core.c revision 1724813d9f2c7ff702b46d3e4a4f6d9b10a8f8c2
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 * Maximum bandwidth available for all -deadline tasks and groups 301 * (if group scheduling is configured) on each CPU. 302 * 303 * default: 5% 304 */ 305unsigned int sysctl_sched_dl_period = 1000000; 306int sysctl_sched_dl_runtime = 50000; 307 308 309 310/* 311 * __task_rq_lock - lock the rq @p resides on. 312 */ 313static inline struct rq *__task_rq_lock(struct task_struct *p) 314 __acquires(rq->lock) 315{ 316 struct rq *rq; 317 318 lockdep_assert_held(&p->pi_lock); 319 320 for (;;) { 321 rq = task_rq(p); 322 raw_spin_lock(&rq->lock); 323 if (likely(rq == task_rq(p))) 324 return rq; 325 raw_spin_unlock(&rq->lock); 326 } 327} 328 329/* 330 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 331 */ 332static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 333 __acquires(p->pi_lock) 334 __acquires(rq->lock) 335{ 336 struct rq *rq; 337 338 for (;;) { 339 raw_spin_lock_irqsave(&p->pi_lock, *flags); 340 rq = task_rq(p); 341 raw_spin_lock(&rq->lock); 342 if (likely(rq == task_rq(p))) 343 return rq; 344 raw_spin_unlock(&rq->lock); 345 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 346 } 347} 348 349static void __task_rq_unlock(struct rq *rq) 350 __releases(rq->lock) 351{ 352 raw_spin_unlock(&rq->lock); 353} 354 355static inline void 356task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 357 __releases(rq->lock) 358 __releases(p->pi_lock) 359{ 360 raw_spin_unlock(&rq->lock); 361 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 362} 363 364/* 365 * this_rq_lock - lock this runqueue and disable interrupts. 366 */ 367static struct rq *this_rq_lock(void) 368 __acquires(rq->lock) 369{ 370 struct rq *rq; 371 372 local_irq_disable(); 373 rq = this_rq(); 374 raw_spin_lock(&rq->lock); 375 376 return rq; 377} 378 379#ifdef CONFIG_SCHED_HRTICK 380/* 381 * Use HR-timers to deliver accurate preemption points. 382 */ 383 384static void hrtick_clear(struct rq *rq) 385{ 386 if (hrtimer_active(&rq->hrtick_timer)) 387 hrtimer_cancel(&rq->hrtick_timer); 388} 389 390/* 391 * High-resolution timer tick. 392 * Runs from hardirq context with interrupts disabled. 393 */ 394static enum hrtimer_restart hrtick(struct hrtimer *timer) 395{ 396 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 397 398 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 399 400 raw_spin_lock(&rq->lock); 401 update_rq_clock(rq); 402 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 403 raw_spin_unlock(&rq->lock); 404 405 return HRTIMER_NORESTART; 406} 407 408#ifdef CONFIG_SMP 409 410static int __hrtick_restart(struct rq *rq) 411{ 412 struct hrtimer *timer = &rq->hrtick_timer; 413 ktime_t time = hrtimer_get_softexpires(timer); 414 415 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 416} 417 418/* 419 * called from hardirq (IPI) context 420 */ 421static void __hrtick_start(void *arg) 422{ 423 struct rq *rq = arg; 424 425 raw_spin_lock(&rq->lock); 426 __hrtick_restart(rq); 427 rq->hrtick_csd_pending = 0; 428 raw_spin_unlock(&rq->lock); 429} 430 431/* 432 * Called to set the hrtick timer state. 433 * 434 * called with rq->lock held and irqs disabled 435 */ 436void hrtick_start(struct rq *rq, u64 delay) 437{ 438 struct hrtimer *timer = &rq->hrtick_timer; 439 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 440 441 hrtimer_set_expires(timer, time); 442 443 if (rq == this_rq()) { 444 __hrtick_restart(rq); 445 } else if (!rq->hrtick_csd_pending) { 446 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 447 rq->hrtick_csd_pending = 1; 448 } 449} 450 451static int 452hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 453{ 454 int cpu = (int)(long)hcpu; 455 456 switch (action) { 457 case CPU_UP_CANCELED: 458 case CPU_UP_CANCELED_FROZEN: 459 case CPU_DOWN_PREPARE: 460 case CPU_DOWN_PREPARE_FROZEN: 461 case CPU_DEAD: 462 case CPU_DEAD_FROZEN: 463 hrtick_clear(cpu_rq(cpu)); 464 return NOTIFY_OK; 465 } 466 467 return NOTIFY_DONE; 468} 469 470static __init void init_hrtick(void) 471{ 472 hotcpu_notifier(hotplug_hrtick, 0); 473} 474#else 475/* 476 * Called to set the hrtick timer state. 477 * 478 * called with rq->lock held and irqs disabled 479 */ 480void hrtick_start(struct rq *rq, u64 delay) 481{ 482 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 483 HRTIMER_MODE_REL_PINNED, 0); 484} 485 486static inline void init_hrtick(void) 487{ 488} 489#endif /* CONFIG_SMP */ 490 491static void init_rq_hrtick(struct rq *rq) 492{ 493#ifdef CONFIG_SMP 494 rq->hrtick_csd_pending = 0; 495 496 rq->hrtick_csd.flags = 0; 497 rq->hrtick_csd.func = __hrtick_start; 498 rq->hrtick_csd.info = rq; 499#endif 500 501 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 502 rq->hrtick_timer.function = hrtick; 503} 504#else /* CONFIG_SCHED_HRTICK */ 505static inline void hrtick_clear(struct rq *rq) 506{ 507} 508 509static inline void init_rq_hrtick(struct rq *rq) 510{ 511} 512 513static inline void init_hrtick(void) 514{ 515} 516#endif /* CONFIG_SCHED_HRTICK */ 517 518/* 519 * resched_task - mark a task 'to be rescheduled now'. 520 * 521 * On UP this means the setting of the need_resched flag, on SMP it 522 * might also involve a cross-CPU call to trigger the scheduler on 523 * the target CPU. 524 */ 525void resched_task(struct task_struct *p) 526{ 527 int cpu; 528 529 lockdep_assert_held(&task_rq(p)->lock); 530 531 if (test_tsk_need_resched(p)) 532 return; 533 534 set_tsk_need_resched(p); 535 536 cpu = task_cpu(p); 537 if (cpu == smp_processor_id()) { 538 set_preempt_need_resched(); 539 return; 540 } 541 542 /* NEED_RESCHED must be visible before we test polling */ 543 smp_mb(); 544 if (!tsk_is_polling(p)) 545 smp_send_reschedule(cpu); 546} 547 548void resched_cpu(int cpu) 549{ 550 struct rq *rq = cpu_rq(cpu); 551 unsigned long flags; 552 553 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 554 return; 555 resched_task(cpu_curr(cpu)); 556 raw_spin_unlock_irqrestore(&rq->lock, flags); 557} 558 559#ifdef CONFIG_SMP 560#ifdef CONFIG_NO_HZ_COMMON 561/* 562 * In the semi idle case, use the nearest busy cpu for migrating timers 563 * from an idle cpu. This is good for power-savings. 564 * 565 * We don't do similar optimization for completely idle system, as 566 * selecting an idle cpu will add more delays to the timers than intended 567 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 568 */ 569int get_nohz_timer_target(void) 570{ 571 int cpu = smp_processor_id(); 572 int i; 573 struct sched_domain *sd; 574 575 rcu_read_lock(); 576 for_each_domain(cpu, sd) { 577 for_each_cpu(i, sched_domain_span(sd)) { 578 if (!idle_cpu(i)) { 579 cpu = i; 580 goto unlock; 581 } 582 } 583 } 584unlock: 585 rcu_read_unlock(); 586 return cpu; 587} 588/* 589 * When add_timer_on() enqueues a timer into the timer wheel of an 590 * idle CPU then this timer might expire before the next timer event 591 * which is scheduled to wake up that CPU. In case of a completely 592 * idle system the next event might even be infinite time into the 593 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 594 * leaves the inner idle loop so the newly added timer is taken into 595 * account when the CPU goes back to idle and evaluates the timer 596 * wheel for the next timer event. 597 */ 598static void wake_up_idle_cpu(int cpu) 599{ 600 struct rq *rq = cpu_rq(cpu); 601 602 if (cpu == smp_processor_id()) 603 return; 604 605 /* 606 * This is safe, as this function is called with the timer 607 * wheel base lock of (cpu) held. When the CPU is on the way 608 * to idle and has not yet set rq->curr to idle then it will 609 * be serialized on the timer wheel base lock and take the new 610 * timer into account automatically. 611 */ 612 if (rq->curr != rq->idle) 613 return; 614 615 /* 616 * We can set TIF_RESCHED on the idle task of the other CPU 617 * lockless. The worst case is that the other CPU runs the 618 * idle task through an additional NOOP schedule() 619 */ 620 set_tsk_need_resched(rq->idle); 621 622 /* NEED_RESCHED must be visible before we test polling */ 623 smp_mb(); 624 if (!tsk_is_polling(rq->idle)) 625 smp_send_reschedule(cpu); 626} 627 628static bool wake_up_full_nohz_cpu(int cpu) 629{ 630 if (tick_nohz_full_cpu(cpu)) { 631 if (cpu != smp_processor_id() || 632 tick_nohz_tick_stopped()) 633 smp_send_reschedule(cpu); 634 return true; 635 } 636 637 return false; 638} 639 640void wake_up_nohz_cpu(int cpu) 641{ 642 if (!wake_up_full_nohz_cpu(cpu)) 643 wake_up_idle_cpu(cpu); 644} 645 646static inline bool got_nohz_idle_kick(void) 647{ 648 int cpu = smp_processor_id(); 649 650 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 651 return false; 652 653 if (idle_cpu(cpu) && !need_resched()) 654 return true; 655 656 /* 657 * We can't run Idle Load Balance on this CPU for this time so we 658 * cancel it and clear NOHZ_BALANCE_KICK 659 */ 660 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 661 return false; 662} 663 664#else /* CONFIG_NO_HZ_COMMON */ 665 666static inline bool got_nohz_idle_kick(void) 667{ 668 return false; 669} 670 671#endif /* CONFIG_NO_HZ_COMMON */ 672 673#ifdef CONFIG_NO_HZ_FULL 674bool sched_can_stop_tick(void) 675{ 676 struct rq *rq; 677 678 rq = this_rq(); 679 680 /* Make sure rq->nr_running update is visible after the IPI */ 681 smp_rmb(); 682 683 /* More than one running task need preemption */ 684 if (rq->nr_running > 1) 685 return false; 686 687 return true; 688} 689#endif /* CONFIG_NO_HZ_FULL */ 690 691void sched_avg_update(struct rq *rq) 692{ 693 s64 period = sched_avg_period(); 694 695 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 696 /* 697 * Inline assembly required to prevent the compiler 698 * optimising this loop into a divmod call. 699 * See __iter_div_u64_rem() for another example of this. 700 */ 701 asm("" : "+rm" (rq->age_stamp)); 702 rq->age_stamp += period; 703 rq->rt_avg /= 2; 704 } 705} 706 707#endif /* CONFIG_SMP */ 708 709#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 710 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 711/* 712 * Iterate task_group tree rooted at *from, calling @down when first entering a 713 * node and @up when leaving it for the final time. 714 * 715 * Caller must hold rcu_lock or sufficient equivalent. 716 */ 717int walk_tg_tree_from(struct task_group *from, 718 tg_visitor down, tg_visitor up, void *data) 719{ 720 struct task_group *parent, *child; 721 int ret; 722 723 parent = from; 724 725down: 726 ret = (*down)(parent, data); 727 if (ret) 728 goto out; 729 list_for_each_entry_rcu(child, &parent->children, siblings) { 730 parent = child; 731 goto down; 732 733up: 734 continue; 735 } 736 ret = (*up)(parent, data); 737 if (ret || parent == from) 738 goto out; 739 740 child = parent; 741 parent = parent->parent; 742 if (parent) 743 goto up; 744out: 745 return ret; 746} 747 748int tg_nop(struct task_group *tg, void *data) 749{ 750 return 0; 751} 752#endif 753 754static void set_load_weight(struct task_struct *p) 755{ 756 int prio = p->static_prio - MAX_RT_PRIO; 757 struct load_weight *load = &p->se.load; 758 759 /* 760 * SCHED_IDLE tasks get minimal weight: 761 */ 762 if (p->policy == SCHED_IDLE) { 763 load->weight = scale_load(WEIGHT_IDLEPRIO); 764 load->inv_weight = WMULT_IDLEPRIO; 765 return; 766 } 767 768 load->weight = scale_load(prio_to_weight[prio]); 769 load->inv_weight = prio_to_wmult[prio]; 770} 771 772static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 773{ 774 update_rq_clock(rq); 775 sched_info_queued(rq, p); 776 p->sched_class->enqueue_task(rq, p, flags); 777} 778 779static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 780{ 781 update_rq_clock(rq); 782 sched_info_dequeued(rq, p); 783 p->sched_class->dequeue_task(rq, p, flags); 784} 785 786void activate_task(struct rq *rq, struct task_struct *p, int flags) 787{ 788 if (task_contributes_to_load(p)) 789 rq->nr_uninterruptible--; 790 791 enqueue_task(rq, p, flags); 792} 793 794void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 795{ 796 if (task_contributes_to_load(p)) 797 rq->nr_uninterruptible++; 798 799 dequeue_task(rq, p, flags); 800} 801 802static void update_rq_clock_task(struct rq *rq, s64 delta) 803{ 804/* 805 * In theory, the compile should just see 0 here, and optimize out the call 806 * to sched_rt_avg_update. But I don't trust it... 807 */ 808#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 809 s64 steal = 0, irq_delta = 0; 810#endif 811#ifdef CONFIG_IRQ_TIME_ACCOUNTING 812 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 813 814 /* 815 * Since irq_time is only updated on {soft,}irq_exit, we might run into 816 * this case when a previous update_rq_clock() happened inside a 817 * {soft,}irq region. 818 * 819 * When this happens, we stop ->clock_task and only update the 820 * prev_irq_time stamp to account for the part that fit, so that a next 821 * update will consume the rest. This ensures ->clock_task is 822 * monotonic. 823 * 824 * It does however cause some slight miss-attribution of {soft,}irq 825 * time, a more accurate solution would be to update the irq_time using 826 * the current rq->clock timestamp, except that would require using 827 * atomic ops. 828 */ 829 if (irq_delta > delta) 830 irq_delta = delta; 831 832 rq->prev_irq_time += irq_delta; 833 delta -= irq_delta; 834#endif 835#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 836 if (static_key_false((¶virt_steal_rq_enabled))) { 837 u64 st; 838 839 steal = paravirt_steal_clock(cpu_of(rq)); 840 steal -= rq->prev_steal_time_rq; 841 842 if (unlikely(steal > delta)) 843 steal = delta; 844 845 st = steal_ticks(steal); 846 steal = st * TICK_NSEC; 847 848 rq->prev_steal_time_rq += steal; 849 850 delta -= steal; 851 } 852#endif 853 854 rq->clock_task += delta; 855 856#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 857 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 858 sched_rt_avg_update(rq, irq_delta + steal); 859#endif 860} 861 862void sched_set_stop_task(int cpu, struct task_struct *stop) 863{ 864 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 865 struct task_struct *old_stop = cpu_rq(cpu)->stop; 866 867 if (stop) { 868 /* 869 * Make it appear like a SCHED_FIFO task, its something 870 * userspace knows about and won't get confused about. 871 * 872 * Also, it will make PI more or less work without too 873 * much confusion -- but then, stop work should not 874 * rely on PI working anyway. 875 */ 876 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 877 878 stop->sched_class = &stop_sched_class; 879 } 880 881 cpu_rq(cpu)->stop = stop; 882 883 if (old_stop) { 884 /* 885 * Reset it back to a normal scheduling class so that 886 * it can die in pieces. 887 */ 888 old_stop->sched_class = &rt_sched_class; 889 } 890} 891 892/* 893 * __normal_prio - return the priority that is based on the static prio 894 */ 895static inline int __normal_prio(struct task_struct *p) 896{ 897 return p->static_prio; 898} 899 900/* 901 * Calculate the expected normal priority: i.e. priority 902 * without taking RT-inheritance into account. Might be 903 * boosted by interactivity modifiers. Changes upon fork, 904 * setprio syscalls, and whenever the interactivity 905 * estimator recalculates. 906 */ 907static inline int normal_prio(struct task_struct *p) 908{ 909 int prio; 910 911 if (task_has_dl_policy(p)) 912 prio = MAX_DL_PRIO-1; 913 else if (task_has_rt_policy(p)) 914 prio = MAX_RT_PRIO-1 - p->rt_priority; 915 else 916 prio = __normal_prio(p); 917 return prio; 918} 919 920/* 921 * Calculate the current priority, i.e. the priority 922 * taken into account by the scheduler. This value might 923 * be boosted by RT tasks, or might be boosted by 924 * interactivity modifiers. Will be RT if the task got 925 * RT-boosted. If not then it returns p->normal_prio. 926 */ 927static int effective_prio(struct task_struct *p) 928{ 929 p->normal_prio = normal_prio(p); 930 /* 931 * If we are RT tasks or we were boosted to RT priority, 932 * keep the priority unchanged. Otherwise, update priority 933 * to the normal priority: 934 */ 935 if (!rt_prio(p->prio)) 936 return p->normal_prio; 937 return p->prio; 938} 939 940/** 941 * task_curr - is this task currently executing on a CPU? 942 * @p: the task in question. 943 * 944 * Return: 1 if the task is currently executing. 0 otherwise. 945 */ 946inline int task_curr(const struct task_struct *p) 947{ 948 return cpu_curr(task_cpu(p)) == p; 949} 950 951static inline void check_class_changed(struct rq *rq, struct task_struct *p, 952 const struct sched_class *prev_class, 953 int oldprio) 954{ 955 if (prev_class != p->sched_class) { 956 if (prev_class->switched_from) 957 prev_class->switched_from(rq, p); 958 p->sched_class->switched_to(rq, p); 959 } else if (oldprio != p->prio || dl_task(p)) 960 p->sched_class->prio_changed(rq, p, oldprio); 961} 962 963void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 964{ 965 const struct sched_class *class; 966 967 if (p->sched_class == rq->curr->sched_class) { 968 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 969 } else { 970 for_each_class(class) { 971 if (class == rq->curr->sched_class) 972 break; 973 if (class == p->sched_class) { 974 resched_task(rq->curr); 975 break; 976 } 977 } 978 } 979 980 /* 981 * A queue event has occurred, and we're going to schedule. In 982 * this case, we can save a useless back to back clock update. 983 */ 984 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 985 rq->skip_clock_update = 1; 986} 987 988#ifdef CONFIG_SMP 989void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 990{ 991#ifdef CONFIG_SCHED_DEBUG 992 /* 993 * We should never call set_task_cpu() on a blocked task, 994 * ttwu() will sort out the placement. 995 */ 996 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 997 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 998 999#ifdef CONFIG_LOCKDEP 1000 /* 1001 * The caller should hold either p->pi_lock or rq->lock, when changing 1002 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1003 * 1004 * sched_move_task() holds both and thus holding either pins the cgroup, 1005 * see task_group(). 1006 * 1007 * Furthermore, all task_rq users should acquire both locks, see 1008 * task_rq_lock(). 1009 */ 1010 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1011 lockdep_is_held(&task_rq(p)->lock))); 1012#endif 1013#endif 1014 1015 trace_sched_migrate_task(p, new_cpu); 1016 1017 if (task_cpu(p) != new_cpu) { 1018 if (p->sched_class->migrate_task_rq) 1019 p->sched_class->migrate_task_rq(p, new_cpu); 1020 p->se.nr_migrations++; 1021 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1022 } 1023 1024 __set_task_cpu(p, new_cpu); 1025} 1026 1027static void __migrate_swap_task(struct task_struct *p, int cpu) 1028{ 1029 if (p->on_rq) { 1030 struct rq *src_rq, *dst_rq; 1031 1032 src_rq = task_rq(p); 1033 dst_rq = cpu_rq(cpu); 1034 1035 deactivate_task(src_rq, p, 0); 1036 set_task_cpu(p, cpu); 1037 activate_task(dst_rq, p, 0); 1038 check_preempt_curr(dst_rq, p, 0); 1039 } else { 1040 /* 1041 * Task isn't running anymore; make it appear like we migrated 1042 * it before it went to sleep. This means on wakeup we make the 1043 * previous cpu our targer instead of where it really is. 1044 */ 1045 p->wake_cpu = cpu; 1046 } 1047} 1048 1049struct migration_swap_arg { 1050 struct task_struct *src_task, *dst_task; 1051 int src_cpu, dst_cpu; 1052}; 1053 1054static int migrate_swap_stop(void *data) 1055{ 1056 struct migration_swap_arg *arg = data; 1057 struct rq *src_rq, *dst_rq; 1058 int ret = -EAGAIN; 1059 1060 src_rq = cpu_rq(arg->src_cpu); 1061 dst_rq = cpu_rq(arg->dst_cpu); 1062 1063 double_raw_lock(&arg->src_task->pi_lock, 1064 &arg->dst_task->pi_lock); 1065 double_rq_lock(src_rq, dst_rq); 1066 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1067 goto unlock; 1068 1069 if (task_cpu(arg->src_task) != arg->src_cpu) 1070 goto unlock; 1071 1072 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1073 goto unlock; 1074 1075 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1076 goto unlock; 1077 1078 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1079 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1080 1081 ret = 0; 1082 1083unlock: 1084 double_rq_unlock(src_rq, dst_rq); 1085 raw_spin_unlock(&arg->dst_task->pi_lock); 1086 raw_spin_unlock(&arg->src_task->pi_lock); 1087 1088 return ret; 1089} 1090 1091/* 1092 * Cross migrate two tasks 1093 */ 1094int migrate_swap(struct task_struct *cur, struct task_struct *p) 1095{ 1096 struct migration_swap_arg arg; 1097 int ret = -EINVAL; 1098 1099 arg = (struct migration_swap_arg){ 1100 .src_task = cur, 1101 .src_cpu = task_cpu(cur), 1102 .dst_task = p, 1103 .dst_cpu = task_cpu(p), 1104 }; 1105 1106 if (arg.src_cpu == arg.dst_cpu) 1107 goto out; 1108 1109 /* 1110 * These three tests are all lockless; this is OK since all of them 1111 * will be re-checked with proper locks held further down the line. 1112 */ 1113 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1114 goto out; 1115 1116 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1117 goto out; 1118 1119 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1120 goto out; 1121 1122 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1123 1124out: 1125 return ret; 1126} 1127 1128struct migration_arg { 1129 struct task_struct *task; 1130 int dest_cpu; 1131}; 1132 1133static int migration_cpu_stop(void *data); 1134 1135/* 1136 * wait_task_inactive - wait for a thread to unschedule. 1137 * 1138 * If @match_state is nonzero, it's the @p->state value just checked and 1139 * not expected to change. If it changes, i.e. @p might have woken up, 1140 * then return zero. When we succeed in waiting for @p to be off its CPU, 1141 * we return a positive number (its total switch count). If a second call 1142 * a short while later returns the same number, the caller can be sure that 1143 * @p has remained unscheduled the whole time. 1144 * 1145 * The caller must ensure that the task *will* unschedule sometime soon, 1146 * else this function might spin for a *long* time. This function can't 1147 * be called with interrupts off, or it may introduce deadlock with 1148 * smp_call_function() if an IPI is sent by the same process we are 1149 * waiting to become inactive. 1150 */ 1151unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1152{ 1153 unsigned long flags; 1154 int running, on_rq; 1155 unsigned long ncsw; 1156 struct rq *rq; 1157 1158 for (;;) { 1159 /* 1160 * We do the initial early heuristics without holding 1161 * any task-queue locks at all. We'll only try to get 1162 * the runqueue lock when things look like they will 1163 * work out! 1164 */ 1165 rq = task_rq(p); 1166 1167 /* 1168 * If the task is actively running on another CPU 1169 * still, just relax and busy-wait without holding 1170 * any locks. 1171 * 1172 * NOTE! Since we don't hold any locks, it's not 1173 * even sure that "rq" stays as the right runqueue! 1174 * But we don't care, since "task_running()" will 1175 * return false if the runqueue has changed and p 1176 * is actually now running somewhere else! 1177 */ 1178 while (task_running(rq, p)) { 1179 if (match_state && unlikely(p->state != match_state)) 1180 return 0; 1181 cpu_relax(); 1182 } 1183 1184 /* 1185 * Ok, time to look more closely! We need the rq 1186 * lock now, to be *sure*. If we're wrong, we'll 1187 * just go back and repeat. 1188 */ 1189 rq = task_rq_lock(p, &flags); 1190 trace_sched_wait_task(p); 1191 running = task_running(rq, p); 1192 on_rq = p->on_rq; 1193 ncsw = 0; 1194 if (!match_state || p->state == match_state) 1195 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1196 task_rq_unlock(rq, p, &flags); 1197 1198 /* 1199 * If it changed from the expected state, bail out now. 1200 */ 1201 if (unlikely(!ncsw)) 1202 break; 1203 1204 /* 1205 * Was it really running after all now that we 1206 * checked with the proper locks actually held? 1207 * 1208 * Oops. Go back and try again.. 1209 */ 1210 if (unlikely(running)) { 1211 cpu_relax(); 1212 continue; 1213 } 1214 1215 /* 1216 * It's not enough that it's not actively running, 1217 * it must be off the runqueue _entirely_, and not 1218 * preempted! 1219 * 1220 * So if it was still runnable (but just not actively 1221 * running right now), it's preempted, and we should 1222 * yield - it could be a while. 1223 */ 1224 if (unlikely(on_rq)) { 1225 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1226 1227 set_current_state(TASK_UNINTERRUPTIBLE); 1228 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1229 continue; 1230 } 1231 1232 /* 1233 * Ahh, all good. It wasn't running, and it wasn't 1234 * runnable, which means that it will never become 1235 * running in the future either. We're all done! 1236 */ 1237 break; 1238 } 1239 1240 return ncsw; 1241} 1242 1243/*** 1244 * kick_process - kick a running thread to enter/exit the kernel 1245 * @p: the to-be-kicked thread 1246 * 1247 * Cause a process which is running on another CPU to enter 1248 * kernel-mode, without any delay. (to get signals handled.) 