 entry->procname = procname; 4945 entry->data = data; 4946 entry->maxlen = maxlen; 4947 entry->mode = mode; 4948 entry->proc_handler = proc_handler; 4949 4950 if (load_idx) { 4951 entry->extra1 = &min_load_idx; 4952 entry->extra2 = &max_load_idx; 4953 } 4954} 4955 4956static struct ctl_table * 4957sd_alloc_ctl_domain_table(struct sched_domain *sd) 4958{ 4959 struct ctl_table *table = sd_alloc_ctl_entry(13); 4960 4961 if (table == NULL) 4962 return NULL; 4963 4964 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4965 sizeof(long), 0644, proc_doulongvec_minmax, false); 4966 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4967 sizeof(long), 0644, proc_doulongvec_minmax, false); 4968 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4969 sizeof(int), 0644, proc_dointvec_minmax, true); 4970 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4971 sizeof(int), 0644, proc_dointvec_minmax, true); 4972 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4973 sizeof(int), 0644, proc_dointvec_minmax, true); 4974 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4975 sizeof(int), 0644, proc_dointvec_minmax, true); 4976 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4977 sizeof(int), 0644, proc_dointvec_minmax, true); 4978 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4979 sizeof(int), 0644, proc_dointvec_minmax, false); 4980 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4981 sizeof(int), 0644, proc_dointvec_minmax, false); 4982 set_table_entry(&table[9], "cache_nice_tries", 4983 &sd->cache_nice_tries, 4984 sizeof(int), 0644, proc_dointvec_minmax, false); 4985 set_table_entry(&table[10], "flags", &sd->flags, 4986 sizeof(int), 0644, proc_dointvec_minmax, false); 4987 set_table_entry(&table[11], "name", sd->name, 4988 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4989 /* &table[12] is terminator */ 4990 4991 return table; 4992} 4993 4994static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4995{ 4996 struct ctl_table *entry, *table; 4997 struct sched_domain *sd; 4998 int domain_num = 0, i; 4999 char buf[32]; 5000 5001 for_each_domain(cpu, sd) 5002 domain_num++; 5003 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5004 if (table == NULL) 5005 return NULL; 5006 5007 i = 0; 5008 for_each_domain(cpu, sd) { 5009 snprintf(buf, 32, "domain%d", i); 5010 entry->procname = kstrdup(buf, GFP_KERNEL); 5011 entry->mode = 0555; 5012 entry->child = sd_alloc_ctl_domain_table(sd); 5013 entry++; 5014 i++; 5015 } 5016 return table; 5017} 5018 5019static struct ctl_table_header *sd_sysctl_header; 5020static void register_sched_domain_sysctl(void) 5021{ 5022 int i, cpu_num = num_possible_cpus(); 5023 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5024 char buf[32]; 5025 5026 WARN_ON(sd_ctl_dir[0].child); 5027 sd_ctl_dir[0].child = entry; 5028 5029 if (entry == NULL) 5030 return; 5031 5032 for_each_possible_cpu(i) { 5033 snprintf(buf, 32, "cpu%d", i); 5034 entry->procname = kstrdup(buf, GFP_KERNEL); 5035 entry->mode = 0555; 5036 entry->child = sd_alloc_ctl_cpu_table(i); 5037 entry++; 5038 } 5039 5040 WARN_ON(sd_sysctl_header); 5041 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5042} 5043 5044/* may be called multiple times per register */ 5045static void unregister_sched_domain_sysctl(void) 5046{ 5047 if (sd_sysctl_header) 5048 unregister_sysctl_table(sd_sysctl_header); 5049 sd_sysctl_header = NULL; 5050 if (sd_ctl_dir[0].child) 5051 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5052} 5053#else 5054static void register_sched_domain_sysctl(void) 5055{ 5056} 5057static void unregister_sched_domain_sysctl(void) 5058{ 5059} 5060#endif 5061 5062static void set_rq_online(struct rq *rq) 5063{ 5064 if (!rq->online) { 5065 const struct sched_class *class; 5066 5067 cpumask_set_cpu(rq->cpu, rq->rd->online); 5068 rq->online = 1; 5069 5070 for_each_class(class) { 5071 if (class->rq_online) 5072 class->rq_online(rq); 5073 } 5074 } 5075} 5076 5077static void set_rq_offline(struct rq *rq) 5078{ 5079 if (rq->online) { 5080 const struct sched_class *class; 5081 5082 for_each_class(class) { 5083 if (class->rq_offline) 5084 class->rq_offline(rq); 5085 } 5086 5087 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5088 rq->online = 0; 5089 } 5090} 5091 5092/* 5093 * migration_call - callback that gets triggered when a CPU is added. 5094 * Here we can start up the necessary migration thread for the new CPU. 5095 */ 5096static int __cpuinit 5097migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5098{ 5099 int cpu = (long)hcpu; 5100 unsigned long flags; 5101 struct rq *rq = cpu_rq(cpu); 5102 5103 switch (action & ~CPU_TASKS_FROZEN) { 5104 5105 case CPU_UP_PREPARE: 5106 rq->calc_load_update = calc_load_update; 5107 break; 5108 5109 case CPU_ONLINE: 5110 /* Update our root-domain */ 5111 raw_spin_lock_irqsave(&rq->lock, flags); 5112 if (rq->rd) { 5113 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5114 5115 set_rq_online(rq); 5116 } 5117 raw_spin_unlock_irqrestore(&rq->lock, flags); 5118 break; 5119 5120#ifdef CONFIG_HOTPLUG_CPU 5121 case CPU_DYING: 5122 sched_ttwu_pending(); 5123 /* Update our root-domain */ 5124 raw_spin_lock_irqsave(&rq->lock, flags); 5125 if (rq->rd) { 5126 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5127 set_rq_offline(rq); 5128 } 5129 migrate_tasks(cpu); 5130 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5131 raw_spin_unlock_irqrestore(&rq->lock, flags); 5132 break; 5133 5134 case CPU_DEAD: 5135 calc_load_migrate(rq); 5136 break; 5137#endif 5138 } 5139 5140 update_max_interval(); 5141 5142 return NOTIFY_OK; 5143} 5144 5145/* 5146 * Register at high priority so that task migration (migrate_all_tasks) 5147 * happens before everything else. This has to be lower priority than 5148 * the notifier in the perf_event subsystem, though. 5149 */ 5150static struct notifier_block __cpuinitdata migration_notifier = { 5151 .notifier_call = migration_call, 5152 .priority = CPU_PRI_MIGRATION, 5153}; 5154 5155static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5156 unsigned long action, void *hcpu) 5157{ 5158 switch (action & ~CPU_TASKS_FROZEN) { 5159 case CPU_STARTING: 5160 case CPU_DOWN_FAILED: 5161 set_cpu_active((long)hcpu, true); 5162 return NOTIFY_OK; 5163 default: 5164 return NOTIFY_DONE; 5165 } 5166} 5167 5168static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5169 unsigned long action, void *hcpu) 5170{ 5171 switch (action & ~CPU_TASKS_FROZEN) { 5172 case CPU_DOWN_PREPARE: 5173 set_cpu_active((long)hcpu, false); 5174 return NOTIFY_OK; 5175 default: 5176 return NOTIFY_DONE; 5177 } 5178} 5179 5180static int __init migration_init(void) 5181{ 5182 void *cpu = (void *)(long)smp_processor_id(); 5183 int err; 5184 5185 /* Initialize migration for the boot CPU */ 5186 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5187 BUG_ON(err == NOTIFY_BAD); 5188 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5189 register_cpu_notifier(&migration_notifier); 5190 5191 /* Register cpu active notifiers */ 5192 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5193 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5194 5195 return 0; 5196} 5197early_initcall(migration_init); 5198#endif 5199 5200#ifdef CONFIG_SMP 5201 5202static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5203 5204#ifdef CONFIG_SCHED_DEBUG 5205 5206static __read_mostly int sched_debug_enabled; 5207 5208static int __init sched_debug_setup(char *str) 5209{ 5210 sched_debug_enabled = 1; 5211 5212 return 0; 5213} 5214early_param("sched_debug", sched_debug_setup); 5215 5216static inline bool sched_debug(void) 5217{ 5218 return sched_debug_enabled; 5219} 5220 5221static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5222 struct cpumask *groupmask) 5223{ 5224 struct sched_group *group = sd->groups; 5225 char str[256]; 5226 5227 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5228 cpumask_clear(groupmask); 5229 5230 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5231 5232 if (!(sd->flags & SD_LOAD_BALANCE)) { 5233 printk("does not load-balance\n"); 5234 if (sd->parent) 5235 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5236 " has parent"); 5237 return -1; 5238 } 5239 5240 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5241 5242 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5243 printk(KERN_ERR "ERROR: domain->span does not contain " 5244 "CPU%d\n", cpu); 5245 } 5246 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5247 printk(KERN_ERR "ERROR: domain->groups does not contain" 5248 " CPU%d\n", cpu); 5249 } 5250 5251 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5252 do { 5253 if (!group) { 5254 printk("\n"); 5255 printk(KERN_ERR "ERROR: group is NULL\n"); 5256 break; 5257 } 5258 5259 /* 5260 * Even though we initialize ->power to something semi-sane, 5261 * we leave power_orig unset. This allows us to detect if 5262 * domain iteration is still funny without causing /0 traps. 5263 */ 5264 if (!group->sgp->power_orig) { 5265 printk(KERN_CONT "\n"); 5266 printk(KERN_ERR "ERROR: domain->cpu_power not " 5267 "set\n"); 5268 break; 5269 } 5270 5271 if (!cpumask_weight(sched_group_cpus(group))) { 5272 printk(KERN_CONT "\n"); 5273 printk(KERN_ERR "ERROR: empty group\n"); 5274 break; 5275 } 5276 5277 if (!(sd->flags & SD_OVERLAP) && 5278 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5279 printk(KERN_CONT "\n"); 5280 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5281 break; 5282 } 5283 5284 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5285 5286 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5287 5288 printk(KERN_CONT " %s", str); 5289 if (group->sgp->power != SCHED_POWER_SCALE) { 5290 printk(KERN_CONT " (cpu_power = %d)", 5291 group->sgp->power); 5292 } 5293 5294 group = group->next; 5295 } while (group != sd->groups); 5296 printk(KERN_CONT "\n"); 5297 5298 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5299 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5300 5301 if (sd->parent && 5302 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5303 printk(KERN_ERR "ERROR: parent span is not a superset " 5304 "of domain->span\n"); 5305 return 0; 5306} 5307 5308static void sched_domain_debug(struct sched_domain *sd, int cpu) 5309{ 5310 int level = 0; 5311 5312 if (!sched_debug_enabled) 5313 return; 5314 5315 if (!sd) { 5316 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5317 return; 5318 } 5319 5320 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5321 5322 for (;;) { 5323 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5324 break; 5325 level++; 5326 sd = sd->parent; 5327 if (!sd) 5328 break; 5329 } 5330} 5331#else /* !CONFIG_SCHED_DEBUG */ 5332# define sched_domain_debug(sd, cpu) do { } while (0) 5333static inline bool sched_debug(void) 5334{ 5335 return false; 5336} 5337#endif /* CONFIG_SCHED_DEBUG */ 5338 5339static int sd_degenerate(struct sched_domain *sd) 5340{ 5341 if (cpumask_weight(sched_domain_span(sd)) == 1) 5342 return 1; 5343 5344 /* Following flags need at least 2 groups */ 5345 if (sd->flags & (SD_LOAD_BALANCE | 5346 SD_BALANCE_NEWIDLE | 5347 SD_BALANCE_FORK | 5348 SD_BALANCE_EXEC | 5349 SD_SHARE_CPUPOWER | 5350 SD_SHARE_PKG_RESOURCES)) { 5351 if (sd->groups != sd->groups->next) 5352 return 0; 5353 } 5354 5355 /* Following flags don't use groups */ 5356 if (sd->flags & (SD_WAKE_AFFINE)) 5357 return 0; 5358 5359 return 1; 5360} 5361 5362static int 5363sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5364{ 5365 unsigned long cflags = sd->flags, pflags = parent->flags; 5366 5367 if (sd_degenerate(parent)) 5368 return 1; 5369 5370 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5371 return 0; 5372 5373 /* Flags needing groups don't count if only 1 group in parent */ 5374 if (parent->groups == parent->groups->next) { 5375 pflags &= ~(SD_LOAD_BALANCE | 5376 SD_BALANCE_NEWIDLE | 5377 SD_BALANCE_FORK | 5378 SD_BALANCE_EXEC | 5379 SD_SHARE_CPUPOWER | 5380 SD_SHARE_PKG_RESOURCES); 5381 if (nr_node_ids == 1) 5382 pflags &= ~SD_SERIALIZE; 5383 } 5384 if (~cflags & pflags) 5385 return 0; 5386 5387 return 1; 5388} 5389 5390static void free_rootdomain(struct rcu_head *rcu) 5391{ 5392 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5393 5394 cpupri_cleanup(&rd->cpupri); 5395 free_cpumask_var(rd->rto_mask); 5396 free_cpumask_var(rd->online); 5397 free_cpumask_var(rd->span); 5398 kfree(rd); 5399} 5400 5401static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5402{ 5403 struct root_domain *old_rd = NULL; 5404 unsigned long flags; 5405 5406 raw_spin_lock_irqsave(&rq->lock, flags); 5407 5408 if (rq->rd) { 5409 old_rd = rq->rd; 5410 5411 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5412 set_rq_offline(rq); 5413 5414 cpumask_clear_cpu(rq->cpu, old_rd->span); 5415 5416 /* 5417 * If we dont want to free the old_rt yet then 5418 * set old_rd to NULL to skip the freeing later 5419 * in this function: 5420 */ 5421 if (!atomic_dec_and_test(&old_rd->refcount)) 5422 old_rd = NULL; 5423 } 5424 5425 atomic_inc(&rd->refcount); 5426 rq->rd = rd; 5427 5428 cpumask_set_cpu(rq->cpu, rd->span); 5429 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5430 set_rq_online(rq); 5431 5432 raw_spin_unlock_irqrestore(&rq->lock, flags); 5433 5434 if (old_rd) 5435 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5436} 5437 5438static int init_rootdomain(struct root_domain *rd) 5439{ 5440 memset(rd, 0, sizeof(*rd)); 5441 5442 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5443 goto out; 5444 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5445 goto free_span; 5446 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5447 goto free_online; 5448 5449 if (cpupri_init(&rd->cpupri) != 0) 5450 goto free_rto_mask; 5451 return 0; 5452 5453free_rto_mask: 5454 free_cpumask_var(rd->rto_mask); 5455free_online: 5456 free_cpumask_var(rd->online); 5457free_span: 5458 free_cpumask_var(rd->span); 5459out: 5460 return -ENOMEM; 5461} 5462 5463/* 5464 * By default the system creates a single root-domain with all cpus as 5465 * members (mimicking the global state we have today). 5466 */ 5467struct root_domain def_root_domain; 5468 5469static void init_defrootdomain(void) 5470{ 5471 init_rootdomain(&def_root_domain); 5472 5473 atomic_set(&def_root_domain.refcount, 1); 5474} 5475 5476static struct root_domain *alloc_rootdomain(void) 5477{ 5478 struct root_domain *rd; 5479 5480 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5481 if (!rd) 5482 return NULL; 5483 5484 if (init_rootdomain(rd) != 0) { 5485 kfree(rd); 5486 return NULL; 5487 } 5488 5489 return rd; 5490} 5491 5492static void free_sched_groups(struct sched_group *sg, int free_sgp) 5493{ 5494 struct sched_group *tmp, *first; 5495 5496 if (!sg) 5497 return; 5498 5499 first = sg; 5500 do { 5501 tmp = sg->next; 5502 5503 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5504 kfree(sg->sgp); 5505 5506 kfree(sg); 5507 sg = tmp; 5508 } while (sg != first); 5509} 5510 5511static void free_sched_domain(struct rcu_head *rcu) 5512{ 5513 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5514 5515 /* 5516 * If its an overlapping domain it has private groups, iterate and 5517 * nuke them all. 5518 */ 5519 if (sd->flags & SD_OVERLAP) { 5520 free_sched_groups(sd->groups, 1); 5521 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5522 kfree(sd->groups->sgp); 5523 kfree(sd->groups); 5524 } 5525 kfree(sd); 5526} 5527 5528static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5529{ 5530 call_rcu(&sd->rcu, free_sched_domain); 5531} 5532 5533static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5534{ 5535 for (; sd; sd = sd->parent) 5536 destroy_sched_domain(sd, cpu); 5537} 5538 5539/* 5540 * Keep a special pointer to the highest sched_domain that has 5541 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5542 * allows us to avoid some pointer chasing select_idle_sibling(). 5543 * 5544 * Also keep a unique ID per domain (we use the first cpu number in 5545 * the cpumask of the domain), this allows us to quickly tell if 5546 * two cpus are in the same cache domain, see cpus_share_cache(). 5547 */ 5548DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5549DEFINE_PER_CPU(int, sd_llc_id); 5550 5551static void update_top_cache_domain(int cpu) 5552{ 5553 struct sched_domain *sd; 5554 int id = cpu; 5555 5556 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5557 if (sd) 5558 id = cpumask_first(sched_domain_span(sd)); 5559 5560 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5561 per_cpu(sd_llc_id, cpu) = id; 5562} 5563 5564/* 5565 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5566 * hold the hotplug lock. 5567 */ 5568static void 5569cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5570{ 5571 struct rq *rq = cpu_rq(cpu); 5572 struct sched_domain *tmp; 5573 5574 /* Remove the sched domains which do not contribute to scheduling. */ 5575 for (tmp = sd; tmp; ) { 5576 struct sched_domain *parent = tmp->parent; 5577 if (!parent) 5578 break; 5579 5580 if (sd_parent_degenerate(tmp, parent)) { 5581 tmp->parent = parent->parent; 5582 if (parent->parent) 5583 parent->parent->child = tmp; 5584 destroy_sched_domain(parent, cpu); 5585 } else 5586 tmp = tmp->parent; 5587 } 5588 5589 if (sd && sd_degenerate(sd)) { 5590 tmp = sd; 5591 sd = sd->parent; 5592 destroy_sched_domain(tmp, cpu); 5593 if (sd) 5594 sd->child = NULL; 5595 } 5596 5597 sched_domain_debug(sd, cpu); 5598 5599 rq_attach_root(rq, rd); 5600 tmp = rq->sd; 5601 rcu_assign_pointer(rq->sd, sd); 5602 destroy_sched_domains(tmp, cpu); 5603 5604 update_top_cache_domain(cpu); 5605} 5606 5607/* cpus with isolated domains */ 5608static cpumask_var_t cpu_isolated_map; 5609 5610/* Setup the mask of cpus configured for isolated domains */ 5611static int __init isolated_cpu_setup(char *str) 5612{ 5613 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5614 cpulist_parse(str, cpu_isolated_map); 5615 return 1; 5616} 5617 5618__setup("isolcpus=", isolated_cpu_setup); 5619 5620static const struct cpumask *cpu_cpu_mask(int cpu) 5621{ 5622 return cpumask_of_node(cpu_to_node(cpu)); 5623} 5624 5625struct sd_data { 5626 struct sched_domain **__percpu sd; 5627 struct sched_group **__percpu sg; 5628 struct sched_group_power **__percpu sgp; 5629}; 5630 5631struct s_data { 5632 struct sched_domain ** __percpu sd; 5633 struct root_domain *rd; 5634}; 5635 5636enum s_alloc { 5637 sa_rootdomain, 5638 sa_sd, 5639 sa_sd_storage, 5640 sa_none, 5641}; 5642 5643struct sched_domain_topology_level; 5644 5645typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5646typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5647 5648#define SDTL_OVERLAP 0x01 5649 5650struct sched_domain_topology_level { 5651 sched_domain_init_f init; 5652 sched_domain_mask_f mask; 5653 int flags; 5654 int numa_level; 5655 struct sd_data data; 5656}; 5657 5658/* 5659 * Build an iteration mask that can exclude certain CPUs from the upwards 5660 * domain traversal. 5661 * 5662 * Asymmetric node setups can result in situations where the domain tree is of 5663 * unequal depth, make sure to skip domains that already cover the entire 5664 * range. 5665 * 5666 * In that case build_sched_domains() will have terminated the iteration early 5667 * and our sibling sd spans will be empty. Domains should always include the 5668 * cpu they're built on, so check that. 5669 * 5670 */ 5671static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5672{ 5673 const struct cpumask *span = sched_domain_span(sd); 5674 struct sd_data *sdd = sd->private; 5675 struct sched_domain *sibling; 5676 int i; 5677 5678 for_each_cpu(i, span) { 5679 sibling = *per_cpu_ptr(sdd->sd, i); 5680 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5681 continue; 5682 5683 cpumask_set_cpu(i, sched_group_mask(sg)); 5684 } 5685} 5686 5687/* 5688 * Return the canonical balance cpu for this group, this is the first cpu 5689 * of this group that's also in the iteration mask. 5690 */ 5691int group_balance_cpu(struct sched_group *sg) 5692{ 5693 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5694} 5695 5696static int 5697build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5698{ 5699 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5700 const struct cpumask *span = sched_domain_span(sd); 5701 struct cpumask *covered = sched_domains_tmpmask; 5702 struct sd_data *sdd = sd->private; 5703 struct sched_domain *child; 5704 int i; 5705 5706 cpumask_clear(covered); 5707 5708 for_each_cpu(i, span) { 5709 struct cpumask *sg_span; 5710 5711 if (cpumask_test_cpu(i, covered)) 5712 continue; 5713 5714 child = *per_cpu_ptr(sdd->sd, i); 5715 5716 /* See the comment near build_group_mask(). */ 5717 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5718 continue; 5719 5720 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5721 GFP_KERNEL, cpu_to_node(cpu)); 5722 5723 if (!sg) 5724 goto fail; 5725 5726 sg_span = sched_group_cpus(sg); 5727 if (child->child) { 5728 child = child->child; 5729 cpumask_copy(sg_span, sched_domain_span(child)); 5730 } else 5731 cpumask_set_cpu(i, sg_span); 5732 5733 cpumask_or(covered, covered, sg_span); 5734 5735 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5736 if (atomic_inc_return(&sg->sgp->ref) == 1) 5737 build_group_mask(sd, sg); 5738 5739 /* 5740 * Initialize sgp->power such that even if we mess up the 5741 * domains and no possible iteration will get us here, we won't 5742 * die on a /0 trap. 5743 */ 5744 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5745 5746 /* 5747 * Make sure the first group of this domain contains the 5748 * canonical balance cpu. Otherwise the sched_domain iteration 5749 * breaks. See update_sg_lb_stats(). 