core.c revision 6201b4d61fbf194df6371fb3376c5026cb8f5eec
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(cpu_of(rq), &rq->hrtick_csd, 0); 436 rq->hrtick_csd_pending = 1; 437 } 438} 439 440static int 441hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 442{ 443 int cpu = (int)(long)hcpu; 444 445 switch (action) { 446 case CPU_UP_CANCELED: 447 case CPU_UP_CANCELED_FROZEN: 448 case CPU_DOWN_PREPARE: 449 case CPU_DOWN_PREPARE_FROZEN: 450 case CPU_DEAD: 451 case CPU_DEAD_FROZEN: 452 hrtick_clear(cpu_rq(cpu)); 453 return NOTIFY_OK; 454 } 455 456 return NOTIFY_DONE; 457} 458 459static __init void init_hrtick(void) 460{ 461 hotcpu_notifier(hotplug_hrtick, 0); 462} 463#else 464/* 465 * Called to set the hrtick timer state. 466 * 467 * called with rq->lock held and irqs disabled 468 */ 469void hrtick_start(struct rq *rq, u64 delay) 470{ 471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 472 HRTIMER_MODE_REL_PINNED, 0); 473} 474 475static inline void init_hrtick(void) 476{ 477} 478#endif /* CONFIG_SMP */ 479 480static void init_rq_hrtick(struct rq *rq) 481{ 482#ifdef CONFIG_SMP 483 rq->hrtick_csd_pending = 0; 484 485 rq->hrtick_csd.flags = 0; 486 rq->hrtick_csd.func = __hrtick_start; 487 rq->hrtick_csd.info = rq; 488#endif 489 490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 491 rq->hrtick_timer.function = hrtick; 492} 493#else /* CONFIG_SCHED_HRTICK */ 494static inline void hrtick_clear(struct rq *rq) 495{ 496} 497 498static inline void init_rq_hrtick(struct rq *rq) 499{ 500} 501 502static inline void init_hrtick(void) 503{ 504} 505#endif /* CONFIG_SCHED_HRTICK */ 506 507/* 508 * resched_task - mark a task 'to be rescheduled now'. 509 * 510 * On UP this means the setting of the need_resched flag, on SMP it 511 * might also involve a cross-CPU call to trigger the scheduler on 512 * the target CPU. 513 */ 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 u64 st; 830 831 steal = paravirt_steal_clock(cpu_of(rq)); 832 steal -= rq->prev_steal_time_rq; 833 834 if (unlikely(steal > delta)) 835 steal = delta; 836 837 st = steal_ticks(steal); 838 steal = st * TICK_NSEC; 839 840 rq->prev_steal_time_rq += steal; 841 842 delta -= steal; 843 } 844#endif 845 846 rq->clock_task += delta; 847 848#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 849 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 850 sched_rt_avg_update(rq, irq_delta + steal); 851#endif 852} 853 854void sched_set_stop_task(int cpu, struct task_struct *stop) 855{ 856 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 857 struct task_struct *old_stop = cpu_rq(cpu)->stop; 858 859 if (stop) { 860 /* 861 * Make it appear like a SCHED_FIFO task, its something 862 * userspace knows about and won't get confused about. 863 * 864 * Also, it will make PI more or less work without too 865 * much confusion -- but then, stop work should not 866 * rely on PI working anyway. 867 */ 868 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 869 870 stop->sched_class = &stop_sched_class; 871 } 872 873 cpu_rq(cpu)->stop = stop; 874 875 if (old_stop) { 876 /* 877 * Reset it back to a normal scheduling class so that 878 * it can die in pieces. 879 */ 880 old_stop->sched_class = &rt_sched_class; 881 } 882} 883 884/* 885 * __normal_prio - return the priority that is based on the static prio 886 */ 887static inline int __normal_prio(struct task_struct *p) 888{ 889 return p->static_prio; 890} 891 892/* 893 * Calculate the expected normal priority: i.e. priority 894 * without taking RT-inheritance into account. Might be 895 * boosted by interactivity modifiers. Changes upon fork, 896 * setprio syscalls, and whenever the interactivity 897 * estimator recalculates. 898 */ 899static inline int normal_prio(struct task_struct *p) 900{ 901 int prio; 902 903 if (task_has_dl_policy(p)) 904 prio = MAX_DL_PRIO-1; 905 else if (task_has_rt_policy(p)) 906 prio = MAX_RT_PRIO-1 - p->rt_priority; 907 else 908 prio = __normal_prio(p); 909 return prio; 910} 911 912/* 913 * Calculate the current priority, i.e. the priority 914 * taken into account by the scheduler. This value might 915 * be boosted by RT tasks, or might be boosted by 916 * interactivity modifiers. Will be RT if the task got 917 * RT-boosted. If not then it returns p->normal_prio. 918 */ 919static int effective_prio(struct task_struct *p) 920{ 921 p->normal_prio = normal_prio(p); 922 /* 923 * If we are RT tasks or we were boosted to RT priority, 924 * keep the priority unchanged. Otherwise, update priority 925 * to the normal priority: 926 */ 927 if (!rt_prio(p->prio)) 928 return p->normal_prio; 929 return p->prio; 930} 931 932/** 933 * task_curr - is this task currently executing on a CPU? 934 * @p: the task in question. 935 * 936 * Return: 1 if the task is currently executing. 0 otherwise. 937 */ 938inline int task_curr(const struct task_struct *p) 939{ 940 return cpu_curr(task_cpu(p)) == p; 941} 942 943static inline void check_class_changed(struct rq *rq, struct task_struct *p, 944 const struct sched_class *prev_class, 945 int oldprio) 946{ 947 if (prev_class != p->sched_class) { 948 if (prev_class->switched_from) 949 prev_class->switched_from(rq, p); 950 p->sched_class->switched_to(rq, p); 951 } else if (oldprio != p->prio || dl_task(p)) 952 p->sched_class->prio_changed(rq, p, oldprio); 953} 954 955void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 956{ 957 const struct sched_class *class; 958 959 if (p->sched_class == rq->curr->sched_class) { 960 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 961 } else { 962 for_each_class(class) { 963 if (class == rq->curr->sched_class) 964 break; 965 if (class == p->sched_class) { 966 resched_task(rq->curr); 967 break; 968 } 969 } 970 } 971 972 /* 973 * A queue event has occurred, and we're going to schedule. In 974 * this case, we can save a useless back to back clock update. 975 */ 976 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 977 rq->skip_clock_update = 1; 978} 979 980#ifdef CONFIG_SMP 981void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 982{ 983#ifdef CONFIG_SCHED_DEBUG 984 /* 985 * We should never call set_task_cpu() on a blocked task, 986 * ttwu() will sort out the placement. 987 */ 988 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 989 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 990 991#ifdef CONFIG_LOCKDEP 992 /* 993 * The caller should hold either p->pi_lock or rq->lock, when changing 994 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 995 * 996 * sched_move_task() holds both and thus holding either pins the cgroup, 997 * see task_group(). 998 * 999 * Furthermore, all task_rq users should acquire both locks, see 1000 * task_rq_lock(). 1001 */ 1002 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1003 lockdep_is_held(&task_rq(p)->lock))); 1004#endif 1005#endif 1006 1007 trace_sched_migrate_task(p, new_cpu); 1008 1009 if (task_cpu(p) != new_cpu) { 1010 if (p->sched_class->migrate_task_rq) 1011 p->sched_class->migrate_task_rq(p, new_cpu); 1012 p->se.nr_migrations++; 1013 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1014 } 1015 1016 __set_task_cpu(p, new_cpu); 1017} 1018 1019static void __migrate_swap_task(struct task_struct *p, int cpu) 1020{ 1021 if (p->on_rq) { 1022 struct rq *src_rq, *dst_rq; 1023 1024 src_rq = task_rq(p); 1025 dst_rq = cpu_rq(cpu); 1026 1027 deactivate_task(src_rq, p, 0); 1028 set_task_cpu(p, cpu); 1029 activate_task(dst_rq, p, 0); 1030 check_preempt_curr(dst_rq, p, 0); 1031 } else { 1032 /* 1033 * Task isn't running anymore; make it appear like we migrated 1034 * it before it went to sleep. This means on wakeup we make the 1035 * previous cpu our targer instead of where it really is. 1036 */ 1037 p->wake_cpu = cpu; 1038 } 1039} 1040 1041struct migration_swap_arg { 1042 struct task_struct *src_task, *dst_task; 1043 int src_cpu, dst_cpu; 1044}; 1045 1046static int migrate_swap_stop(void *data) 1047{ 1048 struct migration_swap_arg *arg = data; 1049 struct rq *src_rq, *dst_rq; 1050 int ret = -EAGAIN; 1051 1052 src_rq = cpu_rq(arg->src_cpu); 1053 dst_rq = cpu_rq(arg->dst_cpu); 1054 1055 double_raw_lock(&arg->src_task->pi_lock, 1056 &arg->dst_task->pi_lock); 1057 double_rq_lock(src_rq, dst_rq); 1058 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1059 goto unlock; 1060 1061 if (task_cpu(arg->src_task) != arg->src_cpu) 1062 goto unlock; 1063 1064 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1065 goto unlock; 1066 1067 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1068 goto unlock; 1069 1070 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1071 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1072 1073 ret = 0; 1074 1075unlock: 1076 double_rq_unlock(src_rq, dst_rq); 1077 raw_spin_unlock(&arg->dst_task->pi_lock); 1078 raw_spin_unlock(&arg->src_task->pi_lock); 1079 1080 return ret; 1081} 1082 1083/* 1084 * Cross migrate two tasks 1085 */ 1086int migrate_swap(struct task_struct *cur, struct task_struct *p) 1087{ 1088 struct migration_swap_arg arg; 1089 int ret = -EINVAL; 1090 1091 arg = (struct migration_swap_arg){ 1092 .src_task = cur, 1093 .src_cpu = task_cpu(cur), 1094 .dst_task = p, 1095 .dst_cpu = task_cpu(p), 1096 }; 1097 1098 if (arg.src_cpu == arg.dst_cpu) 1099 goto out; 1100 1101 /* 1102 * These three tests are all lockless; this is OK since all of them 1103 * will be re-checked with proper locks held further down the line. 1104 */ 1105 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1106 goto out; 1107 1108 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1109 goto out; 1110 1111 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1112 goto out; 1113 1114 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1115 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1116 1117out: 1118 return ret; 1119} 1120 1121struct migration_arg { 1122 struct task_struct *task; 1123 int dest_cpu; 1124}; 1125 1126static int migration_cpu_stop(void *data); 1127 1128/* 1129 * wait_task_inactive - wait for a thread to unschedule. 1130 * 1131 * If @match_state is nonzero, it's the @p->state value just checked and 1132 * not expected to change. If it changes, i.e. @p might have woken up, 1133 * then return zero. When we succeed in waiting for @p to be off its CPU, 1134 * we return a positive number (its total switch count). If a second call 1135 * a short while later returns the same number, the caller can be sure that 1136 * @p has remained unscheduled the whole time. 1137 * 1138 * The caller must ensure that the task *will* unschedule sometime soon, 1139 * else this function might spin for a *long* time. This function can't 1140 * be called with interrupts off, or it may introduce deadlock with 1141 * smp_call_function() if an IPI is sent by the same process we are 1142 * waiting to become inactive. 1143 */ 1144unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1145{ 1146 unsigned long flags; 1147 int running, on_rq; 1148 unsigned long ncsw; 1149 struct rq *rq; 1150 1151 for (;;) { 1152 /* 1153 * We do the initial early heuristics without holding 1154 * any task-queue locks at all. We'll only try to get 1155 * the runqueue lock when things look like they will 1156 * work out! 1157 */ 1158 rq = task_rq(p); 1159 1160 /* 1161 * If the task is actively running on another CPU 1162 * still, just relax and busy-wait without holding 1163 * any locks. 1164 * 1165 * NOTE! Since we don't hold any locks, it's not 1166 * even sure that "rq" stays as the right runqueue! 1167 * But we don't care, since "task_running()" will 1168 * return false if the runqueue has changed and p 1169 * is actually now running somewhere else! 1170 */ 1171 while (task_running(rq, p)) { 1172 if (match_state && unlikely(p->state != match_state)) 1173 return 0; 1174 cpu_relax(); 1175 } 1176 1177 /* 1178 * Ok, time to look more closely! We need the rq 1179 * lock now, to be *sure*. If we're wrong, we'll 1180 * just go back and repeat. 1181 */ 1182 rq = task_rq_lock(p, &flags); 1183 trace_sched_wait_task(p); 1184 running = task_running(rq, p); 1185 on_rq = p->on_rq; 1186 ncsw = 0; 1187 if (!match_state || p->state == match_state) 1188 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1189 task_rq_unlock(rq, p, &flags); 1190 1191 /* 1192 * If it changed from the expected state, bail out now. 1193 */ 1194 if (unlikely(!ncsw)) 1195 break; 1196 1197 /* 1198 * Was it really running after all now that we 1199 * checked with the proper locks actually held? 1200 * 1201 * Oops. Go back and try again.. 1202 */ 1203 if (unlikely(running)) { 1204 cpu_relax(); 1205 continue; 1206 } 1207 1208 /* 1209 * It's not enough that it's not actively running, 1210 * it must be off the runqueue _entirely_, and not 1211 * preempted! 1212 * 1213 * So if it was still runnable (but just not actively 1214 * running right now), it's preempted, and we should 1215 * yield - it could be a while. 1216 */ 1217 if (unlikely(on_rq)) { 1218 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1219 1220 set_current_state(TASK_UNINTERRUPTIBLE); 1221 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1222 continue; 1223 } 1224 1225 /* 1226 * Ahh, all good. It wasn't running, and it wasn't 1227 * runnable, which means that it will never become 1228 * running in the future either. We're all done! 1229 */ 1230 break; 1231 } 1232 1233 return ncsw; 1234} 1235 1236/*** 1237 * kick_process - kick a running thread to enter/exit the kernel 1238 * @p: the to-be-kicked thread 1239 * 1240 * Cause a process which is running on another CPU to enter 1241 * kernel-mode, without any delay. (to get signals handled.) 1242 * 1243 * NOTE: this function doesn't have to take the runqueue lock, 1244 * because all it wants to ensure is that the remote task enters 1245 * the kernel. If the IPI races and the task has been migrated 1246 * to another CPU then no harm is done and the purpose has been 1247 * achieved as well. 1248 */ 1249void kick_process(struct task_struct *p) 1250{ 1251 int cpu; 1252 1253 preempt_disable(); 1254 cpu = task_cpu(p); 1255 if ((cpu != smp_processor_id()) && task_curr(p)) 1256 smp_send_reschedule(cpu); 1257 preempt_enable(); 1258} 1259EXPORT_SYMBOL_GPL(kick_process); 1260#endif /* CONFIG_SMP */ 1261 1262#ifdef CONFIG_SMP 1263/* 1264 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1265 */ 1266static int select_fallback_rq(int cpu, struct task_struct *p) 1267{ 1268 int nid = cpu_to_node(cpu); 1269 const struct cpumask *nodemask = NULL; 1270 enum { cpuset, possible, fail } state = cpuset; 1271 int dest_cpu; 1272 1273 /* 1274 * If the node that the cpu is on has been offlined, cpu_to_node() 1275 * will return -1. There is no cpu on the node, and we should 1276 * select the cpu on the other node. 1277 */ 1278 if (nid != -1) { 1279 nodemask = cpumask_of_node(nid); 1280 1281 /* Look for allowed, online CPU in same node. */ 1282 for_each_cpu(dest_cpu, nodemask) { 1283 if (!cpu_online(dest_cpu)) 1284 continue; 1285 if (!cpu_active(dest_cpu)) 1286 continue; 1287 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1288 return dest_cpu; 1289 } 1290 } 1291 1292 for (;;) { 1293 /* Any allowed, online CPU? */ 1294 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1295 if (!cpu_online(dest_cpu)) 1296 continue; 1297 if (!cpu_active(dest_cpu)) 1298 continue; 1299 goto out; 1300 } 1301 1302 switch (state) { 1303 case cpuset: 1304 /* No more Mr. Nice Guy. */ 1305 cpuset_cpus_allowed_fallback(p); 1306 state = possible; 1307 break; 1308 1309 case possible: 1310 do_set_cpus_allowed(p, cpu_possible_mask); 1311 state = fail; 1312 break; 1313 1314 case fail: 1315 BUG(); 1316 break; 1317 } 1318 } 1319 1320out: 1321 if (state != cpuset) { 1322 /* 1323 * Don't tell them about moving exiting tasks or 1324 * kernel threads (both mm NULL), since they never 1325 * leave kernel. 1326 */ 1327 if (p->mm && printk_ratelimit()) { 1328 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1329 task_pid_nr(p), p->comm, cpu); 1330 } 1331 } 1332 1333 return dest_cpu; 1334} 1335 1336/* 1337 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1338 */ 1339static inline 1340int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1341{ 1342 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1343 1344 /* 1345 * In order not to call set_task_cpu() on a blocking task we need 1346 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1347 * cpu. 1348 * 1349 * Since this is common to all placement strategies, this lives here. 1350 * 1351 * [ this allows ->select_task() to simply return task_cpu(p) and 1352 * not worry about this generic constraint ] 1353 */ 1354 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1355 !cpu_online(cpu))) 1356 cpu = select_fallback_rq(task_cpu(p), p); 1357 1358 return cpu; 1359} 1360 1361static void update_avg(u64 *avg, u64 sample) 1362{ 1363 s64 diff = sample - *avg; 1364 *avg += diff >> 3; 1365} 1366#endif 1367 1368static void 1369ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1370{ 1371#ifdef CONFIG_SCHEDSTATS 1372 struct rq *rq = this_rq(); 1373 1374#ifdef CONFIG_SMP 1375 int this_cpu = smp_processor_id(); 1376 1377 if (cpu == this_cpu) { 1378 schedstat_inc(rq, ttwu_local); 1379 schedstat_inc(p, se.statistics.nr_wakeups_local); 1380 } else { 1381 struct sched_domain *sd; 1382 1383 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1384 rcu_read_lock(); 1385 for_each_domain(this_cpu, sd) { 1386 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1387 schedstat_inc(sd, ttwu_wake_remote); 1388 break; 1389 } 1390 } 1391 rcu_read_unlock(); 1392 } 1393 1394 if (wake_flags & WF_MIGRATED) 1395 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1396 1397#endif /* CONFIG_SMP */ 1398 1399 schedstat_inc(rq, ttwu_count); 1400 schedstat_inc(p, se.statistics.nr_wakeups); 1401 1402 if (wake_flags & WF_SYNC) 1403 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1404 1405#endif /* CONFIG_SCHEDSTATS */ 1406} 1407 1408static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1409{ 1410 activate_task(rq, p, en_flags); 1411 p->on_rq = 1; 1412 1413 /* if a worker is waking up, notify workqueue */ 1414 if (p->flags & PF_WQ_WORKER) 1415 wq_worker_waking_up(p, cpu_of(rq)); 1416} 1417 1418/* 1419 * Mark the task runnable and perform wakeup-preemption. 1420 */ 1421static void 1422ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1423{ 1424 check_preempt_curr(rq, p, wake_flags); 1425 trace_sched_wakeup(p, true); 1426 1427 p->state = TASK_RUNNING; 1428#ifdef CONFIG_SMP 1429 if (p->sched_class->task_woken) 1430 p->sched_class->task_woken(rq, p); 1431 1432 if (rq->idle_stamp) { 1433 u64 delta = rq_clock(rq) - rq->idle_stamp; 1434 u64 max = 2*rq->max_idle_balance_cost; 1435 1436 update_avg(&rq->avg_idle, delta); 1437 1438 if (rq->avg_idle > max) 1439 rq->avg_idle = max; 1440 1441 rq->idle_stamp = 0; 1442 } 1443#endif 1444} 1445 1446static void 1447ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1448{ 1449#ifdef CONFIG_SMP 1450 if (p->sched_contributes_to_load) 1451 rq->nr_uninterruptible--; 1452#endif 1453 1454 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1455 ttwu_do_wakeup(rq, p, wake_flags); 1456} 1457 1458/* 1459 * Called in case the task @p isn't fully descheduled from its runqueue, 1460 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1461 * since all we need to do is flip p->state to TASK_RUNNING, since 1462 * the task is still ->on_rq. 1463 */ 1464static int ttwu_remote(struct task_struct *p, int wake_flags) 1465{ 1466 struct rq *rq; 1467 int ret = 0; 1468 1469 rq = __task_rq_lock(p); 1470 if (p->on_rq) { 1471 /* check_preempt_curr() may use rq clock */ 1472 update_rq_clock(rq); 1473 ttwu_do_wakeup(rq, p, wake_flags); 1474 ret = 1; 1475 } 1476 __task_rq_unlock(rq); 1477 1478 return ret; 1479} 1480 1481#ifdef CONFIG_SMP 1482static void sched_ttwu_pending(void) 1483{ 1484 struct rq *rq = this_rq(); 1485 struct llist_node *llist = llist_del_all(&rq->wake_list); 1486 struct task_struct *p; 1487 1488 raw_spin_lock(&rq->lock); 1489 1490 while (llist) { 1491 p = llist_entry(llist, struct task_struct, wake_entry); 1492 llist = llist_next(llist); 1493 ttwu_do_activate(rq, p, 0); 1494 } 1495 1496 raw_spin_unlock(&rq->lock); 1497} 1498 1499void scheduler_ipi(void) 1500{ 1501 /* 1502 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1503 * TIF_NEED_RESCHED remotely (for the first time) will also send 1504 * this IPI. 1505 */ 1506 preempt_fold_need_resched(); 1507 1508 if (llist_empty(&this_rq()->wake_list) 1509 && !tick_nohz_full_cpu(smp_processor_id()) 1510 && !got_nohz_idle_kick()) 1511 return; 1512 1513 /* 1514 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1515 * traditionally all their work was done from the interrupt return 1516 * path. Now that we actually do some work, we need to make sure 1517 * we do call them. 1518 * 1519 * Some archs already do call them, luckily irq_enter/exit nest 1520 * properly. 1521 * 1522 * Arguably we should visit all archs and update all handlers, 1523 * however a fair share of IPIs are still resched only so this would 1524 * somewhat pessimize the simple resched case. 1525 */ 1526 irq_enter(); 1527 tick_nohz_full_check(); 1528 sched_ttwu_pending(); 1529 1530 /* 1531 * Check if someone kicked us for doing the nohz idle load balance. 1532 */ 1533 if (unlikely(got_nohz_idle_kick())) { 1534 this_rq()->idle_balance = 1; 1535 raise_softirq_irqoff(SCHED_SOFTIRQ); 1536 } 1537 irq_exit(); 1538} 1539 1540static void ttwu_queue_remote(struct task_struct *p, int cpu) 1541{ 1542 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1543 smp_send_reschedule(cpu); 1544} 1545 1546bool cpus_share_cache(int this_cpu, int that_cpu) 1547{ 1548 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1549} 1550#endif /* CONFIG_SMP */ 1551 1552static void ttwu_queue(struct task_struct *p, int cpu) 1553{ 1554 struct rq *rq = cpu_rq(cpu); 1555 1556#if defined(CONFIG_SMP) 1557 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1558 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1559 ttwu_queue_remote(p, cpu); 1560 return; 1561 } 1562#endif 1563 1564 raw_spin_lock(&rq->lock); 1565 ttwu_do_activate(rq, p, 0); 1566 raw_spin_unlock(&rq->lock); 1567} 1568 1569/** 1570 * try_to_wake_up - wake up a thread 1571 * @p: the thread to be awakened 1572 * @state: the mask of task states that can be woken 1573 * @wake_flags: wake modifier flags (WF_*) 1574 * 1575 * Put it on the run-queue if it's not already there. The "current" 1576 * thread is always on the run-queue (except when the actual 1577 * re-schedule is in progress), and as such you're allowed to do 1578 * the simpler "current->state = TASK_RUNNING" to mark yourself 1579 * runnable without the overhead of this. 1580 * 1581 * Return: %true if @p was woken up, %false if it was already running. 1582 * or @state didn't match @p's state. 1583 */ 1584static int 1585try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1586{ 1587 unsigned long flags; 1588 int cpu, success = 0; 1589 1590 /* 1591 * If we are going to wake up a thread waiting for CONDITION we 1592 * need to ensure that CONDITION=1 done by the caller can not be 1593 * reordered with p->state check below. This pairs with mb() in 1594 * set_current_state() the waiting thread does. 1595 */ 1596 smp_mb__before_spinlock(); 1597 raw_spin_lock_irqsave(&p->pi_lock, flags); 1598 if (!(p->state & state)) 1599 goto out; 1600 1601 success = 1; /* we're going to change ->state */ 1602 cpu = task_cpu(p); 1603 1604 if (p->on_rq && ttwu_remote(p, wake_flags)) 1605 goto stat; 1606 1607#ifdef CONFIG_SMP 1608 /* 1609 * If the owning (remote) cpu is still in the middle of schedule() with 1610 * this task as prev, wait until its done referencing the task. 1611 */ 1612 while (p->on_cpu) 1613 cpu_relax(); 1614 /* 1615 * Pairs with the smp_wmb() in finish_lock_switch(). 1616 */ 1617 smp_rmb(); 1618 1619 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1620 p->state = TASK_WAKING; 1621 1622 if (p->sched_class->task_waking) 1623 p->sched_class->task_waking(p); 1624 1625 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1626 if (task_cpu(p) != cpu) { 1627 wake_flags |= WF_MIGRATED; 1628 set_task_cpu(p, cpu); 1629 } 1630#endif /* CONFIG_SMP */ 1631 1632 ttwu_queue(p, cpu); 1633stat: 1634 ttwu_stat(p, cpu, wake_flags); 1635out: 1636 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1637 1638 return success; 1639} 1640 1641/** 1642 * try_to_wake_up_local - try to wake up a local task with rq lock held 1643 * @p: the thread to be awakened 1644 * 1645 * Put @p on the run-queue if it's not already there. The caller must 1646 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1647 * the current task. 1648 */ 1649static void try_to_wake_up_local(struct task_struct *p) 1650{ 1651 struct rq *rq = task_rq(p); 1652 1653 if (WARN_ON_ONCE(rq != this_rq()) || 1654 WARN_ON_ONCE(p == current)) 1655 return; 1656 1657 lockdep_assert_held(&rq->lock); 1658 1659 if (!raw_spin_trylock(&p->pi_lock)) { 1660 raw_spin_unlock(&rq->lock); 1661 raw_spin_lock(&p->pi_lock); 1662 raw_spin_lock(&rq->lock); 1663 } 1664 1665 if (!