core.c revision b8780c363d808a726a34793caa900923d32b6b80
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 * __task_rq_lock - lock the rq @p resides on. 301 */ 302static inline struct rq *__task_rq_lock(struct task_struct *p) 303 __acquires(rq->lock) 304{ 305 struct rq *rq; 306 307 lockdep_assert_held(&p->pi_lock); 308 309 for (;;) { 310 rq = task_rq(p); 311 raw_spin_lock(&rq->lock); 312 if (likely(rq == task_rq(p))) 313 return rq; 314 raw_spin_unlock(&rq->lock); 315 } 316} 317 318/* 319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 320 */ 321static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 322 __acquires(p->pi_lock) 323 __acquires(rq->lock) 324{ 325 struct rq *rq; 326 327 for (;;) { 328 raw_spin_lock_irqsave(&p->pi_lock, *flags); 329 rq = task_rq(p); 330 raw_spin_lock(&rq->lock); 331 if (likely(rq == task_rq(p))) 332 return rq; 333 raw_spin_unlock(&rq->lock); 334 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 335 } 336} 337 338static void __task_rq_unlock(struct rq *rq) 339 __releases(rq->lock) 340{ 341 raw_spin_unlock(&rq->lock); 342} 343 344static inline void 345task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 346 __releases(rq->lock) 347 __releases(p->pi_lock) 348{ 349 raw_spin_unlock(&rq->lock); 350 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 351} 352 353/* 354 * this_rq_lock - lock this runqueue and disable interrupts. 355 */ 356static struct rq *this_rq_lock(void) 357 __acquires(rq->lock) 358{ 359 struct rq *rq; 360 361 local_irq_disable(); 362 rq = this_rq(); 363 raw_spin_lock(&rq->lock); 364 365 return rq; 366} 367 368#ifdef CONFIG_SCHED_HRTICK 369/* 370 * Use HR-timers to deliver accurate preemption points. 371 */ 372 373static void hrtick_clear(struct rq *rq) 374{ 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377} 378 379/* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383static enum hrtimer_restart hrtick(struct hrtimer *timer) 384{ 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395} 396 397#ifdef CONFIG_SMP 398 399static int __hrtick_restart(struct rq *rq) 400{ 401 struct hrtimer *timer = &rq->hrtick_timer; 402 ktime_t time = hrtimer_get_softexpires(timer); 403 404 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 405} 406 407/* 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 __hrtick_restart(rq); 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 __hrtick_restart(rq); 434 } else if (!rq->hrtick_csd_pending) { 435 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 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 */ 514void resched_task(struct task_struct *p) 515{ 516 int cpu; 517 518 lockdep_assert_held(&task_rq(p)->lock); 519 520 if (test_tsk_need_resched(p)) 521 return; 522 523 set_tsk_need_resched(p); 524 525 cpu = task_cpu(p); 526 if (cpu == smp_processor_id()) { 527 set_preempt_need_resched(); 528 return; 529 } 530 531 /* NEED_RESCHED must be visible before we test polling */ 532 smp_mb(); 533 if (!tsk_is_polling(p)) 534 smp_send_reschedule(cpu); 535} 536 537void resched_cpu(int cpu) 538{ 539 struct rq *rq = cpu_rq(cpu); 540 unsigned long flags; 541 542 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 543 return; 544 resched_task(cpu_curr(cpu)); 545 raw_spin_unlock_irqrestore(&rq->lock, flags); 546} 547 548#ifdef CONFIG_SMP 549#ifdef CONFIG_NO_HZ_COMMON 550/* 551 * In the semi idle case, use the nearest busy cpu for migrating timers 552 * from an idle cpu. This is good for power-savings. 553 * 554 * We don't do similar optimization for completely idle system, as 555 * selecting an idle cpu will add more delays to the timers than intended 556 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 557 */ 558int get_nohz_timer_target(int pinned) 559{ 560 int cpu = smp_processor_id(); 561 int i; 562 struct sched_domain *sd; 563 564 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 565 return cpu; 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 */ 590static void 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 bool wake_up_full_nohz_cpu(int cpu) 621{ 622 if (tick_nohz_full_cpu(cpu)) { 623 if (cpu != smp_processor_id() || 624 tick_nohz_tick_stopped()) 625 smp_send_reschedule(cpu); 626 return true; 627 } 628 629 return false; 630} 631 632void wake_up_nohz_cpu(int cpu) 633{ 634 if (!wake_up_full_nohz_cpu(cpu)) 635 wake_up_idle_cpu(cpu); 636} 637 638static inline bool got_nohz_idle_kick(void) 639{ 640 int cpu = smp_processor_id(); 641 642 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 643 return false; 644 645 if (idle_cpu(cpu) && !need_resched()) 646 return true; 647 648 /* 649 * We can't run Idle Load Balance on this CPU for this time so we 650 * cancel it and clear NOHZ_BALANCE_KICK 651 */ 652 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 653 return false; 654} 655 656#else /* CONFIG_NO_HZ_COMMON */ 657 658static inline bool got_nohz_idle_kick(void) 659{ 660 return false; 661} 662 663#endif /* CONFIG_NO_HZ_COMMON */ 664 665#ifdef CONFIG_NO_HZ_FULL 666bool sched_can_stop_tick(void) 667{ 668 struct rq *rq; 669 670 rq = this_rq(); 671 672 /* Make sure rq->nr_running update is visible after the IPI */ 673 smp_rmb(); 674 675 /* More than one running task need preemption */ 676 if (rq->nr_running > 1) 677 return false; 678 679 return true; 680} 681#endif /* CONFIG_NO_HZ_FULL */ 682 683void sched_avg_update(struct rq *rq) 684{ 685 s64 period = sched_avg_period(); 686 687 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 688 /* 689 * Inline assembly required to prevent the compiler 690 * optimising this loop into a divmod call. 691 * See __iter_div_u64_rem() for another example of this. 692 */ 693 asm("" : "+rm" (rq->age_stamp)); 694 rq->age_stamp += period; 695 rq->rt_avg /= 2; 696 } 697} 698 699#endif /* CONFIG_SMP */ 700 701#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 702 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 703/* 704 * Iterate task_group tree rooted at *from, calling @down when first entering a 705 * node and @up when leaving it for the final time. 706 * 707 * Caller must hold rcu_lock or sufficient equivalent. 708 */ 709int walk_tg_tree_from(struct task_group *from, 710 tg_visitor down, tg_visitor up, void *data) 711{ 712 struct task_group *parent, *child; 713 int ret; 714 715 parent = from; 716 717down: 718 ret = (*down)(parent, data); 719 if (ret) 720 goto out; 721 list_for_each_entry_rcu(child, &parent->children, siblings) { 722 parent = child; 723 goto down; 724 725up: 726 continue; 727 } 728 ret = (*up)(parent, data); 729 if (ret || parent == from) 730 goto out; 731 732 child = parent; 733 parent = parent->parent; 734 if (parent) 735 goto up; 736out: 737 return ret; 738} 739 740int tg_nop(struct task_group *tg, void *data) 741{ 742 return 0; 743} 744#endif 745 746static void set_load_weight(struct task_struct *p) 747{ 748 int prio = p->static_prio - MAX_RT_PRIO; 749 struct load_weight *load = &p->se.load; 750 751 /* 752 * SCHED_IDLE tasks get minimal weight: 753 */ 754 if (p->policy == SCHED_IDLE) { 755 load->weight = scale_load(WEIGHT_IDLEPRIO); 756 load->inv_weight = WMULT_IDLEPRIO; 757 return; 758 } 759 760 load->weight = scale_load(prio_to_weight[prio]); 761 load->inv_weight = prio_to_wmult[prio]; 762} 763 764static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 765{ 766 update_rq_clock(rq); 767 sched_info_queued(rq, p); 768 p->sched_class->enqueue_task(rq, p, flags); 769} 770 771static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 772{ 773 update_rq_clock(rq); 774 sched_info_dequeued(rq, p); 775 p->sched_class->dequeue_task(rq, p, flags); 776} 777 778void activate_task(struct rq *rq, struct task_struct *p, int flags) 779{ 780 if (task_contributes_to_load(p)) 781 rq->nr_uninterruptible--; 782 783 enqueue_task(rq, p, flags); 784} 785 786void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 787{ 788 if (task_contributes_to_load(p)) 789 rq->nr_uninterruptible++; 790 791 dequeue_task(rq, p, flags); 792} 793 794static void update_rq_clock_task(struct rq *rq, s64 delta) 795{ 796/* 797 * In theory, the compile should just see 0 here, and optimize out the call 798 * to sched_rt_avg_update. But I don't trust it... 799 */ 800#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 801 s64 steal = 0, irq_delta = 0; 802#endif 803#ifdef CONFIG_IRQ_TIME_ACCOUNTING 804 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 805 806 /* 807 * Since irq_time is only updated on {soft,}irq_exit, we might run into 808 * this case when a previous update_rq_clock() happened inside a 809 * {soft,}irq region. 810 * 811 * When this happens, we stop ->clock_task and only update the 812 * prev_irq_time stamp to account for the part that fit, so that a next 813 * update will consume the rest. This ensures ->clock_task is 814 * monotonic. 815 * 816 * It does however cause some slight miss-attribution of {soft,}irq 817 * time, a more accurate solution would be to update the irq_time using 818 * the current rq->clock timestamp, except that would require using 819 * atomic ops. 820 */ 821 if (irq_delta > delta) 822 irq_delta = delta; 823 824 rq->prev_irq_time += irq_delta; 825 delta -= irq_delta; 826#endif 827#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 828 if (static_key_false((¶virt_steal_rq_enabled))) { 829 steal = paravirt_steal_clock(cpu_of(rq)); 830 steal -= rq->prev_steal_time_rq; 831 832 if (unlikely(steal > delta)) 833 steal = delta; 834 835 rq->prev_steal_time_rq += steal; 836 delta -= steal; 837 } 838#endif 839 840 rq->clock_task += delta; 841 842#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 843 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 844 sched_rt_avg_update(rq, irq_delta + steal); 845#endif 846} 847 848void sched_set_stop_task(int cpu, struct task_struct *stop) 849{ 850 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 851 struct task_struct *old_stop = cpu_rq(cpu)->stop; 852 853 if (stop) { 854 /* 855 * Make it appear like a SCHED_FIFO task, its something 856 * userspace knows about and won't get confused about. 857 * 858 * Also, it will make PI more or less work without too 859 * much confusion -- but then, stop work should not 860 * rely on PI working anyway. 861 */ 862 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 863 864 stop->sched_class = &stop_sched_class; 865 } 866 867 cpu_rq(cpu)->stop = stop; 868 869 if (old_stop) { 870 /* 871 * Reset it back to a normal scheduling class so that 872 * it can die in pieces. 873 */ 874 old_stop->sched_class = &rt_sched_class; 875 } 876} 877 878/* 879 * __normal_prio - return the priority that is based on the static prio 880 */ 881static inline int __normal_prio(struct task_struct *p) 882{ 883 return p->static_prio; 884} 885 886/* 887 * Calculate the expected normal priority: i.e. priority 888 * without taking RT-inheritance into account. Might be 889 * boosted by interactivity modifiers. Changes upon fork, 890 * setprio syscalls, and whenever the interactivity 891 * estimator recalculates. 892 */ 893static inline int normal_prio(struct task_struct *p) 894{ 895 int prio; 896 897 if (task_has_dl_policy(p)) 898 prio = MAX_DL_PRIO-1; 899 else if (task_has_rt_policy(p)) 900 prio = MAX_RT_PRIO-1 - p->rt_priority; 901 else 902 prio = __normal_prio(p); 903 return prio; 904} 905 906/* 907 * Calculate the current priority, i.e. the priority 908 * taken into account by the scheduler. This value might 909 * be boosted by RT tasks, or might be boosted by 910 * interactivity modifiers. Will be RT if the task got 911 * RT-boosted. If not then it returns p->normal_prio. 912 */ 913static int effective_prio(struct task_struct *p) 914{ 915 p->normal_prio = normal_prio(p); 916 /* 917 * If we are RT tasks or we were boosted to RT priority, 918 * keep the priority unchanged. Otherwise, update priority 919 * to the normal priority: 920 */ 921 if (!rt_prio(p->prio)) 922 return p->normal_prio; 923 return p->prio; 924} 925 926/** 927 * task_curr - is this task currently executing on a CPU? 928 * @p: the task in question. 929 * 930 * Return: 1 if the task is currently executing. 0 otherwise. 931 */ 932inline int task_curr(const struct task_struct *p) 933{ 934 return cpu_curr(task_cpu(p)) == p; 935} 936 937static inline void check_class_changed(struct rq *rq, struct task_struct *p, 938 const struct sched_class *prev_class, 939 int oldprio) 940{ 941 if (prev_class != p->sched_class) { 942 if (prev_class->switched_from) 943 prev_class->switched_from(rq, p); 944 p->sched_class->switched_to(rq, p); 945 } else if (oldprio != p->prio || dl_task(p)) 946 p->sched_class->prio_changed(rq, p, oldprio); 947} 948 949void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 950{ 951 const struct sched_class *class; 952 953 if (p->sched_class == rq->curr->sched_class) { 954 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 955 } else { 956 for_each_class(class) { 957 if (class == rq->curr->sched_class) 958 break; 959 if (class == p->sched_class) { 960 resched_task(rq->curr); 961 break; 962 } 963 } 964 } 965 966 /* 967 * A queue event has occurred, and we're going to schedule. In 968 * this case, we can save a useless back to back clock update. 969 */ 970 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 971 rq->skip_clock_update = 1; 972} 973 974#ifdef CONFIG_SMP 975void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 976{ 977#ifdef CONFIG_SCHED_DEBUG 978 /* 979 * We should never call set_task_cpu() on a blocked task, 980 * ttwu() will sort out the placement. 981 */ 982 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 983 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 984 985#ifdef CONFIG_LOCKDEP 986 /* 987 * The caller should hold either p->pi_lock or rq->lock, when changing 988 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 989 * 990 * sched_move_task() holds both and thus holding either pins the cgroup, 991 * see task_group(). 992 * 993 * Furthermore, all task_rq users should acquire both locks, see 994 * task_rq_lock(). 995 */ 996 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 997 lockdep_is_held(&task_rq(p)->lock))); 998#endif 999#endif 1000 1001 trace_sched_migrate_task(p, new_cpu); 1002 1003 if (task_cpu(p) != new_cpu) { 1004 if (p->sched_class->migrate_task_rq) 1005 p->sched_class->migrate_task_rq(p, new_cpu); 1006 p->se.nr_migrations++; 1007 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1008 } 1009 1010 __set_task_cpu(p, new_cpu); 1011} 1012 1013static void __migrate_swap_task(struct task_struct *p, int cpu) 1014{ 1015 if (p->on_rq) { 1016 struct rq *src_rq, *dst_rq; 1017 1018 src_rq = task_rq(p); 1019 dst_rq = cpu_rq(cpu); 1020 1021 deactivate_task(src_rq, p, 0); 1022 set_task_cpu(p, cpu); 1023 activate_task(dst_rq, p, 0); 1024 check_preempt_curr(dst_rq, p, 0); 1025 } else { 1026 /* 1027 * Task isn't running anymore; make it appear like we migrated 1028 * it before it went to sleep. This means on wakeup we make the 1029 * previous cpu our targer instead of where it really is. 1030 */ 1031 p->wake_cpu = cpu; 1032 } 1033} 1034 1035struct migration_swap_arg { 1036 struct task_struct *src_task, *dst_task; 1037 int src_cpu, dst_cpu; 1038}; 1039 1040static int migrate_swap_stop(void *data) 1041{ 1042 struct migration_swap_arg *arg = data; 1043 struct rq *src_rq, *dst_rq; 1044 int ret = -EAGAIN; 1045 1046 src_rq = cpu_rq(arg->src_cpu); 1047 dst_rq = cpu_rq(arg->dst_cpu); 1048 1049 double_raw_lock(&arg->src_task->pi_lock, 1050 &arg->dst_task->pi_lock); 1051 double_rq_lock(src_rq, dst_rq); 1052 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1053 goto unlock; 1054 1055 if (task_cpu(arg->src_task) != arg->src_cpu) 1056 goto unlock; 1057 1058 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1059 goto unlock; 1060 1061 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1062 goto unlock; 1063 1064 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1065 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1066 1067 ret = 0; 1068 1069unlock: 1070 double_rq_unlock(src_rq, dst_rq); 1071 raw_spin_unlock(&arg->dst_task->pi_lock); 1072 raw_spin_unlock(&arg->src_task->pi_lock); 1073 1074 return ret; 1075} 1076 1077/* 1078 * Cross migrate two tasks 1079 */ 1080int migrate_swap(struct task_struct *cur, struct task_struct *p) 1081{ 1082 struct migration_swap_arg arg; 1083 int ret = -EINVAL; 1084 1085 arg = (struct migration_swap_arg){ 1086 .src_task = cur, 1087 .src_cpu = task_cpu(cur), 1088 .dst_task = p, 1089 .dst_cpu = task_cpu(p), 1090 }; 1091 1092 if (arg.src_cpu == arg.dst_cpu) 1093 goto out; 1094 1095 /* 1096 * These three tests are all lockless; this is OK since all of them 1097 * will be re-checked with proper locks held further down the line. 1098 */ 1099 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1100 goto out; 1101 1102 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1103 goto out; 1104 1105 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1106 goto out; 1107 1108 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1109 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1110 1111out: 1112 return ret; 1113} 1114 1115struct migration_arg { 1116 struct task_struct *task; 1117 int dest_cpu; 1118}; 1119 1120static int migration_cpu_stop(void *data); 1121 1122/* 1123 * wait_task_inactive - wait for a thread to unschedule. 1124 * 1125 * If @match_state is nonzero, it's the @p->state value just checked and 1126 * not expected to change. If it changes, i.e. @p might have woken up, 1127 * then return zero. When we succeed in waiting for @p to be off its CPU, 1128 * we return a positive number (its total switch count). If a second call 1129 * a short while later returns the same number, the caller can be sure that 1130 * @p has remained unscheduled the whole time. 1131 * 1132 * The caller must ensure that the task *will* unschedule sometime soon, 1133 * else this function might spin for a *long* time. This function can't 1134 * be called with interrupts off, or it may introduce deadlock with 1135 * smp_call_function() if an IPI is sent by the same process we are 1136 * waiting to become inactive. 1137 */ 1138unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1139{ 1140 unsigned long flags; 1141 int running, on_rq; 1142 unsigned long ncsw; 1143 struct rq *rq; 1144 1145 for (;;) { 1146 /* 1147 * We do the initial early heuristics without holding 1148 * any task-queue locks at all. We'll only try to get 1149 * the runqueue lock when things look like they will 1150 * work out! 1151 */ 1152 rq = task_rq(p); 1153 1154 /* 1155 * If the task is actively running on another CPU 1156 * still, just relax and busy-wait without holding 1157 * any locks. 1158 * 1159 * NOTE! Since we don't hold any locks, it's not 1160 * even sure that "rq" stays as the right runqueue! 1161 * But we don't care, since "task_running()" will 1162 * return false if the runqueue has changed and p 1163 * is actually now running somewhere else! 1164 */ 1165 while (task_running(rq, p)) { 1166 if (match_state && unlikely(p->state != match_state)) 1167 return 0; 1168 cpu_relax(); 1169 } 1170 1171 /* 1172 * Ok, time to look more closely! We need the rq 1173 * lock now, to be *sure*. If we're wrong, we'll 1174 * just go back and repeat. 1175 */ 1176 rq = task_rq_lock(p, &flags); 1177 trace_sched_wait_task(p); 1178 running = task_running(rq, p); 1179 on_rq = p->on_rq; 1180 ncsw = 0; 1181 if (!match_state || p->state == match_state) 1182 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1183 task_rq_unlock(rq, p, &flags); 1184 1185 /* 1186 * If it changed from the expected state, bail out now. 1187 */ 1188 if (unlikely(!ncsw)) 1189 break; 1190 1191 /* 1192 * Was it really running after all now that we 1193 * checked with the proper locks actually held? 1194 * 1195 * Oops. Go back and try again.. 1196 */ 1197 if (unlikely(running)) { 1198 cpu_relax(); 1199 continue; 1200 } 1201 1202 /* 1203 * It's not enough that it's not actively running, 1204 * it must be off the runqueue _entirely_, and not 1205 * preempted! 1206 * 1207 * So if it was still runnable (but just not actively 1208 * running right now), it's preempted, and we should 1209 * yield - it could be a while. 1210 */ 1211 if (unlikely(on_rq)) { 1212 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1213 1214 set_current_state(TASK_UNINTERRUPTIBLE); 1215 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1216 continue; 1217 } 1218 1219 /* 1220 * Ahh, all good. It wasn't running, and it wasn't 1221 * runnable, which means that it will never become 1222 * running in the future either. We're all done! 1223 */ 1224 break; 1225 } 1226 1227 return ncsw; 1228} 1229 1230/*** 1231 * kick_process - kick a running thread to enter/exit the kernel 1232 * @p: the to-be-kicked thread 1233 * 1234 * Cause a process which is running on another CPU to enter 1235 * kernel-mode, without any delay. (to get signals handled.) 1236 * 1237 * NOTE: this function doesn't have to take the runqueue lock, 1238 * because all it wants to ensure is that the remote task enters 1239 * the kernel. If the IPI races and the task has been migrated 1240 * to another CPU then no harm is done and the purpose has been 1241 * achieved as well. 1242 */ 1243void kick_process(struct task_struct *p) 1244{ 1245 int cpu; 1246 1247 preempt_disable(); 1248 cpu = task_cpu(p); 1249 if ((cpu != smp_processor_id()) && task_curr(p)) 1250 smp_send_reschedule(cpu); 1251 preempt_enable(); 1252} 1253EXPORT_SYMBOL_GPL(kick_process); 1254#endif /* CONFIG_SMP */ 1255 1256#ifdef CONFIG_SMP 1257/* 1258 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1259 */ 1260static int select_fallback_rq(int cpu, struct task_struct *p) 1261{ 1262 int nid = cpu_to_node(cpu); 1263 const struct cpumask *nodemask = NULL; 1264 enum { cpuset, possible, fail } state = cpuset; 1265 int dest_cpu; 1266 1267 /* 1268 * If the node that the cpu is on has been offlined, cpu_to_node() 1269 * will return -1. There is no cpu on the node, and we should 1270 * select the cpu on the other node. 1271 */ 1272 if (nid != -1) { 1273 nodemask = cpumask_of_node(nid); 1274 1275 /* Look for allowed, online CPU in same node. */ 1276 for_each_cpu(dest_cpu, nodemask) { 1277 if (!cpu_online(dest_cpu)) 1278 continue; 1279 if (!cpu_active(dest_cpu)) 1280 continue; 1281 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1282 return dest_cpu; 1283 } 1284 } 1285 1286 for (;;) { 1287 /* Any allowed, online CPU? */ 1288 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1289 if (!cpu_online(dest_cpu)) 1290 continue; 1291 if (!cpu_active(dest_cpu)) 1292 continue; 1293 goto out; 1294 } 1295 1296 switch (state) { 1297 case cpuset: 1298 /* No more Mr. Nice Guy. */ 1299 cpuset_cpus_allowed_fallback(p); 1300 state = possible; 1301 break; 1302 1303 case possible: 1304 do_set_cpus_allowed(p, cpu_possible_mask); 1305 state = fail; 1306 break; 1307 1308 case fail: 1309 BUG(); 1310 break; 1311 } 1312 } 1313 1314out: 1315 if (state != cpuset) { 1316 /* 1317 * Don't tell them about moving exiting tasks or 1318 * kernel threads (both mm NULL), since they never 1319 * leave kernel. 1320 */ 1321 if (p->mm && printk_ratelimit()) { 1322 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1323 task_pid_nr(p), p->comm, cpu); 1324 } 1325 } 1326 1327 return dest_cpu; 1328} 1329 1330/* 1331 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1332 */ 1333static inline 1334int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1335{ 1336 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1337 1338 /* 1339 * In order not to call set_task_cpu() on a blocking task we need 1340 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1341 * cpu. 1342 * 1343 * Since this is common to all placement strategies, this lives here. 1344 * 1345 * [ this allows ->select_task() to simply return task_cpu(p) and 1346 * not worry about this generic constraint ] 1347 */ 1348 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1349 !cpu_online(cpu))) 1350 cpu = select_fallback_rq(task_cpu(p), p); 1351 1352 return cpu; 1353} 1354 1355static void update_avg(u64 *avg, u64 sample) 1356{ 1357 s64 diff = sample - *avg; 1358 *avg += diff >> 3; 1359} 1360#endif 1361 1362static void 1363ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1364{ 1365#ifdef CONFIG_SCHEDSTATS 1366 struct rq *rq = this_rq(); 1367 1368#ifdef CONFIG_SMP 1369 int this_cpu = smp_processor_id(); 1370 1371 if (cpu == this_cpu) { 1372 schedstat_inc(rq, ttwu_local); 1373 schedstat_inc(p, se.statistics.nr_wakeups_local); 1374 } else { 1375 struct sched_domain *sd; 1376 1377 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1378 rcu_read_lock(); 1379 for_each_domain(this_cpu, sd) { 1380 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1381 schedstat_inc(sd, ttwu_wake_remote); 1382 break; 1383 } 1384 } 1385 rcu_read_unlock(); 1386 } 1387 1388 if (wake_flags & WF_MIGRATED) 1389 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1390 1391#endif /* CONFIG_SMP */ 1392 1393 schedstat_inc(rq, ttwu_count); 1394 schedstat_inc(p, se.statistics.nr_wakeups); 1395 1396 if (wake_flags & WF_SYNC) 1397 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1398 1399#endif /* CONFIG_SCHEDSTATS */ 1400} 1401 1402static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1403{ 1404 activate_task(rq, p, en_flags); 1405 p->on_rq = 1; 1406 1407 /* if a worker is waking up, notify workqueue */ 1408 if (p->flags & PF_WQ_WORKER) 1409 wq_worker_waking_up(p, cpu_of(rq)); 1410} 1411 1412/* 1413 * Mark the task runnable and perform wakeup-preemption. 