5750 */ 5751 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5752 group_balance_cpu(sg) == cpu) 5753 groups = sg; 5754 5755 if (!first) 5756 first = sg; 5757 if (last) 5758 last->next = sg; 5759 last = sg; 5760 last->next = first; 5761 } 5762 sd->groups = groups; 5763 5764 return 0; 5765 5766fail: 5767 free_sched_groups(first, 0); 5768 5769 return -ENOMEM; 5770} 5771 5772static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5773{ 5774 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5775 struct sched_domain *child = sd->child; 5776 5777 if (child) 5778 cpu = cpumask_first(sched_domain_span(child)); 5779 5780 if (sg) { 5781 *sg = *per_cpu_ptr(sdd->sg, cpu); 5782 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5783 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5784 } 5785 5786 return cpu; 5787} 5788 5789/* 5790 * build_sched_groups will build a circular linked list of the groups 5791 * covered by the given span, and will set each group's ->cpumask correctly, 5792 * and ->cpu_power to 0. 5793 * 5794 * Assumes the sched_domain tree is fully constructed 5795 */ 5796static int 5797build_sched_groups(struct sched_domain *sd, int cpu) 5798{ 5799 struct sched_group *first = NULL, *last = NULL; 5800 struct sd_data *sdd = sd->private; 5801 const struct cpumask *span = sched_domain_span(sd); 5802 struct cpumask *covered; 5803 int i; 5804 5805 get_group(cpu, sdd, &sd->groups); 5806 atomic_inc(&sd->groups->ref); 5807 5808 if (cpu != cpumask_first(sched_domain_span(sd))) 5809 return 0; 5810 5811 lockdep_assert_held(&sched_domains_mutex); 5812 covered = sched_domains_tmpmask; 5813 5814 cpumask_clear(covered); 5815 5816 for_each_cpu(i, span) { 5817 struct sched_group *sg; 5818 int group = get_group(i, sdd, &sg); 5819 int j; 5820 5821 if (cpumask_test_cpu(i, covered)) 5822 continue; 5823 5824 cpumask_clear(sched_group_cpus(sg)); 5825 sg->sgp->power = 0; 5826 cpumask_setall(sched_group_mask(sg)); 5827 5828 for_each_cpu(j, span) { 5829 if (get_group(j, sdd, NULL) != group) 5830 continue; 5831 5832 cpumask_set_cpu(j, covered); 5833 cpumask_set_cpu(j, sched_group_cpus(sg)); 5834 } 5835 5836 if (!first) 5837 first = sg; 5838 if (last) 5839 last->next = sg; 5840 last = sg; 5841 } 5842 last->next = first; 5843 5844 return 0; 5845} 5846 5847/* 5848 * Initialize sched groups cpu_power. 5849 * 5850 * cpu_power indicates the capacity of sched group, which is used while 5851 * distributing the load between different sched groups in a sched domain. 5852 * Typically cpu_power for all the groups in a sched domain will be same unless 5853 * there are asymmetries in the topology. If there are asymmetries, group 5854 * having more cpu_power will pickup more load compared to the group having 5855 * less cpu_power. 5856 */ 5857static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5858{ 5859 struct sched_group *sg = sd->groups; 5860 5861 WARN_ON(!sd || !sg); 5862 5863 do { 5864 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5865 sg = sg->next; 5866 } while (sg != sd->groups); 5867 5868 if (cpu != group_balance_cpu(sg)) 5869 return; 5870 5871 update_group_power(sd, cpu); 5872 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5873} 5874 5875int __weak arch_sd_sibling_asym_packing(void) 5876{ 5877 return 0*SD_ASYM_PACKING; 5878} 5879 5880/* 5881 * Initializers for schedule domains 5882 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5883 */ 5884 5885#ifdef CONFIG_SCHED_DEBUG 5886# define SD_INIT_NAME(sd, type) sd->name = #type 5887#else 5888# define SD_INIT_NAME(sd, type) do { } while (0) 5889#endif 5890 5891#define SD_INIT_FUNC(type) \ 5892static noinline struct sched_domain * \ 5893sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5894{ \ 5895 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5896 *sd = SD_##type##_INIT; \ 5897 SD_INIT_NAME(sd, type); \ 5898 sd->private = &tl->data; \ 5899 return sd; \ 5900} 5901 5902SD_INIT_FUNC(CPU) 5903#ifdef CONFIG_SCHED_SMT 5904 SD_INIT_FUNC(SIBLING) 5905#endif 5906#ifdef CONFIG_SCHED_MC 5907 SD_INIT_FUNC(MC) 5908#endif 5909#ifdef CONFIG_SCHED_BOOK 5910 SD_INIT_FUNC(BOOK) 5911#endif 5912 5913static int default_relax_domain_level = -1; 5914int sched_domain_level_max; 5915 5916static int __init setup_relax_domain_level(char *str) 5917{ 5918 if (kstrtoint(str, 0, &default_relax_domain_level)) 5919 pr_warn("Unable to set relax_domain_level\n"); 5920 5921 return 1; 5922} 5923__setup("relax_domain_level=", setup_relax_domain_level); 5924 5925static void set_domain_attribute(struct sched_domain *sd, 5926 struct sched_domain_attr *attr) 5927{ 5928 int request; 5929 5930 if (!attr || attr->relax_domain_level < 0) { 5931 if (default_relax_domain_level < 0) 5932 return; 5933 else 5934 request = default_relax_domain_level; 5935 } else 5936 request = attr->relax_domain_level; 5937 if (request < sd->level) { 5938 /* turn off idle balance on this domain */ 5939 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5940 } else { 5941 /* turn on idle balance on this domain */ 5942 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5943 } 5944} 5945 5946static void __sdt_free(const struct cpumask *cpu_map); 5947static int __sdt_alloc(const struct cpumask *cpu_map); 5948 5949static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5950 const struct cpumask *cpu_map) 5951{ 5952 switch (what) { 5953 case sa_rootdomain: 5954 if (!atomic_read(&d->rd->refcount)) 5955 free_rootdomain(&d->rd->rcu); /* fall through */ 5956 case sa_sd: 5957 free_percpu(d->sd); /* fall through */ 5958 case sa_sd_storage: 5959 __sdt_free(cpu_map); /* fall through */ 5960 case sa_none: 5961 break; 5962 } 5963} 5964 5965static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5966 const struct cpumask *cpu_map) 5967{ 5968 memset(d, 0, sizeof(*d)); 5969 5970 if (__sdt_alloc(cpu_map)) 5971 return sa_sd_storage; 5972 d->sd = alloc_percpu(struct sched_domain *); 5973 if (!d->sd) 5974 return sa_sd_storage; 5975 d->rd = alloc_rootdomain(); 5976 if (!d->rd) 5977 return sa_sd; 5978 return sa_rootdomain; 5979} 5980 5981/* 5982 * NULL the sd_data elements we've used to build the sched_domain and 5983 * sched_group structure so that the subsequent __free_domain_allocs() 5984 * will not free the data we're using. 5985 */ 5986static void claim_allocations(int cpu, struct sched_domain *sd) 5987{ 5988 struct sd_data *sdd = sd->private; 5989 5990 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5991 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5992 5993 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5994 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5995 5996 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5997 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5998} 5999 6000#ifdef CONFIG_SCHED_SMT 6001static const struct cpumask *cpu_smt_mask(int cpu) 6002{ 6003 return topology_thread_cpumask(cpu); 6004} 6005#endif 6006 6007/* 6008 * Topology list, bottom-up. 6009 */ 6010static struct sched_domain_topology_level default_topology[] = { 6011#ifdef CONFIG_SCHED_SMT 6012 { sd_init_SIBLING, cpu_smt_mask, }, 6013#endif 6014#ifdef CONFIG_SCHED_MC 6015 { sd_init_MC, cpu_coregroup_mask, }, 6016#endif 6017#ifdef CONFIG_SCHED_BOOK 6018 { sd_init_BOOK, cpu_book_mask, }, 6019#endif 6020 { sd_init_CPU, cpu_cpu_mask, }, 6021 { NULL, }, 6022}; 6023 6024static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6025 6026#ifdef CONFIG_NUMA 6027 6028static int sched_domains_numa_levels; 6029static int *sched_domains_numa_distance; 6030static struct cpumask ***sched_domains_numa_masks; 6031static int sched_domains_curr_level; 6032 6033static inline int sd_local_flags(int level) 6034{ 6035 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 6036 return 0; 6037 6038 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 6039} 6040 6041static struct sched_domain * 6042sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 6043{ 6044 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6045 int level = tl->numa_level; 6046 int sd_weight = cpumask_weight( 6047 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 6048 6049 *sd = (struct sched_domain){ 6050 .min_interval = sd_weight, 6051 .max_interval = 2*sd_weight, 6052 .busy_factor = 32, 6053 .imbalance_pct = 125, 6054 .cache_nice_tries = 2, 6055 .busy_idx = 3, 6056 .idle_idx = 2, 6057 .newidle_idx = 0, 6058 .wake_idx = 0, 6059 .forkexec_idx = 0, 6060 6061 .flags = 1*SD_LOAD_BALANCE 6062 | 1*SD_BALANCE_NEWIDLE 6063 | 0*SD_BALANCE_EXEC 6064 | 0*SD_BALANCE_FORK 6065 | 0*SD_BALANCE_WAKE 6066 | 0*SD_WAKE_AFFINE 6067 | 0*SD_SHARE_CPUPOWER 6068 | 0*SD_SHARE_PKG_RESOURCES 6069 | 1*SD_SERIALIZE 6070 | 0*SD_PREFER_SIBLING 6071 | sd_local_flags(level) 6072 , 6073 .last_balance = jiffies, 6074 .balance_interval = sd_weight, 6075 }; 6076 SD_INIT_NAME(sd, NUMA); 6077 sd->private = &tl->data; 6078 6079 /* 6080 * Ugly hack to pass state to sd_numa_mask()... 6081 */ 6082 sched_domains_curr_level = tl->numa_level; 6083 6084 return sd; 6085} 6086 6087static const struct cpumask *sd_numa_mask(int cpu) 6088{ 6089 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6090} 6091 6092static void sched_numa_warn(const char *str) 6093{ 6094 static int done = false; 6095 int i,j; 6096 6097 if (done) 6098 return; 6099 6100 done = true; 6101 6102 printk(KERN_WARNING "ERROR: %s\n\n", str); 6103 6104 for (i = 0; i < nr_node_ids; i++) { 6105 printk(KERN_WARNING " "); 6106 for (j = 0; j < nr_node_ids; j++) 6107 printk(KERN_CONT "%02d ", node_distance(i,j)); 6108 printk(KERN_CONT "\n"); 6109 } 6110 printk(KERN_WARNING "\n"); 6111} 6112 6113static bool find_numa_distance(int distance) 6114{ 6115 int i; 6116 6117 if (distance == node_distance(0, 0)) 6118 return true; 6119 6120 for (i = 0; i < sched_domains_numa_levels; i++) { 6121 if (sched_domains_numa_distance[i] == distance) 6122 return true; 6123 } 6124 6125 return false; 6126} 6127 6128static void sched_init_numa(void) 6129{ 6130 int next_distance, curr_distance = node_distance(0, 0); 6131 struct sched_domain_topology_level *tl; 6132 int level = 0; 6133 int i, j, k; 6134 6135 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6136 if (!sched_domains_numa_distance) 6137 return; 6138 6139 /* 6140 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6141 * unique distances in the node_distance() table. 6142 * 6143 * Assumes node_distance(0,j) includes all distances in 6144 * node_distance(i,j) in order to avoid cubic time. 6145 */ 6146 next_distance = curr_distance; 6147 for (i = 0; i < nr_node_ids; i++) { 6148 for (j = 0; j < nr_node_ids; j++) { 6149 for (k = 0; k < nr_node_ids; k++) { 6150 int distance = node_distance(i, k); 6151 6152 if (distance > curr_distance && 6153 (distance < next_distance || 6154 next_distance == curr_distance)) 6155 next_distance = distance; 6156 6157 /* 6158 * While not a strong assumption it would be nice to know 6159 * about cases where if node A is connected to B, B is not 6160 * equally connected to A. 6161 */ 6162 if (sched_debug() && node_distance(k, i) != distance) 6163 sched_numa_warn("Node-distance not symmetric"); 6164 6165 if (sched_debug() && i && !find_numa_distance(distance)) 6166 sched_numa_warn("Node-0 not representative"); 6167 } 6168 if (next_distance != curr_distance) { 6169 sched_domains_numa_distance[level++] = next_distance; 6170 sched_domains_numa_levels = level; 6171 curr_distance = next_distance; 6172 } else break; 6173 } 6174 6175 /* 6176 * In case of sched_debug() we verify the above assumption. 