core.c revision 37e117c07b89194aae7062bc63bde1104c03db02
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75#include <linux/context_tracking.h> 76 77#include <asm/switch_to.h> 78#include <asm/tlb.h> 79#include <asm/irq_regs.h> 80#include <asm/mutex.h> 81#ifdef CONFIG_PARAVIRT 82#include <asm/paravirt.h> 83#endif 84 85#include "sched.h" 86#include "../workqueue_internal.h" 87#include "../smpboot.h" 88 89#define CREATE_TRACE_POINTS 90#include <trace/events/sched.h> 91 92void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 93{ 94 unsigned long delta; 95 ktime_t soft, hard, now; 96 97 for (;;) { 98 if (hrtimer_active(period_timer)) 99 break; 100 101 now = hrtimer_cb_get_time(period_timer); 102 hrtimer_forward(period_timer, now, period); 103 104 soft = hrtimer_get_softexpires(period_timer); 105 hard = hrtimer_get_expires(period_timer); 106 delta = ktime_to_ns(ktime_sub(hard, soft)); 107 __hrtimer_start_range_ns(period_timer, soft, delta, 108 HRTIMER_MODE_ABS_PINNED, 0); 109 } 110} 111 112DEFINE_MUTEX(sched_domains_mutex); 113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115static void update_rq_clock_task(struct rq *rq, s64 delta); 116 117void update_rq_clock(struct rq *rq) 118{ 119 s64 delta; 120 121 if (rq->skip_clock_update > 0) 122 return; 123 124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 125 rq->clock += delta; 126 update_rq_clock_task(rq, delta); 127} 128 129/* 130 * Debugging: various feature bits 131 */ 132 133#define SCHED_FEAT(name, enabled) \ 134 (1UL << __SCHED_FEAT_##name) * enabled | 135 136const_debug unsigned int sysctl_sched_features = 137#include "features.h" 138 0; 139 140#undef SCHED_FEAT 141 142#ifdef CONFIG_SCHED_DEBUG 143#define SCHED_FEAT(name, enabled) \ 144 #name , 145 146static const char * const sched_feat_names[] = { 147#include "features.h" 148}; 149 150#undef SCHED_FEAT 151 152static int sched_feat_show(struct seq_file *m, void *v) 153{ 154 int i; 155 156 for (i = 0; i < __SCHED_FEAT_NR; i++) { 157 if (!(sysctl_sched_features & (1UL << i))) 158 seq_puts(m, "NO_"); 159 seq_printf(m, "%s ", sched_feat_names[i]); 160 } 161 seq_puts(m, "\n"); 162 163 return 0; 164} 165 166#ifdef HAVE_JUMP_LABEL 167 168#define jump_label_key__true STATIC_KEY_INIT_TRUE 169#define jump_label_key__false STATIC_KEY_INIT_FALSE 170 171#define SCHED_FEAT(name, enabled) \ 172 jump_label_key__##enabled , 173 174struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 175#include "features.h" 176}; 177 178#undef SCHED_FEAT 179 180static void sched_feat_disable(int i) 181{ 182 if (static_key_enabled(&sched_feat_keys[i])) 183 static_key_slow_dec(&sched_feat_keys[i]); 184} 185 186static void sched_feat_enable(int i) 187{ 188 if (!static_key_enabled(&sched_feat_keys[i])) 189 static_key_slow_inc(&sched_feat_keys[i]); 190} 191#else 192static void sched_feat_disable(int i) { }; 193static void sched_feat_enable(int i) { }; 194#endif /* HAVE_JUMP_LABEL */ 195 196static int sched_feat_set(char *cmp) 197{ 198 int i; 199 int neg = 0; 200 201 if (strncmp(cmp, "NO_", 3) == 0) { 202 neg = 1; 203 cmp += 3; 204 } 205 206 for (i = 0; i < __SCHED_FEAT_NR; i++) { 207 if (strcmp(cmp, sched_feat_names[i]) == 0) { 208 if (neg) { 209 sysctl_sched_features &= ~(1UL << i); 210 sched_feat_disable(i); 211 } else { 212 sysctl_sched_features |= (1UL << i); 213 sched_feat_enable(i); 214 } 215 break; 216 } 217 } 218 219 return i; 220} 221 222static ssize_t 223sched_feat_write(struct file *filp, const char __user *ubuf, 224 size_t cnt, loff_t *ppos) 225{ 226 char buf[64]; 227 char *cmp; 228 int i; 229 230 if (cnt > 63) 231 cnt = 63; 232 233 if (copy_from_user(&buf, ubuf, cnt)) 234 return -EFAULT; 235 236 buf[cnt] = 0; 237 cmp = strstrip(buf); 238 239 i = sched_feat_set(cmp); 240 if (i == __SCHED_FEAT_NR) 241 return -EINVAL; 242 243 *ppos += cnt; 244 245 return cnt; 246} 247 248static int sched_feat_open(struct inode *inode, struct file *filp) 249{ 250 return single_open(filp, sched_feat_show, NULL); 251} 252 253static const struct file_operations sched_feat_fops = { 254 .open = sched_feat_open, 255 .write = sched_feat_write, 256 .read = seq_read, 257 .llseek = seq_lseek, 258 .release = single_release, 259}; 260 261static __init int sched_init_debug(void) 262{ 263 debugfs_create_file("sched_features", 0644, NULL, NULL, 264 &sched_feat_fops); 265 266 return 0; 267} 268late_initcall(sched_init_debug); 269#endif /* CONFIG_SCHED_DEBUG */ 270 271/* 272 * Number of tasks to iterate in a single balance run. 273 * Limited because this is done with IRQs disabled. 274 */ 275const_debug unsigned int sysctl_sched_nr_migrate = 32; 276 277/* 278 * period over which we average the RT time consumption, measured 279 * in ms. 280 * 281 * default: 1s 282 */ 283const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 284 285/* 286 * period over which we measure -rt task cpu usage in us. 287 * default: 1s 288 */ 289unsigned int sysctl_sched_rt_period = 1000000; 290 291__read_mostly int scheduler_running; 292 293/* 294 * part of the period that we allow rt tasks to run in us. 295 * default: 0.95s 296 */ 297int sysctl_sched_rt_runtime = 950000; 298 299/* 300 * __task_rq_lock - lock the rq @p resides on. 301 */ 302static inline struct rq *__task_rq_lock(struct task_struct *p) 303 __acquires(rq->lock) 304{ 305 struct rq *rq; 306 307 lockdep_assert_held(&p->pi_lock); 308 309 for (;;) { 310 rq = task_rq(p); 311 raw_spin_lock(&rq->lock); 312 if (likely(rq == task_rq(p))) 313 return rq; 314 raw_spin_unlock(&rq->lock); 315 } 316} 317 318/* 319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 320 */ 321static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 322 __acquires(p->pi_lock) 323 __acquires(rq->lock) 324{ 325 struct rq *rq; 326 327 for (;;) { 328 raw_spin_lock_irqsave(&p->pi_lock, *flags); 329 rq = task_rq(p); 330 raw_spin_lock(&rq->lock); 331 if (likely(rq == task_rq(p))) 332 return rq; 333 raw_spin_unlock(&rq->lock); 334 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 335 } 336} 337 338static void __task_rq_unlock(struct rq *rq) 339 __releases(rq->lock) 340{ 341 raw_spin_unlock(&rq->lock); 342} 343 344static inline void 345task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 346 __releases(rq->lock) 347 __releases(p->pi_lock) 348{ 349 raw_spin_unlock(&rq->lock); 350 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 351} 352 353/* 354 * this_rq_lock - lock this runqueue and disable interrupts. 