1414 */ 1415static void 1416ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1417{ 1418 check_preempt_curr(rq, p, wake_flags); 1419 trace_sched_wakeup(p, true); 1420 1421 p->state = TASK_RUNNING; 1422#ifdef CONFIG_SMP 1423 if (p->sched_class->task_woken) 1424 p->sched_class->task_woken(rq, p); 1425 1426 if (rq->idle_stamp) { 1427 u64 delta = rq_clock(rq) - rq->idle_stamp; 1428 u64 max = 2*rq->max_idle_balance_cost; 1429 1430 update_avg(&rq->avg_idle, delta); 1431 1432 if (rq->avg_idle > max) 1433 rq->avg_idle = max; 1434 1435 rq->idle_stamp = 0; 1436 } 1437#endif 1438} 1439 1440static void 1441ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1442{ 1443#ifdef CONFIG_SMP 1444 if (p->sched_contributes_to_load) 1445 rq->nr_uninterruptible--; 1446#endif 1447 1448 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1449 ttwu_do_wakeup(rq, p, wake_flags); 1450} 1451 1452/* 1453 * Called in case the task @p isn't fully descheduled from its runqueue, 1454 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1455 * since all we need to do is flip p->state to TASK_RUNNING, since 1456 * the task is still ->on_rq. 1457 */ 1458static int ttwu_remote(struct task_struct *p, int wake_flags) 1459{ 1460 struct rq *rq; 1461 int ret = 0; 1462 1463 rq = __task_rq_lock(p); 1464 if (p->on_rq) { 1465 /* check_preempt_curr() may use rq clock */ 1466 update_rq_clock(rq); 1467 ttwu_do_wakeup(rq, p, wake_flags); 1468 ret = 1; 1469 } 1470 __task_rq_unlock(rq); 1471 1472 return ret; 1473} 1474 1475#ifdef CONFIG_SMP 1476static void sched_ttwu_pending(void) 1477{ 1478 struct rq *rq = this_rq(); 1479 struct llist_node *llist = llist_del_all(&rq->wake_list); 1480 struct task_struct *p; 1481 1482 raw_spin_lock(&rq->lock); 1483 1484 while (llist) { 1485 p = llist_entry(llist, struct task_struct, wake_entry); 1486 llist = llist_next(llist); 1487 ttwu_do_activate(rq, p, 0); 1488 } 1489 1490 raw_spin_unlock(&rq->lock); 1491} 1492 1493void scheduler_ipi(void) 1494{ 1495 /* 1496 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1497 * TIF_NEED_RESCHED remotely (for the first time) will also send 1498 * this IPI. 1499 */ 1500 preempt_fold_need_resched(); 1501 1502 if (llist_empty(&this_rq()->wake_list) 1503 && !tick_nohz_full_cpu(smp_processor_id()) 1504 && !got_nohz_idle_kick()) 1505 return; 1506 1507 /* 1508 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1509 * traditionally all their work was done from the interrupt return 1510 * path. Now that we actually do some work, we need to make sure 1511 * we do call them. 1512 * 1513 * Some archs already do call them, luckily irq_enter/exit nest 1514 * properly. 1515 * 1516 * Arguably we should visit all archs and update all handlers, 1517 * however a fair share of IPIs are still resched only so this would 1518 * somewhat pessimize the simple resched case. 1519 */ 1520 irq_enter(); 1521 tick_nohz_full_check(); 1522 sched_ttwu_pending(); 1523 1524 /* 1525 * Check if someone kicked us for doing the nohz idle load balance. 1526 */ 1527 if (unlikely(got_nohz_idle_kick())) { 1528 this_rq()->idle_balance = 1; 1529 raise_softirq_irqoff(SCHED_SOFTIRQ); 1530 } 1531 irq_exit(); 1532} 1533 1534static void ttwu_queue_remote(struct task_struct *p, int cpu) 1535{ 1536 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1537 smp_send_reschedule(cpu); 1538} 1539 1540bool cpus_share_cache(int this_cpu, int that_cpu) 1541{ 1542 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1543} 1544#endif /* CONFIG_SMP */ 1545 1546static void ttwu_queue(struct task_struct *p, int cpu) 1547{ 1548 struct rq *rq = cpu_rq(cpu); 1549 1550#if defined(CONFIG_SMP) 1551 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1552 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1553 ttwu_queue_remote(p, cpu); 1554 return; 1555 } 1556#endif 1557 1558 raw_spin_lock(&rq->lock); 1559 ttwu_do_activate(rq, p, 0); 1560 raw_spin_unlock(&rq->lock); 1561} 1562 1563/** 1564 * try_to_wake_up - wake up a thread 1565 * @p: the thread to be awakened 1566 * @state: the mask of task states that can be woken 1567 * @wake_flags: wake modifier flags (WF_*) 1568 * 1569 * Put it on the run-queue if it's not already there. The "current" 1570 * thread is always on the run-queue (except when the actual 1571 * re-schedule is in progress), and as such you're allowed to do 1572 * the simpler "current->state = TASK_RUNNING" to mark yourself 1573 * runnable without the overhead of this. 1574 * 1575 * Return: %true if @p was woken up, %false if it was already running. 1576 * or @state didn't match @p's state. 1577 */ 1578static int 1579try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1580{ 1581 unsigned long flags; 1582 int cpu, success = 0; 1583 1584 /* 1585 * If we are going to wake up a thread waiting for CONDITION we 1586 * need to ensure that CONDITION=1 done by the caller can not be 1587 * reordered with p->state check below. This pairs with mb() in 1588 * set_current_state() the waiting thread does. 1589 */ 1590 smp_mb__before_spinlock(); 1591 raw_spin_lock_irqsave(&p->pi_lock, flags); 1592 if (!(p->state & state)) 1593 goto out; 1594 1595 success = 1; /* we're going to change ->state */ 1596 cpu = task_cpu(p); 1597 1598 if (p->on_rq && ttwu_remote(p, wake_flags)) 1599 goto stat; 1600 1601#ifdef CONFIG_SMP 1602 /* 1603 * If the owning (remote) cpu is still in the middle of schedule() with 1604 * this task as prev, wait until its done referencing the task. 1605 */ 1606 while (p->on_cpu) 1607 cpu_relax(); 1608 /* 1609 * Pairs with the smp_wmb() in finish_lock_switch(). 1610 */ 1611 smp_rmb(); 1612 1613 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1614 p->state = TASK_WAKING; 1615 1616 if (p->sched_class->task_waking) 1617 p->sched_class->task_waking(p); 1618 1619 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1620 if (task_cpu(p) != cpu) { 1621 wake_flags |= WF_MIGRATED; 1622 set_task_cpu(p, cpu); 1623 } 1624#endif /* CONFIG_SMP */ 1625 1626 ttwu_queue(p, cpu); 1627stat: 1628 ttwu_stat(p, cpu, wake_flags); 1629out: 1630 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1631 1632 return success; 1633} 1634 1635/** 1636 * try_to_wake_up_local - try to wake up a local task with rq lock held 1637 * @p: the thread to be awakened 1638 * 1639 * Put @p on the run-queue if it's not already there. The caller must 1640 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1641 * the current task. 1642 */ 1643static void try_to_wake_up_local(struct task_struct *p) 1644{ 1645 struct rq *rq = task_rq(p); 1646 1647 if (WARN_ON_ONCE(rq != this_rq()) || 1648 WARN_ON_ONCE(p == current)) 1649 return; 1650 1651 lockdep_assert_held(&rq->lock); 1652 1653 if (!raw_spin_trylock(&p->pi_lock)) { 1654 raw_spin_unlock(&rq->lock); 1655 raw_spin_lock(&p->pi_lock); 1656 raw_spin_lock(&rq->lock); 1657 } 1658 1659 if (!(p->state & TASK_NORMAL)) 1660 goto out; 1661 1662 if (!p->on_rq) 1663 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1664 1665 ttwu_do_wakeup(rq, p, 0); 1666 ttwu_stat(p, smp_processor_id(), 0); 1667out: 1668 raw_spin_unlock(&p->pi_lock); 1669} 1670 1671/** 1672 * wake_up_process - Wake up a specific process 1673 * @p: The process to be woken up. 1674 * 1675 * Attempt to wake up the nominated process and move it to the set of runnable 1676 * processes. 1677 * 1678 * Return: 1 if the process was woken up, 0 if it was already running. 1679 * 1680 * It may be assumed that this function implies a write memory barrier before 1681 * changing the task state if and only if any tasks are woken up. 1682 */ 1683int wake_up_process(struct task_struct *p) 1684{ 1685 WARN_ON(task_is_stopped_or_traced(p)); 1686 return try_to_wake_up(p, TASK_NORMAL, 0); 1687} 1688EXPORT_SYMBOL(wake_up_process); 1689 1690int wake_up_state(struct task_struct *p, unsigned int state) 1691{ 1692 return try_to_wake_up(p, state, 0); 1693} 1694 1695/* 1696 * Perform scheduler related setup for a newly forked process p. 1697 * p is forked by current. 1698 * 1699 * __sched_fork() is basic setup used by init_idle() too: 1700 */ 1701static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1702{ 1703 p->on_rq = 0; 1704 1705 p->se.on_rq = 0; 1706 p->se.exec_start = 0; 1707 p->se.sum_exec_runtime = 0; 1708 p->se.prev_sum_exec_runtime = 0; 1709 p->se.nr_migrations = 0; 1710 p->se.vruntime = 0; 1711 INIT_LIST_HEAD(&p->se.group_node); 1712 1713#ifdef CONFIG_SCHEDSTATS 1714 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1715#endif 1716 1717 RB_CLEAR_NODE(&p->dl.rb_node); 1718 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1719 p->dl.dl_runtime = p->dl.runtime = 0; 1720 p->dl.dl_deadline = p->dl.deadline = 0; 1721 p->dl.dl_period = 0; 1722 p->dl.flags = 0; 1723 1724 INIT_LIST_HEAD(&p->rt.run_list); 1725 1726#ifdef CONFIG_PREEMPT_NOTIFIERS 1727 INIT_HLIST_HEAD(&p->preempt_notifiers); 1728#endif 1729 1730#ifdef CONFIG_NUMA_BALANCING 1731 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1732 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1733 p->mm->numa_scan_seq = 0; 1734 } 1735 1736 if (clone_flags & CLONE_VM) 1737 p->numa_preferred_nid = current->numa_preferred_nid; 1738 else 1739 p->numa_preferred_nid = -1; 1740 1741 p->node_stamp = 0ULL; 1742 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1743 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1744 p->numa_work.next = &p->numa_work; 1745 p->numa_faults_memory = NULL; 1746 p->numa_faults_buffer_memory = NULL; 1747 p->last_task_numa_placement = 0; 1748 p->last_sum_exec_runtime = 0; 1749 1750 INIT_LIST_HEAD(&p->numa_entry); 1751 p->numa_group = NULL; 1752#endif /* CONFIG_NUMA_BALANCING */ 1753} 1754 1755#ifdef CONFIG_NUMA_BALANCING 1756#ifdef CONFIG_SCHED_DEBUG 1757void set_numabalancing_state(bool enabled) 1758{ 1759 if (enabled) 1760 sched_feat_set("NUMA"); 1761 else 1762 sched_feat_set("NO_NUMA"); 1763} 1764#else 1765__read_mostly bool numabalancing_enabled; 1766 1767void set_numabalancing_state(bool enabled) 1768{ 1769 numabalancing_enabled = enabled; 1770} 1771#endif /* CONFIG_SCHED_DEBUG */ 1772 1773#ifdef CONFIG_PROC_SYSCTL 1774int sysctl_numa_balancing(struct ctl_table *table, int write, 1775 void __user *buffer, size_t *lenp, loff_t *ppos) 1776{ 1777 struct ctl_table t; 1778 int err; 1779 int state = numabalancing_enabled; 1780 1781 if (write && !capable(CAP_SYS_ADMIN)) 1782 return -EPERM; 1783 1784 t = *table; 1785 t.data = &state; 1786 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1787 if (err < 0) 1788 return err; 1789 if (write) 1790 set_numabalancing_state(state); 1791 return err; 1792} 1793#endif 1794#endif 1795 1796/* 1797 * fork()/clone()-time setup: 1798 */ 1799int sched_fork(unsigned long clone_flags, struct task_struct *p) 1800{ 1801 unsigned long flags; 1802 int cpu = get_cpu(); 1803 1804 __sched_fork(clone_flags, p); 1805 /* 1806 * We mark the process as running here. This guarantees that 1807 * nobody will actually run it, and a signal or other external 1808 * event cannot wake it up and insert it on the runqueue either. 1809 */ 1810 p->state = TASK_RUNNING; 1811 1812 /* 1813 * Make sure we do not leak PI boosting priority to the child. 1814 */ 1815 p->prio = current->normal_prio; 1816 1817 /* 1818 * Revert to default priority/policy on fork if requested. 1819 */ 1820 if (unlikely(p->sched_reset_on_fork)) { 1821 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1822 p->policy = SCHED_NORMAL; 1823 p->static_prio = NICE_TO_PRIO(0); 1824 p->rt_priority = 0; 1825 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1826 p->static_prio = NICE_TO_PRIO(0); 1827 1828 p->prio = p->normal_prio = __normal_prio(p); 1829 set_load_weight(p); 1830 1831 /* 1832 * We don't need the reset flag anymore after the fork. It has 1833 * fulfilled its duty: 1834 */ 1835 p->sched_reset_on_fork = 0; 1836 } 1837 1838 if (dl_prio(p->prio)) { 1839 put_cpu(); 1840 return -EAGAIN; 1841 } else if (rt_prio(p->prio)) { 1842 p->sched_class = &rt_sched_class; 1843 } else { 1844 p->sched_class = &fair_sched_class; 1845 } 1846 1847 if (p->sched_class->task_fork) 1848 p->sched_class->task_fork(p); 1849 1850 /* 1851 * The child is not yet in the pid-hash so no cgroup attach races, 1852 * and the cgroup is pinned to this child due to cgroup_fork() 1853 * is ran before sched_fork(). 1854 * 1855 * Silence PROVE_RCU. 1856 */ 1857 raw_spin_lock_irqsave(&p->pi_lock, flags); 1858 set_task_cpu(p, cpu); 1859 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1860 1861#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1862 if (likely(sched_info_on())) 1863 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1864#endif 1865#if defined(CONFIG_SMP) 1866 p->on_cpu = 0; 1867#endif 1868 init_task_preempt_count(p); 1869#ifdef CONFIG_SMP 1870 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1871 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1872#endif 1873 1874 put_cpu(); 1875 return 0; 1876} 1877 1878unsigned long to_ratio(u64 period, u64 runtime) 1879{ 1880 if (runtime == RUNTIME_INF) 1881 return 1ULL << 20; 1882 1883 /* 1884 * Doing this here saves a lot of checks in all 1885 * the calling paths, and returning zero seems 1886 * safe for them anyway. 1887 */ 1888 if (period == 0) 1889 return 0; 1890 1891 return div64_u64(runtime << 20, period); 1892} 1893 1894#ifdef CONFIG_SMP 1895inline struct dl_bw *dl_bw_of(int i) 1896{ 1897 return &cpu_rq(i)->rd->dl_bw; 1898} 1899 1900static inline int dl_bw_cpus(int i) 1901{ 1902 struct root_domain *rd = cpu_rq(i)->rd; 1903 int cpus = 0; 1904 1905 for_each_cpu_and(i, rd->span, cpu_active_mask) 1906 cpus++; 1907 1908 return cpus; 1909} 1910#else 1911inline struct dl_bw *dl_bw_of(int i) 1912{ 1913 return &cpu_rq(i)->dl.dl_bw; 1914} 1915 1916static inline int dl_bw_cpus(int i) 1917{ 1918 return 1; 1919} 1920#endif 1921 1922static inline 1923void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1924{ 1925 dl_b->total_bw -= tsk_bw; 1926} 1927 1928static inline 1929void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1930{ 1931 dl_b->total_bw += tsk_bw; 1932} 1933 1934static inline 1935bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1936{ 1937 return dl_b->bw != -1 && 1938 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1939} 1940 1941/* 1942 * We must be sure that accepting a new task (or allowing changing the 1943 * parameters of an existing one) is consistent with the bandwidth 1944 * constraints. If yes, this function also accordingly updates the currently 1945 * allocated bandwidth to reflect the new situation. 1946 * 1947 * This function is called while holding p's rq->lock. 1948 */ 1949static int dl_overflow(struct task_struct *p, int policy, 1950 const struct sched_attr *attr) 1951{ 1952 1953 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1954 u64 period = attr->sched_period ?: attr->sched_deadline; 1955 u64 runtime = attr->sched_runtime; 1956 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1957 int cpus, err = -1; 1958 1959 if (new_bw == p->dl.dl_bw) 1960 return 0; 1961 1962 /* 1963 * Either if a task, enters, leave, or stays -deadline but changes 1964 * its parameters, we may need to update accordingly the total 1965 * allocated bandwidth of the container. 1966 */ 1967 raw_spin_lock(&dl_b->lock); 1968 cpus = dl_bw_cpus(task_cpu(p)); 1969 if (dl_policy(policy) && !task_has_dl_policy(p) && 1970 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1971 __dl_add(dl_b, new_bw); 1972 err = 0; 1973 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1974 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1975 __dl_clear(dl_b, p->dl.dl_bw); 1976 __dl_add(dl_b, new_bw); 1977 err = 0; 1978 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1979 __dl_clear(dl_b, p->dl.dl_bw); 1980 err = 0; 1981 } 1982 raw_spin_unlock(&dl_b->lock); 1983 1984 return err; 1985} 1986 1987extern void init_dl_bw(struct dl_bw *dl_b); 1988 1989/* 1990 * wake_up_new_task - wake up a newly created task for the first time. 1991 * 1992 * This function will do some initial scheduler statistics housekeeping 1993 * that must be done for every newly created context, then puts the task 1994 * on the runqueue and wakes it. 1995 */ 1996void wake_up_new_task(struct task_struct *p) 1997{ 1998 unsigned long flags; 1999 struct rq *rq; 2000 2001 raw_spin_lock_irqsave(&p->pi_lock, flags); 2002#ifdef CONFIG_SMP 2003 /* 2004 * Fork balancing, do it here and not earlier because: 2005 * - cpus_allowed can change in the fork path 2006 * - any previously selected cpu might disappear through hotplug 2007 */ 2008 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2009#endif 2010 2011 /* Initialize new task's runnable average */ 2012 init_task_runnable_average(p); 2013 rq = __task_rq_lock(p); 2014 activate_task(rq, p, 0); 2015 p->on_rq = 1; 2016 trace_sched_wakeup_new(p, true); 2017 check_preempt_curr(rq, p, WF_FORK); 2018#ifdef CONFIG_SMP 2019 if (p->sched_class->task_woken) 2020 p->sched_class->task_woken(rq, p); 2021#endif 2022 task_rq_unlock(rq, p, &flags); 2023} 2024 2025#ifdef CONFIG_PREEMPT_NOTIFIERS 2026 2027/** 2028 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2029 * @notifier: notifier struct to register 2030 */ 2031void preempt_notifier_register(struct preempt_notifier *notifier) 2032{ 2033 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2034} 2035EXPORT_SYMBOL_GPL(preempt_notifier_register); 2036 2037/** 2038 * preempt_notifier_unregister - no longer interested in preemption notifications 2039 * @notifier: notifier struct to unregister 2040 * 2041 * This is safe to call from within a preemption notifier. 2042 */ 2043void preempt_notifier_unregister(struct preempt_notifier *notifier) 2044{ 2045 hlist_del(¬ifier->link); 2046} 2047EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2048 2049static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2050{ 2051 struct preempt_notifier *notifier; 2052 2053 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2054 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2055} 2056 2057static void 2058fire_sched_out_preempt_notifiers(struct task_struct *curr, 2059 struct task_struct *next) 2060{ 2061 struct preempt_notifier *notifier; 2062 2063 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2064 notifier->ops->sched_out(notifier, next); 2065} 2066 2067#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2068 2069static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2070{ 2071} 2072 2073static void 2074fire_sched_out_preempt_notifiers(struct task_struct *curr, 2075 struct task_struct *next) 2076{ 2077} 2078 2079#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2080 2081/** 2082 * prepare_task_switch - prepare to switch tasks 2083 * @rq: the runqueue preparing to switch 2084 * @prev: the current task that is being switched out 2085 * @next: the task we are going to switch to. 2086 * 2087 * This is called with the rq lock held and interrupts off. It must 2088 * be paired with a subsequent finish_task_switch after the context 2089 * switch. 2090 * 2091 * prepare_task_switch sets up locking and calls architecture specific 2092 * hooks. 2093 */ 2094static inline void 2095prepare_task_switch(struct rq *rq, struct task_struct *prev, 2096 struct task_struct *next) 2097{ 2098 trace_sched_switch(prev, next); 2099 sched_info_switch(rq, prev, next); 2100 perf_event_task_sched_out(prev, next); 2101 fire_sched_out_preempt_notifiers(prev, next); 2102 prepare_lock_switch(rq, next); 2103 prepare_arch_switch(next); 2104} 2105 2106/** 2107 * finish_task_switch - clean up after a task-switch 2108 * @rq: runqueue associated with task-switch 2109 * @prev: the thread we just switched away from. 2110 * 2111 * finish_task_switch must be called after the context switch, paired 2112 * with a prepare_task_switch call before the context switch. 2113 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2114 * and do any other architecture-specific cleanup actions. 2115 * 2116 * Note that we may have delayed dropping an mm in context_switch(). If 2117 * so, we finish that here outside of the runqueue lock. (Doing it 2118 * with the lock held can cause deadlocks; see schedule() for 2119 * details.) 2120 */ 2121static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2122 __releases(rq->lock) 2123{ 2124 struct mm_struct *mm = rq->prev_mm; 2125 long prev_state; 2126 2127 rq->prev_mm = NULL; 2128 2129 /* 2130 * A task struct has one reference for the use as "current". 2131 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2132 * schedule one last time. The schedule call will never return, and 2133 * the scheduled task must drop that reference. 2134 * The test for TASK_DEAD must occur while the runqueue locks are 2135 * still held, otherwise prev could be scheduled on another cpu, die 2136 * there before we look at prev->state, and then the reference would 2137 * be dropped twice. 2138 * Manfred Spraul <manfred@colorfullife.com> 2139 */ 2140 prev_state = prev->state; 2141 vtime_task_switch(prev); 2142 finish_arch_switch(prev); 2143 perf_event_task_sched_in(prev, current); 2144 finish_lock_switch(rq, prev); 2145 finish_arch_post_lock_switch(); 2146 2147 fire_sched_in_preempt_notifiers(current); 2148 if (mm) 2149 mmdrop(mm); 2150 if (unlikely(prev_state == TASK_DEAD)) { 2151 if (prev->sched_class->task_dead) 2152 prev->sched_class->task_dead(prev); 2153 2154 /* 2155 * Remove function-return probe instances associated with this 2156 * task and put them back on the free list. 2157 */ 2158 kprobe_flush_task(prev); 2159 put_task_struct(prev); 2160 } 2161 2162 tick_nohz_task_switch(current); 2163} 2164 2165#ifdef CONFIG_SMP 2166 2167/* rq->lock is NOT held, but preemption is disabled */ 2168static inline void post_schedule(struct rq *rq) 2169{ 2170 if (rq->post_schedule) { 2171 unsigned long flags; 2172 2173 raw_spin_lock_irqsave(&rq->lock, flags); 2174 if (rq->curr->sched_class->post_schedule) 2175 rq->curr->sched_class->post_schedule(rq); 2176 raw_spin_unlock_irqrestore(&rq->lock, flags); 2177 2178 rq->post_schedule = 0; 2179 } 2180} 2181 2182#else 2183 2184static inline void post_schedule(struct rq *rq) 2185{ 2186} 2187 2188#endif 2189 2190/** 2191 * schedule_tail - first thing a freshly forked thread must call. 2192 * @prev: the thread we just switched away from. 2193 */ 2194asmlinkage void schedule_tail(struct task_struct *prev) 2195 __releases(rq->lock) 2196{ 2197 struct rq *rq = this_rq(); 2198 2199 finish_task_switch(rq, prev); 2200 2201 /* 2202 * FIXME: do we need to worry about rq being invalidated by the 2203 * task_switch? 2204 */ 2205 post_schedule(rq); 2206 2207#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2208 /* In this case, finish_task_switch does not reenable preemption */ 2209 preempt_enable(); 2210#endif 2211 if (current->set_child_tid) 2212 put_user(task_pid_vnr(current), current->set_child_tid); 2213} 2214 2215/* 2216 * context_switch - switch to the new MM and the new 2217 * thread's register state. 2218 */ 2219static inline void 2220context_switch(struct rq *rq, struct task_struct *prev, 2221 struct task_struct *next) 2222{ 2223 struct mm_struct *mm, *oldmm; 2224 2225 prepare_task_switch(rq, prev, next); 2226 2227 mm = next->mm; 2228 oldmm = prev->active_mm; 2229 /* 2230 * For paravirt, this is coupled with an exit in switch_to to 2231 * combine the page table reload and the switch backend into 2232 * one hypercall. 2233 */ 2234 arch_start_context_switch(prev); 2235 2236 if (!mm) { 2237 next->active_mm = oldmm; 2238 atomic_inc(&oldmm->mm_count); 2239 enter_lazy_tlb(oldmm, next); 2240 } else 2241 switch_mm(oldmm, mm, next); 2242 2243 if (!prev->mm) { 2244 prev->active_mm = NULL; 2245 rq->prev_mm = oldmm; 2246 } 2247 /* 2248 * Since the runqueue lock will be released by the next 2249 * task (which is an invalid locking op but in the case 2250 * of the scheduler it's an obvious special-case), so we 2251 * do an early lockdep release here: 2252 */ 2253#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2254 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2255#endif 2256 2257 context_tracking_task_switch(prev, next); 2258 /* Here we just switch the register state and the stack. */ 2259 switch_to(prev, next, prev); 2260 2261 barrier(); 2262 /* 2263 * this_rq must be evaluated again because prev may have moved 2264 * CPUs since it called schedule(), thus the 'rq' on its stack 2265 * frame will be invalid. 2266 */ 2267 finish_task_switch(this_rq(), prev); 2268} 2269 2270/* 2271 * nr_running and nr_context_switches: 2272 * 2273 * externally visible scheduler statistics: current number of runnable 2274 * threads, total number of context switches performed since bootup. 2275 */ 2276unsigned long nr_running(void) 2277{ 2278 unsigned long i, sum = 0; 2279 2280 for_each_online_cpu(i) 2281 sum += cpu_rq(i)->nr_running; 2282 2283 return sum; 2284} 2285 2286unsigned long long nr_context_switches(void) 2287{ 2288 int i; 2289 unsigned long long sum = 0; 2290 2291 for_each_possible_cpu(i) 2292 sum += cpu_rq(i)->nr_switches; 2293 2294 return sum; 2295} 2296 2297unsigned long nr_iowait(void) 2298{ 2299 unsigned long i, sum = 0; 2300 2301 for_each_possible_cpu(i) 2302 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2303 2304 return sum; 2305} 2306 2307unsigned long nr_iowait_cpu(int cpu) 2308{ 2309 struct rq *this = cpu_rq(cpu); 2310 return atomic_read(&this->nr_iowait); 2311} 2312 2313#ifdef CONFIG_SMP 2314 2315/* 2316 * sched_exec - execve() is a valuable balancing opportunity, because at 2317 * this point the task has the smallest effective memory and cache footprint. 2318 */ 2319void sched_exec(void) 2320{ 2321 struct task_struct *p = current; 2322 unsigned long flags; 2323 int dest_cpu; 2324 2325 raw_spin_lock_irqsave(&p->pi_lock, flags); 2326 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2327 if (dest_cpu == smp_processor_id()) 2328 goto unlock; 2329 2330 if (likely(cpu_active(dest_cpu))) { 2331 struct migration_arg arg = { p, dest_cpu }; 2332 2333 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2334 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2335 return; 2336 } 2337unlock: 2338 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2339} 2340 2341#endif 2342 2343DEFINE_PER_CPU(struct kernel_stat, kstat); 2344DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2345 2346EXPORT_PER_CPU_SYMBOL(kstat); 2347EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2348 2349/* 2350 * Return any ns on the sched_clock that have not yet been accounted in 2351 * @p in case that task is currently running. 