core.c revision 6d35ab48090b10c5ea5604ed5d6e91f302dc6060
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 = NULL; 1749 p->numa_faults_buffer = NULL; 1750 1751 INIT_LIST_HEAD(&p->numa_entry); 1752 p->numa_group = NULL; 1753#endif /* CONFIG_NUMA_BALANCING */ 1754} 1755 1756#ifdef CONFIG_NUMA_BALANCING 1757#ifdef CONFIG_SCHED_DEBUG 1758void set_numabalancing_state(bool enabled) 1759{ 1760 if (enabled) 1761 sched_feat_set("NUMA"); 1762 else 1763 sched_feat_set("NO_NUMA"); 1764} 1765#else 1766__read_mostly bool numabalancing_enabled; 1767 1768void set_numabalancing_state(bool enabled) 1769{ 1770 numabalancing_enabled = enabled; 1771} 1772#endif /* CONFIG_SCHED_DEBUG */ 1773 1774#ifdef CONFIG_PROC_SYSCTL 1775int sysctl_numa_balancing(struct ctl_table *table, int write, 1776 void __user *buffer, size_t *lenp, loff_t *ppos) 1777{ 1778 struct ctl_table t; 1779 int err; 1780 int state = numabalancing_enabled; 1781 1782 if (write && !capable(CAP_SYS_ADMIN)) 1783 return -EPERM; 1784 1785 t = *table; 1786 t.data = &state; 1787 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1788 if (err < 0) 1789 return err; 1790 if (write) 1791 set_numabalancing_state(state); 1792 return err; 1793} 1794#endif 1795#endif 1796 1797/* 1798 * fork()/clone()-time setup: 1799 */ 1800int sched_fork(unsigned long clone_flags, struct task_struct *p) 1801{ 1802 unsigned long flags; 1803 int cpu = get_cpu(); 1804 1805 __sched_fork(clone_flags, p); 1806 /* 1807 * We mark the process as running here. This guarantees that 1808 * nobody will actually run it, and a signal or other external 1809 * event cannot wake it up and insert it on the runqueue either. 1810 */ 1811 p->state = TASK_RUNNING; 1812 1813 /* 1814 * Make sure we do not leak PI boosting priority to the child. 1815 */ 1816 p->prio = current->normal_prio; 1817 1818 /* 1819 * Revert to default priority/policy on fork if requested. 1820 */ 1821 if (unlikely(p->sched_reset_on_fork)) { 1822 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1823 p->policy = SCHED_NORMAL; 1824 p->static_prio = NICE_TO_PRIO(0); 1825 p->rt_priority = 0; 1826 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1827 p->static_prio = NICE_TO_PRIO(0); 1828 1829 p->prio = p->normal_prio = __normal_prio(p); 1830 set_load_weight(p); 1831 1832 /* 1833 * We don't need the reset flag anymore after the fork. It has 1834 * fulfilled its duty: 1835 */ 1836 p->sched_reset_on_fork = 0; 1837 } 1838 1839 if (dl_prio(p->prio)) { 1840 put_cpu(); 1841 return -EAGAIN; 1842 } else if (rt_prio(p->prio)) { 1843 p->sched_class = &rt_sched_class; 1844 } else { 1845 p->sched_class = &fair_sched_class; 1846 } 1847 1848 if (p->sched_class->task_fork) 1849 p->sched_class->task_fork(p); 1850 1851 /* 1852 * The child is not yet in the pid-hash so no cgroup attach races, 1853 * and the cgroup is pinned to this child due to cgroup_fork() 1854 * is ran before sched_fork(). 1855 * 1856 * Silence PROVE_RCU. 1857 */ 1858 raw_spin_lock_irqsave(&p->pi_lock, flags); 1859 set_task_cpu(p, cpu); 1860 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1861 1862#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1863 if (likely(sched_info_on())) 1864 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1865#endif 1866#if defined(CONFIG_SMP) 1867 p->on_cpu = 0; 1868#endif 1869 init_task_preempt_count(p); 1870#ifdef CONFIG_SMP 1871 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1872 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1873#endif 1874 1875 put_cpu(); 1876 return 0; 1877} 1878 1879unsigned long to_ratio(u64 period, u64 runtime) 1880{ 1881 if (runtime == RUNTIME_INF) 1882 return 1ULL << 20; 1883 1884 /* 1885 * Doing this here saves a lot of checks in all 1886 * the calling paths, and returning zero seems 1887 * safe for them anyway. 1888 */ 1889 if (period == 0) 1890 return 0; 1891 1892 return div64_u64(runtime << 20, period); 1893} 1894 1895#ifdef CONFIG_SMP 1896inline struct dl_bw *dl_bw_of(int i) 1897{ 1898 return &cpu_rq(i)->rd->dl_bw; 1899} 1900 1901static inline int dl_bw_cpus(int i) 1902{ 1903 struct root_domain *rd = cpu_rq(i)->rd; 1904 int cpus = 0; 1905 1906 for_each_cpu_and(i, rd->span, cpu_active_mask) 1907 cpus++; 1908 1909 return cpus; 1910} 1911#else 1912inline struct dl_bw *dl_bw_of(int i) 1913{ 1914 return &cpu_rq(i)->dl.dl_bw; 1915} 1916 1917static inline int dl_bw_cpus(int i) 1918{ 1919 return 1; 1920} 1921#endif 1922 1923static inline 1924void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1925{ 1926 dl_b->total_bw -= tsk_bw; 1927} 1928 1929static inline 1930void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1931{ 1932 dl_b->total_bw += tsk_bw; 1933} 1934 1935static inline 1936bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1937{ 1938 return dl_b->bw != -1 && 1939 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1940} 1941 1942/* 1943 * We must be sure that accepting a new task (or allowing changing the 1944 * parameters of an existing one) is consistent with the bandwidth 1945 * constraints. If yes, this function also accordingly updates the currently 1946 * allocated bandwidth to reflect the new situation. 1947 * 1948 * This function is called while holding p's rq->lock. 1949 */ 1950static int dl_overflow(struct task_struct *p, int policy, 1951 const struct sched_attr *attr) 1952{ 1953 1954 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1955 u64 period = attr->sched_period ?: attr->sched_deadline; 1956 u64 runtime = attr->sched_runtime; 1957 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1958 int cpus, err = -1; 1959 1960 if (new_bw == p->dl.dl_bw) 1961 return 0; 1962 1963 /* 1964 * Either if a task, enters, leave, or stays -deadline but changes 1965 * its parameters, we may need to update accordingly the total 1966 * allocated bandwidth of the container. 1967 */ 1968 raw_spin_lock(&dl_b->lock); 1969 cpus = dl_bw_cpus(task_cpu(p)); 1970 if (dl_policy(policy) && !task_has_dl_policy(p) && 1971 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1972 __dl_add(dl_b, new_bw); 1973 err = 0; 1974 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1975 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1976 __dl_clear(dl_b, p->dl.dl_bw); 1977 __dl_add(dl_b, new_bw); 1978 err = 0; 1979 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1980 __dl_clear(dl_b, p->dl.dl_bw); 1981 err = 0; 1982 } 1983 raw_spin_unlock(&dl_b->lock); 1984 1985 return err; 1986} 1987 1988extern void init_dl_bw(struct dl_bw *dl_b); 1989 1990/* 1991 * wake_up_new_task - wake up a newly created task for the first time. 1992 * 1993 * This function will do some initial scheduler statistics housekeeping 1994 * that must be done for every newly created context, then puts the task 1995 * on the runqueue and wakes it. 1996 */ 1997void wake_up_new_task(struct task_struct *p) 1998{ 1999 unsigned long flags; 2000 struct rq *rq; 2001 2002 raw_spin_lock_irqsave(&p->pi_lock, flags); 2003#ifdef CONFIG_SMP 2004 /* 2005 * Fork balancing, do it here and not earlier because: 2006 * - cpus_allowed can change in the fork path 2007 * - any previously selected cpu might disappear through hotplug 2008 */ 2009 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2010#endif 2011 2012 /* Initialize new task's runnable average */ 2013 init_task_runnable_average(p); 2014 rq = __task_rq_lock(p); 2015 activate_task(rq, p, 0); 2016 p->on_rq = 1; 2017 trace_sched_wakeup_new(p, true); 2018 check_preempt_curr(rq, p, WF_FORK); 2019#ifdef CONFIG_SMP 2020 if (p->sched_class->task_woken) 2021 p->sched_class->task_woken(rq, p); 2022#endif 2023 task_rq_unlock(rq, p, &flags); 2024} 2025 2026#ifdef CONFIG_PREEMPT_NOTIFIERS 2027 2028/** 2029 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2030 * @notifier: notifier struct to register 2031 */ 2032void preempt_notifier_register(struct preempt_notifier *notifier) 2033{ 2034 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2035} 2036EXPORT_SYMBOL_GPL(preempt_notifier_register); 2037 2038/** 2039 * preempt_notifier_unregister - no longer interested in preemption notifications 2040 * @notifier: notifier struct to unregister 2041 * 2042 * This is safe to call from within a preemption notifier. 2043 */ 2044void preempt_notifier_unregister(struct preempt_notifier *notifier) 2045{ 2046 hlist_del(¬ifier->link); 2047} 2048EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2049 2050static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2051{ 2052 struct preempt_notifier *notifier; 2053 2054 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2055 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2056} 2057 2058static void 2059fire_sched_out_preempt_notifiers(struct task_struct *curr, 2060 struct task_struct *next) 2061{ 2062 struct preempt_notifier *notifier; 2063 2064 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2065 notifier->ops->sched_out(notifier, next); 2066} 2067 2068#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2069 2070static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2071{ 2072} 2073 2074static void 2075fire_sched_out_preempt_notifiers(struct task_struct *curr, 2076 struct task_struct *next) 2077{ 2078} 2079 2080#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2081 2082/** 2083 * prepare_task_switch - prepare to switch tasks 2084 * @rq: the runqueue preparing to switch 2085 * @prev: the current task that is being switched out 2086 * @next: the task we are going to switch to. 2087 * 2088 * This is called with the rq lock held and interrupts off. It must 2089 * be paired with a subsequent finish_task_switch after the context 2090 * switch. 2091 * 2092 * prepare_task_switch sets up locking and calls architecture specific 2093 * hooks. 2094 */ 2095static inline void 2096prepare_task_switch(struct rq *rq, struct task_struct *prev, 2097 struct task_struct *next) 2098{ 2099 trace_sched_switch(prev, next); 2100 sched_info_switch(rq, prev, next); 2101 perf_event_task_sched_out(prev, next); 2102 fire_sched_out_preempt_notifiers(prev, next); 2103 prepare_lock_switch(rq, next); 2104 prepare_arch_switch(next); 2105} 2106 2107/** 2108 * finish_task_switch - clean up after a task-switch 2109 * @rq: runqueue associated with task-switch 2110 * @prev: the thread we just switched away from. 2111 * 2112 * finish_task_switch must be called after the context switch, paired 2113 * with a prepare_task_switch call before the context switch. 2114 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2115 * and do any other architecture-specific cleanup actions. 2116 * 2117 * Note that we may have delayed dropping an mm in context_switch(). If 2118 * so, we finish that here outside of the runqueue lock. (Doing it 2119 * with the lock held can cause deadlocks; see schedule() for 2120 * details.) 2121 */ 2122static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2123 __releases(rq->lock) 2124{ 2125 struct mm_struct *mm = rq->prev_mm; 2126 long prev_state; 2127 2128 rq->prev_mm = NULL; 2129 2130 /* 2131 * A task struct has one reference for the use as "current". 2132 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2133 * schedule one last time. The schedule call will never return, and 2134 * the scheduled task must drop that reference. 2135 * The test for TASK_DEAD must occur while the runqueue locks are 2136 * still held, otherwise prev could be scheduled on another cpu, die 2137 * there before we look at prev->state, and then the reference would 2138 * be dropped twice. 2139 * Manfred Spraul <manfred@colorfullife.com> 2140 */ 2141 prev_state = prev->state; 2142 vtime_task_switch(prev); 2143 finish_arch_switch(prev); 2144 perf_event_task_sched_in(prev, current); 2145 finish_lock_switch(rq, prev); 2146 finish_arch_post_lock_switch(); 2147 2148 fire_sched_in_preempt_notifiers(current); 2149 if (mm) 2150 mmdrop(mm); 2151 if (unlikely(prev_state == TASK_DEAD)) { 2152 task_numa_free(prev); 2153 2154 if (prev->sched_class->task_dead) 2155 prev->sched_class->task_dead(prev); 2156 2157 /* 2158 * Remove function-return probe instances associated with this 2159 * task and put them back on the free list. 2160 */ 2161 kprobe_flush_task(prev); 2162 put_task_struct(prev); 2163 } 2164 2165 tick_nohz_task_switch(current); 2166} 2167 2168#ifdef CONFIG_SMP 2169 2170/* assumes rq->lock is held */ 2171static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 2172{ 2173 if (prev->sched_class->pre_schedule) 2174 prev->sched_class->pre_schedule(rq, prev); 2175} 2176 2177/* rq->lock is NOT held, but preemption is disabled */ 2178static inline void post_schedule(struct rq *rq) 2179{ 2180 if (rq->post_schedule) { 2181 unsigned long flags; 2182 2183 raw_spin_lock_irqsave(&rq->lock, flags); 2184 if (rq->curr->sched_class->post_schedule) 2185 rq->curr->sched_class->post_schedule(rq); 2186 raw_spin_unlock_irqrestore(&rq->lock, flags); 2187 2188 rq->post_schedule = 0; 2189 } 2190} 2191 2192#else 2193 2194static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2195{ 2196} 2197 2198static inline void post_schedule(struct rq *rq) 2199{ 2200} 2201 2202#endif 2203 2204/** 2205 * schedule_tail - first thing a freshly forked thread must call. 2206 * @prev: the thread we just switched away from. 2207 */ 2208asmlinkage void schedule_tail(struct task_struct *prev) 2209 __releases(rq->lock) 2210{ 2211 struct rq *rq = this_rq(); 2212 2213 finish_task_switch(rq, prev); 2214 2215 /* 2216 * FIXME: do we need to worry about rq being invalidated by the 2217 * task_switch? 2218 */ 2219 post_schedule(rq); 2220 2221#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2222 /* In this case, finish_task_switch does not reenable preemption */ 2223 preempt_enable(); 2224#endif 2225 if (current->set_child_tid) 2226 put_user(task_pid_vnr(current), current->set_child_tid); 2227} 2228 2229/* 2230 * context_switch - switch to the new MM and the new 2231 * thread's register state. 2232 */ 2233static inline void 2234context_switch(struct rq *rq, struct task_struct *prev, 2235 struct task_struct *next) 2236{ 2237 struct mm_struct *mm, *oldmm; 2238 2239 prepare_task_switch(rq, prev, next); 2240 2241 mm = next->mm; 2242 oldmm = prev->active_mm; 2243 /* 2244 * For paravirt, this is coupled with an exit in switch_to to 2245 * combine the page table reload and the switch backend into 2246 * one hypercall. 2247 */ 2248 arch_start_context_switch(prev); 2249 2250 if (!mm) { 2251 next->active_mm = oldmm; 2252 atomic_inc(&oldmm->mm_count); 2253 enter_lazy_tlb(oldmm, next); 2254 } else 2255 switch_mm(oldmm, mm, next); 2256 2257 if (!prev->mm) { 2258 prev->active_mm = NULL; 2259 rq->prev_mm = oldmm; 2260 } 2261 /* 2262 * Since the runqueue lock will be released by the next 2263 * task (which is an invalid locking op but in the case 2264 * of the scheduler it's an obvious special-case), so we 2265 * do an early lockdep release here: 2266 */ 2267#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2268 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2269#endif 2270 2271 context_tracking_task_switch(prev, next); 2272 /* Here we just switch the register state and the stack. */ 2273 switch_to(prev, next, prev); 2274 2275 barrier(); 2276 /* 2277 * this_rq must be evaluated again because prev may have moved 2278 * CPUs since it called schedule(), thus the 'rq' on its stack 2279 * frame will be invalid. 2280 */ 2281 finish_task_switch(this_rq(), prev); 2282} 2283 2284/* 2285 * nr_running and nr_context_switches: 2286 * 2287 * externally visible scheduler statistics: current number of runnable 2288 * threads, total number of context switches performed since bootup. 2289 */ 2290unsigned long nr_running(void) 2291{ 2292 unsigned long i, sum = 0; 2293 2294 for_each_online_cpu(i) 2295 sum += cpu_rq(i)->nr_running; 2296 2297 return sum; 2298} 2299 2300unsigned long long nr_context_switches(void) 2301{ 2302 int i; 2303 unsigned long long sum = 0; 2304 2305 for_each_possible_cpu(i) 2306 sum += cpu_rq(i)->nr_switches; 2307 2308 return sum; 2309} 2310 2311unsigned long nr_iowait(void) 2312{ 2313 unsigned long i, sum = 0; 2314 2315 for_each_possible_cpu(i) 2316 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2317 2318 return sum; 2319} 2320 2321unsigned long nr_iowait_cpu(int cpu) 2322{ 2323 struct rq *this = cpu_rq(cpu); 2324 return atomic_read(&this->nr_iowait); 2325} 2326 2327#ifdef CONFIG_SMP 2328 2329/* 2330 * sched_exec - execve() is a valuable balancing opportunity, because at 2331 * this point the task has the smallest effective memory and cache footprint. 2332 */ 2333void sched_exec(void) 2334{ 2335 struct task_struct *p = current; 2336 unsigned long flags; 2337 int dest_cpu; 2338 2339 raw_spin_lock_irqsave(&p->pi_lock, flags); 2340 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2341 if (dest_cpu == smp_processor_id()) 2342 goto unlock; 2343 2344 if (likely(cpu_active(dest_cpu))) { 2345 struct migration_arg arg = { p, dest_cpu }; 2346 2347 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2348 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2349 return; 2350 } 2351unlock: 2352 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2353} 2354 2355#endif 2356 2357DEFINE_PER_CPU(struct kernel_stat, kstat); 2358DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2359 2360EXPORT_PER_CPU_SYMBOL(kstat); 2361EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2362 2363/* 2364 * Return any ns on the sched_clock that have not yet been accounted in 2365 * @p in case that task is currently running. 2366 * 2367 * Called with task_rq_lock() held on @rq. 2368 */ 2369static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2370{ 2371 u64 ns = 0; 2372 2373 if (task_current(rq, p)) { 2374 update_rq_clock(rq); 2375 ns = rq_clock_task(rq) - p->se.exec_start; 2376 if ((s64)ns < 0) 2377 ns = 0; 2378 } 2379 2380 return ns; 2381} 2382 2383unsigned long long task_delta_exec(struct task_struct *p) 2384{ 2385 unsigned long flags; 2386 struct rq *rq; 2387 u64 ns = 0; 2388 2389 rq = task_rq_lock(p, &flags); 2390 ns = do_task_delta_exec(p, rq); 2391 task_rq_unlock(rq, p, &flags); 2392 2393 return ns; 2394} 2395 2396/* 2397 * Return accounted runtime for the task. 2398 * In case the task is currently running, return the runtime plus current's 2399 * pending runtime that have not been accounted yet. 2400 */ 2401unsigned long long task_sched_runtime(struct task_struct *p) 2402{ 2403 unsigned long flags; 2404 struct rq *rq; 2405 u64 ns = 0; 2406 2407#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2408 /* 2409 * 64-bit doesn't need locks to atomically read a 64bit value. 2410 * So we have a optimization chance when the task's delta_exec is 0. 2411 * Reading ->on_cpu is racy, but this is ok. 2412 * 2413 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2414 * If we race with it entering cpu, unaccounted time is 0. This is 2415 * indistinguishable from the read occurring a few cycles earlier. 2416 */ 2417 if (!p->on_cpu) 2418 return p->se.sum_exec_runtime; 2419#endif 2420 2421 rq = task_rq_lock(p, &flags); 2422 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2423 task_rq_unlock(rq, p, &flags); 2424 2425 return ns; 2426} 2427 2428/* 2429 * This function gets called by the timer code, with HZ frequency. 2430 * We call it with interrupts disabled. 2431 */ 2432void scheduler_tick(void) 2433{ 2434 int cpu = smp_processor_id(); 2435 struct rq *rq = cpu_rq(cpu); 2436 struct task_struct *curr = rq->curr; 2437 2438 sched_clock_tick(); 2439 2440 raw_spin_lock(&rq->lock); 2441 update_rq_clock(rq); 2442 curr->sched_class->task_tick(rq, curr, 0); 2443 update_cpu_load_active(rq); 2444 raw_spin_unlock(&rq->lock); 2445 2446 perf_event_task_tick(); 2447 2448#ifdef CONFIG_SMP 2449 rq->idle_balance = idle_cpu(cpu); 2450 trigger_load_balance(rq); 2451#endif 2452 rq_last_tick_reset(rq); 2453} 2454 2455#ifdef CONFIG_NO_HZ_FULL 2456/** 2457 * scheduler_tick_max_deferment 2458 * 2459 * Keep at least one tick per second when a single 2460 * active task is running because the scheduler doesn't 2461 * yet completely support full dynticks environment. 2462 * 2463 * This makes sure that uptime, CFS vruntime, load 2464 * balancing, etc... continue to move forward, even 2465 * with a very low granularity. 2466 * 2467 * Return: Maximum deferment in nanoseconds. 2468 */ 2469u64 scheduler_tick_max_deferment(void) 2470{ 2471 struct rq *rq = this_rq(); 2472 unsigned long next, now = ACCESS_ONCE(jiffies); 2473 2474 next = rq->last_sched_tick + HZ; 2475 2476 if (time_before_eq(next, now)) 2477 return 0; 2478 2479 return jiffies_to_nsecs(next - now); 2480} 2481#endif 2482 2483notrace unsigned long get_parent_ip(unsigned long addr) 2484{ 2485 if (in_lock_functions(addr)) { 2486 addr = CALLER_ADDR2; 2487 if (in_lock_functions(addr)) 2488 addr = CALLER_ADDR3; 2489 } 2490 return addr; 2491} 2492 2493#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2494 defined(CONFIG_PREEMPT_TRACER)) 2495 2496void __kprobes preempt_count_add(int val) 2497{ 2498#ifdef CONFIG_DEBUG_PREEMPT 2499 /* 2500 * Underflow? 2501 */ 2502 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2503 return; 2504#endif 2505 __preempt_count_add(val); 2506#ifdef CONFIG_DEBUG_PREEMPT 2507 /* 2508 * Spinlock count overflowing soon? 2509 */ 2510 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2511 PREEMPT_MASK - 10); 2512#endif 2513 if (preempt_count() == val) 2514 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2515} 2516EXPORT_SYMBOL(preempt_count_add); 2517 2518void __kprobes preempt_count_sub(int val) 2519{ 2520#ifdef CONFIG_DEBUG_PREEMPT 2521 /* 2522 * Underflow? 