core.c revision 81a44c5441d7f7d2c3dc9105f4d65ad0d5818617
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 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1112 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1113 1114out: 1115 return ret; 1116} 1117 1118struct migration_arg { 1119 struct task_struct *task; 1120 int dest_cpu; 1121}; 1122 1123static int migration_cpu_stop(void *data); 1124 1125/* 1126 * wait_task_inactive - wait for a thread to unschedule. 1127 * 1128 * If @match_state is nonzero, it's the @p->state value just checked and 1129 * not expected to change. If it changes, i.e. @p might have woken up, 1130 * then return zero. When we succeed in waiting for @p to be off its CPU, 1131 * we return a positive number (its total switch count). If a second call 1132 * a short while later returns the same number, the caller can be sure that 1133 * @p has remained unscheduled the whole time. 1134 * 1135 * The caller must ensure that the task *will* unschedule sometime soon, 1136 * else this function might spin for a *long* time. This function can't 1137 * be called with interrupts off, or it may introduce deadlock with 1138 * smp_call_function() if an IPI is sent by the same process we are 1139 * waiting to become inactive. 1140 */ 1141unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1142{ 1143 unsigned long flags; 1144 int running, on_rq; 1145 unsigned long ncsw; 1146 struct rq *rq; 1147 1148 for (;;) { 1149 /* 1150 * We do the initial early heuristics without holding 1151 * any task-queue locks at all. We'll only try to get 1152 * the runqueue lock when things look like they will 1153 * work out! 1154 */ 1155 rq = task_rq(p); 1156 1157 /* 1158 * If the task is actively running on another CPU 1159 * still, just relax and busy-wait without holding 1160 * any locks. 1161 * 1162 * NOTE! Since we don't hold any locks, it's not 1163 * even sure that "rq" stays as the right runqueue! 1164 * But we don't care, since "task_running()" will 1165 * return false if the runqueue has changed and p 1166 * is actually now running somewhere else! 1167 */ 1168 while (task_running(rq, p)) { 1169 if (match_state && unlikely(p->state != match_state)) 1170 return 0; 1171 cpu_relax(); 1172 } 1173 1174 /* 1175 * Ok, time to look more closely! We need the rq 1176 * lock now, to be *sure*. If we're wrong, we'll 1177 * just go back and repeat. 1178 */ 1179 rq = task_rq_lock(p, &flags); 1180 trace_sched_wait_task(p); 1181 running = task_running(rq, p); 1182 on_rq = p->on_rq; 1183 ncsw = 0; 1184 if (!match_state || p->state == match_state) 1185 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1186 task_rq_unlock(rq, p, &flags); 1187 1188 /* 1189 * If it changed from the expected state, bail out now. 1190 */ 1191 if (unlikely(!ncsw)) 1192 break; 1193 1194 /* 1195 * Was it really running after all now that we 1196 * checked with the proper locks actually held? 1197 * 1198 * Oops. Go back and try again.. 1199 */ 1200 if (unlikely(running)) { 1201 cpu_relax(); 1202 continue; 1203 } 1204 1205 /* 1206 * It's not enough that it's not actively running, 1207 * it must be off the runqueue _entirely_, and not 1208 * preempted! 1209 * 1210 * So if it was still runnable (but just not actively 1211 * running right now), it's preempted, and we should 1212 * yield - it could be a while. 1213 */ 1214 if (unlikely(on_rq)) { 1215 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1216 1217 set_current_state(TASK_UNINTERRUPTIBLE); 1218 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1219 continue; 1220 } 1221 1222 /* 1223 * Ahh, all good. It wasn't running, and it wasn't 1224 * runnable, which means that it will never become 1225 * running in the future either. We're all done! 1226 */ 1227 break; 1228 } 1229 1230 return ncsw; 1231} 1232 1233/*** 1234 * kick_process - kick a running thread to enter/exit the kernel 1235 * @p: the to-be-kicked thread 1236 * 1237 * Cause a process which is running on another CPU to enter 1238 * kernel-mode, without any delay. (to get signals handled.) 1239 * 1240 * NOTE: this function doesn't have to take the runqueue lock, 1241 * because all it wants to ensure is that the remote task enters 1242 * the kernel. If the IPI races and the task has been migrated 1243 * to another CPU then no harm is done and the purpose has been 1244 * achieved as well. 1245 */ 1246void kick_process(struct task_struct *p) 1247{ 1248 int cpu; 1249 1250 preempt_disable(); 1251 cpu = task_cpu(p); 1252 if ((cpu != smp_processor_id()) && task_curr(p)) 1253 smp_send_reschedule(cpu); 1254 preempt_enable(); 1255} 1256EXPORT_SYMBOL_GPL(kick_process); 1257#endif /* CONFIG_SMP */ 1258 1259#ifdef CONFIG_SMP 1260/* 1261 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1262 */ 1263static int select_fallback_rq(int cpu, struct task_struct *p) 1264{ 1265 int nid = cpu_to_node(cpu); 1266 const struct cpumask *nodemask = NULL; 1267 enum { cpuset, possible, fail } state = cpuset; 1268 int dest_cpu; 1269 1270 /* 1271 * If the node that the cpu is on has been offlined, cpu_to_node() 1272 * will return -1. There is no cpu on the node, and we should 1273 * select the cpu on the other node. 1274 */ 1275 if (nid != -1) { 1276 nodemask = cpumask_of_node(nid); 1277 1278 /* Look for allowed, online CPU in same node. */ 1279 for_each_cpu(dest_cpu, nodemask) { 1280 if (!cpu_online(dest_cpu)) 1281 continue; 1282 if (!cpu_active(dest_cpu)) 1283 continue; 1284 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1285 return dest_cpu; 1286 } 1287 } 1288 1289 for (;;) { 1290 /* Any allowed, online CPU? */ 1291 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1292 if (!cpu_online(dest_cpu)) 1293 continue; 1294 if (!cpu_active(dest_cpu)) 1295 continue; 1296 goto out; 1297 } 1298 1299 switch (state) { 1300 case cpuset: 1301 /* No more Mr. Nice Guy. */ 1302 cpuset_cpus_allowed_fallback(p); 1303 state = possible; 1304 break; 1305 1306 case possible: 1307 do_set_cpus_allowed(p, cpu_possible_mask); 1308 state = fail; 1309 break; 1310 1311 case fail: 1312 BUG(); 1313 break; 1314 } 1315 } 1316 1317out: 1318 if (state != cpuset) { 1319 /* 1320 * Don't tell them about moving exiting tasks or 1321 * kernel threads (both mm NULL), since they never 1322 * leave kernel. 1323 */ 1324 if (p->mm && printk_ratelimit()) { 1325 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1326 task_pid_nr(p), p->comm, cpu); 1327 } 1328 } 1329 1330 return dest_cpu; 1331} 1332 1333/* 1334 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1335 */ 1336static inline 1337int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1338{ 1339 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1340 1341 /* 1342 * In order not to call set_task_cpu() on a blocking task we need 1343 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1344 * cpu. 1345 * 1346 * Since this is common to all placement strategies, this lives here. 1347 * 1348 * [ this allows ->select_task() to simply return task_cpu(p) and 1349 * not worry about this generic constraint ] 1350 */ 1351 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1352 !cpu_online(cpu))) 1353 cpu = select_fallback_rq(task_cpu(p), p); 1354 1355 return cpu; 1356} 1357 1358static void update_avg(u64 *avg, u64 sample) 1359{ 1360 s64 diff = sample - *avg; 1361 *avg += diff >> 3; 1362} 1363#endif 1364 1365static void 1366ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1367{ 1368#ifdef CONFIG_SCHEDSTATS 1369 struct rq *rq = this_rq(); 1370 1371#ifdef CONFIG_SMP 1372 int this_cpu = smp_processor_id(); 1373 1374 if (cpu == this_cpu) { 1375 schedstat_inc(rq, ttwu_local); 1376 schedstat_inc(p, se.statistics.nr_wakeups_local); 1377 } else { 1378 struct sched_domain *sd; 1379 1380 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1381 rcu_read_lock(); 1382 for_each_domain(this_cpu, sd) { 1383 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1384 schedstat_inc(sd, ttwu_wake_remote); 1385 break; 1386 } 1387 } 1388 rcu_read_unlock(); 1389 } 1390 1391 if (wake_flags & WF_MIGRATED) 1392 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1393 1394#endif /* CONFIG_SMP */ 1395 1396 schedstat_inc(rq, ttwu_count); 1397 schedstat_inc(p, se.statistics.nr_wakeups); 1398 1399 if (wake_flags & WF_SYNC) 1400 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1401 1402#endif /* CONFIG_SCHEDSTATS */ 1403} 1404 1405static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1406{ 1407 activate_task(rq, p, en_flags); 1408 p->on_rq = 1; 1409 1410 /* if a worker is waking up, notify workqueue */ 1411 if (p->flags & PF_WQ_WORKER) 1412 wq_worker_waking_up(p, cpu_of(rq)); 1413} 1414 1415/* 1416 * Mark the task runnable and perform wakeup-preemption. 1417 */ 1418static void 1419ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1420{ 1421 check_preempt_curr(rq, p, wake_flags); 1422 trace_sched_wakeup(p, true); 1423 1424 p->state = TASK_RUNNING; 1425#ifdef CONFIG_SMP 1426 if (p->sched_class->task_woken) 1427 p->sched_class->task_woken(rq, p); 1428 1429 if (rq->idle_stamp) { 1430 u64 delta = rq_clock(rq) - rq->idle_stamp; 1431 u64 max = 2*rq->max_idle_balance_cost; 1432 1433 update_avg(&rq->avg_idle, delta); 1434 1435 if (rq->avg_idle > max) 1436 rq->avg_idle = max; 1437 1438 rq->idle_stamp = 0; 1439 } 1440#endif 1441} 1442 1443static void 1444ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1445{ 1446#ifdef CONFIG_SMP 1447 if (p->sched_contributes_to_load) 1448 rq->nr_uninterruptible--; 1449#endif 1450 1451 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1452 ttwu_do_wakeup(rq, p, wake_flags); 1453} 1454 1455/* 1456 * Called in case the task @p isn't fully descheduled from its runqueue, 1457 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1458 * since all we need to do is flip p->state to TASK_RUNNING, since 1459 * the task is still ->on_rq. 1460 */ 1461static int ttwu_remote(struct task_struct *p, int wake_flags) 1462{ 1463 struct rq *rq; 1464 int ret = 0; 1465 1466 rq = __task_rq_lock(p); 1467 if (p->on_rq) { 1468 /* check_preempt_curr() may use rq clock */ 1469 update_rq_clock(rq); 1470 ttwu_do_wakeup(rq, p, wake_flags); 1471 ret = 1; 1472 } 1473 __task_rq_unlock(rq); 1474 1475 return ret; 1476} 1477 1478#ifdef CONFIG_SMP 1479static void sched_ttwu_pending(void) 1480{ 1481 struct rq *rq = this_rq(); 1482 struct llist_node *llist = llist_del_all(&rq->wake_list); 1483 struct task_struct *p; 1484 1485 raw_spin_lock(&rq->lock); 1486 1487 while (llist) { 1488 p = llist_entry(llist, struct task_struct, wake_entry); 1489 llist = llist_next(llist); 1490 ttwu_do_activate(rq, p, 0); 1491 } 1492 1493 raw_spin_unlock(&rq->lock); 1494} 1495 1496void scheduler_ipi(void) 1497{ 1498 /* 1499 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1500 * TIF_NEED_RESCHED remotely (for the first time) will also send 1501 * this IPI. 1502 */ 1503 preempt_fold_need_resched(); 1504 1505 if (llist_empty(&this_rq()->wake_list) 1506 && !tick_nohz_full_cpu(smp_processor_id()) 1507 && !got_nohz_idle_kick()) 1508 return; 1509 1510 /* 1511 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1512 * traditionally all their work was done from the interrupt return 1513 * path. Now that we actually do some work, we need to make sure 1514 * we do call them. 1515 * 1516 * Some archs already do call them, luckily irq_enter/exit nest 1517 * properly. 1518 * 1519 * Arguably we should visit all archs and update all handlers, 1520 * however a fair share of IPIs are still resched only so this would 1521 * somewhat pessimize the simple resched case. 1522 */ 1523 irq_enter(); 1524 tick_nohz_full_check(); 1525 sched_ttwu_pending(); 1526 1527 /* 1528 * Check if someone kicked us for doing the nohz idle load balance. 1529 */ 1530 if (unlikely(got_nohz_idle_kick())) { 1531 this_rq()->idle_balance = 1; 1532 raise_softirq_irqoff(SCHED_SOFTIRQ); 1533 } 1534 irq_exit(); 1535} 1536 1537static void ttwu_queue_remote(struct task_struct *p, int cpu) 1538{ 1539 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1540 smp_send_reschedule(cpu); 1541} 1542 1543bool cpus_share_cache(int this_cpu, int that_cpu) 1544{ 1545 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1546} 1547#endif /* CONFIG_SMP */ 1548 1549static void ttwu_queue(struct task_struct *p, int cpu) 1550{ 1551 struct rq *rq = cpu_rq(cpu); 1552 1553#if defined(CONFIG_SMP) 1554 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1555 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1556 ttwu_queue_remote(p, cpu); 1557 return; 1558 } 1559#endif 1560 1561 raw_spin_lock(&rq->lock); 1562 ttwu_do_activate(rq, p, 0); 1563 raw_spin_unlock(&rq->lock); 1564} 1565 1566/** 1567 * try_to_wake_up - wake up a thread 1568 * @p: the thread to be awakened 1569 * @state: the mask of task states that can be woken 1570 * @wake_flags: wake modifier flags (WF_*) 1571 * 1572 * Put it on the run-queue if it's not already there. The "current" 1573 * thread is always on the run-queue (except when the actual 1574 * re-schedule is in progress), and as such you're allowed to do 1575 * the simpler "current->state = TASK_RUNNING" to mark yourself 1576 * runnable without the overhead of this. 1577 * 1578 * Return: %true if @p was woken up, %false if it was already running. 1579 * or @state didn't match @p's state. 1580 */ 1581static int 1582try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1583{ 1584 unsigned long flags; 1585 int cpu, success = 0; 1586 1587 /* 1588 * If we are going to wake up a thread waiting for CONDITION we 1589 * need to ensure that CONDITION=1 done by the caller can not be 1590 * reordered with p->state check below. This pairs with mb() in 1591 * set_current_state() the waiting thread does. 1592 */ 1593 smp_mb__before_spinlock(); 1594 raw_spin_lock_irqsave(&p->pi_lock, flags); 1595 if (!(p->state & state)) 1596 goto out; 1597 1598 success = 1; /* we're going to change ->state */ 1599 cpu = task_cpu(p); 1600 1601 if (p->on_rq && ttwu_remote(p, wake_flags)) 1602 goto stat; 1603 1604#ifdef CONFIG_SMP 1605 /* 1606 * If the owning (remote) cpu is still in the middle of schedule() with 1607 * this task as prev, wait until its done referencing the task. 1608 */ 1609 while (p->on_cpu) 1610 cpu_relax(); 1611 /* 1612 * Pairs with the smp_wmb() in finish_lock_switch(). 1613 */ 1614 smp_rmb(); 1615 1616 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1617 p->state = TASK_WAKING; 1618 1619 if (p->sched_class->task_waking) 1620 p->sched_class->task_waking(p); 1621 1622 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1623 if (task_cpu(p) != cpu) { 1624 wake_flags |= WF_MIGRATED; 1625 set_task_cpu(p, cpu); 1626 } 1627#endif /* CONFIG_SMP */ 1628 1629 ttwu_queue(p, cpu); 1630stat: 1631 ttwu_stat(p, cpu, wake_flags); 1632out: 1633 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1634 1635 return success; 1636} 1637 1638/** 1639 * try_to_wake_up_local - try to wake up a local task with rq lock held 1640 * @p: the thread to be awakened 1641 * 1642 * Put @p on the run-queue if it's not already there. The caller must 1643 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1644 * the current task. 1645 */ 1646static void try_to_wake_up_local(struct task_struct *p) 1647{ 1648 struct rq *rq = task_rq(p); 1649 1650 if (WARN_ON_ONCE(rq != this_rq()) || 1651 WARN_ON_ONCE(p == current)) 1652 return; 1653 1654 lockdep_assert_held(&rq->lock); 1655 1656 if (!raw_spin_trylock(&p->pi_lock)) { 1657 raw_spin_unlock(&rq->lock); 1658 raw_spin_lock(&p->pi_lock); 1659 raw_spin_lock(&rq->lock); 1660 } 1661 1662 if (!(p->state & TASK_NORMAL)) 1663 goto out; 1664 1665 if (!p->on_rq) 1666 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1667 1668 ttwu_do_wakeup(rq, p, 0); 1669 ttwu_stat(p, smp_processor_id(), 0); 1670out: 1671 raw_spin_unlock(&p->pi_lock); 1672} 1673 1674/** 1675 * wake_up_process - Wake up a specific process 1676 * @p: The process to be woken up. 1677 * 1678 * Attempt to wake up the nominated process and move it to the set of runnable 1679 * processes. 1680 * 1681 * Return: 1 if the process was woken up, 0 if it was already running. 1682 * 1683 * It may be assumed that this function implies a write memory barrier before 1684 * changing the task state if and only if any tasks are woken up. 1685 */ 1686int wake_up_process(struct task_struct *p) 1687{ 1688 WARN_ON(task_is_stopped_or_traced(p)); 1689 return try_to_wake_up(p, TASK_NORMAL, 0); 1690} 1691EXPORT_SYMBOL(wake_up_process); 1692 1693int wake_up_state(struct task_struct *p, unsigned int state) 1694{ 1695 return try_to_wake_up(p, state, 0); 1696} 1697 1698/* 1699 * Perform scheduler related setup for a newly forked process p. 1700 * p is forked by current. 1701 * 1702 * __sched_fork() is basic setup used by init_idle() too: 1703 */ 1704static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1705{ 1706 p->on_rq = 0; 1707 1708 p->se.on_rq = 0; 1709 p->se.exec_start = 0; 1710 p->se.sum_exec_runtime = 0; 1711 p->se.prev_sum_exec_runtime = 0; 1712 p->se.nr_migrations = 0; 1713 p->se.vruntime = 0; 1714 INIT_LIST_HEAD(&p->se.group_node); 1715 1716#ifdef CONFIG_SCHEDSTATS 1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1718#endif 1719 1720 RB_CLEAR_NODE(&p->dl.rb_node); 1721 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1722 p->dl.dl_runtime = p->dl.runtime = 0; 1723 p->dl.dl_deadline = p->dl.deadline = 0; 1724 p->dl.dl_period = 0; 1725 p->dl.flags = 0; 1726 1727 INIT_LIST_HEAD(&p->rt.run_list); 1728 1729#ifdef CONFIG_PREEMPT_NOTIFIERS 1730 INIT_HLIST_HEAD(&p->preempt_notifiers); 1731#endif 1732 1733#ifdef CONFIG_NUMA_BALANCING 1734 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1735 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1736 p->mm->numa_scan_seq = 0; 1737 } 1738 1739 if (clone_flags & CLONE_VM) 1740 p->numa_preferred_nid = current->numa_preferred_nid; 1741 else 1742 p->numa_preferred_nid = -1; 1743 1744 p->node_stamp = 0ULL; 1745 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1746 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1747 p->numa_work.next = &p->numa_work; 1748 p->numa_faults_memory = NULL; 1749 p->numa_faults_buffer_memory = NULL; 1750 p->last_task_numa_placement = 0; 1751 p->last_sum_exec_runtime = 0; 1752 1753 INIT_LIST_HEAD(&p->numa_entry); 1754 p->numa_group = NULL; 1755#endif /* CONFIG_NUMA_BALANCING */ 1756} 1757 1758#ifdef CONFIG_NUMA_BALANCING 1759#ifdef CONFIG_SCHED_DEBUG 1760void set_numabalancing_state(bool enabled) 1761{ 1762 if (enabled) 1763 sched_feat_set("NUMA"); 1764 else 1765 sched_feat_set("NO_NUMA"); 1766} 1767#else 1768__read_mostly bool numabalancing_enabled; 1769 1770void set_numabalancing_state(bool enabled) 1771{ 1772 numabalancing_enabled = enabled; 1773} 1774#endif /* CONFIG_SCHED_DEBUG */ 1775 1776#ifdef CONFIG_PROC_SYSCTL 1777int sysctl_numa_balancing(struct ctl_table *table, int write, 1778 void __user *buffer, size_t *lenp, loff_t *ppos) 1779{ 1780 struct ctl_table t; 1781 int err; 1782 int state = numabalancing_enabled; 1783 1784 if (write && !capable(CAP_SYS_ADMIN)) 1785 return -EPERM; 1786 1787 t = *table; 1788 t.data = &state; 1789 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1790 if (err < 0) 1791 return err; 1792 if (write) 1793 set_numabalancing_state(state); 1794 return err; 1795} 1796#endif 1797#endif 1798 1799/* 1800 * fork()/clone()-time setup: 1801 */ 1802int sched_fork(unsigned long clone_flags, struct task_struct *p) 1803{ 1804 unsigned long flags; 1805 int cpu = get_cpu(); 1806 1807 __sched_fork(clone_flags, p); 1808 /* 1809 * We mark the process as running here. This guarantees that 1810 * nobody will actually run it, and a signal or other external 1811 * event cannot wake it up and insert it on the runqueue either. 1812 */ 1813 p->state = TASK_RUNNING; 1814 1815 /* 1816 * Make sure we do not leak PI boosting priority to the child. 1817 */ 1818 p->prio = current->normal_prio; 1819 1820 /* 1821 * Revert to default priority/policy on fork if requested. 1822 */ 1823 if (unlikely(p->sched_reset_on_fork)) { 1824 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1825 p->policy = SCHED_NORMAL; 1826 p->static_prio = NICE_TO_PRIO(0); 1827 p->rt_priority = 0; 1828 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1829 p->static_prio = NICE_TO_PRIO(0); 1830 1831 p->prio = p->normal_prio = __normal_prio(p); 1832 set_load_weight(p); 1833 1834 /* 1835 * We don't need the reset flag anymore after the fork. It has 1836 * fulfilled its duty: 1837 */ 1838 p->sched_reset_on_fork = 0; 1839 } 1840 1841 if (dl_prio(p->prio)) { 1842 put_cpu(); 1843 return -EAGAIN; 1844 } else if (rt_prio(p->prio)) { 1845 p->sched_class = &rt_sched_class; 1846 } else { 1847 p->sched_class = &fair_sched_class; 1848 } 1849 1850 if (p->sched_class->task_fork) 1851 p->sched_class->task_fork(p); 1852 1853 /* 1854 * The child is not yet in the pid-hash so no cgroup attach races, 1855 * and the cgroup is pinned to this child due to cgroup_fork() 1856 * is ran before sched_fork(). 1857 * 1858 * Silence PROVE_RCU. 1859 */ 1860 raw_spin_lock_irqsave(&p->pi_lock, flags); 1861 set_task_cpu(p, cpu); 1862 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1863 1864#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1865 if (likely(sched_info_on())) 1866 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1867#endif 1868#if defined(CONFIG_SMP) 1869 p->on_cpu = 0; 1870#endif 1871 init_task_preempt_count(p); 1872#ifdef CONFIG_SMP 1873 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1874 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1875#endif 1876 1877 put_cpu(); 1878 return 0; 1879} 1880 1881unsigned long to_ratio(u64 period, u64 runtime) 1882{ 1883 if (runtime == RUNTIME_INF) 1884 return 1ULL << 20; 1885 1886 /* 1887 * Doing this here saves a lot of checks in all 1888 * the calling paths, and returning zero seems 1889 * safe for them anyway. 1890 */ 1891 if (period == 0) 1892 return 0; 1893 1894 return div64_u64(runtime << 20, period); 1895} 1896 1897#ifdef CONFIG_SMP 1898inline struct dl_bw *dl_bw_of(int i) 1899{ 1900 return &cpu_rq(i)->rd->dl_bw; 1901} 1902 1903static inline int dl_bw_cpus(int i) 1904{ 1905 struct root_domain *rd = cpu_rq(i)->rd; 1906 int cpus = 0; 1907 1908 for_each_cpu_and(i, rd->span, cpu_active_mask) 1909 cpus++; 1910 1911 return cpus; 1912} 1913#else 1914inline struct dl_bw *dl_bw_of(int i) 1915{ 1916 return &cpu_rq(i)->dl.dl_bw; 1917} 1918 1919static inline int dl_bw_cpus(int i) 1920{ 1921 return 1; 1922} 1923#endif 1924 1925static inline 1926void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1927{ 1928 dl_b->total_bw -= tsk_bw; 1929} 1930 1931static inline 1932void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1933{ 1934 dl_b->total_bw += tsk_bw; 1935} 1936 1937static inline 1938bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1939{ 1940 return dl_b->bw != -1 && 1941 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1942} 1943 1944/* 1945 * We must be sure that accepting a new task (or allowing changing the 1946 * parameters of an existing one) is consistent with the bandwidth 1947 * constraints. If yes, this function also accordingly updates the currently 1948 * allocated bandwidth to reflect the new situation. 1949 * 1950 * This function is called while holding p's rq->lock. 1951 */ 1952static int dl_overflow(struct task_struct *p, int policy, 1953 const struct sched_attr *attr) 1954{ 1955 1956 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1957 u64 period = attr->sched_period; 1958 u64 runtime = attr->sched_runtime; 1959 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1960 int cpus, err = -1; 1961 1962 if (new_bw == p->dl.dl_bw) 1963 return 0; 1964 1965 /* 1966 * Either if a task, enters, leave, or stays -deadline but changes 1967 * its parameters, we may need to update accordingly the total 1968 * allocated bandwidth of the container. 