2352 * 2353 * Called with task_rq_lock() held on @rq. 2354 */ 2355static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2356{ 2357 u64 ns = 0; 2358 2359 if (task_current(rq, p)) { 2360 update_rq_clock(rq); 2361 ns = rq_clock_task(rq) - p->se.exec_start; 2362 if ((s64)ns < 0) 2363 ns = 0; 2364 } 2365 2366 return ns; 2367} 2368 2369unsigned long long task_delta_exec(struct task_struct *p) 2370{ 2371 unsigned long flags; 2372 struct rq *rq; 2373 u64 ns = 0; 2374 2375 rq = task_rq_lock(p, &flags); 2376 ns = do_task_delta_exec(p, rq); 2377 task_rq_unlock(rq, p, &flags); 2378 2379 return ns; 2380} 2381 2382/* 2383 * Return accounted runtime for the task. 2384 * In case the task is currently running, return the runtime plus current's 2385 * pending runtime that have not been accounted yet. 2386 */ 2387unsigned long long task_sched_runtime(struct task_struct *p) 2388{ 2389 unsigned long flags; 2390 struct rq *rq; 2391 u64 ns = 0; 2392 2393#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2394 /* 2395 * 64-bit doesn't need locks to atomically read a 64bit value. 2396 * So we have a optimization chance when the task's delta_exec is 0. 2397 * Reading ->on_cpu is racy, but this is ok. 2398 * 2399 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2400 * If we race with it entering cpu, unaccounted time is 0. This is 2401 * indistinguishable from the read occurring a few cycles earlier. 2402 */ 2403 if (!p->on_cpu) 2404 return p->se.sum_exec_runtime; 2405#endif 2406 2407 rq = task_rq_lock(p, &flags); 2408 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2409 task_rq_unlock(rq, p, &flags); 2410 2411 return ns; 2412} 2413 2414/* 2415 * This function gets called by the timer code, with HZ frequency. 2416 * We call it with interrupts disabled. 2417 */ 2418void scheduler_tick(void) 2419{ 2420 int cpu = smp_processor_id(); 2421 struct rq *rq = cpu_rq(cpu); 2422 struct task_struct *curr = rq->curr; 2423 2424 sched_clock_tick(); 2425 2426 raw_spin_lock(&rq->lock); 2427 update_rq_clock(rq); 2428 curr->sched_class->task_tick(rq, curr, 0); 2429 update_cpu_load_active(rq); 2430 raw_spin_unlock(&rq->lock); 2431 2432 perf_event_task_tick(); 2433 2434#ifdef CONFIG_SMP 2435 rq->idle_balance = idle_cpu(cpu); 2436 trigger_load_balance(rq); 2437#endif 2438 rq_last_tick_reset(rq); 2439} 2440 2441#ifdef CONFIG_NO_HZ_FULL 2442/** 2443 * scheduler_tick_max_deferment 2444 * 2445 * Keep at least one tick per second when a single 2446 * active task is running because the scheduler doesn't 2447 * yet completely support full dynticks environment. 2448 * 2449 * This makes sure that uptime, CFS vruntime, load 2450 * balancing, etc... continue to move forward, even 2451 * with a very low granularity. 2452 * 2453 * Return: Maximum deferment in nanoseconds. 2454 */ 2455u64 scheduler_tick_max_deferment(void) 2456{ 2457 struct rq *rq = this_rq(); 2458 unsigned long next, now = ACCESS_ONCE(jiffies); 2459 2460 next = rq->last_sched_tick + HZ; 2461 2462 if (time_before_eq(next, now)) 2463 return 0; 2464 2465 return jiffies_to_nsecs(next - now); 2466} 2467#endif 2468 2469notrace unsigned long get_parent_ip(unsigned long addr) 2470{ 2471 if (in_lock_functions(addr)) { 2472 addr = CALLER_ADDR2; 2473 if (in_lock_functions(addr)) 2474 addr = CALLER_ADDR3; 2475 } 2476 return addr; 2477} 2478 2479#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2480 defined(CONFIG_PREEMPT_TRACER)) 2481 2482void __kprobes preempt_count_add(int val) 2483{ 2484#ifdef CONFIG_DEBUG_PREEMPT 2485 /* 2486 * Underflow? 2487 */ 2488 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2489 return; 2490#endif 2491 __preempt_count_add(val); 2492#ifdef CONFIG_DEBUG_PREEMPT 2493 /* 2494 * Spinlock count overflowing soon? 2495 */ 2496 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2497 PREEMPT_MASK - 10); 2498#endif 2499 if (preempt_count() == val) { 2500 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2501#ifdef CONFIG_DEBUG_PREEMPT 2502 current->preempt_disable_ip = ip; 2503#endif 2504 trace_preempt_off(CALLER_ADDR0, ip); 2505 } 2506} 2507EXPORT_SYMBOL(preempt_count_add); 2508 2509void __kprobes preempt_count_sub(int val) 2510{ 2511#ifdef CONFIG_DEBUG_PREEMPT 2512 /* 2513 * Underflow? 2514 */ 2515 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2516 return; 2517 /* 2518 * Is the spinlock portion underflowing? 2519 */ 2520 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2521 !(preempt_count() & PREEMPT_MASK))) 2522 return; 2523#endif 2524 2525 if (preempt_count() == val) 2526 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2527 __preempt_count_sub(val); 2528} 2529EXPORT_SYMBOL(preempt_count_sub); 2530 2531#endif 2532 2533/* 2534 * Print scheduling while atomic bug: 2535 */ 2536static noinline void __schedule_bug(struct task_struct *prev) 2537{ 2538 if (oops_in_progress) 2539 return; 2540 2541 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2542 prev->comm, prev->pid, preempt_count()); 2543 2544 debug_show_held_locks(prev); 2545 print_modules(); 2546 if (irqs_disabled()) 2547 print_irqtrace_events(prev); 2548#ifdef CONFIG_DEBUG_PREEMPT 2549 if (in_atomic_preempt_off()) { 2550 pr_err("Preemption disabled at:"); 2551 print_ip_sym(current->preempt_disable_ip); 2552 pr_cont("\n"); 2553 } 2554#endif 2555 dump_stack(); 2556 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2557} 2558 2559/* 2560 * Various schedule()-time debugging checks and statistics: 2561 */ 2562static inline void schedule_debug(struct task_struct *prev) 2563{ 2564 /* 2565 * Test if we are atomic. Since do_exit() needs to call into 2566 * schedule() atomically, we ignore that path. Otherwise whine 2567 * if we are scheduling when we should not. 2568 */ 2569 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2570 __schedule_bug(prev); 2571 rcu_sleep_check(); 2572 2573 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2574 2575 schedstat_inc(this_rq(), sched_count); 2576} 2577 2578/* 2579 * Pick up the highest-prio task: 2580 */ 2581static inline struct task_struct * 2582pick_next_task(struct rq *rq, struct task_struct *prev) 2583{ 2584 const struct sched_class *class = &fair_sched_class; 2585 struct task_struct *p; 2586 2587 /* 2588 * Optimization: we know that if all tasks are in 2589 * the fair class we can call that function directly: 2590 */ 2591 if (likely(prev->sched_class == class && 2592 rq->nr_running == rq->cfs.h_nr_running)) { 2593 p = fair_sched_class.pick_next_task(rq, prev); 2594 if (likely(p && p != RETRY_TASK)) 2595 return p; 2596 } 2597 2598again: 2599 for_each_class(class) { 2600 p = class->pick_next_task(rq, prev); 2601 if (p) { 2602 if (unlikely(p == RETRY_TASK)) 2603 goto again; 2604 return p; 2605 } 2606 } 2607 2608 BUG(); /* the idle class will always have a runnable task */ 2609} 2610 2611/* 2612 * __schedule() is the main scheduler function. 2613 * 2614 * The main means of driving the scheduler and thus entering this function are: 2615 * 2616 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2617 * 2618 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2619 * paths. For example, see arch/x86/entry_64.S. 2620 * 2621 * To drive preemption between tasks, the scheduler sets the flag in timer 2622 * interrupt handler scheduler_tick(). 2623 * 2624 * 3. Wakeups don't really cause entry into schedule(). They add a 2625 * task to the run-queue and that's it. 2626 * 2627 * Now, if the new task added to the run-queue preempts the current 2628 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2629 * called on the nearest possible occasion: 2630 * 2631 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2632 * 2633 * - in syscall or exception context, at the next outmost 2634 * preempt_enable(). (this might be as soon as the wake_up()'s 2635 * spin_unlock()!) 2636 * 2637 * - in IRQ context, return from interrupt-handler to 2638 * preemptible context 2639 * 2640 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2641 * then at the next: 2642 * 2643 * - cond_resched() call 2644 * - explicit schedule() call 2645 * - return from syscall or exception to user-space 2646 * - return from interrupt-handler to user-space 2647 */ 2648static void __sched __schedule(void) 2649{ 2650 struct task_struct *prev, *next; 2651 unsigned long *switch_count; 2652 struct rq *rq; 2653 int cpu; 2654 2655need_resched: 2656 preempt_disable(); 2657 cpu = smp_processor_id(); 2658 rq = cpu_rq(cpu); 2659 rcu_note_context_switch(cpu); 2660 prev = rq->curr; 2661 2662 schedule_debug(prev); 2663 2664 if (sched_feat(HRTICK)) 2665 hrtick_clear(rq); 2666 2667 /* 2668 * Make sure that signal_pending_state()->signal_pending() below 2669 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2670 * done by the caller to avoid the race with signal_wake_up(). 2671 */ 2672 smp_mb__before_spinlock(); 2673 raw_spin_lock_irq(&rq->lock); 2674 2675 switch_count = &prev->nivcsw; 2676 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2677 if (unlikely(signal_pending_state(prev->state, prev))) { 2678 prev->state = TASK_RUNNING; 2679 } else { 2680 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2681 prev->on_rq = 0; 2682 2683 /* 2684 * If a worker went to sleep, notify and ask workqueue 2685 * whether it wants to wake up a task to maintain 2686 * concurrency. 2687 */ 2688 if (prev->flags & PF_WQ_WORKER) { 2689 struct task_struct *to_wakeup; 2690 2691 to_wakeup = wq_worker_sleeping(prev, cpu); 2692 if (to_wakeup) 2693 try_to_wake_up_local(to_wakeup); 2694 } 2695 } 2696 switch_count = &prev->nvcsw; 2697 } 2698 2699 if (prev->on_rq || rq->skip_clock_update < 0) 2700 update_rq_clock(rq); 2701 2702 next = pick_next_task(rq, prev); 2703 clear_tsk_need_resched(prev); 2704 clear_preempt_need_resched(); 2705 rq->skip_clock_update = 0; 2706 2707 if (likely(prev != next)) { 2708 rq->nr_switches++; 2709 rq->curr = next; 2710 ++*switch_count; 2711 2712 context_switch(rq, prev, next); /* unlocks the rq */ 2713 /* 2714 * The context switch have flipped the stack from under us 2715 * and restored the local variables which were saved when 2716 * this task called schedule() in the past. prev == current 2717 * is still correct, but it can be moved to another cpu/rq. 2718 */ 2719 cpu = smp_processor_id(); 2720 rq = cpu_rq(cpu); 2721 } else 2722 raw_spin_unlock_irq(&rq->lock); 2723 2724 post_schedule(rq); 2725 2726 sched_preempt_enable_no_resched(); 2727 if (need_resched()) 2728 goto need_resched; 2729} 2730 2731static inline void sched_submit_work(struct task_struct *tsk) 2732{ 2733 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2734 return; 2735 /* 2736 * If we are going to sleep and we have plugged IO queued, 2737 * make sure to submit it to avoid deadlocks. 2738 */ 2739 if (blk_needs_flush_plug(tsk)) 2740 blk_schedule_flush_plug(tsk); 2741} 2742 2743asmlinkage void __sched schedule(void) 2744{ 2745 struct task_struct *tsk = current; 2746 2747 sched_submit_work(tsk); 2748 __schedule(); 2749} 2750EXPORT_SYMBOL(schedule); 2751 2752#ifdef CONFIG_CONTEXT_TRACKING 2753asmlinkage void __sched schedule_user(void) 2754{ 2755 /* 2756 * If we come here after a random call to set_need_resched(), 2757 * or we have been woken up remotely but the IPI has not yet arrived, 2758 * we haven't yet exited the RCU idle mode. Do it here manually until 2759 * we find a better solution. 2760 */ 2761 user_exit(); 2762 schedule(); 2763 user_enter(); 2764} 2765#endif 2766 2767/** 2768 * schedule_preempt_disabled - called with preemption disabled 2769 * 2770 * Returns with preemption disabled. Note: preempt_count must be 1 2771 */ 2772void __sched schedule_preempt_disabled(void) 2773{ 2774 sched_preempt_enable_no_resched(); 2775 schedule(); 2776 preempt_disable(); 2777} 2778 2779#ifdef CONFIG_PREEMPT 2780/* 2781 * this is the entry point to schedule() from in-kernel preemption 2782 * off of preempt_enable. Kernel preemptions off return from interrupt 2783 * occur there and call schedule directly. 2784 */ 2785asmlinkage void __sched notrace preempt_schedule(void) 2786{ 2787 /* 2788 * If there is a non-zero preempt_count or interrupts are disabled, 2789 * we do not want to preempt the current task. Just return.. 2790 */ 2791 if (likely(!preemptible())) 2792 return; 2793 2794 do { 2795 __preempt_count_add(PREEMPT_ACTIVE); 2796 __schedule(); 2797 __preempt_count_sub(PREEMPT_ACTIVE); 2798 2799 /* 2800 * Check again in case we missed a preemption opportunity 2801 * between schedule and now. 2802 */ 2803 barrier(); 2804 } while (need_resched()); 2805} 2806EXPORT_SYMBOL(preempt_schedule); 2807#endif /* CONFIG_PREEMPT */ 2808 2809/* 2810 * this is the entry point to schedule() from kernel preemption 2811 * off of irq context. 2812 * Note, that this is called and return with irqs disabled. This will 2813 * protect us against recursive calling from irq. 2814 */ 2815asmlinkage void __sched preempt_schedule_irq(void) 2816{ 2817 enum ctx_state prev_state; 2818 2819 /* Catch callers which need to be fixed */ 2820 BUG_ON(preempt_count() || !irqs_disabled()); 2821 2822 prev_state = exception_enter(); 2823 2824 do { 2825 __preempt_count_add(PREEMPT_ACTIVE); 2826 local_irq_enable(); 2827 __schedule(); 2828 local_irq_disable(); 2829 __preempt_count_sub(PREEMPT_ACTIVE); 2830 2831 /* 2832 * Check again in case we missed a preemption opportunity 2833 * between schedule and now. 2834 */ 2835 barrier(); 2836 } while (need_resched()); 2837 2838 exception_exit(prev_state); 2839} 2840 2841int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2842 void *key) 2843{ 2844 return try_to_wake_up(curr->private, mode, wake_flags); 2845} 2846EXPORT_SYMBOL(default_wake_function); 2847 2848#ifdef CONFIG_RT_MUTEXES 2849 2850/* 2851 * rt_mutex_setprio - set the current priority of a task 2852 * @p: task 2853 * @prio: prio value (kernel-internal form) 2854 * 2855 * This function changes the 'effective' priority of a task. It does 2856 * not touch ->normal_prio like __setscheduler(). 2857 * 2858 * Used by the rt_mutex code to implement priority inheritance 2859 * logic. Call site only calls if the priority of the task changed. 2860 */ 2861void rt_mutex_setprio(struct task_struct *p, int prio) 2862{ 2863 int oldprio, on_rq, running, enqueue_flag = 0; 2864 struct rq *rq; 2865 const struct sched_class *prev_class; 2866 2867 BUG_ON(prio > MAX_PRIO); 2868 2869 rq = __task_rq_lock(p); 2870 2871 /* 2872 * Idle task boosting is a nono in general. There is one 2873 * exception, when PREEMPT_RT and NOHZ is active: 2874 * 2875 * The idle task calls get_next_timer_interrupt() and holds 2876 * the timer wheel base->lock on the CPU and another CPU wants 2877 * to access the timer (probably to cancel it). We can safely 2878 * ignore the boosting request, as the idle CPU runs this code 2879 * with interrupts disabled and will complete the lock 2880 * protected section without being interrupted. So there is no 2881 * real need to boost. 2882 */ 2883 if (unlikely(p == rq->idle)) { 2884 WARN_ON(p != rq->curr); 2885 WARN_ON(p->pi_blocked_on); 2886 goto out_unlock; 2887 } 2888 2889 trace_sched_pi_setprio(p, prio); 2890 p->pi_top_task = rt_mutex_get_top_task(p); 2891 oldprio = p->prio; 2892 prev_class = p->sched_class; 2893 on_rq = p->on_rq; 2894 running = task_current(rq, p); 2895 if (on_rq) 2896 dequeue_task(rq, p, 0); 2897 if (running) 2898 p->sched_class->put_prev_task(rq, p); 2899 2900 /* 2901 * Boosting condition are: 2902 * 1. -rt task is running and holds mutex A 2903 * --> -dl task blocks on mutex A 2904 * 2905 * 2. -dl task is running and holds mutex A 2906 * --> -dl task blocks on mutex A and could preempt the 2907 * running task 2908 */ 2909 if (dl_prio(prio)) { 2910 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2911 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2912 p->dl.dl_boosted = 1; 2913 p->dl.dl_throttled = 0; 2914 enqueue_flag = ENQUEUE_REPLENISH; 2915 } else 2916 p->dl.dl_boosted = 0; 2917 p->sched_class = &dl_sched_class; 2918 } else if (rt_prio(prio)) { 2919 if (dl_prio(oldprio)) 2920 p->dl.dl_boosted = 0; 2921 if (oldprio < prio) 2922 enqueue_flag = ENQUEUE_HEAD; 2923 p->sched_class = &rt_sched_class; 2924 } else { 2925 if (dl_prio(oldprio)) 2926 p->dl.dl_boosted = 0; 2927 p->sched_class = &fair_sched_class; 2928 } 2929 2930 p->prio = prio; 2931 2932 if (running) 2933 p->sched_class->set_curr_task(rq); 2934 if (on_rq) 2935 enqueue_task(rq, p, enqueue_flag); 2936 2937 check_class_changed(rq, p, prev_class, oldprio); 2938out_unlock: 2939 __task_rq_unlock(rq); 2940} 2941#endif 2942 2943void set_user_nice(struct task_struct *p, long nice) 2944{ 2945 int old_prio, delta, on_rq; 2946 unsigned long flags; 2947 struct rq *rq; 2948 2949 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 2950 return; 2951 /* 2952 * We have to be careful, if called from sys_setpriority(), 2953 * the task might be in the middle of scheduling on another CPU. 2954 */ 2955 rq = task_rq_lock(p, &flags); 2956 /* 2957 * The RT priorities are set via sched_setscheduler(), but we still 2958 * allow the 'normal' nice value to be set - but as expected 2959 * it wont have any effect on scheduling until the task is 2960 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 2961 */ 2962 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 2963 p->static_prio = NICE_TO_PRIO(nice); 2964 goto out_unlock; 2965 } 2966 on_rq = p->on_rq; 2967 if (on_rq) 2968 dequeue_task(rq, p, 0); 2969 2970 p->static_prio = NICE_TO_PRIO(nice); 2971 set_load_weight(p); 2972 old_prio = p->prio; 2973 p->prio = effective_prio(p); 2974 delta = p->prio - old_prio; 2975 2976 if (on_rq) { 2977 enqueue_task(rq, p, 0); 2978 /* 2979 * If the task increased its priority or is running and 2980 * lowered its priority, then reschedule its CPU: 2981 */ 2982 if (delta < 0 || (delta > 0 && task_running(rq, p))) 2983 resched_task(rq->curr); 2984 } 2985out_unlock: 2986 task_rq_unlock(rq, p, &flags); 2987} 2988EXPORT_SYMBOL(set_user_nice); 2989 2990/* 2991 * can_nice - check if a task can reduce its nice value 2992 * @p: task 2993 * @nice: nice value 2994 */ 2995int can_nice(const struct task_struct *p, const int nice) 2996{ 2997 /* convert nice value [19,-20] to rlimit style value [1,40] */ 2998 int nice_rlim = 20 - nice; 2999 3000 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3001 capable(CAP_SYS_NICE)); 3002} 3003 3004#ifdef __ARCH_WANT_SYS_NICE 3005 3006/* 3007 * sys_nice - change the priority of the current process. 3008 * @increment: priority increment 3009 * 3010 * sys_setpriority is a more generic, but much slower function that 3011 * does similar things. 3012 */ 3013SYSCALL_DEFINE1(nice, int, increment) 3014{ 3015 long nice, retval; 3016 3017 /* 3018 * Setpriority might change our priority at the same moment. 3019 * We don't have to worry. Conceptually one call occurs first 3020 * and we have a single winner. 3021 */ 3022 if (increment < -40) 3023 increment = -40; 3024 if (increment > 40) 3025 increment = 40; 3026 3027 nice = task_nice(current) + increment; 3028 if (nice < MIN_NICE) 3029 nice = MIN_NICE; 3030 if (nice > MAX_NICE) 3031 nice = MAX_NICE; 3032 3033 if (increment < 0 && !can_nice(current, nice)) 3034 return -EPERM; 3035 3036 retval = security_task_setnice(current, nice); 3037 if (retval) 3038 return retval; 3039 3040 set_user_nice(current, nice); 3041 return 0; 3042} 3043 3044#endif 3045 3046/** 3047 * task_prio - return the priority value of a given task. 3048 * @p: the task in question. 3049 * 3050 * Return: The priority value as seen by users in /proc. 3051 * RT tasks are offset by -200. Normal tasks are centered 3052 * around 0, value goes from -16 to +15. 3053 */ 3054int task_prio(const struct task_struct *p) 3055{ 3056 return p->prio - MAX_RT_PRIO; 3057} 3058 3059/** 3060 * idle_cpu - is a given cpu idle currently? 3061 * @cpu: the processor in question. 3062 * 3063 * Return: 1 if the CPU is currently idle. 0 otherwise. 3064 */ 3065int idle_cpu(int cpu) 3066{ 3067 struct rq *rq = cpu_rq(cpu); 3068 3069 if (rq->curr != rq->idle) 3070 return 0; 3071 3072 if (rq->nr_running) 3073 return 0; 3074 3075#ifdef CONFIG_SMP 3076 if (!llist_empty(&rq->wake_list)) 3077 return 0; 3078#endif 3079 3080 return 1; 3081} 3082 3083/** 3084 * idle_task - return the idle task for a given cpu. 3085 * @cpu: the processor in question. 3086 * 3087 * Return: The idle task for the cpu @cpu. 3088 */ 3089struct task_struct *idle_task(int cpu) 3090{ 3091 return cpu_rq(cpu)->idle; 3092} 3093 3094/** 3095 * find_process_by_pid - find a process with a matching PID value. 3096 * @pid: the pid in question. 3097 * 3098 * The task of @pid, if found. %NULL otherwise. 3099 */ 3100static struct task_struct *find_process_by_pid(pid_t pid) 3101{ 3102 return pid ? find_task_by_vpid(pid) : current; 3103} 3104 3105/* 3106 * This function initializes the sched_dl_entity of a newly becoming 3107 * SCHED_DEADLINE task. 3108 * 3109 * Only the static values are considered here, the actual runtime and the 3110 * absolute deadline will be properly calculated when the task is enqueued 3111 * for the first time with its new policy. 3112 */ 3113static void 3114__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3115{ 3116 struct sched_dl_entity *dl_se = &p->dl; 3117 3118 init_dl_task_timer(dl_se); 3119 dl_se->dl_runtime = attr->sched_runtime; 3120 dl_se->dl_deadline = attr->sched_deadline; 3121 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3122 dl_se->flags = attr->sched_flags; 3123 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3124 dl_se->dl_throttled = 0; 3125 dl_se->dl_new = 1; 3126} 3127 3128static void __setscheduler_params(struct task_struct *p, 3129 const struct sched_attr *attr) 3130{ 3131 int policy = attr->sched_policy; 3132 3133 if (policy == -1) /* setparam */ 3134 policy = p->policy; 3135 3136 p->policy = policy; 3137 3138 if (dl_policy(policy)) 3139 __setparam_dl(p, attr); 3140 else if (fair_policy(policy)) 3141 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3142 3143 /* 3144 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3145 * !rt_policy. Always setting this ensures that things like 3146 * getparam()/getattr() don't report silly values for !rt tasks. 3147 */ 3148 p->rt_priority = attr->sched_priority; 3149 p->normal_prio = normal_prio(p); 3150 set_load_weight(p); 3151} 3152 3153/* Actually do priority change: must hold pi & rq lock. */ 3154static void __setscheduler(struct rq *rq, struct task_struct *p, 3155 const struct sched_attr *attr) 3156{ 3157 __setscheduler_params(p, attr); 3158 3159 /* 3160 * If we get here, there was no pi waiters boosting the 3161 * task. It is safe to use the normal prio. 3162 */ 3163 p->prio = normal_prio(p); 3164 3165 if (dl_prio(p->prio)) 3166 p->sched_class = &dl_sched_class; 3167 else if (rt_prio(p->prio)) 3168 p->sched_class = &rt_sched_class; 3169 else 3170 p->sched_class = &fair_sched_class; 3171} 3172 3173static void 3174__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3175{ 3176 struct sched_dl_entity *dl_se = &p->dl; 3177 3178 attr->sched_priority = p->rt_priority; 3179 attr->sched_runtime = dl_se->dl_runtime; 3180 attr->sched_deadline = dl_se->dl_deadline; 3181 attr->sched_period = dl_se->dl_period; 3182 attr->sched_flags = dl_se->flags; 3183} 3184 3185/* 3186 * This function validates the new parameters of a -deadline task. 3187 * We ask for the deadline not being zero, and greater or equal 3188 * than the runtime, as well as the period of being zero or 3189 * greater than deadline. Furthermore, we have to be sure that 3190 * user parameters are above the internal resolution (1us); we 3191 * check sched_runtime only since it is always the smaller one. 3192 */ 3193static bool 3194__checkparam_dl(const struct sched_attr *attr) 3195{ 3196 return attr && attr->sched_deadline != 0 && 3197 (attr->sched_period == 0 || 3198 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3199 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3200 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3201} 3202 3203/* 3204 * check the target process has a UID that matches the current process's 3205 */ 3206static bool check_same_owner(struct task_struct *p) 3207{ 3208 const struct cred *cred = current_cred(), *pcred; 3209 bool match; 3210 3211 rcu_read_lock(); 3212 pcred = __task_cred(p); 3213 match = (uid_eq(cred->euid, pcred->euid) || 3214 uid_eq(cred->euid, pcred->uid)); 3215 rcu_read_unlock(); 3216 return match; 3217} 3218 3219static int __sched_setscheduler(struct task_struct *p, 3220 const struct sched_attr *attr, 3221 bool user) 3222{ 3223 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3224 MAX_RT_PRIO - 1 - attr->sched_priority; 3225 int retval, oldprio, oldpolicy = -1, on_rq, running; 3226 int policy = attr->sched_policy; 3227 unsigned long flags; 3228 const struct sched_class *prev_class; 3229 struct rq *rq; 3230 int reset_on_fork; 3231 3232 /* may grab non-irq protected spin_locks */ 3233 BUG_ON(in_interrupt()); 3234recheck: 3235 /* double check policy once rq lock held */ 3236 if (policy < 0) { 3237 reset_on_fork = p->sched_reset_on_fork; 3238 policy = oldpolicy = p->policy; 3239 } else { 3240 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3241 3242 if (policy != SCHED_DEADLINE && 3243 policy != SCHED_FIFO && policy != SCHED_RR && 3244 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3245 policy != SCHED_IDLE) 3246 return -EINVAL; 3247 } 3248 3249 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3250 return -EINVAL; 3251 3252 /* 3253 * Valid priorities for SCHED_FIFO and SCHED_RR are 3254 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3255 * SCHED_BATCH and SCHED_IDLE is 0. 