(p->state & TASK_NORMAL)) 1666 goto out; 1667 1668 if (!p->on_rq) 1669 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1670 1671 ttwu_do_wakeup(rq, p, 0); 1672 ttwu_stat(p, smp_processor_id(), 0); 1673out: 1674 raw_spin_unlock(&p->pi_lock); 1675} 1676 1677/** 1678 * wake_up_process - Wake up a specific process 1679 * @p: The process to be woken up. 1680 * 1681 * Attempt to wake up the nominated process and move it to the set of runnable 1682 * processes. 1683 * 1684 * Return: 1 if the process was woken up, 0 if it was already running. 1685 * 1686 * It may be assumed that this function implies a write memory barrier before 1687 * changing the task state if and only if any tasks are woken up. 1688 */ 1689int wake_up_process(struct task_struct *p) 1690{ 1691 WARN_ON(task_is_stopped_or_traced(p)); 1692 return try_to_wake_up(p, TASK_NORMAL, 0); 1693} 1694EXPORT_SYMBOL(wake_up_process); 1695 1696int wake_up_state(struct task_struct *p, unsigned int state) 1697{ 1698 return try_to_wake_up(p, state, 0); 1699} 1700 1701/* 1702 * Perform scheduler related setup for a newly forked process p. 1703 * p is forked by current. 1704 * 1705 * __sched_fork() is basic setup used by init_idle() too: 1706 */ 1707static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1708{ 1709 p->on_rq = 0; 1710 1711 p->se.on_rq = 0; 1712 p->se.exec_start = 0; 1713 p->se.sum_exec_runtime = 0; 1714 p->se.prev_sum_exec_runtime = 0; 1715 p->se.nr_migrations = 0; 1716 p->se.vruntime = 0; 1717 INIT_LIST_HEAD(&p->se.group_node); 1718 1719#ifdef CONFIG_SCHEDSTATS 1720 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1721#endif 1722 1723 RB_CLEAR_NODE(&p->dl.rb_node); 1724 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1725 p->dl.dl_runtime = p->dl.runtime = 0; 1726 p->dl.dl_deadline = p->dl.deadline = 0; 1727 p->dl.dl_period = 0; 1728 p->dl.flags = 0; 1729 1730 INIT_LIST_HEAD(&p->rt.run_list); 1731 1732#ifdef CONFIG_PREEMPT_NOTIFIERS 1733 INIT_HLIST_HEAD(&p->preempt_notifiers); 1734#endif 1735 1736#ifdef CONFIG_NUMA_BALANCING 1737 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1738 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1739 p->mm->numa_scan_seq = 0; 1740 } 1741 1742 if (clone_flags & CLONE_VM) 1743 p->numa_preferred_nid = current->numa_preferred_nid; 1744 else 1745 p->numa_preferred_nid = -1; 1746 1747 p->node_stamp = 0ULL; 1748 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1749 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1750 p->numa_work.next = &p->numa_work; 1751 p->numa_faults = NULL; 1752 p->numa_faults_buffer = NULL; 1753 1754 INIT_LIST_HEAD(&p->numa_entry); 1755 p->numa_group = NULL; 1756#endif /* CONFIG_NUMA_BALANCING */ 1757} 1758 1759#ifdef CONFIG_NUMA_BALANCING 1760#ifdef CONFIG_SCHED_DEBUG 1761void set_numabalancing_state(bool enabled) 1762{ 1763 if (enabled) 1764 sched_feat_set("NUMA"); 1765 else 1766 sched_feat_set("NO_NUMA"); 1767} 1768#else 1769__read_mostly bool numabalancing_enabled; 1770 1771void set_numabalancing_state(bool enabled) 1772{ 1773 numabalancing_enabled = enabled; 1774} 1775#endif /* CONFIG_SCHED_DEBUG */ 1776 1777#ifdef CONFIG_PROC_SYSCTL 1778int sysctl_numa_balancing(struct ctl_table *table, int write, 1779 void __user *buffer, size_t *lenp, loff_t *ppos) 1780{ 1781 struct ctl_table t; 1782 int err; 1783 int state = numabalancing_enabled; 1784 1785 if (write && !capable(CAP_SYS_ADMIN)) 1786 return -EPERM; 1787 1788 t = *table; 1789 t.data = &state; 1790 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1791 if (err < 0) 1792 return err; 1793 if (write) 1794 set_numabalancing_state(state); 1795 return err; 1796} 1797#endif 1798#endif 1799 1800/* 1801 * fork()/clone()-time setup: 1802 */ 1803int sched_fork(unsigned long clone_flags, struct task_struct *p) 1804{ 1805 unsigned long flags; 1806 int cpu = get_cpu(); 1807 1808 __sched_fork(clone_flags, p); 1809 /* 1810 * We mark the process as running here. This guarantees that 1811 * nobody will actually run it, and a signal or other external 1812 * event cannot wake it up and insert it on the runqueue either. 1813 */ 1814 p->state = TASK_RUNNING; 1815 1816 /* 1817 * Make sure we do not leak PI boosting priority to the child. 1818 */ 1819 p->prio = current->normal_prio; 1820 1821 /* 1822 * Revert to default priority/policy on fork if requested. 1823 */ 1824 if (unlikely(p->sched_reset_on_fork)) { 1825 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1826 p->policy = SCHED_NORMAL; 1827 p->static_prio = NICE_TO_PRIO(0); 1828 p->rt_priority = 0; 1829 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1830 p->static_prio = NICE_TO_PRIO(0); 1831 1832 p->prio = p->normal_prio = __normal_prio(p); 1833 set_load_weight(p); 1834 1835 /* 1836 * We don't need the reset flag anymore after the fork. It has 1837 * fulfilled its duty: 1838 */ 1839 p->sched_reset_on_fork = 0; 1840 } 1841 1842 if (dl_prio(p->prio)) { 1843 put_cpu(); 1844 return -EAGAIN; 1845 } else if (rt_prio(p->prio)) { 1846 p->sched_class = &rt_sched_class; 1847 } else { 1848 p->sched_class = &fair_sched_class; 1849 } 1850 1851 if (p->sched_class->task_fork) 1852 p->sched_class->task_fork(p); 1853 1854 /* 1855 * The child is not yet in the pid-hash so no cgroup attach races, 1856 * and the cgroup is pinned to this child due to cgroup_fork() 1857 * is ran before sched_fork(). 1858 * 1859 * Silence PROVE_RCU. 1860 */ 1861 raw_spin_lock_irqsave(&p->pi_lock, flags); 1862 set_task_cpu(p, cpu); 1863 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1864 1865#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1866 if (likely(sched_info_on())) 1867 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1868#endif 1869#if defined(CONFIG_SMP) 1870 p->on_cpu = 0; 1871#endif 1872 init_task_preempt_count(p); 1873#ifdef CONFIG_SMP 1874 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1875 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1876#endif 1877 1878 put_cpu(); 1879 return 0; 1880} 1881 1882unsigned long to_ratio(u64 period, u64 runtime) 1883{ 1884 if (runtime == RUNTIME_INF) 1885 return 1ULL << 20; 1886 1887 /* 1888 * Doing this here saves a lot of checks in all 1889 * the calling paths, and returning zero seems 1890 * safe for them anyway. 1891 */ 1892 if (period == 0) 1893 return 0; 1894 1895 return div64_u64(runtime << 20, period); 1896} 1897 1898#ifdef CONFIG_SMP 1899inline struct dl_bw *dl_bw_of(int i) 1900{ 1901 return &cpu_rq(i)->rd->dl_bw; 1902} 1903 1904static inline int dl_bw_cpus(int i) 1905{ 1906 struct root_domain *rd = cpu_rq(i)->rd; 1907 int cpus = 0; 1908 1909 for_each_cpu_and(i, rd->span, cpu_active_mask) 1910 cpus++; 1911 1912 return cpus; 1913} 1914#else 1915inline struct dl_bw *dl_bw_of(int i) 1916{ 1917 return &cpu_rq(i)->dl.dl_bw; 1918} 1919 1920static inline int dl_bw_cpus(int i) 1921{ 1922 return 1; 1923} 1924#endif 1925 1926static inline 1927void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1928{ 1929 dl_b->total_bw -= tsk_bw; 1930} 1931 1932static inline 1933void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1934{ 1935 dl_b->total_bw += tsk_bw; 1936} 1937 1938static inline 1939bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1940{ 1941 return dl_b->bw != -1 && 1942 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1943} 1944 1945/* 1946 * We must be sure that accepting a new task (or allowing changing the 1947 * parameters of an existing one) is consistent with the bandwidth 1948 * constraints. If yes, this function also accordingly updates the currently 1949 * allocated bandwidth to reflect the new situation. 1950 * 1951 * This function is called while holding p's rq->lock. 1952 */ 1953static int dl_overflow(struct task_struct *p, int policy, 1954 const struct sched_attr *attr) 1955{ 1956 1957 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1958 u64 period = attr->sched_period; 1959 u64 runtime = attr->sched_runtime; 1960 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1961 int cpus, err = -1; 1962 1963 if (new_bw == p->dl.dl_bw) 1964 return 0; 1965 1966 /* 1967 * Either if a task, enters, leave, or stays -deadline but changes 1968 * its parameters, we may need to update accordingly the total 1969 * allocated bandwidth of the container. 1970 */ 1971 raw_spin_lock(&dl_b->lock); 1972 cpus = dl_bw_cpus(task_cpu(p)); 1973 if (dl_policy(policy) && !task_has_dl_policy(p) && 1974 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1975 __dl_add(dl_b, new_bw); 1976 err = 0; 1977 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1978 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1979 __dl_clear(dl_b, p->dl.dl_bw); 1980 __dl_add(dl_b, new_bw); 1981 err = 0; 1982 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1983 __dl_clear(dl_b, p->dl.dl_bw); 1984 err = 0; 1985 } 1986 raw_spin_unlock(&dl_b->lock); 1987 1988 return err; 1989} 1990 1991extern void init_dl_bw(struct dl_bw *dl_b); 1992 1993/* 1994 * wake_up_new_task - wake up a newly created task for the first time. 1995 * 1996 * This function will do some initial scheduler statistics housekeeping 1997 * that must be done for every newly created context, then puts the task 1998 * on the runqueue and wakes it. 1999 */ 2000void wake_up_new_task(struct task_struct *p) 2001{ 2002 unsigned long flags; 2003 struct rq *rq; 2004 2005 raw_spin_lock_irqsave(&p->pi_lock, flags); 2006#ifdef CONFIG_SMP 2007 /* 2008 * Fork balancing, do it here and not earlier because: 2009 * - cpus_allowed can change in the fork path 2010 * - any previously selected cpu might disappear through hotplug 2011 */ 2012 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2013#endif 2014 2015 /* Initialize new task's runnable average */ 2016 init_task_runnable_average(p); 2017 rq = __task_rq_lock(p); 2018 activate_task(rq, p, 0); 2019 p->on_rq = 1; 2020 trace_sched_wakeup_new(p, true); 2021 check_preempt_curr(rq, p, WF_FORK); 2022#ifdef CONFIG_SMP 2023 if (p->sched_class->task_woken) 2024 p->sched_class->task_woken(rq, p); 2025#endif 2026 task_rq_unlock(rq, p, &flags); 2027} 2028 2029#ifdef CONFIG_PREEMPT_NOTIFIERS 2030 2031/** 2032 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2033 * @notifier: notifier struct to register 2034 */ 2035void preempt_notifier_register(struct preempt_notifier *notifier) 2036{ 2037 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2038} 2039EXPORT_SYMBOL_GPL(preempt_notifier_register); 2040 2041/** 2042 * preempt_notifier_unregister - no longer interested in preemption notifications 2043 * @notifier: notifier struct to unregister 2044 * 2045 * This is safe to call from within a preemption notifier. 2046 */ 2047void preempt_notifier_unregister(struct preempt_notifier *notifier) 2048{ 2049 hlist_del(¬ifier->link); 2050} 2051EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2052 2053static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2054{ 2055 struct preempt_notifier *notifier; 2056 2057 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2058 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2059} 2060 2061static void 2062fire_sched_out_preempt_notifiers(struct task_struct *curr, 2063 struct task_struct *next) 2064{ 2065 struct preempt_notifier *notifier; 2066 2067 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2068 notifier->ops->sched_out(notifier, next); 2069} 2070 2071#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2072 2073static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2074{ 2075} 2076 2077static void 2078fire_sched_out_preempt_notifiers(struct task_struct *curr, 2079 struct task_struct *next) 2080{ 2081} 2082 2083#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2084 2085/** 2086 * prepare_task_switch - prepare to switch tasks 2087 * @rq: the runqueue preparing to switch 2088 * @prev: the current task that is being switched out 2089 * @next: the task we are going to switch to. 2090 * 2091 * This is called with the rq lock held and interrupts off. It must 2092 * be paired with a subsequent finish_task_switch after the context 2093 * switch. 2094 * 2095 * prepare_task_switch sets up locking and calls architecture specific 2096 * hooks. 2097 */ 2098static inline void 2099prepare_task_switch(struct rq *rq, struct task_struct *prev, 2100 struct task_struct *next) 2101{ 2102 trace_sched_switch(prev, next); 2103 sched_info_switch(rq, prev, next); 2104 perf_event_task_sched_out(prev, next); 2105 fire_sched_out_preempt_notifiers(prev, next); 2106 prepare_lock_switch(rq, next); 2107 prepare_arch_switch(next); 2108} 2109 2110/** 2111 * finish_task_switch - clean up after a task-switch 2112 * @rq: runqueue associated with task-switch 2113 * @prev: the thread we just switched away from. 2114 * 2115 * finish_task_switch must be called after the context switch, paired 2116 * with a prepare_task_switch call before the context switch. 2117 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2118 * and do any other architecture-specific cleanup actions. 2119 * 2120 * Note that we may have delayed dropping an mm in context_switch(). If 2121 * so, we finish that here outside of the runqueue lock. (Doing it 2122 * with the lock held can cause deadlocks; see schedule() for 2123 * details.) 2124 */ 2125static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2126 __releases(rq->lock) 2127{ 2128 struct mm_struct *mm = rq->prev_mm; 2129 long prev_state; 2130 2131 rq->prev_mm = NULL; 2132 2133 /* 2134 * A task struct has one reference for the use as "current". 2135 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2136 * schedule one last time. The schedule call will never return, and 2137 * the scheduled task must drop that reference. 2138 * The test for TASK_DEAD must occur while the runqueue locks are 2139 * still held, otherwise prev could be scheduled on another cpu, die 2140 * there before we look at prev->state, and then the reference would 2141 * be dropped twice. 2142 * Manfred Spraul <manfred@colorfullife.com> 2143 */ 2144 prev_state = prev->state; 2145 vtime_task_switch(prev); 2146 finish_arch_switch(prev); 2147 perf_event_task_sched_in(prev, current); 2148 finish_lock_switch(rq, prev); 2149 finish_arch_post_lock_switch(); 2150 2151 fire_sched_in_preempt_notifiers(current); 2152 if (mm) 2153 mmdrop(mm); 2154 if (unlikely(prev_state == TASK_DEAD)) { 2155 task_numa_free(prev); 2156 2157 if (prev->sched_class->task_dead) 2158 prev->sched_class->task_dead(prev); 2159 2160 /* 2161 * Remove function-return probe instances associated with this 2162 * task and put them back on the free list. 2163 */ 2164 kprobe_flush_task(prev); 2165 put_task_struct(prev); 2166 } 2167 2168 tick_nohz_task_switch(current); 2169} 2170 2171#ifdef CONFIG_SMP 2172 2173/* assumes rq->lock is held */ 2174static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 2175{ 2176 if (prev->sched_class->pre_schedule) 2177 prev->sched_class->pre_schedule(rq, prev); 2178} 2179 2180/* rq->lock is NOT held, but preemption is disabled */ 2181static inline void post_schedule(struct rq *rq) 2182{ 2183 if (rq->post_schedule) { 2184 unsigned long flags; 2185 2186 raw_spin_lock_irqsave(&rq->lock, flags); 2187 if (rq->curr->sched_class->post_schedule) 2188 rq->curr->sched_class->post_schedule(rq); 2189 raw_spin_unlock_irqrestore(&rq->lock, flags); 2190 2191 rq->post_schedule = 0; 2192 } 2193} 2194 2195#else 2196 2197static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2198{ 2199} 2200 2201static inline void post_schedule(struct rq *rq) 2202{ 2203} 2204 2205#endif 2206 2207/** 2208 * schedule_tail - first thing a freshly forked thread must call. 2209 * @prev: the thread we just switched away from. 2210 */ 2211asmlinkage void schedule_tail(struct task_struct *prev) 2212 __releases(rq->lock) 2213{ 2214 struct rq *rq = this_rq(); 2215 2216 finish_task_switch(rq, prev); 2217 2218 /* 2219 * FIXME: do we need to worry about rq being invalidated by the 2220 * task_switch? 2221 */ 2222 post_schedule(rq); 2223 2224#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2225 /* In this case, finish_task_switch does not reenable preemption */ 2226 preempt_enable(); 2227#endif 2228 if (current->set_child_tid) 2229 put_user(task_pid_vnr(current), current->set_child_tid); 2230} 2231 2232/* 2233 * context_switch - switch to the new MM and the new 2234 * thread's register state. 2235 */ 2236static inline void 2237context_switch(struct rq *rq, struct task_struct *prev, 2238 struct task_struct *next) 2239{ 2240 struct mm_struct *mm, *oldmm; 2241 2242 prepare_task_switch(rq, prev, next); 2243 2244 mm = next->mm; 2245 oldmm = prev->active_mm; 2246 /* 2247 * For paravirt, this is coupled with an exit in switch_to to 2248 * combine the page table reload and the switch backend into 2249 * one hypercall. 2250 */ 2251 arch_start_context_switch(prev); 2252 2253 if (!mm) { 2254 next->active_mm = oldmm; 2255 atomic_inc(&oldmm->mm_count); 2256 enter_lazy_tlb(oldmm, next); 2257 } else 2258 switch_mm(oldmm, mm, next); 2259 2260 if (!prev->mm) { 2261 prev->active_mm = NULL; 2262 rq->prev_mm = oldmm; 2263 } 2264 /* 2265 * Since the runqueue lock will be released by the next 2266 * task (which is an invalid locking op but in the case 2267 * of the scheduler it's an obvious special-case), so we 2268 * do an early lockdep release here: 2269 */ 2270#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2271 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2272#endif 2273 2274 context_tracking_task_switch(prev, next); 2275 /* Here we just switch the register state and the stack. */ 2276 switch_to(prev, next, prev); 2277 2278 barrier(); 2279 /* 2280 * this_rq must be evaluated again because prev may have moved 2281 * CPUs since it called schedule(), thus the 'rq' on its stack 2282 * frame will be invalid. 2283 */ 2284 finish_task_switch(this_rq(), prev); 2285} 2286 2287/* 2288 * nr_running and nr_context_switches: 2289 * 2290 * externally visible scheduler statistics: current number of runnable 2291 * threads, total number of context switches performed since bootup. 2292 */ 2293unsigned long nr_running(void) 2294{ 2295 unsigned long i, sum = 0; 2296 2297 for_each_online_cpu(i) 2298 sum += cpu_rq(i)->nr_running; 2299 2300 return sum; 2301} 2302 2303unsigned long long nr_context_switches(void) 2304{ 2305 int i; 2306 unsigned long long sum = 0; 2307 2308 for_each_possible_cpu(i) 2309 sum += cpu_rq(i)->nr_switches; 2310 2311 return sum; 2312} 2313 2314unsigned long nr_iowait(void) 2315{ 2316 unsigned long i, sum = 0; 2317 2318 for_each_possible_cpu(i) 2319 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2320 2321 return sum; 2322} 2323 2324unsigned long nr_iowait_cpu(int cpu) 2325{ 2326 struct rq *this = cpu_rq(cpu); 2327 return atomic_read(&this->nr_iowait); 2328} 2329 2330#ifdef CONFIG_SMP 2331 2332/* 2333 * sched_exec - execve() is a valuable balancing opportunity, because at 2334 * this point the task has the smallest effective memory and cache footprint. 2335 */ 2336void sched_exec(void) 2337{ 2338 struct task_struct *p = current; 2339 unsigned long flags; 2340 int dest_cpu; 2341 2342 raw_spin_lock_irqsave(&p->pi_lock, flags); 2343 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2344 if (dest_cpu == smp_processor_id()) 2345 goto unlock; 2346 2347 if (likely(cpu_active(dest_cpu))) { 2348 struct migration_arg arg = { p, dest_cpu }; 2349 2350 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2351 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2352 return; 2353 } 2354unlock: 2355 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2356} 2357 2358#endif 2359 2360DEFINE_PER_CPU(struct kernel_stat, kstat); 2361DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2362 2363EXPORT_PER_CPU_SYMBOL(kstat); 2364EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2365 2366/* 2367 * Return any ns on the sched_clock that have not yet been accounted in 2368 * @p in case that task is currently running. 2369 * 2370 * Called with task_rq_lock() held on @rq. 2371 */ 2372static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2373{ 2374 u64 ns = 0; 2375 2376 if (task_current(rq, p)) { 2377 update_rq_clock(rq); 2378 ns = rq_clock_task(rq) - p->se.exec_start; 2379 if ((s64)ns < 0) 2380 ns = 0; 2381 } 2382 2383 return ns; 2384} 2385 2386unsigned long long task_delta_exec(struct task_struct *p) 2387{ 2388 unsigned long flags; 2389 struct rq *rq; 2390 u64 ns = 0; 2391 2392 rq = task_rq_lock(p, &flags); 2393 ns = do_task_delta_exec(p, rq); 2394 task_rq_unlock(rq, p, &flags); 2395 2396 return ns; 2397} 2398 2399/* 2400 * Return accounted runtime for the task. 2401 * In case the task is currently running, return the runtime plus current's 2402 * pending runtime that have not been accounted yet. 2403 */ 2404unsigned long long task_sched_runtime(struct task_struct *p) 2405{ 2406 unsigned long flags; 2407 struct rq *rq; 2408 u64 ns = 0; 2409 2410#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2411 /* 2412 * 64-bit doesn't need locks to atomically read a 64bit value. 2413 * So we have a optimization chance when the task's delta_exec is 0. 2414 * Reading ->on_cpu is racy, but this is ok. 2415 * 2416 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2417 * If we race with it entering cpu, unaccounted time is 0. This is 2418 * indistinguishable from the read occurring a few cycles earlier. 2419 */ 2420 if (!p->on_cpu) 2421 return p->se.sum_exec_runtime; 2422#endif 2423 2424 rq = task_rq_lock(p, &flags); 2425 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2426 task_rq_unlock(rq, p, &flags); 2427 2428 return ns; 2429} 2430 2431/* 2432 * This function gets called by the timer code, with HZ frequency. 2433 * We call it with interrupts disabled. 2434 */ 2435void scheduler_tick(void) 2436{ 2437 int cpu = smp_processor_id(); 2438 struct rq *rq = cpu_rq(cpu); 2439 struct task_struct *curr = rq->curr; 2440 2441 sched_clock_tick(); 2442 2443 raw_spin_lock(&rq->lock); 2444 update_rq_clock(rq); 2445 curr->sched_class->task_tick(rq, curr, 0); 2446 update_cpu_load_active(rq); 2447 raw_spin_unlock(&rq->lock); 2448 2449 perf_event_task_tick(); 2450 2451#ifdef CONFIG_SMP 2452 rq->idle_balance = idle_cpu(cpu); 2453 trigger_load_balance(rq); 2454#endif 2455 rq_last_tick_reset(rq); 2456} 2457 2458#ifdef CONFIG_NO_HZ_FULL 2459/** 2460 * scheduler_tick_max_deferment 2461 * 2462 * Keep at least one tick per second when a single 2463 * active task is running because the scheduler doesn't 2464 * yet completely support full dynticks environment. 2465 * 2466 * This makes sure that uptime, CFS vruntime, load 2467 * balancing, etc... continue to move forward, even 2468 * with a very low granularity. 2469 * 2470 * Return: Maximum deferment in nanoseconds. 2471 */ 2472u64 scheduler_tick_max_deferment(void) 2473{ 2474 struct rq *rq = this_rq(); 2475 unsigned long next, now = ACCESS_ONCE(jiffies); 2476 2477 next = rq->last_sched_tick + HZ; 2478 2479 if (time_before_eq(next, now)) 2480 return 0; 2481 2482 return jiffies_to_nsecs(next - now); 2483} 2484#endif 2485 2486notrace unsigned long get_parent_ip(unsigned long addr) 2487{ 2488 if (in_lock_functions(addr)) { 2489 addr = CALLER_ADDR2; 2490 if (in_lock_functions(addr)) 2491 addr = CALLER_ADDR3; 2492 } 2493 return addr; 2494} 2495 2496#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2497 defined(CONFIG_PREEMPT_TRACER)) 2498 2499void __kprobes preempt_count_add(int val) 2500{ 2501#ifdef CONFIG_DEBUG_PREEMPT 2502 /* 2503 * Underflow? 2504 */ 2505 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2506 return; 2507#endif 2508 __preempt_count_add(val); 2509#ifdef CONFIG_DEBUG_PREEMPT 2510 /* 2511 * Spinlock count overflowing soon? 2512 */ 2513 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2514 PREEMPT_MASK - 10); 2515#endif 2516 if (preempt_count() == val) 2517 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2518} 2519EXPORT_SYMBOL(preempt_count_add); 2520 2521void __kprobes preempt_count_sub(int val) 2522{ 2523#ifdef CONFIG_DEBUG_PREEMPT 2524 /* 2525 * Underflow? 2526 */ 2527 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2528 return; 2529 /* 2530 * Is the spinlock portion underflowing? 2531 */ 2532 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2533 !