core.c revision ff1df896aef8e0ec1556a5c44f424bd45bfa2cbe
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75#include <linux/context_tracking.h> 76 77#include <asm/switch_to.h> 78#include <asm/tlb.h> 79#include <asm/irq_regs.h> 80#include <asm/mutex.h> 81#ifdef CONFIG_PARAVIRT 82#include <asm/paravirt.h> 83#endif 84 85#include "sched.h" 86#include "../workqueue_internal.h" 87#include "../smpboot.h" 88 89#define CREATE_TRACE_POINTS 90#include <trace/events/sched.h> 91 92void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 93{ 94 unsigned long delta; 95 ktime_t soft, hard, now; 96 97 for (;;) { 98 if (hrtimer_active(period_timer)) 99 break; 100 101 now = hrtimer_cb_get_time(period_timer); 102 hrtimer_forward(period_timer, now, period); 103 104 soft = hrtimer_get_softexpires(period_timer); 105 hard = hrtimer_get_expires(period_timer); 106 delta = ktime_to_ns(ktime_sub(hard, soft)); 107 __hrtimer_start_range_ns(period_timer, soft, delta, 108 HRTIMER_MODE_ABS_PINNED, 0); 109 } 110} 111 112DEFINE_MUTEX(sched_domains_mutex); 113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115static void update_rq_clock_task(struct rq *rq, s64 delta); 116 117void update_rq_clock(struct rq *rq) 118{ 119 s64 delta; 120 121 if (rq->skip_clock_update > 0) 122 return; 123 124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 125 rq->clock += delta; 126 update_rq_clock_task(rq, delta); 127} 128 129/* 130 * Debugging: various feature bits 131 */ 132 133#define SCHED_FEAT(name, enabled) \ 134 (1UL << __SCHED_FEAT_##name) * enabled | 135 136const_debug unsigned int sysctl_sched_features = 137#include "features.h" 138 0; 139 140#undef SCHED_FEAT 141 142#ifdef CONFIG_SCHED_DEBUG 143#define SCHED_FEAT(name, enabled) \ 144 #name , 145 146static const char * const sched_feat_names[] = { 147#include "features.h" 148}; 149 150#undef SCHED_FEAT 151 152static int sched_feat_show(struct seq_file *m, void *v) 153{ 154 int i; 155 156 for (i = 0; i < __SCHED_FEAT_NR; i++) { 157 if (!(sysctl_sched_features & (1UL << i))) 158 seq_puts(m, "NO_"); 159 seq_printf(m, "%s ", sched_feat_names[i]); 160 } 161 seq_puts(m, "\n"); 162 163 return 0; 164} 165 166#ifdef HAVE_JUMP_LABEL 167 168#define jump_label_key__true STATIC_KEY_INIT_TRUE 169#define jump_label_key__false STATIC_KEY_INIT_FALSE 170 171#define SCHED_FEAT(name, enabled) \ 172 jump_label_key__##enabled , 173 174struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 175#include "features.h" 176}; 177 178#undef SCHED_FEAT 179 180static void sched_feat_disable(int i) 181{ 182 if (static_key_enabled(&sched_feat_keys[i])) 183 static_key_slow_dec(&sched_feat_keys[i]); 184} 185 186static void sched_feat_enable(int i) 187{ 188 if (!static_key_enabled(&sched_feat_keys[i])) 189 static_key_slow_inc(&sched_feat_keys[i]); 190} 191#else 192static void sched_feat_disable(int i) { }; 193static void sched_feat_enable(int i) { }; 194#endif /* HAVE_JUMP_LABEL */ 195 196static int sched_feat_set(char *cmp) 197{ 198 int i; 199 int neg = 0; 200 201 if (strncmp(cmp, "NO_", 3) == 0) { 202 neg = 1; 203 cmp += 3; 204 } 205 206 for (i = 0; i < __SCHED_FEAT_NR; i++) { 207 if (strcmp(cmp, sched_feat_names[i]) == 0) { 208 if (neg) { 209 sysctl_sched_features &= ~(1UL << i); 210 sched_feat_disable(i); 211 } else { 212 sysctl_sched_features |= (1UL << i); 213 sched_feat_enable(i); 214 } 215 break; 216 } 217 } 218 219 return i; 220} 221 222static ssize_t 223sched_feat_write(struct file *filp, const char __user *ubuf, 224 size_t cnt, loff_t *ppos) 225{ 226 char buf[64]; 227 char *cmp; 228 int i; 229 230 if (cnt > 63) 231 cnt = 63; 232 233 if (copy_from_user(&buf, ubuf, cnt)) 234 return -EFAULT; 235 236 buf[cnt] = 0; 237 cmp = strstrip(buf); 238 239 i = sched_feat_set(cmp); 240 if (i == __SCHED_FEAT_NR) 241 return -EINVAL; 242 243 *ppos += cnt; 244 245 return cnt; 246} 247 248static int sched_feat_open(struct inode *inode, struct file *filp) 249{ 250 return single_open(filp, sched_feat_show, NULL); 251} 252 253static const struct file_operations sched_feat_fops = { 254 .open = sched_feat_open, 255 .write = sched_feat_write, 256 .read = seq_read, 257 .llseek = seq_lseek, 258 .release = single_release, 259}; 260 261static __init int sched_init_debug(void) 262{ 263 debugfs_create_file("sched_features", 0644, NULL, NULL, 264 &sched_feat_fops); 265 266 return 0; 267} 268late_initcall(sched_init_debug); 269#endif /* CONFIG_SCHED_DEBUG */ 270 271/* 272 * Number of tasks to iterate in a single balance run. 273 * Limited because this is done with IRQs disabled. 274 */ 275const_debug unsigned int sysctl_sched_nr_migrate = 32; 276 277/* 278 * period over which we average the RT time consumption, measured 279 * in ms. 280 * 281 * default: 1s 282 */ 283const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 284 285/* 286 * period over which we measure -rt task cpu usage in us. 287 * default: 1s 288 */ 289unsigned int sysctl_sched_rt_period = 1000000; 290 291__read_mostly int scheduler_running; 292 293/* 294 * part of the period that we allow rt tasks to run in us. 295 * default: 0.95s 296 */ 297int sysctl_sched_rt_runtime = 950000; 298 299/* 300 * __task_rq_lock - lock the rq @p resides on. 301 */ 302static inline struct rq *__task_rq_lock(struct task_struct *p) 303 __acquires(rq->lock) 304{ 305 struct rq *rq; 306 307 lockdep_assert_held(&p->pi_lock); 308 309 for (;;) { 310 rq = task_rq(p); 311 raw_spin_lock(&rq->lock); 312 if (likely(rq == task_rq(p))) 313 return rq; 314 raw_spin_unlock(&rq->lock); 315 } 316} 317 318/* 319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 320 */ 321static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 322 __acquires(p->pi_lock) 323 __acquires(rq->lock) 324{ 325 struct rq *rq; 326 327 for (;;) { 328 raw_spin_lock_irqsave(&p->pi_lock, *flags); 329 rq = task_rq(p); 330 raw_spin_lock(&rq->lock); 331 if (likely(rq == task_rq(p))) 332 return rq; 333 raw_spin_unlock(&rq->lock); 334 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 335 } 336} 337 338static void __task_rq_unlock(struct rq *rq) 339 __releases(rq->lock) 340{ 341 raw_spin_unlock(&rq->lock); 342} 343 344static inline void 345task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 346 __releases(rq->lock) 347 __releases(p->pi_lock) 348{ 349 raw_spin_unlock(&rq->lock); 350 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 351} 352 353/* 354 * this_rq_lock - lock this runqueue and disable interrupts. 355 */ 356static struct rq *this_rq_lock(void) 357 __acquires(rq->lock) 358{ 359 struct rq *rq; 360 361 local_irq_disable(); 362 rq = this_rq(); 363 raw_spin_lock(&rq->lock); 364 365 return rq; 366} 367 368#ifdef CONFIG_SCHED_HRTICK 369/* 370 * Use HR-timers to deliver accurate preemption points. 371 */ 372 373static void hrtick_clear(struct rq *rq) 374{ 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377} 378 379/* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383static enum hrtimer_restart hrtick(struct hrtimer *timer) 384{ 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395} 396 397#ifdef CONFIG_SMP 398 399static int __hrtick_restart(struct rq *rq) 400{ 401 struct hrtimer *timer = &rq->hrtick_timer; 402 ktime_t time = hrtimer_get_softexpires(timer); 403 404 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 405} 406 407/* 408 * called from hardirq (IPI) context 409 */ 410static void __hrtick_start(void *arg) 411{ 412 struct rq *rq = arg; 413 414 raw_spin_lock(&rq->lock); 415 __hrtick_restart(rq); 416 rq->hrtick_csd_pending = 0; 417 raw_spin_unlock(&rq->lock); 418} 419 420/* 421 * Called to set the hrtick timer state. 422 * 423 * called with rq->lock held and irqs disabled 424 */ 425void hrtick_start(struct rq *rq, u64 delay) 426{ 427 struct hrtimer *timer = &rq->hrtick_timer; 428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 429 430 hrtimer_set_expires(timer, time); 431 432 if (rq == this_rq()) { 433 __hrtick_restart(rq); 434 } else if (!rq->hrtick_csd_pending) { 435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 436 rq->hrtick_csd_pending = 1; 437 } 438} 439 440static int 441hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 442{ 443 int cpu = (int)(long)hcpu; 444 445 switch (action) { 446 case CPU_UP_CANCELED: 447 case CPU_UP_CANCELED_FROZEN: 448 case CPU_DOWN_PREPARE: 449 case CPU_DOWN_PREPARE_FROZEN: 450 case CPU_DEAD: 451 case CPU_DEAD_FROZEN: 452 hrtick_clear(cpu_rq(cpu)); 453 return NOTIFY_OK; 454 } 455 456 return NOTIFY_DONE; 457} 458 459static __init void init_hrtick(void) 460{ 461 hotcpu_notifier(hotplug_hrtick, 0); 462} 463#else 464/* 465 * Called to set the hrtick timer state. 466 * 467 * called with rq->lock held and irqs disabled 468 */ 469void hrtick_start(struct rq *rq, u64 delay) 470{ 471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 472 HRTIMER_MODE_REL_PINNED, 0); 473} 474 475static inline void init_hrtick(void) 476{ 477} 478#endif /* CONFIG_SMP */ 479 480static void init_rq_hrtick(struct rq *rq) 481{ 482#ifdef CONFIG_SMP 483 rq->hrtick_csd_pending = 0; 484 485 rq->hrtick_csd.flags = 0; 486 rq->hrtick_csd.func = __hrtick_start; 487 rq->hrtick_csd.info = rq; 488#endif 489 490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 491 rq->hrtick_timer.function = hrtick; 492} 493#else /* CONFIG_SCHED_HRTICK */ 494static inline void hrtick_clear(struct rq *rq) 495{ 496} 497 498static inline void init_rq_hrtick(struct rq *rq) 499{ 500} 501 502static inline void init_hrtick(void) 503{ 504} 505#endif /* CONFIG_SCHED_HRTICK */ 506 507/* 508 * resched_task - mark a task 'to be rescheduled now'. 509 * 510 * On UP this means the setting of the need_resched flag, on SMP it 511 * might also involve a cross-CPU call to trigger the scheduler on 512 * the target CPU. 513 */ 514void resched_task(struct task_struct *p) 515{ 516 int cpu; 517 518 lockdep_assert_held(&task_rq(p)->lock); 519 520 if (test_tsk_need_resched(p)) 521 return; 522 523 set_tsk_need_resched(p); 524 525 cpu = task_cpu(p); 526 if (cpu == smp_processor_id()) { 527 set_preempt_need_resched(); 528 return; 529 } 530 531 /* NEED_RESCHED must be visible before we test polling */ 532 smp_mb(); 533 if (!tsk_is_polling(p)) 534 smp_send_reschedule(cpu); 535} 536 537void resched_cpu(int cpu) 538{ 539 struct rq *rq = cpu_rq(cpu); 540 unsigned long flags; 541 542 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 543 return; 544 resched_task(cpu_curr(cpu)); 545 raw_spin_unlock_irqrestore(&rq->lock, flags); 546} 547 548#ifdef CONFIG_SMP 549#ifdef CONFIG_NO_HZ_COMMON 550/* 551 * In the semi idle case, use the nearest busy cpu for migrating timers 552 * from an idle cpu. This is good for power-savings. 553 * 554 * We don't do similar optimization for completely idle system, as 555 * selecting an idle cpu will add more delays to the timers than intended 556 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 557 */ 558int get_nohz_timer_target(void) 559{ 560 int cpu = smp_processor_id(); 561 int i; 562 struct sched_domain *sd; 563 564 rcu_read_lock(); 565 for_each_domain(cpu, sd) { 566 for_each_cpu(i, sched_domain_span(sd)) { 567 if (!idle_cpu(i)) { 568 cpu = i; 569 goto unlock; 570 } 571 } 572 } 573unlock: 574 rcu_read_unlock(); 575 return cpu; 576} 577/* 578 * When add_timer_on() enqueues a timer into the timer wheel of an 579 * idle CPU then this timer might expire before the next timer event 580 * which is scheduled to wake up that CPU. In case of a completely 581 * idle system the next event might even be infinite time into the 582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 583 * leaves the inner idle loop so the newly added timer is taken into 584 * account when the CPU goes back to idle and evaluates the timer 585 * wheel for the next timer event. 586 */ 587static void wake_up_idle_cpu(int cpu) 588{ 589 struct rq *rq = cpu_rq(cpu); 590 591 if (cpu == smp_processor_id()) 592 return; 593 594 /* 595 * This is safe, as this function is called with the timer 596 * wheel base lock of (cpu) held. When the CPU is on the way 597 * to idle and has not yet set rq->curr to idle then it will 598 * be serialized on the timer wheel base lock and take the new 599 * timer into account automatically. 600 */ 601 if (rq->curr != rq->idle) 602 return; 603 604 /* 605 * We can set TIF_RESCHED on the idle task of the other CPU 606 * lockless. The worst case is that the other CPU runs the 607 * idle task through an additional NOOP schedule() 608 */ 609 set_tsk_need_resched(rq->idle); 610 611 /* NEED_RESCHED must be visible before we test polling */ 612 smp_mb(); 613 if (!tsk_is_polling(rq->idle)) 614 smp_send_reschedule(cpu); 615} 616 617static bool wake_up_full_nohz_cpu(int cpu) 618{ 619 if (tick_nohz_full_cpu(cpu)) { 620 if (cpu != smp_processor_id() || 621 tick_nohz_tick_stopped()) 622 smp_send_reschedule(cpu); 623 return true; 624 } 625 626 return false; 627} 628 629void wake_up_nohz_cpu(int cpu) 630{ 631 if (!wake_up_full_nohz_cpu(cpu)) 632 wake_up_idle_cpu(cpu); 633} 634 635static inline bool got_nohz_idle_kick(void) 636{ 637 int cpu = smp_processor_id(); 638 639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 640 return false; 641 642 if (idle_cpu(cpu) && !need_resched()) 643 return true; 644 645 /* 646 * We can't run Idle Load Balance on this CPU for this time so we 647 * cancel it and clear NOHZ_BALANCE_KICK 648 */ 649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 650 return false; 651} 652 653#else /* CONFIG_NO_HZ_COMMON */ 654 655static inline bool got_nohz_idle_kick(void) 656{ 657 return false; 658} 659 660#endif /* CONFIG_NO_HZ_COMMON */ 661 662#ifdef CONFIG_NO_HZ_FULL 663bool sched_can_stop_tick(void) 664{ 665 struct rq *rq; 666 667 rq = this_rq(); 668 669 /* Make sure rq->nr_running update is visible after the IPI */ 670 smp_rmb(); 671 672 /* More than one running task need preemption */ 673 if (rq->nr_running > 1) 674 return false; 675 676 return true; 677} 678#endif /* CONFIG_NO_HZ_FULL */ 679 680void sched_avg_update(struct rq *rq) 681{ 682 s64 period = sched_avg_period(); 683 684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 685 /* 686 * Inline assembly required to prevent the compiler 687 * optimising this loop into a divmod call. 688 * See __iter_div_u64_rem() for another example of this. 689 */ 690 asm("" : "+rm" (rq->age_stamp)); 691 rq->age_stamp += period; 692 rq->rt_avg /= 2; 693 } 694} 695 696#endif /* CONFIG_SMP */ 697 698#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 699 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 700/* 701 * Iterate task_group tree rooted at *from, calling @down when first entering a 702 * node and @up when leaving it for the final time. 703 * 704 * Caller must hold rcu_lock or sufficient equivalent. 705 */ 706int walk_tg_tree_from(struct task_group *from, 707 tg_visitor down, tg_visitor up, void *data) 708{ 709 struct task_group *parent, *child; 710 int ret; 711 712 parent = from; 713 714down: 715 ret = (*down)(parent, data); 716 if (ret) 717 goto out; 718 list_for_each_entry_rcu(child, &parent->children, siblings) { 719 parent = child; 720 goto down; 721 722up: 723 continue; 724 } 725 ret = (*up)(parent, data); 726 if (ret || parent == from) 727 goto out; 728 729 child = parent; 730 parent = parent->parent; 731 if (parent) 732 goto up; 733out: 734 return ret; 735} 736 737int tg_nop(struct task_group *tg, void *data) 738{ 739 return 0; 740} 741#endif 742 743static void set_load_weight(struct task_struct *p) 744{ 745 int prio = p->static_prio - MAX_RT_PRIO; 746 struct load_weight *load = &p->se.load; 747 748 /* 749 * SCHED_IDLE tasks get minimal weight: 750 */ 751 if (p->policy == SCHED_IDLE) { 752 load->weight = scale_load(WEIGHT_IDLEPRIO); 753 load->inv_weight = WMULT_IDLEPRIO; 754 return; 755 } 756 757 load->weight = scale_load(prio_to_weight[prio]); 758 load->inv_weight = prio_to_wmult[prio]; 759} 760 761static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 762{ 763 update_rq_clock(rq); 764 sched_info_queued(rq, p); 765 p->sched_class->enqueue_task(rq, p, flags); 766} 767 768static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 769{ 770 update_rq_clock(rq); 771 sched_info_dequeued(rq, p); 772 p->sched_class->dequeue_task(rq, p, flags); 773} 774 775void activate_task(struct rq *rq, struct task_struct *p, int flags) 776{ 777 if (task_contributes_to_load(p)) 778 rq->nr_uninterruptible--; 779 780 enqueue_task(rq, p, flags); 781} 782 783void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 784{ 785 if (task_contributes_to_load(p)) 786 rq->nr_uninterruptible++; 787 788 dequeue_task(rq, p, flags); 789} 790 791static void update_rq_clock_task(struct rq *rq, s64 delta) 792{ 793/* 794 * In theory, the compile should just see 0 here, and optimize out the call 795 * to sched_rt_avg_update. But I don't trust it... 796 */ 797#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 798 s64 steal = 0, irq_delta = 0; 799#endif 800#ifdef CONFIG_IRQ_TIME_ACCOUNTING 801 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 802 803 /* 804 * Since irq_time is only updated on {soft,}irq_exit, we might run into 805 * this case when a previous update_rq_clock() happened inside a 806 * {soft,}irq region. 807 * 808 * When this happens, we stop ->clock_task and only update the 809 * prev_irq_time stamp to account for the part that fit, so that a next 810 * update will consume the rest. This ensures ->clock_task is 811 * monotonic. 812 * 813 * It does however cause some slight miss-attribution of {soft,}irq 814 * time, a more accurate solution would be to update the irq_time using 815 * the current rq->clock timestamp, except that would require using 816 * atomic ops. 817 */ 818 if (irq_delta > delta) 819 irq_delta = delta; 820 821 rq->prev_irq_time += irq_delta; 822 delta -= irq_delta; 823#endif 824#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 825 if (static_key_false((¶virt_steal_rq_enabled))) { 826 u64 st; 827 828 steal = paravirt_steal_clock(cpu_of(rq)); 829 steal -= rq->prev_steal_time_rq; 830 831 if (unlikely(steal > delta)) 832 steal = delta; 833 834 st = steal_ticks(steal); 835 steal = st * TICK_NSEC; 836 837 rq->prev_steal_time_rq += steal; 838 839 delta -= steal; 840 } 841#endif 842 843 rq->clock_task += delta; 844 845#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 846 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 847 sched_rt_avg_update(rq, irq_delta + steal); 848#endif 849} 850 851void sched_set_stop_task(int cpu, struct task_struct *stop) 852{ 853 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 854 struct task_struct *old_stop = cpu_rq(cpu)->stop; 855 856 if (stop) { 857 /* 858 * Make it appear like a SCHED_FIFO task, its something 859 * userspace knows about and won't get confused about. 860 * 861 * Also, it will make PI more or less work without too 862 * much confusion -- but then, stop work should not 863 * rely on PI working anyway. 864 */ 865 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 866 867 stop->sched_class = &stop_sched_class; 868 } 869 870 cpu_rq(cpu)->stop = stop; 871 872 if (old_stop) { 873 /* 874 * Reset it back to a normal scheduling class so that 875 * it can die in pieces. 876 */ 877 old_stop->sched_class = &rt_sched_class; 878 } 879} 880 881/* 882 * __normal_prio - return the priority that is based on the static prio 883 */ 884static inline int __normal_prio(struct task_struct *p) 885{ 886 return p->static_prio; 887} 888 889/* 890 * Calculate the expected normal priority: i.e. priority 891 * without taking RT-inheritance into account. Might be 892 * boosted by interactivity modifiers. Changes upon fork, 893 * setprio syscalls, and whenever the interactivity 894 * estimator recalculates. 895 */ 896static inline int normal_prio(struct task_struct *p) 897{ 898 int prio; 899 900 if (task_has_dl_policy(p)) 901 prio = MAX_DL_PRIO-1; 902 else if (task_has_rt_policy(p)) 903 prio = MAX_RT_PRIO-1 - p->rt_priority; 904 else 905 prio = __normal_prio(p); 906 return prio; 907} 908 909/* 910 * Calculate the current priority, i.e. the priority 911 * taken into account by the scheduler. This value might 912 * be boosted by RT tasks, or might be boosted by 913 * interactivity modifiers. Will be RT if the task got 914 * RT-boosted. If not then it returns p->normal_prio. 915 */ 916static int effective_prio(struct task_struct *p) 917{ 918 p->normal_prio = normal_prio(p); 919 /* 920 * If we are RT tasks or we were boosted to RT priority, 921 * keep the priority unchanged. Otherwise, update priority 922 * to the normal priority: 923 */ 924 if (!rt_prio(p->prio)) 925 return p->normal_prio; 926 return p->prio; 927} 928 929/** 930 * task_curr - is this task currently executing on a CPU? 931 * @p: the task in question. 932 * 933 * Return: 1 if the task is currently executing. 0 otherwise. 934 */ 935inline int task_curr(const struct task_struct *p) 936{ 937 return cpu_curr(task_cpu(p)) == p; 938} 939 940static inline void check_class_changed(struct rq *rq, struct task_struct *p, 941 const struct sched_class *prev_class, 942 int oldprio) 943{ 944 if (prev_class != p->sched_class) { 945 if (prev_class->switched_from) 946 prev_class->switched_from(rq, p); 947 p->sched_class->switched_to(rq, p); 948 } else if (oldprio != p->prio || dl_task(p)) 949 p->sched_class->prio_changed(rq, p, oldprio); 950} 951 952void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 953{ 954 const struct sched_class *class; 955 956 if (p->sched_class == rq->curr->sched_class) { 957 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 958 } else { 959 for_each_class(class) { 960 if (class == rq->curr->sched_class) 961 break; 962 if (class == p->sched_class) { 963 resched_task(rq->curr); 964 break; 965 } 966 } 967 } 968 969 /* 970 * A queue event has occurred, and we're going to schedule. In 971 * this case, we can save a useless back to back clock update. 972 */ 973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 974 rq->skip_clock_update = 1; 975} 976 977#ifdef CONFIG_SMP 978void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 979{ 980#ifdef CONFIG_SCHED_DEBUG 981 /* 982 * We should never call set_task_cpu() on a blocked task, 983 * ttwu() will sort out the placement. 984 */ 985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 986 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 987 988#ifdef CONFIG_LOCKDEP 989 /* 990 * The caller should hold either p->pi_lock or rq->lock, when changing 991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 992 * 993 * sched_move_task() holds both and thus holding either pins the cgroup, 994 * see task_group(). 995 * 996 * Furthermore, all task_rq users should acquire both locks, see 997 * task_rq_lock(). 998 */ 999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1000 lockdep_is_held(&task_rq(p)->lock))); 1001#endif 1002#endif 1003 1004 trace_sched_migrate_task(p, new_cpu); 1005 1006 if (task_cpu(p) != new_cpu) { 1007 if (p->sched_class->migrate_task_rq) 1008 p->sched_class->migrate_task_rq(p, new_cpu); 1009 p->se.nr_migrations++; 1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1011 } 1012 1013 __set_task_cpu(p, new_cpu); 1014} 1015 1016static void __migrate_swap_task(struct task_struct *p, int cpu) 1017{ 1018 if (p->on_rq) { 1019 struct rq *src_rq, *dst_rq; 1020 1021 src_rq = task_rq(p); 1022 dst_rq = cpu_rq(cpu); 1023 1024 deactivate_task(src_rq, p, 0); 1025 set_task_cpu(p, cpu); 1026 activate_task(dst_rq, p, 0); 1027 check_preempt_curr(dst_rq, p, 0); 1028 } else { 1029 /* 1030 * Task isn't running anymore; make it appear like we migrated 1031 * it before it went to sleep. This means on wakeup we make the 1032 * previous cpu our targer instead of where it really is. 1033 */ 1034 p->wake_cpu = cpu; 1035 } 1036} 1037 1038struct migration_swap_arg { 1039 struct task_struct *src_task, *dst_task; 1040 int src_cpu, dst_cpu; 1041}; 1042 1043static int migrate_swap_stop(void *data) 1044{ 1045 struct migration_swap_arg *arg = data; 1046 struct rq *src_rq, *dst_rq; 1047 int ret = -EAGAIN; 1048 1049 src_rq = cpu_rq(arg->src_cpu); 1050 dst_rq = cpu_rq(arg->dst_cpu); 1051 1052 double_raw_lock(&arg->src_task->pi_lock, 1053 &arg->dst_task->pi_lock); 1054 double_rq_lock(src_rq, dst_rq); 1055 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1056 goto unlock; 1057 1058 if (task_cpu(arg->src_task) != arg->src_cpu) 1059 goto unlock; 1060 1061 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1062 goto unlock; 1063 1064 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1065 goto unlock; 1066 1067 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1068 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1069 1070 ret = 0; 1071 1072unlock: 1073 double_rq_unlock(src_rq, dst_rq); 1074 raw_spin_unlock(&arg->dst_task->pi_lock); 1075 raw_spin_unlock(&arg->src_task->pi_lock); 1076 1077 return ret; 1078} 1079 1080/* 1081 * Cross migrate two tasks 1082 */ 1083int migrate_swap(struct task_struct *cur, struct task_struct *p) 1084{ 1085 struct migration_swap_arg arg; 1086 int ret = -EINVAL; 1087 1088 arg = (struct migration_swap_arg){ 1089 .src_task = cur, 1090 .src_cpu = task_cpu(cur), 1091 .dst_task = p, 1092 .dst_cpu = task_cpu(p), 1093 }; 1094 1095 if (arg.src_cpu == arg.dst_cpu) 1096 goto out; 1097 1098 /* 1099 * These three tests are all lockless; this is OK since all of them 1100 * will be re-checked with proper locks held further down the line. 1101 */ 1102 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1103 goto out; 1104 1105 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1106 goto out; 1107 1108 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1109 goto out; 1110 1111 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1112 1113out: 1114 return ret; 1115} 1116 1117struct migration_arg { 1118 struct task_struct *task; 1119 int dest_cpu; 1120}; 1121 1122static int migration_cpu_stop(void *data); 1123 1124/* 1125 * wait_task_inactive - wait for a thread to unschedule. 1126 * 1127 * If @match_state is nonzero, it's the @p->state value just checked and 1128 * not expected to change. If it changes, i.e. @p might have woken up, 1129 * then return zero. When we succeed in waiting for @p to be off its CPU, 1130 * we return a positive number (its total switch count). If a second call 1131 * a short while later returns the same number, the caller can be sure that 1132 * @p has remained unscheduled the whole time. 1133 * 1134 * The caller must ensure that the task *will* unschedule sometime soon, 1135 * else this function might spin for a *long* time. This function can't 1136 * be called with interrupts off, or it may introduce deadlock with 1137 * smp_call_function() if an IPI is sent by the same process we are 1138 * waiting to become inactive. 