6177 */ 6178 if (!sched_debug()) 6179 break; 6180 } 6181 /* 6182 * 'level' contains the number of unique distances, excluding the 6183 * identity distance node_distance(i,i). 6184 * 6185 * The sched_domains_nume_distance[] array includes the actual distance 6186 * numbers. 6187 */ 6188 6189 /* 6190 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6191 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6192 * the array will contain less then 'level' members. This could be 6193 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6194 * in other functions. 6195 * 6196 * We reset it to 'level' at the end of this function. 6197 */ 6198 sched_domains_numa_levels = 0; 6199 6200 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6201 if (!sched_domains_numa_masks) 6202 return; 6203 6204 /* 6205 * Now for each level, construct a mask per node which contains all 6206 * cpus of nodes that are that many hops away from us. 6207 */ 6208 for (i = 0; i < level; i++) { 6209 sched_domains_numa_masks[i] = 6210 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6211 if (!sched_domains_numa_masks[i]) 6212 return; 6213 6214 for (j = 0; j < nr_node_ids; j++) { 6215 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6216 if (!mask) 6217 return; 6218 6219 sched_domains_numa_masks[i][j] = mask; 6220 6221 for (k = 0; k < nr_node_ids; k++) { 6222 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6223 continue; 6224 6225 cpumask_or(mask, mask, cpumask_of_node(k)); 6226 } 6227 } 6228 } 6229 6230 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6231 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6232 if (!tl) 6233 return; 6234 6235 /* 6236 * Copy the default topology bits.. 6237 */ 6238 for (i = 0; default_topology[i].init; i++) 6239 tl[i] = default_topology[i]; 6240 6241 /* 6242 * .. and append 'j' levels of NUMA goodness. 6243 */ 6244 for (j = 0; j < level; i++, j++) { 6245 tl[i] = (struct sched_domain_topology_level){ 6246 .init = sd_numa_init, 6247 .mask = sd_numa_mask, 6248 .flags = SDTL_OVERLAP, 6249 .numa_level = j, 6250 }; 6251 } 6252 6253 sched_domain_topology = tl; 6254 6255 sched_domains_numa_levels = level; 6256} 6257 6258static void sched_domains_numa_masks_set(int cpu) 6259{ 6260 int i, j; 6261 int node = cpu_to_node(cpu); 6262 6263 for (i = 0; i < sched_domains_numa_levels; i++) { 6264 for (j = 0; j < nr_node_ids; j++) { 6265 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6266 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6267 } 6268 } 6269} 6270 6271static void sched_domains_numa_masks_clear(int cpu) 6272{ 6273 int i, j; 6274 for (i = 0; i < sched_domains_numa_levels; i++) { 6275 for (j = 0; j < nr_node_ids; j++) 6276 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6277 } 6278} 6279 6280/* 6281 * Update sched_domains_numa_masks[level][node] array when new cpus 6282 * are onlined. 6283 */ 6284static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6285 unsigned long action, 6286 void *hcpu) 6287{ 6288 int cpu = (long)hcpu; 6289 6290 switch (action & ~CPU_TASKS_FROZEN) { 6291 case CPU_ONLINE: 6292 sched_domains_numa_masks_set(cpu); 6293 break; 6294 6295 case CPU_DEAD: 6296 sched_domains_numa_masks_clear(cpu); 6297 break; 6298 6299 default: 6300 return NOTIFY_DONE; 6301 } 6302 6303 return NOTIFY_OK; 6304} 6305#else 6306static inline void sched_init_numa(void) 6307{ 6308} 6309 6310static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6311 unsigned long action, 6312 void *hcpu) 6313{ 6314 return 0; 6315} 6316#endif /* CONFIG_NUMA */ 6317 6318static int __sdt_alloc(const struct cpumask *cpu_map) 6319{ 6320 struct sched_domain_topology_level *tl; 6321 int j; 6322 6323 for (tl = sched_domain_topology; tl->init; tl++) { 6324 struct sd_data *sdd = &tl->data; 6325 6326 sdd->sd = alloc_percpu(struct sched_domain *); 6327 if (!sdd->sd) 6328 return -ENOMEM; 6329 6330 sdd->sg = alloc_percpu(struct sched_group *); 6331 if (!sdd->sg) 6332 return -ENOMEM; 6333 6334 sdd->sgp = alloc_percpu(struct sched_group_power *); 6335 if (!sdd->sgp) 6336 return -ENOMEM; 6337 6338 for_each_cpu(j, cpu_map) { 6339 struct sched_domain *sd; 6340 struct sched_group *sg; 6341 struct sched_group_power *sgp; 6342 6343 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6344 GFP_KERNEL, cpu_to_node(j)); 6345 if (!sd) 6346 return -ENOMEM; 6347 6348 *per_cpu_ptr(sdd->sd, j) = sd; 6349 6350 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6351 GFP_KERNEL, cpu_to_node(j)); 6352 if (!sg) 6353 return -ENOMEM; 6354 6355 sg->next = sg; 6356 6357 *per_cpu_ptr(sdd->sg, j) = sg; 6358 6359 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6360 GFP_KERNEL, cpu_to_node(j)); 6361 if (!sgp) 6362 return -ENOMEM; 6363 6364 *per_cpu_ptr(sdd->sgp, j) = sgp; 6365 } 6366 } 6367 6368 return 0; 6369} 6370 6371static void __sdt_free(const struct cpumask *cpu_map) 6372{ 6373 struct sched_domain_topology_level *tl; 6374 int j; 6375 6376 for (tl = sched_domain_topology; tl->init; tl++) { 6377 struct sd_data *sdd = &tl->data; 6378 6379 for_each_cpu(j, cpu_map) { 6380 struct sched_domain *sd; 6381 6382 if (sdd->sd) { 6383 sd = *per_cpu_ptr(sdd->sd, j); 6384 if (sd && (sd->flags & SD_OVERLAP)) 6385 free_sched_groups(sd->groups, 0); 6386 kfree(*per_cpu_ptr(sdd->sd, j)); 6387 } 6388 6389 if (sdd->sg) 6390 kfree(*per_cpu_ptr(sdd->sg, j)); 6391 if (sdd->sgp) 6392 kfree(*per_cpu_ptr(sdd->sgp, j)); 6393 } 6394 free_percpu(sdd->sd); 6395 sdd->sd = NULL; 6396 free_percpu(sdd->sg); 6397 sdd->sg = NULL; 6398 free_percpu(sdd->sgp); 6399 sdd->sgp = NULL; 6400 } 6401} 6402 6403struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6404 struct s_data *d, const struct cpumask *cpu_map, 6405 struct sched_domain_attr *attr, struct sched_domain *child, 6406 int cpu) 6407{ 6408 struct sched_domain *sd = tl->init(tl, cpu); 6409 if (!sd) 6410 return child; 6411 6412 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6413 if (child) { 6414 sd->level = child->level + 1; 6415 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6416 child->parent = sd; 6417 } 6418 sd->child = child; 6419 set_domain_attribute(sd, attr); 6420 6421 return sd; 6422} 6423 6424/* 6425 * Build sched domains for a given set of cpus and attach the sched domains 6426 * to the individual cpus 6427 */ 6428static int build_sched_domains(const struct cpumask *cpu_map, 6429 struct sched_domain_attr *attr) 6430{ 6431 enum s_alloc alloc_state = sa_none; 6432 struct sched_domain *sd; 6433 struct s_data d; 6434 int i, ret = -ENOMEM; 6435 6436 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6437 if (alloc_state != sa_rootdomain) 6438 goto error; 6439 6440 /* Set up domains for cpus specified by the cpu_map. */ 6441 for_each_cpu(i, cpu_map) { 6442 struct sched_domain_topology_level *tl; 6443 6444 sd = NULL; 6445 for (tl = sched_domain_topology; tl->init; tl++) { 6446 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6447 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6448 sd->flags |= SD_OVERLAP; 6449 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6450 break; 6451 } 6452 6453 while (sd->child) 6454 sd = sd->child; 6455 6456 *per_cpu_ptr(d.sd, i) = sd; 6457 } 6458 6459 /* Build the groups for the domains */ 6460 for_each_cpu(i, cpu_map) { 6461 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6462 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6463 if (sd->flags & SD_OVERLAP) { 6464 if (build_overlap_sched_groups(sd, i)) 6465 goto error; 6466 } else { 6467 if (build_sched_groups(sd, i)) 6468 goto error; 6469 } 6470 } 6471 } 6472 6473 /* Calculate CPU power for physical packages and nodes */ 6474 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6475 if (!cpumask_test_cpu(i, cpu_map)) 6476 continue; 6477 6478 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6479 claim_allocations(i, sd); 6480 init_sched_groups_power(i, sd); 6481 } 6482 } 6483 6484 /* Attach the domains */ 6485 rcu_read_lock(); 6486 for_each_cpu(i, cpu_map) { 6487 sd = *per_cpu_ptr(d.sd, i); 6488 cpu_attach_domain(sd, d.rd, i); 6489 } 6490 rcu_read_unlock(); 6491 6492 ret = 0; 6493error: 6494 __free_domain_allocs(&d, alloc_state, cpu_map); 6495 return ret; 6496} 6497 6498static cpumask_var_t *doms_cur; /* current sched domains */ 6499static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6500static struct sched_domain_attr *dattr_cur; 6501 /* attribues of custom domains in 'doms_cur' */ 6502 6503/* 6504 * Special case: If a kmalloc of a doms_cur partition (array of 6505 * cpumask) fails, then fallback to a single sched domain, 6506 * as determined by the single cpumask fallback_doms. 6507 */ 6508static cpumask_var_t fallback_doms; 6509 6510/* 6511 * arch_update_cpu_topology lets virtualized architectures update the 6512 * cpu core maps. It is supposed to return 1 if the topology changed 6513 * or 0 if it stayed the same. 6514 */ 6515int __attribute__((weak)) arch_update_cpu_topology(void) 6516{ 6517 return 0; 6518} 6519 6520cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6521{ 6522 int i; 6523 cpumask_var_t *doms; 6524 6525 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6526 if (!doms) 6527 return NULL; 6528 for (i = 0; i < ndoms; i++) { 6529 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6530 free_sched_domains(doms, i); 6531 return NULL; 6532 } 6533 } 6534 return doms; 6535} 6536 6537void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6538{ 6539 unsigned int i; 6540 for (i = 0; i < ndoms; i++) 6541 free_cpumask_var(doms[i]); 6542 kfree(doms); 6543} 6544 6545/* 6546 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6547 * For now this just excludes isolated cpus, but could be used to 6548 * exclude other special cases in the future. 6549 */ 6550static int init_sched_domains(const struct cpumask *cpu_map) 6551{ 6552 int err; 6553 6554 arch_update_cpu_topology(); 6555 ndoms_cur = 1; 6556 doms_cur = alloc_sched_domains(ndoms_cur); 6557 if (!doms_cur) 6558 doms_cur = &fallback_doms; 6559 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6560 err = build_sched_domains(doms_cur[0], NULL); 6561 register_sched_domain_sysctl(); 6562 6563 return err; 6564} 6565 6566/* 6567 * Detach sched domains from a group of cpus specified in cpu_map 6568 * These cpus will now be attached to the NULL domain 6569 */ 6570static void detach_destroy_domains(const struct cpumask *cpu_map) 6571{ 6572 int i; 6573 6574 rcu_read_lock(); 6575 for_each_cpu(i, cpu_map) 6576 cpu_attach_domain(NULL, &def_root_domain, i); 6577 rcu_read_unlock(); 6578} 6579 6580/* handle null as "default" */ 6581static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6582 struct sched_domain_attr *new, int idx_new) 6583{ 6584 struct sched_domain_attr tmp; 6585 6586 /* fast path */ 6587 if (!new && !cur) 6588 return 1; 6589 6590 tmp = SD_ATTR_INIT; 6591 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6592 new ? (new + idx_new) : &tmp, 6593 sizeof(struct sched_domain_attr)); 6594} 6595 6596/* 6597 * Partition sched domains as specified by the 'ndoms_new' 6598 * cpumasks in the array doms_new[] of cpumasks. This compares 6599 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6600 * It destroys each deleted domain and builds each new domain. 