(preempt_count() & PREEMPT_MASK))) 2534 return; 2535#endif 2536 2537 if (preempt_count() == val) 2538 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2539 __preempt_count_sub(val); 2540} 2541EXPORT_SYMBOL(preempt_count_sub); 2542 2543#endif 2544 2545/* 2546 * Print scheduling while atomic bug: 2547 */ 2548static noinline void __schedule_bug(struct task_struct *prev) 2549{ 2550 if (oops_in_progress) 2551 return; 2552 2553 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2554 prev->comm, prev->pid, preempt_count()); 2555 2556 debug_show_held_locks(prev); 2557 print_modules(); 2558 if (irqs_disabled()) 2559 print_irqtrace_events(prev); 2560 dump_stack(); 2561 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2562} 2563 2564/* 2565 * Various schedule()-time debugging checks and statistics: 2566 */ 2567static inline void schedule_debug(struct task_struct *prev) 2568{ 2569 /* 2570 * Test if we are atomic. Since do_exit() needs to call into 2571 * schedule() atomically, we ignore that path. Otherwise whine 2572 * if we are scheduling when we should not. 2573 */ 2574 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2575 __schedule_bug(prev); 2576 rcu_sleep_check(); 2577 2578 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2579 2580 schedstat_inc(this_rq(), sched_count); 2581} 2582 2583static void put_prev_task(struct rq *rq, struct task_struct *prev) 2584{ 2585 if (prev->on_rq || rq->skip_clock_update < 0) 2586 update_rq_clock(rq); 2587 prev->sched_class->put_prev_task(rq, prev); 2588} 2589 2590/* 2591 * Pick up the highest-prio task: 2592 */ 2593static inline struct task_struct * 2594pick_next_task(struct rq *rq) 2595{ 2596 const struct sched_class *class; 2597 struct task_struct *p; 2598 2599 /* 2600 * Optimization: we know that if all tasks are in 2601 * the fair class we can call that function directly: 2602 */ 2603 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2604 p = fair_sched_class.pick_next_task(rq); 2605 if (likely(p)) 2606 return p; 2607 } 2608 2609 for_each_class(class) { 2610 p = class->pick_next_task(rq); 2611 if (p) 2612 return p; 2613 } 2614 2615 BUG(); /* the idle class will always have a runnable task */ 2616} 2617 2618/* 2619 * __schedule() is the main scheduler function. 2620 * 2621 * The main means of driving the scheduler and thus entering this function are: 2622 * 2623 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2624 * 2625 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2626 * paths. For example, see arch/x86/entry_64.S. 2627 * 2628 * To drive preemption between tasks, the scheduler sets the flag in timer 2629 * interrupt handler scheduler_tick(). 2630 * 2631 * 3. Wakeups don't really cause entry into schedule(). They add a 2632 * task to the run-queue and that's it. 2633 * 2634 * Now, if the new task added to the run-queue preempts the current 2635 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2636 * called on the nearest possible occasion: 2637 * 2638 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2639 * 2640 * - in syscall or exception context, at the next outmost 2641 * preempt_enable(). (this might be as soon as the wake_up()'s 2642 * spin_unlock()!) 2643 * 2644 * - in IRQ context, return from interrupt-handler to 2645 * preemptible context 2646 * 2647 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2648 * then at the next: 2649 * 2650 * - cond_resched() call 2651 * - explicit schedule() call 2652 * - return from syscall or exception to user-space 2653 * - return from interrupt-handler to user-space 2654 */ 2655static void __sched __schedule(void) 2656{ 2657 struct task_struct *prev, *next; 2658 unsigned long *switch_count; 2659 struct rq *rq; 2660 int cpu; 2661 2662need_resched: 2663 preempt_disable(); 2664 cpu = smp_processor_id(); 2665 rq = cpu_rq(cpu); 2666 rcu_note_context_switch(cpu); 2667 prev = rq->curr; 2668 2669 schedule_debug(prev); 2670 2671 if (sched_feat(HRTICK)) 2672 hrtick_clear(rq); 2673 2674 /* 2675 * Make sure that signal_pending_state()->signal_pending() below 2676 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2677 * done by the caller to avoid the race with signal_wake_up(). 2678 */ 2679 smp_mb__before_spinlock(); 2680 raw_spin_lock_irq(&rq->lock); 2681 2682 switch_count = &prev->nivcsw; 2683 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2684 if (unlikely(signal_pending_state(prev->state, prev))) { 2685 prev->state = TASK_RUNNING; 2686 } else { 2687 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2688 prev->on_rq = 0; 2689 2690 /* 2691 * If a worker went to sleep, notify and ask workqueue 2692 * whether it wants to wake up a task to maintain 2693 * concurrency. 2694 */ 2695 if (prev->flags & PF_WQ_WORKER) { 2696 struct task_struct *to_wakeup; 2697 2698 to_wakeup = wq_worker_sleeping(prev, cpu); 2699 if (to_wakeup) 2700 try_to_wake_up_local(to_wakeup); 2701 } 2702 } 2703 switch_count = &prev->nvcsw; 2704 } 2705 2706 pre_schedule(rq, prev); 2707 2708 if (unlikely(!rq->nr_running)) 2709 idle_balance(cpu, rq); 2710 2711 put_prev_task(rq, prev); 2712 next = pick_next_task(rq); 2713 clear_tsk_need_resched(prev); 2714 clear_preempt_need_resched(); 2715 rq->skip_clock_update = 0; 2716 2717 if (likely(prev != next)) { 2718 rq->nr_switches++; 2719 rq->curr = next; 2720 ++*switch_count; 2721 2722 context_switch(rq, prev, next); /* unlocks the rq */ 2723 /* 2724 * The context switch have flipped the stack from under us 2725 * and restored the local variables which were saved when 2726 * this task called schedule() in the past. prev == current 2727 * is still correct, but it can be moved to another cpu/rq. 2728 */ 2729 cpu = smp_processor_id(); 2730 rq = cpu_rq(cpu); 2731 } else 2732 raw_spin_unlock_irq(&rq->lock); 2733 2734 post_schedule(rq); 2735 2736 sched_preempt_enable_no_resched(); 2737 if (need_resched()) 2738 goto need_resched; 2739} 2740 2741static inline void sched_submit_work(struct task_struct *tsk) 2742{ 2743 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2744 return; 2745 /* 2746 * If we are going to sleep and we have plugged IO queued, 2747 * make sure to submit it to avoid deadlocks. 2748 */ 2749 if (blk_needs_flush_plug(tsk)) 2750 blk_schedule_flush_plug(tsk); 2751} 2752 2753asmlinkage void __sched schedule(void) 2754{ 2755 struct task_struct *tsk = current; 2756 2757 sched_submit_work(tsk); 2758 __schedule(); 2759} 2760EXPORT_SYMBOL(schedule); 2761 2762#ifdef CONFIG_CONTEXT_TRACKING 2763asmlinkage void __sched schedule_user(void) 2764{ 2765 /* 2766 * If we come here after a random call to set_need_resched(), 2767 * or we have been woken up remotely but the IPI has not yet arrived, 2768 * we haven't yet exited the RCU idle mode. Do it here manually until 2769 * we find a better solution. 2770 */ 2771 user_exit(); 2772 schedule(); 2773 user_enter(); 2774} 2775#endif 2776 2777/** 2778 * schedule_preempt_disabled - called with preemption disabled 2779 * 2780 * Returns with preemption disabled. Note: preempt_count must be 1 2781 */ 2782void __sched schedule_preempt_disabled(void) 2783{ 2784 sched_preempt_enable_no_resched(); 2785 schedule(); 2786 preempt_disable(); 2787} 2788 2789#ifdef CONFIG_PREEMPT 2790/* 2791 * this is the entry point to schedule() from in-kernel preemption 2792 * off of preempt_enable. Kernel preemptions off return from interrupt 2793 * occur there and call schedule directly. 2794 */ 2795asmlinkage void __sched notrace preempt_schedule(void) 2796{ 2797 /* 2798 * If there is a non-zero preempt_count or interrupts are disabled, 2799 * we do not want to preempt the current task. Just return.. 2800 */ 2801 if (likely(!preemptible())) 2802 return; 2803 2804 do { 2805 __preempt_count_add(PREEMPT_ACTIVE); 2806 __schedule(); 2807 __preempt_count_sub(PREEMPT_ACTIVE); 2808 2809 /* 2810 * Check again in case we missed a preemption opportunity 2811 * between schedule and now. 2812 */ 2813 barrier(); 2814 } while (need_resched()); 2815} 2816EXPORT_SYMBOL(preempt_schedule); 2817#endif /* CONFIG_PREEMPT */ 2818 2819/* 2820 * this is the entry point to schedule() from kernel preemption 2821 * off of irq context. 2822 * Note, that this is called and return with irqs disabled. This will 2823 * protect us against recursive calling from irq. 2824 */ 2825asmlinkage void __sched preempt_schedule_irq(void) 2826{ 2827 enum ctx_state prev_state; 2828 2829 /* Catch callers which need to be fixed */ 2830 BUG_ON(preempt_count() || !irqs_disabled()); 2831 2832 prev_state = exception_enter(); 2833 2834 do { 2835 __preempt_count_add(PREEMPT_ACTIVE); 2836 local_irq_enable(); 2837 __schedule(); 2838 local_irq_disable(); 2839 __preempt_count_sub(PREEMPT_ACTIVE); 2840 2841 /* 2842 * Check again in case we missed a preemption opportunity 2843 * between schedule and now. 2844 */ 2845 barrier(); 2846 } while (need_resched()); 2847 2848 exception_exit(prev_state); 2849} 2850 2851int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2852 void *key) 2853{ 2854 return try_to_wake_up(curr->private, mode, wake_flags); 2855} 2856EXPORT_SYMBOL(default_wake_function); 2857 2858static long __sched 2859sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2860{ 2861 unsigned long flags; 2862 wait_queue_t wait; 2863 2864 init_waitqueue_entry(&wait, current); 2865 2866 __set_current_state(state); 2867 2868 spin_lock_irqsave(&q->lock, flags); 2869 __add_wait_queue(q, &wait); 2870 spin_unlock(&q->lock); 2871 timeout = schedule_timeout(timeout); 2872 spin_lock_irq(&q->lock); 2873 __remove_wait_queue(q, &wait); 2874 spin_unlock_irqrestore(&q->lock, flags); 2875 2876 return timeout; 2877} 2878 2879void __sched interruptible_sleep_on(wait_queue_head_t *q) 2880{ 2881 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2882} 2883EXPORT_SYMBOL(interruptible_sleep_on); 2884 2885long __sched 2886interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2887{ 2888 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2889} 2890EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2891 2892void __sched sleep_on(wait_queue_head_t *q) 2893{ 2894 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2895} 2896EXPORT_SYMBOL(sleep_on); 2897 2898long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2899{ 2900 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2901} 2902EXPORT_SYMBOL(sleep_on_timeout); 2903 2904#ifdef CONFIG_RT_MUTEXES 2905 2906/* 2907 * rt_mutex_setprio - set the current priority of a task 2908 * @p: task 2909 * @prio: prio value (kernel-internal form) 2910 * 2911 * This function changes the 'effective' priority of a task. It does 2912 * not touch ->normal_prio like __setscheduler(). 2913 * 2914 * Used by the rt_mutex code to implement priority inheritance logic. 2915 */ 2916void rt_mutex_setprio(struct task_struct *p, int prio) 2917{ 2918 int oldprio, on_rq, running, enqueue_flag = 0; 2919 struct rq *rq; 2920 const struct sched_class *prev_class; 2921 2922 BUG_ON(prio > MAX_PRIO); 2923 2924 rq = __task_rq_lock(p); 2925 2926 /* 2927 * Idle task boosting is a nono in general. There is one 2928 * exception, when PREEMPT_RT and NOHZ is active: 2929 * 2930 * The idle task calls get_next_timer_interrupt() and holds 2931 * the timer wheel base->lock on the CPU and another CPU wants 2932 * to access the timer (probably to cancel it). We can safely 2933 * ignore the boosting request, as the idle CPU runs this code 2934 * with interrupts disabled and will complete the lock 2935 * protected section without being interrupted. So there is no 2936 * real need to boost. 2937 */ 2938 if (unlikely(p == rq->idle)) { 2939 WARN_ON(p != rq->curr); 2940 WARN_ON(p->pi_blocked_on); 2941 goto out_unlock; 2942 } 2943 2944 trace_sched_pi_setprio(p, prio); 2945 p->pi_top_task = rt_mutex_get_top_task(p); 2946 oldprio = p->prio; 2947 prev_class = p->sched_class; 2948 on_rq = p->on_rq; 2949 running = task_current(rq, p); 2950 if (on_rq) 2951 dequeue_task(rq, p, 0); 2952 if (running) 2953 p->sched_class->put_prev_task(rq, p); 2954 2955 /* 2956 * Boosting condition are: 2957 * 1. -rt task is running and holds mutex A 2958 * --> -dl task blocks on mutex A 2959 * 2960 * 2. -dl task is running and holds mutex A 2961 * --> -dl task blocks on mutex A and could preempt the 2962 * running task 2963 */ 2964 if (dl_prio(prio)) { 2965 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2966 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2967 p->dl.dl_boosted = 1; 2968 p->dl.dl_throttled = 0; 2969 enqueue_flag = ENQUEUE_REPLENISH; 2970 } else 2971 p->dl.dl_boosted = 0; 2972 p->sched_class = &dl_sched_class; 2973 } else if (rt_prio(prio)) { 2974 if (dl_prio(oldprio)) 2975 p->dl.dl_boosted = 0; 2976 if (oldprio < prio) 2977 enqueue_flag = ENQUEUE_HEAD; 2978 p->sched_class = &rt_sched_class; 2979 } else { 2980 if (dl_prio(oldprio)) 2981 p->dl.dl_boosted = 0; 2982 p->sched_class = &fair_sched_class; 2983 } 2984 2985 p->prio = prio; 2986 2987 if (running) 2988 p->sched_class->set_curr_task(rq); 2989 if (on_rq) 2990 enqueue_task(rq, p, enqueue_flag); 2991 2992 check_class_changed(rq, p, prev_class, oldprio); 2993out_unlock: 2994 __task_rq_unlock(rq); 2995} 2996#endif 2997 2998void set_user_nice(struct task_struct *p, long nice) 2999{ 3000 int old_prio, delta, on_rq; 3001 unsigned long flags; 3002 struct rq *rq; 3003 3004 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3005 return; 3006 /* 3007 * We have to be careful, if called from sys_setpriority(), 3008 * the task might be in the middle of scheduling on another CPU. 3009 */ 3010 rq = task_rq_lock(p, &flags); 3011 /* 3012 * The RT priorities are set via sched_setscheduler(), but we still 3013 * allow the 'normal' nice value to be set - but as expected 3014 * it wont have any effect on scheduling until the task is 3015 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3016 */ 3017 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3018 p->static_prio = NICE_TO_PRIO(nice); 3019 goto out_unlock; 3020 } 3021 on_rq = p->on_rq; 3022 if (on_rq) 3023 dequeue_task(rq, p, 0); 3024 3025 p->static_prio = NICE_TO_PRIO(nice); 3026 set_load_weight(p); 3027 old_prio = p->prio; 3028 p->prio = effective_prio(p); 3029 delta = p->prio - old_prio; 3030 3031 if (on_rq) { 3032 enqueue_task(rq, p, 0); 3033 /* 3034 * If the task increased its priority or is running and 3035 * lowered its priority, then reschedule its CPU: 3036 */ 3037 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3038 resched_task(rq->curr); 3039 } 3040out_unlock: 3041 task_rq_unlock(rq, p, &flags); 3042} 3043EXPORT_SYMBOL(set_user_nice); 3044 3045/* 3046 * can_nice - check if a task can reduce its nice value 3047 * @p: task 3048 * @nice: nice value 3049 */ 3050int can_nice(const struct task_struct *p, const int nice) 3051{ 3052 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3053 int nice_rlim = 20 - nice; 3054 3055 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3056 capable(CAP_SYS_NICE)); 3057} 3058 3059#ifdef __ARCH_WANT_SYS_NICE 3060 3061/* 3062 * sys_nice - change the priority of the current process. 3063 * @increment: priority increment 3064 * 3065 * sys_setpriority is a more generic, but much slower function that 3066 * does similar things. 3067 */ 3068SYSCALL_DEFINE1(nice, int, increment) 3069{ 3070 long nice, retval; 3071 3072 /* 3073 * Setpriority might change our priority at the same moment. 3074 * We don't have to worry. Conceptually one call occurs first 3075 * and we have a single winner. 3076 */ 3077 if (increment < -40) 3078 increment = -40; 3079 if (increment > 40) 3080 increment = 40; 3081 3082 nice = TASK_NICE(current) + increment; 3083 if (nice < -20) 3084 nice = -20; 3085 if (nice > 19) 3086 nice = 19; 3087 3088 if (increment < 0 && !can_nice(current, nice)) 3089 return -EPERM; 3090 3091 retval = security_task_setnice(current, nice); 3092 if (retval) 3093 return retval; 3094 3095 set_user_nice(current, nice); 3096 return 0; 3097} 3098 3099#endif 3100 3101/** 3102 * task_prio - return the priority value of a given task. 3103 * @p: the task in question. 3104 * 3105 * Return: The priority value as seen by users in /proc. 3106 * RT tasks are offset by -200. Normal tasks are centered 3107 * around 0, value goes from -16 to +15. 3108 */ 3109int task_prio(const struct task_struct *p) 3110{ 3111 return p->prio - MAX_RT_PRIO; 3112} 3113 3114/** 3115 * task_nice - return the nice value of a given task. 3116 * @p: the task in question. 3117 * 3118 * Return: The nice value [ -20 ... 0 ... 19 ]. 3119 */ 3120int task_nice(const struct task_struct *p) 3121{ 3122 return TASK_NICE(p); 3123} 3124EXPORT_SYMBOL(task_nice); 3125 3126/** 3127 * idle_cpu - is a given cpu idle currently? 3128 * @cpu: the processor in question. 3129 * 3130 * Return: 1 if the CPU is currently idle. 0 otherwise. 3131 */ 3132int idle_cpu(int cpu) 3133{ 3134 struct rq *rq = cpu_rq(cpu); 3135 3136 if (rq->curr != rq->idle) 3137 return 0; 3138 3139 if (rq->nr_running) 3140 return 0; 3141 3142#ifdef CONFIG_SMP 3143 if (!llist_empty(&rq->wake_list)) 3144 return 0; 3145#endif 3146 3147 return 1; 3148} 3149 3150/** 3151 * idle_task - return the idle task for a given cpu. 3152 * @cpu: the processor in question. 3153 * 3154 * Return: The idle task for the cpu @cpu. 3155 */ 3156struct task_struct *idle_task(int cpu) 3157{ 3158 return cpu_rq(cpu)->idle; 3159} 3160 3161/** 3162 * find_process_by_pid - find a process with a matching PID value. 3163 * @pid: the pid in question. 3164 * 3165 * The task of @pid, if found. %NULL otherwise. 3166 */ 3167static struct task_struct *find_process_by_pid(pid_t pid) 3168{ 3169 return pid ? find_task_by_vpid(pid) : current; 3170} 3171 3172/* 3173 * This function initializes the sched_dl_entity of a newly becoming 3174 * SCHED_DEADLINE task. 3175 * 3176 * Only the static values are considered here, the actual runtime and the 3177 * absolute deadline will be properly calculated when the task is enqueued 3178 * for the first time with its new policy. 3179 */ 3180static void 3181__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3182{ 3183 struct sched_dl_entity *dl_se = &p->dl; 3184 3185 init_dl_task_timer(dl_se); 3186 dl_se->dl_runtime = attr->sched_runtime; 3187 dl_se->dl_deadline = attr->sched_deadline; 3188 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3189 dl_se->flags = attr->sched_flags; 3190 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3191 dl_se->dl_throttled = 0; 3192 dl_se->dl_new = 1; 3193} 3194 3195/* Actually do priority change: must hold pi & rq lock. */ 3196static void __setscheduler(struct rq *rq, struct task_struct *p, 3197 const struct sched_attr *attr) 3198{ 3199 int policy = attr->sched_policy; 3200 3201 if (policy == -1) /* setparam */ 3202 policy = p->policy; 3203 3204 p->policy = policy; 3205 3206 if (dl_policy(policy)) 3207 __setparam_dl(p, attr); 3208 else if (fair_policy(policy)) 3209 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3210 3211 /* 3212 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3213 * !rt_policy. Always setting this ensures that things like 3214 * getparam()/getattr() don't report silly values for !rt tasks. 3215 */ 3216 p->rt_priority = attr->sched_priority; 3217 3218 p->normal_prio = normal_prio(p); 3219 p->prio = rt_mutex_getprio(p); 3220 3221 if (dl_prio(p->prio)) 3222 p->sched_class = &dl_sched_class; 3223 else if (rt_prio(p->prio)) 3224 p->sched_class = &rt_sched_class; 3225 else 3226 p->sched_class = &fair_sched_class; 3227 3228 set_load_weight(p); 3229} 3230 3231static void 3232__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3233{ 3234 struct sched_dl_entity *dl_se = &p->dl; 3235 3236 attr->sched_priority = p->rt_priority; 3237 attr->sched_runtime = dl_se->dl_runtime; 3238 attr->sched_deadline = dl_se->dl_deadline; 3239 attr->sched_period = dl_se->dl_period; 3240 attr->sched_flags = dl_se->flags; 3241} 3242 3243/* 3244 * This function validates the new parameters of a -deadline task. 3245 * We ask for the deadline not being zero, and greater or equal 3246 * than the runtime, as well as the period of being zero or 3247 * greater than deadline. Furthermore, we have to be sure that 3248 * user parameters are above the internal resolution (1us); we 3249 * check sched_runtime only since it is always the smaller one. 3250 */ 3251static bool 3252__checkparam_dl(const struct sched_attr *attr) 3253{ 3254 return attr && attr->sched_deadline != 0 && 3255 (attr->sched_period == 0 || 3256 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3257 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3258 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3259} 3260 3261/* 3262 * check the target process has a UID that matches the current process's 3263 */ 3264static bool check_same_owner(struct task_struct *p) 3265{ 3266 const struct cred *cred = current_cred(), *pcred; 3267 bool match; 3268 3269 rcu_read_lock(); 3270 pcred = __task_cred(p); 3271 match = (uid_eq(cred->euid, pcred->euid) || 3272 uid_eq(cred->euid, pcred->uid)); 3273 rcu_read_unlock(); 3274 return match; 3275} 3276 3277static int __sched_setscheduler(struct task_struct *p, 3278 const struct sched_attr *attr, 3279 bool user) 3280{ 3281 int retval, oldprio, oldpolicy = -1, on_rq, running; 3282 int policy = attr->sched_policy; 3283 unsigned long flags; 3284 const struct sched_class *prev_class; 3285 struct rq *rq; 3286 int reset_on_fork; 3287 3288 /* may grab non-irq protected spin_locks */ 3289 BUG_ON(in_interrupt()); 3290recheck: 3291 /* double check policy once rq lock held */ 3292 if (policy < 0) { 3293 reset_on_fork = p->sched_reset_on_fork; 3294 policy = oldpolicy = p->policy; 3295 } else { 3296 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3297 3298 if (policy != SCHED_DEADLINE && 3299 policy != SCHED_FIFO && policy != SCHED_RR && 3300 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3301 policy != SCHED_IDLE) 3302 return -EINVAL; 3303 } 3304 3305 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3306 return -EINVAL; 3307 3308 /* 3309 * Valid priorities for SCHED_FIFO and SCHED_RR are 3310 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3311 * SCHED_BATCH and SCHED_IDLE is 0. 3312 */ 3313 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3314 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3315 return -EINVAL; 3316 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3317 (rt_policy(policy) != (attr->sched_priority != 0))) 3318 return -EINVAL; 3319 3320 /* 3321 * Allow unprivileged RT tasks to decrease priority: 3322 */ 3323 if (user && !capable(CAP_SYS_NICE)) { 3324 if (fair_policy(policy)) { 3325 if (attr->sched_nice < TASK_NICE(p) && 3326 !can_nice(p, attr->sched_nice)) 3327 return -EPERM; 3328 } 3329 3330 if (rt_policy(policy)) { 3331 unsigned long rlim_rtprio = 3332 task_rlimit(p, RLIMIT_RTPRIO); 3333 3334 /* can't set/change the rt policy */ 3335 if (policy != p->policy && !rlim_rtprio) 3336 return -EPERM; 3337 3338 /* can't increase priority */ 3339 if (attr->sched_priority > p->rt_priority && 3340 attr->sched_priority > rlim_rtprio) 3341 return -EPERM; 3342 } 3343 3344 /* 3345 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3346 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3347 */ 3348 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3349 if (!can_nice(p, TASK_NICE(p))) 3350 return -EPERM; 3351 } 3352 3353 /* can't change other user's priorities */ 3354 if (!check_same_owner(p)) 3355 return -EPERM; 3356 3357 /* Normal users shall not reset the sched_reset_on_fork flag */ 3358 if (p->sched_reset_on_fork && !reset_on_fork) 3359 return -EPERM; 3360 } 3361 3362 if (user) { 3363 retval = security_task_setscheduler(p); 3364 if (retval) 3365 return retval; 3366 } 3367 3368 /* 3369 * make sure no PI-waiters arrive (or leave) while we are 3370 * changing the priority of the task: 3371 * 3372 * To be able to change p->policy safely, the appropriate 3373 * runqueue lock must be held. 3374 */ 3375 rq = task_rq_lock(p, &flags); 3376 3377 /* 3378 * Changing the policy of the stop threads its a very bad idea 3379 */ 3380 if (p == rq->stop) { 3381 task_rq_unlock(rq, p, &flags); 3382 return -EINVAL; 3383 } 3384 3385 /* 3386 * If not changing anything there's no need to proceed further: 3387 */ 3388 if (unlikely(policy == p->policy)) { 3389 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p)) 3390 goto change; 3391 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3392 goto change; 3393 if (dl_policy(policy)) 3394 goto change; 3395 3396 task_rq_unlock(rq, p, &flags); 3397 return 0; 3398 } 3399change: 3400 3401 if (user) { 3402#ifdef CONFIG_RT_GROUP_SCHED 3403 /* 3404 * Do not allow realtime tasks into groups that have no runtime 3405 * assigned. 3406 */ 3407 if (rt_bandwidth_enabled() && rt_policy(policy) && 3408 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3409 !task_group_is_autogroup(task_group(p))) { 3410 task_rq_unlock(rq, p, &flags); 3411 return -EPERM; 3412 } 3413#endif 3414#ifdef CONFIG_SMP 3415 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3416 cpumask_t *span = rq->rd->span; 3417 3418 /* 3419 * Don't allow tasks with an affinity mask smaller than 3420 * the entire root_domain to become SCHED_DEADLINE. We 3421 * will also fail if there's no bandwidth available. 