1139 */ 1140unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1141{ 1142 unsigned long flags; 1143 int running, on_rq; 1144 unsigned long ncsw; 1145 struct rq *rq; 1146 1147 for (;;) { 1148 /* 1149 * We do the initial early heuristics without holding 1150 * any task-queue locks at all. We'll only try to get 1151 * the runqueue lock when things look like they will 1152 * work out! 1153 */ 1154 rq = task_rq(p); 1155 1156 /* 1157 * If the task is actively running on another CPU 1158 * still, just relax and busy-wait without holding 1159 * any locks. 1160 * 1161 * NOTE! Since we don't hold any locks, it's not 1162 * even sure that "rq" stays as the right runqueue! 1163 * But we don't care, since "task_running()" will 1164 * return false if the runqueue has changed and p 1165 * is actually now running somewhere else! 1166 */ 1167 while (task_running(rq, p)) { 1168 if (match_state && unlikely(p->state != match_state)) 1169 return 0; 1170 cpu_relax(); 1171 } 1172 1173 /* 1174 * Ok, time to look more closely! We need the rq 1175 * lock now, to be *sure*. If we're wrong, we'll 1176 * just go back and repeat. 1177 */ 1178 rq = task_rq_lock(p, &flags); 1179 trace_sched_wait_task(p); 1180 running = task_running(rq, p); 1181 on_rq = p->on_rq; 1182 ncsw = 0; 1183 if (!match_state || p->state == match_state) 1184 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1185 task_rq_unlock(rq, p, &flags); 1186 1187 /* 1188 * If it changed from the expected state, bail out now. 1189 */ 1190 if (unlikely(!ncsw)) 1191 break; 1192 1193 /* 1194 * Was it really running after all now that we 1195 * checked with the proper locks actually held? 1196 * 1197 * Oops. Go back and try again.. 1198 */ 1199 if (unlikely(running)) { 1200 cpu_relax(); 1201 continue; 1202 } 1203 1204 /* 1205 * It's not enough that it's not actively running, 1206 * it must be off the runqueue _entirely_, and not 1207 * preempted! 1208 * 1209 * So if it was still runnable (but just not actively 1210 * running right now), it's preempted, and we should 1211 * yield - it could be a while. 1212 */ 1213 if (unlikely(on_rq)) { 1214 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1215 1216 set_current_state(TASK_UNINTERRUPTIBLE); 1217 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1218 continue; 1219 } 1220 1221 /* 1222 * Ahh, all good. It wasn't running, and it wasn't 1223 * runnable, which means that it will never become 1224 * running in the future either. We're all done! 1225 */ 1226 break; 1227 } 1228 1229 return ncsw; 1230} 1231 1232/*** 1233 * kick_process - kick a running thread to enter/exit the kernel 1234 * @p: the to-be-kicked thread 1235 * 1236 * Cause a process which is running on another CPU to enter 1237 * kernel-mode, without any delay. (to get signals handled.) 1238 * 1239 * NOTE: this function doesn't have to take the runqueue lock, 1240 * because all it wants to ensure is that the remote task enters 1241 * the kernel. If the IPI races and the task has been migrated 1242 * to another CPU then no harm is done and the purpose has been 1243 * achieved as well. 1244 */ 1245void kick_process(struct task_struct *p) 1246{ 1247 int cpu; 1248 1249 preempt_disable(); 1250 cpu = task_cpu(p); 1251 if ((cpu != smp_processor_id()) && task_curr(p)) 1252 smp_send_reschedule(cpu); 1253 preempt_enable(); 1254} 1255EXPORT_SYMBOL_GPL(kick_process); 1256#endif /* CONFIG_SMP */ 1257 1258#ifdef CONFIG_SMP 1259/* 1260 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1261 */ 1262static int select_fallback_rq(int cpu, struct task_struct *p) 1263{ 1264 int nid = cpu_to_node(cpu); 1265 const struct cpumask *nodemask = NULL; 1266 enum { cpuset, possible, fail } state = cpuset; 1267 int dest_cpu; 1268 1269 /* 1270 * If the node that the cpu is on has been offlined, cpu_to_node() 1271 * will return -1. There is no cpu on the node, and we should 1272 * select the cpu on the other node. 1273 */ 1274 if (nid != -1) { 1275 nodemask = cpumask_of_node(nid); 1276 1277 /* Look for allowed, online CPU in same node. */ 1278 for_each_cpu(dest_cpu, nodemask) { 1279 if (!cpu_online(dest_cpu)) 1280 continue; 1281 if (!cpu_active(dest_cpu)) 1282 continue; 1283 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1284 return dest_cpu; 1285 } 1286 } 1287 1288 for (;;) { 1289 /* Any allowed, online CPU? */ 1290 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1291 if (!cpu_online(dest_cpu)) 1292 continue; 1293 if (!cpu_active(dest_cpu)) 1294 continue; 1295 goto out; 1296 } 1297 1298 switch (state) { 1299 case cpuset: 1300 /* No more Mr. Nice Guy. */ 1301 cpuset_cpus_allowed_fallback(p); 1302 state = possible; 1303 break; 1304 1305 case possible: 1306 do_set_cpus_allowed(p, cpu_possible_mask); 1307 state = fail; 1308 break; 1309 1310 case fail: 1311 BUG(); 1312 break; 1313 } 1314 } 1315 1316out: 1317 if (state != cpuset) { 1318 /* 1319 * Don't tell them about moving exiting tasks or 1320 * kernel threads (both mm NULL), since they never 1321 * leave kernel. 1322 */ 1323 if (p->mm && printk_ratelimit()) { 1324 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1325 task_pid_nr(p), p->comm, cpu); 1326 } 1327 } 1328 1329 return dest_cpu; 1330} 1331 1332/* 1333 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1334 */ 1335static inline 1336int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1337{ 1338 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1339 1340 /* 1341 * In order not to call set_task_cpu() on a blocking task we need 1342 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1343 * cpu. 1344 * 1345 * Since this is common to all placement strategies, this lives here. 1346 * 1347 * [ this allows ->select_task() to simply return task_cpu(p) and 1348 * not worry about this generic constraint ] 1349 */ 1350 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1351 !cpu_online(cpu))) 1352 cpu = select_fallback_rq(task_cpu(p), p); 1353 1354 return cpu; 1355} 1356 1357static void update_avg(u64 *avg, u64 sample) 1358{ 1359 s64 diff = sample - *avg; 1360 *avg += diff >> 3; 1361} 1362#endif 1363 1364static void 1365ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1366{ 1367#ifdef CONFIG_SCHEDSTATS 1368 struct rq *rq = this_rq(); 1369 1370#ifdef CONFIG_SMP 1371 int this_cpu = smp_processor_id(); 1372 1373 if (cpu == this_cpu) { 1374 schedstat_inc(rq, ttwu_local); 1375 schedstat_inc(p, se.statistics.nr_wakeups_local); 1376 } else { 1377 struct sched_domain *sd; 1378 1379 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1380 rcu_read_lock(); 1381 for_each_domain(this_cpu, sd) { 1382 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1383 schedstat_inc(sd, ttwu_wake_remote); 1384 break; 1385 } 1386 } 1387 rcu_read_unlock(); 1388 } 1389 1390 if (wake_flags & WF_MIGRATED) 1391 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1392 1393#endif /* CONFIG_SMP */ 1394 1395 schedstat_inc(rq, ttwu_count); 1396 schedstat_inc(p, se.statistics.nr_wakeups); 1397 1398 if (wake_flags & WF_SYNC) 1399 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1400 1401#endif /* CONFIG_SCHEDSTATS */ 1402} 1403 1404static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1405{ 1406 activate_task(rq, p, en_flags); 1407 p->on_rq = 1; 1408 1409 /* if a worker is waking up, notify workqueue */ 1410 if (p->flags & PF_WQ_WORKER) 1411 wq_worker_waking_up(p, cpu_of(rq)); 1412} 1413 1414/* 1415 * Mark the task runnable and perform wakeup-preemption. 1416 */ 1417static void 1418ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1419{ 1420 check_preempt_curr(rq, p, wake_flags); 1421 trace_sched_wakeup(p, true); 1422 1423 p->state = TASK_RUNNING; 1424#ifdef CONFIG_SMP 1425 if (p->sched_class->task_woken) 1426 p->sched_class->task_woken(rq, p); 1427 1428 if (rq->idle_stamp) { 1429 u64 delta = rq_clock(rq) - rq->idle_stamp; 1430 u64 max = 2*rq->max_idle_balance_cost; 1431 1432 update_avg(&rq->avg_idle, delta); 1433 1434 if (rq->avg_idle > max) 1435 rq->avg_idle = max; 1436 1437 rq->idle_stamp = 0; 1438 } 1439#endif 1440} 1441 1442static void 1443ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1444{ 1445#ifdef CONFIG_SMP 1446 if (p->sched_contributes_to_load) 1447 rq->nr_uninterruptible--; 1448#endif 1449 1450 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1451 ttwu_do_wakeup(rq, p, wake_flags); 1452} 1453 1454/* 1455 * Called in case the task @p isn't fully descheduled from its runqueue, 1456 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1457 * since all we need to do is flip p->state to TASK_RUNNING, since 1458 * the task is still ->on_rq. 1459 */ 1460static int ttwu_remote(struct task_struct *p, int wake_flags) 1461{ 1462 struct rq *rq; 1463 int ret = 0; 1464 1465 rq = __task_rq_lock(p); 1466 if (p->on_rq) { 1467 /* check_preempt_curr() may use rq clock */ 1468 update_rq_clock(rq); 1469 ttwu_do_wakeup(rq, p, wake_flags); 1470 ret = 1; 1471 } 1472 __task_rq_unlock(rq); 1473 1474 return ret; 1475} 1476 1477#ifdef CONFIG_SMP 1478static void sched_ttwu_pending(void) 1479{ 1480 struct rq *rq = this_rq(); 1481 struct llist_node *llist = llist_del_all(&rq->wake_list); 1482 struct task_struct *p; 1483 1484 raw_spin_lock(&rq->lock); 1485 1486 while (llist) { 1487 p = llist_entry(llist, struct task_struct, wake_entry); 1488 llist = llist_next(llist); 1489 ttwu_do_activate(rq, p, 0); 1490 } 1491 1492 raw_spin_unlock(&rq->lock); 1493} 1494 1495void scheduler_ipi(void) 1496{ 1497 /* 1498 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1499 * TIF_NEED_RESCHED remotely (for the first time) will also send 1500 * this IPI. 1501 */ 1502 preempt_fold_need_resched(); 1503 1504 if (llist_empty(&this_rq()->wake_list) 1505 && !tick_nohz_full_cpu(smp_processor_id()) 1506 && !got_nohz_idle_kick()) 1507 return; 1508 1509 /* 1510 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1511 * traditionally all their work was done from the interrupt return 1512 * path. Now that we actually do some work, we need to make sure 1513 * we do call them. 1514 * 1515 * Some archs already do call them, luckily irq_enter/exit nest 1516 * properly. 1517 * 1518 * Arguably we should visit all archs and update all handlers, 1519 * however a fair share of IPIs are still resched only so this would 1520 * somewhat pessimize the simple resched case. 1521 */ 1522 irq_enter(); 1523 tick_nohz_full_check(); 1524 sched_ttwu_pending(); 1525 1526 /* 1527 * Check if someone kicked us for doing the nohz idle load balance. 1528 */ 1529 if (unlikely(got_nohz_idle_kick())) { 1530 this_rq()->idle_balance = 1; 1531 raise_softirq_irqoff(SCHED_SOFTIRQ); 1532 } 1533 irq_exit(); 1534} 1535 1536static void ttwu_queue_remote(struct task_struct *p, int cpu) 1537{ 1538 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1539 smp_send_reschedule(cpu); 1540} 1541 1542bool cpus_share_cache(int this_cpu, int that_cpu) 1543{ 1544 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1545} 1546#endif /* CONFIG_SMP */ 1547 1548static void ttwu_queue(struct task_struct *p, int cpu) 1549{ 1550 struct rq *rq = cpu_rq(cpu); 1551 1552#if defined(CONFIG_SMP) 1553 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1554 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1555 ttwu_queue_remote(p, cpu); 1556 return; 1557 } 1558#endif 1559 1560 raw_spin_lock(&rq->lock); 1561 ttwu_do_activate(rq, p, 0); 1562 raw_spin_unlock(&rq->lock); 1563} 1564 1565/** 1566 * try_to_wake_up - wake up a thread 1567 * @p: the thread to be awakened 1568 * @state: the mask of task states that can be woken 1569 * @wake_flags: wake modifier flags (WF_*) 1570 * 1571 * Put it on the run-queue if it's not already there. The "current" 1572 * thread is always on the run-queue (except when the actual 1573 * re-schedule is in progress), and as such you're allowed to do 1574 * the simpler "current->state = TASK_RUNNING" to mark yourself 1575 * runnable without the overhead of this. 1576 * 1577 * Return: %true if @p was woken up, %false if it was already running. 1578 * or @state didn't match @p's state. 1579 */ 1580static int 1581try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1582{ 1583 unsigned long flags; 1584 int cpu, success = 0; 1585 1586 /* 1587 * If we are going to wake up a thread waiting for CONDITION we 1588 * need to ensure that CONDITION=1 done by the caller can not be 1589 * reordered with p->state check below. This pairs with mb() in 1590 * set_current_state() the waiting thread does. 1591 */ 1592 smp_mb__before_spinlock(); 1593 raw_spin_lock_irqsave(&p->pi_lock, flags); 1594 if (!(p->state & state)) 1595 goto out; 1596 1597 success = 1; /* we're going to change ->state */ 1598 cpu = task_cpu(p); 1599 1600 if (p->on_rq && ttwu_remote(p, wake_flags)) 1601 goto stat; 1602 1603#ifdef CONFIG_SMP 1604 /* 1605 * If the owning (remote) cpu is still in the middle of schedule() with 1606 * this task as prev, wait until its done referencing the task. 1607 */ 1608 while (p->on_cpu) 1609 cpu_relax(); 1610 /* 1611 * Pairs with the smp_wmb() in finish_lock_switch(). 1612 */ 1613 smp_rmb(); 1614 1615 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1616 p->state = TASK_WAKING; 1617 1618 if (p->sched_class->task_waking) 1619 p->sched_class->task_waking(p); 1620 1621 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1622 if (task_cpu(p) != cpu) { 1623 wake_flags |= WF_MIGRATED; 1624 set_task_cpu(p, cpu); 1625 } 1626#endif /* CONFIG_SMP */ 1627 1628 ttwu_queue(p, cpu); 1629stat: 1630 ttwu_stat(p, cpu, wake_flags); 1631out: 1632 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1633 1634 return success; 1635} 1636 1637/** 1638 * try_to_wake_up_local - try to wake up a local task with rq lock held 1639 * @p: the thread to be awakened 1640 * 1641 * Put @p on the run-queue if it's not already there. The caller must 1642 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1643 * the current task. 1644 */ 1645static void try_to_wake_up_local(struct task_struct *p) 1646{ 1647 struct rq *rq = task_rq(p); 1648 1649 if (WARN_ON_ONCE(rq != this_rq()) || 1650 WARN_ON_ONCE(p == current)) 1651 return; 1652 1653 lockdep_assert_held(&rq->lock); 1654 1655 if (!raw_spin_trylock(&p->pi_lock)) { 1656 raw_spin_unlock(&rq->lock); 1657 raw_spin_lock(&p->pi_lock); 1658 raw_spin_lock(&rq->lock); 1659 } 1660 1661 if (!(p->state & TASK_NORMAL)) 1662 goto out; 1663 1664 if (!p->on_rq) 1665 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1666 1667 ttwu_do_wakeup(rq, p, 0); 1668 ttwu_stat(p, smp_processor_id(), 0); 1669out: 1670 raw_spin_unlock(&p->pi_lock); 1671} 1672 1673/** 1674 * wake_up_process - Wake up a specific process 1675 * @p: The process to be woken up. 1676 * 1677 * Attempt to wake up the nominated process and move it to the set of runnable 1678 * processes. 1679 * 1680 * Return: 1 if the process was woken up, 0 if it was already running. 1681 * 1682 * It may be assumed that this function implies a write memory barrier before 1683 * changing the task state if and only if any tasks are woken up. 1684 */ 1685int wake_up_process(struct task_struct *p) 1686{ 1687 WARN_ON(task_is_stopped_or_traced(p)); 1688 return try_to_wake_up(p, TASK_NORMAL, 0); 1689} 1690EXPORT_SYMBOL(wake_up_process); 1691 1692int wake_up_state(struct task_struct *p, unsigned int state) 1693{ 1694 return try_to_wake_up(p, state, 0); 1695} 1696 1697/* 1698 * Perform scheduler related setup for a newly forked process p. 1699 * p is forked by current. 1700 * 1701 * __sched_fork() is basic setup used by init_idle() too: 1702 */ 1703static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1704{ 1705 p->on_rq = 0; 1706 1707 p->se.on_rq = 0; 1708 p->se.exec_start = 0; 1709 p->se.sum_exec_runtime = 0; 1710 p->se.prev_sum_exec_runtime = 0; 1711 p->se.nr_migrations = 0; 1712 p->se.vruntime = 0; 1713 INIT_LIST_HEAD(&p->se.group_node); 1714 1715#ifdef CONFIG_SCHEDSTATS 1716 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1717#endif 1718 1719 RB_CLEAR_NODE(&p->dl.rb_node); 1720 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1721 p->dl.dl_runtime = p->dl.runtime = 0; 1722 p->dl.dl_deadline = p->dl.deadline = 0; 1723 p->dl.dl_period = 0; 1724 p->dl.flags = 0; 1725 1726 INIT_LIST_HEAD(&p->rt.run_list); 1727 1728#ifdef CONFIG_PREEMPT_NOTIFIERS 1729 INIT_HLIST_HEAD(&p->preempt_notifiers); 1730#endif 1731 1732#ifdef CONFIG_NUMA_BALANCING 1733 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1734 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1735 p->mm->numa_scan_seq = 0; 1736 } 1737 1738 if (clone_flags & CLONE_VM) 1739 p->numa_preferred_nid = current->numa_preferred_nid; 1740 else 1741 p->numa_preferred_nid = -1; 1742 1743 p->node_stamp = 0ULL; 1744 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1745 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1746 p->numa_work.next = &p->numa_work; 1747 p->numa_faults_memory = NULL; 1748 p->numa_faults_buffer_memory = NULL; 1749 1750 INIT_LIST_HEAD(&p->numa_entry); 1751 p->numa_group = NULL; 1752#endif /* CONFIG_NUMA_BALANCING */ 1753} 1754 1755#ifdef CONFIG_NUMA_BALANCING 1756#ifdef CONFIG_SCHED_DEBUG 1757void set_numabalancing_state(bool enabled) 1758{ 1759 if (enabled) 1760 sched_feat_set("NUMA"); 1761 else 1762 sched_feat_set("NO_NUMA"); 1763} 1764#else 1765__read_mostly bool numabalancing_enabled; 1766 1767void set_numabalancing_state(bool enabled) 1768{ 1769 numabalancing_enabled = enabled; 1770} 1771#endif /* CONFIG_SCHED_DEBUG */ 1772#endif /* CONFIG_NUMA_BALANCING */ 1773 1774/* 1775 * fork()/clone()-time setup: 1776 */ 1777int sched_fork(unsigned long clone_flags, struct task_struct *p) 1778{ 1779 unsigned long flags; 1780 int cpu = get_cpu(); 1781 1782 __sched_fork(clone_flags, p); 1783 /* 1784 * We mark the process as running here. This guarantees that 1785 * nobody will actually run it, and a signal or other external 1786 * event cannot wake it up and insert it on the runqueue either. 1787 */ 1788 p->state = TASK_RUNNING; 1789 1790 /* 1791 * Make sure we do not leak PI boosting priority to the child. 1792 */ 1793 p->prio = current->normal_prio; 1794 1795 /* 1796 * Revert to default priority/policy on fork if requested. 1797 */ 1798 if (unlikely(p->sched_reset_on_fork)) { 1799 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1800 p->policy = SCHED_NORMAL; 1801 p->static_prio = NICE_TO_PRIO(0); 1802 p->rt_priority = 0; 1803 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1804 p->static_prio = NICE_TO_PRIO(0); 1805 1806 p->prio = p->normal_prio = __normal_prio(p); 1807 set_load_weight(p); 1808 1809 /* 1810 * We don't need the reset flag anymore after the fork. It has 1811 * fulfilled its duty: 1812 */ 1813 p->sched_reset_on_fork = 0; 1814 } 1815 1816 if (dl_prio(p->prio)) { 1817 put_cpu(); 1818 return -EAGAIN; 1819 } else if (rt_prio(p->prio)) { 1820 p->sched_class = &rt_sched_class; 1821 } else { 1822 p->sched_class = &fair_sched_class; 1823 } 1824 1825 if (p->sched_class->task_fork) 1826 p->sched_class->task_fork(p); 1827 1828 /* 1829 * The child is not yet in the pid-hash so no cgroup attach races, 1830 * and the cgroup is pinned to this child due to cgroup_fork() 1831 * is ran before sched_fork(). 1832 * 1833 * Silence PROVE_RCU. 1834 */ 1835 raw_spin_lock_irqsave(&p->pi_lock, flags); 1836 set_task_cpu(p, cpu); 1837 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1838 1839#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1840 if (likely(sched_info_on())) 1841 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1842#endif 1843#if defined(CONFIG_SMP) 1844 p->on_cpu = 0; 1845#endif 1846 init_task_preempt_count(p); 1847#ifdef CONFIG_SMP 1848 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1849 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1850#endif 1851 1852 put_cpu(); 1853 return 0; 1854} 1855 1856unsigned long to_ratio(u64 period, u64 runtime) 1857{ 1858 if (runtime == RUNTIME_INF) 1859 return 1ULL << 20; 1860 1861 /* 1862 * Doing this here saves a lot of checks in all 1863 * the calling paths, and returning zero seems 1864 * safe for them anyway. 1865 */ 1866 if (period == 0) 1867 return 0; 1868 1869 return div64_u64(runtime << 20, period); 1870} 1871 1872#ifdef CONFIG_SMP 1873inline struct dl_bw *dl_bw_of(int i) 1874{ 1875 return &cpu_rq(i)->rd->dl_bw; 1876} 1877 1878static inline int dl_bw_cpus(int i) 1879{ 1880 struct root_domain *rd = cpu_rq(i)->rd; 1881 int cpus = 0; 1882 1883 for_each_cpu_and(i, rd->span, cpu_active_mask) 1884 cpus++; 1885 1886 return cpus; 1887} 1888#else 1889inline struct dl_bw *dl_bw_of(int i) 1890{ 1891 return &cpu_rq(i)->dl.dl_bw; 1892} 1893 1894static inline int dl_bw_cpus(int i) 1895{ 1896 return 1; 1897} 1898#endif 1899 1900static inline 1901void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1902{ 1903 dl_b->total_bw -= tsk_bw; 1904} 1905 1906static inline 1907void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1908{ 1909 dl_b->total_bw += tsk_bw; 1910} 1911 1912static inline 1913bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1914{ 1915 return dl_b->bw != -1 && 1916 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1917} 1918 1919/* 1920 * We must be sure that accepting a new task (or allowing changing the 1921 * parameters of an existing one) is consistent with the bandwidth 1922 * constraints. If yes, this function also accordingly updates the currently 1923 * allocated bandwidth to reflect the new situation. 1924 * 1925 * This function is called while holding p's rq->lock. 1926 */ 1927static int dl_overflow(struct task_struct *p, int policy, 1928 const struct sched_attr *attr) 1929{ 1930 1931 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1932 u64 period = attr->sched_period; 1933 u64 runtime = attr->sched_runtime; 1934 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1935 int cpus, err = -1; 1936 1937 if (new_bw == p->dl.dl_bw) 1938 return 0; 1939 1940 /* 1941 * Either if a task, enters, leave, or stays -deadline but changes 1942 * its parameters, we may need to update accordingly the total 1943 * allocated bandwidth of the container. 1944 */ 1945 raw_spin_lock(&dl_b->lock); 1946 cpus = dl_bw_cpus(task_cpu(p)); 1947 if (dl_policy(policy) && !task_has_dl_policy(p) && 1948 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1949 __dl_add(dl_b, new_bw); 1950 err = 0; 1951 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1952 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1953 __dl_clear(dl_b, p->dl.dl_bw); 1954 __dl_add(dl_b, new_bw); 1955 err = 0; 1956 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1957 __dl_clear(dl_b, p->dl.dl_bw); 1958 err = 0; 1959 } 1960 raw_spin_unlock(&dl_b->lock); 1961 1962 return err; 1963} 1964 1965extern void init_dl_bw(struct dl_bw *dl_b); 1966 1967/* 1968 * wake_up_new_task - wake up a newly created task for the first time. 1969 * 1970 * This function will do some initial scheduler statistics housekeeping 1971 * that must be done for every newly created context, then puts the task 1972 * on the runqueue and wakes it. 1973 */ 1974void wake_up_new_task(struct task_struct *p) 1975{ 1976 unsigned long flags; 1977 struct rq *rq; 1978 1979 raw_spin_lock_irqsave(&p->pi_lock, flags); 1980#ifdef CONFIG_SMP 1981 /* 1982 * Fork balancing, do it here and not earlier because: 1983 * - cpus_allowed can change in the fork path 1984 * - any previously selected cpu might disappear through hotplug 1985 */ 1986 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 1987#endif 1988 1989 /* Initialize new task's runnable average */ 1990 init_task_runnable_average(p); 1991 rq = __task_rq_lock(p); 1992 activate_task(rq, p, 0); 1993 p->on_rq = 1; 1994 trace_sched_wakeup_new(p, true); 1995 check_preempt_curr(rq, p, WF_FORK); 1996#ifdef CONFIG_SMP 1997 if (p->sched_class->task_woken) 1998 p->sched_class->task_woken(rq, p); 1999#endif 2000 task_rq_unlock(rq, p, &flags); 2001} 2002 2003#ifdef CONFIG_PREEMPT_NOTIFIERS 2004 2005/** 2006 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2007 * @notifier: notifier struct to register 2008 */ 2009void preempt_notifier_register(struct preempt_notifier *notifier) 2010{ 2011 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2012} 2013EXPORT_SYMBOL_GPL(preempt_notifier_register); 2014 2015/** 2016 * preempt_notifier_unregister - no longer interested in preemption notifications 2017 * @notifier: notifier struct to unregister 2018 * 2019 * This is safe to call from within a preemption notifier. 2020 */ 2021void preempt_notifier_unregister(struct preempt_notifier *notifier) 2022{ 2023 hlist_del(¬ifier->link); 2024} 2025EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2026 2027static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2028{ 2029 struct preempt_notifier *notifier; 2030 2031 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2032 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2033} 2034 2035static void 2036fire_sched_out_preempt_notifiers(struct task_struct *curr, 2037 struct task_struct *next) 2038{ 2039 struct preempt_notifier *notifier; 2040 2041 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2042 notifier->ops->sched_out(notifier, next); 2043} 2044 2045#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2046 2047static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2048{ 2049} 2050 2051static void 2052fire_sched_out_preempt_notifiers(struct task_struct *curr, 2053 struct task_struct *next) 2054{ 2055} 2056 2057#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2058 2059/** 2060 * prepare_task_switch - prepare to switch tasks 2061 * @rq: the runqueue preparing to switch 2062 * @prev: the current task that is being switched out 2063 * @next: the task we are going to switch to. 2064 * 2065 * This is called with the rq lock held and interrupts off. It must 2066 * be paired with a subsequent finish_task_switch after the context 2067 * switch. 2068 * 2069 * prepare_task_switch sets up locking and calls architecture specific 2070 * hooks. 2071 */ 2072static inline void 2073prepare_task_switch(struct rq *rq, struct task_struct *prev, 2074 struct task_struct *next) 2075{ 2076 trace_sched_switch(prev, next); 2077 sched_info_switch(rq, prev, next); 2078 perf_event_task_sched_out(prev, next); 2079 fire_sched_out_preempt_notifiers(prev, next); 2080 prepare_lock_switch(rq, next); 2081 prepare_arch_switch(next); 2082} 2083 2084/** 2085 * finish_task_switch - clean up after a task-switch 2086 * @rq: runqueue associated with task-switch 2087 * @prev: the thread we just switched away from. 2088 * 2089 * finish_task_switch must be called after the context switch, paired 2090 * with a prepare_task_switch call before the context switch. 2091 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2092 * and do any other architecture-specific cleanup actions. 