1969 */ 1970 raw_spin_lock(&dl_b->lock); 1971 cpus = dl_bw_cpus(task_cpu(p)); 1972 if (dl_policy(policy) && !task_has_dl_policy(p) && 1973 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1974 __dl_add(dl_b, new_bw); 1975 err = 0; 1976 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1977 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1978 __dl_clear(dl_b, p->dl.dl_bw); 1979 __dl_add(dl_b, new_bw); 1980 err = 0; 1981 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1982 __dl_clear(dl_b, p->dl.dl_bw); 1983 err = 0; 1984 } 1985 raw_spin_unlock(&dl_b->lock); 1986 1987 return err; 1988} 1989 1990extern void init_dl_bw(struct dl_bw *dl_b); 1991 1992/* 1993 * wake_up_new_task - wake up a newly created task for the first time. 1994 * 1995 * This function will do some initial scheduler statistics housekeeping 1996 * that must be done for every newly created context, then puts the task 1997 * on the runqueue and wakes it. 1998 */ 1999void wake_up_new_task(struct task_struct *p) 2000{ 2001 unsigned long flags; 2002 struct rq *rq; 2003 2004 raw_spin_lock_irqsave(&p->pi_lock, flags); 2005#ifdef CONFIG_SMP 2006 /* 2007 * Fork balancing, do it here and not earlier because: 2008 * - cpus_allowed can change in the fork path 2009 * - any previously selected cpu might disappear through hotplug 2010 */ 2011 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2012#endif 2013 2014 /* Initialize new task's runnable average */ 2015 init_task_runnable_average(p); 2016 rq = __task_rq_lock(p); 2017 activate_task(rq, p, 0); 2018 p->on_rq = 1; 2019 trace_sched_wakeup_new(p, true); 2020 check_preempt_curr(rq, p, WF_FORK); 2021#ifdef CONFIG_SMP 2022 if (p->sched_class->task_woken) 2023 p->sched_class->task_woken(rq, p); 2024#endif 2025 task_rq_unlock(rq, p, &flags); 2026} 2027 2028#ifdef CONFIG_PREEMPT_NOTIFIERS 2029 2030/** 2031 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2032 * @notifier: notifier struct to register 2033 */ 2034void preempt_notifier_register(struct preempt_notifier *notifier) 2035{ 2036 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2037} 2038EXPORT_SYMBOL_GPL(preempt_notifier_register); 2039 2040/** 2041 * preempt_notifier_unregister - no longer interested in preemption notifications 2042 * @notifier: notifier struct to unregister 2043 * 2044 * This is safe to call from within a preemption notifier. 2045 */ 2046void preempt_notifier_unregister(struct preempt_notifier *notifier) 2047{ 2048 hlist_del(¬ifier->link); 2049} 2050EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2051 2052static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2053{ 2054 struct preempt_notifier *notifier; 2055 2056 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2057 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2058} 2059 2060static void 2061fire_sched_out_preempt_notifiers(struct task_struct *curr, 2062 struct task_struct *next) 2063{ 2064 struct preempt_notifier *notifier; 2065 2066 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2067 notifier->ops->sched_out(notifier, next); 2068} 2069 2070#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2071 2072static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2073{ 2074} 2075 2076static void 2077fire_sched_out_preempt_notifiers(struct task_struct *curr, 2078 struct task_struct *next) 2079{ 2080} 2081 2082#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2083 2084/** 2085 * prepare_task_switch - prepare to switch tasks 2086 * @rq: the runqueue preparing to switch 2087 * @prev: the current task that is being switched out 2088 * @next: the task we are going to switch to. 2089 * 2090 * This is called with the rq lock held and interrupts off. It must 2091 * be paired with a subsequent finish_task_switch after the context 2092 * switch. 2093 * 2094 * prepare_task_switch sets up locking and calls architecture specific 2095 * hooks. 2096 */ 2097static inline void 2098prepare_task_switch(struct rq *rq, struct task_struct *prev, 2099 struct task_struct *next) 2100{ 2101 trace_sched_switch(prev, next); 2102 sched_info_switch(rq, prev, next); 2103 perf_event_task_sched_out(prev, next); 2104 fire_sched_out_preempt_notifiers(prev, next); 2105 prepare_lock_switch(rq, next); 2106 prepare_arch_switch(next); 2107} 2108 2109/** 2110 * finish_task_switch - clean up after a task-switch 2111 * @rq: runqueue associated with task-switch 2112 * @prev: the thread we just switched away from. 2113 * 2114 * finish_task_switch must be called after the context switch, paired 2115 * with a prepare_task_switch call before the context switch. 2116 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2117 * and do any other architecture-specific cleanup actions. 2118 * 2119 * Note that we may have delayed dropping an mm in context_switch(). If 2120 * so, we finish that here outside of the runqueue lock. (Doing it 2121 * with the lock held can cause deadlocks; see schedule() for 2122 * details.) 2123 */ 2124static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2125 __releases(rq->lock) 2126{ 2127 struct mm_struct *mm = rq->prev_mm; 2128 long prev_state; 2129 2130 rq->prev_mm = NULL; 2131 2132 /* 2133 * A task struct has one reference for the use as "current". 2134 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2135 * schedule one last time. The schedule call will never return, and 2136 * the scheduled task must drop that reference. 2137 * The test for TASK_DEAD must occur while the runqueue locks are 2138 * still held, otherwise prev could be scheduled on another cpu, die 2139 * there before we look at prev->state, and then the reference would 2140 * be dropped twice. 2141 * Manfred Spraul <manfred@colorfullife.com> 2142 */ 2143 prev_state = prev->state; 2144 vtime_task_switch(prev); 2145 finish_arch_switch(prev); 2146 perf_event_task_sched_in(prev, current); 2147 finish_lock_switch(rq, prev); 2148 finish_arch_post_lock_switch(); 2149 2150 fire_sched_in_preempt_notifiers(current); 2151 if (mm) 2152 mmdrop(mm); 2153 if (unlikely(prev_state == TASK_DEAD)) { 2154 task_numa_free(prev); 2155 2156 if (prev->sched_class->task_dead) 2157 prev->sched_class->task_dead(prev); 2158 2159 /* 2160 * Remove function-return probe instances associated with this 2161 * task and put them back on the free list. 2162 */ 2163 kprobe_flush_task(prev); 2164 put_task_struct(prev); 2165 } 2166 2167 tick_nohz_task_switch(current); 2168} 2169 2170#ifdef CONFIG_SMP 2171 2172/* rq->lock is NOT held, but preemption is disabled */ 2173static inline void post_schedule(struct rq *rq) 2174{ 2175 if (rq->post_schedule) { 2176 unsigned long flags; 2177 2178 raw_spin_lock_irqsave(&rq->lock, flags); 2179 if (rq->curr->sched_class->post_schedule) 2180 rq->curr->sched_class->post_schedule(rq); 2181 raw_spin_unlock_irqrestore(&rq->lock, flags); 2182 2183 rq->post_schedule = 0; 2184 } 2185} 2186 2187#else 2188 2189static inline void post_schedule(struct rq *rq) 2190{ 2191} 2192 2193#endif 2194 2195/** 2196 * schedule_tail - first thing a freshly forked thread must call. 2197 * @prev: the thread we just switched away from. 2198 */ 2199asmlinkage void schedule_tail(struct task_struct *prev) 2200 __releases(rq->lock) 2201{ 2202 struct rq *rq = this_rq(); 2203 2204 finish_task_switch(rq, prev); 2205 2206 /* 2207 * FIXME: do we need to worry about rq being invalidated by the 2208 * task_switch? 2209 */ 2210 post_schedule(rq); 2211 2212#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2213 /* In this case, finish_task_switch does not reenable preemption */ 2214 preempt_enable(); 2215#endif 2216 if (current->set_child_tid) 2217 put_user(task_pid_vnr(current), current->set_child_tid); 2218} 2219 2220/* 2221 * context_switch - switch to the new MM and the new 2222 * thread's register state. 2223 */ 2224static inline void 2225context_switch(struct rq *rq, struct task_struct *prev, 2226 struct task_struct *next) 2227{ 2228 struct mm_struct *mm, *oldmm; 2229 2230 prepare_task_switch(rq, prev, next); 2231 2232 mm = next->mm; 2233 oldmm = prev->active_mm; 2234 /* 2235 * For paravirt, this is coupled with an exit in switch_to to 2236 * combine the page table reload and the switch backend into 2237 * one hypercall. 2238 */ 2239 arch_start_context_switch(prev); 2240 2241 if (!mm) { 2242 next->active_mm = oldmm; 2243 atomic_inc(&oldmm->mm_count); 2244 enter_lazy_tlb(oldmm, next); 2245 } else 2246 switch_mm(oldmm, mm, next); 2247 2248 if (!prev->mm) { 2249 prev->active_mm = NULL; 2250 rq->prev_mm = oldmm; 2251 } 2252 /* 2253 * Since the runqueue lock will be released by the next 2254 * task (which is an invalid locking op but in the case 2255 * of the scheduler it's an obvious special-case), so we 2256 * do an early lockdep release here: 2257 */ 2258#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2259 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2260#endif 2261 2262 context_tracking_task_switch(prev, next); 2263 /* Here we just switch the register state and the stack. */ 2264 switch_to(prev, next, prev); 2265 2266 barrier(); 2267 /* 2268 * this_rq must be evaluated again because prev may have moved 2269 * CPUs since it called schedule(), thus the 'rq' on its stack 2270 * frame will be invalid. 2271 */ 2272 finish_task_switch(this_rq(), prev); 2273} 2274 2275/* 2276 * nr_running and nr_context_switches: 2277 * 2278 * externally visible scheduler statistics: current number of runnable 2279 * threads, total number of context switches performed since bootup. 2280 */ 2281unsigned long nr_running(void) 2282{ 2283 unsigned long i, sum = 0; 2284 2285 for_each_online_cpu(i) 2286 sum += cpu_rq(i)->nr_running; 2287 2288 return sum; 2289} 2290 2291unsigned long long nr_context_switches(void) 2292{ 2293 int i; 2294 unsigned long long sum = 0; 2295 2296 for_each_possible_cpu(i) 2297 sum += cpu_rq(i)->nr_switches; 2298 2299 return sum; 2300} 2301 2302unsigned long nr_iowait(void) 2303{ 2304 unsigned long i, sum = 0; 2305 2306 for_each_possible_cpu(i) 2307 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2308 2309 return sum; 2310} 2311 2312unsigned long nr_iowait_cpu(int cpu) 2313{ 2314 struct rq *this = cpu_rq(cpu); 2315 return atomic_read(&this->nr_iowait); 2316} 2317 2318#ifdef CONFIG_SMP 2319 2320/* 2321 * sched_exec - execve() is a valuable balancing opportunity, because at 2322 * this point the task has the smallest effective memory and cache footprint. 2323 */ 2324void sched_exec(void) 2325{ 2326 struct task_struct *p = current; 2327 unsigned long flags; 2328 int dest_cpu; 2329 2330 raw_spin_lock_irqsave(&p->pi_lock, flags); 2331 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2332 if (dest_cpu == smp_processor_id()) 2333 goto unlock; 2334 2335 if (likely(cpu_active(dest_cpu))) { 2336 struct migration_arg arg = { p, dest_cpu }; 2337 2338 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2339 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2340 return; 2341 } 2342unlock: 2343 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2344} 2345 2346#endif 2347 2348DEFINE_PER_CPU(struct kernel_stat, kstat); 2349DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2350 2351EXPORT_PER_CPU_SYMBOL(kstat); 2352EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2353 2354/* 2355 * Return any ns on the sched_clock that have not yet been accounted in 2356 * @p in case that task is currently running. 2357 * 2358 * Called with task_rq_lock() held on @rq. 2359 */ 2360static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2361{ 2362 u64 ns = 0; 2363 2364 if (task_current(rq, p)) { 2365 update_rq_clock(rq); 2366 ns = rq_clock_task(rq) - p->se.exec_start; 2367 if ((s64)ns < 0) 2368 ns = 0; 2369 } 2370 2371 return ns; 2372} 2373 2374unsigned long long task_delta_exec(struct task_struct *p) 2375{ 2376 unsigned long flags; 2377 struct rq *rq; 2378 u64 ns = 0; 2379 2380 rq = task_rq_lock(p, &flags); 2381 ns = do_task_delta_exec(p, rq); 2382 task_rq_unlock(rq, p, &flags); 2383 2384 return ns; 2385} 2386 2387/* 2388 * Return accounted runtime for the task. 2389 * In case the task is currently running, return the runtime plus current's 2390 * pending runtime that have not been accounted yet. 2391 */ 2392unsigned long long task_sched_runtime(struct task_struct *p) 2393{ 2394 unsigned long flags; 2395 struct rq *rq; 2396 u64 ns = 0; 2397 2398#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2399 /* 2400 * 64-bit doesn't need locks to atomically read a 64bit value. 2401 * So we have a optimization chance when the task's delta_exec is 0. 2402 * Reading ->on_cpu is racy, but this is ok. 2403 * 2404 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2405 * If we race with it entering cpu, unaccounted time is 0. This is 2406 * indistinguishable from the read occurring a few cycles earlier. 2407 */ 2408 if (!p->on_cpu) 2409 return p->se.sum_exec_runtime; 2410#endif 2411 2412 rq = task_rq_lock(p, &flags); 2413 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2414 task_rq_unlock(rq, p, &flags); 2415 2416 return ns; 2417} 2418 2419/* 2420 * This function gets called by the timer code, with HZ frequency. 2421 * We call it with interrupts disabled. 2422 */ 2423void scheduler_tick(void) 2424{ 2425 int cpu = smp_processor_id(); 2426 struct rq *rq = cpu_rq(cpu); 2427 struct task_struct *curr = rq->curr; 2428 2429 sched_clock_tick(); 2430 2431 raw_spin_lock(&rq->lock); 2432 update_rq_clock(rq); 2433 curr->sched_class->task_tick(rq, curr, 0); 2434 update_cpu_load_active(rq); 2435 raw_spin_unlock(&rq->lock); 2436 2437 perf_event_task_tick(); 2438 2439#ifdef CONFIG_SMP 2440 rq->idle_balance = idle_cpu(cpu); 2441 trigger_load_balance(rq); 2442#endif 2443 rq_last_tick_reset(rq); 2444} 2445 2446#ifdef CONFIG_NO_HZ_FULL 2447/** 2448 * scheduler_tick_max_deferment 2449 * 2450 * Keep at least one tick per second when a single 2451 * active task is running because the scheduler doesn't 2452 * yet completely support full dynticks environment. 2453 * 2454 * This makes sure that uptime, CFS vruntime, load 2455 * balancing, etc... continue to move forward, even 2456 * with a very low granularity. 2457 * 2458 * Return: Maximum deferment in nanoseconds. 2459 */ 2460u64 scheduler_tick_max_deferment(void) 2461{ 2462 struct rq *rq = this_rq(); 2463 unsigned long next, now = ACCESS_ONCE(jiffies); 2464 2465 next = rq->last_sched_tick + HZ; 2466 2467 if (time_before_eq(next, now)) 2468 return 0; 2469 2470 return jiffies_to_nsecs(next - now); 2471} 2472#endif 2473 2474notrace unsigned long get_parent_ip(unsigned long addr) 2475{ 2476 if (in_lock_functions(addr)) { 2477 addr = CALLER_ADDR2; 2478 if (in_lock_functions(addr)) 2479 addr = CALLER_ADDR3; 2480 } 2481 return addr; 2482} 2483 2484#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2485 defined(CONFIG_PREEMPT_TRACER)) 2486 2487void __kprobes preempt_count_add(int val) 2488{ 2489#ifdef CONFIG_DEBUG_PREEMPT 2490 /* 2491 * Underflow? 2492 */ 2493 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2494 return; 2495#endif 2496 __preempt_count_add(val); 2497#ifdef CONFIG_DEBUG_PREEMPT 2498 /* 2499 * Spinlock count overflowing soon? 2500 */ 2501 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2502 PREEMPT_MASK - 10); 2503#endif 2504 if (preempt_count() == val) { 2505 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2506#ifdef CONFIG_DEBUG_PREEMPT 2507 current->preempt_disable_ip = ip; 2508#endif 2509 trace_preempt_off(CALLER_ADDR0, ip); 2510 } 2511} 2512EXPORT_SYMBOL(preempt_count_add); 2513 2514void __kprobes preempt_count_sub(int val) 2515{ 2516#ifdef CONFIG_DEBUG_PREEMPT 2517 /* 2518 * Underflow? 2519 */ 2520 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2521 return; 2522 /* 2523 * Is the spinlock portion underflowing? 2524 */ 2525 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2526 !(preempt_count() & PREEMPT_MASK))) 2527 return; 2528#endif 2529 2530 if (preempt_count() == val) 2531 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2532 __preempt_count_sub(val); 2533} 2534EXPORT_SYMBOL(preempt_count_sub); 2535 2536#endif 2537 2538/* 2539 * Print scheduling while atomic bug: 2540 */ 2541static noinline void __schedule_bug(struct task_struct *prev) 2542{ 2543 if (oops_in_progress) 2544 return; 2545 2546 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2547 prev->comm, prev->pid, preempt_count()); 2548 2549 debug_show_held_locks(prev); 2550 print_modules(); 2551 if (irqs_disabled()) 2552 print_irqtrace_events(prev); 2553#ifdef CONFIG_DEBUG_PREEMPT 2554 if (in_atomic_preempt_off()) { 2555 pr_err("Preemption disabled at:"); 2556 print_ip_sym(current->preempt_disable_ip); 2557 pr_cont("\n"); 2558 } 2559#endif 2560 dump_stack(); 2561 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2562} 2563 2564/* 2565 * Various schedule()-time debugging checks and statistics: 2566 */ 2567static inline void schedule_debug(struct task_struct *prev) 2568{ 2569 /* 2570 * Test if we are atomic. Since do_exit() needs to call into 2571 * schedule() atomically, we ignore that path. Otherwise whine 2572 * if we are scheduling when we should not. 2573 */ 2574 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2575 __schedule_bug(prev); 2576 rcu_sleep_check(); 2577 2578 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2579 2580 schedstat_inc(this_rq(), sched_count); 2581} 2582 2583/* 2584 * Pick up the highest-prio task: 2585 */ 2586static inline struct task_struct * 2587pick_next_task(struct rq *rq, struct task_struct *prev) 2588{ 2589 const struct sched_class *class; 2590 struct task_struct *p; 2591 2592 /* 2593 * Optimization: we know that if all tasks are in 2594 * the fair class we can call that function directly: 2595 */ 2596 if (likely(prev->sched_class == &fair_sched_class && 2597 rq->nr_running == rq->cfs.h_nr_running)) { 2598 p = fair_sched_class.pick_next_task(rq, prev); 2599 if (likely(p)) 2600 return p; 2601 } 2602 2603 for_each_class(class) { 2604 p = class->pick_next_task(rq, prev); 2605 if (p) 2606 return p; 2607 } 2608 2609 BUG(); /* the idle class will always have a runnable task */ 2610} 2611 2612/* 2613 * __schedule() is the main scheduler function. 2614 * 2615 * The main means of driving the scheduler and thus entering this function are: 2616 * 2617 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2618 * 2619 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2620 * paths. For example, see arch/x86/entry_64.S. 2621 * 2622 * To drive preemption between tasks, the scheduler sets the flag in timer 2623 * interrupt handler scheduler_tick(). 2624 * 2625 * 3. Wakeups don't really cause entry into schedule(). They add a 2626 * task to the run-queue and that's it. 2627 * 2628 * Now, if the new task added to the run-queue preempts the current 2629 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2630 * called on the nearest possible occasion: 2631 * 2632 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2633 * 2634 * - in syscall or exception context, at the next outmost 2635 * preempt_enable(). (this might be as soon as the wake_up()'s 2636 * spin_unlock()!) 2637 * 2638 * - in IRQ context, return from interrupt-handler to 2639 * preemptible context 2640 * 2641 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2642 * then at the next: 2643 * 2644 * - cond_resched() call 2645 * - explicit schedule() call 2646 * - return from syscall or exception to user-space 2647 * - return from interrupt-handler to user-space 2648 */ 2649static void __sched __schedule(void) 2650{ 2651 struct task_struct *prev, *next; 2652 unsigned long *switch_count; 2653 struct rq *rq; 2654 int cpu; 2655 2656need_resched: 2657 preempt_disable(); 2658 cpu = smp_processor_id(); 2659 rq = cpu_rq(cpu); 2660 rcu_note_context_switch(cpu); 2661 prev = rq->curr; 2662 2663 schedule_debug(prev); 2664 2665 if (sched_feat(HRTICK)) 2666 hrtick_clear(rq); 2667 2668 /* 2669 * Make sure that signal_pending_state()->signal_pending() below 2670 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2671 * done by the caller to avoid the race with signal_wake_up(). 2672 */ 2673 smp_mb__before_spinlock(); 2674 raw_spin_lock_irq(&rq->lock); 2675 2676 switch_count = &prev->nivcsw; 2677 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2678 if (unlikely(signal_pending_state(prev->state, prev))) { 2679 prev->state = TASK_RUNNING; 2680 } else { 2681 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2682 prev->on_rq = 0; 2683 2684 /* 2685 * If a worker went to sleep, notify and ask workqueue 2686 * whether it wants to wake up a task to maintain 2687 * concurrency. 2688 */ 2689 if (prev->flags & PF_WQ_WORKER) { 2690 struct task_struct *to_wakeup; 2691 2692 to_wakeup = wq_worker_sleeping(prev, cpu); 2693 if (to_wakeup) 2694 try_to_wake_up_local(to_wakeup); 2695 } 2696 } 2697 switch_count = &prev->nvcsw; 2698 } 2699 2700 if (prev->on_rq || rq->skip_clock_update < 0) 2701 update_rq_clock(rq); 2702 2703 next = pick_next_task(rq, prev); 2704 clear_tsk_need_resched(prev); 2705 clear_preempt_need_resched(); 2706 rq->skip_clock_update = 0; 2707 2708 if (likely(prev != next)) { 2709 rq->nr_switches++; 2710 rq->curr = next; 2711 ++*switch_count; 2712 2713 context_switch(rq, prev, next); /* unlocks the rq */ 2714 /* 2715 * The context switch have flipped the stack from under us 2716 * and restored the local variables which were saved when 2717 * this task called schedule() in the past. prev == current 2718 * is still correct, but it can be moved to another cpu/rq. 2719 */ 2720 cpu = smp_processor_id(); 2721 rq = cpu_rq(cpu); 2722 } else 2723 raw_spin_unlock_irq(&rq->lock); 2724 2725 post_schedule(rq); 2726 2727 sched_preempt_enable_no_resched(); 2728 if (need_resched()) 2729 goto need_resched; 2730} 2731 2732static inline void sched_submit_work(struct task_struct *tsk) 2733{ 2734 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2735 return; 2736 /* 2737 * If we are going to sleep and we have plugged IO queued, 2738 * make sure to submit it to avoid deadlocks. 2739 */ 2740 if (blk_needs_flush_plug(tsk)) 2741 blk_schedule_flush_plug(tsk); 2742} 2743 2744asmlinkage void __sched schedule(void) 2745{ 2746 struct task_struct *tsk = current; 2747 2748 sched_submit_work(tsk); 2749 __schedule(); 2750} 2751EXPORT_SYMBOL(schedule); 2752 2753#ifdef CONFIG_CONTEXT_TRACKING 2754asmlinkage void __sched schedule_user(void) 2755{ 2756 /* 2757 * If we come here after a random call to set_need_resched(), 2758 * or we have been woken up remotely but the IPI has not yet arrived, 2759 * we haven't yet exited the RCU idle mode. Do it here manually until 2760 * we find a better solution. 2761 */ 2762 user_exit(); 2763 schedule(); 2764 user_enter(); 2765} 2766#endif 2767 2768/** 2769 * schedule_preempt_disabled - called with preemption disabled 2770 * 2771 * Returns with preemption disabled. Note: preempt_count must be 1 2772 */ 2773void __sched schedule_preempt_disabled(void) 2774{ 2775 sched_preempt_enable_no_resched(); 2776 schedule(); 2777 preempt_disable(); 2778} 2779 2780#ifdef CONFIG_PREEMPT 2781/* 2782 * this is the entry point to schedule() from in-kernel preemption 2783 * off of preempt_enable. Kernel preemptions off return from interrupt 2784 * occur there and call schedule directly. 2785 */ 2786asmlinkage void __sched notrace preempt_schedule(void) 2787{ 2788 /* 2789 * If there is a non-zero preempt_count or interrupts are disabled, 2790 * we do not want to preempt the current task. Just return.. 