3256 */ 3257 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3258 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3259 return -EINVAL; 3260 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3261 (rt_policy(policy) != (attr->sched_priority != 0))) 3262 return -EINVAL; 3263 3264 /* 3265 * Allow unprivileged RT tasks to decrease priority: 3266 */ 3267 if (user && !capable(CAP_SYS_NICE)) { 3268 if (fair_policy(policy)) { 3269 if (attr->sched_nice < task_nice(p) && 3270 !can_nice(p, attr->sched_nice)) 3271 return -EPERM; 3272 } 3273 3274 if (rt_policy(policy)) { 3275 unsigned long rlim_rtprio = 3276 task_rlimit(p, RLIMIT_RTPRIO); 3277 3278 /* can't set/change the rt policy */ 3279 if (policy != p->policy && !rlim_rtprio) 3280 return -EPERM; 3281 3282 /* can't increase priority */ 3283 if (attr->sched_priority > p->rt_priority && 3284 attr->sched_priority > rlim_rtprio) 3285 return -EPERM; 3286 } 3287 3288 /* 3289 * Can't set/change SCHED_DEADLINE policy at all for now 3290 * (safest behavior); in the future we would like to allow 3291 * unprivileged DL tasks to increase their relative deadline 3292 * or reduce their runtime (both ways reducing utilization) 3293 */ 3294 if (dl_policy(policy)) 3295 return -EPERM; 3296 3297 /* 3298 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3299 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3300 */ 3301 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3302 if (!can_nice(p, task_nice(p))) 3303 return -EPERM; 3304 } 3305 3306 /* can't change other user's priorities */ 3307 if (!check_same_owner(p)) 3308 return -EPERM; 3309 3310 /* Normal users shall not reset the sched_reset_on_fork flag */ 3311 if (p->sched_reset_on_fork && !reset_on_fork) 3312 return -EPERM; 3313 } 3314 3315 if (user) { 3316 retval = security_task_setscheduler(p); 3317 if (retval) 3318 return retval; 3319 } 3320 3321 /* 3322 * make sure no PI-waiters arrive (or leave) while we are 3323 * changing the priority of the task: 3324 * 3325 * To be able to change p->policy safely, the appropriate 3326 * runqueue lock must be held. 3327 */ 3328 rq = task_rq_lock(p, &flags); 3329 3330 /* 3331 * Changing the policy of the stop threads its a very bad idea 3332 */ 3333 if (p == rq->stop) { 3334 task_rq_unlock(rq, p, &flags); 3335 return -EINVAL; 3336 } 3337 3338 /* 3339 * If not changing anything there's no need to proceed further, 3340 * but store a possible modification of reset_on_fork. 3341 */ 3342 if (unlikely(policy == p->policy)) { 3343 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3344 goto change; 3345 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3346 goto change; 3347 if (dl_policy(policy)) 3348 goto change; 3349 3350 p->sched_reset_on_fork = reset_on_fork; 3351 task_rq_unlock(rq, p, &flags); 3352 return 0; 3353 } 3354change: 3355 3356 if (user) { 3357#ifdef CONFIG_RT_GROUP_SCHED 3358 /* 3359 * Do not allow realtime tasks into groups that have no runtime 3360 * assigned. 3361 */ 3362 if (rt_bandwidth_enabled() && rt_policy(policy) && 3363 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3364 !task_group_is_autogroup(task_group(p))) { 3365 task_rq_unlock(rq, p, &flags); 3366 return -EPERM; 3367 } 3368#endif 3369#ifdef CONFIG_SMP 3370 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3371 cpumask_t *span = rq->rd->span; 3372 3373 /* 3374 * Don't allow tasks with an affinity mask smaller than 3375 * the entire root_domain to become SCHED_DEADLINE. We 3376 * will also fail if there's no bandwidth available. 3377 */ 3378 if (!cpumask_subset(span, &p->cpus_allowed) || 3379 rq->rd->dl_bw.bw == 0) { 3380 task_rq_unlock(rq, p, &flags); 3381 return -EPERM; 3382 } 3383 } 3384#endif 3385 } 3386 3387 /* recheck policy now with rq lock held */ 3388 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3389 policy = oldpolicy = -1; 3390 task_rq_unlock(rq, p, &flags); 3391 goto recheck; 3392 } 3393 3394 /* 3395 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3396 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3397 * is available. 3398 */ 3399 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3400 task_rq_unlock(rq, p, &flags); 3401 return -EBUSY; 3402 } 3403 3404 p->sched_reset_on_fork = reset_on_fork; 3405 oldprio = p->prio; 3406 3407 /* 3408 * Special case for priority boosted tasks. 3409 * 3410 * If the new priority is lower or equal (user space view) 3411 * than the current (boosted) priority, we just store the new 3412 * normal parameters and do not touch the scheduler class and 3413 * the runqueue. This will be done when the task deboost 3414 * itself. 3415 */ 3416 if (rt_mutex_check_prio(p, newprio)) { 3417 __setscheduler_params(p, attr); 3418 task_rq_unlock(rq, p, &flags); 3419 return 0; 3420 } 3421 3422 on_rq = p->on_rq; 3423 running = task_current(rq, p); 3424 if (on_rq) 3425 dequeue_task(rq, p, 0); 3426 if (running) 3427 p->sched_class->put_prev_task(rq, p); 3428 3429 prev_class = p->sched_class; 3430 __setscheduler(rq, p, attr); 3431 3432 if (running) 3433 p->sched_class->set_curr_task(rq); 3434 if (on_rq) { 3435 /* 3436 * We enqueue to tail when the priority of a task is 3437 * increased (user space view). 3438 */ 3439 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3440 } 3441 3442 check_class_changed(rq, p, prev_class, oldprio); 3443 task_rq_unlock(rq, p, &flags); 3444 3445 rt_mutex_adjust_pi(p); 3446 3447 return 0; 3448} 3449 3450static int _sched_setscheduler(struct task_struct *p, int policy, 3451 const struct sched_param *param, bool check) 3452{ 3453 struct sched_attr attr = { 3454 .sched_policy = policy, 3455 .sched_priority = param->sched_priority, 3456 .sched_nice = PRIO_TO_NICE(p->static_prio), 3457 }; 3458 3459 /* 3460 * Fixup the legacy SCHED_RESET_ON_FORK hack 3461 */ 3462 if (policy & SCHED_RESET_ON_FORK) { 3463 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3464 policy &= ~SCHED_RESET_ON_FORK; 3465 attr.sched_policy = policy; 3466 } 3467 3468 return __sched_setscheduler(p, &attr, check); 3469} 3470/** 3471 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3472 * @p: the task in question. 3473 * @policy: new policy. 3474 * @param: structure containing the new RT priority. 3475 * 3476 * Return: 0 on success. An error code otherwise. 3477 * 3478 * NOTE that the task may be already dead. 3479 */ 3480int sched_setscheduler(struct task_struct *p, int policy, 3481 const struct sched_param *param) 3482{ 3483 return _sched_setscheduler(p, policy, param, true); 3484} 3485EXPORT_SYMBOL_GPL(sched_setscheduler); 3486 3487int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3488{ 3489 return __sched_setscheduler(p, attr, true); 3490} 3491EXPORT_SYMBOL_GPL(sched_setattr); 3492 3493/** 3494 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3495 * @p: the task in question. 3496 * @policy: new policy. 3497 * @param: structure containing the new RT priority. 3498 * 3499 * Just like sched_setscheduler, only don't bother checking if the 3500 * current context has permission. For example, this is needed in 3501 * stop_machine(): we create temporary high priority worker threads, 3502 * but our caller might not have that capability. 3503 * 3504 * Return: 0 on success. An error code otherwise. 3505 */ 3506int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3507 const struct sched_param *param) 3508{ 3509 return _sched_setscheduler(p, policy, param, false); 3510} 3511 3512static int 3513do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3514{ 3515 struct sched_param lparam; 3516 struct task_struct *p; 3517 int retval; 3518 3519 if (!param || pid < 0) 3520 return -EINVAL; 3521 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3522 return -EFAULT; 3523 3524 rcu_read_lock(); 3525 retval = -ESRCH; 3526 p = find_process_by_pid(pid); 3527 if (p != NULL) 3528 retval = sched_setscheduler(p, policy, &lparam); 3529 rcu_read_unlock(); 3530 3531 return retval; 3532} 3533 3534/* 3535 * Mimics kernel/events/core.c perf_copy_attr(). 3536 */ 3537static int sched_copy_attr(struct sched_attr __user *uattr, 3538 struct sched_attr *attr) 3539{ 3540 u32 size; 3541 int ret; 3542 3543 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3544 return -EFAULT; 3545 3546 /* 3547 * zero the full structure, so that a short copy will be nice. 3548 */ 3549 memset(attr, 0, sizeof(*attr)); 3550 3551 ret = get_user(size, &uattr->size); 3552 if (ret) 3553 return ret; 3554 3555 if (size > PAGE_SIZE) /* silly large */ 3556 goto err_size; 3557 3558 if (!size) /* abi compat */ 3559 size = SCHED_ATTR_SIZE_VER0; 3560 3561 if (size < SCHED_ATTR_SIZE_VER0) 3562 goto err_size; 3563 3564 /* 3565 * If we're handed a bigger struct than we know of, 3566 * ensure all the unknown bits are 0 - i.e. new 3567 * user-space does not rely on any kernel feature 3568 * extensions we dont know about yet. 3569 */ 3570 if (size > sizeof(*attr)) { 3571 unsigned char __user *addr; 3572 unsigned char __user *end; 3573 unsigned char val; 3574 3575 addr = (void __user *)uattr + sizeof(*attr); 3576 end = (void __user *)uattr + size; 3577 3578 for (; addr < end; addr++) { 3579 ret = get_user(val, addr); 3580 if (ret) 3581 return ret; 3582 if (val) 3583 goto err_size; 3584 } 3585 size = sizeof(*attr); 3586 } 3587 3588 ret = copy_from_user(attr, uattr, size); 3589 if (ret) 3590 return -EFAULT; 3591 3592 /* 3593 * XXX: do we want to be lenient like existing syscalls; or do we want 3594 * to be strict and return an error on out-of-bounds values? 3595 */ 3596 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3597 3598out: 3599 return ret; 3600 3601err_size: 3602 put_user(sizeof(*attr), &uattr->size); 3603 ret = -E2BIG; 3604 goto out; 3605} 3606 3607/** 3608 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3609 * @pid: the pid in question. 3610 * @policy: new policy. 3611 * @param: structure containing the new RT priority. 3612 * 3613 * Return: 0 on success. An error code otherwise. 3614 */ 3615SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3616 struct sched_param __user *, param) 3617{ 3618 /* negative values for policy are not valid */ 3619 if (policy < 0) 3620 return -EINVAL; 3621 3622 return do_sched_setscheduler(pid, policy, param); 3623} 3624 3625/** 3626 * sys_sched_setparam - set/change the RT priority of a thread 3627 * @pid: the pid in question. 3628 * @param: structure containing the new RT priority. 3629 * 3630 * Return: 0 on success. An error code otherwise. 3631 */ 3632SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3633{ 3634 return do_sched_setscheduler(pid, -1, param); 3635} 3636 3637/** 3638 * sys_sched_setattr - same as above, but with extended sched_attr 3639 * @pid: the pid in question. 3640 * @uattr: structure containing the extended parameters. 3641 */ 3642SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3643 unsigned int, flags) 3644{ 3645 struct sched_attr attr; 3646 struct task_struct *p; 3647 int retval; 3648 3649 if (!uattr || pid < 0 || flags) 3650 return -EINVAL; 3651 3652 if (sched_copy_attr(uattr, &attr)) 3653 return -EFAULT; 3654 3655 rcu_read_lock(); 3656 retval = -ESRCH; 3657 p = find_process_by_pid(pid); 3658 if (p != NULL) 3659 retval = sched_setattr(p, &attr); 3660 rcu_read_unlock(); 3661 3662 return retval; 3663} 3664 3665/** 3666 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3667 * @pid: the pid in question. 3668 * 3669 * Return: On success, the policy of the thread. Otherwise, a negative error 3670 * code. 3671 */ 3672SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3673{ 3674 struct task_struct *p; 3675 int retval; 3676 3677 if (pid < 0) 3678 return -EINVAL; 3679 3680 retval = -ESRCH; 3681 rcu_read_lock(); 3682 p = find_process_by_pid(pid); 3683 if (p) { 3684 retval = security_task_getscheduler(p); 3685 if (!retval) 3686 retval = p->policy 3687 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3688 } 3689 rcu_read_unlock(); 3690 return retval; 3691} 3692 3693/** 3694 * sys_sched_getparam - get the RT priority of a thread 3695 * @pid: the pid in question. 3696 * @param: structure containing the RT priority. 3697 * 3698 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3699 * code. 3700 */ 3701SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3702{ 3703 struct sched_param lp; 3704 struct task_struct *p; 3705 int retval; 3706 3707 if (!param || pid < 0) 3708 return -EINVAL; 3709 3710 rcu_read_lock(); 3711 p = find_process_by_pid(pid); 3712 retval = -ESRCH; 3713 if (!p) 3714 goto out_unlock; 3715 3716 retval = security_task_getscheduler(p); 3717 if (retval) 3718 goto out_unlock; 3719 3720 if (task_has_dl_policy(p)) { 3721 retval = -EINVAL; 3722 goto out_unlock; 3723 } 3724 lp.sched_priority = p->rt_priority; 3725 rcu_read_unlock(); 3726 3727 /* 3728 * This one might sleep, we cannot do it with a spinlock held ... 3729 */ 3730 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3731 3732 return retval; 3733 3734out_unlock: 3735 rcu_read_unlock(); 3736 return retval; 3737} 3738 3739static int sched_read_attr(struct sched_attr __user *uattr, 3740 struct sched_attr *attr, 3741 unsigned int usize) 3742{ 3743 int ret; 3744 3745 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3746 return -EFAULT; 3747 3748 /* 3749 * If we're handed a smaller struct than we know of, 3750 * ensure all the unknown bits are 0 - i.e. old 3751 * user-space does not get uncomplete information. 3752 */ 3753 if (usize < sizeof(*attr)) { 3754 unsigned char *addr; 3755 unsigned char *end; 3756 3757 addr = (void *)attr + usize; 3758 end = (void *)attr + sizeof(*attr); 3759 3760 for (; addr < end; addr++) { 3761 if (*addr) 3762 goto err_size; 3763 } 3764 3765 attr->size = usize; 3766 } 3767 3768 ret = copy_to_user(uattr, attr, attr->size); 3769 if (ret) 3770 return -EFAULT; 3771 3772out: 3773 return ret; 3774 3775err_size: 3776 ret = -E2BIG; 3777 goto out; 3778} 3779 3780/** 3781 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3782 * @pid: the pid in question. 3783 * @uattr: structure containing the extended parameters. 3784 * @size: sizeof(attr) for fwd/bwd comp. 3785 */ 3786SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3787 unsigned int, size, unsigned int, flags) 3788{ 3789 struct sched_attr attr = { 3790 .size = sizeof(struct sched_attr), 3791 }; 3792 struct task_struct *p; 3793 int retval; 3794 3795 if (!uattr || pid < 0 || size > PAGE_SIZE || 3796 size < SCHED_ATTR_SIZE_VER0 || flags) 3797 return -EINVAL; 3798 3799 rcu_read_lock(); 3800 p = find_process_by_pid(pid); 3801 retval = -ESRCH; 3802 if (!p) 3803 goto out_unlock; 3804 3805 retval = security_task_getscheduler(p); 3806 if (retval) 3807 goto out_unlock; 3808 3809 attr.sched_policy = p->policy; 3810 if (p->sched_reset_on_fork) 3811 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3812 if (task_has_dl_policy(p)) 3813 __getparam_dl(p, &attr); 3814 else if (task_has_rt_policy(p)) 3815 attr.sched_priority = p->rt_priority; 3816 else 3817 attr.sched_nice = task_nice(p); 3818 3819 rcu_read_unlock(); 3820 3821 retval = sched_read_attr(uattr, &attr, size); 3822 return retval; 3823 3824out_unlock: 3825 rcu_read_unlock(); 3826 return retval; 3827} 3828 3829long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3830{ 3831 cpumask_var_t cpus_allowed, new_mask; 3832 struct task_struct *p; 3833 int retval; 3834 3835 rcu_read_lock(); 3836 3837 p = find_process_by_pid(pid); 3838 if (!p) { 3839 rcu_read_unlock(); 3840 return -ESRCH; 3841 } 3842 3843 /* Prevent p going away */ 3844 get_task_struct(p); 3845 rcu_read_unlock(); 3846 3847 if (p->flags & PF_NO_SETAFFINITY) { 3848 retval = -EINVAL; 3849 goto out_put_task; 3850 } 3851 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3852 retval = -ENOMEM; 3853 goto out_put_task; 3854 } 3855 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3856 retval = -ENOMEM; 3857 goto out_free_cpus_allowed; 3858 } 3859 retval = -EPERM; 3860 if (!check_same_owner(p)) { 3861 rcu_read_lock(); 3862 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3863 rcu_read_unlock(); 3864 goto out_unlock; 3865 } 3866 rcu_read_unlock(); 3867 } 3868 3869 retval = security_task_setscheduler(p); 3870 if (retval) 3871 goto out_unlock; 3872 3873 3874 cpuset_cpus_allowed(p, cpus_allowed); 3875 cpumask_and(new_mask, in_mask, cpus_allowed); 3876 3877 /* 3878 * Since bandwidth control happens on root_domain basis, 3879 * if admission test is enabled, we only admit -deadline 3880 * tasks allowed to run on all the CPUs in the task's 3881 * root_domain. 3882 */ 3883#ifdef CONFIG_SMP 3884 if (task_has_dl_policy(p)) { 3885 const struct cpumask *span = task_rq(p)->rd->span; 3886 3887 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3888 retval = -EBUSY; 3889 goto out_unlock; 3890 } 3891 } 3892#endif 3893again: 3894 retval = set_cpus_allowed_ptr(p, new_mask); 3895 3896 if (!retval) { 3897 cpuset_cpus_allowed(p, cpus_allowed); 3898 if (!cpumask_subset(new_mask, cpus_allowed)) { 3899 /* 3900 * We must have raced with a concurrent cpuset 3901 * update. Just reset the cpus_allowed to the 3902 * cpuset's cpus_allowed 3903 */ 3904 cpumask_copy(new_mask, cpus_allowed); 3905 goto again; 3906 } 3907 } 3908out_unlock: 3909 free_cpumask_var(new_mask); 3910out_free_cpus_allowed: 3911 free_cpumask_var(cpus_allowed); 3912out_put_task: 3913 put_task_struct(p); 3914 return retval; 3915} 3916 3917static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3918 struct cpumask *new_mask) 3919{ 3920 if (len < cpumask_size()) 3921 cpumask_clear(new_mask); 3922 else if (len > cpumask_size()) 3923 len = cpumask_size(); 3924 3925 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3926} 3927 3928/** 3929 * sys_sched_setaffinity - set the cpu affinity of a process 3930 * @pid: pid of the process 3931 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3932 * @user_mask_ptr: user-space pointer to the new cpu mask 3933 * 3934 * Return: 0 on success. An error code otherwise. 3935 */ 3936SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3937 unsigned long __user *, user_mask_ptr) 3938{ 3939 cpumask_var_t new_mask; 3940 int retval; 3941 3942 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3943 return -ENOMEM; 3944 3945 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3946 if (retval == 0) 3947 retval = sched_setaffinity(pid, new_mask); 3948 free_cpumask_var(new_mask); 3949 return retval; 3950} 3951 3952long sched_getaffinity(pid_t pid, struct cpumask *mask) 3953{ 3954 struct task_struct *p; 3955 unsigned long flags; 3956 int retval; 3957 3958 rcu_read_lock(); 3959 3960 retval = -ESRCH; 3961 p = find_process_by_pid(pid); 3962 if (!p) 3963 goto out_unlock; 3964 3965 retval = security_task_getscheduler(p); 3966 if (retval) 3967 goto out_unlock; 3968 3969 raw_spin_lock_irqsave(&p->pi_lock, flags); 3970 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3971 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3972 3973out_unlock: 3974 rcu_read_unlock(); 3975 3976 return retval; 3977} 3978 3979/** 3980 * sys_sched_getaffinity - get the cpu affinity of a process 3981 * @pid: pid of the process 3982 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3983 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3984 * 3985 * Return: 0 on success. An error code otherwise. 3986 */ 3987SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3988 unsigned long __user *, user_mask_ptr) 3989{ 3990 int ret; 3991 cpumask_var_t mask; 3992 3993 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3994 return -EINVAL; 3995 if (len & (sizeof(unsigned long)-1)) 3996 return -EINVAL; 3997 3998 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3999 return -ENOMEM; 4000 4001 ret = sched_getaffinity(pid, mask); 4002 if (ret == 0) { 4003 size_t retlen = min_t(size_t, len, cpumask_size()); 4004 4005 if (copy_to_user(user_mask_ptr, mask, retlen)) 4006 ret = -EFAULT; 4007 else 4008 ret = retlen; 4009 } 4010 free_cpumask_var(mask); 4011 4012 return ret; 4013} 4014 4015/** 4016 * sys_sched_yield - yield the current processor to other threads. 4017 * 4018 * This function yields the current CPU to other tasks. If there are no 4019 * other threads running on this CPU then this function will return. 4020 * 4021 * Return: 0. 4022 */ 4023SYSCALL_DEFINE0(sched_yield) 4024{ 4025 struct rq *rq = this_rq_lock(); 4026 4027 schedstat_inc(rq, yld_count); 4028 current->sched_class->yield_task(rq); 4029 4030 /* 4031 * Since we are going to call schedule() anyway, there's 4032 * no need to preempt or enable interrupts: 4033 */ 4034 __release(rq->lock); 4035 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4036 do_raw_spin_unlock(&rq->lock); 4037 sched_preempt_enable_no_resched(); 4038 4039 schedule(); 4040 4041 return 0; 4042} 4043 4044static void __cond_resched(void) 4045{ 4046 __preempt_count_add(PREEMPT_ACTIVE); 4047 __schedule(); 4048 __preempt_count_sub(PREEMPT_ACTIVE); 4049} 4050 4051int __sched _cond_resched(void) 4052{ 4053 if (should_resched()) { 4054 __cond_resched(); 4055 return 1; 4056 } 4057 return 0; 4058} 4059EXPORT_SYMBOL(_cond_resched); 4060 4061/* 4062 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4063 * call schedule, and on return reacquire the lock. 4064 * 4065 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4066 * operations here to prevent schedule() from being called twice (once via 4067 * spin_unlock(), once by hand). 4068 */ 4069int __cond_resched_lock(spinlock_t *lock) 4070{ 4071 int resched = should_resched(); 4072 int ret = 0; 4073 4074 lockdep_assert_held(lock); 4075 4076 if (spin_needbreak(lock) || resched) { 4077 spin_unlock(lock); 4078 if (resched) 4079 __cond_resched(); 4080 else 4081 cpu_relax(); 4082 ret = 1; 4083 spin_lock(lock); 4084 } 4085 return ret; 4086} 4087EXPORT_SYMBOL(__cond_resched_lock); 4088 4089int __sched __cond_resched_softirq(void) 4090{ 4091 BUG_ON(!in_softirq()); 4092 4093 if (should_resched()) { 4094 local_bh_enable(); 4095 __cond_resched(); 4096 local_bh_disable(); 4097 return 1; 4098 } 4099 return 0; 4100} 4101EXPORT_SYMBOL(__cond_resched_softirq); 4102 4103/** 4104 * yield - yield the current processor to other threads. 4105 * 4106 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4107 * 4108 * The scheduler is at all times free to pick the calling task as the most 4109 * eligible task to run, if removing the yield() call from your code breaks 4110 * it, its already broken. 4111 * 4112 * Typical broken usage is: 4113 * 4114 * while (!event) 4115 * yield(); 4116 * 4117 * where one assumes that yield() will let 'the other' process run that will 4118 * make event true. If the current task is a SCHED_FIFO task that will never 4119 * happen. Never use yield() as a progress guarantee!! 4120 * 4121 * If you want to use yield() to wait for something, use wait_event(). 4122 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4123 * If you still want to use yield(), do not! 4124 */ 4125void __sched yield(void) 4126{ 4127 set_current_state(TASK_RUNNING); 4128 sys_sched_yield(); 4129} 4130EXPORT_SYMBOL(yield); 4131 4132/** 4133 * yield_to - yield the current processor to another thread in 4134 * your thread group, or accelerate that thread toward the 4135 * processor it's on. 4136 * @p: target task 4137 * @preempt: whether task preemption is allowed or not 4138 * 4139 * It's the caller's job to ensure that the target task struct 4140 * can't go away on us before we can do any checks. 4141 * 4142 * Return: 4143 * true (>0) if we indeed boosted the target task. 4144 * false (0) if we failed to boost the target. 4145 * -ESRCH if there's no task to yield to. 4146 */ 4147bool __sched yield_to(struct task_struct *p, bool preempt) 4148{ 4149 struct task_struct *curr = current; 4150 struct rq *rq, *p_rq; 4151 unsigned long flags; 4152 int yielded = 0; 4153 4154 local_irq_save(flags); 4155 rq = this_rq(); 4156 4157again: 4158 p_rq = task_rq(p); 4159 /* 4160 * If we're the only runnable task on the rq and target rq also 4161 * has only one task, there's absolutely no point in yielding. 4162 */ 4163 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4164 yielded = -ESRCH; 4165 goto out_irq; 4166 } 4167 4168 double_rq_lock(rq, p_rq); 4169 if (task_rq(p) != p_rq) { 4170 double_rq_unlock(rq, p_rq); 4171 goto again; 4172 } 4173 4174 if (!curr->sched_class->yield_to_task) 4175 goto out_unlock; 4176 4177 if (curr->sched_class != p->sched_class) 4178 goto out_unlock; 4179 4180 if (task_running(p_rq, p) || p->state) 4181 goto out_unlock; 4182 4183 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4184 if (yielded) { 4185 schedstat_inc(rq, yld_count); 4186 /* 4187 * Make p's CPU reschedule; pick_next_entity takes care of 4188 * fairness. 4189 */ 4190 if (preempt && rq != p_rq) 4191 resched_task(p_rq->curr); 4192 } 4193 4194out_unlock: 4195 double_rq_unlock(rq, p_rq); 4196out_irq: 4197 local_irq_restore(flags); 4198 4199 if (yielded > 0) 4200 schedule(); 4201 4202 return yielded; 4203} 4204EXPORT_SYMBOL_GPL(yield_to); 4205 4206/* 4207 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4208 * that process accounting knows that this is a task in IO wait state. 4209 */ 4210void __sched io_schedule(void) 4211{ 4212 struct rq *rq = raw_rq(); 4213 4214 delayacct_blkio_start(); 4215 atomic_inc(&rq->nr_iowait); 4216 blk_flush_plug(current); 4217 current->in_iowait = 1; 4218 schedule(); 4219 current->in_iowait = 0; 4220 atomic_dec(&rq->nr_iowait); 4221 delayacct_blkio_end(); 4222} 4223EXPORT_SYMBOL(io_schedule); 4224 4225long __sched io_schedule_timeout(long timeout) 4226{ 4227 struct rq *rq = raw_rq(); 4228 long ret; 4229 4230 delayacct_blkio_start(); 4231 atomic_inc(&rq->nr_iowait); 4232 blk_flush_plug(current); 4233 current->in_iowait = 1; 4234 ret = schedule_timeout(timeout); 4235 current->in_iowait = 0; 4236 atomic_dec(&rq->nr_iowait); 4237 delayacct_blkio_end(); 4238 return ret; 4239} 4240 4241/** 4242 * sys_sched_get_priority_max - return maximum RT priority. 