core.c revision db273be2a7d42f92b3471e0f717982928214a650
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75#include <linux/context_tracking.h> 76 77#include <asm/switch_to.h> 78#include <asm/tlb.h> 79#include <asm/irq_regs.h> 80#include <asm/mutex.h> 81#ifdef CONFIG_PARAVIRT 82#include <asm/paravirt.h> 83#endif 84 85#include "sched.h" 86#include "../workqueue_internal.h" 87#include "../smpboot.h" 88 89#define CREATE_TRACE_POINTS 90#include <trace/events/sched.h> 91 92void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 93{ 94 unsigned long delta; 95 ktime_t soft, hard, now; 96 97 for (;;) { 98 if (hrtimer_active(period_timer)) 99 break; 100 101 now = hrtimer_cb_get_time(period_timer); 102 hrtimer_forward(period_timer, now, period); 103 104 soft = hrtimer_get_softexpires(period_timer); 105 hard = hrtimer_get_expires(period_timer); 106 delta = ktime_to_ns(ktime_sub(hard, soft)); 107 __hrtimer_start_range_ns(period_timer, soft, delta, 108 HRTIMER_MODE_ABS_PINNED, 0); 109 } 110} 111 112DEFINE_MUTEX(sched_domains_mutex); 113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115static void update_rq_clock_task(struct rq *rq, s64 delta); 116 117void update_rq_clock(struct rq *rq) 118{ 119 s64 delta; 120 121 if (rq->skip_clock_update > 0) 122 return; 123 124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 125 rq->clock += delta; 126 update_rq_clock_task(rq, delta); 127} 128 129/* 130 * Debugging: various feature bits 131 */ 132 133#define SCHED_FEAT(name, enabled) \ 134 (1UL << __SCHED_FEAT_##name) * enabled | 135 136const_debug unsigned int sysctl_sched_features = 137#include "features.h" 138 0; 139 140#undef SCHED_FEAT 141 142#ifdef CONFIG_SCHED_DEBUG 143#define SCHED_FEAT(name, enabled) \ 144 #name , 145 146static const char * const sched_feat_names[] = { 147#include "features.h" 148}; 149 150#undef SCHED_FEAT 151 152static int sched_feat_show(struct seq_file *m, void *v) 153{ 154 int i; 155 156 for (i = 0; i < __SCHED_FEAT_NR; i++) { 157 if (!(sysctl_sched_features & (1UL << i))) 158 seq_puts(m, "NO_"); 159 seq_printf(m, "%s ", sched_feat_names[i]); 160 } 161 seq_puts(m, "\n"); 162 163 return 0; 164} 165 166#ifdef HAVE_JUMP_LABEL 167 168#define jump_label_key__true STATIC_KEY_INIT_TRUE 169#define jump_label_key__false STATIC_KEY_INIT_FALSE 170 171#define SCHED_FEAT(name, enabled) \ 172 jump_label_key__##enabled , 173 174struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 175#include "features.h" 176}; 177 178#undef SCHED_FEAT 179 180static void sched_feat_disable(int i) 181{ 182 if (static_key_enabled(&sched_feat_keys[i])) 183 static_key_slow_dec(&sched_feat_keys[i]); 184} 185 186static void sched_feat_enable(int i) 187{ 188 if (!static_key_enabled(&sched_feat_keys[i])) 189 static_key_slow_inc(&sched_feat_keys[i]); 190} 191#else 192static void sched_feat_disable(int i) { }; 193static void sched_feat_enable(int i) { }; 194#endif /* HAVE_JUMP_LABEL */ 195 196static int sched_feat_set(char *cmp) 197{ 198 int i; 199 int neg = 0; 200 201 if (strncmp(cmp, "NO_", 3) == 0) { 202 neg = 1; 203 cmp += 3; 204 } 205 206 for (i = 0; i < __SCHED_FEAT_NR; i++) { 207 if (strcmp(cmp, sched_feat_names[i]) == 0) { 208 if (neg) { 209 sysctl_sched_features &= ~(1UL << i); 210 sched_feat_disable(i); 211 } else { 212 sysctl_sched_features |= (1UL << i); 213 sched_feat_enable(i); 214 } 215 break; 216 } 217 } 218 219 return i; 220} 221 222static ssize_t 223sched_feat_write(struct file *filp, const char __user *ubuf, 224 size_t cnt, loff_t *ppos) 225{ 226 char buf[64]; 227 char *cmp; 228 int i; 229 230 if (cnt > 63) 231 cnt = 63; 232 233 if (copy_from_user(&buf, ubuf, cnt)) 234 return -EFAULT; 235 236 buf[cnt] = 0; 237 cmp = strstrip(buf); 238 239 i = sched_feat_set(cmp); 240 if (i == __SCHED_FEAT_NR) 241 return -EINVAL; 242 243 *ppos += cnt; 244 245 return cnt; 246} 247 248static int sched_feat_open(struct inode *inode, struct file *filp) 249{ 250 return single_open(filp, sched_feat_show, NULL); 251} 252 253static const struct file_operations sched_feat_fops = { 254 .open = sched_feat_open, 255 .write = sched_feat_write, 256 .read = seq_read, 257 .llseek = seq_lseek, 258 .release = single_release, 259}; 260 261static __init int sched_init_debug(void) 262{ 263 debugfs_create_file("sched_features", 0644, NULL, NULL, 264 &sched_feat_fops); 265 266 return 0; 267} 268late_initcall(sched_init_debug); 269#endif /* CONFIG_SCHED_DEBUG */ 270 271/* 272 * Number of tasks to iterate in a single balance run. 273 * Limited because this is done with IRQs disabled. 274 */ 275const_debug unsigned int sysctl_sched_nr_migrate = 32; 276 277/* 278 * period over which we average the RT time consumption, measured 279 * in ms. 280 * 281 * default: 1s 282 */ 283const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 284 285/* 286 * period over which we measure -rt task cpu usage in us. 287 * default: 1s 288 */ 289unsigned int sysctl_sched_rt_period = 1000000; 290 291__read_mostly int scheduler_running; 292 293/* 294 * part of the period that we allow rt tasks to run in us. 295 * default: 0.95s 296 */ 297int sysctl_sched_rt_runtime = 950000; 298 299/* 300 * __task_rq_lock - lock the rq @p resides on. 301 */ 302static inline struct rq *__task_rq_lock(struct task_struct *p) 303 __acquires(rq->lock) 304{ 305 struct rq *rq; 306 307 lockdep_assert_held(&p->pi_lock); 308 309 for (;;) { 310 rq = task_rq(p); 311 raw_spin_lock(&rq->lock); 312 if (likely(rq == task_rq(p))) 313 return rq; 314 raw_spin_unlock(&rq->lock); 315 } 316} 317 318/* 319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 320 */ 321static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 322 __acquires(p->pi_lock) 323 __acquires(rq->lock) 324{ 325 struct rq *rq; 326 327 for (;;) { 328 raw_spin_lock_irqsave(&p->pi_lock, *flags); 329 rq = task_rq(p); 330 raw_spin_lock(&rq->lock); 331 if (likely(rq == task_rq(p))) 332 return rq; 333 raw_spin_unlock(&rq->lock); 334 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 335 } 336} 337 338static void __task_rq_unlock(struct rq *rq) 339 __releases(rq->lock) 340{ 341 raw_spin_unlock(&rq->lock); 342} 343 344static inline void 345task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 346 __releases(rq->lock) 347 __releases(p->pi_lock) 348{ 349 raw_spin_unlock(&rq->lock); 350 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 351} 352 353/* 354 * this_rq_lock - lock this runqueue and disable interrupts. 355 */ 356static struct rq *this_rq_lock(void) 357 __acquires(rq->lock) 358{ 359 struct rq *rq; 360 361 local_irq_disable(); 362 rq = this_rq(); 363 raw_spin_lock(&rq->lock); 364 365 return rq; 366} 367 368#ifdef CONFIG_SCHED_HRTICK 369/* 370 * Use HR-timers to deliver accurate preemption points. 371 */ 372 373static void hrtick_clear(struct rq *rq) 374{ 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377} 378 379/* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383static enum hrtimer_restart hrtick(struct hrtimer *timer) 384{ 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395} 396 397#ifdef CONFIG_SMP 398 399static int __hrtick_restart(struct rq *rq) 400{ 401 struct hrtimer *timer = &rq->hrtick_timer; 402 ktime_t time = hrtimer_get_softexpires(timer); 403 404 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 405} 406 407/* 408 * called from hardirq (IPI) context 409 */ 410static void __hrtick_start(void *arg) 411{ 412 struct rq *rq = arg; 413 414 raw_spin_lock(&rq->lock); 415 __hrtick_restart(rq); 416 rq->hrtick_csd_pending = 0; 417 raw_spin_unlock(&rq->lock); 418} 419 420/* 421 * Called to set the hrtick timer state. 422 * 423 * called with rq->lock held and irqs disabled 424 */ 425void hrtick_start(struct rq *rq, u64 delay) 426{ 427 struct hrtimer *timer = &rq->hrtick_timer; 428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 429 430 hrtimer_set_expires(timer, time); 431 432 if (rq == this_rq()) { 433 __hrtick_restart(rq); 434 } else if (!rq->hrtick_csd_pending) { 435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 436 rq->hrtick_csd_pending = 1; 437 } 438} 439 440static int 441hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 442{ 443 int cpu = (int)(long)hcpu; 444 445 switch (action) { 446 case CPU_UP_CANCELED: 447 case CPU_UP_CANCELED_FROZEN: 448 case CPU_DOWN_PREPARE: 449 case CPU_DOWN_PREPARE_FROZEN: 450 case CPU_DEAD: 451 case CPU_DEAD_FROZEN: 452 hrtick_clear(cpu_rq(cpu)); 453 return NOTIFY_OK; 454 } 455 456 return NOTIFY_DONE; 457} 458 459static __init void init_hrtick(void) 460{ 461 hotcpu_notifier(hotplug_hrtick, 0); 462} 463#else 464/* 465 * Called to set the hrtick timer state. 466 * 467 * called with rq->lock held and irqs disabled 468 */ 469void hrtick_start(struct rq *rq, u64 delay) 470{ 471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 472 HRTIMER_MODE_REL_PINNED, 0); 473} 474 475static inline void init_hrtick(void) 476{ 477} 478#endif /* CONFIG_SMP */ 479 480static void init_rq_hrtick(struct rq *rq) 481{ 482#ifdef CONFIG_SMP 483 rq->hrtick_csd_pending = 0; 484 485 rq->hrtick_csd.flags = 0; 486 rq->hrtick_csd.func = __hrtick_start; 487 rq->hrtick_csd.info = rq; 488#endif 489 490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 491 rq->hrtick_timer.function = hrtick; 492} 493#else /* CONFIG_SCHED_HRTICK */ 494static inline void hrtick_clear(struct rq *rq) 495{ 496} 497 498static inline void init_rq_hrtick(struct rq *rq) 499{ 500} 501 502static inline void init_hrtick(void) 503{ 504} 505#endif /* CONFIG_SCHED_HRTICK */ 506 507/* 508 * resched_task - mark a task 'to be rescheduled now'. 509 * 510 * On UP this means the setting of the need_resched flag, on SMP it 511 * might also involve a cross-CPU call to trigger the scheduler on 512 * the target CPU. 513 */ 514void resched_task(struct task_struct *p) 515{ 516 int cpu; 517 518 lockdep_assert_held(&task_rq(p)->lock); 519 520 if (test_tsk_need_resched(p)) 521 return; 522 523 set_tsk_need_resched(p); 524 525 cpu = task_cpu(p); 526 if (cpu == smp_processor_id()) { 527 set_preempt_need_resched(); 528 return; 529 } 530 531 /* NEED_RESCHED must be visible before we test polling */ 532 smp_mb(); 533 if (!tsk_is_polling(p)) 534 smp_send_reschedule(cpu); 535} 536 537void resched_cpu(int cpu) 538{ 539 struct rq *rq = cpu_rq(cpu); 540 unsigned long flags; 541 542 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 543 return; 544 resched_task(cpu_curr(cpu)); 545 raw_spin_unlock_irqrestore(&rq->lock, flags); 546} 547 548#ifdef CONFIG_SMP 549#ifdef CONFIG_NO_HZ_COMMON 550/* 551 * In the semi idle case, use the nearest busy cpu for migrating timers 552 * from an idle cpu. This is good for power-savings. 553 * 554 * We don't do similar optimization for completely idle system, as 555 * selecting an idle cpu will add more delays to the timers than intended 556 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 557 */ 558int get_nohz_timer_target(void) 559{ 560 int cpu = smp_processor_id(); 561 int i; 562 struct sched_domain *sd; 563 564 rcu_read_lock(); 565 for_each_domain(cpu, sd) { 566 for_each_cpu(i, sched_domain_span(sd)) { 567 if (!idle_cpu(i)) { 568 cpu = i; 569 goto unlock; 570 } 571 } 572 } 573unlock: 574 rcu_read_unlock(); 575 return cpu; 576} 577/* 578 * When add_timer_on() enqueues a timer into the timer wheel of an 579 * idle CPU then this timer might expire before the next timer event 580 * which is scheduled to wake up that CPU. In case of a completely 581 * idle system the next event might even be infinite time into the 582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 583 * leaves the inner idle loop so the newly added timer is taken into 584 * account when the CPU goes back to idle and evaluates the timer 585 * wheel for the next timer event. 586 */ 587static void wake_up_idle_cpu(int cpu) 588{ 589 struct rq *rq = cpu_rq(cpu); 590 591 if (cpu == smp_processor_id()) 592 return; 593 594 /* 595 * This is safe, as this function is called with the timer 596 * wheel base lock of (cpu) held. When the CPU is on the way 597 * to idle and has not yet set rq->curr to idle then it will 598 * be serialized on the timer wheel base lock and take the new 599 * timer into account automatically. 600 */ 601 if (rq->curr != rq->idle) 602 return; 603 604 /* 605 * We can set TIF_RESCHED on the idle task of the other CPU 606 * lockless. The worst case is that the other CPU runs the 607 * idle task through an additional NOOP schedule() 608 */ 609 set_tsk_need_resched(rq->idle); 610 611 /* NEED_RESCHED must be visible before we test polling */ 612 smp_mb(); 613 if (!tsk_is_polling(rq->idle)) 614 smp_send_reschedule(cpu); 615} 616 617static bool wake_up_full_nohz_cpu(int cpu) 618{ 619 if (tick_nohz_full_cpu(cpu)) { 620 if (cpu != smp_processor_id() || 621 tick_nohz_tick_stopped()) 622 smp_send_reschedule(cpu); 623 return true; 624 } 625 626 return false; 627} 628 629void wake_up_nohz_cpu(int cpu) 630{ 631 if (!wake_up_full_nohz_cpu(cpu)) 632 wake_up_idle_cpu(cpu); 633} 634 635static inline bool got_nohz_idle_kick(void) 636{ 637 int cpu = smp_processor_id(); 638 639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 640 return false; 641 642 if (idle_cpu(cpu) && !need_resched()) 643 return true; 644 645 /* 646 * We can't run Idle Load Balance on this CPU for this time so we 647 * cancel it and clear NOHZ_BALANCE_KICK 648 */ 649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 650 return false; 651} 652 653#else /* CONFIG_NO_HZ_COMMON */ 654 655static inline bool got_nohz_idle_kick(void) 656{ 657 return false; 658} 659 660#endif /* CONFIG_NO_HZ_COMMON */ 661 662#ifdef CONFIG_NO_HZ_FULL 663bool sched_can_stop_tick(void) 664{ 665 struct rq *rq; 666 667 rq = this_rq(); 668 669 /* Make sure rq->nr_running update is visible after the IPI */ 670 smp_rmb(); 671 672 /* More than one running task need preemption */ 673 if (rq->nr_running > 1) 674 return false; 675 676 return true; 677} 678#endif /* CONFIG_NO_HZ_FULL */ 679 680void sched_avg_update(struct rq *rq) 681{ 682 s64 period = sched_avg_period(); 683 684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 685 /* 686 * Inline assembly required to prevent the compiler 687 * optimising this loop into a divmod call. 688 * See __iter_div_u64_rem() for another example of this. 689 */ 690 asm("" : "+rm" (rq->age_stamp)); 691 rq->age_stamp += period; 692 rq->rt_avg /= 2; 693 } 694} 695 696#endif /* CONFIG_SMP */ 697 698#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 699 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 700/* 701 * Iterate task_group tree rooted at *from, calling @down when first entering a 702 * node and @up when leaving it for the final time. 703 * 704 * Caller must hold rcu_lock or sufficient equivalent. 705 */ 706int walk_tg_tree_from(struct task_group *from, 707 tg_visitor down, tg_visitor up, void *data) 708{ 709 struct task_group *parent, *child; 710 int ret; 711 712 parent = from; 713 714down: 715 ret = (*down)(parent, data); 716 if (ret) 717 goto out; 718 list_for_each_entry_rcu(child, &parent->children, siblings) { 719 parent = child; 720 goto down; 721 722up: 723 continue; 724 } 725 ret = (*up)(parent, data); 726 if (ret || parent == from) 727 goto out; 728 729 child = parent; 730 parent = parent->parent; 731 if (parent) 732 goto up; 733out: 734 return ret; 735} 736 737int tg_nop(struct task_group *tg, void *data) 738{ 739 return 0; 740} 741#endif 742 743static void set_load_weight(struct task_struct *p) 744{ 745 int prio = p->static_prio - MAX_RT_PRIO; 746 struct load_weight *load = &p->se.load; 747 748 /* 749 * SCHED_IDLE tasks get minimal weight: 750 */ 751 if (p->policy == SCHED_IDLE) { 752 load->weight = scale_load(WEIGHT_IDLEPRIO); 753 load->inv_weight = WMULT_IDLEPRIO; 754 return; 755 } 756 757 load->weight = scale_load(prio_to_weight[prio]); 758 load->inv_weight = prio_to_wmult[prio]; 759} 760 761static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 762{ 763 update_rq_clock(rq); 764 sched_info_queued(rq, p); 765 p->sched_class->enqueue_task(rq, p, flags); 766} 767 768static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 769{ 770 update_rq_clock(rq); 771 sched_info_dequeued(rq, p); 772 p->sched_class->dequeue_task(rq, p, flags); 773} 774 775void activate_task(struct rq *rq, struct task_struct *p, int flags) 776{ 777 if (task_contributes_to_load(p)) 778 rq->nr_uninterruptible--; 779 780 enqueue_task(rq, p, flags); 781} 782 783void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 784{ 785 if (task_contributes_to_load(p)) 786 rq->nr_uninterruptible++; 787 788 dequeue_task(rq, p, flags); 789} 790 791static void update_rq_clock_task(struct rq *rq, s64 delta) 792{ 793/* 794 * In theory, the compile should just see 0 here, and optimize out the call 795 * to sched_rt_avg_update. But I don't trust it... 796 */ 797#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 798 s64 steal = 0, irq_delta = 0; 799#endif 800#ifdef CONFIG_IRQ_TIME_ACCOUNTING 801 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 802 803 /* 804 * Since irq_time is only updated on {soft,}irq_exit, we might run into 805 * this case when a previous update_rq_clock() happened inside a 806 * {soft,}irq region. 807 * 808 * When this happens, we stop ->clock_task and only update the 809 * prev_irq_time stamp to account for the part that fit, so that a next 810 * update will consume the rest. This ensures ->clock_task is 811 * monotonic. 812 * 813 * It does however cause some slight miss-attribution of {soft,}irq 814 * time, a more accurate solution would be to update the irq_time using 815 * the current rq->clock timestamp, except that would require using 816 * atomic ops. 817 */ 818 if (irq_delta > delta) 819 irq_delta = delta; 820 821 rq->prev_irq_time += irq_delta; 822 delta -= irq_delta; 823#endif 824#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 825 if (static_key_false((¶virt_steal_rq_enabled))) { 826 u64 st; 827 828 steal = paravirt_steal_clock(cpu_of(rq)); 829 steal -= rq->prev_steal_time_rq; 830 831 if (unlikely(steal > delta)) 832 steal = delta; 833 834 st = steal_ticks(steal); 835 steal = st * TICK_NSEC; 836 837 rq->prev_steal_time_rq += steal; 838 839 delta -= steal; 840 } 841#endif 842 843 rq->clock_task += delta; 844 845#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 846 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 847 sched_rt_avg_update(rq, irq_delta + steal); 848#endif 849} 850 851void sched_set_stop_task(int cpu, struct task_struct *stop) 852{ 853 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 854 struct task_struct *old_stop = cpu_rq(cpu)->stop; 855 856 if (stop) { 857 /* 858 * Make it appear like a SCHED_FIFO task, its something 859 * userspace knows about and won't get confused about. 860 * 861 * Also, it will make PI more or less work without too 862 * much confusion -- but then, stop work should not 863 * rely on PI working anyway. 864 */ 865 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 866 867 stop->sched_class = &stop_sched_class; 868 } 869 870 cpu_rq(cpu)->stop = stop; 871 872 if (old_stop) { 873 /* 874 * Reset it back to a normal scheduling class so that 875 * it can die in pieces. 876 */ 877 old_stop->sched_class = &rt_sched_class; 878 } 879} 880 881/* 882 * __normal_prio - return the priority that is based on the static prio 883 */ 884static inline int __normal_prio(struct task_struct *p) 885{ 886 return p->static_prio; 887} 888 889/* 890 * Calculate the expected normal priority: i.e. priority 891 * without taking RT-inheritance into account. Might be 892 * boosted by interactivity modifiers. Changes upon fork, 893 * setprio syscalls, and whenever the interactivity 894 * estimator recalculates. 895 */ 896static inline int normal_prio(struct task_struct *p) 897{ 898 int prio; 899 900 if (task_has_dl_policy(p)) 901 prio = MAX_DL_PRIO-1; 902 else if (task_has_rt_policy(p)) 903 prio = MAX_RT_PRIO-1 - p->rt_priority; 904 else 905 prio = __normal_prio(p); 906 return prio; 907} 908 909/* 910 * Calculate the current priority, i.e. the priority 911 * taken into account by the scheduler. This value might 912 * be boosted by RT tasks, or might be boosted by 913 * interactivity modifiers. Will be RT if the task got 914 * RT-boosted. If not then it returns p->normal_prio. 915 */ 916static int effective_prio(struct task_struct *p) 917{ 918 p->normal_prio = normal_prio(p); 919 /* 920 * If we are RT tasks or we were boosted to RT priority, 921 * keep the priority unchanged. Otherwise, update priority 922 * to the normal priority: 923 */ 924 if (!rt_prio(p->prio)) 925 return p->normal_prio; 926 return p->prio; 927} 928 929/** 930 * task_curr - is this task currently executing on a CPU? 931 * @p: the task in question. 932 * 933 * Return: 1 if the task is currently executing. 0 otherwise. 934 */ 935inline int task_curr(const struct task_struct *p) 936{ 937 return cpu_curr(task_cpu(p)) == p; 938} 939 940static inline void check_class_changed(struct rq *rq, struct task_struct *p, 941 const struct sched_class *prev_class, 942 int oldprio) 943{ 944 if (prev_class != p->sched_class) { 945 if (prev_class->switched_from) 946 prev_class->switched_from(rq, p); 947 p->sched_class->switched_to(rq, p); 948 } else if (oldprio != p->prio || dl_task(p)) 949 p->sched_class->prio_changed(rq, p, oldprio); 950} 951 952void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 953{ 954 const struct sched_class *class; 955 956 if (p->sched_class == rq->curr->sched_class) { 957 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 958 } else { 959 for_each_class(class) { 960 if (class == rq->curr->sched_class) 961 break; 962 if (class == p->sched_class) { 963 resched_task(rq->curr); 964 break; 965 } 966 } 967 } 968 969 /* 970 * A queue event has occurred, and we're going to schedule. In 971 * this case, we can save a useless back to back clock update. 972 */ 973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 974 rq->skip_clock_update = 1; 975} 976 977#ifdef CONFIG_SMP 978void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 979{ 980#ifdef CONFIG_SCHED_DEBUG 981 /* 982 * We should never call set_task_cpu() on a blocked task, 983 * ttwu() will sort out the placement. 984 */ 985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 986 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 987 988#ifdef CONFIG_LOCKDEP 989 /* 990 * The caller should hold either p->pi_lock or rq->lock, when changing 991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 992 * 993 * sched_move_task() holds both and thus holding either pins the cgroup, 994 * see task_group(). 995 * 996 * Furthermore, all task_rq users should acquire both locks, see 997 * task_rq_lock(). 998 */ 999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1000 lockdep_is_held(&task_rq(p)->lock))); 1001#endif 1002#endif 1003 1004 trace_sched_migrate_task(p, new_cpu); 1005 1006 if (task_cpu(p) != new_cpu) { 1007 if (p->sched_class->migrate_task_rq) 1008 p->sched_class->migrate_task_rq(p, new_cpu); 1009 p->se.nr_migrations++; 1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1011 } 1012 1013 __set_task_cpu(p, new_cpu); 1014} 1015 1016static void __migrate_swap_task(struct task_struct *p, int cpu) 1017{ 1018 if (p->on_rq) { 1019 struct rq *src_rq, *dst_rq; 1020 1021 src_rq = task_rq(p); 1022 dst_rq = cpu_rq(cpu); 1023 1024 deactivate_task(src_rq, p, 0); 1025 set_task_cpu(p, cpu); 1026 activate_task(dst_rq, p, 0); 1027 check_preempt_curr(dst_rq, p, 0); 1028 } else { 1029 /* 1030 * Task isn't running anymore; make it appear like we migrated 1031 * it before it went to sleep. This means on wakeup we make the 1032 * previous cpu our targer instead of where it really is. 1033 */ 1034 p->wake_cpu = cpu; 1035 } 1036} 1037 1038struct migration_swap_arg { 1039 struct task_struct *src_task, *dst_task; 1040 int src_cpu, dst_cpu; 1041}; 1042 1043static int migrate_swap_stop(void *data) 1044{ 1045 struct migration_swap_arg *arg = data; 1046 struct rq *src_rq, *dst_rq; 1047 int ret = -EAGAIN; 1048 1049 src_rq = cpu_rq(arg->src_cpu); 1050 dst_rq = cpu_rq(arg->dst_cpu); 1051 1052 double_raw_lock(&arg->src_task->pi_lock, 1053 &arg->dst_task->pi_lock); 1054 double_rq_lock(src_rq, dst_rq); 1055 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1056 goto unlock; 1057 1058 if (task_cpu(arg->src_task) != arg->src_cpu) 1059 goto unlock; 1060 1061 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1062 goto unlock; 1063 1064 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1065 goto unlock; 1066 1067 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1068 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1069 1070 ret = 0; 1071 1072unlock: 1073 double_rq_unlock(src_rq, dst_rq); 1074 raw_spin_unlock(&arg->dst_task->pi_lock); 1075 raw_spin_unlock(&arg->src_task->pi_lock); 1076 1077 return ret; 1078} 1079 1080/* 1081 * Cross migrate two tasks 1082 */ 1083int migrate_swap(struct task_struct *cur, struct task_struct *p) 1084{ 1085 struct migration_swap_arg arg; 1086 int ret = -EINVAL; 1087 1088 arg = (struct migration_swap_arg){ 1089 .src_task = cur, 1090 .src_cpu = task_cpu(cur), 1091 .dst_task = p, 1092 .dst_cpu = task_cpu(p), 1093 }; 1094 1095 if (arg.src_cpu == arg.dst_cpu) 1096 goto out; 1097 1098 /* 1099 * These three tests are all lockless; this is OK since all of them 1100 * will be re-checked with proper locks held further down the line. 1101 */ 1102 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1103 goto out; 1104 1105 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1106 goto out; 1107 1108 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1109 goto out; 1110 1111 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1112 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1113 1114out: 1115 return ret; 1116} 1117 1118struct migration_arg { 1119 struct task_struct *task; 1120 int dest_cpu; 1121}; 1122 1123static int migration_cpu_stop(void *data); 1124 1125/* 1126 * wait_task_inactive - wait for a thread to unschedule. 1127 * 1128 * If @match_state is nonzero, it's the @p->state value just checked and 1129 * not expected to change. If it changes, i.e. @p might have woken up, 1130 * then return zero. When we succeed in waiting for @p to be off its CPU, 1131 * we return a positive number (its total switch count). If a second call 1132 * a short while later returns the same number, the caller can be sure that 1133 * @p has remained unscheduled the whole time. 1134 * 1135 * The caller must ensure that the task *will* unschedule sometime soon, 1136 * else this function might spin for a *long* time. This function can't 1137 * be called with interrupts off, or it may introduce deadlock with 1138 * smp_call_function() if an IPI is sent by the same process we are 1139 * waiting to become inactive. 