core.c revision 3c4017c13f91069194fce3160944efec50f15a6e
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75#include <linux/context_tracking.h> 76 77#include <asm/switch_to.h> 78#include <asm/tlb.h> 79#include <asm/irq_regs.h> 80#include <asm/mutex.h> 81#ifdef CONFIG_PARAVIRT 82#include <asm/paravirt.h> 83#endif 84 85#include "sched.h" 86#include "../workqueue_internal.h" 87#include "../smpboot.h" 88 89#define CREATE_TRACE_POINTS 90#include <trace/events/sched.h> 91 92void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 93{ 94 unsigned long delta; 95 ktime_t soft, hard, now; 96 97 for (;;) { 98 if (hrtimer_active(period_timer)) 99 break; 100 101 now = hrtimer_cb_get_time(period_timer); 102 hrtimer_forward(period_timer, now, period); 103 104 soft = hrtimer_get_softexpires(period_timer); 105 hard = hrtimer_get_expires(period_timer); 106 delta = ktime_to_ns(ktime_sub(hard, soft)); 107 __hrtimer_start_range_ns(period_timer, soft, delta, 108 HRTIMER_MODE_ABS_PINNED, 0); 109 } 110} 111 112DEFINE_MUTEX(sched_domains_mutex); 113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115static void update_rq_clock_task(struct rq *rq, s64 delta); 116 117void update_rq_clock(struct rq *rq) 118{ 119 s64 delta; 120 121 if (rq->skip_clock_update > 0) 122 return; 123 124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 125 rq->clock += delta; 126 update_rq_clock_task(rq, delta); 127} 128 129/* 130 * Debugging: various feature bits 131 */ 132 133#define SCHED_FEAT(name, enabled) \ 134 (1UL << __SCHED_FEAT_##name) * enabled | 135 136const_debug unsigned int sysctl_sched_features = 137#include "features.h" 138 0; 139 140#undef SCHED_FEAT 141 142#ifdef CONFIG_SCHED_DEBUG 143#define SCHED_FEAT(name, enabled) \ 144 #name , 145 146static const char * const sched_feat_names[] = { 147#include "features.h" 148}; 149 150#undef SCHED_FEAT 151 152static int sched_feat_show(struct seq_file *m, void *v) 153{ 154 int i; 155 156 for (i = 0; i < __SCHED_FEAT_NR; i++) { 157 if (!(sysctl_sched_features & (1UL << i))) 158 seq_puts(m, "NO_"); 159 seq_printf(m, "%s ", sched_feat_names[i]); 160 } 161 seq_puts(m, "\n"); 162 163 return 0; 164} 165 166#ifdef HAVE_JUMP_LABEL 167 168#define jump_label_key__true STATIC_KEY_INIT_TRUE 169#define jump_label_key__false STATIC_KEY_INIT_FALSE 170 171#define SCHED_FEAT(name, enabled) \ 172 jump_label_key__##enabled , 173 174struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 175#include "features.h" 176}; 177 178#undef SCHED_FEAT 179 180static void sched_feat_disable(int i) 181{ 182 if (static_key_enabled(&sched_feat_keys[i])) 183 static_key_slow_dec(&sched_feat_keys[i]); 184} 185 186static void sched_feat_enable(int i) 187{ 188 if (!static_key_enabled(&sched_feat_keys[i])) 189 static_key_slow_inc(&sched_feat_keys[i]); 190} 191#else 192static void sched_feat_disable(int i) { }; 193static void sched_feat_enable(int i) { }; 194#endif /* HAVE_JUMP_LABEL */ 195 196static int sched_feat_set(char *cmp) 197{ 198 int i; 199 int neg = 0; 200 201 if (strncmp(cmp, "NO_", 3) == 0) { 202 neg = 1; 203 cmp += 3; 204 } 205 206 for (i = 0; i < __SCHED_FEAT_NR; i++) { 207 if (strcmp(cmp, sched_feat_names[i]) == 0) { 208 if (neg) { 209 sysctl_sched_features &= ~(1UL << i); 210 sched_feat_disable(i); 211 } else { 212 sysctl_sched_features |= (1UL << i); 213 sched_feat_enable(i); 214 } 215 break; 216 } 217 } 218 219 return i; 220} 221 222static ssize_t 223sched_feat_write(struct file *filp, const char __user *ubuf, 224 size_t cnt, loff_t *ppos) 225{ 226 char buf[64]; 227 char *cmp; 228 int i; 229 230 if (cnt > 63) 231 cnt = 63; 232 233 if (copy_from_user(&buf, ubuf, cnt)) 234 return -EFAULT; 235 236 buf[cnt] = 0; 237 cmp = strstrip(buf); 238 239 i = sched_feat_set(cmp); 240 if (i == __SCHED_FEAT_NR) 241 return -EINVAL; 242 243 *ppos += cnt; 244 245 return cnt; 246} 247 248static int sched_feat_open(struct inode *inode, struct file *filp) 249{ 250 return single_open(filp, sched_feat_show, NULL); 251} 252 253static const struct file_operations sched_feat_fops = { 254 .open = sched_feat_open, 255 .write = sched_feat_write, 256 .read = seq_read, 257 .llseek = seq_lseek, 258 .release = single_release, 259}; 260 261static __init int sched_init_debug(void) 262{ 263 debugfs_create_file("sched_features", 0644, NULL, NULL, 264 &sched_feat_fops); 265 266 return 0; 267} 268late_initcall(sched_init_debug); 269#endif /* CONFIG_SCHED_DEBUG */ 270 271/* 272 * Number of tasks to iterate in a single balance run. 273 * Limited because this is done with IRQs disabled. 274 */ 275const_debug unsigned int sysctl_sched_nr_migrate = 32; 276 277/* 278 * period over which we average the RT time consumption, measured 279 * in ms. 280 * 281 * default: 1s 282 */ 283const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 284 285/* 286 * period over which we measure -rt task cpu usage in us. 287 * default: 1s 288 */ 289unsigned int sysctl_sched_rt_period = 1000000; 290 291__read_mostly int scheduler_running; 292 293/* 294 * part of the period that we allow rt tasks to run in us. 295 * default: 0.95s 296 */ 297int sysctl_sched_rt_runtime = 950000; 298 299/* 300 * __task_rq_lock - lock the rq @p resides on. 301 */ 302static inline struct rq *__task_rq_lock(struct task_struct *p) 303 __acquires(rq->lock) 304{ 305 struct rq *rq; 306 307 lockdep_assert_held(&p->pi_lock); 308 309 for (;;) { 310 rq = task_rq(p); 311 raw_spin_lock(&rq->lock); 312 if (likely(rq == task_rq(p))) 313 return rq; 314 raw_spin_unlock(&rq->lock); 315 } 316} 317 318/* 319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 320 */ 321static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 322 __acquires(p->pi_lock) 323 __acquires(rq->lock) 324{ 325 struct rq *rq; 326 327 for (;;) { 328 raw_spin_lock_irqsave(&p->pi_lock, *flags); 329 rq = task_rq(p); 330 raw_spin_lock(&rq->lock); 331 if (likely(rq == task_rq(p))) 332 return rq; 333 raw_spin_unlock(&rq->lock); 334 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 335 } 336} 337 338static void __task_rq_unlock(struct rq *rq) 339 __releases(rq->lock) 340{ 341 raw_spin_unlock(&rq->lock); 342} 343 344static inline void 345task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 346 __releases(rq->lock) 347 __releases(p->pi_lock) 348{ 349 raw_spin_unlock(&rq->lock); 350 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 351} 352 353/* 354 * this_rq_lock - lock this runqueue and disable interrupts. 355 */ 356static struct rq *this_rq_lock(void) 357 __acquires(rq->lock) 358{ 359 struct rq *rq; 360 361 local_irq_disable(); 362 rq = this_rq(); 363 raw_spin_lock(&rq->lock); 364 365 return rq; 366} 367 368#ifdef CONFIG_SCHED_HRTICK 369/* 370 * Use HR-timers to deliver accurate preemption points. 371 */ 372 373static void hrtick_clear(struct rq *rq) 374{ 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377} 378 379/* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383static enum hrtimer_restart hrtick(struct hrtimer *timer) 384{ 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395} 396 397#ifdef CONFIG_SMP 398 399static int __hrtick_restart(struct rq *rq) 400{ 401 struct hrtimer *timer = &rq->hrtick_timer; 402 ktime_t time = hrtimer_get_softexpires(timer); 403 404 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 405} 406 407/* 408 * called from hardirq (IPI) context 409 */ 410static void __hrtick_start(void *arg) 411{ 412 struct rq *rq = arg; 413 414 raw_spin_lock(&rq->lock); 415 __hrtick_restart(rq); 416 rq->hrtick_csd_pending = 0; 417 raw_spin_unlock(&rq->lock); 418} 419 420/* 421 * Called to set the hrtick timer state. 422 * 423 * called with rq->lock held and irqs disabled 424 */ 425void hrtick_start(struct rq *rq, u64 delay) 426{ 427 struct hrtimer *timer = &rq->hrtick_timer; 428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 429 430 hrtimer_set_expires(timer, time); 431 432 if (rq == this_rq()) { 433 __hrtick_restart(rq); 434 } else if (!rq->hrtick_csd_pending) { 435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 436 rq->hrtick_csd_pending = 1; 437 } 438} 439 440static int 441hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 442{ 443 int cpu = (int)(long)hcpu; 444 445 switch (action) { 446 case CPU_UP_CANCELED: 447 case CPU_UP_CANCELED_FROZEN: 448 case CPU_DOWN_PREPARE: 449 case CPU_DOWN_PREPARE_FROZEN: 450 case CPU_DEAD: 451 case CPU_DEAD_FROZEN: 452 hrtick_clear(cpu_rq(cpu)); 453 return NOTIFY_OK; 454 } 455 456 return NOTIFY_DONE; 457} 458 459static __init void init_hrtick(void) 460{ 461 hotcpu_notifier(hotplug_hrtick, 0); 462} 463#else 464/* 465 * Called to set the hrtick timer state. 466 * 467 * called with rq->lock held and irqs disabled 468 */ 469void hrtick_start(struct rq *rq, u64 delay) 470{ 471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 472 HRTIMER_MODE_REL_PINNED, 0); 473} 474 475static inline void init_hrtick(void) 476{ 477} 478#endif /* CONFIG_SMP */ 479 480static void init_rq_hrtick(struct rq *rq) 481{ 482#ifdef CONFIG_SMP 483 rq->hrtick_csd_pending = 0; 484 485 rq->hrtick_csd.flags = 0; 486 rq->hrtick_csd.func = __hrtick_start; 487 rq->hrtick_csd.info = rq; 488#endif 489 490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 491 rq->hrtick_timer.function = hrtick; 492} 493#else /* CONFIG_SCHED_HRTICK */ 494static inline void hrtick_clear(struct rq *rq) 495{ 496} 497 498static inline void init_rq_hrtick(struct rq *rq) 499{ 500} 501 502static inline void init_hrtick(void) 503{ 504} 505#endif /* CONFIG_SCHED_HRTICK */ 506 507/* 508 * resched_task - mark a task 'to be rescheduled now'. 509 * 510 * On UP this means the setting of the need_resched flag, on SMP it 511 * might also involve a cross-CPU call to trigger the scheduler on 512 * the target CPU. 513 */ 514void resched_task(struct task_struct *p) 515{ 516 int cpu; 517 518 lockdep_assert_held(&task_rq(p)->lock); 519 520 if (test_tsk_need_resched(p)) 521 return; 522 523 set_tsk_need_resched(p); 524 525 cpu = task_cpu(p); 526 if (cpu == smp_processor_id()) { 527 set_preempt_need_resched(); 528 return; 529 } 530 531 /* NEED_RESCHED must be visible before we test polling */ 532 smp_mb(); 533 if (!tsk_is_polling(p)) 534 smp_send_reschedule(cpu); 535} 536 537void resched_cpu(int cpu) 538{ 539 struct rq *rq = cpu_rq(cpu); 540 unsigned long flags; 541 542 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 543 return; 544 resched_task(cpu_curr(cpu)); 545 raw_spin_unlock_irqrestore(&rq->lock, flags); 546} 547 548#ifdef CONFIG_SMP 549#ifdef CONFIG_NO_HZ_COMMON 550/* 551 * In the semi idle case, use the nearest busy cpu for migrating timers 552 * from an idle cpu. This is good for power-savings. 553 * 554 * We don't do similar optimization for completely idle system, as 555 * selecting an idle cpu will add more delays to the timers than intended 556 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 557 */ 558int get_nohz_timer_target(void) 559{ 560 int cpu = smp_processor_id(); 561 int i; 562 struct sched_domain *sd; 563 564 rcu_read_lock(); 565 for_each_domain(cpu, sd) { 566 for_each_cpu(i, sched_domain_span(sd)) { 567 if (!idle_cpu(i)) { 568 cpu = i; 569 goto unlock; 570 } 571 } 572 } 573unlock: 574 rcu_read_unlock(); 575 return cpu; 576} 577/* 578 * When add_timer_on() enqueues a timer into the timer wheel of an 579 * idle CPU then this timer might expire before the next timer event 580 * which is scheduled to wake up that CPU. In case of a completely 581 * idle system the next event might even be infinite time into the 582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 583 * leaves the inner idle loop so the newly added timer is taken into 584 * account when the CPU goes back to idle and evaluates the timer 585 * wheel for the next timer event. 586 */ 587static void wake_up_idle_cpu(int cpu) 588{ 589 struct rq *rq = cpu_rq(cpu); 590 591 if (cpu == smp_processor_id()) 592 return; 593 594 /* 595 * This is safe, as this function is called with the timer 596 * wheel base lock of (cpu) held. When the CPU is on the way 597 * to idle and has not yet set rq->curr to idle then it will 598 * be serialized on the timer wheel base lock and take the new 599 * timer into account automatically. 600 */ 601 if (rq->curr != rq->idle) 602 return; 603 604 /* 605 * We can set TIF_RESCHED on the idle task of the other CPU 606 * lockless. The worst case is that the other CPU runs the 607 * idle task through an additional NOOP schedule() 608 */ 609 set_tsk_need_resched(rq->idle); 610 611 /* NEED_RESCHED must be visible before we test polling */ 612 smp_mb(); 613 if (!tsk_is_polling(rq->idle)) 614 smp_send_reschedule(cpu); 615} 616 617static bool wake_up_full_nohz_cpu(int cpu) 618{ 619 if (tick_nohz_full_cpu(cpu)) { 620 if (cpu != smp_processor_id() || 621 tick_nohz_tick_stopped()) 622 smp_send_reschedule(cpu); 623 return true; 624 } 625 626 return false; 627} 628 629void wake_up_nohz_cpu(int cpu) 630{ 631 if (!wake_up_full_nohz_cpu(cpu)) 632 wake_up_idle_cpu(cpu); 633} 634 635static inline bool got_nohz_idle_kick(void) 636{ 637 int cpu = smp_processor_id(); 638 639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 640 return false; 641 642 if (idle_cpu(cpu) && !need_resched()) 643 return true; 644 645 /* 646 * We can't run Idle Load Balance on this CPU for this time so we 647 * cancel it and clear NOHZ_BALANCE_KICK 648 */ 649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 650 return false; 651} 652 653#else /* CONFIG_NO_HZ_COMMON */ 654 655static inline bool got_nohz_idle_kick(void) 656{ 657 return false; 658} 659 660#endif /* CONFIG_NO_HZ_COMMON */ 661 662#ifdef CONFIG_NO_HZ_FULL 663bool sched_can_stop_tick(void) 664{ 665 struct rq *rq; 666 667 rq = this_rq(); 668 669 /* Make sure rq->nr_running update is visible after the IPI */ 670 smp_rmb(); 671 672 /* More than one running task need preemption */ 673 if (rq->nr_running > 1) 674 return false; 675 676 return true; 677} 678#endif /* CONFIG_NO_HZ_FULL */ 679 680void sched_avg_update(struct rq *rq) 681{ 682 s64 period = sched_avg_period(); 683 684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 685 /* 686 * Inline assembly required to prevent the compiler 687 * optimising this loop into a divmod call. 688 * See __iter_div_u64_rem() for another example of this. 689 */ 690 asm("" : "+rm" (rq->age_stamp)); 691 rq->age_stamp += period; 692 rq->rt_avg /= 2; 693 } 694} 695 696#endif /* CONFIG_SMP */ 697 698#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 699 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 700/* 701 * Iterate task_group tree rooted at *from, calling @down when first entering a 702 * node and @up when leaving it for the final time. 703 * 704 * Caller must hold rcu_lock or sufficient equivalent. 705 */ 706int walk_tg_tree_from(struct task_group *from, 707 tg_visitor down, tg_visitor up, void *data) 708{ 709 struct task_group *parent, *child; 710 int ret; 711 712 parent = from; 713 714down: 715 ret = (*down)(parent, data); 716 if (ret) 717 goto out; 718 list_for_each_entry_rcu(child, &parent->children, siblings) { 719 parent = child; 720 goto down; 721 722up: 723 continue; 724 } 725 ret = (*up)(parent, data); 726 if (ret || parent == from) 727 goto out; 728 729 child = parent; 730 parent = parent->parent; 731 if (parent) 732 goto up; 733out: 734 return ret; 735} 736 737int tg_nop(struct task_group *tg, void *data) 738{ 739 return 0; 740} 741#endif 742 743static void set_load_weight(struct task_struct *p) 744{ 745 int prio = p->static_prio - MAX_RT_PRIO; 746 struct load_weight *load = &p->se.load; 747 748 /* 749 * SCHED_IDLE tasks get minimal weight: 750 */ 751 if (p->policy == SCHED_IDLE) { 752 load->weight = scale_load(WEIGHT_IDLEPRIO); 753 load->inv_weight = WMULT_IDLEPRIO; 754 return; 755 } 756 757 load->weight = scale_load(prio_to_weight[prio]); 758 load->inv_weight = prio_to_wmult[prio]; 759} 760 761static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 762{ 763 update_rq_clock(rq); 764 sched_info_queued(rq, p); 765 p->sched_class->enqueue_task(rq, p, flags); 766} 767 768static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 769{ 770 update_rq_clock(rq); 771 sched_info_dequeued(rq, p); 772 p->sched_class->dequeue_task(rq, p, flags); 773} 774 775void activate_task(struct rq *rq, struct task_struct *p, int flags) 776{ 777 if (task_contributes_to_load(p)) 778 rq->nr_uninterruptible--; 779 780 enqueue_task(rq, p, flags); 781} 782 783void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 784{ 785 if (task_contributes_to_load(p)) 786 rq->nr_uninterruptible++; 787 788 dequeue_task(rq, p, flags); 789} 790 791static void update_rq_clock_task(struct rq *rq, s64 delta) 792{ 793/* 794 * In theory, the compile should just see 0 here, and optimize out the call 795 * to sched_rt_avg_update. But I don't trust it... 796 */ 797#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 798 s64 steal = 0, irq_delta = 0; 799#endif 800#ifdef CONFIG_IRQ_TIME_ACCOUNTING 801 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 802 803 /* 804 * Since irq_time is only updated on {soft,}irq_exit, we might run into 805 * this case when a previous update_rq_clock() happened inside a 806 * {soft,}irq region. 807 * 808 * When this happens, we stop ->clock_task and only update the 809 * prev_irq_time stamp to account for the part that fit, so that a next 810 * update will consume the rest. This ensures ->clock_task is 811 * monotonic. 812 * 813 * It does however cause some slight miss-attribution of {soft,}irq 814 * time, a more accurate solution would be to update the irq_time using 815 * the current rq->clock timestamp, except that would require using 816 * atomic ops. 817 */ 818 if (irq_delta > delta) 819 irq_delta = delta; 820 821 rq->prev_irq_time += irq_delta; 822 delta -= irq_delta; 823#endif 824#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 825 if (static_key_false((¶virt_steal_rq_enabled))) { 826 u64 st; 827 828 steal = paravirt_steal_clock(cpu_of(rq)); 829 steal -= rq->prev_steal_time_rq; 830 831 if (unlikely(steal > delta)) 832 steal = delta; 833 834 st = steal_ticks(steal); 835 steal = st * TICK_NSEC; 836 837 rq->prev_steal_time_rq += steal; 838 839 delta -= steal; 840 } 841#endif 842 843 rq->clock_task += delta; 844 845#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 846 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 847 sched_rt_avg_update(rq, irq_delta + steal); 848#endif 849} 850 851void sched_set_stop_task(int cpu, struct task_struct *stop) 852{ 853 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 854 struct task_struct *old_stop = cpu_rq(cpu)->stop; 855 856 if (stop) { 857 /* 858 * Make it appear like a SCHED_FIFO task, its something 859 * userspace knows about and won't get confused about. 860 * 861 * Also, it will make PI more or less work without too 862 * much confusion -- but then, stop work should not 863 * rely on PI working anyway. 864 */ 865 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 866 867 stop->sched_class = &stop_sched_class; 868 } 869 870 cpu_rq(cpu)->stop = stop; 871 872 if (old_stop) { 873 /* 874 * Reset it back to a normal scheduling class so that 875 * it can die in pieces. 876 */ 877 old_stop->sched_class = &rt_sched_class; 878 } 879} 880 881/* 882 * __normal_prio - return the priority that is based on the static prio 883 */ 884static inline int __normal_prio(struct task_struct *p) 885{ 886 return p->static_prio; 887} 888 889/* 890 * Calculate the expected normal priority: i.e. priority 891 * without taking RT-inheritance into account. Might be 892 * boosted by interactivity modifiers. Changes upon fork, 893 * setprio syscalls, and whenever the interactivity 894 * estimator recalculates. 895 */ 896static inline int normal_prio(struct task_struct *p) 897{ 898 int prio; 899 900 if (task_has_dl_policy(p)) 901 prio = MAX_DL_PRIO-1; 902 else if (task_has_rt_policy(p)) 903 prio = MAX_RT_PRIO-1 - p->rt_priority; 904 else 905 prio = __normal_prio(p); 906 return prio; 907} 908 909/* 910 * Calculate the current priority, i.e. the priority 911 * taken into account by the scheduler. This value might 912 * be boosted by RT tasks, or might be boosted by 913 * interactivity modifiers. Will be RT if the task got 914 * RT-boosted. If not then it returns p->normal_prio. 915 */ 916static int effective_prio(struct task_struct *p) 917{ 918 p->normal_prio = normal_prio(p); 919 /* 920 * If we are RT tasks or we were boosted to RT priority, 921 * keep the priority unchanged. Otherwise, update priority 922 * to the normal priority: 923 */ 924 if (!rt_prio(p->prio)) 925 return p->normal_prio; 926 return p->prio; 927} 928 929/** 930 * task_curr - is this task currently executing on a CPU? 931 * @p: the task in question. 932 * 933 * Return: 1 if the task is currently executing. 0 otherwise. 934 */ 935inline int task_curr(const struct task_struct *p) 936{ 937 return cpu_curr(task_cpu(p)) == p; 938} 939 940static inline void check_class_changed(struct rq *rq, struct task_struct *p, 941 const struct sched_class *prev_class, 942 int oldprio) 943{ 944 if (prev_class != p->sched_class) { 945 if (prev_class->switched_from) 946 prev_class->switched_from(rq, p); 947 p->sched_class->switched_to(rq, p); 948 } else if (oldprio != p->prio || dl_task(p)) 949 p->sched_class->prio_changed(rq, p, oldprio); 950} 951 952void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 953{ 954 const struct sched_class *class; 955 956 if (p->sched_class == rq->curr->sched_class) { 957 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 958 } else { 959 for_each_class(class) { 960 if (class == rq->curr->sched_class) 961 break; 962 if (class == p->sched_class) { 963 resched_task(rq->curr); 964 break; 965 } 966 } 967 } 968 969 /* 970 * A queue event has occurred, and we're going to schedule. In 971 * this case, we can save a useless back to back clock update. 972 */ 973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 974 rq->skip_clock_update = 1; 975} 976 977#ifdef CONFIG_SMP 978void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 979{ 980#ifdef CONFIG_SCHED_DEBUG 981 /* 982 * We should never call set_task_cpu() on a blocked task, 983 * ttwu() will sort out the placement. 984 */ 985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 986 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 987 988#ifdef CONFIG_LOCKDEP 989 /* 990 * The caller should hold either p->pi_lock or rq->lock, when changing 991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 992 * 993 * sched_move_task() holds both and thus holding either pins the cgroup, 994 * see task_group(). 995 * 996 * Furthermore, all task_rq users should acquire both locks, see 997 * task_rq_lock(). 998 */ 999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1000 lockdep_is_held(&task_rq(p)->lock))); 1001#endif 1002#endif 1003 1004 trace_sched_migrate_task(p, new_cpu); 1005 1006 if (task_cpu(p) != new_cpu) { 1007 if (p->sched_class->migrate_task_rq) 1008 p->sched_class->migrate_task_rq(p, new_cpu); 1009 p->se.nr_migrations++; 1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1011 } 1012 1013 __set_task_cpu(p, new_cpu); 1014} 1015 1016static void __migrate_swap_task(struct task_struct *p, int cpu) 1017{ 1018 if (p->on_rq) { 1019 struct rq *src_rq, *dst_rq; 1020 1021 src_rq = task_rq(p); 1022 dst_rq = cpu_rq(cpu); 1023 1024 deactivate_task(src_rq, p, 0); 1025 set_task_cpu(p, cpu); 1026 activate_task(dst_rq, p, 0); 1027 check_preempt_curr(dst_rq, p, 0); 1028 } else { 1029 /* 1030 * Task isn't running anymore; make it appear like we migrated 1031 * it before it went to sleep. This means on wakeup we make the 1032 * previous cpu our targer instead of where it really is. 1033 */ 1034 p->wake_cpu = cpu; 1035 } 1036} 1037 1038struct migration_swap_arg { 1039 struct task_struct *src_task, *dst_task; 1040 int src_cpu, dst_cpu; 1041}; 1042 1043static int migrate_swap_stop(void *data) 1044{ 1045 struct migration_swap_arg *arg = data; 1046 struct rq *src_rq, *dst_rq; 1047 int ret = -EAGAIN; 1048 1049 src_rq = cpu_rq(arg->src_cpu); 1050 dst_rq = cpu_rq(arg->dst_cpu); 1051 1052 double_raw_lock(&arg->src_task->pi_lock, 1053 &arg->dst_task->pi_lock); 1054 double_rq_lock(src_rq, dst_rq); 1055 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1056 goto unlock; 1057 1058 if (task_cpu(arg->src_task) != arg->src_cpu) 1059 goto unlock; 1060 1061 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1062 goto unlock; 1063 1064 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1065 goto unlock; 1066 1067 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1068 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1069 1070 ret = 0; 1071 1072unlock: 1073 double_rq_unlock(src_rq, dst_rq); 1074 raw_spin_unlock(&arg->dst_task->pi_lock); 1075 raw_spin_unlock(&arg->src_task->pi_lock); 1076 1077 return ret; 1078} 1079 1080/* 1081 * Cross migrate two tasks 1082 */ 1083int migrate_swap(struct task_struct *cur, struct task_struct *p) 1084{ 1085 struct migration_swap_arg arg; 1086 int ret = -EINVAL; 1087 1088 arg = (struct migration_swap_arg){ 1089 .src_task = cur, 1090 .src_cpu = task_cpu(cur), 1091 .dst_task = p, 1092 .dst_cpu = task_cpu(p), 1093 }; 1094 1095 if (arg.src_cpu == arg.dst_cpu) 1096 goto out; 1097 1098 /* 1099 * These three tests are all lockless; this is OK since all of them 1100 * will be re-checked with proper locks held further down the line. 1101 */ 1102 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1103 goto out; 1104 1105 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1106 goto out; 1107 1108 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1109 goto out; 1110 1111 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1112 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1113 1114out: 1115 return ret; 1116} 1117 1118struct migration_arg { 1119 struct task_struct *task; 1120 int dest_cpu; 1121}; 1122 1123static int migration_cpu_stop(void *data); 1124 1125/* 1126 * wait_task_inactive - wait for a thread to unschedule. 1127 * 1128 * If @match_state is nonzero, it's the @p->state value just checked and 1129 * not expected to change. If it changes, i.e. @p might have woken up, 1130 * then return zero. When we succeed in waiting for @p to be off its CPU, 1131 * we return a positive number (its total switch count). If a second call 1132 * a short while later returns the same number, the caller can be sure that 1133 * @p has remained unscheduled the whole time. 1134 * 1135 * The caller must ensure that the task *will* unschedule sometime soon, 1136 * else this function might spin for a *long* time. This function can't 1137 * be called with interrupts off, or it may introduce deadlock with 1138 * smp_call_function() if an IPI is sent by the same process we are 1139 * waiting to become inactive. 