1249 * 1250 * NOTE: this function doesn't have to take the runqueue lock, 1251 * because all it wants to ensure is that the remote task enters 1252 * the kernel. If the IPI races and the task has been migrated 1253 * to another CPU then no harm is done and the purpose has been 1254 * achieved as well. 1255 */ 1256void kick_process(struct task_struct *p) 1257{ 1258 int cpu; 1259 1260 preempt_disable(); 1261 cpu = task_cpu(p); 1262 if ((cpu != smp_processor_id()) && task_curr(p)) 1263 smp_send_reschedule(cpu); 1264 preempt_enable(); 1265} 1266EXPORT_SYMBOL_GPL(kick_process); 1267#endif /* CONFIG_SMP */ 1268 1269#ifdef CONFIG_SMP 1270/* 1271 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1272 */ 1273static int select_fallback_rq(int cpu, struct task_struct *p) 1274{ 1275 int nid = cpu_to_node(cpu); 1276 const struct cpumask *nodemask = NULL; 1277 enum { cpuset, possible, fail } state = cpuset; 1278 int dest_cpu; 1279 1280 /* 1281 * If the node that the cpu is on has been offlined, cpu_to_node() 1282 * will return -1. There is no cpu on the node, and we should 1283 * select the cpu on the other node. 1284 */ 1285 if (nid != -1) { 1286 nodemask = cpumask_of_node(nid); 1287 1288 /* Look for allowed, online CPU in same node. */ 1289 for_each_cpu(dest_cpu, nodemask) { 1290 if (!cpu_online(dest_cpu)) 1291 continue; 1292 if (!cpu_active(dest_cpu)) 1293 continue; 1294 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1295 return dest_cpu; 1296 } 1297 } 1298 1299 for (;;) { 1300 /* Any allowed, online CPU? */ 1301 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1302 if (!cpu_online(dest_cpu)) 1303 continue; 1304 if (!cpu_active(dest_cpu)) 1305 continue; 1306 goto out; 1307 } 1308 1309 switch (state) { 1310 case cpuset: 1311 /* No more Mr. Nice Guy. */ 1312 cpuset_cpus_allowed_fallback(p); 1313 state = possible; 1314 break; 1315 1316 case possible: 1317 do_set_cpus_allowed(p, cpu_possible_mask); 1318 state = fail; 1319 break; 1320 1321 case fail: 1322 BUG(); 1323 break; 1324 } 1325 } 1326 1327out: 1328 if (state != cpuset) { 1329 /* 1330 * Don't tell them about moving exiting tasks or 1331 * kernel threads (both mm NULL), since they never 1332 * leave kernel. 1333 */ 1334 if (p->mm && printk_ratelimit()) { 1335 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1336 task_pid_nr(p), p->comm, cpu); 1337 } 1338 } 1339 1340 return dest_cpu; 1341} 1342 1343/* 1344 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1345 */ 1346static inline 1347int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1348{ 1349 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1350 1351 /* 1352 * In order not to call set_task_cpu() on a blocking task we need 1353 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1354 * cpu. 1355 * 1356 * Since this is common to all placement strategies, this lives here. 1357 * 1358 * [ this allows ->select_task() to simply return task_cpu(p) and 1359 * not worry about this generic constraint ] 1360 */ 1361 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1362 !cpu_online(cpu))) 1363 cpu = select_fallback_rq(task_cpu(p), p); 1364 1365 return cpu; 1366} 1367 1368static void update_avg(u64 *avg, u64 sample) 1369{ 1370 s64 diff = sample - *avg; 1371 *avg += diff >> 3; 1372} 1373#endif 1374 1375static void 1376ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1377{ 1378#ifdef CONFIG_SCHEDSTATS 1379 struct rq *rq = this_rq(); 1380 1381#ifdef CONFIG_SMP 1382 int this_cpu = smp_processor_id(); 1383 1384 if (cpu == this_cpu) { 1385 schedstat_inc(rq, ttwu_local); 1386 schedstat_inc(p, se.statistics.nr_wakeups_local); 1387 } else { 1388 struct sched_domain *sd; 1389 1390 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1391 rcu_read_lock(); 1392 for_each_domain(this_cpu, sd) { 1393 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1394 schedstat_inc(sd, ttwu_wake_remote); 1395 break; 1396 } 1397 } 1398 rcu_read_unlock(); 1399 } 1400 1401 if (wake_flags & WF_MIGRATED) 1402 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1403 1404#endif /* CONFIG_SMP */ 1405 1406 schedstat_inc(rq, ttwu_count); 1407 schedstat_inc(p, se.statistics.nr_wakeups); 1408 1409 if (wake_flags & WF_SYNC) 1410 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1411 1412#endif /* CONFIG_SCHEDSTATS */ 1413} 1414 1415static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1416{ 1417 activate_task(rq, p, en_flags); 1418 p->on_rq = 1; 1419 1420 /* if a worker is waking up, notify workqueue */ 1421 if (p->flags & PF_WQ_WORKER) 1422 wq_worker_waking_up(p, cpu_of(rq)); 1423} 1424 1425/* 1426 * Mark the task runnable and perform wakeup-preemption. 1427 */ 1428static void 1429ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1430{ 1431 check_preempt_curr(rq, p, wake_flags); 1432 trace_sched_wakeup(p, true); 1433 1434 p->state = TASK_RUNNING; 1435#ifdef CONFIG_SMP 1436 if (p->sched_class->task_woken) 1437 p->sched_class->task_woken(rq, p); 1438 1439 if (rq->idle_stamp) { 1440 u64 delta = rq_clock(rq) - rq->idle_stamp; 1441 u64 max = 2*rq->max_idle_balance_cost; 1442 1443 update_avg(&rq->avg_idle, delta); 1444 1445 if (rq->avg_idle > max) 1446 rq->avg_idle = max; 1447 1448 rq->idle_stamp = 0; 1449 } 1450#endif 1451} 1452 1453static void 1454ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1455{ 1456#ifdef CONFIG_SMP 1457 if (p->sched_contributes_to_load) 1458 rq->nr_uninterruptible--; 1459#endif 1460 1461 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1462 ttwu_do_wakeup(rq, p, wake_flags); 1463} 1464 1465/* 1466 * Called in case the task @p isn't fully descheduled from its runqueue, 1467 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1468 * since all we need to do is flip p->state to TASK_RUNNING, since 1469 * the task is still ->on_rq. 1470 */ 1471static int ttwu_remote(struct task_struct *p, int wake_flags) 1472{ 1473 struct rq *rq; 1474 int ret = 0; 1475 1476 rq = __task_rq_lock(p); 1477 if (p->on_rq) { 1478 /* check_preempt_curr() may use rq clock */ 1479 update_rq_clock(rq); 1480 ttwu_do_wakeup(rq, p, wake_flags); 1481 ret = 1; 1482 } 1483 __task_rq_unlock(rq); 1484 1485 return ret; 1486} 1487 1488#ifdef CONFIG_SMP 1489static void sched_ttwu_pending(void) 1490{ 1491 struct rq *rq = this_rq(); 1492 struct llist_node *llist = llist_del_all(&rq->wake_list); 1493 struct task_struct *p; 1494 1495 raw_spin_lock(&rq->lock); 1496 1497 while (llist) { 1498 p = llist_entry(llist, struct task_struct, wake_entry); 1499 llist = llist_next(llist); 1500 ttwu_do_activate(rq, p, 0); 1501 } 1502 1503 raw_spin_unlock(&rq->lock); 1504} 1505 1506void scheduler_ipi(void) 1507{ 1508 /* 1509 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1510 * TIF_NEED_RESCHED remotely (for the first time) will also send 1511 * this IPI. 1512 */ 1513 if (tif_need_resched()) 1514 set_preempt_need_resched(); 1515 1516 if (llist_empty(&this_rq()->wake_list) 1517 && !tick_nohz_full_cpu(smp_processor_id()) 1518 && !got_nohz_idle_kick()) 1519 return; 1520 1521 /* 1522 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1523 * traditionally all their work was done from the interrupt return 1524 * path. Now that we actually do some work, we need to make sure 1525 * we do call them. 1526 * 1527 * Some archs already do call them, luckily irq_enter/exit nest 1528 * properly. 1529 * 1530 * Arguably we should visit all archs and update all handlers, 1531 * however a fair share of IPIs are still resched only so this would 1532 * somewhat pessimize the simple resched case. 1533 */ 1534 irq_enter(); 1535 tick_nohz_full_check(); 1536 sched_ttwu_pending(); 1537 1538 /* 1539 * Check if someone kicked us for doing the nohz idle load balance. 1540 */ 1541 if (unlikely(got_nohz_idle_kick())) { 1542 this_rq()->idle_balance = 1; 1543 raise_softirq_irqoff(SCHED_SOFTIRQ); 1544 } 1545 irq_exit(); 1546} 1547 1548static void ttwu_queue_remote(struct task_struct *p, int cpu) 1549{ 1550 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1551 smp_send_reschedule(cpu); 1552} 1553 1554bool cpus_share_cache(int this_cpu, int that_cpu) 1555{ 1556 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1557} 1558#endif /* CONFIG_SMP */ 1559 1560static void ttwu_queue(struct task_struct *p, int cpu) 1561{ 1562 struct rq *rq = cpu_rq(cpu); 1563 1564#if defined(CONFIG_SMP) 1565 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1566 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1567 ttwu_queue_remote(p, cpu); 1568 return; 1569 } 1570#endif 1571 1572 raw_spin_lock(&rq->lock); 1573 ttwu_do_activate(rq, p, 0); 1574 raw_spin_unlock(&rq->lock); 1575} 1576 1577/** 1578 * try_to_wake_up - wake up a thread 1579 * @p: the thread to be awakened 1580 * @state: the mask of task states that can be woken 1581 * @wake_flags: wake modifier flags (WF_*) 1582 * 1583 * Put it on the run-queue if it's not already there. The "current" 1584 * thread is always on the run-queue (except when the actual 1585 * re-schedule is in progress), and as such you're allowed to do 1586 * the simpler "current->state = TASK_RUNNING" to mark yourself 1587 * runnable without the overhead of this. 1588 * 1589 * Return: %true if @p was woken up, %false if it was already running. 1590 * or @state didn't match @p's state. 1591 */ 1592static int 1593try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1594{ 1595 unsigned long flags; 1596 int cpu, success = 0; 1597 1598 /* 1599 * If we are going to wake up a thread waiting for CONDITION we 1600 * need to ensure that CONDITION=1 done by the caller can not be 1601 * reordered with p->state check below. This pairs with mb() in 1602 * set_current_state() the waiting thread does. 1603 */ 1604 smp_mb__before_spinlock(); 1605 raw_spin_lock_irqsave(&p->pi_lock, flags); 1606 if (!(p->state & state)) 1607 goto out; 1608 1609 success = 1; /* we're going to change ->state */ 1610 cpu = task_cpu(p); 1611 1612 if (p->on_rq && ttwu_remote(p, wake_flags)) 1613 goto stat; 1614 1615#ifdef CONFIG_SMP 1616 /* 1617 * If the owning (remote) cpu is still in the middle of schedule() with 1618 * this task as prev, wait until its done referencing the task. 1619 */ 1620 while (p->on_cpu) 1621 cpu_relax(); 1622 /* 1623 * Pairs with the smp_wmb() in finish_lock_switch(). 1624 */ 1625 smp_rmb(); 1626 1627 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1628 p->state = TASK_WAKING; 1629 1630 if (p->sched_class->task_waking) 1631 p->sched_class->task_waking(p); 1632 1633 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1634 if (task_cpu(p) != cpu) { 1635 wake_flags |= WF_MIGRATED; 1636 set_task_cpu(p, cpu); 1637 } 1638#endif /* CONFIG_SMP */ 1639 1640 ttwu_queue(p, cpu); 1641stat: 1642 ttwu_stat(p, cpu, wake_flags); 1643out: 1644 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1645 1646 return success; 1647} 1648 1649/** 1650 * try_to_wake_up_local - try to wake up a local task with rq lock held 1651 * @p: the thread to be awakened 1652 * 1653 * Put @p on the run-queue if it's not already there. The caller must 1654 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1655 * the current task. 1656 */ 1657static void try_to_wake_up_local(struct task_struct *p) 1658{ 1659 struct rq *rq = task_rq(p); 1660 1661 if (WARN_ON_ONCE(rq != this_rq()) || 1662 WARN_ON_ONCE(p == current)) 1663 return; 1664 1665 lockdep_assert_held(&rq->lock); 1666 1667 if (!raw_spin_trylock(&p->pi_lock)) { 1668 raw_spin_unlock(&rq->lock); 1669 raw_spin_lock(&p->pi_lock); 1670 raw_spin_lock(&rq->lock); 1671 } 1672 1673 if (!(p->state & TASK_NORMAL)) 1674 goto out; 1675 1676 if (!p->on_rq) 1677 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1678 1679 ttwu_do_wakeup(rq, p, 0); 1680 ttwu_stat(p, smp_processor_id(), 0); 1681out: 1682 raw_spin_unlock(&p->pi_lock); 1683} 1684 1685/** 1686 * wake_up_process - Wake up a specific process 1687 * @p: The process to be woken up. 1688 * 1689 * Attempt to wake up the nominated process and move it to the set of runnable 1690 * processes. 1691 * 1692 * Return: 1 if the process was woken up, 0 if it was already running. 1693 * 1694 * It may be assumed that this function implies a write memory barrier before 1695 * changing the task state if and only if any tasks are woken up. 1696 */ 1697int wake_up_process(struct task_struct *p) 1698{ 1699 WARN_ON(task_is_stopped_or_traced(p)); 1700 return try_to_wake_up(p, TASK_NORMAL, 0); 1701} 1702EXPORT_SYMBOL(wake_up_process); 1703 1704int wake_up_state(struct task_struct *p, unsigned int state) 1705{ 1706 return try_to_wake_up(p, state, 0); 1707} 1708 1709/* 1710 * Perform scheduler related setup for a newly forked process p. 1711 * p is forked by current. 1712 * 1713 * __sched_fork() is basic setup used by init_idle() too: 1714 */ 1715static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1716{ 1717 p->on_rq = 0; 1718 1719 p->se.on_rq = 0; 1720 p->se.exec_start = 0; 1721 p->se.sum_exec_runtime = 0; 1722 p->se.prev_sum_exec_runtime = 0; 1723 p->se.nr_migrations = 0; 1724 p->se.vruntime = 0; 1725 INIT_LIST_HEAD(&p->se.group_node); 1726 1727#ifdef CONFIG_SCHEDSTATS 1728 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1729#endif 1730 1731 RB_CLEAR_NODE(&p->dl.rb_node); 1732 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1733 p->dl.dl_runtime = p->dl.runtime = 0; 1734 p->dl.dl_deadline = p->dl.deadline = 0; 1735 p->dl.dl_period = 0; 1736 p->dl.flags = 0; 1737 1738 INIT_LIST_HEAD(&p->rt.run_list); 1739 1740#ifdef CONFIG_PREEMPT_NOTIFIERS 1741 INIT_HLIST_HEAD(&p->preempt_notifiers); 1742#endif 1743 1744#ifdef CONFIG_NUMA_BALANCING 1745 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1746 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1747 p->mm->numa_scan_seq = 0; 1748 } 1749 1750 if (clone_flags & CLONE_VM) 1751 p->numa_preferred_nid = current->numa_preferred_nid; 1752 else 1753 p->numa_preferred_nid = -1; 1754 1755 p->node_stamp = 0ULL; 1756 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1757 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1758 p->numa_work.next = &p->numa_work; 1759 p->numa_faults = NULL; 1760 p->numa_faults_buffer = NULL; 1761 1762 INIT_LIST_HEAD(&p->numa_entry); 1763 p->numa_group = NULL; 1764#endif /* CONFIG_NUMA_BALANCING */ 1765} 1766 1767#ifdef CONFIG_NUMA_BALANCING 1768#ifdef CONFIG_SCHED_DEBUG 1769void set_numabalancing_state(bool enabled) 1770{ 1771 if (enabled) 1772 sched_feat_set("NUMA"); 1773 else 1774 sched_feat_set("NO_NUMA"); 1775} 1776#else 1777__read_mostly bool numabalancing_enabled; 1778 1779void set_numabalancing_state(bool enabled) 1780{ 1781 numabalancing_enabled = enabled; 1782} 1783#endif /* CONFIG_SCHED_DEBUG */ 1784#endif /* CONFIG_NUMA_BALANCING */ 1785 1786/* 1787 * fork()/clone()-time setup: 1788 */ 1789int sched_fork(unsigned long clone_flags, struct task_struct *p) 1790{ 1791 unsigned long flags; 1792 int cpu = get_cpu(); 1793 1794 __sched_fork(clone_flags, p); 1795 /* 1796 * We mark the process as running here. This guarantees that 1797 * nobody will actually run it, and a signal or other external 1798 * event cannot wake it up and insert it on the runqueue either. 1799 */ 1800 p->state = TASK_RUNNING; 1801 1802 /* 1803 * Make sure we do not leak PI boosting priority to the child. 1804 */ 1805 p->prio = current->normal_prio; 1806 1807 /* 1808 * Revert to default priority/policy on fork if requested. 1809 */ 1810 if (unlikely(p->sched_reset_on_fork)) { 1811 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1812 p->policy = SCHED_NORMAL; 1813 p->static_prio = NICE_TO_PRIO(0); 1814 p->rt_priority = 0; 1815 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1816 p->static_prio = NICE_TO_PRIO(0); 1817 1818 p->prio = p->normal_prio = __normal_prio(p); 1819 set_load_weight(p); 1820 1821 /* 1822 * We don't need the reset flag anymore after the fork. It has 1823 * fulfilled its duty: 1824 */ 1825 p->sched_reset_on_fork = 0; 1826 } 1827 1828 if (dl_prio(p->prio)) { 1829 put_cpu(); 1830 return -EAGAIN; 1831 } else if (rt_prio(p->prio)) { 1832 p->sched_class = &rt_sched_class; 1833 } else { 1834 p->sched_class = &fair_sched_class; 1835 } 1836 1837 if (p->sched_class->task_fork) 1838 p->sched_class->task_fork(p); 1839 1840 /* 1841 * The child is not yet in the pid-hash so no cgroup attach races, 1842 * and the cgroup is pinned to this child due to cgroup_fork() 1843 * is ran before sched_fork(). 1844 * 1845 * Silence PROVE_RCU. 1846 */ 1847 raw_spin_lock_irqsave(&p->pi_lock, flags); 1848 set_task_cpu(p, cpu); 1849 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1850 1851#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1852 if (likely(sched_info_on())) 1853 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1854#endif 1855#if defined(CONFIG_SMP) 1856 p->on_cpu = 0; 1857#endif 1858 init_task_preempt_count(p); 1859#ifdef CONFIG_SMP 1860 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1861 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1862#endif 1863 1864 put_cpu(); 1865 return 0; 1866} 1867 1868unsigned long to_ratio(u64 period, u64 runtime) 1869{ 1870 if (runtime == RUNTIME_INF) 1871 return 1ULL << 20; 1872 1873 /* 1874 * Doing this here saves a lot of checks in all 1875 * the calling paths, and returning zero seems 1876 * safe for them anyway. 1877 */ 1878 if (period == 0) 1879 return 0; 1880 1881 return div64_u64(runtime << 20, period); 1882} 1883 1884#ifdef CONFIG_SMP 1885inline struct dl_bw *dl_bw_of(int i) 1886{ 1887 return &cpu_rq(i)->rd->dl_bw; 1888} 1889 1890static inline int __dl_span_weight(struct rq *rq) 1891{ 1892 return cpumask_weight(rq->rd->span); 1893} 1894#else 1895inline struct dl_bw *dl_bw_of(int i) 1896{ 1897 return &cpu_rq(i)->dl.dl_bw; 1898} 1899 1900static inline int __dl_span_weight(struct rq *rq) 1901{ 1902 return 1; 1903} 1904#endif 1905 1906static inline 1907void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1908{ 1909 dl_b->total_bw -= tsk_bw; 1910} 1911 1912static inline 1913void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1914{ 1915 dl_b->total_bw += tsk_bw; 1916} 1917 1918static inline 1919bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1920{ 1921 return dl_b->bw != -1 && 1922 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1923} 1924 1925/* 1926 * We must be sure that accepting a new task (or allowing changing the 1927 * parameters of an existing one) is consistent with the bandwidth 1928 * constraints. If yes, this function also accordingly updates the currently 1929 * allocated bandwidth to reflect the new situation. 1930 * 1931 * This function is called while holding p's rq->lock. 1932 */ 1933static int dl_overflow(struct task_struct *p, int policy, 1934 const struct sched_attr *attr) 1935{ 1936 1937 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1938 u64 period = attr->sched_period; 1939 u64 runtime = attr->sched_runtime; 1940 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1941 int cpus = __dl_span_weight(task_rq(p)); 1942 int err = -1; 1943 1944 if (new_bw == p->dl.dl_bw) 1945 return 0; 1946 1947 /* 1948 * Either if a task, enters, leave, or stays -deadline but changes 1949 * its parameters, we may need to update accordingly the total 1950 * allocated bandwidth of the container. 1951 */ 1952 raw_spin_lock(&dl_b->lock); 1953 if (dl_policy(policy) && !task_has_dl_policy(p) && 1954 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1955 __dl_add(dl_b, new_bw); 1956 err = 0; 1957 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1958 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1959 __dl_clear(dl_b, p->dl.dl_bw); 1960 __dl_add(dl_b, new_bw); 1961 err = 0; 1962 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1963 __dl_clear(dl_b, p->dl.dl_bw); 1964 err = 0; 1965 } 1966 raw_spin_unlock(&dl_b->lock); 1967 1968 return err; 1969} 1970 1971extern void init_dl_bw(struct dl_bw *dl_b); 1972 1973/* 1974 * wake_up_new_task - wake up a newly created task for the first time. 1975 * 1976 * This function will do some initial scheduler statistics housekeeping 1977 * that must be done for every newly created context, then puts the task 1978 * on the runqueue and wakes it. 1979 */ 1980void wake_up_new_task(struct task_struct *p) 1981{ 1982 unsigned long flags; 1983 struct rq *rq; 1984 1985 raw_spin_lock_irqsave(&p->pi_lock, flags); 1986#ifdef CONFIG_SMP 1987 /* 1988 * Fork balancing, do it here and not earlier because: 1989 * - cpus_allowed can change in the fork path 1990 * - any previously selected cpu might disappear through hotplug 1991 */ 1992 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 1993#endif 1994 1995 /* Initialize new task's runnable average */ 1996 init_task_runnable_average(p); 1997 rq = __task_rq_lock(p); 1998 activate_task(rq, p, 0); 1999 p->on_rq = 1; 2000 trace_sched_wakeup_new(p, true); 2001 check_preempt_curr(rq, p, WF_FORK); 2002#ifdef CONFIG_SMP 2003 if (p->sched_class->task_woken) 2004 p->sched_class->task_woken(rq, p); 2005#endif 2006 task_rq_unlock(rq, p, &flags); 2007} 2008 2009#ifdef CONFIG_PREEMPT_NOTIFIERS 2010 2011/** 2012 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2013 * @notifier: notifier struct to register 2014 */ 2015void preempt_notifier_register(struct preempt_notifier *notifier) 2016{ 2017 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2018} 2019EXPORT_SYMBOL_GPL(preempt_notifier_register); 2020 2021/** 2022 * preempt_notifier_unregister - no longer interested in preemption notifications 2023 * @notifier: notifier struct to unregister 2024 * 2025 * This is safe to call from within a preemption notifier. 2026 */ 2027void preempt_notifier_unregister(struct preempt_notifier *notifier) 2028{ 2029 hlist_del(¬ifier->link); 2030} 2031EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2032 2033static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2034{ 2035 struct preempt_notifier *notifier; 2036 2037 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2038 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2039} 2040 2041static void 2042fire_sched_out_preempt_notifiers(struct task_struct *curr, 2043 struct task_struct *next) 2044{ 2045 struct preempt_notifier *notifier; 2046 2047 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2048 notifier->ops->sched_out(notifier, next); 2049} 2050 2051#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2052 2053static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2054{ 2055} 2056 2057static void 2058fire_sched_out_preempt_notifiers(struct task_struct *curr, 2059 struct task_struct *next) 2060{ 2061} 2062 2063#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2064 2065/** 2066 * prepare_task_switch - prepare to switch tasks 2067 * @rq: the runqueue preparing to switch 2068 * @prev: the current task that is being switched out 2069 * @next: the task we are going to switch to. 2070 * 2071 * This is called with the rq lock held and interrupts off. It must 2072 * be paired with a subsequent finish_task_switch after the context 2073 * switch. 2074 * 2075 * prepare_task_switch sets up locking and calls architecture specific 2076 * hooks. 2077 */ 2078static inline void 2079prepare_task_switch(struct rq *rq, struct task_struct *prev, 2080 struct task_struct *next) 2081{ 2082 trace_sched_switch(prev, next); 2083 sched_info_switch(rq, prev, next); 2084 perf_event_task_sched_out(prev, next); 2085 fire_sched_out_preempt_notifiers(prev, next); 2086 prepare_lock_switch(rq, next); 2087 prepare_arch_switch(next); 2088} 2089 2090/** 2091 * finish_task_switch - clean up after a task-switch 2092 * @rq: runqueue associated with task-switch 2093 * @prev: the thread we just switched away from. 2094 * 2095 * finish_task_switch must be called after the context switch, paired 2096 * with a prepare_task_switch call before the context switch. 2097 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2098 * and do any other architecture-specific cleanup actions. 2099 * 2100 * Note that we may have delayed dropping an mm in context_switch(). If 2101 * so, we finish that here outside of the runqueue lock. (Doing it 2102 * with the lock held can cause deadlocks; see schedule() for 2103 * details.) 2104 */ 2105static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2106 __releases(rq->lock) 2107{ 2108 struct mm_struct *mm = rq->prev_mm; 2109 long prev_state; 2110 2111 rq->prev_mm = NULL; 2112 2113 /* 2114 * A task struct has one reference for the use as "current". 2115 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2116 * schedule one last time. The schedule call will never return, and 2117 * the scheduled task must drop that reference. 2118 * The test for TASK_DEAD must occur while the runqueue locks are 2119 * still held, otherwise prev could be scheduled on another cpu, die 2120 * there before we look at prev->state, and then the reference would 2121 * be dropped twice. 2122 * Manfred Spraul <manfred@colorfullife.com> 2123 */ 2124 prev_state = prev->state; 2125 vtime_task_switch(prev); 2126 finish_arch_switch(prev); 2127 perf_event_task_sched_in(prev, current); 2128 finish_lock_switch(rq, prev); 2129 finish_arch_post_lock_switch(); 2130 2131 fire_sched_in_preempt_notifiers(current); 2132 if (mm) 2133 mmdrop(mm); 2134 if (unlikely(prev_state == TASK_DEAD)) { 2135 task_numa_free(prev); 2136 2137 if (prev->sched_class->task_dead) 2138 prev->sched_class->task_dead(prev); 2139 2140 /* 2141 * Remove function-return probe instances associated with this 2142 * task and put them back on the free list. 2143 */ 2144 kprobe_flush_task(prev); 2145 put_task_struct(prev); 2146 } 2147 2148 tick_nohz_task_switch(current); 2149} 2150 2151#ifdef CONFIG_SMP 2152 2153/* assumes rq->lock is held */ 2154static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 2155{ 2156 if (prev->sched_class->pre_schedule) 2157 prev->sched_class->pre_schedule(rq, prev); 2158} 2159 2160/* rq->lock is NOT held, but preemption is disabled */ 2161static inline void post_schedule(struct rq *rq) 2162{ 2163 if (rq->post_schedule) { 2164 unsigned long flags; 2165 2166 raw_spin_lock_irqsave(&rq->lock, flags); 2167 if (rq->curr->sched_class->post_schedule) 2168 rq->curr->sched_class->post_schedule(rq); 2169 raw_spin_unlock_irqrestore(&rq->lock, flags); 2170 2171 rq->post_schedule = 0; 2172 } 2173} 2174 2175#else 2176 2177static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2178{ 2179} 2180 2181static inline void post_schedule(struct rq *rq) 2182{ 2183} 2184 2185#endif 2186 2187/** 2188 * schedule_tail - first thing a freshly forked thread must call. 2189 * @prev: the thread we just switched away from. 2190 */ 2191asmlinkage void schedule_tail(struct task_struct *prev) 2192 __releases(rq->lock) 2193{ 2194 struct rq *rq = this_rq(); 2195 2196 finish_task_switch(rq, prev); 2197 2198 /* 2199 * FIXME: do we need to worry about rq being invalidated by the 2200 * task_switch? 2201 */ 2202 post_schedule(rq); 2203 2204#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2205 /* In this case, finish_task_switch does not reenable preemption */ 2206 preempt_enable(); 2207#endif 2208 if (current->set_child_tid) 2209 put_user(task_pid_vnr(current), current->set_child_tid); 2210} 2211 2212/* 2213 * context_switch - switch to the new MM and the new 2214 * thread's register state. 