355 */ 356static struct rq *this_rq_lock(void) 357 __acquires(rq->lock) 358{ 359 struct rq *rq; 360 361 local_irq_disable(); 362 rq = this_rq(); 363 raw_spin_lock(&rq->lock); 364 365 return rq; 366} 367 368#ifdef CONFIG_SCHED_HRTICK 369/* 370 * Use HR-timers to deliver accurate preemption points. 371 */ 372 373static void hrtick_clear(struct rq *rq) 374{ 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377} 378 379/* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383static enum hrtimer_restart hrtick(struct hrtimer *timer) 384{ 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395} 396 397#ifdef CONFIG_SMP 398 399static int __hrtick_restart(struct rq *rq) 400{ 401 struct hrtimer *timer = &rq->hrtick_timer; 402 ktime_t time = hrtimer_get_softexpires(timer); 403 404 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 405} 406 407/* 408 * called from hardirq (IPI) context 409 */ 410static void __hrtick_start(void *arg) 411{ 412 struct rq *rq = arg; 413 414 raw_spin_lock(&rq->lock); 415 __hrtick_restart(rq); 416 rq->hrtick_csd_pending = 0; 417 raw_spin_unlock(&rq->lock); 418} 419 420/* 421 * Called to set the hrtick timer state. 422 * 423 * called with rq->lock held and irqs disabled 424 */ 425void hrtick_start(struct rq *rq, u64 delay) 426{ 427 struct hrtimer *timer = &rq->hrtick_timer; 428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 429 430 hrtimer_set_expires(timer, time); 431 432 if (rq == this_rq()) { 433 __hrtick_restart(rq); 434 } else if (!rq->hrtick_csd_pending) { 435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 436 rq->hrtick_csd_pending = 1; 437 } 438} 439 440static int 441hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 442{ 443 int cpu = (int)(long)hcpu; 444 445 switch (action) { 446 case CPU_UP_CANCELED: 447 case CPU_UP_CANCELED_FROZEN: 448 case CPU_DOWN_PREPARE: 449 case CPU_DOWN_PREPARE_FROZEN: 450 case CPU_DEAD: 451 case CPU_DEAD_FROZEN: 452 hrtick_clear(cpu_rq(cpu)); 453 return NOTIFY_OK; 454 } 455 456 return NOTIFY_DONE; 457} 458 459static __init void init_hrtick(void) 460{ 461 hotcpu_notifier(hotplug_hrtick, 0); 462} 463#else 464/* 465 * Called to set the hrtick timer state. 466 * 467 * called with rq->lock held and irqs disabled 468 */ 469void hrtick_start(struct rq *rq, u64 delay) 470{ 471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 472 HRTIMER_MODE_REL_PINNED, 0); 473} 474 475static inline void init_hrtick(void) 476{ 477} 478#endif /* CONFIG_SMP */ 479 480static void init_rq_hrtick(struct rq *rq) 481{ 482#ifdef CONFIG_SMP 483 rq->hrtick_csd_pending = 0; 484 485 rq->hrtick_csd.flags = 0; 486 rq->hrtick_csd.func = __hrtick_start; 487 rq->hrtick_csd.info = rq; 488#endif 489 490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 491 rq->hrtick_timer.function = hrtick; 492} 493#else /* CONFIG_SCHED_HRTICK */ 494static inline void hrtick_clear(struct rq *rq) 495{ 496} 497 498static inline void init_rq_hrtick(struct rq *rq) 499{ 500} 501 502static inline void init_hrtick(void) 503{ 504} 505#endif /* CONFIG_SCHED_HRTICK */ 506 507/* 508 * resched_task - mark a task 'to be rescheduled now'. 509 * 510 * On UP this means the setting of the need_resched flag, on SMP it 511 * might also involve a cross-CPU call to trigger the scheduler on 512 * the target CPU. 513 */ 514void resched_task(struct task_struct *p) 515{ 516 int cpu; 517 518 lockdep_assert_held(&task_rq(p)->lock); 519 520 if (test_tsk_need_resched(p)) 521 return; 522 523 set_tsk_need_resched(p); 524 525 cpu = task_cpu(p); 526 if (cpu == smp_processor_id()) { 527 set_preempt_need_resched(); 528 return; 529 } 530 531 /* NEED_RESCHED must be visible before we test polling */ 532 smp_mb(); 533 if (!tsk_is_polling(p)) 534 smp_send_reschedule(cpu); 535} 536 537void resched_cpu(int cpu) 538{ 539 struct rq *rq = cpu_rq(cpu); 540 unsigned long flags; 541 542 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 543 return; 544 resched_task(cpu_curr(cpu)); 545 raw_spin_unlock_irqrestore(&rq->lock, flags); 546} 547 548#ifdef CONFIG_SMP 549#ifdef CONFIG_NO_HZ_COMMON 550/* 551 * In the semi idle case, use the nearest busy cpu for migrating timers 552 * from an idle cpu. This is good for power-savings. 553 * 554 * We don't do similar optimization for completely idle system, as 555 * selecting an idle cpu will add more delays to the timers than intended 556 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 557 */ 558int get_nohz_timer_target(void) 559{ 560 int cpu = smp_processor_id(); 561 int i; 562 struct sched_domain *sd; 563 564 rcu_read_lock(); 565 for_each_domain(cpu, sd) { 566 for_each_cpu(i, sched_domain_span(sd)) { 567 if (!idle_cpu(i)) { 568 cpu = i; 569 goto unlock; 570 } 571 } 572 } 573unlock: 574 rcu_read_unlock(); 575 return cpu; 576} 577/* 578 * When add_timer_on() enqueues a timer into the timer wheel of an 579 * idle CPU then this timer might expire before the next timer event 580 * which is scheduled to wake up that CPU. In case of a completely 581 * idle system the next event might even be infinite time into the 582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 583 * leaves the inner idle loop so the newly added timer is taken into 584 * account when the CPU goes back to idle and evaluates the timer 585 * wheel for the next timer event. 