1140 */ 1141unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1142{ 1143 unsigned long flags; 1144 int running, on_rq; 1145 unsigned long ncsw; 1146 struct rq *rq; 1147 1148 for (;;) { 1149 /* 1150 * We do the initial early heuristics without holding 1151 * any task-queue locks at all. We'll only try to get 1152 * the runqueue lock when things look like they will 1153 * work out! 1154 */ 1155 rq = task_rq(p); 1156 1157 /* 1158 * If the task is actively running on another CPU 1159 * still, just relax and busy-wait without holding 1160 * any locks. 1161 * 1162 * NOTE! Since we don't hold any locks, it's not 1163 * even sure that "rq" stays as the right runqueue! 1164 * But we don't care, since "task_running()" will 1165 * return false if the runqueue has changed and p 1166 * is actually now running somewhere else! 1167 */ 1168 while (task_running(rq, p)) { 1169 if (match_state && unlikely(p->state != match_state)) 1170 return 0; 1171 cpu_relax(); 1172 } 1173 1174 /* 1175 * Ok, time to look more closely! We need the rq 1176 * lock now, to be *sure*. If we're wrong, we'll 1177 * just go back and repeat. 1178 */ 1179 rq = task_rq_lock(p, &flags); 1180 trace_sched_wait_task(p); 1181 running = task_running(rq, p); 1182 on_rq = p->on_rq; 1183 ncsw = 0; 1184 if (!match_state || p->state == match_state) 1185 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1186 task_rq_unlock(rq, p, &flags); 1187 1188 /* 1189 * If it changed from the expected state, bail out now. 1190 */ 1191 if (unlikely(!ncsw)) 1192 break; 1193 1194 /* 1195 * Was it really running after all now that we 1196 * checked with the proper locks actually held? 1197 * 1198 * Oops. Go back and try again.. 1199 */ 1200 if (unlikely(running)) { 1201 cpu_relax(); 1202 continue; 1203 } 1204 1205 /* 1206 * It's not enough that it's not actively running, 1207 * it must be off the runqueue _entirely_, and not 1208 * preempted! 1209 * 1210 * So if it was still runnable (but just not actively 1211 * running right now), it's preempted, and we should 1212 * yield - it could be a while. 1213 */ 1214 if (unlikely(on_rq)) { 1215 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1216 1217 set_current_state(TASK_UNINTERRUPTIBLE); 1218 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1219 continue; 1220 } 1221 1222 /* 1223 * Ahh, all good. It wasn't running, and it wasn't 1224 * runnable, which means that it will never become 1225 * running in the future either. We're all done! 1226 */ 1227 break; 1228 } 1229 1230 return ncsw; 1231} 1232 1233/*** 1234 * kick_process - kick a running thread to enter/exit the kernel 1235 * @p: the to-be-kicked thread 1236 * 1237 * Cause a process which is running on another CPU to enter 1238 * kernel-mode, without any delay. (to get signals handled.) 1239 * 1240 * NOTE: this function doesn't have to take the runqueue lock, 1241 * because all it wants to ensure is that the remote task enters 1242 * the kernel. If the IPI races and the task has been migrated 1243 * to another CPU then no harm is done and the purpose has been 1244 * achieved as well. 1245 */ 1246void kick_process(struct task_struct *p) 1247{ 1248 int cpu; 1249 1250 preempt_disable(); 1251 cpu = task_cpu(p); 1252 if ((cpu != smp_processor_id()) && task_curr(p)) 1253 smp_send_reschedule(cpu); 1254 preempt_enable(); 1255} 1256EXPORT_SYMBOL_GPL(kick_process); 1257#endif /* CONFIG_SMP */ 1258 1259#ifdef CONFIG_SMP 1260/* 1261 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1262 */ 1263static int select_fallback_rq(int cpu, struct task_struct *p) 1264{ 1265 int nid = cpu_to_node(cpu); 1266 const struct cpumask *nodemask = NULL; 1267 enum { cpuset, possible, fail } state = cpuset; 1268 int dest_cpu; 1269 1270 /* 1271 * If the node that the cpu is on has been offlined, cpu_to_node() 1272 * will return -1. There is no cpu on the node, and we should 1273 * select the cpu on the other node. 1274 */ 1275 if (nid != -1) { 1276 nodemask = cpumask_of_node(nid); 1277 1278 /* Look for allowed, online CPU in same node. */ 1279 for_each_cpu(dest_cpu, nodemask) { 1280 if (!cpu_online(dest_cpu)) 1281 continue; 1282 if (!cpu_active(dest_cpu)) 1283 continue; 1284 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1285 return dest_cpu; 1286 } 1287 } 1288 1289 for (;;) { 1290 /* Any allowed, online CPU? */ 1291 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1292 if (!cpu_online(dest_cpu)) 1293 continue; 1294 if (!cpu_active(dest_cpu)) 1295 continue; 1296 goto out; 1297 } 1298 1299 switch (state) { 1300 case cpuset: 1301 /* No more Mr. Nice Guy. */ 1302 cpuset_cpus_allowed_fallback(p); 1303 state = possible; 1304 break; 1305 1306 case possible: 1307 do_set_cpus_allowed(p, cpu_possible_mask); 1308 state = fail; 1309 break; 1310 1311 case fail: 1312 BUG(); 1313 break; 1314 } 1315 } 1316 1317out: 1318 if (state != cpuset) { 1319 /* 1320 * Don't tell them about moving exiting tasks or 1321 * kernel threads (both mm NULL), since they never 1322 * leave kernel. 1323 */ 1324 if (p->mm && printk_ratelimit()) { 1325 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1326 task_pid_nr(p), p->comm, cpu); 1327 } 1328 } 1329 1330 return dest_cpu; 1331} 1332 1333/* 1334 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1335 */ 1336static inline 1337int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1338{ 1339 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1340 1341 /* 1342 * In order not to call set_task_cpu() on a blocking task we need 1343 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1344 * cpu. 1345 * 1346 * Since this is common to all placement strategies, this lives here. 1347 * 1348 * [ this allows ->select_task() to simply return task_cpu(p) and 1349 * not worry about this generic constraint ] 1350 */ 1351 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1352 !cpu_online(cpu))) 1353 cpu = select_fallback_rq(task_cpu(p), p); 1354 1355 return cpu; 1356} 1357 1358static void update_avg(u64 *avg, u64 sample) 1359{ 1360 s64 diff = sample - *avg; 1361 *avg += diff >> 3; 1362} 1363#endif 1364 1365static void 1366ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1367{ 1368#ifdef CONFIG_SCHEDSTATS 1369 struct rq *rq = this_rq(); 1370 1371#ifdef CONFIG_SMP 1372 int this_cpu = smp_processor_id(); 1373 1374 if (cpu == this_cpu) { 1375 schedstat_inc(rq, ttwu_local); 1376 schedstat_inc(p, se.statistics.nr_wakeups_local); 1377 } else { 1378 struct sched_domain *sd; 1379 1380 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1381 rcu_read_lock(); 1382 for_each_domain(this_cpu, sd) { 1383 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1384 schedstat_inc(sd, ttwu_wake_remote); 1385 break; 1386 } 1387 } 1388 rcu_read_unlock(); 1389 } 1390 1391 if (wake_flags & WF_MIGRATED) 1392 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1393 1394#endif /* CONFIG_SMP */ 1395 1396 schedstat_inc(rq, ttwu_count); 1397 schedstat_inc(p, se.statistics.nr_wakeups); 1398 1399 if (wake_flags & WF_SYNC) 1400 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1401 1402#endif /* CONFIG_SCHEDSTATS */ 1403} 1404 1405static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1406{ 1407 activate_task(rq, p, en_flags); 1408 p->on_rq = 1; 1409 1410 /* if a worker is waking up, notify workqueue */ 1411 if (p->flags & PF_WQ_WORKER) 1412 wq_worker_waking_up(p, cpu_of(rq)); 1413} 1414 1415/* 1416 * Mark the task runnable and perform wakeup-preemption. 1417 */ 1418static void 1419ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1420{ 1421 check_preempt_curr(rq, p, wake_flags); 1422 trace_sched_wakeup(p, true); 1423 1424 p->state = TASK_RUNNING; 1425#ifdef CONFIG_SMP 1426 if (p->sched_class->task_woken) 1427 p->sched_class->task_woken(rq, p); 1428 1429 if (rq->idle_stamp) { 1430 u64 delta = rq_clock(rq) - rq->idle_stamp; 1431 u64 max = 2*rq->max_idle_balance_cost; 1432 1433 update_avg(&rq->avg_idle, delta); 1434 1435 if (rq->avg_idle > max) 1436 rq->avg_idle = max; 1437 1438 rq->idle_stamp = 0; 1439 } 1440#endif 1441} 1442 1443static void 1444ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1445{ 1446#ifdef CONFIG_SMP 1447 if (p->sched_contributes_to_load) 1448 rq->nr_uninterruptible--; 1449#endif 1450 1451 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1452 ttwu_do_wakeup(rq, p, wake_flags); 1453} 1454 1455/* 1456 * Called in case the task @p isn't fully descheduled from its runqueue, 1457 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1458 * since all we need to do is flip p->state to TASK_RUNNING, since 1459 * the task is still ->on_rq. 1460 */ 1461static int ttwu_remote(struct task_struct *p, int wake_flags) 1462{ 1463 struct rq *rq; 1464 int ret = 0; 1465 1466 rq = __task_rq_lock(p); 1467 if (p->on_rq) { 1468 /* check_preempt_curr() may use rq clock */ 1469 update_rq_clock(rq); 1470 ttwu_do_wakeup(rq, p, wake_flags); 1471 ret = 1; 1472 } 1473 __task_rq_unlock(rq); 1474 1475 return ret; 1476} 1477 1478#ifdef CONFIG_SMP 1479static void sched_ttwu_pending(void) 1480{ 1481 struct rq *rq = this_rq(); 1482 struct llist_node *llist = llist_del_all(&rq->wake_list); 1483 struct task_struct *p; 1484 1485 raw_spin_lock(&rq->lock); 1486 1487 while (llist) { 1488 p = llist_entry(llist, struct task_struct, wake_entry); 1489 llist = llist_next(llist); 1490 ttwu_do_activate(rq, p, 0); 1491 } 1492 1493 raw_spin_unlock(&rq->lock); 1494} 1495 1496void scheduler_ipi(void) 1497{ 1498 /* 1499 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1500 * TIF_NEED_RESCHED remotely (for the first time) will also send 1501 * this IPI. 1502 */ 1503 preempt_fold_need_resched(); 1504 1505 if (llist_empty(&this_rq()->wake_list) 1506 && !tick_nohz_full_cpu(smp_processor_id()) 1507 && !got_nohz_idle_kick()) 1508 return; 1509 1510 /* 1511 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1512 * traditionally all their work was done from the interrupt return 1513 * path. Now that we actually do some work, we need to make sure 1514 * we do call them. 1515 * 1516 * Some archs already do call them, luckily irq_enter/exit nest 1517 * properly. 1518 * 1519 * Arguably we should visit all archs and update all handlers, 1520 * however a fair share of IPIs are still resched only so this would 1521 * somewhat pessimize the simple resched case. 1522 */ 1523 irq_enter(); 1524 tick_nohz_full_check(); 1525 sched_ttwu_pending(); 1526 1527 /* 1528 * Check if someone kicked us for doing the nohz idle load balance. 1529 */ 1530 if (unlikely(got_nohz_idle_kick())) { 1531 this_rq()->idle_balance = 1; 1532 raise_softirq_irqoff(SCHED_SOFTIRQ); 1533 } 1534 irq_exit(); 1535} 1536 1537static void ttwu_queue_remote(struct task_struct *p, int cpu) 1538{ 1539 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1540 smp_send_reschedule(cpu); 1541} 1542 1543bool cpus_share_cache(int this_cpu, int that_cpu) 1544{ 1545 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1546} 1547#endif /* CONFIG_SMP */ 1548 1549static void ttwu_queue(struct task_struct *p, int cpu) 1550{ 1551 struct rq *rq = cpu_rq(cpu); 1552 1553#if defined(CONFIG_SMP) 1554 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1555 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1556 ttwu_queue_remote(p, cpu); 1557 return; 1558 } 1559#endif 1560 1561 raw_spin_lock(&rq->lock); 1562 ttwu_do_activate(rq, p, 0); 1563 raw_spin_unlock(&rq->lock); 1564} 1565 1566/** 1567 * try_to_wake_up - wake up a thread 1568 * @p: the thread to be awakened 1569 * @state: the mask of task states that can be woken 1570 * @wake_flags: wake modifier flags (WF_*) 1571 * 1572 * Put it on the run-queue if it's not already there. The "current" 1573 * thread is always on the run-queue (except when the actual 1574 * re-schedule is in progress), and as such you're allowed to do 1575 * the simpler "current->state = TASK_RUNNING" to mark yourself 1576 * runnable without the overhead of this. 1577 * 1578 * Return: %true if @p was woken up, %false if it was already running. 1579 * or @state didn't match @p's state. 1580 */ 1581static int 1582try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1583{ 1584 unsigned long flags; 1585 int cpu, success = 0; 1586 1587 /* 1588 * If we are going to wake up a thread waiting for CONDITION we 1589 * need to ensure that CONDITION=1 done by the caller can not be 1590 * reordered with p->state check below. This pairs with mb() in 1591 * set_current_state() the waiting thread does. 1592 */ 1593 smp_mb__before_spinlock(); 1594 raw_spin_lock_irqsave(&p->pi_lock, flags); 1595 if (!(p->state & state)) 1596 goto out; 1597 1598 success = 1; /* we're going to change ->state */ 1599 cpu = task_cpu(p); 1600 1601 if (p->on_rq && ttwu_remote(p, wake_flags)) 1602 goto stat; 1603 1604#ifdef CONFIG_SMP 1605 /* 1606 * If the owning (remote) cpu is still in the middle of schedule() with 1607 * this task as prev, wait until its done referencing the task. 1608 */ 1609 while (p->on_cpu) 1610 cpu_relax(); 1611 /* 1612 * Pairs with the smp_wmb() in finish_lock_switch(). 1613 */ 1614 smp_rmb(); 1615 1616 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1617 p->state = TASK_WAKING; 1618 1619 if (p->sched_class->task_waking) 1620 p->sched_class->task_waking(p); 1621 1622 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1623 if (task_cpu(p) != cpu) { 1624 wake_flags |= WF_MIGRATED; 1625 set_task_cpu(p, cpu); 1626 } 1627#endif /* CONFIG_SMP */ 1628 1629 ttwu_queue(p, cpu); 1630stat: 1631 ttwu_stat(p, cpu, wake_flags); 1632out: 1633 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1634 1635 return success; 1636} 1637 1638/** 1639 * try_to_wake_up_local - try to wake up a local task with rq lock held 1640 * @p: the thread to be awakened 1641 * 1642 * Put @p on the run-queue if it's not already there. The caller must 1643 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1644 * the current task. 1645 */ 1646static void try_to_wake_up_local(struct task_struct *p) 1647{ 1648 struct rq *rq = task_rq(p); 1649 1650 if (WARN_ON_ONCE(rq != this_rq()) || 1651 WARN_ON_ONCE(p == current)) 1652 return; 1653 1654 lockdep_assert_held(&rq->lock); 1655 1656 if (!raw_spin_trylock(&p->pi_lock)) { 1657 raw_spin_unlock(&rq->lock); 1658 raw_spin_lock(&p->pi_lock); 1659 raw_spin_lock(&rq->lock); 1660 } 1661 1662 if (!(p->state & TASK_NORMAL)) 1663 goto out; 1664 1665 if (!p->on_rq) 1666 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1667 1668 ttwu_do_wakeup(rq, p, 0); 1669 ttwu_stat(p, smp_processor_id(), 0); 1670out: 1671 raw_spin_unlock(&p->pi_lock); 1672} 1673 1674/** 1675 * wake_up_process - Wake up a specific process 1676 * @p: The process to be woken up. 1677 * 1678 * Attempt to wake up the nominated process and move it to the set of runnable 1679 * processes. 1680 * 1681 * Return: 1 if the process was woken up, 0 if it was already running. 1682 * 1683 * It may be assumed that this function implies a write memory barrier before 1684 * changing the task state if and only if any tasks are woken up. 1685 */ 1686int wake_up_process(struct task_struct *p) 1687{ 1688 WARN_ON(task_is_stopped_or_traced(p)); 1689 return try_to_wake_up(p, TASK_NORMAL, 0); 1690} 1691EXPORT_SYMBOL(wake_up_process); 1692 1693int wake_up_state(struct task_struct *p, unsigned int state) 1694{ 1695 return try_to_wake_up(p, state, 0); 1696} 1697 1698/* 1699 * Perform scheduler related setup for a newly forked process p. 1700 * p is forked by current. 1701 * 1702 * __sched_fork() is basic setup used by init_idle() too: 1703 */ 1704static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1705{ 1706 p->on_rq = 0; 1707 1708 p->se.on_rq = 0; 1709 p->se.exec_start = 0; 1710 p->se.sum_exec_runtime = 0; 1711 p->se.prev_sum_exec_runtime = 0; 1712 p->se.nr_migrations = 0; 1713 p->se.vruntime = 0; 1714 INIT_LIST_HEAD(&p->se.group_node); 1715 1716#ifdef CONFIG_SCHEDSTATS 1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1718#endif 1719 1720 RB_CLEAR_NODE(&p->dl.rb_node); 1721 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1722 p->dl.dl_runtime = p->dl.runtime = 0; 1723 p->dl.dl_deadline = p->dl.deadline = 0; 1724 p->dl.dl_period = 0; 1725 p->dl.flags = 0; 1726 1727 INIT_LIST_HEAD(&p->rt.run_list); 1728 1729#ifdef CONFIG_PREEMPT_NOTIFIERS 1730 INIT_HLIST_HEAD(&p->preempt_notifiers); 1731#endif 1732 1733#ifdef CONFIG_NUMA_BALANCING 1734 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1735 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1736 p->mm->numa_scan_seq = 0; 1737 } 1738 1739 if (clone_flags & CLONE_VM) 1740 p->numa_preferred_nid = current->numa_preferred_nid; 1741 else 1742 p->numa_preferred_nid = -1; 1743 1744 p->node_stamp = 0ULL; 1745 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1746 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1747 p->numa_work.next = &p->numa_work; 1748 p->numa_faults_memory = NULL; 1749 p->numa_faults_buffer_memory = NULL; 1750 p->last_task_numa_placement = 0; 1751 p->last_sum_exec_runtime = 0; 1752 1753 INIT_LIST_HEAD(&p->numa_entry); 1754 p->numa_group = NULL; 1755#endif /* CONFIG_NUMA_BALANCING */ 1756} 1757 1758#ifdef CONFIG_NUMA_BALANCING 1759#ifdef CONFIG_SCHED_DEBUG 1760void set_numabalancing_state(bool enabled) 1761{ 1762 if (enabled) 1763 sched_feat_set("NUMA"); 1764 else 1765 sched_feat_set("NO_NUMA"); 1766} 1767#else 1768__read_mostly bool numabalancing_enabled; 1769 1770void set_numabalancing_state(bool enabled) 1771{ 1772 numabalancing_enabled = enabled; 1773} 1774#endif /* CONFIG_SCHED_DEBUG */ 1775 1776#ifdef CONFIG_PROC_SYSCTL 1777int sysctl_numa_balancing(struct ctl_table *table, int write, 1778 void __user *buffer, size_t *lenp, loff_t *ppos) 1779{ 1780 struct ctl_table t; 1781 int err; 1782 int state = numabalancing_enabled; 1783 1784 if (write && !capable(CAP_SYS_ADMIN)) 1785 return -EPERM; 1786 1787 t = *table; 1788 t.data = &state; 1789 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1790 if (err < 0) 1791 return err; 1792 if (write) 1793 set_numabalancing_state(state); 1794 return err; 1795} 1796#endif 1797#endif 1798 1799/* 1800 * fork()/clone()-time setup: 1801 */ 1802int sched_fork(unsigned long clone_flags, struct task_struct *p) 1803{ 1804 unsigned long flags; 1805 int cpu = get_cpu(); 1806 1807 __sched_fork(clone_flags, p); 1808 /* 1809 * We mark the process as running here. This guarantees that 1810 * nobody will actually run it, and a signal or other external 1811 * event cannot wake it up and insert it on the runqueue either. 1812 */ 1813 p->state = TASK_RUNNING; 1814 1815 /* 1816 * Make sure we do not leak PI boosting priority to the child. 1817 */ 1818 p->prio = current->normal_prio; 1819 1820 /* 1821 * Revert to default priority/policy on fork if requested. 1822 */ 1823 if (unlikely(p->sched_reset_on_fork)) { 1824 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1825 p->policy = SCHED_NORMAL; 1826 p->static_prio = NICE_TO_PRIO(0); 1827 p->rt_priority = 0; 1828 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1829 p->static_prio = NICE_TO_PRIO(0); 1830 1831 p->prio = p->normal_prio = __normal_prio(p); 1832 set_load_weight(p); 1833 1834 /* 1835 * We don't need the reset flag anymore after the fork. It has 1836 * fulfilled its duty: 1837 */ 1838 p->sched_reset_on_fork = 0; 1839 } 1840 1841 if (dl_prio(p->prio)) { 1842 put_cpu(); 1843 return -EAGAIN; 1844 } else if (rt_prio(p->prio)) { 1845 p->sched_class = &rt_sched_class; 1846 } else { 1847 p->sched_class = &fair_sched_class; 1848 } 1849 1850 if (p->sched_class->task_fork) 1851 p->sched_class->task_fork(p); 1852 1853 /* 1854 * The child is not yet in the pid-hash so no cgroup attach races, 1855 * and the cgroup is pinned to this child due to cgroup_fork() 1856 * is ran before sched_fork(). 1857 * 1858 * Silence PROVE_RCU. 1859 */ 1860 raw_spin_lock_irqsave(&p->pi_lock, flags); 1861 set_task_cpu(p, cpu); 1862 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1863 1864#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1865 if (likely(sched_info_on())) 1866 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1867#endif 1868#if defined(CONFIG_SMP) 1869 p->on_cpu = 0; 1870#endif 1871 init_task_preempt_count(p); 1872#ifdef CONFIG_SMP 1873 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1874 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1875#endif 1876 1877 put_cpu(); 1878 return 0; 1879} 1880 1881unsigned long to_ratio(u64 period, u64 runtime) 1882{ 1883 if (runtime == RUNTIME_INF) 1884 return 1ULL << 20; 1885 1886 /* 1887 * Doing this here saves a lot of checks in all 1888 * the calling paths, and returning zero seems 1889 * safe for them anyway. 1890 */ 1891 if (period == 0) 1892 return 0; 1893 1894 return div64_u64(runtime << 20, period); 1895} 1896 1897#ifdef CONFIG_SMP 1898inline struct dl_bw *dl_bw_of(int i) 1899{ 1900 return &cpu_rq(i)->rd->dl_bw; 1901} 1902 1903static inline int dl_bw_cpus(int i) 1904{ 1905 struct root_domain *rd = cpu_rq(i)->rd; 1906 int cpus = 0; 1907 1908 for_each_cpu_and(i, rd->span, cpu_active_mask) 1909 cpus++; 1910 1911 return cpus; 1912} 1913#else 1914inline struct dl_bw *dl_bw_of(int i) 1915{ 1916 return &cpu_rq(i)->dl.dl_bw; 1917} 1918 1919static inline int dl_bw_cpus(int i) 1920{ 1921 return 1; 1922} 1923#endif 1924 1925static inline 1926void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1927{ 1928 dl_b->total_bw -= tsk_bw; 1929} 1930 1931static inline 1932void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1933{ 1934 dl_b->total_bw += tsk_bw; 1935} 1936 1937static inline 1938bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1939{ 1940 return dl_b->bw != -1 && 1941 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1942} 1943 1944/* 1945 * We must be sure that accepting a new task (or allowing changing the 1946 * parameters of an existing one) is consistent with the bandwidth 1947 * constraints. If yes, this function also accordingly updates the currently 1948 * allocated bandwidth to reflect the new situation. 1949 * 1950 * This function is called while holding p's rq->lock. 1951 */ 1952static int dl_overflow(struct task_struct *p, int policy, 1953 const struct sched_attr *attr) 1954{ 1955 1956 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1957 u64 period = attr->sched_period; 1958 u64 runtime = attr->sched_runtime; 1959 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1960 int cpus, err = -1; 1961 1962 if (new_bw == p->dl.dl_bw) 1963 return 0; 1964 1965 /* 1966 * Either if a task, enters, leave, or stays -deadline but changes 1967 * its parameters, we may need to update accordingly the total 1968 * allocated bandwidth of the container. 1969 */ 1970 raw_spin_lock(&dl_b->lock); 1971 cpus = dl_bw_cpus(task_cpu(p)); 1972 if (dl_policy(policy) && !task_has_dl_policy(p) && 1973 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1974 __dl_add(dl_b, new_bw); 1975 err = 0; 1976 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1977 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1978 __dl_clear(dl_b, p->dl.dl_bw); 1979 __dl_add(dl_b, new_bw); 1980 err = 0; 1981 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1982 __dl_clear(dl_b, p->dl.dl_bw); 1983 err = 0; 1984 } 1985 raw_spin_unlock(&dl_b->lock); 1986 1987 return err; 1988} 1989 1990extern void init_dl_bw(struct dl_bw *dl_b); 1991 1992/* 1993 * wake_up_new_task - wake up a newly created task for the first time. 1994 * 1995 * This function will do some initial scheduler statistics housekeeping 1996 * that must be done for every newly created context, then puts the task 1997 * on the runqueue and wakes it. 1998 */ 1999void wake_up_new_task(struct task_struct *p) 2000{ 2001 unsigned long flags; 2002 struct rq *rq; 2003 2004 raw_spin_lock_irqsave(&p->pi_lock, flags); 2005#ifdef CONFIG_SMP 2006 /* 2007 * Fork balancing, do it here and not earlier because: 2008 * - cpus_allowed can change in the fork path 2009 * - any previously selected cpu might disappear through hotplug 2010 */ 2011 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2012#endif 2013 2014 /* Initialize new task's runnable average */ 2015 init_task_runnable_average(p); 2016 rq = __task_rq_lock(p); 2017 activate_task(rq, p, 0); 2018 p->on_rq = 1; 2019 trace_sched_wakeup_new(p, true); 2020 check_preempt_curr(rq, p, WF_FORK); 2021#ifdef CONFIG_SMP 2022 if (p->sched_class->task_woken) 2023 p->sched_class->task_woken(rq, p); 2024#endif 2025 task_rq_unlock(rq, p, &flags); 2026} 2027 2028#ifdef CONFIG_PREEMPT_NOTIFIERS 2029 2030/** 2031 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2032 * @notifier: notifier struct to register 2033 */ 2034void preempt_notifier_register(struct preempt_notifier *notifier) 2035{ 2036 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2037} 2038EXPORT_SYMBOL_GPL(preempt_notifier_register); 2039 2040/** 2041 * preempt_notifier_unregister - no longer interested in preemption notifications 2042 * @notifier: notifier struct to unregister 2043 * 2044 * This is safe to call from within a preemption notifier. 2045 */ 2046void preempt_notifier_unregister(struct preempt_notifier *notifier) 2047{ 2048 hlist_del(¬ifier->link); 2049} 2050EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2051 2052static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2053{ 2054 struct preempt_notifier *notifier; 2055 2056 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2057 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2058} 2059 2060static void 2061fire_sched_out_preempt_notifiers(struct task_struct *curr, 2062 struct task_struct *next) 2063{ 2064 struct preempt_notifier *notifier; 2065 2066 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2067 notifier->ops->sched_out(notifier, next); 2068} 2069 2070#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2071 2072static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2073{ 2074} 2075 2076static void 2077fire_sched_out_preempt_notifiers(struct task_struct *curr, 2078 struct task_struct *next) 2079{ 2080} 2081 2082#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2083 2084/** 2085 * prepare_task_switch - prepare to switch tasks 2086 * @rq: the runqueue preparing to switch 2087 * @prev: the current task that is being switched out 2088 * @next: the task we are going to switch to. 2089 * 2090 * This is called with the rq lock held and interrupts off. It must 2091 * be paired with a subsequent finish_task_switch after the context 2092 * switch. 2093 * 2094 * prepare_task_switch sets up locking and calls architecture specific 2095 * hooks. 