4243 * @policy: scheduling class. 4244 * 4245 * Return: On success, this syscall returns the maximum 4246 * rt_priority that can be used by a given scheduling class. 4247 * On failure, a negative error code is returned. 4248 */ 4249SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4250{ 4251 int ret = -EINVAL; 4252 4253 switch (policy) { 4254 case SCHED_FIFO: 4255 case SCHED_RR: 4256 ret = MAX_USER_RT_PRIO-1; 4257 break; 4258 case SCHED_DEADLINE: 4259 case SCHED_NORMAL: 4260 case SCHED_BATCH: 4261 case SCHED_IDLE: 4262 ret = 0; 4263 break; 4264 } 4265 return ret; 4266} 4267 4268/** 4269 * sys_sched_get_priority_min - return minimum RT priority. 4270 * @policy: scheduling class. 4271 * 4272 * Return: On success, this syscall returns the minimum 4273 * rt_priority that can be used by a given scheduling class. 4274 * On failure, a negative error code is returned. 4275 */ 4276SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4277{ 4278 int ret = -EINVAL; 4279 4280 switch (policy) { 4281 case SCHED_FIFO: 4282 case SCHED_RR: 4283 ret = 1; 4284 break; 4285 case SCHED_DEADLINE: 4286 case SCHED_NORMAL: 4287 case SCHED_BATCH: 4288 case SCHED_IDLE: 4289 ret = 0; 4290 } 4291 return ret; 4292} 4293 4294/** 4295 * sys_sched_rr_get_interval - return the default timeslice of a process. 4296 * @pid: pid of the process. 4297 * @interval: userspace pointer to the timeslice value. 4298 * 4299 * this syscall writes the default timeslice value of a given process 4300 * into the user-space timespec buffer. A value of '0' means infinity. 4301 * 4302 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4303 * an error code. 4304 */ 4305SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4306 struct timespec __user *, interval) 4307{ 4308 struct task_struct *p; 4309 unsigned int time_slice; 4310 unsigned long flags; 4311 struct rq *rq; 4312 int retval; 4313 struct timespec t; 4314 4315 if (pid < 0) 4316 return -EINVAL; 4317 4318 retval = -ESRCH; 4319 rcu_read_lock(); 4320 p = find_process_by_pid(pid); 4321 if (!p) 4322 goto out_unlock; 4323 4324 retval = security_task_getscheduler(p); 4325 if (retval) 4326 goto out_unlock; 4327 4328 rq = task_rq_lock(p, &flags); 4329 time_slice = 0; 4330 if (p->sched_class->get_rr_interval) 4331 time_slice = p->sched_class->get_rr_interval(rq, p); 4332 task_rq_unlock(rq, p, &flags); 4333 4334 rcu_read_unlock(); 4335 jiffies_to_timespec(time_slice, &t); 4336 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4337 return retval; 4338 4339out_unlock: 4340 rcu_read_unlock(); 4341 return retval; 4342} 4343 4344static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4345 4346void sched_show_task(struct task_struct *p) 4347{ 4348 unsigned long free = 0; 4349 int ppid; 4350 unsigned state; 4351 4352 state = p->state ? __ffs(p->state) + 1 : 0; 4353 printk(KERN_INFO "%-15.15s %c", p->comm, 4354 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4355#if BITS_PER_LONG == 32 4356 if (state == TASK_RUNNING) 4357 printk(KERN_CONT " running "); 4358 else 4359 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4360#else 4361 if (state == TASK_RUNNING) 4362 printk(KERN_CONT " running task "); 4363 else 4364 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4365#endif 4366#ifdef CONFIG_DEBUG_STACK_USAGE 4367 free = stack_not_used(p); 4368#endif 4369 rcu_read_lock(); 4370 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4371 rcu_read_unlock(); 4372 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4373 task_pid_nr(p), ppid, 4374 (unsigned long)task_thread_info(p)->flags); 4375 4376 print_worker_info(KERN_INFO, p); 4377 show_stack(p, NULL); 4378} 4379 4380void show_state_filter(unsigned long state_filter) 4381{ 4382 struct task_struct *g, *p; 4383 4384#if BITS_PER_LONG == 32 4385 printk(KERN_INFO 4386 " task PC stack pid father\n"); 4387#else 4388 printk(KERN_INFO 4389 " task PC stack pid father\n"); 4390#endif 4391 rcu_read_lock(); 4392 do_each_thread(g, p) { 4393 /* 4394 * reset the NMI-timeout, listing all files on a slow 4395 * console might take a lot of time: 4396 */ 4397 touch_nmi_watchdog(); 4398 if (!state_filter || (p->state & state_filter)) 4399 sched_show_task(p); 4400 } while_each_thread(g, p); 4401 4402 touch_all_softlockup_watchdogs(); 4403 4404#ifdef CONFIG_SCHED_DEBUG 4405 sysrq_sched_debug_show(); 4406#endif 4407 rcu_read_unlock(); 4408 /* 4409 * Only show locks if all tasks are dumped: 4410 */ 4411 if (!state_filter) 4412 debug_show_all_locks(); 4413} 4414 4415void init_idle_bootup_task(struct task_struct *idle) 4416{ 4417 idle->sched_class = &idle_sched_class; 4418} 4419 4420/** 4421 * init_idle - set up an idle thread for a given CPU 4422 * @idle: task in question 4423 * @cpu: cpu the idle task belongs to 4424 * 4425 * NOTE: this function does not set the idle thread's NEED_RESCHED 4426 * flag, to make booting more robust. 4427 */ 4428void init_idle(struct task_struct *idle, int cpu) 4429{ 4430 struct rq *rq = cpu_rq(cpu); 4431 unsigned long flags; 4432 4433 raw_spin_lock_irqsave(&rq->lock, flags); 4434 4435 __sched_fork(0, idle); 4436 idle->state = TASK_RUNNING; 4437 idle->se.exec_start = sched_clock(); 4438 4439 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4440 /* 4441 * We're having a chicken and egg problem, even though we are 4442 * holding rq->lock, the cpu isn't yet set to this cpu so the 4443 * lockdep check in task_group() will fail. 4444 * 4445 * Similar case to sched_fork(). / Alternatively we could 4446 * use task_rq_lock() here and obtain the other rq->lock. 4447 * 4448 * Silence PROVE_RCU 4449 */ 4450 rcu_read_lock(); 4451 __set_task_cpu(idle, cpu); 4452 rcu_read_unlock(); 4453 4454 rq->curr = rq->idle = idle; 4455 idle->on_rq = 1; 4456#if defined(CONFIG_SMP) 4457 idle->on_cpu = 1; 4458#endif 4459 raw_spin_unlock_irqrestore(&rq->lock, flags); 4460 4461 /* Set the preempt count _outside_ the spinlocks! */ 4462 init_idle_preempt_count(idle, cpu); 4463 4464 /* 4465 * The idle tasks have their own, simple scheduling class: 4466 */ 4467 idle->sched_class = &idle_sched_class; 4468 ftrace_graph_init_idle_task(idle, cpu); 4469 vtime_init_idle(idle, cpu); 4470#if defined(CONFIG_SMP) 4471 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4472#endif 4473} 4474 4475#ifdef CONFIG_SMP 4476void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4477{ 4478 if (p->sched_class && p->sched_class->set_cpus_allowed) 4479 p->sched_class->set_cpus_allowed(p, new_mask); 4480 4481 cpumask_copy(&p->cpus_allowed, new_mask); 4482 p->nr_cpus_allowed = cpumask_weight(new_mask); 4483} 4484 4485/* 4486 * This is how migration works: 4487 * 4488 * 1) we invoke migration_cpu_stop() on the target CPU using 4489 * stop_one_cpu(). 4490 * 2) stopper starts to run (implicitly forcing the migrated thread 4491 * off the CPU) 4492 * 3) it checks whether the migrated task is still in the wrong runqueue. 4493 * 4) if it's in the wrong runqueue then the migration thread removes 4494 * it and puts it into the right queue. 4495 * 5) stopper completes and stop_one_cpu() returns and the migration 4496 * is done. 4497 */ 4498 4499/* 4500 * Change a given task's CPU affinity. Migrate the thread to a 4501 * proper CPU and schedule it away if the CPU it's executing on 4502 * is removed from the allowed bitmask. 4503 * 4504 * NOTE: the caller must have a valid reference to the task, the 4505 * task must not exit() & deallocate itself prematurely. The 4506 * call is not atomic; no spinlocks may be held. 4507 */ 4508int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4509{ 4510 unsigned long flags; 4511 struct rq *rq; 4512 unsigned int dest_cpu; 4513 int ret = 0; 4514 4515 rq = task_rq_lock(p, &flags); 4516 4517 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4518 goto out; 4519 4520 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4521 ret = -EINVAL; 4522 goto out; 4523 } 4524 4525 do_set_cpus_allowed(p, new_mask); 4526 4527 /* Can the task run on the task's current CPU? If so, we're done */ 4528 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4529 goto out; 4530 4531 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4532 if (p->on_rq) { 4533 struct migration_arg arg = { p, dest_cpu }; 4534 /* Need help from migration thread: drop lock and wait. */ 4535 task_rq_unlock(rq, p, &flags); 4536 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4537 tlb_migrate_finish(p->mm); 4538 return 0; 4539 } 4540out: 4541 task_rq_unlock(rq, p, &flags); 4542 4543 return ret; 4544} 4545EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4546 4547/* 4548 * Move (not current) task off this cpu, onto dest cpu. We're doing 4549 * this because either it can't run here any more (set_cpus_allowed() 4550 * away from this CPU, or CPU going down), or because we're 4551 * attempting to rebalance this task on exec (sched_exec). 4552 * 4553 * So we race with normal scheduler movements, but that's OK, as long 4554 * as the task is no longer on this CPU. 4555 * 4556 * Returns non-zero if task was successfully migrated. 4557 */ 4558static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4559{ 4560 struct rq *rq_dest, *rq_src; 4561 int ret = 0; 4562 4563 if (unlikely(!cpu_active(dest_cpu))) 4564 return ret; 4565 4566 rq_src = cpu_rq(src_cpu); 4567 rq_dest = cpu_rq(dest_cpu); 4568 4569 raw_spin_lock(&p->pi_lock); 4570 double_rq_lock(rq_src, rq_dest); 4571 /* Already moved. */ 4572 if (task_cpu(p) != src_cpu) 4573 goto done; 4574 /* Affinity changed (again). */ 4575 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4576 goto fail; 4577 4578 /* 4579 * If we're not on a rq, the next wake-up will ensure we're 4580 * placed properly. 4581 */ 4582 if (p->on_rq) { 4583 dequeue_task(rq_src, p, 0); 4584 set_task_cpu(p, dest_cpu); 4585 enqueue_task(rq_dest, p, 0); 4586 check_preempt_curr(rq_dest, p, 0); 4587 } 4588done: 4589 ret = 1; 4590fail: 4591 double_rq_unlock(rq_src, rq_dest); 4592 raw_spin_unlock(&p->pi_lock); 4593 return ret; 4594} 4595 4596#ifdef CONFIG_NUMA_BALANCING 4597/* Migrate current task p to target_cpu */ 4598int migrate_task_to(struct task_struct *p, int target_cpu) 4599{ 4600 struct migration_arg arg = { p, target_cpu }; 4601 int curr_cpu = task_cpu(p); 4602 4603 if (curr_cpu == target_cpu) 4604 return 0; 4605 4606 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4607 return -EINVAL; 4608 4609 /* TODO: This is not properly updating schedstats */ 4610 4611 trace_sched_move_numa(p, curr_cpu, target_cpu); 4612 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4613} 4614 4615/* 4616 * Requeue a task on a given node and accurately track the number of NUMA 4617 * tasks on the runqueues 4618 */ 4619void sched_setnuma(struct task_struct *p, int nid) 4620{ 4621 struct rq *rq; 4622 unsigned long flags; 4623 bool on_rq, running; 4624 4625 rq = task_rq_lock(p, &flags); 4626 on_rq = p->on_rq; 4627 running = task_current(rq, p); 4628 4629 if (on_rq) 4630 dequeue_task(rq, p, 0); 4631 if (running) 4632 p->sched_class->put_prev_task(rq, p); 4633 4634 p->numa_preferred_nid = nid; 4635 4636 if (running) 4637 p->sched_class->set_curr_task(rq); 4638 if (on_rq) 4639 enqueue_task(rq, p, 0); 4640 task_rq_unlock(rq, p, &flags); 4641} 4642#endif 4643 4644/* 4645 * migration_cpu_stop - this will be executed by a highprio stopper thread 4646 * and performs thread migration by bumping thread off CPU then 4647 * 'pushing' onto another runqueue. 4648 */ 4649static int migration_cpu_stop(void *data) 4650{ 4651 struct migration_arg *arg = data; 4652 4653 /* 4654 * The original target cpu might have gone down and we might 4655 * be on another cpu but it doesn't matter. 4656 */ 4657 local_irq_disable(); 4658 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4659 local_irq_enable(); 4660 return 0; 4661} 4662 4663#ifdef CONFIG_HOTPLUG_CPU 4664 4665/* 4666 * Ensures that the idle task is using init_mm right before its cpu goes 4667 * offline. 4668 */ 4669void idle_task_exit(void) 4670{ 4671 struct mm_struct *mm = current->active_mm; 4672 4673 BUG_ON(cpu_online(smp_processor_id())); 4674 4675 if (mm != &init_mm) { 4676 switch_mm(mm, &init_mm, current); 4677 finish_arch_post_lock_switch(); 4678 } 4679 mmdrop(mm); 4680} 4681 4682/* 4683 * Since this CPU is going 'away' for a while, fold any nr_active delta 4684 * we might have. Assumes we're called after migrate_tasks() so that the 4685 * nr_active count is stable. 4686 * 4687 * Also see the comment "Global load-average calculations". 4688 */ 4689static void calc_load_migrate(struct rq *rq) 4690{ 4691 long delta = calc_load_fold_active(rq); 4692 if (delta) 4693 atomic_long_add(delta, &calc_load_tasks); 4694} 4695 4696static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4697{ 4698} 4699 4700static const struct sched_class fake_sched_class = { 4701 .put_prev_task = put_prev_task_fake, 4702}; 4703 4704static struct task_struct fake_task = { 4705 /* 4706 * Avoid pull_{rt,dl}_task() 4707 */ 4708 .prio = MAX_PRIO + 1, 4709 .sched_class = &fake_sched_class, 4710}; 4711 4712/* 4713 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4714 * try_to_wake_up()->select_task_rq(). 4715 * 4716 * Called with rq->lock held even though we'er in stop_machine() and 4717 * there's no concurrency possible, we hold the required locks anyway 4718 * because of lock validation efforts. 4719 */ 4720static void migrate_tasks(unsigned int dead_cpu) 4721{ 4722 struct rq *rq = cpu_rq(dead_cpu); 4723 struct task_struct *next, *stop = rq->stop; 4724 int dest_cpu; 4725 4726 /* 4727 * Fudge the rq selection such that the below task selection loop 4728 * doesn't get stuck on the currently eligible stop task. 4729 * 4730 * We're currently inside stop_machine() and the rq is either stuck 4731 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4732 * either way we should never end up calling schedule() until we're 4733 * done here. 4734 */ 4735 rq->stop = NULL; 4736 4737 /* 4738 * put_prev_task() and pick_next_task() sched 4739 * class method both need to have an up-to-date 4740 * value of rq->clock[_task] 4741 */ 4742 update_rq_clock(rq); 4743 4744 for ( ; ; ) { 4745 /* 4746 * There's this thread running, bail when that's the only 4747 * remaining thread. 4748 */ 4749 if (rq->nr_running == 1) 4750 break; 4751 4752 next = pick_next_task(rq, &fake_task); 4753 BUG_ON(!next); 4754 next->sched_class->put_prev_task(rq, next); 4755 4756 /* Find suitable destination for @next, with force if needed. */ 4757 dest_cpu = select_fallback_rq(dead_cpu, next); 4758 raw_spin_unlock(&rq->lock); 4759 4760 __migrate_task(next, dead_cpu, dest_cpu); 4761 4762 raw_spin_lock(&rq->lock); 4763 } 4764 4765 rq->stop = stop; 4766} 4767 4768#endif /* CONFIG_HOTPLUG_CPU */ 4769 4770#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4771 4772static struct ctl_table sd_ctl_dir[] = { 4773 { 4774 .procname = "sched_domain", 4775 .mode = 0555, 4776 }, 4777 {} 4778}; 4779 4780static struct ctl_table sd_ctl_root[] = { 4781 { 4782 .procname = "kernel", 4783 .mode = 0555, 4784 .child = sd_ctl_dir, 4785 }, 4786 {} 4787}; 4788 4789static struct ctl_table *sd_alloc_ctl_entry(int n) 4790{ 4791 struct ctl_table *entry = 4792 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4793 4794 return entry; 4795} 4796 4797static void sd_free_ctl_entry(struct ctl_table **tablep) 4798{ 4799 struct ctl_table *entry; 4800 4801 /* 4802 * In the intermediate directories, both the child directory and 4803 * procname are dynamically allocated and could fail but the mode 4804 * will always be set. In the lowest directory the names are 4805 * static strings and all have proc handlers. 4806 */ 4807 for (entry = *tablep; entry->mode; entry++) { 4808 if (entry->child) 4809 sd_free_ctl_entry(&entry->child); 4810 if (entry->proc_handler == NULL) 4811 kfree(entry->procname); 4812 } 4813 4814 kfree(*tablep); 4815 *tablep = NULL; 4816} 4817 4818static int min_load_idx = 0; 4819static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4820 4821static void 4822set_table_entry(struct ctl_table *entry, 4823 const char *procname, void *data, int maxlen, 4824 umode_t mode, proc_handler *proc_handler, 4825 bool load_idx) 4826{ 4827 entry->procname = procname; 4828 entry->data = data; 4829 entry->maxlen = maxlen; 4830 entry->mode = mode; 4831 entry->proc_handler = proc_handler; 4832 4833 if (load_idx) { 4834 entry->extra1 = &min_load_idx; 4835 entry->extra2 = &max_load_idx; 4836 } 4837} 4838 4839static struct ctl_table * 4840sd_alloc_ctl_domain_table(struct sched_domain *sd) 4841{ 4842 struct ctl_table *table = sd_alloc_ctl_entry(14); 4843 4844 if (table == NULL) 4845 return NULL; 4846 4847 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4848 sizeof(long), 0644, proc_doulongvec_minmax, false); 4849 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4850 sizeof(long), 0644, proc_doulongvec_minmax, false); 4851 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4852 sizeof(int), 0644, proc_dointvec_minmax, true); 4853 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4854 sizeof(int), 0644, proc_dointvec_minmax, true); 4855 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4856 sizeof(int), 0644, proc_dointvec_minmax, true); 4857 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4858 sizeof(int), 0644, proc_dointvec_minmax, true); 4859 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4860 sizeof(int), 0644, proc_dointvec_minmax, true); 4861 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4862 sizeof(int), 0644, proc_dointvec_minmax, false); 4863 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4864 sizeof(int), 0644, proc_dointvec_minmax, false); 4865 set_table_entry(&table[9], "cache_nice_tries", 4866 &sd->cache_nice_tries, 4867 sizeof(int), 0644, proc_dointvec_minmax, false); 4868 set_table_entry(&table[10], "flags", &sd->flags, 4869 sizeof(int), 0644, proc_dointvec_minmax, false); 4870 set_table_entry(&table[11], "max_newidle_lb_cost", 4871 &sd->max_newidle_lb_cost, 4872 sizeof(long), 0644, proc_doulongvec_minmax, false); 4873 set_table_entry(&table[12], "name", sd->name, 4874 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4875 /* &table[13] is terminator */ 4876 4877 return table; 4878} 4879 4880static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4881{ 4882 struct ctl_table *entry, *table; 4883 struct sched_domain *sd; 4884 int domain_num = 0, i; 4885 char buf[32]; 4886 4887 for_each_domain(cpu, sd) 4888 domain_num++; 4889 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4890 if (table == NULL) 4891 return NULL; 4892 4893 i = 0; 4894 for_each_domain(cpu, sd) { 4895 snprintf(buf, 32, "domain%d", i); 4896 entry->procname = kstrdup(buf, GFP_KERNEL); 4897 entry->mode = 0555; 4898 entry->child = sd_alloc_ctl_domain_table(sd); 4899 entry++; 4900 i++; 4901 } 4902 return table; 4903} 4904 4905static struct ctl_table_header *sd_sysctl_header; 4906static void register_sched_domain_sysctl(void) 4907{ 4908 int i, cpu_num = num_possible_cpus(); 4909 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4910 char buf[32]; 4911 4912 WARN_ON(sd_ctl_dir[0].child); 4913 sd_ctl_dir[0].child = entry; 4914 4915 if (entry == NULL) 4916 return; 4917 4918 for_each_possible_cpu(i) { 4919 snprintf(buf, 32, "cpu%d", i); 4920 entry->procname = kstrdup(buf, GFP_KERNEL); 4921 entry->mode = 0555; 4922 entry->child = sd_alloc_ctl_cpu_table(i); 4923 entry++; 4924 } 4925 4926 WARN_ON(sd_sysctl_header); 4927 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4928} 4929 4930/* may be called multiple times per register */ 4931static void unregister_sched_domain_sysctl(void) 4932{ 4933 if (sd_sysctl_header) 4934 unregister_sysctl_table(sd_sysctl_header); 4935 sd_sysctl_header = NULL; 4936 if (sd_ctl_dir[0].child) 4937 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4938} 4939#else 4940static void register_sched_domain_sysctl(void) 4941{ 4942} 4943static void unregister_sched_domain_sysctl(void) 4944{ 4945} 4946#endif 4947 4948static void set_rq_online(struct rq *rq) 4949{ 4950 if (!rq->online) { 4951 const struct sched_class *class; 4952 4953 cpumask_set_cpu(rq->cpu, rq->rd->online); 4954 rq->online = 1; 4955 4956 for_each_class(class) { 4957 if (class->rq_online) 4958 class->rq_online(rq); 4959 } 4960 } 4961} 4962 4963static void set_rq_offline(struct rq *rq) 4964{ 4965 if (rq->online) { 4966 const struct sched_class *class; 4967 4968 for_each_class(class) { 4969 if (class->rq_offline) 4970 class->rq_offline(rq); 4971 } 4972 4973 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4974 rq->online = 0; 4975 } 4976} 4977 4978/* 4979 * migration_call - callback that gets triggered when a CPU is added. 4980 * Here we can start up the necessary migration thread for the new CPU. 4981 */ 4982static int 4983migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4984{ 4985 int cpu = (long)hcpu; 4986 unsigned long flags; 4987 struct rq *rq = cpu_rq(cpu); 4988 4989 switch (action & ~CPU_TASKS_FROZEN) { 4990 4991 case CPU_UP_PREPARE: 4992 rq->calc_load_update = calc_load_update; 4993 break; 4994 4995 case CPU_ONLINE: 4996 /* Update our root-domain */ 4997 raw_spin_lock_irqsave(&rq->lock, flags); 4998 if (rq->rd) { 4999 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5000 5001 set_rq_online(rq); 5002 } 5003 raw_spin_unlock_irqrestore(&rq->lock, flags); 5004 break; 5005 5006#ifdef CONFIG_HOTPLUG_CPU 5007 case CPU_DYING: 5008 sched_ttwu_pending(); 5009 /* Update our root-domain */ 5010 raw_spin_lock_irqsave(&rq->lock, flags); 5011 if (rq->rd) { 5012 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5013 set_rq_offline(rq); 5014 } 5015 migrate_tasks(cpu); 5016 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5017 raw_spin_unlock_irqrestore(&rq->lock, flags); 5018 break; 5019 5020 case CPU_DEAD: 5021 calc_load_migrate(rq); 5022 break; 5023#endif 5024 } 5025 5026 update_max_interval(); 5027 5028 return NOTIFY_OK; 5029} 5030 5031/* 5032 * Register at high priority so that task migration (migrate_all_tasks) 5033 * happens before everything else. This has to be lower priority than 5034 * the notifier in the perf_event subsystem, though. 5035 */ 5036static struct notifier_block migration_notifier = { 5037 .notifier_call = migration_call, 5038 .priority = CPU_PRI_MIGRATION, 5039}; 5040 5041static int sched_cpu_active(struct notifier_block *nfb, 5042 unsigned long action, void *hcpu) 5043{ 5044 switch (action & ~CPU_TASKS_FROZEN) { 5045 case CPU_STARTING: 5046 case CPU_DOWN_FAILED: 5047 set_cpu_active((long)hcpu, true); 5048 return NOTIFY_OK; 5049 default: 5050 return NOTIFY_DONE; 5051 } 5052} 5053 5054static int sched_cpu_inactive(struct notifier_block *nfb, 5055 unsigned long action, void *hcpu) 5056{ 5057 unsigned long flags; 5058 long cpu = (long)hcpu; 5059 5060 switch (action & ~CPU_TASKS_FROZEN) { 5061 case CPU_DOWN_PREPARE: 5062 set_cpu_active(cpu, false); 5063 5064 /* explicitly allow suspend */ 5065 if (!(action & CPU_TASKS_FROZEN)) { 5066 struct dl_bw *dl_b = dl_bw_of(cpu); 5067 bool overflow; 5068 int cpus; 5069 5070 raw_spin_lock_irqsave(&dl_b->lock, flags); 5071 cpus = dl_bw_cpus(cpu); 5072 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5073 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5074 5075 if (overflow) 5076 return notifier_from_errno(-EBUSY); 5077 } 5078 return NOTIFY_OK; 5079 } 5080 5081 return NOTIFY_DONE; 5082} 5083 5084static int __init migration_init(void) 5085{ 5086 void *cpu = (void *)(long)smp_processor_id(); 5087 int err; 5088 5089 /* Initialize migration for the boot CPU */ 5090 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5091 BUG_ON(err == NOTIFY_BAD); 5092 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5093 register_cpu_notifier(&migration_notifier); 5094 5095 /* Register cpu active notifiers */ 5096 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5097 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5098 5099 return 0; 5100} 5101early_initcall(migration_init); 5102#endif 5103 5104#ifdef CONFIG_SMP 5105 5106static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5107 5108#ifdef CONFIG_SCHED_DEBUG 5109 5110static __read_mostly int sched_debug_enabled; 5111 5112static int __init sched_debug_setup(char *str) 5113{ 5114 sched_debug_enabled = 1; 5115 5116 return 0; 5117} 5118early_param("sched_debug", sched_debug_setup); 5119 5120static inline bool sched_debug(void) 5121{ 5122 return sched_debug_enabled; 5123} 5124 5125static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5126 struct cpumask *groupmask) 5127{ 5128 struct sched_group *group = sd->groups; 5129 char str[256]; 5130 5131 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5132 cpumask_clear(groupmask); 5133 5134 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5135 5136 if (!(sd->flags & SD_LOAD_BALANCE)) { 5137 printk("does not load-balance\n"); 5138 if (sd->parent) 5139 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5140 " has parent"); 5141 return -1; 5142 } 5143 5144 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5145 5146 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5147 printk(KERN_ERR "ERROR: domain->span does not contain " 5148 "CPU%d\n", cpu); 5149 } 5150 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5151 printk(KERN_ERR "ERROR: domain->groups does not contain" 5152 " CPU%d\n", cpu); 5153 } 5154 5155 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5156 do { 5157 if (!group) { 5158 printk("\n"); 5159 printk(KERN_ERR "ERROR: group is NULL\n"); 5160 break; 5161 } 5162 5163 /* 5164 * Even though we initialize ->power to something semi-sane, 5165 * we leave power_orig unset. This allows us to detect if 5166 * domain iteration is still funny without causing /0 traps. 