2093 * 2094 * Note that we may have delayed dropping an mm in context_switch(). If 2095 * so, we finish that here outside of the runqueue lock. (Doing it 2096 * with the lock held can cause deadlocks; see schedule() for 2097 * details.) 2098 */ 2099static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2100 __releases(rq->lock) 2101{ 2102 struct mm_struct *mm = rq->prev_mm; 2103 long prev_state; 2104 2105 rq->prev_mm = NULL; 2106 2107 /* 2108 * A task struct has one reference for the use as "current". 2109 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2110 * schedule one last time. The schedule call will never return, and 2111 * the scheduled task must drop that reference. 2112 * The test for TASK_DEAD must occur while the runqueue locks are 2113 * still held, otherwise prev could be scheduled on another cpu, die 2114 * there before we look at prev->state, and then the reference would 2115 * be dropped twice. 2116 * Manfred Spraul <manfred@colorfullife.com> 2117 */ 2118 prev_state = prev->state; 2119 vtime_task_switch(prev); 2120 finish_arch_switch(prev); 2121 perf_event_task_sched_in(prev, current); 2122 finish_lock_switch(rq, prev); 2123 finish_arch_post_lock_switch(); 2124 2125 fire_sched_in_preempt_notifiers(current); 2126 if (mm) 2127 mmdrop(mm); 2128 if (unlikely(prev_state == TASK_DEAD)) { 2129 task_numa_free(prev); 2130 2131 if (prev->sched_class->task_dead) 2132 prev->sched_class->task_dead(prev); 2133 2134 /* 2135 * Remove function-return probe instances associated with this 2136 * task and put them back on the free list. 2137 */ 2138 kprobe_flush_task(prev); 2139 put_task_struct(prev); 2140 } 2141 2142 tick_nohz_task_switch(current); 2143} 2144 2145#ifdef CONFIG_SMP 2146 2147/* assumes rq->lock is held */ 2148static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 2149{ 2150 if (prev->sched_class->pre_schedule) 2151 prev->sched_class->pre_schedule(rq, prev); 2152} 2153 2154/* rq->lock is NOT held, but preemption is disabled */ 2155static inline void post_schedule(struct rq *rq) 2156{ 2157 if (rq->post_schedule) { 2158 unsigned long flags; 2159 2160 raw_spin_lock_irqsave(&rq->lock, flags); 2161 if (rq->curr->sched_class->post_schedule) 2162 rq->curr->sched_class->post_schedule(rq); 2163 raw_spin_unlock_irqrestore(&rq->lock, flags); 2164 2165 rq->post_schedule = 0; 2166 } 2167} 2168 2169#else 2170 2171static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2172{ 2173} 2174 2175static inline void post_schedule(struct rq *rq) 2176{ 2177} 2178 2179#endif 2180 2181/** 2182 * schedule_tail - first thing a freshly forked thread must call. 2183 * @prev: the thread we just switched away from. 2184 */ 2185asmlinkage void schedule_tail(struct task_struct *prev) 2186 __releases(rq->lock) 2187{ 2188 struct rq *rq = this_rq(); 2189 2190 finish_task_switch(rq, prev); 2191 2192 /* 2193 * FIXME: do we need to worry about rq being invalidated by the 2194 * task_switch? 2195 */ 2196 post_schedule(rq); 2197 2198#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2199 /* In this case, finish_task_switch does not reenable preemption */ 2200 preempt_enable(); 2201#endif 2202 if (current->set_child_tid) 2203 put_user(task_pid_vnr(current), current->set_child_tid); 2204} 2205 2206/* 2207 * context_switch - switch to the new MM and the new 2208 * thread's register state. 2209 */ 2210static inline void 2211context_switch(struct rq *rq, struct task_struct *prev, 2212 struct task_struct *next) 2213{ 2214 struct mm_struct *mm, *oldmm; 2215 2216 prepare_task_switch(rq, prev, next); 2217 2218 mm = next->mm; 2219 oldmm = prev->active_mm; 2220 /* 2221 * For paravirt, this is coupled with an exit in switch_to to 2222 * combine the page table reload and the switch backend into 2223 * one hypercall. 2224 */ 2225 arch_start_context_switch(prev); 2226 2227 if (!mm) { 2228 next->active_mm = oldmm; 2229 atomic_inc(&oldmm->mm_count); 2230 enter_lazy_tlb(oldmm, next); 2231 } else 2232 switch_mm(oldmm, mm, next); 2233 2234 if (!prev->mm) { 2235 prev->active_mm = NULL; 2236 rq->prev_mm = oldmm; 2237 } 2238 /* 2239 * Since the runqueue lock will be released by the next 2240 * task (which is an invalid locking op but in the case 2241 * of the scheduler it's an obvious special-case), so we 2242 * do an early lockdep release here: 2243 */ 2244#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2245 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2246#endif 2247 2248 context_tracking_task_switch(prev, next); 2249 /* Here we just switch the register state and the stack. */ 2250 switch_to(prev, next, prev); 2251 2252 barrier(); 2253 /* 2254 * this_rq must be evaluated again because prev may have moved 2255 * CPUs since it called schedule(), thus the 'rq' on its stack 2256 * frame will be invalid. 2257 */ 2258 finish_task_switch(this_rq(), prev); 2259} 2260 2261/* 2262 * nr_running and nr_context_switches: 2263 * 2264 * externally visible scheduler statistics: current number of runnable 2265 * threads, total number of context switches performed since bootup. 2266 */ 2267unsigned long nr_running(void) 2268{ 2269 unsigned long i, sum = 0; 2270 2271 for_each_online_cpu(i) 2272 sum += cpu_rq(i)->nr_running; 2273 2274 return sum; 2275} 2276 2277unsigned long long nr_context_switches(void) 2278{ 2279 int i; 2280 unsigned long long sum = 0; 2281 2282 for_each_possible_cpu(i) 2283 sum += cpu_rq(i)->nr_switches; 2284 2285 return sum; 2286} 2287 2288unsigned long nr_iowait(void) 2289{ 2290 unsigned long i, sum = 0; 2291 2292 for_each_possible_cpu(i) 2293 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2294 2295 return sum; 2296} 2297 2298unsigned long nr_iowait_cpu(int cpu) 2299{ 2300 struct rq *this = cpu_rq(cpu); 2301 return atomic_read(&this->nr_iowait); 2302} 2303 2304#ifdef CONFIG_SMP 2305 2306/* 2307 * sched_exec - execve() is a valuable balancing opportunity, because at 2308 * this point the task has the smallest effective memory and cache footprint. 2309 */ 2310void sched_exec(void) 2311{ 2312 struct task_struct *p = current; 2313 unsigned long flags; 2314 int dest_cpu; 2315 2316 raw_spin_lock_irqsave(&p->pi_lock, flags); 2317 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2318 if (dest_cpu == smp_processor_id()) 2319 goto unlock; 2320 2321 if (likely(cpu_active(dest_cpu))) { 2322 struct migration_arg arg = { p, dest_cpu }; 2323 2324 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2325 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2326 return; 2327 } 2328unlock: 2329 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2330} 2331 2332#endif 2333 2334DEFINE_PER_CPU(struct kernel_stat, kstat); 2335DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2336 2337EXPORT_PER_CPU_SYMBOL(kstat); 2338EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2339 2340/* 2341 * Return any ns on the sched_clock that have not yet been accounted in 2342 * @p in case that task is currently running. 2343 * 2344 * Called with task_rq_lock() held on @rq. 2345 */ 2346static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2347{ 2348 u64 ns = 0; 2349 2350 if (task_current(rq, p)) { 2351 update_rq_clock(rq); 2352 ns = rq_clock_task(rq) - p->se.exec_start; 2353 if ((s64)ns < 0) 2354 ns = 0; 2355 } 2356 2357 return ns; 2358} 2359 2360unsigned long long task_delta_exec(struct task_struct *p) 2361{ 2362 unsigned long flags; 2363 struct rq *rq; 2364 u64 ns = 0; 2365 2366 rq = task_rq_lock(p, &flags); 2367 ns = do_task_delta_exec(p, rq); 2368 task_rq_unlock(rq, p, &flags); 2369 2370 return ns; 2371} 2372 2373/* 2374 * Return accounted runtime for the task. 2375 * In case the task is currently running, return the runtime plus current's 2376 * pending runtime that have not been accounted yet. 2377 */ 2378unsigned long long task_sched_runtime(struct task_struct *p) 2379{ 2380 unsigned long flags; 2381 struct rq *rq; 2382 u64 ns = 0; 2383 2384#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2385 /* 2386 * 64-bit doesn't need locks to atomically read a 64bit value. 2387 * So we have a optimization chance when the task's delta_exec is 0. 2388 * Reading ->on_cpu is racy, but this is ok. 2389 * 2390 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2391 * If we race with it entering cpu, unaccounted time is 0. This is 2392 * indistinguishable from the read occurring a few cycles earlier. 2393 */ 2394 if (!p->on_cpu) 2395 return p->se.sum_exec_runtime; 2396#endif 2397 2398 rq = task_rq_lock(p, &flags); 2399 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2400 task_rq_unlock(rq, p, &flags); 2401 2402 return ns; 2403} 2404 2405/* 2406 * This function gets called by the timer code, with HZ frequency. 2407 * We call it with interrupts disabled. 2408 */ 2409void scheduler_tick(void) 2410{ 2411 int cpu = smp_processor_id(); 2412 struct rq *rq = cpu_rq(cpu); 2413 struct task_struct *curr = rq->curr; 2414 2415 sched_clock_tick(); 2416 2417 raw_spin_lock(&rq->lock); 2418 update_rq_clock(rq); 2419 curr->sched_class->task_tick(rq, curr, 0); 2420 update_cpu_load_active(rq); 2421 raw_spin_unlock(&rq->lock); 2422 2423 perf_event_task_tick(); 2424 2425#ifdef CONFIG_SMP 2426 rq->idle_balance = idle_cpu(cpu); 2427 trigger_load_balance(rq); 2428#endif 2429 rq_last_tick_reset(rq); 2430} 2431 2432#ifdef CONFIG_NO_HZ_FULL 2433/** 2434 * scheduler_tick_max_deferment 2435 * 2436 * Keep at least one tick per second when a single 2437 * active task is running because the scheduler doesn't 2438 * yet completely support full dynticks environment. 2439 * 2440 * This makes sure that uptime, CFS vruntime, load 2441 * balancing, etc... continue to move forward, even 2442 * with a very low granularity. 2443 * 2444 * Return: Maximum deferment in nanoseconds. 2445 */ 2446u64 scheduler_tick_max_deferment(void) 2447{ 2448 struct rq *rq = this_rq(); 2449 unsigned long next, now = ACCESS_ONCE(jiffies); 2450 2451 next = rq->last_sched_tick + HZ; 2452 2453 if (time_before_eq(next, now)) 2454 return 0; 2455 2456 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2457} 2458#endif 2459 2460notrace unsigned long get_parent_ip(unsigned long addr) 2461{ 2462 if (in_lock_functions(addr)) { 2463 addr = CALLER_ADDR2; 2464 if (in_lock_functions(addr)) 2465 addr = CALLER_ADDR3; 2466 } 2467 return addr; 2468} 2469 2470#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2471 defined(CONFIG_PREEMPT_TRACER)) 2472 2473void __kprobes preempt_count_add(int val) 2474{ 2475#ifdef CONFIG_DEBUG_PREEMPT 2476 /* 2477 * Underflow? 2478 */ 2479 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2480 return; 2481#endif 2482 __preempt_count_add(val); 2483#ifdef CONFIG_DEBUG_PREEMPT 2484 /* 2485 * Spinlock count overflowing soon? 2486 */ 2487 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2488 PREEMPT_MASK - 10); 2489#endif 2490 if (preempt_count() == val) 2491 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2492} 2493EXPORT_SYMBOL(preempt_count_add); 2494 2495void __kprobes preempt_count_sub(int val) 2496{ 2497#ifdef CONFIG_DEBUG_PREEMPT 2498 /* 2499 * Underflow? 2500 */ 2501 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2502 return; 2503 /* 2504 * Is the spinlock portion underflowing? 2505 */ 2506 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2507 !(preempt_count() & PREEMPT_MASK))) 2508 return; 2509#endif 2510 2511 if (preempt_count() == val) 2512 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2513 __preempt_count_sub(val); 2514} 2515EXPORT_SYMBOL(preempt_count_sub); 2516 2517#endif 2518 2519/* 2520 * Print scheduling while atomic bug: 2521 */ 2522static noinline void __schedule_bug(struct task_struct *prev) 2523{ 2524 if (oops_in_progress) 2525 return; 2526 2527 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2528 prev->comm, prev->pid, preempt_count()); 2529 2530 debug_show_held_locks(prev); 2531 print_modules(); 2532 if (irqs_disabled()) 2533 print_irqtrace_events(prev); 2534 dump_stack(); 2535 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2536} 2537 2538/* 2539 * Various schedule()-time debugging checks and statistics: 2540 */ 2541static inline void schedule_debug(struct task_struct *prev) 2542{ 2543 /* 2544 * Test if we are atomic. Since do_exit() needs to call into 2545 * schedule() atomically, we ignore that path. Otherwise whine 2546 * if we are scheduling when we should not. 2547 */ 2548 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2549 __schedule_bug(prev); 2550 rcu_sleep_check(); 2551 2552 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2553 2554 schedstat_inc(this_rq(), sched_count); 2555} 2556 2557static void put_prev_task(struct rq *rq, struct task_struct *prev) 2558{ 2559 if (prev->on_rq || rq->skip_clock_update < 0) 2560 update_rq_clock(rq); 2561 prev->sched_class->put_prev_task(rq, prev); 2562} 2563 2564/* 2565 * Pick up the highest-prio task: 2566 */ 2567static inline struct task_struct * 2568pick_next_task(struct rq *rq) 2569{ 2570 const struct sched_class *class; 2571 struct task_struct *p; 2572 2573 /* 2574 * Optimization: we know that if all tasks are in 2575 * the fair class we can call that function directly: 2576 */ 2577 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2578 p = fair_sched_class.pick_next_task(rq); 2579 if (likely(p)) 2580 return p; 2581 } 2582 2583 for_each_class(class) { 2584 p = class->pick_next_task(rq); 2585 if (p) 2586 return p; 2587 } 2588 2589 BUG(); /* the idle class will always have a runnable task */ 2590} 2591 2592/* 2593 * __schedule() is the main scheduler function. 2594 * 2595 * The main means of driving the scheduler and thus entering this function are: 2596 * 2597 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2598 * 2599 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2600 * paths. For example, see arch/x86/entry_64.S. 2601 * 2602 * To drive preemption between tasks, the scheduler sets the flag in timer 2603 * interrupt handler scheduler_tick(). 2604 * 2605 * 3. Wakeups don't really cause entry into schedule(). They add a 2606 * task to the run-queue and that's it. 2607 * 2608 * Now, if the new task added to the run-queue preempts the current 2609 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2610 * called on the nearest possible occasion: 2611 * 2612 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2613 * 2614 * - in syscall or exception context, at the next outmost 2615 * preempt_enable(). (this might be as soon as the wake_up()'s 2616 * spin_unlock()!) 2617 * 2618 * - in IRQ context, return from interrupt-handler to 2619 * preemptible context 2620 * 2621 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2622 * then at the next: 2623 * 2624 * - cond_resched() call 2625 * - explicit schedule() call 2626 * - return from syscall or exception to user-space 2627 * - return from interrupt-handler to user-space 2628 */ 2629static void __sched __schedule(void) 2630{ 2631 struct task_struct *prev, *next; 2632 unsigned long *switch_count; 2633 struct rq *rq; 2634 int cpu; 2635 2636need_resched: 2637 preempt_disable(); 2638 cpu = smp_processor_id(); 2639 rq = cpu_rq(cpu); 2640 rcu_note_context_switch(cpu); 2641 prev = rq->curr; 2642 2643 schedule_debug(prev); 2644 2645 if (sched_feat(HRTICK)) 2646 hrtick_clear(rq); 2647 2648 /* 2649 * Make sure that signal_pending_state()->signal_pending() below 2650 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2651 * done by the caller to avoid the race with signal_wake_up(). 2652 */ 2653 smp_mb__before_spinlock(); 2654 raw_spin_lock_irq(&rq->lock); 2655 2656 switch_count = &prev->nivcsw; 2657 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2658 if (unlikely(signal_pending_state(prev->state, prev))) { 2659 prev->state = TASK_RUNNING; 2660 } else { 2661 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2662 prev->on_rq = 0; 2663 2664 /* 2665 * If a worker went to sleep, notify and ask workqueue 2666 * whether it wants to wake up a task to maintain 2667 * concurrency. 2668 */ 2669 if (prev->flags & PF_WQ_WORKER) { 2670 struct task_struct *to_wakeup; 2671 2672 to_wakeup = wq_worker_sleeping(prev, cpu); 2673 if (to_wakeup) 2674 try_to_wake_up_local(to_wakeup); 2675 } 2676 } 2677 switch_count = &prev->nvcsw; 2678 } 2679 2680 pre_schedule(rq, prev); 2681 2682 if (unlikely(!rq->nr_running)) 2683 idle_balance(cpu, rq); 2684 2685 put_prev_task(rq, prev); 2686 next = pick_next_task(rq); 2687 clear_tsk_need_resched(prev); 2688 clear_preempt_need_resched(); 2689 rq->skip_clock_update = 0; 2690 2691 if (likely(prev != next)) { 2692 rq->nr_switches++; 2693 rq->curr = next; 2694 ++*switch_count; 2695 2696 context_switch(rq, prev, next); /* unlocks the rq */ 2697 /* 2698 * The context switch have flipped the stack from under us 2699 * and restored the local variables which were saved when 2700 * this task called schedule() in the past. prev == current 2701 * is still correct, but it can be moved to another cpu/rq. 2702 */ 2703 cpu = smp_processor_id(); 2704 rq = cpu_rq(cpu); 2705 } else 2706 raw_spin_unlock_irq(&rq->lock); 2707 2708 post_schedule(rq); 2709 2710 sched_preempt_enable_no_resched(); 2711 if (need_resched()) 2712 goto need_resched; 2713} 2714 2715static inline void sched_submit_work(struct task_struct *tsk) 2716{ 2717 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2718 return; 2719 /* 2720 * If we are going to sleep and we have plugged IO queued, 2721 * make sure to submit it to avoid deadlocks. 2722 */ 2723 if (blk_needs_flush_plug(tsk)) 2724 blk_schedule_flush_plug(tsk); 2725} 2726 2727asmlinkage void __sched schedule(void) 2728{ 2729 struct task_struct *tsk = current; 2730 2731 sched_submit_work(tsk); 2732 __schedule(); 2733} 2734EXPORT_SYMBOL(schedule); 2735 2736#ifdef CONFIG_CONTEXT_TRACKING 2737asmlinkage void __sched schedule_user(void) 2738{ 2739 /* 2740 * If we come here after a random call to set_need_resched(), 2741 * or we have been woken up remotely but the IPI has not yet arrived, 2742 * we haven't yet exited the RCU idle mode. Do it here manually until 2743 * we find a better solution. 2744 */ 2745 user_exit(); 2746 schedule(); 2747 user_enter(); 2748} 2749#endif 2750 2751/** 2752 * schedule_preempt_disabled - called with preemption disabled 2753 * 2754 * Returns with preemption disabled. Note: preempt_count must be 1 2755 */ 2756void __sched schedule_preempt_disabled(void) 2757{ 2758 sched_preempt_enable_no_resched(); 2759 schedule(); 2760 preempt_disable(); 2761} 2762 2763#ifdef CONFIG_PREEMPT 2764/* 2765 * this is the entry point to schedule() from in-kernel preemption 2766 * off of preempt_enable. Kernel preemptions off return from interrupt 2767 * occur there and call schedule directly. 2768 */ 2769asmlinkage void __sched notrace preempt_schedule(void) 2770{ 2771 /* 2772 * If there is a non-zero preempt_count or interrupts are disabled, 2773 * we do not want to preempt the current task. Just return.. 2774 */ 2775 if (likely(!preemptible())) 2776 return; 2777 2778 do { 2779 __preempt_count_add(PREEMPT_ACTIVE); 2780 __schedule(); 2781 __preempt_count_sub(PREEMPT_ACTIVE); 2782 2783 /* 2784 * Check again in case we missed a preemption opportunity 2785 * between schedule and now. 2786 */ 2787 barrier(); 2788 } while (need_resched()); 2789} 2790EXPORT_SYMBOL(preempt_schedule); 2791#endif /* CONFIG_PREEMPT */ 2792 2793/* 2794 * this is the entry point to schedule() from kernel preemption 2795 * off of irq context. 2796 * Note, that this is called and return with irqs disabled. This will 2797 * protect us against recursive calling from irq. 2798 */ 2799asmlinkage void __sched preempt_schedule_irq(void) 2800{ 2801 enum ctx_state prev_state; 2802 2803 /* Catch callers which need to be fixed */ 2804 BUG_ON(preempt_count() || !irqs_disabled()); 2805 2806 prev_state = exception_enter(); 2807 2808 do { 2809 __preempt_count_add(PREEMPT_ACTIVE); 2810 local_irq_enable(); 2811 __schedule(); 2812 local_irq_disable(); 2813 __preempt_count_sub(PREEMPT_ACTIVE); 2814 2815 /* 2816 * Check again in case we missed a preemption opportunity 2817 * between schedule and now. 2818 */ 2819 barrier(); 2820 } while (need_resched()); 2821 2822 exception_exit(prev_state); 2823} 2824 2825int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2826 void *key) 2827{ 2828 return try_to_wake_up(curr->private, mode, wake_flags); 2829} 2830EXPORT_SYMBOL(default_wake_function); 2831 2832static long __sched 2833sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2834{ 2835 unsigned long flags; 2836 wait_queue_t wait; 2837 2838 init_waitqueue_entry(&wait, current); 2839 2840 __set_current_state(state); 2841 2842 spin_lock_irqsave(&q->lock, flags); 2843 __add_wait_queue(q, &wait); 2844 spin_unlock(&q->lock); 2845 timeout = schedule_timeout(timeout); 2846 spin_lock_irq(&q->lock); 2847 __remove_wait_queue(q, &wait); 2848 spin_unlock_irqrestore(&q->lock, flags); 2849 2850 return timeout; 2851} 2852 2853void __sched interruptible_sleep_on(wait_queue_head_t *q) 2854{ 2855 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2856} 2857EXPORT_SYMBOL(interruptible_sleep_on); 2858 2859long __sched 2860interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2861{ 2862 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2863} 2864EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2865 2866void __sched sleep_on(wait_queue_head_t *q) 2867{ 2868 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2869} 2870EXPORT_SYMBOL(sleep_on); 2871 2872long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2873{ 2874 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2875} 2876EXPORT_SYMBOL(sleep_on_timeout); 2877 2878#ifdef CONFIG_RT_MUTEXES 2879 2880/* 2881 * rt_mutex_setprio - set the current priority of a task 2882 * @p: task 2883 * @prio: prio value (kernel-internal form) 2884 * 2885 * This function changes the 'effective' priority of a task. It does 2886 * not touch ->normal_prio like __setscheduler(). 2887 * 2888 * Used by the rt_mutex code to implement priority inheritance logic. 2889 */ 2890void rt_mutex_setprio(struct task_struct *p, int prio) 2891{ 2892 int oldprio, on_rq, running, enqueue_flag = 0; 2893 struct rq *rq; 2894 const struct sched_class *prev_class; 2895 2896 BUG_ON(prio > MAX_PRIO); 2897 2898 rq = __task_rq_lock(p); 2899 2900 /* 2901 * Idle task boosting is a nono in general. There is one 2902 * exception, when PREEMPT_RT and NOHZ is active: 2903 * 2904 * The idle task calls get_next_timer_interrupt() and holds 2905 * the timer wheel base->lock on the CPU and another CPU wants 2906 * to access the timer (probably to cancel it). We can safely 2907 * ignore the boosting request, as the idle CPU runs this code 2908 * with interrupts disabled and will complete the lock 2909 * protected section without being interrupted. So there is no 2910 * real need to boost. 2911 */ 2912 if (unlikely(p == rq->idle)) { 2913 WARN_ON(p != rq->curr); 2914 WARN_ON(p->pi_blocked_on); 2915 goto out_unlock; 2916 } 2917 2918 trace_sched_pi_setprio(p, prio); 2919 p->pi_top_task = rt_mutex_get_top_task(p); 2920 oldprio = p->prio; 2921 prev_class = p->sched_class; 2922 on_rq = p->on_rq; 2923 running = task_current(rq, p); 2924 if (on_rq) 2925 dequeue_task(rq, p, 0); 2926 if (running) 2927 p->sched_class->put_prev_task(rq, p); 2928 2929 /* 2930 * Boosting condition are: 2931 * 1. -rt task is running and holds mutex A 2932 * --> -dl task blocks on mutex A 2933 * 2934 * 2. -dl task is running and holds mutex A 2935 * --> -dl task blocks on mutex A and could preempt the 2936 * running task 2937 */ 2938 if (dl_prio(prio)) { 2939 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2940 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2941 p->dl.dl_boosted = 1; 2942 p->dl.dl_throttled = 0; 2943 enqueue_flag = ENQUEUE_REPLENISH; 2944 } else 2945 p->dl.dl_boosted = 0; 2946 p->sched_class = &dl_sched_class; 2947 } else if (rt_prio(prio)) { 2948 if (dl_prio(oldprio)) 2949 p->dl.dl_boosted = 0; 2950 if (oldprio < prio) 2951 enqueue_flag = ENQUEUE_HEAD; 2952 p->sched_class = &rt_sched_class; 2953 } else { 2954 if (dl_prio(oldprio)) 2955 p->dl.dl_boosted = 0; 2956 p->sched_class = &fair_sched_class; 2957 } 2958 2959 p->prio = prio; 2960 2961 if (running) 2962 p->sched_class->set_curr_task(rq); 2963 if (on_rq) 2964 enqueue_task(rq, p, enqueue_flag); 2965 2966 check_class_changed(rq, p, prev_class, oldprio); 2967out_unlock: 2968 __task_rq_unlock(rq); 2969} 2970#endif 2971 2972void set_user_nice(struct task_struct *p, long nice) 2973{ 2974 int old_prio, delta, on_rq; 2975 unsigned long flags; 2976 struct rq *rq; 2977 2978 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 2979 return; 2980 /* 2981 * We have to be careful, if called from sys_setpriority(), 2982 * the task might be in the middle of scheduling on another CPU. 2983 */ 2984 rq = task_rq_lock(p, &flags); 2985 /* 2986 * The RT priorities are set via sched_setscheduler(), but we still 2987 * allow the 'normal' nice value to be set - but as expected 2988 * it wont have any effect on scheduling until the task is 2989 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 2990 */ 2991 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 2992 p->static_prio = NICE_TO_PRIO(nice); 2993 goto out_unlock; 2994 } 2995 on_rq = p->on_rq; 2996 if (on_rq) 2997 dequeue_task(rq, p, 0); 2998 2999 p->static_prio = NICE_TO_PRIO(nice); 3000 set_load_weight(p); 3001 old_prio = p->prio; 3002 p->prio = effective_prio(p); 3003 delta = p->prio - old_prio; 3004 3005 if (on_rq) { 3006 enqueue_task(rq, p, 0); 3007 /* 3008 * If the task increased its priority or is running and 3009 * lowered its priority, then reschedule its CPU: 3010 */ 3011 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3012 resched_task(rq->curr); 3013 } 3014out_unlock: 3015 task_rq_unlock(rq, p, &flags); 3016} 3017EXPORT_SYMBOL(set_user_nice); 3018 3019/* 3020 * can_nice - check if a task can reduce its nice value 3021 * @p: task 3022 * @nice: nice value 3023 */ 3024int can_nice(const struct task_struct *p, const int nice) 3025{ 3026 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3027 int nice_rlim = 20 - nice; 3028 3029 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3030 capable(CAP_SYS_NICE)); 3031} 3032 3033#ifdef __ARCH_WANT_SYS_NICE 3034 3035/* 3036 * sys_nice - change the priority of the current process. 3037 * @increment: priority increment 3038 * 3039 * sys_setpriority is a more generic, but much slower function that 3040 * does similar things. 3041 */ 3042SYSCALL_DEFINE1(nice, int, increment) 3043{ 3044 long nice, retval; 3045 3046 /* 3047 * Setpriority might change our priority at the same moment. 3048 * We don't have to worry. Conceptually one call occurs first 3049 * and we have a single winner. 