6601 * 6602 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6603 * The masks don't intersect (don't overlap.) We should setup one 6604 * sched domain for each mask. CPUs not in any of the cpumasks will 6605 * not be load balanced. If the same cpumask appears both in the 6606 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6607 * it as it is. 6608 * 6609 * The passed in 'doms_new' should be allocated using 6610 * alloc_sched_domains. This routine takes ownership of it and will 6611 * free_sched_domains it when done with it. If the caller failed the 6612 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6613 * and partition_sched_domains() will fallback to the single partition 6614 * 'fallback_doms', it also forces the domains to be rebuilt. 6615 * 6616 * If doms_new == NULL it will be replaced with cpu_online_mask. 6617 * ndoms_new == 0 is a special case for destroying existing domains, 6618 * and it will not create the default domain. 6619 * 6620 * Call with hotplug lock held 6621 */ 6622void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6623 struct sched_domain_attr *dattr_new) 6624{ 6625 int i, j, n; 6626 int new_topology; 6627 6628 mutex_lock(&sched_domains_mutex); 6629 6630 /* always unregister in case we don't destroy any domains */ 6631 unregister_sched_domain_sysctl(); 6632 6633 /* Let architecture update cpu core mappings. */ 6634 new_topology = arch_update_cpu_topology(); 6635 6636 n = doms_new ? ndoms_new : 0; 6637 6638 /* Destroy deleted domains */ 6639 for (i = 0; i < ndoms_cur; i++) { 6640 for (j = 0; j < n && !new_topology; j++) { 6641 if (cpumask_equal(doms_cur[i], doms_new[j]) 6642 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6643 goto match1; 6644 } 6645 /* no match - a current sched domain not in new doms_new[] */ 6646 detach_destroy_domains(doms_cur[i]); 6647match1: 6648 ; 6649 } 6650 6651 if (doms_new == NULL) { 6652 ndoms_cur = 0; 6653 doms_new = &fallback_doms; 6654 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6655 WARN_ON_ONCE(dattr_new); 6656 } 6657 6658 /* Build new domains */ 6659 for (i = 0; i < ndoms_new; i++) { 6660 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6661 if (cpumask_equal(doms_new[i], doms_cur[j]) 6662 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6663 goto match2; 6664 } 6665 /* no match - add a new doms_new */ 6666 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6667match2: 6668 ; 6669 } 6670 6671 /* Remember the new sched domains */ 6672 if (doms_cur != &fallback_doms) 6673 free_sched_domains(doms_cur, ndoms_cur); 6674 kfree(dattr_cur); /* kfree(NULL) is safe */ 6675 doms_cur = doms_new; 6676 dattr_cur = dattr_new; 6677 ndoms_cur = ndoms_new; 6678 6679 register_sched_domain_sysctl(); 6680 6681 mutex_unlock(&sched_domains_mutex); 6682} 6683 6684static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6685 6686/* 6687 * Update cpusets according to cpu_active mask. If cpusets are 6688 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6689 * around partition_sched_domains(). 6690 * 6691 * If we come here as part of a suspend/resume, don't touch cpusets because we 6692 * want to restore it back to its original state upon resume anyway. 6693 */ 6694static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6695 void *hcpu) 6696{ 6697 switch (action) { 6698 case CPU_ONLINE_FROZEN: 6699 case CPU_DOWN_FAILED_FROZEN: 6700 6701 /* 6702 * num_cpus_frozen tracks how many CPUs are involved in suspend 6703 * resume sequence. As long as this is not the last online 6704 * operation in the resume sequence, just build a single sched 6705 * domain, ignoring cpusets. 6706 */ 6707 num_cpus_frozen--; 6708 if (likely(num_cpus_frozen)) { 6709 partition_sched_domains(1, NULL, NULL); 6710 break; 6711 } 6712 6713 /* 6714 * This is the last CPU online operation. So fall through and 6715 * restore the original sched domains by considering the 6716 * cpuset configurations. 6717 */ 6718 6719 case CPU_ONLINE: 6720 case CPU_DOWN_FAILED: 6721 cpuset_update_active_cpus(true); 6722 break; 6723 default: 6724 return NOTIFY_DONE; 6725 } 6726 return NOTIFY_OK; 6727} 6728 6729static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6730 void *hcpu) 6731{ 6732 switch (action) { 6733 case CPU_DOWN_PREPARE: 6734 cpuset_update_active_cpus(false); 6735 break; 6736 case CPU_DOWN_PREPARE_FROZEN: 6737 num_cpus_frozen++; 6738 partition_sched_domains(1, NULL, NULL); 6739 break; 6740 default: 6741 return NOTIFY_DONE; 6742 } 6743 return NOTIFY_OK; 6744} 6745 6746void __init sched_init_smp(void) 6747{ 6748 cpumask_var_t non_isolated_cpus; 6749 6750 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6751 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6752 6753 sched_init_numa(); 6754 6755 get_online_cpus(); 6756 mutex_lock(&sched_domains_mutex); 6757 init_sched_domains(cpu_active_mask); 6758 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6759 if (cpumask_empty(non_isolated_cpus)) 6760 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6761 mutex_unlock(&sched_domains_mutex); 6762 put_online_cpus(); 6763 6764 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6765 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6766 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6767 6768 /* RT runtime code needs to handle some hotplug events */ 6769 hotcpu_notifier(update_runtime, 0); 6770 6771 init_hrtick(); 6772 6773 /* Move init over to a non-isolated CPU */ 6774 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6775 BUG(); 6776 sched_init_granularity(); 6777 free_cpumask_var(non_isolated_cpus); 6778 6779 init_sched_rt_class(); 6780} 6781#else 6782void __init sched_init_smp(void) 6783{ 6784 sched_init_granularity(); 6785} 6786#endif /* CONFIG_SMP */ 6787 6788const_debug unsigned int sysctl_timer_migration = 1; 6789 6790int in_sched_functions(unsigned long addr) 6791{ 6792 return in_lock_functions(addr) || 6793 (addr >= (unsigned long)__sched_text_start 6794 && addr < (unsigned long)__sched_text_end); 6795} 6796 6797#ifdef CONFIG_CGROUP_SCHED 6798struct task_group root_task_group; 6799LIST_HEAD(task_groups); 6800#endif 6801 6802DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6803 6804void __init sched_init(void) 6805{ 6806 int i, j; 6807 unsigned long alloc_size = 0, ptr; 6808 6809#ifdef CONFIG_FAIR_GROUP_SCHED 6810 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6811#endif 6812#ifdef CONFIG_RT_GROUP_SCHED 6813 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6814#endif 6815#ifdef CONFIG_CPUMASK_OFFSTACK 6816 alloc_size += num_possible_cpus() * cpumask_size(); 6817#endif 6818 if (alloc_size) { 6819 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6820 6821#ifdef CONFIG_FAIR_GROUP_SCHED 6822 root_task_group.se = (struct sched_entity **)ptr; 6823 ptr += nr_cpu_ids * sizeof(void **); 6824 6825 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6826 ptr += nr_cpu_ids * sizeof(void **); 6827 6828#endif /* CONFIG_FAIR_GROUP_SCHED */ 6829#ifdef CONFIG_RT_GROUP_SCHED 6830 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6831 ptr += nr_cpu_ids * sizeof(void **); 6832 6833 root_task_group.rt_rq = (struct rt_rq **)ptr; 6834 ptr += nr_cpu_ids * sizeof(void **); 6835 6836#endif /* CONFIG_RT_GROUP_SCHED */ 6837#ifdef CONFIG_CPUMASK_OFFSTACK 6838 for_each_possible_cpu(i) { 6839 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6840 ptr += cpumask_size(); 6841 } 6842#endif /* CONFIG_CPUMASK_OFFSTACK */ 6843 } 6844 6845#ifdef CONFIG_SMP 6846 init_defrootdomain(); 6847#endif 6848 6849 init_rt_bandwidth(&def_rt_bandwidth, 6850 global_rt_period(), global_rt_runtime()); 6851 6852#ifdef CONFIG_RT_GROUP_SCHED 6853 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6854 global_rt_period(), global_rt_runtime()); 6855#endif /* CONFIG_RT_GROUP_SCHED */ 6856 6857#ifdef CONFIG_CGROUP_SCHED 6858 list_add(&root_task_group.list, &task_groups); 6859 INIT_LIST_HEAD(&root_task_group.children); 6860 INIT_LIST_HEAD(&root_task_group.siblings); 6861 autogroup_init(&init_task); 6862 6863#endif /* CONFIG_CGROUP_SCHED */ 6864 6865#ifdef CONFIG_CGROUP_CPUACCT 6866 root_cpuacct.cpustat = &kernel_cpustat; 6867 root_cpuacct.cpuusage = alloc_percpu(u64); 6868 /* Too early, not expected to fail */ 6869 BUG_ON(!root_cpuacct.cpuusage); 6870#endif 6871 for_each_possible_cpu(i) { 6872 struct rq *rq; 6873 6874 rq = cpu_rq(i); 6875 raw_spin_lock_init(&rq->lock); 6876 rq->nr_running = 0; 6877 rq->calc_load_active = 0; 6878 rq->calc_load_update = jiffies + LOAD_FREQ; 6879 init_cfs_rq(&rq->cfs); 6880 init_rt_rq(&rq->rt, rq); 6881#ifdef CONFIG_FAIR_GROUP_SCHED 6882 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6883 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6884 /* 6885 * How much cpu bandwidth does root_task_group get? 6886 * 6887 * In case of task-groups formed thr' the cgroup filesystem, it 6888 * gets 100% of the cpu resources in the system. This overall 6889 * system cpu resource is divided among the tasks of 6890 * root_task_group and its child task-groups in a fair manner, 6891 * based on each entity's (task or task-group's) weight 6892 * (se->load.weight). 6893 * 6894 * In other words, if root_task_group has 10 tasks of weight 6895 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6896 * then A0's share of the cpu resource is: 6897 * 6898 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6899 * 6900 * We achieve this by letting root_task_group's tasks sit 6901 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6902 */ 6903 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6904 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6905#endif /* CONFIG_FAIR_GROUP_SCHED */ 6906 6907 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6908#ifdef CONFIG_RT_GROUP_SCHED 6909 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6910 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6911#endif 6912 6913 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6914 rq->cpu_load[j] = 0; 6915 6916 rq->last_load_update_tick = jiffies; 6917 6918#ifdef CONFIG_SMP 6919 rq->sd = NULL; 6920 rq->rd = NULL; 6921 rq->cpu_power = SCHED_POWER_SCALE; 6922 rq->post_schedule = 0; 6923 rq->active_balance = 0; 6924 rq->next_balance = jiffies; 6925 rq->push_cpu = 0; 6926 rq->cpu = i; 6927 rq->online = 0; 6928 rq->idle_stamp = 0; 6929 rq->avg_idle = 2*sysctl_sched_migration_cost; 6930 6931 INIT_LIST_HEAD(&rq->cfs_tasks); 6932 6933 rq_attach_root(rq, &def_root_domain); 6934#ifdef CONFIG_NO_HZ 6935 rq->nohz_flags = 0; 6936#endif 6937#endif 6938 init_rq_hrtick(rq); 6939 atomic_set(&rq->nr_iowait, 0); 6940 } 6941 6942 set_load_weight(&init_task); 6943 6944#ifdef CONFIG_PREEMPT_NOTIFIERS 6945 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6946#endif 6947 6948#ifdef CONFIG_RT_MUTEXES 6949 plist_head_init(&init_task.pi_waiters); 6950#endif 6951 6952 /* 6953 * The boot idle thread does lazy MMU switching as well: 6954 */ 6955 atomic_inc(&init_mm.mm_count); 6956 enter_lazy_tlb(&init_mm, current); 6957 6958 /* 6959 * Make us the idle thread. Technically, schedule() should not be 6960 * called from this thread, however somewhere below it might be, 6961 * but because we are the idle thread, we just pick up running again 6962 * when this runqueue becomes "idle". 