5167 */ 5168 if (!group->sgp->power_orig) { 5169 printk(KERN_CONT "\n"); 5170 printk(KERN_ERR "ERROR: domain->cpu_power not " 5171 "set\n"); 5172 break; 5173 } 5174 5175 if (!cpumask_weight(sched_group_cpus(group))) { 5176 printk(KERN_CONT "\n"); 5177 printk(KERN_ERR "ERROR: empty group\n"); 5178 break; 5179 } 5180 5181 if (!(sd->flags & SD_OVERLAP) && 5182 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5183 printk(KERN_CONT "\n"); 5184 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5185 break; 5186 } 5187 5188 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5189 5190 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5191 5192 printk(KERN_CONT " %s", str); 5193 if (group->sgp->power != SCHED_POWER_SCALE) { 5194 printk(KERN_CONT " (cpu_power = %d)", 5195 group->sgp->power); 5196 } 5197 5198 group = group->next; 5199 } while (group != sd->groups); 5200 printk(KERN_CONT "\n"); 5201 5202 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5203 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5204 5205 if (sd->parent && 5206 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5207 printk(KERN_ERR "ERROR: parent span is not a superset " 5208 "of domain->span\n"); 5209 return 0; 5210} 5211 5212static void sched_domain_debug(struct sched_domain *sd, int cpu) 5213{ 5214 int level = 0; 5215 5216 if (!sched_debug_enabled) 5217 return; 5218 5219 if (!sd) { 5220 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5221 return; 5222 } 5223 5224 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5225 5226 for (;;) { 5227 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5228 break; 5229 level++; 5230 sd = sd->parent; 5231 if (!sd) 5232 break; 5233 } 5234} 5235#else /* !CONFIG_SCHED_DEBUG */ 5236# define sched_domain_debug(sd, cpu) do { } while (0) 5237static inline bool sched_debug(void) 5238{ 5239 return false; 5240} 5241#endif /* CONFIG_SCHED_DEBUG */ 5242 5243static int sd_degenerate(struct sched_domain *sd) 5244{ 5245 if (cpumask_weight(sched_domain_span(sd)) == 1) 5246 return 1; 5247 5248 /* Following flags need at least 2 groups */ 5249 if (sd->flags & (SD_LOAD_BALANCE | 5250 SD_BALANCE_NEWIDLE | 5251 SD_BALANCE_FORK | 5252 SD_BALANCE_EXEC | 5253 SD_SHARE_CPUPOWER | 5254 SD_SHARE_PKG_RESOURCES)) { 5255 if (sd->groups != sd->groups->next) 5256 return 0; 5257 } 5258 5259 /* Following flags don't use groups */ 5260 if (sd->flags & (SD_WAKE_AFFINE)) 5261 return 0; 5262 5263 return 1; 5264} 5265 5266static int 5267sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5268{ 5269 unsigned long cflags = sd->flags, pflags = parent->flags; 5270 5271 if (sd_degenerate(parent)) 5272 return 1; 5273 5274 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5275 return 0; 5276 5277 /* Flags needing groups don't count if only 1 group in parent */ 5278 if (parent->groups == parent->groups->next) { 5279 pflags &= ~(SD_LOAD_BALANCE | 5280 SD_BALANCE_NEWIDLE | 5281 SD_BALANCE_FORK | 5282 SD_BALANCE_EXEC | 5283 SD_SHARE_CPUPOWER | 5284 SD_SHARE_PKG_RESOURCES | 5285 SD_PREFER_SIBLING); 5286 if (nr_node_ids == 1) 5287 pflags &= ~SD_SERIALIZE; 5288 } 5289 if (~cflags & pflags) 5290 return 0; 5291 5292 return 1; 5293} 5294 5295static void free_rootdomain(struct rcu_head *rcu) 5296{ 5297 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5298 5299 cpupri_cleanup(&rd->cpupri); 5300 cpudl_cleanup(&rd->cpudl); 5301 free_cpumask_var(rd->dlo_mask); 5302 free_cpumask_var(rd->rto_mask); 5303 free_cpumask_var(rd->online); 5304 free_cpumask_var(rd->span); 5305 kfree(rd); 5306} 5307 5308static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5309{ 5310 struct root_domain *old_rd = NULL; 5311 unsigned long flags; 5312 5313 raw_spin_lock_irqsave(&rq->lock, flags); 5314 5315 if (rq->rd) { 5316 old_rd = rq->rd; 5317 5318 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5319 set_rq_offline(rq); 5320 5321 cpumask_clear_cpu(rq->cpu, old_rd->span); 5322 5323 /* 5324 * If we dont want to free the old_rd yet then 5325 * set old_rd to NULL to skip the freeing later 5326 * in this function: 5327 */ 5328 if (!atomic_dec_and_test(&old_rd->refcount)) 5329 old_rd = NULL; 5330 } 5331 5332 atomic_inc(&rd->refcount); 5333 rq->rd = rd; 5334 5335 cpumask_set_cpu(rq->cpu, rd->span); 5336 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5337 set_rq_online(rq); 5338 5339 raw_spin_unlock_irqrestore(&rq->lock, flags); 5340 5341 if (old_rd) 5342 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5343} 5344 5345static int init_rootdomain(struct root_domain *rd) 5346{ 5347 memset(rd, 0, sizeof(*rd)); 5348 5349 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5350 goto out; 5351 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5352 goto free_span; 5353 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5354 goto free_online; 5355 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5356 goto free_dlo_mask; 5357 5358 init_dl_bw(&rd->dl_bw); 5359 if (cpudl_init(&rd->cpudl) != 0) 5360 goto free_dlo_mask; 5361 5362 if (cpupri_init(&rd->cpupri) != 0) 5363 goto free_rto_mask; 5364 return 0; 5365 5366free_rto_mask: 5367 free_cpumask_var(rd->rto_mask); 5368free_dlo_mask: 5369 free_cpumask_var(rd->dlo_mask); 5370free_online: 5371 free_cpumask_var(rd->online); 5372free_span: 5373 free_cpumask_var(rd->span); 5374out: 5375 return -ENOMEM; 5376} 5377 5378/* 5379 * By default the system creates a single root-domain with all cpus as 5380 * members (mimicking the global state we have today). 5381 */ 5382struct root_domain def_root_domain; 5383 5384static void init_defrootdomain(void) 5385{ 5386 init_rootdomain(&def_root_domain); 5387 5388 atomic_set(&def_root_domain.refcount, 1); 5389} 5390 5391static struct root_domain *alloc_rootdomain(void) 5392{ 5393 struct root_domain *rd; 5394 5395 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5396 if (!rd) 5397 return NULL; 5398 5399 if (init_rootdomain(rd) != 0) { 5400 kfree(rd); 5401 return NULL; 5402 } 5403 5404 return rd; 5405} 5406 5407static void free_sched_groups(struct sched_group *sg, int free_sgp) 5408{ 5409 struct sched_group *tmp, *first; 5410 5411 if (!sg) 5412 return; 5413 5414 first = sg; 5415 do { 5416 tmp = sg->next; 5417 5418 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5419 kfree(sg->sgp); 5420 5421 kfree(sg); 5422 sg = tmp; 5423 } while (sg != first); 5424} 5425 5426static void free_sched_domain(struct rcu_head *rcu) 5427{ 5428 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5429 5430 /* 5431 * If its an overlapping domain it has private groups, iterate and 5432 * nuke them all. 5433 */ 5434 if (sd->flags & SD_OVERLAP) { 5435 free_sched_groups(sd->groups, 1); 5436 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5437 kfree(sd->groups->sgp); 5438 kfree(sd->groups); 5439 } 5440 kfree(sd); 5441} 5442 5443static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5444{ 5445 call_rcu(&sd->rcu, free_sched_domain); 5446} 5447 5448static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5449{ 5450 for (; sd; sd = sd->parent) 5451 destroy_sched_domain(sd, cpu); 5452} 5453 5454/* 5455 * Keep a special pointer to the highest sched_domain that has 5456 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5457 * allows us to avoid some pointer chasing select_idle_sibling(). 5458 * 5459 * Also keep a unique ID per domain (we use the first cpu number in 5460 * the cpumask of the domain), this allows us to quickly tell if 5461 * two cpus are in the same cache domain, see cpus_share_cache(). 5462 */ 5463DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5464DEFINE_PER_CPU(int, sd_llc_size); 5465DEFINE_PER_CPU(int, sd_llc_id); 5466DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5467DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5468DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5469 5470static void update_top_cache_domain(int cpu) 5471{ 5472 struct sched_domain *sd; 5473 struct sched_domain *busy_sd = NULL; 5474 int id = cpu; 5475 int size = 1; 5476 5477 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5478 if (sd) { 5479 id = cpumask_first(sched_domain_span(sd)); 5480 size = cpumask_weight(sched_domain_span(sd)); 5481 busy_sd = sd->parent; /* sd_busy */ 5482 } 5483 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5484 5485 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5486 per_cpu(sd_llc_size, cpu) = size; 5487 per_cpu(sd_llc_id, cpu) = id; 5488 5489 sd = lowest_flag_domain(cpu, SD_NUMA); 5490 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5491 5492 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5493 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5494} 5495 5496/* 5497 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5498 * hold the hotplug lock. 5499 */ 5500static void 5501cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5502{ 5503 struct rq *rq = cpu_rq(cpu); 5504 struct sched_domain *tmp; 5505 5506 /* Remove the sched domains which do not contribute to scheduling. */ 5507 for (tmp = sd; tmp; ) { 5508 struct sched_domain *parent = tmp->parent; 5509 if (!parent) 5510 break; 5511 5512 if (sd_parent_degenerate(tmp, parent)) { 5513 tmp->parent = parent->parent; 5514 if (parent->parent) 5515 parent->parent->child = tmp; 5516 /* 5517 * Transfer SD_PREFER_SIBLING down in case of a 5518 * degenerate parent; the spans match for this 5519 * so the property transfers. 5520 */ 5521 if (parent->flags & SD_PREFER_SIBLING) 5522 tmp->flags |= SD_PREFER_SIBLING; 5523 destroy_sched_domain(parent, cpu); 5524 } else 5525 tmp = tmp->parent; 5526 } 5527 5528 if (sd && sd_degenerate(sd)) { 5529 tmp = sd; 5530 sd = sd->parent; 5531 destroy_sched_domain(tmp, cpu); 5532 if (sd) 5533 sd->child = NULL; 5534 } 5535 5536 sched_domain_debug(sd, cpu); 5537 5538 rq_attach_root(rq, rd); 5539 tmp = rq->sd; 5540 rcu_assign_pointer(rq->sd, sd); 5541 destroy_sched_domains(tmp, cpu); 5542 5543 update_top_cache_domain(cpu); 5544} 5545 5546/* cpus with isolated domains */ 5547static cpumask_var_t cpu_isolated_map; 5548 5549/* Setup the mask of cpus configured for isolated domains */ 5550static int __init isolated_cpu_setup(char *str) 5551{ 5552 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5553 cpulist_parse(str, cpu_isolated_map); 5554 return 1; 5555} 5556 5557__setup("isolcpus=", isolated_cpu_setup); 5558 5559static const struct cpumask *cpu_cpu_mask(int cpu) 5560{ 5561 return cpumask_of_node(cpu_to_node(cpu)); 5562} 5563 5564struct sd_data { 5565 struct sched_domain **__percpu sd; 5566 struct sched_group **__percpu sg; 5567 struct sched_group_power **__percpu sgp; 5568}; 5569 5570struct s_data { 5571 struct sched_domain ** __percpu sd; 5572 struct root_domain *rd; 5573}; 5574 5575enum s_alloc { 5576 sa_rootdomain, 5577 sa_sd, 5578 sa_sd_storage, 5579 sa_none, 5580}; 5581 5582struct sched_domain_topology_level; 5583 5584typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5585typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5586 5587#define SDTL_OVERLAP 0x01 5588 5589struct sched_domain_topology_level { 5590 sched_domain_init_f init; 5591 sched_domain_mask_f mask; 5592 int flags; 5593 int numa_level; 5594 struct sd_data data; 5595}; 5596 5597/* 5598 * Build an iteration mask that can exclude certain CPUs from the upwards 5599 * domain traversal. 5600 * 5601 * Asymmetric node setups can result in situations where the domain tree is of 5602 * unequal depth, make sure to skip domains that already cover the entire 5603 * range. 5604 * 5605 * In that case build_sched_domains() will have terminated the iteration early 5606 * and our sibling sd spans will be empty. Domains should always include the 5607 * cpu they're built on, so check that. 5608 * 5609 */ 5610static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5611{ 5612 const struct cpumask *span = sched_domain_span(sd); 5613 struct sd_data *sdd = sd->private; 5614 struct sched_domain *sibling; 5615 int i; 5616 5617 for_each_cpu(i, span) { 5618 sibling = *per_cpu_ptr(sdd->sd, i); 5619 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5620 continue; 5621 5622 cpumask_set_cpu(i, sched_group_mask(sg)); 5623 } 5624} 5625 5626/* 5627 * Return the canonical balance cpu for this group, this is the first cpu 5628 * of this group that's also in the iteration mask. 5629 */ 5630int group_balance_cpu(struct sched_group *sg) 5631{ 5632 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5633} 5634 5635static int 5636build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5637{ 5638 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5639 const struct cpumask *span = sched_domain_span(sd); 5640 struct cpumask *covered = sched_domains_tmpmask; 5641 struct sd_data *sdd = sd->private; 5642 struct sched_domain *child; 5643 int i; 5644 5645 cpumask_clear(covered); 5646 5647 for_each_cpu(i, span) { 5648 struct cpumask *sg_span; 5649 5650 if (cpumask_test_cpu(i, covered)) 5651 continue; 5652 5653 child = *per_cpu_ptr(sdd->sd, i); 5654 5655 /* See the comment near build_group_mask(). */ 5656 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5657 continue; 5658 5659 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5660 GFP_KERNEL, cpu_to_node(cpu)); 5661 5662 if (!sg) 5663 goto fail; 5664 5665 sg_span = sched_group_cpus(sg); 5666 if (child->child) { 5667 child = child->child; 5668 cpumask_copy(sg_span, sched_domain_span(child)); 5669 } else 5670 cpumask_set_cpu(i, sg_span); 5671 5672 cpumask_or(covered, covered, sg_span); 5673 5674 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5675 if (atomic_inc_return(&sg->sgp->ref) == 1) 5676 build_group_mask(sd, sg); 5677 5678 /* 5679 * Initialize sgp->power such that even if we mess up the 5680 * domains and no possible iteration will get us here, we won't 5681 * die on a /0 trap. 5682 */ 5683 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5684 sg->sgp->power_orig = sg->sgp->power; 5685 5686 /* 5687 * Make sure the first group of this domain contains the 5688 * canonical balance cpu. Otherwise the sched_domain iteration 5689 * breaks. See update_sg_lb_stats(). 5690 */ 5691 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5692 group_balance_cpu(sg) == cpu) 5693 groups = sg; 5694 5695 if (!first) 5696 first = sg; 5697 if (last) 5698 last->next = sg; 5699 last = sg; 5700 last->next = first; 5701 } 5702 sd->groups = groups; 5703 5704 return 0; 5705 5706fail: 5707 free_sched_groups(first, 0); 5708 5709 return -ENOMEM; 5710} 5711 5712static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5713{ 5714 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5715 struct sched_domain *child = sd->child; 5716 5717 if (child) 5718 cpu = cpumask_first(sched_domain_span(child)); 5719 5720 if (sg) { 5721 *sg = *per_cpu_ptr(sdd->sg, cpu); 5722 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5723 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5724 } 5725 5726 return cpu; 5727} 5728 5729/* 5730 * build_sched_groups will build a circular linked list of the groups 5731 * covered by the given span, and will set each group's ->cpumask correctly, 5732 * and ->cpu_power to 0. 5733 * 5734 * Assumes the sched_domain tree is fully constructed 5735 */ 5736static int 5737build_sched_groups(struct sched_domain *sd, int cpu) 5738{ 5739 struct sched_group *first = NULL, *last = NULL; 5740 struct sd_data *sdd = sd->private; 5741 const struct cpumask *span = sched_domain_span(sd); 5742 struct cpumask *covered; 5743 int i; 5744 5745 get_group(cpu, sdd, &sd->groups); 5746 atomic_inc(&sd->groups->ref); 5747 5748 if (cpu != cpumask_first(span)) 5749 return 0; 5750 5751 lockdep_assert_held(&sched_domains_mutex); 5752 covered = sched_domains_tmpmask; 5753 5754 cpumask_clear(covered); 5755 5756 for_each_cpu(i, span) { 5757 struct sched_group *sg; 5758 int group, j; 5759 5760 if (cpumask_test_cpu(i, covered)) 5761 continue; 5762 5763 group = get_group(i, sdd, &sg); 5764 cpumask_clear(sched_group_cpus(sg)); 5765 sg->sgp->power = 0; 5766 cpumask_setall(sched_group_mask(sg)); 5767 5768 for_each_cpu(j, span) { 5769 if (get_group(j, sdd, NULL) != group) 5770 continue; 5771 5772 cpumask_set_cpu(j, covered); 5773 cpumask_set_cpu(j, sched_group_cpus(sg)); 5774 } 5775 5776 if (!first) 5777 first = sg; 5778 if (last) 5779 last->next = sg; 5780 last = sg; 5781 } 5782 last->next = first; 5783 5784 return 0; 5785} 5786 5787/* 5788 * Initialize sched groups cpu_power. 5789 * 5790 * cpu_power indicates the capacity of sched group, which is used while 5791 * distributing the load between different sched groups in a sched domain. 5792 * Typically cpu_power for all the groups in a sched domain will be same unless 5793 * there are asymmetries in the topology. If there are asymmetries, group 5794 * having more cpu_power will pickup more load compared to the group having 5795 * less cpu_power. 5796 */ 5797static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5798{ 5799 struct sched_group *sg = sd->groups; 5800 5801 WARN_ON(!sg); 5802 5803 do { 5804 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5805 sg = sg->next; 5806 } while (sg != sd->groups); 5807 5808 if (cpu != group_balance_cpu(sg)) 5809 return; 5810 5811 update_group_power(sd, cpu); 5812 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5813} 5814 5815int __weak arch_sd_sibling_asym_packing(void) 5816{ 5817 return 0*SD_ASYM_PACKING; 5818} 5819 5820/* 5821 * Initializers for schedule domains 5822 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5823 */ 5824 5825#ifdef CONFIG_SCHED_DEBUG 5826# define SD_INIT_NAME(sd, type) sd->name = #type 5827#else 5828# define SD_INIT_NAME(sd, type) do { } while (0) 5829#endif 5830 5831#define SD_INIT_FUNC(type) \ 5832static noinline struct sched_domain * \ 5833sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5834{ \ 5835 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5836 *sd = SD_##type##_INIT; \ 5837 SD_INIT_NAME(sd, type); \ 5838 sd->private = &tl->data; \ 5839 return sd; \ 5840} 5841 5842SD_INIT_FUNC(CPU) 5843#ifdef CONFIG_SCHED_SMT 5844 SD_INIT_FUNC(SIBLING) 5845#endif 5846#ifdef CONFIG_SCHED_MC 5847 SD_INIT_FUNC(MC) 5848#endif 5849#ifdef CONFIG_SCHED_BOOK 5850 SD_INIT_FUNC(BOOK) 5851#endif 5852 5853static int default_relax_domain_level = -1; 5854int sched_domain_level_max; 5855 5856static int __init setup_relax_domain_level(char *str) 5857{ 5858 if (kstrtoint(str, 0, &default_relax_domain_level)) 5859 pr_warn("Unable to set relax_domain_level\n"); 5860 5861 return 1; 5862} 5863__setup("relax_domain_level=", setup_relax_domain_level); 5864 5865static void set_domain_attribute(struct sched_domain *sd, 5866 struct sched_domain_attr *attr) 5867{ 5868 int request; 5869 5870 if (!attr || attr->relax_domain_level < 0) { 5871 if (default_relax_domain_level < 0) 5872 return; 5873 else 5874 request = default_relax_domain_level; 5875 } else 5876 request = attr->relax_domain_level; 5877 if (request < sd->level) { 5878 /* turn off idle balance on this domain */ 5879 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5880 } else { 5881 /* turn on idle balance on this domain */ 5882 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5883 } 5884} 5885 5886static void __sdt_free(const struct cpumask *cpu_map); 5887static int __sdt_alloc(const struct cpumask *cpu_map); 5888 5889static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5890 const struct cpumask *cpu_map) 5891{ 5892 switch (what) { 5893 case sa_rootdomain: 5894 if (!atomic_read(&d->rd->refcount)) 5895 free_rootdomain(&d->rd->rcu); /* fall through */ 5896 case sa_sd: 5897 free_percpu(d->sd); /* fall through */ 5898 case sa_sd_storage: 5899 __sdt_free(cpu_map); /* fall through */ 5900 case sa_none: 5901 break; 5902 } 5903} 5904 5905static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5906 const struct cpumask *cpu_map) 5907{ 5908 memset(d, 0, sizeof(*d)); 5909 5910 if (__sdt_alloc(cpu_map)) 5911 return sa_sd_storage; 5912 d->sd = alloc_percpu(struct sched_domain *); 5913 if (!d->sd) 5914 return sa_sd_storage; 5915 d->rd = alloc_rootdomain(); 5916 if (!d->rd) 5917 return sa_sd; 5918 return sa_rootdomain; 5919} 5920 5921/* 5922 * NULL the sd_data elements we've used to build the sched_domain and 5923 * sched_group structure so that the subsequent __free_domain_allocs() 5924 * will not free the data we're using. 5925 */ 5926static void claim_allocations(int cpu, struct sched_domain *sd) 5927{ 5928 struct sd_data *sdd = sd->private; 5929 5930 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5931 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5932 5933 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5934 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5935 5936 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5937 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5938} 5939 5940#ifdef CONFIG_SCHED_SMT 5941static const struct cpumask *cpu_smt_mask(int cpu) 5942{ 5943 return topology_thread_cpumask(cpu); 5944} 5945#endif 5946 5947/* 5948 * Topology list, bottom-up. 5949 */ 5950static struct sched_domain_topology_level default_topology[] = { 5951#ifdef CONFIG_SCHED_SMT 5952 { sd_init_SIBLING, cpu_smt_mask, }, 5953#endif 5954#ifdef CONFIG_SCHED_MC 5955 { sd_init_MC, cpu_coregroup_mask, }, 5956#endif 5957#ifdef CONFIG_SCHED_BOOK 5958 { sd_init_BOOK, cpu_book_mask, }, 5959#endif 5960 { sd_init_CPU, cpu_cpu_mask, }, 5961 { NULL, }, 5962}; 5963 5964static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5965 5966#define for_each_sd_topology(tl) \ 5967 for (tl = sched_domain_topology; tl->init; tl++) 5968 5969#ifdef CONFIG_NUMA 5970 5971static int sched_domains_numa_levels; 5972static int *sched_domains_numa_distance; 5973static struct cpumask ***sched_domains_numa_masks; 5974static int sched_domains_curr_level; 5975 5976static inline int sd_local_flags(int level) 5977{ 5978 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5979 return 0; 5980 5981 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5982} 5983 5984static struct sched_domain * 5985sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5986{ 5987 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5988 int level = tl->numa_level; 5989 int sd_weight = cpumask_weight( 5990 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5991 5992 *sd = (struct sched_domain){ 5993 .min_interval = sd_weight, 5994 .max_interval = 2*sd_weight, 5995 .busy_factor = 32, 5996 .imbalance_pct = 125, 5997 .cache_nice_tries = 2, 5998 .busy_idx = 3, 5999 .idle_idx = 2, 6000 .newidle_idx = 0, 6001 .wake_idx = 0, 6002 .forkexec_idx = 0, 6003 6004 .flags = 1*SD_LOAD_BALANCE 6005 | 1*SD_BALANCE_NEWIDLE 6006 | 0*SD_BALANCE_EXEC 6007 | 0*SD_BALANCE_FORK 6008 | 0*SD_BALANCE_WAKE 6009 | 0*SD_WAKE_AFFINE 6010 | 0*SD_SHARE_CPUPOWER 6011 | 0*SD_SHARE_PKG_RESOURCES 6012 | 1*SD_SERIALIZE 6013 | 0*SD_PREFER_SIBLING 6014 | 1*SD_NUMA 6015 | sd_local_flags(level) 6016 , 6017 .last_balance = jiffies, 6018 .balance_interval = sd_weight, 6019 }; 6020 SD_INIT_NAME(sd, NUMA); 6021 sd->private = &tl->data; 6022 6023 /* 6024 * Ugly hack to pass state to sd_numa_mask()... 6025 */ 6026 sched_domains_curr_level = tl->numa_level; 6027 6028 return sd; 6029} 6030 6031static const struct cpumask *sd_numa_mask(int cpu) 6032{ 6033 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6034} 6035 6036static void sched_numa_warn(const char *str) 6037{ 6038 static int done = false; 6039 int i,j; 6040 6041 if (done) 6042 return; 6043 6044 done = true; 6045 6046 printk(KERN_WARNING "ERROR: %s\n\n", str); 6047 6048 for (i = 0; i < nr_node_ids; i++) { 6049 printk(KERN_WARNING " "); 6050 for (j = 0; j < nr_node_ids; j++) 6051 printk(KERN_CONT "%02d ", node_distance(i,j)); 6052 printk(KERN_CONT "\n"); 6053 } 6054 printk(KERN_WARNING "\n"); 6055} 6056 6057static bool find_numa_distance(int distance) 6058{ 6059 int i; 6060 6061 if (distance == node_distance(0, 0)) 6062 return true; 6063 6064 for (i = 0; i < sched_domains_numa_levels; i++) { 6065 if (sched_domains_numa_distance[i] == distance) 6066 return true; 6067 } 6068 6069 return false; 6070} 6071 6072static void sched_init_numa(void) 6073{ 6074 int next_distance, curr_distance = node_distance(0, 0); 6075 struct sched_domain_topology_level *tl; 6076 int level = 0; 6077 int i, j, k; 6078 6079 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6080 if (!sched_domains_numa_distance) 6081 return; 6082 6083 /* 6084 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6085 * unique distances in the node_distance() table. 6086 * 6087 * Assumes node_distance(0,j) includes all distances in 6088 * node_distance(i,j) in order to avoid cubic time. 6089 */ 6090 next_distance = curr_distance; 6091 for (i = 0; i < nr_node_ids; i++) { 6092 for (j = 0; j < nr_node_ids; j++) { 6093 for (k = 0; k < nr_node_ids; k++) { 6094 int distance = node_distance(i, k); 6095 6096 if (distance > curr_distance && 6097 (distance < next_distance || 6098 next_distance == curr_distance)) 6099 next_distance = distance; 6100 6101 /* 6102 * While not a strong assumption it would be nice to know 6103 * about cases where if node A is connected to B, B is not 6104 * equally connected to A. 6105 */ 6106 if (sched_debug() && node_distance(k, i) != distance) 6107 sched_numa_warn("Node-distance not symmetric"); 6108 6109 if (sched_debug() && i && !find_numa_distance(distance)) 6110 sched_numa_warn("Node-0 not representative"); 6111 } 6112 if (next_distance != curr_distance) { 6113 sched_domains_numa_distance[level++] = next_distance; 6114 sched_domains_numa_levels = level; 6115 curr_distance = next_distance; 6116 } else break; 6117 } 6118 6119 /* 6120 * In case of sched_debug() we verify the above assumption. 