2523 */ 2524 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2525 return; 2526 /* 2527 * Is the spinlock portion underflowing? 2528 */ 2529 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2530 !(preempt_count() & PREEMPT_MASK))) 2531 return; 2532#endif 2533 2534 if (preempt_count() == val) 2535 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2536 __preempt_count_sub(val); 2537} 2538EXPORT_SYMBOL(preempt_count_sub); 2539 2540#endif 2541 2542/* 2543 * Print scheduling while atomic bug: 2544 */ 2545static noinline void __schedule_bug(struct task_struct *prev) 2546{ 2547 if (oops_in_progress) 2548 return; 2549 2550 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2551 prev->comm, prev->pid, preempt_count()); 2552 2553 debug_show_held_locks(prev); 2554 print_modules(); 2555 if (irqs_disabled()) 2556 print_irqtrace_events(prev); 2557 dump_stack(); 2558 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2559} 2560 2561/* 2562 * Various schedule()-time debugging checks and statistics: 2563 */ 2564static inline void schedule_debug(struct task_struct *prev) 2565{ 2566 /* 2567 * Test if we are atomic. Since do_exit() needs to call into 2568 * schedule() atomically, we ignore that path. Otherwise whine 2569 * if we are scheduling when we should not. 2570 */ 2571 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2572 __schedule_bug(prev); 2573 rcu_sleep_check(); 2574 2575 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2576 2577 schedstat_inc(this_rq(), sched_count); 2578} 2579 2580static void put_prev_task(struct rq *rq, struct task_struct *prev) 2581{ 2582 if (prev->on_rq || rq->skip_clock_update < 0) 2583 update_rq_clock(rq); 2584 prev->sched_class->put_prev_task(rq, prev); 2585} 2586 2587/* 2588 * Pick up the highest-prio task: 2589 */ 2590static inline struct task_struct * 2591pick_next_task(struct rq *rq) 2592{ 2593 const struct sched_class *class; 2594 struct task_struct *p; 2595 2596 /* 2597 * Optimization: we know that if all tasks are in 2598 * the fair class we can call that function directly: 2599 */ 2600 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2601 p = fair_sched_class.pick_next_task(rq); 2602 if (likely(p)) 2603 return p; 2604 } 2605 2606 for_each_class(class) { 2607 p = class->pick_next_task(rq); 2608 if (p) 2609 return p; 2610 } 2611 2612 BUG(); /* the idle class will always have a runnable task */ 2613} 2614 2615/* 2616 * __schedule() is the main scheduler function. 2617 * 2618 * The main means of driving the scheduler and thus entering this function are: 2619 * 2620 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2621 * 2622 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2623 * paths. For example, see arch/x86/entry_64.S. 2624 * 2625 * To drive preemption between tasks, the scheduler sets the flag in timer 2626 * interrupt handler scheduler_tick(). 2627 * 2628 * 3. Wakeups don't really cause entry into schedule(). They add a 2629 * task to the run-queue and that's it. 2630 * 2631 * Now, if the new task added to the run-queue preempts the current 2632 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2633 * called on the nearest possible occasion: 2634 * 2635 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2636 * 2637 * - in syscall or exception context, at the next outmost 2638 * preempt_enable(). (this might be as soon as the wake_up()'s 2639 * spin_unlock()!) 2640 * 2641 * - in IRQ context, return from interrupt-handler to 2642 * preemptible context 2643 * 2644 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2645 * then at the next: 2646 * 2647 * - cond_resched() call 2648 * - explicit schedule() call 2649 * - return from syscall or exception to user-space 2650 * - return from interrupt-handler to user-space 2651 */ 2652static void __sched __schedule(void) 2653{ 2654 struct task_struct *prev, *next; 2655 unsigned long *switch_count; 2656 struct rq *rq; 2657 int cpu; 2658 2659need_resched: 2660 preempt_disable(); 2661 cpu = smp_processor_id(); 2662 rq = cpu_rq(cpu); 2663 rcu_note_context_switch(cpu); 2664 prev = rq->curr; 2665 2666 schedule_debug(prev); 2667 2668 if (sched_feat(HRTICK)) 2669 hrtick_clear(rq); 2670 2671 /* 2672 * Make sure that signal_pending_state()->signal_pending() below 2673 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2674 * done by the caller to avoid the race with signal_wake_up(). 2675 */ 2676 smp_mb__before_spinlock(); 2677 raw_spin_lock_irq(&rq->lock); 2678 2679 switch_count = &prev->nivcsw; 2680 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2681 if (unlikely(signal_pending_state(prev->state, prev))) { 2682 prev->state = TASK_RUNNING; 2683 } else { 2684 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2685 prev->on_rq = 0; 2686 2687 /* 2688 * If a worker went to sleep, notify and ask workqueue 2689 * whether it wants to wake up a task to maintain 2690 * concurrency. 2691 */ 2692 if (prev->flags & PF_WQ_WORKER) { 2693 struct task_struct *to_wakeup; 2694 2695 to_wakeup = wq_worker_sleeping(prev, cpu); 2696 if (to_wakeup) 2697 try_to_wake_up_local(to_wakeup); 2698 } 2699 } 2700 switch_count = &prev->nvcsw; 2701 } 2702 2703 pre_schedule(rq, prev); 2704 2705 if (unlikely(!rq->nr_running)) 2706 idle_balance(cpu, rq); 2707 2708 put_prev_task(rq, prev); 2709 next = pick_next_task(rq); 2710 clear_tsk_need_resched(prev); 2711 clear_preempt_need_resched(); 2712 rq->skip_clock_update = 0; 2713 2714 if (likely(prev != next)) { 2715 rq->nr_switches++; 2716 rq->curr = next; 2717 ++*switch_count; 2718 2719 context_switch(rq, prev, next); /* unlocks the rq */ 2720 /* 2721 * The context switch have flipped the stack from under us 2722 * and restored the local variables which were saved when 2723 * this task called schedule() in the past. prev == current 2724 * is still correct, but it can be moved to another cpu/rq. 2725 */ 2726 cpu = smp_processor_id(); 2727 rq = cpu_rq(cpu); 2728 } else 2729 raw_spin_unlock_irq(&rq->lock); 2730 2731 post_schedule(rq); 2732 2733 sched_preempt_enable_no_resched(); 2734 if (need_resched()) 2735 goto need_resched; 2736} 2737 2738static inline void sched_submit_work(struct task_struct *tsk) 2739{ 2740 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2741 return; 2742 /* 2743 * If we are going to sleep and we have plugged IO queued, 2744 * make sure to submit it to avoid deadlocks. 2745 */ 2746 if (blk_needs_flush_plug(tsk)) 2747 blk_schedule_flush_plug(tsk); 2748} 2749 2750asmlinkage void __sched schedule(void) 2751{ 2752 struct task_struct *tsk = current; 2753 2754 sched_submit_work(tsk); 2755 __schedule(); 2756} 2757EXPORT_SYMBOL(schedule); 2758 2759#ifdef CONFIG_CONTEXT_TRACKING 2760asmlinkage void __sched schedule_user(void) 2761{ 2762 /* 2763 * If we come here after a random call to set_need_resched(), 2764 * or we have been woken up remotely but the IPI has not yet arrived, 2765 * we haven't yet exited the RCU idle mode. Do it here manually until 2766 * we find a better solution. 2767 */ 2768 user_exit(); 2769 schedule(); 2770 user_enter(); 2771} 2772#endif 2773 2774/** 2775 * schedule_preempt_disabled - called with preemption disabled 2776 * 2777 * Returns with preemption disabled. Note: preempt_count must be 1 2778 */ 2779void __sched schedule_preempt_disabled(void) 2780{ 2781 sched_preempt_enable_no_resched(); 2782 schedule(); 2783 preempt_disable(); 2784} 2785 2786#ifdef CONFIG_PREEMPT 2787/* 2788 * this is the entry point to schedule() from in-kernel preemption 2789 * off of preempt_enable. Kernel preemptions off return from interrupt 2790 * occur there and call schedule directly. 2791 */ 2792asmlinkage void __sched notrace preempt_schedule(void) 2793{ 2794 /* 2795 * If there is a non-zero preempt_count or interrupts are disabled, 2796 * we do not want to preempt the current task. Just return.. 2797 */ 2798 if (likely(!preemptible())) 2799 return; 2800 2801 do { 2802 __preempt_count_add(PREEMPT_ACTIVE); 2803 __schedule(); 2804 __preempt_count_sub(PREEMPT_ACTIVE); 2805 2806 /* 2807 * Check again in case we missed a preemption opportunity 2808 * between schedule and now. 2809 */ 2810 barrier(); 2811 } while (need_resched()); 2812} 2813EXPORT_SYMBOL(preempt_schedule); 2814#endif /* CONFIG_PREEMPT */ 2815 2816/* 2817 * this is the entry point to schedule() from kernel preemption 2818 * off of irq context. 2819 * Note, that this is called and return with irqs disabled. This will 2820 * protect us against recursive calling from irq. 2821 */ 2822asmlinkage void __sched preempt_schedule_irq(void) 2823{ 2824 enum ctx_state prev_state; 2825 2826 /* Catch callers which need to be fixed */ 2827 BUG_ON(preempt_count() || !irqs_disabled()); 2828 2829 prev_state = exception_enter(); 2830 2831 do { 2832 __preempt_count_add(PREEMPT_ACTIVE); 2833 local_irq_enable(); 2834 __schedule(); 2835 local_irq_disable(); 2836 __preempt_count_sub(PREEMPT_ACTIVE); 2837 2838 /* 2839 * Check again in case we missed a preemption opportunity 2840 * between schedule and now. 2841 */ 2842 barrier(); 2843 } while (need_resched()); 2844 2845 exception_exit(prev_state); 2846} 2847 2848int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2849 void *key) 2850{ 2851 return try_to_wake_up(curr->private, mode, wake_flags); 2852} 2853EXPORT_SYMBOL(default_wake_function); 2854 2855static long __sched 2856sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2857{ 2858 unsigned long flags; 2859 wait_queue_t wait; 2860 2861 init_waitqueue_entry(&wait, current); 2862 2863 __set_current_state(state); 2864 2865 spin_lock_irqsave(&q->lock, flags); 2866 __add_wait_queue(q, &wait); 2867 spin_unlock(&q->lock); 2868 timeout = schedule_timeout(timeout); 2869 spin_lock_irq(&q->lock); 2870 __remove_wait_queue(q, &wait); 2871 spin_unlock_irqrestore(&q->lock, flags); 2872 2873 return timeout; 2874} 2875 2876void __sched interruptible_sleep_on(wait_queue_head_t *q) 2877{ 2878 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2879} 2880EXPORT_SYMBOL(interruptible_sleep_on); 2881 2882long __sched 2883interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2884{ 2885 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2886} 2887EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2888 2889void __sched sleep_on(wait_queue_head_t *q) 2890{ 2891 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2892} 2893EXPORT_SYMBOL(sleep_on); 2894 2895long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2896{ 2897 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2898} 2899EXPORT_SYMBOL(sleep_on_timeout); 2900 2901#ifdef CONFIG_RT_MUTEXES 2902 2903/* 2904 * rt_mutex_setprio - set the current priority of a task 2905 * @p: task 2906 * @prio: prio value (kernel-internal form) 2907 * 2908 * This function changes the 'effective' priority of a task. It does 2909 * not touch ->normal_prio like __setscheduler(). 2910 * 2911 * Used by the rt_mutex code to implement priority inheritance logic. 2912 */ 2913void rt_mutex_setprio(struct task_struct *p, int prio) 2914{ 2915 int oldprio, on_rq, running, enqueue_flag = 0; 2916 struct rq *rq; 2917 const struct sched_class *prev_class; 2918 2919 BUG_ON(prio > MAX_PRIO); 2920 2921 rq = __task_rq_lock(p); 2922 2923 /* 2924 * Idle task boosting is a nono in general. There is one 2925 * exception, when PREEMPT_RT and NOHZ is active: 2926 * 2927 * The idle task calls get_next_timer_interrupt() and holds 2928 * the timer wheel base->lock on the CPU and another CPU wants 2929 * to access the timer (probably to cancel it). We can safely 2930 * ignore the boosting request, as the idle CPU runs this code 2931 * with interrupts disabled and will complete the lock 2932 * protected section without being interrupted. So there is no 2933 * real need to boost. 2934 */ 2935 if (unlikely(p == rq->idle)) { 2936 WARN_ON(p != rq->curr); 2937 WARN_ON(p->pi_blocked_on); 2938 goto out_unlock; 2939 } 2940 2941 trace_sched_pi_setprio(p, prio); 2942 p->pi_top_task = rt_mutex_get_top_task(p); 2943 oldprio = p->prio; 2944 prev_class = p->sched_class; 2945 on_rq = p->on_rq; 2946 running = task_current(rq, p); 2947 if (on_rq) 2948 dequeue_task(rq, p, 0); 2949 if (running) 2950 p->sched_class->put_prev_task(rq, p); 2951 2952 /* 2953 * Boosting condition are: 2954 * 1. -rt task is running and holds mutex A 2955 * --> -dl task blocks on mutex A 2956 * 2957 * 2. -dl task is running and holds mutex A 2958 * --> -dl task blocks on mutex A and could preempt the 2959 * running task 2960 */ 2961 if (dl_prio(prio)) { 2962 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2963 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2964 p->dl.dl_boosted = 1; 2965 p->dl.dl_throttled = 0; 2966 enqueue_flag = ENQUEUE_REPLENISH; 2967 } else 2968 p->dl.dl_boosted = 0; 2969 p->sched_class = &dl_sched_class; 2970 } else if (rt_prio(prio)) { 2971 if (dl_prio(oldprio)) 2972 p->dl.dl_boosted = 0; 2973 if (oldprio < prio) 2974 enqueue_flag = ENQUEUE_HEAD; 2975 p->sched_class = &rt_sched_class; 2976 } else { 2977 if (dl_prio(oldprio)) 2978 p->dl.dl_boosted = 0; 2979 p->sched_class = &fair_sched_class; 2980 } 2981 2982 p->prio = prio; 2983 2984 if (running) 2985 p->sched_class->set_curr_task(rq); 2986 if (on_rq) 2987 enqueue_task(rq, p, enqueue_flag); 2988 2989 check_class_changed(rq, p, prev_class, oldprio); 2990out_unlock: 2991 __task_rq_unlock(rq); 2992} 2993#endif 2994 2995void set_user_nice(struct task_struct *p, long nice) 2996{ 2997 int old_prio, delta, on_rq; 2998 unsigned long flags; 2999 struct rq *rq; 3000 3001 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3002 return; 3003 /* 3004 * We have to be careful, if called from sys_setpriority(), 3005 * the task might be in the middle of scheduling on another CPU. 3006 */ 3007 rq = task_rq_lock(p, &flags); 3008 /* 3009 * The RT priorities are set via sched_setscheduler(), but we still 3010 * allow the 'normal' nice value to be set - but as expected 3011 * it wont have any effect on scheduling until the task is 3012 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3013 */ 3014 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3015 p->static_prio = NICE_TO_PRIO(nice); 3016 goto out_unlock; 3017 } 3018 on_rq = p->on_rq; 3019 if (on_rq) 3020 dequeue_task(rq, p, 0); 3021 3022 p->static_prio = NICE_TO_PRIO(nice); 3023 set_load_weight(p); 3024 old_prio = p->prio; 3025 p->prio = effective_prio(p); 3026 delta = p->prio - old_prio; 3027 3028 if (on_rq) { 3029 enqueue_task(rq, p, 0); 3030 /* 3031 * If the task increased its priority or is running and 3032 * lowered its priority, then reschedule its CPU: 3033 */ 3034 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3035 resched_task(rq->curr); 3036 } 3037out_unlock: 3038 task_rq_unlock(rq, p, &flags); 3039} 3040EXPORT_SYMBOL(set_user_nice); 3041 3042/* 3043 * can_nice - check if a task can reduce its nice value 3044 * @p: task 3045 * @nice: nice value 3046 */ 3047int can_nice(const struct task_struct *p, const int nice) 3048{ 3049 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3050 int nice_rlim = 20 - nice; 3051 3052 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3053 capable(CAP_SYS_NICE)); 3054} 3055 3056#ifdef __ARCH_WANT_SYS_NICE 3057 3058/* 3059 * sys_nice - change the priority of the current process. 3060 * @increment: priority increment 3061 * 3062 * sys_setpriority is a more generic, but much slower function that 3063 * does similar things. 3064 */ 3065SYSCALL_DEFINE1(nice, int, increment) 3066{ 3067 long nice, retval; 3068 3069 /* 3070 * Setpriority might change our priority at the same moment. 3071 * We don't have to worry. Conceptually one call occurs first 3072 * and we have a single winner. 3073 */ 3074 if (increment < -40) 3075 increment = -40; 3076 if (increment > 40) 3077 increment = 40; 3078 3079 nice = TASK_NICE(current) + increment; 3080 if (nice < -20) 3081 nice = -20; 3082 if (nice > 19) 3083 nice = 19; 3084 3085 if (increment < 0 && !can_nice(current, nice)) 3086 return -EPERM; 3087 3088 retval = security_task_setnice(current, nice); 3089 if (retval) 3090 return retval; 3091 3092 set_user_nice(current, nice); 3093 return 0; 3094} 3095 3096#endif 3097 3098/** 3099 * task_prio - return the priority value of a given task. 3100 * @p: the task in question. 3101 * 3102 * Return: The priority value as seen by users in /proc. 3103 * RT tasks are offset by -200. Normal tasks are centered 3104 * around 0, value goes from -16 to +15. 3105 */ 3106int task_prio(const struct task_struct *p) 3107{ 3108 return p->prio - MAX_RT_PRIO; 3109} 3110 3111/** 3112 * task_nice - return the nice value of a given task. 3113 * @p: the task in question. 3114 * 3115 * Return: The nice value [ -20 ... 0 ... 19 ]. 3116 */ 3117int task_nice(const struct task_struct *p) 3118{ 3119 return TASK_NICE(p); 3120} 3121EXPORT_SYMBOL(task_nice); 3122 3123/** 3124 * idle_cpu - is a given cpu idle currently? 3125 * @cpu: the processor in question. 3126 * 3127 * Return: 1 if the CPU is currently idle. 0 otherwise. 3128 */ 3129int idle_cpu(int cpu) 3130{ 3131 struct rq *rq = cpu_rq(cpu); 3132 3133 if (rq->curr != rq->idle) 3134 return 0; 3135 3136 if (rq->nr_running) 3137 return 0; 3138 3139#ifdef CONFIG_SMP 3140 if (!llist_empty(&rq->wake_list)) 3141 return 0; 3142#endif 3143 3144 return 1; 3145} 3146 3147/** 3148 * idle_task - return the idle task for a given cpu. 3149 * @cpu: the processor in question. 3150 * 3151 * Return: The idle task for the cpu @cpu. 3152 */ 3153struct task_struct *idle_task(int cpu) 3154{ 3155 return cpu_rq(cpu)->idle; 3156} 3157 3158/** 3159 * find_process_by_pid - find a process with a matching PID value. 3160 * @pid: the pid in question. 3161 * 3162 * The task of @pid, if found. %NULL otherwise. 3163 */ 3164static struct task_struct *find_process_by_pid(pid_t pid) 3165{ 3166 return pid ? find_task_by_vpid(pid) : current; 3167} 3168 3169/* 3170 * This function initializes the sched_dl_entity of a newly becoming 3171 * SCHED_DEADLINE task. 3172 * 3173 * Only the static values are considered here, the actual runtime and the 3174 * absolute deadline will be properly calculated when the task is enqueued 3175 * for the first time with its new policy. 3176 */ 3177static void 3178__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3179{ 3180 struct sched_dl_entity *dl_se = &p->dl; 3181 3182 init_dl_task_timer(dl_se); 3183 dl_se->dl_runtime = attr->sched_runtime; 3184 dl_se->dl_deadline = attr->sched_deadline; 3185 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3186 dl_se->flags = attr->sched_flags; 3187 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3188 dl_se->dl_throttled = 0; 3189 dl_se->dl_new = 1; 3190} 3191 3192/* Actually do priority change: must hold pi & rq lock. */ 3193static void __setscheduler(struct rq *rq, struct task_struct *p, 3194 const struct sched_attr *attr) 3195{ 3196 int policy = attr->sched_policy; 3197 3198 if (policy == -1) /* setparam */ 3199 policy = p->policy; 3200 3201 p->policy = policy; 3202 3203 if (dl_policy(policy)) 3204 __setparam_dl(p, attr); 3205 else if (fair_policy(policy)) 3206 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3207 3208 /* 3209 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3210 * !rt_policy. Always setting this ensures that things like 3211 * getparam()/getattr() don't report silly values for !rt tasks. 3212 */ 3213 p->rt_priority = attr->sched_priority; 3214 3215 p->normal_prio = normal_prio(p); 3216 p->prio = rt_mutex_getprio(p); 3217 3218 if (dl_prio(p->prio)) 3219 p->sched_class = &dl_sched_class; 3220 else if (rt_prio(p->prio)) 3221 p->sched_class = &rt_sched_class; 3222 else 3223 p->sched_class = &fair_sched_class; 3224 3225 set_load_weight(p); 3226} 3227 3228static void 3229__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3230{ 3231 struct sched_dl_entity *dl_se = &p->dl; 3232 3233 attr->sched_priority = p->rt_priority; 3234 attr->sched_runtime = dl_se->dl_runtime; 3235 attr->sched_deadline = dl_se->dl_deadline; 3236 attr->sched_period = dl_se->dl_period; 3237 attr->sched_flags = dl_se->flags; 3238} 3239 3240/* 3241 * This function validates the new parameters of a -deadline task. 3242 * We ask for the deadline not being zero, and greater or equal 3243 * than the runtime, as well as the period of being zero or 3244 * greater than deadline. Furthermore, we have to be sure that 3245 * user parameters are above the internal resolution (1us); we 3246 * check sched_runtime only since it is always the smaller one. 3247 */ 3248static bool 3249__checkparam_dl(const struct sched_attr *attr) 3250{ 3251 return attr && attr->sched_deadline != 0 && 3252 (attr->sched_period == 0 || 3253 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3254 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3255 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3256} 3257 3258/* 3259 * check the target process has a UID that matches the current process's 3260 */ 3261static bool check_same_owner(struct task_struct *p) 3262{ 3263 const struct cred *cred = current_cred(), *pcred; 3264 bool match; 3265 3266 rcu_read_lock(); 3267 pcred = __task_cred(p); 3268 match = (uid_eq(cred->euid, pcred->euid) || 3269 uid_eq(cred->euid, pcred->uid)); 3270 rcu_read_unlock(); 3271 return match; 3272} 3273 3274static int __sched_setscheduler(struct task_struct *p, 3275 const struct sched_attr *attr, 3276 bool user) 3277{ 3278 int retval, oldprio, oldpolicy = -1, on_rq, running; 3279 int policy = attr->sched_policy; 3280 unsigned long flags; 3281 const struct sched_class *prev_class; 3282 struct rq *rq; 3283 int reset_on_fork; 3284 3285 /* may grab non-irq protected spin_locks */ 3286 BUG_ON(in_interrupt()); 3287recheck: 3288 /* double check policy once rq lock held */ 3289 if (policy < 0) { 3290 reset_on_fork = p->sched_reset_on_fork; 3291 policy = oldpolicy = p->policy; 3292 } else { 3293 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3294 3295 if (policy != SCHED_DEADLINE && 3296 policy != SCHED_FIFO && policy != SCHED_RR && 3297 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3298 policy != SCHED_IDLE) 3299 return -EINVAL; 3300 } 3301 3302 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3303 return -EINVAL; 3304 3305 /* 3306 * Valid priorities for SCHED_FIFO and SCHED_RR are 3307 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3308 * SCHED_BATCH and SCHED_IDLE is 0. 