2096 */ 2097static inline void 2098prepare_task_switch(struct rq *rq, struct task_struct *prev, 2099 struct task_struct *next) 2100{ 2101 trace_sched_switch(prev, next); 2102 sched_info_switch(rq, prev, next); 2103 perf_event_task_sched_out(prev, next); 2104 fire_sched_out_preempt_notifiers(prev, next); 2105 prepare_lock_switch(rq, next); 2106 prepare_arch_switch(next); 2107} 2108 2109/** 2110 * finish_task_switch - clean up after a task-switch 2111 * @rq: runqueue associated with task-switch 2112 * @prev: the thread we just switched away from. 2113 * 2114 * finish_task_switch must be called after the context switch, paired 2115 * with a prepare_task_switch call before the context switch. 2116 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2117 * and do any other architecture-specific cleanup actions. 2118 * 2119 * Note that we may have delayed dropping an mm in context_switch(). If 2120 * so, we finish that here outside of the runqueue lock. (Doing it 2121 * with the lock held can cause deadlocks; see schedule() for 2122 * details.) 2123 */ 2124static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2125 __releases(rq->lock) 2126{ 2127 struct mm_struct *mm = rq->prev_mm; 2128 long prev_state; 2129 2130 rq->prev_mm = NULL; 2131 2132 /* 2133 * A task struct has one reference for the use as "current". 2134 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2135 * schedule one last time. The schedule call will never return, and 2136 * the scheduled task must drop that reference. 2137 * The test for TASK_DEAD must occur while the runqueue locks are 2138 * still held, otherwise prev could be scheduled on another cpu, die 2139 * there before we look at prev->state, and then the reference would 2140 * be dropped twice. 2141 * Manfred Spraul <manfred@colorfullife.com> 2142 */ 2143 prev_state = prev->state; 2144 vtime_task_switch(prev); 2145 finish_arch_switch(prev); 2146 perf_event_task_sched_in(prev, current); 2147 finish_lock_switch(rq, prev); 2148 finish_arch_post_lock_switch(); 2149 2150 fire_sched_in_preempt_notifiers(current); 2151 if (mm) 2152 mmdrop(mm); 2153 if (unlikely(prev_state == TASK_DEAD)) { 2154 task_numa_free(prev); 2155 2156 if (prev->sched_class->task_dead) 2157 prev->sched_class->task_dead(prev); 2158 2159 /* 2160 * Remove function-return probe instances associated with this 2161 * task and put them back on the free list. 2162 */ 2163 kprobe_flush_task(prev); 2164 put_task_struct(prev); 2165 } 2166 2167 tick_nohz_task_switch(current); 2168} 2169 2170#ifdef CONFIG_SMP 2171 2172/* assumes rq->lock is held */ 2173static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 2174{ 2175 if (prev->sched_class->pre_schedule) 2176 prev->sched_class->pre_schedule(rq, prev); 2177} 2178 2179/* rq->lock is NOT held, but preemption is disabled */ 2180static inline void post_schedule(struct rq *rq) 2181{ 2182 if (rq->post_schedule) { 2183 unsigned long flags; 2184 2185 raw_spin_lock_irqsave(&rq->lock, flags); 2186 if (rq->curr->sched_class->post_schedule) 2187 rq->curr->sched_class->post_schedule(rq); 2188 raw_spin_unlock_irqrestore(&rq->lock, flags); 2189 2190 rq->post_schedule = 0; 2191 } 2192} 2193 2194#else 2195 2196static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2197{ 2198} 2199 2200static inline void post_schedule(struct rq *rq) 2201{ 2202} 2203 2204#endif 2205 2206/** 2207 * schedule_tail - first thing a freshly forked thread must call. 2208 * @prev: the thread we just switched away from. 2209 */ 2210asmlinkage void schedule_tail(struct task_struct *prev) 2211 __releases(rq->lock) 2212{ 2213 struct rq *rq = this_rq(); 2214 2215 finish_task_switch(rq, prev); 2216 2217 /* 2218 * FIXME: do we need to worry about rq being invalidated by the 2219 * task_switch? 2220 */ 2221 post_schedule(rq); 2222 2223#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2224 /* In this case, finish_task_switch does not reenable preemption */ 2225 preempt_enable(); 2226#endif 2227 if (current->set_child_tid) 2228 put_user(task_pid_vnr(current), current->set_child_tid); 2229} 2230 2231/* 2232 * context_switch - switch to the new MM and the new 2233 * thread's register state. 2234 */ 2235static inline void 2236context_switch(struct rq *rq, struct task_struct *prev, 2237 struct task_struct *next) 2238{ 2239 struct mm_struct *mm, *oldmm; 2240 2241 prepare_task_switch(rq, prev, next); 2242 2243 mm = next->mm; 2244 oldmm = prev->active_mm; 2245 /* 2246 * For paravirt, this is coupled with an exit in switch_to to 2247 * combine the page table reload and the switch backend into 2248 * one hypercall. 2249 */ 2250 arch_start_context_switch(prev); 2251 2252 if (!mm) { 2253 next->active_mm = oldmm; 2254 atomic_inc(&oldmm->mm_count); 2255 enter_lazy_tlb(oldmm, next); 2256 } else 2257 switch_mm(oldmm, mm, next); 2258 2259 if (!prev->mm) { 2260 prev->active_mm = NULL; 2261 rq->prev_mm = oldmm; 2262 } 2263 /* 2264 * Since the runqueue lock will be released by the next 2265 * task (which is an invalid locking op but in the case 2266 * of the scheduler it's an obvious special-case), so we 2267 * do an early lockdep release here: 2268 */ 2269#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2270 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2271#endif 2272 2273 context_tracking_task_switch(prev, next); 2274 /* Here we just switch the register state and the stack. */ 2275 switch_to(prev, next, prev); 2276 2277 barrier(); 2278 /* 2279 * this_rq must be evaluated again because prev may have moved 2280 * CPUs since it called schedule(), thus the 'rq' on its stack 2281 * frame will be invalid. 2282 */ 2283 finish_task_switch(this_rq(), prev); 2284} 2285 2286/* 2287 * nr_running and nr_context_switches: 2288 * 2289 * externally visible scheduler statistics: current number of runnable 2290 * threads, total number of context switches performed since bootup. 2291 */ 2292unsigned long nr_running(void) 2293{ 2294 unsigned long i, sum = 0; 2295 2296 for_each_online_cpu(i) 2297 sum += cpu_rq(i)->nr_running; 2298 2299 return sum; 2300} 2301 2302unsigned long long nr_context_switches(void) 2303{ 2304 int i; 2305 unsigned long long sum = 0; 2306 2307 for_each_possible_cpu(i) 2308 sum += cpu_rq(i)->nr_switches; 2309 2310 return sum; 2311} 2312 2313unsigned long nr_iowait(void) 2314{ 2315 unsigned long i, sum = 0; 2316 2317 for_each_possible_cpu(i) 2318 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2319 2320 return sum; 2321} 2322 2323unsigned long nr_iowait_cpu(int cpu) 2324{ 2325 struct rq *this = cpu_rq(cpu); 2326 return atomic_read(&this->nr_iowait); 2327} 2328 2329#ifdef CONFIG_SMP 2330 2331/* 2332 * sched_exec - execve() is a valuable balancing opportunity, because at 2333 * this point the task has the smallest effective memory and cache footprint. 2334 */ 2335void sched_exec(void) 2336{ 2337 struct task_struct *p = current; 2338 unsigned long flags; 2339 int dest_cpu; 2340 2341 raw_spin_lock_irqsave(&p->pi_lock, flags); 2342 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2343 if (dest_cpu == smp_processor_id()) 2344 goto unlock; 2345 2346 if (likely(cpu_active(dest_cpu))) { 2347 struct migration_arg arg = { p, dest_cpu }; 2348 2349 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2350 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2351 return; 2352 } 2353unlock: 2354 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2355} 2356 2357#endif 2358 2359DEFINE_PER_CPU(struct kernel_stat, kstat); 2360DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2361 2362EXPORT_PER_CPU_SYMBOL(kstat); 2363EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2364 2365/* 2366 * Return any ns on the sched_clock that have not yet been accounted in 2367 * @p in case that task is currently running. 2368 * 2369 * Called with task_rq_lock() held on @rq. 2370 */ 2371static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2372{ 2373 u64 ns = 0; 2374 2375 if (task_current(rq, p)) { 2376 update_rq_clock(rq); 2377 ns = rq_clock_task(rq) - p->se.exec_start; 2378 if ((s64)ns < 0) 2379 ns = 0; 2380 } 2381 2382 return ns; 2383} 2384 2385unsigned long long task_delta_exec(struct task_struct *p) 2386{ 2387 unsigned long flags; 2388 struct rq *rq; 2389 u64 ns = 0; 2390 2391 rq = task_rq_lock(p, &flags); 2392 ns = do_task_delta_exec(p, rq); 2393 task_rq_unlock(rq, p, &flags); 2394 2395 return ns; 2396} 2397 2398/* 2399 * Return accounted runtime for the task. 2400 * In case the task is currently running, return the runtime plus current's 2401 * pending runtime that have not been accounted yet. 2402 */ 2403unsigned long long task_sched_runtime(struct task_struct *p) 2404{ 2405 unsigned long flags; 2406 struct rq *rq; 2407 u64 ns = 0; 2408 2409#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2410 /* 2411 * 64-bit doesn't need locks to atomically read a 64bit value. 2412 * So we have a optimization chance when the task's delta_exec is 0. 2413 * Reading ->on_cpu is racy, but this is ok. 2414 * 2415 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2416 * If we race with it entering cpu, unaccounted time is 0. This is 2417 * indistinguishable from the read occurring a few cycles earlier. 2418 */ 2419 if (!p->on_cpu) 2420 return p->se.sum_exec_runtime; 2421#endif 2422 2423 rq = task_rq_lock(p, &flags); 2424 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2425 task_rq_unlock(rq, p, &flags); 2426 2427 return ns; 2428} 2429 2430/* 2431 * This function gets called by the timer code, with HZ frequency. 2432 * We call it with interrupts disabled. 2433 */ 2434void scheduler_tick(void) 2435{ 2436 int cpu = smp_processor_id(); 2437 struct rq *rq = cpu_rq(cpu); 2438 struct task_struct *curr = rq->curr; 2439 2440 sched_clock_tick(); 2441 2442 raw_spin_lock(&rq->lock); 2443 update_rq_clock(rq); 2444 curr->sched_class->task_tick(rq, curr, 0); 2445 update_cpu_load_active(rq); 2446 raw_spin_unlock(&rq->lock); 2447 2448 perf_event_task_tick(); 2449 2450#ifdef CONFIG_SMP 2451 rq->idle_balance = idle_cpu(cpu); 2452 trigger_load_balance(rq); 2453#endif 2454 rq_last_tick_reset(rq); 2455} 2456 2457#ifdef CONFIG_NO_HZ_FULL 2458/** 2459 * scheduler_tick_max_deferment 2460 * 2461 * Keep at least one tick per second when a single 2462 * active task is running because the scheduler doesn't 2463 * yet completely support full dynticks environment. 2464 * 2465 * This makes sure that uptime, CFS vruntime, load 2466 * balancing, etc... continue to move forward, even 2467 * with a very low granularity. 2468 * 2469 * Return: Maximum deferment in nanoseconds. 2470 */ 2471u64 scheduler_tick_max_deferment(void) 2472{ 2473 struct rq *rq = this_rq(); 2474 unsigned long next, now = ACCESS_ONCE(jiffies); 2475 2476 next = rq->last_sched_tick + HZ; 2477 2478 if (time_before_eq(next, now)) 2479 return 0; 2480 2481 return jiffies_to_nsecs(next - now); 2482} 2483#endif 2484 2485notrace unsigned long get_parent_ip(unsigned long addr) 2486{ 2487 if (in_lock_functions(addr)) { 2488 addr = CALLER_ADDR2; 2489 if (in_lock_functions(addr)) 2490 addr = CALLER_ADDR3; 2491 } 2492 return addr; 2493} 2494 2495#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2496 defined(CONFIG_PREEMPT_TRACER)) 2497 2498void __kprobes preempt_count_add(int val) 2499{ 2500#ifdef CONFIG_DEBUG_PREEMPT 2501 /* 2502 * Underflow? 2503 */ 2504 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2505 return; 2506#endif 2507 __preempt_count_add(val); 2508#ifdef CONFIG_DEBUG_PREEMPT 2509 /* 2510 * Spinlock count overflowing soon? 2511 */ 2512 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2513 PREEMPT_MASK - 10); 2514#endif 2515 if (preempt_count() == val) 2516 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2517} 2518EXPORT_SYMBOL(preempt_count_add); 2519 2520void __kprobes preempt_count_sub(int val) 2521{ 2522#ifdef CONFIG_DEBUG_PREEMPT 2523 /* 2524 * Underflow? 2525 */ 2526 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2527 return; 2528 /* 2529 * Is the spinlock portion underflowing? 2530 */ 2531 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2532 !(preempt_count() & PREEMPT_MASK))) 2533 return; 2534#endif 2535 2536 if (preempt_count() == val) 2537 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2538 __preempt_count_sub(val); 2539} 2540EXPORT_SYMBOL(preempt_count_sub); 2541 2542#endif 2543 2544/* 2545 * Print scheduling while atomic bug: 2546 */ 2547static noinline void __schedule_bug(struct task_struct *prev) 2548{ 2549 if (oops_in_progress) 2550 return; 2551 2552 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2553 prev->comm, prev->pid, preempt_count()); 2554 2555 debug_show_held_locks(prev); 2556 print_modules(); 2557 if (irqs_disabled()) 2558 print_irqtrace_events(prev); 2559 dump_stack(); 2560 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2561} 2562 2563/* 2564 * Various schedule()-time debugging checks and statistics: 2565 */ 2566static inline void schedule_debug(struct task_struct *prev) 2567{ 2568 /* 2569 * Test if we are atomic. Since do_exit() needs to call into 2570 * schedule() atomically, we ignore that path. Otherwise whine 2571 * if we are scheduling when we should not. 2572 */ 2573 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2574 __schedule_bug(prev); 2575 rcu_sleep_check(); 2576 2577 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2578 2579 schedstat_inc(this_rq(), sched_count); 2580} 2581 2582static void put_prev_task(struct rq *rq, struct task_struct *prev) 2583{ 2584 if (prev->on_rq || rq->skip_clock_update < 0) 2585 update_rq_clock(rq); 2586 prev->sched_class->put_prev_task(rq, prev); 2587} 2588 2589/* 2590 * Pick up the highest-prio task: 2591 */ 2592static inline struct task_struct * 2593pick_next_task(struct rq *rq) 2594{ 2595 const struct sched_class *class; 2596 struct task_struct *p; 2597 2598 /* 2599 * Optimization: we know that if all tasks are in 2600 * the fair class we can call that function directly: 2601 */ 2602 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2603 p = fair_sched_class.pick_next_task(rq); 2604 if (likely(p)) 2605 return p; 2606 } 2607 2608 for_each_class(class) { 2609 p = class->pick_next_task(rq); 2610 if (p) 2611 return p; 2612 } 2613 2614 BUG(); /* the idle class will always have a runnable task */ 2615} 2616 2617/* 2618 * __schedule() is the main scheduler function. 2619 * 2620 * The main means of driving the scheduler and thus entering this function are: 2621 * 2622 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2623 * 2624 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2625 * paths. For example, see arch/x86/entry_64.S. 2626 * 2627 * To drive preemption between tasks, the scheduler sets the flag in timer 2628 * interrupt handler scheduler_tick(). 2629 * 2630 * 3. Wakeups don't really cause entry into schedule(). They add a 2631 * task to the run-queue and that's it. 2632 * 2633 * Now, if the new task added to the run-queue preempts the current 2634 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2635 * called on the nearest possible occasion: 2636 * 2637 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2638 * 2639 * - in syscall or exception context, at the next outmost 2640 * preempt_enable(). (this might be as soon as the wake_up()'s 2641 * spin_unlock()!) 2642 * 2643 * - in IRQ context, return from interrupt-handler to 2644 * preemptible context 2645 * 2646 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2647 * then at the next: 2648 * 2649 * - cond_resched() call 2650 * - explicit schedule() call 2651 * - return from syscall or exception to user-space 2652 * - return from interrupt-handler to user-space 2653 */ 2654static void __sched __schedule(void) 2655{ 2656 struct task_struct *prev, *next; 2657 unsigned long *switch_count; 2658 struct rq *rq; 2659 int cpu; 2660 2661need_resched: 2662 preempt_disable(); 2663 cpu = smp_processor_id(); 2664 rq = cpu_rq(cpu); 2665 rcu_note_context_switch(cpu); 2666 prev = rq->curr; 2667 2668 schedule_debug(prev); 2669 2670 if (sched_feat(HRTICK)) 2671 hrtick_clear(rq); 2672 2673 /* 2674 * Make sure that signal_pending_state()->signal_pending() below 2675 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2676 * done by the caller to avoid the race with signal_wake_up(). 2677 */ 2678 smp_mb__before_spinlock(); 2679 raw_spin_lock_irq(&rq->lock); 2680 2681 switch_count = &prev->nivcsw; 2682 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2683 if (unlikely(signal_pending_state(prev->state, prev))) { 2684 prev->state = TASK_RUNNING; 2685 } else { 2686 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2687 prev->on_rq = 0; 2688 2689 /* 2690 * If a worker went to sleep, notify and ask workqueue 2691 * whether it wants to wake up a task to maintain 2692 * concurrency. 2693 */ 2694 if (prev->flags & PF_WQ_WORKER) { 2695 struct task_struct *to_wakeup; 2696 2697 to_wakeup = wq_worker_sleeping(prev, cpu); 2698 if (to_wakeup) 2699 try_to_wake_up_local(to_wakeup); 2700 } 2701 } 2702 switch_count = &prev->nvcsw; 2703 } 2704 2705 pre_schedule(rq, prev); 2706 2707 if (unlikely(!rq->nr_running)) { 2708 /* 2709 * We must set idle_stamp _before_ calling idle_balance(), such 2710 * that we measure the duration of idle_balance() as idle time. 2711 */ 2712 rq->idle_stamp = rq_clock(rq); 2713 if (idle_balance(rq)) 2714 rq->idle_stamp = 0; 2715 } 2716 2717 put_prev_task(rq, prev); 2718 next = pick_next_task(rq); 2719 clear_tsk_need_resched(prev); 2720 clear_preempt_need_resched(); 2721 rq->skip_clock_update = 0; 2722 2723 if (likely(prev != next)) { 2724 rq->nr_switches++; 2725 rq->curr = next; 2726 ++*switch_count; 2727 2728 context_switch(rq, prev, next); /* unlocks the rq */ 2729 /* 2730 * The context switch have flipped the stack from under us 2731 * and restored the local variables which were saved when 2732 * this task called schedule() in the past. prev == current 2733 * is still correct, but it can be moved to another cpu/rq. 2734 */ 2735 cpu = smp_processor_id(); 2736 rq = cpu_rq(cpu); 2737 } else 2738 raw_spin_unlock_irq(&rq->lock); 2739 2740 post_schedule(rq); 2741 2742 sched_preempt_enable_no_resched(); 2743 if (need_resched()) 2744 goto need_resched; 2745} 2746 2747static inline void sched_submit_work(struct task_struct *tsk) 2748{ 2749 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2750 return; 2751 /* 2752 * If we are going to sleep and we have plugged IO queued, 2753 * make sure to submit it to avoid deadlocks. 2754 */ 2755 if (blk_needs_flush_plug(tsk)) 2756 blk_schedule_flush_plug(tsk); 2757} 2758 2759asmlinkage void __sched schedule(void) 2760{ 2761 struct task_struct *tsk = current; 2762 2763 sched_submit_work(tsk); 2764 __schedule(); 2765} 2766EXPORT_SYMBOL(schedule); 2767 2768#ifdef CONFIG_CONTEXT_TRACKING 2769asmlinkage void __sched schedule_user(void) 2770{ 2771 /* 2772 * If we come here after a random call to set_need_resched(), 2773 * or we have been woken up remotely but the IPI has not yet arrived, 2774 * we haven't yet exited the RCU idle mode. Do it here manually until 2775 * we find a better solution. 2776 */ 2777 user_exit(); 2778 schedule(); 2779 user_enter(); 2780} 2781#endif 2782 2783/** 2784 * schedule_preempt_disabled - called with preemption disabled 2785 * 2786 * Returns with preemption disabled. Note: preempt_count must be 1 2787 */ 2788void __sched schedule_preempt_disabled(void) 2789{ 2790 sched_preempt_enable_no_resched(); 2791 schedule(); 2792 preempt_disable(); 2793} 2794 2795#ifdef CONFIG_PREEMPT 2796/* 2797 * this is the entry point to schedule() from in-kernel preemption 2798 * off of preempt_enable. Kernel preemptions off return from interrupt 2799 * occur there and call schedule directly. 2800 */ 2801asmlinkage void __sched notrace preempt_schedule(void) 2802{ 2803 /* 2804 * If there is a non-zero preempt_count or interrupts are disabled, 2805 * we do not want to preempt the current task. Just return.. 2806 */ 2807 if (likely(!preemptible())) 2808 return; 2809 2810 do { 2811 __preempt_count_add(PREEMPT_ACTIVE); 2812 __schedule(); 2813 __preempt_count_sub(PREEMPT_ACTIVE); 2814 2815 /* 2816 * Check again in case we missed a preemption opportunity 2817 * between schedule and now. 2818 */ 2819 barrier(); 2820 } while (need_resched()); 2821} 2822EXPORT_SYMBOL(preempt_schedule); 2823#endif /* CONFIG_PREEMPT */ 2824 2825/* 2826 * this is the entry point to schedule() from kernel preemption 2827 * off of irq context. 2828 * Note, that this is called and return with irqs disabled. This will 2829 * protect us against recursive calling from irq. 2830 */ 2831asmlinkage void __sched preempt_schedule_irq(void) 2832{ 2833 enum ctx_state prev_state; 2834 2835 /* Catch callers which need to be fixed */ 2836 BUG_ON(preempt_count() || !irqs_disabled()); 2837 2838 prev_state = exception_enter(); 2839 2840 do { 2841 __preempt_count_add(PREEMPT_ACTIVE); 2842 local_irq_enable(); 2843 __schedule(); 2844 local_irq_disable(); 2845 __preempt_count_sub(PREEMPT_ACTIVE); 2846 2847 /* 2848 * Check again in case we missed a preemption opportunity 2849 * between schedule and now. 2850 */ 2851 barrier(); 2852 } while (need_resched()); 2853 2854 exception_exit(prev_state); 2855} 2856 2857int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2858 void *key) 2859{ 2860 return try_to_wake_up(curr->private, mode, wake_flags); 2861} 2862EXPORT_SYMBOL(default_wake_function); 2863 2864static long __sched 2865sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2866{ 2867 unsigned long flags; 2868 wait_queue_t wait; 2869 2870 init_waitqueue_entry(&wait, current); 2871 2872 __set_current_state(state); 2873 2874 spin_lock_irqsave(&q->lock, flags); 2875 __add_wait_queue(q, &wait); 2876 spin_unlock(&q->lock); 2877 timeout = schedule_timeout(timeout); 2878 spin_lock_irq(&q->lock); 2879 __remove_wait_queue(q, &wait); 2880 spin_unlock_irqrestore(&q->lock, flags); 2881 2882 return timeout; 2883} 2884 2885void __sched interruptible_sleep_on(wait_queue_head_t *q) 2886{ 2887 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2888} 2889EXPORT_SYMBOL(interruptible_sleep_on); 2890 2891long __sched 2892interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2893{ 2894 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2895} 2896EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2897 2898void __sched sleep_on(wait_queue_head_t *q) 2899{ 2900 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2901} 2902EXPORT_SYMBOL(sleep_on); 2903 2904long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2905{ 2906 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2907} 2908EXPORT_SYMBOL(sleep_on_timeout); 2909 2910#ifdef CONFIG_RT_MUTEXES 2911 2912/* 2913 * rt_mutex_setprio - set the current priority of a task 2914 * @p: task 2915 * @prio: prio value (kernel-internal form) 2916 * 2917 * This function changes the 'effective' priority of a task. It does 2918 * not touch ->normal_prio like __setscheduler(). 2919 * 2920 * Used by the rt_mutex code to implement priority inheritance logic. 2921 */ 2922void rt_mutex_setprio(struct task_struct *p, int prio) 2923{ 2924 int oldprio, on_rq, running, enqueue_flag = 0; 2925 struct rq *rq; 2926 const struct sched_class *prev_class; 2927 2928 BUG_ON(prio > MAX_PRIO); 2929 2930 rq = __task_rq_lock(p); 2931 2932 /* 2933 * Idle task boosting is a nono in general. There is one 2934 * exception, when PREEMPT_RT and NOHZ is active: 2935 * 2936 * The idle task calls get_next_timer_interrupt() and holds 2937 * the timer wheel base->lock on the CPU and another CPU wants 2938 * to access the timer (probably to cancel it). We can safely 2939 * ignore the boosting request, as the idle CPU runs this code 2940 * with interrupts disabled and will complete the lock 2941 * protected section without being interrupted. So there is no 2942 * real need to boost. 2943 */ 2944 if (unlikely(p == rq->idle)) { 2945 WARN_ON(p != rq->curr); 2946 WARN_ON(p->pi_blocked_on); 2947 goto out_unlock; 2948 } 2949 2950 trace_sched_pi_setprio(p, prio); 2951 p->pi_top_task = rt_mutex_get_top_task(p); 2952 oldprio = p->prio; 2953 prev_class = p->sched_class; 2954 on_rq = p->on_rq; 2955 running = task_current(rq, p); 2956 if (on_rq) 2957 dequeue_task(rq, p, 0); 2958 if (running) 2959 p->sched_class->put_prev_task(rq, p); 2960 2961 /* 2962 * Boosting condition are: 2963 * 1. -rt task is running and holds mutex A 2964 * --> -dl task blocks on mutex A 2965 * 2966 * 2. -dl task is running and holds mutex A 2967 * --> -dl task blocks on mutex A and could preempt the 2968 * running task 2969 */ 2970 if (dl_prio(prio)) { 2971 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2972 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2973 p->dl.dl_boosted = 1; 2974 p->dl.dl_throttled = 0; 2975 enqueue_flag = ENQUEUE_REPLENISH; 2976 } else 2977 p->dl.dl_boosted = 0; 2978 p->sched_class = &dl_sched_class; 2979 } else if (rt_prio(prio)) { 2980 if (dl_prio(oldprio)) 2981 p->dl.dl_boosted = 0; 2982 if (oldprio < prio) 2983 enqueue_flag = ENQUEUE_HEAD; 2984 p->sched_class = &rt_sched_class; 2985 } else { 2986 if (dl_prio(oldprio)) 2987 p->dl.dl_boosted = 0; 2988 p->sched_class = &fair_sched_class; 2989 } 2990 2991 p->prio = prio; 2992 2993 if (running) 2994 p->sched_class->set_curr_task(rq); 2995 if (on_rq) 2996 enqueue_task(rq, p, enqueue_flag); 2997 2998 check_class_changed(rq, p, prev_class, oldprio); 2999out_unlock: 3000 __task_rq_unlock(rq); 3001} 3002#endif 3003 3004void set_user_nice(struct task_struct *p, long nice) 3005{ 3006 int old_prio, delta, on_rq; 3007 unsigned long flags; 3008 struct rq *rq; 3009 3010 if (task_nice(p) == nice || nice < -20 || nice > 19) 3011 return; 3012 /* 3013 * We have to be careful, if called from sys_setpriority(), 3014 * the task might be in the middle of scheduling on another CPU. 3015 */ 3016 rq = task_rq_lock(p, &flags); 3017 /* 3018 * The RT priorities are set via sched_setscheduler(), but we still 3019 * allow the 'normal' nice value to be set - but as expected 3020 * it wont have any effect on scheduling until the task is 3021 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3022 */ 3023 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3024 p->static_prio = NICE_TO_PRIO(nice); 3025 goto out_unlock; 3026 } 3027 on_rq = p->on_rq; 3028 if (on_rq) 3029 dequeue_task(rq, p, 0); 3030 3031 p->static_prio = NICE_TO_PRIO(nice); 3032 set_load_weight(p); 3033 old_prio = p->prio; 3034 p->prio = effective_prio(p); 3035 delta = p->prio - old_prio; 3036 3037 if (on_rq) { 3038 enqueue_task(rq, p, 0); 3039 /* 3040 * If the task increased its priority or is running and 3041 * lowered its priority, then reschedule its CPU: 3042 */ 3043 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3044 resched_task(rq->curr); 3045 } 3046out_unlock: 3047 task_rq_unlock(rq, p, &flags); 3048} 3049EXPORT_SYMBOL(set_user_nice); 3050 3051/* 3052 * can_nice - check if a task can reduce its nice value 3053 * @p: task 3054 * @nice: nice value 3055 */ 3056int can_nice(const struct task_struct *p, const int nice) 3057{ 3058 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3059 int nice_rlim = 20 - nice; 3060 3061 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3062 capable(CAP_SYS_NICE)); 3063} 3064 3065#ifdef __ARCH_WANT_SYS_NICE 3066 3067/* 3068 * sys_nice - change the priority of the current process. 