586 */ 587static void wake_up_idle_cpu(int cpu) 588{ 589 struct rq *rq = cpu_rq(cpu); 590 591 if (cpu == smp_processor_id()) 592 return; 593 594 /* 595 * This is safe, as this function is called with the timer 596 * wheel base lock of (cpu) held. When the CPU is on the way 597 * to idle and has not yet set rq->curr to idle then it will 598 * be serialized on the timer wheel base lock and take the new 599 * timer into account automatically. 600 */ 601 if (rq->curr != rq->idle) 602 return; 603 604 /* 605 * We can set TIF_RESCHED on the idle task of the other CPU 606 * lockless. The worst case is that the other CPU runs the 607 * idle task through an additional NOOP schedule() 608 */ 609 set_tsk_need_resched(rq->idle); 610 611 /* NEED_RESCHED must be visible before we test polling */ 612 smp_mb(); 613 if (!tsk_is_polling(rq->idle)) 614 smp_send_reschedule(cpu); 615} 616 617static bool wake_up_full_nohz_cpu(int cpu) 618{ 619 if (tick_nohz_full_cpu(cpu)) { 620 if (cpu != smp_processor_id() || 621 tick_nohz_tick_stopped()) 622 smp_send_reschedule(cpu); 623 return true; 624 } 625 626 return false; 627} 628 629void wake_up_nohz_cpu(int cpu) 630{ 631 if (!wake_up_full_nohz_cpu(cpu)) 632 wake_up_idle_cpu(cpu); 633} 634 635static inline bool got_nohz_idle_kick(void) 636{ 637 int cpu = smp_processor_id(); 638 639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 640 return false; 641 642 if (idle_cpu(cpu) && !need_resched()) 643 return true; 644 645 /* 646 * We can't run Idle Load Balance on this CPU for this time so we 647 * cancel it and clear NOHZ_BALANCE_KICK 648 */ 649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 650 return false; 651} 652 653#else /* CONFIG_NO_HZ_COMMON */ 654 655static inline bool got_nohz_idle_kick(void) 656{ 657 return false; 658} 659 660#endif /* CONFIG_NO_HZ_COMMON */ 661 662#ifdef CONFIG_NO_HZ_FULL 663bool sched_can_stop_tick(void) 664{ 665 struct rq *rq; 666 667 rq = this_rq(); 668 669 /* Make sure rq->nr_running update is visible after the IPI */ 670 smp_rmb(); 671 672 /* More than one running task need preemption */ 673 if (rq->nr_running > 1) 674 return false; 675 676 return true; 677} 678#endif /* CONFIG_NO_HZ_FULL */ 679 680void sched_avg_update(struct rq *rq) 681{ 682 s64 period = sched_avg_period(); 683 684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 685 /* 686 * Inline assembly required to prevent the compiler 687 * optimising this loop into a divmod call. 688 * See __iter_div_u64_rem() for another example of this. 689 */ 690 asm("" : "+rm" (rq->age_stamp)); 691 rq->age_stamp += period; 692 rq->rt_avg /= 2; 693 } 694} 695 696#endif /* CONFIG_SMP */ 697 698#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 699 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 700/* 701 * Iterate task_group tree rooted at *from, calling @down when first entering a 702 * node and @up when leaving it for the final time. 703 * 704 * Caller must hold rcu_lock or sufficient equivalent. 705 */ 706int walk_tg_tree_from(struct task_group *from, 707 tg_visitor down, tg_visitor up, void *data) 708{ 709 struct task_group *parent, *child; 710 int ret; 711 712 parent = from; 713 714down: 715 ret = (*down)(parent, data); 716 if (ret) 717 goto out; 718 list_for_each_entry_rcu(child, &parent->children, siblings) { 719 parent = child; 720 goto down; 721 722up: 723 continue; 724 } 725 ret = (*up)(parent, data); 726 if (ret || parent == from) 727 goto out; 728 729 child = parent; 730 parent = parent->parent; 731 if (parent) 732 goto up; 733out: 734 return ret; 735} 736 737int tg_nop(struct task_group *tg, void *data) 738{ 739 return 0; 740} 741#endif 742 743static void set_load_weight(struct task_struct *p) 744{ 745 int prio = p->static_prio - MAX_RT_PRIO; 746 struct load_weight *load = &p->se.load; 747 748 /* 749 * SCHED_IDLE tasks get minimal weight: 750 */ 751 if (p->policy == SCHED_IDLE) { 752 load->weight = scale_load(WEIGHT_IDLEPRIO); 753 load->inv_weight = WMULT_IDLEPRIO; 754 return; 755 } 756 757 load->weight = scale_load(prio_to_weight[prio]); 758 load->inv_weight = prio_to_wmult[prio]; 759} 760 761static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 762{ 763 update_rq_clock(rq); 764 sched_info_queued(rq, p); 765 p->sched_class->enqueue_task(rq, p, flags); 766} 767 768static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 769{ 770 update_rq_clock(rq); 771 sched_info_dequeued(rq, p); 772 p->sched_class->dequeue_task(rq, p, flags); 773} 774 775void activate_task(struct rq *rq, struct task_struct *p, int flags) 776{ 777 if (task_contributes_to_load(p)) 778 rq->nr_uninterruptible--; 779 780 enqueue_task(rq, p, flags); 781} 782 783void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 784{ 785 if (task_contributes_to_load(p)) 786 rq->nr_uninterruptible++; 787 788 dequeue_task(rq, p, flags); 789} 790 791static void update_rq_clock_task(struct rq *rq, s64 delta) 792{ 793/* 794 * In theory, the compile should just see 0 here, and optimize out the call 795 * to sched_rt_avg_update. But I don't trust it... 796 */ 797#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 798 s64 steal = 0, irq_delta = 0; 799#endif 800#ifdef CONFIG_IRQ_TIME_ACCOUNTING 801 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 802 803 /* 804 * Since irq_time is only updated on {soft,}irq_exit, we might run into 805 * this case when a previous update_rq_clock() happened inside a 806 * {soft,}irq region. 