2215 */ 2216static inline void 2217context_switch(struct rq *rq, struct task_struct *prev, 2218 struct task_struct *next) 2219{ 2220 struct mm_struct *mm, *oldmm; 2221 2222 prepare_task_switch(rq, prev, next); 2223 2224 mm = next->mm; 2225 oldmm = prev->active_mm; 2226 /* 2227 * For paravirt, this is coupled with an exit in switch_to to 2228 * combine the page table reload and the switch backend into 2229 * one hypercall. 2230 */ 2231 arch_start_context_switch(prev); 2232 2233 if (!mm) { 2234 next->active_mm = oldmm; 2235 atomic_inc(&oldmm->mm_count); 2236 enter_lazy_tlb(oldmm, next); 2237 } else 2238 switch_mm(oldmm, mm, next); 2239 2240 if (!prev->mm) { 2241 prev->active_mm = NULL; 2242 rq->prev_mm = oldmm; 2243 } 2244 /* 2245 * Since the runqueue lock will be released by the next 2246 * task (which is an invalid locking op but in the case 2247 * of the scheduler it's an obvious special-case), so we 2248 * do an early lockdep release here: 2249 */ 2250#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2251 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2252#endif 2253 2254 context_tracking_task_switch(prev, next); 2255 /* Here we just switch the register state and the stack. */ 2256 switch_to(prev, next, prev); 2257 2258 barrier(); 2259 /* 2260 * this_rq must be evaluated again because prev may have moved 2261 * CPUs since it called schedule(), thus the 'rq' on its stack 2262 * frame will be invalid. 2263 */ 2264 finish_task_switch(this_rq(), prev); 2265} 2266 2267/* 2268 * nr_running and nr_context_switches: 2269 * 2270 * externally visible scheduler statistics: current number of runnable 2271 * threads, total number of context switches performed since bootup. 2272 */ 2273unsigned long nr_running(void) 2274{ 2275 unsigned long i, sum = 0; 2276 2277 for_each_online_cpu(i) 2278 sum += cpu_rq(i)->nr_running; 2279 2280 return sum; 2281} 2282 2283unsigned long long nr_context_switches(void) 2284{ 2285 int i; 2286 unsigned long long sum = 0; 2287 2288 for_each_possible_cpu(i) 2289 sum += cpu_rq(i)->nr_switches; 2290 2291 return sum; 2292} 2293 2294unsigned long nr_iowait(void) 2295{ 2296 unsigned long i, sum = 0; 2297 2298 for_each_possible_cpu(i) 2299 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2300 2301 return sum; 2302} 2303 2304unsigned long nr_iowait_cpu(int cpu) 2305{ 2306 struct rq *this = cpu_rq(cpu); 2307 return atomic_read(&this->nr_iowait); 2308} 2309 2310#ifdef CONFIG_SMP 2311 2312/* 2313 * sched_exec - execve() is a valuable balancing opportunity, because at 2314 * this point the task has the smallest effective memory and cache footprint. 2315 */ 2316void sched_exec(void) 2317{ 2318 struct task_struct *p = current; 2319 unsigned long flags; 2320 int dest_cpu; 2321 2322 raw_spin_lock_irqsave(&p->pi_lock, flags); 2323 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2324 if (dest_cpu == smp_processor_id()) 2325 goto unlock; 2326 2327 if (likely(cpu_active(dest_cpu))) { 2328 struct migration_arg arg = { p, dest_cpu }; 2329 2330 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2331 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2332 return; 2333 } 2334unlock: 2335 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2336} 2337 2338#endif 2339 2340DEFINE_PER_CPU(struct kernel_stat, kstat); 2341DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2342 2343EXPORT_PER_CPU_SYMBOL(kstat); 2344EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2345 2346/* 2347 * Return any ns on the sched_clock that have not yet been accounted in 2348 * @p in case that task is currently running. 2349 * 2350 * Called with task_rq_lock() held on @rq. 2351 */ 2352static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2353{ 2354 u64 ns = 0; 2355 2356 if (task_current(rq, p)) { 2357 update_rq_clock(rq); 2358 ns = rq_clock_task(rq) - p->se.exec_start; 2359 if ((s64)ns < 0) 2360 ns = 0; 2361 } 2362 2363 return ns; 2364} 2365 2366unsigned long long task_delta_exec(struct task_struct *p) 2367{ 2368 unsigned long flags; 2369 struct rq *rq; 2370 u64 ns = 0; 2371 2372 rq = task_rq_lock(p, &flags); 2373 ns = do_task_delta_exec(p, rq); 2374 task_rq_unlock(rq, p, &flags); 2375 2376 return ns; 2377} 2378 2379/* 2380 * Return accounted runtime for the task. 2381 * In case the task is currently running, return the runtime plus current's 2382 * pending runtime that have not been accounted yet. 2383 */ 2384unsigned long long task_sched_runtime(struct task_struct *p) 2385{ 2386 unsigned long flags; 2387 struct rq *rq; 2388 u64 ns = 0; 2389 2390#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2391 /* 2392 * 64-bit doesn't need locks to atomically read a 64bit value. 2393 * So we have a optimization chance when the task's delta_exec is 0. 2394 * Reading ->on_cpu is racy, but this is ok. 2395 * 2396 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2397 * If we race with it entering cpu, unaccounted time is 0. This is 2398 * indistinguishable from the read occurring a few cycles earlier. 2399 */ 2400 if (!p->on_cpu) 2401 return p->se.sum_exec_runtime; 2402#endif 2403 2404 rq = task_rq_lock(p, &flags); 2405 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2406 task_rq_unlock(rq, p, &flags); 2407 2408 return ns; 2409} 2410 2411/* 2412 * This function gets called by the timer code, with HZ frequency. 2413 * We call it with interrupts disabled. 2414 */ 2415void scheduler_tick(void) 2416{ 2417 int cpu = smp_processor_id(); 2418 struct rq *rq = cpu_rq(cpu); 2419 struct task_struct *curr = rq->curr; 2420 2421 sched_clock_tick(); 2422 2423 raw_spin_lock(&rq->lock); 2424 update_rq_clock(rq); 2425 curr->sched_class->task_tick(rq, curr, 0); 2426 update_cpu_load_active(rq); 2427 raw_spin_unlock(&rq->lock); 2428 2429 perf_event_task_tick(); 2430 2431#ifdef CONFIG_SMP 2432 rq->idle_balance = idle_cpu(cpu); 2433 trigger_load_balance(rq, cpu); 2434#endif 2435 rq_last_tick_reset(rq); 2436} 2437 2438#ifdef CONFIG_NO_HZ_FULL 2439/** 2440 * scheduler_tick_max_deferment 2441 * 2442 * Keep at least one tick per second when a single 2443 * active task is running because the scheduler doesn't 2444 * yet completely support full dynticks environment. 2445 * 2446 * This makes sure that uptime, CFS vruntime, load 2447 * balancing, etc... continue to move forward, even 2448 * with a very low granularity. 2449 * 2450 * Return: Maximum deferment in nanoseconds. 2451 */ 2452u64 scheduler_tick_max_deferment(void) 2453{ 2454 struct rq *rq = this_rq(); 2455 unsigned long next, now = ACCESS_ONCE(jiffies); 2456 2457 next = rq->last_sched_tick + HZ; 2458 2459 if (time_before_eq(next, now)) 2460 return 0; 2461 2462 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2463} 2464#endif 2465 2466notrace unsigned long get_parent_ip(unsigned long addr) 2467{ 2468 if (in_lock_functions(addr)) { 2469 addr = CALLER_ADDR2; 2470 if (in_lock_functions(addr)) 2471 addr = CALLER_ADDR3; 2472 } 2473 return addr; 2474} 2475 2476#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2477 defined(CONFIG_PREEMPT_TRACER)) 2478 2479void __kprobes preempt_count_add(int val) 2480{ 2481#ifdef CONFIG_DEBUG_PREEMPT 2482 /* 2483 * Underflow? 2484 */ 2485 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2486 return; 2487#endif 2488 __preempt_count_add(val); 2489#ifdef CONFIG_DEBUG_PREEMPT 2490 /* 2491 * Spinlock count overflowing soon? 2492 */ 2493 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2494 PREEMPT_MASK - 10); 2495#endif 2496 if (preempt_count() == val) 2497 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2498} 2499EXPORT_SYMBOL(preempt_count_add); 2500 2501void __kprobes preempt_count_sub(int val) 2502{ 2503#ifdef CONFIG_DEBUG_PREEMPT 2504 /* 2505 * Underflow? 2506 */ 2507 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2508 return; 2509 /* 2510 * Is the spinlock portion underflowing? 2511 */ 2512 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2513 !(preempt_count() & PREEMPT_MASK))) 2514 return; 2515#endif 2516 2517 if (preempt_count() == val) 2518 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2519 __preempt_count_sub(val); 2520} 2521EXPORT_SYMBOL(preempt_count_sub); 2522 2523#endif 2524 2525/* 2526 * Print scheduling while atomic bug: 2527 */ 2528static noinline void __schedule_bug(struct task_struct *prev) 2529{ 2530 if (oops_in_progress) 2531 return; 2532 2533 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2534 prev->comm, prev->pid, preempt_count()); 2535 2536 debug_show_held_locks(prev); 2537 print_modules(); 2538 if (irqs_disabled()) 2539 print_irqtrace_events(prev); 2540 dump_stack(); 2541 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2542} 2543 2544/* 2545 * Various schedule()-time debugging checks and statistics: 2546 */ 2547static inline void schedule_debug(struct task_struct *prev) 2548{ 2549 /* 2550 * Test if we are atomic. Since do_exit() needs to call into 2551 * schedule() atomically, we ignore that path. Otherwise whine 2552 * if we are scheduling when we should not. 2553 */ 2554 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2555 __schedule_bug(prev); 2556 rcu_sleep_check(); 2557 2558 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2559 2560 schedstat_inc(this_rq(), sched_count); 2561} 2562 2563static void put_prev_task(struct rq *rq, struct task_struct *prev) 2564{ 2565 if (prev->on_rq || rq->skip_clock_update < 0) 2566 update_rq_clock(rq); 2567 prev->sched_class->put_prev_task(rq, prev); 2568} 2569 2570/* 2571 * Pick up the highest-prio task: 2572 */ 2573static inline struct task_struct * 2574pick_next_task(struct rq *rq) 2575{ 2576 const struct sched_class *class; 2577 struct task_struct *p; 2578 2579 /* 2580 * Optimization: we know that if all tasks are in 2581 * the fair class we can call that function directly: 2582 */ 2583 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2584 p = fair_sched_class.pick_next_task(rq); 2585 if (likely(p)) 2586 return p; 2587 } 2588 2589 for_each_class(class) { 2590 p = class->pick_next_task(rq); 2591 if (p) 2592 return p; 2593 } 2594 2595 BUG(); /* the idle class will always have a runnable task */ 2596} 2597 2598/* 2599 * __schedule() is the main scheduler function. 2600 * 2601 * The main means of driving the scheduler and thus entering this function are: 2602 * 2603 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2604 * 2605 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2606 * paths. For example, see arch/x86/entry_64.S. 2607 * 2608 * To drive preemption between tasks, the scheduler sets the flag in timer 2609 * interrupt handler scheduler_tick(). 2610 * 2611 * 3. Wakeups don't really cause entry into schedule(). They add a 2612 * task to the run-queue and that's it. 2613 * 2614 * Now, if the new task added to the run-queue preempts the current 2615 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2616 * called on the nearest possible occasion: 2617 * 2618 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2619 * 2620 * - in syscall or exception context, at the next outmost 2621 * preempt_enable(). (this might be as soon as the wake_up()'s 2622 * spin_unlock()!) 2623 * 2624 * - in IRQ context, return from interrupt-handler to 2625 * preemptible context 2626 * 2627 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2628 * then at the next: 2629 * 2630 * - cond_resched() call 2631 * - explicit schedule() call 2632 * - return from syscall or exception to user-space 2633 * - return from interrupt-handler to user-space 2634 */ 2635static void __sched __schedule(void) 2636{ 2637 struct task_struct *prev, *next; 2638 unsigned long *switch_count; 2639 struct rq *rq; 2640 int cpu; 2641 2642need_resched: 2643 preempt_disable(); 2644 cpu = smp_processor_id(); 2645 rq = cpu_rq(cpu); 2646 rcu_note_context_switch(cpu); 2647 prev = rq->curr; 2648 2649 schedule_debug(prev); 2650 2651 if (sched_feat(HRTICK)) 2652 hrtick_clear(rq); 2653 2654 /* 2655 * Make sure that signal_pending_state()->signal_pending() below 2656 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2657 * done by the caller to avoid the race with signal_wake_up(). 2658 */ 2659 smp_mb__before_spinlock(); 2660 raw_spin_lock_irq(&rq->lock); 2661 2662 switch_count = &prev->nivcsw; 2663 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2664 if (unlikely(signal_pending_state(prev->state, prev))) { 2665 prev->state = TASK_RUNNING; 2666 } else { 2667 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2668 prev->on_rq = 0; 2669 2670 /* 2671 * If a worker went to sleep, notify and ask workqueue 2672 * whether it wants to wake up a task to maintain 2673 * concurrency. 2674 */ 2675 if (prev->flags & PF_WQ_WORKER) { 2676 struct task_struct *to_wakeup; 2677 2678 to_wakeup = wq_worker_sleeping(prev, cpu); 2679 if (to_wakeup) 2680 try_to_wake_up_local(to_wakeup); 2681 } 2682 } 2683 switch_count = &prev->nvcsw; 2684 } 2685 2686 pre_schedule(rq, prev); 2687 2688 if (unlikely(!rq->nr_running)) 2689 idle_balance(cpu, rq); 2690 2691 put_prev_task(rq, prev); 2692 next = pick_next_task(rq); 2693 clear_tsk_need_resched(prev); 2694 clear_preempt_need_resched(); 2695 rq->skip_clock_update = 0; 2696 2697 if (likely(prev != next)) { 2698 rq->nr_switches++; 2699 rq->curr = next; 2700 ++*switch_count; 2701 2702 context_switch(rq, prev, next); /* unlocks the rq */ 2703 /* 2704 * The context switch have flipped the stack from under us 2705 * and restored the local variables which were saved when 2706 * this task called schedule() in the past. prev == current 2707 * is still correct, but it can be moved to another cpu/rq. 2708 */ 2709 cpu = smp_processor_id(); 2710 rq = cpu_rq(cpu); 2711 } else 2712 raw_spin_unlock_irq(&rq->lock); 2713 2714 post_schedule(rq); 2715 2716 sched_preempt_enable_no_resched(); 2717 if (need_resched()) 2718 goto need_resched; 2719} 2720 2721static inline void sched_submit_work(struct task_struct *tsk) 2722{ 2723 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2724 return; 2725 /* 2726 * If we are going to sleep and we have plugged IO queued, 2727 * make sure to submit it to avoid deadlocks. 2728 */ 2729 if (blk_needs_flush_plug(tsk)) 2730 blk_schedule_flush_plug(tsk); 2731} 2732 2733asmlinkage void __sched schedule(void) 2734{ 2735 struct task_struct *tsk = current; 2736 2737 sched_submit_work(tsk); 2738 __schedule(); 2739} 2740EXPORT_SYMBOL(schedule); 2741 2742#ifdef CONFIG_CONTEXT_TRACKING 2743asmlinkage void __sched schedule_user(void) 2744{ 2745 /* 2746 * If we come here after a random call to set_need_resched(), 2747 * or we have been woken up remotely but the IPI has not yet arrived, 2748 * we haven't yet exited the RCU idle mode. Do it here manually until 2749 * we find a better solution. 2750 */ 2751 user_exit(); 2752 schedule(); 2753 user_enter(); 2754} 2755#endif 2756 2757/** 2758 * schedule_preempt_disabled - called with preemption disabled 2759 * 2760 * Returns with preemption disabled. Note: preempt_count must be 1 2761 */ 2762void __sched schedule_preempt_disabled(void) 2763{ 2764 sched_preempt_enable_no_resched(); 2765 schedule(); 2766 preempt_disable(); 2767} 2768 2769#ifdef CONFIG_PREEMPT 2770/* 2771 * this is the entry point to schedule() from in-kernel preemption 2772 * off of preempt_enable. Kernel preemptions off return from interrupt 2773 * occur there and call schedule directly. 2774 */ 2775asmlinkage void __sched notrace preempt_schedule(void) 2776{ 2777 /* 2778 * If there is a non-zero preempt_count or interrupts are disabled, 2779 * we do not want to preempt the current task. Just return.. 2780 */ 2781 if (likely(!preemptible())) 2782 return; 2783 2784 do { 2785 __preempt_count_add(PREEMPT_ACTIVE); 2786 __schedule(); 2787 __preempt_count_sub(PREEMPT_ACTIVE); 2788 2789 /* 2790 * Check again in case we missed a preemption opportunity 2791 * between schedule and now. 2792 */ 2793 barrier(); 2794 } while (need_resched()); 2795} 2796EXPORT_SYMBOL(preempt_schedule); 2797#endif /* CONFIG_PREEMPT */ 2798 2799/* 2800 * this is the entry point to schedule() from kernel preemption 2801 * off of irq context. 2802 * Note, that this is called and return with irqs disabled. This will 2803 * protect us against recursive calling from irq. 2804 */ 2805asmlinkage void __sched preempt_schedule_irq(void) 2806{ 2807 enum ctx_state prev_state; 2808 2809 /* Catch callers which need to be fixed */ 2810 BUG_ON(preempt_count() || !irqs_disabled()); 2811 2812 prev_state = exception_enter(); 2813 2814 do { 2815 __preempt_count_add(PREEMPT_ACTIVE); 2816 local_irq_enable(); 2817 __schedule(); 2818 local_irq_disable(); 2819 __preempt_count_sub(PREEMPT_ACTIVE); 2820 2821 /* 2822 * Check again in case we missed a preemption opportunity 2823 * between schedule and now. 2824 */ 2825 barrier(); 2826 } while (need_resched()); 2827 2828 exception_exit(prev_state); 2829} 2830 2831int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2832 void *key) 2833{ 2834 return try_to_wake_up(curr->private, mode, wake_flags); 2835} 2836EXPORT_SYMBOL(default_wake_function); 2837 2838static long __sched 2839sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2840{ 2841 unsigned long flags; 2842 wait_queue_t wait; 2843 2844 init_waitqueue_entry(&wait, current); 2845 2846 __set_current_state(state); 2847 2848 spin_lock_irqsave(&q->lock, flags); 2849 __add_wait_queue(q, &wait); 2850 spin_unlock(&q->lock); 2851 timeout = schedule_timeout(timeout); 2852 spin_lock_irq(&q->lock); 2853 __remove_wait_queue(q, &wait); 2854 spin_unlock_irqrestore(&q->lock, flags); 2855 2856 return timeout; 2857} 2858 2859void __sched interruptible_sleep_on(wait_queue_head_t *q) 2860{ 2861 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2862} 2863EXPORT_SYMBOL(interruptible_sleep_on); 2864 2865long __sched 2866interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2867{ 2868 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2869} 2870EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2871 2872void __sched sleep_on(wait_queue_head_t *q) 2873{ 2874 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2875} 2876EXPORT_SYMBOL(sleep_on); 2877 2878long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2879{ 2880 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2881} 2882EXPORT_SYMBOL(sleep_on_timeout); 2883 2884#ifdef CONFIG_RT_MUTEXES 2885 2886/* 2887 * rt_mutex_setprio - set the current priority of a task 2888 * @p: task 2889 * @prio: prio value (kernel-internal form) 2890 * 2891 * This function changes the 'effective' priority of a task. It does 2892 * not touch ->normal_prio like __setscheduler(). 2893 * 2894 * Used by the rt_mutex code to implement priority inheritance logic. 2895 */ 2896void rt_mutex_setprio(struct task_struct *p, int prio) 2897{ 2898 int oldprio, on_rq, running, enqueue_flag = 0; 2899 struct rq *rq; 2900 const struct sched_class *prev_class; 2901 2902 BUG_ON(prio > MAX_PRIO); 2903 2904 rq = __task_rq_lock(p); 2905 2906 /* 2907 * Idle task boosting is a nono in general. There is one 2908 * exception, when PREEMPT_RT and NOHZ is active: 2909 * 2910 * The idle task calls get_next_timer_interrupt() and holds 2911 * the timer wheel base->lock on the CPU and another CPU wants 2912 * to access the timer (probably to cancel it). We can safely 2913 * ignore the boosting request, as the idle CPU runs this code 2914 * with interrupts disabled and will complete the lock 2915 * protected section without being interrupted. So there is no 2916 * real need to boost. 2917 */ 2918 if (unlikely(p == rq->idle)) { 2919 WARN_ON(p != rq->curr); 2920 WARN_ON(p->pi_blocked_on); 2921 goto out_unlock; 2922 } 2923 2924 trace_sched_pi_setprio(p, prio); 2925 p->pi_top_task = rt_mutex_get_top_task(p); 2926 oldprio = p->prio; 2927 prev_class = p->sched_class; 2928 on_rq = p->on_rq; 2929 running = task_current(rq, p); 2930 if (on_rq) 2931 dequeue_task(rq, p, 0); 2932 if (running) 2933 p->sched_class->put_prev_task(rq, p); 2934 2935 /* 2936 * Boosting condition are: 2937 * 1. -rt task is running and holds mutex A 2938 * --> -dl task blocks on mutex A 2939 * 2940 * 2. -dl task is running and holds mutex A 2941 * --> -dl task blocks on mutex A and could preempt the 2942 * running task 2943 */ 2944 if (dl_prio(prio)) { 2945 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2946 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2947 p->dl.dl_boosted = 1; 2948 p->dl.dl_throttled = 0; 2949 enqueue_flag = ENQUEUE_REPLENISH; 2950 } else 2951 p->dl.dl_boosted = 0; 2952 p->sched_class = &dl_sched_class; 2953 } else if (rt_prio(prio)) { 2954 if (dl_prio(oldprio)) 2955 p->dl.dl_boosted = 0; 2956 if (oldprio < prio) 2957 enqueue_flag = ENQUEUE_HEAD; 2958 p->sched_class = &rt_sched_class; 2959 } else { 2960 if (dl_prio(oldprio)) 2961 p->dl.dl_boosted = 0; 2962 p->sched_class = &fair_sched_class; 2963 } 2964 2965 p->prio = prio; 2966 2967 if (running) 2968 p->sched_class->set_curr_task(rq); 2969 if (on_rq) 2970 enqueue_task(rq, p, enqueue_flag); 2971 2972 check_class_changed(rq, p, prev_class, oldprio); 2973out_unlock: 2974 __task_rq_unlock(rq); 2975} 2976#endif 2977 2978void set_user_nice(struct task_struct *p, long nice) 2979{ 2980 int old_prio, delta, on_rq; 2981 unsigned long flags; 2982 struct rq *rq; 2983 2984 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 2985 return; 2986 /* 2987 * We have to be careful, if called from sys_setpriority(), 2988 * the task might be in the middle of scheduling on another CPU. 2989 */ 2990 rq = task_rq_lock(p, &flags); 2991 /* 2992 * The RT priorities are set via sched_setscheduler(), but we still 2993 * allow the 'normal' nice value to be set - but as expected 2994 * it wont have any effect on scheduling until the task is 2995 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 2996 */ 2997 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 2998 p->static_prio = NICE_TO_PRIO(nice); 2999 goto out_unlock; 3000 } 3001 on_rq = p->on_rq; 3002 if (on_rq) 3003 dequeue_task(rq, p, 0); 3004 3005 p->static_prio = NICE_TO_PRIO(nice); 3006 set_load_weight(p); 3007 old_prio = p->prio; 3008 p->prio = effective_prio(p); 3009 delta = p->prio - old_prio; 3010 3011 if (on_rq) { 3012 enqueue_task(rq, p, 0); 3013 /* 3014 * If the task increased its priority or is running and 3015 * lowered its priority, then reschedule its CPU: 3016 */ 3017 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3018 resched_task(rq->curr); 3019 } 3020out_unlock: 3021 task_rq_unlock(rq, p, &flags); 3022} 3023EXPORT_SYMBOL(set_user_nice); 3024 3025/* 3026 * can_nice - check if a task can reduce its nice value 3027 * @p: task 3028 * @nice: nice value 3029 */ 3030int can_nice(const struct task_struct *p, const int nice) 3031{ 3032 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3033 int nice_rlim = 20 - nice; 3034 3035 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3036 capable(CAP_SYS_NICE)); 3037} 3038 3039#ifdef __ARCH_WANT_SYS_NICE 3040 3041/* 3042 * sys_nice - change the priority of the current process. 3043 * @increment: priority increment 3044 * 3045 * sys_setpriority is a more generic, but much slower function that 3046 * does similar things. 3047 */ 3048SYSCALL_DEFINE1(nice, int, increment) 3049{ 3050 long nice, retval; 3051 3052 /* 3053 * Setpriority might change our priority at the same moment. 3054 * We don't have to worry. Conceptually one call occurs first 3055 * and we have a single winner. 3056 */ 3057 if (increment < -40) 3058 increment = -40; 3059 if (increment > 40) 3060 increment = 40; 3061 3062 nice = TASK_NICE(current) + increment; 3063 if (nice < -20) 3064 nice = -20; 3065 if (nice > 19) 3066 nice = 19; 3067 3068 if (increment < 0 && !can_nice(current, nice)) 3069 return -EPERM; 3070 3071 retval = security_task_setnice(current, nice); 3072 if (retval) 3073 return retval; 3074 3075 set_user_nice(current, nice); 3076 return 0; 3077} 3078 3079#endif 3080 3081/** 3082 * task_prio - return the priority value of a given task. 3083 * @p: the task in question. 3084 * 3085 * Return: The priority value as seen by users in /proc. 3086 * RT tasks are offset by -200. Normal tasks are centered 3087 * around 0, value goes from -16 to +15. 3088 */ 3089int task_prio(const struct task_struct *p) 3090{ 3091 return p->prio - MAX_RT_PRIO; 3092} 3093 3094/** 3095 * task_nice - return the nice value of a given task. 3096 * @p: the task in question. 3097 * 3098 * Return: The nice value [ -20 ... 0 ... 19 ]. 3099 */ 3100int task_nice(const struct task_struct *p) 3101{ 3102 return TASK_NICE(p); 3103} 3104EXPORT_SYMBOL(task_nice); 3105 3106/** 3107 * idle_cpu - is a given cpu idle currently? 3108 * @cpu: the processor in question. 3109 * 3110 * Return: 1 if the CPU is currently idle. 0 otherwise. 3111 */ 3112int idle_cpu(int cpu) 3113{ 3114 struct rq *rq = cpu_rq(cpu); 3115 3116 if (rq->curr != rq->idle) 3117 return 0; 3118 3119 if (rq->nr_running) 3120 return 0; 3121 3122#ifdef CONFIG_SMP 3123 if (!llist_empty(&rq->wake_list)) 3124 return 0; 3125#endif 3126 3127 return 1; 3128} 3129 3130/** 3131 * idle_task - return the idle task for a given cpu. 3132 * @cpu: the processor in question. 3133 * 3134 * Return: The idle task for the cpu @cpu. 3135 */ 3136struct task_struct *idle_task(int cpu) 3137{ 3138 return cpu_rq(cpu)->idle; 3139} 3140 3141/** 3142 * find_process_by_pid - find a process with a matching PID value. 3143 * @pid: the pid in question. 3144 * 3145 * The task of @pid, if found. %NULL otherwise. 3146 */ 3147static struct task_struct *find_process_by_pid(pid_t pid) 3148{ 3149 return pid ? find_task_by_vpid(pid) : current; 3150} 3151 3152/* 3153 * This function initializes the sched_dl_entity of a newly becoming 3154 * SCHED_DEADLINE task. 3155 * 3156 * Only the static values are considered here, the actual runtime and the 3157 * absolute deadline will be properly calculated when the task is enqueued 3158 * for the first time with its new policy. 