3422 */ 3423 if (!cpumask_subset(span, &p->cpus_allowed) || 3424 rq->rd->dl_bw.bw == 0) { 3425 task_rq_unlock(rq, p, &flags); 3426 return -EPERM; 3427 } 3428 } 3429#endif 3430 } 3431 3432 /* recheck policy now with rq lock held */ 3433 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3434 policy = oldpolicy = -1; 3435 task_rq_unlock(rq, p, &flags); 3436 goto recheck; 3437 } 3438 3439 /* 3440 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3441 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3442 * is available. 3443 */ 3444 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3445 task_rq_unlock(rq, p, &flags); 3446 return -EBUSY; 3447 } 3448 3449 on_rq = p->on_rq; 3450 running = task_current(rq, p); 3451 if (on_rq) 3452 dequeue_task(rq, p, 0); 3453 if (running) 3454 p->sched_class->put_prev_task(rq, p); 3455 3456 p->sched_reset_on_fork = reset_on_fork; 3457 3458 oldprio = p->prio; 3459 prev_class = p->sched_class; 3460 __setscheduler(rq, p, attr); 3461 3462 if (running) 3463 p->sched_class->set_curr_task(rq); 3464 if (on_rq) 3465 enqueue_task(rq, p, 0); 3466 3467 check_class_changed(rq, p, prev_class, oldprio); 3468 task_rq_unlock(rq, p, &flags); 3469 3470 rt_mutex_adjust_pi(p); 3471 3472 return 0; 3473} 3474 3475static int _sched_setscheduler(struct task_struct *p, int policy, 3476 const struct sched_param *param, bool check) 3477{ 3478 struct sched_attr attr = { 3479 .sched_policy = policy, 3480 .sched_priority = param->sched_priority, 3481 .sched_nice = PRIO_TO_NICE(p->static_prio), 3482 }; 3483 3484 /* 3485 * Fixup the legacy SCHED_RESET_ON_FORK hack 3486 */ 3487 if (policy & SCHED_RESET_ON_FORK) { 3488 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3489 policy &= ~SCHED_RESET_ON_FORK; 3490 attr.sched_policy = policy; 3491 } 3492 3493 return __sched_setscheduler(p, &attr, check); 3494} 3495/** 3496 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3497 * @p: the task in question. 3498 * @policy: new policy. 3499 * @param: structure containing the new RT priority. 3500 * 3501 * Return: 0 on success. An error code otherwise. 3502 * 3503 * NOTE that the task may be already dead. 3504 */ 3505int sched_setscheduler(struct task_struct *p, int policy, 3506 const struct sched_param *param) 3507{ 3508 return _sched_setscheduler(p, policy, param, true); 3509} 3510EXPORT_SYMBOL_GPL(sched_setscheduler); 3511 3512int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3513{ 3514 return __sched_setscheduler(p, attr, true); 3515} 3516EXPORT_SYMBOL_GPL(sched_setattr); 3517 3518/** 3519 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3520 * @p: the task in question. 3521 * @policy: new policy. 3522 * @param: structure containing the new RT priority. 3523 * 3524 * Just like sched_setscheduler, only don't bother checking if the 3525 * current context has permission. For example, this is needed in 3526 * stop_machine(): we create temporary high priority worker threads, 3527 * but our caller might not have that capability. 3528 * 3529 * Return: 0 on success. An error code otherwise. 3530 */ 3531int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3532 const struct sched_param *param) 3533{ 3534 return _sched_setscheduler(p, policy, param, false); 3535} 3536 3537static int 3538do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3539{ 3540 struct sched_param lparam; 3541 struct task_struct *p; 3542 int retval; 3543 3544 if (!param || pid < 0) 3545 return -EINVAL; 3546 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3547 return -EFAULT; 3548 3549 rcu_read_lock(); 3550 retval = -ESRCH; 3551 p = find_process_by_pid(pid); 3552 if (p != NULL) 3553 retval = sched_setscheduler(p, policy, &lparam); 3554 rcu_read_unlock(); 3555 3556 return retval; 3557} 3558 3559/* 3560 * Mimics kernel/events/core.c perf_copy_attr(). 3561 */ 3562static int sched_copy_attr(struct sched_attr __user *uattr, 3563 struct sched_attr *attr) 3564{ 3565 u32 size; 3566 int ret; 3567 3568 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3569 return -EFAULT; 3570 3571 /* 3572 * zero the full structure, so that a short copy will be nice. 3573 */ 3574 memset(attr, 0, sizeof(*attr)); 3575 3576 ret = get_user(size, &uattr->size); 3577 if (ret) 3578 return ret; 3579 3580 if (size > PAGE_SIZE) /* silly large */ 3581 goto err_size; 3582 3583 if (!size) /* abi compat */ 3584 size = SCHED_ATTR_SIZE_VER0; 3585 3586 if (size < SCHED_ATTR_SIZE_VER0) 3587 goto err_size; 3588 3589 /* 3590 * If we're handed a bigger struct than we know of, 3591 * ensure all the unknown bits are 0 - i.e. new 3592 * user-space does not rely on any kernel feature 3593 * extensions we dont know about yet. 3594 */ 3595 if (size > sizeof(*attr)) { 3596 unsigned char __user *addr; 3597 unsigned char __user *end; 3598 unsigned char val; 3599 3600 addr = (void __user *)uattr + sizeof(*attr); 3601 end = (void __user *)uattr + size; 3602 3603 for (; addr < end; addr++) { 3604 ret = get_user(val, addr); 3605 if (ret) 3606 return ret; 3607 if (val) 3608 goto err_size; 3609 } 3610 size = sizeof(*attr); 3611 } 3612 3613 ret = copy_from_user(attr, uattr, size); 3614 if (ret) 3615 return -EFAULT; 3616 3617 /* 3618 * XXX: do we want to be lenient like existing syscalls; or do we want 3619 * to be strict and return an error on out-of-bounds values? 3620 */ 3621 attr->sched_nice = clamp(attr->sched_nice, -20, 19); 3622 3623out: 3624 return ret; 3625 3626err_size: 3627 put_user(sizeof(*attr), &uattr->size); 3628 ret = -E2BIG; 3629 goto out; 3630} 3631 3632/** 3633 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3634 * @pid: the pid in question. 3635 * @policy: new policy. 3636 * @param: structure containing the new RT priority. 3637 * 3638 * Return: 0 on success. An error code otherwise. 3639 */ 3640SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3641 struct sched_param __user *, param) 3642{ 3643 /* negative values for policy are not valid */ 3644 if (policy < 0) 3645 return -EINVAL; 3646 3647 return do_sched_setscheduler(pid, policy, param); 3648} 3649 3650/** 3651 * sys_sched_setparam - set/change the RT priority of a thread 3652 * @pid: the pid in question. 3653 * @param: structure containing the new RT priority. 3654 * 3655 * Return: 0 on success. An error code otherwise. 3656 */ 3657SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3658{ 3659 return do_sched_setscheduler(pid, -1, param); 3660} 3661 3662/** 3663 * sys_sched_setattr - same as above, but with extended sched_attr 3664 * @pid: the pid in question. 3665 * @uattr: structure containing the extended parameters. 3666 */ 3667SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr) 3668{ 3669 struct sched_attr attr; 3670 struct task_struct *p; 3671 int retval; 3672 3673 if (!uattr || pid < 0) 3674 return -EINVAL; 3675 3676 if (sched_copy_attr(uattr, &attr)) 3677 return -EFAULT; 3678 3679 rcu_read_lock(); 3680 retval = -ESRCH; 3681 p = find_process_by_pid(pid); 3682 if (p != NULL) 3683 retval = sched_setattr(p, &attr); 3684 rcu_read_unlock(); 3685 3686 return retval; 3687} 3688 3689/** 3690 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3691 * @pid: the pid in question. 3692 * 3693 * Return: On success, the policy of the thread. Otherwise, a negative error 3694 * code. 3695 */ 3696SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3697{ 3698 struct task_struct *p; 3699 int retval; 3700 3701 if (pid < 0) 3702 return -EINVAL; 3703 3704 retval = -ESRCH; 3705 rcu_read_lock(); 3706 p = find_process_by_pid(pid); 3707 if (p) { 3708 retval = security_task_getscheduler(p); 3709 if (!retval) 3710 retval = p->policy 3711 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3712 } 3713 rcu_read_unlock(); 3714 return retval; 3715} 3716 3717/** 3718 * sys_sched_getparam - get the RT priority of a thread 3719 * @pid: the pid in question. 3720 * @param: structure containing the RT priority. 3721 * 3722 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3723 * code. 3724 */ 3725SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3726{ 3727 struct sched_param lp; 3728 struct task_struct *p; 3729 int retval; 3730 3731 if (!param || pid < 0) 3732 return -EINVAL; 3733 3734 rcu_read_lock(); 3735 p = find_process_by_pid(pid); 3736 retval = -ESRCH; 3737 if (!p) 3738 goto out_unlock; 3739 3740 retval = security_task_getscheduler(p); 3741 if (retval) 3742 goto out_unlock; 3743 3744 if (task_has_dl_policy(p)) { 3745 retval = -EINVAL; 3746 goto out_unlock; 3747 } 3748 lp.sched_priority = p->rt_priority; 3749 rcu_read_unlock(); 3750 3751 /* 3752 * This one might sleep, we cannot do it with a spinlock held ... 3753 */ 3754 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3755 3756 return retval; 3757 3758out_unlock: 3759 rcu_read_unlock(); 3760 return retval; 3761} 3762 3763static int sched_read_attr(struct sched_attr __user *uattr, 3764 struct sched_attr *attr, 3765 unsigned int usize) 3766{ 3767 int ret; 3768 3769 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3770 return -EFAULT; 3771 3772 /* 3773 * If we're handed a smaller struct than we know of, 3774 * ensure all the unknown bits are 0 - i.e. old 3775 * user-space does not get uncomplete information. 3776 */ 3777 if (usize < sizeof(*attr)) { 3778 unsigned char *addr; 3779 unsigned char *end; 3780 3781 addr = (void *)attr + usize; 3782 end = (void *)attr + sizeof(*attr); 3783 3784 for (; addr < end; addr++) { 3785 if (*addr) 3786 goto err_size; 3787 } 3788 3789 attr->size = usize; 3790 } 3791 3792 ret = copy_to_user(uattr, attr, usize); 3793 if (ret) 3794 return -EFAULT; 3795 3796out: 3797 return ret; 3798 3799err_size: 3800 ret = -E2BIG; 3801 goto out; 3802} 3803 3804/** 3805 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3806 * @pid: the pid in question. 3807 * @uattr: structure containing the extended parameters. 3808 * @size: sizeof(attr) for fwd/bwd comp. 3809 */ 3810SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3811 unsigned int, size) 3812{ 3813 struct sched_attr attr = { 3814 .size = sizeof(struct sched_attr), 3815 }; 3816 struct task_struct *p; 3817 int retval; 3818 3819 if (!uattr || pid < 0 || size > PAGE_SIZE || 3820 size < SCHED_ATTR_SIZE_VER0) 3821 return -EINVAL; 3822 3823 rcu_read_lock(); 3824 p = find_process_by_pid(pid); 3825 retval = -ESRCH; 3826 if (!p) 3827 goto out_unlock; 3828 3829 retval = security_task_getscheduler(p); 3830 if (retval) 3831 goto out_unlock; 3832 3833 attr.sched_policy = p->policy; 3834 if (p->sched_reset_on_fork) 3835 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3836 if (task_has_dl_policy(p)) 3837 __getparam_dl(p, &attr); 3838 else if (task_has_rt_policy(p)) 3839 attr.sched_priority = p->rt_priority; 3840 else 3841 attr.sched_nice = TASK_NICE(p); 3842 3843 rcu_read_unlock(); 3844 3845 retval = sched_read_attr(uattr, &attr, size); 3846 return retval; 3847 3848out_unlock: 3849 rcu_read_unlock(); 3850 return retval; 3851} 3852 3853long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3854{ 3855 cpumask_var_t cpus_allowed, new_mask; 3856 struct task_struct *p; 3857 int retval; 3858 3859 rcu_read_lock(); 3860 3861 p = find_process_by_pid(pid); 3862 if (!p) { 3863 rcu_read_unlock(); 3864 return -ESRCH; 3865 } 3866 3867 /* Prevent p going away */ 3868 get_task_struct(p); 3869 rcu_read_unlock(); 3870 3871 if (p->flags & PF_NO_SETAFFINITY) { 3872 retval = -EINVAL; 3873 goto out_put_task; 3874 } 3875 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3876 retval = -ENOMEM; 3877 goto out_put_task; 3878 } 3879 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3880 retval = -ENOMEM; 3881 goto out_free_cpus_allowed; 3882 } 3883 retval = -EPERM; 3884 if (!check_same_owner(p)) { 3885 rcu_read_lock(); 3886 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3887 rcu_read_unlock(); 3888 goto out_unlock; 3889 } 3890 rcu_read_unlock(); 3891 } 3892 3893 retval = security_task_setscheduler(p); 3894 if (retval) 3895 goto out_unlock; 3896 3897 3898 cpuset_cpus_allowed(p, cpus_allowed); 3899 cpumask_and(new_mask, in_mask, cpus_allowed); 3900 3901 /* 3902 * Since bandwidth control happens on root_domain basis, 3903 * if admission test is enabled, we only admit -deadline 3904 * tasks allowed to run on all the CPUs in the task's 3905 * root_domain. 3906 */ 3907#ifdef CONFIG_SMP 3908 if (task_has_dl_policy(p)) { 3909 const struct cpumask *span = task_rq(p)->rd->span; 3910 3911 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3912 retval = -EBUSY; 3913 goto out_unlock; 3914 } 3915 } 3916#endif 3917again: 3918 retval = set_cpus_allowed_ptr(p, new_mask); 3919 3920 if (!retval) { 3921 cpuset_cpus_allowed(p, cpus_allowed); 3922 if (!cpumask_subset(new_mask, cpus_allowed)) { 3923 /* 3924 * We must have raced with a concurrent cpuset 3925 * update. Just reset the cpus_allowed to the 3926 * cpuset's cpus_allowed 3927 */ 3928 cpumask_copy(new_mask, cpus_allowed); 3929 goto again; 3930 } 3931 } 3932out_unlock: 3933 free_cpumask_var(new_mask); 3934out_free_cpus_allowed: 3935 free_cpumask_var(cpus_allowed); 3936out_put_task: 3937 put_task_struct(p); 3938 return retval; 3939} 3940 3941static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3942 struct cpumask *new_mask) 3943{ 3944 if (len < cpumask_size()) 3945 cpumask_clear(new_mask); 3946 else if (len > cpumask_size()) 3947 len = cpumask_size(); 3948 3949 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3950} 3951 3952/** 3953 * sys_sched_setaffinity - set the cpu affinity of a process 3954 * @pid: pid of the process 3955 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3956 * @user_mask_ptr: user-space pointer to the new cpu mask 3957 * 3958 * Return: 0 on success. An error code otherwise. 3959 */ 3960SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3961 unsigned long __user *, user_mask_ptr) 3962{ 3963 cpumask_var_t new_mask; 3964 int retval; 3965 3966 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3967 return -ENOMEM; 3968 3969 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3970 if (retval == 0) 3971 retval = sched_setaffinity(pid, new_mask); 3972 free_cpumask_var(new_mask); 3973 return retval; 3974} 3975 3976long sched_getaffinity(pid_t pid, struct cpumask *mask) 3977{ 3978 struct task_struct *p; 3979 unsigned long flags; 3980 int retval; 3981 3982 rcu_read_lock(); 3983 3984 retval = -ESRCH; 3985 p = find_process_by_pid(pid); 3986 if (!p) 3987 goto out_unlock; 3988 3989 retval = security_task_getscheduler(p); 3990 if (retval) 3991 goto out_unlock; 3992 3993 raw_spin_lock_irqsave(&p->pi_lock, flags); 3994 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3995 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3996 3997out_unlock: 3998 rcu_read_unlock(); 3999 4000 return retval; 4001} 4002 4003/** 4004 * sys_sched_getaffinity - get the cpu affinity of a process 4005 * @pid: pid of the process 4006 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4007 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4008 * 4009 * Return: 0 on success. An error code otherwise. 4010 */ 4011SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4012 unsigned long __user *, user_mask_ptr) 4013{ 4014 int ret; 4015 cpumask_var_t mask; 4016 4017 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4018 return -EINVAL; 4019 if (len & (sizeof(unsigned long)-1)) 4020 return -EINVAL; 4021 4022 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4023 return -ENOMEM; 4024 4025 ret = sched_getaffinity(pid, mask); 4026 if (ret == 0) { 4027 size_t retlen = min_t(size_t, len, cpumask_size()); 4028 4029 if (copy_to_user(user_mask_ptr, mask, retlen)) 4030 ret = -EFAULT; 4031 else 4032 ret = retlen; 4033 } 4034 free_cpumask_var(mask); 4035 4036 return ret; 4037} 4038 4039/** 4040 * sys_sched_yield - yield the current processor to other threads. 4041 * 4042 * This function yields the current CPU to other tasks. If there are no 4043 * other threads running on this CPU then this function will return. 4044 * 4045 * Return: 0. 4046 */ 4047SYSCALL_DEFINE0(sched_yield) 4048{ 4049 struct rq *rq = this_rq_lock(); 4050 4051 schedstat_inc(rq, yld_count); 4052 current->sched_class->yield_task(rq); 4053 4054 /* 4055 * Since we are going to call schedule() anyway, there's 4056 * no need to preempt or enable interrupts: 4057 */ 4058 __release(rq->lock); 4059 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4060 do_raw_spin_unlock(&rq->lock); 4061 sched_preempt_enable_no_resched(); 4062 4063 schedule(); 4064 4065 return 0; 4066} 4067 4068static void __cond_resched(void) 4069{ 4070 __preempt_count_add(PREEMPT_ACTIVE); 4071 __schedule(); 4072 __preempt_count_sub(PREEMPT_ACTIVE); 4073} 4074 4075int __sched _cond_resched(void) 4076{ 4077 if (should_resched()) { 4078 __cond_resched(); 4079 return 1; 4080 } 4081 return 0; 4082} 4083EXPORT_SYMBOL(_cond_resched); 4084 4085/* 4086 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4087 * call schedule, and on return reacquire the lock. 4088 * 4089 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4090 * operations here to prevent schedule() from being called twice (once via 4091 * spin_unlock(), once by hand). 4092 */ 4093int __cond_resched_lock(spinlock_t *lock) 4094{ 4095 int resched = should_resched(); 4096 int ret = 0; 4097 4098 lockdep_assert_held(lock); 4099 4100 if (spin_needbreak(lock) || resched) { 4101 spin_unlock(lock); 4102 if (resched) 4103 __cond_resched(); 4104 else 4105 cpu_relax(); 4106 ret = 1; 4107 spin_lock(lock); 4108 } 4109 return ret; 4110} 4111EXPORT_SYMBOL(__cond_resched_lock); 4112 4113int __sched __cond_resched_softirq(void) 4114{ 4115 BUG_ON(!in_softirq()); 4116 4117 if (should_resched()) { 4118 local_bh_enable(); 4119 __cond_resched(); 4120 local_bh_disable(); 4121 return 1; 4122 } 4123 return 0; 4124} 4125EXPORT_SYMBOL(__cond_resched_softirq); 4126 4127/** 4128 * yield - yield the current processor to other threads. 4129 * 4130 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4131 * 4132 * The scheduler is at all times free to pick the calling task as the most 4133 * eligible task to run, if removing the yield() call from your code breaks 4134 * it, its already broken. 4135 * 4136 * Typical broken usage is: 4137 * 4138 * while (!event) 4139 * yield(); 4140 * 4141 * where one assumes that yield() will let 'the other' process run that will 4142 * make event true. If the current task is a SCHED_FIFO task that will never 4143 * happen. Never use yield() as a progress guarantee!! 4144 * 4145 * If you want to use yield() to wait for something, use wait_event(). 4146 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4147 * If you still want to use yield(), do not! 4148 */ 4149void __sched yield(void) 4150{ 4151 set_current_state(TASK_RUNNING); 4152 sys_sched_yield(); 4153} 4154EXPORT_SYMBOL(yield); 4155 4156/** 4157 * yield_to - yield the current processor to another thread in 4158 * your thread group, or accelerate that thread toward the 4159 * processor it's on. 4160 * @p: target task 4161 * @preempt: whether task preemption is allowed or not 4162 * 4163 * It's the caller's job to ensure that the target task struct 4164 * can't go away on us before we can do any checks. 4165 * 4166 * Return: 4167 * true (>0) if we indeed boosted the target task. 4168 * false (0) if we failed to boost the target. 4169 * -ESRCH if there's no task to yield to. 4170 */ 4171bool __sched yield_to(struct task_struct *p, bool preempt) 4172{ 4173 struct task_struct *curr = current; 4174 struct rq *rq, *p_rq; 4175 unsigned long flags; 4176 int yielded = 0; 4177 4178 local_irq_save(flags); 4179 rq = this_rq(); 4180 4181again: 4182 p_rq = task_rq(p); 4183 /* 4184 * If we're the only runnable task on the rq and target rq also 4185 * has only one task, there's absolutely no point in yielding. 4186 */ 4187 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4188 yielded = -ESRCH; 4189 goto out_irq; 4190 } 4191 4192 double_rq_lock(rq, p_rq); 4193 if (task_rq(p) != p_rq) { 4194 double_rq_unlock(rq, p_rq); 4195 goto again; 4196 } 4197 4198 if (!curr->sched_class->yield_to_task) 4199 goto out_unlock; 4200 4201 if (curr->sched_class != p->sched_class) 4202 goto out_unlock; 4203 4204 if (task_running(p_rq, p) || p->state) 4205 goto out_unlock; 4206 4207 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4208 if (yielded) { 4209 schedstat_inc(rq, yld_count); 4210 /* 4211 * Make p's CPU reschedule; pick_next_entity takes care of 4212 * fairness. 4213 */ 4214 if (preempt && rq != p_rq) 4215 resched_task(p_rq->curr); 4216 } 4217 4218out_unlock: 4219 double_rq_unlock(rq, p_rq); 4220out_irq: 4221 local_irq_restore(flags); 4222 4223 if (yielded > 0) 4224 schedule(); 4225 4226 return yielded; 4227} 4228EXPORT_SYMBOL_GPL(yield_to); 4229 4230/* 4231 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4232 * that process accounting knows that this is a task in IO wait state. 4233 */ 4234void __sched io_schedule(void) 4235{ 4236 struct rq *rq = raw_rq(); 4237 4238 delayacct_blkio_start(); 4239 atomic_inc(&rq->nr_iowait); 4240 blk_flush_plug(current); 4241 current->in_iowait = 1; 4242 schedule(); 4243 current->in_iowait = 0; 4244 atomic_dec(&rq->nr_iowait); 4245 delayacct_blkio_end(); 4246} 4247EXPORT_SYMBOL(io_schedule); 4248 4249long __sched io_schedule_timeout(long timeout) 4250{ 4251 struct rq *rq = raw_rq(); 4252 long ret; 4253 4254 delayacct_blkio_start(); 4255 atomic_inc(&rq->nr_iowait); 4256 blk_flush_plug(current); 4257 current->in_iowait = 1; 4258 ret = schedule_timeout(timeout); 4259 current->in_iowait = 0; 4260 atomic_dec(&rq->nr_iowait); 4261 delayacct_blkio_end(); 4262 return ret; 4263} 4264 4265/** 4266 * sys_sched_get_priority_max - return maximum RT priority. 4267 * @policy: scheduling class. 4268 * 4269 * Return: On success, this syscall returns the maximum 4270 * rt_priority that can be used by a given scheduling class. 4271 * On failure, a negative error code is returned. 4272 */ 4273SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4274{ 4275 int ret = -EINVAL; 4276 4277 switch (policy) { 4278 case SCHED_FIFO: 4279 case SCHED_RR: 4280 ret = MAX_USER_RT_PRIO-1; 4281 break; 4282 case SCHED_DEADLINE: 4283 case SCHED_NORMAL: 4284 case SCHED_BATCH: 4285 case SCHED_IDLE: 4286 ret = 0; 4287 break; 4288 } 4289 return ret; 4290} 4291 4292/** 4293 * sys_sched_get_priority_min - return minimum RT priority. 4294 * @policy: scheduling class. 4295 * 4296 * Return: On success, this syscall returns the minimum 4297 * rt_priority that can be used by a given scheduling class. 4298 * On failure, a negative error code is returned. 4299 */ 4300SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4301{ 4302 int ret = -EINVAL; 4303 4304 switch (policy) { 4305 case SCHED_FIFO: 4306 case SCHED_RR: 4307 ret = 1; 4308 break; 4309 case SCHED_DEADLINE: 4310 case SCHED_NORMAL: 4311 case SCHED_BATCH: 4312 case SCHED_IDLE: 4313 ret = 0; 4314 } 4315 return ret; 4316} 4317 4318/** 4319 * sys_sched_rr_get_interval - return the default timeslice of a process. 4320 * @pid: pid of the process. 4321 * @interval: userspace pointer to the timeslice value. 4322 * 4323 * this syscall writes the default timeslice value of a given process 4324 * into the user-space timespec buffer. A value of '0' means infinity. 