3309 */ 3310 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3311 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3312 return -EINVAL; 3313 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3314 (rt_policy(policy) != (attr->sched_priority != 0))) 3315 return -EINVAL; 3316 3317 /* 3318 * Allow unprivileged RT tasks to decrease priority: 3319 */ 3320 if (user && !capable(CAP_SYS_NICE)) { 3321 if (fair_policy(policy)) { 3322 if (attr->sched_nice < TASK_NICE(p) && 3323 !can_nice(p, attr->sched_nice)) 3324 return -EPERM; 3325 } 3326 3327 if (rt_policy(policy)) { 3328 unsigned long rlim_rtprio = 3329 task_rlimit(p, RLIMIT_RTPRIO); 3330 3331 /* can't set/change the rt policy */ 3332 if (policy != p->policy && !rlim_rtprio) 3333 return -EPERM; 3334 3335 /* can't increase priority */ 3336 if (attr->sched_priority > p->rt_priority && 3337 attr->sched_priority > rlim_rtprio) 3338 return -EPERM; 3339 } 3340 3341 /* 3342 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3343 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3344 */ 3345 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3346 if (!can_nice(p, TASK_NICE(p))) 3347 return -EPERM; 3348 } 3349 3350 /* can't change other user's priorities */ 3351 if (!check_same_owner(p)) 3352 return -EPERM; 3353 3354 /* Normal users shall not reset the sched_reset_on_fork flag */ 3355 if (p->sched_reset_on_fork && !reset_on_fork) 3356 return -EPERM; 3357 } 3358 3359 if (user) { 3360 retval = security_task_setscheduler(p); 3361 if (retval) 3362 return retval; 3363 } 3364 3365 /* 3366 * make sure no PI-waiters arrive (or leave) while we are 3367 * changing the priority of the task: 3368 * 3369 * To be able to change p->policy safely, the appropriate 3370 * runqueue lock must be held. 3371 */ 3372 rq = task_rq_lock(p, &flags); 3373 3374 /* 3375 * Changing the policy of the stop threads its a very bad idea 3376 */ 3377 if (p == rq->stop) { 3378 task_rq_unlock(rq, p, &flags); 3379 return -EINVAL; 3380 } 3381 3382 /* 3383 * If not changing anything there's no need to proceed further: 3384 */ 3385 if (unlikely(policy == p->policy)) { 3386 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p)) 3387 goto change; 3388 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3389 goto change; 3390 if (dl_policy(policy)) 3391 goto change; 3392 3393 task_rq_unlock(rq, p, &flags); 3394 return 0; 3395 } 3396change: 3397 3398 if (user) { 3399#ifdef CONFIG_RT_GROUP_SCHED 3400 /* 3401 * Do not allow realtime tasks into groups that have no runtime 3402 * assigned. 3403 */ 3404 if (rt_bandwidth_enabled() && rt_policy(policy) && 3405 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3406 !task_group_is_autogroup(task_group(p))) { 3407 task_rq_unlock(rq, p, &flags); 3408 return -EPERM; 3409 } 3410#endif 3411#ifdef CONFIG_SMP 3412 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3413 cpumask_t *span = rq->rd->span; 3414 3415 /* 3416 * Don't allow tasks with an affinity mask smaller than 3417 * the entire root_domain to become SCHED_DEADLINE. We 3418 * will also fail if there's no bandwidth available. 3419 */ 3420 if (!cpumask_subset(span, &p->cpus_allowed) || 3421 rq->rd->dl_bw.bw == 0) { 3422 task_rq_unlock(rq, p, &flags); 3423 return -EPERM; 3424 } 3425 } 3426#endif 3427 } 3428 3429 /* recheck policy now with rq lock held */ 3430 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3431 policy = oldpolicy = -1; 3432 task_rq_unlock(rq, p, &flags); 3433 goto recheck; 3434 } 3435 3436 /* 3437 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3438 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3439 * is available. 3440 */ 3441 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3442 task_rq_unlock(rq, p, &flags); 3443 return -EBUSY; 3444 } 3445 3446 on_rq = p->on_rq; 3447 running = task_current(rq, p); 3448 if (on_rq) 3449 dequeue_task(rq, p, 0); 3450 if (running) 3451 p->sched_class->put_prev_task(rq, p); 3452 3453 p->sched_reset_on_fork = reset_on_fork; 3454 3455 oldprio = p->prio; 3456 prev_class = p->sched_class; 3457 __setscheduler(rq, p, attr); 3458 3459 if (running) 3460 p->sched_class->set_curr_task(rq); 3461 if (on_rq) 3462 enqueue_task(rq, p, 0); 3463 3464 check_class_changed(rq, p, prev_class, oldprio); 3465 task_rq_unlock(rq, p, &flags); 3466 3467 rt_mutex_adjust_pi(p); 3468 3469 return 0; 3470} 3471 3472static int _sched_setscheduler(struct task_struct *p, int policy, 3473 const struct sched_param *param, bool check) 3474{ 3475 struct sched_attr attr = { 3476 .sched_policy = policy, 3477 .sched_priority = param->sched_priority, 3478 .sched_nice = PRIO_TO_NICE(p->static_prio), 3479 }; 3480 3481 /* 3482 * Fixup the legacy SCHED_RESET_ON_FORK hack 3483 */ 3484 if (policy & SCHED_RESET_ON_FORK) { 3485 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3486 policy &= ~SCHED_RESET_ON_FORK; 3487 attr.sched_policy = policy; 3488 } 3489 3490 return __sched_setscheduler(p, &attr, check); 3491} 3492/** 3493 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3494 * @p: the task in question. 3495 * @policy: new policy. 3496 * @param: structure containing the new RT priority. 3497 * 3498 * Return: 0 on success. An error code otherwise. 3499 * 3500 * NOTE that the task may be already dead. 3501 */ 3502int sched_setscheduler(struct task_struct *p, int policy, 3503 const struct sched_param *param) 3504{ 3505 return _sched_setscheduler(p, policy, param, true); 3506} 3507EXPORT_SYMBOL_GPL(sched_setscheduler); 3508 3509int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3510{ 3511 return __sched_setscheduler(p, attr, true); 3512} 3513EXPORT_SYMBOL_GPL(sched_setattr); 3514 3515/** 3516 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3517 * @p: the task in question. 3518 * @policy: new policy. 3519 * @param: structure containing the new RT priority. 3520 * 3521 * Just like sched_setscheduler, only don't bother checking if the 3522 * current context has permission. For example, this is needed in 3523 * stop_machine(): we create temporary high priority worker threads, 3524 * but our caller might not have that capability. 3525 * 3526 * Return: 0 on success. An error code otherwise. 3527 */ 3528int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3529 const struct sched_param *param) 3530{ 3531 return _sched_setscheduler(p, policy, param, false); 3532} 3533 3534static int 3535do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3536{ 3537 struct sched_param lparam; 3538 struct task_struct *p; 3539 int retval; 3540 3541 if (!param || pid < 0) 3542 return -EINVAL; 3543 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3544 return -EFAULT; 3545 3546 rcu_read_lock(); 3547 retval = -ESRCH; 3548 p = find_process_by_pid(pid); 3549 if (p != NULL) 3550 retval = sched_setscheduler(p, policy, &lparam); 3551 rcu_read_unlock(); 3552 3553 return retval; 3554} 3555 3556/* 3557 * Mimics kernel/events/core.c perf_copy_attr(). 3558 */ 3559static int sched_copy_attr(struct sched_attr __user *uattr, 3560 struct sched_attr *attr) 3561{ 3562 u32 size; 3563 int ret; 3564 3565 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3566 return -EFAULT; 3567 3568 /* 3569 * zero the full structure, so that a short copy will be nice. 3570 */ 3571 memset(attr, 0, sizeof(*attr)); 3572 3573 ret = get_user(size, &uattr->size); 3574 if (ret) 3575 return ret; 3576 3577 if (size > PAGE_SIZE) /* silly large */ 3578 goto err_size; 3579 3580 if (!size) /* abi compat */ 3581 size = SCHED_ATTR_SIZE_VER0; 3582 3583 if (size < SCHED_ATTR_SIZE_VER0) 3584 goto err_size; 3585 3586 /* 3587 * If we're handed a bigger struct than we know of, 3588 * ensure all the unknown bits are 0 - i.e. new 3589 * user-space does not rely on any kernel feature 3590 * extensions we dont know about yet. 3591 */ 3592 if (size > sizeof(*attr)) { 3593 unsigned char __user *addr; 3594 unsigned char __user *end; 3595 unsigned char val; 3596 3597 addr = (void __user *)uattr + sizeof(*attr); 3598 end = (void __user *)uattr + size; 3599 3600 for (; addr < end; addr++) { 3601 ret = get_user(val, addr); 3602 if (ret) 3603 return ret; 3604 if (val) 3605 goto err_size; 3606 } 3607 size = sizeof(*attr); 3608 } 3609 3610 ret = copy_from_user(attr, uattr, size); 3611 if (ret) 3612 return -EFAULT; 3613 3614 /* 3615 * XXX: do we want to be lenient like existing syscalls; or do we want 3616 * to be strict and return an error on out-of-bounds values? 3617 */ 3618 attr->sched_nice = clamp(attr->sched_nice, -20, 19); 3619 3620out: 3621 return ret; 3622 3623err_size: 3624 put_user(sizeof(*attr), &uattr->size); 3625 ret = -E2BIG; 3626 goto out; 3627} 3628 3629/** 3630 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3631 * @pid: the pid in question. 3632 * @policy: new policy. 3633 * @param: structure containing the new RT priority. 3634 * 3635 * Return: 0 on success. An error code otherwise. 3636 */ 3637SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3638 struct sched_param __user *, param) 3639{ 3640 /* negative values for policy are not valid */ 3641 if (policy < 0) 3642 return -EINVAL; 3643 3644 return do_sched_setscheduler(pid, policy, param); 3645} 3646 3647/** 3648 * sys_sched_setparam - set/change the RT priority of a thread 3649 * @pid: the pid in question. 3650 * @param: structure containing the new RT priority. 3651 * 3652 * Return: 0 on success. An error code otherwise. 3653 */ 3654SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3655{ 3656 return do_sched_setscheduler(pid, -1, param); 3657} 3658 3659/** 3660 * sys_sched_setattr - same as above, but with extended sched_attr 3661 * @pid: the pid in question. 3662 * @uattr: structure containing the extended parameters. 3663 */ 3664SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3665 unsigned int, flags) 3666{ 3667 struct sched_attr attr; 3668 struct task_struct *p; 3669 int retval; 3670 3671 if (!uattr || pid < 0 || flags) 3672 return -EINVAL; 3673 3674 if (sched_copy_attr(uattr, &attr)) 3675 return -EFAULT; 3676 3677 rcu_read_lock(); 3678 retval = -ESRCH; 3679 p = find_process_by_pid(pid); 3680 if (p != NULL) 3681 retval = sched_setattr(p, &attr); 3682 rcu_read_unlock(); 3683 3684 return retval; 3685} 3686 3687/** 3688 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3689 * @pid: the pid in question. 3690 * 3691 * Return: On success, the policy of the thread. Otherwise, a negative error 3692 * code. 3693 */ 3694SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3695{ 3696 struct task_struct *p; 3697 int retval; 3698 3699 if (pid < 0) 3700 return -EINVAL; 3701 3702 retval = -ESRCH; 3703 rcu_read_lock(); 3704 p = find_process_by_pid(pid); 3705 if (p) { 3706 retval = security_task_getscheduler(p); 3707 if (!retval) 3708 retval = p->policy 3709 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3710 } 3711 rcu_read_unlock(); 3712 return retval; 3713} 3714 3715/** 3716 * sys_sched_getparam - get the RT priority of a thread 3717 * @pid: the pid in question. 3718 * @param: structure containing the RT priority. 3719 * 3720 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3721 * code. 3722 */ 3723SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3724{ 3725 struct sched_param lp; 3726 struct task_struct *p; 3727 int retval; 3728 3729 if (!param || pid < 0) 3730 return -EINVAL; 3731 3732 rcu_read_lock(); 3733 p = find_process_by_pid(pid); 3734 retval = -ESRCH; 3735 if (!p) 3736 goto out_unlock; 3737 3738 retval = security_task_getscheduler(p); 3739 if (retval) 3740 goto out_unlock; 3741 3742 if (task_has_dl_policy(p)) { 3743 retval = -EINVAL; 3744 goto out_unlock; 3745 } 3746 lp.sched_priority = p->rt_priority; 3747 rcu_read_unlock(); 3748 3749 /* 3750 * This one might sleep, we cannot do it with a spinlock held ... 3751 */ 3752 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3753 3754 return retval; 3755 3756out_unlock: 3757 rcu_read_unlock(); 3758 return retval; 3759} 3760 3761static int sched_read_attr(struct sched_attr __user *uattr, 3762 struct sched_attr *attr, 3763 unsigned int usize) 3764{ 3765 int ret; 3766 3767 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3768 return -EFAULT; 3769 3770 /* 3771 * If we're handed a smaller struct than we know of, 3772 * ensure all the unknown bits are 0 - i.e. old 3773 * user-space does not get uncomplete information. 3774 */ 3775 if (usize < sizeof(*attr)) { 3776 unsigned char *addr; 3777 unsigned char *end; 3778 3779 addr = (void *)attr + usize; 3780 end = (void *)attr + sizeof(*attr); 3781 3782 for (; addr < end; addr++) { 3783 if (*addr) 3784 goto err_size; 3785 } 3786 3787 attr->size = usize; 3788 } 3789 3790 ret = copy_to_user(uattr, attr, attr->size); 3791 if (ret) 3792 return -EFAULT; 3793 3794out: 3795 return ret; 3796 3797err_size: 3798 ret = -E2BIG; 3799 goto out; 3800} 3801 3802/** 3803 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3804 * @pid: the pid in question. 3805 * @uattr: structure containing the extended parameters. 3806 * @size: sizeof(attr) for fwd/bwd comp. 3807 */ 3808SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3809 unsigned int, size, unsigned int, flags) 3810{ 3811 struct sched_attr attr = { 3812 .size = sizeof(struct sched_attr), 3813 }; 3814 struct task_struct *p; 3815 int retval; 3816 3817 if (!uattr || pid < 0 || size > PAGE_SIZE || 3818 size < SCHED_ATTR_SIZE_VER0 || flags) 3819 return -EINVAL; 3820 3821 rcu_read_lock(); 3822 p = find_process_by_pid(pid); 3823 retval = -ESRCH; 3824 if (!p) 3825 goto out_unlock; 3826 3827 retval = security_task_getscheduler(p); 3828 if (retval) 3829 goto out_unlock; 3830 3831 attr.sched_policy = p->policy; 3832 if (p->sched_reset_on_fork) 3833 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3834 if (task_has_dl_policy(p)) 3835 __getparam_dl(p, &attr); 3836 else if (task_has_rt_policy(p)) 3837 attr.sched_priority = p->rt_priority; 3838 else 3839 attr.sched_nice = TASK_NICE(p); 3840 3841 rcu_read_unlock(); 3842 3843 retval = sched_read_attr(uattr, &attr, size); 3844 return retval; 3845 3846out_unlock: 3847 rcu_read_unlock(); 3848 return retval; 3849} 3850 3851long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3852{ 3853 cpumask_var_t cpus_allowed, new_mask; 3854 struct task_struct *p; 3855 int retval; 3856 3857 rcu_read_lock(); 3858 3859 p = find_process_by_pid(pid); 3860 if (!p) { 3861 rcu_read_unlock(); 3862 return -ESRCH; 3863 } 3864 3865 /* Prevent p going away */ 3866 get_task_struct(p); 3867 rcu_read_unlock(); 3868 3869 if (p->flags & PF_NO_SETAFFINITY) { 3870 retval = -EINVAL; 3871 goto out_put_task; 3872 } 3873 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3874 retval = -ENOMEM; 3875 goto out_put_task; 3876 } 3877 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3878 retval = -ENOMEM; 3879 goto out_free_cpus_allowed; 3880 } 3881 retval = -EPERM; 3882 if (!check_same_owner(p)) { 3883 rcu_read_lock(); 3884 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3885 rcu_read_unlock(); 3886 goto out_unlock; 3887 } 3888 rcu_read_unlock(); 3889 } 3890 3891 retval = security_task_setscheduler(p); 3892 if (retval) 3893 goto out_unlock; 3894 3895 3896 cpuset_cpus_allowed(p, cpus_allowed); 3897 cpumask_and(new_mask, in_mask, cpus_allowed); 3898 3899 /* 3900 * Since bandwidth control happens on root_domain basis, 3901 * if admission test is enabled, we only admit -deadline 3902 * tasks allowed to run on all the CPUs in the task's 3903 * root_domain. 3904 */ 3905#ifdef CONFIG_SMP 3906 if (task_has_dl_policy(p)) { 3907 const struct cpumask *span = task_rq(p)->rd->span; 3908 3909 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3910 retval = -EBUSY; 3911 goto out_unlock; 3912 } 3913 } 3914#endif 3915again: 3916 retval = set_cpus_allowed_ptr(p, new_mask); 3917 3918 if (!retval) { 3919 cpuset_cpus_allowed(p, cpus_allowed); 3920 if (!cpumask_subset(new_mask, cpus_allowed)) { 3921 /* 3922 * We must have raced with a concurrent cpuset 3923 * update. Just reset the cpus_allowed to the 3924 * cpuset's cpus_allowed 3925 */ 3926 cpumask_copy(new_mask, cpus_allowed); 3927 goto again; 3928 } 3929 } 3930out_unlock: 3931 free_cpumask_var(new_mask); 3932out_free_cpus_allowed: 3933 free_cpumask_var(cpus_allowed); 3934out_put_task: 3935 put_task_struct(p); 3936 return retval; 3937} 3938 3939static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3940 struct cpumask *new_mask) 3941{ 3942 if (len < cpumask_size()) 3943 cpumask_clear(new_mask); 3944 else if (len > cpumask_size()) 3945 len = cpumask_size(); 3946 3947 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3948} 3949 3950/** 3951 * sys_sched_setaffinity - set the cpu affinity of a process 3952 * @pid: pid of the process 3953 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3954 * @user_mask_ptr: user-space pointer to the new cpu mask 3955 * 3956 * Return: 0 on success. An error code otherwise. 3957 */ 3958SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3959 unsigned long __user *, user_mask_ptr) 3960{ 3961 cpumask_var_t new_mask; 3962 int retval; 3963 3964 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3965 return -ENOMEM; 3966 3967 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3968 if (retval == 0) 3969 retval = sched_setaffinity(pid, new_mask); 3970 free_cpumask_var(new_mask); 3971 return retval; 3972} 3973 3974long sched_getaffinity(pid_t pid, struct cpumask *mask) 3975{ 3976 struct task_struct *p; 3977 unsigned long flags; 3978 int retval; 3979 3980 rcu_read_lock(); 3981 3982 retval = -ESRCH; 3983 p = find_process_by_pid(pid); 3984 if (!p) 3985 goto out_unlock; 3986 3987 retval = security_task_getscheduler(p); 3988 if (retval) 3989 goto out_unlock; 3990 3991 raw_spin_lock_irqsave(&p->pi_lock, flags); 3992 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3993 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3994 3995out_unlock: 3996 rcu_read_unlock(); 3997 3998 return retval; 3999} 4000 4001/** 4002 * sys_sched_getaffinity - get the cpu affinity of a process 4003 * @pid: pid of the process 4004 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4005 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4006 * 4007 * Return: 0 on success. An error code otherwise. 4008 */ 4009SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4010 unsigned long __user *, user_mask_ptr) 4011{ 4012 int ret; 4013 cpumask_var_t mask; 4014 4015 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4016 return -EINVAL; 4017 if (len & (sizeof(unsigned long)-1)) 4018 return -EINVAL; 4019 4020 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4021 return -ENOMEM; 4022 4023 ret = sched_getaffinity(pid, mask); 4024 if (ret == 0) { 4025 size_t retlen = min_t(size_t, len, cpumask_size()); 4026 4027 if (copy_to_user(user_mask_ptr, mask, retlen)) 4028 ret = -EFAULT; 4029 else 4030 ret = retlen; 4031 } 4032 free_cpumask_var(mask); 4033 4034 return ret; 4035} 4036 4037/** 4038 * sys_sched_yield - yield the current processor to other threads. 4039 * 4040 * This function yields the current CPU to other tasks. If there are no 4041 * other threads running on this CPU then this function will return. 4042 * 4043 * Return: 0. 4044 */ 4045SYSCALL_DEFINE0(sched_yield) 4046{ 4047 struct rq *rq = this_rq_lock(); 4048 4049 schedstat_inc(rq, yld_count); 4050 current->sched_class->yield_task(rq); 4051 4052 /* 4053 * Since we are going to call schedule() anyway, there's 4054 * no need to preempt or enable interrupts: 4055 */ 4056 __release(rq->lock); 4057 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4058 do_raw_spin_unlock(&rq->lock); 4059 sched_preempt_enable_no_resched(); 4060 4061 schedule(); 4062 4063 return 0; 4064} 4065 4066static void __cond_resched(void) 4067{ 4068 __preempt_count_add(PREEMPT_ACTIVE); 4069 __schedule(); 4070 __preempt_count_sub(PREEMPT_ACTIVE); 4071} 4072 4073int __sched _cond_resched(void) 4074{ 4075 if (should_resched()) { 4076 __cond_resched(); 4077 return 1; 4078 } 4079 return 0; 4080} 4081EXPORT_SYMBOL(_cond_resched); 4082 4083/* 4084 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4085 * call schedule, and on return reacquire the lock. 4086 * 4087 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4088 * operations here to prevent schedule() from being called twice (once via 4089 * spin_unlock(), once by hand). 