807 * 808 * When this happens, we stop ->clock_task and only update the 809 * prev_irq_time stamp to account for the part that fit, so that a next 810 * update will consume the rest. This ensures ->clock_task is 811 * monotonic. 812 * 813 * It does however cause some slight miss-attribution of {soft,}irq 814 * time, a more accurate solution would be to update the irq_time using 815 * the current rq->clock timestamp, except that would require using 816 * atomic ops. 817 */ 818 if (irq_delta > delta) 819 irq_delta = delta; 820 821 rq->prev_irq_time += irq_delta; 822 delta -= irq_delta; 823#endif 824#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 825 if (static_key_false((¶virt_steal_rq_enabled))) { 826 u64 st; 827 828 steal = paravirt_steal_clock(cpu_of(rq)); 829 steal -= rq->prev_steal_time_rq; 830 831 if (unlikely(steal > delta)) 832 steal = delta; 833 834 st = steal_ticks(steal); 835 steal = st * TICK_NSEC; 836 837 rq->prev_steal_time_rq += steal; 838 839 delta -= steal; 840 } 841#endif 842 843 rq->clock_task += delta; 844 845#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 846 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 847 sched_rt_avg_update(rq, irq_delta + steal); 848#endif 849} 850 851void sched_set_stop_task(int cpu, struct task_struct *stop) 852{ 853 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 854 struct task_struct *old_stop = cpu_rq(cpu)->stop; 855 856 if (stop) { 857 /* 858 * Make it appear like a SCHED_FIFO task, its something 859 * userspace knows about and won't get confused about. 860 * 861 * Also, it will make PI more or less work without too 862 * much confusion -- but then, stop work should not 863 * rely on PI working anyway. 864 */ 865 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 866 867 stop->sched_class = &stop_sched_class; 868 } 869 870 cpu_rq(cpu)->stop = stop; 871 872 if (old_stop) { 873 /* 874 * Reset it back to a normal scheduling class so that 875 * it can die in pieces. 876 */ 877 old_stop->sched_class = &rt_sched_class; 878 } 879} 880 881/* 882 * __normal_prio - return the priority that is based on the static prio 883 */ 884static inline int __normal_prio(struct task_struct *p) 885{ 886 return p->static_prio; 887} 888 889/* 890 * Calculate the expected normal priority: i.e. priority 891 * without taking RT-inheritance into account. Might be 892 * boosted by interactivity modifiers. Changes upon fork, 893 * setprio syscalls, and whenever the interactivity 894 * estimator recalculates. 895 */ 896static inline int normal_prio(struct task_struct *p) 897{ 898 int prio; 899 900 if (task_has_dl_policy(p)) 901 prio = MAX_DL_PRIO-1; 902 else if (task_has_rt_policy(p)) 903 prio = MAX_RT_PRIO-1 - p->rt_priority; 904 else 905 prio = __normal_prio(p); 906 return prio; 907} 908 909/* 910 * Calculate the current priority, i.e. the priority 911 * taken into account by the scheduler. This value might 912 * be boosted by RT tasks, or might be boosted by 913 * interactivity modifiers. Will be RT if the task got 914 * RT-boosted. If not then it returns p->normal_prio. 915 */ 916static int effective_prio(struct task_struct *p) 917{ 918 p->normal_prio = normal_prio(p); 919 /* 920 * If we are RT tasks or we were boosted to RT priority, 921 * keep the priority unchanged. Otherwise, update priority 922 * to the normal priority: 923 */ 924 if (!rt_prio(p->prio)) 925 return p->normal_prio; 926 return p->prio; 927} 928 929/** 930 * task_curr - is this task currently executing on a CPU? 931 * @p: the task in question. 932 * 933 * Return: 1 if the task is currently executing. 0 otherwise. 934 */ 935inline int task_curr(const struct task_struct *p) 936{ 937 return cpu_curr(task_cpu(p)) == p; 938} 939 940static inline void check_class_changed(struct rq *rq, struct task_struct *p, 941 const struct sched_class *prev_class, 942 int oldprio) 943{ 944 if (prev_class != p->sched_class) { 945 if (prev_class->switched_from) 946 prev_class->switched_from(rq, p); 947 p->sched_class->switched_to(rq, p); 948 } else if (oldprio != p->prio || dl_task(p)) 949 p->sched_class->prio_changed(rq, p, oldprio); 950} 951 952void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 953{ 954 const struct sched_class *class; 955 956 if (p->sched_class == rq->curr->sched_class) { 957 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 958 } else { 959 for_each_class(class) { 960 if (class == rq->curr->sched_class) 961 break; 962 if (class == p->sched_class) { 963 resched_task(rq->curr); 964 break; 965 } 966 } 967 } 968 969 /* 970 * A queue event has occurred, and we're going to schedule. In 971 * this case, we can save a useless back to back clock update. 972 */ 973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 974 rq->skip_clock_update = 1; 975} 976 977#ifdef CONFIG_SMP 978void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 979{ 980#ifdef CONFIG_SCHED_DEBUG 981 /* 982 * We should never call set_task_cpu() on a blocked task, 983 * ttwu() will sort out the placement. 984 */ 985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 986 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 987 988#ifdef CONFIG_LOCKDEP 989 /* 990 * The caller should hold either p->pi_lock or rq->lock, when changing 991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 992 * 993 * sched_move_task() holds both and thus holding either pins the cgroup, 994 * see task_group(). 995 * 996 * Furthermore, all task_rq users should acquire both locks, see 997 * task_rq_lock(). 