3069 * @increment: priority increment 3070 * 3071 * sys_setpriority is a more generic, but much slower function that 3072 * does similar things. 3073 */ 3074SYSCALL_DEFINE1(nice, int, increment) 3075{ 3076 long nice, retval; 3077 3078 /* 3079 * Setpriority might change our priority at the same moment. 3080 * We don't have to worry. Conceptually one call occurs first 3081 * and we have a single winner. 3082 */ 3083 if (increment < -40) 3084 increment = -40; 3085 if (increment > 40) 3086 increment = 40; 3087 3088 nice = task_nice(current) + increment; 3089 if (nice < -20) 3090 nice = -20; 3091 if (nice > 19) 3092 nice = 19; 3093 3094 if (increment < 0 && !can_nice(current, nice)) 3095 return -EPERM; 3096 3097 retval = security_task_setnice(current, nice); 3098 if (retval) 3099 return retval; 3100 3101 set_user_nice(current, nice); 3102 return 0; 3103} 3104 3105#endif 3106 3107/** 3108 * task_prio - return the priority value of a given task. 3109 * @p: the task in question. 3110 * 3111 * Return: The priority value as seen by users in /proc. 3112 * RT tasks are offset by -200. Normal tasks are centered 3113 * around 0, value goes from -16 to +15. 3114 */ 3115int task_prio(const struct task_struct *p) 3116{ 3117 return p->prio - MAX_RT_PRIO; 3118} 3119 3120/** 3121 * idle_cpu - is a given cpu idle currently? 3122 * @cpu: the processor in question. 3123 * 3124 * Return: 1 if the CPU is currently idle. 0 otherwise. 3125 */ 3126int idle_cpu(int cpu) 3127{ 3128 struct rq *rq = cpu_rq(cpu); 3129 3130 if (rq->curr != rq->idle) 3131 return 0; 3132 3133 if (rq->nr_running) 3134 return 0; 3135 3136#ifdef CONFIG_SMP 3137 if (!llist_empty(&rq->wake_list)) 3138 return 0; 3139#endif 3140 3141 return 1; 3142} 3143 3144/** 3145 * idle_task - return the idle task for a given cpu. 3146 * @cpu: the processor in question. 3147 * 3148 * Return: The idle task for the cpu @cpu. 3149 */ 3150struct task_struct *idle_task(int cpu) 3151{ 3152 return cpu_rq(cpu)->idle; 3153} 3154 3155/** 3156 * find_process_by_pid - find a process with a matching PID value. 3157 * @pid: the pid in question. 3158 * 3159 * The task of @pid, if found. %NULL otherwise. 3160 */ 3161static struct task_struct *find_process_by_pid(pid_t pid) 3162{ 3163 return pid ? find_task_by_vpid(pid) : current; 3164} 3165 3166/* 3167 * This function initializes the sched_dl_entity of a newly becoming 3168 * SCHED_DEADLINE task. 3169 * 3170 * Only the static values are considered here, the actual runtime and the 3171 * absolute deadline will be properly calculated when the task is enqueued 3172 * for the first time with its new policy. 3173 */ 3174static void 3175__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3176{ 3177 struct sched_dl_entity *dl_se = &p->dl; 3178 3179 init_dl_task_timer(dl_se); 3180 dl_se->dl_runtime = attr->sched_runtime; 3181 dl_se->dl_deadline = attr->sched_deadline; 3182 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3183 dl_se->flags = attr->sched_flags; 3184 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3185 dl_se->dl_throttled = 0; 3186 dl_se->dl_new = 1; 3187} 3188 3189/* Actually do priority change: must hold pi & rq lock. */ 3190static void __setscheduler(struct rq *rq, struct task_struct *p, 3191 const struct sched_attr *attr) 3192{ 3193 int policy = attr->sched_policy; 3194 3195 if (policy == -1) /* setparam */ 3196 policy = p->policy; 3197 3198 p->policy = policy; 3199 3200 if (dl_policy(policy)) 3201 __setparam_dl(p, attr); 3202 else if (fair_policy(policy)) 3203 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3204 3205 /* 3206 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3207 * !rt_policy. Always setting this ensures that things like 3208 * getparam()/getattr() don't report silly values for !rt tasks. 3209 */ 3210 p->rt_priority = attr->sched_priority; 3211 3212 p->normal_prio = normal_prio(p); 3213 p->prio = rt_mutex_getprio(p); 3214 3215 if (dl_prio(p->prio)) 3216 p->sched_class = &dl_sched_class; 3217 else if (rt_prio(p->prio)) 3218 p->sched_class = &rt_sched_class; 3219 else 3220 p->sched_class = &fair_sched_class; 3221 3222 set_load_weight(p); 3223} 3224 3225static void 3226__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3227{ 3228 struct sched_dl_entity *dl_se = &p->dl; 3229 3230 attr->sched_priority = p->rt_priority; 3231 attr->sched_runtime = dl_se->dl_runtime; 3232 attr->sched_deadline = dl_se->dl_deadline; 3233 attr->sched_period = dl_se->dl_period; 3234 attr->sched_flags = dl_se->flags; 3235} 3236 3237/* 3238 * This function validates the new parameters of a -deadline task. 3239 * We ask for the deadline not being zero, and greater or equal 3240 * than the runtime, as well as the period of being zero or 3241 * greater than deadline. Furthermore, we have to be sure that 3242 * user parameters are above the internal resolution (1us); we 3243 * check sched_runtime only since it is always the smaller one. 3244 */ 3245static bool 3246__checkparam_dl(const struct sched_attr *attr) 3247{ 3248 return attr && attr->sched_deadline != 0 && 3249 (attr->sched_period == 0 || 3250 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3251 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3252 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3253} 3254 3255/* 3256 * check the target process has a UID that matches the current process's 3257 */ 3258static bool check_same_owner(struct task_struct *p) 3259{ 3260 const struct cred *cred = current_cred(), *pcred; 3261 bool match; 3262 3263 rcu_read_lock(); 3264 pcred = __task_cred(p); 3265 match = (uid_eq(cred->euid, pcred->euid) || 3266 uid_eq(cred->euid, pcred->uid)); 3267 rcu_read_unlock(); 3268 return match; 3269} 3270 3271static int __sched_setscheduler(struct task_struct *p, 3272 const struct sched_attr *attr, 3273 bool user) 3274{ 3275 int retval, oldprio, oldpolicy = -1, on_rq, running; 3276 int policy = attr->sched_policy; 3277 unsigned long flags; 3278 const struct sched_class *prev_class; 3279 struct rq *rq; 3280 int reset_on_fork; 3281 3282 /* may grab non-irq protected spin_locks */ 3283 BUG_ON(in_interrupt()); 3284recheck: 3285 /* double check policy once rq lock held */ 3286 if (policy < 0) { 3287 reset_on_fork = p->sched_reset_on_fork; 3288 policy = oldpolicy = p->policy; 3289 } else { 3290 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3291 3292 if (policy != SCHED_DEADLINE && 3293 policy != SCHED_FIFO && policy != SCHED_RR && 3294 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3295 policy != SCHED_IDLE) 3296 return -EINVAL; 3297 } 3298 3299 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3300 return -EINVAL; 3301 3302 /* 3303 * Valid priorities for SCHED_FIFO and SCHED_RR are 3304 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3305 * SCHED_BATCH and SCHED_IDLE is 0. 3306 */ 3307 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3308 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3309 return -EINVAL; 3310 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3311 (rt_policy(policy) != (attr->sched_priority != 0))) 3312 return -EINVAL; 3313 3314 /* 3315 * Allow unprivileged RT tasks to decrease priority: 3316 */ 3317 if (user && !capable(CAP_SYS_NICE)) { 3318 if (fair_policy(policy)) { 3319 if (attr->sched_nice < task_nice(p) && 3320 !can_nice(p, attr->sched_nice)) 3321 return -EPERM; 3322 } 3323 3324 if (rt_policy(policy)) { 3325 unsigned long rlim_rtprio = 3326 task_rlimit(p, RLIMIT_RTPRIO); 3327 3328 /* can't set/change the rt policy */ 3329 if (policy != p->policy && !rlim_rtprio) 3330 return -EPERM; 3331 3332 /* can't increase priority */ 3333 if (attr->sched_priority > p->rt_priority && 3334 attr->sched_priority > rlim_rtprio) 3335 return -EPERM; 3336 } 3337 3338 /* 3339 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3340 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3341 */ 3342 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3343 if (!can_nice(p, task_nice(p))) 3344 return -EPERM; 3345 } 3346 3347 /* can't change other user's priorities */ 3348 if (!check_same_owner(p)) 3349 return -EPERM; 3350 3351 /* Normal users shall not reset the sched_reset_on_fork flag */ 3352 if (p->sched_reset_on_fork && !reset_on_fork) 3353 return -EPERM; 3354 } 3355 3356 if (user) { 3357 retval = security_task_setscheduler(p); 3358 if (retval) 3359 return retval; 3360 } 3361 3362 /* 3363 * make sure no PI-waiters arrive (or leave) while we are 3364 * changing the priority of the task: 3365 * 3366 * To be able to change p->policy safely, the appropriate 3367 * runqueue lock must be held. 3368 */ 3369 rq = task_rq_lock(p, &flags); 3370 3371 /* 3372 * Changing the policy of the stop threads its a very bad idea 3373 */ 3374 if (p == rq->stop) { 3375 task_rq_unlock(rq, p, &flags); 3376 return -EINVAL; 3377 } 3378 3379 /* 3380 * If not changing anything there's no need to proceed further: 3381 */ 3382 if (unlikely(policy == p->policy)) { 3383 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3384 goto change; 3385 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3386 goto change; 3387 if (dl_policy(policy)) 3388 goto change; 3389 3390 task_rq_unlock(rq, p, &flags); 3391 return 0; 3392 } 3393change: 3394 3395 if (user) { 3396#ifdef CONFIG_RT_GROUP_SCHED 3397 /* 3398 * Do not allow realtime tasks into groups that have no runtime 3399 * assigned. 3400 */ 3401 if (rt_bandwidth_enabled() && rt_policy(policy) && 3402 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3403 !task_group_is_autogroup(task_group(p))) { 3404 task_rq_unlock(rq, p, &flags); 3405 return -EPERM; 3406 } 3407#endif 3408#ifdef CONFIG_SMP 3409 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3410 cpumask_t *span = rq->rd->span; 3411 3412 /* 3413 * Don't allow tasks with an affinity mask smaller than 3414 * the entire root_domain to become SCHED_DEADLINE. We 3415 * will also fail if there's no bandwidth available. 3416 */ 3417 if (!cpumask_subset(span, &p->cpus_allowed) || 3418 rq->rd->dl_bw.bw == 0) { 3419 task_rq_unlock(rq, p, &flags); 3420 return -EPERM; 3421 } 3422 } 3423#endif 3424 } 3425 3426 /* recheck policy now with rq lock held */ 3427 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3428 policy = oldpolicy = -1; 3429 task_rq_unlock(rq, p, &flags); 3430 goto recheck; 3431 } 3432 3433 /* 3434 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3435 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3436 * is available. 3437 */ 3438 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3439 task_rq_unlock(rq, p, &flags); 3440 return -EBUSY; 3441 } 3442 3443 on_rq = p->on_rq; 3444 running = task_current(rq, p); 3445 if (on_rq) 3446 dequeue_task(rq, p, 0); 3447 if (running) 3448 p->sched_class->put_prev_task(rq, p); 3449 3450 p->sched_reset_on_fork = reset_on_fork; 3451 3452 oldprio = p->prio; 3453 prev_class = p->sched_class; 3454 __setscheduler(rq, p, attr); 3455 3456 if (running) 3457 p->sched_class->set_curr_task(rq); 3458 if (on_rq) 3459 enqueue_task(rq, p, 0); 3460 3461 check_class_changed(rq, p, prev_class, oldprio); 3462 task_rq_unlock(rq, p, &flags); 3463 3464 rt_mutex_adjust_pi(p); 3465 3466 return 0; 3467} 3468 3469static int _sched_setscheduler(struct task_struct *p, int policy, 3470 const struct sched_param *param, bool check) 3471{ 3472 struct sched_attr attr = { 3473 .sched_policy = policy, 3474 .sched_priority = param->sched_priority, 3475 .sched_nice = PRIO_TO_NICE(p->static_prio), 3476 }; 3477 3478 /* 3479 * Fixup the legacy SCHED_RESET_ON_FORK hack 3480 */ 3481 if (policy & SCHED_RESET_ON_FORK) { 3482 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3483 policy &= ~SCHED_RESET_ON_FORK; 3484 attr.sched_policy = policy; 3485 } 3486 3487 return __sched_setscheduler(p, &attr, check); 3488} 3489/** 3490 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3491 * @p: the task in question. 3492 * @policy: new policy. 3493 * @param: structure containing the new RT priority. 3494 * 3495 * Return: 0 on success. An error code otherwise. 3496 * 3497 * NOTE that the task may be already dead. 3498 */ 3499int sched_setscheduler(struct task_struct *p, int policy, 3500 const struct sched_param *param) 3501{ 3502 return _sched_setscheduler(p, policy, param, true); 3503} 3504EXPORT_SYMBOL_GPL(sched_setscheduler); 3505 3506int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3507{ 3508 return __sched_setscheduler(p, attr, true); 3509} 3510EXPORT_SYMBOL_GPL(sched_setattr); 3511 3512/** 3513 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3514 * @p: the task in question. 3515 * @policy: new policy. 3516 * @param: structure containing the new RT priority. 3517 * 3518 * Just like sched_setscheduler, only don't bother checking if the 3519 * current context has permission. For example, this is needed in 3520 * stop_machine(): we create temporary high priority worker threads, 3521 * but our caller might not have that capability. 3522 * 3523 * Return: 0 on success. An error code otherwise. 3524 */ 3525int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3526 const struct sched_param *param) 3527{ 3528 return _sched_setscheduler(p, policy, param, false); 3529} 3530 3531static int 3532do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3533{ 3534 struct sched_param lparam; 3535 struct task_struct *p; 3536 int retval; 3537 3538 if (!param || pid < 0) 3539 return -EINVAL; 3540 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3541 return -EFAULT; 3542 3543 rcu_read_lock(); 3544 retval = -ESRCH; 3545 p = find_process_by_pid(pid); 3546 if (p != NULL) 3547 retval = sched_setscheduler(p, policy, &lparam); 3548 rcu_read_unlock(); 3549 3550 return retval; 3551} 3552 3553/* 3554 * Mimics kernel/events/core.c perf_copy_attr(). 3555 */ 3556static int sched_copy_attr(struct sched_attr __user *uattr, 3557 struct sched_attr *attr) 3558{ 3559 u32 size; 3560 int ret; 3561 3562 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3563 return -EFAULT; 3564 3565 /* 3566 * zero the full structure, so that a short copy will be nice. 3567 */ 3568 memset(attr, 0, sizeof(*attr)); 3569 3570 ret = get_user(size, &uattr->size); 3571 if (ret) 3572 return ret; 3573 3574 if (size > PAGE_SIZE) /* silly large */ 3575 goto err_size; 3576 3577 if (!size) /* abi compat */ 3578 size = SCHED_ATTR_SIZE_VER0; 3579 3580 if (size < SCHED_ATTR_SIZE_VER0) 3581 goto err_size; 3582 3583 /* 3584 * If we're handed a bigger struct than we know of, 3585 * ensure all the unknown bits are 0 - i.e. new 3586 * user-space does not rely on any kernel feature 3587 * extensions we dont know about yet. 3588 */ 3589 if (size > sizeof(*attr)) { 3590 unsigned char __user *addr; 3591 unsigned char __user *end; 3592 unsigned char val; 3593 3594 addr = (void __user *)uattr + sizeof(*attr); 3595 end = (void __user *)uattr + size; 3596 3597 for (; addr < end; addr++) { 3598 ret = get_user(val, addr); 3599 if (ret) 3600 return ret; 3601 if (val) 3602 goto err_size; 3603 } 3604 size = sizeof(*attr); 3605 } 3606 3607 ret = copy_from_user(attr, uattr, size); 3608 if (ret) 3609 return -EFAULT; 3610 3611 /* 3612 * XXX: do we want to be lenient like existing syscalls; or do we want 3613 * to be strict and return an error on out-of-bounds values? 3614 */ 3615 attr->sched_nice = clamp(attr->sched_nice, -20, 19); 3616 3617out: 3618 return ret; 3619 3620err_size: 3621 put_user(sizeof(*attr), &uattr->size); 3622 ret = -E2BIG; 3623 goto out; 3624} 3625 3626/** 3627 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3628 * @pid: the pid in question. 3629 * @policy: new policy. 3630 * @param: structure containing the new RT priority. 3631 * 3632 * Return: 0 on success. An error code otherwise. 3633 */ 3634SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3635 struct sched_param __user *, param) 3636{ 3637 /* negative values for policy are not valid */ 3638 if (policy < 0) 3639 return -EINVAL; 3640 3641 return do_sched_setscheduler(pid, policy, param); 3642} 3643 3644/** 3645 * sys_sched_setparam - set/change the RT priority of a thread 3646 * @pid: the pid in question. 3647 * @param: structure containing the new RT priority. 3648 * 3649 * Return: 0 on success. An error code otherwise. 3650 */ 3651SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3652{ 3653 return do_sched_setscheduler(pid, -1, param); 3654} 3655 3656/** 3657 * sys_sched_setattr - same as above, but with extended sched_attr 3658 * @pid: the pid in question. 3659 * @uattr: structure containing the extended parameters. 3660 */ 3661SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr) 3662{ 3663 struct sched_attr attr; 3664 struct task_struct *p; 3665 int retval; 3666 3667 if (!uattr || pid < 0) 3668 return -EINVAL; 3669 3670 if (sched_copy_attr(uattr, &attr)) 3671 return -EFAULT; 3672 3673 rcu_read_lock(); 3674 retval = -ESRCH; 3675 p = find_process_by_pid(pid); 3676 if (p != NULL) 3677 retval = sched_setattr(p, &attr); 3678 rcu_read_unlock(); 3679 3680 return retval; 3681} 3682 3683/** 3684 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3685 * @pid: the pid in question. 3686 * 3687 * Return: On success, the policy of the thread. Otherwise, a negative error 3688 * code. 3689 */ 3690SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3691{ 3692 struct task_struct *p; 3693 int retval; 3694 3695 if (pid < 0) 3696 return -EINVAL; 3697 3698 retval = -ESRCH; 3699 rcu_read_lock(); 3700 p = find_process_by_pid(pid); 3701 if (p) { 3702 retval = security_task_getscheduler(p); 3703 if (!retval) 3704 retval = p->policy 3705 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3706 } 3707 rcu_read_unlock(); 3708 return retval; 3709} 3710 3711/** 3712 * sys_sched_getparam - get the RT priority of a thread 3713 * @pid: the pid in question. 3714 * @param: structure containing the RT priority. 3715 * 3716 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3717 * code. 3718 */ 3719SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3720{ 3721 struct sched_param lp; 3722 struct task_struct *p; 3723 int retval; 3724 3725 if (!param || pid < 0) 3726 return -EINVAL; 3727 3728 rcu_read_lock(); 3729 p = find_process_by_pid(pid); 3730 retval = -ESRCH; 3731 if (!p) 3732 goto out_unlock; 3733 3734 retval = security_task_getscheduler(p); 3735 if (retval) 3736 goto out_unlock; 3737 3738 if (task_has_dl_policy(p)) { 3739 retval = -EINVAL; 3740 goto out_unlock; 3741 } 3742 lp.sched_priority = p->rt_priority; 3743 rcu_read_unlock(); 3744 3745 /* 3746 * This one might sleep, we cannot do it with a spinlock held ... 3747 */ 3748 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3749 3750 return retval; 3751 3752out_unlock: 3753 rcu_read_unlock(); 3754 return retval; 3755} 3756 3757static int sched_read_attr(struct sched_attr __user *uattr, 3758 struct sched_attr *attr, 3759 unsigned int usize) 3760{ 3761 int ret; 3762 3763 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3764 return -EFAULT; 3765 3766 /* 3767 * If we're handed a smaller struct than we know of, 3768 * ensure all the unknown bits are 0 - i.e. old 3769 * user-space does not get uncomplete information. 3770 */ 3771 if (usize < sizeof(*attr)) { 3772 unsigned char *addr; 3773 unsigned char *end; 3774 3775 addr = (void *)attr + usize; 3776 end = (void *)attr + sizeof(*attr); 3777 3778 for (; addr < end; addr++) { 3779 if (*addr) 3780 goto err_size; 3781 } 3782 3783 attr->size = usize; 3784 } 3785 3786 ret = copy_to_user(uattr, attr, usize); 3787 if (ret) 3788 return -EFAULT; 3789 3790out: 3791 return ret; 3792 3793err_size: 3794 ret = -E2BIG; 3795 goto out; 3796} 3797 3798/** 3799 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3800 * @pid: the pid in question. 3801 * @uattr: structure containing the extended parameters. 3802 * @size: sizeof(attr) for fwd/bwd comp. 3803 */ 3804SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3805 unsigned int, size) 3806{ 3807 struct sched_attr attr = { 3808 .size = sizeof(struct sched_attr), 3809 }; 3810 struct task_struct *p; 3811 int retval; 3812 3813 if (!uattr || pid < 0 || size > PAGE_SIZE || 3814 size < SCHED_ATTR_SIZE_VER0) 3815 return -EINVAL; 3816 3817 rcu_read_lock(); 3818 p = find_process_by_pid(pid); 3819 retval = -ESRCH; 3820 if (!p) 3821 goto out_unlock; 3822 3823 retval = security_task_getscheduler(p); 3824 if (retval) 3825 goto out_unlock; 3826 3827 attr.sched_policy = p->policy; 3828 if (p->sched_reset_on_fork) 3829 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3830 if (task_has_dl_policy(p)) 3831 __getparam_dl(p, &attr); 3832 else if (task_has_rt_policy(p)) 3833 attr.sched_priority = p->rt_priority; 3834 else 3835 attr.sched_nice = task_nice(p); 3836 3837 rcu_read_unlock(); 3838 3839 retval = sched_read_attr(uattr, &attr, size); 3840 return retval; 3841 3842out_unlock: 3843 rcu_read_unlock(); 3844 return retval; 3845} 3846 3847long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3848{ 3849 cpumask_var_t cpus_allowed, new_mask; 3850 struct task_struct *p; 3851 int retval; 3852 3853 rcu_read_lock(); 3854 3855 p = find_process_by_pid(pid); 3856 if (!p) { 3857 rcu_read_unlock(); 3858 return -ESRCH; 3859 } 3860 3861 /* Prevent p going away */ 3862 get_task_struct(p); 3863 rcu_read_unlock(); 3864 3865 if (p->flags & PF_NO_SETAFFINITY) { 3866 retval = -EINVAL; 3867 goto out_put_task; 3868 } 3869 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3870 retval = -ENOMEM; 3871 goto out_put_task; 3872 } 3873 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3874 retval = -ENOMEM; 3875 goto out_free_cpus_allowed; 3876 } 3877 retval = -EPERM; 3878 if (!check_same_owner(p)) { 3879 rcu_read_lock(); 3880 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3881 rcu_read_unlock(); 3882 goto out_unlock; 3883 } 3884 rcu_read_unlock(); 3885 } 3886 3887 retval = security_task_setscheduler(p); 3888 if (retval) 3889 goto out_unlock; 3890 3891 3892 cpuset_cpus_allowed(p, cpus_allowed); 3893 cpumask_and(new_mask, in_mask, cpus_allowed); 3894 3895 /* 3896 * Since bandwidth control happens on root_domain basis, 3897 * if admission test is enabled, we only admit -deadline 3898 * tasks allowed to run on all the CPUs in the task's 3899 * root_domain. 3900 */ 3901#ifdef CONFIG_SMP 3902 if (task_has_dl_policy(p)) { 3903 const struct cpumask *span = task_rq(p)->rd->span; 3904 3905 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3906 retval = -EBUSY; 3907 goto out_unlock; 3908 } 3909 } 3910#endif 3911again: 3912 retval = set_cpus_allowed_ptr(p, new_mask); 3913 3914 if (!retval) { 3915 cpuset_cpus_allowed(p, cpus_allowed); 3916 if (!cpumask_subset(new_mask, cpus_allowed)) { 3917 /* 3918 * We must have raced with a concurrent cpuset 3919 * update. Just reset the cpus_allowed to the 3920 * cpuset's cpus_allowed 3921 */ 3922 cpumask_copy(new_mask, cpus_allowed); 3923 goto again; 3924 } 3925 } 3926out_unlock: 3927 free_cpumask_var(new_mask); 3928out_free_cpus_allowed: 3929 free_cpumask_var(cpus_allowed); 3930out_put_task: 3931 put_task_struct(p); 3932 return retval; 3933} 3934 3935static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3936 struct cpumask *new_mask) 3937{ 3938 if (len < cpumask_size()) 3939 cpumask_clear(new_mask); 3940 else if (len > cpumask_size()) 3941 len = cpumask_size(); 3942 3943 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3944} 3945 3946/** 3947 * sys_sched_setaffinity - set the cpu affinity of a process 3948 * @pid: pid of the process 3949 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3950 * @user_mask_ptr: user-space pointer to the new cpu mask 3951 * 3952 * Return: 0 on success. An error code otherwise. 3953 */ 3954SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3955 unsigned long __user *, user_mask_ptr) 3956{ 3957 cpumask_var_t new_mask; 3958 int retval; 3959 3960 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3961 return -ENOMEM; 3962 3963 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3964 if (retval == 0) 3965 retval = sched_setaffinity(pid, new_mask); 3966 free_cpumask_var(new_mask); 3967 return retval; 3968} 3969 3970long sched_getaffinity(pid_t pid, struct cpumask *mask) 3971{ 3972 struct task_struct *p; 3973 unsigned long flags; 3974 int retval; 3975 3976 rcu_read_lock(); 3977 3978 retval = -ESRCH; 3979 p = find_process_by_pid(pid); 3980 if (!p) 3981 goto out_unlock; 3982 3983 retval = security_task_getscheduler(p); 3984 if (retval) 3985 goto out_unlock; 3986 3987 raw_spin_lock_irqsave(&p->pi_lock, flags); 3988 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3989 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3990 3991out_unlock: 3992 rcu_read_unlock(); 3993 3994 return retval; 3995} 3996 3997/** 3998 * sys_sched_getaffinity - get the cpu affinity of a process 3999 * @pid: pid of the process 4000 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4001 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4002 * 4003 * Return: 0 on success. An error code otherwise. 4004 */ 4005SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4006 unsigned long __user *, user_mask_ptr) 4007{ 4008 int ret; 4009 cpumask_var_t mask; 4010 4011 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4012 return -EINVAL; 4013 if (len & (sizeof(unsigned long)-1)) 4014 return -EINVAL; 4015 4016 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4017 return -ENOMEM; 4018 4019 ret = sched_getaffinity(pid, mask); 4020 if (ret == 0) { 4021 size_t retlen = min_t(size_t, len, cpumask_size()); 4022 4023 if (copy_to_user(user_mask_ptr, mask, retlen)) 4024 ret = -EFAULT; 4025 else 4026 ret = retlen; 4027 } 4028 free_cpumask_var(mask); 4029 4030 return ret; 4031} 4032 4033/** 4034 * sys_sched_yield - yield the current processor to other threads. 4035 * 4036 * This function yields the current CPU to other tasks. If there are no 4037 * other threads running on this CPU then this function will return. 4038 * 4039 * Return: 0. 4040 */ 4041SYSCALL_DEFINE0(sched_yield) 4042{ 4043 struct rq *rq = this_rq_lock(); 4044 4045 schedstat_inc(rq, yld_count); 4046 current->sched_class->yield_task(rq); 4047 4048 /* 4049 * Since we are going to call schedule() anyway, there's 4050 * no need to preempt or enable interrupts: 4051 */ 4052 __release(rq->lock); 4053 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4054 do_raw_spin_unlock(&rq->lock); 4055 sched_preempt_enable_no_resched(); 4056 4057 schedule(); 4058 4059 return 0; 4060} 4061 4062static void __cond_resched(void) 4063{ 4064 __preempt_count_add(PREEMPT_ACTIVE); 4065 __schedule(); 4066 __preempt_count_sub(PREEMPT_ACTIVE); 4067} 4068 4069int __sched _cond_resched(void) 4070{ 4071 if (should_resched()) { 4072 __cond_resched(); 4073 return 1; 4074 } 4075 return 0; 4076} 4077EXPORT_SYMBOL(_cond_resched); 4078 4079/* 4080 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4081 * call schedule, and on return reacquire the lock. 4082 * 4083 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4084 * operations here to prevent schedule() from being called twice (once via 4085 * spin_unlock(), once by hand). 