1140 */ 1141unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1142{ 1143 unsigned long flags; 1144 int running, on_rq; 1145 unsigned long ncsw; 1146 struct rq *rq; 1147 1148 for (;;) { 1149 /* 1150 * We do the initial early heuristics without holding 1151 * any task-queue locks at all. We'll only try to get 1152 * the runqueue lock when things look like they will 1153 * work out! 1154 */ 1155 rq = task_rq(p); 1156 1157 /* 1158 * If the task is actively running on another CPU 1159 * still, just relax and busy-wait without holding 1160 * any locks. 1161 * 1162 * NOTE! Since we don't hold any locks, it's not 1163 * even sure that "rq" stays as the right runqueue! 1164 * But we don't care, since "task_running()" will 1165 * return false if the runqueue has changed and p 1166 * is actually now running somewhere else! 1167 */ 1168 while (task_running(rq, p)) { 1169 if (match_state && unlikely(p->state != match_state)) 1170 return 0; 1171 cpu_relax(); 1172 } 1173 1174 /* 1175 * Ok, time to look more closely! We need the rq 1176 * lock now, to be *sure*. If we're wrong, we'll 1177 * just go back and repeat. 1178 */ 1179 rq = task_rq_lock(p, &flags); 1180 trace_sched_wait_task(p); 1181 running = task_running(rq, p); 1182 on_rq = p->on_rq; 1183 ncsw = 0; 1184 if (!match_state || p->state == match_state) 1185 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1186 task_rq_unlock(rq, p, &flags); 1187 1188 /* 1189 * If it changed from the expected state, bail out now. 1190 */ 1191 if (unlikely(!ncsw)) 1192 break; 1193 1194 /* 1195 * Was it really running after all now that we 1196 * checked with the proper locks actually held? 1197 * 1198 * Oops. Go back and try again.. 1199 */ 1200 if (unlikely(running)) { 1201 cpu_relax(); 1202 continue; 1203 } 1204 1205 /* 1206 * It's not enough that it's not actively running, 1207 * it must be off the runqueue _entirely_, and not 1208 * preempted! 1209 * 1210 * So if it was still runnable (but just not actively 1211 * running right now), it's preempted, and we should 1212 * yield - it could be a while. 1213 */ 1214 if (unlikely(on_rq)) { 1215 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1216 1217 set_current_state(TASK_UNINTERRUPTIBLE); 1218 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1219 continue; 1220 } 1221 1222 /* 1223 * Ahh, all good. It wasn't running, and it wasn't 1224 * runnable, which means that it will never become 1225 * running in the future either. We're all done! 1226 */ 1227 break; 1228 } 1229 1230 return ncsw; 1231} 1232 1233/*** 1234 * kick_process - kick a running thread to enter/exit the kernel 1235 * @p: the to-be-kicked thread 1236 * 1237 * Cause a process which is running on another CPU to enter 1238 * kernel-mode, without any delay. (to get signals handled.) 1239 * 1240 * NOTE: this function doesn't have to take the runqueue lock, 1241 * because all it wants to ensure is that the remote task enters 1242 * the kernel. If the IPI races and the task has been migrated 1243 * to another CPU then no harm is done and the purpose has been 1244 * achieved as well. 1245 */ 1246void kick_process(struct task_struct *p) 1247{ 1248 int cpu; 1249 1250 preempt_disable(); 1251 cpu = task_cpu(p); 1252 if ((cpu != smp_processor_id()) && task_curr(p)) 1253 smp_send_reschedule(cpu); 1254 preempt_enable(); 1255} 1256EXPORT_SYMBOL_GPL(kick_process); 1257#endif /* CONFIG_SMP */ 1258 1259#ifdef CONFIG_SMP 1260/* 1261 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1262 */ 1263static int select_fallback_rq(int cpu, struct task_struct *p) 1264{ 1265 int nid = cpu_to_node(cpu); 1266 const struct cpumask *nodemask = NULL; 1267 enum { cpuset, possible, fail } state = cpuset; 1268 int dest_cpu; 1269 1270 /* 1271 * If the node that the cpu is on has been offlined, cpu_to_node() 1272 * will return -1. There is no cpu on the node, and we should 1273 * select the cpu on the other node. 1274 */ 1275 if (nid != -1) { 1276 nodemask = cpumask_of_node(nid); 1277 1278 /* Look for allowed, online CPU in same node. */ 1279 for_each_cpu(dest_cpu, nodemask) { 1280 if (!cpu_online(dest_cpu)) 1281 continue; 1282 if (!cpu_active(dest_cpu)) 1283 continue; 1284 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1285 return dest_cpu; 1286 } 1287 } 1288 1289 for (;;) { 1290 /* Any allowed, online CPU? */ 1291 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1292 if (!cpu_online(dest_cpu)) 1293 continue; 1294 if (!cpu_active(dest_cpu)) 1295 continue; 1296 goto out; 1297 } 1298 1299 switch (state) { 1300 case cpuset: 1301 /* No more Mr. Nice Guy. */ 1302 cpuset_cpus_allowed_fallback(p); 1303 state = possible; 1304 break; 1305 1306 case possible: 1307 do_set_cpus_allowed(p, cpu_possible_mask); 1308 state = fail; 1309 break; 1310 1311 case fail: 1312 BUG(); 1313 break; 1314 } 1315 } 1316 1317out: 1318 if (state != cpuset) { 1319 /* 1320 * Don't tell them about moving exiting tasks or 1321 * kernel threads (both mm NULL), since they never 1322 * leave kernel. 1323 */ 1324 if (p->mm && printk_ratelimit()) { 1325 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1326 task_pid_nr(p), p->comm, cpu); 1327 } 1328 } 1329 1330 return dest_cpu; 1331} 1332 1333/* 1334 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1335 */ 1336static inline 1337int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1338{ 1339 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1340 1341 /* 1342 * In order not to call set_task_cpu() on a blocking task we need 1343 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1344 * cpu. 1345 * 1346 * Since this is common to all placement strategies, this lives here. 1347 * 1348 * [ this allows ->select_task() to simply return task_cpu(p) and 1349 * not worry about this generic constraint ] 1350 */ 1351 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1352 !cpu_online(cpu))) 1353 cpu = select_fallback_rq(task_cpu(p), p); 1354 1355 return cpu; 1356} 1357 1358static void update_avg(u64 *avg, u64 sample) 1359{ 1360 s64 diff = sample - *avg; 1361 *avg += diff >> 3; 1362} 1363#endif 1364 1365static void 1366ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1367{ 1368#ifdef CONFIG_SCHEDSTATS 1369 struct rq *rq = this_rq(); 1370 1371#ifdef CONFIG_SMP 1372 int this_cpu = smp_processor_id(); 1373 1374 if (cpu == this_cpu) { 1375 schedstat_inc(rq, ttwu_local); 1376 schedstat_inc(p, se.statistics.nr_wakeups_local); 1377 } else { 1378 struct sched_domain *sd; 1379 1380 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1381 rcu_read_lock(); 1382 for_each_domain(this_cpu, sd) { 1383 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1384 schedstat_inc(sd, ttwu_wake_remote); 1385 break; 1386 } 1387 } 1388 rcu_read_unlock(); 1389 } 1390 1391 if (wake_flags & WF_MIGRATED) 1392 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1393 1394#endif /* CONFIG_SMP */ 1395 1396 schedstat_inc(rq, ttwu_count); 1397 schedstat_inc(p, se.statistics.nr_wakeups); 1398 1399 if (wake_flags & WF_SYNC) 1400 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1401 1402#endif /* CONFIG_SCHEDSTATS */ 1403} 1404 1405static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1406{ 1407 activate_task(rq, p, en_flags); 1408 p->on_rq = 1; 1409 1410 /* if a worker is waking up, notify workqueue */ 1411 if (p->flags & PF_WQ_WORKER) 1412 wq_worker_waking_up(p, cpu_of(rq)); 1413} 1414 1415/* 1416 * Mark the task runnable and perform wakeup-preemption. 1417 */ 1418static void 1419ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1420{ 1421 check_preempt_curr(rq, p, wake_flags); 1422 trace_sched_wakeup(p, true); 1423 1424 p->state = TASK_RUNNING; 1425#ifdef CONFIG_SMP 1426 if (p->sched_class->task_woken) 1427 p->sched_class->task_woken(rq, p); 1428 1429 if (rq->idle_stamp) { 1430 u64 delta = rq_clock(rq) - rq->idle_stamp; 1431 u64 max = 2*rq->max_idle_balance_cost; 1432 1433 update_avg(&rq->avg_idle, delta); 1434 1435 if (rq->avg_idle > max) 1436 rq->avg_idle = max; 1437 1438 rq->idle_stamp = 0; 1439 } 1440#endif 1441} 1442 1443static void 1444ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1445{ 1446#ifdef CONFIG_SMP 1447 if (p->sched_contributes_to_load) 1448 rq->nr_uninterruptible--; 1449#endif 1450 1451 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1452 ttwu_do_wakeup(rq, p, wake_flags); 1453} 1454 1455/* 1456 * Called in case the task @p isn't fully descheduled from its runqueue, 1457 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1458 * since all we need to do is flip p->state to TASK_RUNNING, since 1459 * the task is still ->on_rq. 1460 */ 1461static int ttwu_remote(struct task_struct *p, int wake_flags) 1462{ 1463 struct rq *rq; 1464 int ret = 0; 1465 1466 rq = __task_rq_lock(p); 1467 if (p->on_rq) { 1468 /* check_preempt_curr() may use rq clock */ 1469 update_rq_clock(rq); 1470 ttwu_do_wakeup(rq, p, wake_flags); 1471 ret = 1; 1472 } 1473 __task_rq_unlock(rq); 1474 1475 return ret; 1476} 1477 1478#ifdef CONFIG_SMP 1479static void sched_ttwu_pending(void) 1480{ 1481 struct rq *rq = this_rq(); 1482 struct llist_node *llist = llist_del_all(&rq->wake_list); 1483 struct task_struct *p; 1484 1485 raw_spin_lock(&rq->lock); 1486 1487 while (llist) { 1488 p = llist_entry(llist, struct task_struct, wake_entry); 1489 llist = llist_next(llist); 1490 ttwu_do_activate(rq, p, 0); 1491 } 1492 1493 raw_spin_unlock(&rq->lock); 1494} 1495 1496void scheduler_ipi(void) 1497{ 1498 /* 1499 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1500 * TIF_NEED_RESCHED remotely (for the first time) will also send 1501 * this IPI. 1502 */ 1503 preempt_fold_need_resched(); 1504 1505 if (llist_empty(&this_rq()->wake_list) 1506 && !tick_nohz_full_cpu(smp_processor_id()) 1507 && !got_nohz_idle_kick()) 1508 return; 1509 1510 /* 1511 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1512 * traditionally all their work was done from the interrupt return 1513 * path. Now that we actually do some work, we need to make sure 1514 * we do call them. 1515 * 1516 * Some archs already do call them, luckily irq_enter/exit nest 1517 * properly. 1518 * 1519 * Arguably we should visit all archs and update all handlers, 1520 * however a fair share of IPIs are still resched only so this would 1521 * somewhat pessimize the simple resched case. 1522 */ 1523 irq_enter(); 1524 tick_nohz_full_check(); 1525 sched_ttwu_pending(); 1526 1527 /* 1528 * Check if someone kicked us for doing the nohz idle load balance. 1529 */ 1530 if (unlikely(got_nohz_idle_kick())) { 1531 this_rq()->idle_balance = 1; 1532 raise_softirq_irqoff(SCHED_SOFTIRQ); 1533 } 1534 irq_exit(); 1535} 1536 1537static void ttwu_queue_remote(struct task_struct *p, int cpu) 1538{ 1539 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1540 smp_send_reschedule(cpu); 1541} 1542 1543bool cpus_share_cache(int this_cpu, int that_cpu) 1544{ 1545 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1546} 1547#endif /* CONFIG_SMP */ 1548 1549static void ttwu_queue(struct task_struct *p, int cpu) 1550{ 1551 struct rq *rq = cpu_rq(cpu); 1552 1553#if defined(CONFIG_SMP) 1554 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1555 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1556 ttwu_queue_remote(p, cpu); 1557 return; 1558 } 1559#endif 1560 1561 raw_spin_lock(&rq->lock); 1562 ttwu_do_activate(rq, p, 0); 1563 raw_spin_unlock(&rq->lock); 1564} 1565 1566/** 1567 * try_to_wake_up - wake up a thread 1568 * @p: the thread to be awakened 1569 * @state: the mask of task states that can be woken 1570 * @wake_flags: wake modifier flags (WF_*) 1571 * 1572 * Put it on the run-queue if it's not already there. The "current" 1573 * thread is always on the run-queue (except when the actual 1574 * re-schedule is in progress), and as such you're allowed to do 1575 * the simpler "current->state = TASK_RUNNING" to mark yourself 1576 * runnable without the overhead of this. 1577 * 1578 * Return: %true if @p was woken up, %false if it was already running. 1579 * or @state didn't match @p's state. 1580 */ 1581static int 1582try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1583{ 1584 unsigned long flags; 1585 int cpu, success = 0; 1586 1587 /* 1588 * If we are going to wake up a thread waiting for CONDITION we 1589 * need to ensure that CONDITION=1 done by the caller can not be 1590 * reordered with p->state check below. This pairs with mb() in 1591 * set_current_state() the waiting thread does. 1592 */ 1593 smp_mb__before_spinlock(); 1594 raw_spin_lock_irqsave(&p->pi_lock, flags); 1595 if (!(p->state & state)) 1596 goto out; 1597 1598 success = 1; /* we're going to change ->state */ 1599 cpu = task_cpu(p); 1600 1601 if (p->on_rq && ttwu_remote(p, wake_flags)) 1602 goto stat; 1603 1604#ifdef CONFIG_SMP 1605 /* 1606 * If the owning (remote) cpu is still in the middle of schedule() with 1607 * this task as prev, wait until its done referencing the task. 1608 */ 1609 while (p->on_cpu) 1610 cpu_relax(); 1611 /* 1612 * Pairs with the smp_wmb() in finish_lock_switch(). 1613 */ 1614 smp_rmb(); 1615 1616 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1617 p->state = TASK_WAKING; 1618 1619 if (p->sched_class->task_waking) 1620 p->sched_class->task_waking(p); 1621 1622 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1623 if (task_cpu(p) != cpu) { 1624 wake_flags |= WF_MIGRATED; 1625 set_task_cpu(p, cpu); 1626 } 1627#endif /* CONFIG_SMP */ 1628 1629 ttwu_queue(p, cpu); 1630stat: 1631 ttwu_stat(p, cpu, wake_flags); 1632out: 1633 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1634 1635 return success; 1636} 1637 1638/** 1639 * try_to_wake_up_local - try to wake up a local task with rq lock held 1640 * @p: the thread to be awakened 1641 * 1642 * Put @p on the run-queue if it's not already there. The caller must 1643 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1644 * the current task. 1645 */ 1646static void try_to_wake_up_local(struct task_struct *p) 1647{ 1648 struct rq *rq = task_rq(p); 1649 1650 if (WARN_ON_ONCE(rq != this_rq()) || 1651 WARN_ON_ONCE(p == current)) 1652 return; 1653 1654 lockdep_assert_held(&rq->lock); 1655 1656 if (!raw_spin_trylock(&p->pi_lock)) { 1657 raw_spin_unlock(&rq->lock); 1658 raw_spin_lock(&p->pi_lock); 1659 raw_spin_lock(&rq->lock); 1660 } 1661 1662 if (!(p->state & TASK_NORMAL)) 1663 goto out; 1664 1665 if (!p->on_rq) 1666 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1667 1668 ttwu_do_wakeup(rq, p, 0); 1669 ttwu_stat(p, smp_processor_id(), 0); 1670out: 1671 raw_spin_unlock(&p->pi_lock); 1672} 1673 1674/** 1675 * wake_up_process - Wake up a specific process 1676 * @p: The process to be woken up. 1677 * 1678 * Attempt to wake up the nominated process and move it to the set of runnable 1679 * processes. 1680 * 1681 * Return: 1 if the process was woken up, 0 if it was already running. 1682 * 1683 * It may be assumed that this function implies a write memory barrier before 1684 * changing the task state if and only if any tasks are woken up. 1685 */ 1686int wake_up_process(struct task_struct *p) 1687{ 1688 WARN_ON(task_is_stopped_or_traced(p)); 1689 return try_to_wake_up(p, TASK_NORMAL, 0); 1690} 1691EXPORT_SYMBOL(wake_up_process); 1692 1693int wake_up_state(struct task_struct *p, unsigned int state) 1694{ 1695 return try_to_wake_up(p, state, 0); 1696} 1697 1698/* 1699 * Perform scheduler related setup for a newly forked process p. 1700 * p is forked by current. 1701 * 1702 * __sched_fork() is basic setup used by init_idle() too: 1703 */ 1704static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1705{ 1706 p->on_rq = 0; 1707 1708 p->se.on_rq = 0; 1709 p->se.exec_start = 0; 1710 p->se.sum_exec_runtime = 0; 1711 p->se.prev_sum_exec_runtime = 0; 1712 p->se.nr_migrations = 0; 1713 p->se.vruntime = 0; 1714 INIT_LIST_HEAD(&p->se.group_node); 1715 1716#ifdef CONFIG_SCHEDSTATS 1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1718#endif 1719 1720 RB_CLEAR_NODE(&p->dl.rb_node); 1721 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1722 p->dl.dl_runtime = p->dl.runtime = 0; 1723 p->dl.dl_deadline = p->dl.deadline = 0; 1724 p->dl.dl_period = 0; 1725 p->dl.flags = 0; 1726 1727 INIT_LIST_HEAD(&p->rt.run_list); 1728 1729#ifdef CONFIG_PREEMPT_NOTIFIERS 1730 INIT_HLIST_HEAD(&p->preempt_notifiers); 1731#endif 1732 1733#ifdef CONFIG_NUMA_BALANCING 1734 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1735 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1736 p->mm->numa_scan_seq = 0; 1737 } 1738 1739 if (clone_flags & CLONE_VM) 1740 p->numa_preferred_nid = current->numa_preferred_nid; 1741 else 1742 p->numa_preferred_nid = -1; 1743 1744 p->node_stamp = 0ULL; 1745 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1746 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1747 p->numa_work.next = &p->numa_work; 1748 p->numa_faults_memory = NULL; 1749 p->numa_faults_buffer_memory = NULL; 1750 p->last_task_numa_placement = 0; 1751 p->last_sum_exec_runtime = 0; 1752 1753 INIT_LIST_HEAD(&p->numa_entry); 1754 p->numa_group = NULL; 1755#endif /* CONFIG_NUMA_BALANCING */ 1756} 1757 1758#ifdef CONFIG_NUMA_BALANCING 1759#ifdef CONFIG_SCHED_DEBUG 1760void set_numabalancing_state(bool enabled) 1761{ 1762 if (enabled) 1763 sched_feat_set("NUMA"); 1764 else 1765 sched_feat_set("NO_NUMA"); 1766} 1767#else 1768__read_mostly bool numabalancing_enabled; 1769 1770void set_numabalancing_state(bool enabled) 1771{ 1772 numabalancing_enabled = enabled; 1773} 1774#endif /* CONFIG_SCHED_DEBUG */ 1775 1776#ifdef CONFIG_PROC_SYSCTL 1777int sysctl_numa_balancing(struct ctl_table *table, int write, 1778 void __user *buffer, size_t *lenp, loff_t *ppos) 1779{ 1780 struct ctl_table t; 1781 int err; 1782 int state = numabalancing_enabled; 1783 1784 if (write && !capable(CAP_SYS_ADMIN)) 1785 return -EPERM; 1786 1787 t = *table; 1788 t.data = &state; 1789 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1790 if (err < 0) 1791 return err; 1792 if (write) 1793 set_numabalancing_state(state); 1794 return err; 1795} 1796#endif 1797#endif 1798 1799/* 1800 * fork()/clone()-time setup: 1801 */ 1802int sched_fork(unsigned long clone_flags, struct task_struct *p) 1803{ 1804 unsigned long flags; 1805 int cpu = get_cpu(); 1806 1807 __sched_fork(clone_flags, p); 1808 /* 1809 * We mark the process as running here. This guarantees that 1810 * nobody will actually run it, and a signal or other external 1811 * event cannot wake it up and insert it on the runqueue either. 1812 */ 1813 p->state = TASK_RUNNING; 1814 1815 /* 1816 * Make sure we do not leak PI boosting priority to the child. 1817 */ 1818 p->prio = current->normal_prio; 1819 1820 /* 1821 * Revert to default priority/policy on fork if requested. 1822 */ 1823 if (unlikely(p->sched_reset_on_fork)) { 1824 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1825 p->policy = SCHED_NORMAL; 1826 p->static_prio = NICE_TO_PRIO(0); 1827 p->rt_priority = 0; 1828 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1829 p->static_prio = NICE_TO_PRIO(0); 1830 1831 p->prio = p->normal_prio = __normal_prio(p); 1832 set_load_weight(p); 1833 1834 /* 1835 * We don't need the reset flag anymore after the fork. It has 1836 * fulfilled its duty: 1837 */ 1838 p->sched_reset_on_fork = 0; 1839 } 1840 1841 if (dl_prio(p->prio)) { 1842 put_cpu(); 1843 return -EAGAIN; 1844 } else if (rt_prio(p->prio)) { 1845 p->sched_class = &rt_sched_class; 1846 } else { 1847 p->sched_class = &fair_sched_class; 1848 } 1849 1850 if (p->sched_class->task_fork) 1851 p->sched_class->task_fork(p); 1852 1853 /* 1854 * The child is not yet in the pid-hash so no cgroup attach races, 1855 * and the cgroup is pinned to this child due to cgroup_fork() 1856 * is ran before sched_fork(). 1857 * 1858 * Silence PROVE_RCU. 1859 */ 1860 raw_spin_lock_irqsave(&p->pi_lock, flags); 1861 set_task_cpu(p, cpu); 1862 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1863 1864#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1865 if (likely(sched_info_on())) 1866 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1867#endif 1868#if defined(CONFIG_SMP) 1869 p->on_cpu = 0; 1870#endif 1871 init_task_preempt_count(p); 1872#ifdef CONFIG_SMP 1873 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1874 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1875#endif 1876 1877 put_cpu(); 1878 return 0; 1879} 1880 1881unsigned long to_ratio(u64 period, u64 runtime) 1882{ 1883 if (runtime == RUNTIME_INF) 1884 return 1ULL << 20; 1885 1886 /* 1887 * Doing this here saves a lot of checks in all 1888 * the calling paths, and returning zero seems 1889 * safe for them anyway. 1890 */ 1891 if (period == 0) 1892 return 0; 1893 1894 return div64_u64(runtime << 20, period); 1895} 1896 1897#ifdef CONFIG_SMP 1898inline struct dl_bw *dl_bw_of(int i) 1899{ 1900 return &cpu_rq(i)->rd->dl_bw; 1901} 1902 1903static inline int dl_bw_cpus(int i) 1904{ 1905 struct root_domain *rd = cpu_rq(i)->rd; 1906 int cpus = 0; 1907 1908 for_each_cpu_and(i, rd->span, cpu_active_mask) 1909 cpus++; 1910 1911 return cpus; 1912} 1913#else 1914inline struct dl_bw *dl_bw_of(int i) 1915{ 1916 return &cpu_rq(i)->dl.dl_bw; 1917} 1918 1919static inline int dl_bw_cpus(int i) 1920{ 1921 return 1; 1922} 1923#endif 1924 1925static inline 1926void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1927{ 1928 dl_b->total_bw -= tsk_bw; 1929} 1930 1931static inline 1932void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1933{ 1934 dl_b->total_bw += tsk_bw; 1935} 1936 1937static inline 1938bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1939{ 1940 return dl_b->bw != -1 && 1941 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1942} 1943 1944/* 1945 * We must be sure that accepting a new task (or allowing changing the 1946 * parameters of an existing one) is consistent with the bandwidth 1947 * constraints. If yes, this function also accordingly updates the currently 1948 * allocated bandwidth to reflect the new situation. 1949 * 1950 * This function is called while holding p's rq->lock. 1951 */ 1952static int dl_overflow(struct task_struct *p, int policy, 1953 const struct sched_attr *attr) 1954{ 1955 1956 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1957 u64 period = attr->sched_period; 1958 u64 runtime = attr->sched_runtime; 1959 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1960 int cpus, err = -1; 1961 1962 if (new_bw == p->dl.dl_bw) 1963 return 0; 1964 1965 /* 1966 * Either if a task, enters, leave, or stays -deadline but changes 1967 * its parameters, we may need to update accordingly the total 1968 * allocated bandwidth of the container. 1969 */ 1970 raw_spin_lock(&dl_b->lock); 1971 cpus = dl_bw_cpus(task_cpu(p)); 1972 if (dl_policy(policy) && !task_has_dl_policy(p) && 1973 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1974 __dl_add(dl_b, new_bw); 1975 err = 0; 1976 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1977 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1978 __dl_clear(dl_b, p->dl.dl_bw); 1979 __dl_add(dl_b, new_bw); 1980 err = 0; 1981 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1982 __dl_clear(dl_b, p->dl.dl_bw); 1983 err = 0; 1984 } 1985 raw_spin_unlock(&dl_b->lock); 1986 1987 return err; 1988} 1989 1990extern void init_dl_bw(struct dl_bw *dl_b); 1991 1992/* 1993 * wake_up_new_task - wake up a newly created task for the first time. 1994 * 1995 * This function will do some initial scheduler statistics housekeeping 1996 * that must be done for every newly created context, then puts the task 1997 * on the runqueue and wakes it. 1998 */ 1999void wake_up_new_task(struct task_struct *p) 2000{ 2001 unsigned long flags; 2002 struct rq *rq; 2003 2004 raw_spin_lock_irqsave(&p->pi_lock, flags); 2005#ifdef CONFIG_SMP 2006 /* 2007 * Fork balancing, do it here and not earlier because: 2008 * - cpus_allowed can change in the fork path 2009 * - any previously selected cpu might disappear through hotplug 2010 */ 2011 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2012#endif 2013 2014 /* Initialize new task's runnable average */ 2015 init_task_runnable_average(p); 2016 rq = __task_rq_lock(p); 2017 activate_task(rq, p, 0); 2018 p->on_rq = 1; 2019 trace_sched_wakeup_new(p, true); 2020 check_preempt_curr(rq, p, WF_FORK); 2021#ifdef CONFIG_SMP 2022 if (p->sched_class->task_woken) 2023 p->sched_class->task_woken(rq, p); 2024#endif 2025 task_rq_unlock(rq, p, &flags); 2026} 2027 2028#ifdef CONFIG_PREEMPT_NOTIFIERS 2029 2030/** 2031 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2032 * @notifier: notifier struct to register 2033 */ 2034void preempt_notifier_register(struct preempt_notifier *notifier) 2035{ 2036 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2037} 2038EXPORT_SYMBOL_GPL(preempt_notifier_register); 2039 2040/** 2041 * preempt_notifier_unregister - no longer interested in preemption notifications 2042 * @notifier: notifier struct to unregister 2043 * 2044 * This is safe to call from within a preemption notifier. 2045 */ 2046void preempt_notifier_unregister(struct preempt_notifier *notifier) 2047{ 2048 hlist_del(¬ifier->link); 2049} 2050EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2051 2052static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2053{ 2054 struct preempt_notifier *notifier; 2055 2056 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2057 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2058} 2059 2060static void 2061fire_sched_out_preempt_notifiers(struct task_struct *curr, 2062 struct task_struct *next) 2063{ 2064 struct preempt_notifier *notifier; 2065 2066 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2067 notifier->ops->sched_out(notifier, next); 2068} 2069 2070#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2071 2072static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2073{ 2074} 2075 2076static void 2077fire_sched_out_preempt_notifiers(struct task_struct *curr, 2078 struct task_struct *next) 2079{ 2080} 2081 2082#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2083 2084/** 2085 * prepare_task_switch - prepare to switch tasks 2086 * @rq: the runqueue preparing to switch 2087 * @prev: the current task that is being switched out 2088 * @next: the task we are going to switch to. 2089 * 2090 * This is called with the rq lock held and interrupts off. It must 2091 * be paired with a subsequent finish_task_switch after the context 2092 * switch. 2093 * 2094 * prepare_task_switch sets up locking and calls architecture specific 2095 * hooks. 