4090 */ 4091int __cond_resched_lock(spinlock_t *lock) 4092{ 4093 int resched = should_resched(); 4094 int ret = 0; 4095 4096 lockdep_assert_held(lock); 4097 4098 if (spin_needbreak(lock) || resched) { 4099 spin_unlock(lock); 4100 if (resched) 4101 __cond_resched(); 4102 else 4103 cpu_relax(); 4104 ret = 1; 4105 spin_lock(lock); 4106 } 4107 return ret; 4108} 4109EXPORT_SYMBOL(__cond_resched_lock); 4110 4111int __sched __cond_resched_softirq(void) 4112{ 4113 BUG_ON(!in_softirq()); 4114 4115 if (should_resched()) { 4116 local_bh_enable(); 4117 __cond_resched(); 4118 local_bh_disable(); 4119 return 1; 4120 } 4121 return 0; 4122} 4123EXPORT_SYMBOL(__cond_resched_softirq); 4124 4125/** 4126 * yield - yield the current processor to other threads. 4127 * 4128 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4129 * 4130 * The scheduler is at all times free to pick the calling task as the most 4131 * eligible task to run, if removing the yield() call from your code breaks 4132 * it, its already broken. 4133 * 4134 * Typical broken usage is: 4135 * 4136 * while (!event) 4137 * yield(); 4138 * 4139 * where one assumes that yield() will let 'the other' process run that will 4140 * make event true. If the current task is a SCHED_FIFO task that will never 4141 * happen. Never use yield() as a progress guarantee!! 4142 * 4143 * If you want to use yield() to wait for something, use wait_event(). 4144 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4145 * If you still want to use yield(), do not! 4146 */ 4147void __sched yield(void) 4148{ 4149 set_current_state(TASK_RUNNING); 4150 sys_sched_yield(); 4151} 4152EXPORT_SYMBOL(yield); 4153 4154/** 4155 * yield_to - yield the current processor to another thread in 4156 * your thread group, or accelerate that thread toward the 4157 * processor it's on. 4158 * @p: target task 4159 * @preempt: whether task preemption is allowed or not 4160 * 4161 * It's the caller's job to ensure that the target task struct 4162 * can't go away on us before we can do any checks. 4163 * 4164 * Return: 4165 * true (>0) if we indeed boosted the target task. 4166 * false (0) if we failed to boost the target. 4167 * -ESRCH if there's no task to yield to. 4168 */ 4169bool __sched yield_to(struct task_struct *p, bool preempt) 4170{ 4171 struct task_struct *curr = current; 4172 struct rq *rq, *p_rq; 4173 unsigned long flags; 4174 int yielded = 0; 4175 4176 local_irq_save(flags); 4177 rq = this_rq(); 4178 4179again: 4180 p_rq = task_rq(p); 4181 /* 4182 * If we're the only runnable task on the rq and target rq also 4183 * has only one task, there's absolutely no point in yielding. 4184 */ 4185 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4186 yielded = -ESRCH; 4187 goto out_irq; 4188 } 4189 4190 double_rq_lock(rq, p_rq); 4191 if (task_rq(p) != p_rq) { 4192 double_rq_unlock(rq, p_rq); 4193 goto again; 4194 } 4195 4196 if (!curr->sched_class->yield_to_task) 4197 goto out_unlock; 4198 4199 if (curr->sched_class != p->sched_class) 4200 goto out_unlock; 4201 4202 if (task_running(p_rq, p) || p->state) 4203 goto out_unlock; 4204 4205 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4206 if (yielded) { 4207 schedstat_inc(rq, yld_count); 4208 /* 4209 * Make p's CPU reschedule; pick_next_entity takes care of 4210 * fairness. 4211 */ 4212 if (preempt && rq != p_rq) 4213 resched_task(p_rq->curr); 4214 } 4215 4216out_unlock: 4217 double_rq_unlock(rq, p_rq); 4218out_irq: 4219 local_irq_restore(flags); 4220 4221 if (yielded > 0) 4222 schedule(); 4223 4224 return yielded; 4225} 4226EXPORT_SYMBOL_GPL(yield_to); 4227 4228/* 4229 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4230 * that process accounting knows that this is a task in IO wait state. 4231 */ 4232void __sched io_schedule(void) 4233{ 4234 struct rq *rq = raw_rq(); 4235 4236 delayacct_blkio_start(); 4237 atomic_inc(&rq->nr_iowait); 4238 blk_flush_plug(current); 4239 current->in_iowait = 1; 4240 schedule(); 4241 current->in_iowait = 0; 4242 atomic_dec(&rq->nr_iowait); 4243 delayacct_blkio_end(); 4244} 4245EXPORT_SYMBOL(io_schedule); 4246 4247long __sched io_schedule_timeout(long timeout) 4248{ 4249 struct rq *rq = raw_rq(); 4250 long ret; 4251 4252 delayacct_blkio_start(); 4253 atomic_inc(&rq->nr_iowait); 4254 blk_flush_plug(current); 4255 current->in_iowait = 1; 4256 ret = schedule_timeout(timeout); 4257 current->in_iowait = 0; 4258 atomic_dec(&rq->nr_iowait); 4259 delayacct_blkio_end(); 4260 return ret; 4261} 4262 4263/** 4264 * sys_sched_get_priority_max - return maximum RT priority. 4265 * @policy: scheduling class. 4266 * 4267 * Return: On success, this syscall returns the maximum 4268 * rt_priority that can be used by a given scheduling class. 4269 * On failure, a negative error code is returned. 4270 */ 4271SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4272{ 4273 int ret = -EINVAL; 4274 4275 switch (policy) { 4276 case SCHED_FIFO: 4277 case SCHED_RR: 4278 ret = MAX_USER_RT_PRIO-1; 4279 break; 4280 case SCHED_DEADLINE: 4281 case SCHED_NORMAL: 4282 case SCHED_BATCH: 4283 case SCHED_IDLE: 4284 ret = 0; 4285 break; 4286 } 4287 return ret; 4288} 4289 4290/** 4291 * sys_sched_get_priority_min - return minimum RT priority. 4292 * @policy: scheduling class. 4293 * 4294 * Return: On success, this syscall returns the minimum 4295 * rt_priority that can be used by a given scheduling class. 4296 * On failure, a negative error code is returned. 4297 */ 4298SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4299{ 4300 int ret = -EINVAL; 4301 4302 switch (policy) { 4303 case SCHED_FIFO: 4304 case SCHED_RR: 4305 ret = 1; 4306 break; 4307 case SCHED_DEADLINE: 4308 case SCHED_NORMAL: 4309 case SCHED_BATCH: 4310 case SCHED_IDLE: 4311 ret = 0; 4312 } 4313 return ret; 4314} 4315 4316/** 4317 * sys_sched_rr_get_interval - return the default timeslice of a process. 4318 * @pid: pid of the process. 4319 * @interval: userspace pointer to the timeslice value. 4320 * 4321 * this syscall writes the default timeslice value of a given process 4322 * into the user-space timespec buffer. A value of '0' means infinity. 4323 * 4324 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4325 * an error code. 4326 */ 4327SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4328 struct timespec __user *, interval) 4329{ 4330 struct task_struct *p; 4331 unsigned int time_slice; 4332 unsigned long flags; 4333 struct rq *rq; 4334 int retval; 4335 struct timespec t; 4336 4337 if (pid < 0) 4338 return -EINVAL; 4339 4340 retval = -ESRCH; 4341 rcu_read_lock(); 4342 p = find_process_by_pid(pid); 4343 if (!p) 4344 goto out_unlock; 4345 4346 retval = security_task_getscheduler(p); 4347 if (retval) 4348 goto out_unlock; 4349 4350 rq = task_rq_lock(p, &flags); 4351 time_slice = 0; 4352 if (p->sched_class->get_rr_interval) 4353 time_slice = p->sched_class->get_rr_interval(rq, p); 4354 task_rq_unlock(rq, p, &flags); 4355 4356 rcu_read_unlock(); 4357 jiffies_to_timespec(time_slice, &t); 4358 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4359 return retval; 4360 4361out_unlock: 4362 rcu_read_unlock(); 4363 return retval; 4364} 4365 4366static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4367 4368void sched_show_task(struct task_struct *p) 4369{ 4370 unsigned long free = 0; 4371 int ppid; 4372 unsigned state; 4373 4374 state = p->state ? __ffs(p->state) + 1 : 0; 4375 printk(KERN_INFO "%-15.15s %c", p->comm, 4376 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4377#if BITS_PER_LONG == 32 4378 if (state == TASK_RUNNING) 4379 printk(KERN_CONT " running "); 4380 else 4381 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4382#else 4383 if (state == TASK_RUNNING) 4384 printk(KERN_CONT " running task "); 4385 else 4386 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4387#endif 4388#ifdef CONFIG_DEBUG_STACK_USAGE 4389 free = stack_not_used(p); 4390#endif 4391 rcu_read_lock(); 4392 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4393 rcu_read_unlock(); 4394 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4395 task_pid_nr(p), ppid, 4396 (unsigned long)task_thread_info(p)->flags); 4397 4398 print_worker_info(KERN_INFO, p); 4399 show_stack(p, NULL); 4400} 4401 4402void show_state_filter(unsigned long state_filter) 4403{ 4404 struct task_struct *g, *p; 4405 4406#if BITS_PER_LONG == 32 4407 printk(KERN_INFO 4408 " task PC stack pid father\n"); 4409#else 4410 printk(KERN_INFO 4411 " task PC stack pid father\n"); 4412#endif 4413 rcu_read_lock(); 4414 do_each_thread(g, p) { 4415 /* 4416 * reset the NMI-timeout, listing all files on a slow 4417 * console might take a lot of time: 4418 */ 4419 touch_nmi_watchdog(); 4420 if (!state_filter || (p->state & state_filter)) 4421 sched_show_task(p); 4422 } while_each_thread(g, p); 4423 4424 touch_all_softlockup_watchdogs(); 4425 4426#ifdef CONFIG_SCHED_DEBUG 4427 sysrq_sched_debug_show(); 4428#endif 4429 rcu_read_unlock(); 4430 /* 4431 * Only show locks if all tasks are dumped: 4432 */ 4433 if (!state_filter) 4434 debug_show_all_locks(); 4435} 4436 4437void init_idle_bootup_task(struct task_struct *idle) 4438{ 4439 idle->sched_class = &idle_sched_class; 4440} 4441 4442/** 4443 * init_idle - set up an idle thread for a given CPU 4444 * @idle: task in question 4445 * @cpu: cpu the idle task belongs to 4446 * 4447 * NOTE: this function does not set the idle thread's NEED_RESCHED 4448 * flag, to make booting more robust. 4449 */ 4450void init_idle(struct task_struct *idle, int cpu) 4451{ 4452 struct rq *rq = cpu_rq(cpu); 4453 unsigned long flags; 4454 4455 raw_spin_lock_irqsave(&rq->lock, flags); 4456 4457 __sched_fork(0, idle); 4458 idle->state = TASK_RUNNING; 4459 idle->se.exec_start = sched_clock(); 4460 4461 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4462 /* 4463 * We're having a chicken and egg problem, even though we are 4464 * holding rq->lock, the cpu isn't yet set to this cpu so the 4465 * lockdep check in task_group() will fail. 4466 * 4467 * Similar case to sched_fork(). / Alternatively we could 4468 * use task_rq_lock() here and obtain the other rq->lock. 4469 * 4470 * Silence PROVE_RCU 4471 */ 4472 rcu_read_lock(); 4473 __set_task_cpu(idle, cpu); 4474 rcu_read_unlock(); 4475 4476 rq->curr = rq->idle = idle; 4477#if defined(CONFIG_SMP) 4478 idle->on_cpu = 1; 4479#endif 4480 raw_spin_unlock_irqrestore(&rq->lock, flags); 4481 4482 /* Set the preempt count _outside_ the spinlocks! */ 4483 init_idle_preempt_count(idle, cpu); 4484 4485 /* 4486 * The idle tasks have their own, simple scheduling class: 4487 */ 4488 idle->sched_class = &idle_sched_class; 4489 ftrace_graph_init_idle_task(idle, cpu); 4490 vtime_init_idle(idle, cpu); 4491#if defined(CONFIG_SMP) 4492 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4493#endif 4494} 4495 4496#ifdef CONFIG_SMP 4497void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4498{ 4499 if (p->sched_class && p->sched_class->set_cpus_allowed) 4500 p->sched_class->set_cpus_allowed(p, new_mask); 4501 4502 cpumask_copy(&p->cpus_allowed, new_mask); 4503 p->nr_cpus_allowed = cpumask_weight(new_mask); 4504} 4505 4506/* 4507 * This is how migration works: 4508 * 4509 * 1) we invoke migration_cpu_stop() on the target CPU using 4510 * stop_one_cpu(). 4511 * 2) stopper starts to run (implicitly forcing the migrated thread 4512 * off the CPU) 4513 * 3) it checks whether the migrated task is still in the wrong runqueue. 4514 * 4) if it's in the wrong runqueue then the migration thread removes 4515 * it and puts it into the right queue. 4516 * 5) stopper completes and stop_one_cpu() returns and the migration 4517 * is done. 4518 */ 4519 4520/* 4521 * Change a given task's CPU affinity. Migrate the thread to a 4522 * proper CPU and schedule it away if the CPU it's executing on 4523 * is removed from the allowed bitmask. 4524 * 4525 * NOTE: the caller must have a valid reference to the task, the 4526 * task must not exit() & deallocate itself prematurely. The 4527 * call is not atomic; no spinlocks may be held. 4528 */ 4529int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4530{ 4531 unsigned long flags; 4532 struct rq *rq; 4533 unsigned int dest_cpu; 4534 int ret = 0; 4535 4536 rq = task_rq_lock(p, &flags); 4537 4538 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4539 goto out; 4540 4541 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4542 ret = -EINVAL; 4543 goto out; 4544 } 4545 4546 do_set_cpus_allowed(p, new_mask); 4547 4548 /* Can the task run on the task's current CPU? If so, we're done */ 4549 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4550 goto out; 4551 4552 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4553 if (p->on_rq) { 4554 struct migration_arg arg = { p, dest_cpu }; 4555 /* Need help from migration thread: drop lock and wait. */ 4556 task_rq_unlock(rq, p, &flags); 4557 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4558 tlb_migrate_finish(p->mm); 4559 return 0; 4560 } 4561out: 4562 task_rq_unlock(rq, p, &flags); 4563 4564 return ret; 4565} 4566EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4567 4568/* 4569 * Move (not current) task off this cpu, onto dest cpu. We're doing 4570 * this because either it can't run here any more (set_cpus_allowed() 4571 * away from this CPU, or CPU going down), or because we're 4572 * attempting to rebalance this task on exec (sched_exec). 4573 * 4574 * So we race with normal scheduler movements, but that's OK, as long 4575 * as the task is no longer on this CPU. 4576 * 4577 * Returns non-zero if task was successfully migrated. 4578 */ 4579static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4580{ 4581 struct rq *rq_dest, *rq_src; 4582 int ret = 0; 4583 4584 if (unlikely(!cpu_active(dest_cpu))) 4585 return ret; 4586 4587 rq_src = cpu_rq(src_cpu); 4588 rq_dest = cpu_rq(dest_cpu); 4589 4590 raw_spin_lock(&p->pi_lock); 4591 double_rq_lock(rq_src, rq_dest); 4592 /* Already moved. */ 4593 if (task_cpu(p) != src_cpu) 4594 goto done; 4595 /* Affinity changed (again). */ 4596 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4597 goto fail; 4598 4599 /* 4600 * If we're not on a rq, the next wake-up will ensure we're 4601 * placed properly. 4602 */ 4603 if (p->on_rq) { 4604 dequeue_task(rq_src, p, 0); 4605 set_task_cpu(p, dest_cpu); 4606 enqueue_task(rq_dest, p, 0); 4607 check_preempt_curr(rq_dest, p, 0); 4608 } 4609done: 4610 ret = 1; 4611fail: 4612 double_rq_unlock(rq_src, rq_dest); 4613 raw_spin_unlock(&p->pi_lock); 4614 return ret; 4615} 4616 4617#ifdef CONFIG_NUMA_BALANCING 4618/* Migrate current task p to target_cpu */ 4619int migrate_task_to(struct task_struct *p, int target_cpu) 4620{ 4621 struct migration_arg arg = { p, target_cpu }; 4622 int curr_cpu = task_cpu(p); 4623 4624 if (curr_cpu == target_cpu) 4625 return 0; 4626 4627 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4628 return -EINVAL; 4629 4630 /* TODO: This is not properly updating schedstats */ 4631 4632 trace_sched_move_numa(p, curr_cpu, target_cpu); 4633 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4634} 4635 4636/* 4637 * Requeue a task on a given node and accurately track the number of NUMA 4638 * tasks on the runqueues 4639 */ 4640void sched_setnuma(struct task_struct *p, int nid) 4641{ 4642 struct rq *rq; 4643 unsigned long flags; 4644 bool on_rq, running; 4645 4646 rq = task_rq_lock(p, &flags); 4647 on_rq = p->on_rq; 4648 running = task_current(rq, p); 4649 4650 if (on_rq) 4651 dequeue_task(rq, p, 0); 4652 if (running) 4653 p->sched_class->put_prev_task(rq, p); 4654 4655 p->numa_preferred_nid = nid; 4656 4657 if (running) 4658 p->sched_class->set_curr_task(rq); 4659 if (on_rq) 4660 enqueue_task(rq, p, 0); 4661 task_rq_unlock(rq, p, &flags); 4662} 4663#endif 4664 4665/* 4666 * migration_cpu_stop - this will be executed by a highprio stopper thread 4667 * and performs thread migration by bumping thread off CPU then 4668 * 'pushing' onto another runqueue. 4669 */ 4670static int migration_cpu_stop(void *data) 4671{ 4672 struct migration_arg *arg = data; 4673 4674 /* 4675 * The original target cpu might have gone down and we might 4676 * be on another cpu but it doesn't matter. 4677 */ 4678 local_irq_disable(); 4679 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4680 local_irq_enable(); 4681 return 0; 4682} 4683 4684#ifdef CONFIG_HOTPLUG_CPU 4685 4686/* 4687 * Ensures that the idle task is using init_mm right before its cpu goes 4688 * offline. 4689 */ 4690void idle_task_exit(void) 4691{ 4692 struct mm_struct *mm = current->active_mm; 4693 4694 BUG_ON(cpu_online(smp_processor_id())); 4695 4696 if (mm != &init_mm) 4697 switch_mm(mm, &init_mm, current); 4698 mmdrop(mm); 4699} 4700 4701/* 4702 * Since this CPU is going 'away' for a while, fold any nr_active delta 4703 * we might have. Assumes we're called after migrate_tasks() so that the 4704 * nr_active count is stable. 4705 * 4706 * Also see the comment "Global load-average calculations". 4707 */ 4708static void calc_load_migrate(struct rq *rq) 4709{ 4710 long delta = calc_load_fold_active(rq); 4711 if (delta) 4712 atomic_long_add(delta, &calc_load_tasks); 4713} 4714 4715/* 4716 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4717 * try_to_wake_up()->select_task_rq(). 4718 * 4719 * Called with rq->lock held even though we'er in stop_machine() and 4720 * there's no concurrency possible, we hold the required locks anyway 4721 * because of lock validation efforts. 4722 */ 4723static void migrate_tasks(unsigned int dead_cpu) 4724{ 4725 struct rq *rq = cpu_rq(dead_cpu); 4726 struct task_struct *next, *stop = rq->stop; 4727 int dest_cpu; 4728 4729 /* 4730 * Fudge the rq selection such that the below task selection loop 4731 * doesn't get stuck on the currently eligible stop task. 4732 * 4733 * We're currently inside stop_machine() and the rq is either stuck 4734 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4735 * either way we should never end up calling schedule() until we're 4736 * done here. 4737 */ 4738 rq->stop = NULL; 4739 4740 /* 4741 * put_prev_task() and pick_next_task() sched 4742 * class method both need to have an up-to-date 4743 * value of rq->clock[_task] 4744 */ 4745 update_rq_clock(rq); 4746 4747 for ( ; ; ) { 4748 /* 4749 * There's this thread running, bail when that's the only 4750 * remaining thread. 4751 */ 4752 if (rq->nr_running == 1) 4753 break; 4754 4755 next = pick_next_task(rq); 4756 BUG_ON(!next); 4757 next->sched_class->put_prev_task(rq, next); 4758 4759 /* Find suitable destination for @next, with force if needed. */ 4760 dest_cpu = select_fallback_rq(dead_cpu, next); 4761 raw_spin_unlock(&rq->lock); 4762 4763 __migrate_task(next, dead_cpu, dest_cpu); 4764 4765 raw_spin_lock(&rq->lock); 4766 } 4767 4768 rq->stop = stop; 4769} 4770 4771#endif /* CONFIG_HOTPLUG_CPU */ 4772 4773#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4774 4775static struct ctl_table sd_ctl_dir[] = { 4776 { 4777 .procname = "sched_domain", 4778 .mode = 0555, 4779 }, 4780 {} 4781}; 4782 4783static struct ctl_table sd_ctl_root[] = { 4784 { 4785 .procname = "kernel", 4786 .mode = 0555, 4787 .child = sd_ctl_dir, 4788 }, 4789 {} 4790}; 4791 4792static struct ctl_table *sd_alloc_ctl_entry(int n) 4793{ 4794 struct ctl_table *entry = 4795 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4796 4797 return entry; 4798} 4799 4800static void sd_free_ctl_entry(struct ctl_table **tablep) 4801{ 4802 struct ctl_table *entry; 4803 4804 /* 4805 * In the intermediate directories, both the child directory and 4806 * procname are dynamically allocated and could fail but the mode 4807 * will always be set. In the lowest directory the names are 4808 * static strings and all have proc handlers. 