6963 */ 6964 init_idle(current, smp_processor_id()); 6965 6966 calc_load_update = jiffies + LOAD_FREQ; 6967 6968 /* 6969 * During early bootup we pretend to be a normal task: 6970 */ 6971 current->sched_class = &fair_sched_class; 6972 6973#ifdef CONFIG_SMP 6974 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6975 /* May be allocated at isolcpus cmdline parse time */ 6976 if (cpu_isolated_map == NULL) 6977 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6978 idle_thread_set_boot_cpu(); 6979#endif 6980 init_sched_fair_class(); 6981 6982 scheduler_running = 1; 6983} 6984 6985#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6986static inline int preempt_count_equals(int preempt_offset) 6987{ 6988 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6989 6990 return (nested == preempt_offset); 6991} 6992 6993void __might_sleep(const char *file, int line, int preempt_offset) 6994{ 6995 static unsigned long prev_jiffy; /* ratelimiting */ 6996 6997 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6998 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6999 system_state != SYSTEM_RUNNING || oops_in_progress) 7000 return; 7001 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7002 return; 7003 prev_jiffy = jiffies; 7004 7005 printk(KERN_ERR 7006 "BUG: sleeping function called from invalid context at %s:%d\n", 7007 file, line); 7008 printk(KERN_ERR 7009 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7010 in_atomic(), irqs_disabled(), 7011 current->pid, current->comm); 7012 7013 debug_show_held_locks(current); 7014 if (irqs_disabled()) 7015 print_irqtrace_events(current); 7016 dump_stack(); 7017} 7018EXPORT_SYMBOL(__might_sleep); 7019#endif 7020 7021#ifdef CONFIG_MAGIC_SYSRQ 7022static void normalize_task(struct rq *rq, struct task_struct *p) 7023{ 7024 const struct sched_class *prev_class = p->sched_class; 7025 int old_prio = p->prio; 7026 int on_rq; 7027 7028 on_rq = p->on_rq; 7029 if (on_rq) 7030 dequeue_task(rq, p, 0); 7031 __setscheduler(rq, p, SCHED_NORMAL, 0); 7032 if (on_rq) { 7033 enqueue_task(rq, p, 0); 7034 resched_task(rq->curr); 7035 } 7036 7037 check_class_changed(rq, p, prev_class, old_prio); 7038} 7039 7040void normalize_rt_tasks(void) 7041{ 7042 struct task_struct *g, *p; 7043 unsigned long flags; 7044 struct rq *rq; 7045 7046 read_lock_irqsave(&tasklist_lock, flags); 7047 do_each_thread(g, p) { 7048 /* 7049 * Only normalize user tasks: 7050 */ 7051 if (!p->mm) 7052 continue; 7053 7054 p->se.exec_start = 0; 7055#ifdef CONFIG_SCHEDSTATS 7056 p->se.statistics.wait_start = 0; 7057 p->se.statistics.sleep_start = 0; 7058 p->se.statistics.block_start = 0; 7059#endif 7060 7061 if (!rt_task(p)) { 7062 /* 7063 * Renice negative nice level userspace 7064 * tasks back to 0: 7065 */ 7066 if (TASK_NICE(p) < 0 && p->mm) 7067 set_user_nice(p, 0); 7068 continue; 7069 } 7070 7071 raw_spin_lock(&p->pi_lock); 7072 rq = __task_rq_lock(p); 7073 7074 normalize_task(rq, p); 7075 7076 __task_rq_unlock(rq); 7077 raw_spin_unlock(&p->pi_lock); 7078 } while_each_thread(g, p); 7079 7080 read_unlock_irqrestore(&tasklist_lock, flags); 7081} 7082 7083#endif /* CONFIG_MAGIC_SYSRQ */ 7084 7085#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7086/* 7087 * These functions are only useful for the IA64 MCA handling, or kdb. 7088 * 7089 * They can only be called when the whole system has been 7090 * stopped - every CPU needs to be quiescent, and no scheduling 7091 * activity can take place. Using them for anything else would 7092 * be a serious bug, and as a result, they aren't even visible 7093 * under any other configuration. 7094 */ 7095 7096/** 7097 * curr_task - return the current task for a given cpu. 7098 * @cpu: the processor in question. 7099 * 7100 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7101 */ 7102struct task_struct *curr_task(int cpu) 7103{ 7104 return cpu_curr(cpu); 7105} 7106 7107#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7108 7109#ifdef CONFIG_IA64 7110/** 7111 * set_curr_task - set the current task for a given cpu. 7112 * @cpu: the processor in question. 7113 * @p: the task pointer to set. 7114 * 7115 * Description: This function must only be used when non-maskable interrupts 7116 * are serviced on a separate stack. It allows the architecture to switch the 7117 * notion of the current task on a cpu in a non-blocking manner. This function 7118 * must be called with all CPU's synchronized, and interrupts disabled, the 7119 * and caller must save the original value of the current task (see 7120 * curr_task() above) and restore that value before reenabling interrupts and 7121 * re-starting the system. 7122 * 7123 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7124 */ 7125void set_curr_task(int cpu, struct task_struct *p) 7126{ 7127 cpu_curr(cpu) = p; 7128} 7129 7130#endif 7131 7132#ifdef CONFIG_CGROUP_SCHED 7133/* task_group_lock serializes the addition/removal of task groups */ 7134static DEFINE_SPINLOCK(task_group_lock); 7135 7136static void free_sched_group(struct task_group *tg) 7137{ 7138 free_fair_sched_group(tg); 7139 free_rt_sched_group(tg); 7140 autogroup_free(tg); 7141 kfree(tg); 7142} 7143 7144/* allocate runqueue etc for a new task group */ 7145struct task_group *sched_create_group(struct task_group *parent) 7146{ 7147 struct task_group *tg; 7148 unsigned long flags; 7149 7150 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7151 if (!tg) 7152 return ERR_PTR(-ENOMEM); 7153 7154 if (!alloc_fair_sched_group(tg, parent)) 7155 goto err; 7156 7157 if (!alloc_rt_sched_group(tg, parent)) 7158 goto err; 7159 7160 spin_lock_irqsave(&task_group_lock, flags); 7161 list_add_rcu(&tg->list, &task_groups); 7162 7163 WARN_ON(!parent); /* root should already exist */ 7164 7165 tg->parent = parent; 7166 INIT_LIST_HEAD(&tg->children); 7167 list_add_rcu(&tg->siblings, &parent->children); 7168 spin_unlock_irqrestore(&task_group_lock, flags); 7169 7170 return tg; 7171 7172err: 7173 free_sched_group(tg); 7174 return ERR_PTR(-ENOMEM); 7175} 7176 7177/* rcu callback to free various structures associated with a task group */ 7178static void free_sched_group_rcu(struct rcu_head *rhp) 7179{ 7180 /* now it should be safe to free those cfs_rqs */ 7181 free_sched_group(container_of(rhp, struct task_group, rcu)); 7182} 7183 7184/* Destroy runqueue etc associated with a task group */ 7185void sched_destroy_group(struct task_group *tg) 7186{ 7187 unsigned long flags; 7188 int i; 7189 7190 /* end participation in shares distribution */ 7191 for_each_possible_cpu(i) 7192 unregister_fair_sched_group(tg, i); 7193 7194 spin_lock_irqsave(&task_group_lock, flags); 7195 list_del_rcu(&tg->list); 7196 list_del_rcu(&tg->siblings); 7197 spin_unlock_irqrestore(&task_group_lock, flags); 7198 7199 /* wait for possible concurrent references to cfs_rqs complete */ 7200 call_rcu(&tg->rcu, free_sched_group_rcu); 7201} 7202 7203/* change task's runqueue when it moves between groups. 7204 * The caller of this function should have put the task in its new group 7205 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7206 * reflect its new group. 7207 */ 7208void sched_move_task(struct task_struct *tsk) 7209{ 7210 struct task_group *tg; 7211 int on_rq, running; 7212 unsigned long flags; 7213 struct rq *rq; 7214 7215 rq = task_rq_lock(tsk, &flags); 7216 7217 running = task_current(rq, tsk); 7218 on_rq = tsk->on_rq; 7219 7220 if (on_rq) 7221 dequeue_task(rq, tsk, 0); 7222 if (unlikely(running)) 7223 tsk->sched_class->put_prev_task(rq, tsk); 7224 7225 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 7226 lockdep_is_held(&tsk->sighand->siglock)), 7227 struct task_group, css); 7228 tg = autogroup_task_group(tsk, tg); 7229 tsk->sched_task_group = tg; 7230 7231#ifdef CONFIG_FAIR_GROUP_SCHED 7232 if (tsk->sched_class->task_move_group) 7233 tsk->sched_class->task_move_group(tsk, on_rq); 7234 else 7235#endif 7236 set_task_rq(tsk, task_cpu(tsk)); 7237 7238 if (unlikely(running)) 7239 tsk->sched_class->set_curr_task(rq); 7240 if (on_rq) 7241 enqueue_task(rq, tsk, 0); 7242 7243 task_rq_unlock(rq, tsk, &flags); 7244} 7245#endif /* CONFIG_CGROUP_SCHED */ 7246 7247#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7248static unsigned long to_ratio(u64 period, u64 runtime) 7249{ 7250 if (runtime == RUNTIME_INF) 7251 return 1ULL << 20; 7252 7253 return div64_u64(runtime << 20, period); 7254} 7255#endif 7256 7257#ifdef CONFIG_RT_GROUP_SCHED 7258/* 7259 * Ensure that the real time constraints are schedulable. 7260 */ 7261static DEFINE_MUTEX(rt_constraints_mutex); 7262 7263/* Must be called with tasklist_lock held */ 7264static inline int tg_has_rt_tasks(struct task_group *tg) 7265{ 7266 struct task_struct *g, *p; 7267 7268 do_each_thread(g, p) { 7269 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7270 return 1; 7271 } while_each_thread(g, p); 7272 7273 return 0; 7274} 7275 7276struct rt_schedulable_data { 7277 struct task_group *tg; 7278 u64 rt_period; 7279 u64 rt_runtime; 7280}; 7281 7282static int tg_rt_schedulable(struct task_group *tg, void *data) 7283{ 7284 struct rt_schedulable_data *d = data; 7285 struct task_group *child; 7286 unsigned long total, sum = 0; 7287 u64 period, runtime; 7288 7289 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7290 runtime = tg->rt_bandwidth.rt_runtime; 7291 7292 if (tg == d->tg) { 7293 period = d->rt_period; 7294 runtime = d->rt_runtime; 7295 } 7296 7297 /* 7298 * Cannot have more runtime than the period. 7299 */ 7300 if (runtime > period && runtime != RUNTIME_INF) 7301 return -EINVAL; 7302 7303 /* 7304 * Ensure we don't starve existing RT tasks. 7305 */ 7306 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7307 return -EBUSY; 7308 7309 total = to_ratio(period, runtime); 7310 7311 /* 7312 * Nobody can have more than the global setting allows. 