2096 */ 2097static inline void 2098prepare_task_switch(struct rq *rq, struct task_struct *prev, 2099 struct task_struct *next) 2100{ 2101 trace_sched_switch(prev, next); 2102 sched_info_switch(rq, prev, next); 2103 perf_event_task_sched_out(prev, next); 2104 fire_sched_out_preempt_notifiers(prev, next); 2105 prepare_lock_switch(rq, next); 2106 prepare_arch_switch(next); 2107} 2108 2109/** 2110 * finish_task_switch - clean up after a task-switch 2111 * @rq: runqueue associated with task-switch 2112 * @prev: the thread we just switched away from. 2113 * 2114 * finish_task_switch must be called after the context switch, paired 2115 * with a prepare_task_switch call before the context switch. 2116 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2117 * and do any other architecture-specific cleanup actions. 2118 * 2119 * Note that we may have delayed dropping an mm in context_switch(). If 2120 * so, we finish that here outside of the runqueue lock. (Doing it 2121 * with the lock held can cause deadlocks; see schedule() for 2122 * details.) 2123 */ 2124static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2125 __releases(rq->lock) 2126{ 2127 struct mm_struct *mm = rq->prev_mm; 2128 long prev_state; 2129 2130 rq->prev_mm = NULL; 2131 2132 /* 2133 * A task struct has one reference for the use as "current". 2134 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2135 * schedule one last time. The schedule call will never return, and 2136 * the scheduled task must drop that reference. 2137 * The test for TASK_DEAD must occur while the runqueue locks are 2138 * still held, otherwise prev could be scheduled on another cpu, die 2139 * there before we look at prev->state, and then the reference would 2140 * be dropped twice. 2141 * Manfred Spraul <manfred@colorfullife.com> 2142 */ 2143 prev_state = prev->state; 2144 vtime_task_switch(prev); 2145 finish_arch_switch(prev); 2146 perf_event_task_sched_in(prev, current); 2147 finish_lock_switch(rq, prev); 2148 finish_arch_post_lock_switch(); 2149 2150 fire_sched_in_preempt_notifiers(current); 2151 if (mm) 2152 mmdrop(mm); 2153 if (unlikely(prev_state == TASK_DEAD)) { 2154 task_numa_free(prev); 2155 2156 if (prev->sched_class->task_dead) 2157 prev->sched_class->task_dead(prev); 2158 2159 /* 2160 * Remove function-return probe instances associated with this 2161 * task and put them back on the free list. 2162 */ 2163 kprobe_flush_task(prev); 2164 put_task_struct(prev); 2165 } 2166 2167 tick_nohz_task_switch(current); 2168} 2169 2170#ifdef CONFIG_SMP 2171 2172/* rq->lock is NOT held, but preemption is disabled */ 2173static inline void post_schedule(struct rq *rq) 2174{ 2175 if (rq->post_schedule) { 2176 unsigned long flags; 2177 2178 raw_spin_lock_irqsave(&rq->lock, flags); 2179 if (rq->curr->sched_class->post_schedule) 2180 rq->curr->sched_class->post_schedule(rq); 2181 raw_spin_unlock_irqrestore(&rq->lock, flags); 2182 2183 rq->post_schedule = 0; 2184 } 2185} 2186 2187#else 2188 2189static inline void post_schedule(struct rq *rq) 2190{ 2191} 2192 2193#endif 2194 2195/** 2196 * schedule_tail - first thing a freshly forked thread must call. 2197 * @prev: the thread we just switched away from. 2198 */ 2199asmlinkage void schedule_tail(struct task_struct *prev) 2200 __releases(rq->lock) 2201{ 2202 struct rq *rq = this_rq(); 2203 2204 finish_task_switch(rq, prev); 2205 2206 /* 2207 * FIXME: do we need to worry about rq being invalidated by the 2208 * task_switch? 2209 */ 2210 post_schedule(rq); 2211 2212#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2213 /* In this case, finish_task_switch does not reenable preemption */ 2214 preempt_enable(); 2215#endif 2216 if (current->set_child_tid) 2217 put_user(task_pid_vnr(current), current->set_child_tid); 2218} 2219 2220/* 2221 * context_switch - switch to the new MM and the new 2222 * thread's register state. 2223 */ 2224static inline void 2225context_switch(struct rq *rq, struct task_struct *prev, 2226 struct task_struct *next) 2227{ 2228 struct mm_struct *mm, *oldmm; 2229 2230 prepare_task_switch(rq, prev, next); 2231 2232 mm = next->mm; 2233 oldmm = prev->active_mm; 2234 /* 2235 * For paravirt, this is coupled with an exit in switch_to to 2236 * combine the page table reload and the switch backend into 2237 * one hypercall. 2238 */ 2239 arch_start_context_switch(prev); 2240 2241 if (!mm) { 2242 next->active_mm = oldmm; 2243 atomic_inc(&oldmm->mm_count); 2244 enter_lazy_tlb(oldmm, next); 2245 } else 2246 switch_mm(oldmm, mm, next); 2247 2248 if (!prev->mm) { 2249 prev->active_mm = NULL; 2250 rq->prev_mm = oldmm; 2251 } 2252 /* 2253 * Since the runqueue lock will be released by the next 2254 * task (which is an invalid locking op but in the case 2255 * of the scheduler it's an obvious special-case), so we 2256 * do an early lockdep release here: 2257 */ 2258#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2259 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2260#endif 2261 2262 context_tracking_task_switch(prev, next); 2263 /* Here we just switch the register state and the stack. */ 2264 switch_to(prev, next, prev); 2265 2266 barrier(); 2267 /* 2268 * this_rq must be evaluated again because prev may have moved 2269 * CPUs since it called schedule(), thus the 'rq' on its stack 2270 * frame will be invalid. 2271 */ 2272 finish_task_switch(this_rq(), prev); 2273} 2274 2275/* 2276 * nr_running and nr_context_switches: 2277 * 2278 * externally visible scheduler statistics: current number of runnable 2279 * threads, total number of context switches performed since bootup. 2280 */ 2281unsigned long nr_running(void) 2282{ 2283 unsigned long i, sum = 0; 2284 2285 for_each_online_cpu(i) 2286 sum += cpu_rq(i)->nr_running; 2287 2288 return sum; 2289} 2290 2291unsigned long long nr_context_switches(void) 2292{ 2293 int i; 2294 unsigned long long sum = 0; 2295 2296 for_each_possible_cpu(i) 2297 sum += cpu_rq(i)->nr_switches; 2298 2299 return sum; 2300} 2301 2302unsigned long nr_iowait(void) 2303{ 2304 unsigned long i, sum = 0; 2305 2306 for_each_possible_cpu(i) 2307 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2308 2309 return sum; 2310} 2311 2312unsigned long nr_iowait_cpu(int cpu) 2313{ 2314 struct rq *this = cpu_rq(cpu); 2315 return atomic_read(&this->nr_iowait); 2316} 2317 2318#ifdef CONFIG_SMP 2319 2320/* 2321 * sched_exec - execve() is a valuable balancing opportunity, because at 2322 * this point the task has the smallest effective memory and cache footprint. 2323 */ 2324void sched_exec(void) 2325{ 2326 struct task_struct *p = current; 2327 unsigned long flags; 2328 int dest_cpu; 2329 2330 raw_spin_lock_irqsave(&p->pi_lock, flags); 2331 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2332 if (dest_cpu == smp_processor_id()) 2333 goto unlock; 2334 2335 if (likely(cpu_active(dest_cpu))) { 2336 struct migration_arg arg = { p, dest_cpu }; 2337 2338 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2339 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2340 return; 2341 } 2342unlock: 2343 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2344} 2345 2346#endif 2347 2348DEFINE_PER_CPU(struct kernel_stat, kstat); 2349DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2350 2351EXPORT_PER_CPU_SYMBOL(kstat); 2352EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2353 2354/* 2355 * Return any ns on the sched_clock that have not yet been accounted in 2356 * @p in case that task is currently running. 2357 * 2358 * Called with task_rq_lock() held on @rq. 2359 */ 2360static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2361{ 2362 u64 ns = 0; 2363 2364 if (task_current(rq, p)) { 2365 update_rq_clock(rq); 2366 ns = rq_clock_task(rq) - p->se.exec_start; 2367 if ((s64)ns < 0) 2368 ns = 0; 2369 } 2370 2371 return ns; 2372} 2373 2374unsigned long long task_delta_exec(struct task_struct *p) 2375{ 2376 unsigned long flags; 2377 struct rq *rq; 2378 u64 ns = 0; 2379 2380 rq = task_rq_lock(p, &flags); 2381 ns = do_task_delta_exec(p, rq); 2382 task_rq_unlock(rq, p, &flags); 2383 2384 return ns; 2385} 2386 2387/* 2388 * Return accounted runtime for the task. 2389 * In case the task is currently running, return the runtime plus current's 2390 * pending runtime that have not been accounted yet. 2391 */ 2392unsigned long long task_sched_runtime(struct task_struct *p) 2393{ 2394 unsigned long flags; 2395 struct rq *rq; 2396 u64 ns = 0; 2397 2398#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2399 /* 2400 * 64-bit doesn't need locks to atomically read a 64bit value. 2401 * So we have a optimization chance when the task's delta_exec is 0. 2402 * Reading ->on_cpu is racy, but this is ok. 2403 * 2404 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2405 * If we race with it entering cpu, unaccounted time is 0. This is 2406 * indistinguishable from the read occurring a few cycles earlier. 2407 */ 2408 if (!p->on_cpu) 2409 return p->se.sum_exec_runtime; 2410#endif 2411 2412 rq = task_rq_lock(p, &flags); 2413 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2414 task_rq_unlock(rq, p, &flags); 2415 2416 return ns; 2417} 2418 2419/* 2420 * This function gets called by the timer code, with HZ frequency. 2421 * We call it with interrupts disabled. 2422 */ 2423void scheduler_tick(void) 2424{ 2425 int cpu = smp_processor_id(); 2426 struct rq *rq = cpu_rq(cpu); 2427 struct task_struct *curr = rq->curr; 2428 2429 sched_clock_tick(); 2430 2431 raw_spin_lock(&rq->lock); 2432 update_rq_clock(rq); 2433 curr->sched_class->task_tick(rq, curr, 0); 2434 update_cpu_load_active(rq); 2435 raw_spin_unlock(&rq->lock); 2436 2437 perf_event_task_tick(); 2438 2439#ifdef CONFIG_SMP 2440 rq->idle_balance = idle_cpu(cpu); 2441 trigger_load_balance(rq); 2442#endif 2443 rq_last_tick_reset(rq); 2444} 2445 2446#ifdef CONFIG_NO_HZ_FULL 2447/** 2448 * scheduler_tick_max_deferment 2449 * 2450 * Keep at least one tick per second when a single 2451 * active task is running because the scheduler doesn't 2452 * yet completely support full dynticks environment. 2453 * 2454 * This makes sure that uptime, CFS vruntime, load 2455 * balancing, etc... continue to move forward, even 2456 * with a very low granularity. 2457 * 2458 * Return: Maximum deferment in nanoseconds. 2459 */ 2460u64 scheduler_tick_max_deferment(void) 2461{ 2462 struct rq *rq = this_rq(); 2463 unsigned long next, now = ACCESS_ONCE(jiffies); 2464 2465 next = rq->last_sched_tick + HZ; 2466 2467 if (time_before_eq(next, now)) 2468 return 0; 2469 2470 return jiffies_to_nsecs(next - now); 2471} 2472#endif 2473 2474notrace unsigned long get_parent_ip(unsigned long addr) 2475{ 2476 if (in_lock_functions(addr)) { 2477 addr = CALLER_ADDR2; 2478 if (in_lock_functions(addr)) 2479 addr = CALLER_ADDR3; 2480 } 2481 return addr; 2482} 2483 2484#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2485 defined(CONFIG_PREEMPT_TRACER)) 2486 2487void __kprobes preempt_count_add(int val) 2488{ 2489#ifdef CONFIG_DEBUG_PREEMPT 2490 /* 2491 * Underflow? 2492 */ 2493 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2494 return; 2495#endif 2496 __preempt_count_add(val); 2497#ifdef CONFIG_DEBUG_PREEMPT 2498 /* 2499 * Spinlock count overflowing soon? 2500 */ 2501 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2502 PREEMPT_MASK - 10); 2503#endif 2504 if (preempt_count() == val) 2505 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2506} 2507EXPORT_SYMBOL(preempt_count_add); 2508 2509void __kprobes preempt_count_sub(int val) 2510{ 2511#ifdef CONFIG_DEBUG_PREEMPT 2512 /* 2513 * Underflow? 2514 */ 2515 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2516 return; 2517 /* 2518 * Is the spinlock portion underflowing? 2519 */ 2520 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2521 !(preempt_count() & PREEMPT_MASK))) 2522 return; 2523#endif 2524 2525 if (preempt_count() == val) 2526 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2527 __preempt_count_sub(val); 2528} 2529EXPORT_SYMBOL(preempt_count_sub); 2530 2531#endif 2532 2533/* 2534 * Print scheduling while atomic bug: 2535 */ 2536static noinline void __schedule_bug(struct task_struct *prev) 2537{ 2538 if (oops_in_progress) 2539 return; 2540 2541 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2542 prev->comm, prev->pid, preempt_count()); 2543 2544 debug_show_held_locks(prev); 2545 print_modules(); 2546 if (irqs_disabled()) 2547 print_irqtrace_events(prev); 2548 dump_stack(); 2549 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2550} 2551 2552/* 2553 * Various schedule()-time debugging checks and statistics: 2554 */ 2555static inline void schedule_debug(struct task_struct *prev) 2556{ 2557 /* 2558 * Test if we are atomic. Since do_exit() needs to call into 2559 * schedule() atomically, we ignore that path. Otherwise whine 2560 * if we are scheduling when we should not. 2561 */ 2562 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2563 __schedule_bug(prev); 2564 rcu_sleep_check(); 2565 2566 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2567 2568 schedstat_inc(this_rq(), sched_count); 2569} 2570 2571/* 2572 * Pick up the highest-prio task: 2573 */ 2574static inline struct task_struct * 2575pick_next_task(struct rq *rq, struct task_struct *prev) 2576{ 2577 const struct sched_class *class; 2578 struct task_struct *p; 2579 2580 /* 2581 * Optimization: we know that if all tasks are in 2582 * the fair class we can call that function directly: 2583 */ 2584 if (likely(prev->sched_class == &fair_sched_class && 2585 rq->nr_running == rq->cfs.h_nr_running)) { 2586 p = fair_sched_class.pick_next_task(rq, prev); 2587 if (likely(p)) 2588 return p; 2589 } 2590 2591 for_each_class(class) { 2592 p = class->pick_next_task(rq, prev); 2593 if (p) 2594 return p; 2595 } 2596 2597 BUG(); /* the idle class will always have a runnable task */ 2598} 2599 2600/* 2601 * __schedule() is the main scheduler function. 2602 * 2603 * The main means of driving the scheduler and thus entering this function are: 2604 * 2605 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2606 * 2607 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2608 * paths. For example, see arch/x86/entry_64.S. 2609 * 2610 * To drive preemption between tasks, the scheduler sets the flag in timer 2611 * interrupt handler scheduler_tick(). 2612 * 2613 * 3. Wakeups don't really cause entry into schedule(). They add a 2614 * task to the run-queue and that's it. 2615 * 2616 * Now, if the new task added to the run-queue preempts the current 2617 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2618 * called on the nearest possible occasion: 2619 * 2620 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2621 * 2622 * - in syscall or exception context, at the next outmost 2623 * preempt_enable(). (this might be as soon as the wake_up()'s 2624 * spin_unlock()!) 2625 * 2626 * - in IRQ context, return from interrupt-handler to 2627 * preemptible context 2628 * 2629 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2630 * then at the next: 2631 * 2632 * - cond_resched() call 2633 * - explicit schedule() call 2634 * - return from syscall or exception to user-space 2635 * - return from interrupt-handler to user-space 2636 */ 2637static void __sched __schedule(void) 2638{ 2639 struct task_struct *prev, *next; 2640 unsigned long *switch_count; 2641 struct rq *rq; 2642 int cpu; 2643 2644need_resched: 2645 preempt_disable(); 2646 cpu = smp_processor_id(); 2647 rq = cpu_rq(cpu); 2648 rcu_note_context_switch(cpu); 2649 prev = rq->curr; 2650 2651 schedule_debug(prev); 2652 2653 if (sched_feat(HRTICK)) 2654 hrtick_clear(rq); 2655 2656 /* 2657 * Make sure that signal_pending_state()->signal_pending() below 2658 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2659 * done by the caller to avoid the race with signal_wake_up(). 2660 */ 2661 smp_mb__before_spinlock(); 2662 raw_spin_lock_irq(&rq->lock); 2663 2664 switch_count = &prev->nivcsw; 2665 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2666 if (unlikely(signal_pending_state(prev->state, prev))) { 2667 prev->state = TASK_RUNNING; 2668 } else { 2669 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2670 prev->on_rq = 0; 2671 2672 /* 2673 * If a worker went to sleep, notify and ask workqueue 2674 * whether it wants to wake up a task to maintain 2675 * concurrency. 2676 */ 2677 if (prev->flags & PF_WQ_WORKER) { 2678 struct task_struct *to_wakeup; 2679 2680 to_wakeup = wq_worker_sleeping(prev, cpu); 2681 if (to_wakeup) 2682 try_to_wake_up_local(to_wakeup); 2683 } 2684 } 2685 switch_count = &prev->nvcsw; 2686 } 2687 2688 if (prev->on_rq || rq->skip_clock_update < 0) 2689 update_rq_clock(rq); 2690 2691 next = pick_next_task(rq, prev); 2692 clear_tsk_need_resched(prev); 2693 clear_preempt_need_resched(); 2694 rq->skip_clock_update = 0; 2695 2696 if (likely(prev != next)) { 2697 rq->nr_switches++; 2698 rq->curr = next; 2699 ++*switch_count; 2700 2701 context_switch(rq, prev, next); /* unlocks the rq */ 2702 /* 2703 * The context switch have flipped the stack from under us 2704 * and restored the local variables which were saved when 2705 * this task called schedule() in the past. prev == current 2706 * is still correct, but it can be moved to another cpu/rq. 2707 */ 2708 cpu = smp_processor_id(); 2709 rq = cpu_rq(cpu); 2710 } else 2711 raw_spin_unlock_irq(&rq->lock); 2712 2713 post_schedule(rq); 2714 2715 sched_preempt_enable_no_resched(); 2716 if (need_resched()) 2717 goto need_resched; 2718} 2719 2720static inline void sched_submit_work(struct task_struct *tsk) 2721{ 2722 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2723 return; 2724 /* 2725 * If we are going to sleep and we have plugged IO queued, 2726 * make sure to submit it to avoid deadlocks. 2727 */ 2728 if (blk_needs_flush_plug(tsk)) 2729 blk_schedule_flush_plug(tsk); 2730} 2731 2732asmlinkage void __sched schedule(void) 2733{ 2734 struct task_struct *tsk = current; 2735 2736 sched_submit_work(tsk); 2737 __schedule(); 2738} 2739EXPORT_SYMBOL(schedule); 2740 2741#ifdef CONFIG_CONTEXT_TRACKING 2742asmlinkage void __sched schedule_user(void) 2743{ 2744 /* 2745 * If we come here after a random call to set_need_resched(), 2746 * or we have been woken up remotely but the IPI has not yet arrived, 2747 * we haven't yet exited the RCU idle mode. Do it here manually until 2748 * we find a better solution. 2749 */ 2750 user_exit(); 2751 schedule(); 2752 user_enter(); 2753} 2754#endif 2755 2756/** 2757 * schedule_preempt_disabled - called with preemption disabled 2758 * 2759 * Returns with preemption disabled. Note: preempt_count must be 1 2760 */ 2761void __sched schedule_preempt_disabled(void) 2762{ 2763 sched_preempt_enable_no_resched(); 2764 schedule(); 2765 preempt_disable(); 2766} 2767 2768#ifdef CONFIG_PREEMPT 2769/* 2770 * this is the entry point to schedule() from in-kernel preemption 2771 * off of preempt_enable. Kernel preemptions off return from interrupt 2772 * occur there and call schedule directly. 2773 */ 2774asmlinkage void __sched notrace preempt_schedule(void) 2775{ 2776 /* 2777 * If there is a non-zero preempt_count or interrupts are disabled, 2778 * we do not want to preempt the current task. Just return.. 2779 */ 2780 if (likely(!preemptible())) 2781 return; 2782 2783 do { 2784 __preempt_count_add(PREEMPT_ACTIVE); 2785 __schedule(); 2786 __preempt_count_sub(PREEMPT_ACTIVE); 2787 2788 /* 2789 * Check again in case we missed a preemption opportunity 2790 * between schedule and now. 2791 */ 2792 barrier(); 2793 } while (need_resched()); 2794} 2795EXPORT_SYMBOL(preempt_schedule); 2796#endif /* CONFIG_PREEMPT */ 2797 2798/* 2799 * this is the entry point to schedule() from kernel preemption 2800 * off of irq context. 2801 * Note, that this is called and return with irqs disabled. This will 2802 * protect us against recursive calling from irq. 2803 */ 2804asmlinkage void __sched preempt_schedule_irq(void) 2805{ 2806 enum ctx_state prev_state; 2807 2808 /* Catch callers which need to be fixed */ 2809 BUG_ON(preempt_count() || !irqs_disabled()); 2810 2811 prev_state = exception_enter(); 2812 2813 do { 2814 __preempt_count_add(PREEMPT_ACTIVE); 2815 local_irq_enable(); 2816 __schedule(); 2817 local_irq_disable(); 2818 __preempt_count_sub(PREEMPT_ACTIVE); 2819 2820 /* 2821 * Check again in case we missed a preemption opportunity 2822 * between schedule and now. 2823 */ 2824 barrier(); 2825 } while (need_resched()); 2826 2827 exception_exit(prev_state); 2828} 2829 2830int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2831 void *key) 2832{ 2833 return try_to_wake_up(curr->private, mode, wake_flags); 2834} 2835EXPORT_SYMBOL(default_wake_function); 2836 2837static long __sched 2838sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2839{ 2840 unsigned long flags; 2841 wait_queue_t wait; 2842 2843 init_waitqueue_entry(&wait, current); 2844 2845 __set_current_state(state); 2846 2847 spin_lock_irqsave(&q->lock, flags); 2848 __add_wait_queue(q, &wait); 2849 spin_unlock(&q->lock); 2850 timeout = schedule_timeout(timeout); 2851 spin_lock_irq(&q->lock); 2852 __remove_wait_queue(q, &wait); 2853 spin_unlock_irqrestore(&q->lock, flags); 2854 2855 return timeout; 2856} 2857 2858void __sched interruptible_sleep_on(wait_queue_head_t *q) 2859{ 2860 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2861} 2862EXPORT_SYMBOL(interruptible_sleep_on); 2863 2864long __sched 2865interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2866{ 2867 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2868} 2869EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2870 2871void __sched sleep_on(wait_queue_head_t *q) 2872{ 2873 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2874} 2875EXPORT_SYMBOL(sleep_on); 2876 2877long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2878{ 2879 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2880} 2881EXPORT_SYMBOL(sleep_on_timeout); 2882 2883#ifdef CONFIG_RT_MUTEXES 2884 2885/* 2886 * rt_mutex_setprio - set the current priority of a task 2887 * @p: task 2888 * @prio: prio value (kernel-internal form) 2889 * 2890 * This function changes the 'effective' priority of a task. It does 2891 * not touch ->normal_prio like __setscheduler(). 2892 * 2893 * Used by the rt_mutex code to implement priority inheritance logic. 2894 */ 2895void rt_mutex_setprio(struct task_struct *p, int prio) 2896{ 2897 int oldprio, on_rq, running, enqueue_flag = 0; 2898 struct rq *rq; 2899 const struct sched_class *prev_class; 2900 2901 BUG_ON(prio > MAX_PRIO); 2902 2903 rq = __task_rq_lock(p); 2904 2905 /* 2906 * Idle task boosting is a nono in general. There is one 2907 * exception, when PREEMPT_RT and NOHZ is active: 2908 * 2909 * The idle task calls get_next_timer_interrupt() and holds 2910 * the timer wheel base->lock on the CPU and another CPU wants 2911 * to access the timer (probably to cancel it). We can safely 2912 * ignore the boosting request, as the idle CPU runs this code 2913 * with interrupts disabled and will complete the lock 2914 * protected section without being interrupted. So there is no 2915 * real need to boost. 2916 */ 2917 if (unlikely(p == rq->idle)) { 2918 WARN_ON(p != rq->curr); 2919 WARN_ON(p->pi_blocked_on); 2920 goto out_unlock; 2921 } 2922 2923 trace_sched_pi_setprio(p, prio); 2924 p->pi_top_task = rt_mutex_get_top_task(p); 2925 oldprio = p->prio; 2926 prev_class = p->sched_class; 2927 on_rq = p->on_rq; 2928 running = task_current(rq, p); 2929 if (on_rq) 2930 dequeue_task(rq, p, 0); 2931 if (running) 2932 p->sched_class->put_prev_task(rq, p); 2933 2934 /* 2935 * Boosting condition are: 2936 * 1. -rt task is running and holds mutex A 2937 * --> -dl task blocks on mutex A 2938 * 2939 * 2. -dl task is running and holds mutex A 2940 * --> -dl task blocks on mutex A and could preempt the 2941 * running task 2942 */ 2943 if (dl_prio(prio)) { 2944 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2945 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2946 p->dl.dl_boosted = 1; 2947 p->dl.dl_throttled = 0; 2948 enqueue_flag = ENQUEUE_REPLENISH; 2949 } else 2950 p->dl.dl_boosted = 0; 2951 p->sched_class = &dl_sched_class; 2952 } else if (rt_prio(prio)) { 2953 if (dl_prio(oldprio)) 2954 p->dl.dl_boosted = 0; 2955 if (oldprio < prio) 2956 enqueue_flag = ENQUEUE_HEAD; 2957 p->sched_class = &rt_sched_class; 2958 } else { 2959 if (dl_prio(oldprio)) 2960 p->dl.dl_boosted = 0; 2961 p->sched_class = &fair_sched_class; 2962 } 2963 2964 p->prio = prio; 2965 2966 if (running) 2967 p->sched_class->set_curr_task(rq); 2968 if (on_rq) 2969 enqueue_task(rq, p, enqueue_flag); 2970 2971 check_class_changed(rq, p, prev_class, oldprio); 2972out_unlock: 2973 __task_rq_unlock(rq); 2974} 2975#endif 2976 2977void set_user_nice(struct task_struct *p, long nice) 2978{ 2979 int old_prio, delta, on_rq; 2980 unsigned long flags; 2981 struct rq *rq; 2982 2983 if (task_nice(p) == nice || nice < -20 || nice > 19) 2984 return; 2985 /* 2986 * We have to be careful, if called from sys_setpriority(), 2987 * the task might be in the middle of scheduling on another CPU. 2988 */ 2989 rq = task_rq_lock(p, &flags); 2990 /* 2991 * The RT priorities are set via sched_setscheduler(), but we still 2992 * allow the 'normal' nice value to be set - but as expected 2993 * it wont have any effect on scheduling until the task is 2994 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 2995 */ 2996 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 2997 p->static_prio = NICE_TO_PRIO(nice); 2998 goto out_unlock; 2999 } 3000 on_rq = p->on_rq; 3001 if (on_rq) 3002 dequeue_task(rq, p, 0); 3003 3004 p->static_prio = NICE_TO_PRIO(nice); 3005 set_load_weight(p); 3006 old_prio = p->prio; 3007 p->prio = effective_prio(p); 3008 delta = p->prio - old_prio; 3009 3010 if (on_rq) { 3011 enqueue_task(rq, p, 0); 3012 /* 3013 * If the task increased its priority or is running and 3014 * lowered its priority, then reschedule its CPU: 3015 */ 3016 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3017 resched_task(rq->curr); 3018 } 3019out_unlock: 3020 task_rq_unlock(rq, p, &flags); 3021} 3022EXPORT_SYMBOL(set_user_nice); 3023 3024/* 3025 * can_nice - check if a task can reduce its nice value 3026 * @p: task 3027 * @nice: nice value 3028 */ 3029int can_nice(const struct task_struct *p, const int nice) 3030{ 3031 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3032 int nice_rlim = 20 - nice; 3033 3034 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3035 capable(CAP_SYS_NICE)); 3036} 3037 3038#ifdef __ARCH_WANT_SYS_NICE 3039 3040/* 3041 * sys_nice - change the priority of the current process. 3042 * @increment: priority increment 3043 * 3044 * sys_setpriority is a more generic, but much slower function that 3045 * does similar things. 3046 */ 3047SYSCALL_DEFINE1(nice, int, increment) 3048{ 3049 long nice, retval; 3050 3051 /* 3052 * Setpriority might change our priority at the same moment. 3053 * We don't have to worry. Conceptually one call occurs first 3054 * and we have a single winner. 