6121 */ 6122 if (!sched_debug()) 6123 break; 6124 } 6125 /* 6126 * 'level' contains the number of unique distances, excluding the 6127 * identity distance node_distance(i,i). 6128 * 6129 * The sched_domains_numa_distance[] array includes the actual distance 6130 * numbers. 6131 */ 6132 6133 /* 6134 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6135 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6136 * the array will contain less then 'level' members. This could be 6137 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6138 * in other functions. 6139 * 6140 * We reset it to 'level' at the end of this function. 6141 */ 6142 sched_domains_numa_levels = 0; 6143 6144 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6145 if (!sched_domains_numa_masks) 6146 return; 6147 6148 /* 6149 * Now for each level, construct a mask per node which contains all 6150 * cpus of nodes that are that many hops away from us. 6151 */ 6152 for (i = 0; i < level; i++) { 6153 sched_domains_numa_masks[i] = 6154 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6155 if (!sched_domains_numa_masks[i]) 6156 return; 6157 6158 for (j = 0; j < nr_node_ids; j++) { 6159 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6160 if (!mask) 6161 return; 6162 6163 sched_domains_numa_masks[i][j] = mask; 6164 6165 for (k = 0; k < nr_node_ids; k++) { 6166 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6167 continue; 6168 6169 cpumask_or(mask, mask, cpumask_of_node(k)); 6170 } 6171 } 6172 } 6173 6174 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6175 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6176 if (!tl) 6177 return; 6178 6179 /* 6180 * Copy the default topology bits.. 6181 */ 6182 for (i = 0; default_topology[i].init; i++) 6183 tl[i] = default_topology[i]; 6184 6185 /* 6186 * .. and append 'j' levels of NUMA goodness. 6187 */ 6188 for (j = 0; j < level; i++, j++) { 6189 tl[i] = (struct sched_domain_topology_level){ 6190 .init = sd_numa_init, 6191 .mask = sd_numa_mask, 6192 .flags = SDTL_OVERLAP, 6193 .numa_level = j, 6194 }; 6195 } 6196 6197 sched_domain_topology = tl; 6198 6199 sched_domains_numa_levels = level; 6200} 6201 6202static void sched_domains_numa_masks_set(int cpu) 6203{ 6204 int i, j; 6205 int node = cpu_to_node(cpu); 6206 6207 for (i = 0; i < sched_domains_numa_levels; i++) { 6208 for (j = 0; j < nr_node_ids; j++) { 6209 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6210 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6211 } 6212 } 6213} 6214 6215static void sched_domains_numa_masks_clear(int cpu) 6216{ 6217 int i, j; 6218 for (i = 0; i < sched_domains_numa_levels; i++) { 6219 for (j = 0; j < nr_node_ids; j++) 6220 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6221 } 6222} 6223 6224/* 6225 * Update sched_domains_numa_masks[level][node] array when new cpus 6226 * are onlined. 6227 */ 6228static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6229 unsigned long action, 6230 void *hcpu) 6231{ 6232 int cpu = (long)hcpu; 6233 6234 switch (action & ~CPU_TASKS_FROZEN) { 6235 case CPU_ONLINE: 6236 sched_domains_numa_masks_set(cpu); 6237 break; 6238 6239 case CPU_DEAD: 6240 sched_domains_numa_masks_clear(cpu); 6241 break; 6242 6243 default: 6244 return NOTIFY_DONE; 6245 } 6246 6247 return NOTIFY_OK; 6248} 6249#else 6250static inline void sched_init_numa(void) 6251{ 6252} 6253 6254static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6255 unsigned long action, 6256 void *hcpu) 6257{ 6258 return 0; 6259} 6260#endif /* CONFIG_NUMA */ 6261 6262static int __sdt_alloc(const struct cpumask *cpu_map) 6263{ 6264 struct sched_domain_topology_level *tl; 6265 int j; 6266 6267 for_each_sd_topology(tl) { 6268 struct sd_data *sdd = &tl->data; 6269 6270 sdd->sd = alloc_percpu(struct sched_domain *); 6271 if (!sdd->sd) 6272 return -ENOMEM; 6273 6274 sdd->sg = alloc_percpu(struct sched_group *); 6275 if (!sdd->sg) 6276 return -ENOMEM; 6277 6278 sdd->sgp = alloc_percpu(struct sched_group_power *); 6279 if (!sdd->sgp) 6280 return -ENOMEM; 6281 6282 for_each_cpu(j, cpu_map) { 6283 struct sched_domain *sd; 6284 struct sched_group *sg; 6285 struct sched_group_power *sgp; 6286 6287 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6288 GFP_KERNEL, cpu_to_node(j)); 6289 if (!sd) 6290 return -ENOMEM; 6291 6292 *per_cpu_ptr(sdd->sd, j) = sd; 6293 6294 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6295 GFP_KERNEL, cpu_to_node(j)); 6296 if (!sg) 6297 return -ENOMEM; 6298 6299 sg->next = sg; 6300 6301 *per_cpu_ptr(sdd->sg, j) = sg; 6302 6303 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6304 GFP_KERNEL, cpu_to_node(j)); 6305 if (!sgp) 6306 return -ENOMEM; 6307 6308 *per_cpu_ptr(sdd->sgp, j) = sgp; 6309 } 6310 } 6311 6312 return 0; 6313} 6314 6315static void __sdt_free(const struct cpumask *cpu_map) 6316{ 6317 struct sched_domain_topology_level *tl; 6318 int j; 6319 6320 for_each_sd_topology(tl) { 6321 struct sd_data *sdd = &tl->data; 6322 6323 for_each_cpu(j, cpu_map) { 6324 struct sched_domain *sd; 6325 6326 if (sdd->sd) { 6327 sd = *per_cpu_ptr(sdd->sd, j); 6328 if (sd && (sd->flags & SD_OVERLAP)) 6329 free_sched_groups(sd->groups, 0); 6330 kfree(*per_cpu_ptr(sdd->sd, j)); 6331 } 6332 6333 if (sdd->sg) 6334 kfree(*per_cpu_ptr(sdd->sg, j)); 6335 if (sdd->sgp) 6336 kfree(*per_cpu_ptr(sdd->sgp, j)); 6337 } 6338 free_percpu(sdd->sd); 6339 sdd->sd = NULL; 6340 free_percpu(sdd->sg); 6341 sdd->sg = NULL; 6342 free_percpu(sdd->sgp); 6343 sdd->sgp = NULL; 6344 } 6345} 6346 6347struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6348 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6349 struct sched_domain *child, int cpu) 6350{ 6351 struct sched_domain *sd = tl->init(tl, cpu); 6352 if (!sd) 6353 return child; 6354 6355 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6356 if (child) { 6357 sd->level = child->level + 1; 6358 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6359 child->parent = sd; 6360 sd->child = child; 6361 } 6362 set_domain_attribute(sd, attr); 6363 6364 return sd; 6365} 6366 6367/* 6368 * Build sched domains for a given set of cpus and attach the sched domains 6369 * to the individual cpus 6370 */ 6371static int build_sched_domains(const struct cpumask *cpu_map, 6372 struct sched_domain_attr *attr) 6373{ 6374 enum s_alloc alloc_state; 6375 struct sched_domain *sd; 6376 struct s_data d; 6377 int i, ret = -ENOMEM; 6378 6379 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6380 if (alloc_state != sa_rootdomain) 6381 goto error; 6382 6383 /* Set up domains for cpus specified by the cpu_map. */ 6384 for_each_cpu(i, cpu_map) { 6385 struct sched_domain_topology_level *tl; 6386 6387 sd = NULL; 6388 for_each_sd_topology(tl) { 6389 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6390 if (tl == sched_domain_topology) 6391 *per_cpu_ptr(d.sd, i) = sd; 6392 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6393 sd->flags |= SD_OVERLAP; 6394 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6395 break; 6396 } 6397 } 6398 6399 /* Build the groups for the domains */ 6400 for_each_cpu(i, cpu_map) { 6401 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6402 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6403 if (sd->flags & SD_OVERLAP) { 6404 if (build_overlap_sched_groups(sd, i)) 6405 goto error; 6406 } else { 6407 if (build_sched_groups(sd, i)) 6408 goto error; 6409 } 6410 } 6411 } 6412 6413 /* Calculate CPU power for physical packages and nodes */ 6414 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6415 if (!cpumask_test_cpu(i, cpu_map)) 6416 continue; 6417 6418 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6419 claim_allocations(i, sd); 6420 init_sched_groups_power(i, sd); 6421 } 6422 } 6423 6424 /* Attach the domains */ 6425 rcu_read_lock(); 6426 for_each_cpu(i, cpu_map) { 6427 sd = *per_cpu_ptr(d.sd, i); 6428 cpu_attach_domain(sd, d.rd, i); 6429 } 6430 rcu_read_unlock(); 6431 6432 ret = 0; 6433error: 6434 __free_domain_allocs(&d, alloc_state, cpu_map); 6435 return ret; 6436} 6437 6438static cpumask_var_t *doms_cur; /* current sched domains */ 6439static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6440static struct sched_domain_attr *dattr_cur; 6441 /* attribues of custom domains in 'doms_cur' */ 6442 6443/* 6444 * Special case: If a kmalloc of a doms_cur partition (array of 6445 * cpumask) fails, then fallback to a single sched domain, 6446 * as determined by the single cpumask fallback_doms. 6447 */ 6448static cpumask_var_t fallback_doms; 6449 6450/* 6451 * arch_update_cpu_topology lets virtualized architectures update the 6452 * cpu core maps. It is supposed to return 1 if the topology changed 6453 * or 0 if it stayed the same. 6454 */ 6455int __attribute__((weak)) arch_update_cpu_topology(void) 6456{ 6457 return 0; 6458} 6459 6460cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6461{ 6462 int i; 6463 cpumask_var_t *doms; 6464 6465 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6466 if (!doms) 6467 return NULL; 6468 for (i = 0; i < ndoms; i++) { 6469 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6470 free_sched_domains(doms, i); 6471 return NULL; 6472 } 6473 } 6474 return doms; 6475} 6476 6477void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6478{ 6479 unsigned int i; 6480 for (i = 0; i < ndoms; i++) 6481 free_cpumask_var(doms[i]); 6482 kfree(doms); 6483} 6484 6485/* 6486 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6487 * For now this just excludes isolated cpus, but could be used to 6488 * exclude other special cases in the future. 6489 */ 6490static int init_sched_domains(const struct cpumask *cpu_map) 6491{ 6492 int err; 6493 6494 arch_update_cpu_topology(); 6495 ndoms_cur = 1; 6496 doms_cur = alloc_sched_domains(ndoms_cur); 6497 if (!doms_cur) 6498 doms_cur = &fallback_doms; 6499 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6500 err = build_sched_domains(doms_cur[0], NULL); 6501 register_sched_domain_sysctl(); 6502 6503 return err; 6504} 6505 6506/* 6507 * Detach sched domains from a group of cpus specified in cpu_map 6508 * These cpus will now be attached to the NULL domain 6509 */ 6510static void detach_destroy_domains(const struct cpumask *cpu_map) 6511{ 6512 int i; 6513 6514 rcu_read_lock(); 6515 for_each_cpu(i, cpu_map) 6516 cpu_attach_domain(NULL, &def_root_domain, i); 6517 rcu_read_unlock(); 6518} 6519 6520/* handle null as "default" */ 6521static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6522 struct sched_domain_attr *new, int idx_new) 6523{ 6524 struct sched_domain_attr tmp; 6525 6526 /* fast path */ 6527 if (!new && !cur) 6528 return 1; 6529 6530 tmp = SD_ATTR_INIT; 6531 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6532 new ? (new + idx_new) : &tmp, 6533 sizeof(struct sched_domain_attr)); 6534} 6535 6536/* 6537 * Partition sched domains as specified by the 'ndoms_new' 6538 * cpumasks in the array doms_new[] of cpumasks. This compares 6539 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6540 * It destroys each deleted domain and builds each new domain. 6541 * 6542 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6543 * The masks don't intersect (don't overlap.) We should setup one 6544 * sched domain for each mask. CPUs not in any of the cpumasks will 6545 * not be load balanced. If the same cpumask appears both in the 6546 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6547 * it as it is. 6548 * 6549 * The passed in 'doms_new' should be allocated using 6550 * alloc_sched_domains. This routine takes ownership of it and will 6551 * free_sched_domains it when done with it. If the caller failed the 6552 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6553 * and partition_sched_domains() will fallback to the single partition 6554 * 'fallback_doms', it also forces the domains to be rebuilt. 6555 * 6556 * If doms_new == NULL it will be replaced with cpu_online_mask. 6557 * ndoms_new == 0 is a special case for destroying existing domains, 6558 * and it will not create the default domain. 6559 * 6560 * Call with hotplug lock held 6561 */ 6562void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6563 struct sched_domain_attr *dattr_new) 6564{ 6565 int i, j, n; 6566 int new_topology; 6567 6568 mutex_lock(&sched_domains_mutex); 6569 6570 /* always unregister in case we don't destroy any domains */ 6571 unregister_sched_domain_sysctl(); 6572 6573 /* Let architecture update cpu core mappings. */ 6574 new_topology = arch_update_cpu_topology(); 6575 6576 n = doms_new ? ndoms_new : 0; 6577 6578 /* Destroy deleted domains */ 6579 for (i = 0; i < ndoms_cur; i++) { 6580 for (j = 0; j < n && !new_topology; j++) { 6581 if (cpumask_equal(doms_cur[i], doms_new[j]) 6582 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6583 goto match1; 6584 } 6585 /* no match - a current sched domain not in new doms_new[] */ 6586 detach_destroy_domains(doms_cur[i]); 6587match1: 6588 ; 6589 } 6590 6591 n = ndoms_cur; 6592 if (doms_new == NULL) { 6593 n = 0; 6594 doms_new = &fallback_doms; 6595 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6596 WARN_ON_ONCE(dattr_new); 6597 } 6598 6599 /* Build new domains */ 6600 for (i = 0; i < ndoms_new; i++) { 6601 for (j = 0; j < n && !new_topology; j++) { 6602 if (cpumask_equal(doms_new[i], doms_cur[j]) 6603 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6604 goto match2; 6605 } 6606 /* no match - add a new doms_new */ 6607 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6608match2: 6609 ; 6610 } 6611 6612 /* Remember the new sched domains */ 6613 if (doms_cur != &fallback_doms) 6614 free_sched_domains(doms_cur, ndoms_cur); 6615 kfree(dattr_cur); /* kfree(NULL) is safe */ 6616 doms_cur = doms_new; 6617 dattr_cur = dattr_new; 6618 ndoms_cur = ndoms_new; 6619 6620 register_sched_domain_sysctl(); 6621 6622 mutex_unlock(&sched_domains_mutex); 6623} 6624 6625static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6626 6627/* 6628 * Update cpusets according to cpu_active mask. If cpusets are 6629 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6630 * around partition_sched_domains(). 6631 * 6632 * If we come here as part of a suspend/resume, don't touch cpusets because we 6633 * want to restore it back to its original state upon resume anyway. 6634 */ 6635static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6636 void *hcpu) 6637{ 6638 switch (action) { 6639 case CPU_ONLINE_FROZEN: 6640 case CPU_DOWN_FAILED_FROZEN: 6641 6642 /* 6643 * num_cpus_frozen tracks how many CPUs are involved in suspend 6644 * resume sequence. As long as this is not the last online 6645 * operation in the resume sequence, just build a single sched 6646 * domain, ignoring cpusets. 6647 */ 6648 num_cpus_frozen--; 6649 if (likely(num_cpus_frozen)) { 6650 partition_sched_domains(1, NULL, NULL); 6651 break; 6652 } 6653 6654 /* 6655 * This is the last CPU online operation. So fall through and 6656 * restore the original sched domains by considering the 6657 * cpuset configurations. 6658 */ 6659 6660 case CPU_ONLINE: 6661 case CPU_DOWN_FAILED: 6662 cpuset_update_active_cpus(true); 6663 break; 6664 default: 6665 return NOTIFY_DONE; 6666 } 6667 return NOTIFY_OK; 6668} 6669 6670static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6671 void *hcpu) 6672{ 6673 switch (action) { 6674 case CPU_DOWN_PREPARE: 6675 cpuset_update_active_cpus(false); 6676 break; 6677 case CPU_DOWN_PREPARE_FROZEN: 6678 num_cpus_frozen++; 6679 partition_sched_domains(1, NULL, NULL); 6680 break; 6681 default: 6682 return NOTIFY_DONE; 6683 } 6684 return NOTIFY_OK; 6685} 6686 6687void __init sched_init_smp(void) 6688{ 6689 cpumask_var_t non_isolated_cpus; 6690 6691 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6692 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6693 6694 sched_init_numa(); 6695 6696 /* 6697 * There's no userspace yet to cause hotplug operations; hence all the 6698 * cpu masks are stable and all blatant races in the below code cannot 6699 * happen. 6700 */ 6701 mutex_lock(&sched_domains_mutex); 6702 init_sched_domains(cpu_active_mask); 6703 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6704 if (cpumask_empty(non_isolated_cpus)) 6705 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6706 mutex_unlock(&sched_domains_mutex); 6707 6708 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6709 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6710 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6711 6712 init_hrtick(); 6713 6714 /* Move init over to a non-isolated CPU */ 6715 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6716 BUG(); 6717 sched_init_granularity(); 6718 free_cpumask_var(non_isolated_cpus); 6719 6720 init_sched_rt_class(); 6721 init_sched_dl_class(); 6722} 6723#else 6724void __init sched_init_smp(void) 6725{ 6726 sched_init_granularity(); 6727} 6728#endif /* CONFIG_SMP */ 6729 6730const_debug unsigned int sysctl_timer_migration = 1; 6731 6732int in_sched_functions(unsigned long addr) 6733{ 6734 return in_lock_functions(addr) || 6735 (addr >= (unsigned long)__sched_text_start 6736 && addr < (unsigned long)__sched_text_end); 6737} 6738 6739#ifdef CONFIG_CGROUP_SCHED 6740/* 6741 * Default task group. 6742 * Every task in system belongs to this group at bootup. 6743 */ 6744struct task_group root_task_group; 6745LIST_HEAD(task_groups); 6746#endif 6747 6748DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6749 6750void __init sched_init(void) 6751{ 6752 int i, j; 6753 unsigned long alloc_size = 0, ptr; 6754 6755#ifdef CONFIG_FAIR_GROUP_SCHED 6756 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6757#endif 6758#ifdef CONFIG_RT_GROUP_SCHED 6759 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6760#endif 6761#ifdef CONFIG_CPUMASK_OFFSTACK 6762 alloc_size += num_possible_cpus() * cpumask_size(); 6763#endif 6764 if (alloc_size) { 6765 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6766 6767#ifdef CONFIG_FAIR_GROUP_SCHED 6768 root_task_group.se = (struct sched_entity **)ptr; 6769 ptr += nr_cpu_ids * sizeof(void **); 6770 6771 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6772 ptr += nr_cpu_ids * sizeof(void **); 6773 6774#endif /* CONFIG_FAIR_GROUP_SCHED */ 6775#ifdef CONFIG_RT_GROUP_SCHED 6776 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6777 ptr += nr_cpu_ids * sizeof(void **); 6778 6779 root_task_group.rt_rq = (struct rt_rq **)ptr; 6780 ptr += nr_cpu_ids * sizeof(void **); 6781 6782#endif /* CONFIG_RT_GROUP_SCHED */ 6783#ifdef CONFIG_CPUMASK_OFFSTACK 6784 for_each_possible_cpu(i) { 6785 per_cpu(load_balance_mask, i) = (void *)ptr; 6786 ptr += cpumask_size(); 6787 } 6788#endif /* CONFIG_CPUMASK_OFFSTACK */ 6789 } 6790 6791 init_rt_bandwidth(&def_rt_bandwidth, 6792 global_rt_period(), global_rt_runtime()); 6793 init_dl_bandwidth(&def_dl_bandwidth, 6794 global_rt_period(), global_rt_runtime()); 6795 6796#ifdef CONFIG_SMP 6797 init_defrootdomain(); 6798#endif 6799 6800#ifdef CONFIG_RT_GROUP_SCHED 6801 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6802 global_rt_period(), global_rt_runtime()); 6803#endif /* CONFIG_RT_GROUP_SCHED */ 6804 6805#ifdef CONFIG_CGROUP_SCHED 6806 list_add(&root_task_group.list, &task_groups); 6807 INIT_LIST_HEAD(&root_task_group.children); 6808 INIT_LIST_HEAD(&root_task_group.siblings); 6809 autogroup_init(&init_task); 6810 6811#endif /* CONFIG_CGROUP_SCHED */ 6812 6813 for_each_possible_cpu(i) { 6814 struct rq *rq; 6815 6816 rq = cpu_rq(i); 6817 raw_spin_lock_init(&rq->lock); 6818 rq->nr_running = 0; 6819 rq->calc_load_active = 0; 6820 rq->calc_load_update = jiffies + LOAD_FREQ; 6821 init_cfs_rq(&rq->cfs); 6822 init_rt_rq(&rq->rt, rq); 6823 init_dl_rq(&rq->dl, rq); 6824#ifdef CONFIG_FAIR_GROUP_SCHED 6825 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6826 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6827 /* 6828 * How much cpu bandwidth does root_task_group get? 6829 * 6830 * In case of task-groups formed thr' the cgroup filesystem, it 6831 * gets 100% of the cpu resources in the system. This overall 6832 * system cpu resource is divided among the tasks of 6833 * root_task_group and its child task-groups in a fair manner, 6834 * based on each entity's (task or task-group's) weight 6835 * (se->load.weight). 6836 * 6837 * In other words, if root_task_group has 10 tasks of weight 6838 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6839 * then A0's share of the cpu resource is: 6840 * 6841 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6842 * 6843 * We achieve this by letting root_task_group's tasks sit 6844 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6845 */ 6846 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6847 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6848#endif /* CONFIG_FAIR_GROUP_SCHED */ 6849 6850 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6851#ifdef CONFIG_RT_GROUP_SCHED 6852 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6853#endif 6854 6855 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6856 rq->cpu_load[j] = 0; 6857 6858 rq->last_load_update_tick = jiffies; 6859 6860#ifdef CONFIG_SMP 6861 rq->sd = NULL; 6862 rq->rd = NULL; 6863 rq->cpu_power = SCHED_POWER_SCALE; 6864 rq->post_schedule = 0; 6865 rq->active_balance = 0; 6866 rq->next_balance = jiffies; 6867 rq->push_cpu = 0; 6868 rq->cpu = i; 6869 rq->online = 0; 6870 rq->idle_stamp = 0; 6871 rq->avg_idle = 2*sysctl_sched_migration_cost; 6872 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6873 6874 INIT_LIST_HEAD(&rq->cfs_tasks); 6875 6876 rq_attach_root(rq, &def_root_domain); 6877#ifdef CONFIG_NO_HZ_COMMON 6878 rq->nohz_flags = 0; 6879#endif 6880#ifdef CONFIG_NO_HZ_FULL 6881 rq->last_sched_tick = 0; 6882#endif 6883#endif 6884 init_rq_hrtick(rq); 6885 atomic_set(&rq->nr_iowait, 0); 6886 } 6887 6888 set_load_weight(&init_task); 6889 6890#ifdef CONFIG_PREEMPT_NOTIFIERS 6891 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6892#endif 6893 6894 /* 6895 * The boot idle thread does lazy MMU switching as well: 6896 */ 6897 atomic_inc(&init_mm.mm_count); 6898 enter_lazy_tlb(&init_mm, current); 6899 6900 /* 6901 * Make us the idle thread. Technically, schedule() should not be 6902 * called from this thread, however somewhere below it might be, 6903 * but because we are the idle thread, we just pick up running again 6904 * when this runqueue becomes "idle". 6905 */ 6906 init_idle(current, smp_processor_id()); 6907 6908 calc_load_update = jiffies + LOAD_FREQ; 6909 6910 /* 6911 * During early bootup we pretend to be a normal task: 6912 */ 6913 current->sched_class = &fair_sched_class; 6914 6915#ifdef CONFIG_SMP 6916 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6917 /* May be allocated at isolcpus cmdline parse time */ 6918 if (cpu_isolated_map == NULL) 6919 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6920 idle_thread_set_boot_cpu(); 6921#endif 6922 init_sched_fair_class(); 6923 6924 scheduler_running = 1; 6925} 6926 6927#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6928static inline int preempt_count_equals(int preempt_offset) 6929{ 6930 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6931 6932 return (nested == preempt_offset); 6933} 6934 6935void __might_sleep(const char *file, int line, int preempt_offset) 6936{ 6937 static unsigned long prev_jiffy; /* ratelimiting */ 6938 6939 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6940 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 6941 !is_idle_task(current)) || 6942 system_state != SYSTEM_RUNNING || oops_in_progress) 6943 return; 6944 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6945 return; 6946 prev_jiffy = jiffies; 6947 6948 printk(KERN_ERR 6949 "BUG: sleeping function called from invalid context at %s:%d\n", 6950 file, line); 6951 printk(KERN_ERR 6952 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6953 in_atomic(), irqs_disabled(), 6954 current->pid, current->comm); 6955 6956 debug_show_held_locks(current); 6957 if (irqs_disabled()) 6958 print_irqtrace_events(current); 6959#ifdef CONFIG_DEBUG_PREEMPT 6960 if (!preempt_count_equals(preempt_offset)) { 6961 pr_err("Preemption disabled at:"); 6962 print_ip_sym(current->preempt_disable_ip); 6963 pr_cont("\n"); 6964 } 6965#endif 6966 dump_stack(); 6967} 6968EXPORT_SYMBOL(__might_sleep); 6969#endif 6970 6971#ifdef CONFIG_MAGIC_SYSRQ 6972static void normalize_task(struct rq *rq, struct task_struct *p) 6973{ 6974 const struct sched_class *prev_class = p->sched_class; 6975 struct sched_attr attr = { 6976 .sched_policy = SCHED_NORMAL, 6977 }; 6978 int old_prio = p->prio; 6979 int on_rq; 6980 6981 on_rq = p->on_rq; 6982 if (on_rq) 6983 dequeue_task(rq, p, 0); 6984 __setscheduler(rq, p, &attr); 6985 if (on_rq) { 6986 enqueue_task(rq, p, 0); 6987 resched_task(rq->curr); 6988 } 6989 6990 check_class_changed(rq, p, prev_class, old_prio); 6991} 6992 6993void normalize_rt_tasks(void) 6994{ 6995 struct task_struct *g, *p; 6996 unsigned long flags; 6997 struct rq *rq; 6998 6999 read_lock_irqsave(&tasklist_lock, flags); 7000 do_each_thread(g, p) { 7001 /* 7002 * Only normalize user tasks: 7003 */ 7004 if (!p->mm) 7005 continue; 7006 7007 p->se.exec_start = 0; 7008#ifdef CONFIG_SCHEDSTATS 7009 p->se.statistics.wait_start = 0; 7010 p->se.statistics.sleep_start = 0; 7011 p->se.statistics.block_start = 0; 7012#endif 7013 7014 if (!dl_task(p) && !rt_task(p)) { 7015 /* 7016 * Renice negative nice level userspace 7017 * tasks back to 0: 7018 */ 7019 if (task_nice(p) < 0 && p->mm) 7020 set_user_nice(p, 0); 7021 continue; 7022 } 7023 7024 raw_spin_lock(&p->pi_lock); 7025 rq = __task_rq_lock(p); 7026 7027 normalize_task(rq, p); 7028 7029 __task_rq_unlock(rq); 7030 raw_spin_unlock(&p->pi_lock); 7031 } while_each_thread(g, p); 7032 7033 read_unlock_irqrestore(&tasklist_lock, flags); 7034} 7035 7036#endif /* CONFIG_MAGIC_SYSRQ */ 7037 7038#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7039/* 7040 * These functions are only useful for the IA64 MCA handling, or kdb. 7041 * 7042 * They can only be called when the whole system has been 7043 * stopped - every CPU needs to be quiescent, and no scheduling 7044 * activity can take place. Using them for anything else would 7045 * be a serious bug, and as a result, they aren't even visible 7046 * under any other configuration. 