3050 */ 3051 if (increment < -40) 3052 increment = -40; 3053 if (increment > 40) 3054 increment = 40; 3055 3056 nice = TASK_NICE(current) + increment; 3057 if (nice < -20) 3058 nice = -20; 3059 if (nice > 19) 3060 nice = 19; 3061 3062 if (increment < 0 && !can_nice(current, nice)) 3063 return -EPERM; 3064 3065 retval = security_task_setnice(current, nice); 3066 if (retval) 3067 return retval; 3068 3069 set_user_nice(current, nice); 3070 return 0; 3071} 3072 3073#endif 3074 3075/** 3076 * task_prio - return the priority value of a given task. 3077 * @p: the task in question. 3078 * 3079 * Return: The priority value as seen by users in /proc. 3080 * RT tasks are offset by -200. Normal tasks are centered 3081 * around 0, value goes from -16 to +15. 3082 */ 3083int task_prio(const struct task_struct *p) 3084{ 3085 return p->prio - MAX_RT_PRIO; 3086} 3087 3088/** 3089 * task_nice - return the nice value of a given task. 3090 * @p: the task in question. 3091 * 3092 * Return: The nice value [ -20 ... 0 ... 19 ]. 3093 */ 3094int task_nice(const struct task_struct *p) 3095{ 3096 return TASK_NICE(p); 3097} 3098EXPORT_SYMBOL(task_nice); 3099 3100/** 3101 * idle_cpu - is a given cpu idle currently? 3102 * @cpu: the processor in question. 3103 * 3104 * Return: 1 if the CPU is currently idle. 0 otherwise. 3105 */ 3106int idle_cpu(int cpu) 3107{ 3108 struct rq *rq = cpu_rq(cpu); 3109 3110 if (rq->curr != rq->idle) 3111 return 0; 3112 3113 if (rq->nr_running) 3114 return 0; 3115 3116#ifdef CONFIG_SMP 3117 if (!llist_empty(&rq->wake_list)) 3118 return 0; 3119#endif 3120 3121 return 1; 3122} 3123 3124/** 3125 * idle_task - return the idle task for a given cpu. 3126 * @cpu: the processor in question. 3127 * 3128 * Return: The idle task for the cpu @cpu. 3129 */ 3130struct task_struct *idle_task(int cpu) 3131{ 3132 return cpu_rq(cpu)->idle; 3133} 3134 3135/** 3136 * find_process_by_pid - find a process with a matching PID value. 3137 * @pid: the pid in question. 3138 * 3139 * The task of @pid, if found. %NULL otherwise. 3140 */ 3141static struct task_struct *find_process_by_pid(pid_t pid) 3142{ 3143 return pid ? find_task_by_vpid(pid) : current; 3144} 3145 3146/* 3147 * This function initializes the sched_dl_entity of a newly becoming 3148 * SCHED_DEADLINE task. 3149 * 3150 * Only the static values are considered here, the actual runtime and the 3151 * absolute deadline will be properly calculated when the task is enqueued 3152 * for the first time with its new policy. 3153 */ 3154static void 3155__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3156{ 3157 struct sched_dl_entity *dl_se = &p->dl; 3158 3159 init_dl_task_timer(dl_se); 3160 dl_se->dl_runtime = attr->sched_runtime; 3161 dl_se->dl_deadline = attr->sched_deadline; 3162 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3163 dl_se->flags = attr->sched_flags; 3164 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3165 dl_se->dl_throttled = 0; 3166 dl_se->dl_new = 1; 3167} 3168 3169/* Actually do priority change: must hold pi & rq lock. */ 3170static void __setscheduler(struct rq *rq, struct task_struct *p, 3171 const struct sched_attr *attr) 3172{ 3173 int policy = attr->sched_policy; 3174 3175 if (policy == -1) /* setparam */ 3176 policy = p->policy; 3177 3178 p->policy = policy; 3179 3180 if (dl_policy(policy)) 3181 __setparam_dl(p, attr); 3182 else if (fair_policy(policy)) 3183 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3184 3185 /* 3186 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3187 * !rt_policy. Always setting this ensures that things like 3188 * getparam()/getattr() don't report silly values for !rt tasks. 3189 */ 3190 p->rt_priority = attr->sched_priority; 3191 3192 p->normal_prio = normal_prio(p); 3193 p->prio = rt_mutex_getprio(p); 3194 3195 if (dl_prio(p->prio)) 3196 p->sched_class = &dl_sched_class; 3197 else if (rt_prio(p->prio)) 3198 p->sched_class = &rt_sched_class; 3199 else 3200 p->sched_class = &fair_sched_class; 3201 3202 set_load_weight(p); 3203} 3204 3205static void 3206__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3207{ 3208 struct sched_dl_entity *dl_se = &p->dl; 3209 3210 attr->sched_priority = p->rt_priority; 3211 attr->sched_runtime = dl_se->dl_runtime; 3212 attr->sched_deadline = dl_se->dl_deadline; 3213 attr->sched_period = dl_se->dl_period; 3214 attr->sched_flags = dl_se->flags; 3215} 3216 3217/* 3218 * This function validates the new parameters of a -deadline task. 3219 * We ask for the deadline not being zero, and greater or equal 3220 * than the runtime, as well as the period of being zero or 3221 * greater than deadline. Furthermore, we have to be sure that 3222 * user parameters are above the internal resolution (1us); we 3223 * check sched_runtime only since it is always the smaller one. 3224 */ 3225static bool 3226__checkparam_dl(const struct sched_attr *attr) 3227{ 3228 return attr && attr->sched_deadline != 0 && 3229 (attr->sched_period == 0 || 3230 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3231 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3232 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3233} 3234 3235/* 3236 * check the target process has a UID that matches the current process's 3237 */ 3238static bool check_same_owner(struct task_struct *p) 3239{ 3240 const struct cred *cred = current_cred(), *pcred; 3241 bool match; 3242 3243 rcu_read_lock(); 3244 pcred = __task_cred(p); 3245 match = (uid_eq(cred->euid, pcred->euid) || 3246 uid_eq(cred->euid, pcred->uid)); 3247 rcu_read_unlock(); 3248 return match; 3249} 3250 3251static int __sched_setscheduler(struct task_struct *p, 3252 const struct sched_attr *attr, 3253 bool user) 3254{ 3255 int retval, oldprio, oldpolicy = -1, on_rq, running; 3256 int policy = attr->sched_policy; 3257 unsigned long flags; 3258 const struct sched_class *prev_class; 3259 struct rq *rq; 3260 int reset_on_fork; 3261 3262 /* may grab non-irq protected spin_locks */ 3263 BUG_ON(in_interrupt()); 3264recheck: 3265 /* double check policy once rq lock held */ 3266 if (policy < 0) { 3267 reset_on_fork = p->sched_reset_on_fork; 3268 policy = oldpolicy = p->policy; 3269 } else { 3270 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3271 3272 if (policy != SCHED_DEADLINE && 3273 policy != SCHED_FIFO && policy != SCHED_RR && 3274 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3275 policy != SCHED_IDLE) 3276 return -EINVAL; 3277 } 3278 3279 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3280 return -EINVAL; 3281 3282 /* 3283 * Valid priorities for SCHED_FIFO and SCHED_RR are 3284 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3285 * SCHED_BATCH and SCHED_IDLE is 0. 3286 */ 3287 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3288 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3289 return -EINVAL; 3290 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3291 (rt_policy(policy) != (attr->sched_priority != 0))) 3292 return -EINVAL; 3293 3294 /* 3295 * Allow unprivileged RT tasks to decrease priority: 3296 */ 3297 if (user && !capable(CAP_SYS_NICE)) { 3298 if (fair_policy(policy)) { 3299 if (attr->sched_nice < TASK_NICE(p) && 3300 !can_nice(p, attr->sched_nice)) 3301 return -EPERM; 3302 } 3303 3304 if (rt_policy(policy)) { 3305 unsigned long rlim_rtprio = 3306 task_rlimit(p, RLIMIT_RTPRIO); 3307 3308 /* can't set/change the rt policy */ 3309 if (policy != p->policy && !rlim_rtprio) 3310 return -EPERM; 3311 3312 /* can't increase priority */ 3313 if (attr->sched_priority > p->rt_priority && 3314 attr->sched_priority > rlim_rtprio) 3315 return -EPERM; 3316 } 3317 3318 /* 3319 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3320 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3321 */ 3322 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3323 if (!can_nice(p, TASK_NICE(p))) 3324 return -EPERM; 3325 } 3326 3327 /* can't change other user's priorities */ 3328 if (!check_same_owner(p)) 3329 return -EPERM; 3330 3331 /* Normal users shall not reset the sched_reset_on_fork flag */ 3332 if (p->sched_reset_on_fork && !reset_on_fork) 3333 return -EPERM; 3334 } 3335 3336 if (user) { 3337 retval = security_task_setscheduler(p); 3338 if (retval) 3339 return retval; 3340 } 3341 3342 /* 3343 * make sure no PI-waiters arrive (or leave) while we are 3344 * changing the priority of the task: 3345 * 3346 * To be able to change p->policy safely, the appropriate 3347 * runqueue lock must be held. 3348 */ 3349 rq = task_rq_lock(p, &flags); 3350 3351 /* 3352 * Changing the policy of the stop threads its a very bad idea 3353 */ 3354 if (p == rq->stop) { 3355 task_rq_unlock(rq, p, &flags); 3356 return -EINVAL; 3357 } 3358 3359 /* 3360 * If not changing anything there's no need to proceed further: 3361 */ 3362 if (unlikely(policy == p->policy)) { 3363 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p)) 3364 goto change; 3365 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3366 goto change; 3367 if (dl_policy(policy)) 3368 goto change; 3369 3370 task_rq_unlock(rq, p, &flags); 3371 return 0; 3372 } 3373change: 3374 3375 if (user) { 3376#ifdef CONFIG_RT_GROUP_SCHED 3377 /* 3378 * Do not allow realtime tasks into groups that have no runtime 3379 * assigned. 3380 */ 3381 if (rt_bandwidth_enabled() && rt_policy(policy) && 3382 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3383 !task_group_is_autogroup(task_group(p))) { 3384 task_rq_unlock(rq, p, &flags); 3385 return -EPERM; 3386 } 3387#endif 3388#ifdef CONFIG_SMP 3389 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3390 cpumask_t *span = rq->rd->span; 3391 3392 /* 3393 * Don't allow tasks with an affinity mask smaller than 3394 * the entire root_domain to become SCHED_DEADLINE. We 3395 * will also fail if there's no bandwidth available. 3396 */ 3397 if (!cpumask_subset(span, &p->cpus_allowed) || 3398 rq->rd->dl_bw.bw == 0) { 3399 task_rq_unlock(rq, p, &flags); 3400 return -EPERM; 3401 } 3402 } 3403#endif 3404 } 3405 3406 /* recheck policy now with rq lock held */ 3407 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3408 policy = oldpolicy = -1; 3409 task_rq_unlock(rq, p, &flags); 3410 goto recheck; 3411 } 3412 3413 /* 3414 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3415 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3416 * is available. 3417 */ 3418 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3419 task_rq_unlock(rq, p, &flags); 3420 return -EBUSY; 3421 } 3422 3423 on_rq = p->on_rq; 3424 running = task_current(rq, p); 3425 if (on_rq) 3426 dequeue_task(rq, p, 0); 3427 if (running) 3428 p->sched_class->put_prev_task(rq, p); 3429 3430 p->sched_reset_on_fork = reset_on_fork; 3431 3432 oldprio = p->prio; 3433 prev_class = p->sched_class; 3434 __setscheduler(rq, p, attr); 3435 3436 if (running) 3437 p->sched_class->set_curr_task(rq); 3438 if (on_rq) 3439 enqueue_task(rq, p, 0); 3440 3441 check_class_changed(rq, p, prev_class, oldprio); 3442 task_rq_unlock(rq, p, &flags); 3443 3444 rt_mutex_adjust_pi(p); 3445 3446 return 0; 3447} 3448 3449static int _sched_setscheduler(struct task_struct *p, int policy, 3450 const struct sched_param *param, bool check) 3451{ 3452 struct sched_attr attr = { 3453 .sched_policy = policy, 3454 .sched_priority = param->sched_priority, 3455 .sched_nice = PRIO_TO_NICE(p->static_prio), 3456 }; 3457 3458 /* 3459 * Fixup the legacy SCHED_RESET_ON_FORK hack 3460 */ 3461 if (policy & SCHED_RESET_ON_FORK) { 3462 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3463 policy &= ~SCHED_RESET_ON_FORK; 3464 attr.sched_policy = policy; 3465 } 3466 3467 return __sched_setscheduler(p, &attr, check); 3468} 3469/** 3470 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3471 * @p: the task in question. 3472 * @policy: new policy. 3473 * @param: structure containing the new RT priority. 3474 * 3475 * Return: 0 on success. An error code otherwise. 3476 * 3477 * NOTE that the task may be already dead. 3478 */ 3479int sched_setscheduler(struct task_struct *p, int policy, 3480 const struct sched_param *param) 3481{ 3482 return _sched_setscheduler(p, policy, param, true); 3483} 3484EXPORT_SYMBOL_GPL(sched_setscheduler); 3485 3486int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3487{ 3488 return __sched_setscheduler(p, attr, true); 3489} 3490EXPORT_SYMBOL_GPL(sched_setattr); 3491 3492/** 3493 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3494 * @p: the task in question. 3495 * @policy: new policy. 3496 * @param: structure containing the new RT priority. 3497 * 3498 * Just like sched_setscheduler, only don't bother checking if the 3499 * current context has permission. For example, this is needed in 3500 * stop_machine(): we create temporary high priority worker threads, 3501 * but our caller might not have that capability. 3502 * 3503 * Return: 0 on success. An error code otherwise. 3504 */ 3505int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3506 const struct sched_param *param) 3507{ 3508 return _sched_setscheduler(p, policy, param, false); 3509} 3510 3511static int 3512do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3513{ 3514 struct sched_param lparam; 3515 struct task_struct *p; 3516 int retval; 3517 3518 if (!param || pid < 0) 3519 return -EINVAL; 3520 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3521 return -EFAULT; 3522 3523 rcu_read_lock(); 3524 retval = -ESRCH; 3525 p = find_process_by_pid(pid); 3526 if (p != NULL) 3527 retval = sched_setscheduler(p, policy, &lparam); 3528 rcu_read_unlock(); 3529 3530 return retval; 3531} 3532 3533/* 3534 * Mimics kernel/events/core.c perf_copy_attr(). 3535 */ 3536static int sched_copy_attr(struct sched_attr __user *uattr, 3537 struct sched_attr *attr) 3538{ 3539 u32 size; 3540 int ret; 3541 3542 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3543 return -EFAULT; 3544 3545 /* 3546 * zero the full structure, so that a short copy will be nice. 3547 */ 3548 memset(attr, 0, sizeof(*attr)); 3549 3550 ret = get_user(size, &uattr->size); 3551 if (ret) 3552 return ret; 3553 3554 if (size > PAGE_SIZE) /* silly large */ 3555 goto err_size; 3556 3557 if (!size) /* abi compat */ 3558 size = SCHED_ATTR_SIZE_VER0; 3559 3560 if (size < SCHED_ATTR_SIZE_VER0) 3561 goto err_size; 3562 3563 /* 3564 * If we're handed a bigger struct than we know of, 3565 * ensure all the unknown bits are 0 - i.e. new 3566 * user-space does not rely on any kernel feature 3567 * extensions we dont know about yet. 3568 */ 3569 if (size > sizeof(*attr)) { 3570 unsigned char __user *addr; 3571 unsigned char __user *end; 3572 unsigned char val; 3573 3574 addr = (void __user *)uattr + sizeof(*attr); 3575 end = (void __user *)uattr + size; 3576 3577 for (; addr < end; addr++) { 3578 ret = get_user(val, addr); 3579 if (ret) 3580 return ret; 3581 if (val) 3582 goto err_size; 3583 } 3584 size = sizeof(*attr); 3585 } 3586 3587 ret = copy_from_user(attr, uattr, size); 3588 if (ret) 3589 return -EFAULT; 3590 3591 /* 3592 * XXX: do we want to be lenient like existing syscalls; or do we want 3593 * to be strict and return an error on out-of-bounds values? 3594 */ 3595 attr->sched_nice = clamp(attr->sched_nice, -20, 19); 3596 3597out: 3598 return ret; 3599 3600err_size: 3601 put_user(sizeof(*attr), &uattr->size); 3602 ret = -E2BIG; 3603 goto out; 3604} 3605 3606/** 3607 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3608 * @pid: the pid in question. 3609 * @policy: new policy. 3610 * @param: structure containing the new RT priority. 3611 * 3612 * Return: 0 on success. An error code otherwise. 3613 */ 3614SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3615 struct sched_param __user *, param) 3616{ 3617 /* negative values for policy are not valid */ 3618 if (policy < 0) 3619 return -EINVAL; 3620 3621 return do_sched_setscheduler(pid, policy, param); 3622} 3623 3624/** 3625 * sys_sched_setparam - set/change the RT priority of a thread 3626 * @pid: the pid in question. 3627 * @param: structure containing the new RT priority. 3628 * 3629 * Return: 0 on success. An error code otherwise. 3630 */ 3631SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3632{ 3633 return do_sched_setscheduler(pid, -1, param); 3634} 3635 3636/** 3637 * sys_sched_setattr - same as above, but with extended sched_attr 3638 * @pid: the pid in question. 3639 * @uattr: structure containing the extended parameters. 3640 */ 3641SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr) 3642{ 3643 struct sched_attr attr; 3644 struct task_struct *p; 3645 int retval; 3646 3647 if (!uattr || pid < 0) 3648 return -EINVAL; 3649 3650 if (sched_copy_attr(uattr, &attr)) 3651 return -EFAULT; 3652 3653 rcu_read_lock(); 3654 retval = -ESRCH; 3655 p = find_process_by_pid(pid); 3656 if (p != NULL) 3657 retval = sched_setattr(p, &attr); 3658 rcu_read_unlock(); 3659 3660 return retval; 3661} 3662 3663/** 3664 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3665 * @pid: the pid in question. 3666 * 3667 * Return: On success, the policy of the thread. Otherwise, a negative error 3668 * code. 3669 */ 3670SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3671{ 3672 struct task_struct *p; 3673 int retval; 3674 3675 if (pid < 0) 3676 return -EINVAL; 3677 3678 retval = -ESRCH; 3679 rcu_read_lock(); 3680 p = find_process_by_pid(pid); 3681 if (p) { 3682 retval = security_task_getscheduler(p); 3683 if (!retval) 3684 retval = p->policy 3685 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3686 } 3687 rcu_read_unlock(); 3688 return retval; 3689} 3690 3691/** 3692 * sys_sched_getparam - get the RT priority of a thread 3693 * @pid: the pid in question. 3694 * @param: structure containing the RT priority. 3695 * 3696 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3697 * code. 3698 */ 3699SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3700{ 3701 struct sched_param lp; 3702 struct task_struct *p; 3703 int retval; 3704 3705 if (!param || pid < 0) 3706 return -EINVAL; 3707 3708 rcu_read_lock(); 3709 p = find_process_by_pid(pid); 3710 retval = -ESRCH; 3711 if (!p) 3712 goto out_unlock; 3713 3714 retval = security_task_getscheduler(p); 3715 if (retval) 3716 goto out_unlock; 3717 3718 if (task_has_dl_policy(p)) { 3719 retval = -EINVAL; 3720 goto out_unlock; 3721 } 3722 lp.sched_priority = p->rt_priority; 3723 rcu_read_unlock(); 3724 3725 /* 3726 * This one might sleep, we cannot do it with a spinlock held ... 3727 */ 3728 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3729 3730 return retval; 3731 3732out_unlock: 3733 rcu_read_unlock(); 3734 return retval; 3735} 3736 3737static int sched_read_attr(struct sched_attr __user *uattr, 3738 struct sched_attr *attr, 3739 unsigned int usize) 3740{ 3741 int ret; 3742 3743 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3744 return -EFAULT; 3745 3746 /* 3747 * If we're handed a smaller struct than we know of, 3748 * ensure all the unknown bits are 0 - i.e. old 3749 * user-space does not get uncomplete information. 3750 */ 3751 if (usize < sizeof(*attr)) { 3752 unsigned char *addr; 3753 unsigned char *end; 3754 3755 addr = (void *)attr + usize; 3756 end = (void *)attr + sizeof(*attr); 3757 3758 for (; addr < end; addr++) { 3759 if (*addr) 3760 goto err_size; 3761 } 3762 3763 attr->size = usize; 3764 } 3765 3766 ret = copy_to_user(uattr, attr, usize); 3767 if (ret) 3768 return -EFAULT; 3769 3770out: 3771 return ret; 3772 3773err_size: 3774 ret = -E2BIG; 3775 goto out; 3776} 3777 3778/** 3779 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3780 * @pid: the pid in question. 3781 * @uattr: structure containing the extended parameters. 3782 * @size: sizeof(attr) for fwd/bwd comp. 3783 */ 3784SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3785 unsigned int, size) 3786{ 3787 struct sched_attr attr = { 3788 .size = sizeof(struct sched_attr), 3789 }; 3790 struct task_struct *p; 3791 int retval; 3792 3793 if (!uattr || pid < 0 || size > PAGE_SIZE || 3794 size < SCHED_ATTR_SIZE_VER0) 3795 return -EINVAL; 3796 3797 rcu_read_lock(); 3798 p = find_process_by_pid(pid); 3799 retval = -ESRCH; 3800 if (!p) 3801 goto out_unlock; 3802 3803 retval = security_task_getscheduler(p); 3804 if (retval) 3805 goto out_unlock; 3806 3807 attr.sched_policy = p->policy; 3808 if (p->sched_reset_on_fork) 3809 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3810 if (task_has_dl_policy(p)) 3811 __getparam_dl(p, &attr); 3812 else if (task_has_rt_policy(p)) 3813 attr.sched_priority = p->rt_priority; 3814 else 3815 attr.sched_nice = TASK_NICE(p); 3816 3817 rcu_read_unlock(); 3818 3819 retval = sched_read_attr(uattr, &attr, size); 3820 return retval; 3821 3822out_unlock: 3823 rcu_read_unlock(); 3824 return retval; 3825} 3826 3827long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3828{ 3829 cpumask_var_t cpus_allowed, new_mask; 3830 struct task_struct *p; 3831 int retval; 3832 3833 rcu_read_lock(); 3834 3835 p = find_process_by_pid(pid); 3836 if (!p) { 3837 rcu_read_unlock(); 3838 return -ESRCH; 3839 } 3840 3841 /* Prevent p going away */ 3842 get_task_struct(p); 3843 rcu_read_unlock(); 3844 3845 if (p->flags & PF_NO_SETAFFINITY) { 3846 retval = -EINVAL; 3847 goto out_put_task; 3848 } 3849 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3850 retval = -ENOMEM; 3851 goto out_put_task; 3852 } 3853 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3854 retval = -ENOMEM; 3855 goto out_free_cpus_allowed; 3856 } 3857 retval = -EPERM; 3858 if (!check_same_owner(p)) { 3859 rcu_read_lock(); 3860 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3861 rcu_read_unlock(); 3862 goto out_unlock; 3863 } 3864 rcu_read_unlock(); 3865 } 3866 3867 retval = security_task_setscheduler(p); 3868 if (retval) 3869 goto out_unlock; 3870 3871 3872 cpuset_cpus_allowed(p, cpus_allowed); 3873 cpumask_and(new_mask, in_mask, cpus_allowed); 3874 3875 /* 3876 * Since bandwidth control happens on root_domain basis, 3877 * if admission test is enabled, we only admit -deadline 3878 * tasks allowed to run on all the CPUs in the task's 3879 * root_domain. 3880 */ 3881#ifdef CONFIG_SMP 3882 if (task_has_dl_policy(p)) { 3883 const struct cpumask *span = task_rq(p)->rd->span; 3884 3885 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3886 retval = -EBUSY; 3887 goto out_unlock; 3888 } 3889 } 3890#endif 3891again: 3892 retval = set_cpus_allowed_ptr(p, new_mask); 3893 3894 if (!retval) { 3895 cpuset_cpus_allowed(p, cpus_allowed); 3896 if (!cpumask_subset(new_mask, cpus_allowed)) { 3897 /* 3898 * We must have raced with a concurrent cpuset 3899 * update. Just reset the cpus_allowed to the 3900 * cpuset's cpus_allowed 3901 */ 3902 cpumask_copy(new_mask, cpus_allowed); 3903 goto again; 3904 } 3905 } 3906out_unlock: 3907 free_cpumask_var(new_mask); 3908out_free_cpus_allowed: 3909 free_cpumask_var(cpus_allowed); 3910out_put_task: 3911 put_task_struct(p); 3912 return retval; 3913} 3914 3915static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3916 struct cpumask *new_mask) 3917{ 3918 if (len < cpumask_size()) 3919 cpumask_clear(new_mask); 3920 else if (len > cpumask_size()) 3921 len = cpumask_size(); 3922 3923 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3924} 3925 3926/** 3927 * sys_sched_setaffinity - set the cpu affinity of a process 3928 * @pid: pid of the process 3929 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3930 * @user_mask_ptr: user-space pointer to the new cpu mask 3931 * 3932 * Return: 0 on success. An error code otherwise. 3933 */ 3934SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3935 unsigned long __user *, user_mask_ptr) 3936{ 3937 cpumask_var_t new_mask; 3938 int retval; 3939 3940 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3941 return -ENOMEM; 3942 3943 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3944 if (retval == 0) 3945 retval = sched_setaffinity(pid, new_mask); 3946 free_cpumask_var(new_mask); 3947 return retval; 3948} 3949 3950long sched_getaffinity(pid_t pid, struct cpumask *mask) 3951{ 3952 struct task_struct *p; 3953 unsigned long flags; 3954 int retval; 3955 3956 rcu_read_lock(); 3957 3958 retval = -ESRCH; 3959 p = find_process_by_pid(pid); 3960 if (!p) 3961 goto out_unlock; 3962 3963 retval = security_task_getscheduler(p); 3964 if (retval) 3965 goto out_unlock; 3966 3967 raw_spin_lock_irqsave(&p->pi_lock, flags); 3968 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3969 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3970 3971out_unlock: 3972 rcu_read_unlock(); 3973 3974 return retval; 3975} 3976 3977/** 3978 * sys_sched_getaffinity - get the cpu affinity of a process 3979 * @pid: pid of the process 3980 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3981 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3982 * 3983 * Return: 0 on success. An error code otherwise. 3984 */ 3985SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3986 unsigned long __user *, user_mask_ptr) 3987{ 3988 int ret; 3989 cpumask_var_t mask; 3990 3991 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3992 return -EINVAL; 3993 if (len & (sizeof(unsigned long)-1)) 3994 return -EINVAL; 3995 3996 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3997 return -ENOMEM; 3998 3999 ret = sched_getaffinity(pid, mask); 4000 if (ret == 0) { 4001 size_t retlen = min_t(size_t, len, cpumask_size()); 4002 4003 if (copy_to_user(user_mask_ptr, mask, retlen)) 4004 ret = -EFAULT; 4005 else 4006 ret = retlen; 4007 } 4008 free_cpumask_var(mask); 4009 4010 return ret; 4011} 4012 4013/** 4014 * sys_sched_yield - yield the current processor to other threads. 4015 * 4016 * This function yields the current CPU to other tasks. If there are no 4017 * other threads running on this CPU then this function will return. 4018 * 4019 * Return: 0. 4020 */ 4021SYSCALL_DEFINE0(sched_yield) 4022{ 4023 struct rq *rq = this_rq_lock(); 4024 4025 schedstat_inc(rq, yld_count); 4026 current->sched_class->yield_task(rq); 4027 4028 /* 4029 * Since we are going to call schedule() anyway, there's 4030 * no need to preempt or enable interrupts: 4031 */ 4032 __release(rq->lock); 4033 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4034 do_raw_spin_unlock(&rq->lock); 4035 sched_preempt_enable_no_resched(); 4036 4037 schedule(); 4038 4039 return 0; 4040} 4041 4042static void __cond_resched(void) 4043{ 4044 __preempt_count_add(PREEMPT_ACTIVE); 4045 __schedule(); 4046 __preempt_count_sub(PREEMPT_ACTIVE); 4047} 4048 4049int __sched _cond_resched(void) 4050{ 4051 if (should_resched()) { 4052 __cond_resched(); 4053 return 1; 4054 } 4055 return 0; 4056} 4057EXPORT_SYMBOL(_cond_resched); 4058 4059/* 4060 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4061 * call schedule, and on return reacquire the lock. 4062 * 4063 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4064 * operations here to prevent schedule() from being called twice (once via 4065 * spin_unlock(), once by hand). 