4809 */ 4810 for (entry = *tablep; entry->mode; entry++) { 4811 if (entry->child) 4812 sd_free_ctl_entry(&entry->child); 4813 if (entry->proc_handler == NULL) 4814 kfree(entry->procname); 4815 } 4816 4817 kfree(*tablep); 4818 *tablep = NULL; 4819} 4820 4821static int min_load_idx = 0; 4822static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4823 4824static void 4825set_table_entry(struct ctl_table *entry, 4826 const char *procname, void *data, int maxlen, 4827 umode_t mode, proc_handler *proc_handler, 4828 bool load_idx) 4829{ 4830 entry->procname = procname; 4831 entry->data = data; 4832 entry->maxlen = maxlen; 4833 entry->mode = mode; 4834 entry->proc_handler = proc_handler; 4835 4836 if (load_idx) { 4837 entry->extra1 = &min_load_idx; 4838 entry->extra2 = &max_load_idx; 4839 } 4840} 4841 4842static struct ctl_table * 4843sd_alloc_ctl_domain_table(struct sched_domain *sd) 4844{ 4845 struct ctl_table *table = sd_alloc_ctl_entry(13); 4846 4847 if (table == NULL) 4848 return NULL; 4849 4850 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4851 sizeof(long), 0644, proc_doulongvec_minmax, false); 4852 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4853 sizeof(long), 0644, proc_doulongvec_minmax, false); 4854 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4855 sizeof(int), 0644, proc_dointvec_minmax, true); 4856 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4857 sizeof(int), 0644, proc_dointvec_minmax, true); 4858 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4859 sizeof(int), 0644, proc_dointvec_minmax, true); 4860 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4861 sizeof(int), 0644, proc_dointvec_minmax, true); 4862 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4863 sizeof(int), 0644, proc_dointvec_minmax, true); 4864 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4865 sizeof(int), 0644, proc_dointvec_minmax, false); 4866 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4867 sizeof(int), 0644, proc_dointvec_minmax, false); 4868 set_table_entry(&table[9], "cache_nice_tries", 4869 &sd->cache_nice_tries, 4870 sizeof(int), 0644, proc_dointvec_minmax, false); 4871 set_table_entry(&table[10], "flags", &sd->flags, 4872 sizeof(int), 0644, proc_dointvec_minmax, false); 4873 set_table_entry(&table[11], "name", sd->name, 4874 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4875 /* &table[12] is terminator */ 4876 4877 return table; 4878} 4879 4880static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4881{ 4882 struct ctl_table *entry, *table; 4883 struct sched_domain *sd; 4884 int domain_num = 0, i; 4885 char buf[32]; 4886 4887 for_each_domain(cpu, sd) 4888 domain_num++; 4889 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4890 if (table == NULL) 4891 return NULL; 4892 4893 i = 0; 4894 for_each_domain(cpu, sd) { 4895 snprintf(buf, 32, "domain%d", i); 4896 entry->procname = kstrdup(buf, GFP_KERNEL); 4897 entry->mode = 0555; 4898 entry->child = sd_alloc_ctl_domain_table(sd); 4899 entry++; 4900 i++; 4901 } 4902 return table; 4903} 4904 4905static struct ctl_table_header *sd_sysctl_header; 4906static void register_sched_domain_sysctl(void) 4907{ 4908 int i, cpu_num = num_possible_cpus(); 4909 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4910 char buf[32]; 4911 4912 WARN_ON(sd_ctl_dir[0].child); 4913 sd_ctl_dir[0].child = entry; 4914 4915 if (entry == NULL) 4916 return; 4917 4918 for_each_possible_cpu(i) { 4919 snprintf(buf, 32, "cpu%d", i); 4920 entry->procname = kstrdup(buf, GFP_KERNEL); 4921 entry->mode = 0555; 4922 entry->child = sd_alloc_ctl_cpu_table(i); 4923 entry++; 4924 } 4925 4926 WARN_ON(sd_sysctl_header); 4927 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4928} 4929 4930/* may be called multiple times per register */ 4931static void unregister_sched_domain_sysctl(void) 4932{ 4933 if (sd_sysctl_header) 4934 unregister_sysctl_table(sd_sysctl_header); 4935 sd_sysctl_header = NULL; 4936 if (sd_ctl_dir[0].child) 4937 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4938} 4939#else 4940static void register_sched_domain_sysctl(void) 4941{ 4942} 4943static void unregister_sched_domain_sysctl(void) 4944{ 4945} 4946#endif 4947 4948static void set_rq_online(struct rq *rq) 4949{ 4950 if (!rq->online) { 4951 const struct sched_class *class; 4952 4953 cpumask_set_cpu(rq->cpu, rq->rd->online); 4954 rq->online = 1; 4955 4956 for_each_class(class) { 4957 if (class->rq_online) 4958 class->rq_online(rq); 4959 } 4960 } 4961} 4962 4963static void set_rq_offline(struct rq *rq) 4964{ 4965 if (rq->online) { 4966 const struct sched_class *class; 4967 4968 for_each_class(class) { 4969 if (class->rq_offline) 4970 class->rq_offline(rq); 4971 } 4972 4973 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4974 rq->online = 0; 4975 } 4976} 4977 4978/* 4979 * migration_call - callback that gets triggered when a CPU is added. 4980 * Here we can start up the necessary migration thread for the new CPU. 4981 */ 4982static int 4983migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4984{ 4985 int cpu = (long)hcpu; 4986 unsigned long flags; 4987 struct rq *rq = cpu_rq(cpu); 4988 4989 switch (action & ~CPU_TASKS_FROZEN) { 4990 4991 case CPU_UP_PREPARE: 4992 rq->calc_load_update = calc_load_update; 4993 break; 4994 4995 case CPU_ONLINE: 4996 /* Update our root-domain */ 4997 raw_spin_lock_irqsave(&rq->lock, flags); 4998 if (rq->rd) { 4999 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5000 5001 set_rq_online(rq); 5002 } 5003 raw_spin_unlock_irqrestore(&rq->lock, flags); 5004 break; 5005 5006#ifdef CONFIG_HOTPLUG_CPU 5007 case CPU_DYING: 5008 sched_ttwu_pending(); 5009 /* Update our root-domain */ 5010 raw_spin_lock_irqsave(&rq->lock, flags); 5011 if (rq->rd) { 5012 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5013 set_rq_offline(rq); 5014 } 5015 migrate_tasks(cpu); 5016 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5017 raw_spin_unlock_irqrestore(&rq->lock, flags); 5018 break; 5019 5020 case CPU_DEAD: 5021 calc_load_migrate(rq); 5022 break; 5023#endif 5024 } 5025 5026 update_max_interval(); 5027 5028 return NOTIFY_OK; 5029} 5030 5031/* 5032 * Register at high priority so that task migration (migrate_all_tasks) 5033 * happens before everything else. This has to be lower priority than 5034 * the notifier in the perf_event subsystem, though. 5035 */ 5036static struct notifier_block migration_notifier = { 5037 .notifier_call = migration_call, 5038 .priority = CPU_PRI_MIGRATION, 5039}; 5040 5041static int sched_cpu_active(struct notifier_block *nfb, 5042 unsigned long action, void *hcpu) 5043{ 5044 switch (action & ~CPU_TASKS_FROZEN) { 5045 case CPU_STARTING: 5046 case CPU_DOWN_FAILED: 5047 set_cpu_active((long)hcpu, true); 5048 return NOTIFY_OK; 5049 default: 5050 return NOTIFY_DONE; 5051 } 5052} 5053 5054static int sched_cpu_inactive(struct notifier_block *nfb, 5055 unsigned long action, void *hcpu) 5056{ 5057 unsigned long flags; 5058 long cpu = (long)hcpu; 5059 5060 switch (action & ~CPU_TASKS_FROZEN) { 5061 case CPU_DOWN_PREPARE: 5062 set_cpu_active(cpu, false); 5063 5064 /* explicitly allow suspend */ 5065 if (!(action & CPU_TASKS_FROZEN)) { 5066 struct dl_bw *dl_b = dl_bw_of(cpu); 5067 bool overflow; 5068 int cpus; 5069 5070 raw_spin_lock_irqsave(&dl_b->lock, flags); 5071 cpus = dl_bw_cpus(cpu); 5072 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5073 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5074 5075 if (overflow) 5076 return notifier_from_errno(-EBUSY); 5077 } 5078 return NOTIFY_OK; 5079 } 5080 5081 return NOTIFY_DONE; 5082} 5083 5084static int __init migration_init(void) 5085{ 5086 void *cpu = (void *)(long)smp_processor_id(); 5087 int err; 5088 5089 /* Initialize migration for the boot CPU */ 5090 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5091 BUG_ON(err == NOTIFY_BAD); 5092 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5093 register_cpu_notifier(&migration_notifier); 5094 5095 /* Register cpu active notifiers */ 5096 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5097 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5098 5099 return 0; 5100} 5101early_initcall(migration_init); 5102#endif 5103 5104#ifdef CONFIG_SMP 5105 5106static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5107 5108#ifdef CONFIG_SCHED_DEBUG 5109 5110static __read_mostly int sched_debug_enabled; 5111 5112static int __init sched_debug_setup(char *str) 5113{ 5114 sched_debug_enabled = 1; 5115 5116 return 0; 5117} 5118early_param("sched_debug", sched_debug_setup); 5119 5120static inline bool sched_debug(void) 5121{ 5122 return sched_debug_enabled; 5123} 5124 5125static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5126 struct cpumask *groupmask) 5127{ 5128 struct sched_group *group = sd->groups; 5129 char str[256]; 5130 5131 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5132 cpumask_clear(groupmask); 5133 5134 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5135 5136 if (!(sd->flags & SD_LOAD_BALANCE)) { 5137 printk("does not load-balance\n"); 5138 if (sd->parent) 5139 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5140 " has parent"); 5141 return -1; 5142 } 5143 5144 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5145 5146 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5147 printk(KERN_ERR "ERROR: domain->span does not contain " 5148 "CPU%d\n", cpu); 5149 } 5150 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5151 printk(KERN_ERR "ERROR: domain->groups does not contain" 5152 " CPU%d\n", cpu); 5153 } 5154 5155 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5156 do { 5157 if (!group) { 5158 printk("\n"); 5159 printk(KERN_ERR "ERROR: group is NULL\n"); 5160 break; 5161 } 5162 5163 /* 5164 * Even though we initialize ->power to something semi-sane, 5165 * we leave power_orig unset. This allows us to detect if 5166 * domain iteration is still funny without causing /0 traps. 5167 */ 5168 if (!group->sgp->power_orig) { 5169 printk(KERN_CONT "\n"); 5170 printk(KERN_ERR "ERROR: domain->cpu_power not " 5171 "set\n"); 5172 break; 5173 } 5174 5175 if (!cpumask_weight(sched_group_cpus(group))) { 5176 printk(KERN_CONT "\n"); 5177 printk(KERN_ERR "ERROR: empty group\n"); 5178 break; 5179 } 5180 5181 if (!(sd->flags & SD_OVERLAP) && 5182 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5183 printk(KERN_CONT "\n"); 5184 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5185 break; 5186 } 5187 5188 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5189 5190 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5191 5192 printk(KERN_CONT " %s", str); 5193 if (group->sgp->power != SCHED_POWER_SCALE) { 5194 printk(KERN_CONT " (cpu_power = %d)", 5195 group->sgp->power); 5196 } 5197 5198 group = group->next; 5199 } while (group != sd->groups); 5200 printk(KERN_CONT "\n"); 5201 5202 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5203 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5204 5205 if (sd->parent && 5206 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5207 printk(KERN_ERR "ERROR: parent span is not a superset " 5208 "of domain->span\n"); 5209 return 0; 5210} 5211 5212static void sched_domain_debug(struct sched_domain *sd, int cpu) 5213{ 5214 int level = 0; 5215 5216 if (!sched_debug_enabled) 5217 return; 5218 5219 if (!sd) { 5220 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5221 return; 5222 } 5223 5224 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5225 5226 for (;;) { 5227 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5228 break; 5229 level++; 5230 sd = sd->parent; 5231 if (!sd) 5232 break; 5233 } 5234} 5235#else /* !CONFIG_SCHED_DEBUG */ 5236# define sched_domain_debug(sd, cpu) do { } while (0) 5237static inline bool sched_debug(void) 5238{ 5239 return false; 5240} 5241#endif /* CONFIG_SCHED_DEBUG */ 5242 5243static int sd_degenerate(struct sched_domain *sd) 5244{ 5245 if (cpumask_weight(sched_domain_span(sd)) == 1) 5246 return 1; 5247 5248 /* Following flags need at least 2 groups */ 5249 if (sd->flags & (SD_LOAD_BALANCE | 5250 SD_BALANCE_NEWIDLE | 5251 SD_BALANCE_FORK | 5252 SD_BALANCE_EXEC | 5253 SD_SHARE_CPUPOWER | 5254 SD_SHARE_PKG_RESOURCES)) { 5255 if (sd->groups != sd->groups->next) 5256 return 0; 5257 } 5258 5259 /* Following flags don't use groups */ 5260 if (sd->flags & (SD_WAKE_AFFINE)) 5261 return 0; 5262 5263 return 1; 5264} 5265 5266static int 5267sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5268{ 5269 unsigned long cflags = sd->flags, pflags = parent->flags; 5270 5271 if (sd_degenerate(parent)) 5272 return 1; 5273 5274 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5275 return 0; 5276 5277 /* Flags needing groups don't count if only 1 group in parent */ 5278 if (parent->groups == parent->groups->next) { 5279 pflags &= ~(SD_LOAD_BALANCE | 5280 SD_BALANCE_NEWIDLE | 5281 SD_BALANCE_FORK | 5282 SD_BALANCE_EXEC | 5283 SD_SHARE_CPUPOWER | 5284 SD_SHARE_PKG_RESOURCES | 5285 SD_PREFER_SIBLING); 5286 if (nr_node_ids == 1) 5287 pflags &= ~SD_SERIALIZE; 5288 } 5289 if (~cflags & pflags) 5290 return 0; 5291 5292 return 1; 5293} 5294 5295static void free_rootdomain(struct rcu_head *rcu) 5296{ 5297 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5298 5299 cpupri_cleanup(&rd->cpupri); 5300 cpudl_cleanup(&rd->cpudl); 5301 free_cpumask_var(rd->dlo_mask); 5302 free_cpumask_var(rd->rto_mask); 5303 free_cpumask_var(rd->online); 5304 free_cpumask_var(rd->span); 5305 kfree(rd); 5306} 5307 5308static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5309{ 5310 struct root_domain *old_rd = NULL; 5311 unsigned long flags; 5312 5313 raw_spin_lock_irqsave(&rq->lock, flags); 5314 5315 if (rq->rd) { 5316 old_rd = rq->rd; 5317 5318 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5319 set_rq_offline(rq); 5320 5321 cpumask_clear_cpu(rq->cpu, old_rd->span); 5322 5323 /* 5324 * If we dont want to free the old_rd yet then 5325 * set old_rd to NULL to skip the freeing later 5326 * in this function: 5327 */ 5328 if (!atomic_dec_and_test(&old_rd->refcount)) 5329 old_rd = NULL; 5330 } 5331 5332 atomic_inc(&rd->refcount); 5333 rq->rd = rd; 5334 5335 cpumask_set_cpu(rq->cpu, rd->span); 5336 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5337 set_rq_online(rq); 5338 5339 raw_spin_unlock_irqrestore(&rq->lock, flags); 5340 5341 if (old_rd) 5342 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5343} 5344 5345static int init_rootdomain(struct root_domain *rd) 5346{ 5347 memset(rd, 0, sizeof(*rd)); 5348 5349 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5350 goto out; 5351 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5352 goto free_span; 5353 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5354 goto free_online; 5355 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5356 goto free_dlo_mask; 5357 5358 init_dl_bw(&rd->dl_bw); 5359 if (cpudl_init(&rd->cpudl) != 0) 5360 goto free_dlo_mask; 5361 5362 if (cpupri_init(&rd->cpupri) != 0) 5363 goto free_rto_mask; 5364 return 0; 5365 5366free_rto_mask: 5367 free_cpumask_var(rd->rto_mask); 5368free_dlo_mask: 5369 free_cpumask_var(rd->dlo_mask); 5370free_online: 5371 free_cpumask_var(rd->online); 5372free_span: 5373 free_cpumask_var(rd->span); 5374out: 5375 return -ENOMEM; 5376} 5377 5378/* 5379 * By default the system creates a single root-domain with all cpus as 5380 * members (mimicking the global state we have today). 5381 */ 5382struct root_domain def_root_domain; 5383 5384static void init_defrootdomain(void) 5385{ 5386 init_rootdomain(&def_root_domain); 5387 5388 atomic_set(&def_root_domain.refcount, 1); 5389} 5390 5391static struct root_domain *alloc_rootdomain(void) 5392{ 5393 struct root_domain *rd; 5394 5395 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5396 if (!rd) 5397 return NULL; 5398 5399 if (init_rootdomain(rd) != 0) { 5400 kfree(rd); 5401 return NULL; 5402 } 5403 5404 return rd; 5405} 5406 5407static void free_sched_groups(struct sched_group *sg, int free_sgp) 5408{ 5409 struct sched_group *tmp, *first; 5410 5411 if (!sg) 5412 return; 5413 5414 first = sg; 5415 do { 5416 tmp = sg->next; 5417 5418 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5419 kfree(sg->sgp); 5420 5421 kfree(sg); 5422 sg = tmp; 5423 } while (sg != first); 5424} 5425 5426static void free_sched_domain(struct rcu_head *rcu) 5427{ 5428 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5429 5430 /* 5431 * If its an overlapping domain it has private groups, iterate and 5432 * nuke them all. 5433 */ 5434 if (sd->flags & SD_OVERLAP) { 5435 free_sched_groups(sd->groups, 1); 5436 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5437 kfree(sd->groups->sgp); 5438 kfree(sd->groups); 5439 } 5440 kfree(sd); 5441} 5442 5443static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5444{ 5445 call_rcu(&sd->rcu, free_sched_domain); 5446} 5447 5448static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5449{ 5450 for (; sd; sd = sd->parent) 5451 destroy_sched_domain(sd, cpu); 5452} 5453 5454/* 5455 * Keep a special pointer to the highest sched_domain that has 5456 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5457 * allows us to avoid some pointer chasing select_idle_sibling(). 5458 * 5459 * Also keep a unique ID per domain (we use the first cpu number in 5460 * the cpumask of the domain), this allows us to quickly tell if 5461 * two cpus are in the same cache domain, see cpus_share_cache(). 5462 */ 5463DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5464DEFINE_PER_CPU(int, sd_llc_size); 5465DEFINE_PER_CPU(int, sd_llc_id); 5466DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5467DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5468DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5469 5470static void update_top_cache_domain(int cpu) 5471{ 5472 struct sched_domain *sd; 5473 struct sched_domain *busy_sd = NULL; 5474 int id = cpu; 5475 int size = 1; 5476 5477 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5478 if (sd) { 5479 id = cpumask_first(sched_domain_span(sd)); 5480 size = cpumask_weight(sched_domain_span(sd)); 5481 busy_sd = sd->parent; /* sd_busy */ 5482 } 5483 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5484 5485 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5486 per_cpu(sd_llc_size, cpu) = size; 5487 per_cpu(sd_llc_id, cpu) = id; 5488 5489 sd = lowest_flag_domain(cpu, SD_NUMA); 5490 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5491 5492 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5493 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5494} 5495 5496/* 5497 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5498 * hold the hotplug lock. 5499 */ 5500static void 5501cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5502{ 5503 struct rq *rq = cpu_rq(cpu); 5504 struct sched_domain *tmp; 5505 5506 /* Remove the sched domains which do not contribute to scheduling. */ 5507 for (tmp = sd; tmp; ) { 5508 struct sched_domain *parent = tmp->parent; 5509 if (!parent) 5510 break; 5511 5512 if (sd_parent_degenerate(tmp, parent)) { 5513 tmp->parent = parent->parent; 5514 if (parent->parent) 5515 parent->parent->child = tmp; 5516 /* 5517 * Transfer SD_PREFER_SIBLING down in case of a 5518 * degenerate parent; the spans match for this 5519 * so the property transfers. 5520 */ 5521 if (parent->flags & SD_PREFER_SIBLING) 5522 tmp->flags |= SD_PREFER_SIBLING; 5523 destroy_sched_domain(parent, cpu); 5524 } else 5525 tmp = tmp->parent; 5526 } 5527 5528 if (sd && sd_degenerate(sd)) { 5529 tmp = sd; 5530 sd = sd->parent; 5531 destroy_sched_domain(tmp, cpu); 5532 if (sd) 5533 sd->child = NULL; 5534 } 5535 5536 sched_domain_debug(sd, cpu); 5537 5538 rq_attach_root(rq, rd); 5539 tmp = rq->sd; 5540 rcu_assign_pointer(rq->sd, sd); 5541 destroy_sched_domains(tmp, cpu); 5542 5543 update_top_cache_domain(cpu); 5544} 5545 5546/* cpus with isolated domains */ 5547static cpumask_var_t cpu_isolated_map; 5548 5549/* Setup the mask of cpus configured for isolated domains */ 5550static int __init isolated_cpu_setup(char *str) 5551{ 5552 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5553 cpulist_parse(str, cpu_isolated_map); 5554 return 1; 5555} 5556 5557__setup("isolcpus=", isolated_cpu_setup); 5558 5559static const struct cpumask *cpu_cpu_mask(int cpu) 5560{ 5561 return cpumask_of_node(cpu_to_node(cpu)); 5562} 5563 5564struct sd_data { 5565 struct sched_domain **__percpu sd; 5566 struct sched_group **__percpu sg; 5567 struct sched_group_power **__percpu sgp; 5568}; 5569 5570struct s_data { 5571 struct sched_domain ** __percpu sd; 5572 struct root_domain *rd; 5573}; 5574 5575enum s_alloc { 5576 sa_rootdomain, 5577 sa_sd, 5578 sa_sd_storage, 5579 sa_none, 5580}; 5581 5582struct sched_domain_topology_level; 5583 5584typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5585typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5586 5587#define SDTL_OVERLAP 0x01 5588 5589struct sched_domain_topology_level { 5590 sched_domain_init_f init; 5591 sched_domain_mask_f mask; 5592 int flags; 5593 int numa_level; 5594 struct sd_data data; 5595}; 5596 5597/* 5598 * Build an iteration mask that can exclude certain CPUs from the upwards 5599 * domain traversal. 5600 * 5601 * Asymmetric node setups can result in situations where the domain tree is of 5602 * unequal depth, make sure to skip domains that already cover the entire 5603 * range. 5604 * 5605 * In that case build_sched_domains() will have terminated the iteration early 5606 * and our sibling sd spans will be empty. Domains should always include the 5607 * cpu they're built on, so check that. 5608 * 5609 */ 5610static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5611{ 5612 const struct cpumask *span = sched_domain_span(sd); 5613 struct sd_data *sdd = sd->private; 5614 struct sched_domain *sibling; 5615 int i; 5616 5617 for_each_cpu(i, span) { 5618 sibling = *per_cpu_ptr(sdd->sd, i); 5619 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5620 continue; 5621 5622 cpumask_set_cpu(i, sched_group_mask(sg)); 5623 } 5624} 5625 5626/* 5627 * Return the canonical balance cpu for this group, this is the first cpu 5628 * of this group that's also in the iteration mask. 5629 */ 5630int group_balance_cpu(struct sched_group *sg) 5631{ 5632 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5633} 5634 5635static int 5636build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5637{ 5638 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5639 const struct cpumask *span = sched_domain_span(sd); 5640 struct cpumask *covered = sched_domains_tmpmask; 5641 struct sd_data *sdd = sd->private; 5642 struct sched_domain *child; 5643 int i; 5644 5645 cpumask_clear(covered); 5646 5647 for_each_cpu(i, span) { 5648 struct cpumask *sg_span; 5649 5650 if (cpumask_test_cpu(i, covered)) 5651 continue; 5652 5653 child = *per_cpu_ptr(sdd->sd, i); 5654 5655 /* See the comment near build_group_mask(). */ 5656 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5657 continue; 5658 5659 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5660 GFP_KERNEL, cpu_to_node(cpu)); 5661 5662 if (!sg) 5663 goto fail; 5664 5665 sg_span = sched_group_cpus(sg); 5666 if (child->child) { 5667 child = child->child; 5668 cpumask_copy(sg_span, sched_domain_span(child)); 5669 } else 5670 cpumask_set_cpu(i, sg_span); 5671 5672 cpumask_or(covered, covered, sg_span); 5673 5674 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5675 if (atomic_inc_return(&sg->sgp->ref) == 1) 5676 build_group_mask(sd, sg); 5677 5678 /* 5679 * Initialize sgp->power such that even if we mess up the 5680 * domains and no possible iteration will get us here, we won't 5681 * die on a /0 trap. 5682 */ 5683 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5684 sg->sgp->power_orig = sg->sgp->power; 5685 5686 /* 5687 * Make sure the first group of this domain contains the 5688 * canonical balance cpu. Otherwise the sched_domain iteration 5689 * breaks. See update_sg_lb_stats(). 