2791 */ 2792 if (likely(!preemptible())) 2793 return; 2794 2795 do { 2796 __preempt_count_add(PREEMPT_ACTIVE); 2797 __schedule(); 2798 __preempt_count_sub(PREEMPT_ACTIVE); 2799 2800 /* 2801 * Check again in case we missed a preemption opportunity 2802 * between schedule and now. 2803 */ 2804 barrier(); 2805 } while (need_resched()); 2806} 2807EXPORT_SYMBOL(preempt_schedule); 2808#endif /* CONFIG_PREEMPT */ 2809 2810/* 2811 * this is the entry point to schedule() from kernel preemption 2812 * off of irq context. 2813 * Note, that this is called and return with irqs disabled. This will 2814 * protect us against recursive calling from irq. 2815 */ 2816asmlinkage void __sched preempt_schedule_irq(void) 2817{ 2818 enum ctx_state prev_state; 2819 2820 /* Catch callers which need to be fixed */ 2821 BUG_ON(preempt_count() || !irqs_disabled()); 2822 2823 prev_state = exception_enter(); 2824 2825 do { 2826 __preempt_count_add(PREEMPT_ACTIVE); 2827 local_irq_enable(); 2828 __schedule(); 2829 local_irq_disable(); 2830 __preempt_count_sub(PREEMPT_ACTIVE); 2831 2832 /* 2833 * Check again in case we missed a preemption opportunity 2834 * between schedule and now. 2835 */ 2836 barrier(); 2837 } while (need_resched()); 2838 2839 exception_exit(prev_state); 2840} 2841 2842int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2843 void *key) 2844{ 2845 return try_to_wake_up(curr->private, mode, wake_flags); 2846} 2847EXPORT_SYMBOL(default_wake_function); 2848 2849static long __sched 2850sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2851{ 2852 unsigned long flags; 2853 wait_queue_t wait; 2854 2855 init_waitqueue_entry(&wait, current); 2856 2857 __set_current_state(state); 2858 2859 spin_lock_irqsave(&q->lock, flags); 2860 __add_wait_queue(q, &wait); 2861 spin_unlock(&q->lock); 2862 timeout = schedule_timeout(timeout); 2863 spin_lock_irq(&q->lock); 2864 __remove_wait_queue(q, &wait); 2865 spin_unlock_irqrestore(&q->lock, flags); 2866 2867 return timeout; 2868} 2869 2870void __sched interruptible_sleep_on(wait_queue_head_t *q) 2871{ 2872 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2873} 2874EXPORT_SYMBOL(interruptible_sleep_on); 2875 2876long __sched 2877interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2878{ 2879 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2880} 2881EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2882 2883void __sched sleep_on(wait_queue_head_t *q) 2884{ 2885 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2886} 2887EXPORT_SYMBOL(sleep_on); 2888 2889long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2890{ 2891 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2892} 2893EXPORT_SYMBOL(sleep_on_timeout); 2894 2895#ifdef CONFIG_RT_MUTEXES 2896 2897/* 2898 * rt_mutex_setprio - set the current priority of a task 2899 * @p: task 2900 * @prio: prio value (kernel-internal form) 2901 * 2902 * This function changes the 'effective' priority of a task. It does 2903 * not touch ->normal_prio like __setscheduler(). 2904 * 2905 * Used by the rt_mutex code to implement priority inheritance logic. 2906 */ 2907void rt_mutex_setprio(struct task_struct *p, int prio) 2908{ 2909 int oldprio, on_rq, running, enqueue_flag = 0; 2910 struct rq *rq; 2911 const struct sched_class *prev_class; 2912 2913 BUG_ON(prio > MAX_PRIO); 2914 2915 rq = __task_rq_lock(p); 2916 2917 /* 2918 * Idle task boosting is a nono in general. There is one 2919 * exception, when PREEMPT_RT and NOHZ is active: 2920 * 2921 * The idle task calls get_next_timer_interrupt() and holds 2922 * the timer wheel base->lock on the CPU and another CPU wants 2923 * to access the timer (probably to cancel it). We can safely 2924 * ignore the boosting request, as the idle CPU runs this code 2925 * with interrupts disabled and will complete the lock 2926 * protected section without being interrupted. So there is no 2927 * real need to boost. 2928 */ 2929 if (unlikely(p == rq->idle)) { 2930 WARN_ON(p != rq->curr); 2931 WARN_ON(p->pi_blocked_on); 2932 goto out_unlock; 2933 } 2934 2935 trace_sched_pi_setprio(p, prio); 2936 p->pi_top_task = rt_mutex_get_top_task(p); 2937 oldprio = p->prio; 2938 prev_class = p->sched_class; 2939 on_rq = p->on_rq; 2940 running = task_current(rq, p); 2941 if (on_rq) 2942 dequeue_task(rq, p, 0); 2943 if (running) 2944 p->sched_class->put_prev_task(rq, p); 2945 2946 /* 2947 * Boosting condition are: 2948 * 1. -rt task is running and holds mutex A 2949 * --> -dl task blocks on mutex A 2950 * 2951 * 2. -dl task is running and holds mutex A 2952 * --> -dl task blocks on mutex A and could preempt the 2953 * running task 2954 */ 2955 if (dl_prio(prio)) { 2956 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2957 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2958 p->dl.dl_boosted = 1; 2959 p->dl.dl_throttled = 0; 2960 enqueue_flag = ENQUEUE_REPLENISH; 2961 } else 2962 p->dl.dl_boosted = 0; 2963 p->sched_class = &dl_sched_class; 2964 } else if (rt_prio(prio)) { 2965 if (dl_prio(oldprio)) 2966 p->dl.dl_boosted = 0; 2967 if (oldprio < prio) 2968 enqueue_flag = ENQUEUE_HEAD; 2969 p->sched_class = &rt_sched_class; 2970 } else { 2971 if (dl_prio(oldprio)) 2972 p->dl.dl_boosted = 0; 2973 p->sched_class = &fair_sched_class; 2974 } 2975 2976 p->prio = prio; 2977 2978 if (running) 2979 p->sched_class->set_curr_task(rq); 2980 if (on_rq) 2981 enqueue_task(rq, p, enqueue_flag); 2982 2983 check_class_changed(rq, p, prev_class, oldprio); 2984out_unlock: 2985 __task_rq_unlock(rq); 2986} 2987#endif 2988 2989void set_user_nice(struct task_struct *p, long nice) 2990{ 2991 int old_prio, delta, on_rq; 2992 unsigned long flags; 2993 struct rq *rq; 2994 2995 if (task_nice(p) == nice || nice < -20 || nice > 19) 2996 return; 2997 /* 2998 * We have to be careful, if called from sys_setpriority(), 2999 * the task might be in the middle of scheduling on another CPU. 3000 */ 3001 rq = task_rq_lock(p, &flags); 3002 /* 3003 * The RT priorities are set via sched_setscheduler(), but we still 3004 * allow the 'normal' nice value to be set - but as expected 3005 * it wont have any effect on scheduling until the task is 3006 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3007 */ 3008 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3009 p->static_prio = NICE_TO_PRIO(nice); 3010 goto out_unlock; 3011 } 3012 on_rq = p->on_rq; 3013 if (on_rq) 3014 dequeue_task(rq, p, 0); 3015 3016 p->static_prio = NICE_TO_PRIO(nice); 3017 set_load_weight(p); 3018 old_prio = p->prio; 3019 p->prio = effective_prio(p); 3020 delta = p->prio - old_prio; 3021 3022 if (on_rq) { 3023 enqueue_task(rq, p, 0); 3024 /* 3025 * If the task increased its priority or is running and 3026 * lowered its priority, then reschedule its CPU: 3027 */ 3028 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3029 resched_task(rq->curr); 3030 } 3031out_unlock: 3032 task_rq_unlock(rq, p, &flags); 3033} 3034EXPORT_SYMBOL(set_user_nice); 3035 3036/* 3037 * can_nice - check if a task can reduce its nice value 3038 * @p: task 3039 * @nice: nice value 3040 */ 3041int can_nice(const struct task_struct *p, const int nice) 3042{ 3043 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3044 int nice_rlim = 20 - nice; 3045 3046 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3047 capable(CAP_SYS_NICE)); 3048} 3049 3050#ifdef __ARCH_WANT_SYS_NICE 3051 3052/* 3053 * sys_nice - change the priority of the current process. 3054 * @increment: priority increment 3055 * 3056 * sys_setpriority is a more generic, but much slower function that 3057 * does similar things. 3058 */ 3059SYSCALL_DEFINE1(nice, int, increment) 3060{ 3061 long nice, retval; 3062 3063 /* 3064 * Setpriority might change our priority at the same moment. 3065 * We don't have to worry. Conceptually one call occurs first 3066 * and we have a single winner. 3067 */ 3068 if (increment < -40) 3069 increment = -40; 3070 if (increment > 40) 3071 increment = 40; 3072 3073 nice = task_nice(current) + increment; 3074 if (nice < -20) 3075 nice = -20; 3076 if (nice > 19) 3077 nice = 19; 3078 3079 if (increment < 0 && !can_nice(current, nice)) 3080 return -EPERM; 3081 3082 retval = security_task_setnice(current, nice); 3083 if (retval) 3084 return retval; 3085 3086 set_user_nice(current, nice); 3087 return 0; 3088} 3089 3090#endif 3091 3092/** 3093 * task_prio - return the priority value of a given task. 3094 * @p: the task in question. 3095 * 3096 * Return: The priority value as seen by users in /proc. 3097 * RT tasks are offset by -200. Normal tasks are centered 3098 * around 0, value goes from -16 to +15. 3099 */ 3100int task_prio(const struct task_struct *p) 3101{ 3102 return p->prio - MAX_RT_PRIO; 3103} 3104 3105/** 3106 * idle_cpu - is a given cpu idle currently? 3107 * @cpu: the processor in question. 3108 * 3109 * Return: 1 if the CPU is currently idle. 0 otherwise. 3110 */ 3111int idle_cpu(int cpu) 3112{ 3113 struct rq *rq = cpu_rq(cpu); 3114 3115 if (rq->curr != rq->idle) 3116 return 0; 3117 3118 if (rq->nr_running) 3119 return 0; 3120 3121#ifdef CONFIG_SMP 3122 if (!llist_empty(&rq->wake_list)) 3123 return 0; 3124#endif 3125 3126 return 1; 3127} 3128 3129/** 3130 * idle_task - return the idle task for a given cpu. 3131 * @cpu: the processor in question. 3132 * 3133 * Return: The idle task for the cpu @cpu. 3134 */ 3135struct task_struct *idle_task(int cpu) 3136{ 3137 return cpu_rq(cpu)->idle; 3138} 3139 3140/** 3141 * find_process_by_pid - find a process with a matching PID value. 3142 * @pid: the pid in question. 3143 * 3144 * The task of @pid, if found. %NULL otherwise. 3145 */ 3146static struct task_struct *find_process_by_pid(pid_t pid) 3147{ 3148 return pid ? find_task_by_vpid(pid) : current; 3149} 3150 3151/* 3152 * This function initializes the sched_dl_entity of a newly becoming 3153 * SCHED_DEADLINE task. 3154 * 3155 * Only the static values are considered here, the actual runtime and the 3156 * absolute deadline will be properly calculated when the task is enqueued 3157 * for the first time with its new policy. 3158 */ 3159static void 3160__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3161{ 3162 struct sched_dl_entity *dl_se = &p->dl; 3163 3164 init_dl_task_timer(dl_se); 3165 dl_se->dl_runtime = attr->sched_runtime; 3166 dl_se->dl_deadline = attr->sched_deadline; 3167 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3168 dl_se->flags = attr->sched_flags; 3169 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3170 dl_se->dl_throttled = 0; 3171 dl_se->dl_new = 1; 3172} 3173 3174/* Actually do priority change: must hold pi & rq lock. */ 3175static void __setscheduler(struct rq *rq, struct task_struct *p, 3176 const struct sched_attr *attr) 3177{ 3178 int policy = attr->sched_policy; 3179 3180 if (policy == -1) /* setparam */ 3181 policy = p->policy; 3182 3183 p->policy = policy; 3184 3185 if (dl_policy(policy)) 3186 __setparam_dl(p, attr); 3187 else if (fair_policy(policy)) 3188 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3189 3190 /* 3191 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3192 * !rt_policy. Always setting this ensures that things like 3193 * getparam()/getattr() don't report silly values for !rt tasks. 3194 */ 3195 p->rt_priority = attr->sched_priority; 3196 3197 p->normal_prio = normal_prio(p); 3198 p->prio = rt_mutex_getprio(p); 3199 3200 if (dl_prio(p->prio)) 3201 p->sched_class = &dl_sched_class; 3202 else if (rt_prio(p->prio)) 3203 p->sched_class = &rt_sched_class; 3204 else 3205 p->sched_class = &fair_sched_class; 3206 3207 set_load_weight(p); 3208} 3209 3210static void 3211__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3212{ 3213 struct sched_dl_entity *dl_se = &p->dl; 3214 3215 attr->sched_priority = p->rt_priority; 3216 attr->sched_runtime = dl_se->dl_runtime; 3217 attr->sched_deadline = dl_se->dl_deadline; 3218 attr->sched_period = dl_se->dl_period; 3219 attr->sched_flags = dl_se->flags; 3220} 3221 3222/* 3223 * This function validates the new parameters of a -deadline task. 3224 * We ask for the deadline not being zero, and greater or equal 3225 * than the runtime, as well as the period of being zero or 3226 * greater than deadline. Furthermore, we have to be sure that 3227 * user parameters are above the internal resolution (1us); we 3228 * check sched_runtime only since it is always the smaller one. 3229 */ 3230static bool 3231__checkparam_dl(const struct sched_attr *attr) 3232{ 3233 return attr && attr->sched_deadline != 0 && 3234 (attr->sched_period == 0 || 3235 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3236 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3237 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3238} 3239 3240/* 3241 * check the target process has a UID that matches the current process's 3242 */ 3243static bool check_same_owner(struct task_struct *p) 3244{ 3245 const struct cred *cred = current_cred(), *pcred; 3246 bool match; 3247 3248 rcu_read_lock(); 3249 pcred = __task_cred(p); 3250 match = (uid_eq(cred->euid, pcred->euid) || 3251 uid_eq(cred->euid, pcred->uid)); 3252 rcu_read_unlock(); 3253 return match; 3254} 3255 3256static int __sched_setscheduler(struct task_struct *p, 3257 const struct sched_attr *attr, 3258 bool user) 3259{ 3260 int retval, oldprio, oldpolicy = -1, on_rq, running; 3261 int policy = attr->sched_policy; 3262 unsigned long flags; 3263 const struct sched_class *prev_class; 3264 struct rq *rq; 3265 int reset_on_fork; 3266 3267 /* may grab non-irq protected spin_locks */ 3268 BUG_ON(in_interrupt()); 3269recheck: 3270 /* double check policy once rq lock held */ 3271 if (policy < 0) { 3272 reset_on_fork = p->sched_reset_on_fork; 3273 policy = oldpolicy = p->policy; 3274 } else { 3275 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3276 3277 if (policy != SCHED_DEADLINE && 3278 policy != SCHED_FIFO && policy != SCHED_RR && 3279 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3280 policy != SCHED_IDLE) 3281 return -EINVAL; 3282 } 3283 3284 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3285 return -EINVAL; 3286 3287 /* 3288 * Valid priorities for SCHED_FIFO and SCHED_RR are 3289 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3290 * SCHED_BATCH and SCHED_IDLE is 0. 3291 */ 3292 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3293 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3294 return -EINVAL; 3295 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3296 (rt_policy(policy) != (attr->sched_priority != 0))) 3297 return -EINVAL; 3298 3299 /* 3300 * Allow unprivileged RT tasks to decrease priority: 3301 */ 3302 if (user && !capable(CAP_SYS_NICE)) { 3303 if (fair_policy(policy)) { 3304 if (attr->sched_nice < task_nice(p) && 3305 !can_nice(p, attr->sched_nice)) 3306 return -EPERM; 3307 } 3308 3309 if (rt_policy(policy)) { 3310 unsigned long rlim_rtprio = 3311 task_rlimit(p, RLIMIT_RTPRIO); 3312 3313 /* can't set/change the rt policy */ 3314 if (policy != p->policy && !rlim_rtprio) 3315 return -EPERM; 3316 3317 /* can't increase priority */ 3318 if (attr->sched_priority > p->rt_priority && 3319 attr->sched_priority > rlim_rtprio) 3320 return -EPERM; 3321 } 3322 3323 /* 3324 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3325 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3326 */ 3327 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3328 if (!can_nice(p, task_nice(p))) 3329 return -EPERM; 3330 } 3331 3332 /* can't change other user's priorities */ 3333 if (!check_same_owner(p)) 3334 return -EPERM; 3335 3336 /* Normal users shall not reset the sched_reset_on_fork flag */ 3337 if (p->sched_reset_on_fork && !reset_on_fork) 3338 return -EPERM; 3339 } 3340 3341 if (user) { 3342 retval = security_task_setscheduler(p); 3343 if (retval) 3344 return retval; 3345 } 3346 3347 /* 3348 * make sure no PI-waiters arrive (or leave) while we are 3349 * changing the priority of the task: 3350 * 3351 * To be able to change p->policy safely, the appropriate 3352 * runqueue lock must be held. 3353 */ 3354 rq = task_rq_lock(p, &flags); 3355 3356 /* 3357 * Changing the policy of the stop threads its a very bad idea 3358 */ 3359 if (p == rq->stop) { 3360 task_rq_unlock(rq, p, &flags); 3361 return -EINVAL; 3362 } 3363 3364 /* 3365 * If not changing anything there's no need to proceed further, 3366 * but store a possible modification of reset_on_fork. 3367 */ 3368 if (unlikely(policy == p->policy)) { 3369 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3370 goto change; 3371 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3372 goto change; 3373 if (dl_policy(policy)) 3374 goto change; 3375 3376 p->sched_reset_on_fork = reset_on_fork; 3377 task_rq_unlock(rq, p, &flags); 3378 return 0; 3379 } 3380change: 3381 3382 if (user) { 3383#ifdef CONFIG_RT_GROUP_SCHED 3384 /* 3385 * Do not allow realtime tasks into groups that have no runtime 3386 * assigned. 3387 */ 3388 if (rt_bandwidth_enabled() && rt_policy(policy) && 3389 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3390 !task_group_is_autogroup(task_group(p))) { 3391 task_rq_unlock(rq, p, &flags); 3392 return -EPERM; 3393 } 3394#endif 3395#ifdef CONFIG_SMP 3396 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3397 cpumask_t *span = rq->rd->span; 3398 3399 /* 3400 * Don't allow tasks with an affinity mask smaller than 3401 * the entire root_domain to become SCHED_DEADLINE. We 3402 * will also fail if there's no bandwidth available. 3403 */ 3404 if (!cpumask_subset(span, &p->cpus_allowed) || 3405 rq->rd->dl_bw.bw == 0) { 3406 task_rq_unlock(rq, p, &flags); 3407 return -EPERM; 3408 } 3409 } 3410#endif 3411 } 3412 3413 /* recheck policy now with rq lock held */ 3414 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3415 policy = oldpolicy = -1; 3416 task_rq_unlock(rq, p, &flags); 3417 goto recheck; 3418 } 3419 3420 /* 3421 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3422 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3423 * is available. 3424 */ 3425 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3426 task_rq_unlock(rq, p, &flags); 3427 return -EBUSY; 3428 } 3429 3430 on_rq = p->on_rq; 3431 running = task_current(rq, p); 3432 if (on_rq) 3433 dequeue_task(rq, p, 0); 3434 if (running) 3435 p->sched_class->put_prev_task(rq, p); 3436 3437 p->sched_reset_on_fork = reset_on_fork; 3438 3439 oldprio = p->prio; 3440 prev_class = p->sched_class; 3441 __setscheduler(rq, p, attr); 3442 3443 if (running) 3444 p->sched_class->set_curr_task(rq); 3445 if (on_rq) { 3446 /* 3447 * We enqueue to tail when the priority of a task is 3448 * increased (user space view). 3449 */ 3450 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3451 } 3452 3453 check_class_changed(rq, p, prev_class, oldprio); 3454 task_rq_unlock(rq, p, &flags); 3455 3456 rt_mutex_adjust_pi(p); 3457 3458 return 0; 3459} 3460 3461static int _sched_setscheduler(struct task_struct *p, int policy, 3462 const struct sched_param *param, bool check) 3463{ 3464 struct sched_attr attr = { 3465 .sched_policy = policy, 3466 .sched_priority = param->sched_priority, 3467 .sched_nice = PRIO_TO_NICE(p->static_prio), 3468 }; 3469 3470 /* 3471 * Fixup the legacy SCHED_RESET_ON_FORK hack 3472 */ 3473 if (policy & SCHED_RESET_ON_FORK) { 3474 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3475 policy &= ~SCHED_RESET_ON_FORK; 3476 attr.sched_policy = policy; 3477 } 3478 3479 return __sched_setscheduler(p, &attr, check); 3480} 3481/** 3482 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3483 * @p: the task in question. 3484 * @policy: new policy. 3485 * @param: structure containing the new RT priority. 3486 * 3487 * Return: 0 on success. An error code otherwise. 3488 * 3489 * NOTE that the task may be already dead. 3490 */ 3491int sched_setscheduler(struct task_struct *p, int policy, 3492 const struct sched_param *param) 3493{ 3494 return _sched_setscheduler(p, policy, param, true); 3495} 3496EXPORT_SYMBOL_GPL(sched_setscheduler); 3497 3498int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3499{ 3500 return __sched_setscheduler(p, attr, true); 3501} 3502EXPORT_SYMBOL_GPL(sched_setattr); 3503 3504/** 3505 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3506 * @p: the task in question. 3507 * @policy: new policy. 3508 * @param: structure containing the new RT priority. 3509 * 3510 * Just like sched_setscheduler, only don't bother checking if the 3511 * current context has permission. For example, this is needed in 3512 * stop_machine(): we create temporary high priority worker threads, 3513 * but our caller might not have that capability. 3514 * 3515 * Return: 0 on success. An error code otherwise. 3516 */ 3517int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3518 const struct sched_param *param) 3519{ 3520 return _sched_setscheduler(p, policy, param, false); 3521} 3522 3523static int 3524do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3525{ 3526 struct sched_param lparam; 3527 struct task_struct *p; 3528 int retval; 3529 3530 if (!param || pid < 0) 3531 return -EINVAL; 3532 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3533 return -EFAULT; 3534 3535 rcu_read_lock(); 3536 retval = -ESRCH; 3537 p = find_process_by_pid(pid); 3538 if (p != NULL) 3539 retval = sched_setscheduler(p, policy, &lparam); 3540 rcu_read_unlock(); 3541 3542 return retval; 3543} 3544 3545/* 3546 * Mimics kernel/events/core.c perf_copy_attr(). 3547 */ 3548static int sched_copy_attr(struct sched_attr __user *uattr, 3549 struct sched_attr *attr) 3550{ 3551 u32 size; 3552 int ret; 3553 3554 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3555 return -EFAULT; 3556 3557 /* 3558 * zero the full structure, so that a short copy will be nice. 3559 */ 3560 memset(attr, 0, sizeof(*attr)); 3561 3562 ret = get_user(size, &uattr->size); 3563 if (ret) 3564 return ret; 3565 3566 if (size > PAGE_SIZE) /* silly large */ 3567 goto err_size; 3568 3569 if (!size) /* abi compat */ 3570 size = SCHED_ATTR_SIZE_VER0; 3571 3572 if (size < SCHED_ATTR_SIZE_VER0) 3573 goto err_size; 3574 3575 /* 3576 * If we're handed a bigger struct than we know of, 3577 * ensure all the unknown bits are 0 - i.e. new 3578 * user-space does not rely on any kernel feature 3579 * extensions we dont know about yet. 3580 */ 3581 if (size > sizeof(*attr)) { 3582 unsigned char __user *addr; 3583 unsigned char __user *end; 3584 unsigned char val; 3585 3586 addr = (void __user *)uattr + sizeof(*attr); 3587 end = (void __user *)uattr + size; 3588 3589 for (; addr < end; addr++) { 3590 ret = get_user(val, addr); 3591 if (ret) 3592 return ret; 3593 if (val) 3594 goto err_size; 3595 } 3596 size = sizeof(*attr); 3597 } 3598 3599 ret = copy_from_user(attr, uattr, size); 3600 if (ret) 3601 return -EFAULT; 3602 3603 /* 3604 * XXX: do we want to be lenient like existing syscalls; or do we want 3605 * to be strict and return an error on out-of-bounds values? 3606 */ 3607 attr->sched_nice = clamp(attr->sched_nice, -20, 19); 3608 3609out: 3610 return ret; 3611 3612err_size: 3613 put_user(sizeof(*attr), &uattr->size); 3614 ret = -E2BIG; 3615 goto out; 3616} 3617 3618/** 3619 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3620 * @pid: the pid in question. 3621 * @policy: new policy. 3622 * @param: structure containing the new RT priority. 3623 * 3624 * Return: 0 on success. An error code otherwise. 3625 */ 3626SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3627 struct sched_param __user *, param) 3628{ 3629 /* negative values for policy are not valid */ 3630 if (policy < 0) 3631 return -EINVAL; 3632 3633 return do_sched_setscheduler(pid, policy, param); 3634} 3635 3636/** 3637 * sys_sched_setparam - set/change the RT priority of a thread 3638 * @pid: the pid in question. 3639 * @param: structure containing the new RT priority. 