3159 */ 3160static void 3161__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3162{ 3163 struct sched_dl_entity *dl_se = &p->dl; 3164 3165 init_dl_task_timer(dl_se); 3166 dl_se->dl_runtime = attr->sched_runtime; 3167 dl_se->dl_deadline = attr->sched_deadline; 3168 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3169 dl_se->flags = attr->sched_flags; 3170 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3171 dl_se->dl_throttled = 0; 3172 dl_se->dl_new = 1; 3173} 3174 3175/* Actually do priority change: must hold pi & rq lock. */ 3176static void __setscheduler(struct rq *rq, struct task_struct *p, 3177 const struct sched_attr *attr) 3178{ 3179 int policy = attr->sched_policy; 3180 3181 p->policy = policy; 3182 3183 if (dl_policy(policy)) 3184 __setparam_dl(p, attr); 3185 else if (rt_policy(policy)) 3186 p->rt_priority = attr->sched_priority; 3187 else 3188 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3189 3190 p->normal_prio = normal_prio(p); 3191 p->prio = rt_mutex_getprio(p); 3192 3193 if (dl_prio(p->prio)) 3194 p->sched_class = &dl_sched_class; 3195 else if (rt_prio(p->prio)) 3196 p->sched_class = &rt_sched_class; 3197 else 3198 p->sched_class = &fair_sched_class; 3199 3200 set_load_weight(p); 3201} 3202 3203static void 3204__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3205{ 3206 struct sched_dl_entity *dl_se = &p->dl; 3207 3208 attr->sched_priority = p->rt_priority; 3209 attr->sched_runtime = dl_se->dl_runtime; 3210 attr->sched_deadline = dl_se->dl_deadline; 3211 attr->sched_period = dl_se->dl_period; 3212 attr->sched_flags = dl_se->flags; 3213} 3214 3215/* 3216 * This function validates the new parameters of a -deadline task. 3217 * We ask for the deadline not being zero, and greater or equal 3218 * than the runtime, as well as the period of being zero or 3219 * greater than deadline. Furthermore, we have to be sure that 3220 * user parameters are above the internal resolution (1us); we 3221 * check sched_runtime only since it is always the smaller one. 3222 */ 3223static bool 3224__checkparam_dl(const struct sched_attr *attr) 3225{ 3226 return attr && attr->sched_deadline != 0 && 3227 (attr->sched_period == 0 || 3228 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3229 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3230 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3231} 3232 3233/* 3234 * check the target process has a UID that matches the current process's 3235 */ 3236static bool check_same_owner(struct task_struct *p) 3237{ 3238 const struct cred *cred = current_cred(), *pcred; 3239 bool match; 3240 3241 rcu_read_lock(); 3242 pcred = __task_cred(p); 3243 match = (uid_eq(cred->euid, pcred->euid) || 3244 uid_eq(cred->euid, pcred->uid)); 3245 rcu_read_unlock(); 3246 return match; 3247} 3248 3249static int __sched_setscheduler(struct task_struct *p, 3250 const struct sched_attr *attr, 3251 bool user) 3252{ 3253 int retval, oldprio, oldpolicy = -1, on_rq, running; 3254 int policy = attr->sched_policy; 3255 unsigned long flags; 3256 const struct sched_class *prev_class; 3257 struct rq *rq; 3258 int reset_on_fork; 3259 3260 /* may grab non-irq protected spin_locks */ 3261 BUG_ON(in_interrupt()); 3262recheck: 3263 /* double check policy once rq lock held */ 3264 if (policy < 0) { 3265 reset_on_fork = p->sched_reset_on_fork; 3266 policy = oldpolicy = p->policy; 3267 } else { 3268 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3269 policy &= ~SCHED_RESET_ON_FORK; 3270 3271 if (policy != SCHED_DEADLINE && 3272 policy != SCHED_FIFO && policy != SCHED_RR && 3273 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3274 policy != SCHED_IDLE) 3275 return -EINVAL; 3276 } 3277 3278 /* 3279 * Valid priorities for SCHED_FIFO and SCHED_RR are 3280 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3281 * SCHED_BATCH and SCHED_IDLE is 0. 3282 */ 3283 if (attr->sched_priority < 0 || 3284 (p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3285 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3286 return -EINVAL; 3287 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3288 (rt_policy(policy) != (attr->sched_priority != 0))) 3289 return -EINVAL; 3290 3291 /* 3292 * Allow unprivileged RT tasks to decrease priority: 3293 */ 3294 if (user && !capable(CAP_SYS_NICE)) { 3295 if (fair_policy(policy)) { 3296 if (!can_nice(p, attr->sched_nice)) 3297 return -EPERM; 3298 } 3299 3300 if (rt_policy(policy)) { 3301 unsigned long rlim_rtprio = 3302 task_rlimit(p, RLIMIT_RTPRIO); 3303 3304 /* can't set/change the rt policy */ 3305 if (policy != p->policy && !rlim_rtprio) 3306 return -EPERM; 3307 3308 /* can't increase priority */ 3309 if (attr->sched_priority > p->rt_priority && 3310 attr->sched_priority > rlim_rtprio) 3311 return -EPERM; 3312 } 3313 3314 /* 3315 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3316 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3317 */ 3318 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3319 if (!can_nice(p, TASK_NICE(p))) 3320 return -EPERM; 3321 } 3322 3323 /* can't change other user's priorities */ 3324 if (!check_same_owner(p)) 3325 return -EPERM; 3326 3327 /* Normal users shall not reset the sched_reset_on_fork flag */ 3328 if (p->sched_reset_on_fork && !reset_on_fork) 3329 return -EPERM; 3330 } 3331 3332 if (user) { 3333 retval = security_task_setscheduler(p); 3334 if (retval) 3335 return retval; 3336 } 3337 3338 /* 3339 * make sure no PI-waiters arrive (or leave) while we are 3340 * changing the priority of the task: 3341 * 3342 * To be able to change p->policy safely, the appropriate 3343 * runqueue lock must be held. 3344 */ 3345 rq = task_rq_lock(p, &flags); 3346 3347 /* 3348 * Changing the policy of the stop threads its a very bad idea 3349 */ 3350 if (p == rq->stop) { 3351 task_rq_unlock(rq, p, &flags); 3352 return -EINVAL; 3353 } 3354 3355 /* 3356 * If not changing anything there's no need to proceed further: 3357 */ 3358 if (unlikely(policy == p->policy)) { 3359 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p)) 3360 goto change; 3361 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3362 goto change; 3363 if (dl_policy(policy)) 3364 goto change; 3365 3366 task_rq_unlock(rq, p, &flags); 3367 return 0; 3368 } 3369change: 3370 3371 if (user) { 3372#ifdef CONFIG_RT_GROUP_SCHED 3373 /* 3374 * Do not allow realtime tasks into groups that have no runtime 3375 * assigned. 3376 */ 3377 if (rt_bandwidth_enabled() && rt_policy(policy) && 3378 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3379 !task_group_is_autogroup(task_group(p))) { 3380 task_rq_unlock(rq, p, &flags); 3381 return -EPERM; 3382 } 3383#endif 3384#ifdef CONFIG_SMP 3385 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3386 cpumask_t *span = rq->rd->span; 3387 3388 /* 3389 * Don't allow tasks with an affinity mask smaller than 3390 * the entire root_domain to become SCHED_DEADLINE. We 3391 * will also fail if there's no bandwidth available. 3392 */ 3393 if (!cpumask_subset(span, &p->cpus_allowed) || 3394 rq->rd->dl_bw.bw == 0) { 3395 task_rq_unlock(rq, p, &flags); 3396 return -EPERM; 3397 } 3398 } 3399#endif 3400 } 3401 3402 /* recheck policy now with rq lock held */ 3403 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3404 policy = oldpolicy = -1; 3405 task_rq_unlock(rq, p, &flags); 3406 goto recheck; 3407 } 3408 3409 /* 3410 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3411 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3412 * is available. 3413 */ 3414 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3415 task_rq_unlock(rq, p, &flags); 3416 return -EBUSY; 3417 } 3418 3419 on_rq = p->on_rq; 3420 running = task_current(rq, p); 3421 if (on_rq) 3422 dequeue_task(rq, p, 0); 3423 if (running) 3424 p->sched_class->put_prev_task(rq, p); 3425 3426 p->sched_reset_on_fork = reset_on_fork; 3427 3428 oldprio = p->prio; 3429 prev_class = p->sched_class; 3430 __setscheduler(rq, p, attr); 3431 3432 if (running) 3433 p->sched_class->set_curr_task(rq); 3434 if (on_rq) 3435 enqueue_task(rq, p, 0); 3436 3437 check_class_changed(rq, p, prev_class, oldprio); 3438 task_rq_unlock(rq, p, &flags); 3439 3440 rt_mutex_adjust_pi(p); 3441 3442 return 0; 3443} 3444 3445/** 3446 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3447 * @p: the task in question. 3448 * @policy: new policy. 3449 * @param: structure containing the new RT priority. 3450 * 3451 * Return: 0 on success. An error code otherwise. 3452 * 3453 * NOTE that the task may be already dead. 3454 */ 3455int sched_setscheduler(struct task_struct *p, int policy, 3456 const struct sched_param *param) 3457{ 3458 struct sched_attr attr = { 3459 .sched_policy = policy, 3460 .sched_priority = param->sched_priority 3461 }; 3462 return __sched_setscheduler(p, &attr, true); 3463} 3464EXPORT_SYMBOL_GPL(sched_setscheduler); 3465 3466int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3467{ 3468 return __sched_setscheduler(p, attr, true); 3469} 3470EXPORT_SYMBOL_GPL(sched_setattr); 3471 3472/** 3473 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3474 * @p: the task in question. 3475 * @policy: new policy. 3476 * @param: structure containing the new RT priority. 3477 * 3478 * Just like sched_setscheduler, only don't bother checking if the 3479 * current context has permission. For example, this is needed in 3480 * stop_machine(): we create temporary high priority worker threads, 3481 * but our caller might not have that capability. 3482 * 3483 * Return: 0 on success. An error code otherwise. 3484 */ 3485int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3486 const struct sched_param *param) 3487{ 3488 struct sched_attr attr = { 3489 .sched_policy = policy, 3490 .sched_priority = param->sched_priority 3491 }; 3492 return __sched_setscheduler(p, &attr, false); 3493} 3494 3495static int 3496do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3497{ 3498 struct sched_param lparam; 3499 struct task_struct *p; 3500 int retval; 3501 3502 if (!param || pid < 0) 3503 return -EINVAL; 3504 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3505 return -EFAULT; 3506 3507 rcu_read_lock(); 3508 retval = -ESRCH; 3509 p = find_process_by_pid(pid); 3510 if (p != NULL) 3511 retval = sched_setscheduler(p, policy, &lparam); 3512 rcu_read_unlock(); 3513 3514 return retval; 3515} 3516 3517/* 3518 * Mimics kernel/events/core.c perf_copy_attr(). 3519 */ 3520static int sched_copy_attr(struct sched_attr __user *uattr, 3521 struct sched_attr *attr) 3522{ 3523 u32 size; 3524 int ret; 3525 3526 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3527 return -EFAULT; 3528 3529 /* 3530 * zero the full structure, so that a short copy will be nice. 3531 */ 3532 memset(attr, 0, sizeof(*attr)); 3533 3534 ret = get_user(size, &uattr->size); 3535 if (ret) 3536 return ret; 3537 3538 if (size > PAGE_SIZE) /* silly large */ 3539 goto err_size; 3540 3541 if (!size) /* abi compat */ 3542 size = SCHED_ATTR_SIZE_VER0; 3543 3544 if (size < SCHED_ATTR_SIZE_VER0) 3545 goto err_size; 3546 3547 /* 3548 * If we're handed a bigger struct than we know of, 3549 * ensure all the unknown bits are 0 - i.e. new 3550 * user-space does not rely on any kernel feature 3551 * extensions we dont know about yet. 3552 */ 3553 if (size > sizeof(*attr)) { 3554 unsigned char __user *addr; 3555 unsigned char __user *end; 3556 unsigned char val; 3557 3558 addr = (void __user *)uattr + sizeof(*attr); 3559 end = (void __user *)uattr + size; 3560 3561 for (; addr < end; addr++) { 3562 ret = get_user(val, addr); 3563 if (ret) 3564 return ret; 3565 if (val) 3566 goto err_size; 3567 } 3568 size = sizeof(*attr); 3569 } 3570 3571 ret = copy_from_user(attr, uattr, size); 3572 if (ret) 3573 return -EFAULT; 3574 3575 /* 3576 * XXX: do we want to be lenient like existing syscalls; or do we want 3577 * to be strict and return an error on out-of-bounds values? 3578 */ 3579 attr->sched_nice = clamp(attr->sched_nice, -20, 19); 3580 3581out: 3582 return ret; 3583 3584err_size: 3585 put_user(sizeof(*attr), &uattr->size); 3586 ret = -E2BIG; 3587 goto out; 3588} 3589 3590/** 3591 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3592 * @pid: the pid in question. 3593 * @policy: new policy. 3594 * @param: structure containing the new RT priority. 3595 * 3596 * Return: 0 on success. An error code otherwise. 3597 */ 3598SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3599 struct sched_param __user *, param) 3600{ 3601 /* negative values for policy are not valid */ 3602 if (policy < 0) 3603 return -EINVAL; 3604 3605 return do_sched_setscheduler(pid, policy, param); 3606} 3607 3608/** 3609 * sys_sched_setparam - set/change the RT priority of a thread 3610 * @pid: the pid in question. 3611 * @param: structure containing the new RT priority. 3612 * 3613 * Return: 0 on success. An error code otherwise. 3614 */ 3615SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3616{ 3617 return do_sched_setscheduler(pid, -1, param); 3618} 3619 3620/** 3621 * sys_sched_setattr - same as above, but with extended sched_attr 3622 * @pid: the pid in question. 3623 * @attr: structure containing the extended parameters. 3624 */ 3625SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr) 3626{ 3627 struct sched_attr attr; 3628 struct task_struct *p; 3629 int retval; 3630 3631 if (!uattr || pid < 0) 3632 return -EINVAL; 3633 3634 if (sched_copy_attr(uattr, &attr)) 3635 return -EFAULT; 3636 3637 rcu_read_lock(); 3638 retval = -ESRCH; 3639 p = find_process_by_pid(pid); 3640 if (p != NULL) 3641 retval = sched_setattr(p, &attr); 3642 rcu_read_unlock(); 3643 3644 return retval; 3645} 3646 3647/** 3648 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3649 * @pid: the pid in question. 3650 * 3651 * Return: On success, the policy of the thread. Otherwise, a negative error 3652 * code. 3653 */ 3654SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3655{ 3656 struct task_struct *p; 3657 int retval; 3658 3659 if (pid < 0) 3660 return -EINVAL; 3661 3662 retval = -ESRCH; 3663 rcu_read_lock(); 3664 p = find_process_by_pid(pid); 3665 if (p) { 3666 retval = security_task_getscheduler(p); 3667 if (!retval) 3668 retval = p->policy 3669 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3670 } 3671 rcu_read_unlock(); 3672 return retval; 3673} 3674 3675/** 3676 * sys_sched_getparam - get the RT priority of a thread 3677 * @pid: the pid in question. 3678 * @param: structure containing the RT priority. 3679 * 3680 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3681 * code. 3682 */ 3683SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3684{ 3685 struct sched_param lp; 3686 struct task_struct *p; 3687 int retval; 3688 3689 if (!param || pid < 0) 3690 return -EINVAL; 3691 3692 rcu_read_lock(); 3693 p = find_process_by_pid(pid); 3694 retval = -ESRCH; 3695 if (!p) 3696 goto out_unlock; 3697 3698 retval = security_task_getscheduler(p); 3699 if (retval) 3700 goto out_unlock; 3701 3702 if (task_has_dl_policy(p)) { 3703 retval = -EINVAL; 3704 goto out_unlock; 3705 } 3706 lp.sched_priority = p->rt_priority; 3707 rcu_read_unlock(); 3708 3709 /* 3710 * This one might sleep, we cannot do it with a spinlock held ... 3711 */ 3712 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3713 3714 return retval; 3715 3716out_unlock: 3717 rcu_read_unlock(); 3718 return retval; 3719} 3720 3721static int sched_read_attr(struct sched_attr __user *uattr, 3722 struct sched_attr *attr, 3723 unsigned int usize) 3724{ 3725 int ret; 3726 3727 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3728 return -EFAULT; 3729 3730 /* 3731 * If we're handed a smaller struct than we know of, 3732 * ensure all the unknown bits are 0 - i.e. old 3733 * user-space does not get uncomplete information. 3734 */ 3735 if (usize < sizeof(*attr)) { 3736 unsigned char *addr; 3737 unsigned char *end; 3738 3739 addr = (void *)attr + usize; 3740 end = (void *)attr + sizeof(*attr); 3741 3742 for (; addr < end; addr++) { 3743 if (*addr) 3744 goto err_size; 3745 } 3746 3747 attr->size = usize; 3748 } 3749 3750 ret = copy_to_user(uattr, attr, usize); 3751 if (ret) 3752 return -EFAULT; 3753 3754out: 3755 return ret; 3756 3757err_size: 3758 ret = -E2BIG; 3759 goto out; 3760} 3761 3762/** 3763 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3764 * @pid: the pid in question. 3765 * @attr: structure containing the extended parameters. 3766 * @size: sizeof(attr) for fwd/bwd comp. 3767 */ 3768SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3769 unsigned int, size) 3770{ 3771 struct sched_attr attr = { 3772 .size = sizeof(struct sched_attr), 3773 }; 3774 struct task_struct *p; 3775 int retval; 3776 3777 if (!uattr || pid < 0 || size > PAGE_SIZE || 3778 size < SCHED_ATTR_SIZE_VER0) 3779 return -EINVAL; 3780 3781 rcu_read_lock(); 3782 p = find_process_by_pid(pid); 3783 retval = -ESRCH; 3784 if (!p) 3785 goto out_unlock; 3786 3787 retval = security_task_getscheduler(p); 3788 if (retval) 3789 goto out_unlock; 3790 3791 attr.sched_policy = p->policy; 3792 if (task_has_dl_policy(p)) 3793 __getparam_dl(p, &attr); 3794 else if (task_has_rt_policy(p)) 3795 attr.sched_priority = p->rt_priority; 3796 else 3797 attr.sched_nice = TASK_NICE(p); 3798 3799 rcu_read_unlock(); 3800 3801 retval = sched_read_attr(uattr, &attr, size); 3802 return retval; 3803 3804out_unlock: 3805 rcu_read_unlock(); 3806 return retval; 3807} 3808 3809long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3810{ 3811 cpumask_var_t cpus_allowed, new_mask; 3812 struct task_struct *p; 3813 int retval; 3814 3815 rcu_read_lock(); 3816 3817 p = find_process_by_pid(pid); 3818 if (!p) { 3819 rcu_read_unlock(); 3820 return -ESRCH; 3821 } 3822 3823 /* Prevent p going away */ 3824 get_task_struct(p); 3825 rcu_read_unlock(); 3826 3827 if (p->flags & PF_NO_SETAFFINITY) { 3828 retval = -EINVAL; 3829 goto out_put_task; 3830 } 3831 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3832 retval = -ENOMEM; 3833 goto out_put_task; 3834 } 3835 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3836 retval = -ENOMEM; 3837 goto out_free_cpus_allowed; 3838 } 3839 retval = -EPERM; 3840 if (!check_same_owner(p)) { 3841 rcu_read_lock(); 3842 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3843 rcu_read_unlock(); 3844 goto out_unlock; 3845 } 3846 rcu_read_unlock(); 3847 } 3848 3849 retval = security_task_setscheduler(p); 3850 if (retval) 3851 goto out_unlock; 3852 3853 3854 cpuset_cpus_allowed(p, cpus_allowed); 3855 cpumask_and(new_mask, in_mask, cpus_allowed); 3856 3857 /* 3858 * Since bandwidth control happens on root_domain basis, 3859 * if admission test is enabled, we only admit -deadline 3860 * tasks allowed to run on all the CPUs in the task's 3861 * root_domain. 3862 */ 3863#ifdef CONFIG_SMP 3864 if (task_has_dl_policy(p)) { 3865 const struct cpumask *span = task_rq(p)->rd->span; 3866 3867 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3868 retval = -EBUSY; 3869 goto out_unlock; 3870 } 3871 } 3872#endif 3873again: 3874 retval = set_cpus_allowed_ptr(p, new_mask); 3875 3876 if (!retval) { 3877 cpuset_cpus_allowed(p, cpus_allowed); 3878 if (!cpumask_subset(new_mask, cpus_allowed)) { 3879 /* 3880 * We must have raced with a concurrent cpuset 3881 * update. Just reset the cpus_allowed to the 3882 * cpuset's cpus_allowed 3883 */ 3884 cpumask_copy(new_mask, cpus_allowed); 3885 goto again; 3886 } 3887 } 3888out_unlock: 3889 free_cpumask_var(new_mask); 3890out_free_cpus_allowed: 3891 free_cpumask_var(cpus_allowed); 3892out_put_task: 3893 put_task_struct(p); 3894 return retval; 3895} 3896 3897static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3898 struct cpumask *new_mask) 3899{ 3900 if (len < cpumask_size()) 3901 cpumask_clear(new_mask); 3902 else if (len > cpumask_size()) 3903 len = cpumask_size(); 3904 3905 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3906} 3907 3908/** 3909 * sys_sched_setaffinity - set the cpu affinity of a process 3910 * @pid: pid of the process 3911 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3912 * @user_mask_ptr: user-space pointer to the new cpu mask 3913 * 3914 * Return: 0 on success. An error code otherwise. 3915 */ 3916SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3917 unsigned long __user *, user_mask_ptr) 3918{ 3919 cpumask_var_t new_mask; 3920 int retval; 3921 3922 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3923 return -ENOMEM; 3924 3925 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3926 if (retval == 0) 3927 retval = sched_setaffinity(pid, new_mask); 3928 free_cpumask_var(new_mask); 3929 return retval; 3930} 3931 3932long sched_getaffinity(pid_t pid, struct cpumask *mask) 3933{ 3934 struct task_struct *p; 3935 unsigned long flags; 3936 int retval; 3937 3938 rcu_read_lock(); 3939 3940 retval = -ESRCH; 3941 p = find_process_by_pid(pid); 3942 if (!p) 3943 goto out_unlock; 3944 3945 retval = security_task_getscheduler(p); 3946 if (retval) 3947 goto out_unlock; 3948 3949 raw_spin_lock_irqsave(&p->pi_lock, flags); 3950 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3951 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3952 3953out_unlock: 3954 rcu_read_unlock(); 3955 3956 return retval; 3957} 3958 3959/** 3960 * sys_sched_getaffinity - get the cpu affinity of a process 3961 * @pid: pid of the process 3962 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3963 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3964 * 3965 * Return: 0 on success. An error code otherwise. 3966 */ 3967SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3968 unsigned long __user *, user_mask_ptr) 3969{ 3970 int ret; 3971 cpumask_var_t mask; 3972 3973 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3974 return -EINVAL; 3975 if (len & (sizeof(unsigned long)-1)) 3976 return -EINVAL; 3977 3978 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3979 return -ENOMEM; 3980 3981 ret = sched_getaffinity(pid, mask); 3982 if (ret == 0) { 3983 size_t retlen = min_t(size_t, len, cpumask_size()); 3984 3985 if (copy_to_user(user_mask_ptr, mask, retlen)) 3986 ret = -EFAULT; 3987 else 3988 ret = retlen; 3989 } 3990 free_cpumask_var(mask); 3991 3992 return ret; 3993} 3994 3995/** 3996 * sys_sched_yield - yield the current processor to other threads. 3997 * 3998 * This function yields the current CPU to other tasks. If there are no 3999 * other threads running on this CPU then this function will return. 4000 * 4001 * Return: 0. 4002 */ 4003SYSCALL_DEFINE0(sched_yield) 4004{ 4005 struct rq *rq = this_rq_lock(); 4006 4007 schedstat_inc(rq, yld_count); 4008 current->sched_class->yield_task(rq); 4009 4010 /* 4011 * Since we are going to call schedule() anyway, there's 4012 * no need to preempt or enable interrupts: 4013 */ 4014 __release(rq->lock); 4015 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4016 do_raw_spin_unlock(&rq->lock); 4017 sched_preempt_enable_no_resched(); 4018 4019 schedule(); 4020 4021 return 0; 4022} 4023 4024static void __cond_resched(void) 4025{ 4026 __preempt_count_add(PREEMPT_ACTIVE); 4027 __schedule(); 4028 __preempt_count_sub(PREEMPT_ACTIVE); 4029} 4030 4031int __sched _cond_resched(void) 4032{ 4033 if (should_resched()) { 4034 __cond_resched(); 4035 return 1; 4036 } 4037 return 0; 4038} 4039EXPORT_SYMBOL(_cond_resched); 4040 4041/* 4042 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4043 * call schedule, and on return reacquire the lock. 4044 * 4045 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4046 * operations here to prevent schedule() from being called twice (once via 4047 * spin_unlock(), once by hand). 4048 */ 4049int __cond_resched_lock(spinlock_t *lock) 4050{ 4051 int resched = should_resched(); 4052 int ret = 0; 4053 4054 lockdep_assert_held(lock); 4055 4056 if (spin_needbreak(lock) || resched) { 4057 spin_unlock(lock); 4058 if (resched) 4059 __cond_resched(); 4060 else 4061 cpu_relax(); 4062 ret = 1; 4063 spin_lock(lock); 4064 } 4065 return ret; 4066} 4067EXPORT_SYMBOL(__cond_resched_lock); 4068 4069int __sched __cond_resched_softirq(void) 4070{ 4071 BUG_ON(!in_softirq()); 4072 4073 if (should_resched()) { 4074 local_bh_enable(); 4075 __cond_resched(); 4076 local_bh_disable(); 4077 return 1; 4078 } 4079 return 0; 4080} 4081EXPORT_SYMBOL(__cond_resched_softirq); 4082 4083/** 4084 * yield - yield the current processor to other threads. 4085 * 4086 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4087 * 4088 * The scheduler is at all times free to pick the calling task as the most 4089 * eligible task to run, if removing the yield() call from your code breaks 4090 * it, its already broken. 4091 * 4092 * Typical broken usage is: 4093 * 4094 * while (!event) 4095 * yield(); 4096 * 4097 * where one assumes that yield() will let 'the other' process run that will 4098 * make event true. If the current task is a SCHED_FIFO task that will never 4099 * happen. Never use yield() as a progress guarantee!! 4100 * 4101 * If you want to use yield() to wait for something, use wait_event(). 4102 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4103 * If you still want to use yield(), do not! 4104 */ 4105void __sched yield(void) 4106{ 4107 set_current_state(TASK_RUNNING); 4108 sys_sched_yield(); 4109} 4110EXPORT_SYMBOL(yield); 4111 4112/** 4113 * yield_to - yield the current processor to another thread in 4114 * your thread group, or accelerate that thread toward the 4115 * processor it's on. 4116 * @p: target task 4117 * @preempt: whether task preemption is allowed or not 4118 * 4119 * It's the caller's job to ensure that the target task struct 4120 * can't go away on us before we can do any checks. 4121 * 4122 * Return: 4123 * true (>0) if we indeed boosted the target task. 4124 * false (0) if we failed to boost the target. 4125 * -ESRCH if there's no task to yield to. 4126 */ 4127bool __sched yield_to(struct task_struct *p, bool preempt) 4128{ 4129 struct task_struct *curr = current; 4130 struct rq *rq, *p_rq; 4131 unsigned long flags; 4132 int yielded = 0; 4133 4134 local_irq_save(flags); 4135 rq = this_rq(); 4136 4137again: 4138 p_rq = task_rq(p); 4139 /* 4140 * If we're the only runnable task on the rq and target rq also 4141 * has only one task, there's absolutely no point in yielding. 