7047 */ 7048 7049/** 7050 * curr_task - return the current task for a given cpu. 7051 * @cpu: the processor in question. 7052 * 7053 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7054 * 7055 * Return: The current task for @cpu. 7056 */ 7057struct task_struct *curr_task(int cpu) 7058{ 7059 return cpu_curr(cpu); 7060} 7061 7062#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7063 7064#ifdef CONFIG_IA64 7065/** 7066 * set_curr_task - set the current task for a given cpu. 7067 * @cpu: the processor in question. 7068 * @p: the task pointer to set. 7069 * 7070 * Description: This function must only be used when non-maskable interrupts 7071 * are serviced on a separate stack. It allows the architecture to switch the 7072 * notion of the current task on a cpu in a non-blocking manner. This function 7073 * must be called with all CPU's synchronized, and interrupts disabled, the 7074 * and caller must save the original value of the current task (see 7075 * curr_task() above) and restore that value before reenabling interrupts and 7076 * re-starting the system. 7077 * 7078 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7079 */ 7080void set_curr_task(int cpu, struct task_struct *p) 7081{ 7082 cpu_curr(cpu) = p; 7083} 7084 7085#endif 7086 7087#ifdef CONFIG_CGROUP_SCHED 7088/* task_group_lock serializes the addition/removal of task groups */ 7089static DEFINE_SPINLOCK(task_group_lock); 7090 7091static void free_sched_group(struct task_group *tg) 7092{ 7093 free_fair_sched_group(tg); 7094 free_rt_sched_group(tg); 7095 autogroup_free(tg); 7096 kfree(tg); 7097} 7098 7099/* allocate runqueue etc for a new task group */ 7100struct task_group *sched_create_group(struct task_group *parent) 7101{ 7102 struct task_group *tg; 7103 7104 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7105 if (!tg) 7106 return ERR_PTR(-ENOMEM); 7107 7108 if (!alloc_fair_sched_group(tg, parent)) 7109 goto err; 7110 7111 if (!alloc_rt_sched_group(tg, parent)) 7112 goto err; 7113 7114 return tg; 7115 7116err: 7117 free_sched_group(tg); 7118 return ERR_PTR(-ENOMEM); 7119} 7120 7121void sched_online_group(struct task_group *tg, struct task_group *parent) 7122{ 7123 unsigned long flags; 7124 7125 spin_lock_irqsave(&task_group_lock, flags); 7126 list_add_rcu(&tg->list, &task_groups); 7127 7128 WARN_ON(!parent); /* root should already exist */ 7129 7130 tg->parent = parent; 7131 INIT_LIST_HEAD(&tg->children); 7132 list_add_rcu(&tg->siblings, &parent->children); 7133 spin_unlock_irqrestore(&task_group_lock, flags); 7134} 7135 7136/* rcu callback to free various structures associated with a task group */ 7137static void free_sched_group_rcu(struct rcu_head *rhp) 7138{ 7139 /* now it should be safe to free those cfs_rqs */ 7140 free_sched_group(container_of(rhp, struct task_group, rcu)); 7141} 7142 7143/* Destroy runqueue etc associated with a task group */ 7144void sched_destroy_group(struct task_group *tg) 7145{ 7146 /* wait for possible concurrent references to cfs_rqs complete */ 7147 call_rcu(&tg->rcu, free_sched_group_rcu); 7148} 7149 7150void sched_offline_group(struct task_group *tg) 7151{ 7152 unsigned long flags; 7153 int i; 7154 7155 /* end participation in shares distribution */ 7156 for_each_possible_cpu(i) 7157 unregister_fair_sched_group(tg, i); 7158 7159 spin_lock_irqsave(&task_group_lock, flags); 7160 list_del_rcu(&tg->list); 7161 list_del_rcu(&tg->siblings); 7162 spin_unlock_irqrestore(&task_group_lock, flags); 7163} 7164 7165/* change task's runqueue when it moves between groups. 7166 * The caller of this function should have put the task in its new group 7167 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7168 * reflect its new group. 7169 */ 7170void sched_move_task(struct task_struct *tsk) 7171{ 7172 struct task_group *tg; 7173 int on_rq, running; 7174 unsigned long flags; 7175 struct rq *rq; 7176 7177 rq = task_rq_lock(tsk, &flags); 7178 7179 running = task_current(rq, tsk); 7180 on_rq = tsk->on_rq; 7181 7182 if (on_rq) 7183 dequeue_task(rq, tsk, 0); 7184 if (unlikely(running)) 7185 tsk->sched_class->put_prev_task(rq, tsk); 7186 7187 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7188 lockdep_is_held(&tsk->sighand->siglock)), 7189 struct task_group, css); 7190 tg = autogroup_task_group(tsk, tg); 7191 tsk->sched_task_group = tg; 7192 7193#ifdef CONFIG_FAIR_GROUP_SCHED 7194 if (tsk->sched_class->task_move_group) 7195 tsk->sched_class->task_move_group(tsk, on_rq); 7196 else 7197#endif 7198 set_task_rq(tsk, task_cpu(tsk)); 7199 7200 if (unlikely(running)) 7201 tsk->sched_class->set_curr_task(rq); 7202 if (on_rq) 7203 enqueue_task(rq, tsk, 0); 7204 7205 task_rq_unlock(rq, tsk, &flags); 7206} 7207#endif /* CONFIG_CGROUP_SCHED */ 7208 7209#ifdef CONFIG_RT_GROUP_SCHED 7210/* 7211 * Ensure that the real time constraints are schedulable. 7212 */ 7213static DEFINE_MUTEX(rt_constraints_mutex); 7214 7215/* Must be called with tasklist_lock held */ 7216static inline int tg_has_rt_tasks(struct task_group *tg) 7217{ 7218 struct task_struct *g, *p; 7219 7220 do_each_thread(g, p) { 7221 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7222 return 1; 7223 } while_each_thread(g, p); 7224 7225 return 0; 7226} 7227 7228struct rt_schedulable_data { 7229 struct task_group *tg; 7230 u64 rt_period; 7231 u64 rt_runtime; 7232}; 7233 7234static int tg_rt_schedulable(struct task_group *tg, void *data) 7235{ 7236 struct rt_schedulable_data *d = data; 7237 struct task_group *child; 7238 unsigned long total, sum = 0; 7239 u64 period, runtime; 7240 7241 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7242 runtime = tg->rt_bandwidth.rt_runtime; 7243 7244 if (tg == d->tg) { 7245 period = d->rt_period; 7246 runtime = d->rt_runtime; 7247 } 7248 7249 /* 7250 * Cannot have more runtime than the period. 7251 */ 7252 if (runtime > period && runtime != RUNTIME_INF) 7253 return -EINVAL; 7254 7255 /* 7256 * Ensure we don't starve existing RT tasks. 7257 */ 7258 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7259 return -EBUSY; 7260 7261 total = to_ratio(period, runtime); 7262 7263 /* 7264 * Nobody can have more than the global setting allows. 7265 */ 7266 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7267 return -EINVAL; 7268 7269 /* 7270 * The sum of our children's runtime should not exceed our own. 7271 */ 7272 list_for_each_entry_rcu(child, &tg->children, siblings) { 7273 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7274 runtime = child->rt_bandwidth.rt_runtime; 7275 7276 if (child == d->tg) { 7277 period = d->rt_period; 7278 runtime = d->rt_runtime; 7279 } 7280 7281 sum += to_ratio(period, runtime); 7282 } 7283 7284 if (sum > total) 7285 return -EINVAL; 7286 7287 return 0; 7288} 7289 7290static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7291{ 7292 int ret; 7293 7294 struct rt_schedulable_data data = { 7295 .tg = tg, 7296 .rt_period = period, 7297 .rt_runtime = runtime, 7298 }; 7299 7300 rcu_read_lock(); 7301 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7302 rcu_read_unlock(); 7303 7304 return ret; 7305} 7306 7307static int tg_set_rt_bandwidth(struct task_group *tg, 7308 u64 rt_period, u64 rt_runtime) 7309{ 7310 int i, err = 0; 7311 7312 mutex_lock(&rt_constraints_mutex); 7313 read_lock(&tasklist_lock); 7314 err = __rt_schedulable(tg, rt_period, rt_runtime); 7315 if (err) 7316 goto unlock; 7317 7318 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7319 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7320 tg->rt_bandwidth.rt_runtime = rt_runtime; 7321 7322 for_each_possible_cpu(i) { 7323 struct rt_rq *rt_rq = tg->rt_rq[i]; 7324 7325 raw_spin_lock(&rt_rq->rt_runtime_lock); 7326 rt_rq->rt_runtime = rt_runtime; 7327 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7328 } 7329 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7330unlock: 7331 read_unlock(&tasklist_lock); 7332 mutex_unlock(&rt_constraints_mutex); 7333 7334 return err; 7335} 7336 7337static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7338{ 7339 u64 rt_runtime, rt_period; 7340 7341 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7342 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7343 if (rt_runtime_us < 0) 7344 rt_runtime = RUNTIME_INF; 7345 7346 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7347} 7348 7349static long sched_group_rt_runtime(struct task_group *tg) 7350{ 7351 u64 rt_runtime_us; 7352 7353 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7354 return -1; 7355 7356 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7357 do_div(rt_runtime_us, NSEC_PER_USEC); 7358 return rt_runtime_us; 7359} 7360 7361static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7362{ 7363 u64 rt_runtime, rt_period; 7364 7365 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7366 rt_runtime = tg->rt_bandwidth.rt_runtime; 7367 7368 if (rt_period == 0) 7369 return -EINVAL; 7370 7371 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7372} 7373 7374static long sched_group_rt_period(struct task_group *tg) 7375{ 7376 u64 rt_period_us; 7377 7378 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7379 do_div(rt_period_us, NSEC_PER_USEC); 7380 return rt_period_us; 7381} 7382#endif /* CONFIG_RT_GROUP_SCHED */ 7383 7384#ifdef CONFIG_RT_GROUP_SCHED 7385static int sched_rt_global_constraints(void) 7386{ 7387 int ret = 0; 7388 7389 mutex_lock(&rt_constraints_mutex); 7390 read_lock(&tasklist_lock); 7391 ret = __rt_schedulable(NULL, 0, 0); 7392 read_unlock(&tasklist_lock); 7393 mutex_unlock(&rt_constraints_mutex); 7394 7395 return ret; 7396} 7397 7398static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7399{ 7400 /* Don't accept realtime tasks when there is no way for them to run */ 7401 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7402 return 0; 7403 7404 return 1; 7405} 7406 7407#else /* !CONFIG_RT_GROUP_SCHED */ 7408static int sched_rt_global_constraints(void) 7409{ 7410 unsigned long flags; 7411 int i, ret = 0; 7412 7413 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7414 for_each_possible_cpu(i) { 7415 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7416 7417 raw_spin_lock(&rt_rq->rt_runtime_lock); 7418 rt_rq->rt_runtime = global_rt_runtime(); 7419 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7420 } 7421 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7422 7423 return ret; 7424} 7425#endif /* CONFIG_RT_GROUP_SCHED */ 7426 7427static int sched_dl_global_constraints(void) 7428{ 7429 u64 runtime = global_rt_runtime(); 7430 u64 period = global_rt_period(); 7431 u64 new_bw = to_ratio(period, runtime); 7432 int cpu, ret = 0; 7433 unsigned long flags; 7434 7435 /* 7436 * Here we want to check the bandwidth not being set to some 7437 * value smaller than the currently allocated bandwidth in 7438 * any of the root_domains. 7439 * 7440 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7441 * cycling on root_domains... Discussion on different/better 7442 * solutions is welcome! 7443 */ 7444 for_each_possible_cpu(cpu) { 7445 struct dl_bw *dl_b = dl_bw_of(cpu); 7446 7447 raw_spin_lock_irqsave(&dl_b->lock, flags); 7448 if (new_bw < dl_b->total_bw) 7449 ret = -EBUSY; 7450 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7451 7452 if (ret) 7453 break; 7454 } 7455 7456 return ret; 7457} 7458 7459static void sched_dl_do_global(void) 7460{ 7461 u64 new_bw = -1; 7462 int cpu; 7463 unsigned long flags; 7464 7465 def_dl_bandwidth.dl_period = global_rt_period(); 7466 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7467 7468 if (global_rt_runtime() != RUNTIME_INF) 7469 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7470 7471 /* 7472 * FIXME: As above... 7473 */ 7474 for_each_possible_cpu(cpu) { 7475 struct dl_bw *dl_b = dl_bw_of(cpu); 7476 7477 raw_spin_lock_irqsave(&dl_b->lock, flags); 7478 dl_b->bw = new_bw; 7479 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7480 } 7481} 7482 7483static int sched_rt_global_validate(void) 7484{ 7485 if (sysctl_sched_rt_period <= 0) 7486 return -EINVAL; 7487 7488 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7489 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7490 return -EINVAL; 7491 7492 return 0; 7493} 7494 7495static void sched_rt_do_global(void) 7496{ 7497 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7498 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7499} 7500 7501int sched_rt_handler(struct ctl_table *table, int write, 7502 void __user *buffer, size_t *lenp, 7503 loff_t *ppos) 7504{ 7505 int old_period, old_runtime; 7506 static DEFINE_MUTEX(mutex); 7507 int ret; 7508 7509 mutex_lock(&mutex); 7510 old_period = sysctl_sched_rt_period; 7511 old_runtime = sysctl_sched_rt_runtime; 7512 7513 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7514 7515 if (!ret && write) { 7516 ret = sched_rt_global_validate(); 7517 if (ret) 7518 goto undo; 7519 7520 ret = sched_rt_global_constraints(); 7521 if (ret) 7522 goto undo; 7523 7524 ret = sched_dl_global_constraints(); 7525 if (ret) 7526 goto undo; 7527 7528 sched_rt_do_global(); 7529 sched_dl_do_global(); 7530 } 7531 if (0) { 7532undo: 7533 sysctl_sched_rt_period = old_period; 7534 sysctl_sched_rt_runtime = old_runtime; 7535 } 7536 mutex_unlock(&mutex); 7537 7538 return ret; 7539} 7540 7541int sched_rr_handler(struct ctl_table *table, int write, 7542 void __user *buffer, size_t *lenp, 7543 loff_t *ppos) 7544{ 7545 int ret; 7546 static DEFINE_MUTEX(mutex); 7547 7548 mutex_lock(&mutex); 7549 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7550 /* make sure that internally we keep jiffies */ 7551 /* also, writing zero resets timeslice to default */ 7552 if (!ret && write) { 7553 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7554 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7555 } 7556 mutex_unlock(&mutex); 7557 return ret; 7558} 7559 7560#ifdef CONFIG_CGROUP_SCHED 7561 7562static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7563{ 7564 return css ? container_of(css, struct task_group, css) : NULL; 7565} 7566 7567static struct cgroup_subsys_state * 7568cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7569{ 7570 struct task_group *parent = css_tg(parent_css); 7571 struct task_group *tg; 7572 7573 if (!parent) { 7574 /* This is early initialization for the top cgroup */ 7575 return &root_task_group.css; 7576 } 7577 7578 tg = sched_create_group(parent); 7579 if (IS_ERR(tg)) 7580 return ERR_PTR(-ENOMEM); 7581 7582 return &tg->css; 7583} 7584 7585static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7586{ 7587 struct task_group *tg = css_tg(css); 7588 struct task_group *parent = css_tg(css_parent(css)); 7589 7590 if (parent) 7591 sched_online_group(tg, parent); 7592 return 0; 7593} 7594 7595static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7596{ 7597 struct task_group *tg = css_tg(css); 7598 7599 sched_destroy_group(tg); 7600} 7601 7602static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7603{ 7604 struct task_group *tg = css_tg(css); 7605 7606 sched_offline_group(tg); 7607} 7608 7609static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7610 struct cgroup_taskset *tset) 7611{ 7612 struct task_struct *task; 7613 7614 cgroup_taskset_for_each(task, tset) { 7615#ifdef CONFIG_RT_GROUP_SCHED 7616 if (!sched_rt_can_attach(css_tg(css), task)) 7617 return -EINVAL; 7618#else 7619 /* We don't support RT-tasks being in separate groups */ 7620 if (task->sched_class != &fair_sched_class) 7621 return -EINVAL; 7622#endif 7623 } 7624 return 0; 7625} 7626 7627static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7628 struct cgroup_taskset *tset) 7629{ 7630 struct task_struct *task; 7631 7632 cgroup_taskset_for_each(task, tset) 7633 sched_move_task(task); 7634} 7635 7636static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7637 struct cgroup_subsys_state *old_css, 7638 struct task_struct *task) 7639{ 7640 /* 7641 * cgroup_exit() is called in the copy_process() failure path. 7642 * Ignore this case since the task hasn't ran yet, this avoids 7643 * trying to poke a half freed task state from generic code. 7644 */ 7645 if (!(task->flags & PF_EXITING)) 7646 return; 7647 7648 sched_move_task(task); 7649} 7650 7651#ifdef CONFIG_FAIR_GROUP_SCHED 7652static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7653 struct cftype *cftype, u64 shareval) 7654{ 7655 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7656} 7657 7658static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7659 struct cftype *cft) 7660{ 7661 struct task_group *tg = css_tg(css); 7662 7663 return (u64) scale_load_down(tg->shares); 7664} 7665 7666#ifdef CONFIG_CFS_BANDWIDTH 7667static DEFINE_MUTEX(cfs_constraints_mutex); 7668 7669const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7670const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7671 7672static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7673 7674static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7675{ 7676 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7677 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7678 7679 if (tg == &root_task_group) 7680 return -EINVAL; 7681 7682 /* 7683 * Ensure we have at some amount of bandwidth every period. This is 7684 * to prevent reaching a state of large arrears when throttled via 7685 * entity_tick() resulting in prolonged exit starvation. 7686 */ 7687 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7688 return -EINVAL; 7689 7690 /* 7691 * Likewise, bound things on the otherside by preventing insane quota 7692 * periods. This also allows us to normalize in computing quota 7693 * feasibility. 7694 */ 7695 if (period > max_cfs_quota_period) 7696 return -EINVAL; 7697 7698 mutex_lock(&cfs_constraints_mutex); 7699 ret = __cfs_schedulable(tg, period, quota); 7700 if (ret) 7701 goto out_unlock; 7702 7703 runtime_enabled = quota != RUNTIME_INF; 7704 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7705 /* 7706 * If we need to toggle cfs_bandwidth_used, off->on must occur 7707 * before making related changes, and on->off must occur afterwards 7708 */ 7709 if (runtime_enabled && !runtime_was_enabled) 7710 cfs_bandwidth_usage_inc(); 7711 raw_spin_lock_irq(&cfs_b->lock); 7712 cfs_b->period = ns_to_ktime(period); 7713 cfs_b->quota = quota; 7714 7715 __refill_cfs_bandwidth_runtime(cfs_b); 7716 /* restart the period timer (if active) to handle new period expiry */ 7717 if (runtime_enabled && cfs_b->timer_active) { 7718 /* force a reprogram */ 7719 cfs_b->timer_active = 0; 7720 __start_cfs_bandwidth(cfs_b); 7721 } 7722 raw_spin_unlock_irq(&cfs_b->lock); 7723 7724 for_each_possible_cpu(i) { 7725 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7726 struct rq *rq = cfs_rq->rq; 7727 7728 raw_spin_lock_irq(&rq->lock); 7729 cfs_rq->runtime_enabled = runtime_enabled; 7730 cfs_rq->runtime_remaining = 0; 7731 7732 if (cfs_rq->throttled) 7733 unthrottle_cfs_rq(cfs_rq); 7734 raw_spin_unlock_irq(&rq->lock); 7735 } 7736 if (runtime_was_enabled && !runtime_enabled) 7737 cfs_bandwidth_usage_dec(); 7738out_unlock: 7739 mutex_unlock(&cfs_constraints_mutex); 7740 7741 return ret; 7742} 7743 7744int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7745{ 7746 u64 quota, period; 7747 7748 period = ktime_to_ns(tg->cfs_bandwidth.period); 7749 if (cfs_quota_us < 0) 7750 quota = RUNTIME_INF; 7751 else 7752 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7753 7754 return tg_set_cfs_bandwidth(tg, period, quota); 7755} 7756 7757long tg_get_cfs_quota(struct task_group *tg) 7758{ 7759 u64 quota_us; 7760 7761 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7762 return -1; 7763 7764 quota_us = tg->cfs_bandwidth.quota; 7765 do_div(quota_us, NSEC_PER_USEC); 7766 7767 return quota_us; 7768} 7769 7770int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7771{ 7772 u64 quota, period; 7773 7774 period = (u64)cfs_period_us * NSEC_PER_USEC; 7775 quota = tg->cfs_bandwidth.quota; 7776 7777 return tg_set_cfs_bandwidth(tg, period, quota); 7778} 7779 7780long tg_get_cfs_period(struct task_group *tg) 7781{ 7782 u64 cfs_period_us; 7783 7784 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7785 do_div(cfs_period_us, NSEC_PER_USEC); 7786 7787 return cfs_period_us; 7788} 7789 7790static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7791 struct cftype *cft) 7792{ 7793 return tg_get_cfs_quota(css_tg(css)); 7794} 7795 7796static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7797 struct cftype *cftype, s64 cfs_quota_us) 7798{ 7799 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7800} 7801 7802static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7803 struct cftype *cft) 7804{ 7805 return tg_get_cfs_period(css_tg(css)); 7806} 7807 7808static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7809 struct cftype *cftype, u64 cfs_period_us) 7810{ 7811 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7812} 7813 7814struct cfs_schedulable_data { 7815 struct task_group *tg; 7816 u64 period, quota; 7817}; 7818 7819/* 7820 * normalize group quota/period to be quota/max_period 7821 * note: units are usecs 7822 */ 7823static u64 normalize_cfs_quota(struct task_group *tg, 7824 struct cfs_schedulable_data *d) 7825{ 7826 u64 quota, period; 7827 7828 if (tg == d->tg) { 7829 period = d->period; 7830 quota = d->quota; 7831 } else { 7832 period = tg_get_cfs_period(tg); 7833 quota = tg_get_cfs_quota(tg); 7834 } 7835 7836 /* note: these should typically be equivalent */ 7837 if (quota == RUNTIME_INF || quota == -1) 7838 return RUNTIME_INF; 7839 7840 return to_ratio(period, quota); 7841} 7842 7843static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7844{ 7845 struct cfs_schedulable_data *d = data; 7846 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7847 s64 quota = 0, parent_quota = -1; 7848 7849 if (!tg->parent) { 7850 quota = RUNTIME_INF; 7851 } else { 7852 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7853 7854 quota = normalize_cfs_quota(tg, d); 7855 parent_quota = parent_b->hierarchal_quota; 7856 7857 /* 7858 * ensure max(child_quota) <= parent_quota, inherit when no 7859 * limit is set 7860 */ 7861 if (quota == RUNTIME_INF) 7862 quota = parent_quota; 7863 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7864 return -EINVAL; 7865 } 7866 cfs_b->hierarchal_quota = quota; 7867 7868 return 0; 7869} 7870 7871static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7872{ 7873 int ret; 7874 struct cfs_schedulable_data data = { 7875 .tg = tg, 7876 .period = period, 7877 .quota = quota, 7878 }; 7879 7880 if (quota != RUNTIME_INF) { 7881 do_div(data.period, NSEC_PER_USEC); 7882 do_div(data.quota, NSEC_PER_USEC); 7883 } 7884 7885 rcu_read_lock(); 7886 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7887 rcu_read_unlock(); 7888 7889 return ret; 7890} 7891 7892static int cpu_stats_show(struct seq_file *sf, void *v) 7893{ 7894 struct task_group *tg = css_tg(seq_css(sf)); 7895 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7896 7897 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7898 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7899 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7900 7901 return 0; 7902} 7903#endif /* CONFIG_CFS_BANDWIDTH */ 7904#endif /* CONFIG_FAIR_GROUP_SCHED */ 7905 7906#ifdef CONFIG_RT_GROUP_SCHED 7907static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7908 struct cftype *cft, s64 val) 7909{ 7910 return sched_group_set_rt_runtime(css_tg(css), val); 7911} 7912 7913static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7914 struct cftype *cft) 7915{ 7916 return sched_group_rt_runtime(css_tg(css)); 7917} 7918 7919static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7920 struct cftype *cftype, u64 rt_period_us) 7921{ 7922 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7923} 7924 7925static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7926 struct cftype *cft) 7927{ 7928 return sched_group_rt_period(css_tg(css)); 7929} 7930#endif /* CONFIG_RT_GROUP_SCHED */ 7931 7932static struct cftype cpu_files[] = { 7933#ifdef CONFIG_FAIR_GROUP_SCHED 7934 { 7935 .name = "shares", 7936 .read_u64 = cpu_shares_read_u64, 7937 .write_u64 = cpu_shares_write_u64, 7938 }, 7939#endif 7940#ifdef CONFIG_CFS_BANDWIDTH 7941 { 7942 .name = "cfs_quota_us", 7943 .read_s64 = cpu_cfs_quota_read_s64, 7944 .write_s64 = cpu_cfs_quota_write_s64, 7945 }, 7946 { 7947 .name = "cfs_period_us", 7948 .read_u64 = cpu_cfs_period_read_u64, 7949 .write_u64 = cpu_cfs_period_write_u64, 7950 }, 7951 { 7952 .name = "stat", 7953 .seq_show = cpu_stats_show, 7954 }, 7955#endif 7956#ifdef CONFIG_RT_GROUP_SCHED 7957 { 7958 .name = "rt_runtime_us", 7959 .read_s64 = cpu_rt_runtime_read, 7960 .write_s64 = cpu_rt_runtime_write, 7961 }, 7962 { 7963 .name = "rt_period_us", 7964 .read_u64 = cpu_rt_period_read_uint, 7965 .write_u64 = cpu_rt_period_write_uint, 7966 }, 7967#endif 7968 { } /* terminate */ 7969}; 7970 7971struct cgroup_subsys cpu_cgrp_subsys = { 7972 .css_alloc = cpu_cgroup_css_alloc, 7973 .css_free = cpu_cgroup_css_free, 7974 .css_online = cpu_cgroup_css_online, 7975 .css_offline = cpu_cgroup_css_offline, 7976 .can_attach = cpu_cgroup_can_attach, 7977 .attach = cpu_cgroup_attach, 7978 .exit = cpu_cgroup_exit, 7979 .base_cftypes = cpu_files, 7980 .early_init = 1, 7981}; 7982 7983#endif /* CONFIG_CGROUP_SCHED */ 7984 7985void dump_cpu_task(int cpu) 7986{ 7987 pr_info("Task dump for CPU %d:\n", cpu); 7988 sched_show_task(cpu_curr(cpu)); 7989} 7990