3640 * 3641 * Return: 0 on success. An error code otherwise. 3642 */ 3643SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3644{ 3645 return do_sched_setscheduler(pid, -1, param); 3646} 3647 3648/** 3649 * sys_sched_setattr - same as above, but with extended sched_attr 3650 * @pid: the pid in question. 3651 * @uattr: structure containing the extended parameters. 3652 */ 3653SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr) 3654{ 3655 struct sched_attr attr; 3656 struct task_struct *p; 3657 int retval; 3658 3659 if (!uattr || pid < 0) 3660 return -EINVAL; 3661 3662 if (sched_copy_attr(uattr, &attr)) 3663 return -EFAULT; 3664 3665 rcu_read_lock(); 3666 retval = -ESRCH; 3667 p = find_process_by_pid(pid); 3668 if (p != NULL) 3669 retval = sched_setattr(p, &attr); 3670 rcu_read_unlock(); 3671 3672 return retval; 3673} 3674 3675/** 3676 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3677 * @pid: the pid in question. 3678 * 3679 * Return: On success, the policy of the thread. Otherwise, a negative error 3680 * code. 3681 */ 3682SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3683{ 3684 struct task_struct *p; 3685 int retval; 3686 3687 if (pid < 0) 3688 return -EINVAL; 3689 3690 retval = -ESRCH; 3691 rcu_read_lock(); 3692 p = find_process_by_pid(pid); 3693 if (p) { 3694 retval = security_task_getscheduler(p); 3695 if (!retval) 3696 retval = p->policy 3697 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3698 } 3699 rcu_read_unlock(); 3700 return retval; 3701} 3702 3703/** 3704 * sys_sched_getparam - get the RT priority of a thread 3705 * @pid: the pid in question. 3706 * @param: structure containing the RT priority. 3707 * 3708 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3709 * code. 3710 */ 3711SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3712{ 3713 struct sched_param lp; 3714 struct task_struct *p; 3715 int retval; 3716 3717 if (!param || pid < 0) 3718 return -EINVAL; 3719 3720 rcu_read_lock(); 3721 p = find_process_by_pid(pid); 3722 retval = -ESRCH; 3723 if (!p) 3724 goto out_unlock; 3725 3726 retval = security_task_getscheduler(p); 3727 if (retval) 3728 goto out_unlock; 3729 3730 if (task_has_dl_policy(p)) { 3731 retval = -EINVAL; 3732 goto out_unlock; 3733 } 3734 lp.sched_priority = p->rt_priority; 3735 rcu_read_unlock(); 3736 3737 /* 3738 * This one might sleep, we cannot do it with a spinlock held ... 3739 */ 3740 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3741 3742 return retval; 3743 3744out_unlock: 3745 rcu_read_unlock(); 3746 return retval; 3747} 3748 3749static int sched_read_attr(struct sched_attr __user *uattr, 3750 struct sched_attr *attr, 3751 unsigned int usize) 3752{ 3753 int ret; 3754 3755 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3756 return -EFAULT; 3757 3758 /* 3759 * If we're handed a smaller struct than we know of, 3760 * ensure all the unknown bits are 0 - i.e. old 3761 * user-space does not get uncomplete information. 3762 */ 3763 if (usize < sizeof(*attr)) { 3764 unsigned char *addr; 3765 unsigned char *end; 3766 3767 addr = (void *)attr + usize; 3768 end = (void *)attr + sizeof(*attr); 3769 3770 for (; addr < end; addr++) { 3771 if (*addr) 3772 goto err_size; 3773 } 3774 3775 attr->size = usize; 3776 } 3777 3778 ret = copy_to_user(uattr, attr, usize); 3779 if (ret) 3780 return -EFAULT; 3781 3782out: 3783 return ret; 3784 3785err_size: 3786 ret = -E2BIG; 3787 goto out; 3788} 3789 3790/** 3791 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3792 * @pid: the pid in question. 3793 * @uattr: structure containing the extended parameters. 3794 * @size: sizeof(attr) for fwd/bwd comp. 3795 */ 3796SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3797 unsigned int, size) 3798{ 3799 struct sched_attr attr = { 3800 .size = sizeof(struct sched_attr), 3801 }; 3802 struct task_struct *p; 3803 int retval; 3804 3805 if (!uattr || pid < 0 || size > PAGE_SIZE || 3806 size < SCHED_ATTR_SIZE_VER0) 3807 return -EINVAL; 3808 3809 rcu_read_lock(); 3810 p = find_process_by_pid(pid); 3811 retval = -ESRCH; 3812 if (!p) 3813 goto out_unlock; 3814 3815 retval = security_task_getscheduler(p); 3816 if (retval) 3817 goto out_unlock; 3818 3819 attr.sched_policy = p->policy; 3820 if (p->sched_reset_on_fork) 3821 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3822 if (task_has_dl_policy(p)) 3823 __getparam_dl(p, &attr); 3824 else if (task_has_rt_policy(p)) 3825 attr.sched_priority = p->rt_priority; 3826 else 3827 attr.sched_nice = task_nice(p); 3828 3829 rcu_read_unlock(); 3830 3831 retval = sched_read_attr(uattr, &attr, size); 3832 return retval; 3833 3834out_unlock: 3835 rcu_read_unlock(); 3836 return retval; 3837} 3838 3839long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3840{ 3841 cpumask_var_t cpus_allowed, new_mask; 3842 struct task_struct *p; 3843 int retval; 3844 3845 rcu_read_lock(); 3846 3847 p = find_process_by_pid(pid); 3848 if (!p) { 3849 rcu_read_unlock(); 3850 return -ESRCH; 3851 } 3852 3853 /* Prevent p going away */ 3854 get_task_struct(p); 3855 rcu_read_unlock(); 3856 3857 if (p->flags & PF_NO_SETAFFINITY) { 3858 retval = -EINVAL; 3859 goto out_put_task; 3860 } 3861 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3862 retval = -ENOMEM; 3863 goto out_put_task; 3864 } 3865 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3866 retval = -ENOMEM; 3867 goto out_free_cpus_allowed; 3868 } 3869 retval = -EPERM; 3870 if (!check_same_owner(p)) { 3871 rcu_read_lock(); 3872 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3873 rcu_read_unlock(); 3874 goto out_unlock; 3875 } 3876 rcu_read_unlock(); 3877 } 3878 3879 retval = security_task_setscheduler(p); 3880 if (retval) 3881 goto out_unlock; 3882 3883 3884 cpuset_cpus_allowed(p, cpus_allowed); 3885 cpumask_and(new_mask, in_mask, cpus_allowed); 3886 3887 /* 3888 * Since bandwidth control happens on root_domain basis, 3889 * if admission test is enabled, we only admit -deadline 3890 * tasks allowed to run on all the CPUs in the task's 3891 * root_domain. 3892 */ 3893#ifdef CONFIG_SMP 3894 if (task_has_dl_policy(p)) { 3895 const struct cpumask *span = task_rq(p)->rd->span; 3896 3897 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3898 retval = -EBUSY; 3899 goto out_unlock; 3900 } 3901 } 3902#endif 3903again: 3904 retval = set_cpus_allowed_ptr(p, new_mask); 3905 3906 if (!retval) { 3907 cpuset_cpus_allowed(p, cpus_allowed); 3908 if (!cpumask_subset(new_mask, cpus_allowed)) { 3909 /* 3910 * We must have raced with a concurrent cpuset 3911 * update. Just reset the cpus_allowed to the 3912 * cpuset's cpus_allowed 3913 */ 3914 cpumask_copy(new_mask, cpus_allowed); 3915 goto again; 3916 } 3917 } 3918out_unlock: 3919 free_cpumask_var(new_mask); 3920out_free_cpus_allowed: 3921 free_cpumask_var(cpus_allowed); 3922out_put_task: 3923 put_task_struct(p); 3924 return retval; 3925} 3926 3927static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3928 struct cpumask *new_mask) 3929{ 3930 if (len < cpumask_size()) 3931 cpumask_clear(new_mask); 3932 else if (len > cpumask_size()) 3933 len = cpumask_size(); 3934 3935 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3936} 3937 3938/** 3939 * sys_sched_setaffinity - set the cpu affinity of a process 3940 * @pid: pid of the process 3941 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3942 * @user_mask_ptr: user-space pointer to the new cpu mask 3943 * 3944 * Return: 0 on success. An error code otherwise. 3945 */ 3946SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3947 unsigned long __user *, user_mask_ptr) 3948{ 3949 cpumask_var_t new_mask; 3950 int retval; 3951 3952 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3953 return -ENOMEM; 3954 3955 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3956 if (retval == 0) 3957 retval = sched_setaffinity(pid, new_mask); 3958 free_cpumask_var(new_mask); 3959 return retval; 3960} 3961 3962long sched_getaffinity(pid_t pid, struct cpumask *mask) 3963{ 3964 struct task_struct *p; 3965 unsigned long flags; 3966 int retval; 3967 3968 rcu_read_lock(); 3969 3970 retval = -ESRCH; 3971 p = find_process_by_pid(pid); 3972 if (!p) 3973 goto out_unlock; 3974 3975 retval = security_task_getscheduler(p); 3976 if (retval) 3977 goto out_unlock; 3978 3979 raw_spin_lock_irqsave(&p->pi_lock, flags); 3980 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3981 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3982 3983out_unlock: 3984 rcu_read_unlock(); 3985 3986 return retval; 3987} 3988 3989/** 3990 * sys_sched_getaffinity - get the cpu affinity of a process 3991 * @pid: pid of the process 3992 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3993 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3994 * 3995 * Return: 0 on success. An error code otherwise. 3996 */ 3997SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3998 unsigned long __user *, user_mask_ptr) 3999{ 4000 int ret; 4001 cpumask_var_t mask; 4002 4003 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4004 return -EINVAL; 4005 if (len & (sizeof(unsigned long)-1)) 4006 return -EINVAL; 4007 4008 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4009 return -ENOMEM; 4010 4011 ret = sched_getaffinity(pid, mask); 4012 if (ret == 0) { 4013 size_t retlen = min_t(size_t, len, cpumask_size()); 4014 4015 if (copy_to_user(user_mask_ptr, mask, retlen)) 4016 ret = -EFAULT; 4017 else 4018 ret = retlen; 4019 } 4020 free_cpumask_var(mask); 4021 4022 return ret; 4023} 4024 4025/** 4026 * sys_sched_yield - yield the current processor to other threads. 4027 * 4028 * This function yields the current CPU to other tasks. If there are no 4029 * other threads running on this CPU then this function will return. 4030 * 4031 * Return: 0. 4032 */ 4033SYSCALL_DEFINE0(sched_yield) 4034{ 4035 struct rq *rq = this_rq_lock(); 4036 4037 schedstat_inc(rq, yld_count); 4038 current->sched_class->yield_task(rq); 4039 4040 /* 4041 * Since we are going to call schedule() anyway, there's 4042 * no need to preempt or enable interrupts: 4043 */ 4044 __release(rq->lock); 4045 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4046 do_raw_spin_unlock(&rq->lock); 4047 sched_preempt_enable_no_resched(); 4048 4049 schedule(); 4050 4051 return 0; 4052} 4053 4054static void __cond_resched(void) 4055{ 4056 __preempt_count_add(PREEMPT_ACTIVE); 4057 __schedule(); 4058 __preempt_count_sub(PREEMPT_ACTIVE); 4059} 4060 4061int __sched _cond_resched(void) 4062{ 4063 if (should_resched()) { 4064 __cond_resched(); 4065 return 1; 4066 } 4067 return 0; 4068} 4069EXPORT_SYMBOL(_cond_resched); 4070 4071/* 4072 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4073 * call schedule, and on return reacquire the lock. 4074 * 4075 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4076 * operations here to prevent schedule() from being called twice (once via 4077 * spin_unlock(), once by hand). 4078 */ 4079int __cond_resched_lock(spinlock_t *lock) 4080{ 4081 int resched = should_resched(); 4082 int ret = 0; 4083 4084 lockdep_assert_held(lock); 4085 4086 if (spin_needbreak(lock) || resched) { 4087 spin_unlock(lock); 4088 if (resched) 4089 __cond_resched(); 4090 else 4091 cpu_relax(); 4092 ret = 1; 4093 spin_lock(lock); 4094 } 4095 return ret; 4096} 4097EXPORT_SYMBOL(__cond_resched_lock); 4098 4099int __sched __cond_resched_softirq(void) 4100{ 4101 BUG_ON(!in_softirq()); 4102 4103 if (should_resched()) { 4104 local_bh_enable(); 4105 __cond_resched(); 4106 local_bh_disable(); 4107 return 1; 4108 } 4109 return 0; 4110} 4111EXPORT_SYMBOL(__cond_resched_softirq); 4112 4113/** 4114 * yield - yield the current processor to other threads. 4115 * 4116 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4117 * 4118 * The scheduler is at all times free to pick the calling task as the most 4119 * eligible task to run, if removing the yield() call from your code breaks 4120 * it, its already broken. 4121 * 4122 * Typical broken usage is: 4123 * 4124 * while (!event) 4125 * yield(); 4126 * 4127 * where one assumes that yield() will let 'the other' process run that will 4128 * make event true. If the current task is a SCHED_FIFO task that will never 4129 * happen. Never use yield() as a progress guarantee!! 4130 * 4131 * If you want to use yield() to wait for something, use wait_event(). 4132 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4133 * If you still want to use yield(), do not! 4134 */ 4135void __sched yield(void) 4136{ 4137 set_current_state(TASK_RUNNING); 4138 sys_sched_yield(); 4139} 4140EXPORT_SYMBOL(yield); 4141 4142/** 4143 * yield_to - yield the current processor to another thread in 4144 * your thread group, or accelerate that thread toward the 4145 * processor it's on. 4146 * @p: target task 4147 * @preempt: whether task preemption is allowed or not 4148 * 4149 * It's the caller's job to ensure that the target task struct 4150 * can't go away on us before we can do any checks. 4151 * 4152 * Return: 4153 * true (>0) if we indeed boosted the target task. 4154 * false (0) if we failed to boost the target. 4155 * -ESRCH if there's no task to yield to. 4156 */ 4157bool __sched yield_to(struct task_struct *p, bool preempt) 4158{ 4159 struct task_struct *curr = current; 4160 struct rq *rq, *p_rq; 4161 unsigned long flags; 4162 int yielded = 0; 4163 4164 local_irq_save(flags); 4165 rq = this_rq(); 4166 4167again: 4168 p_rq = task_rq(p); 4169 /* 4170 * If we're the only runnable task on the rq and target rq also 4171 * has only one task, there's absolutely no point in yielding. 4172 */ 4173 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4174 yielded = -ESRCH; 4175 goto out_irq; 4176 } 4177 4178 double_rq_lock(rq, p_rq); 4179 if (task_rq(p) != p_rq) { 4180 double_rq_unlock(rq, p_rq); 4181 goto again; 4182 } 4183 4184 if (!curr->sched_class->yield_to_task) 4185 goto out_unlock; 4186 4187 if (curr->sched_class != p->sched_class) 4188 goto out_unlock; 4189 4190 if (task_running(p_rq, p) || p->state) 4191 goto out_unlock; 4192 4193 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4194 if (yielded) { 4195 schedstat_inc(rq, yld_count); 4196 /* 4197 * Make p's CPU reschedule; pick_next_entity takes care of 4198 * fairness. 4199 */ 4200 if (preempt && rq != p_rq) 4201 resched_task(p_rq->curr); 4202 } 4203 4204out_unlock: 4205 double_rq_unlock(rq, p_rq); 4206out_irq: 4207 local_irq_restore(flags); 4208 4209 if (yielded > 0) 4210 schedule(); 4211 4212 return yielded; 4213} 4214EXPORT_SYMBOL_GPL(yield_to); 4215 4216/* 4217 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4218 * that process accounting knows that this is a task in IO wait state. 4219 */ 4220void __sched io_schedule(void) 4221{ 4222 struct rq *rq = raw_rq(); 4223 4224 delayacct_blkio_start(); 4225 atomic_inc(&rq->nr_iowait); 4226 blk_flush_plug(current); 4227 current->in_iowait = 1; 4228 schedule(); 4229 current->in_iowait = 0; 4230 atomic_dec(&rq->nr_iowait); 4231 delayacct_blkio_end(); 4232} 4233EXPORT_SYMBOL(io_schedule); 4234 4235long __sched io_schedule_timeout(long timeout) 4236{ 4237 struct rq *rq = raw_rq(); 4238 long ret; 4239 4240 delayacct_blkio_start(); 4241 atomic_inc(&rq->nr_iowait); 4242 blk_flush_plug(current); 4243 current->in_iowait = 1; 4244 ret = schedule_timeout(timeout); 4245 current->in_iowait = 0; 4246 atomic_dec(&rq->nr_iowait); 4247 delayacct_blkio_end(); 4248 return ret; 4249} 4250 4251/** 4252 * sys_sched_get_priority_max - return maximum RT priority. 4253 * @policy: scheduling class. 4254 * 4255 * Return: On success, this syscall returns the maximum 4256 * rt_priority that can be used by a given scheduling class. 4257 * On failure, a negative error code is returned. 4258 */ 4259SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4260{ 4261 int ret = -EINVAL; 4262 4263 switch (policy) { 4264 case SCHED_FIFO: 4265 case SCHED_RR: 4266 ret = MAX_USER_RT_PRIO-1; 4267 break; 4268 case SCHED_DEADLINE: 4269 case SCHED_NORMAL: 4270 case SCHED_BATCH: 4271 case SCHED_IDLE: 4272 ret = 0; 4273 break; 4274 } 4275 return ret; 4276} 4277 4278/** 4279 * sys_sched_get_priority_min - return minimum RT priority. 4280 * @policy: scheduling class. 4281 * 4282 * Return: On success, this syscall returns the minimum 4283 * rt_priority that can be used by a given scheduling class. 4284 * On failure, a negative error code is returned. 4285 */ 4286SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4287{ 4288 int ret = -EINVAL; 4289 4290 switch (policy) { 4291 case SCHED_FIFO: 4292 case SCHED_RR: 4293 ret = 1; 4294 break; 4295 case SCHED_DEADLINE: 4296 case SCHED_NORMAL: 4297 case SCHED_BATCH: 4298 case SCHED_IDLE: 4299 ret = 0; 4300 } 4301 return ret; 4302} 4303 4304/** 4305 * sys_sched_rr_get_interval - return the default timeslice of a process. 4306 * @pid: pid of the process. 4307 * @interval: userspace pointer to the timeslice value. 4308 * 4309 * this syscall writes the default timeslice value of a given process 4310 * into the user-space timespec buffer. A value of '0' means infinity. 4311 * 4312 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4313 * an error code. 4314 */ 4315SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4316 struct timespec __user *, interval) 4317{ 4318 struct task_struct *p; 4319 unsigned int time_slice; 4320 unsigned long flags; 4321 struct rq *rq; 4322 int retval; 4323 struct timespec t; 4324 4325 if (pid < 0) 4326 return -EINVAL; 4327 4328 retval = -ESRCH; 4329 rcu_read_lock(); 4330 p = find_process_by_pid(pid); 4331 if (!p) 4332 goto out_unlock; 4333 4334 retval = security_task_getscheduler(p); 4335 if (retval) 4336 goto out_unlock; 4337 4338 rq = task_rq_lock(p, &flags); 4339 time_slice = 0; 4340 if (p->sched_class->get_rr_interval) 4341 time_slice = p->sched_class->get_rr_interval(rq, p); 4342 task_rq_unlock(rq, p, &flags); 4343 4344 rcu_read_unlock(); 4345 jiffies_to_timespec(time_slice, &t); 4346 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4347 return retval; 4348 4349out_unlock: 4350 rcu_read_unlock(); 4351 return retval; 4352} 4353 4354static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4355 4356void sched_show_task(struct task_struct *p) 4357{ 4358 unsigned long free = 0; 4359 int ppid; 4360 unsigned state; 4361 4362 state = p->state ? __ffs(p->state) + 1 : 0; 4363 printk(KERN_INFO "%-15.15s %c", p->comm, 4364 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4365#if BITS_PER_LONG == 32 4366 if (state == TASK_RUNNING) 4367 printk(KERN_CONT " running "); 4368 else 4369 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4370#else 4371 if (state == TASK_RUNNING) 4372 printk(KERN_CONT " running task "); 4373 else 4374 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4375#endif 4376#ifdef CONFIG_DEBUG_STACK_USAGE 4377 free = stack_not_used(p); 4378#endif 4379 rcu_read_lock(); 4380 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4381 rcu_read_unlock(); 4382 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4383 task_pid_nr(p), ppid, 4384 (unsigned long)task_thread_info(p)->flags); 4385 4386 print_worker_info(KERN_INFO, p); 4387 show_stack(p, NULL); 4388} 4389 4390void show_state_filter(unsigned long state_filter) 4391{ 4392 struct task_struct *g, *p; 4393 4394#if BITS_PER_LONG == 32 4395 printk(KERN_INFO 4396 " task PC stack pid father\n"); 4397#else 4398 printk(KERN_INFO 4399 " task PC stack pid father\n"); 4400#endif 4401 rcu_read_lock(); 4402 do_each_thread(g, p) { 4403 /* 4404 * reset the NMI-timeout, listing all files on a slow 4405 * console might take a lot of time: 4406 */ 4407 touch_nmi_watchdog(); 4408 if (!state_filter || (p->state & state_filter)) 4409 sched_show_task(p); 4410 } while_each_thread(g, p); 4411 4412 touch_all_softlockup_watchdogs(); 4413 4414#ifdef CONFIG_SCHED_DEBUG 4415 sysrq_sched_debug_show(); 4416#endif 4417 rcu_read_unlock(); 4418 /* 4419 * Only show locks if all tasks are dumped: 4420 */ 4421 if (!state_filter) 4422 debug_show_all_locks(); 4423} 4424 4425void init_idle_bootup_task(struct task_struct *idle) 4426{ 4427 idle->sched_class = &idle_sched_class; 4428} 4429 4430/** 4431 * init_idle - set up an idle thread for a given CPU 4432 * @idle: task in question 4433 * @cpu: cpu the idle task belongs to 4434 * 4435 * NOTE: this function does not set the idle thread's NEED_RESCHED 4436 * flag, to make booting more robust. 4437 */ 4438void init_idle(struct task_struct *idle, int cpu) 4439{ 4440 struct rq *rq = cpu_rq(cpu); 4441 unsigned long flags; 4442 4443 raw_spin_lock_irqsave(&rq->lock, flags); 4444 4445 __sched_fork(0, idle); 4446 idle->state = TASK_RUNNING; 4447 idle->se.exec_start = sched_clock(); 4448 4449 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4450 /* 4451 * We're having a chicken and egg problem, even though we are 4452 * holding rq->lock, the cpu isn't yet set to this cpu so the 4453 * lockdep check in task_group() will fail. 4454 * 4455 * Similar case to sched_fork(). / Alternatively we could 4456 * use task_rq_lock() here and obtain the other rq->lock. 4457 * 4458 * Silence PROVE_RCU 4459 */ 4460 rcu_read_lock(); 4461 __set_task_cpu(idle, cpu); 4462 rcu_read_unlock(); 4463 4464 rq->curr = rq->idle = idle; 4465 idle->on_rq = 1; 4466#if defined(CONFIG_SMP) 4467 idle->on_cpu = 1; 4468#endif 4469 raw_spin_unlock_irqrestore(&rq->lock, flags); 4470 4471 /* Set the preempt count _outside_ the spinlocks! */ 4472 init_idle_preempt_count(idle, cpu); 4473 4474 /* 4475 * The idle tasks have their own, simple scheduling class: 4476 */ 4477 idle->sched_class = &idle_sched_class; 4478 ftrace_graph_init_idle_task(idle, cpu); 4479 vtime_init_idle(idle, cpu); 4480#if defined(CONFIG_SMP) 4481 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4482#endif 4483} 4484 4485#ifdef CONFIG_SMP 4486void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4487{ 4488 if (p->sched_class && p->sched_class->set_cpus_allowed) 4489 p->sched_class->set_cpus_allowed(p, new_mask); 4490 4491 cpumask_copy(&p->cpus_allowed, new_mask); 4492 p->nr_cpus_allowed = cpumask_weight(new_mask); 4493} 4494 4495/* 4496 * This is how migration works: 4497 * 4498 * 1) we invoke migration_cpu_stop() on the target CPU using 4499 * stop_one_cpu(). 4500 * 2) stopper starts to run (implicitly forcing the migrated thread 4501 * off the CPU) 4502 * 3) it checks whether the migrated task is still in the wrong runqueue. 4503 * 4) if it's in the wrong runqueue then the migration thread removes 4504 * it and puts it into the right queue. 4505 * 5) stopper completes and stop_one_cpu() returns and the migration 4506 * is done. 4507 */ 4508 4509/* 4510 * Change a given task's CPU affinity. Migrate the thread to a 4511 * proper CPU and schedule it away if the CPU it's executing on 4512 * is removed from the allowed bitmask. 4513 * 4514 * NOTE: the caller must have a valid reference to the task, the 4515 * task must not exit() & deallocate itself prematurely. The 4516 * call is not atomic; no spinlocks may be held. 4517 */ 4518int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4519{ 4520 unsigned long flags; 4521 struct rq *rq; 4522 unsigned int dest_cpu; 4523 int ret = 0; 4524 4525 rq = task_rq_lock(p, &flags); 4526 4527 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4528 goto out; 4529 4530 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4531 ret = -EINVAL; 4532 goto out; 4533 } 4534 4535 do_set_cpus_allowed(p, new_mask); 4536 4537 /* Can the task run on the task's current CPU? If so, we're done */ 4538 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4539 goto out; 4540 4541 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4542 if (p->on_rq) { 4543 struct migration_arg arg = { p, dest_cpu }; 4544 /* Need help from migration thread: drop lock and wait. */ 4545 task_rq_unlock(rq, p, &flags); 4546 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4547 tlb_migrate_finish(p->mm); 4548 return 0; 4549 } 4550out: 4551 task_rq_unlock(rq, p, &flags); 4552 4553 return ret; 4554} 4555EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4556 4557/* 4558 * Move (not current) task off this cpu, onto dest cpu. We're doing 4559 * this because either it can't run here any more (set_cpus_allowed() 4560 * away from this CPU, or CPU going down), or because we're 4561 * attempting to rebalance this task on exec (sched_exec). 