7313 */ 7314 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7315 return -EINVAL; 7316 7317 /* 7318 * The sum of our children's runtime should not exceed our own. 7319 */ 7320 list_for_each_entry_rcu(child, &tg->children, siblings) { 7321 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7322 runtime = child->rt_bandwidth.rt_runtime; 7323 7324 if (child == d->tg) { 7325 period = d->rt_period; 7326 runtime = d->rt_runtime; 7327 } 7328 7329 sum += to_ratio(period, runtime); 7330 } 7331 7332 if (sum > total) 7333 return -EINVAL; 7334 7335 return 0; 7336} 7337 7338static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7339{ 7340 int ret; 7341 7342 struct rt_schedulable_data data = { 7343 .tg = tg, 7344 .rt_period = period, 7345 .rt_runtime = runtime, 7346 }; 7347 7348 rcu_read_lock(); 7349 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7350 rcu_read_unlock(); 7351 7352 return ret; 7353} 7354 7355static int tg_set_rt_bandwidth(struct task_group *tg, 7356 u64 rt_period, u64 rt_runtime) 7357{ 7358 int i, err = 0; 7359 7360 mutex_lock(&rt_constraints_mutex); 7361 read_lock(&tasklist_lock); 7362 err = __rt_schedulable(tg, rt_period, rt_runtime); 7363 if (err) 7364 goto unlock; 7365 7366 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7367 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7368 tg->rt_bandwidth.rt_runtime = rt_runtime; 7369 7370 for_each_possible_cpu(i) { 7371 struct rt_rq *rt_rq = tg->rt_rq[i]; 7372 7373 raw_spin_lock(&rt_rq->rt_runtime_lock); 7374 rt_rq->rt_runtime = rt_runtime; 7375 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7376 } 7377 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7378unlock: 7379 read_unlock(&tasklist_lock); 7380 mutex_unlock(&rt_constraints_mutex); 7381 7382 return err; 7383} 7384 7385int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7386{ 7387 u64 rt_runtime, rt_period; 7388 7389 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7390 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7391 if (rt_runtime_us < 0) 7392 rt_runtime = RUNTIME_INF; 7393 7394 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7395} 7396 7397long sched_group_rt_runtime(struct task_group *tg) 7398{ 7399 u64 rt_runtime_us; 7400 7401 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7402 return -1; 7403 7404 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7405 do_div(rt_runtime_us, NSEC_PER_USEC); 7406 return rt_runtime_us; 7407} 7408 7409int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7410{ 7411 u64 rt_runtime, rt_period; 7412 7413 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7414 rt_runtime = tg->rt_bandwidth.rt_runtime; 7415 7416 if (rt_period == 0) 7417 return -EINVAL; 7418 7419 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7420} 7421 7422long sched_group_rt_period(struct task_group *tg) 7423{ 7424 u64 rt_period_us; 7425 7426 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7427 do_div(rt_period_us, NSEC_PER_USEC); 7428 return rt_period_us; 7429} 7430 7431static int sched_rt_global_constraints(void) 7432{ 7433 u64 runtime, period; 7434 int ret = 0; 7435 7436 if (sysctl_sched_rt_period <= 0) 7437 return -EINVAL; 7438 7439 runtime = global_rt_runtime(); 7440 period = global_rt_period(); 7441 7442 /* 7443 * Sanity check on the sysctl variables. 7444 */ 7445 if (runtime > period && runtime != RUNTIME_INF) 7446 return -EINVAL; 7447 7448 mutex_lock(&rt_constraints_mutex); 7449 read_lock(&tasklist_lock); 7450 ret = __rt_schedulable(NULL, 0, 0); 7451 read_unlock(&tasklist_lock); 7452 mutex_unlock(&rt_constraints_mutex); 7453 7454 return ret; 7455} 7456 7457int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7458{ 7459 /* Don't accept realtime tasks when there is no way for them to run */ 7460 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7461 return 0; 7462 7463 return 1; 7464} 7465 7466#else /* !CONFIG_RT_GROUP_SCHED */ 7467static int sched_rt_global_constraints(void) 7468{ 7469 unsigned long flags; 7470 int i; 7471 7472 if (sysctl_sched_rt_period <= 0) 7473 return -EINVAL; 7474 7475 /* 7476 * There's always some RT tasks in the root group 7477 * -- migration, kstopmachine etc.. 7478 */ 7479 if (sysctl_sched_rt_runtime == 0) 7480 return -EBUSY; 7481 7482 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7483 for_each_possible_cpu(i) { 7484 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7485 7486 raw_spin_lock(&rt_rq->rt_runtime_lock); 7487 rt_rq->rt_runtime = global_rt_runtime(); 7488 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7489 } 7490 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7491 7492 return 0; 7493} 7494#endif /* CONFIG_RT_GROUP_SCHED */ 7495 7496int sched_rr_handler(struct ctl_table *table, int write, 7497 void __user *buffer, size_t *lenp, 7498 loff_t *ppos) 7499{ 7500 int ret; 7501 static DEFINE_MUTEX(mutex); 7502 7503 mutex_lock(&mutex); 7504 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7505 /* make sure that internally we keep jiffies */ 7506 /* also, writing zero resets timeslice to default */ 7507 if (!ret && write) { 7508 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7509 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7510 } 7511 mutex_unlock(&mutex); 7512 return ret; 7513} 7514 7515int sched_rt_handler(struct ctl_table *table, int write, 7516 void __user *buffer, size_t *lenp, 7517 loff_t *ppos) 7518{ 7519 int ret; 7520 int old_period, old_runtime; 7521 static DEFINE_MUTEX(mutex); 7522 7523 mutex_lock(&mutex); 7524 old_period = sysctl_sched_rt_period; 7525 old_runtime = sysctl_sched_rt_runtime; 7526 7527 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7528 7529 if (!ret && write) { 7530 ret = sched_rt_global_constraints(); 7531 if (ret) { 7532 sysctl_sched_rt_period = old_period; 7533 sysctl_sched_rt_runtime = old_runtime; 7534 } else { 7535 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7536 def_rt_bandwidth.rt_period = 7537 ns_to_ktime(global_rt_period()); 7538 } 7539 } 7540 mutex_unlock(&mutex); 7541 7542 return ret; 7543} 7544 7545#ifdef CONFIG_CGROUP_SCHED 7546 7547/* return corresponding task_group object of a cgroup */ 7548static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7549{ 7550 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7551 struct task_group, css); 7552} 7553 7554static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp) 7555{ 7556 struct task_group *tg, *parent; 7557 7558 if (!cgrp->parent) { 7559 /* This is early initialization for the top cgroup */ 7560 return &root_task_group.css; 7561 } 7562 7563 parent = cgroup_tg(cgrp->parent); 7564 tg = sched_create_group(parent); 7565 if (IS_ERR(tg)) 7566 return ERR_PTR(-ENOMEM); 7567 7568 return &tg->css; 7569} 7570 7571static void cpu_cgroup_css_free(struct cgroup *cgrp) 7572{ 7573 struct task_group *tg = cgroup_tg(cgrp); 7574 7575 sched_destroy_group(tg); 7576} 7577 7578static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7579 struct cgroup_taskset *tset) 7580{ 7581 struct task_struct *task; 7582 7583 cgroup_taskset_for_each(task, cgrp, tset) { 7584#ifdef CONFIG_RT_GROUP_SCHED 7585 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7586 return -EINVAL; 7587#else 7588 /* We don't support RT-tasks being in separate groups */ 7589 if (task->sched_class != &fair_sched_class) 7590 return -EINVAL; 7591#endif 7592 } 7593 return 0; 7594} 7595 7596static void cpu_cgroup_attach(struct cgroup *cgrp, 7597 struct cgroup_taskset *tset) 7598{ 7599 struct task_struct *task; 7600 7601 cgroup_taskset_for_each(task, cgrp, tset) 7602 sched_move_task(task); 7603} 7604 7605static void 7606cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7607 struct task_struct *task) 7608{ 7609 /* 7610 * cgroup_exit() is called in the copy_process() failure path. 7611 * Ignore this case since the task hasn't ran yet, this avoids 7612 * trying to poke a half freed task state from generic code. 7613 */ 7614 if (!(task->flags & PF_EXITING)) 7615 return; 7616 7617 sched_move_task(task); 7618} 7619 7620#ifdef CONFIG_FAIR_GROUP_SCHED 7621static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7622 u64 shareval) 7623{ 7624 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7625} 7626 7627static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7628{ 7629 struct task_group *tg = cgroup_tg(cgrp); 7630 7631 return (u64) scale_load_down(tg->shares); 7632} 7633 7634#ifdef CONFIG_CFS_BANDWIDTH 7635static DEFINE_MUTEX(cfs_constraints_mutex); 7636 7637const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7638const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7639 7640static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7641 7642static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7643{ 7644 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7645 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7646 7647 if (tg == &root_task_group) 7648 return -EINVAL; 7649 7650 /* 7651 * Ensure we have at some amount of bandwidth every period. This is 7652 * to prevent reaching a state of large arrears when throttled via 7653 * entity_tick() resulting in prolonged exit starvation. 7654 */ 7655 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7656 return -EINVAL; 7657 7658 /* 7659 * Likewise, bound things on the otherside by preventing insane quota 7660 * periods. This also allows us to normalize in computing quota 7661 * feasibility. 7662 */ 7663 if (period > max_cfs_quota_period) 7664 return -EINVAL; 7665 7666 mutex_lock(&cfs_constraints_mutex); 7667 ret = __cfs_schedulable(tg, period, quota); 7668 if (ret) 7669 goto out_unlock; 7670 7671 runtime_enabled = quota != RUNTIME_INF; 7672 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7673 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7674 raw_spin_lock_irq(&cfs_b->lock); 7675 cfs_b->period = ns_to_ktime(period); 7676 cfs_b->quota = quota; 7677 7678 __refill_cfs_bandwidth_runtime(cfs_b); 7679 /* restart the period timer (if active) to handle new period expiry */ 7680 if (runtime_enabled && cfs_b->timer_active) { 7681 /* force a reprogram */ 7682 cfs_b->timer_active = 0; 7683 __start_cfs_bandwidth(cfs_b); 7684 } 7685 raw_spin_unlock_irq(&cfs_b->lock); 7686 7687 for_each_possible_cpu(i) { 7688 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7689 struct rq *rq = cfs_rq->rq; 7690 7691 raw_spin_lock_irq(&rq->lock); 7692 cfs_rq->runtime_enabled = runtime_enabled; 7693 cfs_rq->runtime_remaining = 0; 7694 7695 if (cfs_rq->throttled) 7696 unthrottle_cfs_rq(cfs_rq); 7697 raw_spin_unlock_irq(&rq->lock); 7698 } 7699out_unlock: 7700 mutex_unlock(&cfs_constraints_mutex); 7701 7702 return ret; 7703} 7704 7705int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7706{ 7707 u64 quota, period; 7708 7709 period = ktime_to_ns(tg->cfs_bandwidth.period); 7710 if (cfs_quota_us < 0) 7711 quota = RUNTIME_INF; 7712 else 7713 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7714 7715 return tg_set_cfs_bandwidth(tg, period, quota); 7716} 7717 7718long tg_get_cfs_quota(struct task_group *tg) 7719{ 7720 u64 quota_us; 7721 7722 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7723 return -1; 7724 7725 