4325 * 4326 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4327 * an error code. 4328 */ 4329SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4330 struct timespec __user *, interval) 4331{ 4332 struct task_struct *p; 4333 unsigned int time_slice; 4334 unsigned long flags; 4335 struct rq *rq; 4336 int retval; 4337 struct timespec t; 4338 4339 if (pid < 0) 4340 return -EINVAL; 4341 4342 retval = -ESRCH; 4343 rcu_read_lock(); 4344 p = find_process_by_pid(pid); 4345 if (!p) 4346 goto out_unlock; 4347 4348 retval = security_task_getscheduler(p); 4349 if (retval) 4350 goto out_unlock; 4351 4352 rq = task_rq_lock(p, &flags); 4353 time_slice = 0; 4354 if (p->sched_class->get_rr_interval) 4355 time_slice = p->sched_class->get_rr_interval(rq, p); 4356 task_rq_unlock(rq, p, &flags); 4357 4358 rcu_read_unlock(); 4359 jiffies_to_timespec(time_slice, &t); 4360 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4361 return retval; 4362 4363out_unlock: 4364 rcu_read_unlock(); 4365 return retval; 4366} 4367 4368static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4369 4370void sched_show_task(struct task_struct *p) 4371{ 4372 unsigned long free = 0; 4373 int ppid; 4374 unsigned state; 4375 4376 state = p->state ? __ffs(p->state) + 1 : 0; 4377 printk(KERN_INFO "%-15.15s %c", p->comm, 4378 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4379#if BITS_PER_LONG == 32 4380 if (state == TASK_RUNNING) 4381 printk(KERN_CONT " running "); 4382 else 4383 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4384#else 4385 if (state == TASK_RUNNING) 4386 printk(KERN_CONT " running task "); 4387 else 4388 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4389#endif 4390#ifdef CONFIG_DEBUG_STACK_USAGE 4391 free = stack_not_used(p); 4392#endif 4393 rcu_read_lock(); 4394 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4395 rcu_read_unlock(); 4396 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4397 task_pid_nr(p), ppid, 4398 (unsigned long)task_thread_info(p)->flags); 4399 4400 print_worker_info(KERN_INFO, p); 4401 show_stack(p, NULL); 4402} 4403 4404void show_state_filter(unsigned long state_filter) 4405{ 4406 struct task_struct *g, *p; 4407 4408#if BITS_PER_LONG == 32 4409 printk(KERN_INFO 4410 " task PC stack pid father\n"); 4411#else 4412 printk(KERN_INFO 4413 " task PC stack pid father\n"); 4414#endif 4415 rcu_read_lock(); 4416 do_each_thread(g, p) { 4417 /* 4418 * reset the NMI-timeout, listing all files on a slow 4419 * console might take a lot of time: 4420 */ 4421 touch_nmi_watchdog(); 4422 if (!state_filter || (p->state & state_filter)) 4423 sched_show_task(p); 4424 } while_each_thread(g, p); 4425 4426 touch_all_softlockup_watchdogs(); 4427 4428#ifdef CONFIG_SCHED_DEBUG 4429 sysrq_sched_debug_show(); 4430#endif 4431 rcu_read_unlock(); 4432 /* 4433 * Only show locks if all tasks are dumped: 4434 */ 4435 if (!state_filter) 4436 debug_show_all_locks(); 4437} 4438 4439void init_idle_bootup_task(struct task_struct *idle) 4440{ 4441 idle->sched_class = &idle_sched_class; 4442} 4443 4444/** 4445 * init_idle - set up an idle thread for a given CPU 4446 * @idle: task in question 4447 * @cpu: cpu the idle task belongs to 4448 * 4449 * NOTE: this function does not set the idle thread's NEED_RESCHED 4450 * flag, to make booting more robust. 4451 */ 4452void init_idle(struct task_struct *idle, int cpu) 4453{ 4454 struct rq *rq = cpu_rq(cpu); 4455 unsigned long flags; 4456 4457 raw_spin_lock_irqsave(&rq->lock, flags); 4458 4459 __sched_fork(0, idle); 4460 idle->state = TASK_RUNNING; 4461 idle->se.exec_start = sched_clock(); 4462 4463 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4464 /* 4465 * We're having a chicken and egg problem, even though we are 4466 * holding rq->lock, the cpu isn't yet set to this cpu so the 4467 * lockdep check in task_group() will fail. 4468 * 4469 * Similar case to sched_fork(). / Alternatively we could 4470 * use task_rq_lock() here and obtain the other rq->lock. 4471 * 4472 * Silence PROVE_RCU 4473 */ 4474 rcu_read_lock(); 4475 __set_task_cpu(idle, cpu); 4476 rcu_read_unlock(); 4477 4478 rq->curr = rq->idle = idle; 4479#if defined(CONFIG_SMP) 4480 idle->on_cpu = 1; 4481#endif 4482 raw_spin_unlock_irqrestore(&rq->lock, flags); 4483 4484 /* Set the preempt count _outside_ the spinlocks! */ 4485 init_idle_preempt_count(idle, cpu); 4486 4487 /* 4488 * The idle tasks have their own, simple scheduling class: 4489 */ 4490 idle->sched_class = &idle_sched_class; 4491 ftrace_graph_init_idle_task(idle, cpu); 4492 vtime_init_idle(idle, cpu); 4493#if defined(CONFIG_SMP) 4494 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4495#endif 4496} 4497 4498#ifdef CONFIG_SMP 4499void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4500{ 4501 if (p->sched_class && p->sched_class->set_cpus_allowed) 4502 p->sched_class->set_cpus_allowed(p, new_mask); 4503 4504 cpumask_copy(&p->cpus_allowed, new_mask); 4505 p->nr_cpus_allowed = cpumask_weight(new_mask); 4506} 4507 4508/* 4509 * This is how migration works: 4510 * 4511 * 1) we invoke migration_cpu_stop() on the target CPU using 4512 * stop_one_cpu(). 4513 * 2) stopper starts to run (implicitly forcing the migrated thread 4514 * off the CPU) 4515 * 3) it checks whether the migrated task is still in the wrong runqueue. 4516 * 4) if it's in the wrong runqueue then the migration thread removes 4517 * it and puts it into the right queue. 4518 * 5) stopper completes and stop_one_cpu() returns and the migration 4519 * is done. 4520 */ 4521 4522/* 4523 * Change a given task's CPU affinity. Migrate the thread to a 4524 * proper CPU and schedule it away if the CPU it's executing on 4525 * is removed from the allowed bitmask. 4526 * 4527 * NOTE: the caller must have a valid reference to the task, the 4528 * task must not exit() & deallocate itself prematurely. The 4529 * call is not atomic; no spinlocks may be held. 4530 */ 4531int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4532{ 4533 unsigned long flags; 4534 struct rq *rq; 4535 unsigned int dest_cpu; 4536 int ret = 0; 4537 4538 rq = task_rq_lock(p, &flags); 4539 4540 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4541 goto out; 4542 4543 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4544 ret = -EINVAL; 4545 goto out; 4546 } 4547 4548 do_set_cpus_allowed(p, new_mask); 4549 4550 /* Can the task run on the task's current CPU? If so, we're done */ 4551 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4552 goto out; 4553 4554 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4555 if (p->on_rq) { 4556 struct migration_arg arg = { p, dest_cpu }; 4557 /* Need help from migration thread: drop lock and wait. */ 4558 task_rq_unlock(rq, p, &flags); 4559 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4560 tlb_migrate_finish(p->mm); 4561 return 0; 4562 } 4563out: 4564 task_rq_unlock(rq, p, &flags); 4565 4566 return ret; 4567} 4568EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4569 4570/* 4571 * Move (not current) task off this cpu, onto dest cpu. We're doing 4572 * this because either it can't run here any more (set_cpus_allowed() 4573 * away from this CPU, or CPU going down), or because we're 4574 * attempting to rebalance this task on exec (sched_exec). 4575 * 4576 * So we race with normal scheduler movements, but that's OK, as long 4577 * as the task is no longer on this CPU. 4578 * 4579 * Returns non-zero if task was successfully migrated. 4580 */ 4581static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4582{ 4583 struct rq *rq_dest, *rq_src; 4584 int ret = 0; 4585 4586 if (unlikely(!cpu_active(dest_cpu))) 4587 return ret; 4588 4589 rq_src = cpu_rq(src_cpu); 4590 rq_dest = cpu_rq(dest_cpu); 4591 4592 raw_spin_lock(&p->pi_lock); 4593 double_rq_lock(rq_src, rq_dest); 4594 /* Already moved. */ 4595 if (task_cpu(p) != src_cpu) 4596 goto done; 4597 /* Affinity changed (again). */ 4598 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4599 goto fail; 4600 4601 /* 4602 * If we're not on a rq, the next wake-up will ensure we're 4603 * placed properly. 4604 */ 4605 if (p->on_rq) { 4606 dequeue_task(rq_src, p, 0); 4607 set_task_cpu(p, dest_cpu); 4608 enqueue_task(rq_dest, p, 0); 4609 check_preempt_curr(rq_dest, p, 0); 4610 } 4611done: 4612 ret = 1; 4613fail: 4614 double_rq_unlock(rq_src, rq_dest); 4615 raw_spin_unlock(&p->pi_lock); 4616 return ret; 4617} 4618 4619#ifdef CONFIG_NUMA_BALANCING 4620/* Migrate current task p to target_cpu */ 4621int migrate_task_to(struct task_struct *p, int target_cpu) 4622{ 4623 struct migration_arg arg = { p, target_cpu }; 4624 int curr_cpu = task_cpu(p); 4625 4626 if (curr_cpu == target_cpu) 4627 return 0; 4628 4629 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4630 return -EINVAL; 4631 4632 /* TODO: This is not properly updating schedstats */ 4633 4634 trace_sched_move_numa(p, curr_cpu, target_cpu); 4635 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4636} 4637 4638/* 4639 * Requeue a task on a given node and accurately track the number of NUMA 4640 * tasks on the runqueues 4641 */ 4642void sched_setnuma(struct task_struct *p, int nid) 4643{ 4644 struct rq *rq; 4645 unsigned long flags; 4646 bool on_rq, running; 4647 4648 rq = task_rq_lock(p, &flags); 4649 on_rq = p->on_rq; 4650 running = task_current(rq, p); 4651 4652 if (on_rq) 4653 dequeue_task(rq, p, 0); 4654 if (running) 4655 p->sched_class->put_prev_task(rq, p); 4656 4657 p->numa_preferred_nid = nid; 4658 4659 if (running) 4660 p->sched_class->set_curr_task(rq); 4661 if (on_rq) 4662 enqueue_task(rq, p, 0); 4663 task_rq_unlock(rq, p, &flags); 4664} 4665#endif 4666 4667/* 4668 * migration_cpu_stop - this will be executed by a highprio stopper thread 4669 * and performs thread migration by bumping thread off CPU then 4670 * 'pushing' onto another runqueue. 4671 */ 4672static int migration_cpu_stop(void *data) 4673{ 4674 struct migration_arg *arg = data; 4675 4676 /* 4677 * The original target cpu might have gone down and we might 4678 * be on another cpu but it doesn't matter. 4679 */ 4680 local_irq_disable(); 4681 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4682 local_irq_enable(); 4683 return 0; 4684} 4685 4686#ifdef CONFIG_HOTPLUG_CPU 4687 4688/* 4689 * Ensures that the idle task is using init_mm right before its cpu goes 4690 * offline. 4691 */ 4692void idle_task_exit(void) 4693{ 4694 struct mm_struct *mm = current->active_mm; 4695 4696 BUG_ON(cpu_online(smp_processor_id())); 4697 4698 if (mm != &init_mm) 4699 switch_mm(mm, &init_mm, current); 4700 mmdrop(mm); 4701} 4702 4703/* 4704 * Since this CPU is going 'away' for a while, fold any nr_active delta 4705 * we might have. Assumes we're called after migrate_tasks() so that the 4706 * nr_active count is stable. 4707 * 4708 * Also see the comment "Global load-average calculations". 4709 */ 4710static void calc_load_migrate(struct rq *rq) 4711{ 4712 long delta = calc_load_fold_active(rq); 4713 if (delta) 4714 atomic_long_add(delta, &calc_load_tasks); 4715} 4716 4717/* 4718 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4719 * try_to_wake_up()->select_task_rq(). 4720 * 4721 * Called with rq->lock held even though we'er in stop_machine() and 4722 * there's no concurrency possible, we hold the required locks anyway 4723 * because of lock validation efforts. 4724 */ 4725static void migrate_tasks(unsigned int dead_cpu) 4726{ 4727 struct rq *rq = cpu_rq(dead_cpu); 4728 struct task_struct *next, *stop = rq->stop; 4729 int dest_cpu; 4730 4731 /* 4732 * Fudge the rq selection such that the below task selection loop 4733 * doesn't get stuck on the currently eligible stop task. 4734 * 4735 * We're currently inside stop_machine() and the rq is either stuck 4736 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4737 * either way we should never end up calling schedule() until we're 4738 * done here. 4739 */ 4740 rq->stop = NULL; 4741 4742 /* 4743 * put_prev_task() and pick_next_task() sched 4744 * class method both need to have an up-to-date 4745 * value of rq->clock[_task] 4746 */ 4747 update_rq_clock(rq); 4748 4749 for ( ; ; ) { 4750 /* 4751 * There's this thread running, bail when that's the only 4752 * remaining thread. 4753 */ 4754 if (rq->nr_running == 1) 4755 break; 4756 4757 next = pick_next_task(rq); 4758 BUG_ON(!next); 4759 next->sched_class->put_prev_task(rq, next); 4760 4761 /* Find suitable destination for @next, with force if needed. */ 4762 dest_cpu = select_fallback_rq(dead_cpu, next); 4763 raw_spin_unlock(&rq->lock); 4764 4765 __migrate_task(next, dead_cpu, dest_cpu); 4766 4767 raw_spin_lock(&rq->lock); 4768 } 4769 4770 rq->stop = stop; 4771} 4772 4773#endif /* CONFIG_HOTPLUG_CPU */ 4774 4775#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4776 4777static struct ctl_table sd_ctl_dir[] = { 4778 { 4779 .procname = "sched_domain", 4780 .mode = 0555, 4781 }, 4782 {} 4783}; 4784 4785static struct ctl_table sd_ctl_root[] = { 4786 { 4787 .procname = "kernel", 4788 .mode = 0555, 4789 .child = sd_ctl_dir, 4790 }, 4791 {} 4792}; 4793 4794static struct ctl_table *sd_alloc_ctl_entry(int n) 4795{ 4796 struct ctl_table *entry = 4797 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4798 4799 return entry; 4800} 4801 4802static void sd_free_ctl_entry(struct ctl_table **tablep) 4803{ 4804 struct ctl_table *entry; 4805 4806 /* 4807 * In the intermediate directories, both the child directory and 4808 * procname are dynamically allocated and could fail but the mode 4809 * will always be set. In the lowest directory the names are 4810 * static strings and all have proc handlers. 4811 */ 4812 for (entry = *tablep; entry->mode; entry++) { 4813 if (entry->child) 4814 sd_free_ctl_entry(&entry->child); 4815 if (entry->proc_handler == NULL) 4816 kfree(entry->procname); 4817 } 4818 4819 kfree(*tablep); 4820 *tablep = NULL; 4821} 4822 4823static int min_load_idx = 0; 4824static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4825 4826static void 4827set_table_entry(struct ctl_table *entry, 4828 const char *procname, void *data, int maxlen, 4829 umode_t mode, proc_handler *proc_handler, 4830 bool load_idx) 4831{ 4832 entry->procname = procname; 4833 entry->data = data; 4834 entry->maxlen = maxlen; 4835 entry->mode = mode; 4836 entry->proc_handler = proc_handler; 4837 4838 if (load_idx) { 4839 entry->extra1 = &min_load_idx; 4840 entry->extra2 = &max_load_idx; 4841 } 4842} 4843 4844static struct ctl_table * 4845sd_alloc_ctl_domain_table(struct sched_domain *sd) 4846{ 4847 struct ctl_table *table = sd_alloc_ctl_entry(13); 4848 4849 if (table == NULL) 4850 return NULL; 4851 4852 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4853 sizeof(long), 0644, proc_doulongvec_minmax, false); 4854 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4855 sizeof(long), 0644, proc_doulongvec_minmax, false); 4856 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4857 sizeof(int), 0644, proc_dointvec_minmax, true); 4858 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4859 sizeof(int), 0644, proc_dointvec_minmax, true); 4860 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4861 sizeof(int), 0644, proc_dointvec_minmax, true); 4862 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4863 sizeof(int), 0644, proc_dointvec_minmax, true); 4864 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4865 sizeof(int), 0644, proc_dointvec_minmax, true); 4866 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4867 sizeof(int), 0644, proc_dointvec_minmax, false); 4868 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4869 sizeof(int), 0644, proc_dointvec_minmax, false); 4870 set_table_entry(&table[9], "cache_nice_tries", 4871 &sd->cache_nice_tries, 4872 sizeof(int), 0644, proc_dointvec_minmax, false); 4873 set_table_entry(&table[10], "flags", &sd->flags, 4874 sizeof(int), 0644, proc_dointvec_minmax, false); 4875 set_table_entry(&table[11], "name", sd->name, 4876 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4877 /* &table[12] is terminator */ 4878 4879 return table; 4880} 4881 4882static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4883{ 4884 struct ctl_table *entry, *table; 4885 struct sched_domain *sd; 4886 int domain_num = 0, i; 4887 char buf[32]; 4888 4889 for_each_domain(cpu, sd) 4890 domain_num++; 4891 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4892 if (table == NULL) 4893 return NULL; 4894 4895 i = 0; 4896 for_each_domain(cpu, sd) { 4897 snprintf(buf, 32, "domain%d", i); 4898 entry->procname = kstrdup(buf, GFP_KERNEL); 4899 entry->mode = 0555; 4900 entry->child = sd_alloc_ctl_domain_table(sd); 4901 entry++; 4902 i++; 4903 } 4904 return table; 4905} 4906 4907static struct ctl_table_header *sd_sysctl_header; 4908static void register_sched_domain_sysctl(void) 4909{ 4910 int i, cpu_num = num_possible_cpus(); 4911 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4912 char buf[32]; 4913 4914 WARN_ON(sd_ctl_dir[0].child); 4915 sd_ctl_dir[0].child = entry; 4916 4917 if (entry == NULL) 4918 return; 4919 4920 for_each_possible_cpu(i) { 4921 snprintf(buf, 32, "cpu%d", i); 4922 entry->procname = kstrdup(buf, GFP_KERNEL); 4923 entry->mode = 0555; 4924 entry->child = sd_alloc_ctl_cpu_table(i); 4925 entry++; 4926 } 4927 4928 WARN_ON(sd_sysctl_header); 4929 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4930} 4931 4932/* may be called multiple times per register */ 4933static void unregister_sched_domain_sysctl(void) 4934{ 4935 if (sd_sysctl_header) 4936 unregister_sysctl_table(sd_sysctl_header); 4937 sd_sysctl_header = NULL; 4938 if (sd_ctl_dir[0].child) 4939 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4940} 4941#else 4942static void register_sched_domain_sysctl(void) 4943{ 4944} 4945static void unregister_sched_domain_sysctl(void) 4946{ 4947} 4948#endif 4949 4950static void set_rq_online(struct rq *rq) 4951{ 4952 if (!rq->online) { 4953 const struct sched_class *class; 4954 4955 cpumask_set_cpu(rq->cpu, rq->rd->online); 4956 rq->online = 1; 4957 4958 for_each_class(class) { 4959 if (class->rq_online) 4960 class->rq_online(rq); 4961 } 4962 } 4963} 4964 4965static void set_rq_offline(struct rq *rq) 4966{ 4967 if (rq->online) { 4968 const struct sched_class *class; 4969 4970 for_each_class(class) { 4971 if (class->rq_offline) 4972 class->rq_offline(rq); 4973 } 4974 4975 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4976 rq->online = 0; 4977 } 4978} 4979 4980/* 4981 * migration_call - callback that gets triggered when a CPU is added. 4982 * Here we can start up the necessary migration thread for the new CPU. 4983 */ 4984static int 4985migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4986{ 4987 int cpu = (long)hcpu; 4988 unsigned long flags; 4989 struct rq *rq = cpu_rq(cpu); 4990 4991 switch (action & ~CPU_TASKS_FROZEN) { 4992 4993 case CPU_UP_PREPARE: 4994 rq->calc_load_update = calc_load_update; 4995 break; 4996 4997 case CPU_ONLINE: 4998 /* Update our root-domain */ 4999 raw_spin_lock_irqsave(&rq->lock, flags); 5000 if (rq->rd) { 5001 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5002 5003 set_rq_online(rq); 5004 } 5005 raw_spin_unlock_irqrestore(&rq->lock, flags); 5006 break; 5007 5008#ifdef CONFIG_HOTPLUG_CPU 5009 case CPU_DYING: 5010 sched_ttwu_pending(); 5011 /* Update our root-domain */ 5012 raw_spin_lock_irqsave(&rq->lock, flags); 5013 if (rq->rd) { 5014 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5015 set_rq_offline(rq); 5016 } 5017 migrate_tasks(cpu); 5018 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5019 raw_spin_unlock_irqrestore(&rq->lock, flags); 5020 break; 5021 5022 case CPU_DEAD: 5023 calc_load_migrate(rq); 5024 break; 5025#endif 5026 } 5027 5028 update_max_interval(); 5029 5030 return NOTIFY_OK; 5031} 5032 5033/* 5034 * Register at high priority so that task migration (migrate_all_tasks) 5035 * happens before everything else. This has to be lower priority than 5036 * the notifier in the perf_event subsystem, though. 5037 */ 5038static struct notifier_block migration_notifier = { 5039 .notifier_call = migration_call, 5040 .priority = CPU_PRI_MIGRATION, 5041}; 5042 5043static int sched_cpu_active(struct notifier_block *nfb, 5044 unsigned long action, void *hcpu) 5045{ 5046 switch (action & ~CPU_TASKS_FROZEN) { 5047 case CPU_STARTING: 5048 case CPU_DOWN_FAILED: 5049 set_cpu_active((long)hcpu, true); 5050 return NOTIFY_OK; 5051 default: 5052 return NOTIFY_DONE; 5053 } 5054} 5055 5056static int sched_cpu_inactive(struct notifier_block *nfb, 5057 unsigned long action, void *hcpu) 5058{ 5059 unsigned long flags; 5060 long cpu = (long)hcpu; 5061 5062 switch (action & ~CPU_TASKS_FROZEN) { 5063 case CPU_DOWN_PREPARE: 5064 set_cpu_active(cpu, false); 5065 5066 /* explicitly allow suspend */ 5067 if (!(action & CPU_TASKS_FROZEN)) { 5068 struct dl_bw *dl_b = dl_bw_of(cpu); 5069 bool overflow; 5070 int cpus; 5071 5072 raw_spin_lock_irqsave(&dl_b->lock, flags); 5073 cpus = dl_bw_cpus(cpu); 5074 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5075 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5076 5077 if (overflow) 5078 return notifier_from_errno(-EBUSY); 5079 } 5080 return NOTIFY_OK; 5081 } 5082 5083 return NOTIFY_DONE; 5084} 5085 5086static int __init migration_init(void) 5087{ 5088 void *cpu = (void *)(long)smp_processor_id(); 5089 int err; 5090 5091 /* Initialize migration for the boot CPU */ 5092 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5093 BUG_ON(err == NOTIFY_BAD); 5094 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5095 register_cpu_notifier(&migration_notifier); 5096 5097 /* Register cpu active notifiers */ 5098 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5099 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5100 5101 return 0; 5102} 5103early_initcall(migration_init); 5104#endif 5105 5106#ifdef CONFIG_SMP 5107 5108static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5109 5110#ifdef CONFIG_SCHED_DEBUG 5111 5112static __read_mostly int sched_debug_enabled; 5113 5114static int __init sched_debug_setup(char *str) 5115{ 5116 sched_debug_enabled = 1; 5117 5118 return 0; 5119} 5120early_param("sched_debug", sched_debug_setup); 5121 5122static inline bool sched_debug(void) 5123{ 5124 return sched_debug_enabled; 5125} 5126 5127static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5128 struct cpumask *groupmask) 5129{ 5130 struct sched_group *group = sd->groups; 5131 char str[256]; 5132 5133 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5134 cpumask_clear(groupmask); 5135 5136 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5137 5138 if (!(sd->flags & SD_LOAD_BALANCE)) { 5139 printk("does not load-balance\n"); 5140 if (sd->parent) 5141 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5142 " has parent"); 5143 return -1; 5144 } 5145 5146 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5147 5148 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5149 printk(KERN_ERR "ERROR: domain->span does not contain " 5150 "CPU%d\n", cpu); 5151 } 5152 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5153 printk(KERN_ERR "ERROR: domain->groups does not contain" 5154 " CPU%d\n", cpu); 5155 } 5156 5157 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5158 do { 5159 if (!group) { 5160 printk("\n"); 5161 printk(KERN_ERR "ERROR: group is NULL\n"); 5162 break; 5163 } 5164 5165 /* 5166 * Even though we initialize ->power to something semi-sane, 5167 * we leave power_orig unset. This allows us to detect if 5168 * domain iteration is still funny without causing /0 traps. 5169 */ 5170 if (!group->sgp->power_orig) { 5171 printk(KERN_CONT "\n"); 5172 printk(KERN_ERR "ERROR: domain->cpu_power not " 5173 "set\n"); 5174 break; 5175 } 5176 5177 if (!cpumask_weight(sched_group_cpus(group))) { 5178 printk(KERN_CONT "\n"); 5179 printk(KERN_ERR "ERROR: empty group\n"); 5180 break; 5181 } 5182 5183 if (!