4562 * 4563 * So we race with normal scheduler movements, but that's OK, as long 4564 * as the task is no longer on this CPU. 4565 * 4566 * Returns non-zero if task was successfully migrated. 4567 */ 4568static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4569{ 4570 struct rq *rq_dest, *rq_src; 4571 int ret = 0; 4572 4573 if (unlikely(!cpu_active(dest_cpu))) 4574 return ret; 4575 4576 rq_src = cpu_rq(src_cpu); 4577 rq_dest = cpu_rq(dest_cpu); 4578 4579 raw_spin_lock(&p->pi_lock); 4580 double_rq_lock(rq_src, rq_dest); 4581 /* Already moved. */ 4582 if (task_cpu(p) != src_cpu) 4583 goto done; 4584 /* Affinity changed (again). */ 4585 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4586 goto fail; 4587 4588 /* 4589 * If we're not on a rq, the next wake-up will ensure we're 4590 * placed properly. 4591 */ 4592 if (p->on_rq) { 4593 dequeue_task(rq_src, p, 0); 4594 set_task_cpu(p, dest_cpu); 4595 enqueue_task(rq_dest, p, 0); 4596 check_preempt_curr(rq_dest, p, 0); 4597 } 4598done: 4599 ret = 1; 4600fail: 4601 double_rq_unlock(rq_src, rq_dest); 4602 raw_spin_unlock(&p->pi_lock); 4603 return ret; 4604} 4605 4606#ifdef CONFIG_NUMA_BALANCING 4607/* Migrate current task p to target_cpu */ 4608int migrate_task_to(struct task_struct *p, int target_cpu) 4609{ 4610 struct migration_arg arg = { p, target_cpu }; 4611 int curr_cpu = task_cpu(p); 4612 4613 if (curr_cpu == target_cpu) 4614 return 0; 4615 4616 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4617 return -EINVAL; 4618 4619 /* TODO: This is not properly updating schedstats */ 4620 4621 trace_sched_move_numa(p, curr_cpu, target_cpu); 4622 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4623} 4624 4625/* 4626 * Requeue a task on a given node and accurately track the number of NUMA 4627 * tasks on the runqueues 4628 */ 4629void sched_setnuma(struct task_struct *p, int nid) 4630{ 4631 struct rq *rq; 4632 unsigned long flags; 4633 bool on_rq, running; 4634 4635 rq = task_rq_lock(p, &flags); 4636 on_rq = p->on_rq; 4637 running = task_current(rq, p); 4638 4639 if (on_rq) 4640 dequeue_task(rq, p, 0); 4641 if (running) 4642 p->sched_class->put_prev_task(rq, p); 4643 4644 p->numa_preferred_nid = nid; 4645 4646 if (running) 4647 p->sched_class->set_curr_task(rq); 4648 if (on_rq) 4649 enqueue_task(rq, p, 0); 4650 task_rq_unlock(rq, p, &flags); 4651} 4652#endif 4653 4654/* 4655 * migration_cpu_stop - this will be executed by a highprio stopper thread 4656 * and performs thread migration by bumping thread off CPU then 4657 * 'pushing' onto another runqueue. 4658 */ 4659static int migration_cpu_stop(void *data) 4660{ 4661 struct migration_arg *arg = data; 4662 4663 /* 4664 * The original target cpu might have gone down and we might 4665 * be on another cpu but it doesn't matter. 4666 */ 4667 local_irq_disable(); 4668 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4669 local_irq_enable(); 4670 return 0; 4671} 4672 4673#ifdef CONFIG_HOTPLUG_CPU 4674 4675/* 4676 * Ensures that the idle task is using init_mm right before its cpu goes 4677 * offline. 4678 */ 4679void idle_task_exit(void) 4680{ 4681 struct mm_struct *mm = current->active_mm; 4682 4683 BUG_ON(cpu_online(smp_processor_id())); 4684 4685 if (mm != &init_mm) 4686 switch_mm(mm, &init_mm, current); 4687 mmdrop(mm); 4688} 4689 4690/* 4691 * Since this CPU is going 'away' for a while, fold any nr_active delta 4692 * we might have. Assumes we're called after migrate_tasks() so that the 4693 * nr_active count is stable. 4694 * 4695 * Also see the comment "Global load-average calculations". 4696 */ 4697static void calc_load_migrate(struct rq *rq) 4698{ 4699 long delta = calc_load_fold_active(rq); 4700 if (delta) 4701 atomic_long_add(delta, &calc_load_tasks); 4702} 4703 4704static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4705{ 4706} 4707 4708static const struct sched_class fake_sched_class = { 4709 .put_prev_task = put_prev_task_fake, 4710}; 4711 4712static struct task_struct fake_task = { 4713 /* 4714 * Avoid pull_{rt,dl}_task() 4715 */ 4716 .prio = MAX_PRIO + 1, 4717 .sched_class = &fake_sched_class, 4718}; 4719 4720/* 4721 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4722 * try_to_wake_up()->select_task_rq(). 4723 * 4724 * Called with rq->lock held even though we'er in stop_machine() and 4725 * there's no concurrency possible, we hold the required locks anyway 4726 * because of lock validation efforts. 4727 */ 4728static void migrate_tasks(unsigned int dead_cpu) 4729{ 4730 struct rq *rq = cpu_rq(dead_cpu); 4731 struct task_struct *next, *stop = rq->stop; 4732 int dest_cpu; 4733 4734 /* 4735 * Fudge the rq selection such that the below task selection loop 4736 * doesn't get stuck on the currently eligible stop task. 4737 * 4738 * We're currently inside stop_machine() and the rq is either stuck 4739 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4740 * either way we should never end up calling schedule() until we're 4741 * done here. 4742 */ 4743 rq->stop = NULL; 4744 4745 /* 4746 * put_prev_task() and pick_next_task() sched 4747 * class method both need to have an up-to-date 4748 * value of rq->clock[_task] 4749 */ 4750 update_rq_clock(rq); 4751 4752 for ( ; ; ) { 4753 /* 4754 * There's this thread running, bail when that's the only 4755 * remaining thread. 4756 */ 4757 if (rq->nr_running == 1) 4758 break; 4759 4760 next = pick_next_task(rq, &fake_task); 4761 BUG_ON(!next); 4762 next->sched_class->put_prev_task(rq, next); 4763 4764 /* Find suitable destination for @next, with force if needed. */ 4765 dest_cpu = select_fallback_rq(dead_cpu, next); 4766 raw_spin_unlock(&rq->lock); 4767 4768 __migrate_task(next, dead_cpu, dest_cpu); 4769 4770 raw_spin_lock(&rq->lock); 4771 } 4772 4773 rq->stop = stop; 4774} 4775 4776#endif /* CONFIG_HOTPLUG_CPU */ 4777 4778#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4779 4780static struct ctl_table sd_ctl_dir[] = { 4781 { 4782 .procname = "sched_domain", 4783 .mode = 0555, 4784 }, 4785 {} 4786}; 4787 4788static struct ctl_table sd_ctl_root[] = { 4789 { 4790 .procname = "kernel", 4791 .mode = 0555, 4792 .child = sd_ctl_dir, 4793 }, 4794 {} 4795}; 4796 4797static struct ctl_table *sd_alloc_ctl_entry(int n) 4798{ 4799 struct ctl_table *entry = 4800 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4801 4802 return entry; 4803} 4804 4805static void sd_free_ctl_entry(struct ctl_table **tablep) 4806{ 4807 struct ctl_table *entry; 4808 4809 /* 4810 * In the intermediate directories, both the child directory and 4811 * procname are dynamically allocated and could fail but the mode 4812 * will always be set. In the lowest directory the names are 4813 * static strings and all have proc handlers. 4814 */ 4815 for (entry = *tablep; entry->mode; entry++) { 4816 if (entry->child) 4817 sd_free_ctl_entry(&entry->child); 4818 if (entry->proc_handler == NULL) 4819 kfree(entry->procname); 4820 } 4821 4822 kfree(*tablep); 4823 *tablep = NULL; 4824} 4825 4826static int min_load_idx = 0; 4827static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4828 4829static void 4830set_table_entry(struct ctl_table *entry, 4831 const char *procname, void *data, int maxlen, 4832 umode_t mode, proc_handler *proc_handler, 4833 bool load_idx) 4834{ 4835 entry->procname = procname; 4836 entry->data = data; 4837 entry->maxlen = maxlen; 4838 entry->mode = mode; 4839 entry->proc_handler = proc_handler; 4840 4841 if (load_idx) { 4842 entry->extra1 = &min_load_idx; 4843 entry->extra2 = &max_load_idx; 4844 } 4845} 4846 4847static struct ctl_table * 4848sd_alloc_ctl_domain_table(struct sched_domain *sd) 4849{ 4850 struct ctl_table *table = sd_alloc_ctl_entry(14); 4851 4852 if (table == NULL) 4853 return NULL; 4854 4855 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4856 sizeof(long), 0644, proc_doulongvec_minmax, false); 4857 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4858 sizeof(long), 0644, proc_doulongvec_minmax, false); 4859 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4860 sizeof(int), 0644, proc_dointvec_minmax, true); 4861 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4862 sizeof(int), 0644, proc_dointvec_minmax, true); 4863 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4864 sizeof(int), 0644, proc_dointvec_minmax, true); 4865 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4866 sizeof(int), 0644, proc_dointvec_minmax, true); 4867 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4868 sizeof(int), 0644, proc_dointvec_minmax, true); 4869 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4870 sizeof(int), 0644, proc_dointvec_minmax, false); 4871 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4872 sizeof(int), 0644, proc_dointvec_minmax, false); 4873 set_table_entry(&table[9], "cache_nice_tries", 4874 &sd->cache_nice_tries, 4875 sizeof(int), 0644, proc_dointvec_minmax, false); 4876 set_table_entry(&table[10], "flags", &sd->flags, 4877 sizeof(int), 0644, proc_dointvec_minmax, false); 4878 set_table_entry(&table[11], "max_newidle_lb_cost", 4879 &sd->max_newidle_lb_cost, 4880 sizeof(long), 0644, proc_doulongvec_minmax, false); 4881 set_table_entry(&table[12], "name", sd->name, 4882 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4883 /* &table[13] is terminator */ 4884 4885 return table; 4886} 4887 4888static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4889{ 4890 struct ctl_table *entry, *table; 4891 struct sched_domain *sd; 4892 int domain_num = 0, i; 4893 char buf[32]; 4894 4895 for_each_domain(cpu, sd) 4896 domain_num++; 4897 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4898 if (table == NULL) 4899 return NULL; 4900 4901 i = 0; 4902 for_each_domain(cpu, sd) { 4903 snprintf(buf, 32, "domain%d", i); 4904 entry->procname = kstrdup(buf, GFP_KERNEL); 4905 entry->mode = 0555; 4906 entry->child = sd_alloc_ctl_domain_table(sd); 4907 entry++; 4908 i++; 4909 } 4910 return table; 4911} 4912 4913static struct ctl_table_header *sd_sysctl_header; 4914static void register_sched_domain_sysctl(void) 4915{ 4916 int i, cpu_num = num_possible_cpus(); 4917 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4918 char buf[32]; 4919 4920 WARN_ON(sd_ctl_dir[0].child); 4921 sd_ctl_dir[0].child = entry; 4922 4923 if (entry == NULL) 4924 return; 4925 4926 for_each_possible_cpu(i) { 4927 snprintf(buf, 32, "cpu%d", i); 4928 entry->procname = kstrdup(buf, GFP_KERNEL); 4929 entry->mode = 0555; 4930 entry->child = sd_alloc_ctl_cpu_table(i); 4931 entry++; 4932 } 4933 4934 WARN_ON(sd_sysctl_header); 4935 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4936} 4937 4938/* may be called multiple times per register */ 4939static void unregister_sched_domain_sysctl(void) 4940{ 4941 if (sd_sysctl_header) 4942 unregister_sysctl_table(sd_sysctl_header); 4943 sd_sysctl_header = NULL; 4944 if (sd_ctl_dir[0].child) 4945 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4946} 4947#else 4948static void register_sched_domain_sysctl(void) 4949{ 4950} 4951static void unregister_sched_domain_sysctl(void) 4952{ 4953} 4954#endif 4955 4956static void set_rq_online(struct rq *rq) 4957{ 4958 if (!rq->online) { 4959 const struct sched_class *class; 4960 4961 cpumask_set_cpu(rq->cpu, rq->rd->online); 4962 rq->online = 1; 4963 4964 for_each_class(class) { 4965 if (class->rq_online) 4966 class->rq_online(rq); 4967 } 4968 } 4969} 4970 4971static void set_rq_offline(struct rq *rq) 4972{ 4973 if (rq->online) { 4974 const struct sched_class *class; 4975 4976 for_each_class(class) { 4977 if (class->rq_offline) 4978 class->rq_offline(rq); 4979 } 4980 4981 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4982 rq->online = 0; 4983 } 4984} 4985 4986/* 4987 * migration_call - callback that gets triggered when a CPU is added. 4988 * Here we can start up the necessary migration thread for the new CPU. 4989 */ 4990static int 4991migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4992{ 4993 int cpu = (long)hcpu; 4994 unsigned long flags; 4995 struct rq *rq = cpu_rq(cpu); 4996 4997 switch (action & ~CPU_TASKS_FROZEN) { 4998 4999 case CPU_UP_PREPARE: 5000 rq->calc_load_update = calc_load_update; 5001 break; 5002 5003 case CPU_ONLINE: 5004 /* Update our root-domain */ 5005 raw_spin_lock_irqsave(&rq->lock, flags); 5006 if (rq->rd) { 5007 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5008 5009 set_rq_online(rq); 5010 } 5011 raw_spin_unlock_irqrestore(&rq->lock, flags); 5012 break; 5013 5014#ifdef CONFIG_HOTPLUG_CPU 5015 case CPU_DYING: 5016 sched_ttwu_pending(); 5017 /* Update our root-domain */ 5018 raw_spin_lock_irqsave(&rq->lock, flags); 5019 if (rq->rd) { 5020 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5021 set_rq_offline(rq); 5022 } 5023 migrate_tasks(cpu); 5024 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5025 raw_spin_unlock_irqrestore(&rq->lock, flags); 5026 break; 5027 5028 case CPU_DEAD: 5029 calc_load_migrate(rq); 5030 break; 5031#endif 5032 } 5033 5034 update_max_interval(); 5035 5036 return NOTIFY_OK; 5037} 5038 5039/* 5040 * Register at high priority so that task migration (migrate_all_tasks) 5041 * happens before everything else. This has to be lower priority than 5042 * the notifier in the perf_event subsystem, though. 5043 */ 5044static struct notifier_block migration_notifier = { 5045 .notifier_call = migration_call, 5046 .priority = CPU_PRI_MIGRATION, 5047}; 5048 5049static int sched_cpu_active(struct notifier_block *nfb, 5050 unsigned long action, void *hcpu) 5051{ 5052 switch (action & ~CPU_TASKS_FROZEN) { 5053 case CPU_STARTING: 5054 case CPU_DOWN_FAILED: 5055 set_cpu_active((long)hcpu, true); 5056 return NOTIFY_OK; 5057 default: 5058 return NOTIFY_DONE; 5059 } 5060} 5061 5062static int sched_cpu_inactive(struct notifier_block *nfb, 5063 unsigned long action, void *hcpu) 5064{ 5065 unsigned long flags; 5066 long cpu = (long)hcpu; 5067 5068 switch (action & ~CPU_TASKS_FROZEN) { 5069 case CPU_DOWN_PREPARE: 5070 set_cpu_active(cpu, false); 5071 5072 /* explicitly allow suspend */ 5073 if (!(action & CPU_TASKS_FROZEN)) { 5074 struct dl_bw *dl_b = dl_bw_of(cpu); 5075 bool overflow; 5076 int cpus; 5077 5078 raw_spin_lock_irqsave(&dl_b->lock, flags); 5079 cpus = dl_bw_cpus(cpu); 5080 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5081 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5082 5083 if (overflow) 5084 return notifier_from_errno(-EBUSY); 5085 } 5086 return NOTIFY_OK; 5087 } 5088 5089 return NOTIFY_DONE; 5090} 5091 5092static int __init migration_init(void) 5093{ 5094 void *cpu = (void *)(long)smp_processor_id(); 5095 int err; 5096 5097 /* Initialize migration for the boot CPU */ 5098 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5099 BUG_ON(err == NOTIFY_BAD); 5100 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5101 register_cpu_notifier(&migration_notifier); 5102 5103 /* Register cpu active notifiers */ 5104 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5105 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5106 5107 return 0; 5108} 5109early_initcall(migration_init); 5110#endif 5111 5112#ifdef CONFIG_SMP 5113 5114static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5115 5116#ifdef CONFIG_SCHED_DEBUG 5117 5118static __read_mostly int sched_debug_enabled; 5119 5120static int __init sched_debug_setup(char *str) 5121{ 5122 sched_debug_enabled = 1; 5123 5124 return 0; 5125} 5126early_param("sched_debug", sched_debug_setup); 5127 5128static inline bool sched_debug(void) 5129{ 5130 return sched_debug_enabled; 5131} 5132 5133static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5134 struct cpumask *groupmask) 5135{ 5136 struct sched_group *group = sd->groups; 5137 char str[256]; 5138 5139 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5140 cpumask_clear(groupmask); 5141 5142 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5143 5144 if (!(sd->flags & SD_LOAD_BALANCE)) { 5145 printk("does not load-balance\n"); 5146 if (sd->parent) 5147 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5148 " has parent"); 5149 return -1; 5150 } 5151 5152 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5153 5154 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5155 printk(KERN_ERR "ERROR: domain->span does not contain " 5156 "CPU%d\n", cpu); 5157 } 5158 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5159 printk(KERN_ERR "ERROR: domain->groups does not contain" 5160 " CPU%d\n", cpu); 5161 } 5162 5163 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5164 do { 5165 if (!group) { 5166 printk("\n"); 5167 printk(KERN_ERR "ERROR: group is NULL\n"); 5168 break; 5169 } 5170 5171 /* 5172 * Even though we initialize ->power to something semi-sane, 5173 * we leave power_orig unset. This allows us to detect if 5174 * domain iteration is still funny without causing /0 traps. 5175 */ 5176 if (!group->sgp->power_orig) { 5177 printk(KERN_CONT "\n"); 5178 printk(KERN_ERR "ERROR: domain->cpu_power not " 5179 "set\n"); 5180 break; 5181 } 5182 5183 if (!cpumask_weight(sched_group_cpus(group))) { 5184 printk(KERN_CONT "\n"); 5185 printk(KERN_ERR "ERROR: empty group\n"); 5186 break; 5187 } 5188 5189 if (!(sd->flags & SD_OVERLAP) && 5190 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5191 printk(KERN_CONT "\n"); 5192 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5193 break; 5194 } 5195 5196 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5197 5198 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5199 5200 printk(KERN_CONT " %s", str); 5201 if (group->sgp->power != SCHED_POWER_SCALE) { 5202 printk(KERN_CONT " (cpu_power = %d)", 5203 group->sgp->power); 5204 } 5205 5206 group = group->next; 5207 } while (group != sd->groups); 5208 printk(KERN_CONT "\n"); 5209 5210 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5211 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5212 5213 if (sd->parent && 5214 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5215 printk(KERN_ERR "ERROR: parent span is not a superset " 5216 "of domain->span\n"); 5217 return 0; 5218} 5219 5220static void sched_domain_debug(struct sched_domain *sd, int cpu) 5221{ 5222 int level = 0; 5223 5224 if (!sched_debug_enabled) 5225 return; 5226 5227 if (!sd) { 5228 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5229 return; 5230 } 5231 5232 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5233 5234 for (;;) { 5235 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5236 break; 5237 level++; 5238 sd = sd->parent; 5239 if (!sd) 5240 break; 5241 } 5242} 5243#else /* !CONFIG_SCHED_DEBUG */ 5244# define sched_domain_debug(sd, cpu) do { } while (0) 5245static inline bool sched_debug(void) 5246{ 5247 return false; 5248} 5249#endif /* CONFIG_SCHED_DEBUG */ 5250 5251static int sd_degenerate(struct sched_domain *sd) 5252{ 5253 if (cpumask_weight(sched_domain_span(sd)) == 1) 5254 return 1; 5255 5256 /* Following flags need at least 2 groups */ 5257 if (sd->flags & (SD_LOAD_BALANCE | 5258 SD_BALANCE_NEWIDLE | 5259 SD_BALANCE_FORK | 5260 SD_BALANCE_EXEC | 5261 SD_SHARE_CPUPOWER | 5262 SD_SHARE_PKG_RESOURCES)) { 5263 if (sd->groups != sd->groups->next) 5264 return 0; 5265 } 5266 5267 /* Following flags don't use groups */ 5268 if (sd->flags & (SD_WAKE_AFFINE)) 5269 return 0; 5270 5271 return 1; 5272} 5273 5274static int 5275sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5276{ 5277 unsigned long cflags = sd->flags, pflags = parent->flags; 5278 5279 if (sd_degenerate(parent)) 5280 return 1; 5281 5282 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5283 return 0; 5284 5285 /* Flags needing groups don't count if only 1 group in parent */ 5286 if (parent->groups == parent->groups->next) { 5287 pflags &= ~(SD_LOAD_BALANCE | 5288 SD_BALANCE_NEWIDLE | 5289 SD_BALANCE_FORK | 5290 SD_BALANCE_EXEC | 5291 SD_SHARE_CPUPOWER | 5292 SD_SHARE_PKG_RESOURCES | 5293 SD_PREFER_SIBLING); 5294 if (nr_node_ids == 1) 5295 pflags &= ~SD_SERIALIZE; 5296 } 5297 if (~cflags & pflags) 5298 return 0; 5299 5300 return 1; 5301} 5302 5303static void free_rootdomain(struct rcu_head *rcu) 5304{ 5305 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5306 5307 cpupri_cleanup(&rd->cpupri); 5308 cpudl_cleanup(&rd->cpudl); 5309 free_cpumask_var(rd->dlo_mask); 5310 free_cpumask_var(rd->rto_mask); 5311 free_cpumask_var(rd->online); 5312 free_cpumask_var(rd->span); 5313 kfree(rd); 5314} 5315 5316static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5317{ 5318 struct root_domain *old_rd = NULL; 5319 unsigned long flags; 5320 5321 raw_spin_lock_irqsave(&rq->lock, flags); 5322 5323 if (rq->rd) { 5324 old_rd = rq->rd; 5325 5326 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5327 set_rq_offline(rq); 5328 5329 cpumask_clear_cpu(rq->cpu, old_rd->span); 5330 5331 /* 5332 * If we dont want to free the old_rd yet then 5333 * set old_rd to NULL to skip the freeing later 5334 * in this function: 5335 */ 5336 if (!atomic_dec_and_test(&old_rd->refcount)) 5337 old_rd = NULL; 5338 } 5339 5340 atomic_inc(&rd->refcount); 5341 rq->rd = rd; 5342 5343 cpumask_set_cpu(rq->cpu, rd->span); 5344 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5345 set_rq_online(rq); 5346 5347 raw_spin_unlock_irqrestore(&rq->lock, flags); 5348 5349 if (old_rd) 5350 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5351} 5352 5353static int init_rootdomain(struct root_domain *rd) 5354{ 5355 memset(rd, 0, sizeof(*rd)); 5356 5357 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5358 goto out; 5359 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5360 goto free_span; 5361 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5362 goto free_online; 5363 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5364 goto free_dlo_mask; 5365 5366 init_dl_bw(&rd->dl_bw); 5367 if (cpudl_init(&rd->cpudl) != 0) 5368 goto free_dlo_mask; 5369 5370 if (cpupri_init(&rd->cpupri) != 0) 5371 goto free_rto_mask; 5372 return 0; 5373 5374free_rto_mask: 5375 free_cpumask_var(rd->rto_mask); 5376free_dlo_mask: 5377 free_cpumask_var(rd->dlo_mask); 5378free_online: 5379 free_cpumask_var(rd->online); 5380free_span: 5381 free_cpumask_var(rd->span); 5382out: 5383 return -ENOMEM; 5384} 5385 5386/* 5387 * By default the system creates a single root-domain with all cpus as 5388 * members (mimicking the global state we have today). 