3055 */ 3056 if (increment < -40) 3057 increment = -40; 3058 if (increment > 40) 3059 increment = 40; 3060 3061 nice = task_nice(current) + increment; 3062 if (nice < -20) 3063 nice = -20; 3064 if (nice > 19) 3065 nice = 19; 3066 3067 if (increment < 0 && !can_nice(current, nice)) 3068 return -EPERM; 3069 3070 retval = security_task_setnice(current, nice); 3071 if (retval) 3072 return retval; 3073 3074 set_user_nice(current, nice); 3075 return 0; 3076} 3077 3078#endif 3079 3080/** 3081 * task_prio - return the priority value of a given task. 3082 * @p: the task in question. 3083 * 3084 * Return: The priority value as seen by users in /proc. 3085 * RT tasks are offset by -200. Normal tasks are centered 3086 * around 0, value goes from -16 to +15. 3087 */ 3088int task_prio(const struct task_struct *p) 3089{ 3090 return p->prio - MAX_RT_PRIO; 3091} 3092 3093/** 3094 * idle_cpu - is a given cpu idle currently? 3095 * @cpu: the processor in question. 3096 * 3097 * Return: 1 if the CPU is currently idle. 0 otherwise. 3098 */ 3099int idle_cpu(int cpu) 3100{ 3101 struct rq *rq = cpu_rq(cpu); 3102 3103 if (rq->curr != rq->idle) 3104 return 0; 3105 3106 if (rq->nr_running) 3107 return 0; 3108 3109#ifdef CONFIG_SMP 3110 if (!llist_empty(&rq->wake_list)) 3111 return 0; 3112#endif 3113 3114 return 1; 3115} 3116 3117/** 3118 * idle_task - return the idle task for a given cpu. 3119 * @cpu: the processor in question. 3120 * 3121 * Return: The idle task for the cpu @cpu. 3122 */ 3123struct task_struct *idle_task(int cpu) 3124{ 3125 return cpu_rq(cpu)->idle; 3126} 3127 3128/** 3129 * find_process_by_pid - find a process with a matching PID value. 3130 * @pid: the pid in question. 3131 * 3132 * The task of @pid, if found. %NULL otherwise. 3133 */ 3134static struct task_struct *find_process_by_pid(pid_t pid) 3135{ 3136 return pid ? find_task_by_vpid(pid) : current; 3137} 3138 3139/* 3140 * This function initializes the sched_dl_entity of a newly becoming 3141 * SCHED_DEADLINE task. 3142 * 3143 * Only the static values are considered here, the actual runtime and the 3144 * absolute deadline will be properly calculated when the task is enqueued 3145 * for the first time with its new policy. 3146 */ 3147static void 3148__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3149{ 3150 struct sched_dl_entity *dl_se = &p->dl; 3151 3152 init_dl_task_timer(dl_se); 3153 dl_se->dl_runtime = attr->sched_runtime; 3154 dl_se->dl_deadline = attr->sched_deadline; 3155 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3156 dl_se->flags = attr->sched_flags; 3157 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3158 dl_se->dl_throttled = 0; 3159 dl_se->dl_new = 1; 3160} 3161 3162/* Actually do priority change: must hold pi & rq lock. */ 3163static void __setscheduler(struct rq *rq, struct task_struct *p, 3164 const struct sched_attr *attr) 3165{ 3166 int policy = attr->sched_policy; 3167 3168 if (policy == -1) /* setparam */ 3169 policy = p->policy; 3170 3171 p->policy = policy; 3172 3173 if (dl_policy(policy)) 3174 __setparam_dl(p, attr); 3175 else if (fair_policy(policy)) 3176 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3177 3178 /* 3179 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3180 * !rt_policy. Always setting this ensures that things like 3181 * getparam()/getattr() don't report silly values for !rt tasks. 3182 */ 3183 p->rt_priority = attr->sched_priority; 3184 3185 p->normal_prio = normal_prio(p); 3186 p->prio = rt_mutex_getprio(p); 3187 3188 if (dl_prio(p->prio)) 3189 p->sched_class = &dl_sched_class; 3190 else if (rt_prio(p->prio)) 3191 p->sched_class = &rt_sched_class; 3192 else 3193 p->sched_class = &fair_sched_class; 3194 3195 set_load_weight(p); 3196} 3197 3198static void 3199__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3200{ 3201 struct sched_dl_entity *dl_se = &p->dl; 3202 3203 attr->sched_priority = p->rt_priority; 3204 attr->sched_runtime = dl_se->dl_runtime; 3205 attr->sched_deadline = dl_se->dl_deadline; 3206 attr->sched_period = dl_se->dl_period; 3207 attr->sched_flags = dl_se->flags; 3208} 3209 3210/* 3211 * This function validates the new parameters of a -deadline task. 3212 * We ask for the deadline not being zero, and greater or equal 3213 * than the runtime, as well as the period of being zero or 3214 * greater than deadline. Furthermore, we have to be sure that 3215 * user parameters are above the internal resolution (1us); we 3216 * check sched_runtime only since it is always the smaller one. 3217 */ 3218static bool 3219__checkparam_dl(const struct sched_attr *attr) 3220{ 3221 return attr && attr->sched_deadline != 0 && 3222 (attr->sched_period == 0 || 3223 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3224 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3225 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3226} 3227 3228/* 3229 * check the target process has a UID that matches the current process's 3230 */ 3231static bool check_same_owner(struct task_struct *p) 3232{ 3233 const struct cred *cred = current_cred(), *pcred; 3234 bool match; 3235 3236 rcu_read_lock(); 3237 pcred = __task_cred(p); 3238 match = (uid_eq(cred->euid, pcred->euid) || 3239 uid_eq(cred->euid, pcred->uid)); 3240 rcu_read_unlock(); 3241 return match; 3242} 3243 3244static int __sched_setscheduler(struct task_struct *p, 3245 const struct sched_attr *attr, 3246 bool user) 3247{ 3248 int retval, oldprio, oldpolicy = -1, on_rq, running; 3249 int policy = attr->sched_policy; 3250 unsigned long flags; 3251 const struct sched_class *prev_class; 3252 struct rq *rq; 3253 int reset_on_fork; 3254 3255 /* may grab non-irq protected spin_locks */ 3256 BUG_ON(in_interrupt()); 3257recheck: 3258 /* double check policy once rq lock held */ 3259 if (policy < 0) { 3260 reset_on_fork = p->sched_reset_on_fork; 3261 policy = oldpolicy = p->policy; 3262 } else { 3263 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3264 3265 if (policy != SCHED_DEADLINE && 3266 policy != SCHED_FIFO && policy != SCHED_RR && 3267 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3268 policy != SCHED_IDLE) 3269 return -EINVAL; 3270 } 3271 3272 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3273 return -EINVAL; 3274 3275 /* 3276 * Valid priorities for SCHED_FIFO and SCHED_RR are 3277 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3278 * SCHED_BATCH and SCHED_IDLE is 0. 3279 */ 3280 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3281 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3282 return -EINVAL; 3283 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3284 (rt_policy(policy) != (attr->sched_priority != 0))) 3285 return -EINVAL; 3286 3287 /* 3288 * Allow unprivileged RT tasks to decrease priority: 3289 */ 3290 if (user && !capable(CAP_SYS_NICE)) { 3291 if (fair_policy(policy)) { 3292 if (attr->sched_nice < task_nice(p) && 3293 !can_nice(p, attr->sched_nice)) 3294 return -EPERM; 3295 } 3296 3297 if (rt_policy(policy)) { 3298 unsigned long rlim_rtprio = 3299 task_rlimit(p, RLIMIT_RTPRIO); 3300 3301 /* can't set/change the rt policy */ 3302 if (policy != p->policy && !rlim_rtprio) 3303 return -EPERM; 3304 3305 /* can't increase priority */ 3306 if (attr->sched_priority > p->rt_priority && 3307 attr->sched_priority > rlim_rtprio) 3308 return -EPERM; 3309 } 3310 3311 /* 3312 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3313 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3314 */ 3315 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3316 if (!can_nice(p, task_nice(p))) 3317 return -EPERM; 3318 } 3319 3320 /* can't change other user's priorities */ 3321 if (!check_same_owner(p)) 3322 return -EPERM; 3323 3324 /* Normal users shall not reset the sched_reset_on_fork flag */ 3325 if (p->sched_reset_on_fork && !reset_on_fork) 3326 return -EPERM; 3327 } 3328 3329 if (user) { 3330 retval = security_task_setscheduler(p); 3331 if (retval) 3332 return retval; 3333 } 3334 3335 /* 3336 * make sure no PI-waiters arrive (or leave) while we are 3337 * changing the priority of the task: 3338 * 3339 * To be able to change p->policy safely, the appropriate 3340 * runqueue lock must be held. 3341 */ 3342 rq = task_rq_lock(p, &flags); 3343 3344 /* 3345 * Changing the policy of the stop threads its a very bad idea 3346 */ 3347 if (p == rq->stop) { 3348 task_rq_unlock(rq, p, &flags); 3349 return -EINVAL; 3350 } 3351 3352 /* 3353 * If not changing anything there's no need to proceed further: 3354 */ 3355 if (unlikely(policy == p->policy)) { 3356 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3357 goto change; 3358 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3359 goto change; 3360 if (dl_policy(policy)) 3361 goto change; 3362 3363 task_rq_unlock(rq, p, &flags); 3364 return 0; 3365 } 3366change: 3367 3368 if (user) { 3369#ifdef CONFIG_RT_GROUP_SCHED 3370 /* 3371 * Do not allow realtime tasks into groups that have no runtime 3372 * assigned. 3373 */ 3374 if (rt_bandwidth_enabled() && rt_policy(policy) && 3375 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3376 !task_group_is_autogroup(task_group(p))) { 3377 task_rq_unlock(rq, p, &flags); 3378 return -EPERM; 3379 } 3380#endif 3381#ifdef CONFIG_SMP 3382 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3383 cpumask_t *span = rq->rd->span; 3384 3385 /* 3386 * Don't allow tasks with an affinity mask smaller than 3387 * the entire root_domain to become SCHED_DEADLINE. We 3388 * will also fail if there's no bandwidth available. 3389 */ 3390 if (!cpumask_subset(span, &p->cpus_allowed) || 3391 rq->rd->dl_bw.bw == 0) { 3392 task_rq_unlock(rq, p, &flags); 3393 return -EPERM; 3394 } 3395 } 3396#endif 3397 } 3398 3399 /* recheck policy now with rq lock held */ 3400 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3401 policy = oldpolicy = -1; 3402 task_rq_unlock(rq, p, &flags); 3403 goto recheck; 3404 } 3405 3406 /* 3407 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3408 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3409 * is available. 3410 */ 3411 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3412 task_rq_unlock(rq, p, &flags); 3413 return -EBUSY; 3414 } 3415 3416 on_rq = p->on_rq; 3417 running = task_current(rq, p); 3418 if (on_rq) 3419 dequeue_task(rq, p, 0); 3420 if (running) 3421 p->sched_class->put_prev_task(rq, p); 3422 3423 p->sched_reset_on_fork = reset_on_fork; 3424 3425 oldprio = p->prio; 3426 prev_class = p->sched_class; 3427 __setscheduler(rq, p, attr); 3428 3429 if (running) 3430 p->sched_class->set_curr_task(rq); 3431 if (on_rq) 3432 enqueue_task(rq, p, 0); 3433 3434 check_class_changed(rq, p, prev_class, oldprio); 3435 task_rq_unlock(rq, p, &flags); 3436 3437 rt_mutex_adjust_pi(p); 3438 3439 return 0; 3440} 3441 3442static int _sched_setscheduler(struct task_struct *p, int policy, 3443 const struct sched_param *param, bool check) 3444{ 3445 struct sched_attr attr = { 3446 .sched_policy = policy, 3447 .sched_priority = param->sched_priority, 3448 .sched_nice = PRIO_TO_NICE(p->static_prio), 3449 }; 3450 3451 /* 3452 * Fixup the legacy SCHED_RESET_ON_FORK hack 3453 */ 3454 if (policy & SCHED_RESET_ON_FORK) { 3455 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3456 policy &= ~SCHED_RESET_ON_FORK; 3457 attr.sched_policy = policy; 3458 } 3459 3460 return __sched_setscheduler(p, &attr, check); 3461} 3462/** 3463 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3464 * @p: the task in question. 3465 * @policy: new policy. 3466 * @param: structure containing the new RT priority. 3467 * 3468 * Return: 0 on success. An error code otherwise. 3469 * 3470 * NOTE that the task may be already dead. 3471 */ 3472int sched_setscheduler(struct task_struct *p, int policy, 3473 const struct sched_param *param) 3474{ 3475 return _sched_setscheduler(p, policy, param, true); 3476} 3477EXPORT_SYMBOL_GPL(sched_setscheduler); 3478 3479int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3480{ 3481 return __sched_setscheduler(p, attr, true); 3482} 3483EXPORT_SYMBOL_GPL(sched_setattr); 3484 3485/** 3486 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3487 * @p: the task in question. 3488 * @policy: new policy. 3489 * @param: structure containing the new RT priority. 3490 * 3491 * Just like sched_setscheduler, only don't bother checking if the 3492 * current context has permission. For example, this is needed in 3493 * stop_machine(): we create temporary high priority worker threads, 3494 * but our caller might not have that capability. 3495 * 3496 * Return: 0 on success. An error code otherwise. 3497 */ 3498int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3499 const struct sched_param *param) 3500{ 3501 return _sched_setscheduler(p, policy, param, false); 3502} 3503 3504static int 3505do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3506{ 3507 struct sched_param lparam; 3508 struct task_struct *p; 3509 int retval; 3510 3511 if (!param || pid < 0) 3512 return -EINVAL; 3513 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3514 return -EFAULT; 3515 3516 rcu_read_lock(); 3517 retval = -ESRCH; 3518 p = find_process_by_pid(pid); 3519 if (p != NULL) 3520 retval = sched_setscheduler(p, policy, &lparam); 3521 rcu_read_unlock(); 3522 3523 return retval; 3524} 3525 3526/* 3527 * Mimics kernel/events/core.c perf_copy_attr(). 3528 */ 3529static int sched_copy_attr(struct sched_attr __user *uattr, 3530 struct sched_attr *attr) 3531{ 3532 u32 size; 3533 int ret; 3534 3535 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3536 return -EFAULT; 3537 3538 /* 3539 * zero the full structure, so that a short copy will be nice. 3540 */ 3541 memset(attr, 0, sizeof(*attr)); 3542 3543 ret = get_user(size, &uattr->size); 3544 if (ret) 3545 return ret; 3546 3547 if (size > PAGE_SIZE) /* silly large */ 3548 goto err_size; 3549 3550 if (!size) /* abi compat */ 3551 size = SCHED_ATTR_SIZE_VER0; 3552 3553 if (size < SCHED_ATTR_SIZE_VER0) 3554 goto err_size; 3555 3556 /* 3557 * If we're handed a bigger struct than we know of, 3558 * ensure all the unknown bits are 0 - i.e. new 3559 * user-space does not rely on any kernel feature 3560 * extensions we dont know about yet. 3561 */ 3562 if (size > sizeof(*attr)) { 3563 unsigned char __user *addr; 3564 unsigned char __user *end; 3565 unsigned char val; 3566 3567 addr = (void __user *)uattr + sizeof(*attr); 3568 end = (void __user *)uattr + size; 3569 3570 for (; addr < end; addr++) { 3571 ret = get_user(val, addr); 3572 if (ret) 3573 return ret; 3574 if (val) 3575 goto err_size; 3576 } 3577 size = sizeof(*attr); 3578 } 3579 3580 ret = copy_from_user(attr, uattr, size); 3581 if (ret) 3582 return -EFAULT; 3583 3584 /* 3585 * XXX: do we want to be lenient like existing syscalls; or do we want 3586 * to be strict and return an error on out-of-bounds values? 3587 */ 3588 attr->sched_nice = clamp(attr->sched_nice, -20, 19); 3589 3590out: 3591 return ret; 3592 3593err_size: 3594 put_user(sizeof(*attr), &uattr->size); 3595 ret = -E2BIG; 3596 goto out; 3597} 3598 3599/** 3600 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3601 * @pid: the pid in question. 3602 * @policy: new policy. 3603 * @param: structure containing the new RT priority. 3604 * 3605 * Return: 0 on success. An error code otherwise. 3606 */ 3607SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3608 struct sched_param __user *, param) 3609{ 3610 /* negative values for policy are not valid */ 3611 if (policy < 0) 3612 return -EINVAL; 3613 3614 return do_sched_setscheduler(pid, policy, param); 3615} 3616 3617/** 3618 * sys_sched_setparam - set/change the RT priority of a thread 3619 * @pid: the pid in question. 3620 * @param: structure containing the new RT priority. 3621 * 3622 * Return: 0 on success. An error code otherwise. 3623 */ 3624SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3625{ 3626 return do_sched_setscheduler(pid, -1, param); 3627} 3628 3629/** 3630 * sys_sched_setattr - same as above, but with extended sched_attr 3631 * @pid: the pid in question. 3632 * @uattr: structure containing the extended parameters. 3633 */ 3634SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr) 3635{ 3636 struct sched_attr attr; 3637 struct task_struct *p; 3638 int retval; 3639 3640 if (!uattr || pid < 0) 3641 return -EINVAL; 3642 3643 if (sched_copy_attr(uattr, &attr)) 3644 return -EFAULT; 3645 3646 rcu_read_lock(); 3647 retval = -ESRCH; 3648 p = find_process_by_pid(pid); 3649 if (p != NULL) 3650 retval = sched_setattr(p, &attr); 3651 rcu_read_unlock(); 3652 3653 return retval; 3654} 3655 3656/** 3657 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3658 * @pid: the pid in question. 3659 * 3660 * Return: On success, the policy of the thread. Otherwise, a negative error 3661 * code. 3662 */ 3663SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3664{ 3665 struct task_struct *p; 3666 int retval; 3667 3668 if (pid < 0) 3669 return -EINVAL; 3670 3671 retval = -ESRCH; 3672 rcu_read_lock(); 3673 p = find_process_by_pid(pid); 3674 if (p) { 3675 retval = security_task_getscheduler(p); 3676 if (!retval) 3677 retval = p->policy 3678 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3679 } 3680 rcu_read_unlock(); 3681 return retval; 3682} 3683 3684/** 3685 * sys_sched_getparam - get the RT priority of a thread 3686 * @pid: the pid in question. 3687 * @param: structure containing the RT priority. 3688 * 3689 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3690 * code. 3691 */ 3692SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3693{ 3694 struct sched_param lp; 3695 struct task_struct *p; 3696 int retval; 3697 3698 if (!param || pid < 0) 3699 return -EINVAL; 3700 3701 rcu_read_lock(); 3702 p = find_process_by_pid(pid); 3703 retval = -ESRCH; 3704 if (!p) 3705 goto out_unlock; 3706 3707 retval = security_task_getscheduler(p); 3708 if (retval) 3709 goto out_unlock; 3710 3711 if (task_has_dl_policy(p)) { 3712 retval = -EINVAL; 3713 goto out_unlock; 3714 } 3715 lp.sched_priority = p->rt_priority; 3716 rcu_read_unlock(); 3717 3718 /* 3719 * This one might sleep, we cannot do it with a spinlock held ... 3720 */ 3721 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3722 3723 return retval; 3724 3725out_unlock: 3726 rcu_read_unlock(); 3727 return retval; 3728} 3729 3730static int sched_read_attr(struct sched_attr __user *uattr, 3731 struct sched_attr *attr, 3732 unsigned int usize) 3733{ 3734 int ret; 3735 3736 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3737 return -EFAULT; 3738 3739 /* 3740 * If we're handed a smaller struct than we know of, 3741 * ensure all the unknown bits are 0 - i.e. old 3742 * user-space does not get uncomplete information. 3743 */ 3744 if (usize < sizeof(*attr)) { 3745 unsigned char *addr; 3746 unsigned char *end; 3747 3748 addr = (void *)attr + usize; 3749 end = (void *)attr + sizeof(*attr); 3750 3751 for (; addr < end; addr++) { 3752 if (*addr) 3753 goto err_size; 3754 } 3755 3756 attr->size = usize; 3757 } 3758 3759 ret = copy_to_user(uattr, attr, usize); 3760 if (ret) 3761 return -EFAULT; 3762 3763out: 3764 return ret; 3765 3766err_size: 3767 ret = -E2BIG; 3768 goto out; 3769} 3770 3771/** 3772 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3773 * @pid: the pid in question. 3774 * @uattr: structure containing the extended parameters. 3775 * @size: sizeof(attr) for fwd/bwd comp. 3776 */ 3777SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3778 unsigned int, size) 3779{ 3780 struct sched_attr attr = { 3781 .size = sizeof(struct sched_attr), 3782 }; 3783 struct task_struct *p; 3784 int retval; 3785 3786 if (!uattr || pid < 0 || size > PAGE_SIZE || 3787 size < SCHED_ATTR_SIZE_VER0) 3788 return -EINVAL; 3789 3790 rcu_read_lock(); 3791 p = find_process_by_pid(pid); 3792 retval = -ESRCH; 3793 if (!p) 3794 goto out_unlock; 3795 3796 retval = security_task_getscheduler(p); 3797 if (retval) 3798 goto out_unlock; 3799 3800 attr.sched_policy = p->policy; 3801 if (p->sched_reset_on_fork) 3802 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3803 if (task_has_dl_policy(p)) 3804 __getparam_dl(p, &attr); 3805 else if (task_has_rt_policy(p)) 3806 attr.sched_priority = p->rt_priority; 3807 else 3808 attr.sched_nice = task_nice(p); 3809 3810 rcu_read_unlock(); 3811 3812 retval = sched_read_attr(uattr, &attr, size); 3813 return retval; 3814 3815out_unlock: 3816 rcu_read_unlock(); 3817 return retval; 3818} 3819 3820long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3821{ 3822 cpumask_var_t cpus_allowed, new_mask; 3823 struct task_struct *p; 3824 int retval; 3825 3826 rcu_read_lock(); 3827 3828 p = find_process_by_pid(pid); 3829 if (!p) { 3830 rcu_read_unlock(); 3831 return -ESRCH; 3832 } 3833 3834 /* Prevent p going away */ 3835 get_task_struct(p); 3836 rcu_read_unlock(); 3837 3838 if (p->flags & PF_NO_SETAFFINITY) { 3839 retval = -EINVAL; 3840 goto out_put_task; 3841 } 3842 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3843 retval = -ENOMEM; 3844 goto out_put_task; 3845 } 3846 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3847 retval = -ENOMEM; 3848 goto out_free_cpus_allowed; 3849 } 3850 retval = -EPERM; 3851 if (!check_same_owner(p)) { 3852 rcu_read_lock(); 3853 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3854 rcu_read_unlock(); 3855 goto out_unlock; 3856 } 3857 rcu_read_unlock(); 3858 } 3859 3860 retval = security_task_setscheduler(p); 3861 if (retval) 3862 goto out_unlock; 3863 3864 3865 cpuset_cpus_allowed(p, cpus_allowed); 3866 cpumask_and(new_mask, in_mask, cpus_allowed); 3867 3868 /* 3869 * Since bandwidth control happens on root_domain basis, 3870 * if admission test is enabled, we only admit -deadline 3871 * tasks allowed to run on all the CPUs in the task's 3872 * root_domain. 3873 */ 3874#ifdef CONFIG_SMP 3875 if (task_has_dl_policy(p)) { 3876 const struct cpumask *span = task_rq(p)->rd->span; 3877 3878 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3879 retval = -EBUSY; 3880 goto out_unlock; 3881 } 3882 } 3883#endif 3884again: 3885 retval = set_cpus_allowed_ptr(p, new_mask); 3886 3887 if (!retval) { 3888 cpuset_cpus_allowed(p, cpus_allowed); 3889 if (!cpumask_subset(new_mask, cpus_allowed)) { 3890 /* 3891 * We must have raced with a concurrent cpuset 3892 * update. Just reset the cpus_allowed to the 3893 * cpuset's cpus_allowed 3894 */ 3895 cpumask_copy(new_mask, cpus_allowed); 3896 goto again; 3897 } 3898 } 3899out_unlock: 3900 free_cpumask_var(new_mask); 3901out_free_cpus_allowed: 3902 free_cpumask_var(cpus_allowed); 3903out_put_task: 3904 put_task_struct(p); 3905 return retval; 3906} 3907 3908static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3909 struct cpumask *new_mask) 3910{ 3911 if (len < cpumask_size()) 3912 cpumask_clear(new_mask); 3913 else if (len > cpumask_size()) 3914 len = cpumask_size(); 3915 3916 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3917} 3918 3919/** 3920 * sys_sched_setaffinity - set the cpu affinity of a process 3921 * @pid: pid of the process 3922 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3923 * @user_mask_ptr: user-space pointer to the new cpu mask 3924 * 3925 * Return: 0 on success. An error code otherwise. 3926 */ 3927SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3928 unsigned long __user *, user_mask_ptr) 3929{ 3930 cpumask_var_t new_mask; 3931 int retval; 3932 3933 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3934 return -ENOMEM; 3935 3936 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3937 if (retval == 0) 3938 retval = sched_setaffinity(pid, new_mask); 3939 free_cpumask_var(new_mask); 3940 return retval; 3941} 3942 3943long sched_getaffinity(pid_t pid, struct cpumask *mask) 3944{ 3945 struct task_struct *p; 3946 unsigned long flags; 3947 int retval; 3948 3949 rcu_read_lock(); 3950 3951 retval = -ESRCH; 3952 p = find_process_by_pid(pid); 3953 if (!p) 3954 goto out_unlock; 3955 3956 retval = security_task_getscheduler(p); 3957 if (retval) 3958 goto out_unlock; 3959 3960 raw_spin_lock_irqsave(&p->pi_lock, flags); 3961 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3962 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3963 3964out_unlock: 3965 rcu_read_unlock(); 3966 3967 return retval; 3968} 3969 3970/** 3971 * sys_sched_getaffinity - get the cpu affinity of a process 3972 * @pid: pid of the process 3973 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3974 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3975 * 3976 * Return: 0 on success. An error code otherwise. 3977 */ 3978SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3979 unsigned long __user *, user_mask_ptr) 3980{ 3981 int ret; 3982 cpumask_var_t mask; 3983 3984 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3985 return -EINVAL; 3986 if (len & (sizeof(unsigned long)-1)) 3987 return -EINVAL; 3988 3989 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3990 return -ENOMEM; 3991 3992 ret = sched_getaffinity(pid, mask); 3993 if (ret == 0) { 3994 size_t retlen = min_t(size_t, len, cpumask_size()); 3995 3996 if (copy_to_user(user_mask_ptr, mask, retlen)) 3997 ret = -EFAULT; 3998 else 3999 ret = retlen; 4000 } 4001 free_cpumask_var(mask); 4002 4003 return ret; 4004} 4005 4006/** 4007 * sys_sched_yield - yield the current processor to other threads. 4008 * 4009 * This function yields the current CPU to other tasks. If there are no 4010 * other threads running on this CPU then this function will return. 4011 * 4012 * Return: 0. 4013 */ 4014SYSCALL_DEFINE0(sched_yield) 4015{ 4016 struct rq *rq = this_rq_lock(); 4017 4018 schedstat_inc(rq, yld_count); 4019 current->sched_class->yield_task(rq); 4020 4021 /* 4022 * Since we are going to call schedule() anyway, there's 4023 * no need to preempt or enable interrupts: 4024 */ 4025 __release(rq->lock); 4026 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4027 do_raw_spin_unlock(&rq->lock); 4028 sched_preempt_enable_no_resched(); 4029 4030 schedule(); 4031 4032 return 0; 4033} 4034 4035static void __cond_resched(void) 4036{ 4037 __preempt_count_add(PREEMPT_ACTIVE); 4038 __schedule(); 4039 __preempt_count_sub(PREEMPT_ACTIVE); 4040} 4041 4042int __sched _cond_resched(void) 4043{ 4044 if (should_resched()) { 4045 __cond_resched(); 4046 return 1; 4047 } 4048 return 0; 4049} 4050EXPORT_SYMBOL(_cond_resched); 4051 4052/* 4053 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4054 * call schedule, and on return reacquire the lock. 4055 * 4056 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4057 * operations here to prevent schedule() from being called twice (once via 4058 * spin_unlock(), once by hand). 