(sd->flags & SD_OVERLAP) && 5184 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5185 printk(KERN_CONT "\n"); 5186 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5187 break; 5188 } 5189 5190 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5191 5192 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5193 5194 printk(KERN_CONT " %s", str); 5195 if (group->sgp->power != SCHED_POWER_SCALE) { 5196 printk(KERN_CONT " (cpu_power = %d)", 5197 group->sgp->power); 5198 } 5199 5200 group = group->next; 5201 } while (group != sd->groups); 5202 printk(KERN_CONT "\n"); 5203 5204 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5205 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5206 5207 if (sd->parent && 5208 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5209 printk(KERN_ERR "ERROR: parent span is not a superset " 5210 "of domain->span\n"); 5211 return 0; 5212} 5213 5214static void sched_domain_debug(struct sched_domain *sd, int cpu) 5215{ 5216 int level = 0; 5217 5218 if (!sched_debug_enabled) 5219 return; 5220 5221 if (!sd) { 5222 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5223 return; 5224 } 5225 5226 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5227 5228 for (;;) { 5229 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5230 break; 5231 level++; 5232 sd = sd->parent; 5233 if (!sd) 5234 break; 5235 } 5236} 5237#else /* !CONFIG_SCHED_DEBUG */ 5238# define sched_domain_debug(sd, cpu) do { } while (0) 5239static inline bool sched_debug(void) 5240{ 5241 return false; 5242} 5243#endif /* CONFIG_SCHED_DEBUG */ 5244 5245static int sd_degenerate(struct sched_domain *sd) 5246{ 5247 if (cpumask_weight(sched_domain_span(sd)) == 1) 5248 return 1; 5249 5250 /* Following flags need at least 2 groups */ 5251 if (sd->flags & (SD_LOAD_BALANCE | 5252 SD_BALANCE_NEWIDLE | 5253 SD_BALANCE_FORK | 5254 SD_BALANCE_EXEC | 5255 SD_SHARE_CPUPOWER | 5256 SD_SHARE_PKG_RESOURCES)) { 5257 if (sd->groups != sd->groups->next) 5258 return 0; 5259 } 5260 5261 /* Following flags don't use groups */ 5262 if (sd->flags & (SD_WAKE_AFFINE)) 5263 return 0; 5264 5265 return 1; 5266} 5267 5268static int 5269sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5270{ 5271 unsigned long cflags = sd->flags, pflags = parent->flags; 5272 5273 if (sd_degenerate(parent)) 5274 return 1; 5275 5276 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5277 return 0; 5278 5279 /* Flags needing groups don't count if only 1 group in parent */ 5280 if (parent->groups == parent->groups->next) { 5281 pflags &= ~(SD_LOAD_BALANCE | 5282 SD_BALANCE_NEWIDLE | 5283 SD_BALANCE_FORK | 5284 SD_BALANCE_EXEC | 5285 SD_SHARE_CPUPOWER | 5286 SD_SHARE_PKG_RESOURCES | 5287 SD_PREFER_SIBLING); 5288 if (nr_node_ids == 1) 5289 pflags &= ~SD_SERIALIZE; 5290 } 5291 if (~cflags & pflags) 5292 return 0; 5293 5294 return 1; 5295} 5296 5297static void free_rootdomain(struct rcu_head *rcu) 5298{ 5299 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5300 5301 cpupri_cleanup(&rd->cpupri); 5302 cpudl_cleanup(&rd->cpudl); 5303 free_cpumask_var(rd->dlo_mask); 5304 free_cpumask_var(rd->rto_mask); 5305 free_cpumask_var(rd->online); 5306 free_cpumask_var(rd->span); 5307 kfree(rd); 5308} 5309 5310static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5311{ 5312 struct root_domain *old_rd = NULL; 5313 unsigned long flags; 5314 5315 raw_spin_lock_irqsave(&rq->lock, flags); 5316 5317 if (rq->rd) { 5318 old_rd = rq->rd; 5319 5320 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5321 set_rq_offline(rq); 5322 5323 cpumask_clear_cpu(rq->cpu, old_rd->span); 5324 5325 /* 5326 * If we dont want to free the old_rd yet then 5327 * set old_rd to NULL to skip the freeing later 5328 * in this function: 5329 */ 5330 if (!atomic_dec_and_test(&old_rd->refcount)) 5331 old_rd = NULL; 5332 } 5333 5334 atomic_inc(&rd->refcount); 5335 rq->rd = rd; 5336 5337 cpumask_set_cpu(rq->cpu, rd->span); 5338 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5339 set_rq_online(rq); 5340 5341 raw_spin_unlock_irqrestore(&rq->lock, flags); 5342 5343 if (old_rd) 5344 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5345} 5346 5347static int init_rootdomain(struct root_domain *rd) 5348{ 5349 memset(rd, 0, sizeof(*rd)); 5350 5351 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5352 goto out; 5353 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5354 goto free_span; 5355 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5356 goto free_online; 5357 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5358 goto free_dlo_mask; 5359 5360 init_dl_bw(&rd->dl_bw); 5361 if (cpudl_init(&rd->cpudl) != 0) 5362 goto free_dlo_mask; 5363 5364 if (cpupri_init(&rd->cpupri) != 0) 5365 goto free_rto_mask; 5366 return 0; 5367 5368free_rto_mask: 5369 free_cpumask_var(rd->rto_mask); 5370free_dlo_mask: 5371 free_cpumask_var(rd->dlo_mask); 5372free_online: 5373 free_cpumask_var(rd->online); 5374free_span: 5375 free_cpumask_var(rd->span); 5376out: 5377 return -ENOMEM; 5378} 5379 5380/* 5381 * By default the system creates a single root-domain with all cpus as 5382 * members (mimicking the global state we have today). 5383 */ 5384struct root_domain def_root_domain; 5385 5386static void init_defrootdomain(void) 5387{ 5388 init_rootdomain(&def_root_domain); 5389 5390 atomic_set(&def_root_domain.refcount, 1); 5391} 5392 5393static struct root_domain *alloc_rootdomain(void) 5394{ 5395 struct root_domain *rd; 5396 5397 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5398 if (!rd) 5399 return NULL; 5400 5401 if (init_rootdomain(rd) != 0) { 5402 kfree(rd); 5403 return NULL; 5404 } 5405 5406 return rd; 5407} 5408 5409static void free_sched_groups(struct sched_group *sg, int free_sgp) 5410{ 5411 struct sched_group *tmp, *first; 5412 5413 if (!sg) 5414 return; 5415 5416 first = sg; 5417 do { 5418 tmp = sg->next; 5419 5420 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5421 kfree(sg->sgp); 5422 5423 kfree(sg); 5424 sg = tmp; 5425 } while (sg != first); 5426} 5427 5428static void free_sched_domain(struct rcu_head *rcu) 5429{ 5430 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5431 5432 /* 5433 * If its an overlapping domain it has private groups, iterate and 5434 * nuke them all. 5435 */ 5436 if (sd->flags & SD_OVERLAP) { 5437 free_sched_groups(sd->groups, 1); 5438 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5439 kfree(sd->groups->sgp); 5440 kfree(sd->groups); 5441 } 5442 kfree(sd); 5443} 5444 5445static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5446{ 5447 call_rcu(&sd->rcu, free_sched_domain); 5448} 5449 5450static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5451{ 5452 for (; sd; sd = sd->parent) 5453 destroy_sched_domain(sd, cpu); 5454} 5455 5456/* 5457 * Keep a special pointer to the highest sched_domain that has 5458 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5459 * allows us to avoid some pointer chasing select_idle_sibling(). 5460 * 5461 * Also keep a unique ID per domain (we use the first cpu number in 5462 * the cpumask of the domain), this allows us to quickly tell if 5463 * two cpus are in the same cache domain, see cpus_share_cache(). 5464 */ 5465DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5466DEFINE_PER_CPU(int, sd_llc_size); 5467DEFINE_PER_CPU(int, sd_llc_id); 5468DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5469DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5470DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5471 5472static void update_top_cache_domain(int cpu) 5473{ 5474 struct sched_domain *sd; 5475 struct sched_domain *busy_sd = NULL; 5476 int id = cpu; 5477 int size = 1; 5478 5479 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5480 if (sd) { 5481 id = cpumask_first(sched_domain_span(sd)); 5482 size = cpumask_weight(sched_domain_span(sd)); 5483 busy_sd = sd->parent; /* sd_busy */ 5484 } 5485 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5486 5487 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5488 per_cpu(sd_llc_size, cpu) = size; 5489 per_cpu(sd_llc_id, cpu) = id; 5490 5491 sd = lowest_flag_domain(cpu, SD_NUMA); 5492 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5493 5494 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5495 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5496} 5497 5498/* 5499 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5500 * hold the hotplug lock. 5501 */ 5502static void 5503cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5504{ 5505 struct rq *rq = cpu_rq(cpu); 5506 struct sched_domain *tmp; 5507 5508 /* Remove the sched domains which do not contribute to scheduling. */ 5509 for (tmp = sd; tmp; ) { 5510 struct sched_domain *parent = tmp->parent; 5511 if (!parent) 5512 break; 5513 5514 if (sd_parent_degenerate(tmp, parent)) { 5515 tmp->parent = parent->parent; 5516 if (parent->parent) 5517 parent->parent->child = tmp; 5518 /* 5519 * Transfer SD_PREFER_SIBLING down in case of a 5520 * degenerate parent; the spans match for this 5521 * so the property transfers. 5522 */ 5523 if (parent->flags & SD_PREFER_SIBLING) 5524 tmp->flags |= SD_PREFER_SIBLING; 5525 destroy_sched_domain(parent, cpu); 5526 } else 5527 tmp = tmp->parent; 5528 } 5529 5530 if (sd && sd_degenerate(sd)) { 5531 tmp = sd; 5532 sd = sd->parent; 5533 destroy_sched_domain(tmp, cpu); 5534 if (sd) 5535 sd->child = NULL; 5536 } 5537 5538 sched_domain_debug(sd, cpu); 5539 5540 rq_attach_root(rq, rd); 5541 tmp = rq->sd; 5542 rcu_assign_pointer(rq->sd, sd); 5543 destroy_sched_domains(tmp, cpu); 5544 5545 update_top_cache_domain(cpu); 5546} 5547 5548/* cpus with isolated domains */ 5549static cpumask_var_t cpu_isolated_map; 5550 5551/* Setup the mask of cpus configured for isolated domains */ 5552static int __init isolated_cpu_setup(char *str) 5553{ 5554 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5555 cpulist_parse(str, cpu_isolated_map); 5556 return 1; 5557} 5558 5559__setup("isolcpus=", isolated_cpu_setup); 5560 5561static const struct cpumask *cpu_cpu_mask(int cpu) 5562{ 5563 return cpumask_of_node(cpu_to_node(cpu)); 5564} 5565 5566struct sd_data { 5567 struct sched_domain **__percpu sd; 5568 struct sched_group **__percpu sg; 5569 struct sched_group_power **__percpu sgp; 5570}; 5571 5572struct s_data { 5573 struct sched_domain ** __percpu sd; 5574 struct root_domain *rd; 5575}; 5576 5577enum s_alloc { 5578 sa_rootdomain, 5579 sa_sd, 5580 sa_sd_storage, 5581 sa_none, 5582}; 5583 5584struct sched_domain_topology_level; 5585 5586typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5587typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5588 5589#define SDTL_OVERLAP 0x01 5590 5591struct sched_domain_topology_level { 5592 sched_domain_init_f init; 5593 sched_domain_mask_f mask; 5594 int flags; 5595 int numa_level; 5596 struct sd_data data; 5597}; 5598 5599/* 5600 * Build an iteration mask that can exclude certain CPUs from the upwards 5601 * domain traversal. 5602 * 5603 * Asymmetric node setups can result in situations where the domain tree is of 5604 * unequal depth, make sure to skip domains that already cover the entire 5605 * range. 5606 * 5607 * In that case build_sched_domains() will have terminated the iteration early 5608 * and our sibling sd spans will be empty. Domains should always include the 5609 * cpu they're built on, so check that. 5610 * 5611 */ 5612static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5613{ 5614 const struct cpumask *span = sched_domain_span(sd); 5615 struct sd_data *sdd = sd->private; 5616 struct sched_domain *sibling; 5617 int i; 5618 5619 for_each_cpu(i, span) { 5620 sibling = *per_cpu_ptr(sdd->sd, i); 5621 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5622 continue; 5623 5624 cpumask_set_cpu(i, sched_group_mask(sg)); 5625 } 5626} 5627 5628/* 5629 * Return the canonical balance cpu for this group, this is the first cpu 5630 * of this group that's also in the iteration mask. 5631 */ 5632int group_balance_cpu(struct sched_group *sg) 5633{ 5634 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5635} 5636 5637static int 5638build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5639{ 5640 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5641 const struct cpumask *span = sched_domain_span(sd); 5642 struct cpumask *covered = sched_domains_tmpmask; 5643 struct sd_data *sdd = sd->private; 5644 struct sched_domain *child; 5645 int i; 5646 5647 cpumask_clear(covered); 5648 5649 for_each_cpu(i, span) { 5650 struct cpumask *sg_span; 5651 5652 if (cpumask_test_cpu(i, covered)) 5653 continue; 5654 5655 child = *per_cpu_ptr(sdd->sd, i); 5656 5657 /* See the comment near build_group_mask(). */ 5658 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5659 continue; 5660 5661 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5662 GFP_KERNEL, cpu_to_node(cpu)); 5663 5664 if (!sg) 5665 goto fail; 5666 5667 sg_span = sched_group_cpus(sg); 5668 if (child->child) { 5669 child = child->child; 5670 cpumask_copy(sg_span, sched_domain_span(child)); 5671 } else 5672 cpumask_set_cpu(i, sg_span); 5673 5674 cpumask_or(covered, covered, sg_span); 5675 5676 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5677 if (atomic_inc_return(&sg->sgp->ref) == 1) 5678 build_group_mask(sd, sg); 5679 5680 /* 5681 * Initialize sgp->power such that even if we mess up the 5682 * domains and no possible iteration will get us here, we won't 5683 * die on a /0 trap. 5684 */ 5685 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5686 sg->sgp->power_orig = sg->sgp->power; 5687 5688 /* 5689 * Make sure the first group of this domain contains the 5690 * canonical balance cpu. Otherwise the sched_domain iteration 5691 * breaks. See update_sg_lb_stats(). 5692 */ 5693 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5694 group_balance_cpu(sg) == cpu) 5695 groups = sg; 5696 5697 if (!first) 5698 first = sg; 5699 if (last) 5700 last->next = sg; 5701 last = sg; 5702 last->next = first; 5703 } 5704 sd->groups = groups; 5705 5706 return 0; 5707 5708fail: 5709 free_sched_groups(first, 0); 5710 5711 return -ENOMEM; 5712} 5713 5714static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5715{ 5716 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5717 struct sched_domain *child = sd->child; 5718 5719 if (child) 5720 cpu = cpumask_first(sched_domain_span(child)); 5721 5722 if (sg) { 5723 *sg = *per_cpu_ptr(sdd->sg, cpu); 5724 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5725 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5726 } 5727 5728 return cpu; 5729} 5730 5731/* 5732 * build_sched_groups will build a circular linked list of the groups 5733 * covered by the given span, and will set each group's ->cpumask correctly, 5734 * and ->cpu_power to 0. 5735 * 5736 * Assumes the sched_domain tree is fully constructed 5737 */ 5738static int 5739build_sched_groups(struct sched_domain *sd, int cpu) 5740{ 5741 struct sched_group *first = NULL, *last = NULL; 5742 struct sd_data *sdd = sd->private; 5743 const struct cpumask *span = sched_domain_span(sd); 5744 struct cpumask *covered; 5745 int i; 5746 5747 get_group(cpu, sdd, &sd->groups); 5748 atomic_inc(&sd->groups->ref); 5749 5750 if (cpu != cpumask_first(span)) 5751 return 0; 5752 5753 lockdep_assert_held(&sched_domains_mutex); 5754 covered = sched_domains_tmpmask; 5755 5756 cpumask_clear(covered); 5757 5758 for_each_cpu(i, span) { 5759 struct sched_group *sg; 5760 int group, j; 5761 5762 if (cpumask_test_cpu(i, covered)) 5763 continue; 5764 5765 group = get_group(i, sdd, &sg); 5766 cpumask_clear(sched_group_cpus(sg)); 5767 sg->sgp->power = 0; 5768 cpumask_setall(sched_group_mask(sg)); 5769 5770 for_each_cpu(j, span) { 5771 if (get_group(j, sdd, NULL) != group) 5772 continue; 5773 5774 cpumask_set_cpu(j, covered); 5775 cpumask_set_cpu(j, sched_group_cpus(sg)); 5776 } 5777 5778 if (!first) 5779 first = sg; 5780 if (last) 5781 last->next = sg; 5782 last = sg; 5783 } 5784 last->next = first; 5785 5786 return 0; 5787} 5788 5789/* 5790 * Initialize sched groups cpu_power. 5791 * 5792 * cpu_power indicates the capacity of sched group, which is used while 5793 * distributing the load between different sched groups in a sched domain. 5794 * Typically cpu_power for all the groups in a sched domain will be same unless 5795 * there are asymmetries in the topology. If there are asymmetries, group 5796 * having more cpu_power will pickup more load compared to the group having 5797 * less cpu_power. 5798 */ 5799static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5800{ 5801 struct sched_group *sg = sd->groups; 5802 5803 WARN_ON(!sg); 5804 5805 do { 5806 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5807 sg = sg->next; 5808 } while (sg != sd->groups); 5809 5810 if (cpu != group_balance_cpu(sg)) 5811 return; 5812 5813 update_group_power(sd, cpu); 5814 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5815} 5816 5817int __weak arch_sd_sibling_asym_packing(void) 5818{ 5819 return 0*SD_ASYM_PACKING; 5820} 5821 5822/* 5823 * Initializers for schedule domains 5824 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5825 */ 5826 5827#ifdef CONFIG_SCHED_DEBUG 5828# define SD_INIT_NAME(sd, type) sd->name = #type 5829#else 5830# define SD_INIT_NAME(sd, type) do { } while (0) 5831#endif 5832 5833#define SD_INIT_FUNC(type) \ 5834static noinline struct sched_domain * \ 5835sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5836{ \ 5837 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5838 *sd = SD_##type##_INIT; \ 5839 SD_INIT_NAME(sd, type); \ 5840 sd->private = &tl->data; \ 5841 return sd; \ 5842} 5843 5844SD_INIT_FUNC(CPU) 5845#ifdef CONFIG_SCHED_SMT 5846 SD_INIT_FUNC(SIBLING) 5847#endif 5848#ifdef CONFIG_SCHED_MC 5849 SD_INIT_FUNC(MC) 5850#endif 5851#ifdef CONFIG_SCHED_BOOK 5852 SD_INIT_FUNC(BOOK) 5853#endif 5854 5855static int default_relax_domain_level = -1; 5856int sched_domain_level_max; 5857 5858static int __init setup_relax_domain_level(char *str) 5859{ 5860 if (kstrtoint(str, 0, &default_relax_domain_level)) 5861 pr_warn("Unable to set relax_domain_level\n"); 5862 5863 return 1; 5864} 5865__setup("relax_domain_level=", setup_relax_domain_level); 5866 5867static void set_domain_attribute(struct sched_domain *sd, 5868 struct sched_domain_attr *attr) 5869{ 5870 int request; 5871 5872 if (!attr || attr->relax_domain_level < 0) { 5873 if (default_relax_domain_level < 0) 5874 return; 5875 else 5876 request = default_relax_domain_level; 5877 } else 5878 request = attr->relax_domain_level; 5879 if (request < sd->level) { 5880 /* turn off idle balance on this domain */ 5881 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5882 } else { 5883 /* turn on idle balance on this domain */ 5884 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5885 } 5886} 5887 5888static void __sdt_free(const struct cpumask *cpu_map); 5889static int __sdt_alloc(const struct cpumask *cpu_map); 5890 5891static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5892 const struct cpumask *cpu_map) 5893{ 5894 switch (what) { 5895 case sa_rootdomain: 5896 if (!atomic_read(&d->rd->refcount)) 5897 free_rootdomain(&d->rd->rcu); /* fall through */ 5898 case sa_sd: 5899 free_percpu(d->sd); /* fall through */ 5900 case sa_sd_storage: 5901 __sdt_free(cpu_map); /* fall through */ 5902 case sa_none: 5903 break; 5904 } 5905} 5906 5907static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5908 const struct cpumask *cpu_map) 5909{ 5910 memset(d, 0, sizeof(*d)); 5911 5912 if (__sdt_alloc(cpu_map)) 5913 return sa_sd_storage; 5914 d->sd = alloc_percpu(struct sched_domain *); 5915 if (!d->sd) 5916 return sa_sd_storage; 5917 d->rd = alloc_rootdomain(); 5918 if (!d->rd) 5919 return sa_sd; 5920 return sa_rootdomain; 5921} 5922 5923/* 5924 * NULL the sd_data elements we've used to build the sched_domain and 5925 * sched_group structure so that the subsequent __free_domain_allocs() 5926 * will not free the data we're using. 5927 */ 5928static void claim_allocations(int cpu, struct sched_domain *sd) 5929{ 5930 struct sd_data *sdd = sd->private; 5931 5932 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5933 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5934 5935 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5936 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5937 5938 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5939 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5940} 5941 5942#ifdef CONFIG_SCHED_SMT 5943static const struct cpumask *cpu_smt_mask(int cpu) 5944{ 5945 return topology_thread_cpumask(cpu); 5946} 5947#endif 5948 5949/* 5950 * Topology list, bottom-up. 5951 */ 5952static struct sched_domain_topology_level default_topology[] = { 5953#ifdef CONFIG_SCHED_SMT 5954 { sd_init_SIBLING, cpu_smt_mask, }, 5955#endif 5956#ifdef CONFIG_SCHED_MC 5957 { sd_init_MC, cpu_coregroup_mask, }, 5958#endif 5959#ifdef CONFIG_SCHED_BOOK 5960 { sd_init_BOOK, cpu_book_mask, }, 5961#endif 5962 { sd_init_CPU, cpu_cpu_mask, }, 5963 { NULL, }, 5964}; 5965 5966static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5967 5968#define for_each_sd_topology(tl) \ 5969 for (tl = sched_domain_topology; tl->init; tl++) 5970 5971#ifdef CONFIG_NUMA 5972 5973static int sched_domains_numa_levels; 5974static int *sched_domains_numa_distance; 5975static struct cpumask ***sched_domains_numa_masks; 5976static int sched_domains_curr_level; 5977 5978static inline int sd_local_flags(int level) 5979{ 5980 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5981 return 0; 5982 5983 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5984} 5985 5986static struct sched_domain * 5987sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5988{ 5989 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5990 int level = tl->numa_level; 5991 int sd_weight = cpumask_weight( 5992 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5993 5994 *sd = (struct sched_domain){ 5995 .min_interval = sd_weight, 5996 .max_interval = 2*sd_weight, 5997 .busy_factor = 32, 5998 .imbalance_pct = 125, 5999 .cache_nice_tries = 2, 6000 .busy_idx = 3, 6001 .idle_idx = 2, 6002 .newidle_idx = 0, 6003 .wake_idx = 0, 6004 .forkexec_idx = 0, 6005 6006 .flags = 1*SD_LOAD_BALANCE 6007 | 1*SD_BALANCE_NEWIDLE 6008 | 0*SD_BALANCE_EXEC 6009 | 0*SD_BALANCE_FORK 6010 | 0*SD_BALANCE_WAKE 6011 | 0*SD_WAKE_AFFINE 6012 | 0*SD_SHARE_CPUPOWER 6013 | 0*SD_SHARE_PKG_RESOURCES 6014 | 1*SD_SERIALIZE 6015 | 0*SD_PREFER_SIBLING 6016 | 1*SD_NUMA 6017 | sd_local_flags(level) 6018 , 6019 .last_balance = jiffies, 6020 .balance_interval = sd_weight, 6021 }; 6022 SD_INIT_NAME(sd, NUMA); 6023 sd->private = &tl->data; 6024 6025 /* 6026 * Ugly hack to pass state to sd_numa_mask()... 6027 */ 6028 sched_domains_curr_level = tl->numa_level; 6029 6030 return sd; 6031} 6032 6033static const struct cpumask *sd_numa_mask(int cpu) 6034{ 6035 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6036} 6037 6038static void sched_numa_warn(const char *str) 6039{ 6040 static int done = false; 6041 int i,j; 6042 6043 if (done) 6044 return; 6045 6046 done = true; 6047 6048 printk(KERN_WARNING "ERROR: %s\n\n", str); 6049 6050 for (i = 0; i < nr_node_ids; i++) { 6051 printk(KERN_WARNING " "); 6052 for (j = 0; j < nr_node_ids; j++) 6053 printk(KERN_CONT "%02d ", node_distance(i,j)); 6054 printk(KERN_CONT "\n"); 6055 } 6056 printk(KERN_WARNING "\n"); 6057} 6058 6059static bool find_numa_distance(int distance) 6060{ 6061 int i; 6062 6063 if (distance == node_distance(0, 0)) 6064 return true; 6065 6066 for (i = 0; i < sched_domains_numa_levels; i++) { 6067 if (sched_domains_numa_distance[i] == distance) 6068 return true; 6069 } 6070 6071 return false; 6072} 6073 6074static void sched_init_numa(void) 6075{ 6076 int next_distance, curr_distance = node_distance(0, 0); 6077 struct sched_domain_topology_level *tl; 6078 int level = 0; 6079 int i, j, k; 6080 6081 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6082 if (!sched_domains_numa_distance) 6083 return; 6084 6085 /* 6086 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6087 * unique distances in the node_distance() table. 6088 * 6089 * Assumes node_distance(0,j) includes all distances in 6090 * node_distance(i,j) in order to avoid cubic time. 6091 */ 6092 next_distance = curr_distance; 6093 for (i = 0; i < nr_node_ids; i++) { 6094 for (j = 0; j < nr_node_ids; j++) { 6095 for (k = 0; k < nr_node_ids; k++) { 6096 int distance = node_distance(i, k); 6097 6098 if (distance > curr_distance && 6099 (distance < next_distance || 6100 next_distance == curr_distance)) 6101 next_distance = distance; 6102 6103 /* 6104 * While not a strong assumption it would be nice to know 6105 * about cases where if node A is connected to B, B is not 6106 * equally connected to A. 6107 */ 6108 if (sched_debug() && node_distance(k, i) != distance) 6109 sched_numa_warn("Node-distance not symmetric"); 6110 6111 if (sched_debug() && i && !find_numa_distance(distance)) 6112 sched_numa_warn("Node-0 not representative"); 6113 } 6114 if (next_distance != curr_distance) { 6115 sched_domains_numa_distance[level++] = next_distance; 6116 sched_domains_numa_levels = level; 6117 curr_distance = next_distance; 6118 } else break; 6119 } 6120 6121 /* 6122 * In case of sched_debug() we verify the above assumption. 6123 */ 6124 if (!sched_debug()) 6125 break; 6126 } 6127 /* 6128 * 'level' contains the number of unique distances, excluding the 6129 * identity distance node_distance(i,i). 6130 * 6131 * The sched_domains_numa_distance[] array includes the actual distance 6132 * numbers. 6133 */ 6134 6135 /* 6136 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6137 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6138 * the array will contain less then 'level' members. This could be 6139 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6140 * in other functions. 6141 * 6142 * We reset it to 'level' at the end of this function. 6143 */ 6144 sched_domains_numa_levels = 0; 6145 6146 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6147 if (!sched_domains_numa_masks) 6148 return; 6149 6150 /* 6151 * Now for each level, construct a mask per node which contains all 6152 * cpus of nodes that are that many hops away from us. 6153 */ 6154 for (i = 0; i < level; i++) { 6155 sched_domains_numa_masks[i] = 6156 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6157 if (!sched_domains_numa_masks[i]) 6158 return; 6159 6160 for (j = 0; j < nr_node_ids; j++) { 6161 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6162 if (!mask) 6163 return; 6164 6165 sched_domains_numa_masks[i][j] = mask; 6166 6167 for (k = 0; k < nr_node_ids; k++) { 6168 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6169 continue; 6170 6171 cpumask_or(mask, mask, cpumask_of_node(k)); 6172 } 6173 } 6174 } 6175 6176 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6177 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6178 if (!tl) 6179 return; 6180 6181 /* 6182 * Copy the default topology bits.. 6183 */ 6184 for (i = 0; default_topology[i].init; i++) 6185 tl[i] = default_topology[i]; 6186 6187 /* 6188 * .. and append 'j' levels of NUMA goodness. 