4066 */ 4067int __cond_resched_lock(spinlock_t *lock) 4068{ 4069 int resched = should_resched(); 4070 int ret = 0; 4071 4072 lockdep_assert_held(lock); 4073 4074 if (spin_needbreak(lock) || resched) { 4075 spin_unlock(lock); 4076 if (resched) 4077 __cond_resched(); 4078 else 4079 cpu_relax(); 4080 ret = 1; 4081 spin_lock(lock); 4082 } 4083 return ret; 4084} 4085EXPORT_SYMBOL(__cond_resched_lock); 4086 4087int __sched __cond_resched_softirq(void) 4088{ 4089 BUG_ON(!in_softirq()); 4090 4091 if (should_resched()) { 4092 local_bh_enable(); 4093 __cond_resched(); 4094 local_bh_disable(); 4095 return 1; 4096 } 4097 return 0; 4098} 4099EXPORT_SYMBOL(__cond_resched_softirq); 4100 4101/** 4102 * yield - yield the current processor to other threads. 4103 * 4104 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4105 * 4106 * The scheduler is at all times free to pick the calling task as the most 4107 * eligible task to run, if removing the yield() call from your code breaks 4108 * it, its already broken. 4109 * 4110 * Typical broken usage is: 4111 * 4112 * while (!event) 4113 * yield(); 4114 * 4115 * where one assumes that yield() will let 'the other' process run that will 4116 * make event true. If the current task is a SCHED_FIFO task that will never 4117 * happen. Never use yield() as a progress guarantee!! 4118 * 4119 * If you want to use yield() to wait for something, use wait_event(). 4120 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4121 * If you still want to use yield(), do not! 4122 */ 4123void __sched yield(void) 4124{ 4125 set_current_state(TASK_RUNNING); 4126 sys_sched_yield(); 4127} 4128EXPORT_SYMBOL(yield); 4129 4130/** 4131 * yield_to - yield the current processor to another thread in 4132 * your thread group, or accelerate that thread toward the 4133 * processor it's on. 4134 * @p: target task 4135 * @preempt: whether task preemption is allowed or not 4136 * 4137 * It's the caller's job to ensure that the target task struct 4138 * can't go away on us before we can do any checks. 4139 * 4140 * Return: 4141 * true (>0) if we indeed boosted the target task. 4142 * false (0) if we failed to boost the target. 4143 * -ESRCH if there's no task to yield to. 4144 */ 4145bool __sched yield_to(struct task_struct *p, bool preempt) 4146{ 4147 struct task_struct *curr = current; 4148 struct rq *rq, *p_rq; 4149 unsigned long flags; 4150 int yielded = 0; 4151 4152 local_irq_save(flags); 4153 rq = this_rq(); 4154 4155again: 4156 p_rq = task_rq(p); 4157 /* 4158 * If we're the only runnable task on the rq and target rq also 4159 * has only one task, there's absolutely no point in yielding. 4160 */ 4161 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4162 yielded = -ESRCH; 4163 goto out_irq; 4164 } 4165 4166 double_rq_lock(rq, p_rq); 4167 if (task_rq(p) != p_rq) { 4168 double_rq_unlock(rq, p_rq); 4169 goto again; 4170 } 4171 4172 if (!curr->sched_class->yield_to_task) 4173 goto out_unlock; 4174 4175 if (curr->sched_class != p->sched_class) 4176 goto out_unlock; 4177 4178 if (task_running(p_rq, p) || p->state) 4179 goto out_unlock; 4180 4181 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4182 if (yielded) { 4183 schedstat_inc(rq, yld_count); 4184 /* 4185 * Make p's CPU reschedule; pick_next_entity takes care of 4186 * fairness. 4187 */ 4188 if (preempt && rq != p_rq) 4189 resched_task(p_rq->curr); 4190 } 4191 4192out_unlock: 4193 double_rq_unlock(rq, p_rq); 4194out_irq: 4195 local_irq_restore(flags); 4196 4197 if (yielded > 0) 4198 schedule(); 4199 4200 return yielded; 4201} 4202EXPORT_SYMBOL_GPL(yield_to); 4203 4204/* 4205 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4206 * that process accounting knows that this is a task in IO wait state. 4207 */ 4208void __sched io_schedule(void) 4209{ 4210 struct rq *rq = raw_rq(); 4211 4212 delayacct_blkio_start(); 4213 atomic_inc(&rq->nr_iowait); 4214 blk_flush_plug(current); 4215 current->in_iowait = 1; 4216 schedule(); 4217 current->in_iowait = 0; 4218 atomic_dec(&rq->nr_iowait); 4219 delayacct_blkio_end(); 4220} 4221EXPORT_SYMBOL(io_schedule); 4222 4223long __sched io_schedule_timeout(long timeout) 4224{ 4225 struct rq *rq = raw_rq(); 4226 long ret; 4227 4228 delayacct_blkio_start(); 4229 atomic_inc(&rq->nr_iowait); 4230 blk_flush_plug(current); 4231 current->in_iowait = 1; 4232 ret = schedule_timeout(timeout); 4233 current->in_iowait = 0; 4234 atomic_dec(&rq->nr_iowait); 4235 delayacct_blkio_end(); 4236 return ret; 4237} 4238 4239/** 4240 * sys_sched_get_priority_max - return maximum RT priority. 4241 * @policy: scheduling class. 4242 * 4243 * Return: On success, this syscall returns the maximum 4244 * rt_priority that can be used by a given scheduling class. 4245 * On failure, a negative error code is returned. 4246 */ 4247SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4248{ 4249 int ret = -EINVAL; 4250 4251 switch (policy) { 4252 case SCHED_FIFO: 4253 case SCHED_RR: 4254 ret = MAX_USER_RT_PRIO-1; 4255 break; 4256 case SCHED_DEADLINE: 4257 case SCHED_NORMAL: 4258 case SCHED_BATCH: 4259 case SCHED_IDLE: 4260 ret = 0; 4261 break; 4262 } 4263 return ret; 4264} 4265 4266/** 4267 * sys_sched_get_priority_min - return minimum RT priority. 4268 * @policy: scheduling class. 4269 * 4270 * Return: On success, this syscall returns the minimum 4271 * rt_priority that can be used by a given scheduling class. 4272 * On failure, a negative error code is returned. 4273 */ 4274SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4275{ 4276 int ret = -EINVAL; 4277 4278 switch (policy) { 4279 case SCHED_FIFO: 4280 case SCHED_RR: 4281 ret = 1; 4282 break; 4283 case SCHED_DEADLINE: 4284 case SCHED_NORMAL: 4285 case SCHED_BATCH: 4286 case SCHED_IDLE: 4287 ret = 0; 4288 } 4289 return ret; 4290} 4291 4292/** 4293 * sys_sched_rr_get_interval - return the default timeslice of a process. 4294 * @pid: pid of the process. 4295 * @interval: userspace pointer to the timeslice value. 4296 * 4297 * this syscall writes the default timeslice value of a given process 4298 * into the user-space timespec buffer. A value of '0' means infinity. 4299 * 4300 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4301 * an error code. 4302 */ 4303SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4304 struct timespec __user *, interval) 4305{ 4306 struct task_struct *p; 4307 unsigned int time_slice; 4308 unsigned long flags; 4309 struct rq *rq; 4310 int retval; 4311 struct timespec t; 4312 4313 if (pid < 0) 4314 return -EINVAL; 4315 4316 retval = -ESRCH; 4317 rcu_read_lock(); 4318 p = find_process_by_pid(pid); 4319 if (!p) 4320 goto out_unlock; 4321 4322 retval = security_task_getscheduler(p); 4323 if (retval) 4324 goto out_unlock; 4325 4326 rq = task_rq_lock(p, &flags); 4327 time_slice = 0; 4328 if (p->sched_class->get_rr_interval) 4329 time_slice = p->sched_class->get_rr_interval(rq, p); 4330 task_rq_unlock(rq, p, &flags); 4331 4332 rcu_read_unlock(); 4333 jiffies_to_timespec(time_slice, &t); 4334 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4335 return retval; 4336 4337out_unlock: 4338 rcu_read_unlock(); 4339 return retval; 4340} 4341 4342static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4343 4344void sched_show_task(struct task_struct *p) 4345{ 4346 unsigned long free = 0; 4347 int ppid; 4348 unsigned state; 4349 4350 state = p->state ? __ffs(p->state) + 1 : 0; 4351 printk(KERN_INFO "%-15.15s %c", p->comm, 4352 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4353#if BITS_PER_LONG == 32 4354 if (state == TASK_RUNNING) 4355 printk(KERN_CONT " running "); 4356 else 4357 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4358#else 4359 if (state == TASK_RUNNING) 4360 printk(KERN_CONT " running task "); 4361 else 4362 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4363#endif 4364#ifdef CONFIG_DEBUG_STACK_USAGE 4365 free = stack_not_used(p); 4366#endif 4367 rcu_read_lock(); 4368 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4369 rcu_read_unlock(); 4370 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4371 task_pid_nr(p), ppid, 4372 (unsigned long)task_thread_info(p)->flags); 4373 4374 print_worker_info(KERN_INFO, p); 4375 show_stack(p, NULL); 4376} 4377 4378void show_state_filter(unsigned long state_filter) 4379{ 4380 struct task_struct *g, *p; 4381 4382#if BITS_PER_LONG == 32 4383 printk(KERN_INFO 4384 " task PC stack pid father\n"); 4385#else 4386 printk(KERN_INFO 4387 " task PC stack pid father\n"); 4388#endif 4389 rcu_read_lock(); 4390 do_each_thread(g, p) { 4391 /* 4392 * reset the NMI-timeout, listing all files on a slow 4393 * console might take a lot of time: 4394 */ 4395 touch_nmi_watchdog(); 4396 if (!state_filter || (p->state & state_filter)) 4397 sched_show_task(p); 4398 } while_each_thread(g, p); 4399 4400 touch_all_softlockup_watchdogs(); 4401 4402#ifdef CONFIG_SCHED_DEBUG 4403 sysrq_sched_debug_show(); 4404#endif 4405 rcu_read_unlock(); 4406 /* 4407 * Only show locks if all tasks are dumped: 4408 */ 4409 if (!state_filter) 4410 debug_show_all_locks(); 4411} 4412 4413void init_idle_bootup_task(struct task_struct *idle) 4414{ 4415 idle->sched_class = &idle_sched_class; 4416} 4417 4418/** 4419 * init_idle - set up an idle thread for a given CPU 4420 * @idle: task in question 4421 * @cpu: cpu the idle task belongs to 4422 * 4423 * NOTE: this function does not set the idle thread's NEED_RESCHED 4424 * flag, to make booting more robust. 4425 */ 4426void init_idle(struct task_struct *idle, int cpu) 4427{ 4428 struct rq *rq = cpu_rq(cpu); 4429 unsigned long flags; 4430 4431 raw_spin_lock_irqsave(&rq->lock, flags); 4432 4433 __sched_fork(0, idle); 4434 idle->state = TASK_RUNNING; 4435 idle->se.exec_start = sched_clock(); 4436 4437 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4438 /* 4439 * We're having a chicken and egg problem, even though we are 4440 * holding rq->lock, the cpu isn't yet set to this cpu so the 4441 * lockdep check in task_group() will fail. 4442 * 4443 * Similar case to sched_fork(). / Alternatively we could 4444 * use task_rq_lock() here and obtain the other rq->lock. 4445 * 4446 * Silence PROVE_RCU 4447 */ 4448 rcu_read_lock(); 4449 __set_task_cpu(idle, cpu); 4450 rcu_read_unlock(); 4451 4452 rq->curr = rq->idle = idle; 4453#if defined(CONFIG_SMP) 4454 idle->on_cpu = 1; 4455#endif 4456 raw_spin_unlock_irqrestore(&rq->lock, flags); 4457 4458 /* Set the preempt count _outside_ the spinlocks! */ 4459 init_idle_preempt_count(idle, cpu); 4460 4461 /* 4462 * The idle tasks have their own, simple scheduling class: 4463 */ 4464 idle->sched_class = &idle_sched_class; 4465 ftrace_graph_init_idle_task(idle, cpu); 4466 vtime_init_idle(idle, cpu); 4467#if defined(CONFIG_SMP) 4468 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4469#endif 4470} 4471 4472#ifdef CONFIG_SMP 4473void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4474{ 4475 if (p->sched_class && p->sched_class->set_cpus_allowed) 4476 p->sched_class->set_cpus_allowed(p, new_mask); 4477 4478 cpumask_copy(&p->cpus_allowed, new_mask); 4479 p->nr_cpus_allowed = cpumask_weight(new_mask); 4480} 4481 4482/* 4483 * This is how migration works: 4484 * 4485 * 1) we invoke migration_cpu_stop() on the target CPU using 4486 * stop_one_cpu(). 4487 * 2) stopper starts to run (implicitly forcing the migrated thread 4488 * off the CPU) 4489 * 3) it checks whether the migrated task is still in the wrong runqueue. 4490 * 4) if it's in the wrong runqueue then the migration thread removes 4491 * it and puts it into the right queue. 4492 * 5) stopper completes and stop_one_cpu() returns and the migration 4493 * is done. 4494 */ 4495 4496/* 4497 * Change a given task's CPU affinity. Migrate the thread to a 4498 * proper CPU and schedule it away if the CPU it's executing on 4499 * is removed from the allowed bitmask. 4500 * 4501 * NOTE: the caller must have a valid reference to the task, the 4502 * task must not exit() & deallocate itself prematurely. The 4503 * call is not atomic; no spinlocks may be held. 4504 */ 4505int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4506{ 4507 unsigned long flags; 4508 struct rq *rq; 4509 unsigned int dest_cpu; 4510 int ret = 0; 4511 4512 rq = task_rq_lock(p, &flags); 4513 4514 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4515 goto out; 4516 4517 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4518 ret = -EINVAL; 4519 goto out; 4520 } 4521 4522 do_set_cpus_allowed(p, new_mask); 4523 4524 /* Can the task run on the task's current CPU? If so, we're done */ 4525 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4526 goto out; 4527 4528 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4529 if (p->on_rq) { 4530 struct migration_arg arg = { p, dest_cpu }; 4531 /* Need help from migration thread: drop lock and wait. */ 4532 task_rq_unlock(rq, p, &flags); 4533 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4534 tlb_migrate_finish(p->mm); 4535 return 0; 4536 } 4537out: 4538 task_rq_unlock(rq, p, &flags); 4539 4540 return ret; 4541} 4542EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4543 4544/* 4545 * Move (not current) task off this cpu, onto dest cpu. We're doing 4546 * this because either it can't run here any more (set_cpus_allowed() 4547 * away from this CPU, or CPU going down), or because we're 4548 * attempting to rebalance this task on exec (sched_exec). 4549 * 4550 * So we race with normal scheduler movements, but that's OK, as long 4551 * as the task is no longer on this CPU. 4552 * 4553 * Returns non-zero if task was successfully migrated. 4554 */ 4555static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4556{ 4557 struct rq *rq_dest, *rq_src; 4558 int ret = 0; 4559 4560 if (unlikely(!cpu_active(dest_cpu))) 4561 return ret; 4562 4563 rq_src = cpu_rq(src_cpu); 4564 rq_dest = cpu_rq(dest_cpu); 4565 4566 raw_spin_lock(&p->pi_lock); 4567 double_rq_lock(rq_src, rq_dest); 4568 /* Already moved. */ 4569 if (task_cpu(p) != src_cpu) 4570 goto done; 4571 /* Affinity changed (again). */ 4572 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4573 goto fail; 4574 4575 /* 4576 * If we're not on a rq, the next wake-up will ensure we're 4577 * placed properly. 4578 */ 4579 if (p->on_rq) { 4580 dequeue_task(rq_src, p, 0); 4581 set_task_cpu(p, dest_cpu); 4582 enqueue_task(rq_dest, p, 0); 4583 check_preempt_curr(rq_dest, p, 0); 4584 } 4585done: 4586 ret = 1; 4587fail: 4588 double_rq_unlock(rq_src, rq_dest); 4589 raw_spin_unlock(&p->pi_lock); 4590 return ret; 4591} 4592 4593#ifdef CONFIG_NUMA_BALANCING 4594/* Migrate current task p to target_cpu */ 4595int migrate_task_to(struct task_struct *p, int target_cpu) 4596{ 4597 struct migration_arg arg = { p, target_cpu }; 4598 int curr_cpu = task_cpu(p); 4599 4600 if (curr_cpu == target_cpu) 4601 return 0; 4602 4603 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4604 return -EINVAL; 4605 4606 /* TODO: This is not properly updating schedstats */ 4607 4608 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4609} 4610 4611/* 4612 * Requeue a task on a given node and accurately track the number of NUMA 4613 * tasks on the runqueues 4614 */ 4615void sched_setnuma(struct task_struct *p, int nid) 4616{ 4617 struct rq *rq; 4618 unsigned long flags; 4619 bool on_rq, running; 4620 4621 rq = task_rq_lock(p, &flags); 4622 on_rq = p->on_rq; 4623 running = task_current(rq, p); 4624 4625 if (on_rq) 4626 dequeue_task(rq, p, 0); 4627 if (running) 4628 p->sched_class->put_prev_task(rq, p); 4629 4630 p->numa_preferred_nid = nid; 4631 4632 if (running) 4633 p->sched_class->set_curr_task(rq); 4634 if (on_rq) 4635 enqueue_task(rq, p, 0); 4636 task_rq_unlock(rq, p, &flags); 4637} 4638#endif 4639 4640/* 4641 * migration_cpu_stop - this will be executed by a highprio stopper thread 4642 * and performs thread migration by bumping thread off CPU then 4643 * 'pushing' onto another runqueue. 4644 */ 4645static int migration_cpu_stop(void *data) 4646{ 4647 struct migration_arg *arg = data; 4648 4649 /* 4650 * The original target cpu might have gone down and we might 4651 * be on another cpu but it doesn't matter. 4652 */ 4653 local_irq_disable(); 4654 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4655 local_irq_enable(); 4656 return 0; 4657} 4658 4659#ifdef CONFIG_HOTPLUG_CPU 4660 4661/* 4662 * Ensures that the idle task is using init_mm right before its cpu goes 4663 * offline. 4664 */ 4665void idle_task_exit(void) 4666{ 4667 struct mm_struct *mm = current->active_mm; 4668 4669 BUG_ON(cpu_online(smp_processor_id())); 4670 4671 if (mm != &init_mm) 4672 switch_mm(mm, &init_mm, current); 4673 mmdrop(mm); 4674} 4675 4676/* 4677 * Since this CPU is going 'away' for a while, fold any nr_active delta 4678 * we might have. Assumes we're called after migrate_tasks() so that the 4679 * nr_active count is stable. 4680 * 4681 * Also see the comment "Global load-average calculations". 4682 */ 4683static void calc_load_migrate(struct rq *rq) 4684{ 4685 long delta = calc_load_fold_active(rq); 4686 if (delta) 4687 atomic_long_add(delta, &calc_load_tasks); 4688} 4689 4690/* 4691 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4692 * try_to_wake_up()->select_task_rq(). 4693 * 4694 * Called with rq->lock held even though we'er in stop_machine() and 4695 * there's no concurrency possible, we hold the required locks anyway 4696 * because of lock validation efforts. 4697 */ 4698static void migrate_tasks(unsigned int dead_cpu) 4699{ 4700 struct rq *rq = cpu_rq(dead_cpu); 4701 struct task_struct *next, *stop = rq->stop; 4702 int dest_cpu; 4703 4704 /* 4705 * Fudge the rq selection such that the below task selection loop 4706 * doesn't get stuck on the currently eligible stop task. 4707 * 4708 * We're currently inside stop_machine() and the rq is either stuck 4709 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4710 * either way we should never end up calling schedule() until we're 4711 * done here. 4712 */ 4713 rq->stop = NULL; 4714 4715 /* 4716 * put_prev_task() and pick_next_task() sched 4717 * class method both need to have an up-to-date 4718 * value of rq->clock[_task] 4719 */ 4720 update_rq_clock(rq); 4721 4722 for ( ; ; ) { 4723 /* 4724 * There's this thread running, bail when that's the only 4725 * remaining thread. 4726 */ 4727 if (rq->nr_running == 1) 4728 break; 4729 4730 next = pick_next_task(rq); 4731 BUG_ON(!next); 4732 next->sched_class->put_prev_task(rq, next); 4733 4734 /* Find suitable destination for @next, with force if needed. */ 4735 dest_cpu = select_fallback_rq(dead_cpu, next); 4736 raw_spin_unlock(&rq->lock); 4737 4738 __migrate_task(next, dead_cpu, dest_cpu); 4739 4740 raw_spin_lock(&rq->lock); 4741 } 4742 4743 rq->stop = stop; 4744} 4745 4746#endif /* CONFIG_HOTPLUG_CPU */ 4747 4748#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4749 4750static struct ctl_table sd_ctl_dir[] = { 4751 { 4752 .procname = "sched_domain", 4753 .mode = 0555, 4754 }, 4755 {} 4756}; 4757 4758static struct ctl_table sd_ctl_root[] = { 4759 { 4760 .procname = "kernel", 4761 .mode = 0555, 4762 .child = sd_ctl_dir, 4763 }, 4764 {} 4765}; 4766 4767static struct ctl_table *sd_alloc_ctl_entry(int n) 4768{ 4769 struct ctl_table *entry = 4770 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4771 4772 return entry; 4773} 4774 4775static void sd_free_ctl_entry(struct ctl_table **tablep) 4776{ 4777 struct ctl_table *entry; 4778 4779 /* 4780 * In the intermediate directories, both the child directory and 4781 * procname are dynamically allocated and could fail but the mode 4782 * will always be set. In the lowest directory the names are 4783 * static strings and all have proc handlers. 4784 */ 4785 for (entry = *tablep; entry->mode; entry++) { 4786 if (entry->child) 4787 sd_free_ctl_entry(&entry->child); 4788 if (entry->proc_handler == NULL) 4789 kfree(entry->procname); 4790 } 4791 4792 kfree(*tablep); 4793 *tablep = NULL; 4794} 4795 4796static int min_load_idx = 0; 4797static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4798 4799static void 4800set_table_entry(struct ctl_table *entry, 4801 const char *procname, void *data, int maxlen, 4802 umode_t mode, proc_handler *proc_handler, 4803 bool load_idx) 4804{ 4805 entry->procname = procname; 4806 entry->data = data; 4807 entry->maxlen = maxlen; 4808 entry->mode = mode; 4809 entry->proc_handler = proc_handler; 4810 4811 if (load_idx) { 4812 entry->extra1 = &min_load_idx; 4813 entry->extra2 = &max_load_idx; 4814 } 4815} 4816 4817static struct ctl_table * 4818sd_alloc_ctl_domain_table(struct sched_domain *sd) 4819{ 4820 struct ctl_table *table = sd_alloc_ctl_entry(13); 4821 4822 if (table == NULL) 4823 return NULL; 4824 4825 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4826 sizeof(long), 0644, proc_doulongvec_minmax, false); 4827 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4828 sizeof(long), 0644, proc_doulongvec_minmax, false); 4829 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4830 sizeof(int), 0644, proc_dointvec_minmax, true); 4831 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4832 sizeof(int), 0644, proc_dointvec_minmax, true); 4833 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4834 sizeof(int), 0644, proc_dointvec_minmax, true); 4835 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4836 sizeof(int), 0644, proc_dointvec_minmax, true); 4837 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4838 sizeof(int), 0644, proc_dointvec_minmax, true); 4839 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4840 sizeof(int), 0644, proc_dointvec_minmax, false); 4841 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4842 sizeof(int), 0644, proc_dointvec_minmax, false); 4843 set_table_entry(&table[9], "cache_nice_tries", 4844 &sd->cache_nice_tries, 4845 sizeof(int), 0644, proc_dointvec_minmax, false); 4846 set_table_entry(&table[10], "flags", &sd->flags, 4847 sizeof(int), 0644, proc_dointvec_minmax, false); 4848 set_table_entry(&table[11], "name", sd->name, 4849 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4850 /* &table[12] is terminator */ 4851 4852 return table; 4853} 4854 4855static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4856{ 4857 struct ctl_table *entry, *table; 4858 struct sched_domain *sd; 4859 int domain_num = 0, i; 4860 char buf[32]; 4861 4862 for_each_domain(cpu, sd) 4863 domain_num++; 4864 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4865 if (table == NULL) 4866 return NULL; 4867 4868 i = 0; 4869 for_each_domain(cpu, sd) { 4870 snprintf(buf, 32, "domain%d", i); 4871 entry->procname = kstrdup(buf, GFP_KERNEL); 4872 entry->mode = 0555; 4873 entry->child = sd_alloc_ctl_domain_table(sd); 4874 entry++; 4875 i++; 4876 } 4877 return table; 4878} 4879 4880static struct ctl_table_header *sd_sysctl_header; 4881static void register_sched_domain_sysctl(void) 4882{ 4883 int i, cpu_num = num_possible_cpus(); 4884 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4885 char buf[32]; 4886 4887 WARN_ON(sd_ctl_dir[0].child); 4888 sd_ctl_dir[0].child = entry; 4889 4890 if (entry == NULL) 4891 return; 4892 4893 for_each_possible_cpu(i) { 4894 snprintf(buf, 32, "cpu%d", i); 4895 entry->procname = kstrdup(buf, GFP_KERNEL); 4896 entry->mode = 0555; 4897 entry->child = sd_alloc_ctl_cpu_table(i); 4898 entry++; 4899 } 4900 4901 WARN_ON(sd_sysctl_header); 4902 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4903} 4904 4905/* may be called multiple times per register */ 4906static void unregister_sched_domain_sysctl(void) 4907{ 4908 if (sd_sysctl_header) 4909 unregister_sysctl_table(sd_sysctl_header); 4910 sd_sysctl_header = NULL; 4911 if (sd_ctl_dir[0].child) 4912 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4913} 4914#else 4915static void register_sched_domain_sysctl(void) 4916{ 4917} 4918static void unregister_sched_domain_sysctl(void) 4919{ 4920} 4921#endif 4922 4923static void set_rq_online(struct rq *rq) 4924{ 4925 if (!rq->online) { 4926 const struct sched_class *class; 4927 4928 cpumask_set_cpu(rq->cpu, rq->rd->online); 4929 rq->online = 1; 4930 4931 for_each_class(class) { 4932 if (class->rq_online) 4933 class->rq_online(rq); 4934 } 4935 } 4936} 4937 4938static void set_rq_offline(struct rq *rq) 4939{ 4940 if (rq->online) { 4941 const struct sched_class *class; 4942 4943 for_each_class(class) { 4944 if (class->rq_offline) 4945 class->rq_offline(rq); 4946 } 4947 4948 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4949 rq->online = 0; 4950 } 4951} 4952 4953/* 4954 * migration_call - callback that gets triggered when a CPU is added. 4955 * Here we can start up the necessary migration thread for the new CPU. 4956 */ 4957static int 4958migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4959{ 4960 int cpu = (long)hcpu; 4961 unsigned long flags; 4962 struct rq *rq = cpu_rq(cpu); 4963 4964 switch (action & ~CPU_TASKS_FROZEN) { 4965 4966 case CPU_UP_PREPARE: 4967 rq->calc_load_update = calc_load_update; 4968 break; 4969 4970 case CPU_ONLINE: 4971 /* Update our root-domain */ 4972 raw_spin_lock_irqsave(&rq->lock, flags); 4973 if (rq->rd) { 4974 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4975 4976 set_rq_online(rq); 4977 } 4978 raw_spin_unlock_irqrestore(&rq->lock, flags); 4979 break; 4980 4981#ifdef CONFIG_HOTPLUG_CPU 4982 case CPU_DYING: 4983 sched_ttwu_pending(); 4984 /* Update our root-domain */ 4985 raw_spin_lock_irqsave(&rq->lock, flags); 4986 if (rq->rd) { 4987 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4988 set_rq_offline(rq); 4989 } 4990 migrate_tasks(cpu); 4991 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4992 raw_spin_unlock_irqrestore(&rq->lock, flags); 4993 break; 4994 4995 case CPU_DEAD: 4996 calc_load_migrate(rq); 4997 break; 4998#endif 4999 } 5000 5001 update_max_interval(); 5002 5003 return NOTIFY_OK; 5004} 5005 5006/* 5007 * Register at high priority so that task migration (migrate_all_tasks) 5008 * happens before everything else. This has to be lower priority than 5009 * the notifier in the perf_event subsystem, though. 5010 */ 5011static struct notifier_block migration_notifier = { 5012 .notifier_call = migration_call, 5013 .priority = CPU_PRI_MIGRATION, 5014}; 5015 5016static int sched_cpu_active(struct notifier_block *nfb, 5017 unsigned long action, void *hcpu) 5018{ 5019 switch (action & ~CPU_TASKS_FROZEN) { 5020 case CPU_STARTING: 5021 case CPU_DOWN_FAILED: 5022 set_cpu_active((long)hcpu, true); 5023 return NOTIFY_OK; 5024 default: 5025 return NOTIFY_DONE; 5026 } 5027} 5028 5029static int sched_cpu_inactive(struct notifier_block *nfb, 5030 unsigned long action, void *hcpu) 5031{ 5032 unsigned long flags; 5033 long cpu = (long)hcpu; 5034 5035 switch (action & ~CPU_TASKS_FROZEN) { 5036 case CPU_DOWN_PREPARE: 5037 set_cpu_active(cpu, false); 5038 5039 /* explicitly allow suspend */ 5040 if (!