4086 */ 4087int __cond_resched_lock(spinlock_t *lock) 4088{ 4089 int resched = should_resched(); 4090 int ret = 0; 4091 4092 lockdep_assert_held(lock); 4093 4094 if (spin_needbreak(lock) || resched) { 4095 spin_unlock(lock); 4096 if (resched) 4097 __cond_resched(); 4098 else 4099 cpu_relax(); 4100 ret = 1; 4101 spin_lock(lock); 4102 } 4103 return ret; 4104} 4105EXPORT_SYMBOL(__cond_resched_lock); 4106 4107int __sched __cond_resched_softirq(void) 4108{ 4109 BUG_ON(!in_softirq()); 4110 4111 if (should_resched()) { 4112 local_bh_enable(); 4113 __cond_resched(); 4114 local_bh_disable(); 4115 return 1; 4116 } 4117 return 0; 4118} 4119EXPORT_SYMBOL(__cond_resched_softirq); 4120 4121/** 4122 * yield - yield the current processor to other threads. 4123 * 4124 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4125 * 4126 * The scheduler is at all times free to pick the calling task as the most 4127 * eligible task to run, if removing the yield() call from your code breaks 4128 * it, its already broken. 4129 * 4130 * Typical broken usage is: 4131 * 4132 * while (!event) 4133 * yield(); 4134 * 4135 * where one assumes that yield() will let 'the other' process run that will 4136 * make event true. If the current task is a SCHED_FIFO task that will never 4137 * happen. Never use yield() as a progress guarantee!! 4138 * 4139 * If you want to use yield() to wait for something, use wait_event(). 4140 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4141 * If you still want to use yield(), do not! 4142 */ 4143void __sched yield(void) 4144{ 4145 set_current_state(TASK_RUNNING); 4146 sys_sched_yield(); 4147} 4148EXPORT_SYMBOL(yield); 4149 4150/** 4151 * yield_to - yield the current processor to another thread in 4152 * your thread group, or accelerate that thread toward the 4153 * processor it's on. 4154 * @p: target task 4155 * @preempt: whether task preemption is allowed or not 4156 * 4157 * It's the caller's job to ensure that the target task struct 4158 * can't go away on us before we can do any checks. 4159 * 4160 * Return: 4161 * true (>0) if we indeed boosted the target task. 4162 * false (0) if we failed to boost the target. 4163 * -ESRCH if there's no task to yield to. 4164 */ 4165bool __sched yield_to(struct task_struct *p, bool preempt) 4166{ 4167 struct task_struct *curr = current; 4168 struct rq *rq, *p_rq; 4169 unsigned long flags; 4170 int yielded = 0; 4171 4172 local_irq_save(flags); 4173 rq = this_rq(); 4174 4175again: 4176 p_rq = task_rq(p); 4177 /* 4178 * If we're the only runnable task on the rq and target rq also 4179 * has only one task, there's absolutely no point in yielding. 4180 */ 4181 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4182 yielded = -ESRCH; 4183 goto out_irq; 4184 } 4185 4186 double_rq_lock(rq, p_rq); 4187 if (task_rq(p) != p_rq) { 4188 double_rq_unlock(rq, p_rq); 4189 goto again; 4190 } 4191 4192 if (!curr->sched_class->yield_to_task) 4193 goto out_unlock; 4194 4195 if (curr->sched_class != p->sched_class) 4196 goto out_unlock; 4197 4198 if (task_running(p_rq, p) || p->state) 4199 goto out_unlock; 4200 4201 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4202 if (yielded) { 4203 schedstat_inc(rq, yld_count); 4204 /* 4205 * Make p's CPU reschedule; pick_next_entity takes care of 4206 * fairness. 4207 */ 4208 if (preempt && rq != p_rq) 4209 resched_task(p_rq->curr); 4210 } 4211 4212out_unlock: 4213 double_rq_unlock(rq, p_rq); 4214out_irq: 4215 local_irq_restore(flags); 4216 4217 if (yielded > 0) 4218 schedule(); 4219 4220 return yielded; 4221} 4222EXPORT_SYMBOL_GPL(yield_to); 4223 4224/* 4225 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4226 * that process accounting knows that this is a task in IO wait state. 4227 */ 4228void __sched io_schedule(void) 4229{ 4230 struct rq *rq = raw_rq(); 4231 4232 delayacct_blkio_start(); 4233 atomic_inc(&rq->nr_iowait); 4234 blk_flush_plug(current); 4235 current->in_iowait = 1; 4236 schedule(); 4237 current->in_iowait = 0; 4238 atomic_dec(&rq->nr_iowait); 4239 delayacct_blkio_end(); 4240} 4241EXPORT_SYMBOL(io_schedule); 4242 4243long __sched io_schedule_timeout(long timeout) 4244{ 4245 struct rq *rq = raw_rq(); 4246 long ret; 4247 4248 delayacct_blkio_start(); 4249 atomic_inc(&rq->nr_iowait); 4250 blk_flush_plug(current); 4251 current->in_iowait = 1; 4252 ret = schedule_timeout(timeout); 4253 current->in_iowait = 0; 4254 atomic_dec(&rq->nr_iowait); 4255 delayacct_blkio_end(); 4256 return ret; 4257} 4258 4259/** 4260 * sys_sched_get_priority_max - return maximum RT priority. 4261 * @policy: scheduling class. 4262 * 4263 * Return: On success, this syscall returns the maximum 4264 * rt_priority that can be used by a given scheduling class. 4265 * On failure, a negative error code is returned. 4266 */ 4267SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4268{ 4269 int ret = -EINVAL; 4270 4271 switch (policy) { 4272 case SCHED_FIFO: 4273 case SCHED_RR: 4274 ret = MAX_USER_RT_PRIO-1; 4275 break; 4276 case SCHED_DEADLINE: 4277 case SCHED_NORMAL: 4278 case SCHED_BATCH: 4279 case SCHED_IDLE: 4280 ret = 0; 4281 break; 4282 } 4283 return ret; 4284} 4285 4286/** 4287 * sys_sched_get_priority_min - return minimum RT priority. 4288 * @policy: scheduling class. 4289 * 4290 * Return: On success, this syscall returns the minimum 4291 * rt_priority that can be used by a given scheduling class. 4292 * On failure, a negative error code is returned. 4293 */ 4294SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4295{ 4296 int ret = -EINVAL; 4297 4298 switch (policy) { 4299 case SCHED_FIFO: 4300 case SCHED_RR: 4301 ret = 1; 4302 break; 4303 case SCHED_DEADLINE: 4304 case SCHED_NORMAL: 4305 case SCHED_BATCH: 4306 case SCHED_IDLE: 4307 ret = 0; 4308 } 4309 return ret; 4310} 4311 4312/** 4313 * sys_sched_rr_get_interval - return the default timeslice of a process. 4314 * @pid: pid of the process. 4315 * @interval: userspace pointer to the timeslice value. 4316 * 4317 * this syscall writes the default timeslice value of a given process 4318 * into the user-space timespec buffer. A value of '0' means infinity. 4319 * 4320 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4321 * an error code. 4322 */ 4323SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4324 struct timespec __user *, interval) 4325{ 4326 struct task_struct *p; 4327 unsigned int time_slice; 4328 unsigned long flags; 4329 struct rq *rq; 4330 int retval; 4331 struct timespec t; 4332 4333 if (pid < 0) 4334 return -EINVAL; 4335 4336 retval = -ESRCH; 4337 rcu_read_lock(); 4338 p = find_process_by_pid(pid); 4339 if (!p) 4340 goto out_unlock; 4341 4342 retval = security_task_getscheduler(p); 4343 if (retval) 4344 goto out_unlock; 4345 4346 rq = task_rq_lock(p, &flags); 4347 time_slice = 0; 4348 if (p->sched_class->get_rr_interval) 4349 time_slice = p->sched_class->get_rr_interval(rq, p); 4350 task_rq_unlock(rq, p, &flags); 4351 4352 rcu_read_unlock(); 4353 jiffies_to_timespec(time_slice, &t); 4354 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4355 return retval; 4356 4357out_unlock: 4358 rcu_read_unlock(); 4359 return retval; 4360} 4361 4362static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4363 4364void sched_show_task(struct task_struct *p) 4365{ 4366 unsigned long free = 0; 4367 int ppid; 4368 unsigned state; 4369 4370 state = p->state ? __ffs(p->state) + 1 : 0; 4371 printk(KERN_INFO "%-15.15s %c", p->comm, 4372 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4373#if BITS_PER_LONG == 32 4374 if (state == TASK_RUNNING) 4375 printk(KERN_CONT " running "); 4376 else 4377 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4378#else 4379 if (state == TASK_RUNNING) 4380 printk(KERN_CONT " running task "); 4381 else 4382 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4383#endif 4384#ifdef CONFIG_DEBUG_STACK_USAGE 4385 free = stack_not_used(p); 4386#endif 4387 rcu_read_lock(); 4388 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4389 rcu_read_unlock(); 4390 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4391 task_pid_nr(p), ppid, 4392 (unsigned long)task_thread_info(p)->flags); 4393 4394 print_worker_info(KERN_INFO, p); 4395 show_stack(p, NULL); 4396} 4397 4398void show_state_filter(unsigned long state_filter) 4399{ 4400 struct task_struct *g, *p; 4401 4402#if BITS_PER_LONG == 32 4403 printk(KERN_INFO 4404 " task PC stack pid father\n"); 4405#else 4406 printk(KERN_INFO 4407 " task PC stack pid father\n"); 4408#endif 4409 rcu_read_lock(); 4410 do_each_thread(g, p) { 4411 /* 4412 * reset the NMI-timeout, listing all files on a slow 4413 * console might take a lot of time: 4414 */ 4415 touch_nmi_watchdog(); 4416 if (!state_filter || (p->state & state_filter)) 4417 sched_show_task(p); 4418 } while_each_thread(g, p); 4419 4420 touch_all_softlockup_watchdogs(); 4421 4422#ifdef CONFIG_SCHED_DEBUG 4423 sysrq_sched_debug_show(); 4424#endif 4425 rcu_read_unlock(); 4426 /* 4427 * Only show locks if all tasks are dumped: 4428 */ 4429 if (!state_filter) 4430 debug_show_all_locks(); 4431} 4432 4433void init_idle_bootup_task(struct task_struct *idle) 4434{ 4435 idle->sched_class = &idle_sched_class; 4436} 4437 4438/** 4439 * init_idle - set up an idle thread for a given CPU 4440 * @idle: task in question 4441 * @cpu: cpu the idle task belongs to 4442 * 4443 * NOTE: this function does not set the idle thread's NEED_RESCHED 4444 * flag, to make booting more robust. 4445 */ 4446void init_idle(struct task_struct *idle, int cpu) 4447{ 4448 struct rq *rq = cpu_rq(cpu); 4449 unsigned long flags; 4450 4451 raw_spin_lock_irqsave(&rq->lock, flags); 4452 4453 __sched_fork(0, idle); 4454 idle->state = TASK_RUNNING; 4455 idle->se.exec_start = sched_clock(); 4456 4457 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4458 /* 4459 * We're having a chicken and egg problem, even though we are 4460 * holding rq->lock, the cpu isn't yet set to this cpu so the 4461 * lockdep check in task_group() will fail. 4462 * 4463 * Similar case to sched_fork(). / Alternatively we could 4464 * use task_rq_lock() here and obtain the other rq->lock. 4465 * 4466 * Silence PROVE_RCU 4467 */ 4468 rcu_read_lock(); 4469 __set_task_cpu(idle, cpu); 4470 rcu_read_unlock(); 4471 4472 rq->curr = rq->idle = idle; 4473#if defined(CONFIG_SMP) 4474 idle->on_cpu = 1; 4475#endif 4476 raw_spin_unlock_irqrestore(&rq->lock, flags); 4477 4478 /* Set the preempt count _outside_ the spinlocks! */ 4479 init_idle_preempt_count(idle, cpu); 4480 4481 /* 4482 * The idle tasks have their own, simple scheduling class: 4483 */ 4484 idle->sched_class = &idle_sched_class; 4485 ftrace_graph_init_idle_task(idle, cpu); 4486 vtime_init_idle(idle, cpu); 4487#if defined(CONFIG_SMP) 4488 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4489#endif 4490} 4491 4492#ifdef CONFIG_SMP 4493void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4494{ 4495 if (p->sched_class && p->sched_class->set_cpus_allowed) 4496 p->sched_class->set_cpus_allowed(p, new_mask); 4497 4498 cpumask_copy(&p->cpus_allowed, new_mask); 4499 p->nr_cpus_allowed = cpumask_weight(new_mask); 4500} 4501 4502/* 4503 * This is how migration works: 4504 * 4505 * 1) we invoke migration_cpu_stop() on the target CPU using 4506 * stop_one_cpu(). 4507 * 2) stopper starts to run (implicitly forcing the migrated thread 4508 * off the CPU) 4509 * 3) it checks whether the migrated task is still in the wrong runqueue. 4510 * 4) if it's in the wrong runqueue then the migration thread removes 4511 * it and puts it into the right queue. 4512 * 5) stopper completes and stop_one_cpu() returns and the migration 4513 * is done. 4514 */ 4515 4516/* 4517 * Change a given task's CPU affinity. Migrate the thread to a 4518 * proper CPU and schedule it away if the CPU it's executing on 4519 * is removed from the allowed bitmask. 4520 * 4521 * NOTE: the caller must have a valid reference to the task, the 4522 * task must not exit() & deallocate itself prematurely. The 4523 * call is not atomic; no spinlocks may be held. 4524 */ 4525int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4526{ 4527 unsigned long flags; 4528 struct rq *rq; 4529 unsigned int dest_cpu; 4530 int ret = 0; 4531 4532 rq = task_rq_lock(p, &flags); 4533 4534 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4535 goto out; 4536 4537 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4538 ret = -EINVAL; 4539 goto out; 4540 } 4541 4542 do_set_cpus_allowed(p, new_mask); 4543 4544 /* Can the task run on the task's current CPU? If so, we're done */ 4545 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4546 goto out; 4547 4548 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4549 if (p->on_rq) { 4550 struct migration_arg arg = { p, dest_cpu }; 4551 /* Need help from migration thread: drop lock and wait. */ 4552 task_rq_unlock(rq, p, &flags); 4553 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4554 tlb_migrate_finish(p->mm); 4555 return 0; 4556 } 4557out: 4558 task_rq_unlock(rq, p, &flags); 4559 4560 return ret; 4561} 4562EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4563 4564/* 4565 * Move (not current) task off this cpu, onto dest cpu. We're doing 4566 * this because either it can't run here any more (set_cpus_allowed() 4567 * away from this CPU, or CPU going down), or because we're 4568 * attempting to rebalance this task on exec (sched_exec). 4569 * 4570 * So we race with normal scheduler movements, but that's OK, as long 4571 * as the task is no longer on this CPU. 4572 * 4573 * Returns non-zero if task was successfully migrated. 4574 */ 4575static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4576{ 4577 struct rq *rq_dest, *rq_src; 4578 int ret = 0; 4579 4580 if (unlikely(!cpu_active(dest_cpu))) 4581 return ret; 4582 4583 rq_src = cpu_rq(src_cpu); 4584 rq_dest = cpu_rq(dest_cpu); 4585 4586 raw_spin_lock(&p->pi_lock); 4587 double_rq_lock(rq_src, rq_dest); 4588 /* Already moved. */ 4589 if (task_cpu(p) != src_cpu) 4590 goto done; 4591 /* Affinity changed (again). */ 4592 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4593 goto fail; 4594 4595 /* 4596 * If we're not on a rq, the next wake-up will ensure we're 4597 * placed properly. 4598 */ 4599 if (p->on_rq) { 4600 dequeue_task(rq_src, p, 0); 4601 set_task_cpu(p, dest_cpu); 4602 enqueue_task(rq_dest, p, 0); 4603 check_preempt_curr(rq_dest, p, 0); 4604 } 4605done: 4606 ret = 1; 4607fail: 4608 double_rq_unlock(rq_src, rq_dest); 4609 raw_spin_unlock(&p->pi_lock); 4610 return ret; 4611} 4612 4613#ifdef CONFIG_NUMA_BALANCING 4614/* Migrate current task p to target_cpu */ 4615int migrate_task_to(struct task_struct *p, int target_cpu) 4616{ 4617 struct migration_arg arg = { p, target_cpu }; 4618 int curr_cpu = task_cpu(p); 4619 4620 if (curr_cpu == target_cpu) 4621 return 0; 4622 4623 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4624 return -EINVAL; 4625 4626 /* TODO: This is not properly updating schedstats */ 4627 4628 trace_sched_move_numa(p, curr_cpu, target_cpu); 4629 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4630} 4631 4632/* 4633 * Requeue a task on a given node and accurately track the number of NUMA 4634 * tasks on the runqueues 4635 */ 4636void sched_setnuma(struct task_struct *p, int nid) 4637{ 4638 struct rq *rq; 4639 unsigned long flags; 4640 bool on_rq, running; 4641 4642 rq = task_rq_lock(p, &flags); 4643 on_rq = p->on_rq; 4644 running = task_current(rq, p); 4645 4646 if (on_rq) 4647 dequeue_task(rq, p, 0); 4648 if (running) 4649 p->sched_class->put_prev_task(rq, p); 4650 4651 p->numa_preferred_nid = nid; 4652 4653 if (running) 4654 p->sched_class->set_curr_task(rq); 4655 if (on_rq) 4656 enqueue_task(rq, p, 0); 4657 task_rq_unlock(rq, p, &flags); 4658} 4659#endif 4660 4661/* 4662 * migration_cpu_stop - this will be executed by a highprio stopper thread 4663 * and performs thread migration by bumping thread off CPU then 4664 * 'pushing' onto another runqueue. 4665 */ 4666static int migration_cpu_stop(void *data) 4667{ 4668 struct migration_arg *arg = data; 4669 4670 /* 4671 * The original target cpu might have gone down and we might 4672 * be on another cpu but it doesn't matter. 4673 */ 4674 local_irq_disable(); 4675 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4676 local_irq_enable(); 4677 return 0; 4678} 4679 4680#ifdef CONFIG_HOTPLUG_CPU 4681 4682/* 4683 * Ensures that the idle task is using init_mm right before its cpu goes 4684 * offline. 4685 */ 4686void idle_task_exit(void) 4687{ 4688 struct mm_struct *mm = current->active_mm; 4689 4690 BUG_ON(cpu_online(smp_processor_id())); 4691 4692 if (mm != &init_mm) 4693 switch_mm(mm, &init_mm, current); 4694 mmdrop(mm); 4695} 4696 4697/* 4698 * Since this CPU is going 'away' for a while, fold any nr_active delta 4699 * we might have. Assumes we're called after migrate_tasks() so that the 4700 * nr_active count is stable. 4701 * 4702 * Also see the comment "Global load-average calculations". 4703 */ 4704static void calc_load_migrate(struct rq *rq) 4705{ 4706 long delta = calc_load_fold_active(rq); 4707 if (delta) 4708 atomic_long_add(delta, &calc_load_tasks); 4709} 4710 4711/* 4712 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4713 * try_to_wake_up()->select_task_rq(). 4714 * 4715 * Called with rq->lock held even though we'er in stop_machine() and 4716 * there's no concurrency possible, we hold the required locks anyway 4717 * because of lock validation efforts. 4718 */ 4719static void migrate_tasks(unsigned int dead_cpu) 4720{ 4721 struct rq *rq = cpu_rq(dead_cpu); 4722 struct task_struct *next, *stop = rq->stop; 4723 int dest_cpu; 4724 4725 /* 4726 * Fudge the rq selection such that the below task selection loop 4727 * doesn't get stuck on the currently eligible stop task. 4728 * 4729 * We're currently inside stop_machine() and the rq is either stuck 4730 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4731 * either way we should never end up calling schedule() until we're 4732 * done here. 4733 */ 4734 rq->stop = NULL; 4735 4736 /* 4737 * put_prev_task() and pick_next_task() sched 4738 * class method both need to have an up-to-date 4739 * value of rq->clock[_task] 4740 */ 4741 update_rq_clock(rq); 4742 4743 for ( ; ; ) { 4744 /* 4745 * There's this thread running, bail when that's the only 4746 * remaining thread. 4747 */ 4748 if (rq->nr_running == 1) 4749 break; 4750 4751 next = pick_next_task(rq); 4752 BUG_ON(!next); 4753 next->sched_class->put_prev_task(rq, next); 4754 4755 /* Find suitable destination for @next, with force if needed. */ 4756 dest_cpu = select_fallback_rq(dead_cpu, next); 4757 raw_spin_unlock(&rq->lock); 4758 4759 __migrate_task(next, dead_cpu, dest_cpu); 4760 4761 raw_spin_lock(&rq->lock); 4762 } 4763 4764 rq->stop = stop; 4765} 4766 4767#endif /* CONFIG_HOTPLUG_CPU */ 4768 4769#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4770 4771static struct ctl_table sd_ctl_dir[] = { 4772 { 4773 .procname = "sched_domain", 4774 .mode = 0555, 4775 }, 4776 {} 4777}; 4778 4779static struct ctl_table sd_ctl_root[] = { 4780 { 4781 .procname = "kernel", 4782 .mode = 0555, 4783 .child = sd_ctl_dir, 4784 }, 4785 {} 4786}; 4787 4788static struct ctl_table *sd_alloc_ctl_entry(int n) 4789{ 4790 struct ctl_table *entry = 4791 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4792 4793 return entry; 4794} 4795 4796static void sd_free_ctl_entry(struct ctl_table **tablep) 4797{ 4798 struct ctl_table *entry; 4799 4800 /* 4801 * In the intermediate directories, both the child directory and 4802 * procname are dynamically allocated and could fail but the mode 4803 * will always be set. In the lowest directory the names are 4804 * static strings and all have proc handlers. 4805 */ 4806 for (entry = *tablep; entry->mode; entry++) { 4807 if (entry->child) 4808 sd_free_ctl_entry(&entry->child); 4809 if (entry->proc_handler == NULL) 4810 kfree(entry->procname); 4811 } 4812 4813 kfree(*tablep); 4814 *tablep = NULL; 4815} 4816 4817static int min_load_idx = 0; 4818static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4819 4820static void 4821set_table_entry(struct ctl_table *entry, 4822 const char *procname, void *data, int maxlen, 4823 umode_t mode, proc_handler *proc_handler, 4824 bool load_idx) 4825{ 4826 entry->procname = procname; 4827 entry->data = data; 4828 entry->maxlen = maxlen; 4829 entry->mode = mode; 4830 entry->proc_handler = proc_handler; 4831 4832 if (load_idx) { 4833 entry->extra1 = &min_load_idx; 4834 entry->extra2 = &max_load_idx; 4835 } 4836} 4837 4838static struct ctl_table * 4839sd_alloc_ctl_domain_table(struct sched_domain *sd) 4840{ 4841 struct ctl_table *table = sd_alloc_ctl_entry(13); 4842 4843 if (table == NULL) 4844 return NULL; 4845 4846 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4847 sizeof(long), 0644, proc_doulongvec_minmax, false); 4848 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4849 sizeof(long), 0644, proc_doulongvec_minmax, false); 4850 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4851 sizeof(int), 0644, proc_dointvec_minmax, true); 4852 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4853 sizeof(int), 0644, proc_dointvec_minmax, true); 4854 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4855 sizeof(int), 0644, proc_dointvec_minmax, true); 4856 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4857 sizeof(int), 0644, proc_dointvec_minmax, true); 4858 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4859 sizeof(int), 0644, proc_dointvec_minmax, true); 4860 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4861 sizeof(int), 0644, proc_dointvec_minmax, false); 4862 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4863 sizeof(int), 0644, proc_dointvec_minmax, false); 4864 set_table_entry(&table[9], "cache_nice_tries", 4865 &sd->cache_nice_tries, 4866 sizeof(int), 0644, proc_dointvec_minmax, false); 4867 set_table_entry(&table[10], "flags", &sd->flags, 4868 sizeof(int), 0644, proc_dointvec_minmax, false); 4869 set_table_entry(&table[11], "name", sd->name, 4870 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4871 /* &table[12] is terminator */ 4872 4873 return table; 4874} 4875 4876static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4877{ 4878 struct ctl_table *entry, *table; 4879 struct sched_domain *sd; 4880 int domain_num = 0, i; 4881 char buf[32]; 4882 4883 for_each_domain(cpu, sd) 4884 domain_num++; 4885 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4886 if (table == NULL) 4887 return NULL; 4888 4889 i = 0; 4890 for_each_domain(cpu, sd) { 4891 snprintf(buf, 32, "domain%d", i); 4892 entry->procname = kstrdup(buf, GFP_KERNEL); 4893 entry->mode = 0555; 4894 entry->child = sd_alloc_ctl_domain_table(sd); 4895 entry++; 4896 i++; 4897 } 4898 return table; 4899} 4900 4901static struct ctl_table_header *sd_sysctl_header; 4902static void register_sched_domain_sysctl(void) 4903{ 4904 int i, cpu_num = num_possible_cpus(); 4905 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4906 char buf[32]; 4907 4908 WARN_ON(sd_ctl_dir[0].child); 4909 sd_ctl_dir[0].child = entry; 4910 4911 if (entry == NULL) 4912 return; 4913 4914 for_each_possible_cpu(i) { 4915 snprintf(buf, 32, "cpu%d", i); 4916 entry->procname = kstrdup(buf, GFP_KERNEL); 4917 entry->mode = 0555; 4918 entry->child = sd_alloc_ctl_cpu_table(i); 4919 entry++; 4920 } 4921 4922 WARN_ON(sd_sysctl_header); 4923 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4924} 4925 4926/* may be called multiple times per register */ 4927static void unregister_sched_domain_sysctl(void) 4928{ 4929 if (sd_sysctl_header) 4930 unregister_sysctl_table(sd_sysctl_header); 4931 sd_sysctl_header = NULL; 4932 if (sd_ctl_dir[0].child) 4933 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4934} 4935#else 4936static void register_sched_domain_sysctl(void) 4937{ 4938} 4939static void unregister_sched_domain_sysctl(void) 4940{ 4941} 4942#endif 4943 4944static void set_rq_online(struct rq *rq) 4945{ 4946 if (!rq->online) { 4947 const struct sched_class *class; 4948 4949 cpumask_set_cpu(rq->cpu, rq->rd->online); 4950 rq->online = 1; 4951 4952 for_each_class(class) { 4953 if (class->rq_online) 4954 class->rq_online(rq); 4955 } 4956 } 4957} 4958 4959static void set_rq_offline(struct rq *rq) 4960{ 4961 if (rq->online) { 4962 const struct sched_class *class; 4963 4964 for_each_class(class) { 4965 if (class->rq_offline) 4966 class->rq_offline(rq); 4967 } 4968 4969 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4970 rq->online = 0; 4971 } 4972} 4973 4974/* 4975 * migration_call - callback that gets triggered when a CPU is added. 4976 * Here we can start up the necessary migration thread for the new CPU. 4977 */ 4978static int 4979migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4980{ 4981 int cpu = (long)hcpu; 4982 unsigned long flags; 4983 struct rq *rq = cpu_rq(cpu); 4984 4985 switch (action & ~CPU_TASKS_FROZEN) { 4986 4987 case CPU_UP_PREPARE: 4988 rq->calc_load_update = calc_load_update; 4989 break; 4990 4991 case CPU_ONLINE: 4992 /* Update our root-domain */ 4993 raw_spin_lock_irqsave(&rq->lock, flags); 4994 if (rq->rd) { 4995 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4996 4997 set_rq_online(rq); 4998 } 4999 raw_spin_unlock_irqrestore(&rq->lock, flags); 5000 break; 5001 5002#ifdef CONFIG_HOTPLUG_CPU 5003 case CPU_DYING: 5004 sched_ttwu_pending(); 5005 /* Update our root-domain */ 5006 raw_spin_lock_irqsave(&rq->lock, flags); 5007 if (rq->rd) { 5008 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5009 set_rq_offline(rq); 5010 } 5011 migrate_tasks(cpu); 5012 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5013 raw_spin_unlock_irqrestore(&rq->lock, flags); 5014 break; 5015 5016 case CPU_DEAD: 5017 calc_load_migrate(rq); 5018 break; 5019#endif 5020 } 5021 5022 update_max_interval(); 5023 5024 return NOTIFY_OK; 5025} 5026 5027/* 5028 * Register at high priority so that task migration (migrate_all_tasks) 5029 * happens before everything else. This has to be lower priority than 5030 * the notifier in the perf_event subsystem, though. 