4059 */ 4060int __cond_resched_lock(spinlock_t *lock) 4061{ 4062 int resched = should_resched(); 4063 int ret = 0; 4064 4065 lockdep_assert_held(lock); 4066 4067 if (spin_needbreak(lock) || resched) { 4068 spin_unlock(lock); 4069 if (resched) 4070 __cond_resched(); 4071 else 4072 cpu_relax(); 4073 ret = 1; 4074 spin_lock(lock); 4075 } 4076 return ret; 4077} 4078EXPORT_SYMBOL(__cond_resched_lock); 4079 4080int __sched __cond_resched_softirq(void) 4081{ 4082 BUG_ON(!in_softirq()); 4083 4084 if (should_resched()) { 4085 local_bh_enable(); 4086 __cond_resched(); 4087 local_bh_disable(); 4088 return 1; 4089 } 4090 return 0; 4091} 4092EXPORT_SYMBOL(__cond_resched_softirq); 4093 4094/** 4095 * yield - yield the current processor to other threads. 4096 * 4097 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4098 * 4099 * The scheduler is at all times free to pick the calling task as the most 4100 * eligible task to run, if removing the yield() call from your code breaks 4101 * it, its already broken. 4102 * 4103 * Typical broken usage is: 4104 * 4105 * while (!event) 4106 * yield(); 4107 * 4108 * where one assumes that yield() will let 'the other' process run that will 4109 * make event true. If the current task is a SCHED_FIFO task that will never 4110 * happen. Never use yield() as a progress guarantee!! 4111 * 4112 * If you want to use yield() to wait for something, use wait_event(). 4113 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4114 * If you still want to use yield(), do not! 4115 */ 4116void __sched yield(void) 4117{ 4118 set_current_state(TASK_RUNNING); 4119 sys_sched_yield(); 4120} 4121EXPORT_SYMBOL(yield); 4122 4123/** 4124 * yield_to - yield the current processor to another thread in 4125 * your thread group, or accelerate that thread toward the 4126 * processor it's on. 4127 * @p: target task 4128 * @preempt: whether task preemption is allowed or not 4129 * 4130 * It's the caller's job to ensure that the target task struct 4131 * can't go away on us before we can do any checks. 4132 * 4133 * Return: 4134 * true (>0) if we indeed boosted the target task. 4135 * false (0) if we failed to boost the target. 4136 * -ESRCH if there's no task to yield to. 4137 */ 4138bool __sched yield_to(struct task_struct *p, bool preempt) 4139{ 4140 struct task_struct *curr = current; 4141 struct rq *rq, *p_rq; 4142 unsigned long flags; 4143 int yielded = 0; 4144 4145 local_irq_save(flags); 4146 rq = this_rq(); 4147 4148again: 4149 p_rq = task_rq(p); 4150 /* 4151 * If we're the only runnable task on the rq and target rq also 4152 * has only one task, there's absolutely no point in yielding. 4153 */ 4154 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4155 yielded = -ESRCH; 4156 goto out_irq; 4157 } 4158 4159 double_rq_lock(rq, p_rq); 4160 if (task_rq(p) != p_rq) { 4161 double_rq_unlock(rq, p_rq); 4162 goto again; 4163 } 4164 4165 if (!curr->sched_class->yield_to_task) 4166 goto out_unlock; 4167 4168 if (curr->sched_class != p->sched_class) 4169 goto out_unlock; 4170 4171 if (task_running(p_rq, p) || p->state) 4172 goto out_unlock; 4173 4174 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4175 if (yielded) { 4176 schedstat_inc(rq, yld_count); 4177 /* 4178 * Make p's CPU reschedule; pick_next_entity takes care of 4179 * fairness. 4180 */ 4181 if (preempt && rq != p_rq) 4182 resched_task(p_rq->curr); 4183 } 4184 4185out_unlock: 4186 double_rq_unlock(rq, p_rq); 4187out_irq: 4188 local_irq_restore(flags); 4189 4190 if (yielded > 0) 4191 schedule(); 4192 4193 return yielded; 4194} 4195EXPORT_SYMBOL_GPL(yield_to); 4196 4197/* 4198 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4199 * that process accounting knows that this is a task in IO wait state. 4200 */ 4201void __sched io_schedule(void) 4202{ 4203 struct rq *rq = raw_rq(); 4204 4205 delayacct_blkio_start(); 4206 atomic_inc(&rq->nr_iowait); 4207 blk_flush_plug(current); 4208 current->in_iowait = 1; 4209 schedule(); 4210 current->in_iowait = 0; 4211 atomic_dec(&rq->nr_iowait); 4212 delayacct_blkio_end(); 4213} 4214EXPORT_SYMBOL(io_schedule); 4215 4216long __sched io_schedule_timeout(long timeout) 4217{ 4218 struct rq *rq = raw_rq(); 4219 long ret; 4220 4221 delayacct_blkio_start(); 4222 atomic_inc(&rq->nr_iowait); 4223 blk_flush_plug(current); 4224 current->in_iowait = 1; 4225 ret = schedule_timeout(timeout); 4226 current->in_iowait = 0; 4227 atomic_dec(&rq->nr_iowait); 4228 delayacct_blkio_end(); 4229 return ret; 4230} 4231 4232/** 4233 * sys_sched_get_priority_max - return maximum RT priority. 4234 * @policy: scheduling class. 4235 * 4236 * Return: On success, this syscall returns the maximum 4237 * rt_priority that can be used by a given scheduling class. 4238 * On failure, a negative error code is returned. 4239 */ 4240SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4241{ 4242 int ret = -EINVAL; 4243 4244 switch (policy) { 4245 case SCHED_FIFO: 4246 case SCHED_RR: 4247 ret = MAX_USER_RT_PRIO-1; 4248 break; 4249 case SCHED_DEADLINE: 4250 case SCHED_NORMAL: 4251 case SCHED_BATCH: 4252 case SCHED_IDLE: 4253 ret = 0; 4254 break; 4255 } 4256 return ret; 4257} 4258 4259/** 4260 * sys_sched_get_priority_min - return minimum RT priority. 4261 * @policy: scheduling class. 4262 * 4263 * Return: On success, this syscall returns the minimum 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_min, int, policy) 4268{ 4269 int ret = -EINVAL; 4270 4271 switch (policy) { 4272 case SCHED_FIFO: 4273 case SCHED_RR: 4274 ret = 1; 4275 break; 4276 case SCHED_DEADLINE: 4277 case SCHED_NORMAL: 4278 case SCHED_BATCH: 4279 case SCHED_IDLE: 4280 ret = 0; 4281 } 4282 return ret; 4283} 4284 4285/** 4286 * sys_sched_rr_get_interval - return the default timeslice of a process. 4287 * @pid: pid of the process. 4288 * @interval: userspace pointer to the timeslice value. 4289 * 4290 * this syscall writes the default timeslice value of a given process 4291 * into the user-space timespec buffer. A value of '0' means infinity. 4292 * 4293 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4294 * an error code. 4295 */ 4296SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4297 struct timespec __user *, interval) 4298{ 4299 struct task_struct *p; 4300 unsigned int time_slice; 4301 unsigned long flags; 4302 struct rq *rq; 4303 int retval; 4304 struct timespec t; 4305 4306 if (pid < 0) 4307 return -EINVAL; 4308 4309 retval = -ESRCH; 4310 rcu_read_lock(); 4311 p = find_process_by_pid(pid); 4312 if (!p) 4313 goto out_unlock; 4314 4315 retval = security_task_getscheduler(p); 4316 if (retval) 4317 goto out_unlock; 4318 4319 rq = task_rq_lock(p, &flags); 4320 time_slice = 0; 4321 if (p->sched_class->get_rr_interval) 4322 time_slice = p->sched_class->get_rr_interval(rq, p); 4323 task_rq_unlock(rq, p, &flags); 4324 4325 rcu_read_unlock(); 4326 jiffies_to_timespec(time_slice, &t); 4327 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4328 return retval; 4329 4330out_unlock: 4331 rcu_read_unlock(); 4332 return retval; 4333} 4334 4335static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4336 4337void sched_show_task(struct task_struct *p) 4338{ 4339 unsigned long free = 0; 4340 int ppid; 4341 unsigned state; 4342 4343 state = p->state ? __ffs(p->state) + 1 : 0; 4344 printk(KERN_INFO "%-15.15s %c", p->comm, 4345 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4346#if BITS_PER_LONG == 32 4347 if (state == TASK_RUNNING) 4348 printk(KERN_CONT " running "); 4349 else 4350 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4351#else 4352 if (state == TASK_RUNNING) 4353 printk(KERN_CONT " running task "); 4354 else 4355 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4356#endif 4357#ifdef CONFIG_DEBUG_STACK_USAGE 4358 free = stack_not_used(p); 4359#endif 4360 rcu_read_lock(); 4361 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4362 rcu_read_unlock(); 4363 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4364 task_pid_nr(p), ppid, 4365 (unsigned long)task_thread_info(p)->flags); 4366 4367 print_worker_info(KERN_INFO, p); 4368 show_stack(p, NULL); 4369} 4370 4371void show_state_filter(unsigned long state_filter) 4372{ 4373 struct task_struct *g, *p; 4374 4375#if BITS_PER_LONG == 32 4376 printk(KERN_INFO 4377 " task PC stack pid father\n"); 4378#else 4379 printk(KERN_INFO 4380 " task PC stack pid father\n"); 4381#endif 4382 rcu_read_lock(); 4383 do_each_thread(g, p) { 4384 /* 4385 * reset the NMI-timeout, listing all files on a slow 4386 * console might take a lot of time: 4387 */ 4388 touch_nmi_watchdog(); 4389 if (!state_filter || (p->state & state_filter)) 4390 sched_show_task(p); 4391 } while_each_thread(g, p); 4392 4393 touch_all_softlockup_watchdogs(); 4394 4395#ifdef CONFIG_SCHED_DEBUG 4396 sysrq_sched_debug_show(); 4397#endif 4398 rcu_read_unlock(); 4399 /* 4400 * Only show locks if all tasks are dumped: 4401 */ 4402 if (!state_filter) 4403 debug_show_all_locks(); 4404} 4405 4406void init_idle_bootup_task(struct task_struct *idle) 4407{ 4408 idle->sched_class = &idle_sched_class; 4409} 4410 4411/** 4412 * init_idle - set up an idle thread for a given CPU 4413 * @idle: task in question 4414 * @cpu: cpu the idle task belongs to 4415 * 4416 * NOTE: this function does not set the idle thread's NEED_RESCHED 4417 * flag, to make booting more robust. 4418 */ 4419void init_idle(struct task_struct *idle, int cpu) 4420{ 4421 struct rq *rq = cpu_rq(cpu); 4422 unsigned long flags; 4423 4424 raw_spin_lock_irqsave(&rq->lock, flags); 4425 4426 __sched_fork(0, idle); 4427 idle->state = TASK_RUNNING; 4428 idle->se.exec_start = sched_clock(); 4429 4430 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4431 /* 4432 * We're having a chicken and egg problem, even though we are 4433 * holding rq->lock, the cpu isn't yet set to this cpu so the 4434 * lockdep check in task_group() will fail. 4435 * 4436 * Similar case to sched_fork(). / Alternatively we could 4437 * use task_rq_lock() here and obtain the other rq->lock. 4438 * 4439 * Silence PROVE_RCU 4440 */ 4441 rcu_read_lock(); 4442 __set_task_cpu(idle, cpu); 4443 rcu_read_unlock(); 4444 4445 rq->curr = rq->idle = idle; 4446 idle->on_rq = 1; 4447#if defined(CONFIG_SMP) 4448 idle->on_cpu = 1; 4449#endif 4450 raw_spin_unlock_irqrestore(&rq->lock, flags); 4451 4452 /* Set the preempt count _outside_ the spinlocks! */ 4453 init_idle_preempt_count(idle, cpu); 4454 4455 /* 4456 * The idle tasks have their own, simple scheduling class: 4457 */ 4458 idle->sched_class = &idle_sched_class; 4459 ftrace_graph_init_idle_task(idle, cpu); 4460 vtime_init_idle(idle, cpu); 4461#if defined(CONFIG_SMP) 4462 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4463#endif 4464} 4465 4466#ifdef CONFIG_SMP 4467void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4468{ 4469 if (p->sched_class && p->sched_class->set_cpus_allowed) 4470 p->sched_class->set_cpus_allowed(p, new_mask); 4471 4472 cpumask_copy(&p->cpus_allowed, new_mask); 4473 p->nr_cpus_allowed = cpumask_weight(new_mask); 4474} 4475 4476/* 4477 * This is how migration works: 4478 * 4479 * 1) we invoke migration_cpu_stop() on the target CPU using 4480 * stop_one_cpu(). 4481 * 2) stopper starts to run (implicitly forcing the migrated thread 4482 * off the CPU) 4483 * 3) it checks whether the migrated task is still in the wrong runqueue. 4484 * 4) if it's in the wrong runqueue then the migration thread removes 4485 * it and puts it into the right queue. 4486 * 5) stopper completes and stop_one_cpu() returns and the migration 4487 * is done. 4488 */ 4489 4490/* 4491 * Change a given task's CPU affinity. Migrate the thread to a 4492 * proper CPU and schedule it away if the CPU it's executing on 4493 * is removed from the allowed bitmask. 4494 * 4495 * NOTE: the caller must have a valid reference to the task, the 4496 * task must not exit() & deallocate itself prematurely. The 4497 * call is not atomic; no spinlocks may be held. 4498 */ 4499int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4500{ 4501 unsigned long flags; 4502 struct rq *rq; 4503 unsigned int dest_cpu; 4504 int ret = 0; 4505 4506 rq = task_rq_lock(p, &flags); 4507 4508 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4509 goto out; 4510 4511 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4512 ret = -EINVAL; 4513 goto out; 4514 } 4515 4516 do_set_cpus_allowed(p, new_mask); 4517 4518 /* Can the task run on the task's current CPU? If so, we're done */ 4519 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4520 goto out; 4521 4522 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4523 if (p->on_rq) { 4524 struct migration_arg arg = { p, dest_cpu }; 4525 /* Need help from migration thread: drop lock and wait. */ 4526 task_rq_unlock(rq, p, &flags); 4527 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4528 tlb_migrate_finish(p->mm); 4529 return 0; 4530 } 4531out: 4532 task_rq_unlock(rq, p, &flags); 4533 4534 return ret; 4535} 4536EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4537 4538/* 4539 * Move (not current) task off this cpu, onto dest cpu. We're doing 4540 * this because either it can't run here any more (set_cpus_allowed() 4541 * away from this CPU, or CPU going down), or because we're 4542 * attempting to rebalance this task on exec (sched_exec). 4543 * 4544 * So we race with normal scheduler movements, but that's OK, as long 4545 * as the task is no longer on this CPU. 4546 * 4547 * Returns non-zero if task was successfully migrated. 4548 */ 4549static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4550{ 4551 struct rq *rq_dest, *rq_src; 4552 int ret = 0; 4553 4554 if (unlikely(!cpu_active(dest_cpu))) 4555 return ret; 4556 4557 rq_src = cpu_rq(src_cpu); 4558 rq_dest = cpu_rq(dest_cpu); 4559 4560 raw_spin_lock(&p->pi_lock); 4561 double_rq_lock(rq_src, rq_dest); 4562 /* Already moved. */ 4563 if (task_cpu(p) != src_cpu) 4564 goto done; 4565 /* Affinity changed (again). */ 4566 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4567 goto fail; 4568 4569 /* 4570 * If we're not on a rq, the next wake-up will ensure we're 4571 * placed properly. 4572 */ 4573 if (p->on_rq) { 4574 dequeue_task(rq_src, p, 0); 4575 set_task_cpu(p, dest_cpu); 4576 enqueue_task(rq_dest, p, 0); 4577 check_preempt_curr(rq_dest, p, 0); 4578 } 4579done: 4580 ret = 1; 4581fail: 4582 double_rq_unlock(rq_src, rq_dest); 4583 raw_spin_unlock(&p->pi_lock); 4584 return ret; 4585} 4586 4587#ifdef CONFIG_NUMA_BALANCING 4588/* Migrate current task p to target_cpu */ 4589int migrate_task_to(struct task_struct *p, int target_cpu) 4590{ 4591 struct migration_arg arg = { p, target_cpu }; 4592 int curr_cpu = task_cpu(p); 4593 4594 if (curr_cpu == target_cpu) 4595 return 0; 4596 4597 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4598 return -EINVAL; 4599 4600 /* TODO: This is not properly updating schedstats */ 4601 4602 trace_sched_move_numa(p, curr_cpu, target_cpu); 4603 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4604} 4605 4606/* 4607 * Requeue a task on a given node and accurately track the number of NUMA 4608 * tasks on the runqueues 4609 */ 4610void sched_setnuma(struct task_struct *p, int nid) 4611{ 4612 struct rq *rq; 4613 unsigned long flags; 4614 bool on_rq, running; 4615 4616 rq = task_rq_lock(p, &flags); 4617 on_rq = p->on_rq; 4618 running = task_current(rq, p); 4619 4620 if (on_rq) 4621 dequeue_task(rq, p, 0); 4622 if (running) 4623 p->sched_class->put_prev_task(rq, p); 4624 4625 p->numa_preferred_nid = nid; 4626 4627 if (running) 4628 p->sched_class->set_curr_task(rq); 4629 if (on_rq) 4630 enqueue_task(rq, p, 0); 4631 task_rq_unlock(rq, p, &flags); 4632} 4633#endif 4634 4635/* 4636 * migration_cpu_stop - this will be executed by a highprio stopper thread 4637 * and performs thread migration by bumping thread off CPU then 4638 * 'pushing' onto another runqueue. 4639 */ 4640static int migration_cpu_stop(void *data) 4641{ 4642 struct migration_arg *arg = data; 4643 4644 /* 4645 * The original target cpu might have gone down and we might 4646 * be on another cpu but it doesn't matter. 4647 */ 4648 local_irq_disable(); 4649 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4650 local_irq_enable(); 4651 return 0; 4652} 4653 4654#ifdef CONFIG_HOTPLUG_CPU 4655 4656/* 4657 * Ensures that the idle task is using init_mm right before its cpu goes 4658 * offline. 4659 */ 4660void idle_task_exit(void) 4661{ 4662 struct mm_struct *mm = current->active_mm; 4663 4664 BUG_ON(cpu_online(smp_processor_id())); 4665 4666 if (mm != &init_mm) 4667 switch_mm(mm, &init_mm, current); 4668 mmdrop(mm); 4669} 4670 4671/* 4672 * Since this CPU is going 'away' for a while, fold any nr_active delta 4673 * we might have. Assumes we're called after migrate_tasks() so that the 4674 * nr_active count is stable. 4675 * 4676 * Also see the comment "Global load-average calculations". 4677 */ 4678static void calc_load_migrate(struct rq *rq) 4679{ 4680 long delta = calc_load_fold_active(rq); 4681 if (delta) 4682 atomic_long_add(delta, &calc_load_tasks); 4683} 4684 4685static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4686{ 4687} 4688 4689static const struct sched_class fake_sched_class = { 4690 .put_prev_task = put_prev_task_fake, 4691}; 4692 4693static struct task_struct fake_task = { 4694 /* 4695 * Avoid pull_{rt,dl}_task() 4696 */ 4697 .prio = MAX_PRIO + 1, 4698 .sched_class = &fake_sched_class, 4699}; 4700 4701/* 4702 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4703 * try_to_wake_up()->select_task_rq(). 4704 * 4705 * Called with rq->lock held even though we'er in stop_machine() and 4706 * there's no concurrency possible, we hold the required locks anyway 4707 * because of lock validation efforts. 4708 */ 4709static void migrate_tasks(unsigned int dead_cpu) 4710{ 4711 struct rq *rq = cpu_rq(dead_cpu); 4712 struct task_struct *next, *stop = rq->stop; 4713 int dest_cpu; 4714 4715 /* 4716 * Fudge the rq selection such that the below task selection loop 4717 * doesn't get stuck on the currently eligible stop task. 4718 * 4719 * We're currently inside stop_machine() and the rq is either stuck 4720 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4721 * either way we should never end up calling schedule() until we're 4722 * done here. 4723 */ 4724 rq->stop = NULL; 4725 4726 /* 4727 * put_prev_task() and pick_next_task() sched 4728 * class method both need to have an up-to-date 4729 * value of rq->clock[_task] 4730 */ 4731 update_rq_clock(rq); 4732 4733 for ( ; ; ) { 4734 /* 4735 * There's this thread running, bail when that's the only 4736 * remaining thread. 4737 */ 4738 if (rq->nr_running == 1) 4739 break; 4740 4741 next = pick_next_task(rq, &fake_task); 4742 BUG_ON(!next); 4743 next->sched_class->put_prev_task(rq, next); 4744 4745 /* Find suitable destination for @next, with force if needed. */ 4746 dest_cpu = select_fallback_rq(dead_cpu, next); 4747 raw_spin_unlock(&rq->lock); 4748 4749 __migrate_task(next, dead_cpu, dest_cpu); 4750 4751 raw_spin_lock(&rq->lock); 4752 } 4753 4754 rq->stop = stop; 4755} 4756 4757#endif /* CONFIG_HOTPLUG_CPU */ 4758 4759#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4760 4761static struct ctl_table sd_ctl_dir[] = { 4762 { 4763 .procname = "sched_domain", 4764 .mode = 0555, 4765 }, 4766 {} 4767}; 4768 4769static struct ctl_table sd_ctl_root[] = { 4770 { 4771 .procname = "kernel", 4772 .mode = 0555, 4773 .child = sd_ctl_dir, 4774 }, 4775 {} 4776}; 4777 4778static struct ctl_table *sd_alloc_ctl_entry(int n) 4779{ 4780 struct ctl_table *entry = 4781 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4782 4783 return entry; 4784} 4785 4786static void sd_free_ctl_entry(struct ctl_table **tablep) 4787{ 4788 struct ctl_table *entry; 4789 4790 /* 4791 * In the intermediate directories, both the child directory and 4792 * procname are dynamically allocated and could fail but the mode 4793 * will always be set. In the lowest directory the names are 4794 * static strings and all have proc handlers. 4795 */ 4796 for (entry = *tablep; entry->mode; entry++) { 4797 if (entry->child) 4798 sd_free_ctl_entry(&entry->child); 4799 if (entry->proc_handler == NULL) 4800 kfree(entry->procname); 4801 } 4802 4803 kfree(*tablep); 4804 *tablep = NULL; 4805} 4806 4807static int min_load_idx = 0; 4808static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4809 4810static void 4811set_table_entry(struct ctl_table *entry, 4812 const char *procname, void *data, int maxlen, 4813 umode_t mode, proc_handler *proc_handler, 4814 bool load_idx) 4815{ 4816 entry->procname = procname; 4817 entry->data = data; 4818 entry->maxlen = maxlen; 4819 entry->mode = mode; 4820 entry->proc_handler = proc_handler; 4821 4822 if (load_idx) { 4823 entry->extra1 = &min_load_idx; 4824 entry->extra2 = &max_load_idx; 4825 } 4826} 4827 4828static struct ctl_table * 4829sd_alloc_ctl_domain_table(struct sched_domain *sd) 4830{ 4831 struct ctl_table *table = sd_alloc_ctl_entry(14); 4832 4833 if (table == NULL) 4834 return NULL; 4835 4836 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4837 sizeof(long), 0644, proc_doulongvec_minmax, false); 4838 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4839 sizeof(long), 0644, proc_doulongvec_minmax, false); 4840 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4841 sizeof(int), 0644, proc_dointvec_minmax, true); 4842 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4843 sizeof(int), 0644, proc_dointvec_minmax, true); 4844 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4845 sizeof(int), 0644, proc_dointvec_minmax, true); 4846 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4847 sizeof(int), 0644, proc_dointvec_minmax, true); 4848 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4849 sizeof(int), 0644, proc_dointvec_minmax, true); 4850 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4851 sizeof(int), 0644, proc_dointvec_minmax, false); 4852 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4853 sizeof(int), 0644, proc_dointvec_minmax, false); 4854 set_table_entry(&table[9], "cache_nice_tries", 4855 &sd->cache_nice_tries, 4856 sizeof(int), 0644, proc_dointvec_minmax, false); 4857 set_table_entry(&table[10], "flags", &sd->flags, 4858 sizeof(int), 0644, proc_dointvec_minmax, false); 4859 set_table_entry(&table[11], "max_newidle_lb_cost", 4860 &sd->max_newidle_lb_cost, 4861 sizeof(long), 0644, proc_doulongvec_minmax, false); 4862 set_table_entry(&table[12], "name", sd->name, 4863 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4864 /* &table[13] is terminator */ 4865 4866 return table; 4867} 4868 4869static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4870{ 4871 struct ctl_table *entry, *table; 4872 struct sched_domain *sd; 4873 int domain_num = 0, i; 4874 char buf[32]; 4875 4876 for_each_domain(cpu, sd) 4877 domain_num++; 4878 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4879 if (table == NULL) 4880 return NULL; 4881 4882 i = 0; 4883 for_each_domain(cpu, sd) { 4884 snprintf(buf, 32, "domain%d", i); 4885 entry->procname = kstrdup(buf, GFP_KERNEL); 4886 entry->mode = 0555; 4887 entry->child = sd_alloc_ctl_domain_table(sd); 4888 entry++; 4889 i++; 4890 } 4891 return table; 4892} 4893 4894static struct ctl_table_header *sd_sysctl_header; 4895static void register_sched_domain_sysctl(void) 4896{ 4897 int i, cpu_num = num_possible_cpus(); 4898 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4899 char buf[32]; 4900 4901 WARN_ON(sd_ctl_dir[0].child); 4902 sd_ctl_dir[0].child = entry; 4903 4904 if (entry == NULL) 4905 return; 4906 4907 for_each_possible_cpu(i) { 4908 snprintf(buf, 32, "cpu%d", i); 4909 entry->procname = kstrdup(buf, GFP_KERNEL); 4910 entry->mode = 0555; 4911 entry->child = sd_alloc_ctl_cpu_table(i); 4912 entry++; 4913 } 4914 4915 WARN_ON(sd_sysctl_header); 4916 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4917} 4918 4919/* may be called multiple times per register */ 4920static void unregister_sched_domain_sysctl(void) 4921{ 4922 if (sd_sysctl_header) 4923 unregister_sysctl_table(sd_sysctl_header); 4924 sd_sysctl_header = NULL; 4925 if (sd_ctl_dir[0].child) 4926 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4927} 4928#else 4929static void register_sched_domain_sysctl(void) 4930{ 4931} 4932static void unregister_sched_domain_sysctl(void) 4933{ 4934} 4935#endif 4936 4937static void set_rq_online(struct rq *rq) 4938{ 4939 if (!rq->online) { 4940 const struct sched_class *class; 4941 4942 cpumask_set_cpu(rq->cpu, rq->rd->online); 4943 rq->online = 1; 4944 4945 for_each_class(class) { 4946 if (class->rq_online) 4947 class->rq_online(rq); 4948 } 4949 } 4950} 4951 4952static void set_rq_offline(struct rq *rq) 4953{ 4954 if (rq->online) { 4955 const struct sched_class *class; 4956 4957 for_each_class(class) { 4958 if (class->rq_offline) 4959 class->rq_offline(rq); 4960 } 4961 4962 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4963 rq->online = 0; 4964 } 4965} 4966 4967/* 4968 * migration_call - callback that gets triggered when a CPU is added. 4969 * Here we can start up the necessary migration thread for the new CPU. 4970 */ 4971static int 4972migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4973{ 4974 int cpu = (long)hcpu; 4975 unsigned long flags; 4976 struct rq *rq = cpu_rq(cpu); 4977 4978 switch (action & ~CPU_TASKS_FROZEN) { 4979 4980 case CPU_UP_PREPARE: 4981 rq->calc_load_update = calc_load_update; 4982 break; 4983 4984 case CPU_ONLINE: 4985 /* Update our root-domain */ 4986 raw_spin_lock_irqsave(&rq->lock, flags); 4987 if (rq->rd) { 4988 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4989 4990 set_rq_online(rq); 4991 } 4992 raw_spin_unlock_irqrestore(&rq->lock, flags); 4993 break; 4994 4995#ifdef CONFIG_HOTPLUG_CPU 4996 case CPU_DYING: 4997 sched_ttwu_pending(); 4998 /* Update our root-domain */ 4999 raw_spin_lock_irqsave(&rq->lock, flags); 5000 if (rq->rd) { 5001 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5002 set_rq_offline(rq); 5003 } 5004 migrate_tasks(cpu); 5005 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5006 raw_spin_unlock_irqrestore(&rq->lock, flags); 5007 break; 5008 5009 case CPU_DEAD: 5010 calc_load_migrate(rq); 5011 break; 5012#endif 5013 } 5014 5015 update_max_interval(); 5016 5017 return NOTIFY_OK; 5018} 5019 5020/* 5021 * Register at high priority so that task migration (migrate_all_tasks) 5022 * happens before everything else. This has to be lower priority than 5023 * the notifier in the perf_event subsystem, though. 5024 */ 5025static struct notifier_block migration_notifier = { 5026 .notifier_call = migration_call, 5027 .priority = CPU_PRI_MIGRATION, 5028}; 5029 5030static int sched_cpu_active(struct notifier_block *nfb, 5031 unsigned long action, void *hcpu) 5032{ 5033 switch (action & ~CPU_TASKS_FROZEN) { 5034 case CPU_STARTING: 5035 case CPU_DOWN_FAILED: 5036 set_cpu_active((long)hcpu, true); 5037 return NOTIFY_OK; 5038 default: 5039 return NOTIFY_DONE; 5040 } 5041} 5042 5043static int sched_cpu_inactive(struct notifier_block *nfb, 5044 unsigned long action, void *hcpu) 5045{ 5046 unsigned long flags; 5047 long cpu = (long)hcpu; 5048 5049 switch (action & ~CPU_TASKS_FROZEN) { 5050 case CPU_DOWN_PREPARE: 5051 set_cpu_active(cpu, false); 5052 5053 /* explicitly allow suspend */ 5054 if (!