(action & CPU_TASKS_FROZEN)) { 5041 struct dl_bw *dl_b = dl_bw_of(cpu); 5042 bool overflow; 5043 int cpus; 5044 5045 raw_spin_lock_irqsave(&dl_b->lock, flags); 5046 cpus = dl_bw_cpus(cpu); 5047 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5048 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5049 5050 if (overflow) 5051 return notifier_from_errno(-EBUSY); 5052 } 5053 return NOTIFY_OK; 5054 } 5055 5056 return NOTIFY_DONE; 5057} 5058 5059static int __init migration_init(void) 5060{ 5061 void *cpu = (void *)(long)smp_processor_id(); 5062 int err; 5063 5064 /* Initialize migration for the boot CPU */ 5065 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5066 BUG_ON(err == NOTIFY_BAD); 5067 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5068 register_cpu_notifier(&migration_notifier); 5069 5070 /* Register cpu active notifiers */ 5071 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5072 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5073 5074 return 0; 5075} 5076early_initcall(migration_init); 5077#endif 5078 5079#ifdef CONFIG_SMP 5080 5081static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5082 5083#ifdef CONFIG_SCHED_DEBUG 5084 5085static __read_mostly int sched_debug_enabled; 5086 5087static int __init sched_debug_setup(char *str) 5088{ 5089 sched_debug_enabled = 1; 5090 5091 return 0; 5092} 5093early_param("sched_debug", sched_debug_setup); 5094 5095static inline bool sched_debug(void) 5096{ 5097 return sched_debug_enabled; 5098} 5099 5100static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5101 struct cpumask *groupmask) 5102{ 5103 struct sched_group *group = sd->groups; 5104 char str[256]; 5105 5106 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5107 cpumask_clear(groupmask); 5108 5109 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5110 5111 if (!(sd->flags & SD_LOAD_BALANCE)) { 5112 printk("does not load-balance\n"); 5113 if (sd->parent) 5114 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5115 " has parent"); 5116 return -1; 5117 } 5118 5119 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5120 5121 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5122 printk(KERN_ERR "ERROR: domain->span does not contain " 5123 "CPU%d\n", cpu); 5124 } 5125 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5126 printk(KERN_ERR "ERROR: domain->groups does not contain" 5127 " CPU%d\n", cpu); 5128 } 5129 5130 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5131 do { 5132 if (!group) { 5133 printk("\n"); 5134 printk(KERN_ERR "ERROR: group is NULL\n"); 5135 break; 5136 } 5137 5138 /* 5139 * Even though we initialize ->power to something semi-sane, 5140 * we leave power_orig unset. This allows us to detect if 5141 * domain iteration is still funny without causing /0 traps. 5142 */ 5143 if (!group->sgp->power_orig) { 5144 printk(KERN_CONT "\n"); 5145 printk(KERN_ERR "ERROR: domain->cpu_power not " 5146 "set\n"); 5147 break; 5148 } 5149 5150 if (!cpumask_weight(sched_group_cpus(group))) { 5151 printk(KERN_CONT "\n"); 5152 printk(KERN_ERR "ERROR: empty group\n"); 5153 break; 5154 } 5155 5156 if (!(sd->flags & SD_OVERLAP) && 5157 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5158 printk(KERN_CONT "\n"); 5159 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5160 break; 5161 } 5162 5163 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5164 5165 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5166 5167 printk(KERN_CONT " %s", str); 5168 if (group->sgp->power != SCHED_POWER_SCALE) { 5169 printk(KERN_CONT " (cpu_power = %d)", 5170 group->sgp->power); 5171 } 5172 5173 group = group->next; 5174 } while (group != sd->groups); 5175 printk(KERN_CONT "\n"); 5176 5177 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5178 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5179 5180 if (sd->parent && 5181 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5182 printk(KERN_ERR "ERROR: parent span is not a superset " 5183 "of domain->span\n"); 5184 return 0; 5185} 5186 5187static void sched_domain_debug(struct sched_domain *sd, int cpu) 5188{ 5189 int level = 0; 5190 5191 if (!sched_debug_enabled) 5192 return; 5193 5194 if (!sd) { 5195 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5196 return; 5197 } 5198 5199 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5200 5201 for (;;) { 5202 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5203 break; 5204 level++; 5205 sd = sd->parent; 5206 if (!sd) 5207 break; 5208 } 5209} 5210#else /* !CONFIG_SCHED_DEBUG */ 5211# define sched_domain_debug(sd, cpu) do { } while (0) 5212static inline bool sched_debug(void) 5213{ 5214 return false; 5215} 5216#endif /* CONFIG_SCHED_DEBUG */ 5217 5218static int sd_degenerate(struct sched_domain *sd) 5219{ 5220 if (cpumask_weight(sched_domain_span(sd)) == 1) 5221 return 1; 5222 5223 /* Following flags need at least 2 groups */ 5224 if (sd->flags & (SD_LOAD_BALANCE | 5225 SD_BALANCE_NEWIDLE | 5226 SD_BALANCE_FORK | 5227 SD_BALANCE_EXEC | 5228 SD_SHARE_CPUPOWER | 5229 SD_SHARE_PKG_RESOURCES)) { 5230 if (sd->groups != sd->groups->next) 5231 return 0; 5232 } 5233 5234 /* Following flags don't use groups */ 5235 if (sd->flags & (SD_WAKE_AFFINE)) 5236 return 0; 5237 5238 return 1; 5239} 5240 5241static int 5242sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5243{ 5244 unsigned long cflags = sd->flags, pflags = parent->flags; 5245 5246 if (sd_degenerate(parent)) 5247 return 1; 5248 5249 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5250 return 0; 5251 5252 /* Flags needing groups don't count if only 1 group in parent */ 5253 if (parent->groups == parent->groups->next) { 5254 pflags &= ~(SD_LOAD_BALANCE | 5255 SD_BALANCE_NEWIDLE | 5256 SD_BALANCE_FORK | 5257 SD_BALANCE_EXEC | 5258 SD_SHARE_CPUPOWER | 5259 SD_SHARE_PKG_RESOURCES | 5260 SD_PREFER_SIBLING); 5261 if (nr_node_ids == 1) 5262 pflags &= ~SD_SERIALIZE; 5263 } 5264 if (~cflags & pflags) 5265 return 0; 5266 5267 return 1; 5268} 5269 5270static void free_rootdomain(struct rcu_head *rcu) 5271{ 5272 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5273 5274 cpupri_cleanup(&rd->cpupri); 5275 cpudl_cleanup(&rd->cpudl); 5276 free_cpumask_var(rd->dlo_mask); 5277 free_cpumask_var(rd->rto_mask); 5278 free_cpumask_var(rd->online); 5279 free_cpumask_var(rd->span); 5280 kfree(rd); 5281} 5282 5283static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5284{ 5285 struct root_domain *old_rd = NULL; 5286 unsigned long flags; 5287 5288 raw_spin_lock_irqsave(&rq->lock, flags); 5289 5290 if (rq->rd) { 5291 old_rd = rq->rd; 5292 5293 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5294 set_rq_offline(rq); 5295 5296 cpumask_clear_cpu(rq->cpu, old_rd->span); 5297 5298 /* 5299 * If we dont want to free the old_rd yet then 5300 * set old_rd to NULL to skip the freeing later 5301 * in this function: 5302 */ 5303 if (!atomic_dec_and_test(&old_rd->refcount)) 5304 old_rd = NULL; 5305 } 5306 5307 atomic_inc(&rd->refcount); 5308 rq->rd = rd; 5309 5310 cpumask_set_cpu(rq->cpu, rd->span); 5311 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5312 set_rq_online(rq); 5313 5314 raw_spin_unlock_irqrestore(&rq->lock, flags); 5315 5316 if (old_rd) 5317 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5318} 5319 5320static int init_rootdomain(struct root_domain *rd) 5321{ 5322 memset(rd, 0, sizeof(*rd)); 5323 5324 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5325 goto out; 5326 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5327 goto free_span; 5328 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5329 goto free_online; 5330 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5331 goto free_dlo_mask; 5332 5333 init_dl_bw(&rd->dl_bw); 5334 if (cpudl_init(&rd->cpudl) != 0) 5335 goto free_dlo_mask; 5336 5337 if (cpupri_init(&rd->cpupri) != 0) 5338 goto free_rto_mask; 5339 return 0; 5340 5341free_rto_mask: 5342 free_cpumask_var(rd->rto_mask); 5343free_dlo_mask: 5344 free_cpumask_var(rd->dlo_mask); 5345free_online: 5346 free_cpumask_var(rd->online); 5347free_span: 5348 free_cpumask_var(rd->span); 5349out: 5350 return -ENOMEM; 5351} 5352 5353/* 5354 * By default the system creates a single root-domain with all cpus as 5355 * members (mimicking the global state we have today). 5356 */ 5357struct root_domain def_root_domain; 5358 5359static void init_defrootdomain(void) 5360{ 5361 init_rootdomain(&def_root_domain); 5362 5363 atomic_set(&def_root_domain.refcount, 1); 5364} 5365 5366static struct root_domain *alloc_rootdomain(void) 5367{ 5368 struct root_domain *rd; 5369 5370 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5371 if (!rd) 5372 return NULL; 5373 5374 if (init_rootdomain(rd) != 0) { 5375 kfree(rd); 5376 return NULL; 5377 } 5378 5379 return rd; 5380} 5381 5382static void free_sched_groups(struct sched_group *sg, int free_sgp) 5383{ 5384 struct sched_group *tmp, *first; 5385 5386 if (!sg) 5387 return; 5388 5389 first = sg; 5390 do { 5391 tmp = sg->next; 5392 5393 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5394 kfree(sg->sgp); 5395 5396 kfree(sg); 5397 sg = tmp; 5398 } while (sg != first); 5399} 5400 5401static void free_sched_domain(struct rcu_head *rcu) 5402{ 5403 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5404 5405 /* 5406 * If its an overlapping domain it has private groups, iterate and 5407 * nuke them all. 5408 */ 5409 if (sd->flags & SD_OVERLAP) { 5410 free_sched_groups(sd->groups, 1); 5411 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5412 kfree(sd->groups->sgp); 5413 kfree(sd->groups); 5414 } 5415 kfree(sd); 5416} 5417 5418static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5419{ 5420 call_rcu(&sd->rcu, free_sched_domain); 5421} 5422 5423static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5424{ 5425 for (; sd; sd = sd->parent) 5426 destroy_sched_domain(sd, cpu); 5427} 5428 5429/* 5430 * Keep a special pointer to the highest sched_domain that has 5431 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5432 * allows us to avoid some pointer chasing select_idle_sibling(). 5433 * 5434 * Also keep a unique ID per domain (we use the first cpu number in 5435 * the cpumask of the domain), this allows us to quickly tell if 5436 * two cpus are in the same cache domain, see cpus_share_cache(). 5437 */ 5438DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5439DEFINE_PER_CPU(int, sd_llc_size); 5440DEFINE_PER_CPU(int, sd_llc_id); 5441DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5442DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5443DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5444 5445static void update_top_cache_domain(int cpu) 5446{ 5447 struct sched_domain *sd; 5448 struct sched_domain *busy_sd = NULL; 5449 int id = cpu; 5450 int size = 1; 5451 5452 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5453 if (sd) { 5454 id = cpumask_first(sched_domain_span(sd)); 5455 size = cpumask_weight(sched_domain_span(sd)); 5456 busy_sd = sd->parent; /* sd_busy */ 5457 } 5458 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5459 5460 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5461 per_cpu(sd_llc_size, cpu) = size; 5462 per_cpu(sd_llc_id, cpu) = id; 5463 5464 sd = lowest_flag_domain(cpu, SD_NUMA); 5465 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5466 5467 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5468 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5469} 5470 5471/* 5472 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5473 * hold the hotplug lock. 5474 */ 5475static void 5476cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5477{ 5478 struct rq *rq = cpu_rq(cpu); 5479 struct sched_domain *tmp; 5480 5481 /* Remove the sched domains which do not contribute to scheduling. */ 5482 for (tmp = sd; tmp; ) { 5483 struct sched_domain *parent = tmp->parent; 5484 if (!parent) 5485 break; 5486 5487 if (sd_parent_degenerate(tmp, parent)) { 5488 tmp->parent = parent->parent; 5489 if (parent->parent) 5490 parent->parent->child = tmp; 5491 /* 5492 * Transfer SD_PREFER_SIBLING down in case of a 5493 * degenerate parent; the spans match for this 5494 * so the property transfers. 5495 */ 5496 if (parent->flags & SD_PREFER_SIBLING) 5497 tmp->flags |= SD_PREFER_SIBLING; 5498 destroy_sched_domain(parent, cpu); 5499 } else 5500 tmp = tmp->parent; 5501 } 5502 5503 if (sd && sd_degenerate(sd)) { 5504 tmp = sd; 5505 sd = sd->parent; 5506 destroy_sched_domain(tmp, cpu); 5507 if (sd) 5508 sd->child = NULL; 5509 } 5510 5511 sched_domain_debug(sd, cpu); 5512 5513 rq_attach_root(rq, rd); 5514 tmp = rq->sd; 5515 rcu_assign_pointer(rq->sd, sd); 5516 destroy_sched_domains(tmp, cpu); 5517 5518 update_top_cache_domain(cpu); 5519} 5520 5521/* cpus with isolated domains */ 5522static cpumask_var_t cpu_isolated_map; 5523 5524/* Setup the mask of cpus configured for isolated domains */ 5525static int __init isolated_cpu_setup(char *str) 5526{ 5527 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5528 cpulist_parse(str, cpu_isolated_map); 5529 return 1; 5530} 5531 5532__setup("isolcpus=", isolated_cpu_setup); 5533 5534static const struct cpumask *cpu_cpu_mask(int cpu) 5535{ 5536 return cpumask_of_node(cpu_to_node(cpu)); 5537} 5538 5539struct sd_data { 5540 struct sched_domain **__percpu sd; 5541 struct sched_group **__percpu sg; 5542 struct sched_group_power **__percpu sgp; 5543}; 5544 5545struct s_data { 5546 struct sched_domain ** __percpu sd; 5547 struct root_domain *rd; 5548}; 5549 5550enum s_alloc { 5551 sa_rootdomain, 5552 sa_sd, 5553 sa_sd_storage, 5554 sa_none, 5555}; 5556 5557struct sched_domain_topology_level; 5558 5559typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5560typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5561 5562#define SDTL_OVERLAP 0x01 5563 5564struct sched_domain_topology_level { 5565 sched_domain_init_f init; 5566 sched_domain_mask_f mask; 5567 int flags; 5568 int numa_level; 5569 struct sd_data data; 5570}; 5571 5572/* 5573 * Build an iteration mask that can exclude certain CPUs from the upwards 5574 * domain traversal. 5575 * 5576 * Asymmetric node setups can result in situations where the domain tree is of 5577 * unequal depth, make sure to skip domains that already cover the entire 5578 * range. 5579 * 5580 * In that case build_sched_domains() will have terminated the iteration early 5581 * and our sibling sd spans will be empty. Domains should always include the 5582 * cpu they're built on, so check that. 5583 * 5584 */ 5585static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5586{ 5587 const struct cpumask *span = sched_domain_span(sd); 5588 struct sd_data *sdd = sd->private; 5589 struct sched_domain *sibling; 5590 int i; 5591 5592 for_each_cpu(i, span) { 5593 sibling = *per_cpu_ptr(sdd->sd, i); 5594 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5595 continue; 5596 5597 cpumask_set_cpu(i, sched_group_mask(sg)); 5598 } 5599} 5600 5601/* 5602 * Return the canonical balance cpu for this group, this is the first cpu 5603 * of this group that's also in the iteration mask. 5604 */ 5605int group_balance_cpu(struct sched_group *sg) 5606{ 5607 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5608} 5609 5610static int 5611build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5612{ 5613 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5614 const struct cpumask *span = sched_domain_span(sd); 5615 struct cpumask *covered = sched_domains_tmpmask; 5616 struct sd_data *sdd = sd->private; 5617 struct sched_domain *child; 5618 int i; 5619 5620 cpumask_clear(covered); 5621 5622 for_each_cpu(i, span) { 5623 struct cpumask *sg_span; 5624 5625 if (cpumask_test_cpu(i, covered)) 5626 continue; 5627 5628 child = *per_cpu_ptr(sdd->sd, i); 5629 5630 /* See the comment near build_group_mask(). */ 5631 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5632 continue; 5633 5634 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5635 GFP_KERNEL, cpu_to_node(cpu)); 5636 5637 if (!sg) 5638 goto fail; 5639 5640 sg_span = sched_group_cpus(sg); 5641 if (child->child) { 5642 child = child->child; 5643 cpumask_copy(sg_span, sched_domain_span(child)); 5644 } else 5645 cpumask_set_cpu(i, sg_span); 5646 5647 cpumask_or(covered, covered, sg_span); 5648 5649 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5650 if (atomic_inc_return(&sg->sgp->ref) == 1) 5651 build_group_mask(sd, sg); 5652 5653 /* 5654 * Initialize sgp->power such that even if we mess up the 5655 * domains and no possible iteration will get us here, we won't 5656 * die on a /0 trap. 5657 */ 5658 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5659 sg->sgp->power_orig = sg->sgp->power; 5660 5661 /* 5662 * Make sure the first group of this domain contains the 5663 * canonical balance cpu. Otherwise the sched_domain iteration 5664 * breaks. See update_sg_lb_stats(). 5665 */ 5666 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5667 group_balance_cpu(sg) == cpu) 5668 groups = sg; 5669 5670 if (!first) 5671 first = sg; 5672 if (last) 5673 last->next = sg; 5674 last = sg; 5675 last->next = first; 5676 } 5677 sd->groups = groups; 5678 5679 return 0; 5680 5681fail: 5682 free_sched_groups(first, 0); 5683 5684 return -ENOMEM; 5685} 5686 5687static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5688{ 5689 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5690 struct sched_domain *child = sd->child; 5691 5692 if (child) 5693 cpu = cpumask_first(sched_domain_span(child)); 5694 5695 if (sg) { 5696 *sg = *per_cpu_ptr(sdd->sg, cpu); 5697 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5698 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5699 } 5700 5701 return cpu; 5702} 5703 5704/* 5705 * build_sched_groups will build a circular linked list of the groups 5706 * covered by the given span, and will set each group's ->cpumask correctly, 5707 * and ->cpu_power to 0. 5708 * 5709 * Assumes the sched_domain tree is fully constructed 5710 */ 5711static int 5712build_sched_groups(struct sched_domain *sd, int cpu) 5713{ 5714 struct sched_group *first = NULL, *last = NULL; 5715 struct sd_data *sdd = sd->private; 5716 const struct cpumask *span = sched_domain_span(sd); 5717 struct cpumask *covered; 5718 int i; 5719 5720 get_group(cpu, sdd, &sd->groups); 5721 atomic_inc(&sd->groups->ref); 5722 5723 if (cpu != cpumask_first(span)) 5724 return 0; 5725 5726 lockdep_assert_held(&sched_domains_mutex); 5727 covered = sched_domains_tmpmask; 5728 5729 cpumask_clear(covered); 5730 5731 for_each_cpu(i, span) { 5732 struct sched_group *sg; 5733 int group, j; 5734 5735 if (cpumask_test_cpu(i, covered)) 5736 continue; 5737 5738 group = get_group(i, sdd, &sg); 5739 cpumask_clear(sched_group_cpus(sg)); 5740 sg->sgp->power = 0; 5741 cpumask_setall(sched_group_mask(sg)); 5742 5743 for_each_cpu(j, span) { 5744 if (get_group(j, sdd, NULL) != group) 5745 continue; 5746 5747 cpumask_set_cpu(j, covered); 5748 cpumask_set_cpu(j, sched_group_cpus(sg)); 5749 } 5750 5751 if (!first) 5752 first = sg; 5753 if (last) 5754 last->next = sg; 5755 last = sg; 5756 } 5757 last->next = first; 5758 5759 return 0; 5760} 5761 5762/* 5763 * Initialize sched groups cpu_power. 5764 * 5765 * cpu_power indicates the capacity of sched group, which is used while 5766 * distributing the load between different sched groups in a sched domain. 5767 * Typically cpu_power for all the groups in a sched domain will be same unless 5768 * there are asymmetries in the topology. If there are asymmetries, group 5769 * having more cpu_power will pickup more load compared to the group having 5770 * less cpu_power. 5771 */ 5772static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5773{ 5774 struct sched_group *sg = sd->groups; 5775 5776 WARN_ON(!sg); 5777 5778 do { 5779 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5780 sg = sg->next; 5781 } while (sg != sd->groups); 5782 5783 if (cpu != group_balance_cpu(sg)) 5784 return; 5785 5786 update_group_power(sd, cpu); 5787 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5788} 5789 5790int __weak arch_sd_sibling_asym_packing(void) 5791{ 5792 return 0*SD_ASYM_PACKING; 5793} 5794 5795/* 5796 * Initializers for schedule domains 5797 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5798 */ 5799 5800#ifdef CONFIG_SCHED_DEBUG 5801# define SD_INIT_NAME(sd, type) sd->name = #type 5802#else 5803# define SD_INIT_NAME(sd, type) do { } while (0) 5804#endif 5805 5806#define SD_INIT_FUNC(type) \ 5807static noinline struct sched_domain * \ 5808sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5809{ \ 5810 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5811 *sd = SD_##type##_INIT; \ 5812 SD_INIT_NAME(sd, type); \ 5813 sd->private = &tl->data; \ 5814 return sd; \ 5815} 5816 5817SD_INIT_FUNC(CPU) 5818#ifdef CONFIG_SCHED_SMT 5819 SD_INIT_FUNC(SIBLING) 5820#endif 5821#ifdef CONFIG_SCHED_MC 5822 SD_INIT_FUNC(MC) 5823#endif 5824#ifdef CONFIG_SCHED_BOOK 5825 SD_INIT_FUNC(BOOK) 5826#endif 5827 5828static int default_relax_domain_level = -1; 5829int sched_domain_level_max; 5830 5831static int __init setup_relax_domain_level(char *str) 5832{ 5833 if (kstrtoint(str, 0, &default_relax_domain_level)) 5834 pr_warn("Unable to set relax_domain_level\n"); 5835 5836 return 1; 5837} 5838__setup("relax_domain_level=", setup_relax_domain_level); 5839 5840static void set_domain_attribute(struct sched_domain *sd, 5841 struct sched_domain_attr *attr) 5842{ 5843 int request; 5844 5845 if (!attr || attr->relax_domain_level < 0) { 5846 if (default_relax_domain_level < 0) 5847 return; 5848 else 5849 request = default_relax_domain_level; 5850 } else 5851 request = attr->relax_domain_level; 5852 if (request < sd->level) { 5853 /* turn off idle balance on this domain */ 5854 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5855 } else { 5856 /* turn on idle balance on this domain */ 5857 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5858 } 5859} 5860 5861static void __sdt_free(const struct cpumask *cpu_map); 5862static int __sdt_alloc(const struct cpumask *cpu_map); 5863 5864static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5865 const struct cpumask *cpu_map) 5866{ 5867 switch (what) { 5868 case sa_rootdomain: 5869 if (!atomic_read(&d->rd->refcount)) 5870 free_rootdomain(&d->rd->rcu); /* fall through */ 5871 case sa_sd: 5872 free_percpu(d->sd); /* fall through */ 5873 case sa_sd_storage: 5874 __sdt_free(cpu_map); /* fall through */ 5875 case sa_none: 5876 break; 5877 } 5878} 5879 5880static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5881 const struct cpumask *cpu_map) 5882{ 5883 memset(d, 0, sizeof(*d)); 5884 5885 if (__sdt_alloc(cpu_map)) 5886 return sa_sd_storage; 5887 d->sd = alloc_percpu(struct sched_domain *); 5888 if (!d->sd) 5889 return sa_sd_storage; 5890 d->rd = alloc_rootdomain(); 5891 if (!d->rd) 5892 return sa_sd; 5893 return sa_rootdomain; 5894} 5895 5896/* 5897 * NULL the sd_data elements we've used to build the sched_domain and 5898 * sched_group structure so that the subsequent __free_domain_allocs() 5899 * will not free the data we're using. 5900 */ 5901static void claim_allocations(int cpu, struct sched_domain *sd) 5902{ 5903 struct sd_data *sdd = sd->private; 5904 5905 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5906 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5907 5908 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5909 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5910 5911 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5912 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5913} 5914 5915#ifdef CONFIG_SCHED_SMT 5916static const struct cpumask *cpu_smt_mask(int cpu) 5917{ 5918 return topology_thread_cpumask(cpu); 5919} 5920#endif 5921 5922/* 5923 * Topology list, bottom-up. 5924 */ 5925static struct sched_domain_topology_level default_topology[] = { 5926#ifdef CONFIG_SCHED_SMT 5927 { sd_init_SIBLING, cpu_smt_mask, }, 5928#endif 5929#ifdef CONFIG_SCHED_MC 5930 { sd_init_MC, cpu_coregroup_mask, }, 5931#endif 5932#ifdef CONFIG_SCHED_BOOK 5933 { sd_init_BOOK, cpu_book_mask, }, 5934#endif 5935 { sd_init_CPU, cpu_cpu_mask, }, 5936 { NULL, }, 5937}; 5938 5939static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5940 5941#define for_each_sd_topology(tl) \ 5942 for (tl = sched_domain_topology; tl->init; tl++) 5943 5944#ifdef CONFIG_NUMA 5945 5946static int sched_domains_numa_levels; 5947static int *sched_domains_numa_distance; 5948static struct cpumask ***sched_domains_numa_masks; 5949static int sched_domains_curr_level; 5950 5951static inline int sd_local_flags(int level) 5952{ 5953 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5954 return 0; 5955 5956 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5957} 5958 5959static struct sched_domain * 5960sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5961{ 5962 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5963 int level = tl->numa_level; 5964 int sd_weight = cpumask_weight( 5965 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5966 5967 *sd = (struct sched_domain){ 5968 .min_interval = sd_weight, 5969 .max_interval = 2*sd_weight, 5970 .busy_factor = 32, 5971 .imbalance_pct = 125, 5972 .cache_nice_tries = 2, 5973 .busy_idx = 3, 5974 .idle_idx = 2, 5975 .newidle_idx = 0, 5976 .wake_idx = 0, 5977 .forkexec_idx = 0, 5978 5979 .flags = 1*SD_LOAD_BALANCE 5980 | 1*SD_BALANCE_NEWIDLE 5981 | 0*SD_BALANCE_EXEC 5982 | 0*SD_BALANCE_FORK 5983 | 0*SD_BALANCE_WAKE 5984 | 0*SD_WAKE_AFFINE 5985 | 0*SD_SHARE_CPUPOWER 5986 | 0*SD_SHARE_PKG_RESOURCES 5987 | 1*SD_SERIALIZE 5988 | 0*SD_PREFER_SIBLING 5989 | 1*SD_NUMA 5990 | sd_local_flags(level) 5991 , 5992 .last_balance = jiffies, 5993 .balance_interval = sd_weight, 5994 }; 5995 SD_INIT_NAME(sd, NUMA); 5996 sd->private = &tl->data; 5997 5998 /* 5999 * Ugly hack to pass state to sd_numa_mask()... 