5031 */ 5032static struct notifier_block migration_notifier = { 5033 .notifier_call = migration_call, 5034 .priority = CPU_PRI_MIGRATION, 5035}; 5036 5037static int sched_cpu_active(struct notifier_block *nfb, 5038 unsigned long action, void *hcpu) 5039{ 5040 switch (action & ~CPU_TASKS_FROZEN) { 5041 case CPU_STARTING: 5042 case CPU_DOWN_FAILED: 5043 set_cpu_active((long)hcpu, true); 5044 return NOTIFY_OK; 5045 default: 5046 return NOTIFY_DONE; 5047 } 5048} 5049 5050static int sched_cpu_inactive(struct notifier_block *nfb, 5051 unsigned long action, void *hcpu) 5052{ 5053 unsigned long flags; 5054 long cpu = (long)hcpu; 5055 5056 switch (action & ~CPU_TASKS_FROZEN) { 5057 case CPU_DOWN_PREPARE: 5058 set_cpu_active(cpu, false); 5059 5060 /* explicitly allow suspend */ 5061 if (!(action & CPU_TASKS_FROZEN)) { 5062 struct dl_bw *dl_b = dl_bw_of(cpu); 5063 bool overflow; 5064 int cpus; 5065 5066 raw_spin_lock_irqsave(&dl_b->lock, flags); 5067 cpus = dl_bw_cpus(cpu); 5068 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5069 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5070 5071 if (overflow) 5072 return notifier_from_errno(-EBUSY); 5073 } 5074 return NOTIFY_OK; 5075 } 5076 5077 return NOTIFY_DONE; 5078} 5079 5080static int __init migration_init(void) 5081{ 5082 void *cpu = (void *)(long)smp_processor_id(); 5083 int err; 5084 5085 /* Initialize migration for the boot CPU */ 5086 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5087 BUG_ON(err == NOTIFY_BAD); 5088 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5089 register_cpu_notifier(&migration_notifier); 5090 5091 /* Register cpu active notifiers */ 5092 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5093 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5094 5095 return 0; 5096} 5097early_initcall(migration_init); 5098#endif 5099 5100#ifdef CONFIG_SMP 5101 5102static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5103 5104#ifdef CONFIG_SCHED_DEBUG 5105 5106static __read_mostly int sched_debug_enabled; 5107 5108static int __init sched_debug_setup(char *str) 5109{ 5110 sched_debug_enabled = 1; 5111 5112 return 0; 5113} 5114early_param("sched_debug", sched_debug_setup); 5115 5116static inline bool sched_debug(void) 5117{ 5118 return sched_debug_enabled; 5119} 5120 5121static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5122 struct cpumask *groupmask) 5123{ 5124 struct sched_group *group = sd->groups; 5125 char str[256]; 5126 5127 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5128 cpumask_clear(groupmask); 5129 5130 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5131 5132 if (!(sd->flags & SD_LOAD_BALANCE)) { 5133 printk("does not load-balance\n"); 5134 if (sd->parent) 5135 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5136 " has parent"); 5137 return -1; 5138 } 5139 5140 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5141 5142 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5143 printk(KERN_ERR "ERROR: domain->span does not contain " 5144 "CPU%d\n", cpu); 5145 } 5146 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5147 printk(KERN_ERR "ERROR: domain->groups does not contain" 5148 " CPU%d\n", cpu); 5149 } 5150 5151 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5152 do { 5153 if (!group) { 5154 printk("\n"); 5155 printk(KERN_ERR "ERROR: group is NULL\n"); 5156 break; 5157 } 5158 5159 /* 5160 * Even though we initialize ->power to something semi-sane, 5161 * we leave power_orig unset. This allows us to detect if 5162 * domain iteration is still funny without causing /0 traps. 5163 */ 5164 if (!group->sgp->power_orig) { 5165 printk(KERN_CONT "\n"); 5166 printk(KERN_ERR "ERROR: domain->cpu_power not " 5167 "set\n"); 5168 break; 5169 } 5170 5171 if (!cpumask_weight(sched_group_cpus(group))) { 5172 printk(KERN_CONT "\n"); 5173 printk(KERN_ERR "ERROR: empty group\n"); 5174 break; 5175 } 5176 5177 if (!(sd->flags & SD_OVERLAP) && 5178 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5179 printk(KERN_CONT "\n"); 5180 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5181 break; 5182 } 5183 5184 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5185 5186 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5187 5188 printk(KERN_CONT " %s", str); 5189 if (group->sgp->power != SCHED_POWER_SCALE) { 5190 printk(KERN_CONT " (cpu_power = %d)", 5191 group->sgp->power); 5192 } 5193 5194 group = group->next; 5195 } while (group != sd->groups); 5196 printk(KERN_CONT "\n"); 5197 5198 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5199 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5200 5201 if (sd->parent && 5202 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5203 printk(KERN_ERR "ERROR: parent span is not a superset " 5204 "of domain->span\n"); 5205 return 0; 5206} 5207 5208static void sched_domain_debug(struct sched_domain *sd, int cpu) 5209{ 5210 int level = 0; 5211 5212 if (!sched_debug_enabled) 5213 return; 5214 5215 if (!sd) { 5216 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5217 return; 5218 } 5219 5220 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5221 5222 for (;;) { 5223 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5224 break; 5225 level++; 5226 sd = sd->parent; 5227 if (!sd) 5228 break; 5229 } 5230} 5231#else /* !CONFIG_SCHED_DEBUG */ 5232# define sched_domain_debug(sd, cpu) do { } while (0) 5233static inline bool sched_debug(void) 5234{ 5235 return false; 5236} 5237#endif /* CONFIG_SCHED_DEBUG */ 5238 5239static int sd_degenerate(struct sched_domain *sd) 5240{ 5241 if (cpumask_weight(sched_domain_span(sd)) == 1) 5242 return 1; 5243 5244 /* Following flags need at least 2 groups */ 5245 if (sd->flags & (SD_LOAD_BALANCE | 5246 SD_BALANCE_NEWIDLE | 5247 SD_BALANCE_FORK | 5248 SD_BALANCE_EXEC | 5249 SD_SHARE_CPUPOWER | 5250 SD_SHARE_PKG_RESOURCES)) { 5251 if (sd->groups != sd->groups->next) 5252 return 0; 5253 } 5254 5255 /* Following flags don't use groups */ 5256 if (sd->flags & (SD_WAKE_AFFINE)) 5257 return 0; 5258 5259 return 1; 5260} 5261 5262static int 5263sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5264{ 5265 unsigned long cflags = sd->flags, pflags = parent->flags; 5266 5267 if (sd_degenerate(parent)) 5268 return 1; 5269 5270 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5271 return 0; 5272 5273 /* Flags needing groups don't count if only 1 group in parent */ 5274 if (parent->groups == parent->groups->next) { 5275 pflags &= ~(SD_LOAD_BALANCE | 5276 SD_BALANCE_NEWIDLE | 5277 SD_BALANCE_FORK | 5278 SD_BALANCE_EXEC | 5279 SD_SHARE_CPUPOWER | 5280 SD_SHARE_PKG_RESOURCES | 5281 SD_PREFER_SIBLING); 5282 if (nr_node_ids == 1) 5283 pflags &= ~SD_SERIALIZE; 5284 } 5285 if (~cflags & pflags) 5286 return 0; 5287 5288 return 1; 5289} 5290 5291static void free_rootdomain(struct rcu_head *rcu) 5292{ 5293 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5294 5295 cpupri_cleanup(&rd->cpupri); 5296 cpudl_cleanup(&rd->cpudl); 5297 free_cpumask_var(rd->dlo_mask); 5298 free_cpumask_var(rd->rto_mask); 5299 free_cpumask_var(rd->online); 5300 free_cpumask_var(rd->span); 5301 kfree(rd); 5302} 5303 5304static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5305{ 5306 struct root_domain *old_rd = NULL; 5307 unsigned long flags; 5308 5309 raw_spin_lock_irqsave(&rq->lock, flags); 5310 5311 if (rq->rd) { 5312 old_rd = rq->rd; 5313 5314 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5315 set_rq_offline(rq); 5316 5317 cpumask_clear_cpu(rq->cpu, old_rd->span); 5318 5319 /* 5320 * If we dont want to free the old_rd yet then 5321 * set old_rd to NULL to skip the freeing later 5322 * in this function: 5323 */ 5324 if (!atomic_dec_and_test(&old_rd->refcount)) 5325 old_rd = NULL; 5326 } 5327 5328 atomic_inc(&rd->refcount); 5329 rq->rd = rd; 5330 5331 cpumask_set_cpu(rq->cpu, rd->span); 5332 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5333 set_rq_online(rq); 5334 5335 raw_spin_unlock_irqrestore(&rq->lock, flags); 5336 5337 if (old_rd) 5338 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5339} 5340 5341static int init_rootdomain(struct root_domain *rd) 5342{ 5343 memset(rd, 0, sizeof(*rd)); 5344 5345 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5346 goto out; 5347 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5348 goto free_span; 5349 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5350 goto free_online; 5351 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5352 goto free_dlo_mask; 5353 5354 init_dl_bw(&rd->dl_bw); 5355 if (cpudl_init(&rd->cpudl) != 0) 5356 goto free_dlo_mask; 5357 5358 if (cpupri_init(&rd->cpupri) != 0) 5359 goto free_rto_mask; 5360 return 0; 5361 5362free_rto_mask: 5363 free_cpumask_var(rd->rto_mask); 5364free_dlo_mask: 5365 free_cpumask_var(rd->dlo_mask); 5366free_online: 5367 free_cpumask_var(rd->online); 5368free_span: 5369 free_cpumask_var(rd->span); 5370out: 5371 return -ENOMEM; 5372} 5373 5374/* 5375 * By default the system creates a single root-domain with all cpus as 5376 * members (mimicking the global state we have today). 5377 */ 5378struct root_domain def_root_domain; 5379 5380static void init_defrootdomain(void) 5381{ 5382 init_rootdomain(&def_root_domain); 5383 5384 atomic_set(&def_root_domain.refcount, 1); 5385} 5386 5387static struct root_domain *alloc_rootdomain(void) 5388{ 5389 struct root_domain *rd; 5390 5391 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5392 if (!rd) 5393 return NULL; 5394 5395 if (init_rootdomain(rd) != 0) { 5396 kfree(rd); 5397 return NULL; 5398 } 5399 5400 return rd; 5401} 5402 5403static void free_sched_groups(struct sched_group *sg, int free_sgp) 5404{ 5405 struct sched_group *tmp, *first; 5406 5407 if (!sg) 5408 return; 5409 5410 first = sg; 5411 do { 5412 tmp = sg->next; 5413 5414 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5415 kfree(sg->sgp); 5416 5417 kfree(sg); 5418 sg = tmp; 5419 } while (sg != first); 5420} 5421 5422static void free_sched_domain(struct rcu_head *rcu) 5423{ 5424 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5425 5426 /* 5427 * If its an overlapping domain it has private groups, iterate and 5428 * nuke them all. 5429 */ 5430 if (sd->flags & SD_OVERLAP) { 5431 free_sched_groups(sd->groups, 1); 5432 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5433 kfree(sd->groups->sgp); 5434 kfree(sd->groups); 5435 } 5436 kfree(sd); 5437} 5438 5439static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5440{ 5441 call_rcu(&sd->rcu, free_sched_domain); 5442} 5443 5444static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5445{ 5446 for (; sd; sd = sd->parent) 5447 destroy_sched_domain(sd, cpu); 5448} 5449 5450/* 5451 * Keep a special pointer to the highest sched_domain that has 5452 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5453 * allows us to avoid some pointer chasing select_idle_sibling(). 5454 * 5455 * Also keep a unique ID per domain (we use the first cpu number in 5456 * the cpumask of the domain), this allows us to quickly tell if 5457 * two cpus are in the same cache domain, see cpus_share_cache(). 5458 */ 5459DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5460DEFINE_PER_CPU(int, sd_llc_size); 5461DEFINE_PER_CPU(int, sd_llc_id); 5462DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5463DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5464DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5465 5466static void update_top_cache_domain(int cpu) 5467{ 5468 struct sched_domain *sd; 5469 struct sched_domain *busy_sd = NULL; 5470 int id = cpu; 5471 int size = 1; 5472 5473 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5474 if (sd) { 5475 id = cpumask_first(sched_domain_span(sd)); 5476 size = cpumask_weight(sched_domain_span(sd)); 5477 busy_sd = sd->parent; /* sd_busy */ 5478 } 5479 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5480 5481 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5482 per_cpu(sd_llc_size, cpu) = size; 5483 per_cpu(sd_llc_id, cpu) = id; 5484 5485 sd = lowest_flag_domain(cpu, SD_NUMA); 5486 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5487 5488 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5489 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5490} 5491 5492/* 5493 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5494 * hold the hotplug lock. 5495 */ 5496static void 5497cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5498{ 5499 struct rq *rq = cpu_rq(cpu); 5500 struct sched_domain *tmp; 5501 5502 /* Remove the sched domains which do not contribute to scheduling. */ 5503 for (tmp = sd; tmp; ) { 5504 struct sched_domain *parent = tmp->parent; 5505 if (!parent) 5506 break; 5507 5508 if (sd_parent_degenerate(tmp, parent)) { 5509 tmp->parent = parent->parent; 5510 if (parent->parent) 5511 parent->parent->child = tmp; 5512 /* 5513 * Transfer SD_PREFER_SIBLING down in case of a 5514 * degenerate parent; the spans match for this 5515 * so the property transfers. 5516 */ 5517 if (parent->flags & SD_PREFER_SIBLING) 5518 tmp->flags |= SD_PREFER_SIBLING; 5519 destroy_sched_domain(parent, cpu); 5520 } else 5521 tmp = tmp->parent; 5522 } 5523 5524 if (sd && sd_degenerate(sd)) { 5525 tmp = sd; 5526 sd = sd->parent; 5527 destroy_sched_domain(tmp, cpu); 5528 if (sd) 5529 sd->child = NULL; 5530 } 5531 5532 sched_domain_debug(sd, cpu); 5533 5534 rq_attach_root(rq, rd); 5535 tmp = rq->sd; 5536 rcu_assign_pointer(rq->sd, sd); 5537 destroy_sched_domains(tmp, cpu); 5538 5539 update_top_cache_domain(cpu); 5540} 5541 5542/* cpus with isolated domains */ 5543static cpumask_var_t cpu_isolated_map; 5544 5545/* Setup the mask of cpus configured for isolated domains */ 5546static int __init isolated_cpu_setup(char *str) 5547{ 5548 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5549 cpulist_parse(str, cpu_isolated_map); 5550 return 1; 5551} 5552 5553__setup("isolcpus=", isolated_cpu_setup); 5554 5555static const struct cpumask *cpu_cpu_mask(int cpu) 5556{ 5557 return cpumask_of_node(cpu_to_node(cpu)); 5558} 5559 5560struct sd_data { 5561 struct sched_domain **__percpu sd; 5562 struct sched_group **__percpu sg; 5563 struct sched_group_power **__percpu sgp; 5564}; 5565 5566struct s_data { 5567 struct sched_domain ** __percpu sd; 5568 struct root_domain *rd; 5569}; 5570 5571enum s_alloc { 5572 sa_rootdomain, 5573 sa_sd, 5574 sa_sd_storage, 5575 sa_none, 5576}; 5577 5578struct sched_domain_topology_level; 5579 5580typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5581typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5582 5583#define SDTL_OVERLAP 0x01 5584 5585struct sched_domain_topology_level { 5586 sched_domain_init_f init; 5587 sched_domain_mask_f mask; 5588 int flags; 5589 int numa_level; 5590 struct sd_data data; 5591}; 5592 5593/* 5594 * Build an iteration mask that can exclude certain CPUs from the upwards 5595 * domain traversal. 5596 * 5597 * Asymmetric node setups can result in situations where the domain tree is of 5598 * unequal depth, make sure to skip domains that already cover the entire 5599 * range. 5600 * 5601 * In that case build_sched_domains() will have terminated the iteration early 5602 * and our sibling sd spans will be empty. Domains should always include the 5603 * cpu they're built on, so check that. 5604 * 5605 */ 5606static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5607{ 5608 const struct cpumask *span = sched_domain_span(sd); 5609 struct sd_data *sdd = sd->private; 5610 struct sched_domain *sibling; 5611 int i; 5612 5613 for_each_cpu(i, span) { 5614 sibling = *per_cpu_ptr(sdd->sd, i); 5615 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5616 continue; 5617 5618 cpumask_set_cpu(i, sched_group_mask(sg)); 5619 } 5620} 5621 5622/* 5623 * Return the canonical balance cpu for this group, this is the first cpu 5624 * of this group that's also in the iteration mask. 5625 */ 5626int group_balance_cpu(struct sched_group *sg) 5627{ 5628 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5629} 5630 5631static int 5632build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5633{ 5634 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5635 const struct cpumask *span = sched_domain_span(sd); 5636 struct cpumask *covered = sched_domains_tmpmask; 5637 struct sd_data *sdd = sd->private; 5638 struct sched_domain *child; 5639 int i; 5640 5641 cpumask_clear(covered); 5642 5643 for_each_cpu(i, span) { 5644 struct cpumask *sg_span; 5645 5646 if (cpumask_test_cpu(i, covered)) 5647 continue; 5648 5649 child = *per_cpu_ptr(sdd->sd, i); 5650 5651 /* See the comment near build_group_mask(). */ 5652 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5653 continue; 5654 5655 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5656 GFP_KERNEL, cpu_to_node(cpu)); 5657 5658 if (!sg) 5659 goto fail; 5660 5661 sg_span = sched_group_cpus(sg); 5662 if (child->child) { 5663 child = child->child; 5664 cpumask_copy(sg_span, sched_domain_span(child)); 5665 } else 5666 cpumask_set_cpu(i, sg_span); 5667 5668 cpumask_or(covered, covered, sg_span); 5669 5670 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5671 if (atomic_inc_return(&sg->sgp->ref) == 1) 5672 build_group_mask(sd, sg); 5673 5674 /* 5675 * Initialize sgp->power such that even if we mess up the 5676 * domains and no possible iteration will get us here, we won't 5677 * die on a /0 trap. 5678 */ 5679 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5680 sg->sgp->power_orig = sg->sgp->power; 5681 5682 /* 5683 * Make sure the first group of this domain contains the 5684 * canonical balance cpu. Otherwise the sched_domain iteration 5685 * breaks. See update_sg_lb_stats(). 5686 */ 5687 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5688 group_balance_cpu(sg) == cpu) 5689 groups = sg; 5690 5691 if (!first) 5692 first = sg; 5693 if (last) 5694 last->next = sg; 5695 last = sg; 5696 last->next = first; 5697 } 5698 sd->groups = groups; 5699 5700 return 0; 5701 5702fail: 5703 free_sched_groups(first, 0); 5704 5705 return -ENOMEM; 5706} 5707 5708static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5709{ 5710 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5711 struct sched_domain *child = sd->child; 5712 5713 if (child) 5714 cpu = cpumask_first(sched_domain_span(child)); 5715 5716 if (sg) { 5717 *sg = *per_cpu_ptr(sdd->sg, cpu); 5718 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5719 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5720 } 5721 5722 return cpu; 5723} 5724 5725/* 5726 * build_sched_groups will build a circular linked list of the groups 5727 * covered by the given span, and will set each group's ->cpumask correctly, 5728 * and ->cpu_power to 0. 5729 * 5730 * Assumes the sched_domain tree is fully constructed 5731 */ 5732static int 5733build_sched_groups(struct sched_domain *sd, int cpu) 5734{ 5735 struct sched_group *first = NULL, *last = NULL; 5736 struct sd_data *sdd = sd->private; 5737 const struct cpumask *span = sched_domain_span(sd); 5738 struct cpumask *covered; 5739 int i; 5740 5741 get_group(cpu, sdd, &sd->groups); 5742 atomic_inc(&sd->groups->ref); 5743 5744 if (cpu != cpumask_first(span)) 5745 return 0; 5746 5747 lockdep_assert_held(&sched_domains_mutex); 5748 covered = sched_domains_tmpmask; 5749 5750 cpumask_clear(covered); 5751 5752 for_each_cpu(i, span) { 5753 struct sched_group *sg; 5754 int group, j; 5755 5756 if (cpumask_test_cpu(i, covered)) 5757 continue; 5758 5759 group = get_group(i, sdd, &sg); 5760 cpumask_clear(sched_group_cpus(sg)); 5761 sg->sgp->power = 0; 5762 cpumask_setall(sched_group_mask(sg)); 5763 5764 for_each_cpu(j, span) { 5765 if (get_group(j, sdd, NULL) != group) 5766 continue; 5767 5768 cpumask_set_cpu(j, covered); 5769 cpumask_set_cpu(j, sched_group_cpus(sg)); 5770 } 5771 5772 if (!first) 5773 first = sg; 5774 if (last) 5775 last->next = sg; 5776 last = sg; 5777 } 5778 last->next = first; 5779 5780 return 0; 5781} 5782 5783/* 5784 * Initialize sched groups cpu_power. 5785 * 5786 * cpu_power indicates the capacity of sched group, which is used while 5787 * distributing the load between different sched groups in a sched domain. 5788 * Typically cpu_power for all the groups in a sched domain will be same unless 5789 * there are asymmetries in the topology. If there are asymmetries, group 5790 * having more cpu_power will pickup more load compared to the group having 5791 * less cpu_power. 5792 */ 5793static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5794{ 5795 struct sched_group *sg = sd->groups; 5796 5797 WARN_ON(!sg); 5798 5799 do { 5800 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5801 sg = sg->next; 5802 } while (sg != sd->groups); 5803 5804 if (cpu != group_balance_cpu(sg)) 5805 return; 5806 5807 update_group_power(sd, cpu); 5808 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5809} 5810 5811int __weak arch_sd_sibling_asym_packing(void) 5812{ 5813 return 0*SD_ASYM_PACKING; 5814} 5815 5816/* 5817 * Initializers for schedule domains 5818 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5819 */ 5820 5821#ifdef CONFIG_SCHED_DEBUG 5822# define SD_INIT_NAME(sd, type) sd->name = #type 5823#else 5824# define SD_INIT_NAME(sd, type) do { } while (0) 5825#endif 5826 5827#define SD_INIT_FUNC(type) \ 5828static noinline struct sched_domain * \ 5829sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5830{ \ 5831 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5832 *sd = SD_##type##_INIT; \ 5833 SD_INIT_NAME(sd, type); \ 5834 sd->private = &tl->data; \ 5835 return sd; \ 5836} 5837 5838SD_INIT_FUNC(CPU) 5839#ifdef CONFIG_SCHED_SMT 5840 SD_INIT_FUNC(SIBLING) 5841#endif 5842#ifdef CONFIG_SCHED_MC 5843 SD_INIT_FUNC(MC) 5844#endif 5845#ifdef CONFIG_SCHED_BOOK 5846 SD_INIT_FUNC(BOOK) 5847#endif 5848 5849static int default_relax_domain_level = -1; 5850int sched_domain_level_max; 5851 5852static int __init setup_relax_domain_level(char *str) 5853{ 5854 if (kstrtoint(str, 0, &default_relax_domain_level)) 5855 pr_warn("Unable to set relax_domain_level\n"); 5856 5857 return 1; 5858} 5859__setup("relax_domain_level=", setup_relax_domain_level); 5860 5861static void set_domain_attribute(struct sched_domain *sd, 5862 struct sched_domain_attr *attr) 5863{ 5864 int request; 5865 5866 if (!attr || attr->relax_domain_level < 0) { 5867 if (default_relax_domain_level < 0) 5868 return; 5869 else 5870 request = default_relax_domain_level; 5871 } else 5872 request = attr->relax_domain_level; 5873 if (request < sd->level) { 5874 /* turn off idle balance on this domain */ 5875 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5876 } else { 5877 /* turn on idle balance on this domain */ 5878 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5879 } 5880} 5881 5882static void __sdt_free(const struct cpumask *cpu_map); 5883static int __sdt_alloc(const struct cpumask *cpu_map); 5884 5885static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5886 const struct cpumask *cpu_map) 5887{ 5888 switch (what) { 5889 case sa_rootdomain: 5890 if (!atomic_read(&d->rd->refcount)) 5891 free_rootdomain(&d->rd->rcu); /* fall through */ 5892 case sa_sd: 5893 free_percpu(d->sd); /* fall through */ 5894 case sa_sd_storage: 5895 __sdt_free(cpu_map); /* fall through */ 5896 case sa_none: 5897 break; 5898 } 5899} 5900 5901static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5902 const struct cpumask *cpu_map) 5903{ 5904 memset(d, 0, sizeof(*d)); 5905 5906 if (__sdt_alloc(cpu_map)) 5907 return sa_sd_storage; 5908 d->sd = alloc_percpu(struct sched_domain *); 5909 if (!d->sd) 5910 return sa_sd_storage; 5911 d->rd = alloc_rootdomain(); 5912 if (!d->rd) 5913 return sa_sd; 5914 return sa_rootdomain; 5915} 5916 5917/* 5918 * NULL the sd_data elements we've used to build the sched_domain and 5919 * sched_group structure so that the subsequent __free_domain_allocs() 5920 * will not free the data we're using. 5921 */ 5922static void claim_allocations(int cpu, struct sched_domain *sd) 5923{ 5924 struct sd_data *sdd = sd->private; 5925 5926 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5927 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5928 5929 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5930 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5931 5932 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5933 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5934} 5935 5936#ifdef CONFIG_SCHED_SMT 5937static const struct cpumask *cpu_smt_mask(int cpu) 5938{ 5939 return topology_thread_cpumask(cpu); 5940} 5941#endif 5942 5943/* 5944 * Topology list, bottom-up. 5945 */ 5946static struct sched_domain_topology_level default_topology[] = { 5947#ifdef CONFIG_SCHED_SMT 5948 { sd_init_SIBLING, cpu_smt_mask, }, 5949#endif 5950#ifdef CONFIG_SCHED_MC 5951 { sd_init_MC, cpu_coregroup_mask, }, 5952#endif 5953#ifdef CONFIG_SCHED_BOOK 5954 { sd_init_BOOK, cpu_book_mask, }, 5955#endif 5956 { sd_init_CPU, cpu_cpu_mask, }, 5957 { NULL, }, 5958}; 5959 5960static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5961 5962#define for_each_sd_topology(tl) \ 5963 for (tl = sched_domain_topology; tl->init; tl++) 5964 5965#ifdef CONFIG_NUMA 5966 5967static int sched_domains_numa_levels; 5968static int *sched_domains_numa_distance; 5969static struct cpumask ***sched_domains_numa_masks; 5970static int sched_domains_curr_level; 5971 5972static inline int sd_local_flags(int level) 5973{ 5974 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5975 return 0; 5976 5977 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5978} 5979 5980static struct sched_domain * 5981sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5982{ 5983 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5984 int level = tl->numa_level; 5985 int sd_weight = cpumask_weight( 5986 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5987 5988 *sd = (struct sched_domain){ 5989 .min_interval = sd_weight, 5990 .max_interval = 2*sd_weight, 5991 .busy_factor = 32, 5992 .imbalance_pct = 125, 5993 .cache_nice_tries = 2, 5994 .busy_idx = 3, 5995 .idle_idx = 2, 5996 .newidle_idx = 0, 5997 .wake_idx = 0, 5998 .forkexec_idx = 0, 5999 6000 .flags = 1*SD_LOAD_BALANCE 6001 | 1*SD_BALANCE_NEWIDLE 6002 | 0*SD_BALANCE_EXEC 6003 | 0*SD_BALANCE_FORK 6004 | 0*SD_BALANCE_WAKE 6005 | 0*SD_WAKE_AFFINE 6006 | 0*SD_SHARE_CPUPOWER 6007 | 0*SD_SHARE_PKG_RESOURCES 6008 | 1*SD_SERIALIZE 6009 | 0*SD_PREFER_SIBLING 6010 | 1*SD_NUMA 6011 | sd_local_flags(level) 6012 , 6013 .last_balance = jiffies, 6014 .balance_interval = sd_weight, 6015 }; 6016 SD_INIT_NAME(sd, NUMA); 6017 sd->private = &tl->data; 6018 6019 /* 6020 * Ugly hack to pass state to sd_numa_mask()... 