998 */ 999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1000 lockdep_is_held(&task_rq(p)->lock))); 1001#endif 1002#endif 1003 1004 trace_sched_migrate_task(p, new_cpu); 1005 1006 if (task_cpu(p) != new_cpu) { 1007 if (p->sched_class->migrate_task_rq) 1008 p->sched_class->migrate_task_rq(p, new_cpu); 1009 p->se.nr_migrations++; 1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1011 } 1012 1013 __set_task_cpu(p, new_cpu); 1014} 1015 1016static void __migrate_swap_task(struct task_struct *p, int cpu) 1017{ 1018 if (p->on_rq) { 1019 struct rq *src_rq, *dst_rq; 1020 1021 src_rq = task_rq(p); 1022 dst_rq = cpu_rq(cpu); 1023 1024 deactivate_task(src_rq, p, 0); 1025 set_task_cpu(p, cpu); 1026 activate_task(dst_rq, p, 0); 1027 check_preempt_curr(dst_rq, p, 0); 1028 } else { 1029 /* 1030 * Task isn't running anymore; make it appear like we migrated 1031 * it before it went to sleep. This means on wakeup we make the 1032 * previous cpu our targer instead of where it really is. 1033 */ 1034 p->wake_cpu = cpu; 1035 } 1036} 1037 1038struct migration_swap_arg { 1039 struct task_struct *src_task, *dst_task; 1040 int src_cpu, dst_cpu; 1041}; 1042 1043static int migrate_swap_stop(void *data) 1044{ 1045 struct migration_swap_arg *arg = data; 1046 struct rq *src_rq, *dst_rq; 1047 int ret = -EAGAIN; 1048 1049 src_rq = cpu_rq(arg->src_cpu); 1050 dst_rq = cpu_rq(arg->dst_cpu); 1051 1052 double_raw_lock(&arg->src_task->pi_lock, 1053 &arg->dst_task->pi_lock); 1054 double_rq_lock(src_rq, dst_rq); 1055 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1056 goto unlock; 1057 1058 if (task_cpu(arg->src_task) != arg->src_cpu) 1059 goto unlock; 1060 1061 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1062 goto unlock; 1063 1064 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1065 goto unlock; 1066 1067 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1068 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1069 1070 ret = 0; 1071 1072unlock: 1073 double_rq_unlock(src_rq, dst_rq); 1074 raw_spin_unlock(&arg->dst_task->pi_lock); 1075 raw_spin_unlock(&arg->src_task->pi_lock); 1076 1077 return ret; 1078} 1079 1080/* 1081 * Cross migrate two tasks 1082 */ 1083int migrate_swap(struct task_struct *cur, struct task_struct *p) 1084{ 1085 struct migration_swap_arg arg; 1086 int ret = -EINVAL; 1087 1088 arg = (struct migration_swap_arg){ 1089 .src_task = cur, 1090 .src_cpu = task_cpu(cur), 1091 .dst_task = p, 1092 .dst_cpu = task_cpu(p), 1093 }; 1094 1095 if (arg.src_cpu == arg.dst_cpu) 1096 goto out; 1097 1098 /* 1099 * These three tests are all lockless; this is OK since all of them 1100 * will be re-checked with proper locks held further down the line. 1101 */ 1102 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1103 goto out; 1104 1105 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1106 goto out; 1107 1108 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1109 goto out; 1110 1111 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1112 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1113 1114out: 1115 return ret; 1116} 1117 1118struct migration_arg { 1119 struct task_struct *task; 1120 int dest_cpu; 1121}; 1122 1123static int migration_cpu_stop(void *data); 1124 1125/* 1126 * wait_task_inactive - wait for a thread to unschedule. 1127 * 1128 * If @match_state is nonzero, it's the @p->state value just checked and 1129 * not expected to change. If it changes, i.e. @p might have woken up, 1130 * then return zero. When we succeed in waiting for @p to be off its CPU, 1131 * we return a positive number (its total switch count). If a second call 1132 * a short while later returns the same number, the caller can be sure that 1133 * @p has remained unscheduled the whole time. 1134 * 1135 * The caller must ensure that the task *will* unschedule sometime soon, 1136 * else this function might spin for a *long* time. This function can't 1137 * be called with interrupts off, or it may introduce deadlock with 1138 * smp_call_function() if an IPI is sent by the same process we are 1139 * waiting to become inactive. 1140 */ 1141unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1142{ 1143 unsigned long flags; 1144 int running, on_rq; 1145 unsigned long ncsw; 1146 struct rq *rq; 1147 1148 for (;;) { 1149 /* 1150 * We do the initial early heuristics without holding 1151 * any task-queue locks at all. We'll only try to get 1152 * the runqueue lock when things look like they will 1153 * work out! 1154 */ 1155 rq = task_rq(p); 1156 1157 /* 1158 * If the task is actively running on another CPU 1159 * still, just relax and busy-wait without holding 1160 * any locks. 1161 * 1162 * NOTE! Since we don't hold any locks, it's not 1163 * even sure that "rq" stays as the right runqueue! 1164 * But we don't care, since "task_running()" will 1165 * return false if the runqueue has changed and p 1166 * is actually now running somewhere else! 1167 */ 1168 while (task_running(rq, p)) { 1169 if (match_state && unlikely(p->state != match_state)) 1170 return 0; 1171 cpu_relax(); 1172 } 1173 1174 /* 1175 * Ok, time to look more closely! We need the rq 1176 * lock now, to be *sure*. If we're wrong, we'll 1177 * just go back and repeat. 1178 */ 1179 rq = task_rq_lock(p, &flags); 1180 trace_sched_wait_task(p); 1181 running = task_running(rq, p); 1182 on_rq = p->on_rq; 1183 ncsw = 0; 1184 if (!match_state || p->state == match_state) 1185 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1186 task_rq_unlock(rq, p, &flags); 1187 1188 /* 1189 * If it changed from the expected state, bail out now. 1190 */ 1191 if (unlikely(!ncsw)) 1192 break; 1193 1194 /* 1195 * Was it really running after all now that we 1196 * checked with the proper locks actually held? 1197 * 1198 * Oops. Go back and try again.. 1199 */ 1200 if (unlikely(running)) { 1201 cpu_relax(); 1202 continue; 1203 } 1204 1205 /* 1206 * It's not enough that it's not actively running, 1207 * it must be off the runqueue _entirely_, and not 1208 * preempted! 