5389 */ 5390struct root_domain def_root_domain; 5391 5392static void init_defrootdomain(void) 5393{ 5394 init_rootdomain(&def_root_domain); 5395 5396 atomic_set(&def_root_domain.refcount, 1); 5397} 5398 5399static struct root_domain *alloc_rootdomain(void) 5400{ 5401 struct root_domain *rd; 5402 5403 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5404 if (!rd) 5405 return NULL; 5406 5407 if (init_rootdomain(rd) != 0) { 5408 kfree(rd); 5409 return NULL; 5410 } 5411 5412 return rd; 5413} 5414 5415static void free_sched_groups(struct sched_group *sg, int free_sgp) 5416{ 5417 struct sched_group *tmp, *first; 5418 5419 if (!sg) 5420 return; 5421 5422 first = sg; 5423 do { 5424 tmp = sg->next; 5425 5426 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5427 kfree(sg->sgp); 5428 5429 kfree(sg); 5430 sg = tmp; 5431 } while (sg != first); 5432} 5433 5434static void free_sched_domain(struct rcu_head *rcu) 5435{ 5436 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5437 5438 /* 5439 * If its an overlapping domain it has private groups, iterate and 5440 * nuke them all. 5441 */ 5442 if (sd->flags & SD_OVERLAP) { 5443 free_sched_groups(sd->groups, 1); 5444 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5445 kfree(sd->groups->sgp); 5446 kfree(sd->groups); 5447 } 5448 kfree(sd); 5449} 5450 5451static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5452{ 5453 call_rcu(&sd->rcu, free_sched_domain); 5454} 5455 5456static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5457{ 5458 for (; sd; sd = sd->parent) 5459 destroy_sched_domain(sd, cpu); 5460} 5461 5462/* 5463 * Keep a special pointer to the highest sched_domain that has 5464 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5465 * allows us to avoid some pointer chasing select_idle_sibling(). 5466 * 5467 * Also keep a unique ID per domain (we use the first cpu number in 5468 * the cpumask of the domain), this allows us to quickly tell if 5469 * two cpus are in the same cache domain, see cpus_share_cache(). 5470 */ 5471DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5472DEFINE_PER_CPU(int, sd_llc_size); 5473DEFINE_PER_CPU(int, sd_llc_id); 5474DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5475DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5476DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5477 5478static void update_top_cache_domain(int cpu) 5479{ 5480 struct sched_domain *sd; 5481 struct sched_domain *busy_sd = NULL; 5482 int id = cpu; 5483 int size = 1; 5484 5485 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5486 if (sd) { 5487 id = cpumask_first(sched_domain_span(sd)); 5488 size = cpumask_weight(sched_domain_span(sd)); 5489 busy_sd = sd->parent; /* sd_busy */ 5490 } 5491 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5492 5493 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5494 per_cpu(sd_llc_size, cpu) = size; 5495 per_cpu(sd_llc_id, cpu) = id; 5496 5497 sd = lowest_flag_domain(cpu, SD_NUMA); 5498 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5499 5500 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5501 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5502} 5503 5504/* 5505 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5506 * hold the hotplug lock. 5507 */ 5508static void 5509cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5510{ 5511 struct rq *rq = cpu_rq(cpu); 5512 struct sched_domain *tmp; 5513 5514 /* Remove the sched domains which do not contribute to scheduling. */ 5515 for (tmp = sd; tmp; ) { 5516 struct sched_domain *parent = tmp->parent; 5517 if (!parent) 5518 break; 5519 5520 if (sd_parent_degenerate(tmp, parent)) { 5521 tmp->parent = parent->parent; 5522 if (parent->parent) 5523 parent->parent->child = tmp; 5524 /* 5525 * Transfer SD_PREFER_SIBLING down in case of a 5526 * degenerate parent; the spans match for this 5527 * so the property transfers. 5528 */ 5529 if (parent->flags & SD_PREFER_SIBLING) 5530 tmp->flags |= SD_PREFER_SIBLING; 5531 destroy_sched_domain(parent, cpu); 5532 } else 5533 tmp = tmp->parent; 5534 } 5535 5536 if (sd && sd_degenerate(sd)) { 5537 tmp = sd; 5538 sd = sd->parent; 5539 destroy_sched_domain(tmp, cpu); 5540 if (sd) 5541 sd->child = NULL; 5542 } 5543 5544 sched_domain_debug(sd, cpu); 5545 5546 rq_attach_root(rq, rd); 5547 tmp = rq->sd; 5548 rcu_assign_pointer(rq->sd, sd); 5549 destroy_sched_domains(tmp, cpu); 5550 5551 update_top_cache_domain(cpu); 5552} 5553 5554/* cpus with isolated domains */ 5555static cpumask_var_t cpu_isolated_map; 5556 5557/* Setup the mask of cpus configured for isolated domains */ 5558static int __init isolated_cpu_setup(char *str) 5559{ 5560 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5561 cpulist_parse(str, cpu_isolated_map); 5562 return 1; 5563} 5564 5565__setup("isolcpus=", isolated_cpu_setup); 5566 5567static const struct cpumask *cpu_cpu_mask(int cpu) 5568{ 5569 return cpumask_of_node(cpu_to_node(cpu)); 5570} 5571 5572struct sd_data { 5573 struct sched_domain **__percpu sd; 5574 struct sched_group **__percpu sg; 5575 struct sched_group_power **__percpu sgp; 5576}; 5577 5578struct s_data { 5579 struct sched_domain ** __percpu sd; 5580 struct root_domain *rd; 5581}; 5582 5583enum s_alloc { 5584 sa_rootdomain, 5585 sa_sd, 5586 sa_sd_storage, 5587 sa_none, 5588}; 5589 5590struct sched_domain_topology_level; 5591 5592typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5593typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5594 5595#define SDTL_OVERLAP 0x01 5596 5597struct sched_domain_topology_level { 5598 sched_domain_init_f init; 5599 sched_domain_mask_f mask; 5600 int flags; 5601 int numa_level; 5602 struct sd_data data; 5603}; 5604 5605/* 5606 * Build an iteration mask that can exclude certain CPUs from the upwards 5607 * domain traversal. 5608 * 5609 * Asymmetric node setups can result in situations where the domain tree is of 5610 * unequal depth, make sure to skip domains that already cover the entire 5611 * range. 5612 * 5613 * In that case build_sched_domains() will have terminated the iteration early 5614 * and our sibling sd spans will be empty. Domains should always include the 5615 * cpu they're built on, so check that. 5616 * 5617 */ 5618static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5619{ 5620 const struct cpumask *span = sched_domain_span(sd); 5621 struct sd_data *sdd = sd->private; 5622 struct sched_domain *sibling; 5623 int i; 5624 5625 for_each_cpu(i, span) { 5626 sibling = *per_cpu_ptr(sdd->sd, i); 5627 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5628 continue; 5629 5630 cpumask_set_cpu(i, sched_group_mask(sg)); 5631 } 5632} 5633 5634/* 5635 * Return the canonical balance cpu for this group, this is the first cpu 5636 * of this group that's also in the iteration mask. 5637 */ 5638int group_balance_cpu(struct sched_group *sg) 5639{ 5640 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5641} 5642 5643static int 5644build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5645{ 5646 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5647 const struct cpumask *span = sched_domain_span(sd); 5648 struct cpumask *covered = sched_domains_tmpmask; 5649 struct sd_data *sdd = sd->private; 5650 struct sched_domain *child; 5651 int i; 5652 5653 cpumask_clear(covered); 5654 5655 for_each_cpu(i, span) { 5656 struct cpumask *sg_span; 5657 5658 if (cpumask_test_cpu(i, covered)) 5659 continue; 5660 5661 child = *per_cpu_ptr(sdd->sd, i); 5662 5663 /* See the comment near build_group_mask(). */ 5664 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5665 continue; 5666 5667 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5668 GFP_KERNEL, cpu_to_node(cpu)); 5669 5670 if (!sg) 5671 goto fail; 5672 5673 sg_span = sched_group_cpus(sg); 5674 if (child->child) { 5675 child = child->child; 5676 cpumask_copy(sg_span, sched_domain_span(child)); 5677 } else 5678 cpumask_set_cpu(i, sg_span); 5679 5680 cpumask_or(covered, covered, sg_span); 5681 5682 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5683 if (atomic_inc_return(&sg->sgp->ref) == 1) 5684 build_group_mask(sd, sg); 5685 5686 /* 5687 * Initialize sgp->power such that even if we mess up the 5688 * domains and no possible iteration will get us here, we won't 5689 * die on a /0 trap. 5690 */ 5691 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5692 sg->sgp->power_orig = sg->sgp->power; 5693 5694 /* 5695 * Make sure the first group of this domain contains the 5696 * canonical balance cpu. Otherwise the sched_domain iteration 5697 * breaks. See update_sg_lb_stats(). 5698 */ 5699 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5700 group_balance_cpu(sg) == cpu) 5701 groups = sg; 5702 5703 if (!first) 5704 first = sg; 5705 if (last) 5706 last->next = sg; 5707 last = sg; 5708 last->next = first; 5709 } 5710 sd->groups = groups; 5711 5712 return 0; 5713 5714fail: 5715 free_sched_groups(first, 0); 5716 5717 return -ENOMEM; 5718} 5719 5720static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5721{ 5722 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5723 struct sched_domain *child = sd->child; 5724 5725 if (child) 5726 cpu = cpumask_first(sched_domain_span(child)); 5727 5728 if (sg) { 5729 *sg = *per_cpu_ptr(sdd->sg, cpu); 5730 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5731 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5732 } 5733 5734 return cpu; 5735} 5736 5737/* 5738 * build_sched_groups will build a circular linked list of the groups 5739 * covered by the given span, and will set each group's ->cpumask correctly, 5740 * and ->cpu_power to 0. 5741 * 5742 * Assumes the sched_domain tree is fully constructed 5743 */ 5744static int 5745build_sched_groups(struct sched_domain *sd, int cpu) 5746{ 5747 struct sched_group *first = NULL, *last = NULL; 5748 struct sd_data *sdd = sd->private; 5749 const struct cpumask *span = sched_domain_span(sd); 5750 struct cpumask *covered; 5751 int i; 5752 5753 get_group(cpu, sdd, &sd->groups); 5754 atomic_inc(&sd->groups->ref); 5755 5756 if (cpu != cpumask_first(span)) 5757 return 0; 5758 5759 lockdep_assert_held(&sched_domains_mutex); 5760 covered = sched_domains_tmpmask; 5761 5762 cpumask_clear(covered); 5763 5764 for_each_cpu(i, span) { 5765 struct sched_group *sg; 5766 int group, j; 5767 5768 if (cpumask_test_cpu(i, covered)) 5769 continue; 5770 5771 group = get_group(i, sdd, &sg); 5772 cpumask_clear(sched_group_cpus(sg)); 5773 sg->sgp->power = 0; 5774 cpumask_setall(sched_group_mask(sg)); 5775 5776 for_each_cpu(j, span) { 5777 if (get_group(j, sdd, NULL) != group) 5778 continue; 5779 5780 cpumask_set_cpu(j, covered); 5781 cpumask_set_cpu(j, sched_group_cpus(sg)); 5782 } 5783 5784 if (!first) 5785 first = sg; 5786 if (last) 5787 last->next = sg; 5788 last = sg; 5789 } 5790 last->next = first; 5791 5792 return 0; 5793} 5794 5795/* 5796 * Initialize sched groups cpu_power. 5797 * 5798 * cpu_power indicates the capacity of sched group, which is used while 5799 * distributing the load between different sched groups in a sched domain. 5800 * Typically cpu_power for all the groups in a sched domain will be same unless 5801 * there are asymmetries in the topology. If there are asymmetries, group 5802 * having more cpu_power will pickup more load compared to the group having 5803 * less cpu_power. 5804 */ 5805static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5806{ 5807 struct sched_group *sg = sd->groups; 5808 5809 WARN_ON(!sg); 5810 5811 do { 5812 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5813 sg = sg->next; 5814 } while (sg != sd->groups); 5815 5816 if (cpu != group_balance_cpu(sg)) 5817 return; 5818 5819 update_group_power(sd, cpu); 5820 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5821} 5822 5823int __weak arch_sd_sibling_asym_packing(void) 5824{ 5825 return 0*SD_ASYM_PACKING; 5826} 5827 5828/* 5829 * Initializers for schedule domains 5830 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5831 */ 5832 5833#ifdef CONFIG_SCHED_DEBUG 5834# define SD_INIT_NAME(sd, type) sd->name = #type 5835#else 5836# define SD_INIT_NAME(sd, type) do { } while (0) 5837#endif 5838 5839#define SD_INIT_FUNC(type) \ 5840static noinline struct sched_domain * \ 5841sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5842{ \ 5843 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5844 *sd = SD_##type##_INIT; \ 5845 SD_INIT_NAME(sd, type); \ 5846 sd->private = &tl->data; \ 5847 return sd; \ 5848} 5849 5850SD_INIT_FUNC(CPU) 5851#ifdef CONFIG_SCHED_SMT 5852 SD_INIT_FUNC(SIBLING) 5853#endif 5854#ifdef CONFIG_SCHED_MC 5855 SD_INIT_FUNC(MC) 5856#endif 5857#ifdef CONFIG_SCHED_BOOK 5858 SD_INIT_FUNC(BOOK) 5859#endif 5860 5861static int default_relax_domain_level = -1; 5862int sched_domain_level_max; 5863 5864static int __init setup_relax_domain_level(char *str) 5865{ 5866 if (kstrtoint(str, 0, &default_relax_domain_level)) 5867 pr_warn("Unable to set relax_domain_level\n"); 5868 5869 return 1; 5870} 5871__setup("relax_domain_level=", setup_relax_domain_level); 5872 5873static void set_domain_attribute(struct sched_domain *sd, 5874 struct sched_domain_attr *attr) 5875{ 5876 int request; 5877 5878 if (!attr || attr->relax_domain_level < 0) { 5879 if (default_relax_domain_level < 0) 5880 return; 5881 else 5882 request = default_relax_domain_level; 5883 } else 5884 request = attr->relax_domain_level; 5885 if (request < sd->level) { 5886 /* turn off idle balance on this domain */ 5887 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5888 } else { 5889 /* turn on idle balance on this domain */ 5890 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5891 } 5892} 5893 5894static void __sdt_free(const struct cpumask *cpu_map); 5895static int __sdt_alloc(const struct cpumask *cpu_map); 5896 5897static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5898 const struct cpumask *cpu_map) 5899{ 5900 switch (what) { 5901 case sa_rootdomain: 5902 if (!atomic_read(&d->rd->refcount)) 5903 free_rootdomain(&d->rd->rcu); /* fall through */ 5904 case sa_sd: 5905 free_percpu(d->sd); /* fall through */ 5906 case sa_sd_storage: 5907 __sdt_free(cpu_map); /* fall through */ 5908 case sa_none: 5909 break; 5910 } 5911} 5912 5913static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5914 const struct cpumask *cpu_map) 5915{ 5916 memset(d, 0, sizeof(*d)); 5917 5918 if (__sdt_alloc(cpu_map)) 5919 return sa_sd_storage; 5920 d->sd = alloc_percpu(struct sched_domain *); 5921 if (!d->sd) 5922 return sa_sd_storage; 5923 d->rd = alloc_rootdomain(); 5924 if (!d->rd) 5925 return sa_sd; 5926 return sa_rootdomain; 5927} 5928 5929/* 5930 * NULL the sd_data elements we've used to build the sched_domain and 5931 * sched_group structure so that the subsequent __free_domain_allocs() 5932 * will not free the data we're using. 5933 */ 5934static void claim_allocations(int cpu, struct sched_domain *sd) 5935{ 5936 struct sd_data *sdd = sd->private; 5937 5938 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5939 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5940 5941 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5942 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5943 5944 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5945 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5946} 5947 5948#ifdef CONFIG_SCHED_SMT 5949static const struct cpumask *cpu_smt_mask(int cpu) 5950{ 5951 return topology_thread_cpumask(cpu); 5952} 5953#endif 5954 5955/* 5956 * Topology list, bottom-up. 5957 */ 5958static struct sched_domain_topology_level default_topology[] = { 5959#ifdef CONFIG_SCHED_SMT 5960 { sd_init_SIBLING, cpu_smt_mask, }, 5961#endif 5962#ifdef CONFIG_SCHED_MC 5963 { sd_init_MC, cpu_coregroup_mask, }, 5964#endif 5965#ifdef CONFIG_SCHED_BOOK 5966 { sd_init_BOOK, cpu_book_mask, }, 5967#endif 5968 { sd_init_CPU, cpu_cpu_mask, }, 5969 { NULL, }, 5970}; 5971 5972static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5973 5974#define for_each_sd_topology(tl) \ 5975 for (tl = sched_domain_topology; tl->init; tl++) 5976 5977#ifdef CONFIG_NUMA 5978 5979static int sched_domains_numa_levels; 5980static int *sched_domains_numa_distance; 5981static struct cpumask ***sched_domains_numa_masks; 5982static int sched_domains_curr_level; 5983 5984static inline int sd_local_flags(int level) 5985{ 5986 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5987 return 0; 5988 5989 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5990} 5991 5992static struct sched_domain * 5993sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5994{ 5995 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5996 int level = tl->numa_level; 5997 int sd_weight = cpumask_weight( 5998 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5999 6000 *sd = (struct sched_domain){ 6001 .min_interval = sd_weight, 6002 .max_interval = 2*sd_weight, 6003 .busy_factor = 32, 6004 .imbalance_pct = 125, 6005 .cache_nice_tries = 2, 6006 .busy_idx = 3, 6007 .idle_idx = 2, 6008 .newidle_idx = 0, 6009 .wake_idx = 0, 6010 .forkexec_idx = 0, 6011 6012 .flags = 1*SD_LOAD_BALANCE 6013 | 1*SD_BALANCE_NEWIDLE 6014 | 0*SD_BALANCE_EXEC 6015 | 0*SD_BALANCE_FORK 6016 | 0*SD_BALANCE_WAKE 6017 | 0*SD_WAKE_AFFINE 6018 | 0*SD_SHARE_CPUPOWER 6019 | 0*SD_SHARE_PKG_RESOURCES 6020 | 1*SD_SERIALIZE 6021 | 0*SD_PREFER_SIBLING 6022 | 1*SD_NUMA 6023 | sd_local_flags(level) 6024 , 6025 .last_balance = jiffies, 6026 .balance_interval = sd_weight, 6027 }; 6028 SD_INIT_NAME(sd, NUMA); 6029 sd->private = &tl->data; 6030 6031 /* 6032 * Ugly hack to pass state to sd_numa_mask()... 6033 */ 6034 sched_domains_curr_level = tl->numa_level; 6035 6036 return sd; 6037} 6038 6039static const struct cpumask *sd_numa_mask(int cpu) 6040{ 6041 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6042} 6043 6044static void sched_numa_warn(const char *str) 6045{ 6046 static int done = false; 6047 int i,j; 6048 6049 if (done) 6050 return; 6051 6052 done = true; 6053 6054 printk(KERN_WARNING "ERROR: %s\n\n", str); 6055 6056 for (i = 0; i < nr_node_ids; i++) { 6057 printk(KERN_WARNING " "); 6058 for (j = 0; j < nr_node_ids; j++) 6059 printk(KERN_CONT "%02d ", node_distance(i,j)); 6060 printk(KERN_CONT "\n"); 6061 } 6062 printk(KERN_WARNING "\n"); 6063} 6064 6065static bool find_numa_distance(int distance) 6066{ 6067 int i; 6068 6069 if (distance == node_distance(0, 0)) 6070 return true; 6071 6072 for (i = 0; i < sched_domains_numa_levels; i++) { 6073 if (sched_domains_numa_distance[i] == distance) 6074 return true; 6075 } 6076 6077 return false; 6078} 6079 6080static void sched_init_numa(void) 6081{ 6082 int next_distance, curr_distance = node_distance(0, 0); 6083 struct sched_domain_topology_level *tl; 6084 int level = 0; 6085 int i, j, k; 6086 6087 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6088 if (!sched_domains_numa_distance) 6089 return; 6090 6091 /* 6092 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6093 * unique distances in the node_distance() table. 6094 * 6095 * Assumes node_distance(0,j) includes all distances in 6096 * node_distance(i,j) in order to avoid cubic time. 6097 */ 6098 next_distance = curr_distance; 6099 for (i = 0; i < nr_node_ids; i++) { 6100 for (j = 0; j < nr_node_ids; j++) { 6101 for (k = 0; k < nr_node_ids; k++) { 6102 int distance = node_distance(i, k); 6103 6104 if (distance > curr_distance && 6105 (distance < next_distance || 6106 next_distance == curr_distance)) 6107 next_distance = distance; 6108 6109 /* 6110 * While not a strong assumption it would be nice to know 6111 * about cases where if node A is connected to B, B is not 6112 * equally connected to A. 6113 */ 6114 if (sched_debug() && node_distance(k, i) != distance) 6115 sched_numa_warn("Node-distance not symmetric"); 6116 6117 if (sched_debug() && i && !find_numa_distance(distance)) 6118 sched_numa_warn("Node-0 not representative"); 6119 } 6120 if (next_distance != curr_distance) { 6121 sched_domains_numa_distance[level++] = next_distance; 6122 sched_domains_numa_levels = level; 6123 curr_distance = next_distance; 6124 } else break; 6125 } 6126 6127 /* 6128 * In case of sched_debug() we verify the above assumption. 6129 */ 6130 if (!sched_debug()) 6131 break; 6132 } 6133 /* 6134 * 'level' contains the number of unique distances, excluding the 6135 * identity distance node_distance(i,i). 6136 * 6137 * The sched_domains_numa_distance[] array includes the actual distance 6138 * numbers. 6139 */ 6140 6141 /* 6142 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6143 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6144 * the array will contain less then 'level' members. This could be 6145 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6146 * in other functions. 6147 * 6148 * We reset it to 'level' at the end of this function. 6149 */ 6150 sched_domains_numa_levels = 0; 6151 6152 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6153 if (!sched_domains_numa_masks) 6154 return; 6155 6156 /* 6157 * Now for each level, construct a mask per node which contains all 6158 * cpus of nodes that are that many hops away from us. 6159 */ 6160 for (i = 0; i < level; i++) { 6161 sched_domains_numa_masks[i] = 6162 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6163 if (!sched_domains_numa_masks[i]) 6164 return; 6165 6166 for (j = 0; j < nr_node_ids; j++) { 6167 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6168 if (!mask) 6169 return; 6170 6171 sched_domains_numa_masks[i][j] = mask; 6172 6173 for (k = 0; k < nr_node_ids; k++) { 6174 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6175 continue; 6176 6177 cpumask_or(mask, mask, cpumask_of_node(k)); 6178 } 6179 } 6180 } 6181 6182 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6183 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6184 if (!tl) 6185 return; 6186 6187 /* 6188 * Copy the default topology bits.. 6189 */ 6190 for (i = 0; default_topology[i].init; i++) 6191 tl[i] = default_topology[i]; 6192 6193 /* 6194 * .. and append 'j' levels of NUMA goodness. 6195 */ 6196 for (j = 0; j < level; i++, j++) { 6197 tl[i] = (struct sched_domain_topology_level){ 6198 .init = sd_numa_init, 6199 .mask = sd_numa_mask, 6200 .flags = SDTL_OVERLAP, 6201 .numa_level = j, 6202 }; 6203 } 6204 6205 sched_domain_topology = tl; 6206 6207 sched_domains_numa_levels = level; 6208} 6209 6210static void sched_domains_numa_masks_set(int cpu) 6211{ 6212 int i, j; 6213 int node = cpu_to_node(cpu); 6214 6215 for (i = 0; i < sched_domains_numa_levels; i++) { 6216 for (j = 0; j < nr_node_ids; j++) { 6217 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6218 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6219 } 6220 } 6221} 6222 6223static void sched_domains_numa_masks_clear(int cpu) 6224{ 6225 int i, j; 6226 for (i = 0; i < sched_domains_numa_levels; i++) { 6227 for (j = 0; j < nr_node_ids; j++) 6228 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6229 } 6230} 6231 6232/* 6233 * Update sched_domains_numa_masks[level][node] array when new cpus 6234 * are onlined. 