(action & CPU_TASKS_FROZEN)) { 5055 struct dl_bw *dl_b = dl_bw_of(cpu); 5056 bool overflow; 5057 int cpus; 5058 5059 raw_spin_lock_irqsave(&dl_b->lock, flags); 5060 cpus = dl_bw_cpus(cpu); 5061 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5062 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5063 5064 if (overflow) 5065 return notifier_from_errno(-EBUSY); 5066 } 5067 return NOTIFY_OK; 5068 } 5069 5070 return NOTIFY_DONE; 5071} 5072 5073static int __init migration_init(void) 5074{ 5075 void *cpu = (void *)(long)smp_processor_id(); 5076 int err; 5077 5078 /* Initialize migration for the boot CPU */ 5079 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5080 BUG_ON(err == NOTIFY_BAD); 5081 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5082 register_cpu_notifier(&migration_notifier); 5083 5084 /* Register cpu active notifiers */ 5085 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5086 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5087 5088 return 0; 5089} 5090early_initcall(migration_init); 5091#endif 5092 5093#ifdef CONFIG_SMP 5094 5095static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5096 5097#ifdef CONFIG_SCHED_DEBUG 5098 5099static __read_mostly int sched_debug_enabled; 5100 5101static int __init sched_debug_setup(char *str) 5102{ 5103 sched_debug_enabled = 1; 5104 5105 return 0; 5106} 5107early_param("sched_debug", sched_debug_setup); 5108 5109static inline bool sched_debug(void) 5110{ 5111 return sched_debug_enabled; 5112} 5113 5114static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5115 struct cpumask *groupmask) 5116{ 5117 struct sched_group *group = sd->groups; 5118 char str[256]; 5119 5120 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5121 cpumask_clear(groupmask); 5122 5123 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5124 5125 if (!(sd->flags & SD_LOAD_BALANCE)) { 5126 printk("does not load-balance\n"); 5127 if (sd->parent) 5128 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5129 " has parent"); 5130 return -1; 5131 } 5132 5133 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5134 5135 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5136 printk(KERN_ERR "ERROR: domain->span does not contain " 5137 "CPU%d\n", cpu); 5138 } 5139 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5140 printk(KERN_ERR "ERROR: domain->groups does not contain" 5141 " CPU%d\n", cpu); 5142 } 5143 5144 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5145 do { 5146 if (!group) { 5147 printk("\n"); 5148 printk(KERN_ERR "ERROR: group is NULL\n"); 5149 break; 5150 } 5151 5152 /* 5153 * Even though we initialize ->power to something semi-sane, 5154 * we leave power_orig unset. This allows us to detect if 5155 * domain iteration is still funny without causing /0 traps. 5156 */ 5157 if (!group->sgp->power_orig) { 5158 printk(KERN_CONT "\n"); 5159 printk(KERN_ERR "ERROR: domain->cpu_power not " 5160 "set\n"); 5161 break; 5162 } 5163 5164 if (!cpumask_weight(sched_group_cpus(group))) { 5165 printk(KERN_CONT "\n"); 5166 printk(KERN_ERR "ERROR: empty group\n"); 5167 break; 5168 } 5169 5170 if (!(sd->flags & SD_OVERLAP) && 5171 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5172 printk(KERN_CONT "\n"); 5173 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5174 break; 5175 } 5176 5177 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5178 5179 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5180 5181 printk(KERN_CONT " %s", str); 5182 if (group->sgp->power != SCHED_POWER_SCALE) { 5183 printk(KERN_CONT " (cpu_power = %d)", 5184 group->sgp->power); 5185 } 5186 5187 group = group->next; 5188 } while (group != sd->groups); 5189 printk(KERN_CONT "\n"); 5190 5191 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5192 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5193 5194 if (sd->parent && 5195 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5196 printk(KERN_ERR "ERROR: parent span is not a superset " 5197 "of domain->span\n"); 5198 return 0; 5199} 5200 5201static void sched_domain_debug(struct sched_domain *sd, int cpu) 5202{ 5203 int level = 0; 5204 5205 if (!sched_debug_enabled) 5206 return; 5207 5208 if (!sd) { 5209 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5210 return; 5211 } 5212 5213 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5214 5215 for (;;) { 5216 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5217 break; 5218 level++; 5219 sd = sd->parent; 5220 if (!sd) 5221 break; 5222 } 5223} 5224#else /* !CONFIG_SCHED_DEBUG */ 5225# define sched_domain_debug(sd, cpu) do { } while (0) 5226static inline bool sched_debug(void) 5227{ 5228 return false; 5229} 5230#endif /* CONFIG_SCHED_DEBUG */ 5231 5232static int sd_degenerate(struct sched_domain *sd) 5233{ 5234 if (cpumask_weight(sched_domain_span(sd)) == 1) 5235 return 1; 5236 5237 /* Following flags need at least 2 groups */ 5238 if (sd->flags & (SD_LOAD_BALANCE | 5239 SD_BALANCE_NEWIDLE | 5240 SD_BALANCE_FORK | 5241 SD_BALANCE_EXEC | 5242 SD_SHARE_CPUPOWER | 5243 SD_SHARE_PKG_RESOURCES)) { 5244 if (sd->groups != sd->groups->next) 5245 return 0; 5246 } 5247 5248 /* Following flags don't use groups */ 5249 if (sd->flags & (SD_WAKE_AFFINE)) 5250 return 0; 5251 5252 return 1; 5253} 5254 5255static int 5256sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5257{ 5258 unsigned long cflags = sd->flags, pflags = parent->flags; 5259 5260 if (sd_degenerate(parent)) 5261 return 1; 5262 5263 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5264 return 0; 5265 5266 /* Flags needing groups don't count if only 1 group in parent */ 5267 if (parent->groups == parent->groups->next) { 5268 pflags &= ~(SD_LOAD_BALANCE | 5269 SD_BALANCE_NEWIDLE | 5270 SD_BALANCE_FORK | 5271 SD_BALANCE_EXEC | 5272 SD_SHARE_CPUPOWER | 5273 SD_SHARE_PKG_RESOURCES | 5274 SD_PREFER_SIBLING); 5275 if (nr_node_ids == 1) 5276 pflags &= ~SD_SERIALIZE; 5277 } 5278 if (~cflags & pflags) 5279 return 0; 5280 5281 return 1; 5282} 5283 5284static void free_rootdomain(struct rcu_head *rcu) 5285{ 5286 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5287 5288 cpupri_cleanup(&rd->cpupri); 5289 cpudl_cleanup(&rd->cpudl); 5290 free_cpumask_var(rd->dlo_mask); 5291 free_cpumask_var(rd->rto_mask); 5292 free_cpumask_var(rd->online); 5293 free_cpumask_var(rd->span); 5294 kfree(rd); 5295} 5296 5297static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5298{ 5299 struct root_domain *old_rd = NULL; 5300 unsigned long flags; 5301 5302 raw_spin_lock_irqsave(&rq->lock, flags); 5303 5304 if (rq->rd) { 5305 old_rd = rq->rd; 5306 5307 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5308 set_rq_offline(rq); 5309 5310 cpumask_clear_cpu(rq->cpu, old_rd->span); 5311 5312 /* 5313 * If we dont want to free the old_rd yet then 5314 * set old_rd to NULL to skip the freeing later 5315 * in this function: 5316 */ 5317 if (!atomic_dec_and_test(&old_rd->refcount)) 5318 old_rd = NULL; 5319 } 5320 5321 atomic_inc(&rd->refcount); 5322 rq->rd = rd; 5323 5324 cpumask_set_cpu(rq->cpu, rd->span); 5325 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5326 set_rq_online(rq); 5327 5328 raw_spin_unlock_irqrestore(&rq->lock, flags); 5329 5330 if (old_rd) 5331 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5332} 5333 5334static int init_rootdomain(struct root_domain *rd) 5335{ 5336 memset(rd, 0, sizeof(*rd)); 5337 5338 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5339 goto out; 5340 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5341 goto free_span; 5342 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5343 goto free_online; 5344 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5345 goto free_dlo_mask; 5346 5347 init_dl_bw(&rd->dl_bw); 5348 if (cpudl_init(&rd->cpudl) != 0) 5349 goto free_dlo_mask; 5350 5351 if (cpupri_init(&rd->cpupri) != 0) 5352 goto free_rto_mask; 5353 return 0; 5354 5355free_rto_mask: 5356 free_cpumask_var(rd->rto_mask); 5357free_dlo_mask: 5358 free_cpumask_var(rd->dlo_mask); 5359free_online: 5360 free_cpumask_var(rd->online); 5361free_span: 5362 free_cpumask_var(rd->span); 5363out: 5364 return -ENOMEM; 5365} 5366 5367/* 5368 * By default the system creates a single root-domain with all cpus as 5369 * members (mimicking the global state we have today). 5370 */ 5371struct root_domain def_root_domain; 5372 5373static void init_defrootdomain(void) 5374{ 5375 init_rootdomain(&def_root_domain); 5376 5377 atomic_set(&def_root_domain.refcount, 1); 5378} 5379 5380static struct root_domain *alloc_rootdomain(void) 5381{ 5382 struct root_domain *rd; 5383 5384 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5385 if (!rd) 5386 return NULL; 5387 5388 if (init_rootdomain(rd) != 0) { 5389 kfree(rd); 5390 return NULL; 5391 } 5392 5393 return rd; 5394} 5395 5396static void free_sched_groups(struct sched_group *sg, int free_sgp) 5397{ 5398 struct sched_group *tmp, *first; 5399 5400 if (!sg) 5401 return; 5402 5403 first = sg; 5404 do { 5405 tmp = sg->next; 5406 5407 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5408 kfree(sg->sgp); 5409 5410 kfree(sg); 5411 sg = tmp; 5412 } while (sg != first); 5413} 5414 5415static void free_sched_domain(struct rcu_head *rcu) 5416{ 5417 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5418 5419 /* 5420 * If its an overlapping domain it has private groups, iterate and 5421 * nuke them all. 5422 */ 5423 if (sd->flags & SD_OVERLAP) { 5424 free_sched_groups(sd->groups, 1); 5425 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5426 kfree(sd->groups->sgp); 5427 kfree(sd->groups); 5428 } 5429 kfree(sd); 5430} 5431 5432static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5433{ 5434 call_rcu(&sd->rcu, free_sched_domain); 5435} 5436 5437static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5438{ 5439 for (; sd; sd = sd->parent) 5440 destroy_sched_domain(sd, cpu); 5441} 5442 5443/* 5444 * Keep a special pointer to the highest sched_domain that has 5445 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5446 * allows us to avoid some pointer chasing select_idle_sibling(). 5447 * 5448 * Also keep a unique ID per domain (we use the first cpu number in 5449 * the cpumask of the domain), this allows us to quickly tell if 5450 * two cpus are in the same cache domain, see cpus_share_cache(). 5451 */ 5452DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5453DEFINE_PER_CPU(int, sd_llc_size); 5454DEFINE_PER_CPU(int, sd_llc_id); 5455DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5456DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5457DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5458 5459static void update_top_cache_domain(int cpu) 5460{ 5461 struct sched_domain *sd; 5462 struct sched_domain *busy_sd = NULL; 5463 int id = cpu; 5464 int size = 1; 5465 5466 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5467 if (sd) { 5468 id = cpumask_first(sched_domain_span(sd)); 5469 size = cpumask_weight(sched_domain_span(sd)); 5470 busy_sd = sd->parent; /* sd_busy */ 5471 } 5472 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5473 5474 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5475 per_cpu(sd_llc_size, cpu) = size; 5476 per_cpu(sd_llc_id, cpu) = id; 5477 5478 sd = lowest_flag_domain(cpu, SD_NUMA); 5479 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5480 5481 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5482 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5483} 5484 5485/* 5486 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5487 * hold the hotplug lock. 5488 */ 5489static void 5490cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5491{ 5492 struct rq *rq = cpu_rq(cpu); 5493 struct sched_domain *tmp; 5494 5495 /* Remove the sched domains which do not contribute to scheduling. */ 5496 for (tmp = sd; tmp; ) { 5497 struct sched_domain *parent = tmp->parent; 5498 if (!parent) 5499 break; 5500 5501 if (sd_parent_degenerate(tmp, parent)) { 5502 tmp->parent = parent->parent; 5503 if (parent->parent) 5504 parent->parent->child = tmp; 5505 /* 5506 * Transfer SD_PREFER_SIBLING down in case of a 5507 * degenerate parent; the spans match for this 5508 * so the property transfers. 5509 */ 5510 if (parent->flags & SD_PREFER_SIBLING) 5511 tmp->flags |= SD_PREFER_SIBLING; 5512 destroy_sched_domain(parent, cpu); 5513 } else 5514 tmp = tmp->parent; 5515 } 5516 5517 if (sd && sd_degenerate(sd)) { 5518 tmp = sd; 5519 sd = sd->parent; 5520 destroy_sched_domain(tmp, cpu); 5521 if (sd) 5522 sd->child = NULL; 5523 } 5524 5525 sched_domain_debug(sd, cpu); 5526 5527 rq_attach_root(rq, rd); 5528 tmp = rq->sd; 5529 rcu_assign_pointer(rq->sd, sd); 5530 destroy_sched_domains(tmp, cpu); 5531 5532 update_top_cache_domain(cpu); 5533} 5534 5535/* cpus with isolated domains */ 5536static cpumask_var_t cpu_isolated_map; 5537 5538/* Setup the mask of cpus configured for isolated domains */ 5539static int __init isolated_cpu_setup(char *str) 5540{ 5541 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5542 cpulist_parse(str, cpu_isolated_map); 5543 return 1; 5544} 5545 5546__setup("isolcpus=", isolated_cpu_setup); 5547 5548static const struct cpumask *cpu_cpu_mask(int cpu) 5549{ 5550 return cpumask_of_node(cpu_to_node(cpu)); 5551} 5552 5553struct sd_data { 5554 struct sched_domain **__percpu sd; 5555 struct sched_group **__percpu sg; 5556 struct sched_group_power **__percpu sgp; 5557}; 5558 5559struct s_data { 5560 struct sched_domain ** __percpu sd; 5561 struct root_domain *rd; 5562}; 5563 5564enum s_alloc { 5565 sa_rootdomain, 5566 sa_sd, 5567 sa_sd_storage, 5568 sa_none, 5569}; 5570 5571struct sched_domain_topology_level; 5572 5573typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5574typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5575 5576#define SDTL_OVERLAP 0x01 5577 5578struct sched_domain_topology_level { 5579 sched_domain_init_f init; 5580 sched_domain_mask_f mask; 5581 int flags; 5582 int numa_level; 5583 struct sd_data data; 5584}; 5585 5586/* 5587 * Build an iteration mask that can exclude certain CPUs from the upwards 5588 * domain traversal. 5589 * 5590 * Asymmetric node setups can result in situations where the domain tree is of 5591 * unequal depth, make sure to skip domains that already cover the entire 5592 * range. 5593 * 5594 * In that case build_sched_domains() will have terminated the iteration early 5595 * and our sibling sd spans will be empty. Domains should always include the 5596 * cpu they're built on, so check that. 5597 * 5598 */ 5599static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5600{ 5601 const struct cpumask *span = sched_domain_span(sd); 5602 struct sd_data *sdd = sd->private; 5603 struct sched_domain *sibling; 5604 int i; 5605 5606 for_each_cpu(i, span) { 5607 sibling = *per_cpu_ptr(sdd->sd, i); 5608 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5609 continue; 5610 5611 cpumask_set_cpu(i, sched_group_mask(sg)); 5612 } 5613} 5614 5615/* 5616 * Return the canonical balance cpu for this group, this is the first cpu 5617 * of this group that's also in the iteration mask. 5618 */ 5619int group_balance_cpu(struct sched_group *sg) 5620{ 5621 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5622} 5623 5624static int 5625build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5626{ 5627 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5628 const struct cpumask *span = sched_domain_span(sd); 5629 struct cpumask *covered = sched_domains_tmpmask; 5630 struct sd_data *sdd = sd->private; 5631 struct sched_domain *child; 5632 int i; 5633 5634 cpumask_clear(covered); 5635 5636 for_each_cpu(i, span) { 5637 struct cpumask *sg_span; 5638 5639 if (cpumask_test_cpu(i, covered)) 5640 continue; 5641 5642 child = *per_cpu_ptr(sdd->sd, i); 5643 5644 /* See the comment near build_group_mask(). */ 5645 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5646 continue; 5647 5648 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5649 GFP_KERNEL, cpu_to_node(cpu)); 5650 5651 if (!sg) 5652 goto fail; 5653 5654 sg_span = sched_group_cpus(sg); 5655 if (child->child) { 5656 child = child->child; 5657 cpumask_copy(sg_span, sched_domain_span(child)); 5658 } else 5659 cpumask_set_cpu(i, sg_span); 5660 5661 cpumask_or(covered, covered, sg_span); 5662 5663 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5664 if (atomic_inc_return(&sg->sgp->ref) == 1) 5665 build_group_mask(sd, sg); 5666 5667 /* 5668 * Initialize sgp->power such that even if we mess up the 5669 * domains and no possible iteration will get us here, we won't 5670 * die on a /0 trap. 5671 */ 5672 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5673 sg->sgp->power_orig = sg->sgp->power; 5674 5675 /* 5676 * Make sure the first group of this domain contains the 5677 * canonical balance cpu. Otherwise the sched_domain iteration 5678 * breaks. See update_sg_lb_stats(). 5679 */ 5680 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5681 group_balance_cpu(sg) == cpu) 5682 groups = sg; 5683 5684 if (!first) 5685 first = sg; 5686 if (last) 5687 last->next = sg; 5688 last = sg; 5689 last->next = first; 5690 } 5691 sd->groups = groups; 5692 5693 return 0; 5694 5695fail: 5696 free_sched_groups(first, 0); 5697 5698 return -ENOMEM; 5699} 5700 5701static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5702{ 5703 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5704 struct sched_domain *child = sd->child; 5705 5706 if (child) 5707 cpu = cpumask_first(sched_domain_span(child)); 5708 5709 if (sg) { 5710 *sg = *per_cpu_ptr(sdd->sg, cpu); 5711 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5712 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5713 } 5714 5715 return cpu; 5716} 5717 5718/* 5719 * build_sched_groups will build a circular linked list of the groups 5720 * covered by the given span, and will set each group's ->cpumask correctly, 5721 * and ->cpu_power to 0. 5722 * 5723 * Assumes the sched_domain tree is fully constructed 5724 */ 5725static int 5726build_sched_groups(struct sched_domain *sd, int cpu) 5727{ 5728 struct sched_group *first = NULL, *last = NULL; 5729 struct sd_data *sdd = sd->private; 5730 const struct cpumask *span = sched_domain_span(sd); 5731 struct cpumask *covered; 5732 int i; 5733 5734 get_group(cpu, sdd, &sd->groups); 5735 atomic_inc(&sd->groups->ref); 5736 5737 if (cpu != cpumask_first(span)) 5738 return 0; 5739 5740 lockdep_assert_held(&sched_domains_mutex); 5741 covered = sched_domains_tmpmask; 5742 5743 cpumask_clear(covered); 5744 5745 for_each_cpu(i, span) { 5746 struct sched_group *sg; 5747 int group, j; 5748 5749 if (cpumask_test_cpu(i, covered)) 5750 continue; 5751 5752 group = get_group(i, sdd, &sg); 5753 cpumask_clear(sched_group_cpus(sg)); 5754 sg->sgp->power = 0; 5755 cpumask_setall(sched_group_mask(sg)); 5756 5757 for_each_cpu(j, span) { 5758 if (get_group(j, sdd, NULL) != group) 5759 continue; 5760 5761 cpumask_set_cpu(j, covered); 5762 cpumask_set_cpu(j, sched_group_cpus(sg)); 5763 } 5764 5765 if (!first) 5766 first = sg; 5767 if (last) 5768 last->next = sg; 5769 last = sg; 5770 } 5771 last->next = first; 5772 5773 return 0; 5774} 5775 5776/* 5777 * Initialize sched groups cpu_power. 5778 * 5779 * cpu_power indicates the capacity of sched group, which is used while 5780 * distributing the load between different sched groups in a sched domain. 5781 * Typically cpu_power for all the groups in a sched domain will be same unless 5782 * there are asymmetries in the topology. If there are asymmetries, group 5783 * having more cpu_power will pickup more load compared to the group having 5784 * less cpu_power. 5785 */ 5786static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5787{ 5788 struct sched_group *sg = sd->groups; 5789 5790 WARN_ON(!sg); 5791 5792 do { 5793 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5794 sg = sg->next; 5795 } while (sg != sd->groups); 5796 5797 if (cpu != group_balance_cpu(sg)) 5798 return; 5799 5800 update_group_power(sd, cpu); 5801 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5802} 5803 5804int __weak arch_sd_sibling_asym_packing(void) 5805{ 5806 return 0*SD_ASYM_PACKING; 5807} 5808 5809/* 5810 * Initializers for schedule domains 5811 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5812 */ 5813 5814#ifdef CONFIG_SCHED_DEBUG 5815# define SD_INIT_NAME(sd, type) sd->name = #type 5816#else 5817# define SD_INIT_NAME(sd, type) do { } while (0) 5818#endif 5819 5820#define SD_INIT_FUNC(type) \ 5821static noinline struct sched_domain * \ 5822sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5823{ \ 5824 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5825 *sd = SD_##type##_INIT; \ 5826 SD_INIT_NAME(sd, type); \ 5827 sd->private = &tl->data; \ 5828 return sd; \ 5829} 5830 5831SD_INIT_FUNC(CPU) 5832#ifdef CONFIG_SCHED_SMT 5833 SD_INIT_FUNC(SIBLING) 5834#endif 5835#ifdef CONFIG_SCHED_MC 5836 SD_INIT_FUNC(MC) 5837#endif 5838#ifdef CONFIG_SCHED_BOOK 5839 SD_INIT_FUNC(BOOK) 5840#endif 5841 5842static int default_relax_domain_level = -1; 5843int sched_domain_level_max; 5844 5845static int __init setup_relax_domain_level(char *str) 5846{ 5847 if (kstrtoint(str, 0, &default_relax_domain_level)) 5848 pr_warn("Unable to set relax_domain_level\n"); 5849 5850 return 1; 5851} 5852__setup("relax_domain_level=", setup_relax_domain_level); 5853 5854static void set_domain_attribute(struct sched_domain *sd, 5855 struct sched_domain_attr *attr) 5856{ 5857 int request; 5858 5859 if (!attr || attr->relax_domain_level < 0) { 5860 if (default_relax_domain_level < 0) 5861 return; 5862 else 5863 request = default_relax_domain_level; 5864 } else 5865 request = attr->relax_domain_level; 5866 if (request < sd->level) { 5867 /* turn off idle balance on this domain */ 5868 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5869 } else { 5870 /* turn on idle balance on this domain */ 5871 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5872 } 5873} 5874 5875static void __sdt_free(const struct cpumask *cpu_map); 5876static int __sdt_alloc(const struct cpumask *cpu_map); 5877 5878static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5879 const struct cpumask *cpu_map) 5880{ 5881 switch (what) { 5882 case sa_rootdomain: 5883 if (!atomic_read(&d->rd->refcount)) 5884 free_rootdomain(&d->rd->rcu); /* fall through */ 5885 case sa_sd: 5886 free_percpu(d->sd); /* fall through */ 5887 case sa_sd_storage: 5888 __sdt_free(cpu_map); /* fall through */ 5889 case sa_none: 5890 break; 5891 } 5892} 5893 5894static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5895 const struct cpumask *cpu_map) 5896{ 5897 memset(d, 0, sizeof(*d)); 5898 5899 if (__sdt_alloc(cpu_map)) 5900 return sa_sd_storage; 5901 d->sd = alloc_percpu(struct sched_domain *); 5902 if (!d->sd) 5903 return sa_sd_storage; 5904 d->rd = alloc_rootdomain(); 5905 if (!d->rd) 5906 return sa_sd; 5907 return sa_rootdomain; 5908} 5909 5910/* 5911 * NULL the sd_data elements we've used to build the sched_domain and 5912 * sched_group structure so that the subsequent __free_domain_allocs() 5913 * will not free the data we're using. 5914 */ 5915static void claim_allocations(int cpu, struct sched_domain *sd) 5916{ 5917 struct sd_data *sdd = sd->private; 5918 5919 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5920 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5921 5922 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5923 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5924 5925 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5926 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5927} 5928 5929#ifdef CONFIG_SCHED_SMT 5930static const struct cpumask *cpu_smt_mask(int cpu) 5931{ 5932 return topology_thread_cpumask(cpu); 5933} 5934#endif 5935 5936/* 5937 * Topology list, bottom-up. 5938 */ 5939static struct sched_domain_topology_level default_topology[] = { 5940#ifdef CONFIG_SCHED_SMT 5941 { sd_init_SIBLING, cpu_smt_mask, }, 5942#endif 5943#ifdef CONFIG_SCHED_MC 5944 { sd_init_MC, cpu_coregroup_mask, }, 5945#endif 5946#ifdef CONFIG_SCHED_BOOK 5947 { sd_init_BOOK, cpu_book_mask, }, 5948#endif 5949 { sd_init_CPU, cpu_cpu_mask, }, 5950 { NULL, }, 5951}; 5952 5953static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5954 5955#define for_each_sd_topology(tl) \ 5956 for (tl = sched_domain_topology; tl->init; tl++) 5957 5958#ifdef CONFIG_NUMA 5959 5960static int sched_domains_numa_levels; 5961static int *sched_domains_numa_distance; 5962static struct cpumask ***sched_domains_numa_masks; 5963static int sched_domains_curr_level; 5964 5965static inline int sd_local_flags(int level) 5966{ 5967 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5968 return 0; 5969 5970 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5971} 5972 5973static struct sched_domain * 5974sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5975{ 5976 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5977 int level = tl->numa_level; 5978 int sd_weight = cpumask_weight( 5979 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5980 5981 *sd = (struct sched_domain){ 5982 .min_interval = sd_weight, 5983 .max_interval = 2*sd_weight, 5984 .busy_factor = 32, 5985 .imbalance_pct = 125, 5986 .cache_nice_tries = 2, 5987 .busy_idx = 3, 5988 .idle_idx = 2, 5989 .newidle_idx = 0, 5990 .wake_idx = 0, 5991 .forkexec_idx = 0, 5992 5993 .flags = 1*SD_LOAD_BALANCE 5994 | 1*SD_BALANCE_NEWIDLE 5995 | 0*SD_BALANCE_EXEC 5996 | 0*SD_BALANCE_FORK 5997 | 0*SD_BALANCE_WAKE 5998 | 0*SD_WAKE_AFFINE 5999 | 0*SD_SHARE_CPUPOWER 6000 | 0*SD_SHARE_PKG_RESOURCES 6001 | 1*SD_SERIALIZE 6002 | 0*SD_PREFER_SIBLING 6003 | 1*SD_NUMA 6004 | sd_local_flags(level) 6005 , 6006 .last_balance = jiffies, 6007 .balance_interval = sd_weight, 6008 }; 6009 SD_INIT_NAME(sd, NUMA); 6010 sd->private = &tl->data; 6011 6012 /* 6013 * Ugly hack to pass state to sd_numa_mask()... 