4142 */ 4143 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4144 yielded = -ESRCH; 4145 goto out_irq; 4146 } 4147 4148 double_rq_lock(rq, p_rq); 4149 if (task_rq(p) != p_rq) { 4150 double_rq_unlock(rq, p_rq); 4151 goto again; 4152 } 4153 4154 if (!curr->sched_class->yield_to_task) 4155 goto out_unlock; 4156 4157 if (curr->sched_class != p->sched_class) 4158 goto out_unlock; 4159 4160 if (task_running(p_rq, p) || p->state) 4161 goto out_unlock; 4162 4163 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4164 if (yielded) { 4165 schedstat_inc(rq, yld_count); 4166 /* 4167 * Make p's CPU reschedule; pick_next_entity takes care of 4168 * fairness. 4169 */ 4170 if (preempt && rq != p_rq) 4171 resched_task(p_rq->curr); 4172 } 4173 4174out_unlock: 4175 double_rq_unlock(rq, p_rq); 4176out_irq: 4177 local_irq_restore(flags); 4178 4179 if (yielded > 0) 4180 schedule(); 4181 4182 return yielded; 4183} 4184EXPORT_SYMBOL_GPL(yield_to); 4185 4186/* 4187 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4188 * that process accounting knows that this is a task in IO wait state. 4189 */ 4190void __sched io_schedule(void) 4191{ 4192 struct rq *rq = raw_rq(); 4193 4194 delayacct_blkio_start(); 4195 atomic_inc(&rq->nr_iowait); 4196 blk_flush_plug(current); 4197 current->in_iowait = 1; 4198 schedule(); 4199 current->in_iowait = 0; 4200 atomic_dec(&rq->nr_iowait); 4201 delayacct_blkio_end(); 4202} 4203EXPORT_SYMBOL(io_schedule); 4204 4205long __sched io_schedule_timeout(long timeout) 4206{ 4207 struct rq *rq = raw_rq(); 4208 long ret; 4209 4210 delayacct_blkio_start(); 4211 atomic_inc(&rq->nr_iowait); 4212 blk_flush_plug(current); 4213 current->in_iowait = 1; 4214 ret = schedule_timeout(timeout); 4215 current->in_iowait = 0; 4216 atomic_dec(&rq->nr_iowait); 4217 delayacct_blkio_end(); 4218 return ret; 4219} 4220 4221/** 4222 * sys_sched_get_priority_max - return maximum RT priority. 4223 * @policy: scheduling class. 4224 * 4225 * Return: On success, this syscall returns the maximum 4226 * rt_priority that can be used by a given scheduling class. 4227 * On failure, a negative error code is returned. 4228 */ 4229SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4230{ 4231 int ret = -EINVAL; 4232 4233 switch (policy) { 4234 case SCHED_FIFO: 4235 case SCHED_RR: 4236 ret = MAX_USER_RT_PRIO-1; 4237 break; 4238 case SCHED_DEADLINE: 4239 case SCHED_NORMAL: 4240 case SCHED_BATCH: 4241 case SCHED_IDLE: 4242 ret = 0; 4243 break; 4244 } 4245 return ret; 4246} 4247 4248/** 4249 * sys_sched_get_priority_min - return minimum RT priority. 4250 * @policy: scheduling class. 4251 * 4252 * Return: On success, this syscall returns the minimum 4253 * rt_priority that can be used by a given scheduling class. 4254 * On failure, a negative error code is returned. 4255 */ 4256SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4257{ 4258 int ret = -EINVAL; 4259 4260 switch (policy) { 4261 case SCHED_FIFO: 4262 case SCHED_RR: 4263 ret = 1; 4264 break; 4265 case SCHED_DEADLINE: 4266 case SCHED_NORMAL: 4267 case SCHED_BATCH: 4268 case SCHED_IDLE: 4269 ret = 0; 4270 } 4271 return ret; 4272} 4273 4274/** 4275 * sys_sched_rr_get_interval - return the default timeslice of a process. 4276 * @pid: pid of the process. 4277 * @interval: userspace pointer to the timeslice value. 4278 * 4279 * this syscall writes the default timeslice value of a given process 4280 * into the user-space timespec buffer. A value of '0' means infinity. 4281 * 4282 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4283 * an error code. 4284 */ 4285SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4286 struct timespec __user *, interval) 4287{ 4288 struct task_struct *p; 4289 unsigned int time_slice; 4290 unsigned long flags; 4291 struct rq *rq; 4292 int retval; 4293 struct timespec t; 4294 4295 if (pid < 0) 4296 return -EINVAL; 4297 4298 retval = -ESRCH; 4299 rcu_read_lock(); 4300 p = find_process_by_pid(pid); 4301 if (!p) 4302 goto out_unlock; 4303 4304 retval = security_task_getscheduler(p); 4305 if (retval) 4306 goto out_unlock; 4307 4308 rq = task_rq_lock(p, &flags); 4309 time_slice = p->sched_class->get_rr_interval(rq, p); 4310 task_rq_unlock(rq, p, &flags); 4311 4312 rcu_read_unlock(); 4313 jiffies_to_timespec(time_slice, &t); 4314 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4315 return retval; 4316 4317out_unlock: 4318 rcu_read_unlock(); 4319 return retval; 4320} 4321 4322static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4323 4324void sched_show_task(struct task_struct *p) 4325{ 4326 unsigned long free = 0; 4327 int ppid; 4328 unsigned state; 4329 4330 state = p->state ? __ffs(p->state) + 1 : 0; 4331 printk(KERN_INFO "%-15.15s %c", p->comm, 4332 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4333#if BITS_PER_LONG == 32 4334 if (state == TASK_RUNNING) 4335 printk(KERN_CONT " running "); 4336 else 4337 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4338#else 4339 if (state == TASK_RUNNING) 4340 printk(KERN_CONT " running task "); 4341 else 4342 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4343#endif 4344#ifdef CONFIG_DEBUG_STACK_USAGE 4345 free = stack_not_used(p); 4346#endif 4347 rcu_read_lock(); 4348 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4349 rcu_read_unlock(); 4350 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4351 task_pid_nr(p), ppid, 4352 (unsigned long)task_thread_info(p)->flags); 4353 4354 print_worker_info(KERN_INFO, p); 4355 show_stack(p, NULL); 4356} 4357 4358void show_state_filter(unsigned long state_filter) 4359{ 4360 struct task_struct *g, *p; 4361 4362#if BITS_PER_LONG == 32 4363 printk(KERN_INFO 4364 " task PC stack pid father\n"); 4365#else 4366 printk(KERN_INFO 4367 " task PC stack pid father\n"); 4368#endif 4369 rcu_read_lock(); 4370 do_each_thread(g, p) { 4371 /* 4372 * reset the NMI-timeout, listing all files on a slow 4373 * console might take a lot of time: 4374 */ 4375 touch_nmi_watchdog(); 4376 if (!state_filter || (p->state & state_filter)) 4377 sched_show_task(p); 4378 } while_each_thread(g, p); 4379 4380 touch_all_softlockup_watchdogs(); 4381 4382#ifdef CONFIG_SCHED_DEBUG 4383 sysrq_sched_debug_show(); 4384#endif 4385 rcu_read_unlock(); 4386 /* 4387 * Only show locks if all tasks are dumped: 4388 */ 4389 if (!state_filter) 4390 debug_show_all_locks(); 4391} 4392 4393void init_idle_bootup_task(struct task_struct *idle) 4394{ 4395 idle->sched_class = &idle_sched_class; 4396} 4397 4398/** 4399 * init_idle - set up an idle thread for a given CPU 4400 * @idle: task in question 4401 * @cpu: cpu the idle task belongs to 4402 * 4403 * NOTE: this function does not set the idle thread's NEED_RESCHED 4404 * flag, to make booting more robust. 4405 */ 4406void init_idle(struct task_struct *idle, int cpu) 4407{ 4408 struct rq *rq = cpu_rq(cpu); 4409 unsigned long flags; 4410 4411 raw_spin_lock_irqsave(&rq->lock, flags); 4412 4413 __sched_fork(0, idle); 4414 idle->state = TASK_RUNNING; 4415 idle->se.exec_start = sched_clock(); 4416 4417 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4418 /* 4419 * We're having a chicken and egg problem, even though we are 4420 * holding rq->lock, the cpu isn't yet set to this cpu so the 4421 * lockdep check in task_group() will fail. 4422 * 4423 * Similar case to sched_fork(). / Alternatively we could 4424 * use task_rq_lock() here and obtain the other rq->lock. 4425 * 4426 * Silence PROVE_RCU 4427 */ 4428 rcu_read_lock(); 4429 __set_task_cpu(idle, cpu); 4430 rcu_read_unlock(); 4431 4432 rq->curr = rq->idle = idle; 4433#if defined(CONFIG_SMP) 4434 idle->on_cpu = 1; 4435#endif 4436 raw_spin_unlock_irqrestore(&rq->lock, flags); 4437 4438 /* Set the preempt count _outside_ the spinlocks! */ 4439 init_idle_preempt_count(idle, cpu); 4440 4441 /* 4442 * The idle tasks have their own, simple scheduling class: 4443 */ 4444 idle->sched_class = &idle_sched_class; 4445 ftrace_graph_init_idle_task(idle, cpu); 4446 vtime_init_idle(idle, cpu); 4447#if defined(CONFIG_SMP) 4448 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4449#endif 4450} 4451 4452#ifdef CONFIG_SMP 4453void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4454{ 4455 if (p->sched_class && p->sched_class->set_cpus_allowed) 4456 p->sched_class->set_cpus_allowed(p, new_mask); 4457 4458 cpumask_copy(&p->cpus_allowed, new_mask); 4459 p->nr_cpus_allowed = cpumask_weight(new_mask); 4460} 4461 4462/* 4463 * This is how migration works: 4464 * 4465 * 1) we invoke migration_cpu_stop() on the target CPU using 4466 * stop_one_cpu(). 4467 * 2) stopper starts to run (implicitly forcing the migrated thread 4468 * off the CPU) 4469 * 3) it checks whether the migrated task is still in the wrong runqueue. 4470 * 4) if it's in the wrong runqueue then the migration thread removes 4471 * it and puts it into the right queue. 4472 * 5) stopper completes and stop_one_cpu() returns and the migration 4473 * is done. 4474 */ 4475 4476/* 4477 * Change a given task's CPU affinity. Migrate the thread to a 4478 * proper CPU and schedule it away if the CPU it's executing on 4479 * is removed from the allowed bitmask. 4480 * 4481 * NOTE: the caller must have a valid reference to the task, the 4482 * task must not exit() & deallocate itself prematurely. The 4483 * call is not atomic; no spinlocks may be held. 4484 */ 4485int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4486{ 4487 unsigned long flags; 4488 struct rq *rq; 4489 unsigned int dest_cpu; 4490 int ret = 0; 4491 4492 rq = task_rq_lock(p, &flags); 4493 4494 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4495 goto out; 4496 4497 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4498 ret = -EINVAL; 4499 goto out; 4500 } 4501 4502 do_set_cpus_allowed(p, new_mask); 4503 4504 /* Can the task run on the task's current CPU? If so, we're done */ 4505 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4506 goto out; 4507 4508 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4509 if (p->on_rq) { 4510 struct migration_arg arg = { p, dest_cpu }; 4511 /* Need help from migration thread: drop lock and wait. */ 4512 task_rq_unlock(rq, p, &flags); 4513 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4514 tlb_migrate_finish(p->mm); 4515 return 0; 4516 } 4517out: 4518 task_rq_unlock(rq, p, &flags); 4519 4520 return ret; 4521} 4522EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4523 4524/* 4525 * When dealing with a -deadline task, we have to check if moving it to 4526 * a new CPU is possible or not. In fact, this is only true iff there 4527 * is enough bandwidth available on such CPU, otherwise we want the 4528 * whole migration procedure to fail over. 4529 */ 4530static inline 4531bool set_task_cpu_dl(struct task_struct *p, unsigned int cpu) 4532{ 4533 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 4534 struct dl_bw *cpu_b = dl_bw_of(cpu); 4535 int ret = 1; 4536 u64 bw; 4537 4538 if (dl_b == cpu_b) 4539 return 1; 4540 4541 raw_spin_lock(&dl_b->lock); 4542 raw_spin_lock(&cpu_b->lock); 4543 4544 bw = cpu_b->bw * cpumask_weight(cpu_rq(cpu)->rd->span); 4545 if (dl_bandwidth_enabled() && 4546 bw < cpu_b->total_bw + p->dl.dl_bw) { 4547 ret = 0; 4548 goto unlock; 4549 } 4550 dl_b->total_bw -= p->dl.dl_bw; 4551 cpu_b->total_bw += p->dl.dl_bw; 4552 4553unlock: 4554 raw_spin_unlock(&cpu_b->lock); 4555 raw_spin_unlock(&dl_b->lock); 4556 4557 return ret; 4558} 4559 4560/* 4561 * Move (not current) task off this cpu, onto dest cpu. We're doing 4562 * this because either it can't run here any more (set_cpus_allowed() 4563 * away from this CPU, or CPU going down), or because we're 4564 * attempting to rebalance this task on exec (sched_exec). 4565 * 4566 * So we race with normal scheduler movements, but that's OK, as long 4567 * as the task is no longer on this CPU. 4568 * 4569 * Returns non-zero if task was successfully migrated. 4570 */ 4571static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4572{ 4573 struct rq *rq_dest, *rq_src; 4574 int ret = 0; 4575 4576 if (unlikely(!cpu_active(dest_cpu))) 4577 return ret; 4578 4579 rq_src = cpu_rq(src_cpu); 4580 rq_dest = cpu_rq(dest_cpu); 4581 4582 raw_spin_lock(&p->pi_lock); 4583 double_rq_lock(rq_src, rq_dest); 4584 /* Already moved. */ 4585 if (task_cpu(p) != src_cpu) 4586 goto done; 4587 /* Affinity changed (again). */ 4588 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4589 goto fail; 4590 4591 /* 4592 * If p is -deadline, proceed only if there is enough 4593 * bandwidth available on dest_cpu 4594 */ 4595 if (unlikely(dl_task(p)) && !set_task_cpu_dl(p, dest_cpu)) 4596 goto fail; 4597 4598 /* 4599 * If we're not on a rq, the next wake-up will ensure we're 4600 * placed properly. 4601 */ 4602 if (p->on_rq) { 4603 dequeue_task(rq_src, p, 0); 4604 set_task_cpu(p, dest_cpu); 4605 enqueue_task(rq_dest, p, 0); 4606 check_preempt_curr(rq_dest, p, 0); 4607 } 4608done: 4609 ret = 1; 4610fail: 4611 double_rq_unlock(rq_src, rq_dest); 4612 raw_spin_unlock(&p->pi_lock); 4613 return ret; 4614} 4615 4616#ifdef CONFIG_NUMA_BALANCING 4617/* Migrate current task p to target_cpu */ 4618int migrate_task_to(struct task_struct *p, int target_cpu) 4619{ 4620 struct migration_arg arg = { p, target_cpu }; 4621 int curr_cpu = task_cpu(p); 4622 4623 if (curr_cpu == target_cpu) 4624 return 0; 4625 4626 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4627 return -EINVAL; 4628 4629 /* TODO: This is not properly updating schedstats */ 4630 4631 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4632} 4633 4634/* 4635 * Requeue a task on a given node and accurately track the number of NUMA 4636 * tasks on the runqueues 4637 */ 4638void sched_setnuma(struct task_struct *p, int nid) 4639{ 4640 struct rq *rq; 4641 unsigned long flags; 4642 bool on_rq, running; 4643 4644 rq = task_rq_lock(p, &flags); 4645 on_rq = p->on_rq; 4646 running = task_current(rq, p); 4647 4648 if (on_rq) 4649 dequeue_task(rq, p, 0); 4650 if (running) 4651 p->sched_class->put_prev_task(rq, p); 4652 4653 p->numa_preferred_nid = nid; 4654 4655 if (running) 4656 p->sched_class->set_curr_task(rq); 4657 if (on_rq) 4658 enqueue_task(rq, p, 0); 4659 task_rq_unlock(rq, p, &flags); 4660} 4661#endif 4662 4663/* 4664 * migration_cpu_stop - this will be executed by a highprio stopper thread 4665 * and performs thread migration by bumping thread off CPU then 4666 * 'pushing' onto another runqueue. 4667 */ 4668static int migration_cpu_stop(void *data) 4669{ 4670 struct migration_arg *arg = data; 4671 4672 /* 4673 * The original target cpu might have gone down and we might 4674 * be on another cpu but it doesn't matter. 4675 */ 4676 local_irq_disable(); 4677 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4678 local_irq_enable(); 4679 return 0; 4680} 4681 4682#ifdef CONFIG_HOTPLUG_CPU 4683 4684/* 4685 * Ensures that the idle task is using init_mm right before its cpu goes 4686 * offline. 4687 */ 4688void idle_task_exit(void) 4689{ 4690 struct mm_struct *mm = current->active_mm; 4691 4692 BUG_ON(cpu_online(smp_processor_id())); 4693 4694 if (mm != &init_mm) 4695 switch_mm(mm, &init_mm, current); 4696 mmdrop(mm); 4697} 4698 4699/* 4700 * Since this CPU is going 'away' for a while, fold any nr_active delta 4701 * we might have. Assumes we're called after migrate_tasks() so that the 4702 * nr_active count is stable. 4703 * 4704 * Also see the comment "Global load-average calculations". 4705 */ 4706static void calc_load_migrate(struct rq *rq) 4707{ 4708 long delta = calc_load_fold_active(rq); 4709 if (delta) 4710 atomic_long_add(delta, &calc_load_tasks); 4711} 4712 4713/* 4714 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4715 * try_to_wake_up()->select_task_rq(). 4716 * 4717 * Called with rq->lock held even though we'er in stop_machine() and 4718 * there's no concurrency possible, we hold the required locks anyway 4719 * because of lock validation efforts. 4720 */ 4721static void migrate_tasks(unsigned int dead_cpu) 4722{ 4723 struct rq *rq = cpu_rq(dead_cpu); 4724 struct task_struct *next, *stop = rq->stop; 4725 int dest_cpu; 4726 4727 /* 4728 * Fudge the rq selection such that the below task selection loop 4729 * doesn't get stuck on the currently eligible stop task. 4730 * 4731 * We're currently inside stop_machine() and the rq is either stuck 4732 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4733 * either way we should never end up calling schedule() until we're 4734 * done here. 4735 */ 4736 rq->stop = NULL; 4737 4738 /* 4739 * put_prev_task() and pick_next_task() sched 4740 * class method both need to have an up-to-date 4741 * value of rq->clock[_task] 4742 */ 4743 update_rq_clock(rq); 4744 4745 for ( ; ; ) { 4746 /* 4747 * There's this thread running, bail when that's the only 4748 * remaining thread. 4749 */ 4750 if (rq->nr_running == 1) 4751 break; 4752 4753 next = pick_next_task(rq); 4754 BUG_ON(!next); 4755 next->sched_class->put_prev_task(rq, next); 4756 4757 /* Find suitable destination for @next, with force if needed. */ 4758 dest_cpu = select_fallback_rq(dead_cpu, next); 4759 raw_spin_unlock(&rq->lock); 4760 4761 __migrate_task(next, dead_cpu, dest_cpu); 4762 4763 raw_spin_lock(&rq->lock); 4764 } 4765 4766 rq->stop = stop; 4767} 4768 4769#endif /* CONFIG_HOTPLUG_CPU */ 4770 4771#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4772 4773static struct ctl_table sd_ctl_dir[] = { 4774 { 4775 .procname = "sched_domain", 4776 .mode = 0555, 4777 }, 4778 {} 4779}; 4780 4781static struct ctl_table sd_ctl_root[] = { 4782 { 4783 .procname = "kernel", 4784 .mode = 0555, 4785 .child = sd_ctl_dir, 4786 }, 4787 {} 4788}; 4789 4790static struct ctl_table *sd_alloc_ctl_entry(int n) 4791{ 4792 struct ctl_table *entry = 4793 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4794 4795 return entry; 4796} 4797 4798static void sd_free_ctl_entry(struct ctl_table **tablep) 4799{ 4800 struct ctl_table *entry; 4801 4802 /* 4803 * In the intermediate directories, both the child directory and 4804 * procname are dynamically allocated and could fail but the mode 4805 * will always be set. In the lowest directory the names are 4806 * static strings and all have proc handlers. 4807 */ 4808 for (entry = *tablep; entry->mode; entry++) { 4809 if (entry->child) 4810 sd_free_ctl_entry(&entry->child); 4811 if (entry->proc_handler == NULL) 4812 kfree(entry->procname); 4813 } 4814 4815 kfree(*tablep); 4816 *tablep = NULL; 4817} 4818 4819static int min_load_idx = 0; 4820static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4821 4822static void 4823set_table_entry(struct ctl_table *entry, 4824 const char *procname, void *data, int maxlen, 4825 umode_t mode, proc_handler *proc_handler, 4826 bool load_idx) 4827{ 4828 entry->procname = procname; 4829 entry->data = data; 4830 entry->maxlen = maxlen; 4831 entry->mode = mode; 4832 entry->proc_handler = proc_handler; 4833 4834 if (load_idx) { 4835 entry->extra1 = &min_load_idx; 4836 entry->extra2 = &max_load_idx; 4837 } 4838} 4839 4840static struct ctl_table * 4841sd_alloc_ctl_domain_table(struct sched_domain *sd) 4842{ 4843 struct ctl_table *table = sd_alloc_ctl_entry(13); 4844 4845 if (table == NULL) 4846 return NULL; 4847 4848 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4849 sizeof(long), 0644, proc_doulongvec_minmax, false); 4850 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4851 sizeof(long), 0644, proc_doulongvec_minmax, false); 4852 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4853 sizeof(int), 0644, proc_dointvec_minmax, true); 4854 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4855 sizeof(int), 0644, proc_dointvec_minmax, true); 4856 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4857 sizeof(int), 0644, proc_dointvec_minmax, true); 4858 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4859 sizeof(int), 0644, proc_dointvec_minmax, true); 4860 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4861 sizeof(int), 0644, proc_dointvec_minmax, true); 4862 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4863 sizeof(int), 0644, proc_dointvec_minmax, false); 4864 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4865 sizeof(int), 0644, proc_dointvec_minmax, false); 4866 set_table_entry(&table[9], "cache_nice_tries", 4867 &sd->cache_nice_tries, 4868 sizeof(int), 0644, proc_dointvec_minmax, false); 4869 set_table_entry(&table[10], "flags", &sd->flags, 4870 sizeof(int), 0644, proc_dointvec_minmax, false); 4871 set_table_entry(&table[11], "name", sd->name, 4872 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4873 /* &table[12] is terminator */ 4874 4875 return table; 4876} 4877 4878static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4879{ 4880 struct ctl_table *entry, *table; 4881 struct sched_domain *sd; 4882 int domain_num = 0, i; 4883 char buf[32]; 4884 4885 for_each_domain(cpu, sd) 4886 domain_num++; 4887 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4888 if (table == NULL) 4889 return NULL; 4890 4891 i = 0; 4892 for_each_domain(cpu, sd) { 4893 snprintf(buf, 32, "domain%d", i); 4894 entry->procname = kstrdup(buf, GFP_KERNEL); 4895 entry->mode = 0555; 4896 entry->child = sd_alloc_ctl_domain_table(sd); 4897 entry++; 4898 i++; 4899 } 4900 return table; 4901} 4902 4903static struct ctl_table_header *sd_sysctl_header; 4904static void register_sched_domain_sysctl(void) 4905{ 4906 int i, cpu_num = num_possible_cpus(); 4907 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4908 char buf[32]; 4909 4910 WARN_ON(sd_ctl_dir[0].child); 4911 sd_ctl_dir[0].child = entry; 4912 4913 if (entry == NULL) 4914 return; 4915 4916 for_each_possible_cpu(i) { 4917 snprintf(buf, 32, "cpu%d", i); 4918 entry->procname = kstrdup(buf, GFP_KERNEL); 4919 entry->mode = 0555; 4920 entry->child = sd_alloc_ctl_cpu_table(i); 4921 entry++; 4922 } 4923 4924 WARN_ON(sd_sysctl_header); 4925 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4926} 4927 4928/* may be called multiple times per register */ 4929static void unregister_sched_domain_sysctl(void) 4930{ 4931 if (sd_sysctl_header) 4932 unregister_sysctl_table(sd_sysctl_header); 4933 sd_sysctl_header = NULL; 4934 if (sd_ctl_dir[0].child) 4935 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4936} 4937#else 4938static void register_sched_domain_sysctl(void) 4939{ 4940} 4941static void unregister_sched_domain_sysctl(void) 4942{ 4943} 4944#endif 4945 4946static void set_rq_online(struct rq *rq) 4947{ 4948 if (!rq->online) { 4949 const struct sched_class *class; 4950 4951 cpumask_set_cpu(rq->cpu, rq->rd->online); 4952 rq->online = 1; 4953 4954 for_each_class(class) { 4955 if (class->rq_online) 4956 class->rq_online(rq); 4957 } 4958 } 4959} 4960 4961static void set_rq_offline(struct rq *rq) 4962{ 4963 if (rq->online) { 4964 const struct sched_class *class; 4965 4966 for_each_class(class) { 4967 if (class->rq_offline) 4968 class->rq_offline(rq); 4969 } 4970 4971 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4972 rq->online = 0; 4973 } 4974} 4975 4976/* 4977 * migration_call - callback that gets triggered when a CPU is added. 4978 * Here we can start up the necessary migration thread for the new CPU. 4979 */ 4980static int 4981migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4982{ 4983 int cpu = (long)hcpu; 4984 unsigned long flags; 4985 struct rq *rq = cpu_rq(cpu); 4986 4987 switch (action & ~CPU_TASKS_FROZEN) { 4988 4989 case CPU_UP_PREPARE: 4990 rq->calc_load_update = calc_load_update; 4991 break; 4992 4993 case CPU_ONLINE: 4994 /* Update our root-domain */ 4995 raw_spin_lock_irqsave(&rq->lock, flags); 4996 if (rq->rd) { 4997 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4998 4999 set_rq_online(rq); 5000 } 5001 raw_spin_unlock_irqrestore(&rq->lock, flags); 5002 break; 5003 5004#ifdef CONFIG_HOTPLUG_CPU 5005 case CPU_DYING: 5006 sched_ttwu_pending(); 5007 /* Update our root-domain */ 5008 raw_spin_lock_irqsave(&rq->lock, flags); 5009 if (rq->rd) { 5010 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5011 set_rq_offline(rq); 5012 } 5013 migrate_tasks(cpu); 5014 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5015 raw_spin_unlock_irqrestore(&rq->lock, flags); 5016 break; 5017 5018 case CPU_DEAD: 5019 calc_load_migrate(rq); 5020 break; 5021#endif 5022 } 5023 5024 update_max_interval(); 5025 5026 return NOTIFY_OK; 5027} 5028 5029/* 5030 * Register at high priority so that task migration (migrate_all_tasks) 5031 * happens before everything else. This has to be lower priority than 5032 * the notifier in the perf_event subsystem, though. 5033 */ 5034static struct notifier_block migration_notifier = { 5035 .notifier_call = migration_call, 5036 .priority = CPU_PRI_MIGRATION, 5037}; 5038 5039static int sched_cpu_active(struct notifier_block *nfb, 5040 unsigned long action, void *hcpu) 5041{ 5042 switch (action & ~CPU_TASKS_FROZEN) { 5043 case CPU_STARTING: 5044 case CPU_DOWN_FAILED: 5045 set_cpu_active((long)hcpu, true); 5046 return NOTIFY_OK; 5047 default: 5048 return NOTIFY_DONE; 5049 } 5050} 5051 5052static int sched_cpu_inactive(struct notifier_block *nfb, 5053 unsigned long action, void *hcpu) 5054{ 5055 switch (action & ~CPU_TASKS_FROZEN) { 5056 case CPU_DOWN_PREPARE: 5057 set_cpu_active((long)hcpu, false); 5058 return NOTIFY_OK; 5059 default: 5060 return NOTIFY_DONE; 5061 } 5062} 5063 5064static int __init migration_init(void) 5065{ 5066 void *cpu = (void *)(long)smp_processor_id(); 5067 int err; 5068 5069 /* Initialize migration for the boot CPU */ 5070 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5071 BUG_ON(err == NOTIFY_BAD); 5072 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5073 register_cpu_notifier(&migration_notifier); 5074 5075 /* Register cpu active notifiers */ 5076 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5077 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5078 5079 return 0; 5080} 5081early_initcall(migration_init); 5082#endif 5083 5084#ifdef CONFIG_SMP 5085 5086static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5087 5088#ifdef CONFIG_SCHED_DEBUG 5089 5090static __read_mostly int sched_debug_enabled; 5091 5092static int __init sched_debug_setup(char *str) 5093{ 5094 sched_debug_enabled = 1; 5095 5096 return 0; 5097} 5098early_param("sched_debug", sched_debug_setup); 5099 5100static inline bool sched_debug(void) 5101{ 5102 return sched_debug_enabled; 5103} 5104 5105static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5106 struct cpumask *groupmask) 5107{ 5108 struct sched_group *group = sd->groups; 5109 char str[256]; 5110 5111 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5112 cpumask_clear(groupmask); 5113 5114 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5115 5116 if (!