1209 * 1210 * So if it was still runnable (but just not actively 1211 * running right now), it's preempted, and we should 1212 * yield - it could be a while. 1213 */ 1214 if (unlikely(on_rq)) { 1215 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1216 1217 set_current_state(TASK_UNINTERRUPTIBLE); 1218 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1219 continue; 1220 } 1221 1222 /* 1223 * Ahh, all good. It wasn't running, and it wasn't 1224 * runnable, which means that it will never become 1225 * running in the future either. We're all done! 1226 */ 1227 break; 1228 } 1229 1230 return ncsw; 1231} 1232 1233/*** 1234 * kick_process - kick a running thread to enter/exit the kernel 1235 * @p: the to-be-kicked thread 1236 * 1237 * Cause a process which is running on another CPU to enter 1238 * kernel-mode, without any delay. (to get signals handled.) 1239 * 1240 * NOTE: this function doesn't have to take the runqueue lock, 1241 * because all it wants to ensure is that the remote task enters 1242 * the kernel. If the IPI races and the task has been migrated 1243 * to another CPU then no harm is done and the purpose has been 1244 * achieved as well. 1245 */ 1246void kick_process(struct task_struct *p) 1247{ 1248 int cpu; 1249 1250 preempt_disable(); 1251 cpu = task_cpu(p); 1252 if ((cpu != smp_processor_id()) && task_curr(p)) 1253 smp_send_reschedule(cpu); 1254 preempt_enable(); 1255} 1256EXPORT_SYMBOL_GPL(kick_process); 1257#endif /* CONFIG_SMP */ 1258 1259#ifdef CONFIG_SMP 1260/* 1261 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1262 */ 1263static int select_fallback_rq(int cpu, struct task_struct *p) 1264{ 1265 int nid = cpu_to_node(cpu); 1266 const struct cpumask *nodemask = NULL; 1267 enum { cpuset, possible, fail } state = cpuset; 1268 int dest_cpu; 1269 1270 /* 1271 * If the node that the cpu is on has been offlined, cpu_to_node() 1272 * will return -1. There is no cpu on the node, and we should 1273 * select the cpu on the other node. 1274 */ 1275 if (nid != -1) { 1276 nodemask = cpumask_of_node(nid); 1277 1278 /* Look for allowed, online CPU in same node. */ 1279 for_each_cpu(dest_cpu, nodemask) { 1280 if (!cpu_online(dest_cpu)) 1281 continue; 1282 if (!cpu_active(dest_cpu)) 1283 continue; 1284 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1285 return dest_cpu; 1286 } 1287 } 1288 1289 for (;;) { 1290 /* Any allowed, online CPU? */ 1291 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1292 if (!cpu_online(dest_cpu)) 1293 continue; 1294 if (!cpu_active(dest_cpu)) 1295 continue; 1296 goto out; 1297 } 1298 1299 switch (state) { 1300 case cpuset: 1301 /* No more Mr. Nice Guy. */ 1302 cpuset_cpus_allowed_fallback(p); 1303 state = possible; 1304 break; 1305 1306 case possible: 1307 do_set_cpus_allowed(p, cpu_possible_mask); 1308 state = fail; 1309 break; 1310 1311 case fail: 1312 BUG(); 1313 break; 1314 } 1315 } 1316 1317out: 1318 if (state != cpuset) { 1319 /* 1320 * Don't tell them about moving exiting tasks or 1321 * kernel threads (both mm NULL), since they never 1322 * leave kernel. 1323 */ 1324 if (p->mm && printk_ratelimit()) { 1325 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1326 task_pid_nr(p), p->comm, cpu); 1327 } 1328 } 1329 1330 return dest_cpu; 1331} 1332 1333/* 1334 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1335 */ 1336static inline 1337int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1338{ 1339 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1340 1341 /* 1342 * In order not to call set_task_cpu() on a blocking task we need 1343 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1344 * cpu. 1345 * 1346 * Since this is common to all placement strategies, this lives here. 1347 * 1348 * [ this allows ->select_task() to simply return task_cpu(p) and 1349 * not worry about this generic constraint ] 1350 */ 1351 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1352 !cpu_online(cpu))) 1353 cpu = select_fallback_rq(task_cpu(p), p); 1354 1355 return cpu; 1356} 1357 1358static void update_avg(u64 *avg, u64 sample) 1359{ 1360 s64 diff = sample - *avg; 1361 *avg += diff >> 3; 1362} 1363#endif 1364 1365static void 1366ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1367{ 1368#ifdef CONFIG_SCHEDSTATS 1369 struct rq *rq = this_rq(); 1370 1371#ifdef CONFIG_SMP 1372 int this_cpu = smp_processor_id(); 1373 1374 if (cpu == this_cpu) { 1375 schedstat_inc(rq, ttwu_local); 1376 schedstat_inc(p, se.statistics.nr_wakeups_local); 1377 } else { 1378 struct sched_domain *sd; 1379 1380 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1381 rcu_read_lock(); 1382 for_each_domain(this_cpu, sd) { 1383 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1384 schedstat_inc(sd, ttwu_wake_remote); 1385 break; 1386 } 1387 } 1388 rcu_read_unlock(); 1389 } 1390 1391 if (wake_flags & WF_MIGRATED) 1392 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1393 1394#endif /* CONFIG_SMP */ 1395 1396 schedstat_inc(rq, ttwu_count); 1397 schedstat_inc(p, se.statistics.nr_wakeups); 1398 1399 if (wake_flags & WF_SYNC) 1400 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1401 1402#endif /* CONFIG_SCHEDSTATS */ 1403} 1404 1405static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1406{ 1407 activate_task(rq, p, en_flags); 1408 p->on_rq = 1; 1409 1410 /* if a worker is waking up, notify workqueue */ 1411 if (p->flags & PF_WQ_WORKER) 1412 wq_worker_waking_up(p, cpu_of(rq)); 1413} 1414 1415/* 1416 * Mark the task runnable and perform wakeup-preemption. 