6014 */ 6015 sched_domains_curr_level = tl->numa_level; 6016 6017 return sd; 6018} 6019 6020static const struct cpumask *sd_numa_mask(int cpu) 6021{ 6022 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6023} 6024 6025static void sched_numa_warn(const char *str) 6026{ 6027 static int done = false; 6028 int i,j; 6029 6030 if (done) 6031 return; 6032 6033 done = true; 6034 6035 printk(KERN_WARNING "ERROR: %s\n\n", str); 6036 6037 for (i = 0; i < nr_node_ids; i++) { 6038 printk(KERN_WARNING " "); 6039 for (j = 0; j < nr_node_ids; j++) 6040 printk(KERN_CONT "%02d ", node_distance(i,j)); 6041 printk(KERN_CONT "\n"); 6042 } 6043 printk(KERN_WARNING "\n"); 6044} 6045 6046static bool find_numa_distance(int distance) 6047{ 6048 int i; 6049 6050 if (distance == node_distance(0, 0)) 6051 return true; 6052 6053 for (i = 0; i < sched_domains_numa_levels; i++) { 6054 if (sched_domains_numa_distance[i] == distance) 6055 return true; 6056 } 6057 6058 return false; 6059} 6060 6061static void sched_init_numa(void) 6062{ 6063 int next_distance, curr_distance = node_distance(0, 0); 6064 struct sched_domain_topology_level *tl; 6065 int level = 0; 6066 int i, j, k; 6067 6068 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6069 if (!sched_domains_numa_distance) 6070 return; 6071 6072 /* 6073 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6074 * unique distances in the node_distance() table. 6075 * 6076 * Assumes node_distance(0,j) includes all distances in 6077 * node_distance(i,j) in order to avoid cubic time. 6078 */ 6079 next_distance = curr_distance; 6080 for (i = 0; i < nr_node_ids; i++) { 6081 for (j = 0; j < nr_node_ids; j++) { 6082 for (k = 0; k < nr_node_ids; k++) { 6083 int distance = node_distance(i, k); 6084 6085 if (distance > curr_distance && 6086 (distance < next_distance || 6087 next_distance == curr_distance)) 6088 next_distance = distance; 6089 6090 /* 6091 * While not a strong assumption it would be nice to know 6092 * about cases where if node A is connected to B, B is not 6093 * equally connected to A. 6094 */ 6095 if (sched_debug() && node_distance(k, i) != distance) 6096 sched_numa_warn("Node-distance not symmetric"); 6097 6098 if (sched_debug() && i && !find_numa_distance(distance)) 6099 sched_numa_warn("Node-0 not representative"); 6100 } 6101 if (next_distance != curr_distance) { 6102 sched_domains_numa_distance[level++] = next_distance; 6103 sched_domains_numa_levels = level; 6104 curr_distance = next_distance; 6105 } else break; 6106 } 6107 6108 /* 6109 * In case of sched_debug() we verify the above assumption. 6110 */ 6111 if (!sched_debug()) 6112 break; 6113 } 6114 /* 6115 * 'level' contains the number of unique distances, excluding the 6116 * identity distance node_distance(i,i). 6117 * 6118 * The sched_domains_numa_distance[] array includes the actual distance 6119 * numbers. 6120 */ 6121 6122 /* 6123 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6124 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6125 * the array will contain less then 'level' members. This could be 6126 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6127 * in other functions. 6128 * 6129 * We reset it to 'level' at the end of this function. 6130 */ 6131 sched_domains_numa_levels = 0; 6132 6133 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6134 if (!sched_domains_numa_masks) 6135 return; 6136 6137 /* 6138 * Now for each level, construct a mask per node which contains all 6139 * cpus of nodes that are that many hops away from us. 6140 */ 6141 for (i = 0; i < level; i++) { 6142 sched_domains_numa_masks[i] = 6143 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6144 if (!sched_domains_numa_masks[i]) 6145 return; 6146 6147 for (j = 0; j < nr_node_ids; j++) { 6148 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6149 if (!mask) 6150 return; 6151 6152 sched_domains_numa_masks[i][j] = mask; 6153 6154 for (k = 0; k < nr_node_ids; k++) { 6155 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6156 continue; 6157 6158 cpumask_or(mask, mask, cpumask_of_node(k)); 6159 } 6160 } 6161 } 6162 6163 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6164 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6165 if (!tl) 6166 return; 6167 6168 /* 6169 * Copy the default topology bits.. 6170 */ 6171 for (i = 0; default_topology[i].init; i++) 6172 tl[i] = default_topology[i]; 6173 6174 /* 6175 * .. and append 'j' levels of NUMA goodness. 6176 */ 6177 for (j = 0; j < level; i++, j++) { 6178 tl[i] = (struct sched_domain_topology_level){ 6179 .init = sd_numa_init, 6180 .mask = sd_numa_mask, 6181 .flags = SDTL_OVERLAP, 6182 .numa_level = j, 6183 }; 6184 } 6185 6186 sched_domain_topology = tl; 6187 6188 sched_domains_numa_levels = level; 6189} 6190 6191static void sched_domains_numa_masks_set(int cpu) 6192{ 6193 int i, j; 6194 int node = cpu_to_node(cpu); 6195 6196 for (i = 0; i < sched_domains_numa_levels; i++) { 6197 for (j = 0; j < nr_node_ids; j++) { 6198 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6199 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6200 } 6201 } 6202} 6203 6204static void sched_domains_numa_masks_clear(int cpu) 6205{ 6206 int i, j; 6207 for (i = 0; i < sched_domains_numa_levels; i++) { 6208 for (j = 0; j < nr_node_ids; j++) 6209 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6210 } 6211} 6212 6213/* 6214 * Update sched_domains_numa_masks[level][node] array when new cpus 6215 * are onlined. 6216 */ 6217static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6218 unsigned long action, 6219 void *hcpu) 6220{ 6221 int cpu = (long)hcpu; 6222 6223 switch (action & ~CPU_TASKS_FROZEN) { 6224 case CPU_ONLINE: 6225 sched_domains_numa_masks_set(cpu); 6226 break; 6227 6228 case CPU_DEAD: 6229 sched_domains_numa_masks_clear(cpu); 6230 break; 6231 6232 default: 6233 return NOTIFY_DONE; 6234 } 6235 6236 return NOTIFY_OK; 6237} 6238#else 6239static inline void sched_init_numa(void) 6240{ 6241} 6242 6243static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6244 unsigned long action, 6245 void *hcpu) 6246{ 6247 return 0; 6248} 6249#endif /* CONFIG_NUMA */ 6250 6251static int __sdt_alloc(const struct cpumask *cpu_map) 6252{ 6253 struct sched_domain_topology_level *tl; 6254 int j; 6255 6256 for_each_sd_topology(tl) { 6257 struct sd_data *sdd = &tl->data; 6258 6259 sdd->sd = alloc_percpu(struct sched_domain *); 6260 if (!sdd->sd) 6261 return -ENOMEM; 6262 6263 sdd->sg = alloc_percpu(struct sched_group *); 6264 if (!sdd->sg) 6265 return -ENOMEM; 6266 6267 sdd->sgp = alloc_percpu(struct sched_group_power *); 6268 if (!sdd->sgp) 6269 return -ENOMEM; 6270 6271 for_each_cpu(j, cpu_map) { 6272 struct sched_domain *sd; 6273 struct sched_group *sg; 6274 struct sched_group_power *sgp; 6275 6276 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6277 GFP_KERNEL, cpu_to_node(j)); 6278 if (!sd) 6279 return -ENOMEM; 6280 6281 *per_cpu_ptr(sdd->sd, j) = sd; 6282 6283 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6284 GFP_KERNEL, cpu_to_node(j)); 6285 if (!sg) 6286 return -ENOMEM; 6287 6288 sg->next = sg; 6289 6290 *per_cpu_ptr(sdd->sg, j) = sg; 6291 6292 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6293 GFP_KERNEL, cpu_to_node(j)); 6294 if (!sgp) 6295 return -ENOMEM; 6296 6297 *per_cpu_ptr(sdd->sgp, j) = sgp; 6298 } 6299 } 6300 6301 return 0; 6302} 6303 6304static void __sdt_free(const struct cpumask *cpu_map) 6305{ 6306 struct sched_domain_topology_level *tl; 6307 int j; 6308 6309 for_each_sd_topology(tl) { 6310 struct sd_data *sdd = &tl->data; 6311 6312 for_each_cpu(j, cpu_map) { 6313 struct sched_domain *sd; 6314 6315 if (sdd->sd) { 6316 sd = *per_cpu_ptr(sdd->sd, j); 6317 if (sd && (sd->flags & SD_OVERLAP)) 6318 free_sched_groups(sd->groups, 0); 6319 kfree(*per_cpu_ptr(sdd->sd, j)); 6320 } 6321 6322 if (sdd->sg) 6323 kfree(*per_cpu_ptr(sdd->sg, j)); 6324 if (sdd->sgp) 6325 kfree(*per_cpu_ptr(sdd->sgp, j)); 6326 } 6327 free_percpu(sdd->sd); 6328 sdd->sd = NULL; 6329 free_percpu(sdd->sg); 6330 sdd->sg = NULL; 6331 free_percpu(sdd->sgp); 6332 sdd->sgp = NULL; 6333 } 6334} 6335 6336struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6337 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6338 struct sched_domain *child, int cpu) 6339{ 6340 struct sched_domain *sd = tl->init(tl, cpu); 6341 if (!sd) 6342 return child; 6343 6344 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6345 if (child) { 6346 sd->level = child->level + 1; 6347 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6348 child->parent = sd; 6349 sd->child = child; 6350 } 6351 set_domain_attribute(sd, attr); 6352 6353 return sd; 6354} 6355 6356/* 6357 * Build sched domains for a given set of cpus and attach the sched domains 6358 * to the individual cpus 6359 */ 6360static int build_sched_domains(const struct cpumask *cpu_map, 6361 struct sched_domain_attr *attr) 6362{ 6363 enum s_alloc alloc_state; 6364 struct sched_domain *sd; 6365 struct s_data d; 6366 int i, ret = -ENOMEM; 6367 6368 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6369 if (alloc_state != sa_rootdomain) 6370 goto error; 6371 6372 /* Set up domains for cpus specified by the cpu_map. */ 6373 for_each_cpu(i, cpu_map) { 6374 struct sched_domain_topology_level *tl; 6375 6376 sd = NULL; 6377 for_each_sd_topology(tl) { 6378 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6379 if (tl == sched_domain_topology) 6380 *per_cpu_ptr(d.sd, i) = sd; 6381 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6382 sd->flags |= SD_OVERLAP; 6383 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6384 break; 6385 } 6386 } 6387 6388 /* Build the groups for the domains */ 6389 for_each_cpu(i, cpu_map) { 6390 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6391 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6392 if (sd->flags & SD_OVERLAP) { 6393 if (build_overlap_sched_groups(sd, i)) 6394 goto error; 6395 } else { 6396 if (build_sched_groups(sd, i)) 6397 goto error; 6398 } 6399 } 6400 } 6401 6402 /* Calculate CPU power for physical packages and nodes */ 6403 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6404 if (!cpumask_test_cpu(i, cpu_map)) 6405 continue; 6406 6407 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6408 claim_allocations(i, sd); 6409 init_sched_groups_power(i, sd); 6410 } 6411 } 6412 6413 /* Attach the domains */ 6414 rcu_read_lock(); 6415 for_each_cpu(i, cpu_map) { 6416 sd = *per_cpu_ptr(d.sd, i); 6417 cpu_attach_domain(sd, d.rd, i); 6418 } 6419 rcu_read_unlock(); 6420 6421 ret = 0; 6422error: 6423 __free_domain_allocs(&d, alloc_state, cpu_map); 6424 return ret; 6425} 6426 6427static cpumask_var_t *doms_cur; /* current sched domains */ 6428static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6429static struct sched_domain_attr *dattr_cur; 6430 /* attribues of custom domains in 'doms_cur' */ 6431 6432/* 6433 * Special case: If a kmalloc of a doms_cur partition (array of 6434 * cpumask) fails, then fallback to a single sched domain, 6435 * as determined by the single cpumask fallback_doms. 6436 */ 6437static cpumask_var_t fallback_doms; 6438 6439/* 6440 * arch_update_cpu_topology lets virtualized architectures update the 6441 * cpu core maps. It is supposed to return 1 if the topology changed 6442 * or 0 if it stayed the same. 6443 */ 6444int __attribute__((weak)) arch_update_cpu_topology(void) 6445{ 6446 return 0; 6447} 6448 6449cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6450{ 6451 int i; 6452 cpumask_var_t *doms; 6453 6454 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6455 if (!doms) 6456 return NULL; 6457 for (i = 0; i < ndoms; i++) { 6458 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6459 free_sched_domains(doms, i); 6460 return NULL; 6461 } 6462 } 6463 return doms; 6464} 6465 6466void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6467{ 6468 unsigned int i; 6469 for (i = 0; i < ndoms; i++) 6470 free_cpumask_var(doms[i]); 6471 kfree(doms); 6472} 6473 6474/* 6475 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6476 * For now this just excludes isolated cpus, but could be used to 6477 * exclude other special cases in the future. 6478 */ 6479static int init_sched_domains(const struct cpumask *cpu_map) 6480{ 6481 int err; 6482 6483 arch_update_cpu_topology(); 6484 ndoms_cur = 1; 6485 doms_cur = alloc_sched_domains(ndoms_cur); 6486 if (!doms_cur) 6487 doms_cur = &fallback_doms; 6488 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6489 err = build_sched_domains(doms_cur[0], NULL); 6490 register_sched_domain_sysctl(); 6491 6492 return err; 6493} 6494 6495/* 6496 * Detach sched domains from a group of cpus specified in cpu_map 6497 * These cpus will now be attached to the NULL domain 6498 */ 6499static void detach_destroy_domains(const struct cpumask *cpu_map) 6500{ 6501 int i; 6502 6503 rcu_read_lock(); 6504 for_each_cpu(i, cpu_map) 6505 cpu_attach_domain(NULL, &def_root_domain, i); 6506 rcu_read_unlock(); 6507} 6508 6509/* handle null as "default" */ 6510static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6511 struct sched_domain_attr *new, int idx_new) 6512{ 6513 struct sched_domain_attr tmp; 6514 6515 /* fast path */ 6516 if (!new && !cur) 6517 return 1; 6518 6519 tmp = SD_ATTR_INIT; 6520 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6521 new ? (new + idx_new) : &tmp, 6522 sizeof(struct sched_domain_attr)); 6523} 6524 6525/* 6526 * Partition sched domains as specified by the 'ndoms_new' 6527 * cpumasks in the array doms_new[] of cpumasks. This compares 6528 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6529 * It destroys each deleted domain and builds each new domain. 6530 * 6531 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6532 * The masks don't intersect (don't overlap.) We should setup one 6533 * sched domain for each mask. CPUs not in any of the cpumasks will 6534 * not be load balanced. If the same cpumask appears both in the 6535 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6536 * it as it is. 6537 * 6538 * The passed in 'doms_new' should be allocated using 6539 * alloc_sched_domains. This routine takes ownership of it and will 6540 * free_sched_domains it when done with it. If the caller failed the 6541 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6542 * and partition_sched_domains() will fallback to the single partition 6543 * 'fallback_doms', it also forces the domains to be rebuilt. 6544 * 6545 * If doms_new == NULL it will be replaced with cpu_online_mask. 6546 * ndoms_new == 0 is a special case for destroying existing domains, 6547 * and it will not create the default domain. 6548 * 6549 * Call with hotplug lock held 6550 */ 6551void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6552 struct sched_domain_attr *dattr_new) 6553{ 6554 int i, j, n; 6555 int new_topology; 6556 6557 mutex_lock(&sched_domains_mutex); 6558 6559 /* always unregister in case we don't destroy any domains */ 6560 unregister_sched_domain_sysctl(); 6561 6562 /* Let architecture update cpu core mappings. */ 6563 new_topology = arch_update_cpu_topology(); 6564 6565 n = doms_new ? ndoms_new : 0; 6566 6567 /* Destroy deleted domains */ 6568 for (i = 0; i < ndoms_cur; i++) { 6569 for (j = 0; j < n && !new_topology; j++) { 6570 if (cpumask_equal(doms_cur[i], doms_new[j]) 6571 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6572 goto match1; 6573 } 6574 /* no match - a current sched domain not in new doms_new[] */ 6575 detach_destroy_domains(doms_cur[i]); 6576match1: 6577 ; 6578 } 6579 6580 n = ndoms_cur; 6581 if (doms_new == NULL) { 6582 n = 0; 6583 doms_new = &fallback_doms; 6584 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6585 WARN_ON_ONCE(dattr_new); 6586 } 6587 6588 /* Build new domains */ 6589 for (i = 0; i < ndoms_new; i++) { 6590 for (j = 0; j < n && !new_topology; j++) { 6591 if (cpumask_equal(doms_new[i], doms_cur[j]) 6592 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6593 goto match2; 6594 } 6595 /* no match - add a new doms_new */ 6596 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6597match2: 6598 ; 6599 } 6600 6601 /* Remember the new sched domains */ 6602 if (doms_cur != &fallback_doms) 6603 free_sched_domains(doms_cur, ndoms_cur); 6604 kfree(dattr_cur); /* kfree(NULL) is safe */ 6605 doms_cur = doms_new; 6606 dattr_cur = dattr_new; 6607 ndoms_cur = ndoms_new; 6608 6609 register_sched_domain_sysctl(); 6610 6611 mutex_unlock(&sched_domains_mutex); 6612} 6613 6614static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6615 6616/* 6617 * Update cpusets according to cpu_active mask. If cpusets are 6618 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6619 * around partition_sched_domains(). 6620 * 6621 * If we come here as part of a suspend/resume, don't touch cpusets because we 6622 * want to restore it back to its original state upon resume anyway. 6623 */ 6624static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6625 void *hcpu) 6626{ 6627 switch (action) { 6628 case CPU_ONLINE_FROZEN: 6629 case CPU_DOWN_FAILED_FROZEN: 6630 6631 /* 6632 * num_cpus_frozen tracks how many CPUs are involved in suspend 6633 * resume sequence. As long as this is not the last online 6634 * operation in the resume sequence, just build a single sched 6635 * domain, ignoring cpusets. 6636 */ 6637 num_cpus_frozen--; 6638 if (likely(num_cpus_frozen)) { 6639 partition_sched_domains(1, NULL, NULL); 6640 break; 6641 } 6642 6643 /* 6644 * This is the last CPU online operation. So fall through and 6645 * restore the original sched domains by considering the 6646 * cpuset configurations. 6647 */ 6648 6649 case CPU_ONLINE: 6650 case CPU_DOWN_FAILED: 6651 cpuset_update_active_cpus(true); 6652 break; 6653 default: 6654 return NOTIFY_DONE; 6655 } 6656 return NOTIFY_OK; 6657} 6658 6659static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6660 void *hcpu) 6661{ 6662 switch (action) { 6663 case CPU_DOWN_PREPARE: 6664 cpuset_update_active_cpus(false); 6665 break; 6666 case CPU_DOWN_PREPARE_FROZEN: 6667 num_cpus_frozen++; 6668 partition_sched_domains(1, NULL, NULL); 6669 break; 6670 default: 6671 return NOTIFY_DONE; 6672 } 6673 return NOTIFY_OK; 6674} 6675 6676void __init sched_init_smp(void) 6677{ 6678 cpumask_var_t non_isolated_cpus; 6679 6680 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6681 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6682 6683 sched_init_numa(); 6684 6685 /* 6686 * There's no userspace yet to cause hotplug operations; hence all the 6687 * cpu masks are stable and all blatant races in the below code cannot 6688 * happen. 6689 */ 6690 mutex_lock(&sched_domains_mutex); 6691 init_sched_domains(cpu_active_mask); 6692 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6693 if (cpumask_empty(non_isolated_cpus)) 6694 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6695 mutex_unlock(&sched_domains_mutex); 6696 6697 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6698 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6699 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6700 6701 init_hrtick(); 6702 6703 /* Move init over to a non-isolated CPU */ 6704 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6705 BUG(); 6706 sched_init_granularity(); 6707 free_cpumask_var(non_isolated_cpus); 6708 6709 init_sched_rt_class(); 6710 init_sched_dl_class(); 6711} 6712#else 6713void __init sched_init_smp(void) 6714{ 6715 sched_init_granularity(); 6716} 6717#endif /* CONFIG_SMP */ 6718 6719const_debug unsigned int sysctl_timer_migration = 1; 6720 6721int in_sched_functions(unsigned long addr) 6722{ 6723 return in_lock_functions(addr) || 6724 (addr >= (unsigned long)__sched_text_start 6725 && addr < (unsigned long)__sched_text_end); 6726} 6727 6728#ifdef CONFIG_CGROUP_SCHED 6729/* 6730 * Default task group. 6731 * Every task in system belongs to this group at bootup. 6732 */ 6733struct task_group root_task_group; 6734LIST_HEAD(task_groups); 6735#endif 6736 6737DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6738 6739void __init sched_init(void) 6740{ 6741 int i, j; 6742 unsigned long alloc_size = 0, ptr; 6743 6744#ifdef CONFIG_FAIR_GROUP_SCHED 6745 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6746#endif 6747#ifdef CONFIG_RT_GROUP_SCHED 6748 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6749#endif 6750#ifdef CONFIG_CPUMASK_OFFSTACK 6751 alloc_size += num_possible_cpus() * cpumask_size(); 6752#endif 6753 if (alloc_size) { 6754 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6755 6756#ifdef CONFIG_FAIR_GROUP_SCHED 6757 root_task_group.se = (struct sched_entity **)ptr; 6758 ptr += nr_cpu_ids * sizeof(void **); 6759 6760 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6761 ptr += nr_cpu_ids * sizeof(void **); 6762 6763#endif /* CONFIG_FAIR_GROUP_SCHED */ 6764#ifdef CONFIG_RT_GROUP_SCHED 6765 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6766 ptr += nr_cpu_ids * sizeof(void **); 6767 6768 root_task_group.rt_rq = (struct rt_rq **)ptr; 6769 ptr += nr_cpu_ids * sizeof(void **); 6770 6771#endif /* CONFIG_RT_GROUP_SCHED */ 6772#ifdef CONFIG_CPUMASK_OFFSTACK 6773 for_each_possible_cpu(i) { 6774 per_cpu(load_balance_mask, i) = (void *)ptr; 6775 ptr += cpumask_size(); 6776 } 6777#endif /* CONFIG_CPUMASK_OFFSTACK */ 6778 } 6779 6780 init_rt_bandwidth(&def_rt_bandwidth, 6781 global_rt_period(), global_rt_runtime()); 6782 init_dl_bandwidth(&def_dl_bandwidth, 6783 global_rt_period(), global_rt_runtime()); 6784 6785#ifdef CONFIG_SMP 6786 init_defrootdomain(); 6787#endif 6788 6789#ifdef CONFIG_RT_GROUP_SCHED 6790 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6791 global_rt_period(), global_rt_runtime()); 6792#endif /* CONFIG_RT_GROUP_SCHED */ 6793 6794#ifdef CONFIG_CGROUP_SCHED 6795 list_add(&root_task_group.list, &task_groups); 6796 INIT_LIST_HEAD(&root_task_group.children); 6797 INIT_LIST_HEAD(&root_task_group.siblings); 6798 autogroup_init(&init_task); 6799 6800#endif /* CONFIG_CGROUP_SCHED */ 6801 6802 for_each_possible_cpu(i) { 6803 struct rq *rq; 6804 6805 rq = cpu_rq(i); 6806 raw_spin_lock_init(&rq->lock); 6807 rq->nr_running = 0; 6808 rq->calc_load_active = 0; 6809 rq->calc_load_update = jiffies + LOAD_FREQ; 6810 init_cfs_rq(&rq->cfs); 6811 init_rt_rq(&rq->rt, rq); 6812 init_dl_rq(&rq->dl, rq); 6813#ifdef CONFIG_FAIR_GROUP_SCHED 6814 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6815 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6816 /* 6817 * How much cpu bandwidth does root_task_group get? 6818 * 6819 * In case of task-groups formed thr' the cgroup filesystem, it 6820 * gets 100% of the cpu resources in the system. This overall 6821 * system cpu resource is divided among the tasks of 6822 * root_task_group and its child task-groups in a fair manner, 6823 * based on each entity's (task or task-group's) weight 6824 * (se->load.weight). 6825 * 6826 * In other words, if root_task_group has 10 tasks of weight 6827 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6828 * then A0's share of the cpu resource is: 6829 * 6830 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6831 * 6832 * We achieve this by letting root_task_group's tasks sit 6833 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6834 */ 6835 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6836 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6837#endif /* CONFIG_FAIR_GROUP_SCHED */ 6838 6839 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6840#ifdef CONFIG_RT_GROUP_SCHED 6841 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6842 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6843#endif 6844 6845 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6846 rq->cpu_load[j] = 0; 6847 6848 rq->last_load_update_tick = jiffies; 6849 6850#ifdef CONFIG_SMP 6851 rq->sd = NULL; 6852 rq->rd = NULL; 6853 rq->cpu_power = SCHED_POWER_SCALE; 6854 rq->post_schedule = 0; 6855 rq->active_balance = 0; 6856 rq->next_balance = jiffies; 6857 rq->push_cpu = 0; 6858 rq->cpu = i; 6859 rq->online = 0; 6860 rq->idle_stamp = 0; 6861 rq->avg_idle = 2*sysctl_sched_migration_cost; 6862 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6863 6864 INIT_LIST_HEAD(&rq->cfs_tasks); 6865 6866 rq_attach_root(rq, &def_root_domain); 6867#ifdef CONFIG_NO_HZ_COMMON 6868 rq->nohz_flags = 0; 6869#endif 6870#ifdef CONFIG_NO_HZ_FULL 6871 rq->last_sched_tick = 0; 6872#endif 6873#endif 6874 init_rq_hrtick(rq); 6875 atomic_set(&rq->nr_iowait, 0); 6876 } 6877 6878 set_load_weight(&init_task); 6879 6880#ifdef CONFIG_PREEMPT_NOTIFIERS 6881 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6882#endif 6883 6884 /* 6885 * The boot idle thread does lazy MMU switching as well: 6886 */ 6887 atomic_inc(&init_mm.mm_count); 6888 enter_lazy_tlb(&init_mm, current); 6889 6890 /* 6891 * Make us the idle thread. Technically, schedule() should not be 6892 * called from this thread, however somewhere below it might be, 6893 * but because we are the idle thread, we just pick up running again 6894 * when this runqueue becomes "idle". 6895 */ 6896 init_idle(current, smp_processor_id()); 6897 6898 calc_load_update = jiffies + LOAD_FREQ; 6899 6900 /* 6901 * During early bootup we pretend to be a normal task: 6902 */ 6903 current->sched_class = &fair_sched_class; 6904 6905#ifdef CONFIG_SMP 6906 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6907 /* May be allocated at isolcpus cmdline parse time */ 6908 if (cpu_isolated_map == NULL) 6909 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6910 idle_thread_set_boot_cpu(); 6911#endif 6912 init_sched_fair_class(); 6913 6914 scheduler_running = 1; 6915} 6916 6917#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6918static inline int preempt_count_equals(int preempt_offset) 6919{ 6920 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6921 6922 return (nested == preempt_offset); 6923} 6924 6925void __might_sleep(const char *file, int line, int preempt_offset) 6926{ 6927 static unsigned long prev_jiffy; /* ratelimiting */ 6928 6929 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6930 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 6931 !is_idle_task(current)) || 6932 system_state != SYSTEM_RUNNING || oops_in_progress) 6933 return; 6934 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6935 return; 6936 prev_jiffy = jiffies; 6937 6938 printk(KERN_ERR 6939 "BUG: sleeping function called from invalid context at %s:%d\n", 6940 file, line); 6941 printk(KERN_ERR 6942 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6943 in_atomic(), irqs_disabled(), 6944 current->pid, current->comm); 6945 6946 debug_show_held_locks(current); 6947 if (irqs_disabled()) 6948 print_irqtrace_events(current); 6949 dump_stack(); 6950} 6951EXPORT_SYMBOL(__might_sleep); 6952#endif 6953 6954#ifdef CONFIG_MAGIC_SYSRQ 6955static void normalize_task(struct rq *rq, struct task_struct *p) 6956{ 6957 const struct sched_class *prev_class = p->sched_class; 6958 struct sched_attr attr = { 6959 .sched_policy = SCHED_NORMAL, 6960 }; 6961 int old_prio = p->prio; 6962 int on_rq; 6963 6964 on_rq = p->on_rq; 6965 if (on_rq) 6966 dequeue_task(rq, p, 0); 6967 __setscheduler(rq, p, &attr); 6968 if (on_rq) { 6969 enqueue_task(rq, p, 0); 6970 resched_task(rq->curr); 6971 } 6972 6973 check_class_changed(rq, p, prev_class, old_prio); 6974} 6975 6976void normalize_rt_tasks(void) 6977{ 6978 struct task_struct *g, *p; 6979 unsigned long flags; 6980 struct rq *rq; 6981 6982 read_lock_irqsave(&tasklist_lock, flags); 6983 do_each_thread(g, p) { 6984 /* 6985 * Only normalize user tasks: 6986 */ 6987 if (!p->mm) 6988 continue; 6989 6990 p->se.exec_start = 0; 6991#ifdef CONFIG_SCHEDSTATS 6992 p->se.statistics.wait_start = 0; 6993 p->se.statistics.sleep_start = 0; 6994 p->se.statistics.block_start = 0; 6995#endif 6996 6997 if (!dl_task(p) && !rt_task(p)) { 6998 /* 6999 * Renice negative nice level userspace 7000 * tasks back to 0: 7001 */ 7002 if (task_nice(p) < 0 && p->mm) 7003 set_user_nice(p, 0); 7004 continue; 7005 } 7006 7007 raw_spin_lock(&p->pi_lock); 7008 rq = __task_rq_lock(p); 7009 7010 normalize_task(rq, p); 7011 7012 __task_rq_unlock(rq); 7013 raw_spin_unlock(&p->pi_lock); 7014 } while_each_thread(g, p); 7015 7016 read_unlock_irqrestore(&tasklist_lock, flags); 7017} 7018 7019#endif /* CONFIG_MAGIC_SYSRQ */ 7020 7021#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7022/* 7023 * These functions are only useful for the IA64 MCA handling, or kdb. 7024 * 7025 * They can only be called when the whole system has been 7026 * stopped - every CPU needs to be quiescent, and no scheduling 7027 * activity can take place. Using them for anything else would 7028 * be a serious bug, and as a result, they aren't even visible 7029 * under any other configuration. 