(sd->flags & SD_LOAD_BALANCE)) { 5117 printk("does not load-balance\n"); 5118 if (sd->parent) 5119 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5120 " has parent"); 5121 return -1; 5122 } 5123 5124 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5125 5126 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5127 printk(KERN_ERR "ERROR: domain->span does not contain " 5128 "CPU%d\n", cpu); 5129 } 5130 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5131 printk(KERN_ERR "ERROR: domain->groups does not contain" 5132 " CPU%d\n", cpu); 5133 } 5134 5135 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5136 do { 5137 if (!group) { 5138 printk("\n"); 5139 printk(KERN_ERR "ERROR: group is NULL\n"); 5140 break; 5141 } 5142 5143 /* 5144 * Even though we initialize ->power to something semi-sane, 5145 * we leave power_orig unset. This allows us to detect if 5146 * domain iteration is still funny without causing /0 traps. 5147 */ 5148 if (!group->sgp->power_orig) { 5149 printk(KERN_CONT "\n"); 5150 printk(KERN_ERR "ERROR: domain->cpu_power not " 5151 "set\n"); 5152 break; 5153 } 5154 5155 if (!cpumask_weight(sched_group_cpus(group))) { 5156 printk(KERN_CONT "\n"); 5157 printk(KERN_ERR "ERROR: empty group\n"); 5158 break; 5159 } 5160 5161 if (!(sd->flags & SD_OVERLAP) && 5162 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5163 printk(KERN_CONT "\n"); 5164 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5165 break; 5166 } 5167 5168 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5169 5170 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5171 5172 printk(KERN_CONT " %s", str); 5173 if (group->sgp->power != SCHED_POWER_SCALE) { 5174 printk(KERN_CONT " (cpu_power = %d)", 5175 group->sgp->power); 5176 } 5177 5178 group = group->next; 5179 } while (group != sd->groups); 5180 printk(KERN_CONT "\n"); 5181 5182 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5183 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5184 5185 if (sd->parent && 5186 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5187 printk(KERN_ERR "ERROR: parent span is not a superset " 5188 "of domain->span\n"); 5189 return 0; 5190} 5191 5192static void sched_domain_debug(struct sched_domain *sd, int cpu) 5193{ 5194 int level = 0; 5195 5196 if (!sched_debug_enabled) 5197 return; 5198 5199 if (!sd) { 5200 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5201 return; 5202 } 5203 5204 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5205 5206 for (;;) { 5207 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5208 break; 5209 level++; 5210 sd = sd->parent; 5211 if (!sd) 5212 break; 5213 } 5214} 5215#else /* !CONFIG_SCHED_DEBUG */ 5216# define sched_domain_debug(sd, cpu) do { } while (0) 5217static inline bool sched_debug(void) 5218{ 5219 return false; 5220} 5221#endif /* CONFIG_SCHED_DEBUG */ 5222 5223static int sd_degenerate(struct sched_domain *sd) 5224{ 5225 if (cpumask_weight(sched_domain_span(sd)) == 1) 5226 return 1; 5227 5228 /* Following flags need at least 2 groups */ 5229 if (sd->flags & (SD_LOAD_BALANCE | 5230 SD_BALANCE_NEWIDLE | 5231 SD_BALANCE_FORK | 5232 SD_BALANCE_EXEC | 5233 SD_SHARE_CPUPOWER | 5234 SD_SHARE_PKG_RESOURCES)) { 5235 if (sd->groups != sd->groups->next) 5236 return 0; 5237 } 5238 5239 /* Following flags don't use groups */ 5240 if (sd->flags & (SD_WAKE_AFFINE)) 5241 return 0; 5242 5243 return 1; 5244} 5245 5246static int 5247sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5248{ 5249 unsigned long cflags = sd->flags, pflags = parent->flags; 5250 5251 if (sd_degenerate(parent)) 5252 return 1; 5253 5254 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5255 return 0; 5256 5257 /* Flags needing groups don't count if only 1 group in parent */ 5258 if (parent->groups == parent->groups->next) { 5259 pflags &= ~(SD_LOAD_BALANCE | 5260 SD_BALANCE_NEWIDLE | 5261 SD_BALANCE_FORK | 5262 SD_BALANCE_EXEC | 5263 SD_SHARE_CPUPOWER | 5264 SD_SHARE_PKG_RESOURCES | 5265 SD_PREFER_SIBLING); 5266 if (nr_node_ids == 1) 5267 pflags &= ~SD_SERIALIZE; 5268 } 5269 if (~cflags & pflags) 5270 return 0; 5271 5272 return 1; 5273} 5274 5275static void free_rootdomain(struct rcu_head *rcu) 5276{ 5277 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5278 5279 cpupri_cleanup(&rd->cpupri); 5280 cpudl_cleanup(&rd->cpudl); 5281 free_cpumask_var(rd->dlo_mask); 5282 free_cpumask_var(rd->rto_mask); 5283 free_cpumask_var(rd->online); 5284 free_cpumask_var(rd->span); 5285 kfree(rd); 5286} 5287 5288static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5289{ 5290 struct root_domain *old_rd = NULL; 5291 unsigned long flags; 5292 5293 raw_spin_lock_irqsave(&rq->lock, flags); 5294 5295 if (rq->rd) { 5296 old_rd = rq->rd; 5297 5298 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5299 set_rq_offline(rq); 5300 5301 cpumask_clear_cpu(rq->cpu, old_rd->span); 5302 5303 /* 5304 * If we dont want to free the old_rd yet then 5305 * set old_rd to NULL to skip the freeing later 5306 * in this function: 5307 */ 5308 if (!atomic_dec_and_test(&old_rd->refcount)) 5309 old_rd = NULL; 5310 } 5311 5312 atomic_inc(&rd->refcount); 5313 rq->rd = rd; 5314 5315 cpumask_set_cpu(rq->cpu, rd->span); 5316 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5317 set_rq_online(rq); 5318 5319 raw_spin_unlock_irqrestore(&rq->lock, flags); 5320 5321 if (old_rd) 5322 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5323} 5324 5325static int init_rootdomain(struct root_domain *rd) 5326{ 5327 memset(rd, 0, sizeof(*rd)); 5328 5329 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5330 goto out; 5331 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5332 goto free_span; 5333 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5334 goto free_online; 5335 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5336 goto free_dlo_mask; 5337 5338 init_dl_bw(&rd->dl_bw); 5339 if (cpudl_init(&rd->cpudl) != 0) 5340 goto free_dlo_mask; 5341 5342 if (cpupri_init(&rd->cpupri) != 0) 5343 goto free_rto_mask; 5344 return 0; 5345 5346free_rto_mask: 5347 free_cpumask_var(rd->rto_mask); 5348free_dlo_mask: 5349 free_cpumask_var(rd->dlo_mask); 5350free_online: 5351 free_cpumask_var(rd->online); 5352free_span: 5353 free_cpumask_var(rd->span); 5354out: 5355 return -ENOMEM; 5356} 5357 5358/* 5359 * By default the system creates a single root-domain with all cpus as 5360 * members (mimicking the global state we have today). 5361 */ 5362struct root_domain def_root_domain; 5363 5364static void init_defrootdomain(void) 5365{ 5366 init_rootdomain(&def_root_domain); 5367 5368 atomic_set(&def_root_domain.refcount, 1); 5369} 5370 5371static struct root_domain *alloc_rootdomain(void) 5372{ 5373 struct root_domain *rd; 5374 5375 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5376 if (!rd) 5377 return NULL; 5378 5379 if (init_rootdomain(rd) != 0) { 5380 kfree(rd); 5381 return NULL; 5382 } 5383 5384 return rd; 5385} 5386 5387static void free_sched_groups(struct sched_group *sg, int free_sgp) 5388{ 5389 struct sched_group *tmp, *first; 5390 5391 if (!sg) 5392 return; 5393 5394 first = sg; 5395 do { 5396 tmp = sg->next; 5397 5398 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5399 kfree(sg->sgp); 5400 5401 kfree(sg); 5402 sg = tmp; 5403 } while (sg != first); 5404} 5405 5406static void free_sched_domain(struct rcu_head *rcu) 5407{ 5408 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5409 5410 /* 5411 * If its an overlapping domain it has private groups, iterate and 5412 * nuke them all. 5413 */ 5414 if (sd->flags & SD_OVERLAP) { 5415 free_sched_groups(sd->groups, 1); 5416 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5417 kfree(sd->groups->sgp); 5418 kfree(sd->groups); 5419 } 5420 kfree(sd); 5421} 5422 5423static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5424{ 5425 call_rcu(&sd->rcu, free_sched_domain); 5426} 5427 5428static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5429{ 5430 for (; sd; sd = sd->parent) 5431 destroy_sched_domain(sd, cpu); 5432} 5433 5434/* 5435 * Keep a special pointer to the highest sched_domain that has 5436 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5437 * allows us to avoid some pointer chasing select_idle_sibling(). 5438 * 5439 * Also keep a unique ID per domain (we use the first cpu number in 5440 * the cpumask of the domain), this allows us to quickly tell if 5441 * two cpus are in the same cache domain, see cpus_share_cache(). 5442 */ 5443DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5444DEFINE_PER_CPU(int, sd_llc_size); 5445DEFINE_PER_CPU(int, sd_llc_id); 5446DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5447DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5448DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5449 5450static void update_top_cache_domain(int cpu) 5451{ 5452 struct sched_domain *sd; 5453 struct sched_domain *busy_sd = NULL; 5454 int id = cpu; 5455 int size = 1; 5456 5457 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5458 if (sd) { 5459 id = cpumask_first(sched_domain_span(sd)); 5460 size = cpumask_weight(sched_domain_span(sd)); 5461 busy_sd = sd->parent; /* sd_busy */ 5462 } 5463 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5464 5465 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5466 per_cpu(sd_llc_size, cpu) = size; 5467 per_cpu(sd_llc_id, cpu) = id; 5468 5469 sd = lowest_flag_domain(cpu, SD_NUMA); 5470 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5471 5472 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5473 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5474} 5475 5476/* 5477 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5478 * hold the hotplug lock. 5479 */ 5480static void 5481cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5482{ 5483 struct rq *rq = cpu_rq(cpu); 5484 struct sched_domain *tmp; 5485 5486 /* Remove the sched domains which do not contribute to scheduling. */ 5487 for (tmp = sd; tmp; ) { 5488 struct sched_domain *parent = tmp->parent; 5489 if (!parent) 5490 break; 5491 5492 if (sd_parent_degenerate(tmp, parent)) { 5493 tmp->parent = parent->parent; 5494 if (parent->parent) 5495 parent->parent->child = tmp; 5496 /* 5497 * Transfer SD_PREFER_SIBLING down in case of a 5498 * degenerate parent; the spans match for this 5499 * so the property transfers. 5500 */ 5501 if (parent->flags & SD_PREFER_SIBLING) 5502 tmp->flags |= SD_PREFER_SIBLING; 5503 destroy_sched_domain(parent, cpu); 5504 } else 5505 tmp = tmp->parent; 5506 } 5507 5508 if (sd && sd_degenerate(sd)) { 5509 tmp = sd; 5510 sd = sd->parent; 5511 destroy_sched_domain(tmp, cpu); 5512 if (sd) 5513 sd->child = NULL; 5514 } 5515 5516 sched_domain_debug(sd, cpu); 5517 5518 rq_attach_root(rq, rd); 5519 tmp = rq->sd; 5520 rcu_assign_pointer(rq->sd, sd); 5521 destroy_sched_domains(tmp, cpu); 5522 5523 update_top_cache_domain(cpu); 5524} 5525 5526/* cpus with isolated domains */ 5527static cpumask_var_t cpu_isolated_map; 5528 5529/* Setup the mask of cpus configured for isolated domains */ 5530static int __init isolated_cpu_setup(char *str) 5531{ 5532 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5533 cpulist_parse(str, cpu_isolated_map); 5534 return 1; 5535} 5536 5537__setup("isolcpus=", isolated_cpu_setup); 5538 5539static const struct cpumask *cpu_cpu_mask(int cpu) 5540{ 5541 return cpumask_of_node(cpu_to_node(cpu)); 5542} 5543 5544struct sd_data { 5545 struct sched_domain **__percpu sd; 5546 struct sched_group **__percpu sg; 5547 struct sched_group_power **__percpu sgp; 5548}; 5549 5550struct s_data { 5551 struct sched_domain ** __percpu sd; 5552 struct root_domain *rd; 5553}; 5554 5555enum s_alloc { 5556 sa_rootdomain, 5557 sa_sd, 5558 sa_sd_storage, 5559 sa_none, 5560}; 5561 5562struct sched_domain_topology_level; 5563 5564typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5565typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5566 5567#define SDTL_OVERLAP 0x01 5568 5569struct sched_domain_topology_level { 5570 sched_domain_init_f init; 5571 sched_domain_mask_f mask; 5572 int flags; 5573 int numa_level; 5574 struct sd_data data; 5575}; 5576 5577/* 5578 * Build an iteration mask that can exclude certain CPUs from the upwards 5579 * domain traversal. 5580 * 5581 * Asymmetric node setups can result in situations where the domain tree is of 5582 * unequal depth, make sure to skip domains that already cover the entire 5583 * range. 5584 * 5585 * In that case build_sched_domains() will have terminated the iteration early 5586 * and our sibling sd spans will be empty. Domains should always include the 5587 * cpu they're built on, so check that. 5588 * 5589 */ 5590static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5591{ 5592 const struct cpumask *span = sched_domain_span(sd); 5593 struct sd_data *sdd = sd->private; 5594 struct sched_domain *sibling; 5595 int i; 5596 5597 for_each_cpu(i, span) { 5598 sibling = *per_cpu_ptr(sdd->sd, i); 5599 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5600 continue; 5601 5602 cpumask_set_cpu(i, sched_group_mask(sg)); 5603 } 5604} 5605 5606/* 5607 * Return the canonical balance cpu for this group, this is the first cpu 5608 * of this group that's also in the iteration mask. 5609 */ 5610int group_balance_cpu(struct sched_group *sg) 5611{ 5612 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5613} 5614 5615static int 5616build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5617{ 5618 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5619 const struct cpumask *span = sched_domain_span(sd); 5620 struct cpumask *covered = sched_domains_tmpmask; 5621 struct sd_data *sdd = sd->private; 5622 struct sched_domain *child; 5623 int i; 5624 5625 cpumask_clear(covered); 5626 5627 for_each_cpu(i, span) { 5628 struct cpumask *sg_span; 5629 5630 if (cpumask_test_cpu(i, covered)) 5631 continue; 5632 5633 child = *per_cpu_ptr(sdd->sd, i); 5634 5635 /* See the comment near build_group_mask(). */ 5636 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5637 continue; 5638 5639 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5640 GFP_KERNEL, cpu_to_node(cpu)); 5641 5642 if (!sg) 5643 goto fail; 5644 5645 sg_span = sched_group_cpus(sg); 5646 if (child->child) { 5647 child = child->child; 5648 cpumask_copy(sg_span, sched_domain_span(child)); 5649 } else 5650 cpumask_set_cpu(i, sg_span); 5651 5652 cpumask_or(covered, covered, sg_span); 5653 5654 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5655 if (atomic_inc_return(&sg->sgp->ref) == 1) 5656 build_group_mask(sd, sg); 5657 5658 /* 5659 * Initialize sgp->power such that even if we mess up the 5660 * domains and no possible iteration will get us here, we won't 5661 * die on a /0 trap. 5662 */ 5663 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5664 sg->sgp->power_orig = sg->sgp->power; 5665 5666 /* 5667 * Make sure the first group of this domain contains the 5668 * canonical balance cpu. Otherwise the sched_domain iteration 5669 * breaks. See update_sg_lb_stats(). 5670 */ 5671 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5672 group_balance_cpu(sg) == cpu) 5673 groups = sg; 5674 5675 if (!first) 5676 first = sg; 5677 if (last) 5678 last->next = sg; 5679 last = sg; 5680 last->next = first; 5681 } 5682 sd->groups = groups; 5683 5684 return 0; 5685 5686fail: 5687 free_sched_groups(first, 0); 5688 5689 return -ENOMEM; 5690} 5691 5692static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5693{ 5694 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5695 struct sched_domain *child = sd->child; 5696 5697 if (child) 5698 cpu = cpumask_first(sched_domain_span(child)); 5699 5700 if (sg) { 5701 *sg = *per_cpu_ptr(sdd->sg, cpu); 5702 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5703 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5704 } 5705 5706 return cpu; 5707} 5708 5709/* 5710 * build_sched_groups will build a circular linked list of the groups 5711 * covered by the given span, and will set each group's ->cpumask correctly, 5712 * and ->cpu_power to 0. 5713 * 5714 * Assumes the sched_domain tree is fully constructed 5715 */ 5716static int 5717build_sched_groups(struct sched_domain *sd, int cpu) 5718{ 5719 struct sched_group *first = NULL, *last = NULL; 5720 struct sd_data *sdd = sd->private; 5721 const struct cpumask *span = sched_domain_span(sd); 5722 struct cpumask *covered; 5723 int i; 5724 5725 get_group(cpu, sdd, &sd->groups); 5726 atomic_inc(&sd->groups->ref); 5727 5728 if (cpu != cpumask_first(span)) 5729 return 0; 5730 5731 lockdep_assert_held(&sched_domains_mutex); 5732 covered = sched_domains_tmpmask; 5733 5734 cpumask_clear(covered); 5735 5736 for_each_cpu(i, span) { 5737 struct sched_group *sg; 5738 int group, j; 5739 5740 if (cpumask_test_cpu(i, covered)) 5741 continue; 5742 5743 group = get_group(i, sdd, &sg); 5744 cpumask_clear(sched_group_cpus(sg)); 5745 sg->sgp->power = 0; 5746 cpumask_setall(sched_group_mask(sg)); 5747 5748 for_each_cpu(j, span) { 5749 if (get_group(j, sdd, NULL) != group) 5750 continue; 5751 5752 cpumask_set_cpu(j, covered); 5753 cpumask_set_cpu(j, sched_group_cpus(sg)); 5754 } 5755 5756 if (!first) 5757 first = sg; 5758 if (last) 5759 last->next = sg; 5760 last = sg; 5761 } 5762 last->next = first; 5763 5764 return 0; 5765} 5766 5767/* 5768 * Initialize sched groups cpu_power. 5769 * 5770 * cpu_power indicates the capacity of sched group, which is used while 5771 * distributing the load between different sched groups in a sched domain. 5772 * Typically cpu_power for all the groups in a sched domain will be same unless 5773 * there are asymmetries in the topology. If there are asymmetries, group 5774 * having more cpu_power will pickup more load compared to the group having 5775 * less cpu_power. 5776 */ 5777static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5778{ 5779 struct sched_group *sg = sd->groups; 5780 5781 WARN_ON(!sg); 5782 5783 do { 5784 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5785 sg = sg->next; 5786 } while (sg != sd->groups); 5787 5788 if (cpu != group_balance_cpu(sg)) 5789 return; 5790 5791 update_group_power(sd, cpu); 5792 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5793} 5794 5795int __weak arch_sd_sibling_asym_packing(void) 5796{ 5797 return 0*SD_ASYM_PACKING; 5798} 5799 5800/* 5801 * Initializers for schedule domains 5802 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5803 */ 5804 5805#ifdef CONFIG_SCHED_DEBUG 5806# define SD_INIT_NAME(sd, type) sd->name = #type 5807#else 5808# define SD_INIT_NAME(sd, type) do { } while (0) 5809#endif 5810 5811#define SD_INIT_FUNC(type) \ 5812static noinline struct sched_domain * \ 5813sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5814{ \ 5815 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5816 *sd = SD_##type##_INIT; \ 5817 SD_INIT_NAME(sd, type); \ 5818 sd->private = &tl->data; \ 5819 return sd; \ 5820} 5821 5822SD_INIT_FUNC(CPU) 5823#ifdef CONFIG_SCHED_SMT 5824 SD_INIT_FUNC(SIBLING) 5825#endif 5826#ifdef CONFIG_SCHED_MC 5827 SD_INIT_FUNC(MC) 5828#endif 5829#ifdef CONFIG_SCHED_BOOK 5830 SD_INIT_FUNC(BOOK) 5831#endif 5832 5833static int default_relax_domain_level = -1; 5834int sched_domain_level_max; 5835 5836static int __init setup_relax_domain_level(char *str) 5837{ 5838 if (kstrtoint(str, 0, &default_relax_domain_level)) 5839 pr_warn("Unable to set relax_domain_level\n"); 5840 5841 return 1; 5842} 5843__setup("relax_domain_level=", setup_relax_domain_level); 5844 5845static void set_domain_attribute(struct sched_domain *sd, 5846 struct sched_domain_attr *attr) 5847{ 5848 int request; 5849 5850 if (!attr || attr->relax_domain_level < 0) { 5851 if (default_relax_domain_level < 0) 5852 return; 5853 else 5854 request = default_relax_domain_level; 5855 } else 5856 request = attr->relax_domain_level; 5857 if (request < sd->level) { 5858 /* turn off idle balance on this domain */ 5859 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5860 } else { 5861 /* turn on idle balance on this domain */ 5862 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5863 } 5864} 5865 5866static void __sdt_free(const struct cpumask *cpu_map); 5867static int __sdt_alloc(const struct cpumask *cpu_map); 5868 5869static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5870 const struct cpumask *cpu_map) 5871{ 5872 switch (what) { 5873 case sa_rootdomain: 5874 if (!atomic_read(&d->rd->refcount)) 5875 free_rootdomain(&d->rd->rcu); /* fall through */ 5876 case sa_sd: 5877 free_percpu(d->sd); /* fall through */ 5878 case sa_sd_storage: 5879 __sdt_free(cpu_map); /* fall through */ 5880 case sa_none: 5881 break; 5882 } 5883} 5884 5885static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5886 const struct cpumask *cpu_map) 5887{ 5888 memset(d, 0, sizeof(*d)); 5889 5890 if (__sdt_alloc(cpu_map)) 5891 return sa_sd_storage; 5892 d->sd = alloc_percpu(struct sched_domain *); 5893 if (!d->sd) 5894 return sa_sd_storage; 5895 d->rd = alloc_rootdomain(); 5896 if (!d->rd) 5897 return sa_sd; 5898 return sa_rootdomain; 5899} 5900 5901/* 5902 * NULL the sd_data elements we've used to build the sched_domain and 5903 * sched_group structure so that the subsequent __free_domain_allocs() 5904 * will not free the data we're using. 5905 */ 5906static void claim_allocations(int cpu, struct sched_domain *sd) 5907{ 5908 struct sd_data *sdd = sd->private; 5909 5910 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5911 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5912 5913 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5914 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5915 5916 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5917 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5918} 5919 5920#ifdef CONFIG_SCHED_SMT 5921static const struct cpumask *cpu_smt_mask(int cpu) 5922{ 5923 return topology_thread_cpumask(cpu); 5924} 5925#endif 5926 5927/* 5928 * Topology list, bottom-up. 5929 */ 5930static struct sched_domain_topology_level default_topology[] = { 5931#ifdef CONFIG_SCHED_SMT 5932 { sd_init_SIBLING, cpu_smt_mask, }, 5933#endif 5934#ifdef CONFIG_SCHED_MC 5935 { sd_init_MC, cpu_coregroup_mask, }, 5936#endif 5937#ifdef CONFIG_SCHED_BOOK 5938 { sd_init_BOOK, cpu_book_mask, }, 5939#endif 5940 { sd_init_CPU, cpu_cpu_mask, }, 5941 { NULL, }, 5942}; 5943 5944static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5945 5946#define for_each_sd_topology(tl) \ 5947 for (tl = sched_domain_topology; tl->init; tl++) 5948 5949#ifdef CONFIG_NUMA 5950 5951static int sched_domains_numa_levels; 5952static int *sched_domains_numa_distance; 5953static struct cpumask ***sched_domains_numa_masks; 5954static int sched_domains_curr_level; 5955 5956static inline int sd_local_flags(int level) 5957{ 5958 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5959 return 0; 5960 5961 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5962} 5963 5964static struct sched_domain * 5965sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5966{ 5967 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5968 int level = tl->numa_level; 5969 int sd_weight = cpumask_weight( 5970 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5971 5972 *sd = (struct sched_domain){ 5973 .min_interval = sd_weight, 5974 .max_interval = 2*sd_weight, 5975 .busy_factor = 32, 5976 .imbalance_pct = 125, 5977 .cache_nice_tries = 2, 5978 .busy_idx = 3, 5979 .idle_idx = 2, 5980 .newidle_idx = 0, 5981 .wake_idx = 0, 5982 .forkexec_idx = 0, 5983 5984 .flags = 1*SD_LOAD_BALANCE 5985 | 1*SD_BALANCE_NEWIDLE 5986 | 0*SD_BALANCE_EXEC 5987 | 0*SD_BALANCE_FORK 5988 | 0*SD_BALANCE_WAKE 5989 | 0*SD_WAKE_AFFINE 5990 | 0*SD_SHARE_CPUPOWER 5991 | 0*SD_SHARE_PKG_RESOURCES 5992 | 1*SD_SERIALIZE 5993 | 0*SD_PREFER_SIBLING 5994 | 1*SD_NUMA 5995 | sd_local_flags(level) 5996 , 5997 .last_balance = jiffies, 5998 .balance_interval = sd_weight, 5999 }; 6000 SD_INIT_NAME(sd, NUMA); 6001 sd->private = &tl->data; 6002 6003 /* 6004 * Ugly hack to pass state to sd_numa_mask()... 6005 */ 6006 sched_domains_curr_level = tl->numa_level; 6007 6008 return sd; 6009} 6010 6011static const struct cpumask *sd_numa_mask(int cpu) 6012{ 6013 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6014} 6015 6016static void sched_numa_warn(const char *str) 6017{ 6018 static int done = false; 6019 int i,j; 6020 6021 if (done) 6022 return; 6023 6024 done = true; 6025 6026 printk(KERN_WARNING "ERROR: %s\n\n", str); 6027 6028 for (i = 0; i < nr_node_ids; i++) { 6029 printk(KERN_WARNING " "); 6030 for (j = 0; j < nr_node_ids; j++) 6031 printk(KERN_CONT "%02d ", node_distance(i,j)); 6032 printk(KERN_CONT "\n"); 6033 } 6034 printk(KERN_WARNING "\n"); 6035} 6036 6037static bool find_numa_distance(int distance) 6038{ 6039 int i; 6040 6041 if (distance == node_distance(0, 0)) 6042 return true; 6043 6044 for (i = 0; i < sched_domains_numa_levels; i++) { 6045 if (sched_domains_numa_distance[i] == distance) 6046 return true; 6047 } 6048 6049 return false; 6050} 6051 6052static void sched_init_numa(void) 6053{ 6054 int next_distance, curr_distance = node_distance(0, 0); 6055 struct sched_domain_topology_level *tl; 6056 int level = 0; 6057 int i, j, k; 6058 6059 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6060 if (!sched_domains_numa_distance) 6061 return; 6062 6063 /* 6064 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6065 * unique distances in the node_distance() table. 