1417 */ 1418static void 1419ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1420{ 1421 check_preempt_curr(rq, p, wake_flags); 1422 trace_sched_wakeup(p, true); 1423 1424 p->state = TASK_RUNNING; 1425#ifdef CONFIG_SMP 1426 if (p->sched_class->task_woken) 1427 p->sched_class->task_woken(rq, p); 1428 1429 if (rq->idle_stamp) { 1430 u64 delta = rq_clock(rq) - rq->idle_stamp; 1431 u64 max = 2*rq->max_idle_balance_cost; 1432 1433 update_avg(&rq->avg_idle, delta); 1434 1435 if (rq->avg_idle > max) 1436 rq->avg_idle = max; 1437 1438 rq->idle_stamp = 0; 1439 } 1440#endif 1441} 1442 1443static void 1444ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1445{ 1446#ifdef CONFIG_SMP 1447 if (p->sched_contributes_to_load) 1448 rq->nr_uninterruptible--; 1449#endif 1450 1451 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1452 ttwu_do_wakeup(rq, p, wake_flags); 1453} 1454 1455/* 1456 * Called in case the task @p isn't fully descheduled from its runqueue, 1457 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1458 * since all we need to do is flip p->state to TASK_RUNNING, since 1459 * the task is still ->on_rq. 1460 */ 1461static int ttwu_remote(struct task_struct *p, int wake_flags) 1462{ 1463 struct rq *rq; 1464 int ret = 0; 1465 1466 rq = __task_rq_lock(p); 1467 if (p->on_rq) { 1468 /* check_preempt_curr() may use rq clock */ 1469 update_rq_clock(rq); 1470 ttwu_do_wakeup(rq, p, wake_flags); 1471 ret = 1; 1472 } 1473 __task_rq_unlock(rq); 1474 1475 return ret; 1476} 1477 1478#ifdef CONFIG_SMP 1479static void sched_ttwu_pending(void) 1480{ 1481 struct rq *rq = this_rq(); 1482 struct llist_node *llist = llist_del_all(&rq->wake_list); 1483 struct task_struct *p; 1484 1485 raw_spin_lock(&rq->lock); 1486 1487 while (llist) { 1488 p = llist_entry(llist, struct task_struct, wake_entry); 1489 llist = llist_next(llist); 1490 ttwu_do_activate(rq, p, 0); 1491 } 1492 1493 raw_spin_unlock(&rq->lock); 1494} 1495 1496void scheduler_ipi(void) 1497{ 1498 /* 1499 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1500 * TIF_NEED_RESCHED remotely (for the first time) will also send 1501 * this IPI. 1502 */ 1503 preempt_fold_need_resched(); 1504 1505 if (llist_empty(&this_rq()->wake_list) 1506 && !tick_nohz_full_cpu(smp_processor_id()) 1507 && !got_nohz_idle_kick()) 1508 return; 1509 1510 /* 1511 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1512 * traditionally all their work was done from the interrupt return 1513 * path. Now that we actually do some work, we need to make sure 1514 * we do call them. 1515 * 1516 * Some archs already do call them, luckily irq_enter/exit nest 1517 * properly. 1518 * 1519 * Arguably we should visit all archs and update all handlers, 1520 * however a fair share of IPIs are still resched only so this would 1521 * somewhat pessimize the simple resched case. 1522 */ 1523 irq_enter(); 1524 tick_nohz_full_check(); 1525 sched_ttwu_pending(); 1526 1527 /* 1528 * Check if someone kicked us for doing the nohz idle load balance. 1529 */ 1530 if (unlikely(got_nohz_idle_kick())) { 1531 this_rq()->idle_balance = 1; 1532 raise_softirq_irqoff(SCHED_SOFTIRQ); 1533 } 1534 irq_exit(); 1535} 1536 1537static void ttwu_queue_remote(struct task_struct *p, int cpu) 1538{ 1539 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1540 smp_send_reschedule(cpu); 1541} 1542 1543bool cpus_share_cache(int this_cpu, int that_cpu) 1544{ 1545 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1546} 1547#endif /* CONFIG_SMP */ 1548 1549static void ttwu_queue(struct task_struct *p, int cpu) 1550{ 1551 struct rq *rq = cpu_rq(cpu); 1552 1553#if defined(CONFIG_SMP) 1554 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1555 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1556 ttwu_queue_remote(p, cpu); 1557 return; 1558 } 1559#endif 1560 1561 raw_spin_lock(&rq->lock); 1562 ttwu_do_activate(rq, p, 0); 1563 raw_spin_unlock(&rq->lock); 1564} 1565 1566/** 1567 * try_to_wake_up - wake up a thread 1568 * @p: the thread to be awakened 1569 * @state: the mask of task states that can be woken 1570 * @wake_flags: wake modifier flags (WF_*) 1571 * 1572 * Put it on the run-queue if it's not already there. The "current" 1573 * thread is always on the run-queue (except when the actual 1574 * re-schedule is in progress), and as such you're allowed to do 1575 * the simpler "current->state = TASK_RUNNING" to mark yourself 1576 * runnable without the overhead of this. 1577 * 1578 * Return: %true if @p was woken up, %false if it was already running. 1579 * or @state didn't match @p's state. 1580 */ 1581static int 1582try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1583{ 1584 unsigned long flags; 1585 int cpu, success = 0; 1586 1587 /* 1588 * If we are going to wake up a thread waiting for CONDITION we 1589 * need to ensure that CONDITION=1 done by the caller can not be 1590 * reordered with p->state check below. This pairs with mb() in 1591 * set_current_state() the waiting thread does. 1592 */ 1593 smp_mb__before_spinlock(); 1594 raw_spin_lock_irqsave(&p->pi_lock, flags); 1595 if (!(p->state & state)) 1596 goto out; 1597 1598 success = 1; /* we're going to change ->state */ 1599 cpu = task_cpu(p); 1600 1601 if (p->on_rq && ttwu_remote(p, wake_flags)) 1602 goto stat; 1603 1604#ifdef CONFIG_SMP 1605 /* 1606 * If the owning (remote) cpu is still in the middle of schedule() with 1607 * this task as prev, wait until its done referencing the task. 1608 */ 1609 while (p->on_cpu) 1610 cpu_relax(); 1611 /* 1612 * Pairs with the smp_wmb() in finish_lock_switch(). 