7030 */ 7031 7032/** 7033 * curr_task - return the current task for a given cpu. 7034 * @cpu: the processor in question. 7035 * 7036 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7037 * 7038 * Return: The current task for @cpu. 7039 */ 7040struct task_struct *curr_task(int cpu) 7041{ 7042 return cpu_curr(cpu); 7043} 7044 7045#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7046 7047#ifdef CONFIG_IA64 7048/** 7049 * set_curr_task - set the current task for a given cpu. 7050 * @cpu: the processor in question. 7051 * @p: the task pointer to set. 7052 * 7053 * Description: This function must only be used when non-maskable interrupts 7054 * are serviced on a separate stack. It allows the architecture to switch the 7055 * notion of the current task on a cpu in a non-blocking manner. This function 7056 * must be called with all CPU's synchronized, and interrupts disabled, the 7057 * and caller must save the original value of the current task (see 7058 * curr_task() above) and restore that value before reenabling interrupts and 7059 * re-starting the system. 7060 * 7061 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7062 */ 7063void set_curr_task(int cpu, struct task_struct *p) 7064{ 7065 cpu_curr(cpu) = p; 7066} 7067 7068#endif 7069 7070#ifdef CONFIG_CGROUP_SCHED 7071/* task_group_lock serializes the addition/removal of task groups */ 7072static DEFINE_SPINLOCK(task_group_lock); 7073 7074static void free_sched_group(struct task_group *tg) 7075{ 7076 free_fair_sched_group(tg); 7077 free_rt_sched_group(tg); 7078 autogroup_free(tg); 7079 kfree(tg); 7080} 7081 7082/* allocate runqueue etc for a new task group */ 7083struct task_group *sched_create_group(struct task_group *parent) 7084{ 7085 struct task_group *tg; 7086 7087 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7088 if (!tg) 7089 return ERR_PTR(-ENOMEM); 7090 7091 if (!alloc_fair_sched_group(tg, parent)) 7092 goto err; 7093 7094 if (!alloc_rt_sched_group(tg, parent)) 7095 goto err; 7096 7097 return tg; 7098 7099err: 7100 free_sched_group(tg); 7101 return ERR_PTR(-ENOMEM); 7102} 7103 7104void sched_online_group(struct task_group *tg, struct task_group *parent) 7105{ 7106 unsigned long flags; 7107 7108 spin_lock_irqsave(&task_group_lock, flags); 7109 list_add_rcu(&tg->list, &task_groups); 7110 7111 WARN_ON(!parent); /* root should already exist */ 7112 7113 tg->parent = parent; 7114 INIT_LIST_HEAD(&tg->children); 7115 list_add_rcu(&tg->siblings, &parent->children); 7116 spin_unlock_irqrestore(&task_group_lock, flags); 7117} 7118 7119/* rcu callback to free various structures associated with a task group */ 7120static void free_sched_group_rcu(struct rcu_head *rhp) 7121{ 7122 /* now it should be safe to free those cfs_rqs */ 7123 free_sched_group(container_of(rhp, struct task_group, rcu)); 7124} 7125 7126/* Destroy runqueue etc associated with a task group */ 7127void sched_destroy_group(struct task_group *tg) 7128{ 7129 /* wait for possible concurrent references to cfs_rqs complete */ 7130 call_rcu(&tg->rcu, free_sched_group_rcu); 7131} 7132 7133void sched_offline_group(struct task_group *tg) 7134{ 7135 unsigned long flags; 7136 int i; 7137 7138 /* end participation in shares distribution */ 7139 for_each_possible_cpu(i) 7140 unregister_fair_sched_group(tg, i); 7141 7142 spin_lock_irqsave(&task_group_lock, flags); 7143 list_del_rcu(&tg->list); 7144 list_del_rcu(&tg->siblings); 7145 spin_unlock_irqrestore(&task_group_lock, flags); 7146} 7147 7148/* change task's runqueue when it moves between groups. 7149 * The caller of this function should have put the task in its new group 7150 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7151 * reflect its new group. 7152 */ 7153void sched_move_task(struct task_struct *tsk) 7154{ 7155 struct task_group *tg; 7156 int on_rq, running; 7157 unsigned long flags; 7158 struct rq *rq; 7159 7160 rq = task_rq_lock(tsk, &flags); 7161 7162 running = task_current(rq, tsk); 7163 on_rq = tsk->on_rq; 7164 7165 if (on_rq) 7166 dequeue_task(rq, tsk, 0); 7167 if (unlikely(running)) 7168 tsk->sched_class->put_prev_task(rq, tsk); 7169 7170 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 7171 lockdep_is_held(&tsk->sighand->siglock)), 7172 struct task_group, css); 7173 tg = autogroup_task_group(tsk, tg); 7174 tsk->sched_task_group = tg; 7175 7176#ifdef CONFIG_FAIR_GROUP_SCHED 7177 if (tsk->sched_class->task_move_group) 7178 tsk->sched_class->task_move_group(tsk, on_rq); 7179 else 7180#endif 7181 set_task_rq(tsk, task_cpu(tsk)); 7182 7183 if (unlikely(running)) 7184 tsk->sched_class->set_curr_task(rq); 7185 if (on_rq) 7186 enqueue_task(rq, tsk, 0); 7187 7188 task_rq_unlock(rq, tsk, &flags); 7189} 7190#endif /* CONFIG_CGROUP_SCHED */ 7191 7192#ifdef CONFIG_RT_GROUP_SCHED 7193/* 7194 * Ensure that the real time constraints are schedulable. 7195 */ 7196static DEFINE_MUTEX(rt_constraints_mutex); 7197 7198/* Must be called with tasklist_lock held */ 7199static inline int tg_has_rt_tasks(struct task_group *tg) 7200{ 7201 struct task_struct *g, *p; 7202 7203 do_each_thread(g, p) { 7204 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7205 return 1; 7206 } while_each_thread(g, p); 7207 7208 return 0; 7209} 7210 7211struct rt_schedulable_data { 7212 struct task_group *tg; 7213 u64 rt_period; 7214 u64 rt_runtime; 7215}; 7216 7217static int tg_rt_schedulable(struct task_group *tg, void *data) 7218{ 7219 struct rt_schedulable_data *d = data; 7220 struct task_group *child; 7221 unsigned long total, sum = 0; 7222 u64 period, runtime; 7223 7224 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7225 runtime = tg->rt_bandwidth.rt_runtime; 7226 7227 if (tg == d->tg) { 7228 period = d->rt_period; 7229 runtime = d->rt_runtime; 7230 } 7231 7232 /* 7233 * Cannot have more runtime than the period. 7234 */ 7235 if (runtime > period && runtime != RUNTIME_INF) 7236 return -EINVAL; 7237 7238 /* 7239 * Ensure we don't starve existing RT tasks. 7240 */ 7241 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7242 return -EBUSY; 7243 7244 total = to_ratio(period, runtime); 7245 7246 /* 7247 * Nobody can have more than the global setting allows. 7248 */ 7249 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7250 return -EINVAL; 7251 7252 /* 7253 * The sum of our children's runtime should not exceed our own. 7254 */ 7255 list_for_each_entry_rcu(child, &tg->children, siblings) { 7256 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7257 runtime = child->rt_bandwidth.rt_runtime; 7258 7259 if (child == d->tg) { 7260 period = d->rt_period; 7261 runtime = d->rt_runtime; 7262 } 7263 7264 sum += to_ratio(period, runtime); 7265 } 7266 7267 if (sum > total) 7268 return -EINVAL; 7269 7270 return 0; 7271} 7272 7273static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7274{ 7275 int ret; 7276 7277 struct rt_schedulable_data data = { 7278 .tg = tg, 7279 .rt_period = period, 7280 .rt_runtime = runtime, 7281 }; 7282 7283 rcu_read_lock(); 7284 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7285 rcu_read_unlock(); 7286 7287 return ret; 7288} 7289 7290static int tg_set_rt_bandwidth(struct task_group *tg, 7291 u64 rt_period, u64 rt_runtime) 7292{ 7293 int i, err = 0; 7294 7295 mutex_lock(&rt_constraints_mutex); 7296 read_lock(&tasklist_lock); 7297 err = __rt_schedulable(tg, rt_period, rt_runtime); 7298 if (err) 7299 goto unlock; 7300 7301 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7302 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7303 tg->rt_bandwidth.rt_runtime = rt_runtime; 7304 7305 for_each_possible_cpu(i) { 7306 struct rt_rq *rt_rq = tg->rt_rq[i]; 7307 7308 raw_spin_lock(&rt_rq->rt_runtime_lock); 7309 rt_rq->rt_runtime = rt_runtime; 7310 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7311 } 7312 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7313unlock: 7314 read_unlock(&tasklist_lock); 7315 mutex_unlock(&rt_constraints_mutex); 7316 7317 return err; 7318} 7319 7320static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7321{ 7322 u64 rt_runtime, rt_period; 7323 7324 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7325 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7326 if (rt_runtime_us < 0) 7327 rt_runtime = RUNTIME_INF; 7328 7329 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7330} 7331 7332static long sched_group_rt_runtime(struct task_group *tg) 7333{ 7334 u64 rt_runtime_us; 7335 7336 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7337 return -1; 7338 7339 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7340 do_div(rt_runtime_us, NSEC_PER_USEC); 7341 return rt_runtime_us; 7342} 7343 7344static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7345{ 7346 u64 rt_runtime, rt_period; 7347 7348 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7349 rt_runtime = tg->rt_bandwidth.rt_runtime; 7350 7351 if (rt_period == 0) 7352 return -EINVAL; 7353 7354 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7355} 7356 7357static long sched_group_rt_period(struct task_group *tg) 7358{ 7359 u64 rt_period_us; 7360 7361 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7362 do_div(rt_period_us, NSEC_PER_USEC); 7363 return rt_period_us; 7364} 7365#endif /* CONFIG_RT_GROUP_SCHED */ 7366 7367#ifdef CONFIG_RT_GROUP_SCHED 7368static int sched_rt_global_constraints(void) 7369{ 7370 int ret = 0; 7371 7372 mutex_lock(&rt_constraints_mutex); 7373 read_lock(&tasklist_lock); 7374 ret = __rt_schedulable(NULL, 0, 0); 7375 read_unlock(&tasklist_lock); 7376 mutex_unlock(&rt_constraints_mutex); 7377 7378 return ret; 7379} 7380 7381static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7382{ 7383 /* Don't accept realtime tasks when there is no way for them to run */ 7384 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7385 return 0; 7386 7387 return 1; 7388} 7389 7390#else /* !CONFIG_RT_GROUP_SCHED */ 7391static int sched_rt_global_constraints(void) 7392{ 7393 unsigned long flags; 7394 int i, ret = 0; 7395 7396 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7397 for_each_possible_cpu(i) { 7398 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7399 7400 raw_spin_lock(&rt_rq->rt_runtime_lock); 7401 rt_rq->rt_runtime = global_rt_runtime(); 7402 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7403 } 7404 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7405 7406 return ret; 7407} 7408#endif /* CONFIG_RT_GROUP_SCHED */ 7409 7410static int sched_dl_global_constraints(void) 7411{ 7412 u64 runtime = global_rt_runtime(); 7413 u64 period = global_rt_period(); 7414 u64 new_bw = to_ratio(period, runtime); 7415 int cpu, ret = 0; 7416 7417 /* 7418 * Here we want to check the bandwidth not being set to some 7419 * value smaller than the currently allocated bandwidth in 7420 * any of the root_domains. 7421 * 7422 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7423 * cycling on root_domains... Discussion on different/better 7424 * solutions is welcome! 7425 */ 7426 for_each_possible_cpu(cpu) { 7427 struct dl_bw *dl_b = dl_bw_of(cpu); 7428 7429 raw_spin_lock(&dl_b->lock); 7430 if (new_bw < dl_b->total_bw) 7431 ret = -EBUSY; 7432 raw_spin_unlock(&dl_b->lock); 7433 7434 if (ret) 7435 break; 7436 } 7437 7438 return ret; 7439} 7440 7441static void sched_dl_do_global(void) 7442{ 7443 u64 new_bw = -1; 7444 int cpu; 7445 7446 def_dl_bandwidth.dl_period = global_rt_period(); 7447 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7448 7449 if (global_rt_runtime() != RUNTIME_INF) 7450 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7451 7452 /* 7453 * FIXME: As above... 7454 */ 7455 for_each_possible_cpu(cpu) { 7456 struct dl_bw *dl_b = dl_bw_of(cpu); 7457 7458 raw_spin_lock(&dl_b->lock); 7459 dl_b->bw = new_bw; 7460 raw_spin_unlock(&dl_b->lock); 7461 } 7462} 7463 7464static int sched_rt_global_validate(void) 7465{ 7466 if (sysctl_sched_rt_period <= 0) 7467 return -EINVAL; 7468 7469 if (sysctl_sched_rt_runtime > sysctl_sched_rt_period) 7470 return -EINVAL; 7471 7472 return 0; 7473} 7474 7475static void sched_rt_do_global(void) 7476{ 7477 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7478 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7479} 7480 7481int sched_rt_handler(struct ctl_table *table, int write, 7482 void __user *buffer, size_t *lenp, 7483 loff_t *ppos) 7484{ 7485 int old_period, old_runtime; 7486 static DEFINE_MUTEX(mutex); 7487 int ret; 7488 7489 mutex_lock(&mutex); 7490 old_period = sysctl_sched_rt_period; 7491 old_runtime = sysctl_sched_rt_runtime; 7492 7493 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7494 7495 if (!ret && write) { 7496 ret = sched_rt_global_validate(); 7497 if (ret) 7498 goto undo; 7499 7500 ret = sched_rt_global_constraints(); 7501 if (ret) 7502 goto undo; 7503 7504 ret = sched_dl_global_constraints(); 7505 if (ret) 7506 goto undo; 7507 7508 sched_rt_do_global(); 7509 sched_dl_do_global(); 7510 } 7511 if (0) { 7512undo: 7513 sysctl_sched_rt_period = old_period; 7514 sysctl_sched_rt_runtime = old_runtime; 7515 } 7516 mutex_unlock(&mutex); 7517 7518 return ret; 7519} 7520 7521int sched_rr_handler(struct ctl_table *table, int write, 7522 void __user *buffer, size_t *lenp, 7523 loff_t *ppos) 7524{ 7525 int ret; 7526 static DEFINE_MUTEX(mutex); 7527 7528 mutex_lock(&mutex); 7529 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7530 /* make sure that internally we keep jiffies */ 7531 /* also, writing zero resets timeslice to default */ 7532 if (!ret && write) { 7533 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7534 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7535 } 7536 mutex_unlock(&mutex); 7537 return ret; 7538} 7539 7540#ifdef CONFIG_CGROUP_SCHED 7541 7542static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7543{ 7544 return css ? container_of(css, struct task_group, css) : NULL; 7545} 7546 7547static struct cgroup_subsys_state * 7548cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7549{ 7550 struct task_group *parent = css_tg(parent_css); 7551 struct task_group *tg; 7552 7553 if (!parent) { 7554 /* This is early initialization for the top cgroup */ 7555 return &root_task_group.css; 7556 } 7557 7558 tg = sched_create_group(parent); 7559 if (IS_ERR(tg)) 7560 return ERR_PTR(-ENOMEM); 7561 7562 return &tg->css; 7563} 7564 7565static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7566{ 7567 struct task_group *tg = css_tg(css); 7568 struct task_group *parent = css_tg(css_parent(css)); 7569 7570 if (parent) 7571 sched_online_group(tg, parent); 7572 return 0; 7573} 7574 7575static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7576{ 7577 struct task_group *tg = css_tg(css); 7578 7579 sched_destroy_group(tg); 7580} 7581 7582static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7583{ 7584 struct task_group *tg = css_tg(css); 7585 7586 sched_offline_group(tg); 7587} 7588 7589static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7590 struct cgroup_taskset *tset) 7591{ 7592 struct task_struct *task; 7593 7594 cgroup_taskset_for_each(task, css, tset) { 7595#ifdef CONFIG_RT_GROUP_SCHED 7596 if (!sched_rt_can_attach(css_tg(css), task)) 7597 return -EINVAL; 7598#else 7599 /* We don't support RT-tasks being in separate groups */ 7600 if (task->sched_class != &fair_sched_class) 7601 return -EINVAL; 7602#endif 7603 } 7604 return 0; 7605} 7606 7607static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7608 struct cgroup_taskset *tset) 7609{ 7610 struct task_struct *task; 7611 7612 cgroup_taskset_for_each(task, css, tset) 7613 sched_move_task(task); 7614} 7615 7616static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7617 struct cgroup_subsys_state *old_css, 7618 struct task_struct *task) 7619{ 7620 /* 7621 * cgroup_exit() is called in the copy_process() failure path. 7622 * Ignore this case since the task hasn't ran yet, this avoids 7623 * trying to poke a half freed task state from generic code. 7624 */ 7625 if (!(task->flags & PF_EXITING)) 7626 return; 7627 7628 sched_move_task(task); 7629} 7630 7631#ifdef CONFIG_FAIR_GROUP_SCHED 7632static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7633 struct cftype *cftype, u64 shareval) 7634{ 7635 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7636} 7637 7638static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7639 struct cftype *cft) 7640{ 7641 struct task_group *tg = css_tg(css); 7642 7643 return (u64) scale_load_down(tg->shares); 7644} 7645 7646#ifdef CONFIG_CFS_BANDWIDTH 7647static DEFINE_MUTEX(cfs_constraints_mutex); 7648 7649const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7650const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7651 7652static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7653 7654static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7655{ 7656 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7657 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7658 7659 if (tg == &root_task_group) 7660 return -EINVAL; 7661 7662 /* 7663 * Ensure we have at some amount of bandwidth every period. This is 7664 * to prevent reaching a state of large arrears when throttled via 7665 * entity_tick() resulting in prolonged exit starvation. 7666 */ 7667 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7668 return -EINVAL; 7669 7670 /* 7671 * Likewise, bound things on the otherside by preventing insane quota 7672 * periods. This also allows us to normalize in computing quota 7673 * feasibility. 7674 */ 7675 if (period > max_cfs_quota_period) 7676 return -EINVAL; 7677 7678 mutex_lock(&cfs_constraints_mutex); 7679 ret = __cfs_schedulable(tg, period, quota); 7680 if (ret) 7681 goto out_unlock; 7682 7683 runtime_enabled = quota != RUNTIME_INF; 7684 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7685 /* 7686 * If we need to toggle cfs_bandwidth_used, off->on must occur 7687 * before making related changes, and on->off must occur afterwards 7688 */ 7689 if (runtime_enabled && !runtime_was_enabled) 7690 cfs_bandwidth_usage_inc(); 7691 raw_spin_lock_irq(&cfs_b->lock); 7692 cfs_b->period = ns_to_ktime(period); 7693 cfs_b->quota = quota; 7694 7695 __refill_cfs_bandwidth_runtime(cfs_b); 7696 /* restart the period timer (if active) to handle new period expiry */ 7697 if (runtime_enabled && cfs_b->timer_active) { 7698 /* force a reprogram */ 7699 cfs_b->timer_active = 0; 7700 __start_cfs_bandwidth(cfs_b); 7701 } 7702 raw_spin_unlock_irq(&cfs_b->lock); 7703 7704 for_each_possible_cpu(i) { 7705 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7706 struct rq *rq = cfs_rq->rq; 7707 7708 raw_spin_lock_irq(&rq->lock); 7709 cfs_rq->runtime_enabled = runtime_enabled; 7710 cfs_rq->runtime_remaining = 0; 7711 7712 if (cfs_rq->throttled) 7713 unthrottle_cfs_rq(cfs_rq); 7714 raw_spin_unlock_irq(&rq->lock); 7715 } 7716 if (runtime_was_enabled && !runtime_enabled) 7717 cfs_bandwidth_usage_dec(); 7718out_unlock: 7719 mutex_unlock(&cfs_constraints_mutex); 7720 7721 return ret; 7722} 7723 7724int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7725{ 7726 u64 quota, period; 7727 7728 period = ktime_to_ns(tg->cfs_bandwidth.period); 7729 if (cfs_quota_us < 0) 7730 quota = RUNTIME_INF; 7731 else 7732 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7733 7734 return tg_set_cfs_bandwidth(tg, period, quota); 7735} 7736 7737long tg_get_cfs_quota(struct task_group *tg) 7738{ 7739 u64 quota_us; 7740 7741 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7742 return -1; 7743 7744 quota_us = tg->cfs_bandwidth.quota; 7745 do_div(quota_us, NSEC_PER_USEC); 7746 7747 return quota_us; 7748} 7749 7750int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7751{ 7752 u64 quota, period; 7753 7754 period = (u64)cfs_period_us * NSEC_PER_USEC; 7755 quota = tg->cfs_bandwidth.quota; 7756 7757 return tg_set_cfs_bandwidth(tg, period, quota); 7758} 7759 7760long tg_get_cfs_period(struct task_group *tg) 7761{ 7762 u64 cfs_period_us; 7763 7764 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7765 do_div(cfs_period_us, NSEC_PER_USEC); 7766 7767 return cfs_period_us; 7768} 7769 7770static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7771 struct cftype *cft) 7772{ 7773 return tg_get_cfs_quota(css_tg(css)); 7774} 7775 7776static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7777 struct cftype *cftype, s64 cfs_quota_us) 7778{ 7779 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7780} 7781 7782static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7783 struct cftype *cft) 7784{ 7785 return tg_get_cfs_period(css_tg(css)); 7786} 7787 7788static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7789 struct cftype *cftype, u64 cfs_period_us) 7790{ 7791 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7792} 7793 7794struct cfs_schedulable_data { 7795 struct task_group *tg; 7796 u64 period, quota; 7797}; 7798 7799/* 7800 * normalize group quota/period to be quota/max_period 7801 * note: units are usecs 7802 */ 7803static u64 normalize_cfs_quota(struct task_group *tg, 7804 struct cfs_schedulable_data *d) 7805{ 7806 u64 quota, period; 7807 7808 if (tg == d->tg) { 7809 period = d->period; 7810 quota = d->quota; 7811 } else { 7812 period = tg_get_cfs_period(tg); 7813 quota = tg_get_cfs_quota(tg); 7814 } 7815 7816 /* note: these should typically be equivalent */ 7817 if (quota == RUNTIME_INF || quota == -1) 7818 return RUNTIME_INF; 7819 7820 return to_ratio(period, quota); 7821} 7822 7823static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7824{ 7825 struct cfs_schedulable_data *d = data; 7826 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7827 s64 quota = 0, parent_quota = -1; 7828 7829 if (!tg->parent) { 7830 quota = RUNTIME_INF; 7831 } else { 7832 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7833 7834 quota = normalize_cfs_quota(tg, d); 7835 parent_quota = parent_b->hierarchal_quota; 7836 7837 /* 7838 * ensure max(child_quota) <= parent_quota, inherit when no 7839 * limit is set 7840 */ 7841 if (quota == RUNTIME_INF) 7842 quota = parent_quota; 7843 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7844 return -EINVAL; 7845 } 7846 cfs_b->hierarchal_quota = quota; 7847 7848 return 0; 7849} 7850 7851static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7852{ 7853 int ret; 7854 struct cfs_schedulable_data data = { 7855 .tg = tg, 7856 .period = period, 7857 .quota = quota, 7858 }; 7859 7860 if (quota != RUNTIME_INF) { 7861 do_div(data.period, NSEC_PER_USEC); 7862 do_div(data.quota, NSEC_PER_USEC); 7863 } 7864 7865 rcu_read_lock(); 7866 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7867 rcu_read_unlock(); 7868 7869 return ret; 7870} 7871 7872static int cpu_stats_show(struct seq_file *sf, void *v) 7873{ 7874 struct task_group *tg = css_tg(seq_css(sf)); 7875 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7876 7877 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7878 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7879 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7880 7881 return 0; 7882} 7883#endif /* CONFIG_CFS_BANDWIDTH */ 7884#endif /* CONFIG_FAIR_GROUP_SCHED */ 7885 7886#ifdef CONFIG_RT_GROUP_SCHED 7887static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7888 struct cftype *cft, s64 val) 7889{ 7890 return sched_group_set_rt_runtime(css_tg(css), val); 7891} 7892 7893static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7894 struct cftype *cft) 7895{ 7896 return sched_group_rt_runtime(css_tg(css)); 7897} 7898 7899static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7900 struct cftype *cftype, u64 rt_period_us) 7901{ 7902 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7903} 7904 7905static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7906 struct cftype *cft) 7907{ 7908 return sched_group_rt_period(css_tg(css)); 7909} 7910#endif /* CONFIG_RT_GROUP_SCHED */ 7911 7912static struct cftype cpu_files[] = { 7913#ifdef CONFIG_FAIR_GROUP_SCHED 7914 { 7915 .name = "shares", 7916 .read_u64 = cpu_shares_read_u64, 7917 .write_u64 = cpu_shares_write_u64, 7918 }, 7919#endif 7920#ifdef CONFIG_CFS_BANDWIDTH 7921 { 7922 .name = "cfs_quota_us", 7923 .read_s64 = cpu_cfs_quota_read_s64, 7924 .write_s64 = cpu_cfs_quota_write_s64, 7925 }, 7926 { 7927 .name = "cfs_period_us", 7928 .read_u64 = cpu_cfs_period_read_u64, 7929 .write_u64 = cpu_cfs_period_write_u64, 7930 }, 7931 { 7932 .name = "stat", 7933 .seq_show = cpu_stats_show, 7934 }, 7935#endif 7936#ifdef CONFIG_RT_GROUP_SCHED 7937 { 7938 .name = "rt_runtime_us", 7939 .read_s64 = cpu_rt_runtime_read, 7940 .write_s64 = cpu_rt_runtime_write, 7941 }, 7942 { 7943 .name = "rt_period_us", 7944 .read_u64 = cpu_rt_period_read_uint, 7945 .write_u64 = cpu_rt_period_write_uint, 7946 }, 7947#endif 7948 { } /* terminate */ 7949}; 7950 7951struct cgroup_subsys cpu_cgroup_subsys = { 7952 .name = "cpu", 7953 .css_alloc = cpu_cgroup_css_alloc, 7954 .css_free = cpu_cgroup_css_free, 7955 .css_online = cpu_cgroup_css_online, 7956 .css_offline = cpu_cgroup_css_offline, 7957 .can_attach = cpu_cgroup_can_attach, 7958 .attach = cpu_cgroup_attach, 7959 .exit = cpu_cgroup_exit, 7960 .subsys_id = cpu_cgroup_subsys_id, 7961 .base_cftypes = cpu_files, 7962 .early_init = 1, 7963}; 7964 7965#endif /* CONFIG_CGROUP_SCHED */ 7966 7967void dump_cpu_task(int cpu) 7968{ 7969 pr_info("Task dump for CPU %d:\n", cpu); 7970 sched_show_task(cpu_curr(cpu)); 7971} 7972