6021 */ 6022 sched_domains_curr_level = tl->numa_level; 6023 6024 return sd; 6025} 6026 6027static const struct cpumask *sd_numa_mask(int cpu) 6028{ 6029 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6030} 6031 6032static void sched_numa_warn(const char *str) 6033{ 6034 static int done = false; 6035 int i,j; 6036 6037 if (done) 6038 return; 6039 6040 done = true; 6041 6042 printk(KERN_WARNING "ERROR: %s\n\n", str); 6043 6044 for (i = 0; i < nr_node_ids; i++) { 6045 printk(KERN_WARNING " "); 6046 for (j = 0; j < nr_node_ids; j++) 6047 printk(KERN_CONT "%02d ", node_distance(i,j)); 6048 printk(KERN_CONT "\n"); 6049 } 6050 printk(KERN_WARNING "\n"); 6051} 6052 6053static bool find_numa_distance(int distance) 6054{ 6055 int i; 6056 6057 if (distance == node_distance(0, 0)) 6058 return true; 6059 6060 for (i = 0; i < sched_domains_numa_levels; i++) { 6061 if (sched_domains_numa_distance[i] == distance) 6062 return true; 6063 } 6064 6065 return false; 6066} 6067 6068static void sched_init_numa(void) 6069{ 6070 int next_distance, curr_distance = node_distance(0, 0); 6071 struct sched_domain_topology_level *tl; 6072 int level = 0; 6073 int i, j, k; 6074 6075 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6076 if (!sched_domains_numa_distance) 6077 return; 6078 6079 /* 6080 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6081 * unique distances in the node_distance() table. 6082 * 6083 * Assumes node_distance(0,j) includes all distances in 6084 * node_distance(i,j) in order to avoid cubic time. 6085 */ 6086 next_distance = curr_distance; 6087 for (i = 0; i < nr_node_ids; i++) { 6088 for (j = 0; j < nr_node_ids; j++) { 6089 for (k = 0; k < nr_node_ids; k++) { 6090 int distance = node_distance(i, k); 6091 6092 if (distance > curr_distance && 6093 (distance < next_distance || 6094 next_distance == curr_distance)) 6095 next_distance = distance; 6096 6097 /* 6098 * While not a strong assumption it would be nice to know 6099 * about cases where if node A is connected to B, B is not 6100 * equally connected to A. 6101 */ 6102 if (sched_debug() && node_distance(k, i) != distance) 6103 sched_numa_warn("Node-distance not symmetric"); 6104 6105 if (sched_debug() && i && !find_numa_distance(distance)) 6106 sched_numa_warn("Node-0 not representative"); 6107 } 6108 if (next_distance != curr_distance) { 6109 sched_domains_numa_distance[level++] = next_distance; 6110 sched_domains_numa_levels = level; 6111 curr_distance = next_distance; 6112 } else break; 6113 } 6114 6115 /* 6116 * In case of sched_debug() we verify the above assumption. 6117 */ 6118 if (!sched_debug()) 6119 break; 6120 } 6121 /* 6122 * 'level' contains the number of unique distances, excluding the 6123 * identity distance node_distance(i,i). 6124 * 6125 * The sched_domains_numa_distance[] array includes the actual distance 6126 * numbers. 6127 */ 6128 6129 /* 6130 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6131 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6132 * the array will contain less then 'level' members. This could be 6133 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6134 * in other functions. 6135 * 6136 * We reset it to 'level' at the end of this function. 6137 */ 6138 sched_domains_numa_levels = 0; 6139 6140 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6141 if (!sched_domains_numa_masks) 6142 return; 6143 6144 /* 6145 * Now for each level, construct a mask per node which contains all 6146 * cpus of nodes that are that many hops away from us. 6147 */ 6148 for (i = 0; i < level; i++) { 6149 sched_domains_numa_masks[i] = 6150 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6151 if (!sched_domains_numa_masks[i]) 6152 return; 6153 6154 for (j = 0; j < nr_node_ids; j++) { 6155 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6156 if (!mask) 6157 return; 6158 6159 sched_domains_numa_masks[i][j] = mask; 6160 6161 for (k = 0; k < nr_node_ids; k++) { 6162 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6163 continue; 6164 6165 cpumask_or(mask, mask, cpumask_of_node(k)); 6166 } 6167 } 6168 } 6169 6170 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6171 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6172 if (!tl) 6173 return; 6174 6175 /* 6176 * Copy the default topology bits.. 6177 */ 6178 for (i = 0; default_topology[i].init; i++) 6179 tl[i] = default_topology[i]; 6180 6181 /* 6182 * .. and append 'j' levels of NUMA goodness. 6183 */ 6184 for (j = 0; j < level; i++, j++) { 6185 tl[i] = (struct sched_domain_topology_level){ 6186 .init = sd_numa_init, 6187 .mask = sd_numa_mask, 6188 .flags = SDTL_OVERLAP, 6189 .numa_level = j, 6190 }; 6191 } 6192 6193 sched_domain_topology = tl; 6194 6195 sched_domains_numa_levels = level; 6196} 6197 6198static void sched_domains_numa_masks_set(int cpu) 6199{ 6200 int i, j; 6201 int node = cpu_to_node(cpu); 6202 6203 for (i = 0; i < sched_domains_numa_levels; i++) { 6204 for (j = 0; j < nr_node_ids; j++) { 6205 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6206 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6207 } 6208 } 6209} 6210 6211static void sched_domains_numa_masks_clear(int cpu) 6212{ 6213 int i, j; 6214 for (i = 0; i < sched_domains_numa_levels; i++) { 6215 for (j = 0; j < nr_node_ids; j++) 6216 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6217 } 6218} 6219 6220/* 6221 * Update sched_domains_numa_masks[level][node] array when new cpus 6222 * are onlined. 6223 */ 6224static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6225 unsigned long action, 6226 void *hcpu) 6227{ 6228 int cpu = (long)hcpu; 6229 6230 switch (action & ~CPU_TASKS_FROZEN) { 6231 case CPU_ONLINE: 6232 sched_domains_numa_masks_set(cpu); 6233 break; 6234 6235 case CPU_DEAD: 6236 sched_domains_numa_masks_clear(cpu); 6237 break; 6238 6239 default: 6240 return NOTIFY_DONE; 6241 } 6242 6243 return NOTIFY_OK; 6244} 6245#else 6246static inline void sched_init_numa(void) 6247{ 6248} 6249 6250static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6251 unsigned long action, 6252 void *hcpu) 6253{ 6254 return 0; 6255} 6256#endif /* CONFIG_NUMA */ 6257 6258static int __sdt_alloc(const struct cpumask *cpu_map) 6259{ 6260 struct sched_domain_topology_level *tl; 6261 int j; 6262 6263 for_each_sd_topology(tl) { 6264 struct sd_data *sdd = &tl->data; 6265 6266 sdd->sd = alloc_percpu(struct sched_domain *); 6267 if (!sdd->sd) 6268 return -ENOMEM; 6269 6270 sdd->sg = alloc_percpu(struct sched_group *); 6271 if (!sdd->sg) 6272 return -ENOMEM; 6273 6274 sdd->sgp = alloc_percpu(struct sched_group_power *); 6275 if (!sdd->sgp) 6276 return -ENOMEM; 6277 6278 for_each_cpu(j, cpu_map) { 6279 struct sched_domain *sd; 6280 struct sched_group *sg; 6281 struct sched_group_power *sgp; 6282 6283 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6284 GFP_KERNEL, cpu_to_node(j)); 6285 if (!sd) 6286 return -ENOMEM; 6287 6288 *per_cpu_ptr(sdd->sd, j) = sd; 6289 6290 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6291 GFP_KERNEL, cpu_to_node(j)); 6292 if (!sg) 6293 return -ENOMEM; 6294 6295 sg->next = sg; 6296 6297 *per_cpu_ptr(sdd->sg, j) = sg; 6298 6299 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6300 GFP_KERNEL, cpu_to_node(j)); 6301 if (!sgp) 6302 return -ENOMEM; 6303 6304 *per_cpu_ptr(sdd->sgp, j) = sgp; 6305 } 6306 } 6307 6308 return 0; 6309} 6310 6311static void __sdt_free(const struct cpumask *cpu_map) 6312{ 6313 struct sched_domain_topology_level *tl; 6314 int j; 6315 6316 for_each_sd_topology(tl) { 6317 struct sd_data *sdd = &tl->data; 6318 6319 for_each_cpu(j, cpu_map) { 6320 struct sched_domain *sd; 6321 6322 if (sdd->sd) { 6323 sd = *per_cpu_ptr(sdd->sd, j); 6324 if (sd && (sd->flags & SD_OVERLAP)) 6325 free_sched_groups(sd->groups, 0); 6326 kfree(*per_cpu_ptr(sdd->sd, j)); 6327 } 6328 6329 if (sdd->sg) 6330 kfree(*per_cpu_ptr(sdd->sg, j)); 6331 if (sdd->sgp) 6332 kfree(*per_cpu_ptr(sdd->sgp, j)); 6333 } 6334 free_percpu(sdd->sd); 6335 sdd->sd = NULL; 6336 free_percpu(sdd->sg); 6337 sdd->sg = NULL; 6338 free_percpu(sdd->sgp); 6339 sdd->sgp = NULL; 6340 } 6341} 6342 6343struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6344 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6345 struct sched_domain *child, int cpu) 6346{ 6347 struct sched_domain *sd = tl->init(tl, cpu); 6348 if (!sd) 6349 return child; 6350 6351 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6352 if (child) { 6353 sd->level = child->level + 1; 6354 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6355 child->parent = sd; 6356 sd->child = child; 6357 } 6358 set_domain_attribute(sd, attr); 6359 6360 return sd; 6361} 6362 6363/* 6364 * Build sched domains for a given set of cpus and attach the sched domains 6365 * to the individual cpus 6366 */ 6367static int build_sched_domains(const struct cpumask *cpu_map, 6368 struct sched_domain_attr *attr) 6369{ 6370 enum s_alloc alloc_state; 6371 struct sched_domain *sd; 6372 struct s_data d; 6373 int i, ret = -ENOMEM; 6374 6375 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6376 if (alloc_state != sa_rootdomain) 6377 goto error; 6378 6379 /* Set up domains for cpus specified by the cpu_map. */ 6380 for_each_cpu(i, cpu_map) { 6381 struct sched_domain_topology_level *tl; 6382 6383 sd = NULL; 6384 for_each_sd_topology(tl) { 6385 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6386 if (tl == sched_domain_topology) 6387 *per_cpu_ptr(d.sd, i) = sd; 6388 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6389 sd->flags |= SD_OVERLAP; 6390 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6391 break; 6392 } 6393 } 6394 6395 /* Build the groups for the domains */ 6396 for_each_cpu(i, cpu_map) { 6397 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6398 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6399 if (sd->flags & SD_OVERLAP) { 6400 if (build_overlap_sched_groups(sd, i)) 6401 goto error; 6402 } else { 6403 if (build_sched_groups(sd, i)) 6404 goto error; 6405 } 6406 } 6407 } 6408 6409 /* Calculate CPU power for physical packages and nodes */ 6410 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6411 if (!cpumask_test_cpu(i, cpu_map)) 6412 continue; 6413 6414 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6415 claim_allocations(i, sd); 6416 init_sched_groups_power(i, sd); 6417 } 6418 } 6419 6420 /* Attach the domains */ 6421 rcu_read_lock(); 6422 for_each_cpu(i, cpu_map) { 6423 sd = *per_cpu_ptr(d.sd, i); 6424 cpu_attach_domain(sd, d.rd, i); 6425 } 6426 rcu_read_unlock(); 6427 6428 ret = 0; 6429error: 6430 __free_domain_allocs(&d, alloc_state, cpu_map); 6431 return ret; 6432} 6433 6434static cpumask_var_t *doms_cur; /* current sched domains */ 6435static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6436static struct sched_domain_attr *dattr_cur; 6437 /* attribues of custom domains in 'doms_cur' */ 6438 6439/* 6440 * Special case: If a kmalloc of a doms_cur partition (array of 6441 * cpumask) fails, then fallback to a single sched domain, 6442 * as determined by the single cpumask fallback_doms. 6443 */ 6444static cpumask_var_t fallback_doms; 6445 6446/* 6447 * arch_update_cpu_topology lets virtualized architectures update the 6448 * cpu core maps. It is supposed to return 1 if the topology changed 6449 * or 0 if it stayed the same. 6450 */ 6451int __attribute__((weak)) arch_update_cpu_topology(void) 6452{ 6453 return 0; 6454} 6455 6456cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6457{ 6458 int i; 6459 cpumask_var_t *doms; 6460 6461 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6462 if (!doms) 6463 return NULL; 6464 for (i = 0; i < ndoms; i++) { 6465 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6466 free_sched_domains(doms, i); 6467 return NULL; 6468 } 6469 } 6470 return doms; 6471} 6472 6473void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6474{ 6475 unsigned int i; 6476 for (i = 0; i < ndoms; i++) 6477 free_cpumask_var(doms[i]); 6478 kfree(doms); 6479} 6480 6481/* 6482 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6483 * For now this just excludes isolated cpus, but could be used to 6484 * exclude other special cases in the future. 6485 */ 6486static int init_sched_domains(const struct cpumask *cpu_map) 6487{ 6488 int err; 6489 6490 arch_update_cpu_topology(); 6491 ndoms_cur = 1; 6492 doms_cur = alloc_sched_domains(ndoms_cur); 6493 if (!doms_cur) 6494 doms_cur = &fallback_doms; 6495 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6496 err = build_sched_domains(doms_cur[0], NULL); 6497 register_sched_domain_sysctl(); 6498 6499 return err; 6500} 6501 6502/* 6503 * Detach sched domains from a group of cpus specified in cpu_map 6504 * These cpus will now be attached to the NULL domain 6505 */ 6506static void detach_destroy_domains(const struct cpumask *cpu_map) 6507{ 6508 int i; 6509 6510 rcu_read_lock(); 6511 for_each_cpu(i, cpu_map) 6512 cpu_attach_domain(NULL, &def_root_domain, i); 6513 rcu_read_unlock(); 6514} 6515 6516/* handle null as "default" */ 6517static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6518 struct sched_domain_attr *new, int idx_new) 6519{ 6520 struct sched_domain_attr tmp; 6521 6522 /* fast path */ 6523 if (!new && !cur) 6524 return 1; 6525 6526 tmp = SD_ATTR_INIT; 6527 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6528 new ? (new + idx_new) : &tmp, 6529 sizeof(struct sched_domain_attr)); 6530} 6531 6532/* 6533 * Partition sched domains as specified by the 'ndoms_new' 6534 * cpumasks in the array doms_new[] of cpumasks. This compares 6535 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6536 * It destroys each deleted domain and builds each new domain. 6537 * 6538 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6539 * The masks don't intersect (don't overlap.) We should setup one 6540 * sched domain for each mask. CPUs not in any of the cpumasks will 6541 * not be load balanced. If the same cpumask appears both in the 6542 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6543 * it as it is. 6544 * 6545 * The passed in 'doms_new' should be allocated using 6546 * alloc_sched_domains. This routine takes ownership of it and will 6547 * free_sched_domains it when done with it. If the caller failed the 6548 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6549 * and partition_sched_domains() will fallback to the single partition 6550 * 'fallback_doms', it also forces the domains to be rebuilt. 6551 * 6552 * If doms_new == NULL it will be replaced with cpu_online_mask. 6553 * ndoms_new == 0 is a special case for destroying existing domains, 6554 * and it will not create the default domain. 6555 * 6556 * Call with hotplug lock held 6557 */ 6558void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6559 struct sched_domain_attr *dattr_new) 6560{ 6561 int i, j, n; 6562 int new_topology; 6563 6564 mutex_lock(&sched_domains_mutex); 6565 6566 /* always unregister in case we don't destroy any domains */ 6567 unregister_sched_domain_sysctl(); 6568 6569 /* Let architecture update cpu core mappings. */ 6570 new_topology = arch_update_cpu_topology(); 6571 6572 n = doms_new ? ndoms_new : 0; 6573 6574 /* Destroy deleted domains */ 6575 for (i = 0; i < ndoms_cur; i++) { 6576 for (j = 0; j < n && !new_topology; j++) { 6577 if (cpumask_equal(doms_cur[i], doms_new[j]) 6578 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6579 goto match1; 6580 } 6581 /* no match - a current sched domain not in new doms_new[] */ 6582 detach_destroy_domains(doms_cur[i]); 6583match1: 6584 ; 6585 } 6586 6587 n = ndoms_cur; 6588 if (doms_new == NULL) { 6589 n = 0; 6590 doms_new = &fallback_doms; 6591 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6592 WARN_ON_ONCE(dattr_new); 6593 } 6594 6595 /* Build new domains */ 6596 for (i = 0; i < ndoms_new; i++) { 6597 for (j = 0; j < n && !new_topology; j++) { 6598 if (cpumask_equal(doms_new[i], doms_cur[j]) 6599 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6600 goto match2; 6601 } 6602 /* no match - add a new doms_new */ 6603 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6604match2: 6605 ; 6606 } 6607 6608 /* Remember the new sched domains */ 6609 if (doms_cur != &fallback_doms) 6610 free_sched_domains(doms_cur, ndoms_cur); 6611 kfree(dattr_cur); /* kfree(NULL) is safe */ 6612 doms_cur = doms_new; 6613 dattr_cur = dattr_new; 6614 ndoms_cur = ndoms_new; 6615 6616 register_sched_domain_sysctl(); 6617 6618 mutex_unlock(&sched_domains_mutex); 6619} 6620 6621static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6622 6623/* 6624 * Update cpusets according to cpu_active mask. If cpusets are 6625 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6626 * around partition_sched_domains(). 6627 * 6628 * If we come here as part of a suspend/resume, don't touch cpusets because we 6629 * want to restore it back to its original state upon resume anyway. 6630 */ 6631static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6632 void *hcpu) 6633{ 6634 switch (action) { 6635 case CPU_ONLINE_FROZEN: 6636 case CPU_DOWN_FAILED_FROZEN: 6637 6638 /* 6639 * num_cpus_frozen tracks how many CPUs are involved in suspend 6640 * resume sequence. As long as this is not the last online 6641 * operation in the resume sequence, just build a single sched 6642 * domain, ignoring cpusets. 6643 */ 6644 num_cpus_frozen--; 6645 if (likely(num_cpus_frozen)) { 6646 partition_sched_domains(1, NULL, NULL); 6647 break; 6648 } 6649 6650 /* 6651 * This is the last CPU online operation. So fall through and 6652 * restore the original sched domains by considering the 6653 * cpuset configurations. 6654 */ 6655 6656 case CPU_ONLINE: 6657 case CPU_DOWN_FAILED: 6658 cpuset_update_active_cpus(true); 6659 break; 6660 default: 6661 return NOTIFY_DONE; 6662 } 6663 return NOTIFY_OK; 6664} 6665 6666static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6667 void *hcpu) 6668{ 6669 switch (action) { 6670 case CPU_DOWN_PREPARE: 6671 cpuset_update_active_cpus(false); 6672 break; 6673 case CPU_DOWN_PREPARE_FROZEN: 6674 num_cpus_frozen++; 6675 partition_sched_domains(1, NULL, NULL); 6676 break; 6677 default: 6678 return NOTIFY_DONE; 6679 } 6680 return NOTIFY_OK; 6681} 6682 6683void __init sched_init_smp(void) 6684{ 6685 cpumask_var_t non_isolated_cpus; 6686 6687 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6688 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6689 6690 sched_init_numa(); 6691 6692 /* 6693 * There's no userspace yet to cause hotplug operations; hence all the 6694 * cpu masks are stable and all blatant races in the below code cannot 6695 * happen. 6696 */ 6697 mutex_lock(&sched_domains_mutex); 6698 init_sched_domains(cpu_active_mask); 6699 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6700 if (cpumask_empty(non_isolated_cpus)) 6701 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6702 mutex_unlock(&sched_domains_mutex); 6703 6704 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6705 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6706 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6707 6708 init_hrtick(); 6709 6710 /* Move init over to a non-isolated CPU */ 6711 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6712 BUG(); 6713 sched_init_granularity(); 6714 free_cpumask_var(non_isolated_cpus); 6715 6716 init_sched_rt_class(); 6717 init_sched_dl_class(); 6718} 6719#else 6720void __init sched_init_smp(void) 6721{ 6722 sched_init_granularity(); 6723} 6724#endif /* CONFIG_SMP */ 6725 6726const_debug unsigned int sysctl_timer_migration = 1; 6727 6728int in_sched_functions(unsigned long addr) 6729{ 6730 return in_lock_functions(addr) || 6731 (addr >= (unsigned long)__sched_text_start 6732 && addr < (unsigned long)__sched_text_end); 6733} 6734 6735#ifdef CONFIG_CGROUP_SCHED 6736/* 6737 * Default task group. 6738 * Every task in system belongs to this group at bootup. 6739 */ 6740struct task_group root_task_group; 6741LIST_HEAD(task_groups); 6742#endif 6743 6744DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6745 6746void __init sched_init(void) 6747{ 6748 int i, j; 6749 unsigned long alloc_size = 0, ptr; 6750 6751#ifdef CONFIG_FAIR_GROUP_SCHED 6752 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6753#endif 6754#ifdef CONFIG_RT_GROUP_SCHED 6755 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6756#endif 6757#ifdef CONFIG_CPUMASK_OFFSTACK 6758 alloc_size += num_possible_cpus() * cpumask_size(); 6759#endif 6760 if (alloc_size) { 6761 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6762 6763#ifdef CONFIG_FAIR_GROUP_SCHED 6764 root_task_group.se = (struct sched_entity **)ptr; 6765 ptr += nr_cpu_ids * sizeof(void **); 6766 6767 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6768 ptr += nr_cpu_ids * sizeof(void **); 6769 6770#endif /* CONFIG_FAIR_GROUP_SCHED */ 6771#ifdef CONFIG_RT_GROUP_SCHED 6772 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6773 ptr += nr_cpu_ids * sizeof(void **); 6774 6775 root_task_group.rt_rq = (struct rt_rq **)ptr; 6776 ptr += nr_cpu_ids * sizeof(void **); 6777 6778#endif /* CONFIG_RT_GROUP_SCHED */ 6779#ifdef CONFIG_CPUMASK_OFFSTACK 6780 for_each_possible_cpu(i) { 6781 per_cpu(load_balance_mask, i) = (void *)ptr; 6782 ptr += cpumask_size(); 6783 } 6784#endif /* CONFIG_CPUMASK_OFFSTACK */ 6785 } 6786 6787 init_rt_bandwidth(&def_rt_bandwidth, 6788 global_rt_period(), global_rt_runtime()); 6789 init_dl_bandwidth(&def_dl_bandwidth, 6790 global_rt_period(), global_rt_runtime()); 6791 6792#ifdef CONFIG_SMP 6793 init_defrootdomain(); 6794#endif 6795 6796#ifdef CONFIG_RT_GROUP_SCHED 6797 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6798 global_rt_period(), global_rt_runtime()); 6799#endif /* CONFIG_RT_GROUP_SCHED */ 6800 6801#ifdef CONFIG_CGROUP_SCHED 6802 list_add(&root_task_group.list, &task_groups); 6803 INIT_LIST_HEAD(&root_task_group.children); 6804 INIT_LIST_HEAD(&root_task_group.siblings); 6805 autogroup_init(&init_task); 6806 6807#endif /* CONFIG_CGROUP_SCHED */ 6808 6809 for_each_possible_cpu(i) { 6810 struct rq *rq; 6811 6812 rq = cpu_rq(i); 6813 raw_spin_lock_init(&rq->lock); 6814 rq->nr_running = 0; 6815 rq->calc_load_active = 0; 6816 rq->calc_load_update = jiffies + LOAD_FREQ; 6817 init_cfs_rq(&rq->cfs); 6818 init_rt_rq(&rq->rt, rq); 6819 init_dl_rq(&rq->dl, rq); 6820#ifdef CONFIG_FAIR_GROUP_SCHED 6821 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6822 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6823 /* 6824 * How much cpu bandwidth does root_task_group get? 6825 * 6826 * In case of task-groups formed thr' the cgroup filesystem, it 6827 * gets 100% of the cpu resources in the system. This overall 6828 * system cpu resource is divided among the tasks of 6829 * root_task_group and its child task-groups in a fair manner, 6830 * based on each entity's (task or task-group's) weight 6831 * (se->load.weight). 6832 * 6833 * In other words, if root_task_group has 10 tasks of weight 6834 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6835 * then A0's share of the cpu resource is: 6836 * 6837 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6838 * 6839 * We achieve this by letting root_task_group's tasks sit 6840 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6841 */ 6842 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6843 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6844#endif /* CONFIG_FAIR_GROUP_SCHED */ 6845 6846 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6847#ifdef CONFIG_RT_GROUP_SCHED 6848 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6849 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6850#endif 6851 6852 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6853 rq->cpu_load[j] = 0; 6854 6855 rq->last_load_update_tick = jiffies; 6856 6857#ifdef CONFIG_SMP 6858 rq->sd = NULL; 6859 rq->rd = NULL; 6860 rq->cpu_power = SCHED_POWER_SCALE; 6861 rq->post_schedule = 0; 6862 rq->active_balance = 0; 6863 rq->next_balance = jiffies; 6864 rq->push_cpu = 0; 6865 rq->cpu = i; 6866 rq->online = 0; 6867 rq->idle_stamp = 0; 6868 rq->avg_idle = 2*sysctl_sched_migration_cost; 6869 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6870 6871 INIT_LIST_HEAD(&rq->cfs_tasks); 6872 6873 rq_attach_root(rq, &def_root_domain); 6874#ifdef CONFIG_NO_HZ_COMMON 6875 rq->nohz_flags = 0; 6876#endif 6877#ifdef CONFIG_NO_HZ_FULL 6878 rq->last_sched_tick = 0; 6879#endif 6880#endif 6881 init_rq_hrtick(rq); 6882 atomic_set(&rq->nr_iowait, 0); 6883 } 6884 6885 set_load_weight(&init_task); 6886 6887#ifdef CONFIG_PREEMPT_NOTIFIERS 6888 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6889#endif 6890 6891 /* 6892 * The boot idle thread does lazy MMU switching as well: 6893 */ 6894 atomic_inc(&init_mm.mm_count); 6895 enter_lazy_tlb(&init_mm, current); 6896 6897 /* 6898 * Make us the idle thread. Technically, schedule() should not be 6899 * called from this thread, however somewhere below it might be, 6900 * but because we are the idle thread, we just pick up running again 6901 * when this runqueue becomes "idle". 6902 */ 6903 init_idle(current, smp_processor_id()); 6904 6905 calc_load_update = jiffies + LOAD_FREQ; 6906 6907 /* 6908 * During early bootup we pretend to be a normal task: 6909 */ 6910 current->sched_class = &fair_sched_class; 6911 6912#ifdef CONFIG_SMP 6913 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6914 /* May be allocated at isolcpus cmdline parse time */ 6915 if (cpu_isolated_map == NULL) 6916 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6917 idle_thread_set_boot_cpu(); 6918#endif 6919 init_sched_fair_class(); 6920 6921 scheduler_running = 1; 6922} 6923 6924#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6925static inline int preempt_count_equals(int preempt_offset) 6926{ 6927 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6928 6929 return (nested == preempt_offset); 6930} 6931 6932void __might_sleep(const char *file, int line, int preempt_offset) 6933{ 6934 static unsigned long prev_jiffy; /* ratelimiting */ 6935 6936 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6937 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6938 system_state != SYSTEM_RUNNING || oops_in_progress) 6939 return; 6940 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6941 return; 6942 prev_jiffy = jiffies; 6943 6944 printk(KERN_ERR 6945 "BUG: sleeping function called from invalid context at %s:%d\n", 6946 file, line); 6947 printk(KERN_ERR 6948 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6949 in_atomic(), irqs_disabled(), 6950 current->pid, current->comm); 6951 6952 debug_show_held_locks(current); 6953 if (irqs_disabled()) 6954 print_irqtrace_events(current); 6955 dump_stack(); 6956} 6957EXPORT_SYMBOL(__might_sleep); 6958#endif 6959 6960#ifdef CONFIG_MAGIC_SYSRQ 6961static void normalize_task(struct rq *rq, struct task_struct *p) 6962{ 6963 const struct sched_class *prev_class = p->sched_class; 6964 struct sched_attr attr = { 6965 .sched_policy = SCHED_NORMAL, 6966 }; 6967 int old_prio = p->prio; 6968 int on_rq; 6969 6970 on_rq = p->on_rq; 6971 if (on_rq) 6972 dequeue_task(rq, p, 0); 6973 __setscheduler(rq, p, &attr); 6974 if (on_rq) { 6975 enqueue_task(rq, p, 0); 6976 resched_task(rq->curr); 6977 } 6978 6979 check_class_changed(rq, p, prev_class, old_prio); 6980} 6981 6982void normalize_rt_tasks(void) 6983{ 6984 struct task_struct *g, *p; 6985 unsigned long flags; 6986 struct rq *rq; 6987 6988 read_lock_irqsave(&tasklist_lock, flags); 6989 do_each_thread(g, p) { 6990 /* 6991 * Only normalize user tasks: 6992 */ 6993 if (!p->mm) 6994 continue; 6995 6996 p->se.exec_start = 0; 6997#ifdef CONFIG_SCHEDSTATS 6998 p->se.statistics.wait_start = 0; 6999 p->se.statistics.sleep_start = 0; 7000 p->se.statistics.block_start = 0; 7001#endif 7002 7003 if (!dl_task(p) && !rt_task(p)) { 7004 /* 7005 * Renice negative nice level userspace 7006 * tasks back to 0: 7007 */ 7008 if (task_nice(p) < 0 && p->mm) 7009 set_user_nice(p, 0); 7010 continue; 7011 } 7012 7013 raw_spin_lock(&p->pi_lock); 7014 rq = __task_rq_lock(p); 7015 7016 normalize_task(rq, p); 7017 7018 __task_rq_unlock(rq); 7019 raw_spin_unlock(&p->pi_lock); 7020 } while_each_thread(g, p); 7021 7022 read_unlock_irqrestore(&tasklist_lock, flags); 7023} 7024 7025#endif /* CONFIG_MAGIC_SYSRQ */ 7026 7027#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7028/* 7029 * These functions are only useful for the IA64 MCA handling, or kdb. 7030 * 7031 * They can only be called when the whole system has been 7032 * stopped - every CPU needs to be quiescent, and no scheduling 7033 * activity can take place. Using them for anything else would 7034 * be a serious bug, and as a result, they aren't even visible 7035 * under any other configuration. 