6066 * 6067 * Assumes node_distance(0,j) includes all distances in 6068 * node_distance(i,j) in order to avoid cubic time. 6069 */ 6070 next_distance = curr_distance; 6071 for (i = 0; i < nr_node_ids; i++) { 6072 for (j = 0; j < nr_node_ids; j++) { 6073 for (k = 0; k < nr_node_ids; k++) { 6074 int distance = node_distance(i, k); 6075 6076 if (distance > curr_distance && 6077 (distance < next_distance || 6078 next_distance == curr_distance)) 6079 next_distance = distance; 6080 6081 /* 6082 * While not a strong assumption it would be nice to know 6083 * about cases where if node A is connected to B, B is not 6084 * equally connected to A. 6085 */ 6086 if (sched_debug() && node_distance(k, i) != distance) 6087 sched_numa_warn("Node-distance not symmetric"); 6088 6089 if (sched_debug() && i && !find_numa_distance(distance)) 6090 sched_numa_warn("Node-0 not representative"); 6091 } 6092 if (next_distance != curr_distance) { 6093 sched_domains_numa_distance[level++] = next_distance; 6094 sched_domains_numa_levels = level; 6095 curr_distance = next_distance; 6096 } else break; 6097 } 6098 6099 /* 6100 * In case of sched_debug() we verify the above assumption. 6101 */ 6102 if (!sched_debug()) 6103 break; 6104 } 6105 /* 6106 * 'level' contains the number of unique distances, excluding the 6107 * identity distance node_distance(i,i). 6108 * 6109 * The sched_domains_numa_distance[] array includes the actual distance 6110 * numbers. 6111 */ 6112 6113 /* 6114 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6115 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6116 * the array will contain less then 'level' members. This could be 6117 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6118 * in other functions. 6119 * 6120 * We reset it to 'level' at the end of this function. 6121 */ 6122 sched_domains_numa_levels = 0; 6123 6124 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6125 if (!sched_domains_numa_masks) 6126 return; 6127 6128 /* 6129 * Now for each level, construct a mask per node which contains all 6130 * cpus of nodes that are that many hops away from us. 6131 */ 6132 for (i = 0; i < level; i++) { 6133 sched_domains_numa_masks[i] = 6134 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6135 if (!sched_domains_numa_masks[i]) 6136 return; 6137 6138 for (j = 0; j < nr_node_ids; j++) { 6139 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6140 if (!mask) 6141 return; 6142 6143 sched_domains_numa_masks[i][j] = mask; 6144 6145 for (k = 0; k < nr_node_ids; k++) { 6146 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6147 continue; 6148 6149 cpumask_or(mask, mask, cpumask_of_node(k)); 6150 } 6151 } 6152 } 6153 6154 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6155 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6156 if (!tl) 6157 return; 6158 6159 /* 6160 * Copy the default topology bits.. 6161 */ 6162 for (i = 0; default_topology[i].init; i++) 6163 tl[i] = default_topology[i]; 6164 6165 /* 6166 * .. and append 'j' levels of NUMA goodness. 6167 */ 6168 for (j = 0; j < level; i++, j++) { 6169 tl[i] = (struct sched_domain_topology_level){ 6170 .init = sd_numa_init, 6171 .mask = sd_numa_mask, 6172 .flags = SDTL_OVERLAP, 6173 .numa_level = j, 6174 }; 6175 } 6176 6177 sched_domain_topology = tl; 6178 6179 sched_domains_numa_levels = level; 6180} 6181 6182static void sched_domains_numa_masks_set(int cpu) 6183{ 6184 int i, j; 6185 int node = cpu_to_node(cpu); 6186 6187 for (i = 0; i < sched_domains_numa_levels; i++) { 6188 for (j = 0; j < nr_node_ids; j++) { 6189 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6190 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6191 } 6192 } 6193} 6194 6195static void sched_domains_numa_masks_clear(int cpu) 6196{ 6197 int i, j; 6198 for (i = 0; i < sched_domains_numa_levels; i++) { 6199 for (j = 0; j < nr_node_ids; j++) 6200 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6201 } 6202} 6203 6204/* 6205 * Update sched_domains_numa_masks[level][node] array when new cpus 6206 * are onlined. 6207 */ 6208static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6209 unsigned long action, 6210 void *hcpu) 6211{ 6212 int cpu = (long)hcpu; 6213 6214 switch (action & ~CPU_TASKS_FROZEN) { 6215 case CPU_ONLINE: 6216 sched_domains_numa_masks_set(cpu); 6217 break; 6218 6219 case CPU_DEAD: 6220 sched_domains_numa_masks_clear(cpu); 6221 break; 6222 6223 default: 6224 return NOTIFY_DONE; 6225 } 6226 6227 return NOTIFY_OK; 6228} 6229#else 6230static inline void sched_init_numa(void) 6231{ 6232} 6233 6234static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6235 unsigned long action, 6236 void *hcpu) 6237{ 6238 return 0; 6239} 6240#endif /* CONFIG_NUMA */ 6241 6242static int __sdt_alloc(const struct cpumask *cpu_map) 6243{ 6244 struct sched_domain_topology_level *tl; 6245 int j; 6246 6247 for_each_sd_topology(tl) { 6248 struct sd_data *sdd = &tl->data; 6249 6250 sdd->sd = alloc_percpu(struct sched_domain *); 6251 if (!sdd->sd) 6252 return -ENOMEM; 6253 6254 sdd->sg = alloc_percpu(struct sched_group *); 6255 if (!sdd->sg) 6256 return -ENOMEM; 6257 6258 sdd->sgp = alloc_percpu(struct sched_group_power *); 6259 if (!sdd->sgp) 6260 return -ENOMEM; 6261 6262 for_each_cpu(j, cpu_map) { 6263 struct sched_domain *sd; 6264 struct sched_group *sg; 6265 struct sched_group_power *sgp; 6266 6267 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6268 GFP_KERNEL, cpu_to_node(j)); 6269 if (!sd) 6270 return -ENOMEM; 6271 6272 *per_cpu_ptr(sdd->sd, j) = sd; 6273 6274 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6275 GFP_KERNEL, cpu_to_node(j)); 6276 if (!sg) 6277 return -ENOMEM; 6278 6279 sg->next = sg; 6280 6281 *per_cpu_ptr(sdd->sg, j) = sg; 6282 6283 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6284 GFP_KERNEL, cpu_to_node(j)); 6285 if (!sgp) 6286 return -ENOMEM; 6287 6288 *per_cpu_ptr(sdd->sgp, j) = sgp; 6289 } 6290 } 6291 6292 return 0; 6293} 6294 6295static void __sdt_free(const struct cpumask *cpu_map) 6296{ 6297 struct sched_domain_topology_level *tl; 6298 int j; 6299 6300 for_each_sd_topology(tl) { 6301 struct sd_data *sdd = &tl->data; 6302 6303 for_each_cpu(j, cpu_map) { 6304 struct sched_domain *sd; 6305 6306 if (sdd->sd) { 6307 sd = *per_cpu_ptr(sdd->sd, j); 6308 if (sd && (sd->flags & SD_OVERLAP)) 6309 free_sched_groups(sd->groups, 0); 6310 kfree(*per_cpu_ptr(sdd->sd, j)); 6311 } 6312 6313 if (sdd->sg) 6314 kfree(*per_cpu_ptr(sdd->sg, j)); 6315 if (sdd->sgp) 6316 kfree(*per_cpu_ptr(sdd->sgp, j)); 6317 } 6318 free_percpu(sdd->sd); 6319 sdd->sd = NULL; 6320 free_percpu(sdd->sg); 6321 sdd->sg = NULL; 6322 free_percpu(sdd->sgp); 6323 sdd->sgp = NULL; 6324 } 6325} 6326 6327struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6328 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6329 struct sched_domain *child, int cpu) 6330{ 6331 struct sched_domain *sd = tl->init(tl, cpu); 6332 if (!sd) 6333 return child; 6334 6335 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6336 if (child) { 6337 sd->level = child->level + 1; 6338 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6339 child->parent = sd; 6340 sd->child = child; 6341 } 6342 set_domain_attribute(sd, attr); 6343 6344 return sd; 6345} 6346 6347/* 6348 * Build sched domains for a given set of cpus and attach the sched domains 6349 * to the individual cpus 6350 */ 6351static int build_sched_domains(const struct cpumask *cpu_map, 6352 struct sched_domain_attr *attr) 6353{ 6354 enum s_alloc alloc_state; 6355 struct sched_domain *sd; 6356 struct s_data d; 6357 int i, ret = -ENOMEM; 6358 6359 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6360 if (alloc_state != sa_rootdomain) 6361 goto error; 6362 6363 /* Set up domains for cpus specified by the cpu_map. */ 6364 for_each_cpu(i, cpu_map) { 6365 struct sched_domain_topology_level *tl; 6366 6367 sd = NULL; 6368 for_each_sd_topology(tl) { 6369 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6370 if (tl == sched_domain_topology) 6371 *per_cpu_ptr(d.sd, i) = sd; 6372 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6373 sd->flags |= SD_OVERLAP; 6374 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6375 break; 6376 } 6377 } 6378 6379 /* Build the groups for the domains */ 6380 for_each_cpu(i, cpu_map) { 6381 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6382 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6383 if (sd->flags & SD_OVERLAP) { 6384 if (build_overlap_sched_groups(sd, i)) 6385 goto error; 6386 } else { 6387 if (build_sched_groups(sd, i)) 6388 goto error; 6389 } 6390 } 6391 } 6392 6393 /* Calculate CPU power for physical packages and nodes */ 6394 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6395 if (!cpumask_test_cpu(i, cpu_map)) 6396 continue; 6397 6398 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6399 claim_allocations(i, sd); 6400 init_sched_groups_power(i, sd); 6401 } 6402 } 6403 6404 /* Attach the domains */ 6405 rcu_read_lock(); 6406 for_each_cpu(i, cpu_map) { 6407 sd = *per_cpu_ptr(d.sd, i); 6408 cpu_attach_domain(sd, d.rd, i); 6409 } 6410 rcu_read_unlock(); 6411 6412 ret = 0; 6413error: 6414 __free_domain_allocs(&d, alloc_state, cpu_map); 6415 return ret; 6416} 6417 6418static cpumask_var_t *doms_cur; /* current sched domains */ 6419static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6420static struct sched_domain_attr *dattr_cur; 6421 /* attribues of custom domains in 'doms_cur' */ 6422 6423/* 6424 * Special case: If a kmalloc of a doms_cur partition (array of 6425 * cpumask) fails, then fallback to a single sched domain, 6426 * as determined by the single cpumask fallback_doms. 6427 */ 6428static cpumask_var_t fallback_doms; 6429 6430/* 6431 * arch_update_cpu_topology lets virtualized architectures update the 6432 * cpu core maps. It is supposed to return 1 if the topology changed 6433 * or 0 if it stayed the same. 6434 */ 6435int __attribute__((weak)) arch_update_cpu_topology(void) 6436{ 6437 return 0; 6438} 6439 6440cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6441{ 6442 int i; 6443 cpumask_var_t *doms; 6444 6445 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6446 if (!doms) 6447 return NULL; 6448 for (i = 0; i < ndoms; i++) { 6449 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6450 free_sched_domains(doms, i); 6451 return NULL; 6452 } 6453 } 6454 return doms; 6455} 6456 6457void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6458{ 6459 unsigned int i; 6460 for (i = 0; i < ndoms; i++) 6461 free_cpumask_var(doms[i]); 6462 kfree(doms); 6463} 6464 6465/* 6466 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6467 * For now this just excludes isolated cpus, but could be used to 6468 * exclude other special cases in the future. 6469 */ 6470static int init_sched_domains(const struct cpumask *cpu_map) 6471{ 6472 int err; 6473 6474 arch_update_cpu_topology(); 6475 ndoms_cur = 1; 6476 doms_cur = alloc_sched_domains(ndoms_cur); 6477 if (!doms_cur) 6478 doms_cur = &fallback_doms; 6479 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6480 err = build_sched_domains(doms_cur[0], NULL); 6481 register_sched_domain_sysctl(); 6482 6483 return err; 6484} 6485 6486/* 6487 * Detach sched domains from a group of cpus specified in cpu_map 6488 * These cpus will now be attached to the NULL domain 6489 */ 6490static void detach_destroy_domains(const struct cpumask *cpu_map) 6491{ 6492 int i; 6493 6494 rcu_read_lock(); 6495 for_each_cpu(i, cpu_map) 6496 cpu_attach_domain(NULL, &def_root_domain, i); 6497 rcu_read_unlock(); 6498} 6499 6500/* handle null as "default" */ 6501static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6502 struct sched_domain_attr *new, int idx_new) 6503{ 6504 struct sched_domain_attr tmp; 6505 6506 /* fast path */ 6507 if (!new && !cur) 6508 return 1; 6509 6510 tmp = SD_ATTR_INIT; 6511 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6512 new ? (new + idx_new) : &tmp, 6513 sizeof(struct sched_domain_attr)); 6514} 6515 6516/* 6517 * Partition sched domains as specified by the 'ndoms_new' 6518 * cpumasks in the array doms_new[] of cpumasks. This compares 6519 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6520 * It destroys each deleted domain and builds each new domain. 6521 * 6522 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6523 * The masks don't intersect (don't overlap.) We should setup one 6524 * sched domain for each mask. CPUs not in any of the cpumasks will 6525 * not be load balanced. If the same cpumask appears both in the 6526 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6527 * it as it is. 6528 * 6529 * The passed in 'doms_new' should be allocated using 6530 * alloc_sched_domains. This routine takes ownership of it and will 6531 * free_sched_domains it when done with it. If the caller failed the 6532 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6533 * and partition_sched_domains() will fallback to the single partition 6534 * 'fallback_doms', it also forces the domains to be rebuilt. 6535 * 6536 * If doms_new == NULL it will be replaced with cpu_online_mask. 6537 * ndoms_new == 0 is a special case for destroying existing domains, 6538 * and it will not create the default domain. 6539 * 6540 * Call with hotplug lock held 6541 */ 6542void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6543 struct sched_domain_attr *dattr_new) 6544{ 6545 int i, j, n; 6546 int new_topology; 6547 6548 mutex_lock(&sched_domains_mutex); 6549 6550 /* always unregister in case we don't destroy any domains */ 6551 unregister_sched_domain_sysctl(); 6552 6553 /* Let architecture update cpu core mappings. */ 6554 new_topology = arch_update_cpu_topology(); 6555 6556 n = doms_new ? ndoms_new : 0; 6557 6558 /* Destroy deleted domains */ 6559 for (i = 0; i < ndoms_cur; i++) { 6560 for (j = 0; j < n && !new_topology; j++) { 6561 if (cpumask_equal(doms_cur[i], doms_new[j]) 6562 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6563 goto match1; 6564 } 6565 /* no match - a current sched domain not in new doms_new[] */ 6566 detach_destroy_domains(doms_cur[i]); 6567match1: 6568 ; 6569 } 6570 6571 n = ndoms_cur; 6572 if (doms_new == NULL) { 6573 n = 0; 6574 doms_new = &fallback_doms; 6575 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6576 WARN_ON_ONCE(dattr_new); 6577 } 6578 6579 /* Build new domains */ 6580 for (i = 0; i < ndoms_new; i++) { 6581 for (j = 0; j < n && !new_topology; j++) { 6582 if (cpumask_equal(doms_new[i], doms_cur[j]) 6583 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6584 goto match2; 6585 } 6586 /* no match - add a new doms_new */ 6587 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6588match2: 6589 ; 6590 } 6591 6592 /* Remember the new sched domains */ 6593 if (doms_cur != &fallback_doms) 6594 free_sched_domains(doms_cur, ndoms_cur); 6595 kfree(dattr_cur); /* kfree(NULL) is safe */ 6596 doms_cur = doms_new; 6597 dattr_cur = dattr_new; 6598 ndoms_cur = ndoms_new; 6599 6600 register_sched_domain_sysctl(); 6601 6602 mutex_unlock(&sched_domains_mutex); 6603} 6604 6605static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6606 6607/* 6608 * Update cpusets according to cpu_active mask. If cpusets are 6609 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6610 * around partition_sched_domains(). 6611 * 6612 * If we come here as part of a suspend/resume, don't touch cpusets because we 6613 * want to restore it back to its original state upon resume anyway. 6614 */ 6615static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6616 void *hcpu) 6617{ 6618 switch (action) { 6619 case CPU_ONLINE_FROZEN: 6620 case CPU_DOWN_FAILED_FROZEN: 6621 6622 /* 6623 * num_cpus_frozen tracks how many CPUs are involved in suspend 6624 * resume sequence. As long as this is not the last online 6625 * operation in the resume sequence, just build a single sched 6626 * domain, ignoring cpusets. 6627 */ 6628 num_cpus_frozen--; 6629 if (likely(num_cpus_frozen)) { 6630 partition_sched_domains(1, NULL, NULL); 6631 break; 6632 } 6633 6634 /* 6635 * This is the last CPU online operation. So fall through and 6636 * restore the original sched domains by considering the 6637 * cpuset configurations. 6638 */ 6639 6640 case CPU_ONLINE: 6641 case CPU_DOWN_FAILED: 6642 cpuset_update_active_cpus(true); 6643 break; 6644 default: 6645 return NOTIFY_DONE; 6646 } 6647 return NOTIFY_OK; 6648} 6649 6650static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6651 void *hcpu) 6652{ 6653 switch (action) { 6654 case CPU_DOWN_PREPARE: 6655 cpuset_update_active_cpus(false); 6656 break; 6657 case CPU_DOWN_PREPARE_FROZEN: 6658 num_cpus_frozen++; 6659 partition_sched_domains(1, NULL, NULL); 6660 break; 6661 default: 6662 return NOTIFY_DONE; 6663 } 6664 return NOTIFY_OK; 6665} 6666 6667void __init sched_init_smp(void) 6668{ 6669 cpumask_var_t non_isolated_cpus; 6670 6671 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6672 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6673 6674 sched_init_numa(); 6675 6676 /* 6677 * There's no userspace yet to cause hotplug operations; hence all the 6678 * cpu masks are stable and all blatant races in the below code cannot 6679 * happen. 6680 */ 6681 mutex_lock(&sched_domains_mutex); 6682 init_sched_domains(cpu_active_mask); 6683 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6684 if (cpumask_empty(non_isolated_cpus)) 6685 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6686 mutex_unlock(&sched_domains_mutex); 6687 6688 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6689 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6690 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6691 6692 init_hrtick(); 6693 6694 /* Move init over to a non-isolated CPU */ 6695 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6696 BUG(); 6697 sched_init_granularity(); 6698 free_cpumask_var(non_isolated_cpus); 6699 6700 init_sched_rt_class(); 6701 init_sched_dl_class(); 6702} 6703#else 6704void __init sched_init_smp(void) 6705{ 6706 sched_init_granularity(); 6707} 6708#endif /* CONFIG_SMP */ 6709 6710const_debug unsigned int sysctl_timer_migration = 1; 6711 6712int in_sched_functions(unsigned long addr) 6713{ 6714 return in_lock_functions(addr) || 6715 (addr >= (unsigned long)__sched_text_start 6716 && addr < (unsigned long)__sched_text_end); 6717} 6718 6719#ifdef CONFIG_CGROUP_SCHED 6720/* 6721 * Default task group. 6722 * Every task in system belongs to this group at bootup. 6723 */ 6724struct task_group root_task_group; 6725LIST_HEAD(task_groups); 6726#endif 6727 6728DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6729 6730void __init sched_init(void) 6731{ 6732 int i, j; 6733 unsigned long alloc_size = 0, ptr; 6734 6735#ifdef CONFIG_FAIR_GROUP_SCHED 6736 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6737#endif 6738#ifdef CONFIG_RT_GROUP_SCHED 6739 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6740#endif 6741#ifdef CONFIG_CPUMASK_OFFSTACK 6742 alloc_size += num_possible_cpus() * cpumask_size(); 6743#endif 6744 if (alloc_size) { 6745 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6746 6747#ifdef CONFIG_FAIR_GROUP_SCHED 6748 root_task_group.se = (struct sched_entity **)ptr; 6749 ptr += nr_cpu_ids * sizeof(void **); 6750 6751 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6752 ptr += nr_cpu_ids * sizeof(void **); 6753 6754#endif /* CONFIG_FAIR_GROUP_SCHED */ 6755#ifdef CONFIG_RT_GROUP_SCHED 6756 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6757 ptr += nr_cpu_ids * sizeof(void **); 6758 6759 root_task_group.rt_rq = (struct rt_rq **)ptr; 6760 ptr += nr_cpu_ids * sizeof(void **); 6761 6762#endif /* CONFIG_RT_GROUP_SCHED */ 6763#ifdef CONFIG_CPUMASK_OFFSTACK 6764 for_each_possible_cpu(i) { 6765 per_cpu(load_balance_mask, i) = (void *)ptr; 6766 ptr += cpumask_size(); 6767 } 6768#endif /* CONFIG_CPUMASK_OFFSTACK */ 6769 } 6770 6771 init_rt_bandwidth(&def_rt_bandwidth, 6772 global_rt_period(), global_rt_runtime()); 6773 init_dl_bandwidth(&def_dl_bandwidth, 6774 global_rt_period(), global_rt_runtime()); 6775 6776#ifdef CONFIG_SMP 6777 init_defrootdomain(); 6778#endif 6779 6780#ifdef CONFIG_RT_GROUP_SCHED 6781 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6782 global_rt_period(), global_rt_runtime()); 6783#endif /* CONFIG_RT_GROUP_SCHED */ 6784 6785#ifdef CONFIG_CGROUP_SCHED 6786 list_add(&root_task_group.list, &task_groups); 6787 INIT_LIST_HEAD(&root_task_group.children); 6788 INIT_LIST_HEAD(&root_task_group.siblings); 6789 autogroup_init(&init_task); 6790 6791#endif /* CONFIG_CGROUP_SCHED */ 6792 6793 for_each_possible_cpu(i) { 6794 struct rq *rq; 6795 6796 rq = cpu_rq(i); 6797 raw_spin_lock_init(&rq->lock); 6798 rq->nr_running = 0; 6799 rq->calc_load_active = 0; 6800 rq->calc_load_update = jiffies + LOAD_FREQ; 6801 init_cfs_rq(&rq->cfs); 6802 init_rt_rq(&rq->rt, rq); 6803 init_dl_rq(&rq->dl, rq); 6804#ifdef CONFIG_FAIR_GROUP_SCHED 6805 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6806 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6807 /* 6808 * How much cpu bandwidth does root_task_group get? 6809 * 6810 * In case of task-groups formed thr' the cgroup filesystem, it 6811 * gets 100% of the cpu resources in the system. This overall 6812 * system cpu resource is divided among the tasks of 6813 * root_task_group and its child task-groups in a fair manner, 6814 * based on each entity's (task or task-group's) weight 6815 * (se->load.weight). 6816 * 6817 * In other words, if root_task_group has 10 tasks of weight 6818 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6819 * then A0's share of the cpu resource is: 6820 * 6821 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6822 * 6823 * We achieve this by letting root_task_group's tasks sit 6824 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6825 */ 6826 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6827 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6828#endif /* CONFIG_FAIR_GROUP_SCHED */ 6829 6830 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6831#ifdef CONFIG_RT_GROUP_SCHED 6832 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6833 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6834#endif 6835 6836 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6837 rq->cpu_load[j] = 0; 6838 6839 rq->last_load_update_tick = jiffies; 6840 6841#ifdef CONFIG_SMP 6842 rq->sd = NULL; 6843 rq->rd = NULL; 6844 rq->cpu_power = SCHED_POWER_SCALE; 6845 rq->post_schedule = 0; 6846 rq->active_balance = 0; 6847 rq->next_balance = jiffies; 6848 rq->push_cpu = 0; 6849 rq->cpu = i; 6850 rq->online = 0; 6851 rq->idle_stamp = 0; 6852 rq->avg_idle = 2*sysctl_sched_migration_cost; 6853 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6854 6855 INIT_LIST_HEAD(&rq->cfs_tasks); 6856 6857 rq_attach_root(rq, &def_root_domain); 6858#ifdef CONFIG_NO_HZ_COMMON 6859 rq->nohz_flags = 0; 6860#endif 6861#ifdef CONFIG_NO_HZ_FULL 6862 rq->last_sched_tick = 0; 6863#endif 6864#endif 6865 init_rq_hrtick(rq); 6866 atomic_set(&rq->nr_iowait, 0); 6867 } 6868 6869 set_load_weight(&init_task); 6870 6871#ifdef CONFIG_PREEMPT_NOTIFIERS 6872 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6873#endif 6874 6875 /* 6876 * The boot idle thread does lazy MMU switching as well: 6877 */ 6878 atomic_inc(&init_mm.mm_count); 6879 enter_lazy_tlb(&init_mm, current); 6880 6881 /* 6882 * Make us the idle thread. Technically, schedule() should not be 6883 * called from this thread, however somewhere below it might be, 6884 * but because we are the idle thread, we just pick up running again 6885 * when this runqueue becomes "idle". 6886 */ 6887 init_idle(current, smp_processor_id()); 6888 6889 calc_load_update = jiffies + LOAD_FREQ; 6890 6891 /* 6892 * During early bootup we pretend to be a normal task: 6893 */ 6894 current->sched_class = &fair_sched_class; 6895 6896#ifdef CONFIG_SMP 6897 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6898 /* May be allocated at isolcpus cmdline parse time */ 6899 if (cpu_isolated_map == NULL) 6900 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6901 idle_thread_set_boot_cpu(); 6902#endif 6903 init_sched_fair_class(); 6904 6905 scheduler_running = 1; 6906} 6907 6908#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6909static inline int preempt_count_equals(int preempt_offset) 6910{ 6911 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6912 6913 return (nested == preempt_offset); 6914} 6915 6916void __might_sleep(const char *file, int line, int preempt_offset) 6917{ 6918 static unsigned long prev_jiffy; /* ratelimiting */ 6919 6920 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6921 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6922 system_state != SYSTEM_RUNNING || oops_in_progress) 6923 return; 6924 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6925 return; 6926 prev_jiffy = jiffies; 6927 6928 printk(KERN_ERR 6929 "BUG: sleeping function called from invalid context at %s:%d\n", 6930 file, line); 6931 printk(KERN_ERR 6932 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6933 in_atomic(), irqs_disabled(), 6934 current->pid, current->comm); 6935 6936 debug_show_held_locks(current); 6937 if (irqs_disabled()) 6938 print_irqtrace_events(current); 6939 dump_stack(); 6940} 6941EXPORT_SYMBOL(__might_sleep); 6942#endif 6943 6944#ifdef CONFIG_MAGIC_SYSRQ 6945static void normalize_task(struct rq *rq, struct task_struct *p) 6946{ 6947 const struct sched_class *prev_class = p->sched_class; 6948 struct sched_attr attr = { 6949 .sched_policy = SCHED_NORMAL, 6950 }; 6951 int old_prio = p->prio; 6952 int on_rq; 6953 6954 on_rq = p->on_rq; 6955 if (on_rq) 6956 dequeue_task(rq, p, 0); 6957 __setscheduler(rq, p, &attr); 6958 if (on_rq) { 6959 enqueue_task(rq, p, 0); 6960 resched_task(rq->curr); 6961 } 6962 6963 check_class_changed(rq, p, prev_class, old_prio); 6964} 6965 6966void normalize_rt_tasks(void) 6967{ 6968 struct task_struct *g, *p; 6969 unsigned long flags; 6970 struct rq *rq; 6971 6972 read_lock_irqsave(&tasklist_lock, flags); 6973 do_each_thread(g, p) { 6974 /* 6975 * Only normalize user tasks: 6976 */ 6977 if (!p->mm) 6978 continue; 6979 6980 p->se.exec_start = 0; 6981#ifdef CONFIG_SCHEDSTATS 6982 p->se.statistics.wait_start = 0; 6983 p->se.statistics.sleep_start = 0; 6984 p->se.statistics.block_start = 0; 6985#endif 6986 6987 if (!dl_task(p) && !rt_task(p)) { 6988 /* 6989 * Renice negative nice level userspace 6990 * tasks back to 0: 6991 */ 6992 if (TASK_NICE(p) < 0 && p->mm) 6993 set_user_nice(p, 0); 6994 continue; 6995 } 6996 6997 raw_spin_lock(&p->pi_lock); 6998 rq = __task_rq_lock(p); 6999 7000 normalize_task(rq, p); 7001 7002 __task_rq_unlock(rq); 7003 raw_spin_unlock(&p->pi_lock); 7004 } while_each_thread(g, p); 7005 7006 read_unlock_irqrestore(&tasklist_lock, flags); 7007} 7008 7009#endif /* CONFIG_MAGIC_SYSRQ */ 7010 7011#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7012/* 7013 * These functions are only useful for the IA64 MCA handling, or kdb. 7014 * 7015 * They can only be called when the whole system has been 7016 * stopped - every CPU needs to be quiescent, and no scheduling 7017 * activity can take place. Using them for anything else would 7018 * be a serious bug, and as a result, they aren't even visible 7019 * under any other configuration. 