6189 */ 6190 for (j = 0; j < level; i++, j++) { 6191 tl[i] = (struct sched_domain_topology_level){ 6192 .init = sd_numa_init, 6193 .mask = sd_numa_mask, 6194 .flags = SDTL_OVERLAP, 6195 .numa_level = j, 6196 }; 6197 } 6198 6199 sched_domain_topology = tl; 6200 6201 sched_domains_numa_levels = level; 6202} 6203 6204static void sched_domains_numa_masks_set(int cpu) 6205{ 6206 int i, j; 6207 int node = cpu_to_node(cpu); 6208 6209 for (i = 0; i < sched_domains_numa_levels; i++) { 6210 for (j = 0; j < nr_node_ids; j++) { 6211 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6212 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6213 } 6214 } 6215} 6216 6217static void sched_domains_numa_masks_clear(int cpu) 6218{ 6219 int i, j; 6220 for (i = 0; i < sched_domains_numa_levels; i++) { 6221 for (j = 0; j < nr_node_ids; j++) 6222 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6223 } 6224} 6225 6226/* 6227 * Update sched_domains_numa_masks[level][node] array when new cpus 6228 * are onlined. 6229 */ 6230static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6231 unsigned long action, 6232 void *hcpu) 6233{ 6234 int cpu = (long)hcpu; 6235 6236 switch (action & ~CPU_TASKS_FROZEN) { 6237 case CPU_ONLINE: 6238 sched_domains_numa_masks_set(cpu); 6239 break; 6240 6241 case CPU_DEAD: 6242 sched_domains_numa_masks_clear(cpu); 6243 break; 6244 6245 default: 6246 return NOTIFY_DONE; 6247 } 6248 6249 return NOTIFY_OK; 6250} 6251#else 6252static inline void sched_init_numa(void) 6253{ 6254} 6255 6256static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6257 unsigned long action, 6258 void *hcpu) 6259{ 6260 return 0; 6261} 6262#endif /* CONFIG_NUMA */ 6263 6264static int __sdt_alloc(const struct cpumask *cpu_map) 6265{ 6266 struct sched_domain_topology_level *tl; 6267 int j; 6268 6269 for_each_sd_topology(tl) { 6270 struct sd_data *sdd = &tl->data; 6271 6272 sdd->sd = alloc_percpu(struct sched_domain *); 6273 if (!sdd->sd) 6274 return -ENOMEM; 6275 6276 sdd->sg = alloc_percpu(struct sched_group *); 6277 if (!sdd->sg) 6278 return -ENOMEM; 6279 6280 sdd->sgp = alloc_percpu(struct sched_group_power *); 6281 if (!sdd->sgp) 6282 return -ENOMEM; 6283 6284 for_each_cpu(j, cpu_map) { 6285 struct sched_domain *sd; 6286 struct sched_group *sg; 6287 struct sched_group_power *sgp; 6288 6289 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6290 GFP_KERNEL, cpu_to_node(j)); 6291 if (!sd) 6292 return -ENOMEM; 6293 6294 *per_cpu_ptr(sdd->sd, j) = sd; 6295 6296 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6297 GFP_KERNEL, cpu_to_node(j)); 6298 if (!sg) 6299 return -ENOMEM; 6300 6301 sg->next = sg; 6302 6303 *per_cpu_ptr(sdd->sg, j) = sg; 6304 6305 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6306 GFP_KERNEL, cpu_to_node(j)); 6307 if (!sgp) 6308 return -ENOMEM; 6309 6310 *per_cpu_ptr(sdd->sgp, j) = sgp; 6311 } 6312 } 6313 6314 return 0; 6315} 6316 6317static void __sdt_free(const struct cpumask *cpu_map) 6318{ 6319 struct sched_domain_topology_level *tl; 6320 int j; 6321 6322 for_each_sd_topology(tl) { 6323 struct sd_data *sdd = &tl->data; 6324 6325 for_each_cpu(j, cpu_map) { 6326 struct sched_domain *sd; 6327 6328 if (sdd->sd) { 6329 sd = *per_cpu_ptr(sdd->sd, j); 6330 if (sd && (sd->flags & SD_OVERLAP)) 6331 free_sched_groups(sd->groups, 0); 6332 kfree(*per_cpu_ptr(sdd->sd, j)); 6333 } 6334 6335 if (sdd->sg) 6336 kfree(*per_cpu_ptr(sdd->sg, j)); 6337 if (sdd->sgp) 6338 kfree(*per_cpu_ptr(sdd->sgp, j)); 6339 } 6340 free_percpu(sdd->sd); 6341 sdd->sd = NULL; 6342 free_percpu(sdd->sg); 6343 sdd->sg = NULL; 6344 free_percpu(sdd->sgp); 6345 sdd->sgp = NULL; 6346 } 6347} 6348 6349struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6350 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6351 struct sched_domain *child, int cpu) 6352{ 6353 struct sched_domain *sd = tl->init(tl, cpu); 6354 if (!sd) 6355 return child; 6356 6357 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6358 if (child) { 6359 sd->level = child->level + 1; 6360 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6361 child->parent = sd; 6362 sd->child = child; 6363 } 6364 set_domain_attribute(sd, attr); 6365 6366 return sd; 6367} 6368 6369/* 6370 * Build sched domains for a given set of cpus and attach the sched domains 6371 * to the individual cpus 6372 */ 6373static int build_sched_domains(const struct cpumask *cpu_map, 6374 struct sched_domain_attr *attr) 6375{ 6376 enum s_alloc alloc_state; 6377 struct sched_domain *sd; 6378 struct s_data d; 6379 int i, ret = -ENOMEM; 6380 6381 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6382 if (alloc_state != sa_rootdomain) 6383 goto error; 6384 6385 /* Set up domains for cpus specified by the cpu_map. */ 6386 for_each_cpu(i, cpu_map) { 6387 struct sched_domain_topology_level *tl; 6388 6389 sd = NULL; 6390 for_each_sd_topology(tl) { 6391 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6392 if (tl == sched_domain_topology) 6393 *per_cpu_ptr(d.sd, i) = sd; 6394 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6395 sd->flags |= SD_OVERLAP; 6396 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6397 break; 6398 } 6399 } 6400 6401 /* Build the groups for the domains */ 6402 for_each_cpu(i, cpu_map) { 6403 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6404 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6405 if (sd->flags & SD_OVERLAP) { 6406 if (build_overlap_sched_groups(sd, i)) 6407 goto error; 6408 } else { 6409 if (build_sched_groups(sd, i)) 6410 goto error; 6411 } 6412 } 6413 } 6414 6415 /* Calculate CPU power for physical packages and nodes */ 6416 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6417 if (!cpumask_test_cpu(i, cpu_map)) 6418 continue; 6419 6420 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6421 claim_allocations(i, sd); 6422 init_sched_groups_power(i, sd); 6423 } 6424 } 6425 6426 /* Attach the domains */ 6427 rcu_read_lock(); 6428 for_each_cpu(i, cpu_map) { 6429 sd = *per_cpu_ptr(d.sd, i); 6430 cpu_attach_domain(sd, d.rd, i); 6431 } 6432 rcu_read_unlock(); 6433 6434 ret = 0; 6435error: 6436 __free_domain_allocs(&d, alloc_state, cpu_map); 6437 return ret; 6438} 6439 6440static cpumask_var_t *doms_cur; /* current sched domains */ 6441static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6442static struct sched_domain_attr *dattr_cur; 6443 /* attribues of custom domains in 'doms_cur' */ 6444 6445/* 6446 * Special case: If a kmalloc of a doms_cur partition (array of 6447 * cpumask) fails, then fallback to a single sched domain, 6448 * as determined by the single cpumask fallback_doms. 6449 */ 6450static cpumask_var_t fallback_doms; 6451 6452/* 6453 * arch_update_cpu_topology lets virtualized architectures update the 6454 * cpu core maps. It is supposed to return 1 if the topology changed 6455 * or 0 if it stayed the same. 6456 */ 6457int __attribute__((weak)) arch_update_cpu_topology(void) 6458{ 6459 return 0; 6460} 6461 6462cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6463{ 6464 int i; 6465 cpumask_var_t *doms; 6466 6467 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6468 if (!doms) 6469 return NULL; 6470 for (i = 0; i < ndoms; i++) { 6471 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6472 free_sched_domains(doms, i); 6473 return NULL; 6474 } 6475 } 6476 return doms; 6477} 6478 6479void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6480{ 6481 unsigned int i; 6482 for (i = 0; i < ndoms; i++) 6483 free_cpumask_var(doms[i]); 6484 kfree(doms); 6485} 6486 6487/* 6488 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6489 * For now this just excludes isolated cpus, but could be used to 6490 * exclude other special cases in the future. 6491 */ 6492static int init_sched_domains(const struct cpumask *cpu_map) 6493{ 6494 int err; 6495 6496 arch_update_cpu_topology(); 6497 ndoms_cur = 1; 6498 doms_cur = alloc_sched_domains(ndoms_cur); 6499 if (!doms_cur) 6500 doms_cur = &fallback_doms; 6501 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6502 err = build_sched_domains(doms_cur[0], NULL); 6503 register_sched_domain_sysctl(); 6504 6505 return err; 6506} 6507 6508/* 6509 * Detach sched domains from a group of cpus specified in cpu_map 6510 * These cpus will now be attached to the NULL domain 6511 */ 6512static void detach_destroy_domains(const struct cpumask *cpu_map) 6513{ 6514 int i; 6515 6516 rcu_read_lock(); 6517 for_each_cpu(i, cpu_map) 6518 cpu_attach_domain(NULL, &def_root_domain, i); 6519 rcu_read_unlock(); 6520} 6521 6522/* handle null as "default" */ 6523static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6524 struct sched_domain_attr *new, int idx_new) 6525{ 6526 struct sched_domain_attr tmp; 6527 6528 /* fast path */ 6529 if (!new && !cur) 6530 return 1; 6531 6532 tmp = SD_ATTR_INIT; 6533 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6534 new ? (new + idx_new) : &tmp, 6535 sizeof(struct sched_domain_attr)); 6536} 6537 6538/* 6539 * Partition sched domains as specified by the 'ndoms_new' 6540 * cpumasks in the array doms_new[] of cpumasks. This compares 6541 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6542 * It destroys each deleted domain and builds each new domain. 6543 * 6544 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6545 * The masks don't intersect (don't overlap.) We should setup one 6546 * sched domain for each mask. CPUs not in any of the cpumasks will 6547 * not be load balanced. If the same cpumask appears both in the 6548 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6549 * it as it is. 6550 * 6551 * The passed in 'doms_new' should be allocated using 6552 * alloc_sched_domains. This routine takes ownership of it and will 6553 * free_sched_domains it when done with it. If the caller failed the 6554 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6555 * and partition_sched_domains() will fallback to the single partition 6556 * 'fallback_doms', it also forces the domains to be rebuilt. 6557 * 6558 * If doms_new == NULL it will be replaced with cpu_online_mask. 6559 * ndoms_new == 0 is a special case for destroying existing domains, 6560 * and it will not create the default domain. 6561 * 6562 * Call with hotplug lock held 6563 */ 6564void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6565 struct sched_domain_attr *dattr_new) 6566{ 6567 int i, j, n; 6568 int new_topology; 6569 6570 mutex_lock(&sched_domains_mutex); 6571 6572 /* always unregister in case we don't destroy any domains */ 6573 unregister_sched_domain_sysctl(); 6574 6575 /* Let architecture update cpu core mappings. */ 6576 new_topology = arch_update_cpu_topology(); 6577 6578 n = doms_new ? ndoms_new : 0; 6579 6580 /* Destroy deleted domains */ 6581 for (i = 0; i < ndoms_cur; i++) { 6582 for (j = 0; j < n && !new_topology; j++) { 6583 if (cpumask_equal(doms_cur[i], doms_new[j]) 6584 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6585 goto match1; 6586 } 6587 /* no match - a current sched domain not in new doms_new[] */ 6588 detach_destroy_domains(doms_cur[i]); 6589match1: 6590 ; 6591 } 6592 6593 n = ndoms_cur; 6594 if (doms_new == NULL) { 6595 n = 0; 6596 doms_new = &fallback_doms; 6597 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6598 WARN_ON_ONCE(dattr_new); 6599 } 6600 6601 /* Build new domains */ 6602 for (i = 0; i < ndoms_new; i++) { 6603 for (j = 0; j < n && !new_topology; j++) { 6604 if (cpumask_equal(doms_new[i], doms_cur[j]) 6605 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6606 goto match2; 6607 } 6608 /* no match - add a new doms_new */ 6609 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6610match2: 6611 ; 6612 } 6613 6614 /* Remember the new sched domains */ 6615 if (doms_cur != &fallback_doms) 6616 free_sched_domains(doms_cur, ndoms_cur); 6617 kfree(dattr_cur); /* kfree(NULL) is safe */ 6618 doms_cur = doms_new; 6619 dattr_cur = dattr_new; 6620 ndoms_cur = ndoms_new; 6621 6622 register_sched_domain_sysctl(); 6623 6624 mutex_unlock(&sched_domains_mutex); 6625} 6626 6627static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6628 6629/* 6630 * Update cpusets according to cpu_active mask. If cpusets are 6631 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6632 * around partition_sched_domains(). 6633 * 6634 * If we come here as part of a suspend/resume, don't touch cpusets because we 6635 * want to restore it back to its original state upon resume anyway. 6636 */ 6637static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6638 void *hcpu) 6639{ 6640 switch (action) { 6641 case CPU_ONLINE_FROZEN: 6642 case CPU_DOWN_FAILED_FROZEN: 6643 6644 /* 6645 * num_cpus_frozen tracks how many CPUs are involved in suspend 6646 * resume sequence. As long as this is not the last online 6647 * operation in the resume sequence, just build a single sched 6648 * domain, ignoring cpusets. 6649 */ 6650 num_cpus_frozen--; 6651 if (likely(num_cpus_frozen)) { 6652 partition_sched_domains(1, NULL, NULL); 6653 break; 6654 } 6655 6656 /* 6657 * This is the last CPU online operation. So fall through and 6658 * restore the original sched domains by considering the 6659 * cpuset configurations. 6660 */ 6661 6662 case CPU_ONLINE: 6663 case CPU_DOWN_FAILED: 6664 cpuset_update_active_cpus(true); 6665 break; 6666 default: 6667 return NOTIFY_DONE; 6668 } 6669 return NOTIFY_OK; 6670} 6671 6672static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6673 void *hcpu) 6674{ 6675 switch (action) { 6676 case CPU_DOWN_PREPARE: 6677 cpuset_update_active_cpus(false); 6678 break; 6679 case CPU_DOWN_PREPARE_FROZEN: 6680 num_cpus_frozen++; 6681 partition_sched_domains(1, NULL, NULL); 6682 break; 6683 default: 6684 return NOTIFY_DONE; 6685 } 6686 return NOTIFY_OK; 6687} 6688 6689void __init sched_init_smp(void) 6690{ 6691 cpumask_var_t non_isolated_cpus; 6692 6693 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6694 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6695 6696 sched_init_numa(); 6697 6698 /* 6699 * There's no userspace yet to cause hotplug operations; hence all the 6700 * cpu masks are stable and all blatant races in the below code cannot 6701 * happen. 6702 */ 6703 mutex_lock(&sched_domains_mutex); 6704 init_sched_domains(cpu_active_mask); 6705 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6706 if (cpumask_empty(non_isolated_cpus)) 6707 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6708 mutex_unlock(&sched_domains_mutex); 6709 6710 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6711 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6712 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6713 6714 init_hrtick(); 6715 6716 /* Move init over to a non-isolated CPU */ 6717 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6718 BUG(); 6719 sched_init_granularity(); 6720 free_cpumask_var(non_isolated_cpus); 6721 6722 init_sched_rt_class(); 6723 init_sched_dl_class(); 6724} 6725#else 6726void __init sched_init_smp(void) 6727{ 6728 sched_init_granularity(); 6729} 6730#endif /* CONFIG_SMP */ 6731 6732const_debug unsigned int sysctl_timer_migration = 1; 6733 6734int in_sched_functions(unsigned long addr) 6735{ 6736 return in_lock_functions(addr) || 6737 (addr >= (unsigned long)__sched_text_start 6738 && addr < (unsigned long)__sched_text_end); 6739} 6740 6741#ifdef CONFIG_CGROUP_SCHED 6742/* 6743 * Default task group. 6744 * Every task in system belongs to this group at bootup. 6745 */ 6746struct task_group root_task_group; 6747LIST_HEAD(task_groups); 6748#endif 6749 6750DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6751 6752void __init sched_init(void) 6753{ 6754 int i, j; 6755 unsigned long alloc_size = 0, ptr; 6756 6757#ifdef CONFIG_FAIR_GROUP_SCHED 6758 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6759#endif 6760#ifdef CONFIG_RT_GROUP_SCHED 6761 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6762#endif 6763#ifdef CONFIG_CPUMASK_OFFSTACK 6764 alloc_size += num_possible_cpus() * cpumask_size(); 6765#endif 6766 if (alloc_size) { 6767 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6768 6769#ifdef CONFIG_FAIR_GROUP_SCHED 6770 root_task_group.se = (struct sched_entity **)ptr; 6771 ptr += nr_cpu_ids * sizeof(void **); 6772 6773 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6774 ptr += nr_cpu_ids * sizeof(void **); 6775 6776#endif /* CONFIG_FAIR_GROUP_SCHED */ 6777#ifdef CONFIG_RT_GROUP_SCHED 6778 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6779 ptr += nr_cpu_ids * sizeof(void **); 6780 6781 root_task_group.rt_rq = (struct rt_rq **)ptr; 6782 ptr += nr_cpu_ids * sizeof(void **); 6783 6784#endif /* CONFIG_RT_GROUP_SCHED */ 6785#ifdef CONFIG_CPUMASK_OFFSTACK 6786 for_each_possible_cpu(i) { 6787 per_cpu(load_balance_mask, i) = (void *)ptr; 6788 ptr += cpumask_size(); 6789 } 6790#endif /* CONFIG_CPUMASK_OFFSTACK */ 6791 } 6792 6793 init_rt_bandwidth(&def_rt_bandwidth, 6794 global_rt_period(), global_rt_runtime()); 6795 init_dl_bandwidth(&def_dl_bandwidth, 6796 global_rt_period(), global_rt_runtime()); 6797 6798#ifdef CONFIG_SMP 6799 init_defrootdomain(); 6800#endif 6801 6802#ifdef CONFIG_RT_GROUP_SCHED 6803 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6804 global_rt_period(), global_rt_runtime()); 6805#endif /* CONFIG_RT_GROUP_SCHED */ 6806 6807#ifdef CONFIG_CGROUP_SCHED 6808 list_add(&root_task_group.list, &task_groups); 6809 INIT_LIST_HEAD(&root_task_group.children); 6810 INIT_LIST_HEAD(&root_task_group.siblings); 6811 autogroup_init(&init_task); 6812 6813#endif /* CONFIG_CGROUP_SCHED */ 6814 6815 for_each_possible_cpu(i) { 6816 struct rq *rq; 6817 6818 rq = cpu_rq(i); 6819 raw_spin_lock_init(&rq->lock); 6820 rq->nr_running = 0; 6821 rq->calc_load_active = 0; 6822 rq->calc_load_update = jiffies + LOAD_FREQ; 6823 init_cfs_rq(&rq->cfs); 6824 init_rt_rq(&rq->rt, rq); 6825 init_dl_rq(&rq->dl, rq); 6826#ifdef CONFIG_FAIR_GROUP_SCHED 6827 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6828 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6829 /* 6830 * How much cpu bandwidth does root_task_group get? 6831 * 6832 * In case of task-groups formed thr' the cgroup filesystem, it 6833 * gets 100% of the cpu resources in the system. This overall 6834 * system cpu resource is divided among the tasks of 6835 * root_task_group and its child task-groups in a fair manner, 6836 * based on each entity's (task or task-group's) weight 6837 * (se->load.weight). 6838 * 6839 * In other words, if root_task_group has 10 tasks of weight 6840 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6841 * then A0's share of the cpu resource is: 6842 * 6843 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6844 * 6845 * We achieve this by letting root_task_group's tasks sit 6846 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6847 */ 6848 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6849 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6850#endif /* CONFIG_FAIR_GROUP_SCHED */ 6851 6852 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6853#ifdef CONFIG_RT_GROUP_SCHED 6854 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6855 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6856#endif 6857 6858 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6859 rq->cpu_load[j] = 0; 6860 6861 rq->last_load_update_tick = jiffies; 6862 6863#ifdef CONFIG_SMP 6864 rq->sd = NULL; 6865 rq->rd = NULL; 6866 rq->cpu_power = SCHED_POWER_SCALE; 6867 rq->post_schedule = 0; 6868 rq->active_balance = 0; 6869 rq->next_balance = jiffies; 6870 rq->push_cpu = 0; 6871 rq->cpu = i; 6872 rq->online = 0; 6873 rq->idle_stamp = 0; 6874 rq->avg_idle = 2*sysctl_sched_migration_cost; 6875 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6876 6877 INIT_LIST_HEAD(&rq->cfs_tasks); 6878 6879 rq_attach_root(rq, &def_root_domain); 6880#ifdef CONFIG_NO_HZ_COMMON 6881 rq->nohz_flags = 0; 6882#endif 6883#ifdef CONFIG_NO_HZ_FULL 6884 rq->last_sched_tick = 0; 6885#endif 6886#endif 6887 init_rq_hrtick(rq); 6888 atomic_set(&rq->nr_iowait, 0); 6889 } 6890 6891 set_load_weight(&init_task); 6892 6893#ifdef CONFIG_PREEMPT_NOTIFIERS 6894 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6895#endif 6896 6897 /* 6898 * The boot idle thread does lazy MMU switching as well: 6899 */ 6900 atomic_inc(&init_mm.mm_count); 6901 enter_lazy_tlb(&init_mm, current); 6902 6903 /* 6904 * Make us the idle thread. Technically, schedule() should not be 6905 * called from this thread, however somewhere below it might be, 6906 * but because we are the idle thread, we just pick up running again 6907 * when this runqueue becomes "idle". 6908 */ 6909 init_idle(current, smp_processor_id()); 6910 6911 calc_load_update = jiffies + LOAD_FREQ; 6912 6913 /* 6914 * During early bootup we pretend to be a normal task: 6915 */ 6916 current->sched_class = &fair_sched_class; 6917 6918#ifdef CONFIG_SMP 6919 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6920 /* May be allocated at isolcpus cmdline parse time */ 6921 if (cpu_isolated_map == NULL) 6922 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6923 idle_thread_set_boot_cpu(); 6924#endif 6925 init_sched_fair_class(); 6926 6927 scheduler_running = 1; 6928} 6929 6930#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6931static inline int preempt_count_equals(int preempt_offset) 6932{ 6933 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6934 6935 return (nested == preempt_offset); 6936} 6937 6938void __might_sleep(const char *file, int line, int preempt_offset) 6939{ 6940 static unsigned long prev_jiffy; /* ratelimiting */ 6941 6942 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6943 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6944 system_state != SYSTEM_RUNNING || oops_in_progress) 6945 return; 6946 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6947 return; 6948 prev_jiffy = jiffies; 6949 6950 printk(KERN_ERR 6951 "BUG: sleeping function called from invalid context at %s:%d\n", 6952 file, line); 6953 printk(KERN_ERR 6954 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6955 in_atomic(), irqs_disabled(), 6956 current->pid, current->comm); 6957 6958 debug_show_held_locks(current); 6959 if (irqs_disabled()) 6960 print_irqtrace_events(current); 6961 dump_stack(); 6962} 6963EXPORT_SYMBOL(__might_sleep); 6964#endif 6965 6966#ifdef CONFIG_MAGIC_SYSRQ 6967static void normalize_task(struct rq *rq, struct task_struct *p) 6968{ 6969 const struct sched_class *prev_class = p->sched_class; 6970 struct sched_attr attr = { 6971 .sched_policy = SCHED_NORMAL, 6972 }; 6973 int old_prio = p->prio; 6974 int on_rq; 6975 6976 on_rq = p->on_rq; 6977 if (on_rq) 6978 dequeue_task(rq, p, 0); 6979 __setscheduler(rq, p, &attr); 6980 if (on_rq) { 6981 enqueue_task(rq, p, 0); 6982 resched_task(rq->curr); 6983 } 6984 6985 check_class_changed(rq, p, prev_class, old_prio); 6986} 6987 6988void normalize_rt_tasks(void) 6989{ 6990 struct task_struct *g, *p; 6991 unsigned long flags; 6992 struct rq *rq; 6993 6994 read_lock_irqsave(&tasklist_lock, flags); 6995 do_each_thread(g, p) { 6996 /* 6997 * Only normalize user tasks: 6998 */ 6999 if (!p->mm) 7000 continue; 7001 7002 p->se.exec_start = 0; 7003#ifdef CONFIG_SCHEDSTATS 7004 p->se.statistics.wait_start = 0; 7005 p->se.statistics.sleep_start = 0; 7006 p->se.statistics.block_start = 0; 7007#endif 7008 7009 if (!dl_task(p) && !rt_task(p)) { 7010 /* 7011 * Renice negative nice level userspace 7012 * tasks back to 0: 7013 */ 7014 if (TASK_NICE(p) < 0 && p->mm) 7015 set_user_nice(p, 0); 7016 continue; 7017 } 7018 7019 raw_spin_lock(&p->pi_lock); 7020 rq = __task_rq_lock(p); 7021 7022 normalize_task(rq, p); 7023 7024 __task_rq_unlock(rq); 7025 raw_spin_unlock(&p->pi_lock); 7026 } while_each_thread(g, p); 7027 7028 read_unlock_irqrestore(&tasklist_lock, flags); 7029} 7030 7031#endif /* CONFIG_MAGIC_SYSRQ */ 7032 7033#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7034/* 7035 * These functions are only useful for the IA64 MCA handling, or kdb. 7036 * 7037 * They can only be called when the whole system has been 7038 * stopped - every CPU needs to be quiescent, and no scheduling 7039 * activity can take place. Using them for anything else would 7040 * be a serious bug, and as a result, they aren't even visible 7041 * under any other configuration. 7042 */ 7043 7044/** 7045 * curr_task - return the current task for a given cpu. 7046 * @cpu: the processor in question. 7047 * 7048 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7049 * 7050 * Return: The current task for @cpu. 7051 */ 7052struct task_struct *curr_task(int cpu) 7053{ 7054 return cpu_curr(cpu); 7055} 7056 7057#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7058 7059#ifdef CONFIG_IA64 7060/** 7061 * set_curr_task - set the current task for a given cpu. 7062 * @cpu: the processor in question. 