5690 */ 5691 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5692 group_balance_cpu(sg) == cpu) 5693 groups = sg; 5694 5695 if (!first) 5696 first = sg; 5697 if (last) 5698 last->next = sg; 5699 last = sg; 5700 last->next = first; 5701 } 5702 sd->groups = groups; 5703 5704 return 0; 5705 5706fail: 5707 free_sched_groups(first, 0); 5708 5709 return -ENOMEM; 5710} 5711 5712static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5713{ 5714 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5715 struct sched_domain *child = sd->child; 5716 5717 if (child) 5718 cpu = cpumask_first(sched_domain_span(child)); 5719 5720 if (sg) { 5721 *sg = *per_cpu_ptr(sdd->sg, cpu); 5722 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5723 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5724 } 5725 5726 return cpu; 5727} 5728 5729/* 5730 * build_sched_groups will build a circular linked list of the groups 5731 * covered by the given span, and will set each group's ->cpumask correctly, 5732 * and ->cpu_power to 0. 5733 * 5734 * Assumes the sched_domain tree is fully constructed 5735 */ 5736static int 5737build_sched_groups(struct sched_domain *sd, int cpu) 5738{ 5739 struct sched_group *first = NULL, *last = NULL; 5740 struct sd_data *sdd = sd->private; 5741 const struct cpumask *span = sched_domain_span(sd); 5742 struct cpumask *covered; 5743 int i; 5744 5745 get_group(cpu, sdd, &sd->groups); 5746 atomic_inc(&sd->groups->ref); 5747 5748 if (cpu != cpumask_first(span)) 5749 return 0; 5750 5751 lockdep_assert_held(&sched_domains_mutex); 5752 covered = sched_domains_tmpmask; 5753 5754 cpumask_clear(covered); 5755 5756 for_each_cpu(i, span) { 5757 struct sched_group *sg; 5758 int group, j; 5759 5760 if (cpumask_test_cpu(i, covered)) 5761 continue; 5762 5763 group = get_group(i, sdd, &sg); 5764 cpumask_clear(sched_group_cpus(sg)); 5765 sg->sgp->power = 0; 5766 cpumask_setall(sched_group_mask(sg)); 5767 5768 for_each_cpu(j, span) { 5769 if (get_group(j, sdd, NULL) != group) 5770 continue; 5771 5772 cpumask_set_cpu(j, covered); 5773 cpumask_set_cpu(j, sched_group_cpus(sg)); 5774 } 5775 5776 if (!first) 5777 first = sg; 5778 if (last) 5779 last->next = sg; 5780 last = sg; 5781 } 5782 last->next = first; 5783 5784 return 0; 5785} 5786 5787/* 5788 * Initialize sched groups cpu_power. 5789 * 5790 * cpu_power indicates the capacity of sched group, which is used while 5791 * distributing the load between different sched groups in a sched domain. 5792 * Typically cpu_power for all the groups in a sched domain will be same unless 5793 * there are asymmetries in the topology. If there are asymmetries, group 5794 * having more cpu_power will pickup more load compared to the group having 5795 * less cpu_power. 5796 */ 5797static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5798{ 5799 struct sched_group *sg = sd->groups; 5800 5801 WARN_ON(!sg); 5802 5803 do { 5804 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5805 sg = sg->next; 5806 } while (sg != sd->groups); 5807 5808 if (cpu != group_balance_cpu(sg)) 5809 return; 5810 5811 update_group_power(sd, cpu); 5812 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5813} 5814 5815int __weak arch_sd_sibling_asym_packing(void) 5816{ 5817 return 0*SD_ASYM_PACKING; 5818} 5819 5820/* 5821 * Initializers for schedule domains 5822 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5823 */ 5824 5825#ifdef CONFIG_SCHED_DEBUG 5826# define SD_INIT_NAME(sd, type) sd->name = #type 5827#else 5828# define SD_INIT_NAME(sd, type) do { } while (0) 5829#endif 5830 5831#define SD_INIT_FUNC(type) \ 5832static noinline struct sched_domain * \ 5833sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5834{ \ 5835 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5836 *sd = SD_##type##_INIT; \ 5837 SD_INIT_NAME(sd, type); \ 5838 sd->private = &tl->data; \ 5839 return sd; \ 5840} 5841 5842SD_INIT_FUNC(CPU) 5843#ifdef CONFIG_SCHED_SMT 5844 SD_INIT_FUNC(SIBLING) 5845#endif 5846#ifdef CONFIG_SCHED_MC 5847 SD_INIT_FUNC(MC) 5848#endif 5849#ifdef CONFIG_SCHED_BOOK 5850 SD_INIT_FUNC(BOOK) 5851#endif 5852 5853static int default_relax_domain_level = -1; 5854int sched_domain_level_max; 5855 5856static int __init setup_relax_domain_level(char *str) 5857{ 5858 if (kstrtoint(str, 0, &default_relax_domain_level)) 5859 pr_warn("Unable to set relax_domain_level\n"); 5860 5861 return 1; 5862} 5863__setup("relax_domain_level=", setup_relax_domain_level); 5864 5865static void set_domain_attribute(struct sched_domain *sd, 5866 struct sched_domain_attr *attr) 5867{ 5868 int request; 5869 5870 if (!attr || attr->relax_domain_level < 0) { 5871 if (default_relax_domain_level < 0) 5872 return; 5873 else 5874 request = default_relax_domain_level; 5875 } else 5876 request = attr->relax_domain_level; 5877 if (request < sd->level) { 5878 /* turn off idle balance on this domain */ 5879 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5880 } else { 5881 /* turn on idle balance on this domain */ 5882 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5883 } 5884} 5885 5886static void __sdt_free(const struct cpumask *cpu_map); 5887static int __sdt_alloc(const struct cpumask *cpu_map); 5888 5889static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5890 const struct cpumask *cpu_map) 5891{ 5892 switch (what) { 5893 case sa_rootdomain: 5894 if (!atomic_read(&d->rd->refcount)) 5895 free_rootdomain(&d->rd->rcu); /* fall through */ 5896 case sa_sd: 5897 free_percpu(d->sd); /* fall through */ 5898 case sa_sd_storage: 5899 __sdt_free(cpu_map); /* fall through */ 5900 case sa_none: 5901 break; 5902 } 5903} 5904 5905static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5906 const struct cpumask *cpu_map) 5907{ 5908 memset(d, 0, sizeof(*d)); 5909 5910 if (__sdt_alloc(cpu_map)) 5911 return sa_sd_storage; 5912 d->sd = alloc_percpu(struct sched_domain *); 5913 if (!d->sd) 5914 return sa_sd_storage; 5915 d->rd = alloc_rootdomain(); 5916 if (!d->rd) 5917 return sa_sd; 5918 return sa_rootdomain; 5919} 5920 5921/* 5922 * NULL the sd_data elements we've used to build the sched_domain and 5923 * sched_group structure so that the subsequent __free_domain_allocs() 5924 * will not free the data we're using. 5925 */ 5926static void claim_allocations(int cpu, struct sched_domain *sd) 5927{ 5928 struct sd_data *sdd = sd->private; 5929 5930 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5931 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5932 5933 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5934 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5935 5936 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5937 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5938} 5939 5940#ifdef CONFIG_SCHED_SMT 5941static const struct cpumask *cpu_smt_mask(int cpu) 5942{ 5943 return topology_thread_cpumask(cpu); 5944} 5945#endif 5946 5947/* 5948 * Topology list, bottom-up. 5949 */ 5950static struct sched_domain_topology_level default_topology[] = { 5951#ifdef CONFIG_SCHED_SMT 5952 { sd_init_SIBLING, cpu_smt_mask, }, 5953#endif 5954#ifdef CONFIG_SCHED_MC 5955 { sd_init_MC, cpu_coregroup_mask, }, 5956#endif 5957#ifdef CONFIG_SCHED_BOOK 5958 { sd_init_BOOK, cpu_book_mask, }, 5959#endif 5960 { sd_init_CPU, cpu_cpu_mask, }, 5961 { NULL, }, 5962}; 5963 5964static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5965 5966#define for_each_sd_topology(tl) \ 5967 for (tl = sched_domain_topology; tl->init; tl++) 5968 5969#ifdef CONFIG_NUMA 5970 5971static int sched_domains_numa_levels; 5972static int *sched_domains_numa_distance; 5973static struct cpumask ***sched_domains_numa_masks; 5974static int sched_domains_curr_level; 5975 5976static inline int sd_local_flags(int level) 5977{ 5978 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5979 return 0; 5980 5981 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5982} 5983 5984static struct sched_domain * 5985sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5986{ 5987 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5988 int level = tl->numa_level; 5989 int sd_weight = cpumask_weight( 5990 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5991 5992 *sd = (struct sched_domain){ 5993 .min_interval = sd_weight, 5994 .max_interval = 2*sd_weight, 5995 .busy_factor = 32, 5996 .imbalance_pct = 125, 5997 .cache_nice_tries = 2, 5998 .busy_idx = 3, 5999 .idle_idx = 2, 6000 .newidle_idx = 0, 6001 .wake_idx = 0, 6002 .forkexec_idx = 0, 6003 6004 .flags = 1*SD_LOAD_BALANCE 6005 | 1*SD_BALANCE_NEWIDLE 6006 | 0*SD_BALANCE_EXEC 6007 | 0*SD_BALANCE_FORK 6008 | 0*SD_BALANCE_WAKE 6009 | 0*SD_WAKE_AFFINE 6010 | 0*SD_SHARE_CPUPOWER 6011 | 0*SD_SHARE_PKG_RESOURCES 6012 | 1*SD_SERIALIZE 6013 | 0*SD_PREFER_SIBLING 6014 | 1*SD_NUMA 6015 | sd_local_flags(level) 6016 , 6017 .last_balance = jiffies, 6018 .balance_interval = sd_weight, 6019 }; 6020 SD_INIT_NAME(sd, NUMA); 6021 sd->private = &tl->data; 6022 6023 /* 6024 * Ugly hack to pass state to sd_numa_mask()... 6025 */ 6026 sched_domains_curr_level = tl->numa_level; 6027 6028 return sd; 6029} 6030 6031static const struct cpumask *sd_numa_mask(int cpu) 6032{ 6033 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6034} 6035 6036static void sched_numa_warn(const char *str) 6037{ 6038 static int done = false; 6039 int i,j; 6040 6041 if (done) 6042 return; 6043 6044 done = true; 6045 6046 printk(KERN_WARNING "ERROR: %s\n\n", str); 6047 6048 for (i = 0; i < nr_node_ids; i++) { 6049 printk(KERN_WARNING " "); 6050 for (j = 0; j < nr_node_ids; j++) 6051 printk(KERN_CONT "%02d ", node_distance(i,j)); 6052 printk(KERN_CONT "\n"); 6053 } 6054 printk(KERN_WARNING "\n"); 6055} 6056 6057static bool find_numa_distance(int distance) 6058{ 6059 int i; 6060 6061 if (distance == node_distance(0, 0)) 6062 return true; 6063 6064 for (i = 0; i < sched_domains_numa_levels; i++) { 6065 if (sched_domains_numa_distance[i] == distance) 6066 return true; 6067 } 6068 6069 return false; 6070} 6071 6072static void sched_init_numa(void) 6073{ 6074 int next_distance, curr_distance = node_distance(0, 0); 6075 struct sched_domain_topology_level *tl; 6076 int level = 0; 6077 int i, j, k; 6078 6079 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6080 if (!sched_domains_numa_distance) 6081 return; 6082 6083 /* 6084 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6085 * unique distances in the node_distance() table. 6086 * 6087 * Assumes node_distance(0,j) includes all distances in 6088 * node_distance(i,j) in order to avoid cubic time. 6089 */ 6090 next_distance = curr_distance; 6091 for (i = 0; i < nr_node_ids; i++) { 6092 for (j = 0; j < nr_node_ids; j++) { 6093 for (k = 0; k < nr_node_ids; k++) { 6094 int distance = node_distance(i, k); 6095 6096 if (distance > curr_distance && 6097 (distance < next_distance || 6098 next_distance == curr_distance)) 6099 next_distance = distance; 6100 6101 /* 6102 * While not a strong assumption it would be nice to know 6103 * about cases where if node A is connected to B, B is not 6104 * equally connected to A. 6105 */ 6106 if (sched_debug() && node_distance(k, i) != distance) 6107 sched_numa_warn("Node-distance not symmetric"); 6108 6109 if (sched_debug() && i && !find_numa_distance(distance)) 6110 sched_numa_warn("Node-0 not representative"); 6111 } 6112 if (next_distance != curr_distance) { 6113 sched_domains_numa_distance[level++] = next_distance; 6114 sched_domains_numa_levels = level; 6115 curr_distance = next_distance; 6116 } else break; 6117 } 6118 6119 /* 6120 * In case of sched_debug() we verify the above assumption. 6121 */ 6122 if (!sched_debug()) 6123 break; 6124 } 6125 /* 6126 * 'level' contains the number of unique distances, excluding the 6127 * identity distance node_distance(i,i). 6128 * 6129 * The sched_domains_numa_distance[] array includes the actual distance 6130 * numbers. 6131 */ 6132 6133 /* 6134 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6135 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6136 * the array will contain less then 'level' members. This could be 6137 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6138 * in other functions. 6139 * 6140 * We reset it to 'level' at the end of this function. 6141 */ 6142 sched_domains_numa_levels = 0; 6143 6144 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6145 if (!sched_domains_numa_masks) 6146 return; 6147 6148 /* 6149 * Now for each level, construct a mask per node which contains all 6150 * cpus of nodes that are that many hops away from us. 6151 */ 6152 for (i = 0; i < level; i++) { 6153 sched_domains_numa_masks[i] = 6154 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6155 if (!sched_domains_numa_masks[i]) 6156 return; 6157 6158 for (j = 0; j < nr_node_ids; j++) { 6159 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6160 if (!mask) 6161 return; 6162 6163 sched_domains_numa_masks[i][j] = mask; 6164 6165 for (k = 0; k < nr_node_ids; k++) { 6166 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6167 continue; 6168 6169 cpumask_or(mask, mask, cpumask_of_node(k)); 6170 } 6171 } 6172 } 6173 6174 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6175 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6176 if (!tl) 6177 return; 6178 6179 /* 6180 * Copy the default topology bits.. 6181 */ 6182 for (i = 0; default_topology[i].init; i++) 6183 tl[i] = default_topology[i]; 6184 6185 /* 6186 * .. and append 'j' levels of NUMA goodness. 6187 */ 6188 for (j = 0; j < level; i++, j++) { 6189 tl[i] = (struct sched_domain_topology_level){ 6190 .init = sd_numa_init, 6191 .mask = sd_numa_mask, 6192 .flags = SDTL_OVERLAP, 6193 .numa_level = j, 6194 }; 6195 } 6196 6197 sched_domain_topology = tl; 6198 6199 sched_domains_numa_levels = level; 6200} 6201 6202static void sched_domains_numa_masks_set(int cpu) 6203{ 6204 int i, j; 6205 int node = cpu_to_node(cpu); 6206 6207 for (i = 0; i < sched_domains_numa_levels; i++) { 6208 for (j = 0; j < nr_node_ids; j++) { 6209 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6210 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6211 } 6212 } 6213} 6214 6215static void sched_domains_numa_masks_clear(int cpu) 6216{ 6217 int i, j; 6218 for (i = 0; i < sched_domains_numa_levels; i++) { 6219 for (j = 0; j < nr_node_ids; j++) 6220 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6221 } 6222} 6223 6224/* 6225 * Update sched_domains_numa_masks[level][node] array when new cpus 6226 * are onlined. 6227 */ 6228static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6229 unsigned long action, 6230 void *hcpu) 6231{ 6232 int cpu = (long)hcpu; 6233 6234 switch (action & ~CPU_TASKS_FROZEN) { 6235 case CPU_ONLINE: 6236 sched_domains_numa_masks_set(cpu); 6237 break; 6238 6239 case CPU_DEAD: 6240 sched_domains_numa_masks_clear(cpu); 6241 break; 6242 6243 default: 6244 return NOTIFY_DONE; 6245 } 6246 6247 return NOTIFY_OK; 6248} 6249#else 6250static inline void sched_init_numa(void) 6251{ 6252} 6253 6254static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6255 unsigned long action, 6256 void *hcpu) 6257{ 6258 return 0; 6259} 6260#endif /* CONFIG_NUMA */ 6261 6262static int __sdt_alloc(const struct cpumask *cpu_map) 6263{ 6264 struct sched_domain_topology_level *tl; 6265 int j; 6266 6267 for_each_sd_topology(tl) { 6268 struct sd_data *sdd = &tl->data; 6269 6270 sdd->sd = alloc_percpu(struct sched_domain *); 6271 if (!sdd->sd) 6272 return -ENOMEM; 6273 6274 sdd->sg = alloc_percpu(struct sched_group *); 6275 if (!sdd->sg) 6276 return -ENOMEM; 6277 6278 sdd->sgp = alloc_percpu(struct sched_group_power *); 6279 if (!sdd->sgp) 6280 return -ENOMEM; 6281 6282 for_each_cpu(j, cpu_map) { 6283 struct sched_domain *sd; 6284 struct sched_group *sg; 6285 struct sched_group_power *sgp; 6286 6287 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6288 GFP_KERNEL, cpu_to_node(j)); 6289 if (!sd) 6290 return -ENOMEM; 6291 6292 *per_cpu_ptr(sdd->sd, j) = sd; 6293 6294 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6295 GFP_KERNEL, cpu_to_node(j)); 6296 if (!sg) 6297 return -ENOMEM; 6298 6299 sg->next = sg; 6300 6301 *per_cpu_ptr(sdd->sg, j) = sg; 6302 6303 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6304 GFP_KERNEL, cpu_to_node(j)); 6305 if (!sgp) 6306 return -ENOMEM; 6307 6308 *per_cpu_ptr(sdd->sgp, j) = sgp; 6309 } 6310 } 6311 6312 return 0; 6313} 6314 6315static void __sdt_free(const struct cpumask *cpu_map) 6316{ 6317 struct sched_domain_topology_level *tl; 6318 int j; 6319 6320 for_each_sd_topology(tl) { 6321 struct sd_data *sdd = &tl->data; 6322 6323 for_each_cpu(j, cpu_map) { 6324 struct sched_domain *sd; 6325 6326 if (sdd->sd) { 6327 sd = *per_cpu_ptr(sdd->sd, j); 6328 if (sd && (sd->flags & SD_OVERLAP)) 6329 free_sched_groups(sd->groups, 0); 6330 kfree(*per_cpu_ptr(sdd->sd, j)); 6331 } 6332 6333 if (sdd->sg) 6334 kfree(*per_cpu_ptr(sdd->sg, j)); 6335 if (sdd->sgp) 6336 kfree(*per_cpu_ptr(sdd->sgp, j)); 6337 } 6338 free_percpu(sdd->sd); 6339 sdd->sd = NULL; 6340 free_percpu(sdd->sg); 6341 sdd->sg = NULL; 6342 free_percpu(sdd->sgp); 6343 sdd->sgp = NULL; 6344 } 6345} 6346 6347struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6348 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6349 struct sched_domain *child, int cpu) 6350{ 6351 struct sched_domain *sd = tl->init(tl, cpu); 6352 if (!sd) 6353 return child; 6354 6355 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6356 if (child) { 6357 sd->level = child->level + 1; 6358 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6359 child->parent = sd; 6360 sd->child = child; 6361 } 6362 set_domain_attribute(sd, attr); 6363 6364 return sd; 6365} 6366 6367/* 6368 * Build sched domains for a given set of cpus and attach the sched domains 6369 * to the individual cpus 6370 */ 6371static int build_sched_domains(const struct cpumask *cpu_map, 6372 struct sched_domain_attr *attr) 6373{ 6374 enum s_alloc alloc_state; 6375 struct sched_domain *sd; 6376 struct s_data d; 6377 int i, ret = -ENOMEM; 6378 6379 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6380 if (alloc_state != sa_rootdomain) 6381 goto error; 6382 6383 /* Set up domains for cpus specified by the cpu_map. */ 6384 for_each_cpu(i, cpu_map) { 6385 struct sched_domain_topology_level *tl; 6386 6387 sd = NULL; 6388 for_each_sd_topology(tl) { 6389 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6390 if (tl == sched_domain_topology) 6391 *per_cpu_ptr(d.sd, i) = sd; 6392 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6393 sd->flags |= SD_OVERLAP; 6394 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6395 break; 6396 } 6397 } 6398 6399 /* Build the groups for the domains */ 6400 for_each_cpu(i, cpu_map) { 6401 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6402 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6403 if (sd->flags & SD_OVERLAP) { 6404 if (build_overlap_sched_groups(sd, i)) 6405 goto error; 6406 } else { 6407 if (build_sched_groups(sd, i)) 6408 goto error; 6409 } 6410 } 6411 } 6412 6413 /* Calculate CPU power for physical packages and nodes */ 6414 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6415 if (!cpumask_test_cpu(i, cpu_map)) 6416 continue; 6417 6418 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6419 claim_allocations(i, sd); 6420 init_sched_groups_power(i, sd); 6421 } 6422 } 6423 6424 /* Attach the domains */ 6425 rcu_read_lock(); 6426 for_each_cpu(i, cpu_map) { 6427 sd = *per_cpu_ptr(d.sd, i); 6428 cpu_attach_domain(sd, d.rd, i); 6429 } 6430 rcu_read_unlock(); 6431 6432 ret = 0; 6433error: 6434 __free_domain_allocs(&d, alloc_state, cpu_map); 6435 return ret; 6436} 6437 6438static cpumask_var_t *doms_cur; /* current sched domains */ 6439static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6440static struct sched_domain_attr *dattr_cur; 6441 /* attribues of custom domains in 'doms_cur' */ 6442 6443/* 6444 * Special case: If a kmalloc of a doms_cur partition (array of 6445 * cpumask) fails, then fallback to a single sched domain, 6446 * as determined by the single cpumask fallback_doms. 6447 */ 6448static cpumask_var_t fallback_doms; 6449 6450/* 6451 * arch_update_cpu_topology lets virtualized architectures update the 6452 * cpu core maps. It is supposed to return 1 if the topology changed 6453 * or 0 if it stayed the same. 6454 */ 6455int __attribute__((weak)) arch_update_cpu_topology(void) 6456{ 6457 return 0; 6458} 6459 6460cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6461{ 6462 int i; 6463 cpumask_var_t *doms; 6464 6465 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6466 if (!doms) 6467 return NULL; 6468 for (i = 0; i < ndoms; i++) { 6469 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6470 free_sched_domains(doms, i); 6471 return NULL; 6472 } 6473 } 6474 return doms; 6475} 6476 6477void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6478{ 6479 unsigned int i; 6480 for (i = 0; i < ndoms; i++) 6481 free_cpumask_var(doms[i]); 6482 kfree(doms); 6483} 6484 6485/* 6486 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6487 * For now this just excludes isolated cpus, but could be used to 6488 * exclude other special cases in the future. 