7020 */ 7021 7022/** 7023 * curr_task - return the current task for a given cpu. 7024 * @cpu: the processor in question. 7025 * 7026 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7027 * 7028 * Return: The current task for @cpu. 7029 */ 7030struct task_struct *curr_task(int cpu) 7031{ 7032 return cpu_curr(cpu); 7033} 7034 7035#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7036 7037#ifdef CONFIG_IA64 7038/** 7039 * set_curr_task - set the current task for a given cpu. 7040 * @cpu: the processor in question. 7041 * @p: the task pointer to set. 7042 * 7043 * Description: This function must only be used when non-maskable interrupts 7044 * are serviced on a separate stack. It allows the architecture to switch the 7045 * notion of the current task on a cpu in a non-blocking manner. This function 7046 * must be called with all CPU's synchronized, and interrupts disabled, the 7047 * and caller must save the original value of the current task (see 7048 * curr_task() above) and restore that value before reenabling interrupts and 7049 * re-starting the system. 7050 * 7051 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7052 */ 7053void set_curr_task(int cpu, struct task_struct *p) 7054{ 7055 cpu_curr(cpu) = p; 7056} 7057 7058#endif 7059 7060#ifdef CONFIG_CGROUP_SCHED 7061/* task_group_lock serializes the addition/removal of task groups */ 7062static DEFINE_SPINLOCK(task_group_lock); 7063 7064static void free_sched_group(struct task_group *tg) 7065{ 7066 free_fair_sched_group(tg); 7067 free_rt_sched_group(tg); 7068 autogroup_free(tg); 7069 kfree(tg); 7070} 7071 7072/* allocate runqueue etc for a new task group */ 7073struct task_group *sched_create_group(struct task_group *parent) 7074{ 7075 struct task_group *tg; 7076 7077 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7078 if (!tg) 7079 return ERR_PTR(-ENOMEM); 7080 7081 if (!alloc_fair_sched_group(tg, parent)) 7082 goto err; 7083 7084 if (!alloc_rt_sched_group(tg, parent)) 7085 goto err; 7086 7087 return tg; 7088 7089err: 7090 free_sched_group(tg); 7091 return ERR_PTR(-ENOMEM); 7092} 7093 7094void sched_online_group(struct task_group *tg, struct task_group *parent) 7095{ 7096 unsigned long flags; 7097 7098 spin_lock_irqsave(&task_group_lock, flags); 7099 list_add_rcu(&tg->list, &task_groups); 7100 7101 WARN_ON(!parent); /* root should already exist */ 7102 7103 tg->parent = parent; 7104 INIT_LIST_HEAD(&tg->children); 7105 list_add_rcu(&tg->siblings, &parent->children); 7106 spin_unlock_irqrestore(&task_group_lock, flags); 7107} 7108 7109/* rcu callback to free various structures associated with a task group */ 7110static void free_sched_group_rcu(struct rcu_head *rhp) 7111{ 7112 /* now it should be safe to free those cfs_rqs */ 7113 free_sched_group(container_of(rhp, struct task_group, rcu)); 7114} 7115 7116/* Destroy runqueue etc associated with a task group */ 7117void sched_destroy_group(struct task_group *tg) 7118{ 7119 /* wait for possible concurrent references to cfs_rqs complete */ 7120 call_rcu(&tg->rcu, free_sched_group_rcu); 7121} 7122 7123void sched_offline_group(struct task_group *tg) 7124{ 7125 unsigned long flags; 7126 int i; 7127 7128 /* end participation in shares distribution */ 7129 for_each_possible_cpu(i) 7130 unregister_fair_sched_group(tg, i); 7131 7132 spin_lock_irqsave(&task_group_lock, flags); 7133 list_del_rcu(&tg->list); 7134 list_del_rcu(&tg->siblings); 7135 spin_unlock_irqrestore(&task_group_lock, flags); 7136} 7137 7138/* change task's runqueue when it moves between groups. 7139 * The caller of this function should have put the task in its new group 7140 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7141 * reflect its new group. 7142 */ 7143void sched_move_task(struct task_struct *tsk) 7144{ 7145 struct task_group *tg; 7146 int on_rq, running; 7147 unsigned long flags; 7148 struct rq *rq; 7149 7150 rq = task_rq_lock(tsk, &flags); 7151 7152 running = task_current(rq, tsk); 7153 on_rq = tsk->on_rq; 7154 7155 if (on_rq) 7156 dequeue_task(rq, tsk, 0); 7157 if (unlikely(running)) 7158 tsk->sched_class->put_prev_task(rq, tsk); 7159 7160 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 7161 lockdep_is_held(&tsk->sighand->siglock)), 7162 struct task_group, css); 7163 tg = autogroup_task_group(tsk, tg); 7164 tsk->sched_task_group = tg; 7165 7166#ifdef CONFIG_FAIR_GROUP_SCHED 7167 if (tsk->sched_class->task_move_group) 7168 tsk->sched_class->task_move_group(tsk, on_rq); 7169 else 7170#endif 7171 set_task_rq(tsk, task_cpu(tsk)); 7172 7173 if (unlikely(running)) 7174 tsk->sched_class->set_curr_task(rq); 7175 if (on_rq) 7176 enqueue_task(rq, tsk, 0); 7177 7178 task_rq_unlock(rq, tsk, &flags); 7179} 7180#endif /* CONFIG_CGROUP_SCHED */ 7181 7182#ifdef CONFIG_RT_GROUP_SCHED 7183/* 7184 * Ensure that the real time constraints are schedulable. 7185 */ 7186static DEFINE_MUTEX(rt_constraints_mutex); 7187 7188/* Must be called with tasklist_lock held */ 7189static inline int tg_has_rt_tasks(struct task_group *tg) 7190{ 7191 struct task_struct *g, *p; 7192 7193 do_each_thread(g, p) { 7194 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7195 return 1; 7196 } while_each_thread(g, p); 7197 7198 return 0; 7199} 7200 7201struct rt_schedulable_data { 7202 struct task_group *tg; 7203 u64 rt_period; 7204 u64 rt_runtime; 7205}; 7206 7207static int tg_rt_schedulable(struct task_group *tg, void *data) 7208{ 7209 struct rt_schedulable_data *d = data; 7210 struct task_group *child; 7211 unsigned long total, sum = 0; 7212 u64 period, runtime; 7213 7214 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7215 runtime = tg->rt_bandwidth.rt_runtime; 7216 7217 if (tg == d->tg) { 7218 period = d->rt_period; 7219 runtime = d->rt_runtime; 7220 } 7221 7222 /* 7223 * Cannot have more runtime than the period. 7224 */ 7225 if (runtime > period && runtime != RUNTIME_INF) 7226 return -EINVAL; 7227 7228 /* 7229 * Ensure we don't starve existing RT tasks. 7230 */ 7231 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7232 return -EBUSY; 7233 7234 total = to_ratio(period, runtime); 7235 7236 /* 7237 * Nobody can have more than the global setting allows. 7238 */ 7239 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7240 return -EINVAL; 7241 7242 /* 7243 * The sum of our children's runtime should not exceed our own. 7244 */ 7245 list_for_each_entry_rcu(child, &tg->children, siblings) { 7246 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7247 runtime = child->rt_bandwidth.rt_runtime; 7248 7249 if (child == d->tg) { 7250 period = d->rt_period; 7251 runtime = d->rt_runtime; 7252 } 7253 7254 sum += to_ratio(period, runtime); 7255 } 7256 7257 if (sum > total) 7258 return -EINVAL; 7259 7260 return 0; 7261} 7262 7263static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7264{ 7265 int ret; 7266 7267 struct rt_schedulable_data data = { 7268 .tg = tg, 7269 .rt_period = period, 7270 .rt_runtime = runtime, 7271 }; 7272 7273 rcu_read_lock(); 7274 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7275 rcu_read_unlock(); 7276 7277 return ret; 7278} 7279 7280static int tg_set_rt_bandwidth(struct task_group *tg, 7281 u64 rt_period, u64 rt_runtime) 7282{ 7283 int i, err = 0; 7284 7285 mutex_lock(&rt_constraints_mutex); 7286 read_lock(&tasklist_lock); 7287 err = __rt_schedulable(tg, rt_period, rt_runtime); 7288 if (err) 7289 goto unlock; 7290 7291 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7292 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7293 tg->rt_bandwidth.rt_runtime = rt_runtime; 7294 7295 for_each_possible_cpu(i) { 7296 struct rt_rq *rt_rq = tg->rt_rq[i]; 7297 7298 raw_spin_lock(&rt_rq->rt_runtime_lock); 7299 rt_rq->rt_runtime = rt_runtime; 7300 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7301 } 7302 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7303unlock: 7304 read_unlock(&tasklist_lock); 7305 mutex_unlock(&rt_constraints_mutex); 7306 7307 return err; 7308} 7309 7310static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7311{ 7312 u64 rt_runtime, rt_period; 7313 7314 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7315 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7316 if (rt_runtime_us < 0) 7317 rt_runtime = RUNTIME_INF; 7318 7319 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7320} 7321 7322static long sched_group_rt_runtime(struct task_group *tg) 7323{ 7324 u64 rt_runtime_us; 7325 7326 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7327 return -1; 7328 7329 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7330 do_div(rt_runtime_us, NSEC_PER_USEC); 7331 return rt_runtime_us; 7332} 7333 7334static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7335{ 7336 u64 rt_runtime, rt_period; 7337 7338 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7339 rt_runtime = tg->rt_bandwidth.rt_runtime; 7340 7341 if (rt_period == 0) 7342 return -EINVAL; 7343 7344 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7345} 7346 7347static long sched_group_rt_period(struct task_group *tg) 7348{ 7349 u64 rt_period_us; 7350 7351 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7352 do_div(rt_period_us, NSEC_PER_USEC); 7353 return rt_period_us; 7354} 7355#endif /* CONFIG_RT_GROUP_SCHED */ 7356 7357#ifdef CONFIG_RT_GROUP_SCHED 7358static int sched_rt_global_constraints(void) 7359{ 7360 int ret = 0; 7361 7362 mutex_lock(&rt_constraints_mutex); 7363 read_lock(&tasklist_lock); 7364 ret = __rt_schedulable(NULL, 0, 0); 7365 read_unlock(&tasklist_lock); 7366 mutex_unlock(&rt_constraints_mutex); 7367 7368 return ret; 7369} 7370 7371static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7372{ 7373 /* Don't accept realtime tasks when there is no way for them to run */ 7374 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7375 return 0; 7376 7377 return 1; 7378} 7379 7380#else /* !CONFIG_RT_GROUP_SCHED */ 7381static int sched_rt_global_constraints(void) 7382{ 7383 unsigned long flags; 7384 int i, ret = 0; 7385 7386 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7387 for_each_possible_cpu(i) { 7388 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7389 7390 raw_spin_lock(&rt_rq->rt_runtime_lock); 7391 rt_rq->rt_runtime = global_rt_runtime(); 7392 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7393 } 7394 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7395 7396 return ret; 7397} 7398#endif /* CONFIG_RT_GROUP_SCHED */ 7399 7400static int sched_dl_global_constraints(void) 7401{ 7402 u64 runtime = global_rt_runtime(); 7403 u64 period = global_rt_period(); 7404 u64 new_bw = to_ratio(period, runtime); 7405 int cpu, ret = 0; 7406 7407 /* 7408 * Here we want to check the bandwidth not being set to some 7409 * value smaller than the currently allocated bandwidth in 7410 * any of the root_domains. 7411 * 7412 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7413 * cycling on root_domains... Discussion on different/better 7414 * solutions is welcome! 7415 */ 7416 for_each_possible_cpu(cpu) { 7417 struct dl_bw *dl_b = dl_bw_of(cpu); 7418 7419 raw_spin_lock(&dl_b->lock); 7420 if (new_bw < dl_b->total_bw) 7421 ret = -EBUSY; 7422 raw_spin_unlock(&dl_b->lock); 7423 7424 if (ret) 7425 break; 7426 } 7427 7428 return ret; 7429} 7430 7431static void sched_dl_do_global(void) 7432{ 7433 u64 new_bw = -1; 7434 int cpu; 7435 7436 def_dl_bandwidth.dl_period = global_rt_period(); 7437 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7438 7439 if (global_rt_runtime() != RUNTIME_INF) 7440 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7441 7442 /* 7443 * FIXME: As above... 7444 */ 7445 for_each_possible_cpu(cpu) { 7446 struct dl_bw *dl_b = dl_bw_of(cpu); 7447 7448 raw_spin_lock(&dl_b->lock); 7449 dl_b->bw = new_bw; 7450 raw_spin_unlock(&dl_b->lock); 7451 } 7452} 7453 7454static int sched_rt_global_validate(void) 7455{ 7456 if (sysctl_sched_rt_period <= 0) 7457 return -EINVAL; 7458 7459 if (sysctl_sched_rt_runtime > sysctl_sched_rt_period) 7460 return -EINVAL; 7461 7462 return 0; 7463} 7464 7465static void sched_rt_do_global(void) 7466{ 7467 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7468 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7469} 7470 7471int sched_rt_handler(struct ctl_table *table, int write, 7472 void __user *buffer, size_t *lenp, 7473 loff_t *ppos) 7474{ 7475 int old_period, old_runtime; 7476 static DEFINE_MUTEX(mutex); 7477 int ret; 7478 7479 mutex_lock(&mutex); 7480 old_period = sysctl_sched_rt_period; 7481 old_runtime = sysctl_sched_rt_runtime; 7482 7483 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7484 7485 if (!ret && write) { 7486 ret = sched_rt_global_validate(); 7487 if (ret) 7488 goto undo; 7489 7490 ret = sched_rt_global_constraints(); 7491 if (ret) 7492 goto undo; 7493 7494 ret = sched_dl_global_constraints(); 7495 if (ret) 7496 goto undo; 7497 7498 sched_rt_do_global(); 7499 sched_dl_do_global(); 7500 } 7501 if (0) { 7502undo: 7503 sysctl_sched_rt_period = old_period; 7504 sysctl_sched_rt_runtime = old_runtime; 7505 } 7506 mutex_unlock(&mutex); 7507 7508 return ret; 7509} 7510 7511int sched_rr_handler(struct ctl_table *table, int write, 7512 void __user *buffer, size_t *lenp, 7513 loff_t *ppos) 7514{ 7515 int ret; 7516 static DEFINE_MUTEX(mutex); 7517 7518 mutex_lock(&mutex); 7519 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7520 /* make sure that internally we keep jiffies */ 7521 /* also, writing zero resets timeslice to default */ 7522 if (!ret && write) { 7523 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7524 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7525 } 7526 mutex_unlock(&mutex); 7527 return ret; 7528} 7529 7530#ifdef CONFIG_CGROUP_SCHED 7531 7532static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7533{ 7534 return css ? container_of(css, struct task_group, css) : NULL; 7535} 7536 7537static struct cgroup_subsys_state * 7538cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7539{ 7540 struct task_group *parent = css_tg(parent_css); 7541 struct task_group *tg; 7542 7543 if (!parent) { 7544 /* This is early initialization for the top cgroup */ 7545 return &root_task_group.css; 7546 } 7547 7548 tg = sched_create_group(parent); 7549 if (IS_ERR(tg)) 7550 return ERR_PTR(-ENOMEM); 7551 7552 return &tg->css; 7553} 7554 7555static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7556{ 7557 struct task_group *tg = css_tg(css); 7558 struct task_group *parent = css_tg(css_parent(css)); 7559 7560 if (parent) 7561 sched_online_group(tg, parent); 7562 return 0; 7563} 7564 7565static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7566{ 7567 struct task_group *tg = css_tg(css); 7568 7569 sched_destroy_group(tg); 7570} 7571 7572static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7573{ 7574 struct task_group *tg = css_tg(css); 7575 7576 sched_offline_group(tg); 7577} 7578 7579static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7580 struct cgroup_taskset *tset) 7581{ 7582 struct task_struct *task; 7583 7584 cgroup_taskset_for_each(task, css, tset) { 7585#ifdef CONFIG_RT_GROUP_SCHED 7586 if (!sched_rt_can_attach(css_tg(css), task)) 7587 return -EINVAL; 7588#else 7589 /* We don't support RT-tasks being in separate groups */ 7590 if (task->sched_class != &fair_sched_class) 7591 return -EINVAL; 7592#endif 7593 } 7594 return 0; 7595} 7596 7597static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7598 struct cgroup_taskset *tset) 7599{ 7600 struct task_struct *task; 7601 7602 cgroup_taskset_for_each(task, css, tset) 7603 sched_move_task(task); 7604} 7605 7606static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7607 struct cgroup_subsys_state *old_css, 7608 struct task_struct *task) 7609{ 7610 /* 7611 * cgroup_exit() is called in the copy_process() failure path. 7612 * Ignore this case since the task hasn't ran yet, this avoids 7613 * trying to poke a half freed task state from generic code. 7614 */ 7615 if (!(task->flags & PF_EXITING)) 7616 return; 7617 7618 sched_move_task(task); 7619} 7620 7621#ifdef CONFIG_FAIR_GROUP_SCHED 7622static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7623 struct cftype *cftype, u64 shareval) 7624{ 7625 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7626} 7627 7628static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7629 struct cftype *cft) 7630{ 7631 struct task_group *tg = css_tg(css); 7632 7633 return (u64) scale_load_down(tg->shares); 7634} 7635 7636#ifdef CONFIG_CFS_BANDWIDTH 7637static DEFINE_MUTEX(cfs_constraints_mutex); 7638 7639const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7640const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7641 7642static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7643 7644static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7645{ 7646 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7647 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7648 7649 if (tg == &root_task_group) 7650 return -EINVAL; 7651 7652 /* 7653 * Ensure we have at some amount of bandwidth every period. This is 7654 * to prevent reaching a state of large arrears when throttled via 7655 * entity_tick() resulting in prolonged exit starvation. 7656 */ 7657 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7658 return -EINVAL; 7659 7660 /* 7661 * Likewise, bound things on the otherside by preventing insane quota 7662 * periods. This also allows us to normalize in computing quota 7663 * feasibility. 7664 */ 7665 if (period > max_cfs_quota_period) 7666 return -EINVAL; 7667 7668 mutex_lock(&cfs_constraints_mutex); 7669 ret = __cfs_schedulable(tg, period, quota); 7670 if (ret) 7671 goto out_unlock; 7672 7673 runtime_enabled = quota != RUNTIME_INF; 7674 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7675 /* 7676 * If we need to toggle cfs_bandwidth_used, off->on must occur 7677 * before making related changes, and on->off must occur afterwards 7678 */ 7679 if (runtime_enabled && !runtime_was_enabled) 7680 cfs_bandwidth_usage_inc(); 7681 raw_spin_lock_irq(&cfs_b->lock); 7682 cfs_b->period = ns_to_ktime(period); 7683 cfs_b->quota = quota; 7684 7685 __refill_cfs_bandwidth_runtime(cfs_b); 7686 /* restart the period timer (if active) to handle new period expiry */ 7687 if (runtime_enabled && cfs_b->timer_active) { 7688 /* force a reprogram */ 7689 cfs_b->timer_active = 0; 7690 __start_cfs_bandwidth(cfs_b); 7691 } 7692 raw_spin_unlock_irq(&cfs_b->lock); 7693 7694 for_each_possible_cpu(i) { 7695 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7696 struct rq *rq = cfs_rq->rq; 7697 7698 raw_spin_lock_irq(&rq->lock); 7699 cfs_rq->runtime_enabled = runtime_enabled; 7700 cfs_rq->runtime_remaining = 0; 7701 7702 if (cfs_rq->throttled) 7703 unthrottle_cfs_rq(cfs_rq); 7704 raw_spin_unlock_irq(&rq->lock); 7705 } 7706 if (runtime_was_enabled && !runtime_enabled) 7707 cfs_bandwidth_usage_dec(); 7708out_unlock: 7709 mutex_unlock(&cfs_constraints_mutex); 7710 7711 return ret; 7712} 7713 7714int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7715{ 7716 u64 quota, period; 7717 7718 period = ktime_to_ns(tg->cfs_bandwidth.period); 7719 if (cfs_quota_us < 0) 7720 quota = RUNTIME_INF; 7721 else 7722 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7723 7724 return tg_set_cfs_bandwidth(tg, period, quota); 7725} 7726 7727long tg_get_cfs_quota(struct task_group *tg) 7728{ 7729 u64 quota_us; 7730 7731 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7732 return -1; 7733 7734 quota_us = tg->cfs_bandwidth.quota; 7735 do_div(quota_us, NSEC_PER_USEC); 7736 7737 return quota_us; 7738} 7739 7740int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7741{ 7742 u64 quota, period; 7743 7744 period = (u64)cfs_period_us * NSEC_PER_USEC; 7745 quota = tg->cfs_bandwidth.quota; 7746 7747 return tg_set_cfs_bandwidth(tg, period, quota); 7748} 7749 7750long tg_get_cfs_period(struct task_group *tg) 7751{ 7752 u64 cfs_period_us; 7753 7754 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7755 do_div(cfs_period_us, NSEC_PER_USEC); 7756 7757 return cfs_period_us; 7758} 7759 7760static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7761 struct cftype *cft) 7762{ 7763 return tg_get_cfs_quota(css_tg(css)); 7764} 7765 7766static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7767 struct cftype *cftype, s64 cfs_quota_us) 7768{ 7769 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7770} 7771 7772static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7773 struct cftype *cft) 7774{ 7775 return tg_get_cfs_period(css_tg(css)); 7776} 7777 7778static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7779 struct cftype *cftype, u64 cfs_period_us) 7780{ 7781 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7782} 7783 7784struct cfs_schedulable_data { 7785 struct task_group *tg; 7786 u64 period, quota; 7787}; 7788 7789/* 7790 * normalize group quota/period to be quota/max_period 7791 * note: units are usecs 7792 */ 7793static u64 normalize_cfs_quota(struct task_group *tg, 7794 struct cfs_schedulable_data *d) 7795{ 7796 u64 quota, period; 7797 7798 if (tg == d->tg) { 7799 period = d->period; 7800 quota = d->quota; 7801 } else { 7802 period = tg_get_cfs_period(tg); 7803 quota = tg_get_cfs_quota(tg); 7804 } 7805 7806 /* note: these should typically be equivalent */ 7807 if (quota == RUNTIME_INF || quota == -1) 7808 return RUNTIME_INF; 7809 7810 return to_ratio(period, quota); 7811} 7812 7813static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7814{ 7815 struct cfs_schedulable_data *d = data; 7816 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7817 s64 quota = 0, parent_quota = -1; 7818 7819 if (!tg->parent) { 7820 quota = RUNTIME_INF; 7821 } else { 7822 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7823 7824 quota = normalize_cfs_quota(tg, d); 7825 parent_quota = parent_b->hierarchal_quota; 7826 7827 /* 7828 * ensure max(child_quota) <= parent_quota, inherit when no 7829 * limit is set 7830 */ 7831 if (quota == RUNTIME_INF) 7832 quota = parent_quota; 7833 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7834 return -EINVAL; 7835 } 7836 cfs_b->hierarchal_quota = quota; 7837 7838 return 0; 7839} 7840 7841static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7842{ 7843 int ret; 7844 struct cfs_schedulable_data data = { 7845 .tg = tg, 7846 .period = period, 7847 .quota = quota, 7848 }; 7849 7850 if (quota != RUNTIME_INF) { 7851 do_div(data.period, NSEC_PER_USEC); 7852 do_div(data.quota, NSEC_PER_USEC); 7853 } 7854 7855 rcu_read_lock(); 7856 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7857 rcu_read_unlock(); 7858 7859 return ret; 7860} 7861 7862static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft, 7863 struct cgroup_map_cb *cb) 7864{ 7865 struct task_group *tg = css_tg(css); 7866 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7867 7868 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7869 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7870 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7871 7872 return 0; 7873} 7874#endif /* CONFIG_CFS_BANDWIDTH */ 7875#endif /* CONFIG_FAIR_GROUP_SCHED */ 7876 7877#ifdef CONFIG_RT_GROUP_SCHED 7878static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7879 struct cftype *cft, s64 val) 7880{ 7881 return sched_group_set_rt_runtime(css_tg(css), val); 7882} 7883 7884static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7885 struct cftype *cft) 7886{ 7887 return sched_group_rt_runtime(css_tg(css)); 7888} 7889 7890static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7891 struct cftype *cftype, u64 rt_period_us) 7892{ 7893 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7894} 7895 7896static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7897 struct cftype *cft) 7898{ 7899 return sched_group_rt_period(css_tg(css)); 7900} 7901#endif /* CONFIG_RT_GROUP_SCHED */ 7902 7903static struct cftype cpu_files[] = { 7904#ifdef CONFIG_FAIR_GROUP_SCHED 7905 { 7906 .name = "shares", 7907 .read_u64 = cpu_shares_read_u64, 7908 .write_u64 = cpu_shares_write_u64, 7909 }, 7910#endif 7911#ifdef CONFIG_CFS_BANDWIDTH 7912 { 7913 .name = "cfs_quota_us", 7914 .read_s64 = cpu_cfs_quota_read_s64, 7915 .write_s64 = cpu_cfs_quota_write_s64, 7916 }, 7917 { 7918 .name = "cfs_period_us", 7919 .read_u64 = cpu_cfs_period_read_u64, 7920 .write_u64 = cpu_cfs_period_write_u64, 7921 }, 7922 { 7923 .name = "stat", 7924 .read_map = cpu_stats_show, 7925 }, 7926#endif 7927#ifdef CONFIG_RT_GROUP_SCHED 7928 { 7929 .name = "rt_runtime_us", 7930 .read_s64 = cpu_rt_runtime_read, 7931 .write_s64 = cpu_rt_runtime_write, 7932 }, 7933 { 7934 .name = "rt_period_us", 7935 .read_u64 = cpu_rt_period_read_uint, 7936 .write_u64 = cpu_rt_period_write_uint, 7937 }, 7938#endif 7939 { } /* terminate */ 7940}; 7941 7942struct cgroup_subsys cpu_cgroup_subsys = { 7943 .name = "cpu", 7944 .css_alloc = cpu_cgroup_css_alloc, 7945 .css_free = cpu_cgroup_css_free, 7946 .css_online = cpu_cgroup_css_online, 7947 .css_offline = cpu_cgroup_css_offline, 7948 .can_attach = cpu_cgroup_can_attach, 7949 .attach = cpu_cgroup_attach, 7950 .exit = cpu_cgroup_exit, 7951 .subsys_id = cpu_cgroup_subsys_id, 7952 .base_cftypes = cpu_files, 7953 .early_init = 1, 7954}; 7955 7956#endif /* CONFIG_CGROUP_SCHED */ 7957 7958void dump_cpu_task(int cpu) 7959{ 7960 pr_info("Task dump for CPU %d:\n", cpu); 7961 sched_show_task(cpu_curr(cpu)); 7962} 7963