7063 * @p: the task pointer to set. 7064 * 7065 * Description: This function must only be used when non-maskable interrupts 7066 * are serviced on a separate stack. It allows the architecture to switch the 7067 * notion of the current task on a cpu in a non-blocking manner. This function 7068 * must be called with all CPU's synchronized, and interrupts disabled, the 7069 * and caller must save the original value of the current task (see 7070 * curr_task() above) and restore that value before reenabling interrupts and 7071 * re-starting the system. 7072 * 7073 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7074 */ 7075void set_curr_task(int cpu, struct task_struct *p) 7076{ 7077 cpu_curr(cpu) = p; 7078} 7079 7080#endif 7081 7082#ifdef CONFIG_CGROUP_SCHED 7083/* task_group_lock serializes the addition/removal of task groups */ 7084static DEFINE_SPINLOCK(task_group_lock); 7085 7086static void free_sched_group(struct task_group *tg) 7087{ 7088 free_fair_sched_group(tg); 7089 free_rt_sched_group(tg); 7090 autogroup_free(tg); 7091 kfree(tg); 7092} 7093 7094/* allocate runqueue etc for a new task group */ 7095struct task_group *sched_create_group(struct task_group *parent) 7096{ 7097 struct task_group *tg; 7098 7099 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7100 if (!tg) 7101 return ERR_PTR(-ENOMEM); 7102 7103 if (!alloc_fair_sched_group(tg, parent)) 7104 goto err; 7105 7106 if (!alloc_rt_sched_group(tg, parent)) 7107 goto err; 7108 7109 return tg; 7110 7111err: 7112 free_sched_group(tg); 7113 return ERR_PTR(-ENOMEM); 7114} 7115 7116void sched_online_group(struct task_group *tg, struct task_group *parent) 7117{ 7118 unsigned long flags; 7119 7120 spin_lock_irqsave(&task_group_lock, flags); 7121 list_add_rcu(&tg->list, &task_groups); 7122 7123 WARN_ON(!parent); /* root should already exist */ 7124 7125 tg->parent = parent; 7126 INIT_LIST_HEAD(&tg->children); 7127 list_add_rcu(&tg->siblings, &parent->children); 7128 spin_unlock_irqrestore(&task_group_lock, flags); 7129} 7130 7131/* rcu callback to free various structures associated with a task group */ 7132static void free_sched_group_rcu(struct rcu_head *rhp) 7133{ 7134 /* now it should be safe to free those cfs_rqs */ 7135 free_sched_group(container_of(rhp, struct task_group, rcu)); 7136} 7137 7138/* Destroy runqueue etc associated with a task group */ 7139void sched_destroy_group(struct task_group *tg) 7140{ 7141 /* wait for possible concurrent references to cfs_rqs complete */ 7142 call_rcu(&tg->rcu, free_sched_group_rcu); 7143} 7144 7145void sched_offline_group(struct task_group *tg) 7146{ 7147 unsigned long flags; 7148 int i; 7149 7150 /* end participation in shares distribution */ 7151 for_each_possible_cpu(i) 7152 unregister_fair_sched_group(tg, i); 7153 7154 spin_lock_irqsave(&task_group_lock, flags); 7155 list_del_rcu(&tg->list); 7156 list_del_rcu(&tg->siblings); 7157 spin_unlock_irqrestore(&task_group_lock, flags); 7158} 7159 7160/* change task's runqueue when it moves between groups. 7161 * The caller of this function should have put the task in its new group 7162 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7163 * reflect its new group. 7164 */ 7165void sched_move_task(struct task_struct *tsk) 7166{ 7167 struct task_group *tg; 7168 int on_rq, running; 7169 unsigned long flags; 7170 struct rq *rq; 7171 7172 rq = task_rq_lock(tsk, &flags); 7173 7174 running = task_current(rq, tsk); 7175 on_rq = tsk->on_rq; 7176 7177 if (on_rq) 7178 dequeue_task(rq, tsk, 0); 7179 if (unlikely(running)) 7180 tsk->sched_class->put_prev_task(rq, tsk); 7181 7182 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 7183 lockdep_is_held(&tsk->sighand->siglock)), 7184 struct task_group, css); 7185 tg = autogroup_task_group(tsk, tg); 7186 tsk->sched_task_group = tg; 7187 7188#ifdef CONFIG_FAIR_GROUP_SCHED 7189 if (tsk->sched_class->task_move_group) 7190 tsk->sched_class->task_move_group(tsk, on_rq); 7191 else 7192#endif 7193 set_task_rq(tsk, task_cpu(tsk)); 7194 7195 if (unlikely(running)) 7196 tsk->sched_class->set_curr_task(rq); 7197 if (on_rq) 7198 enqueue_task(rq, tsk, 0); 7199 7200 task_rq_unlock(rq, tsk, &flags); 7201} 7202#endif /* CONFIG_CGROUP_SCHED */ 7203 7204#ifdef CONFIG_RT_GROUP_SCHED 7205/* 7206 * Ensure that the real time constraints are schedulable. 7207 */ 7208static DEFINE_MUTEX(rt_constraints_mutex); 7209 7210/* Must be called with tasklist_lock held */ 7211static inline int tg_has_rt_tasks(struct task_group *tg) 7212{ 7213 struct task_struct *g, *p; 7214 7215 do_each_thread(g, p) { 7216 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7217 return 1; 7218 } while_each_thread(g, p); 7219 7220 return 0; 7221} 7222 7223struct rt_schedulable_data { 7224 struct task_group *tg; 7225 u64 rt_period; 7226 u64 rt_runtime; 7227}; 7228 7229static int tg_rt_schedulable(struct task_group *tg, void *data) 7230{ 7231 struct rt_schedulable_data *d = data; 7232 struct task_group *child; 7233 unsigned long total, sum = 0; 7234 u64 period, runtime; 7235 7236 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7237 runtime = tg->rt_bandwidth.rt_runtime; 7238 7239 if (tg == d->tg) { 7240 period = d->rt_period; 7241 runtime = d->rt_runtime; 7242 } 7243 7244 /* 7245 * Cannot have more runtime than the period. 7246 */ 7247 if (runtime > period && runtime != RUNTIME_INF) 7248 return -EINVAL; 7249 7250 /* 7251 * Ensure we don't starve existing RT tasks. 7252 */ 7253 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7254 return -EBUSY; 7255 7256 total = to_ratio(period, runtime); 7257 7258 /* 7259 * Nobody can have more than the global setting allows. 7260 */ 7261 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7262 return -EINVAL; 7263 7264 /* 7265 * The sum of our children's runtime should not exceed our own. 7266 */ 7267 list_for_each_entry_rcu(child, &tg->children, siblings) { 7268 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7269 runtime = child->rt_bandwidth.rt_runtime; 7270 7271 if (child == d->tg) { 7272 period = d->rt_period; 7273 runtime = d->rt_runtime; 7274 } 7275 7276 sum += to_ratio(period, runtime); 7277 } 7278 7279 if (sum > total) 7280 return -EINVAL; 7281 7282 return 0; 7283} 7284 7285static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7286{ 7287 int ret; 7288 7289 struct rt_schedulable_data data = { 7290 .tg = tg, 7291 .rt_period = period, 7292 .rt_runtime = runtime, 7293 }; 7294 7295 rcu_read_lock(); 7296 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7297 rcu_read_unlock(); 7298 7299 return ret; 7300} 7301 7302static int tg_set_rt_bandwidth(struct task_group *tg, 7303 u64 rt_period, u64 rt_runtime) 7304{ 7305 int i, err = 0; 7306 7307 mutex_lock(&rt_constraints_mutex); 7308 read_lock(&tasklist_lock); 7309 err = __rt_schedulable(tg, rt_period, rt_runtime); 7310 if (err) 7311 goto unlock; 7312 7313 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7314 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7315 tg->rt_bandwidth.rt_runtime = rt_runtime; 7316 7317 for_each_possible_cpu(i) { 7318 struct rt_rq *rt_rq = tg->rt_rq[i]; 7319 7320 raw_spin_lock(&rt_rq->rt_runtime_lock); 7321 rt_rq->rt_runtime = rt_runtime; 7322 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7323 } 7324 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7325unlock: 7326 read_unlock(&tasklist_lock); 7327 mutex_unlock(&rt_constraints_mutex); 7328 7329 return err; 7330} 7331 7332static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7333{ 7334 u64 rt_runtime, rt_period; 7335 7336 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7337 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7338 if (rt_runtime_us < 0) 7339 rt_runtime = RUNTIME_INF; 7340 7341 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7342} 7343 7344static long sched_group_rt_runtime(struct task_group *tg) 7345{ 7346 u64 rt_runtime_us; 7347 7348 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7349 return -1; 7350 7351 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7352 do_div(rt_runtime_us, NSEC_PER_USEC); 7353 return rt_runtime_us; 7354} 7355 7356static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7357{ 7358 u64 rt_runtime, rt_period; 7359 7360 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7361 rt_runtime = tg->rt_bandwidth.rt_runtime; 7362 7363 if (rt_period == 0) 7364 return -EINVAL; 7365 7366 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7367} 7368 7369static long sched_group_rt_period(struct task_group *tg) 7370{ 7371 u64 rt_period_us; 7372 7373 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7374 do_div(rt_period_us, NSEC_PER_USEC); 7375 return rt_period_us; 7376} 7377#endif /* CONFIG_RT_GROUP_SCHED */ 7378 7379#ifdef CONFIG_RT_GROUP_SCHED 7380static int sched_rt_global_constraints(void) 7381{ 7382 int ret = 0; 7383 7384 mutex_lock(&rt_constraints_mutex); 7385 read_lock(&tasklist_lock); 7386 ret = __rt_schedulable(NULL, 0, 0); 7387 read_unlock(&tasklist_lock); 7388 mutex_unlock(&rt_constraints_mutex); 7389 7390 return ret; 7391} 7392 7393static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7394{ 7395 /* Don't accept realtime tasks when there is no way for them to run */ 7396 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7397 return 0; 7398 7399 return 1; 7400} 7401 7402#else /* !CONFIG_RT_GROUP_SCHED */ 7403static int sched_rt_global_constraints(void) 7404{ 7405 unsigned long flags; 7406 int i, ret = 0; 7407 7408 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7409 for_each_possible_cpu(i) { 7410 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7411 7412 raw_spin_lock(&rt_rq->rt_runtime_lock); 7413 rt_rq->rt_runtime = global_rt_runtime(); 7414 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7415 } 7416 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7417 7418 return ret; 7419} 7420#endif /* CONFIG_RT_GROUP_SCHED */ 7421 7422static int sched_dl_global_constraints(void) 7423{ 7424 u64 runtime = global_rt_runtime(); 7425 u64 period = global_rt_period(); 7426 u64 new_bw = to_ratio(period, runtime); 7427 int cpu, ret = 0; 7428 7429 /* 7430 * Here we want to check the bandwidth not being set to some 7431 * value smaller than the currently allocated bandwidth in 7432 * any of the root_domains. 7433 * 7434 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7435 * cycling on root_domains... Discussion on different/better 7436 * solutions is welcome! 7437 */ 7438 for_each_possible_cpu(cpu) { 7439 struct dl_bw *dl_b = dl_bw_of(cpu); 7440 7441 raw_spin_lock(&dl_b->lock); 7442 if (new_bw < dl_b->total_bw) 7443 ret = -EBUSY; 7444 raw_spin_unlock(&dl_b->lock); 7445 7446 if (ret) 7447 break; 7448 } 7449 7450 return ret; 7451} 7452 7453static void sched_dl_do_global(void) 7454{ 7455 u64 new_bw = -1; 7456 int cpu; 7457 7458 def_dl_bandwidth.dl_period = global_rt_period(); 7459 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7460 7461 if (global_rt_runtime() != RUNTIME_INF) 7462 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7463 7464 /* 7465 * FIXME: As above... 7466 */ 7467 for_each_possible_cpu(cpu) { 7468 struct dl_bw *dl_b = dl_bw_of(cpu); 7469 7470 raw_spin_lock(&dl_b->lock); 7471 dl_b->bw = new_bw; 7472 raw_spin_unlock(&dl_b->lock); 7473 } 7474} 7475 7476static int sched_rt_global_validate(void) 7477{ 7478 if (sysctl_sched_rt_period <= 0) 7479 return -EINVAL; 7480 7481 if (sysctl_sched_rt_runtime > sysctl_sched_rt_period) 7482 return -EINVAL; 7483 7484 return 0; 7485} 7486 7487static void sched_rt_do_global(void) 7488{ 7489 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7490 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7491} 7492 7493int sched_rt_handler(struct ctl_table *table, int write, 7494 void __user *buffer, size_t *lenp, 7495 loff_t *ppos) 7496{ 7497 int old_period, old_runtime; 7498 static DEFINE_MUTEX(mutex); 7499 int ret; 7500 7501 mutex_lock(&mutex); 7502 old_period = sysctl_sched_rt_period; 7503 old_runtime = sysctl_sched_rt_runtime; 7504 7505 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7506 7507 if (!ret && write) { 7508 ret = sched_rt_global_validate(); 7509 if (ret) 7510 goto undo; 7511 7512 ret = sched_rt_global_constraints(); 7513 if (ret) 7514 goto undo; 7515 7516 ret = sched_dl_global_constraints(); 7517 if (ret) 7518 goto undo; 7519 7520 sched_rt_do_global(); 7521 sched_dl_do_global(); 7522 } 7523 if (0) { 7524undo: 7525 sysctl_sched_rt_period = old_period; 7526 sysctl_sched_rt_runtime = old_runtime; 7527 } 7528 mutex_unlock(&mutex); 7529 7530 return ret; 7531} 7532 7533int sched_rr_handler(struct ctl_table *table, int write, 7534 void __user *buffer, size_t *lenp, 7535 loff_t *ppos) 7536{ 7537 int ret; 7538 static DEFINE_MUTEX(mutex); 7539 7540 mutex_lock(&mutex); 7541 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7542 /* make sure that internally we keep jiffies */ 7543 /* also, writing zero resets timeslice to default */ 7544 if (!ret && write) { 7545 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7546 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7547 } 7548 mutex_unlock(&mutex); 7549 return ret; 7550} 7551 7552#ifdef CONFIG_CGROUP_SCHED 7553 7554static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7555{ 7556 return css ? container_of(css, struct task_group, css) : NULL; 7557} 7558 7559static struct cgroup_subsys_state * 7560cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7561{ 7562 struct task_group *parent = css_tg(parent_css); 7563 struct task_group *tg; 7564 7565 if (!parent) { 7566 /* This is early initialization for the top cgroup */ 7567 return &root_task_group.css; 7568 } 7569 7570 tg = sched_create_group(parent); 7571 if (IS_ERR(tg)) 7572 return ERR_PTR(-ENOMEM); 7573 7574 return &tg->css; 7575} 7576 7577static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7578{ 7579 struct task_group *tg = css_tg(css); 7580 struct task_group *parent = css_tg(css_parent(css)); 7581 7582 if (parent) 7583 sched_online_group(tg, parent); 7584 return 0; 7585} 7586 7587static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7588{ 7589 struct task_group *tg = css_tg(css); 7590 7591 sched_destroy_group(tg); 7592} 7593 7594static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7595{ 7596 struct task_group *tg = css_tg(css); 7597 7598 sched_offline_group(tg); 7599} 7600 7601static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7602 struct cgroup_taskset *tset) 7603{ 7604 struct task_struct *task; 7605 7606 cgroup_taskset_for_each(task, css, tset) { 7607#ifdef CONFIG_RT_GROUP_SCHED 7608 if (!sched_rt_can_attach(css_tg(css), task)) 7609 return -EINVAL; 7610#else 7611 /* We don't support RT-tasks being in separate groups */ 7612 if (task->sched_class != &fair_sched_class) 7613 return -EINVAL; 7614#endif 7615 } 7616 return 0; 7617} 7618 7619static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7620 struct cgroup_taskset *tset) 7621{ 7622 struct task_struct *task; 7623 7624 cgroup_taskset_for_each(task, css, tset) 7625 sched_move_task(task); 7626} 7627 7628static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7629 struct cgroup_subsys_state *old_css, 7630 struct task_struct *task) 7631{ 7632 /* 7633 * cgroup_exit() is called in the copy_process() failure path. 7634 * Ignore this case since the task hasn't ran yet, this avoids 7635 * trying to poke a half freed task state from generic code. 7636 */ 7637 if (!(task->flags & PF_EXITING)) 7638 return; 7639 7640 sched_move_task(task); 7641} 7642 7643#ifdef CONFIG_FAIR_GROUP_SCHED 7644static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7645 struct cftype *cftype, u64 shareval) 7646{ 7647 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7648} 7649 7650static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7651 struct cftype *cft) 7652{ 7653 struct task_group *tg = css_tg(css); 7654 7655 return (u64) scale_load_down(tg->shares); 7656} 7657 7658#ifdef CONFIG_CFS_BANDWIDTH 7659static DEFINE_MUTEX(cfs_constraints_mutex); 7660 7661const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7662const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7663 7664static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7665 7666static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7667{ 7668 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7669 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7670 7671 if (tg == &root_task_group) 7672 return -EINVAL; 7673 7674 /* 7675 * Ensure we have at some amount of bandwidth every period. This is 7676 * to prevent reaching a state of large arrears when throttled via 7677 * entity_tick() resulting in prolonged exit starvation. 7678 */ 7679 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7680 return -EINVAL; 7681 7682 /* 7683 * Likewise, bound things on the otherside by preventing insane quota 7684 * periods. This also allows us to normalize in computing quota 7685 * feasibility. 7686 */ 7687 if (period > max_cfs_quota_period) 7688 return -EINVAL; 7689 7690 mutex_lock(&cfs_constraints_mutex); 7691 ret = __cfs_schedulable(tg, period, quota); 7692 if (ret) 7693 goto out_unlock; 7694 7695 runtime_enabled = quota != RUNTIME_INF; 7696 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7697 /* 7698 * If we need to toggle cfs_bandwidth_used, off->on must occur 7699 * before making related changes, and on->off must occur afterwards 7700 */ 7701 if (runtime_enabled && !runtime_was_enabled) 7702 cfs_bandwidth_usage_inc(); 7703 raw_spin_lock_irq(&cfs_b->lock); 7704 cfs_b->period = ns_to_ktime(period); 7705 cfs_b->quota = quota; 7706 7707 __refill_cfs_bandwidth_runtime(cfs_b); 7708 /* restart the period timer (if active) to handle new period expiry */ 7709 if (runtime_enabled && cfs_b->timer_active) { 7710 /* force a reprogram */ 7711 cfs_b->timer_active = 0; 7712 __start_cfs_bandwidth(cfs_b); 7713 } 7714 raw_spin_unlock_irq(&cfs_b->lock); 7715 7716 for_each_possible_cpu(i) { 7717 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7718 struct rq *rq = cfs_rq->rq; 7719 7720 raw_spin_lock_irq(&rq->lock); 7721 cfs_rq->runtime_enabled = runtime_enabled; 7722 cfs_rq->runtime_remaining = 0; 7723 7724 if (cfs_rq->throttled) 7725 unthrottle_cfs_rq(cfs_rq); 7726 raw_spin_unlock_irq(&rq->lock); 7727 } 7728 if (runtime_was_enabled && !runtime_enabled) 7729 cfs_bandwidth_usage_dec(); 7730out_unlock: 7731 mutex_unlock(&cfs_constraints_mutex); 7732 7733 return ret; 7734} 7735 7736int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7737{ 7738 u64 quota, period; 7739 7740 period = ktime_to_ns(tg->cfs_bandwidth.period); 7741 if (cfs_quota_us < 0) 7742 quota = RUNTIME_INF; 7743 else 7744 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7745 7746 return tg_set_cfs_bandwidth(tg, period, quota); 7747} 7748 7749long tg_get_cfs_quota(struct task_group *tg) 7750{ 7751 u64 quota_us; 7752 7753 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7754 return -1; 7755 7756 quota_us = tg->cfs_bandwidth.quota; 7757 do_div(quota_us, NSEC_PER_USEC); 7758 7759 return quota_us; 7760} 7761 7762int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7763{ 7764 u64 quota, period; 7765 7766 period = (u64)cfs_period_us * NSEC_PER_USEC; 7767 quota = tg->cfs_bandwidth.quota; 7768 7769 return tg_set_cfs_bandwidth(tg, period, quota); 7770} 7771 7772long tg_get_cfs_period(struct task_group *tg) 7773{ 7774 u64 cfs_period_us; 7775 7776 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7777 do_div(cfs_period_us, NSEC_PER_USEC); 7778 7779 return cfs_period_us; 7780} 7781 7782static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7783 struct cftype *cft) 7784{ 7785 return tg_get_cfs_quota(css_tg(css)); 7786} 7787 7788static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7789 struct cftype *cftype, s64 cfs_quota_us) 7790{ 7791 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7792} 7793 7794static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7795 struct cftype *cft) 7796{ 7797 return tg_get_cfs_period(css_tg(css)); 7798} 7799 7800static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7801 struct cftype *cftype, u64 cfs_period_us) 7802{ 7803 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7804} 7805 7806struct cfs_schedulable_data { 7807 struct task_group *tg; 7808 u64 period, quota; 7809}; 7810 7811/* 7812 * normalize group quota/period to be quota/max_period 7813 * note: units are usecs 7814 */ 7815static u64 normalize_cfs_quota(struct task_group *tg, 7816 struct cfs_schedulable_data *d) 7817{ 7818 u64 quota, period; 7819 7820 if (tg == d->tg) { 7821 period = d->period; 7822 quota = d->quota; 7823 } else { 7824 period = tg_get_cfs_period(tg); 7825 quota = tg_get_cfs_quota(tg); 7826 } 7827 7828 /* note: these should typically be equivalent */ 7829 if (quota == RUNTIME_INF || quota == -1) 7830 return RUNTIME_INF; 7831 7832 return to_ratio(period, quota); 7833} 7834 7835static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7836{ 7837 struct cfs_schedulable_data *d = data; 7838 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7839 s64 quota = 0, parent_quota = -1; 7840 7841 if (!tg->parent) { 7842 quota = RUNTIME_INF; 7843 } else { 7844 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7845 7846 quota = normalize_cfs_quota(tg, d); 7847 parent_quota = parent_b->hierarchal_quota; 7848 7849 /* 7850 * ensure max(child_quota) <= parent_quota, inherit when no 7851 * limit is set 7852 */ 7853 if (quota == RUNTIME_INF) 7854 quota = parent_quota; 7855 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7856 return -EINVAL; 7857 } 7858 cfs_b->hierarchal_quota = quota; 7859 7860 return 0; 7861} 7862 7863static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7864{ 7865 int ret; 7866 struct cfs_schedulable_data data = { 7867 .tg = tg, 7868 .period = period, 7869 .quota = quota, 7870 }; 7871 7872 if (quota != RUNTIME_INF) { 7873 do_div(data.period, NSEC_PER_USEC); 7874 do_div(data.quota, NSEC_PER_USEC); 7875 } 7876 7877 rcu_read_lock(); 7878 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7879 rcu_read_unlock(); 7880 7881 return ret; 7882} 7883 7884static int cpu_stats_show(struct seq_file *sf, void *v) 7885{ 7886 struct task_group *tg = css_tg(seq_css(sf)); 7887 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7888 7889 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7890 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7891 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7892 7893 return 0; 7894} 7895#endif /* CONFIG_CFS_BANDWIDTH */ 7896#endif /* CONFIG_FAIR_GROUP_SCHED */ 7897 7898#ifdef CONFIG_RT_GROUP_SCHED 7899static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7900 struct cftype *cft, s64 val) 7901{ 7902 return sched_group_set_rt_runtime(css_tg(css), val); 7903} 7904 7905static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7906 struct cftype *cft) 7907{ 7908 return sched_group_rt_runtime(css_tg(css)); 7909} 7910 7911static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7912 struct cftype *cftype, u64 rt_period_us) 7913{ 7914 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7915} 7916 7917static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7918 struct cftype *cft) 7919{ 7920 return sched_group_rt_period(css_tg(css)); 7921} 7922#endif /* CONFIG_RT_GROUP_SCHED */ 7923 7924static struct cftype cpu_files[] = { 7925#ifdef CONFIG_FAIR_GROUP_SCHED 7926 { 7927 .name = "shares", 7928 .read_u64 = cpu_shares_read_u64, 7929 .write_u64 = cpu_shares_write_u64, 7930 }, 7931#endif 7932#ifdef CONFIG_CFS_BANDWIDTH 7933 { 7934 .name = "cfs_quota_us", 7935 .read_s64 = cpu_cfs_quota_read_s64, 7936 .write_s64 = cpu_cfs_quota_write_s64, 7937 }, 7938 { 7939 .name = "cfs_period_us", 7940 .read_u64 = cpu_cfs_period_read_u64, 7941 .write_u64 = cpu_cfs_period_write_u64, 7942 }, 7943 { 7944 .name = "stat", 7945 .seq_show = cpu_stats_show, 7946 }, 7947#endif 7948#ifdef CONFIG_RT_GROUP_SCHED 7949 { 7950 .name = "rt_runtime_us", 7951 .read_s64 = cpu_rt_runtime_read, 7952 .write_s64 = cpu_rt_runtime_write, 7953 }, 7954 { 7955 .name = "rt_period_us", 7956 .read_u64 = cpu_rt_period_read_uint, 7957 .write_u64 = cpu_rt_period_write_uint, 7958 }, 7959#endif 7960 { } /* terminate */ 7961}; 7962 7963struct cgroup_subsys cpu_cgroup_subsys = { 7964 .name = "cpu", 7965 .css_alloc = cpu_cgroup_css_alloc, 7966 .css_free = cpu_cgroup_css_free, 7967 .css_online = cpu_cgroup_css_online, 7968 .css_offline = cpu_cgroup_css_offline, 7969 .can_attach = cpu_cgroup_can_attach, 7970 .attach = cpu_cgroup_attach, 7971 .exit = cpu_cgroup_exit, 7972 .subsys_id = cpu_cgroup_subsys_id, 7973 .base_cftypes = cpu_files, 7974 .early_init = 1, 7975}; 7976 7977#endif /* CONFIG_CGROUP_SCHED */ 7978 7979void dump_cpu_task(int cpu) 7980{ 7981 pr_info("Task dump for CPU %d:\n", cpu); 7982 sched_show_task(cpu_curr(cpu)); 7983} 7984