6489 */ 6490static int init_sched_domains(const struct cpumask *cpu_map) 6491{ 6492 int err; 6493 6494 arch_update_cpu_topology(); 6495 ndoms_cur = 1; 6496 doms_cur = alloc_sched_domains(ndoms_cur); 6497 if (!doms_cur) 6498 doms_cur = &fallback_doms; 6499 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6500 err = build_sched_domains(doms_cur[0], NULL); 6501 register_sched_domain_sysctl(); 6502 6503 return err; 6504} 6505 6506/* 6507 * Detach sched domains from a group of cpus specified in cpu_map 6508 * These cpus will now be attached to the NULL domain 6509 */ 6510static void detach_destroy_domains(const struct cpumask *cpu_map) 6511{ 6512 int i; 6513 6514 rcu_read_lock(); 6515 for_each_cpu(i, cpu_map) 6516 cpu_attach_domain(NULL, &def_root_domain, i); 6517 rcu_read_unlock(); 6518} 6519 6520/* handle null as "default" */ 6521static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6522 struct sched_domain_attr *new, int idx_new) 6523{ 6524 struct sched_domain_attr tmp; 6525 6526 /* fast path */ 6527 if (!new && !cur) 6528 return 1; 6529 6530 tmp = SD_ATTR_INIT; 6531 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6532 new ? (new + idx_new) : &tmp, 6533 sizeof(struct sched_domain_attr)); 6534} 6535 6536/* 6537 * Partition sched domains as specified by the 'ndoms_new' 6538 * cpumasks in the array doms_new[] of cpumasks. This compares 6539 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6540 * It destroys each deleted domain and builds each new domain. 6541 * 6542 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6543 * The masks don't intersect (don't overlap.) We should setup one 6544 * sched domain for each mask. CPUs not in any of the cpumasks will 6545 * not be load balanced. If the same cpumask appears both in the 6546 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6547 * it as it is. 6548 * 6549 * The passed in 'doms_new' should be allocated using 6550 * alloc_sched_domains. This routine takes ownership of it and will 6551 * free_sched_domains it when done with it. If the caller failed the 6552 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6553 * and partition_sched_domains() will fallback to the single partition 6554 * 'fallback_doms', it also forces the domains to be rebuilt. 6555 * 6556 * If doms_new == NULL it will be replaced with cpu_online_mask. 6557 * ndoms_new == 0 is a special case for destroying existing domains, 6558 * and it will not create the default domain. 6559 * 6560 * Call with hotplug lock held 6561 */ 6562void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6563 struct sched_domain_attr *dattr_new) 6564{ 6565 int i, j, n; 6566 int new_topology; 6567 6568 mutex_lock(&sched_domains_mutex); 6569 6570 /* always unregister in case we don't destroy any domains */ 6571 unregister_sched_domain_sysctl(); 6572 6573 /* Let architecture update cpu core mappings. */ 6574 new_topology = arch_update_cpu_topology(); 6575 6576 n = doms_new ? ndoms_new : 0; 6577 6578 /* Destroy deleted domains */ 6579 for (i = 0; i < ndoms_cur; i++) { 6580 for (j = 0; j < n && !new_topology; j++) { 6581 if (cpumask_equal(doms_cur[i], doms_new[j]) 6582 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6583 goto match1; 6584 } 6585 /* no match - a current sched domain not in new doms_new[] */ 6586 detach_destroy_domains(doms_cur[i]); 6587match1: 6588 ; 6589 } 6590 6591 n = ndoms_cur; 6592 if (doms_new == NULL) { 6593 n = 0; 6594 doms_new = &fallback_doms; 6595 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6596 WARN_ON_ONCE(dattr_new); 6597 } 6598 6599 /* Build new domains */ 6600 for (i = 0; i < ndoms_new; i++) { 6601 for (j = 0; j < n && !new_topology; j++) { 6602 if (cpumask_equal(doms_new[i], doms_cur[j]) 6603 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6604 goto match2; 6605 } 6606 /* no match - add a new doms_new */ 6607 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6608match2: 6609 ; 6610 } 6611 6612 /* Remember the new sched domains */ 6613 if (doms_cur != &fallback_doms) 6614 free_sched_domains(doms_cur, ndoms_cur); 6615 kfree(dattr_cur); /* kfree(NULL) is safe */ 6616 doms_cur = doms_new; 6617 dattr_cur = dattr_new; 6618 ndoms_cur = ndoms_new; 6619 6620 register_sched_domain_sysctl(); 6621 6622 mutex_unlock(&sched_domains_mutex); 6623} 6624 6625static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6626 6627/* 6628 * Update cpusets according to cpu_active mask. If cpusets are 6629 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6630 * around partition_sched_domains(). 6631 * 6632 * If we come here as part of a suspend/resume, don't touch cpusets because we 6633 * want to restore it back to its original state upon resume anyway. 6634 */ 6635static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6636 void *hcpu) 6637{ 6638 switch (action) { 6639 case CPU_ONLINE_FROZEN: 6640 case CPU_DOWN_FAILED_FROZEN: 6641 6642 /* 6643 * num_cpus_frozen tracks how many CPUs are involved in suspend 6644 * resume sequence. As long as this is not the last online 6645 * operation in the resume sequence, just build a single sched 6646 * domain, ignoring cpusets. 6647 */ 6648 num_cpus_frozen--; 6649 if (likely(num_cpus_frozen)) { 6650 partition_sched_domains(1, NULL, NULL); 6651 break; 6652 } 6653 6654 /* 6655 * This is the last CPU online operation. So fall through and 6656 * restore the original sched domains by considering the 6657 * cpuset configurations. 6658 */ 6659 6660 case CPU_ONLINE: 6661 case CPU_DOWN_FAILED: 6662 cpuset_update_active_cpus(true); 6663 break; 6664 default: 6665 return NOTIFY_DONE; 6666 } 6667 return NOTIFY_OK; 6668} 6669 6670static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6671 void *hcpu) 6672{ 6673 switch (action) { 6674 case CPU_DOWN_PREPARE: 6675 cpuset_update_active_cpus(false); 6676 break; 6677 case CPU_DOWN_PREPARE_FROZEN: 6678 num_cpus_frozen++; 6679 partition_sched_domains(1, NULL, NULL); 6680 break; 6681 default: 6682 return NOTIFY_DONE; 6683 } 6684 return NOTIFY_OK; 6685} 6686 6687void __init sched_init_smp(void) 6688{ 6689 cpumask_var_t non_isolated_cpus; 6690 6691 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6692 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6693 6694 sched_init_numa(); 6695 6696 /* 6697 * There's no userspace yet to cause hotplug operations; hence all the 6698 * cpu masks are stable and all blatant races in the below code cannot 6699 * happen. 6700 */ 6701 mutex_lock(&sched_domains_mutex); 6702 init_sched_domains(cpu_active_mask); 6703 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6704 if (cpumask_empty(non_isolated_cpus)) 6705 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6706 mutex_unlock(&sched_domains_mutex); 6707 6708 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6709 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6710 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6711 6712 init_hrtick(); 6713 6714 /* Move init over to a non-isolated CPU */ 6715 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6716 BUG(); 6717 sched_init_granularity(); 6718 free_cpumask_var(non_isolated_cpus); 6719 6720 init_sched_rt_class(); 6721 init_sched_dl_class(); 6722} 6723#else 6724void __init sched_init_smp(void) 6725{ 6726 sched_init_granularity(); 6727} 6728#endif /* CONFIG_SMP */ 6729 6730const_debug unsigned int sysctl_timer_migration = 1; 6731 6732int in_sched_functions(unsigned long addr) 6733{ 6734 return in_lock_functions(addr) || 6735 (addr >= (unsigned long)__sched_text_start 6736 && addr < (unsigned long)__sched_text_end); 6737} 6738 6739#ifdef CONFIG_CGROUP_SCHED 6740/* 6741 * Default task group. 6742 * Every task in system belongs to this group at bootup. 6743 */ 6744struct task_group root_task_group; 6745LIST_HEAD(task_groups); 6746#endif 6747 6748DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6749 6750void __init sched_init(void) 6751{ 6752 int i, j; 6753 unsigned long alloc_size = 0, ptr; 6754 6755#ifdef CONFIG_FAIR_GROUP_SCHED 6756 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6757#endif 6758#ifdef CONFIG_RT_GROUP_SCHED 6759 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6760#endif 6761#ifdef CONFIG_CPUMASK_OFFSTACK 6762 alloc_size += num_possible_cpus() * cpumask_size(); 6763#endif 6764 if (alloc_size) { 6765 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6766 6767#ifdef CONFIG_FAIR_GROUP_SCHED 6768 root_task_group.se = (struct sched_entity **)ptr; 6769 ptr += nr_cpu_ids * sizeof(void **); 6770 6771 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6772 ptr += nr_cpu_ids * sizeof(void **); 6773 6774#endif /* CONFIG_FAIR_GROUP_SCHED */ 6775#ifdef CONFIG_RT_GROUP_SCHED 6776 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6777 ptr += nr_cpu_ids * sizeof(void **); 6778 6779 root_task_group.rt_rq = (struct rt_rq **)ptr; 6780 ptr += nr_cpu_ids * sizeof(void **); 6781 6782#endif /* CONFIG_RT_GROUP_SCHED */ 6783#ifdef CONFIG_CPUMASK_OFFSTACK 6784 for_each_possible_cpu(i) { 6785 per_cpu(load_balance_mask, i) = (void *)ptr; 6786 ptr += cpumask_size(); 6787 } 6788#endif /* CONFIG_CPUMASK_OFFSTACK */ 6789 } 6790 6791 init_rt_bandwidth(&def_rt_bandwidth, 6792 global_rt_period(), global_rt_runtime()); 6793 init_dl_bandwidth(&def_dl_bandwidth, 6794 global_rt_period(), global_rt_runtime()); 6795 6796#ifdef CONFIG_SMP 6797 init_defrootdomain(); 6798#endif 6799 6800#ifdef CONFIG_RT_GROUP_SCHED 6801 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6802 global_rt_period(), global_rt_runtime()); 6803#endif /* CONFIG_RT_GROUP_SCHED */ 6804 6805#ifdef CONFIG_CGROUP_SCHED 6806 list_add(&root_task_group.list, &task_groups); 6807 INIT_LIST_HEAD(&root_task_group.children); 6808 INIT_LIST_HEAD(&root_task_group.siblings); 6809 autogroup_init(&init_task); 6810 6811#endif /* CONFIG_CGROUP_SCHED */ 6812 6813 for_each_possible_cpu(i) { 6814 struct rq *rq; 6815 6816 rq = cpu_rq(i); 6817 raw_spin_lock_init(&rq->lock); 6818 rq->nr_running = 0; 6819 rq->calc_load_active = 0; 6820 rq->calc_load_update = jiffies + LOAD_FREQ; 6821 init_cfs_rq(&rq->cfs); 6822 init_rt_rq(&rq->rt, rq); 6823 init_dl_rq(&rq->dl, rq); 6824#ifdef CONFIG_FAIR_GROUP_SCHED 6825 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6826 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6827 /* 6828 * How much cpu bandwidth does root_task_group get? 6829 * 6830 * In case of task-groups formed thr' the cgroup filesystem, it 6831 * gets 100% of the cpu resources in the system. This overall 6832 * system cpu resource is divided among the tasks of 6833 * root_task_group and its child task-groups in a fair manner, 6834 * based on each entity's (task or task-group's) weight 6835 * (se->load.weight). 6836 * 6837 * In other words, if root_task_group has 10 tasks of weight 6838 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6839 * then A0's share of the cpu resource is: 6840 * 6841 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6842 * 6843 * We achieve this by letting root_task_group's tasks sit 6844 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6845 */ 6846 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6847 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6848#endif /* CONFIG_FAIR_GROUP_SCHED */ 6849 6850 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6851#ifdef CONFIG_RT_GROUP_SCHED 6852 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6853 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6854#endif 6855 6856 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6857 rq->cpu_load[j] = 0; 6858 6859 rq->last_load_update_tick = jiffies; 6860 6861#ifdef CONFIG_SMP 6862 rq->sd = NULL; 6863 rq->rd = NULL; 6864 rq->cpu_power = SCHED_POWER_SCALE; 6865 rq->post_schedule = 0; 6866 rq->active_balance = 0; 6867 rq->next_balance = jiffies; 6868 rq->push_cpu = 0; 6869 rq->cpu = i; 6870 rq->online = 0; 6871 rq->idle_stamp = 0; 6872 rq->avg_idle = 2*sysctl_sched_migration_cost; 6873 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6874 6875 INIT_LIST_HEAD(&rq->cfs_tasks); 6876 6877 rq_attach_root(rq, &def_root_domain); 6878#ifdef CONFIG_NO_HZ_COMMON 6879 rq->nohz_flags = 0; 6880#endif 6881#ifdef CONFIG_NO_HZ_FULL 6882 rq->last_sched_tick = 0; 6883#endif 6884#endif 6885 init_rq_hrtick(rq); 6886 atomic_set(&rq->nr_iowait, 0); 6887 } 6888 6889 set_load_weight(&init_task); 6890 6891#ifdef CONFIG_PREEMPT_NOTIFIERS 6892 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6893#endif 6894 6895 /* 6896 * The boot idle thread does lazy MMU switching as well: 6897 */ 6898 atomic_inc(&init_mm.mm_count); 6899 enter_lazy_tlb(&init_mm, current); 6900 6901 /* 6902 * Make us the idle thread. Technically, schedule() should not be 6903 * called from this thread, however somewhere below it might be, 6904 * but because we are the idle thread, we just pick up running again 6905 * when this runqueue becomes "idle". 6906 */ 6907 init_idle(current, smp_processor_id()); 6908 6909 calc_load_update = jiffies + LOAD_FREQ; 6910 6911 /* 6912 * During early bootup we pretend to be a normal task: 6913 */ 6914 current->sched_class = &fair_sched_class; 6915 6916#ifdef CONFIG_SMP 6917 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6918 /* May be allocated at isolcpus cmdline parse time */ 6919 if (cpu_isolated_map == NULL) 6920 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6921 idle_thread_set_boot_cpu(); 6922#endif 6923 init_sched_fair_class(); 6924 6925 scheduler_running = 1; 6926} 6927 6928#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6929static inline int preempt_count_equals(int preempt_offset) 6930{ 6931 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6932 6933 return (nested == preempt_offset); 6934} 6935 6936void __might_sleep(const char *file, int line, int preempt_offset) 6937{ 6938 static unsigned long prev_jiffy; /* ratelimiting */ 6939 6940 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6941 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6942 system_state != SYSTEM_RUNNING || oops_in_progress) 6943 return; 6944 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6945 return; 6946 prev_jiffy = jiffies; 6947 6948 printk(KERN_ERR 6949 "BUG: sleeping function called from invalid context at %s:%d\n", 6950 file, line); 6951 printk(KERN_ERR 6952 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6953 in_atomic(), irqs_disabled(), 6954 current->pid, current->comm); 6955 6956 debug_show_held_locks(current); 6957 if (irqs_disabled()) 6958 print_irqtrace_events(current); 6959 dump_stack(); 6960} 6961EXPORT_SYMBOL(__might_sleep); 6962#endif 6963 6964#ifdef CONFIG_MAGIC_SYSRQ 6965static void normalize_task(struct rq *rq, struct task_struct *p) 6966{ 6967 const struct sched_class *prev_class = p->sched_class; 6968 struct sched_attr attr = { 6969 .sched_policy = SCHED_NORMAL, 6970 }; 6971 int old_prio = p->prio; 6972 int on_rq; 6973 6974 on_rq = p->on_rq; 6975 if (on_rq) 6976 dequeue_task(rq, p, 0); 6977 __setscheduler(rq, p, &attr); 6978 if (on_rq) { 6979 enqueue_task(rq, p, 0); 6980 resched_task(rq->curr); 6981 } 6982 6983 check_class_changed(rq, p, prev_class, old_prio); 6984} 6985 6986void normalize_rt_tasks(void) 6987{ 6988 struct task_struct *g, *p; 6989 unsigned long flags; 6990 struct rq *rq; 6991 6992 read_lock_irqsave(&tasklist_lock, flags); 6993 do_each_thread(g, p) { 6994 /* 6995 * Only normalize user tasks: 6996 */ 6997 if (!p->mm) 6998 continue; 6999 7000 p->se.exec_start = 0; 7001#ifdef CONFIG_SCHEDSTATS 7002 p->se.statistics.wait_start = 0; 7003 p->se.statistics.sleep_start = 0; 7004 p->se.statistics.block_start = 0; 7005#endif 7006 7007 if (!dl_task(p) && !rt_task(p)) { 7008 /* 7009 * Renice negative nice level userspace 7010 * tasks back to 0: 7011 */ 7012 if (TASK_NICE(p) < 0 && p->mm) 7013 set_user_nice(p, 0); 7014 continue; 7015 } 7016 7017 raw_spin_lock(&p->pi_lock); 7018 rq = __task_rq_lock(p); 7019 7020 normalize_task(rq, p); 7021 7022 __task_rq_unlock(rq); 7023 raw_spin_unlock(&p->pi_lock); 7024 } while_each_thread(g, p); 7025 7026 read_unlock_irqrestore(&tasklist_lock, flags); 7027} 7028 7029#endif /* CONFIG_MAGIC_SYSRQ */ 7030 7031#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7032/* 7033 * These functions are only useful for the IA64 MCA handling, or kdb. 7034 * 7035 * They can only be called when the whole system has been 7036 * stopped - every CPU needs to be quiescent, and no scheduling 7037 * activity can take place. Using them for anything else would 7038 * be a serious bug, and as a result, they aren't even visible 7039 * under any other configuration. 7040 */ 7041 7042/** 7043 * curr_task - return the current task for a given cpu. 7044 * @cpu: the processor in question. 7045 * 7046 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7047 * 7048 * Return: The current task for @cpu. 7049 */ 7050struct task_struct *curr_task(int cpu) 7051{ 7052 return cpu_curr(cpu); 7053} 7054 7055#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7056 7057#ifdef CONFIG_IA64 7058/** 7059 * set_curr_task - set the current task for a given cpu. 7060 * @cpu: the processor in question. 7061 * @p: the task pointer to set. 7062 * 7063 * Description: This function must only be used when non-maskable interrupts 7064 * are serviced on a separate stack. It allows the architecture to switch the 7065 * notion of the current task on a cpu in a non-blocking manner. This function 7066 * must be called with all CPU's synchronized, and interrupts disabled, the 7067 * and caller must save the original value of the current task (see 7068 * curr_task() above) and restore that value before reenabling interrupts and 7069 * re-starting the system. 7070 * 7071 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7072 */ 7073void set_curr_task(int cpu, struct task_struct *p) 7074{ 7075 cpu_curr(cpu) = p; 7076} 7077 7078#endif 7079 7080#ifdef CONFIG_CGROUP_SCHED 7081/* task_group_lock serializes the addition/removal of task groups */ 7082static DEFINE_SPINLOCK(task_group_lock); 7083 7084static void free_sched_group(struct task_group *tg) 7085{ 7086 free_fair_sched_group(tg); 7087 free_rt_sched_group(tg); 7088 autogroup_free(tg); 7089 kfree(tg); 7090} 7091 7092/* allocate runqueue etc for a new task group */ 7093struct task_group *sched_create_group(struct task_group *parent) 7094{ 7095 struct task_group *tg; 7096 7097 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7098 if (!tg) 7099 return ERR_PTR(-ENOMEM); 7100 7101 if (!alloc_fair_sched_group(tg, parent)) 7102 goto err; 7103 7104 if (!alloc_rt_sched_group(tg, parent)) 7105 goto err; 7106 7107 return tg; 7108 7109err: 7110 free_sched_group(tg); 7111 return ERR_PTR(-ENOMEM); 7112} 7113 7114void sched_online_group(struct task_group *tg, struct task_group *parent) 7115{ 7116 unsigned long flags; 7117 7118 spin_lock_irqsave(&task_group_lock, flags); 7119 list_add_rcu(&tg->list, &task_groups); 7120 7121 WARN_ON(!parent); /* root should already exist */ 7122 7123 tg->parent = parent; 7124 INIT_LIST_HEAD(&tg->children); 7125 list_add_rcu(&tg->siblings, &parent->children); 7126 spin_unlock_irqrestore(&task_group_lock, flags); 7127} 7128 7129/* rcu callback to free various structures associated with a task group */ 7130static void free_sched_group_rcu(struct rcu_head *rhp) 7131{ 7132 /* now it should be safe to free those cfs_rqs */ 7133 free_sched_group(container_of(rhp, struct task_group, rcu)); 7134} 7135 7136/* Destroy runqueue etc associated with a task group */ 7137void sched_destroy_group(struct task_group *tg) 7138{ 7139 /* wait for possible concurrent references to cfs_rqs complete */ 7140 call_rcu(&tg->rcu, free_sched_group_rcu); 7141} 7142 7143void sched_offline_group(struct task_group *tg) 7144{ 7145 unsigned long flags; 7146 int i; 7147 7148 /* end participation in shares distribution */ 7149 for_each_possible_cpu(i) 7150 unregister_fair_sched_group(tg, i); 7151 7152 spin_lock_irqsave(&task_group_lock, flags); 7153 list_del_rcu(&tg->list); 7154 list_del_rcu(&tg->siblings); 7155 spin_unlock_irqrestore(&task_group_lock, flags); 7156} 7157 7158/* change task's runqueue when it moves between groups. 7159 * The caller of this function should have put the task in its new group 7160 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7161 * reflect its new group. 