7036 */ 7037 7038/** 7039 * curr_task - return the current task for a given cpu. 7040 * @cpu: the processor in question. 7041 * 7042 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7043 * 7044 * Return: The current task for @cpu. 7045 */ 7046struct task_struct *curr_task(int cpu) 7047{ 7048 return cpu_curr(cpu); 7049} 7050 7051#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7052 7053#ifdef CONFIG_IA64 7054/** 7055 * set_curr_task - set the current task for a given cpu. 7056 * @cpu: the processor in question. 7057 * @p: the task pointer to set. 7058 * 7059 * Description: This function must only be used when non-maskable interrupts 7060 * are serviced on a separate stack. It allows the architecture to switch the 7061 * notion of the current task on a cpu in a non-blocking manner. This function 7062 * must be called with all CPU's synchronized, and interrupts disabled, the 7063 * and caller must save the original value of the current task (see 7064 * curr_task() above) and restore that value before reenabling interrupts and 7065 * re-starting the system. 7066 * 7067 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7068 */ 7069void set_curr_task(int cpu, struct task_struct *p) 7070{ 7071 cpu_curr(cpu) = p; 7072} 7073 7074#endif 7075 7076#ifdef CONFIG_CGROUP_SCHED 7077/* task_group_lock serializes the addition/removal of task groups */ 7078static DEFINE_SPINLOCK(task_group_lock); 7079 7080static void free_sched_group(struct task_group *tg) 7081{ 7082 free_fair_sched_group(tg); 7083 free_rt_sched_group(tg); 7084 autogroup_free(tg); 7085 kfree(tg); 7086} 7087 7088/* allocate runqueue etc for a new task group */ 7089struct task_group *sched_create_group(struct task_group *parent) 7090{ 7091 struct task_group *tg; 7092 7093 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7094 if (!tg) 7095 return ERR_PTR(-ENOMEM); 7096 7097 if (!alloc_fair_sched_group(tg, parent)) 7098 goto err; 7099 7100 if (!alloc_rt_sched_group(tg, parent)) 7101 goto err; 7102 7103 return tg; 7104 7105err: 7106 free_sched_group(tg); 7107 return ERR_PTR(-ENOMEM); 7108} 7109 7110void sched_online_group(struct task_group *tg, struct task_group *parent) 7111{ 7112 unsigned long flags; 7113 7114 spin_lock_irqsave(&task_group_lock, flags); 7115 list_add_rcu(&tg->list, &task_groups); 7116 7117 WARN_ON(!parent); /* root should already exist */ 7118 7119 tg->parent = parent; 7120 INIT_LIST_HEAD(&tg->children); 7121 list_add_rcu(&tg->siblings, &parent->children); 7122 spin_unlock_irqrestore(&task_group_lock, flags); 7123} 7124 7125/* rcu callback to free various structures associated with a task group */ 7126static void free_sched_group_rcu(struct rcu_head *rhp) 7127{ 7128 /* now it should be safe to free those cfs_rqs */ 7129 free_sched_group(container_of(rhp, struct task_group, rcu)); 7130} 7131 7132/* Destroy runqueue etc associated with a task group */ 7133void sched_destroy_group(struct task_group *tg) 7134{ 7135 /* wait for possible concurrent references to cfs_rqs complete */ 7136 call_rcu(&tg->rcu, free_sched_group_rcu); 7137} 7138 7139void sched_offline_group(struct task_group *tg) 7140{ 7141 unsigned long flags; 7142 int i; 7143 7144 /* end participation in shares distribution */ 7145 for_each_possible_cpu(i) 7146 unregister_fair_sched_group(tg, i); 7147 7148 spin_lock_irqsave(&task_group_lock, flags); 7149 list_del_rcu(&tg->list); 7150 list_del_rcu(&tg->siblings); 7151 spin_unlock_irqrestore(&task_group_lock, flags); 7152} 7153 7154/* change task's runqueue when it moves between groups. 7155 * The caller of this function should have put the task in its new group 7156 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7157 * reflect its new group. 7158 */ 7159void sched_move_task(struct task_struct *tsk) 7160{ 7161 struct task_group *tg; 7162 int on_rq, running; 7163 unsigned long flags; 7164 struct rq *rq; 7165 7166 rq = task_rq_lock(tsk, &flags); 7167 7168 running = task_current(rq, tsk); 7169 on_rq = tsk->on_rq; 7170 7171 if (on_rq) 7172 dequeue_task(rq, tsk, 0); 7173 if (unlikely(running)) 7174 tsk->sched_class->put_prev_task(rq, tsk); 7175 7176 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 7177 lockdep_is_held(&tsk->sighand->siglock)), 7178 struct task_group, css); 7179 tg = autogroup_task_group(tsk, tg); 7180 tsk->sched_task_group = tg; 7181 7182#ifdef CONFIG_FAIR_GROUP_SCHED 7183 if (tsk->sched_class->task_move_group) 7184 tsk->sched_class->task_move_group(tsk, on_rq); 7185 else 7186#endif 7187 set_task_rq(tsk, task_cpu(tsk)); 7188 7189 if (unlikely(running)) 7190 tsk->sched_class->set_curr_task(rq); 7191 if (on_rq) 7192 enqueue_task(rq, tsk, 0); 7193 7194 task_rq_unlock(rq, tsk, &flags); 7195} 7196#endif /* CONFIG_CGROUP_SCHED */ 7197 7198#ifdef CONFIG_RT_GROUP_SCHED 7199/* 7200 * Ensure that the real time constraints are schedulable. 7201 */ 7202static DEFINE_MUTEX(rt_constraints_mutex); 7203 7204/* Must be called with tasklist_lock held */ 7205static inline int tg_has_rt_tasks(struct task_group *tg) 7206{ 7207 struct task_struct *g, *p; 7208 7209 do_each_thread(g, p) { 7210 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7211 return 1; 7212 } while_each_thread(g, p); 7213 7214 return 0; 7215} 7216 7217struct rt_schedulable_data { 7218 struct task_group *tg; 7219 u64 rt_period; 7220 u64 rt_runtime; 7221}; 7222 7223static int tg_rt_schedulable(struct task_group *tg, void *data) 7224{ 7225 struct rt_schedulable_data *d = data; 7226 struct task_group *child; 7227 unsigned long total, sum = 0; 7228 u64 period, runtime; 7229 7230 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7231 runtime = tg->rt_bandwidth.rt_runtime; 7232 7233 if (tg == d->tg) { 7234 period = d->rt_period; 7235 runtime = d->rt_runtime; 7236 } 7237 7238 /* 7239 * Cannot have more runtime than the period. 7240 */ 7241 if (runtime > period && runtime != RUNTIME_INF) 7242 return -EINVAL; 7243 7244 /* 7245 * Ensure we don't starve existing RT tasks. 7246 */ 7247 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7248 return -EBUSY; 7249 7250 total = to_ratio(period, runtime); 7251 7252 /* 7253 * Nobody can have more than the global setting allows. 7254 */ 7255 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7256 return -EINVAL; 7257 7258 /* 7259 * The sum of our children's runtime should not exceed our own. 7260 */ 7261 list_for_each_entry_rcu(child, &tg->children, siblings) { 7262 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7263 runtime = child->rt_bandwidth.rt_runtime; 7264 7265 if (child == d->tg) { 7266 period = d->rt_period; 7267 runtime = d->rt_runtime; 7268 } 7269 7270 sum += to_ratio(period, runtime); 7271 } 7272 7273 if (sum > total) 7274 return -EINVAL; 7275 7276 return 0; 7277} 7278 7279static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7280{ 7281 int ret; 7282 7283 struct rt_schedulable_data data = { 7284 .tg = tg, 7285 .rt_period = period, 7286 .rt_runtime = runtime, 7287 }; 7288 7289 rcu_read_lock(); 7290 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7291 rcu_read_unlock(); 7292 7293 return ret; 7294} 7295 7296static int tg_set_rt_bandwidth(struct task_group *tg, 7297 u64 rt_period, u64 rt_runtime) 7298{ 7299 int i, err = 0; 7300 7301 mutex_lock(&rt_constraints_mutex); 7302 read_lock(&tasklist_lock); 7303 err = __rt_schedulable(tg, rt_period, rt_runtime); 7304 if (err) 7305 goto unlock; 7306 7307 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7308 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7309 tg->rt_bandwidth.rt_runtime = rt_runtime; 7310 7311 for_each_possible_cpu(i) { 7312 struct rt_rq *rt_rq = tg->rt_rq[i]; 7313 7314 raw_spin_lock(&rt_rq->rt_runtime_lock); 7315 rt_rq->rt_runtime = rt_runtime; 7316 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7317 } 7318 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7319unlock: 7320 read_unlock(&tasklist_lock); 7321 mutex_unlock(&rt_constraints_mutex); 7322 7323 return err; 7324} 7325 7326static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7327{ 7328 u64 rt_runtime, rt_period; 7329 7330 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7331 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7332 if (rt_runtime_us < 0) 7333 rt_runtime = RUNTIME_INF; 7334 7335 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7336} 7337 7338static long sched_group_rt_runtime(struct task_group *tg) 7339{ 7340 u64 rt_runtime_us; 7341 7342 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7343 return -1; 7344 7345 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7346 do_div(rt_runtime_us, NSEC_PER_USEC); 7347 return rt_runtime_us; 7348} 7349 7350static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7351{ 7352 u64 rt_runtime, rt_period; 7353 7354 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7355 rt_runtime = tg->rt_bandwidth.rt_runtime; 7356 7357 if (rt_period == 0) 7358 return -EINVAL; 7359 7360 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7361} 7362 7363static long sched_group_rt_period(struct task_group *tg) 7364{ 7365 u64 rt_period_us; 7366 7367 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7368 do_div(rt_period_us, NSEC_PER_USEC); 7369 return rt_period_us; 7370} 7371#endif /* CONFIG_RT_GROUP_SCHED */ 7372 7373#ifdef CONFIG_RT_GROUP_SCHED 7374static int sched_rt_global_constraints(void) 7375{ 7376 int ret = 0; 7377 7378 mutex_lock(&rt_constraints_mutex); 7379 read_lock(&tasklist_lock); 7380 ret = __rt_schedulable(NULL, 0, 0); 7381 read_unlock(&tasklist_lock); 7382 mutex_unlock(&rt_constraints_mutex); 7383 7384 return ret; 7385} 7386 7387static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7388{ 7389 /* Don't accept realtime tasks when there is no way for them to run */ 7390 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7391 return 0; 7392 7393 return 1; 7394} 7395 7396#else /* !CONFIG_RT_GROUP_SCHED */ 7397static int sched_rt_global_constraints(void) 7398{ 7399 unsigned long flags; 7400 int i, ret = 0; 7401 7402 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7403 for_each_possible_cpu(i) { 7404 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7405 7406 raw_spin_lock(&rt_rq->rt_runtime_lock); 7407 rt_rq->rt_runtime = global_rt_runtime(); 7408 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7409 } 7410 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7411 7412 return ret; 7413} 7414#endif /* CONFIG_RT_GROUP_SCHED */ 7415 7416static int sched_dl_global_constraints(void) 7417{ 7418 u64 runtime = global_rt_runtime(); 7419 u64 period = global_rt_period(); 7420 u64 new_bw = to_ratio(period, runtime); 7421 int cpu, ret = 0; 7422 7423 /* 7424 * Here we want to check the bandwidth not being set to some 7425 * value smaller than the currently allocated bandwidth in 7426 * any of the root_domains. 7427 * 7428 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7429 * cycling on root_domains... Discussion on different/better 7430 * solutions is welcome! 7431 */ 7432 for_each_possible_cpu(cpu) { 7433 struct dl_bw *dl_b = dl_bw_of(cpu); 7434 7435 raw_spin_lock(&dl_b->lock); 7436 if (new_bw < dl_b->total_bw) 7437 ret = -EBUSY; 7438 raw_spin_unlock(&dl_b->lock); 7439 7440 if (ret) 7441 break; 7442 } 7443 7444 return ret; 7445} 7446 7447static void sched_dl_do_global(void) 7448{ 7449 u64 new_bw = -1; 7450 int cpu; 7451 7452 def_dl_bandwidth.dl_period = global_rt_period(); 7453 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7454 7455 if (global_rt_runtime() != RUNTIME_INF) 7456 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7457 7458 /* 7459 * FIXME: As above... 7460 */ 7461 for_each_possible_cpu(cpu) { 7462 struct dl_bw *dl_b = dl_bw_of(cpu); 7463 7464 raw_spin_lock(&dl_b->lock); 7465 dl_b->bw = new_bw; 7466 raw_spin_unlock(&dl_b->lock); 7467 } 7468} 7469 7470static int sched_rt_global_validate(void) 7471{ 7472 if (sysctl_sched_rt_period <= 0) 7473 return -EINVAL; 7474 7475 if (sysctl_sched_rt_runtime > sysctl_sched_rt_period) 7476 return -EINVAL; 7477 7478 return 0; 7479} 7480 7481static void sched_rt_do_global(void) 7482{ 7483 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7484 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7485} 7486 7487int sched_rt_handler(struct ctl_table *table, int write, 7488 void __user *buffer, size_t *lenp, 7489 loff_t *ppos) 7490{ 7491 int old_period, old_runtime; 7492 static DEFINE_MUTEX(mutex); 7493 int ret; 7494 7495 mutex_lock(&mutex); 7496 old_period = sysctl_sched_rt_period; 7497 old_runtime = sysctl_sched_rt_runtime; 7498 7499 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7500 7501 if (!ret && write) { 7502 ret = sched_rt_global_validate(); 7503 if (ret) 7504 goto undo; 7505 7506 ret = sched_rt_global_constraints(); 7507 if (ret) 7508 goto undo; 7509 7510 ret = sched_dl_global_constraints(); 7511 if (ret) 7512 goto undo; 7513 7514 sched_rt_do_global(); 7515 sched_dl_do_global(); 7516 } 7517 if (0) { 7518undo: 7519 sysctl_sched_rt_period = old_period; 7520 sysctl_sched_rt_runtime = old_runtime; 7521 } 7522 mutex_unlock(&mutex); 7523 7524 return ret; 7525} 7526 7527int sched_rr_handler(struct ctl_table *table, int write, 7528 void __user *buffer, size_t *lenp, 7529 loff_t *ppos) 7530{ 7531 int ret; 7532 static DEFINE_MUTEX(mutex); 7533 7534 mutex_lock(&mutex); 7535 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7536 /* make sure that internally we keep jiffies */ 7537 /* also, writing zero resets timeslice to default */ 7538 if (!ret && write) { 7539 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7540 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7541 } 7542 mutex_unlock(&mutex); 7543 return ret; 7544} 7545 7546#ifdef CONFIG_CGROUP_SCHED 7547 7548static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7549{ 7550 return css ? container_of(css, struct task_group, css) : NULL; 7551} 7552 7553static struct cgroup_subsys_state * 7554cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7555{ 7556 struct task_group *parent = css_tg(parent_css); 7557 struct task_group *tg; 7558 7559 if (!parent) { 7560 /* This is early initialization for the top cgroup */ 7561 return &root_task_group.css; 7562 } 7563 7564 tg = sched_create_group(parent); 7565 if (IS_ERR(tg)) 7566 return ERR_PTR(-ENOMEM); 7567 7568 return &tg->css; 7569} 7570 7571static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7572{ 7573 struct task_group *tg = css_tg(css); 7574 struct task_group *parent = css_tg(css_parent(css)); 7575 7576 if (parent) 7577 sched_online_group(tg, parent); 7578 return 0; 7579} 7580 7581static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7582{ 7583 struct task_group *tg = css_tg(css); 7584 7585 sched_destroy_group(tg); 7586} 7587 7588static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7589{ 7590 struct task_group *tg = css_tg(css); 7591 7592 sched_offline_group(tg); 7593} 7594 7595static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7596 struct cgroup_taskset *tset) 7597{ 7598 struct task_struct *task; 7599 7600 cgroup_taskset_for_each(task, css, tset) { 7601#ifdef CONFIG_RT_GROUP_SCHED 7602 if (!sched_rt_can_attach(css_tg(css), task)) 7603 return -EINVAL; 7604#else 7605 /* We don't support RT-tasks being in separate groups */ 7606 if (task->sched_class != &fair_sched_class) 7607 return -EINVAL; 7608#endif 7609 } 7610 return 0; 7611} 7612 7613static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7614 struct cgroup_taskset *tset) 7615{ 7616 struct task_struct *task; 7617 7618 cgroup_taskset_for_each(task, css, tset) 7619 sched_move_task(task); 7620} 7621 7622static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7623 struct cgroup_subsys_state *old_css, 7624 struct task_struct *task) 7625{ 7626 /* 7627 * cgroup_exit() is called in the copy_process() failure path. 7628 * Ignore this case since the task hasn't ran yet, this avoids 7629 * trying to poke a half freed task state from generic code. 7630 */ 7631 if (!(task->flags & PF_EXITING)) 7632 return; 7633 7634 sched_move_task(task); 7635} 7636 7637#ifdef CONFIG_FAIR_GROUP_SCHED 7638static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7639 struct cftype *cftype, u64 shareval) 7640{ 7641 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7642} 7643 7644static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7645 struct cftype *cft) 7646{ 7647 struct task_group *tg = css_tg(css); 7648 7649 return (u64) scale_load_down(tg->shares); 7650} 7651 7652#ifdef CONFIG_CFS_BANDWIDTH 7653static DEFINE_MUTEX(cfs_constraints_mutex); 7654 7655const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7656const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7657 7658static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7659 7660static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7661{ 7662 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7663 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7664 7665 if (tg == &root_task_group) 7666 return -EINVAL; 7667 7668 /* 7669 * Ensure we have at some amount of bandwidth every period. This is 7670 * to prevent reaching a state of large arrears when throttled via 7671 * entity_tick() resulting in prolonged exit starvation. 7672 */ 7673 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7674 return -EINVAL; 7675 7676 /* 7677 * Likewise, bound things on the otherside by preventing insane quota 7678 * periods. This also allows us to normalize in computing quota 7679 * feasibility. 7680 */ 7681 if (period > max_cfs_quota_period) 7682 return -EINVAL; 7683 7684 mutex_lock(&cfs_constraints_mutex); 7685 ret = __cfs_schedulable(tg, period, quota); 7686 if (ret) 7687 goto out_unlock; 7688 7689 runtime_enabled = quota != RUNTIME_INF; 7690 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7691 /* 7692 * If we need to toggle cfs_bandwidth_used, off->on must occur 7693 * before making related changes, and on->off must occur afterwards 7694 */ 7695 if (runtime_enabled && !runtime_was_enabled) 7696 cfs_bandwidth_usage_inc(); 7697 raw_spin_lock_irq(&cfs_b->lock); 7698 cfs_b->period = ns_to_ktime(period); 7699 cfs_b->quota = quota; 7700 7701 __refill_cfs_bandwidth_runtime(cfs_b); 7702 /* restart the period timer (if active) to handle new period expiry */ 7703 if (runtime_enabled && cfs_b->timer_active) { 7704 /* force a reprogram */ 7705 cfs_b->timer_active = 0; 7706 __start_cfs_bandwidth(cfs_b); 7707 } 7708 raw_spin_unlock_irq(&cfs_b->lock); 7709 7710 for_each_possible_cpu(i) { 7711 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7712 struct rq *rq = cfs_rq->rq; 7713 7714 raw_spin_lock_irq(&rq->lock); 7715 cfs_rq->runtime_enabled = runtime_enabled; 7716 cfs_rq->runtime_remaining = 0; 7717 7718 if (cfs_rq->throttled) 7719 unthrottle_cfs_rq(cfs_rq); 7720 raw_spin_unlock_irq(&rq->lock); 7721 } 7722 if (runtime_was_enabled && !runtime_enabled) 7723 cfs_bandwidth_usage_dec(); 7724out_unlock: 7725 mutex_unlock(&cfs_constraints_mutex); 7726 7727 return ret; 7728} 7729 7730int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7731{ 7732 u64 quota, period; 7733 7734 period = ktime_to_ns(tg->cfs_bandwidth.period); 7735 if (cfs_quota_us < 0) 7736 quota = RUNTIME_INF; 7737 else 7738 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7739 7740 return tg_set_cfs_bandwidth(tg, period, quota); 7741} 7742 7743long tg_get_cfs_quota(struct task_group *tg) 7744{ 7745 u64 quota_us; 7746 7747 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7748 return -1; 7749 7750 quota_us = tg->cfs_bandwidth.quota; 7751 do_div(quota_us, NSEC_PER_USEC); 7752 7753 return quota_us; 7754} 7755 7756int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7757{ 7758 u64 quota, period; 7759 7760 period = (u64)cfs_period_us * NSEC_PER_USEC; 7761 quota = tg->cfs_bandwidth.quota; 7762 7763 return tg_set_cfs_bandwidth(tg, period, quota); 7764} 7765 7766long tg_get_cfs_period(struct task_group *tg) 7767{ 7768 u64 cfs_period_us; 7769 7770 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7771 do_div(cfs_period_us, NSEC_PER_USEC); 7772 7773 return cfs_period_us; 7774} 7775 7776static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7777 struct cftype *cft) 7778{ 7779 return tg_get_cfs_quota(css_tg(css)); 7780} 7781 7782static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7783 struct cftype *cftype, s64 cfs_quota_us) 7784{ 7785 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7786} 7787 7788static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7789 struct cftype *cft) 7790{ 7791 return tg_get_cfs_period(css_tg(css)); 7792} 7793 7794static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7795 struct cftype *cftype, u64 cfs_period_us) 7796{ 7797 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7798} 7799 7800struct cfs_schedulable_data { 7801 struct task_group *tg; 7802 u64 period, quota; 7803}; 7804 7805/* 7806 * normalize group quota/period to be quota/max_period 7807 * note: units are usecs 7808 */ 7809static u64 normalize_cfs_quota(struct task_group *tg, 7810 struct cfs_schedulable_data *d) 7811{ 7812 u64 quota, period; 7813 7814 if (tg == d->tg) { 7815 period = d->period; 7816 quota = d->quota; 7817 } else { 7818 period = tg_get_cfs_period(tg); 7819 quota = tg_get_cfs_quota(tg); 7820 } 7821 7822 /* note: these should typically be equivalent */ 7823 if (quota == RUNTIME_INF || quota == -1) 7824 return RUNTIME_INF; 7825 7826 return to_ratio(period, quota); 7827} 7828 7829static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7830{ 7831 struct cfs_schedulable_data *d = data; 7832 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7833 s64 quota = 0, parent_quota = -1; 7834 7835 if (!tg->parent) { 7836 quota = RUNTIME_INF; 7837 } else { 7838 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7839 7840 quota = normalize_cfs_quota(tg, d); 7841 parent_quota = parent_b->hierarchal_quota; 7842 7843 /* 7844 * ensure max(child_quota) <= parent_quota, inherit when no 7845 * limit is set 7846 */ 7847 if (quota == RUNTIME_INF) 7848 quota = parent_quota; 7849 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7850 return -EINVAL; 7851 } 7852 cfs_b->hierarchal_quota = quota; 7853 7854 return 0; 7855} 7856 7857static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7858{ 7859 int ret; 7860 struct cfs_schedulable_data data = { 7861 .tg = tg, 7862 .period = period, 7863 .quota = quota, 7864 }; 7865 7866 if (quota != RUNTIME_INF) { 7867 do_div(data.period, NSEC_PER_USEC); 7868 do_div(data.quota, NSEC_PER_USEC); 7869 } 7870 7871 rcu_read_lock(); 7872 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7873 rcu_read_unlock(); 7874 7875 return ret; 7876} 7877 7878static int cpu_stats_show(struct seq_file *sf, void *v) 7879{ 7880 struct task_group *tg = css_tg(seq_css(sf)); 7881 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7882 7883 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7884 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7885 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7886 7887 return 0; 7888} 7889#endif /* CONFIG_CFS_BANDWIDTH */ 7890#endif /* CONFIG_FAIR_GROUP_SCHED */ 7891 7892#ifdef CONFIG_RT_GROUP_SCHED 7893static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7894 struct cftype *cft, s64 val) 7895{ 7896 return sched_group_set_rt_runtime(css_tg(css), val); 7897} 7898 7899static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7900 struct cftype *cft) 7901{ 7902 return sched_group_rt_runtime(css_tg(css)); 7903} 7904 7905static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7906 struct cftype *cftype, u64 rt_period_us) 7907{ 7908 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7909} 7910 7911static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7912 struct cftype *cft) 7913{ 7914 return sched_group_rt_period(css_tg(css)); 7915} 7916#endif /* CONFIG_RT_GROUP_SCHED */ 7917 7918static struct cftype cpu_files[] = { 7919#ifdef CONFIG_FAIR_GROUP_SCHED 7920 { 7921 .name = "shares", 7922 .read_u64 = cpu_shares_read_u64, 7923 .write_u64 = cpu_shares_write_u64, 7924 }, 7925#endif 7926#ifdef CONFIG_CFS_BANDWIDTH 7927 { 7928 .name = "cfs_quota_us", 7929 .read_s64 = cpu_cfs_quota_read_s64, 7930 .write_s64 = cpu_cfs_quota_write_s64, 7931 }, 7932 { 7933 .name = "cfs_period_us", 7934 .read_u64 = cpu_cfs_period_read_u64, 7935 .write_u64 = cpu_cfs_period_write_u64, 7936 }, 7937 { 7938 .name = "stat", 7939 .seq_show = cpu_stats_show, 7940 }, 7941#endif 7942#ifdef CONFIG_RT_GROUP_SCHED 7943 { 7944 .name = "rt_runtime_us", 7945 .read_s64 = cpu_rt_runtime_read, 7946 .write_s64 = cpu_rt_runtime_write, 7947 }, 7948 { 7949 .name = "rt_period_us", 7950 .read_u64 = cpu_rt_period_read_uint, 7951 .write_u64 = cpu_rt_period_write_uint, 7952 }, 7953#endif 7954 { } /* terminate */ 7955}; 7956 7957struct cgroup_subsys cpu_cgroup_subsys = { 7958 .name = "cpu", 7959 .css_alloc = cpu_cgroup_css_alloc, 7960 .css_free = cpu_cgroup_css_free, 7961 .css_online = cpu_cgroup_css_online, 7962 .css_offline = cpu_cgroup_css_offline, 7963 .can_attach = cpu_cgroup_can_attach, 7964 .attach = cpu_cgroup_attach, 7965 .exit = cpu_cgroup_exit, 7966 .subsys_id = cpu_cgroup_subsys_id, 7967 .base_cftypes = cpu_files, 7968 .early_init = 1, 7969}; 7970 7971#endif /* CONFIG_CGROUP_SCHED */ 7972 7973void dump_cpu_task(int cpu) 7974{ 7975 pr_info("Task dump for CPU %d:\n", cpu); 7976 sched_show_task(cpu_curr(cpu)); 7977} 7978