6000 */ 6001 sched_domains_curr_level = tl->numa_level; 6002 6003 return sd; 6004} 6005 6006static const struct cpumask *sd_numa_mask(int cpu) 6007{ 6008 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6009} 6010 6011static void sched_numa_warn(const char *str) 6012{ 6013 static int done = false; 6014 int i,j; 6015 6016 if (done) 6017 return; 6018 6019 done = true; 6020 6021 printk(KERN_WARNING "ERROR: %s\n\n", str); 6022 6023 for (i = 0; i < nr_node_ids; i++) { 6024 printk(KERN_WARNING " "); 6025 for (j = 0; j < nr_node_ids; j++) 6026 printk(KERN_CONT "%02d ", node_distance(i,j)); 6027 printk(KERN_CONT "\n"); 6028 } 6029 printk(KERN_WARNING "\n"); 6030} 6031 6032static bool find_numa_distance(int distance) 6033{ 6034 int i; 6035 6036 if (distance == node_distance(0, 0)) 6037 return true; 6038 6039 for (i = 0; i < sched_domains_numa_levels; i++) { 6040 if (sched_domains_numa_distance[i] == distance) 6041 return true; 6042 } 6043 6044 return false; 6045} 6046 6047static void sched_init_numa(void) 6048{ 6049 int next_distance, curr_distance = node_distance(0, 0); 6050 struct sched_domain_topology_level *tl; 6051 int level = 0; 6052 int i, j, k; 6053 6054 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6055 if (!sched_domains_numa_distance) 6056 return; 6057 6058 /* 6059 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6060 * unique distances in the node_distance() table. 6061 * 6062 * Assumes node_distance(0,j) includes all distances in 6063 * node_distance(i,j) in order to avoid cubic time. 6064 */ 6065 next_distance = curr_distance; 6066 for (i = 0; i < nr_node_ids; i++) { 6067 for (j = 0; j < nr_node_ids; j++) { 6068 for (k = 0; k < nr_node_ids; k++) { 6069 int distance = node_distance(i, k); 6070 6071 if (distance > curr_distance && 6072 (distance < next_distance || 6073 next_distance == curr_distance)) 6074 next_distance = distance; 6075 6076 /* 6077 * While not a strong assumption it would be nice to know 6078 * about cases where if node A is connected to B, B is not 6079 * equally connected to A. 6080 */ 6081 if (sched_debug() && node_distance(k, i) != distance) 6082 sched_numa_warn("Node-distance not symmetric"); 6083 6084 if (sched_debug() && i && !find_numa_distance(distance)) 6085 sched_numa_warn("Node-0 not representative"); 6086 } 6087 if (next_distance != curr_distance) { 6088 sched_domains_numa_distance[level++] = next_distance; 6089 sched_domains_numa_levels = level; 6090 curr_distance = next_distance; 6091 } else break; 6092 } 6093 6094 /* 6095 * In case of sched_debug() we verify the above assumption. 6096 */ 6097 if (!sched_debug()) 6098 break; 6099 } 6100 /* 6101 * 'level' contains the number of unique distances, excluding the 6102 * identity distance node_distance(i,i). 6103 * 6104 * The sched_domains_numa_distance[] array includes the actual distance 6105 * numbers. 6106 */ 6107 6108 /* 6109 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6110 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6111 * the array will contain less then 'level' members. This could be 6112 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6113 * in other functions. 6114 * 6115 * We reset it to 'level' at the end of this function. 6116 */ 6117 sched_domains_numa_levels = 0; 6118 6119 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6120 if (!sched_domains_numa_masks) 6121 return; 6122 6123 /* 6124 * Now for each level, construct a mask per node which contains all 6125 * cpus of nodes that are that many hops away from us. 6126 */ 6127 for (i = 0; i < level; i++) { 6128 sched_domains_numa_masks[i] = 6129 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6130 if (!sched_domains_numa_masks[i]) 6131 return; 6132 6133 for (j = 0; j < nr_node_ids; j++) { 6134 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6135 if (!mask) 6136 return; 6137 6138 sched_domains_numa_masks[i][j] = mask; 6139 6140 for (k = 0; k < nr_node_ids; k++) { 6141 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6142 continue; 6143 6144 cpumask_or(mask, mask, cpumask_of_node(k)); 6145 } 6146 } 6147 } 6148 6149 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6150 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6151 if (!tl) 6152 return; 6153 6154 /* 6155 * Copy the default topology bits.. 6156 */ 6157 for (i = 0; default_topology[i].init; i++) 6158 tl[i] = default_topology[i]; 6159 6160 /* 6161 * .. and append 'j' levels of NUMA goodness. 6162 */ 6163 for (j = 0; j < level; i++, j++) { 6164 tl[i] = (struct sched_domain_topology_level){ 6165 .init = sd_numa_init, 6166 .mask = sd_numa_mask, 6167 .flags = SDTL_OVERLAP, 6168 .numa_level = j, 6169 }; 6170 } 6171 6172 sched_domain_topology = tl; 6173 6174 sched_domains_numa_levels = level; 6175} 6176 6177static void sched_domains_numa_masks_set(int cpu) 6178{ 6179 int i, j; 6180 int node = cpu_to_node(cpu); 6181 6182 for (i = 0; i < sched_domains_numa_levels; i++) { 6183 for (j = 0; j < nr_node_ids; j++) { 6184 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6185 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6186 } 6187 } 6188} 6189 6190static void sched_domains_numa_masks_clear(int cpu) 6191{ 6192 int i, j; 6193 for (i = 0; i < sched_domains_numa_levels; i++) { 6194 for (j = 0; j < nr_node_ids; j++) 6195 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6196 } 6197} 6198 6199/* 6200 * Update sched_domains_numa_masks[level][node] array when new cpus 6201 * are onlined. 6202 */ 6203static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6204 unsigned long action, 6205 void *hcpu) 6206{ 6207 int cpu = (long)hcpu; 6208 6209 switch (action & ~CPU_TASKS_FROZEN) { 6210 case CPU_ONLINE: 6211 sched_domains_numa_masks_set(cpu); 6212 break; 6213 6214 case CPU_DEAD: 6215 sched_domains_numa_masks_clear(cpu); 6216 break; 6217 6218 default: 6219 return NOTIFY_DONE; 6220 } 6221 6222 return NOTIFY_OK; 6223} 6224#else 6225static inline void sched_init_numa(void) 6226{ 6227} 6228 6229static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6230 unsigned long action, 6231 void *hcpu) 6232{ 6233 return 0; 6234} 6235#endif /* CONFIG_NUMA */ 6236 6237static int __sdt_alloc(const struct cpumask *cpu_map) 6238{ 6239 struct sched_domain_topology_level *tl; 6240 int j; 6241 6242 for_each_sd_topology(tl) { 6243 struct sd_data *sdd = &tl->data; 6244 6245 sdd->sd = alloc_percpu(struct sched_domain *); 6246 if (!sdd->sd) 6247 return -ENOMEM; 6248 6249 sdd->sg = alloc_percpu(struct sched_group *); 6250 if (!sdd->sg) 6251 return -ENOMEM; 6252 6253 sdd->sgp = alloc_percpu(struct sched_group_power *); 6254 if (!sdd->sgp) 6255 return -ENOMEM; 6256 6257 for_each_cpu(j, cpu_map) { 6258 struct sched_domain *sd; 6259 struct sched_group *sg; 6260 struct sched_group_power *sgp; 6261 6262 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6263 GFP_KERNEL, cpu_to_node(j)); 6264 if (!sd) 6265 return -ENOMEM; 6266 6267 *per_cpu_ptr(sdd->sd, j) = sd; 6268 6269 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6270 GFP_KERNEL, cpu_to_node(j)); 6271 if (!sg) 6272 return -ENOMEM; 6273 6274 sg->next = sg; 6275 6276 *per_cpu_ptr(sdd->sg, j) = sg; 6277 6278 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6279 GFP_KERNEL, cpu_to_node(j)); 6280 if (!sgp) 6281 return -ENOMEM; 6282 6283 *per_cpu_ptr(sdd->sgp, j) = sgp; 6284 } 6285 } 6286 6287 return 0; 6288} 6289 6290static void __sdt_free(const struct cpumask *cpu_map) 6291{ 6292 struct sched_domain_topology_level *tl; 6293 int j; 6294 6295 for_each_sd_topology(tl) { 6296 struct sd_data *sdd = &tl->data; 6297 6298 for_each_cpu(j, cpu_map) { 6299 struct sched_domain *sd; 6300 6301 if (sdd->sd) { 6302 sd = *per_cpu_ptr(sdd->sd, j); 6303 if (sd && (sd->flags & SD_OVERLAP)) 6304 free_sched_groups(sd->groups, 0); 6305 kfree(*per_cpu_ptr(sdd->sd, j)); 6306 } 6307 6308 if (sdd->sg) 6309 kfree(*per_cpu_ptr(sdd->sg, j)); 6310 if (sdd->sgp) 6311 kfree(*per_cpu_ptr(sdd->sgp, j)); 6312 } 6313 free_percpu(sdd->sd); 6314 sdd->sd = NULL; 6315 free_percpu(sdd->sg); 6316 sdd->sg = NULL; 6317 free_percpu(sdd->sgp); 6318 sdd->sgp = NULL; 6319 } 6320} 6321 6322struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6323 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6324 struct sched_domain *child, int cpu) 6325{ 6326 struct sched_domain *sd = tl->init(tl, cpu); 6327 if (!sd) 6328 return child; 6329 6330 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6331 if (child) { 6332 sd->level = child->level + 1; 6333 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6334 child->parent = sd; 6335 sd->child = child; 6336 } 6337 set_domain_attribute(sd, attr); 6338 6339 return sd; 6340} 6341 6342/* 6343 * Build sched domains for a given set of cpus and attach the sched domains 6344 * to the individual cpus 6345 */ 6346static int build_sched_domains(const struct cpumask *cpu_map, 6347 struct sched_domain_attr *attr) 6348{ 6349 enum s_alloc alloc_state; 6350 struct sched_domain *sd; 6351 struct s_data d; 6352 int i, ret = -ENOMEM; 6353 6354 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6355 if (alloc_state != sa_rootdomain) 6356 goto error; 6357 6358 /* Set up domains for cpus specified by the cpu_map. */ 6359 for_each_cpu(i, cpu_map) { 6360 struct sched_domain_topology_level *tl; 6361 6362 sd = NULL; 6363 for_each_sd_topology(tl) { 6364 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6365 if (tl == sched_domain_topology) 6366 *per_cpu_ptr(d.sd, i) = sd; 6367 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6368 sd->flags |= SD_OVERLAP; 6369 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6370 break; 6371 } 6372 } 6373 6374 /* Build the groups for the domains */ 6375 for_each_cpu(i, cpu_map) { 6376 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6377 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6378 if (sd->flags & SD_OVERLAP) { 6379 if (build_overlap_sched_groups(sd, i)) 6380 goto error; 6381 } else { 6382 if (build_sched_groups(sd, i)) 6383 goto error; 6384 } 6385 } 6386 } 6387 6388 /* Calculate CPU power for physical packages and nodes */ 6389 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6390 if (!cpumask_test_cpu(i, cpu_map)) 6391 continue; 6392 6393 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6394 claim_allocations(i, sd); 6395 init_sched_groups_power(i, sd); 6396 } 6397 } 6398 6399 /* Attach the domains */ 6400 rcu_read_lock(); 6401 for_each_cpu(i, cpu_map) { 6402 sd = *per_cpu_ptr(d.sd, i); 6403 cpu_attach_domain(sd, d.rd, i); 6404 } 6405 rcu_read_unlock(); 6406 6407 ret = 0; 6408error: 6409 __free_domain_allocs(&d, alloc_state, cpu_map); 6410 return ret; 6411} 6412 6413static cpumask_var_t *doms_cur; /* current sched domains */ 6414static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6415static struct sched_domain_attr *dattr_cur; 6416 /* attribues of custom domains in 'doms_cur' */ 6417 6418/* 6419 * Special case: If a kmalloc of a doms_cur partition (array of 6420 * cpumask) fails, then fallback to a single sched domain, 6421 * as determined by the single cpumask fallback_doms. 6422 */ 6423static cpumask_var_t fallback_doms; 6424 6425/* 6426 * arch_update_cpu_topology lets virtualized architectures update the 6427 * cpu core maps. It is supposed to return 1 if the topology changed 6428 * or 0 if it stayed the same. 6429 */ 6430int __attribute__((weak)) arch_update_cpu_topology(void) 6431{ 6432 return 0; 6433} 6434 6435cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6436{ 6437 int i; 6438 cpumask_var_t *doms; 6439 6440 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6441 if (!doms) 6442 return NULL; 6443 for (i = 0; i < ndoms; i++) { 6444 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6445 free_sched_domains(doms, i); 6446 return NULL; 6447 } 6448 } 6449 return doms; 6450} 6451 6452void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6453{ 6454 unsigned int i; 6455 for (i = 0; i < ndoms; i++) 6456 free_cpumask_var(doms[i]); 6457 kfree(doms); 6458} 6459 6460/* 6461 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6462 * For now this just excludes isolated cpus, but could be used to 6463 * exclude other special cases in the future. 6464 */ 6465static int init_sched_domains(const struct cpumask *cpu_map) 6466{ 6467 int err; 6468 6469 arch_update_cpu_topology(); 6470 ndoms_cur = 1; 6471 doms_cur = alloc_sched_domains(ndoms_cur); 6472 if (!doms_cur) 6473 doms_cur = &fallback_doms; 6474 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6475 err = build_sched_domains(doms_cur[0], NULL); 6476 register_sched_domain_sysctl(); 6477 6478 return err; 6479} 6480 6481/* 6482 * Detach sched domains from a group of cpus specified in cpu_map 6483 * These cpus will now be attached to the NULL domain 6484 */ 6485static void detach_destroy_domains(const struct cpumask *cpu_map) 6486{ 6487 int i; 6488 6489 rcu_read_lock(); 6490 for_each_cpu(i, cpu_map) 6491 cpu_attach_domain(NULL, &def_root_domain, i); 6492 rcu_read_unlock(); 6493} 6494 6495/* handle null as "default" */ 6496static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6497 struct sched_domain_attr *new, int idx_new) 6498{ 6499 struct sched_domain_attr tmp; 6500 6501 /* fast path */ 6502 if (!new && !cur) 6503 return 1; 6504 6505 tmp = SD_ATTR_INIT; 6506 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6507 new ? (new + idx_new) : &tmp, 6508 sizeof(struct sched_domain_attr)); 6509} 6510 6511/* 6512 * Partition sched domains as specified by the 'ndoms_new' 6513 * cpumasks in the array doms_new[] of cpumasks. This compares 6514 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6515 * It destroys each deleted domain and builds each new domain. 6516 * 6517 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6518 * The masks don't intersect (don't overlap.) We should setup one 6519 * sched domain for each mask. CPUs not in any of the cpumasks will 6520 * not be load balanced. If the same cpumask appears both in the 6521 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6522 * it as it is. 6523 * 6524 * The passed in 'doms_new' should be allocated using 6525 * alloc_sched_domains. This routine takes ownership of it and will 6526 * free_sched_domains it when done with it. If the caller failed the 6527 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6528 * and partition_sched_domains() will fallback to the single partition 6529 * 'fallback_doms', it also forces the domains to be rebuilt. 6530 * 6531 * If doms_new == NULL it will be replaced with cpu_online_mask. 6532 * ndoms_new == 0 is a special case for destroying existing domains, 6533 * and it will not create the default domain. 6534 * 6535 * Call with hotplug lock held 6536 */ 6537void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6538 struct sched_domain_attr *dattr_new) 6539{ 6540 int i, j, n; 6541 int new_topology; 6542 6543 mutex_lock(&sched_domains_mutex); 6544 6545 /* always unregister in case we don't destroy any domains */ 6546 unregister_sched_domain_sysctl(); 6547 6548 /* Let architecture update cpu core mappings. */ 6549 new_topology = arch_update_cpu_topology(); 6550 6551 n = doms_new ? ndoms_new : 0; 6552 6553 /* Destroy deleted domains */ 6554 for (i = 0; i < ndoms_cur; i++) { 6555 for (j = 0; j < n && !new_topology; j++) { 6556 if (cpumask_equal(doms_cur[i], doms_new[j]) 6557 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6558 goto match1; 6559 } 6560 /* no match - a current sched domain not in new doms_new[] */ 6561 detach_destroy_domains(doms_cur[i]); 6562match1: 6563 ; 6564 } 6565 6566 n = ndoms_cur; 6567 if (doms_new == NULL) { 6568 n = 0; 6569 doms_new = &fallback_doms; 6570 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6571 WARN_ON_ONCE(dattr_new); 6572 } 6573 6574 /* Build new domains */ 6575 for (i = 0; i < ndoms_new; i++) { 6576 for (j = 0; j < n && !new_topology; j++) { 6577 if (cpumask_equal(doms_new[i], doms_cur[j]) 6578 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6579 goto match2; 6580 } 6581 /* no match - add a new doms_new */ 6582 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6583match2: 6584 ; 6585 } 6586 6587 /* Remember the new sched domains */ 6588 if (doms_cur != &fallback_doms) 6589 free_sched_domains(doms_cur, ndoms_cur); 6590 kfree(dattr_cur); /* kfree(NULL) is safe */ 6591 doms_cur = doms_new; 6592 dattr_cur = dattr_new; 6593 ndoms_cur = ndoms_new; 6594 6595 register_sched_domain_sysctl(); 6596 6597 mutex_unlock(&sched_domains_mutex); 6598} 6599 6600static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6601 6602/* 6603 * Update cpusets according to cpu_active mask. If cpusets are 6604 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6605 * around partition_sched_domains(). 6606 * 6607 * If we come here as part of a suspend/resume, don't touch cpusets because we 6608 * want to restore it back to its original state upon resume anyway. 6609 */ 6610static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6611 void *hcpu) 6612{ 6613 switch (action) { 6614 case CPU_ONLINE_FROZEN: 6615 case CPU_DOWN_FAILED_FROZEN: 6616 6617 /* 6618 * num_cpus_frozen tracks how many CPUs are involved in suspend 6619 * resume sequence. As long as this is not the last online 6620 * operation in the resume sequence, just build a single sched 6621 * domain, ignoring cpusets. 6622 */ 6623 num_cpus_frozen--; 6624 if (likely(num_cpus_frozen)) { 6625 partition_sched_domains(1, NULL, NULL); 6626 break; 6627 } 6628 6629 /* 6630 * This is the last CPU online operation. So fall through and 6631 * restore the original sched domains by considering the 6632 * cpuset configurations. 6633 */ 6634 6635 case CPU_ONLINE: 6636 case CPU_DOWN_FAILED: 6637 cpuset_update_active_cpus(true); 6638 break; 6639 default: 6640 return NOTIFY_DONE; 6641 } 6642 return NOTIFY_OK; 6643} 6644 6645static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6646 void *hcpu) 6647{ 6648 switch (action) { 6649 case CPU_DOWN_PREPARE: 6650 cpuset_update_active_cpus(false); 6651 break; 6652 case CPU_DOWN_PREPARE_FROZEN: 6653 num_cpus_frozen++; 6654 partition_sched_domains(1, NULL, NULL); 6655 break; 6656 default: 6657 return NOTIFY_DONE; 6658 } 6659 return NOTIFY_OK; 6660} 6661 6662void __init sched_init_smp(void) 6663{ 6664 cpumask_var_t non_isolated_cpus; 6665 6666 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6667 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6668 6669 sched_init_numa(); 6670 6671 /* 6672 * There's no userspace yet to cause hotplug operations; hence all the 6673 * cpu masks are stable and all blatant races in the below code cannot 6674 * happen. 6675 */ 6676 mutex_lock(&sched_domains_mutex); 6677 init_sched_domains(cpu_active_mask); 6678 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6679 if (cpumask_empty(non_isolated_cpus)) 6680 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6681 mutex_unlock(&sched_domains_mutex); 6682 6683 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6684 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6685 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6686 6687 init_hrtick(); 6688 6689 /* Move init over to a non-isolated CPU */ 6690 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6691 BUG(); 6692 sched_init_granularity(); 6693 free_cpumask_var(non_isolated_cpus); 6694 6695 init_sched_rt_class(); 6696 init_sched_dl_class(); 6697} 6698#else 6699void __init sched_init_smp(void) 6700{ 6701 sched_init_granularity(); 6702} 6703#endif /* CONFIG_SMP */ 6704 6705const_debug unsigned int sysctl_timer_migration = 1; 6706 6707int in_sched_functions(unsigned long addr) 6708{ 6709 return in_lock_functions(addr) || 6710 (addr >= (unsigned long)__sched_text_start 6711 && addr < (unsigned long)__sched_text_end); 6712} 6713 6714#ifdef CONFIG_CGROUP_SCHED 6715/* 6716 * Default task group. 6717 * Every task in system belongs to this group at bootup. 6718 */ 6719struct task_group root_task_group; 6720LIST_HEAD(task_groups); 6721#endif 6722 6723DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6724 6725void __init sched_init(void) 6726{ 6727 int i, j; 6728 unsigned long alloc_size = 0, ptr; 6729 6730#ifdef CONFIG_FAIR_GROUP_SCHED 6731 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6732#endif 6733#ifdef CONFIG_RT_GROUP_SCHED 6734 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6735#endif 6736#ifdef CONFIG_CPUMASK_OFFSTACK 6737 alloc_size += num_possible_cpus() * cpumask_size(); 6738#endif 6739 if (alloc_size) { 6740 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6741 6742#ifdef CONFIG_FAIR_GROUP_SCHED 6743 root_task_group.se = (struct sched_entity **)ptr; 6744 ptr += nr_cpu_ids * sizeof(void **); 6745 6746 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6747 ptr += nr_cpu_ids * sizeof(void **); 6748 6749#endif /* CONFIG_FAIR_GROUP_SCHED */ 6750#ifdef CONFIG_RT_GROUP_SCHED 6751 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6752 ptr += nr_cpu_ids * sizeof(void **); 6753 6754 root_task_group.rt_rq = (struct rt_rq **)ptr; 6755 ptr += nr_cpu_ids * sizeof(void **); 6756 6757#endif /* CONFIG_RT_GROUP_SCHED */ 6758#ifdef CONFIG_CPUMASK_OFFSTACK 6759 for_each_possible_cpu(i) { 6760 per_cpu(load_balance_mask, i) = (void *)ptr; 6761 ptr += cpumask_size(); 6762 } 6763#endif /* CONFIG_CPUMASK_OFFSTACK */ 6764 } 6765 6766 init_rt_bandwidth(&def_rt_bandwidth, 6767 global_rt_period(), global_rt_runtime()); 6768 init_dl_bandwidth(&def_dl_bandwidth, 6769 global_rt_period(), global_rt_runtime()); 6770 6771#ifdef CONFIG_SMP 6772 init_defrootdomain(); 6773#endif 6774 6775#ifdef CONFIG_RT_GROUP_SCHED 6776 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6777 global_rt_period(), global_rt_runtime()); 6778#endif /* CONFIG_RT_GROUP_SCHED */ 6779 6780#ifdef CONFIG_CGROUP_SCHED 6781 list_add(&root_task_group.list, &task_groups); 6782 INIT_LIST_HEAD(&root_task_group.children); 6783 INIT_LIST_HEAD(&root_task_group.siblings); 6784 autogroup_init(&init_task); 6785 6786#endif /* CONFIG_CGROUP_SCHED */ 6787 6788 for_each_possible_cpu(i) { 6789 struct rq *rq; 6790 6791 rq = cpu_rq(i); 6792 raw_spin_lock_init(&rq->lock); 6793 rq->nr_running = 0; 6794 rq->calc_load_active = 0; 6795 rq->calc_load_update = jiffies + LOAD_FREQ; 6796 init_cfs_rq(&rq->cfs); 6797 init_rt_rq(&rq->rt, rq); 6798 init_dl_rq(&rq->dl, rq); 6799#ifdef CONFIG_FAIR_GROUP_SCHED 6800 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6801 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6802 /* 6803 * How much cpu bandwidth does root_task_group get? 6804 * 6805 * In case of task-groups formed thr' the cgroup filesystem, it 6806 * gets 100% of the cpu resources in the system. This overall 6807 * system cpu resource is divided among the tasks of 6808 * root_task_group and its child task-groups in a fair manner, 6809 * based on each entity's (task or task-group's) weight 6810 * (se->load.weight). 6811 * 6812 * In other words, if root_task_group has 10 tasks of weight 6813 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6814 * then A0's share of the cpu resource is: 6815 * 6816 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6817 * 6818 * We achieve this by letting root_task_group's tasks sit 6819 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6820 */ 6821 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6822 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6823#endif /* CONFIG_FAIR_GROUP_SCHED */ 6824 6825 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6826#ifdef CONFIG_RT_GROUP_SCHED 6827 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6828 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6829#endif 6830 6831 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6832 rq->cpu_load[j] = 0; 6833 6834 rq->last_load_update_tick = jiffies; 6835 6836#ifdef CONFIG_SMP 6837 rq->sd = NULL; 6838 rq->rd = NULL; 6839 rq->cpu_power = SCHED_POWER_SCALE; 6840 rq->post_schedule = 0; 6841 rq->active_balance = 0; 6842 rq->next_balance = jiffies; 6843 rq->push_cpu = 0; 6844 rq->cpu = i; 6845 rq->online = 0; 6846 rq->idle_stamp = 0; 6847 rq->avg_idle = 2*sysctl_sched_migration_cost; 6848 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6849 6850 INIT_LIST_HEAD(&rq->cfs_tasks); 6851 6852 rq_attach_root(rq, &def_root_domain); 6853#ifdef CONFIG_NO_HZ_COMMON 6854 rq->nohz_flags = 0; 6855#endif 6856#ifdef CONFIG_NO_HZ_FULL 6857 rq->last_sched_tick = 0; 6858#endif 6859#endif 6860 init_rq_hrtick(rq); 6861 atomic_set(&rq->nr_iowait, 0); 6862 } 6863 6864 set_load_weight(&init_task); 6865 6866#ifdef CONFIG_PREEMPT_NOTIFIERS 6867 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6868#endif 6869 6870 /* 6871 * The boot idle thread does lazy MMU switching as well: 6872 */ 6873 atomic_inc(&init_mm.mm_count); 6874 enter_lazy_tlb(&init_mm, current); 6875 6876 /* 6877 * Make us the idle thread. Technically, schedule() should not be 6878 * called from this thread, however somewhere below it might be, 6879 * but because we are the idle thread, we just pick up running again 6880 * when this runqueue becomes "idle". 6881 */ 6882 init_idle(current, smp_processor_id()); 6883 6884 calc_load_update = jiffies + LOAD_FREQ; 6885 6886 /* 6887 * During early bootup we pretend to be a normal task: 6888 */ 6889 current->sched_class = &fair_sched_class; 6890 6891#ifdef CONFIG_SMP 6892 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6893 /* May be allocated at isolcpus cmdline parse time */ 6894 if (cpu_isolated_map == NULL) 6895 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6896 idle_thread_set_boot_cpu(); 6897#endif 6898 init_sched_fair_class(); 6899 6900 scheduler_running = 1; 6901} 6902 6903#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6904static inline int preempt_count_equals(int preempt_offset) 6905{ 6906 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6907 6908 return (nested == preempt_offset); 6909} 6910 6911void __might_sleep(const char *file, int line, int preempt_offset) 6912{ 6913 static unsigned long prev_jiffy; /* ratelimiting */ 6914 6915 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6916 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6917 system_state != SYSTEM_RUNNING || oops_in_progress) 6918 return; 6919 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6920 return; 6921 prev_jiffy = jiffies; 6922 6923 printk(KERN_ERR 6924 "BUG: sleeping function called from invalid context at %s:%d\n", 6925 file, line); 6926 printk(KERN_ERR 6927 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6928 in_atomic(), irqs_disabled(), 6929 current->pid, current->comm); 6930 6931 debug_show_held_locks(current); 6932 if (irqs_disabled()) 6933 print_irqtrace_events(current); 6934 dump_stack(); 6935} 6936EXPORT_SYMBOL(__might_sleep); 6937#endif 6938 6939#ifdef CONFIG_MAGIC_SYSRQ 6940static void normalize_task(struct rq *rq, struct task_struct *p) 6941{ 6942 const struct sched_class *prev_class = p->sched_class; 6943 struct sched_attr attr = { 6944 .sched_policy = SCHED_NORMAL, 6945 }; 6946 int old_prio = p->prio; 6947 int on_rq; 6948 6949 on_rq = p->on_rq; 6950 if (on_rq) 6951 dequeue_task(rq, p, 0); 6952 __setscheduler(rq, p, &attr); 6953 if (on_rq) { 6954 enqueue_task(rq, p, 0); 6955 resched_task(rq->curr); 6956 } 6957 6958 check_class_changed(rq, p, prev_class, old_prio); 6959} 6960 6961void normalize_rt_tasks(void) 6962{ 6963 struct task_struct *g, *p; 6964 unsigned long flags; 6965 struct rq *rq; 6966 6967 read_lock_irqsave(&tasklist_lock, flags); 6968 do_each_thread(g, p) { 6969 /* 6970 * Only normalize user tasks: 6971 */ 6972 if (!p->mm) 6973 continue; 6974 6975 p->se.exec_start = 0; 6976#ifdef CONFIG_SCHEDSTATS 6977 p->se.statistics.wait_start = 0; 6978 p->se.statistics.sleep_start = 0; 6979 p->se.statistics.block_start = 0; 6980#endif 6981 6982 if (!dl_task(p) && !rt_task(p)) { 6983 /* 6984 * Renice negative nice level userspace 6985 * tasks back to 0: 6986 */ 6987 if (TASK_NICE(p) < 0 && p->mm) 6988 set_user_nice(p, 0); 6989 continue; 6990 } 6991 6992 raw_spin_lock(&p->pi_lock); 6993 rq = __task_rq_lock(p); 6994 6995 normalize_task(rq, p); 6996 6997 __task_rq_unlock(rq); 6998 raw_spin_unlock(&p->pi_lock); 6999 } while_each_thread(g, p); 7000 7001 read_unlock_irqrestore(&tasklist_lock, flags); 7002} 7003 7004#endif /* CONFIG_MAGIC_SYSRQ */ 7005 7006#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7007/* 7008 * These functions are only useful for the IA64 MCA handling, or kdb. 7009 * 7010 * They can only be called when the whole system has been 7011 * stopped - every CPU needs to be quiescent, and no scheduling 7012 * activity can take place. Using them for anything else would 7013 * be a serious bug, and as a result, they aren't even visible 7014 * under any other configuration. 