7162 */ 7163void sched_move_task(struct task_struct *tsk) 7164{ 7165 struct task_group *tg; 7166 int on_rq, running; 7167 unsigned long flags; 7168 struct rq *rq; 7169 7170 rq = task_rq_lock(tsk, &flags); 7171 7172 running = task_current(rq, tsk); 7173 on_rq = tsk->on_rq; 7174 7175 if (on_rq) 7176 dequeue_task(rq, tsk, 0); 7177 if (unlikely(running)) 7178 tsk->sched_class->put_prev_task(rq, tsk); 7179 7180 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 7181 lockdep_is_held(&tsk->sighand->siglock)), 7182 struct task_group, css); 7183 tg = autogroup_task_group(tsk, tg); 7184 tsk->sched_task_group = tg; 7185 7186#ifdef CONFIG_FAIR_GROUP_SCHED 7187 if (tsk->sched_class->task_move_group) 7188 tsk->sched_class->task_move_group(tsk, on_rq); 7189 else 7190#endif 7191 set_task_rq(tsk, task_cpu(tsk)); 7192 7193 if (unlikely(running)) 7194 tsk->sched_class->set_curr_task(rq); 7195 if (on_rq) 7196 enqueue_task(rq, tsk, 0); 7197 7198 task_rq_unlock(rq, tsk, &flags); 7199} 7200#endif /* CONFIG_CGROUP_SCHED */ 7201 7202#ifdef CONFIG_RT_GROUP_SCHED 7203/* 7204 * Ensure that the real time constraints are schedulable. 7205 */ 7206static DEFINE_MUTEX(rt_constraints_mutex); 7207 7208/* Must be called with tasklist_lock held */ 7209static inline int tg_has_rt_tasks(struct task_group *tg) 7210{ 7211 struct task_struct *g, *p; 7212 7213 do_each_thread(g, p) { 7214 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7215 return 1; 7216 } while_each_thread(g, p); 7217 7218 return 0; 7219} 7220 7221struct rt_schedulable_data { 7222 struct task_group *tg; 7223 u64 rt_period; 7224 u64 rt_runtime; 7225}; 7226 7227static int tg_rt_schedulable(struct task_group *tg, void *data) 7228{ 7229 struct rt_schedulable_data *d = data; 7230 struct task_group *child; 7231 unsigned long total, sum = 0; 7232 u64 period, runtime; 7233 7234 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7235 runtime = tg->rt_bandwidth.rt_runtime; 7236 7237 if (tg == d->tg) { 7238 period = d->rt_period; 7239 runtime = d->rt_runtime; 7240 } 7241 7242 /* 7243 * Cannot have more runtime than the period. 7244 */ 7245 if (runtime > period && runtime != RUNTIME_INF) 7246 return -EINVAL; 7247 7248 /* 7249 * Ensure we don't starve existing RT tasks. 7250 */ 7251 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7252 return -EBUSY; 7253 7254 total = to_ratio(period, runtime); 7255 7256 /* 7257 * Nobody can have more than the global setting allows. 7258 */ 7259 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7260 return -EINVAL; 7261 7262 /* 7263 * The sum of our children's runtime should not exceed our own. 7264 */ 7265 list_for_each_entry_rcu(child, &tg->children, siblings) { 7266 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7267 runtime = child->rt_bandwidth.rt_runtime; 7268 7269 if (child == d->tg) { 7270 period = d->rt_period; 7271 runtime = d->rt_runtime; 7272 } 7273 7274 sum += to_ratio(period, runtime); 7275 } 7276 7277 if (sum > total) 7278 return -EINVAL; 7279 7280 return 0; 7281} 7282 7283static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7284{ 7285 int ret; 7286 7287 struct rt_schedulable_data data = { 7288 .tg = tg, 7289 .rt_period = period, 7290 .rt_runtime = runtime, 7291 }; 7292 7293 rcu_read_lock(); 7294 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7295 rcu_read_unlock(); 7296 7297 return ret; 7298} 7299 7300static int tg_set_rt_bandwidth(struct task_group *tg, 7301 u64 rt_period, u64 rt_runtime) 7302{ 7303 int i, err = 0; 7304 7305 mutex_lock(&rt_constraints_mutex); 7306 read_lock(&tasklist_lock); 7307 err = __rt_schedulable(tg, rt_period, rt_runtime); 7308 if (err) 7309 goto unlock; 7310 7311 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7312 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7313 tg->rt_bandwidth.rt_runtime = rt_runtime; 7314 7315 for_each_possible_cpu(i) { 7316 struct rt_rq *rt_rq = tg->rt_rq[i]; 7317 7318 raw_spin_lock(&rt_rq->rt_runtime_lock); 7319 rt_rq->rt_runtime = rt_runtime; 7320 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7321 } 7322 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7323unlock: 7324 read_unlock(&tasklist_lock); 7325 mutex_unlock(&rt_constraints_mutex); 7326 7327 return err; 7328} 7329 7330static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7331{ 7332 u64 rt_runtime, rt_period; 7333 7334 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7335 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7336 if (rt_runtime_us < 0) 7337 rt_runtime = RUNTIME_INF; 7338 7339 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7340} 7341 7342static long sched_group_rt_runtime(struct task_group *tg) 7343{ 7344 u64 rt_runtime_us; 7345 7346 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7347 return -1; 7348 7349 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7350 do_div(rt_runtime_us, NSEC_PER_USEC); 7351 return rt_runtime_us; 7352} 7353 7354static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7355{ 7356 u64 rt_runtime, rt_period; 7357 7358 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7359 rt_runtime = tg->rt_bandwidth.rt_runtime; 7360 7361 if (rt_period == 0) 7362 return -EINVAL; 7363 7364 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7365} 7366 7367static long sched_group_rt_period(struct task_group *tg) 7368{ 7369 u64 rt_period_us; 7370 7371 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7372 do_div(rt_period_us, NSEC_PER_USEC); 7373 return rt_period_us; 7374} 7375#endif /* CONFIG_RT_GROUP_SCHED */ 7376 7377#ifdef CONFIG_RT_GROUP_SCHED 7378static int sched_rt_global_constraints(void) 7379{ 7380 int ret = 0; 7381 7382 mutex_lock(&rt_constraints_mutex); 7383 read_lock(&tasklist_lock); 7384 ret = __rt_schedulable(NULL, 0, 0); 7385 read_unlock(&tasklist_lock); 7386 mutex_unlock(&rt_constraints_mutex); 7387 7388 return ret; 7389} 7390 7391static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7392{ 7393 /* Don't accept realtime tasks when there is no way for them to run */ 7394 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7395 return 0; 7396 7397 return 1; 7398} 7399 7400#else /* !CONFIG_RT_GROUP_SCHED */ 7401static int sched_rt_global_constraints(void) 7402{ 7403 unsigned long flags; 7404 int i, ret = 0; 7405 7406 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7407 for_each_possible_cpu(i) { 7408 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7409 7410 raw_spin_lock(&rt_rq->rt_runtime_lock); 7411 rt_rq->rt_runtime = global_rt_runtime(); 7412 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7413 } 7414 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7415 7416 return ret; 7417} 7418#endif /* CONFIG_RT_GROUP_SCHED */ 7419 7420static int sched_dl_global_constraints(void) 7421{ 7422 u64 runtime = global_rt_runtime(); 7423 u64 period = global_rt_period(); 7424 u64 new_bw = to_ratio(period, runtime); 7425 int cpu, ret = 0; 7426 unsigned long flags; 7427 7428 /* 7429 * Here we want to check the bandwidth not being set to some 7430 * value smaller than the currently allocated bandwidth in 7431 * any of the root_domains. 7432 * 7433 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7434 * cycling on root_domains... Discussion on different/better 7435 * solutions is welcome! 7436 */ 7437 for_each_possible_cpu(cpu) { 7438 struct dl_bw *dl_b = dl_bw_of(cpu); 7439 7440 raw_spin_lock_irqsave(&dl_b->lock, flags); 7441 if (new_bw < dl_b->total_bw) 7442 ret = -EBUSY; 7443 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7444 7445 if (ret) 7446 break; 7447 } 7448 7449 return ret; 7450} 7451 7452static void sched_dl_do_global(void) 7453{ 7454 u64 new_bw = -1; 7455 int cpu; 7456 unsigned long flags; 7457 7458 def_dl_bandwidth.dl_period = global_rt_period(); 7459 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7460 7461 if (global_rt_runtime() != RUNTIME_INF) 7462 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7463 7464 /* 7465 * FIXME: As above... 7466 */ 7467 for_each_possible_cpu(cpu) { 7468 struct dl_bw *dl_b = dl_bw_of(cpu); 7469 7470 raw_spin_lock_irqsave(&dl_b->lock, flags); 7471 dl_b->bw = new_bw; 7472 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7473 } 7474} 7475 7476static int sched_rt_global_validate(void) 7477{ 7478 if (sysctl_sched_rt_period <= 0) 7479 return -EINVAL; 7480 7481 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7482 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7483 return -EINVAL; 7484 7485 return 0; 7486} 7487 7488static void sched_rt_do_global(void) 7489{ 7490 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7491 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7492} 7493 7494int sched_rt_handler(struct ctl_table *table, int write, 7495 void __user *buffer, size_t *lenp, 7496 loff_t *ppos) 7497{ 7498 int old_period, old_runtime; 7499 static DEFINE_MUTEX(mutex); 7500 int ret; 7501 7502 mutex_lock(&mutex); 7503 old_period = sysctl_sched_rt_period; 7504 old_runtime = sysctl_sched_rt_runtime; 7505 7506 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7507 7508 if (!ret && write) { 7509 ret = sched_rt_global_validate(); 7510 if (ret) 7511 goto undo; 7512 7513 ret = sched_rt_global_constraints(); 7514 if (ret) 7515 goto undo; 7516 7517 ret = sched_dl_global_constraints(); 7518 if (ret) 7519 goto undo; 7520 7521 sched_rt_do_global(); 7522 sched_dl_do_global(); 7523 } 7524 if (0) { 7525undo: 7526 sysctl_sched_rt_period = old_period; 7527 sysctl_sched_rt_runtime = old_runtime; 7528 } 7529 mutex_unlock(&mutex); 7530 7531 return ret; 7532} 7533 7534int sched_rr_handler(struct ctl_table *table, int write, 7535 void __user *buffer, size_t *lenp, 7536 loff_t *ppos) 7537{ 7538 int ret; 7539 static DEFINE_MUTEX(mutex); 7540 7541 mutex_lock(&mutex); 7542 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7543 /* make sure that internally we keep jiffies */ 7544 /* also, writing zero resets timeslice to default */ 7545 if (!ret && write) { 7546 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7547 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7548 } 7549 mutex_unlock(&mutex); 7550 return ret; 7551} 7552 7553#ifdef CONFIG_CGROUP_SCHED 7554 7555static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7556{ 7557 return css ? container_of(css, struct task_group, css) : NULL; 7558} 7559 7560static struct cgroup_subsys_state * 7561cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7562{ 7563 struct task_group *parent = css_tg(parent_css); 7564 struct task_group *tg; 7565 7566 if (!parent) { 7567 /* This is early initialization for the top cgroup */ 7568 return &root_task_group.css; 7569 } 7570 7571 tg = sched_create_group(parent); 7572 if (IS_ERR(tg)) 7573 return ERR_PTR(-ENOMEM); 7574 7575 return &tg->css; 7576} 7577 7578static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7579{ 7580 struct task_group *tg = css_tg(css); 7581 struct task_group *parent = css_tg(css_parent(css)); 7582 7583 if (parent) 7584 sched_online_group(tg, parent); 7585 return 0; 7586} 7587 7588static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7589{ 7590 struct task_group *tg = css_tg(css); 7591 7592 sched_destroy_group(tg); 7593} 7594 7595static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7596{ 7597 struct task_group *tg = css_tg(css); 7598 7599 sched_offline_group(tg); 7600} 7601 7602static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7603 struct cgroup_taskset *tset) 7604{ 7605 struct task_struct *task; 7606 7607 cgroup_taskset_for_each(task, css, tset) { 7608#ifdef CONFIG_RT_GROUP_SCHED 7609 if (!sched_rt_can_attach(css_tg(css), task)) 7610 return -EINVAL; 7611#else 7612 /* We don't support RT-tasks being in separate groups */ 7613 if (task->sched_class != &fair_sched_class) 7614 return -EINVAL; 7615#endif 7616 } 7617 return 0; 7618} 7619 7620static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7621 struct cgroup_taskset *tset) 7622{ 7623 struct task_struct *task; 7624 7625 cgroup_taskset_for_each(task, css, tset) 7626 sched_move_task(task); 7627} 7628 7629static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7630 struct cgroup_subsys_state *old_css, 7631 struct task_struct *task) 7632{ 7633 /* 7634 * cgroup_exit() is called in the copy_process() failure path. 7635 * Ignore this case since the task hasn't ran yet, this avoids 7636 * trying to poke a half freed task state from generic code. 7637 */ 7638 if (!(task->flags & PF_EXITING)) 7639 return; 7640 7641 sched_move_task(task); 7642} 7643 7644#ifdef CONFIG_FAIR_GROUP_SCHED 7645static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7646 struct cftype *cftype, u64 shareval) 7647{ 7648 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7649} 7650 7651static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7652 struct cftype *cft) 7653{ 7654 struct task_group *tg = css_tg(css); 7655 7656 return (u64) scale_load_down(tg->shares); 7657} 7658 7659#ifdef CONFIG_CFS_BANDWIDTH 7660static DEFINE_MUTEX(cfs_constraints_mutex); 7661 7662const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7663const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7664 7665static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7666 7667static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7668{ 7669 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7670 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7671 7672 if (tg == &root_task_group) 7673 return -EINVAL; 7674 7675 /* 7676 * Ensure we have at some amount of bandwidth every period. This is 7677 * to prevent reaching a state of large arrears when throttled via 7678 * entity_tick() resulting in prolonged exit starvation. 7679 */ 7680 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7681 return -EINVAL; 7682 7683 /* 7684 * Likewise, bound things on the otherside by preventing insane quota 7685 * periods. This also allows us to normalize in computing quota 7686 * feasibility. 7687 */ 7688 if (period > max_cfs_quota_period) 7689 return -EINVAL; 7690 7691 mutex_lock(&cfs_constraints_mutex); 7692 ret = __cfs_schedulable(tg, period, quota); 7693 if (ret) 7694 goto out_unlock; 7695 7696 runtime_enabled = quota != RUNTIME_INF; 7697 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7698 /* 7699 * If we need to toggle cfs_bandwidth_used, off->on must occur 7700 * before making related changes, and on->off must occur afterwards 7701 */ 7702 if (runtime_enabled && !runtime_was_enabled) 7703 cfs_bandwidth_usage_inc(); 7704 raw_spin_lock_irq(&cfs_b->lock); 7705 cfs_b->period = ns_to_ktime(period); 7706 cfs_b->quota = quota; 7707 7708 __refill_cfs_bandwidth_runtime(cfs_b); 7709 /* restart the period timer (if active) to handle new period expiry */ 7710 if (runtime_enabled && cfs_b->timer_active) { 7711 /* force a reprogram */ 7712 cfs_b->timer_active = 0; 7713 __start_cfs_bandwidth(cfs_b); 7714 } 7715 raw_spin_unlock_irq(&cfs_b->lock); 7716 7717 for_each_possible_cpu(i) { 7718 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7719 struct rq *rq = cfs_rq->rq; 7720 7721 raw_spin_lock_irq(&rq->lock); 7722 cfs_rq->runtime_enabled = runtime_enabled; 7723 cfs_rq->runtime_remaining = 0; 7724 7725 if (cfs_rq->throttled) 7726 unthrottle_cfs_rq(cfs_rq); 7727 raw_spin_unlock_irq(&rq->lock); 7728 } 7729 if (runtime_was_enabled && !runtime_enabled) 7730 cfs_bandwidth_usage_dec(); 7731out_unlock: 7732 mutex_unlock(&cfs_constraints_mutex); 7733 7734 return ret; 7735} 7736 7737int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7738{ 7739 u64 quota, period; 7740 7741 period = ktime_to_ns(tg->cfs_bandwidth.period); 7742 if (cfs_quota_us < 0) 7743 quota = RUNTIME_INF; 7744 else 7745 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7746 7747 return tg_set_cfs_bandwidth(tg, period, quota); 7748} 7749 7750long tg_get_cfs_quota(struct task_group *tg) 7751{ 7752 u64 quota_us; 7753 7754 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7755 return -1; 7756 7757 quota_us = tg->cfs_bandwidth.quota; 7758 do_div(quota_us, NSEC_PER_USEC); 7759 7760 return quota_us; 7761} 7762 7763int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7764{ 7765 u64 quota, period; 7766 7767 period = (u64)cfs_period_us * NSEC_PER_USEC; 7768 quota = tg->cfs_bandwidth.quota; 7769 7770 return tg_set_cfs_bandwidth(tg, period, quota); 7771} 7772 7773long tg_get_cfs_period(struct task_group *tg) 7774{ 7775 u64 cfs_period_us; 7776 7777 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7778 do_div(cfs_period_us, NSEC_PER_USEC); 7779 7780 return cfs_period_us; 7781} 7782 7783static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7784 struct cftype *cft) 7785{ 7786 return tg_get_cfs_quota(css_tg(css)); 7787} 7788 7789static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7790 struct cftype *cftype, s64 cfs_quota_us) 7791{ 7792 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7793} 7794 7795static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7796 struct cftype *cft) 7797{ 7798 return tg_get_cfs_period(css_tg(css)); 7799} 7800 7801static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7802 struct cftype *cftype, u64 cfs_period_us) 7803{ 7804 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7805} 7806 7807struct cfs_schedulable_data { 7808 struct task_group *tg; 7809 u64 period, quota; 7810}; 7811 7812/* 7813 * normalize group quota/period to be quota/max_period 7814 * note: units are usecs 7815 */ 7816static u64 normalize_cfs_quota(struct task_group *tg, 7817 struct cfs_schedulable_data *d) 7818{ 7819 u64 quota, period; 7820 7821 if (tg == d->tg) { 7822 period = d->period; 7823 quota = d->quota; 7824 } else { 7825 period = tg_get_cfs_period(tg); 7826 quota = tg_get_cfs_quota(tg); 7827 } 7828 7829 /* note: these should typically be equivalent */ 7830 if (quota == RUNTIME_INF || quota == -1) 7831 return RUNTIME_INF; 7832 7833 return to_ratio(period, quota); 7834} 7835 7836static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7837{ 7838 struct cfs_schedulable_data *d = data; 7839 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7840 s64 quota = 0, parent_quota = -1; 7841 7842 if (!tg->parent) { 7843 quota = RUNTIME_INF; 7844 } else { 7845 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7846 7847 quota = normalize_cfs_quota(tg, d); 7848 parent_quota = parent_b->hierarchal_quota; 7849 7850 /* 7851 * ensure max(child_quota) <= parent_quota, inherit when no 7852 * limit is set 7853 */ 7854 if (quota == RUNTIME_INF) 7855 quota = parent_quota; 7856 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7857 return -EINVAL; 7858 } 7859 cfs_b->hierarchal_quota = quota; 7860 7861 return 0; 7862} 7863 7864static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7865{ 7866 int ret; 7867 struct cfs_schedulable_data data = { 7868 .tg = tg, 7869 .period = period, 7870 .quota = quota, 7871 }; 7872 7873 if (quota != RUNTIME_INF) { 7874 do_div(data.period, NSEC_PER_USEC); 7875 do_div(data.quota, NSEC_PER_USEC); 7876 } 7877 7878 rcu_read_lock(); 7879 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7880 rcu_read_unlock(); 7881 7882 return ret; 7883} 7884 7885static int cpu_stats_show(struct seq_file *sf, void *v) 7886{ 7887 struct task_group *tg = css_tg(seq_css(sf)); 7888 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7889 7890 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7891 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7892 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7893 7894 return 0; 7895} 7896#endif /* CONFIG_CFS_BANDWIDTH */ 7897#endif /* CONFIG_FAIR_GROUP_SCHED */ 7898 7899#ifdef CONFIG_RT_GROUP_SCHED 7900static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7901 struct cftype *cft, s64 val) 7902{ 7903 return sched_group_set_rt_runtime(css_tg(css), val); 7904} 7905 7906static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7907 struct cftype *cft) 7908{ 7909 return sched_group_rt_runtime(css_tg(css)); 7910} 7911 7912static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7913 struct cftype *cftype, u64 rt_period_us) 7914{ 7915 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7916} 7917 7918static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7919 struct cftype *cft) 7920{ 7921 return sched_group_rt_period(css_tg(css)); 7922} 7923#endif /* CONFIG_RT_GROUP_SCHED */ 7924 7925static struct cftype cpu_files[] = { 7926#ifdef CONFIG_FAIR_GROUP_SCHED 7927 { 7928 .name = "shares", 7929 .read_u64 = cpu_shares_read_u64, 7930 .write_u64 = cpu_shares_write_u64, 7931 }, 7932#endif 7933#ifdef CONFIG_CFS_BANDWIDTH 7934 { 7935 .name = "cfs_quota_us", 7936 .read_s64 = cpu_cfs_quota_read_s64, 7937 .write_s64 = cpu_cfs_quota_write_s64, 7938 }, 7939 { 7940 .name = "cfs_period_us", 7941 .read_u64 = cpu_cfs_period_read_u64, 7942 .write_u64 = cpu_cfs_period_write_u64, 7943 }, 7944 { 7945 .name = "stat", 7946 .seq_show = cpu_stats_show, 7947 }, 7948#endif 7949#ifdef CONFIG_RT_GROUP_SCHED 7950 { 7951 .name = "rt_runtime_us", 7952 .read_s64 = cpu_rt_runtime_read, 7953 .write_s64 = cpu_rt_runtime_write, 7954 }, 7955 { 7956 .name = "rt_period_us", 7957 .read_u64 = cpu_rt_period_read_uint, 7958 .write_u64 = cpu_rt_period_write_uint, 7959 }, 7960#endif 7961 { } /* terminate */ 7962}; 7963 7964struct cgroup_subsys cpu_cgroup_subsys = { 7965 .name = "cpu", 7966 .css_alloc = cpu_cgroup_css_alloc, 7967 .css_free = cpu_cgroup_css_free, 7968 .css_online = cpu_cgroup_css_online, 7969 .css_offline = cpu_cgroup_css_offline, 7970 .can_attach = cpu_cgroup_can_attach, 7971 .attach = cpu_cgroup_attach, 7972 .exit = cpu_cgroup_exit, 7973 .subsys_id = cpu_cgroup_subsys_id, 7974 .base_cftypes = cpu_files, 7975 .early_init = 1, 7976}; 7977 7978#endif /* CONFIG_CGROUP_SCHED */ 7979 7980void dump_cpu_task(int cpu) 7981{ 7982 pr_info("Task dump for CPU %d:\n", cpu); 7983 sched_show_task(cpu_curr(cpu)); 7984} 7985