6235 */ 6236static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6237 unsigned long action, 6238 void *hcpu) 6239{ 6240 int cpu = (long)hcpu; 6241 6242 switch (action & ~CPU_TASKS_FROZEN) { 6243 case CPU_ONLINE: 6244 sched_domains_numa_masks_set(cpu); 6245 break; 6246 6247 case CPU_DEAD: 6248 sched_domains_numa_masks_clear(cpu); 6249 break; 6250 6251 default: 6252 return NOTIFY_DONE; 6253 } 6254 6255 return NOTIFY_OK; 6256} 6257#else 6258static inline void sched_init_numa(void) 6259{ 6260} 6261 6262static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6263 unsigned long action, 6264 void *hcpu) 6265{ 6266 return 0; 6267} 6268#endif /* CONFIG_NUMA */ 6269 6270static int __sdt_alloc(const struct cpumask *cpu_map) 6271{ 6272 struct sched_domain_topology_level *tl; 6273 int j; 6274 6275 for_each_sd_topology(tl) { 6276 struct sd_data *sdd = &tl->data; 6277 6278 sdd->sd = alloc_percpu(struct sched_domain *); 6279 if (!sdd->sd) 6280 return -ENOMEM; 6281 6282 sdd->sg = alloc_percpu(struct sched_group *); 6283 if (!sdd->sg) 6284 return -ENOMEM; 6285 6286 sdd->sgp = alloc_percpu(struct sched_group_power *); 6287 if (!sdd->sgp) 6288 return -ENOMEM; 6289 6290 for_each_cpu(j, cpu_map) { 6291 struct sched_domain *sd; 6292 struct sched_group *sg; 6293 struct sched_group_power *sgp; 6294 6295 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6296 GFP_KERNEL, cpu_to_node(j)); 6297 if (!sd) 6298 return -ENOMEM; 6299 6300 *per_cpu_ptr(sdd->sd, j) = sd; 6301 6302 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6303 GFP_KERNEL, cpu_to_node(j)); 6304 if (!sg) 6305 return -ENOMEM; 6306 6307 sg->next = sg; 6308 6309 *per_cpu_ptr(sdd->sg, j) = sg; 6310 6311 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6312 GFP_KERNEL, cpu_to_node(j)); 6313 if (!sgp) 6314 return -ENOMEM; 6315 6316 *per_cpu_ptr(sdd->sgp, j) = sgp; 6317 } 6318 } 6319 6320 return 0; 6321} 6322 6323static void __sdt_free(const struct cpumask *cpu_map) 6324{ 6325 struct sched_domain_topology_level *tl; 6326 int j; 6327 6328 for_each_sd_topology(tl) { 6329 struct sd_data *sdd = &tl->data; 6330 6331 for_each_cpu(j, cpu_map) { 6332 struct sched_domain *sd; 6333 6334 if (sdd->sd) { 6335 sd = *per_cpu_ptr(sdd->sd, j); 6336 if (sd && (sd->flags & SD_OVERLAP)) 6337 free_sched_groups(sd->groups, 0); 6338 kfree(*per_cpu_ptr(sdd->sd, j)); 6339 } 6340 6341 if (sdd->sg) 6342 kfree(*per_cpu_ptr(sdd->sg, j)); 6343 if (sdd->sgp) 6344 kfree(*per_cpu_ptr(sdd->sgp, j)); 6345 } 6346 free_percpu(sdd->sd); 6347 sdd->sd = NULL; 6348 free_percpu(sdd->sg); 6349 sdd->sg = NULL; 6350 free_percpu(sdd->sgp); 6351 sdd->sgp = NULL; 6352 } 6353} 6354 6355struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6356 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6357 struct sched_domain *child, int cpu) 6358{ 6359 struct sched_domain *sd = tl->init(tl, cpu); 6360 if (!sd) 6361 return child; 6362 6363 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6364 if (child) { 6365 sd->level = child->level + 1; 6366 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6367 child->parent = sd; 6368 sd->child = child; 6369 } 6370 set_domain_attribute(sd, attr); 6371 6372 return sd; 6373} 6374 6375/* 6376 * Build sched domains for a given set of cpus and attach the sched domains 6377 * to the individual cpus 6378 */ 6379static int build_sched_domains(const struct cpumask *cpu_map, 6380 struct sched_domain_attr *attr) 6381{ 6382 enum s_alloc alloc_state; 6383 struct sched_domain *sd; 6384 struct s_data d; 6385 int i, ret = -ENOMEM; 6386 6387 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6388 if (alloc_state != sa_rootdomain) 6389 goto error; 6390 6391 /* Set up domains for cpus specified by the cpu_map. */ 6392 for_each_cpu(i, cpu_map) { 6393 struct sched_domain_topology_level *tl; 6394 6395 sd = NULL; 6396 for_each_sd_topology(tl) { 6397 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6398 if (tl == sched_domain_topology) 6399 *per_cpu_ptr(d.sd, i) = sd; 6400 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6401 sd->flags |= SD_OVERLAP; 6402 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6403 break; 6404 } 6405 } 6406 6407 /* Build the groups for the domains */ 6408 for_each_cpu(i, cpu_map) { 6409 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6410 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6411 if (sd->flags & SD_OVERLAP) { 6412 if (build_overlap_sched_groups(sd, i)) 6413 goto error; 6414 } else { 6415 if (build_sched_groups(sd, i)) 6416 goto error; 6417 } 6418 } 6419 } 6420 6421 /* Calculate CPU power for physical packages and nodes */ 6422 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6423 if (!cpumask_test_cpu(i, cpu_map)) 6424 continue; 6425 6426 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6427 claim_allocations(i, sd); 6428 init_sched_groups_power(i, sd); 6429 } 6430 } 6431 6432 /* Attach the domains */ 6433 rcu_read_lock(); 6434 for_each_cpu(i, cpu_map) { 6435 sd = *per_cpu_ptr(d.sd, i); 6436 cpu_attach_domain(sd, d.rd, i); 6437 } 6438 rcu_read_unlock(); 6439 6440 ret = 0; 6441error: 6442 __free_domain_allocs(&d, alloc_state, cpu_map); 6443 return ret; 6444} 6445 6446static cpumask_var_t *doms_cur; /* current sched domains */ 6447static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6448static struct sched_domain_attr *dattr_cur; 6449 /* attribues of custom domains in 'doms_cur' */ 6450 6451/* 6452 * Special case: If a kmalloc of a doms_cur partition (array of 6453 * cpumask) fails, then fallback to a single sched domain, 6454 * as determined by the single cpumask fallback_doms. 6455 */ 6456static cpumask_var_t fallback_doms; 6457 6458/* 6459 * arch_update_cpu_topology lets virtualized architectures update the 6460 * cpu core maps. It is supposed to return 1 if the topology changed 6461 * or 0 if it stayed the same. 6462 */ 6463int __attribute__((weak)) arch_update_cpu_topology(void) 6464{ 6465 return 0; 6466} 6467 6468cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6469{ 6470 int i; 6471 cpumask_var_t *doms; 6472 6473 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6474 if (!doms) 6475 return NULL; 6476 for (i = 0; i < ndoms; i++) { 6477 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6478 free_sched_domains(doms, i); 6479 return NULL; 6480 } 6481 } 6482 return doms; 6483} 6484 6485void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6486{ 6487 unsigned int i; 6488 for (i = 0; i < ndoms; i++) 6489 free_cpumask_var(doms[i]); 6490 kfree(doms); 6491} 6492 6493/* 6494 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6495 * For now this just excludes isolated cpus, but could be used to 6496 * exclude other special cases in the future. 6497 */ 6498static int init_sched_domains(const struct cpumask *cpu_map) 6499{ 6500 int err; 6501 6502 arch_update_cpu_topology(); 6503 ndoms_cur = 1; 6504 doms_cur = alloc_sched_domains(ndoms_cur); 6505 if (!doms_cur) 6506 doms_cur = &fallback_doms; 6507 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6508 err = build_sched_domains(doms_cur[0], NULL); 6509 register_sched_domain_sysctl(); 6510 6511 return err; 6512} 6513 6514/* 6515 * Detach sched domains from a group of cpus specified in cpu_map 6516 * These cpus will now be attached to the NULL domain 6517 */ 6518static void detach_destroy_domains(const struct cpumask *cpu_map) 6519{ 6520 int i; 6521 6522 rcu_read_lock(); 6523 for_each_cpu(i, cpu_map) 6524 cpu_attach_domain(NULL, &def_root_domain, i); 6525 rcu_read_unlock(); 6526} 6527 6528/* handle null as "default" */ 6529static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6530 struct sched_domain_attr *new, int idx_new) 6531{ 6532 struct sched_domain_attr tmp; 6533 6534 /* fast path */ 6535 if (!new && !cur) 6536 return 1; 6537 6538 tmp = SD_ATTR_INIT; 6539 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6540 new ? (new + idx_new) : &tmp, 6541 sizeof(struct sched_domain_attr)); 6542} 6543 6544/* 6545 * Partition sched domains as specified by the 'ndoms_new' 6546 * cpumasks in the array doms_new[] of cpumasks. This compares 6547 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6548 * It destroys each deleted domain and builds each new domain. 6549 * 6550 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6551 * The masks don't intersect (don't overlap.) We should setup one 6552 * sched domain for each mask. CPUs not in any of the cpumasks will 6553 * not be load balanced. If the same cpumask appears both in the 6554 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6555 * it as it is. 6556 * 6557 * The passed in 'doms_new' should be allocated using 6558 * alloc_sched_domains. This routine takes ownership of it and will 6559 * free_sched_domains it when done with it. If the caller failed the 6560 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6561 * and partition_sched_domains() will fallback to the single partition 6562 * 'fallback_doms', it also forces the domains to be rebuilt. 6563 * 6564 * If doms_new == NULL it will be replaced with cpu_online_mask. 6565 * ndoms_new == 0 is a special case for destroying existing domains, 6566 * and it will not create the default domain. 6567 * 6568 * Call with hotplug lock held 6569 */ 6570void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6571 struct sched_domain_attr *dattr_new) 6572{ 6573 int i, j, n; 6574 int new_topology; 6575 6576 mutex_lock(&sched_domains_mutex); 6577 6578 /* always unregister in case we don't destroy any domains */ 6579 unregister_sched_domain_sysctl(); 6580 6581 /* Let architecture update cpu core mappings. */ 6582 new_topology = arch_update_cpu_topology(); 6583 6584 n = doms_new ? ndoms_new : 0; 6585 6586 /* Destroy deleted domains */ 6587 for (i = 0; i < ndoms_cur; i++) { 6588 for (j = 0; j < n && !new_topology; j++) { 6589 if (cpumask_equal(doms_cur[i], doms_new[j]) 6590 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6591 goto match1; 6592 } 6593 /* no match - a current sched domain not in new doms_new[] */ 6594 detach_destroy_domains(doms_cur[i]); 6595match1: 6596 ; 6597 } 6598 6599 n = ndoms_cur; 6600 if (doms_new == NULL) { 6601 n = 0; 6602 doms_new = &fallback_doms; 6603 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6604 WARN_ON_ONCE(dattr_new); 6605 } 6606 6607 /* Build new domains */ 6608 for (i = 0; i < ndoms_new; i++) { 6609 for (j = 0; j < n && !new_topology; j++) { 6610 if (cpumask_equal(doms_new[i], doms_cur[j]) 6611 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6612 goto match2; 6613 } 6614 /* no match - add a new doms_new */ 6615 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6616match2: 6617 ; 6618 } 6619 6620 /* Remember the new sched domains */ 6621 if (doms_cur != &fallback_doms) 6622 free_sched_domains(doms_cur, ndoms_cur); 6623 kfree(dattr_cur); /* kfree(NULL) is safe */ 6624 doms_cur = doms_new; 6625 dattr_cur = dattr_new; 6626 ndoms_cur = ndoms_new; 6627 6628 register_sched_domain_sysctl(); 6629 6630 mutex_unlock(&sched_domains_mutex); 6631} 6632 6633static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6634 6635/* 6636 * Update cpusets according to cpu_active mask. If cpusets are 6637 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6638 * around partition_sched_domains(). 6639 * 6640 * If we come here as part of a suspend/resume, don't touch cpusets because we 6641 * want to restore it back to its original state upon resume anyway. 6642 */ 6643static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6644 void *hcpu) 6645{ 6646 switch (action) { 6647 case CPU_ONLINE_FROZEN: 6648 case CPU_DOWN_FAILED_FROZEN: 6649 6650 /* 6651 * num_cpus_frozen tracks how many CPUs are involved in suspend 6652 * resume sequence. As long as this is not the last online 6653 * operation in the resume sequence, just build a single sched 6654 * domain, ignoring cpusets. 6655 */ 6656 num_cpus_frozen--; 6657 if (likely(num_cpus_frozen)) { 6658 partition_sched_domains(1, NULL, NULL); 6659 break; 6660 } 6661 6662 /* 6663 * This is the last CPU online operation. So fall through and 6664 * restore the original sched domains by considering the 6665 * cpuset configurations. 6666 */ 6667 6668 case CPU_ONLINE: 6669 case CPU_DOWN_FAILED: 6670 cpuset_update_active_cpus(true); 6671 break; 6672 default: 6673 return NOTIFY_DONE; 6674 } 6675 return NOTIFY_OK; 6676} 6677 6678static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6679 void *hcpu) 6680{ 6681 switch (action) { 6682 case CPU_DOWN_PREPARE: 6683 cpuset_update_active_cpus(false); 6684 break; 6685 case CPU_DOWN_PREPARE_FROZEN: 6686 num_cpus_frozen++; 6687 partition_sched_domains(1, NULL, NULL); 6688 break; 6689 default: 6690 return NOTIFY_DONE; 6691 } 6692 return NOTIFY_OK; 6693} 6694 6695void __init sched_init_smp(void) 6696{ 6697 cpumask_var_t non_isolated_cpus; 6698 6699 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6700 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6701 6702 sched_init_numa(); 6703 6704 /* 6705 * There's no userspace yet to cause hotplug operations; hence all the 6706 * cpu masks are stable and all blatant races in the below code cannot 6707 * happen. 6708 */ 6709 mutex_lock(&sched_domains_mutex); 6710 init_sched_domains(cpu_active_mask); 6711 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6712 if (cpumask_empty(non_isolated_cpus)) 6713 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6714 mutex_unlock(&sched_domains_mutex); 6715 6716 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6717 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6718 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6719 6720 init_hrtick(); 6721 6722 /* Move init over to a non-isolated CPU */ 6723 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6724 BUG(); 6725 sched_init_granularity(); 6726 free_cpumask_var(non_isolated_cpus); 6727 6728 init_sched_rt_class(); 6729 init_sched_dl_class(); 6730} 6731#else 6732void __init sched_init_smp(void) 6733{ 6734 sched_init_granularity(); 6735} 6736#endif /* CONFIG_SMP */ 6737 6738const_debug unsigned int sysctl_timer_migration = 1; 6739 6740int in_sched_functions(unsigned long addr) 6741{ 6742 return in_lock_functions(addr) || 6743 (addr >= (unsigned long)__sched_text_start 6744 && addr < (unsigned long)__sched_text_end); 6745} 6746 6747#ifdef CONFIG_CGROUP_SCHED 6748/* 6749 * Default task group. 6750 * Every task in system belongs to this group at bootup. 6751 */ 6752struct task_group root_task_group; 6753LIST_HEAD(task_groups); 6754#endif 6755 6756DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6757 6758void __init sched_init(void) 6759{ 6760 int i, j; 6761 unsigned long alloc_size = 0, ptr; 6762 6763#ifdef CONFIG_FAIR_GROUP_SCHED 6764 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6765#endif 6766#ifdef CONFIG_RT_GROUP_SCHED 6767 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6768#endif 6769#ifdef CONFIG_CPUMASK_OFFSTACK 6770 alloc_size += num_possible_cpus() * cpumask_size(); 6771#endif 6772 if (alloc_size) { 6773 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6774 6775#ifdef CONFIG_FAIR_GROUP_SCHED 6776 root_task_group.se = (struct sched_entity **)ptr; 6777 ptr += nr_cpu_ids * sizeof(void **); 6778 6779 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6780 ptr += nr_cpu_ids * sizeof(void **); 6781 6782#endif /* CONFIG_FAIR_GROUP_SCHED */ 6783#ifdef CONFIG_RT_GROUP_SCHED 6784 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6785 ptr += nr_cpu_ids * sizeof(void **); 6786 6787 root_task_group.rt_rq = (struct rt_rq **)ptr; 6788 ptr += nr_cpu_ids * sizeof(void **); 6789 6790#endif /* CONFIG_RT_GROUP_SCHED */ 6791#ifdef CONFIG_CPUMASK_OFFSTACK 6792 for_each_possible_cpu(i) { 6793 per_cpu(load_balance_mask, i) = (void *)ptr; 6794 ptr += cpumask_size(); 6795 } 6796#endif /* CONFIG_CPUMASK_OFFSTACK */ 6797 } 6798 6799 init_rt_bandwidth(&def_rt_bandwidth, 6800 global_rt_period(), global_rt_runtime()); 6801 init_dl_bandwidth(&def_dl_bandwidth, 6802 global_rt_period(), global_rt_runtime()); 6803 6804#ifdef CONFIG_SMP 6805 init_defrootdomain(); 6806#endif 6807 6808#ifdef CONFIG_RT_GROUP_SCHED 6809 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6810 global_rt_period(), global_rt_runtime()); 6811#endif /* CONFIG_RT_GROUP_SCHED */ 6812 6813#ifdef CONFIG_CGROUP_SCHED 6814 list_add(&root_task_group.list, &task_groups); 6815 INIT_LIST_HEAD(&root_task_group.children); 6816 INIT_LIST_HEAD(&root_task_group.siblings); 6817 autogroup_init(&init_task); 6818 6819#endif /* CONFIG_CGROUP_SCHED */ 6820 6821 for_each_possible_cpu(i) { 6822 struct rq *rq; 6823 6824 rq = cpu_rq(i); 6825 raw_spin_lock_init(&rq->lock); 6826 rq->nr_running = 0; 6827 rq->calc_load_active = 0; 6828 rq->calc_load_update = jiffies + LOAD_FREQ; 6829 init_cfs_rq(&rq->cfs); 6830 init_rt_rq(&rq->rt, rq); 6831 init_dl_rq(&rq->dl, rq); 6832#ifdef CONFIG_FAIR_GROUP_SCHED 6833 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6834 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6835 /* 6836 * How much cpu bandwidth does root_task_group get? 6837 * 6838 * In case of task-groups formed thr' the cgroup filesystem, it 6839 * gets 100% of the cpu resources in the system. This overall 6840 * system cpu resource is divided among the tasks of 6841 * root_task_group and its child task-groups in a fair manner, 6842 * based on each entity's (task or task-group's) weight 6843 * (se->load.weight). 6844 * 6845 * In other words, if root_task_group has 10 tasks of weight 6846 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6847 * then A0's share of the cpu resource is: 6848 * 6849 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6850 * 6851 * We achieve this by letting root_task_group's tasks sit 6852 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6853 */ 6854 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6855 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6856#endif /* CONFIG_FAIR_GROUP_SCHED */ 6857 6858 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6859#ifdef CONFIG_RT_GROUP_SCHED 6860 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6861 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6862#endif 6863 6864 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6865 rq->cpu_load[j] = 0; 6866 6867 rq->last_load_update_tick = jiffies; 6868 6869#ifdef CONFIG_SMP 6870 rq->sd = NULL; 6871 rq->rd = NULL; 6872 rq->cpu_power = SCHED_POWER_SCALE; 6873 rq->post_schedule = 0; 6874 rq->active_balance = 0; 6875 rq->next_balance = jiffies; 6876 rq->push_cpu = 0; 6877 rq->cpu = i; 6878 rq->online = 0; 6879 rq->idle_stamp = 0; 6880 rq->avg_idle = 2*sysctl_sched_migration_cost; 6881 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6882 6883 INIT_LIST_HEAD(&rq->cfs_tasks); 6884 6885 rq_attach_root(rq, &def_root_domain); 6886#ifdef CONFIG_NO_HZ_COMMON 6887 rq->nohz_flags = 0; 6888#endif 6889#ifdef CONFIG_NO_HZ_FULL 6890 rq->last_sched_tick = 0; 6891#endif 6892#endif 6893 init_rq_hrtick(rq); 6894 atomic_set(&rq->nr_iowait, 0); 6895 } 6896 6897 set_load_weight(&init_task); 6898 6899#ifdef CONFIG_PREEMPT_NOTIFIERS 6900 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6901#endif 6902 6903 /* 6904 * The boot idle thread does lazy MMU switching as well: 6905 */ 6906 atomic_inc(&init_mm.mm_count); 6907 enter_lazy_tlb(&init_mm, current); 6908 6909 /* 6910 * Make us the idle thread. Technically, schedule() should not be 6911 * called from this thread, however somewhere below it might be, 6912 * but because we are the idle thread, we just pick up running again 6913 * when this runqueue becomes "idle". 6914 */ 6915 init_idle(current, smp_processor_id()); 6916 6917 calc_load_update = jiffies + LOAD_FREQ; 6918 6919 /* 6920 * During early bootup we pretend to be a normal task: 6921 */ 6922 current->sched_class = &fair_sched_class; 6923 6924#ifdef CONFIG_SMP 6925 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6926 /* May be allocated at isolcpus cmdline parse time */ 6927 if (cpu_isolated_map == NULL) 6928 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6929 idle_thread_set_boot_cpu(); 6930#endif 6931 init_sched_fair_class(); 6932 6933 scheduler_running = 1; 6934} 6935 6936#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6937static inline int preempt_count_equals(int preempt_offset) 6938{ 6939 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6940 6941 return (nested == preempt_offset); 6942} 6943 6944void __might_sleep(const char *file, int line, int preempt_offset) 6945{ 6946 static unsigned long prev_jiffy; /* ratelimiting */ 6947 6948 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6949 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 6950 !is_idle_task(current)) || 6951 system_state != SYSTEM_RUNNING || oops_in_progress) 6952 return; 6953 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6954 return; 6955 prev_jiffy = jiffies; 6956 6957 printk(KERN_ERR 6958 "BUG: sleeping function called from invalid context at %s:%d\n", 6959 file, line); 6960 printk(KERN_ERR 6961 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6962 in_atomic(), irqs_disabled(), 6963 current->pid, current->comm); 6964 6965 debug_show_held_locks(current); 6966 if (irqs_disabled()) 6967 print_irqtrace_events(current); 6968#ifdef CONFIG_DEBUG_PREEMPT 6969 if (!preempt_count_equals(preempt_offset)) { 6970 pr_err("Preemption disabled at:"); 6971 print_ip_sym(current->preempt_disable_ip); 6972 pr_cont("\n"); 6973 } 6974#endif 6975 dump_stack(); 6976} 6977EXPORT_SYMBOL(__might_sleep); 6978#endif 6979 6980#ifdef CONFIG_MAGIC_SYSRQ 6981static void normalize_task(struct rq *rq, struct task_struct *p) 6982{ 6983 const struct sched_class *prev_class = p->sched_class; 6984 struct sched_attr attr = { 6985 .sched_policy = SCHED_NORMAL, 6986 }; 6987 int old_prio = p->prio; 6988 int on_rq; 6989 6990 on_rq = p->on_rq; 6991 if (on_rq) 6992 dequeue_task(rq, p, 0); 6993 __setscheduler(rq, p, &attr); 6994 if (on_rq) { 6995 enqueue_task(rq, p, 0); 6996 resched_task(rq->curr); 6997 } 6998 6999 check_class_changed(rq, p, prev_class, old_prio); 7000} 7001 7002void normalize_rt_tasks(void) 7003{ 7004 struct task_struct *g, *p; 7005 unsigned long flags; 7006 struct rq *rq; 7007 7008 read_lock_irqsave(&tasklist_lock, flags); 7009 do_each_thread(g, p) { 7010 /* 7011 * Only normalize user tasks: 7012 */ 7013 if (!p->mm) 7014 continue; 7015 7016 p->se.exec_start = 0; 7017#ifdef CONFIG_SCHEDSTATS 7018 p->se.statistics.wait_start = 0; 7019 p->se.statistics.sleep_start = 0; 7020 p->se.statistics.block_start = 0; 7021#endif 7022 7023 if (!dl_task(p) && !rt_task(p)) { 7024 /* 7025 * Renice negative nice level userspace 7026 * tasks back to 0: 7027 */ 7028 if (task_nice(p) < 0 && p->mm) 7029 set_user_nice(p, 0); 7030 continue; 7031 } 7032 7033 raw_spin_lock(&p->pi_lock); 7034 rq = __task_rq_lock(p); 7035 7036 normalize_task(rq, p); 7037 7038 __task_rq_unlock(rq); 7039 raw_spin_unlock(&p->pi_lock); 7040 } while_each_thread(g, p); 7041 7042 read_unlock_irqrestore(&tasklist_lock, flags); 7043} 7044 7045#endif /* CONFIG_MAGIC_SYSRQ */ 7046 7047#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7048/* 7049 * These functions are only useful for the IA64 MCA handling, or kdb. 7050 * 7051 * They can only be called when the whole system has been 7052 * stopped - every CPU needs to be quiescent, and no scheduling 7053 * activity can take place. Using them for anything else would 7054 * be a serious bug, and as a result, they aren't even visible 7055 * under any other configuration. 7056 */ 7057 7058/** 7059 * curr_task - return the current task for a given cpu. 7060 * @cpu: the processor in question. 7061 * 7062 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7063 * 7064 * Return: The current task for @cpu. 7065 */ 7066struct task_struct *curr_task(int cpu) 7067{ 7068 return cpu_curr(cpu); 7069} 7070 7071#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7072 7073#ifdef CONFIG_IA64 7074/** 7075 * set_curr_task - set the current task for a given cpu. 7076 * @cpu: the processor in question. 7077 * @p: the task pointer to set. 7078 * 7079 * Description: This function must only be used when non-maskable interrupts 7080 * are serviced on a separate stack. It allows the architecture to switch the 7081 * notion of the current task on a cpu in a non-blocking manner. This function 7082 * must be called with all CPU's synchronized, and interrupts disabled, the 7083 * and caller must save the original value of the current task (see 7084 * curr_task() above) and restore that value before reenabling interrupts and 7085 * re-starting the system. 