quota_us = tg->cfs_bandwidth.quota; 7726 do_div(quota_us, NSEC_PER_USEC); 7727 7728 return quota_us; 7729} 7730 7731int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7732{ 7733 u64 quota, period; 7734 7735 period = (u64)cfs_period_us * NSEC_PER_USEC; 7736 quota = tg->cfs_bandwidth.quota; 7737 7738 return tg_set_cfs_bandwidth(tg, period, quota); 7739} 7740 7741long tg_get_cfs_period(struct task_group *tg) 7742{ 7743 u64 cfs_period_us; 7744 7745 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7746 do_div(cfs_period_us, NSEC_PER_USEC); 7747 7748 return cfs_period_us; 7749} 7750 7751static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7752{ 7753 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7754} 7755 7756static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7757 s64 cfs_quota_us) 7758{ 7759 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7760} 7761 7762static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7763{ 7764 return tg_get_cfs_period(cgroup_tg(cgrp)); 7765} 7766 7767static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7768 u64 cfs_period_us) 7769{ 7770 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7771} 7772 7773struct cfs_schedulable_data { 7774 struct task_group *tg; 7775 u64 period, quota; 7776}; 7777 7778/* 7779 * normalize group quota/period to be quota/max_period 7780 * note: units are usecs 7781 */ 7782static u64 normalize_cfs_quota(struct task_group *tg, 7783 struct cfs_schedulable_data *d) 7784{ 7785 u64 quota, period; 7786 7787 if (tg == d->tg) { 7788 period = d->period; 7789 quota = d->quota; 7790 } else { 7791 period = tg_get_cfs_period(tg); 7792 quota = tg_get_cfs_quota(tg); 7793 } 7794 7795 /* note: these should typically be equivalent */ 7796 if (quota == RUNTIME_INF || quota == -1) 7797 return RUNTIME_INF; 7798 7799 return to_ratio(period, quota); 7800} 7801 7802static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7803{ 7804 struct cfs_schedulable_data *d = data; 7805 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7806 s64 quota = 0, parent_quota = -1; 7807 7808 if (!tg->parent) { 7809 quota = RUNTIME_INF; 7810 } else { 7811 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7812 7813 quota = normalize_cfs_quota(tg, d); 7814 parent_quota = parent_b->hierarchal_quota; 7815 7816 /* 7817 * ensure max(child_quota) <= parent_quota, inherit when no 7818 * limit is set 7819 */ 7820 if (quota == RUNTIME_INF) 7821 quota = parent_quota; 7822 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7823 return -EINVAL; 7824 } 7825 cfs_b->hierarchal_quota = quota; 7826 7827 return 0; 7828} 7829 7830static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7831{ 7832 int ret; 7833 struct cfs_schedulable_data data = { 7834 .tg = tg, 7835 .period = period, 7836 .quota = quota, 7837 }; 7838 7839 if (quota != RUNTIME_INF) { 7840 do_div(data.period, NSEC_PER_USEC); 7841 do_div(data.quota, NSEC_PER_USEC); 7842 } 7843 7844 rcu_read_lock(); 7845 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7846 rcu_read_unlock(); 7847 7848 return ret; 7849} 7850 7851static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7852 struct cgroup_map_cb *cb) 7853{ 7854 struct task_group *tg = cgroup_tg(cgrp); 7855 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7856 7857 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7858 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7859 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7860 7861 return 0; 7862} 7863#endif /* CONFIG_CFS_BANDWIDTH */ 7864#endif /* CONFIG_FAIR_GROUP_SCHED */ 7865 7866#ifdef CONFIG_RT_GROUP_SCHED 7867static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7868 s64 val) 7869{ 7870 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7871} 7872 7873static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7874{ 7875 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7876} 7877 7878static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7879 u64 rt_period_us) 7880{ 7881 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7882} 7883 7884static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7885{ 7886 return sched_group_rt_period(cgroup_tg(cgrp)); 7887} 7888#endif /* CONFIG_RT_GROUP_SCHED */ 7889 7890static struct cftype cpu_files[] = { 7891#ifdef CONFIG_FAIR_GROUP_SCHED 7892 { 7893 .name = "shares", 7894 .read_u64 = cpu_shares_read_u64, 7895 .write_u64 = cpu_shares_write_u64, 7896 }, 7897#endif 7898#ifdef CONFIG_CFS_BANDWIDTH 7899 { 7900 .name = "cfs_quota_us", 7901 .read_s64 = cpu_cfs_quota_read_s64, 7902 .write_s64 = cpu_cfs_quota_write_s64, 7903 }, 7904 { 7905 .name = "cfs_period_us", 7906 .read_u64 = cpu_cfs_period_read_u64, 7907 .write_u64 = cpu_cfs_period_write_u64, 7908 }, 7909 { 7910 .name = "stat", 7911 .read_map = cpu_stats_show, 7912 }, 7913#endif 7914#ifdef CONFIG_RT_GROUP_SCHED 7915 { 7916 .name = "rt_runtime_us", 7917 .read_s64 = cpu_rt_runtime_read, 7918 .write_s64 = cpu_rt_runtime_write, 7919 }, 7920 { 7921 .name = "rt_period_us", 7922 .read_u64 = cpu_rt_period_read_uint, 7923 .write_u64 = cpu_rt_period_write_uint, 7924 }, 7925#endif 7926 { } /* terminate */ 7927}; 7928 7929struct cgroup_subsys cpu_cgroup_subsys = { 7930 .name = "cpu", 7931 .css_alloc = cpu_cgroup_css_alloc, 7932 .css_free = cpu_cgroup_css_free, 7933 .can_attach = cpu_cgroup_can_attach, 7934 .attach = cpu_cgroup_attach, 7935 .exit = cpu_cgroup_exit, 7936 .subsys_id = cpu_cgroup_subsys_id, 7937 .base_cftypes = cpu_files, 7938 .early_init = 1, 7939}; 7940 7941#endif /* CONFIG_CGROUP_SCHED */ 7942 7943#ifdef CONFIG_CGROUP_CPUACCT 7944 7945/* 7946 * CPU accounting code for task groups. 7947 * 7948 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 7949 * (balbir@in.ibm.com). 7950 */ 7951 7952struct cpuacct root_cpuacct; 7953 7954/* create a new cpu accounting group */ 7955static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp) 7956{ 7957 struct cpuacct *ca; 7958 7959 if (!cgrp->parent) 7960 return &root_cpuacct.css; 7961 7962 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 7963 if (!ca) 7964 goto out; 7965 7966 ca->cpuusage = alloc_percpu(u64); 7967 if (!ca->cpuusage) 7968 goto out_free_ca; 7969 7970 ca->cpustat = alloc_percpu(struct kernel_cpustat); 7971 if (!ca->cpustat) 7972 goto out_free_cpuusage; 7973 7974 return &ca->css; 7975 7976out_free_cpuusage: 7977 free_percpu(ca->cpuusage); 7978out_free_ca: 7979 kfree(ca); 7980out: 7981 return ERR_PTR(-ENOMEM); 7982} 7983 7984/* destroy an existing cpu accounting group */ 7985static void cpuacct_css_free(struct cgroup *cgrp) 7986{ 7987 struct cpuacct *ca = cgroup_ca(cgrp); 7988 7989 free_percpu(ca->cpustat); 7990 free_percpu(ca->cpuusage); 7991 kfree(ca); 7992} 7993 7994static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 7995{ 7996 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 7997 u64 data; 7998 7999#ifndef CONFIG_64BIT 8000 /* 8001 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 8002 */ 8003 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8004 data = *cpuusage; 8005 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8006#else 8007 data = *cpuusage; 8008#endif 8009 8010 return data; 8011} 8012 8013static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 8014{ 8015 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8016 8017#ifndef CONFIG_64BIT 8018 /* 8019 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 8020 */ 8021 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8022 *cpuusage = val; 8023 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8024#else 8025 *cpuusage = val; 8026#endif 8027} 8028 8029/* return total cpu usage (in nanoseconds) of a group */ 8030static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 8031{ 8032 struct cpuacct *ca = cgroup_ca(cgrp); 8033 u64 totalcpuusage = 0; 8034 int i; 8035 8036 for_each_present_cpu(i) 8037 totalcpuusage += cpuacct_cpuusage_read(ca, i); 8038 8039 return totalcpuusage; 8040} 8041 8042static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 8043 u64 reset) 8044{ 8045 struct cpuacct *ca = cgroup_ca(cgrp); 8046 int err = 0; 8047 int i; 8048 8049 if (reset) { 8050 err = -EINVAL; 8051 goto out; 8052 } 8053 8054 for_each_present_cpu(i) 8055 cpuacct_cpuusage_write(ca, i, 0); 8056 8057out: 8058 return err; 8059} 8060 8061static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 8062 struct seq_file *m) 8063{ 8064 struct cpuacct *ca = cgroup_ca(cgroup); 8065 u64 percpu; 8066 int i; 8067 8068 for_each_present_cpu(i) { 8069 percpu = cpuacct_cpuusage_read(ca, i); 8070 seq_printf(m, "%llu ", (unsigned long long) percpu); 8071 } 8072 seq_printf(m, "\n"); 8073 return 0; 8074} 8075 8076static const char *cpuacct_stat_desc[] = { 8077 [CPUACCT_STAT_USER] = "user", 8078 [CPUACCT_STAT_SYSTEM] = "system", 8079}; 8080 8081static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8082 struct cgroup_map_cb *cb) 8083{ 8084 struct cpuacct *ca = cgroup_ca(cgrp); 8085 int cpu; 8086 s64 val = 0; 8087 8088 for_each_online_cpu(cpu) { 8089 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8090 val += kcpustat->cpustat[CPUTIME_USER]; 8091 val += kcpustat->cpustat[CPUTIME_NICE]; 8092 } 8093 val = cputime64_to_clock_t(val); 8094 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8095 8096 val = 0; 8097 for_each_online_cpu(cpu) { 8098 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8099 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8100 val += kcpustat->cpustat[CPUTIME_IRQ]; 8101 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8102 } 8103 8104 val = cputime64_to_clock_t(val); 8105 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8106 8107 return 0; 8108} 8109 8110static struct cftype files[] = { 8111 { 8112 .name = "usage", 8113 .read_u64 = cpuusage_read, 8114 .write_u64 = cpuusage_write, 8115 }, 8116 { 8117 .name = "usage_percpu", 8118 .read_seq_string = cpuacct_percpu_seq_read, 8119 }, 8120 { 8121 .name = "stat", 8122 .read_map = cpuacct_stats_show, 8123 }, 8124 { } /* terminate */ 8125}; 8126 8127/* 8128 * charge this task's execution time to its accounting group. 8129 * 8130 * called with rq->lock held. 8131 */ 8132void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8133{ 8134 struct cpuacct *ca; 8135 int cpu; 8136 8137 if (unlikely(!cpuacct_subsys.active)) 8138 return; 8139 8140 cpu = task_cpu(tsk); 8141 8142 rcu_read_lock(); 8143 8144 ca = task_ca(tsk); 8145 8146 for (; ca; ca = parent_ca(ca)) { 8147 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8148 *cpuusage += cputime; 8149 } 8150 8151 rcu_read_unlock(); 8152} 8153 8154struct cgroup_subsys cpuacct_subsys = { 8155 .name = "cpuacct", 8156 .css_alloc = cpuacct_css_alloc, 8157 .css_free = cpuacct_css_free, 8158 .subsys_id = cpuacct_subsys_id, 8159 .base_cftypes = files, 8160}; 8161#endif /* CONFIG_CGROUP_CPUACCT */ 8162 8163void dump_cpu_task(int cpu) 8164{ 8165 pr_info("Task dump for CPU %d:\n", cpu); 8166 sched_show_task(cpu_curr(cpu)); 8167} 8168