7015 */ 7016 7017/** 7018 * curr_task - return the current task for a given cpu. 7019 * @cpu: the processor in question. 7020 * 7021 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7022 * 7023 * Return: The current task for @cpu. 7024 */ 7025struct task_struct *curr_task(int cpu) 7026{ 7027 return cpu_curr(cpu); 7028} 7029 7030#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7031 7032#ifdef CONFIG_IA64 7033/** 7034 * set_curr_task - set the current task for a given cpu. 7035 * @cpu: the processor in question. 7036 * @p: the task pointer to set. 7037 * 7038 * Description: This function must only be used when non-maskable interrupts 7039 * are serviced on a separate stack. It allows the architecture to switch the 7040 * notion of the current task on a cpu in a non-blocking manner. This function 7041 * must be called with all CPU's synchronized, and interrupts disabled, the 7042 * and caller must save the original value of the current task (see 7043 * curr_task() above) and restore that value before reenabling interrupts and 7044 * re-starting the system. 7045 * 7046 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7047 */ 7048void set_curr_task(int cpu, struct task_struct *p) 7049{ 7050 cpu_curr(cpu) = p; 7051} 7052 7053#endif 7054 7055#ifdef CONFIG_CGROUP_SCHED 7056/* task_group_lock serializes the addition/removal of task groups */ 7057static DEFINE_SPINLOCK(task_group_lock); 7058 7059static void free_sched_group(struct task_group *tg) 7060{ 7061 free_fair_sched_group(tg); 7062 free_rt_sched_group(tg); 7063 autogroup_free(tg); 7064 kfree(tg); 7065} 7066 7067/* allocate runqueue etc for a new task group */ 7068struct task_group *sched_create_group(struct task_group *parent) 7069{ 7070 struct task_group *tg; 7071 7072 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7073 if (!tg) 7074 return ERR_PTR(-ENOMEM); 7075 7076 if (!alloc_fair_sched_group(tg, parent)) 7077 goto err; 7078 7079 if (!alloc_rt_sched_group(tg, parent)) 7080 goto err; 7081 7082 return tg; 7083 7084err: 7085 free_sched_group(tg); 7086 return ERR_PTR(-ENOMEM); 7087} 7088 7089void sched_online_group(struct task_group *tg, struct task_group *parent) 7090{ 7091 unsigned long flags; 7092 7093 spin_lock_irqsave(&task_group_lock, flags); 7094 list_add_rcu(&tg->list, &task_groups); 7095 7096 WARN_ON(!parent); /* root should already exist */ 7097 7098 tg->parent = parent; 7099 INIT_LIST_HEAD(&tg->children); 7100 list_add_rcu(&tg->siblings, &parent->children); 7101 spin_unlock_irqrestore(&task_group_lock, flags); 7102} 7103 7104/* rcu callback to free various structures associated with a task group */ 7105static void free_sched_group_rcu(struct rcu_head *rhp) 7106{ 7107 /* now it should be safe to free those cfs_rqs */ 7108 free_sched_group(container_of(rhp, struct task_group, rcu)); 7109} 7110 7111/* Destroy runqueue etc associated with a task group */ 7112void sched_destroy_group(struct task_group *tg) 7113{ 7114 /* wait for possible concurrent references to cfs_rqs complete */ 7115 call_rcu(&tg->rcu, free_sched_group_rcu); 7116} 7117 7118void sched_offline_group(struct task_group *tg) 7119{ 7120 unsigned long flags; 7121 int i; 7122 7123 /* end participation in shares distribution */ 7124 for_each_possible_cpu(i) 7125 unregister_fair_sched_group(tg, i); 7126 7127 spin_lock_irqsave(&task_group_lock, flags); 7128 list_del_rcu(&tg->list); 7129 list_del_rcu(&tg->siblings); 7130 spin_unlock_irqrestore(&task_group_lock, flags); 7131} 7132 7133/* change task's runqueue when it moves between groups. 7134 * The caller of this function should have put the task in its new group 7135 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7136 * reflect its new group. 7137 */ 7138void sched_move_task(struct task_struct *tsk) 7139{ 7140 struct task_group *tg; 7141 int on_rq, running; 7142 unsigned long flags; 7143 struct rq *rq; 7144 7145 rq = task_rq_lock(tsk, &flags); 7146 7147 running = task_current(rq, tsk); 7148 on_rq = tsk->on_rq; 7149 7150 if (on_rq) 7151 dequeue_task(rq, tsk, 0); 7152 if (unlikely(running)) 7153 tsk->sched_class->put_prev_task(rq, tsk); 7154 7155 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 7156 lockdep_is_held(&tsk->sighand->siglock)), 7157 struct task_group, css); 7158 tg = autogroup_task_group(tsk, tg); 7159 tsk->sched_task_group = tg; 7160 7161#ifdef CONFIG_FAIR_GROUP_SCHED 7162 if (tsk->sched_class->task_move_group) 7163 tsk->sched_class->task_move_group(tsk, on_rq); 7164 else 7165#endif 7166 set_task_rq(tsk, task_cpu(tsk)); 7167 7168 if (unlikely(running)) 7169 tsk->sched_class->set_curr_task(rq); 7170 if (on_rq) 7171 enqueue_task(rq, tsk, 0); 7172 7173 task_rq_unlock(rq, tsk, &flags); 7174} 7175#endif /* CONFIG_CGROUP_SCHED */ 7176 7177#ifdef CONFIG_RT_GROUP_SCHED 7178/* 7179 * Ensure that the real time constraints are schedulable. 7180 */ 7181static DEFINE_MUTEX(rt_constraints_mutex); 7182 7183/* Must be called with tasklist_lock held */ 7184static inline int tg_has_rt_tasks(struct task_group *tg) 7185{ 7186 struct task_struct *g, *p; 7187 7188 do_each_thread(g, p) { 7189 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7190 return 1; 7191 } while_each_thread(g, p); 7192 7193 return 0; 7194} 7195 7196struct rt_schedulable_data { 7197 struct task_group *tg; 7198 u64 rt_period; 7199 u64 rt_runtime; 7200}; 7201 7202static int tg_rt_schedulable(struct task_group *tg, void *data) 7203{ 7204 struct rt_schedulable_data *d = data; 7205 struct task_group *child; 7206 unsigned long total, sum = 0; 7207 u64 period, runtime; 7208 7209 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7210 runtime = tg->rt_bandwidth.rt_runtime; 7211 7212 if (tg == d->tg) { 7213 period = d->rt_period; 7214 runtime = d->rt_runtime; 7215 } 7216 7217 /* 7218 * Cannot have more runtime than the period. 7219 */ 7220 if (runtime > period && runtime != RUNTIME_INF) 7221 return -EINVAL; 7222 7223 /* 7224 * Ensure we don't starve existing RT tasks. 7225 */ 7226 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7227 return -EBUSY; 7228 7229 total = to_ratio(period, runtime); 7230 7231 /* 7232 * Nobody can have more than the global setting allows. 7233 */ 7234 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7235 return -EINVAL; 7236 7237 /* 7238 * The sum of our children's runtime should not exceed our own. 7239 */ 7240 list_for_each_entry_rcu(child, &tg->children, siblings) { 7241 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7242 runtime = child->rt_bandwidth.rt_runtime; 7243 7244 if (child == d->tg) { 7245 period = d->rt_period; 7246 runtime = d->rt_runtime; 7247 } 7248 7249 sum += to_ratio(period, runtime); 7250 } 7251 7252 if (sum > total) 7253 return -EINVAL; 7254 7255 return 0; 7256} 7257 7258static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7259{ 7260 int ret; 7261 7262 struct rt_schedulable_data data = { 7263 .tg = tg, 7264 .rt_period = period, 7265 .rt_runtime = runtime, 7266 }; 7267 7268 rcu_read_lock(); 7269 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7270 rcu_read_unlock(); 7271 7272 return ret; 7273} 7274 7275static int tg_set_rt_bandwidth(struct task_group *tg, 7276 u64 rt_period, u64 rt_runtime) 7277{ 7278 int i, err = 0; 7279 7280 mutex_lock(&rt_constraints_mutex); 7281 read_lock(&tasklist_lock); 7282 err = __rt_schedulable(tg, rt_period, rt_runtime); 7283 if (err) 7284 goto unlock; 7285 7286 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7287 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7288 tg->rt_bandwidth.rt_runtime = rt_runtime; 7289 7290 for_each_possible_cpu(i) { 7291 struct rt_rq *rt_rq = tg->rt_rq[i]; 7292 7293 raw_spin_lock(&rt_rq->rt_runtime_lock); 7294 rt_rq->rt_runtime = rt_runtime; 7295 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7296 } 7297 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7298unlock: 7299 read_unlock(&tasklist_lock); 7300 mutex_unlock(&rt_constraints_mutex); 7301 7302 return err; 7303} 7304 7305static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7306{ 7307 u64 rt_runtime, rt_period; 7308 7309 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7310 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7311 if (rt_runtime_us < 0) 7312 rt_runtime = RUNTIME_INF; 7313 7314 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7315} 7316 7317static long sched_group_rt_runtime(struct task_group *tg) 7318{ 7319 u64 rt_runtime_us; 7320 7321 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7322 return -1; 7323 7324 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7325 do_div(rt_runtime_us, NSEC_PER_USEC); 7326 return rt_runtime_us; 7327} 7328 7329static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7330{ 7331 u64 rt_runtime, rt_period; 7332 7333 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7334 rt_runtime = tg->rt_bandwidth.rt_runtime; 7335 7336 if (rt_period == 0) 7337 return -EINVAL; 7338 7339 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7340} 7341 7342static long sched_group_rt_period(struct task_group *tg) 7343{ 7344 u64 rt_period_us; 7345 7346 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7347 do_div(rt_period_us, NSEC_PER_USEC); 7348 return rt_period_us; 7349} 7350#endif /* CONFIG_RT_GROUP_SCHED */ 7351 7352#ifdef CONFIG_RT_GROUP_SCHED 7353static int sched_rt_global_constraints(void) 7354{ 7355 int ret = 0; 7356 7357 mutex_lock(&rt_constraints_mutex); 7358 read_lock(&tasklist_lock); 7359 ret = __rt_schedulable(NULL, 0, 0); 7360 read_unlock(&tasklist_lock); 7361 mutex_unlock(&rt_constraints_mutex); 7362 7363 return ret; 7364} 7365 7366static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7367{ 7368 /* Don't accept realtime tasks when there is no way for them to run */ 7369 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7370 return 0; 7371 7372 return 1; 7373} 7374 7375#else /* !CONFIG_RT_GROUP_SCHED */ 7376static int sched_rt_global_constraints(void) 7377{ 7378 unsigned long flags; 7379 int i, ret = 0; 7380 7381 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7382 for_each_possible_cpu(i) { 7383 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7384 7385 raw_spin_lock(&rt_rq->rt_runtime_lock); 7386 rt_rq->rt_runtime = global_rt_runtime(); 7387 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7388 } 7389 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7390 7391 return ret; 7392} 7393#endif /* CONFIG_RT_GROUP_SCHED */ 7394 7395static int sched_dl_global_constraints(void) 7396{ 7397 u64 runtime = global_rt_runtime(); 7398 u64 period = global_rt_period(); 7399 u64 new_bw = to_ratio(period, runtime); 7400 int cpu, ret = 0; 7401 7402 /* 7403 * Here we want to check the bandwidth not being set to some 7404 * value smaller than the currently allocated bandwidth in 7405 * any of the root_domains. 7406 * 7407 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7408 * cycling on root_domains... Discussion on different/better 7409 * solutions is welcome! 7410 */ 7411 for_each_possible_cpu(cpu) { 7412 struct dl_bw *dl_b = dl_bw_of(cpu); 7413 7414 raw_spin_lock(&dl_b->lock); 7415 if (new_bw < dl_b->total_bw) 7416 ret = -EBUSY; 7417 raw_spin_unlock(&dl_b->lock); 7418 7419 if (ret) 7420 break; 7421 } 7422 7423 return ret; 7424} 7425 7426static void sched_dl_do_global(void) 7427{ 7428 u64 new_bw = -1; 7429 int cpu; 7430 7431 def_dl_bandwidth.dl_period = global_rt_period(); 7432 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7433 7434 if (global_rt_runtime() != RUNTIME_INF) 7435 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7436 7437 /* 7438 * FIXME: As above... 7439 */ 7440 for_each_possible_cpu(cpu) { 7441 struct dl_bw *dl_b = dl_bw_of(cpu); 7442 7443 raw_spin_lock(&dl_b->lock); 7444 dl_b->bw = new_bw; 7445 raw_spin_unlock(&dl_b->lock); 7446 } 7447} 7448 7449static int sched_rt_global_validate(void) 7450{ 7451 if (sysctl_sched_rt_period <= 0) 7452 return -EINVAL; 7453 7454 if (sysctl_sched_rt_runtime > sysctl_sched_rt_period) 7455 return -EINVAL; 7456 7457 return 0; 7458} 7459 7460static void sched_rt_do_global(void) 7461{ 7462 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7463 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7464} 7465 7466int sched_rt_handler(struct ctl_table *table, int write, 7467 void __user *buffer, size_t *lenp, 7468 loff_t *ppos) 7469{ 7470 int old_period, old_runtime; 7471 static DEFINE_MUTEX(mutex); 7472 int ret; 7473 7474 mutex_lock(&mutex); 7475 old_period = sysctl_sched_rt_period; 7476 old_runtime = sysctl_sched_rt_runtime; 7477 7478 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7479 7480 if (!ret && write) { 7481 ret = sched_rt_global_validate(); 7482 if (ret) 7483 goto undo; 7484 7485 ret = sched_rt_global_constraints(); 7486 if (ret) 7487 goto undo; 7488 7489 ret = sched_dl_global_constraints(); 7490 if (ret) 7491 goto undo; 7492 7493 sched_rt_do_global(); 7494 sched_dl_do_global(); 7495 } 7496 if (0) { 7497undo: 7498 sysctl_sched_rt_period = old_period; 7499 sysctl_sched_rt_runtime = old_runtime; 7500 } 7501 mutex_unlock(&mutex); 7502 7503 return ret; 7504} 7505 7506int sched_rr_handler(struct ctl_table *table, int write, 7507 void __user *buffer, size_t *lenp, 7508 loff_t *ppos) 7509{ 7510 int ret; 7511 static DEFINE_MUTEX(mutex); 7512 7513 mutex_lock(&mutex); 7514 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7515 /* make sure that internally we keep jiffies */ 7516 /* also, writing zero resets timeslice to default */ 7517 if (!ret && write) { 7518 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7519 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7520 } 7521 mutex_unlock(&mutex); 7522 return ret; 7523} 7524 7525#ifdef CONFIG_CGROUP_SCHED 7526 7527static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7528{ 7529 return css ? container_of(css, struct task_group, css) : NULL; 7530} 7531 7532static struct cgroup_subsys_state * 7533cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7534{ 7535 struct task_group *parent = css_tg(parent_css); 7536 struct task_group *tg; 7537 7538 if (!parent) { 7539 /* This is early initialization for the top cgroup */ 7540 return &root_task_group.css; 7541 } 7542 7543 tg = sched_create_group(parent); 7544 if (IS_ERR(tg)) 7545 return ERR_PTR(-ENOMEM); 7546 7547 return &tg->css; 7548} 7549 7550static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7551{ 7552 struct task_group *tg = css_tg(css); 7553 struct task_group *parent = css_tg(css_parent(css)); 7554 7555 if (parent) 7556 sched_online_group(tg, parent); 7557 return 0; 7558} 7559 7560static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7561{ 7562 struct task_group *tg = css_tg(css); 7563 7564 sched_destroy_group(tg); 7565} 7566 7567static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7568{ 7569 struct task_group *tg = css_tg(css); 7570 7571 sched_offline_group(tg); 7572} 7573 7574static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7575 struct cgroup_taskset *tset) 7576{ 7577 struct task_struct *task; 7578 7579 cgroup_taskset_for_each(task, css, tset) { 7580#ifdef CONFIG_RT_GROUP_SCHED 7581 if (!sched_rt_can_attach(css_tg(css), task)) 7582 return -EINVAL; 7583#else 7584 /* We don't support RT-tasks being in separate groups */ 7585 if (task->sched_class != &fair_sched_class) 7586 return -EINVAL; 7587#endif 7588 } 7589 return 0; 7590} 7591 7592static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7593 struct cgroup_taskset *tset) 7594{ 7595 struct task_struct *task; 7596 7597 cgroup_taskset_for_each(task, css, tset) 7598 sched_move_task(task); 7599} 7600 7601static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7602 struct cgroup_subsys_state *old_css, 7603 struct task_struct *task) 7604{ 7605 /* 7606 * cgroup_exit() is called in the copy_process() failure path. 7607 * Ignore this case since the task hasn't ran yet, this avoids 7608 * trying to poke a half freed task state from generic code. 7609 */ 7610 if (!(task->flags & PF_EXITING)) 7611 return; 7612 7613 sched_move_task(task); 7614} 7615 7616#ifdef CONFIG_FAIR_GROUP_SCHED 7617static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7618 struct cftype *cftype, u64 shareval) 7619{ 7620 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7621} 7622 7623static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7624 struct cftype *cft) 7625{ 7626 struct task_group *tg = css_tg(css); 7627 7628 return (u64) scale_load_down(tg->shares); 7629} 7630 7631#ifdef CONFIG_CFS_BANDWIDTH 7632static DEFINE_MUTEX(cfs_constraints_mutex); 7633 7634const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7635const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7636 7637static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7638 7639static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7640{ 7641 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7642 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7643 7644 if (tg == &root_task_group) 7645 return -EINVAL; 7646 7647 /* 7648 * Ensure we have at some amount of bandwidth every period. This is 7649 * to prevent reaching a state of large arrears when throttled via 7650 * entity_tick() resulting in prolonged exit starvation. 7651 */ 7652 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7653 return -EINVAL; 7654 7655 /* 7656 * Likewise, bound things on the otherside by preventing insane quota 7657 * periods. This also allows us to normalize in computing quota 7658 * feasibility. 7659 */ 7660 if (period > max_cfs_quota_period) 7661 return -EINVAL; 7662 7663 mutex_lock(&cfs_constraints_mutex); 7664 ret = __cfs_schedulable(tg, period, quota); 7665 if (ret) 7666 goto out_unlock; 7667 7668 runtime_enabled = quota != RUNTIME_INF; 7669 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7670 /* 7671 * If we need to toggle cfs_bandwidth_used, off->on must occur 7672 * before making related changes, and on->off must occur afterwards 7673 */ 7674 if (runtime_enabled && !runtime_was_enabled) 7675 cfs_bandwidth_usage_inc(); 7676 raw_spin_lock_irq(&cfs_b->lock); 7677 cfs_b->period = ns_to_ktime(period); 7678 cfs_b->quota = quota; 7679 7680 __refill_cfs_bandwidth_runtime(cfs_b); 7681 /* restart the period timer (if active) to handle new period expiry */ 7682 if (runtime_enabled && cfs_b->timer_active) { 7683 /* force a reprogram */ 7684 cfs_b->timer_active = 0; 7685 __start_cfs_bandwidth(cfs_b); 7686 } 7687 raw_spin_unlock_irq(&cfs_b->lock); 7688 7689 for_each_possible_cpu(i) { 7690 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7691 struct rq *rq = cfs_rq->rq; 7692 7693 raw_spin_lock_irq(&rq->lock); 7694 cfs_rq->runtime_enabled = runtime_enabled; 7695 cfs_rq->runtime_remaining = 0; 7696 7697 if (cfs_rq->throttled) 7698 unthrottle_cfs_rq(cfs_rq); 7699 raw_spin_unlock_irq(&rq->lock); 7700 } 7701 if (runtime_was_enabled && !runtime_enabled) 7702 cfs_bandwidth_usage_dec(); 7703out_unlock: 7704 mutex_unlock(&cfs_constraints_mutex); 7705 7706 return ret; 7707} 7708 7709int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7710{ 7711 u64 quota, period; 7712 7713 period = ktime_to_ns(tg->cfs_bandwidth.period); 7714 if (cfs_quota_us < 0) 7715 quota = RUNTIME_INF; 7716 else 7717 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7718 7719 return tg_set_cfs_bandwidth(tg, period, quota); 7720} 7721 7722long tg_get_cfs_quota(struct task_group *tg) 7723{ 7724 u64 quota_us; 7725 7726 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7727 return -1; 7728 7729 quota_us = tg->cfs_bandwidth.quota; 7730 do_div(quota_us, NSEC_PER_USEC); 7731 7732 return quota_us; 7733} 7734 7735int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7736{ 7737 u64 quota, period; 7738 7739 period = (u64)cfs_period_us * NSEC_PER_USEC; 7740 quota = tg->cfs_bandwidth.quota; 7741 7742 return tg_set_cfs_bandwidth(tg, period, quota); 7743} 7744 7745long tg_get_cfs_period(struct task_group *tg) 7746{ 7747 u64 cfs_period_us; 7748 7749 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7750 do_div(cfs_period_us, NSEC_PER_USEC); 7751 7752 return cfs_period_us; 7753} 7754 7755static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7756 struct cftype *cft) 7757{ 7758 return tg_get_cfs_quota(css_tg(css)); 7759} 7760 7761static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7762 struct cftype *cftype, s64 cfs_quota_us) 7763{ 7764 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7765} 7766 7767static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7768 struct cftype *cft) 7769{ 7770 return tg_get_cfs_period(css_tg(css)); 7771} 7772 7773static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7774 struct cftype *cftype, u64 cfs_period_us) 7775{ 7776 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7777} 7778 7779struct cfs_schedulable_data { 7780 struct task_group *tg; 7781 u64 period, quota; 7782}; 7783 7784/* 7785 * normalize group quota/period to be quota/max_period 7786 * note: units are usecs 7787 */ 7788static u64 normalize_cfs_quota(struct task_group *tg, 7789 struct cfs_schedulable_data *d) 7790{ 7791 u64 quota, period; 7792 7793 if (tg == d->tg) { 7794 period = d->period; 7795 quota = d->quota; 7796 } else { 7797 period = tg_get_cfs_period(tg); 7798 quota = tg_get_cfs_quota(tg); 7799 } 7800 7801 /* note: these should typically be equivalent */ 7802 if (quota == RUNTIME_INF || quota == -1) 7803 return RUNTIME_INF; 7804 7805 return to_ratio(period, quota); 7806} 7807 7808static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7809{ 7810 struct cfs_schedulable_data *d = data; 7811 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7812 s64 quota = 0, parent_quota = -1; 7813 7814 if (!tg->parent) { 7815 quota = RUNTIME_INF; 7816 } else { 7817 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7818 7819 quota = normalize_cfs_quota(tg, d); 7820 parent_quota = parent_b->hierarchal_quota; 7821 7822 /* 7823 * ensure max(child_quota) <= parent_quota, inherit when no 7824 * limit is set 7825 */ 7826 if (quota == RUNTIME_INF) 7827 quota = parent_quota; 7828 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7829 return -EINVAL; 7830 } 7831 cfs_b->hierarchal_quota = quota; 7832 7833 return 0; 7834} 7835 7836static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7837{ 7838 int ret; 7839 struct cfs_schedulable_data data = { 7840 .tg = tg, 7841 .period = period, 7842 .quota = quota, 7843 }; 7844 7845 if (quota != RUNTIME_INF) { 7846 do_div(data.period, NSEC_PER_USEC); 7847 do_div(data.quota, NSEC_PER_USEC); 7848 } 7849 7850 rcu_read_lock(); 7851 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7852 rcu_read_unlock(); 7853 7854 return ret; 7855} 7856 7857static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft, 7858 struct cgroup_map_cb *cb) 7859{ 7860 struct task_group *tg = css_tg(css); 7861 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7862 7863 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7864 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7865 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7866 7867 return 0; 7868} 7869#endif /* CONFIG_CFS_BANDWIDTH */ 7870#endif /* CONFIG_FAIR_GROUP_SCHED */ 7871 7872#ifdef CONFIG_RT_GROUP_SCHED 7873static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7874 struct cftype *cft, s64 val) 7875{ 7876 return sched_group_set_rt_runtime(css_tg(css), val); 7877} 7878 7879static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7880 struct cftype *cft) 7881{ 7882 return sched_group_rt_runtime(css_tg(css)); 7883} 7884 7885static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7886 struct cftype *cftype, u64 rt_period_us) 7887{ 7888 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7889} 7890 7891static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7892 struct cftype *cft) 7893{ 7894 return sched_group_rt_period(css_tg(css)); 7895} 7896#endif /* CONFIG_RT_GROUP_SCHED */ 7897 7898static struct cftype cpu_files[] = { 7899#ifdef CONFIG_FAIR_GROUP_SCHED 7900 { 7901 .name = "shares", 7902 .read_u64 = cpu_shares_read_u64, 7903 .write_u64 = cpu_shares_write_u64, 7904 }, 7905#endif 7906#ifdef CONFIG_CFS_BANDWIDTH 7907 { 7908 .name = "cfs_quota_us", 7909 .read_s64 = cpu_cfs_quota_read_s64, 7910 .write_s64 = cpu_cfs_quota_write_s64, 7911 }, 7912 { 7913 .name = "cfs_period_us", 7914 .read_u64 = cpu_cfs_period_read_u64, 7915 .write_u64 = cpu_cfs_period_write_u64, 7916 }, 7917 { 7918 .name = "stat", 7919 .read_map = cpu_stats_show, 7920 }, 7921#endif 7922#ifdef CONFIG_RT_GROUP_SCHED 7923 { 7924 .name = "rt_runtime_us", 7925 .read_s64 = cpu_rt_runtime_read, 7926 .write_s64 = cpu_rt_runtime_write, 7927 }, 7928 { 7929 .name = "rt_period_us", 7930 .read_u64 = cpu_rt_period_read_uint, 7931 .write_u64 = cpu_rt_period_write_uint, 7932 }, 7933#endif 7934 { } /* terminate */ 7935}; 7936 7937struct cgroup_subsys cpu_cgroup_subsys = { 7938 .name = "cpu", 7939 .css_alloc = cpu_cgroup_css_alloc, 7940 .css_free = cpu_cgroup_css_free, 7941 .css_online = cpu_cgroup_css_online, 7942 .css_offline = cpu_cgroup_css_offline, 7943 .can_attach = cpu_cgroup_can_attach, 7944 .attach = cpu_cgroup_attach, 7945 .exit = cpu_cgroup_exit, 7946 .subsys_id = cpu_cgroup_subsys_id, 7947 .base_cftypes = cpu_files, 7948 .early_init = 1, 7949}; 7950 7951#endif /* CONFIG_CGROUP_SCHED */ 7952 7953void dump_cpu_task(int cpu) 7954{ 7955 pr_info("Task dump for CPU %d:\n", cpu); 7956 sched_show_task(cpu_curr(cpu)); 7957} 7958