7086 * 7087 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7088 */ 7089void set_curr_task(int cpu, struct task_struct *p) 7090{ 7091 cpu_curr(cpu) = p; 7092} 7093 7094#endif 7095 7096#ifdef CONFIG_CGROUP_SCHED 7097/* task_group_lock serializes the addition/removal of task groups */ 7098static DEFINE_SPINLOCK(task_group_lock); 7099 7100static void free_sched_group(struct task_group *tg) 7101{ 7102 free_fair_sched_group(tg); 7103 free_rt_sched_group(tg); 7104 autogroup_free(tg); 7105 kfree(tg); 7106} 7107 7108/* allocate runqueue etc for a new task group */ 7109struct task_group *sched_create_group(struct task_group *parent) 7110{ 7111 struct task_group *tg; 7112 7113 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7114 if (!tg) 7115 return ERR_PTR(-ENOMEM); 7116 7117 if (!alloc_fair_sched_group(tg, parent)) 7118 goto err; 7119 7120 if (!alloc_rt_sched_group(tg, parent)) 7121 goto err; 7122 7123 return tg; 7124 7125err: 7126 free_sched_group(tg); 7127 return ERR_PTR(-ENOMEM); 7128} 7129 7130void sched_online_group(struct task_group *tg, struct task_group *parent) 7131{ 7132 unsigned long flags; 7133 7134 spin_lock_irqsave(&task_group_lock, flags); 7135 list_add_rcu(&tg->list, &task_groups); 7136 7137 WARN_ON(!parent); /* root should already exist */ 7138 7139 tg->parent = parent; 7140 INIT_LIST_HEAD(&tg->children); 7141 list_add_rcu(&tg->siblings, &parent->children); 7142 spin_unlock_irqrestore(&task_group_lock, flags); 7143} 7144 7145/* rcu callback to free various structures associated with a task group */ 7146static void free_sched_group_rcu(struct rcu_head *rhp) 7147{ 7148 /* now it should be safe to free those cfs_rqs */ 7149 free_sched_group(container_of(rhp, struct task_group, rcu)); 7150} 7151 7152/* Destroy runqueue etc associated with a task group */ 7153void sched_destroy_group(struct task_group *tg) 7154{ 7155 /* wait for possible concurrent references to cfs_rqs complete */ 7156 call_rcu(&tg->rcu, free_sched_group_rcu); 7157} 7158 7159void sched_offline_group(struct task_group *tg) 7160{ 7161 unsigned long flags; 7162 int i; 7163 7164 /* end participation in shares distribution */ 7165 for_each_possible_cpu(i) 7166 unregister_fair_sched_group(tg, i); 7167 7168 spin_lock_irqsave(&task_group_lock, flags); 7169 list_del_rcu(&tg->list); 7170 list_del_rcu(&tg->siblings); 7171 spin_unlock_irqrestore(&task_group_lock, flags); 7172} 7173 7174/* change task's runqueue when it moves between groups. 7175 * The caller of this function should have put the task in its new group 7176 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7177 * reflect its new group. 7178 */ 7179void sched_move_task(struct task_struct *tsk) 7180{ 7181 struct task_group *tg; 7182 int on_rq, running; 7183 unsigned long flags; 7184 struct rq *rq; 7185 7186 rq = task_rq_lock(tsk, &flags); 7187 7188 running = task_current(rq, tsk); 7189 on_rq = tsk->on_rq; 7190 7191 if (on_rq) 7192 dequeue_task(rq, tsk, 0); 7193 if (unlikely(running)) 7194 tsk->sched_class->put_prev_task(rq, tsk); 7195 7196 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 7197 lockdep_is_held(&tsk->sighand->siglock)), 7198 struct task_group, css); 7199 tg = autogroup_task_group(tsk, tg); 7200 tsk->sched_task_group = tg; 7201 7202#ifdef CONFIG_FAIR_GROUP_SCHED 7203 if (tsk->sched_class->task_move_group) 7204 tsk->sched_class->task_move_group(tsk, on_rq); 7205 else 7206#endif 7207 set_task_rq(tsk, task_cpu(tsk)); 7208 7209 if (unlikely(running)) 7210 tsk->sched_class->set_curr_task(rq); 7211 if (on_rq) 7212 enqueue_task(rq, tsk, 0); 7213 7214 task_rq_unlock(rq, tsk, &flags); 7215} 7216#endif /* CONFIG_CGROUP_SCHED */ 7217 7218#ifdef CONFIG_RT_GROUP_SCHED 7219/* 7220 * Ensure that the real time constraints are schedulable. 7221 */ 7222static DEFINE_MUTEX(rt_constraints_mutex); 7223 7224/* Must be called with tasklist_lock held */ 7225static inline int tg_has_rt_tasks(struct task_group *tg) 7226{ 7227 struct task_struct *g, *p; 7228 7229 do_each_thread(g, p) { 7230 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7231 return 1; 7232 } while_each_thread(g, p); 7233 7234 return 0; 7235} 7236 7237struct rt_schedulable_data { 7238 struct task_group *tg; 7239 u64 rt_period; 7240 u64 rt_runtime; 7241}; 7242 7243static int tg_rt_schedulable(struct task_group *tg, void *data) 7244{ 7245 struct rt_schedulable_data *d = data; 7246 struct task_group *child; 7247 unsigned long total, sum = 0; 7248 u64 period, runtime; 7249 7250 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7251 runtime = tg->rt_bandwidth.rt_runtime; 7252 7253 if (tg == d->tg) { 7254 period = d->rt_period; 7255 runtime = d->rt_runtime; 7256 } 7257 7258 /* 7259 * Cannot have more runtime than the period. 7260 */ 7261 if (runtime > period && runtime != RUNTIME_INF) 7262 return -EINVAL; 7263 7264 /* 7265 * Ensure we don't starve existing RT tasks. 7266 */ 7267 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7268 return -EBUSY; 7269 7270 total = to_ratio(period, runtime); 7271 7272 /* 7273 * Nobody can have more than the global setting allows. 7274 */ 7275 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7276 return -EINVAL; 7277 7278 /* 7279 * The sum of our children's runtime should not exceed our own. 7280 */ 7281 list_for_each_entry_rcu(child, &tg->children, siblings) { 7282 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7283 runtime = child->rt_bandwidth.rt_runtime; 7284 7285 if (child == d->tg) { 7286 period = d->rt_period; 7287 runtime = d->rt_runtime; 7288 } 7289 7290 sum += to_ratio(period, runtime); 7291 } 7292 7293 if (sum > total) 7294 return -EINVAL; 7295 7296 return 0; 7297} 7298 7299static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7300{ 7301 int ret; 7302 7303 struct rt_schedulable_data data = { 7304 .tg = tg, 7305 .rt_period = period, 7306 .rt_runtime = runtime, 7307 }; 7308 7309 rcu_read_lock(); 7310 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7311 rcu_read_unlock(); 7312 7313 return ret; 7314} 7315 7316static int tg_set_rt_bandwidth(struct task_group *tg, 7317 u64 rt_period, u64 rt_runtime) 7318{ 7319 int i, err = 0; 7320 7321 mutex_lock(&rt_constraints_mutex); 7322 read_lock(&tasklist_lock); 7323 err = __rt_schedulable(tg, rt_period, rt_runtime); 7324 if (err) 7325 goto unlock; 7326 7327 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7328 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7329 tg->rt_bandwidth.rt_runtime = rt_runtime; 7330 7331 for_each_possible_cpu(i) { 7332 struct rt_rq *rt_rq = tg->rt_rq[i]; 7333 7334 raw_spin_lock(&rt_rq->rt_runtime_lock); 7335 rt_rq->rt_runtime = rt_runtime; 7336 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7337 } 7338 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7339unlock: 7340 read_unlock(&tasklist_lock); 7341 mutex_unlock(&rt_constraints_mutex); 7342 7343 return err; 7344} 7345 7346static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7347{ 7348 u64 rt_runtime, rt_period; 7349 7350 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7351 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7352 if (rt_runtime_us < 0) 7353 rt_runtime = RUNTIME_INF; 7354 7355 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7356} 7357 7358static long sched_group_rt_runtime(struct task_group *tg) 7359{ 7360 u64 rt_runtime_us; 7361 7362 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7363 return -1; 7364 7365 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7366 do_div(rt_runtime_us, NSEC_PER_USEC); 7367 return rt_runtime_us; 7368} 7369 7370static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7371{ 7372 u64 rt_runtime, rt_period; 7373 7374 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7375 rt_runtime = tg->rt_bandwidth.rt_runtime; 7376 7377 if (rt_period == 0) 7378 return -EINVAL; 7379 7380 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7381} 7382 7383static long sched_group_rt_period(struct task_group *tg) 7384{ 7385 u64 rt_period_us; 7386 7387 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7388 do_div(rt_period_us, NSEC_PER_USEC); 7389 return rt_period_us; 7390} 7391#endif /* CONFIG_RT_GROUP_SCHED */ 7392 7393#ifdef CONFIG_RT_GROUP_SCHED 7394static int sched_rt_global_constraints(void) 7395{ 7396 int ret = 0; 7397 7398 mutex_lock(&rt_constraints_mutex); 7399 read_lock(&tasklist_lock); 7400 ret = __rt_schedulable(NULL, 0, 0); 7401 read_unlock(&tasklist_lock); 7402 mutex_unlock(&rt_constraints_mutex); 7403 7404 return ret; 7405} 7406 7407static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7408{ 7409 /* Don't accept realtime tasks when there is no way for them to run */ 7410 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7411 return 0; 7412 7413 return 1; 7414} 7415 7416#else /* !CONFIG_RT_GROUP_SCHED */ 7417static int sched_rt_global_constraints(void) 7418{ 7419 unsigned long flags; 7420 int i, ret = 0; 7421 7422 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7423 for_each_possible_cpu(i) { 7424 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7425 7426 raw_spin_lock(&rt_rq->rt_runtime_lock); 7427 rt_rq->rt_runtime = global_rt_runtime(); 7428 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7429 } 7430 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7431 7432 return ret; 7433} 7434#endif /* CONFIG_RT_GROUP_SCHED */ 7435 7436static int sched_dl_global_constraints(void) 7437{ 7438 u64 runtime = global_rt_runtime(); 7439 u64 period = global_rt_period(); 7440 u64 new_bw = to_ratio(period, runtime); 7441 int cpu, ret = 0; 7442 7443 /* 7444 * Here we want to check the bandwidth not being set to some 7445 * value smaller than the currently allocated bandwidth in 7446 * any of the root_domains. 7447 * 7448 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7449 * cycling on root_domains... Discussion on different/better 7450 * solutions is welcome! 7451 */ 7452 for_each_possible_cpu(cpu) { 7453 struct dl_bw *dl_b = dl_bw_of(cpu); 7454 7455 raw_spin_lock(&dl_b->lock); 7456 if (new_bw < dl_b->total_bw) 7457 ret = -EBUSY; 7458 raw_spin_unlock(&dl_b->lock); 7459 7460 if (ret) 7461 break; 7462 } 7463 7464 return ret; 7465} 7466 7467static void sched_dl_do_global(void) 7468{ 7469 u64 new_bw = -1; 7470 int cpu; 7471 7472 def_dl_bandwidth.dl_period = global_rt_period(); 7473 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7474 7475 if (global_rt_runtime() != RUNTIME_INF) 7476 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7477 7478 /* 7479 * FIXME: As above... 7480 */ 7481 for_each_possible_cpu(cpu) { 7482 struct dl_bw *dl_b = dl_bw_of(cpu); 7483 7484 raw_spin_lock(&dl_b->lock); 7485 dl_b->bw = new_bw; 7486 raw_spin_unlock(&dl_b->lock); 7487 } 7488} 7489 7490static int sched_rt_global_validate(void) 7491{ 7492 if (sysctl_sched_rt_period <= 0) 7493 return -EINVAL; 7494 7495 if (sysctl_sched_rt_runtime > sysctl_sched_rt_period) 7496 return -EINVAL; 7497 7498 return 0; 7499} 7500 7501static void sched_rt_do_global(void) 7502{ 7503 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7504 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7505} 7506 7507int sched_rt_handler(struct ctl_table *table, int write, 7508 void __user *buffer, size_t *lenp, 7509 loff_t *ppos) 7510{ 7511 int old_period, old_runtime; 7512 static DEFINE_MUTEX(mutex); 7513 int ret; 7514 7515 mutex_lock(&mutex); 7516 old_period = sysctl_sched_rt_period; 7517 old_runtime = sysctl_sched_rt_runtime; 7518 7519 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7520 7521 if (!ret && write) { 7522 ret = sched_rt_global_validate(); 7523 if (ret) 7524 goto undo; 7525 7526 ret = sched_rt_global_constraints(); 7527 if (ret) 7528 goto undo; 7529 7530 ret = sched_dl_global_constraints(); 7531 if (ret) 7532 goto undo; 7533 7534 sched_rt_do_global(); 7535 sched_dl_do_global(); 7536 } 7537 if (0) { 7538undo: 7539 sysctl_sched_rt_period = old_period; 7540 sysctl_sched_rt_runtime = old_runtime; 7541 } 7542 mutex_unlock(&mutex); 7543 7544 return ret; 7545} 7546 7547int sched_rr_handler(struct ctl_table *table, int write, 7548 void __user *buffer, size_t *lenp, 7549 loff_t *ppos) 7550{ 7551 int ret; 7552 static DEFINE_MUTEX(mutex); 7553 7554 mutex_lock(&mutex); 7555 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7556 /* make sure that internally we keep jiffies */ 7557 /* also, writing zero resets timeslice to default */ 7558 if (!ret && write) { 7559 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7560 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7561 } 7562 mutex_unlock(&mutex); 7563 return ret; 7564} 7565 7566#ifdef CONFIG_CGROUP_SCHED 7567 7568static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7569{ 7570 return css ? container_of(css, struct task_group, css) : NULL; 7571} 7572 7573static struct cgroup_subsys_state * 7574cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7575{ 7576 struct task_group *parent = css_tg(parent_css); 7577 struct task_group *tg; 7578 7579 if (!parent) { 7580 /* This is early initialization for the top cgroup */ 7581 return &root_task_group.css; 7582 } 7583 7584 tg = sched_create_group(parent); 7585 if (IS_ERR(tg)) 7586 return ERR_PTR(-ENOMEM); 7587 7588 return &tg->css; 7589} 7590 7591static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7592{ 7593 struct task_group *tg = css_tg(css); 7594 struct task_group *parent = css_tg(css_parent(css)); 7595 7596 if (parent) 7597 sched_online_group(tg, parent); 7598 return 0; 7599} 7600 7601static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7602{ 7603 struct task_group *tg = css_tg(css); 7604 7605 sched_destroy_group(tg); 7606} 7607 7608static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7609{ 7610 struct task_group *tg = css_tg(css); 7611 7612 sched_offline_group(tg); 7613} 7614 7615static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7616 struct cgroup_taskset *tset) 7617{ 7618 struct task_struct *task; 7619 7620 cgroup_taskset_for_each(task, css, tset) { 7621#ifdef CONFIG_RT_GROUP_SCHED 7622 if (!sched_rt_can_attach(css_tg(css), task)) 7623 return -EINVAL; 7624#else 7625 /* We don't support RT-tasks being in separate groups */ 7626 if (task->sched_class != &fair_sched_class) 7627 return -EINVAL; 7628#endif 7629 } 7630 return 0; 7631} 7632 7633static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7634 struct cgroup_taskset *tset) 7635{ 7636 struct task_struct *task; 7637 7638 cgroup_taskset_for_each(task, css, tset) 7639 sched_move_task(task); 7640} 7641 7642static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7643 struct cgroup_subsys_state *old_css, 7644 struct task_struct *task) 7645{ 7646 /* 7647 * cgroup_exit() is called in the copy_process() failure path. 7648 * Ignore this case since the task hasn't ran yet, this avoids 7649 * trying to poke a half freed task state from generic code. 7650 */ 7651 if (!(task->flags & PF_EXITING)) 7652 return; 7653 7654 sched_move_task(task); 7655} 7656 7657#ifdef CONFIG_FAIR_GROUP_SCHED 7658static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7659 struct cftype *cftype, u64 shareval) 7660{ 7661 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7662} 7663 7664static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7665 struct cftype *cft) 7666{ 7667 struct task_group *tg = css_tg(css); 7668 7669 return (u64) scale_load_down(tg->shares); 7670} 7671 7672#ifdef CONFIG_CFS_BANDWIDTH 7673static DEFINE_MUTEX(cfs_constraints_mutex); 7674 7675const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7676const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7677 7678static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7679 7680static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7681{ 7682 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7683 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7684 7685 if (tg == &root_task_group) 7686 return -EINVAL; 7687 7688 /* 7689 * Ensure we have at some amount of bandwidth every period. This is 7690 * to prevent reaching a state of large arrears when throttled via 7691 * entity_tick() resulting in prolonged exit starvation. 7692 */ 7693 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7694 return -EINVAL; 7695 7696 /* 7697 * Likewise, bound things on the otherside by preventing insane quota 7698 * periods. This also allows us to normalize in computing quota 7699 * feasibility. 7700 */ 7701 if (period > max_cfs_quota_period) 7702 return -EINVAL; 7703 7704 mutex_lock(&cfs_constraints_mutex); 7705 ret = __cfs_schedulable(tg, period, quota); 7706 if (ret) 7707 goto out_unlock; 7708 7709 runtime_enabled = quota != RUNTIME_INF; 7710 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7711 /* 7712 * If we need to toggle cfs_bandwidth_used, off->on must occur 7713 * before making related changes, and on->off must occur afterwards 7714 */ 7715 if (runtime_enabled && !runtime_was_enabled) 7716 cfs_bandwidth_usage_inc(); 7717 raw_spin_lock_irq(&cfs_b->lock); 7718 cfs_b->period = ns_to_ktime(period); 7719 cfs_b->quota = quota; 7720 7721 __refill_cfs_bandwidth_runtime(cfs_b); 7722 /* restart the period timer (if active) to handle new period expiry */ 7723 if (runtime_enabled && cfs_b->timer_active) { 7724 /* force a reprogram */ 7725 cfs_b->timer_active = 0; 7726 __start_cfs_bandwidth(cfs_b); 7727 } 7728 raw_spin_unlock_irq(&cfs_b->lock); 7729 7730 for_each_possible_cpu(i) { 7731 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7732 struct rq *rq = cfs_rq->rq; 7733 7734 raw_spin_lock_irq(&rq->lock); 7735 cfs_rq->runtime_enabled = runtime_enabled; 7736 cfs_rq->runtime_remaining = 0; 7737 7738 if (cfs_rq->throttled) 7739 unthrottle_cfs_rq(cfs_rq); 7740 raw_spin_unlock_irq(&rq->lock); 7741 } 7742 if (runtime_was_enabled && !runtime_enabled) 7743 cfs_bandwidth_usage_dec(); 7744out_unlock: 7745 mutex_unlock(&cfs_constraints_mutex); 7746 7747 return ret; 7748} 7749 7750int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7751{ 7752 u64 quota, period; 7753 7754 period = ktime_to_ns(tg->cfs_bandwidth.period); 7755 if (cfs_quota_us < 0) 7756 quota = RUNTIME_INF; 7757 else 7758 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7759 7760 return tg_set_cfs_bandwidth(tg, period, quota); 7761} 7762 7763long tg_get_cfs_quota(struct task_group *tg) 7764{ 7765 u64 quota_us; 7766 7767 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7768 return -1; 7769 7770 quota_us = tg->cfs_bandwidth.quota; 7771 do_div(quota_us, NSEC_PER_USEC); 7772 7773 return quota_us; 7774} 7775 7776int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7777{ 7778 u64 quota, period; 7779 7780 period = (u64)cfs_period_us * NSEC_PER_USEC; 7781 quota = tg->cfs_bandwidth.quota; 7782 7783 return tg_set_cfs_bandwidth(tg, period, quota); 7784} 7785 7786long tg_get_cfs_period(struct task_group *tg) 7787{ 7788 u64 cfs_period_us; 7789 7790 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7791 do_div(cfs_period_us, NSEC_PER_USEC); 7792 7793 return cfs_period_us; 7794} 7795 7796static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7797 struct cftype *cft) 7798{ 7799 return tg_get_cfs_quota(css_tg(css)); 7800} 7801 7802static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7803 struct cftype *cftype, s64 cfs_quota_us) 7804{ 7805 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7806} 7807 7808static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7809 struct cftype *cft) 7810{ 7811 return tg_get_cfs_period(css_tg(css)); 7812} 7813 7814static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7815 struct cftype *cftype, u64 cfs_period_us) 7816{ 7817 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7818} 7819 7820struct cfs_schedulable_data { 7821 struct task_group *tg; 7822 u64 period, quota; 7823}; 7824 7825/* 7826 * normalize group quota/period to be quota/max_period 7827 * note: units are usecs 7828 */ 7829static u64 normalize_cfs_quota(struct task_group *tg, 7830 struct cfs_schedulable_data *d) 7831{ 7832 u64 quota, period; 7833 7834 if (tg == d->tg) { 7835 period = d->period; 7836 quota = d->quota; 7837 } else { 7838 period = tg_get_cfs_period(tg); 7839 quota = tg_get_cfs_quota(tg); 7840 } 7841 7842 /* note: these should typically be equivalent */ 7843 if (quota == RUNTIME_INF || quota == -1) 7844 return RUNTIME_INF; 7845 7846 return to_ratio(period, quota); 7847} 7848 7849static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7850{ 7851 struct cfs_schedulable_data *d = data; 7852 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7853 s64 quota = 0, parent_quota = -1; 7854 7855 if (!tg->parent) { 7856 quota = RUNTIME_INF; 7857 } else { 7858 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7859 7860 quota = normalize_cfs_quota(tg, d); 7861 parent_quota = parent_b->hierarchal_quota; 7862 7863 /* 7864 * ensure max(child_quota) <= parent_quota, inherit when no 7865 * limit is set 7866 */ 7867 if (quota == RUNTIME_INF) 7868 quota = parent_quota; 7869 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7870 return -EINVAL; 7871 } 7872 cfs_b->hierarchal_quota = quota; 7873 7874 return 0; 7875} 7876 7877static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7878{ 7879 int ret; 7880 struct cfs_schedulable_data data = { 7881 .tg = tg, 7882 .period = period, 7883 .quota = quota, 7884 }; 7885 7886 if (quota != RUNTIME_INF) { 7887 do_div(data.period, NSEC_PER_USEC); 7888 do_div(data.quota, NSEC_PER_USEC); 7889 } 7890 7891 rcu_read_lock(); 7892 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7893 rcu_read_unlock(); 7894 7895 return ret; 7896} 7897 7898static int cpu_stats_show(struct seq_file *sf, void *v) 7899{ 7900 struct task_group *tg = css_tg(seq_css(sf)); 7901 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7902 7903 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7904 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7905 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7906 7907 return 0; 7908} 7909#endif /* CONFIG_CFS_BANDWIDTH */ 7910#endif /* CONFIG_FAIR_GROUP_SCHED */ 7911 7912#ifdef CONFIG_RT_GROUP_SCHED 7913static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7914 struct cftype *cft, s64 val) 7915{ 7916 return sched_group_set_rt_runtime(css_tg(css), val); 7917} 7918 7919static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7920 struct cftype *cft) 7921{ 7922 return sched_group_rt_runtime(css_tg(css)); 7923} 7924 7925static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7926 struct cftype *cftype, u64 rt_period_us) 7927{ 7928 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7929} 7930 7931static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7932 struct cftype *cft) 7933{ 7934 return sched_group_rt_period(css_tg(css)); 7935} 7936#endif /* CONFIG_RT_GROUP_SCHED */ 7937 7938static struct cftype cpu_files[] = { 7939#ifdef CONFIG_FAIR_GROUP_SCHED 7940 { 7941 .name = "shares", 7942 .read_u64 = cpu_shares_read_u64, 7943 .write_u64 = cpu_shares_write_u64, 7944 }, 7945#endif 7946#ifdef CONFIG_CFS_BANDWIDTH 7947 { 7948 .name = "cfs_quota_us", 7949 .read_s64 = cpu_cfs_quota_read_s64, 7950 .write_s64 = cpu_cfs_quota_write_s64, 7951 }, 7952 { 7953 .name = "cfs_period_us", 7954 .read_u64 = cpu_cfs_period_read_u64, 7955 .write_u64 = cpu_cfs_period_write_u64, 7956 }, 7957 { 7958 .name = "stat", 7959 .seq_show = cpu_stats_show, 7960 }, 7961#endif 7962#ifdef CONFIG_RT_GROUP_SCHED 7963 { 7964 .name = "rt_runtime_us", 7965 .read_s64 = cpu_rt_runtime_read, 7966 .write_s64 = cpu_rt_runtime_write, 7967 }, 7968 { 7969 .name = "rt_period_us", 7970 .read_u64 = cpu_rt_period_read_uint, 7971 .write_u64 = cpu_rt_period_write_uint, 7972 }, 7973#endif 7974 { } /* terminate */ 7975}; 7976 7977struct cgroup_subsys cpu_cgroup_subsys = { 7978 .name = "cpu", 7979 .css_alloc = cpu_cgroup_css_alloc, 7980 .css_free = cpu_cgroup_css_free, 7981 .css_online = cpu_cgroup_css_online, 7982 .css_offline = cpu_cgroup_css_offline, 7983 .can_attach = cpu_cgroup_can_attach, 7984 .attach = cpu_cgroup_attach, 7985 .exit = cpu_cgroup_exit, 7986 .subsys_id = cpu_cgroup_subsys_id, 7987 .base_cftypes = cpu_files, 7988 .early_init = 1, 7989}; 7990 7991#endif /* CONFIG_CGROUP_SCHED */ 7992 7993void dump_cpu_task(int cpu) 7994{ 7995 pr_info("Task dump for CPU %d:\n", cpu); 7996 sched_show_task(cpu_curr(cpu)); 7997} 7998