1613 */ 1614 smp_rmb(); 1615 1616 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1617 p->state = TASK_WAKING; 1618 1619 if (p->sched_class->task_waking) 1620 p->sched_class->task_waking(p); 1621 1622 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1623 if (task_cpu(p) != cpu) { 1624 wake_flags |= WF_MIGRATED; 1625 set_task_cpu(p, cpu); 1626 } 1627#endif /* CONFIG_SMP */ 1628 1629 ttwu_queue(p, cpu); 1630stat: 1631 ttwu_stat(p, cpu, wake_flags); 1632out: 1633 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1634 1635 return success; 1636} 1637 1638/** 1639 * try_to_wake_up_local - try to wake up a local task with rq lock held 1640 * @p: the thread to be awakened 1641 * 1642 * Put @p on the run-queue if it's not already there. The caller must 1643 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1644 * the current task. 1645 */ 1646static void try_to_wake_up_local(struct task_struct *p) 1647{ 1648 struct rq *rq = task_rq(p); 1649 1650 if (WARN_ON_ONCE(rq != this_rq()) || 1651 WARN_ON_ONCE(p == current)) 1652 return; 1653 1654 lockdep_assert_held(&rq->lock); 1655 1656 if (!raw_spin_trylock(&p->pi_lock)) { 1657 raw_spin_unlock(&rq->lock); 1658 raw_spin_lock(&p->pi_lock); 1659 raw_spin_lock(&rq->lock); 1660 } 1661 1662 if (!(p->state & TASK_NORMAL)) 1663 goto out; 1664 1665 if (!p->on_rq) 1666 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1667 1668 ttwu_do_wakeup(rq, p, 0); 1669 ttwu_stat(p, smp_processor_id(), 0); 1670out: 1671 raw_spin_unlock(&p->pi_lock); 1672} 1673 1674/** 1675 * wake_up_process - Wake up a specific process 1676 * @p: The process to be woken up. 1677 * 1678 * Attempt to wake up the nominated process and move it to the set of runnable 1679 * processes. 1680 * 1681 * Return: 1 if the process was woken up, 0 if it was already running. 1682 * 1683 * It may be assumed that this function implies a write memory barrier before 1684 * changing the task state if and only if any tasks are woken up. 1685 */ 1686int wake_up_process(struct task_struct *p) 1687{ 1688 WARN_ON(task_is_stopped_or_traced(p)); 1689 return try_to_wake_up(p, TASK_NORMAL, 0); 1690} 1691EXPORT_SYMBOL(wake_up_process); 1692 1693int wake_up_state(struct task_struct *p, unsigned int state) 1694{ 1695 return try_to_wake_up(p, state, 0); 1696} 1697 1698/* 1699 * Perform scheduler related setup for a newly forked process p. 1700 * p is forked by current. 1701 * 1702 * __sched_fork() is basic setup used by init_idle() too: 1703 */ 1704static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1705{ 1706 p->on_rq = 0; 1707 1708 p->se.on_rq = 0; 1709 p->se.exec_start = 0; 1710 p->se.sum_exec_runtime = 0; 1711 p->se.prev_sum_exec_runtime = 0; 1712 p->se.nr_migrations = 0; 1713 p->se.vruntime = 0; 1714 INIT_LIST_HEAD(&p->se.group_node); 1715 1716#ifdef CONFIG_SCHEDSTATS 1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1718#endif 1719 1720 RB_CLEAR_NODE(&p->dl.rb_node); 1721 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1722 p->dl.dl_runtime = p->dl.runtime = 0; 1723 p->dl.dl_deadline = p->dl.deadline = 0; 1724 p->dl.dl_period = 0; 1725 p->dl.flags = 0; 1726 1727 INIT_LIST_HEAD(&p->rt.run_list); 1728 1729#ifdef CONFIG_PREEMPT_NOTIFIERS 1730 INIT_HLIST_HEAD(&p->preempt_notifiers); 1731#endif 1732 1733#ifdef CONFIG_NUMA_BALANCING 1734 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1735 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1736 p->mm->numa_scan_seq = 0; 1737 } 1738 1739 if (clone_flags & CLONE_VM) 1740 p->numa_preferred_nid = current->numa_preferred_nid; 1741 else 1742 p->numa_preferred_nid = -1; 1743 1744 p->node_stamp = 0ULL; 1745 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1746 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1747 p->numa_work.next = &p->numa_work; 1748 p->numa_faults_memory = NULL; 1749 p->numa_faults_buffer_memory = NULL; 1750 p->last_task_numa_placement = 0; 1751 p->last_sum_exec_runtime = 0; 1752 1753 INIT_LIST_HEAD(&p->numa_entry); 1754 p->numa_group = NULL; 1755#endif /* CONFIG_NUMA_BALANCING */ 1756} 1757 1758#ifdef CONFIG_NUMA_BALANCING 1759#ifdef CONFIG_SCHED_DEBUG 1760void set_numabalancing_state(bool enabled) 1761{ 1762 if (enabled) 1763 sched_feat_set("NUMA"); 1764 else 1765 sched_feat_set("NO_NUMA"); 1766} 1767#else 1768__read_mostly bool numabalancing_enabled; 1769 1770void set_numabalancing_state(bool enabled) 1771{ 1772 numabalancing_enabled = enabled; 1773} 1774#endif /* CONFIG_SCHED_DEBUG */ 1775 1776#ifdef CONFIG_PROC_SYSCTL 1777int sysctl_numa_balancing(struct ctl_table *table, int write, 1778 void __user *buffer, size_t *lenp, loff_t *ppos) 1779{ 1780 struct ctl_table t; 1781 int err; 1782 int state = numabalancing_enabled; 1783 1784 if (write && !capable(CAP_SYS_ADMIN)) 1785 return -EPERM; 1786 1787 t = *table; 1788 t.data = &state; 1789 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1790 if (err < 0) 1791 return err; 1792 if (write) 1793 set_numabalancing_state(state); 1794 return err; 1795} 1796#endif 1797#endif 1798 1799/* 1800 * fork()/clone()-time setup: 1801 */ 1802int sched_fork(unsigned long clone_flags, struct task_struct *p) 1803{ 1804 unsigned long flags; 1805 int cpu = get_cpu(); 1806 1807 __sched_fork(clone_flags, p); 1808 /* 1809 * We mark the process as running here. This guarantees that 1810 * nobody will actually run it, and a signal or other external 1811 * event cannot wake it up and insert it on the runqueue either. 1812 */ 1813 p->state = TASK_RUNNING; 1814 1815 /* 1816 * Make sure we do not leak PI boosting priority to the child. 1817 */ 1818 p->prio = current->normal_prio; 1819 1820 /* 1821 * Revert to default priority/policy on fork if requested. 1822 */ 1823 if (unlikely(p->sched_reset_on_fork)) { 1824 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1825 p->policy = SCHED_NORMAL; 1826 p->static_prio = NICE_TO_PRIO(0); 1827 p->rt_priority = 0; 1828 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1829 p->static_prio = NICE_TO_PRIO(0); 1830 1831 p->prio = p->normal_prio = __normal_prio(p); 1832 set_load_weight(p); 1833 1834 /* 1835 * We don't need the reset flag anymore after the fork. It has 1836 * fulfilled its duty: 1837 */ 1838 p->sched_reset_on_fork = 0; 1839 } 1840 1841 if (dl_prio(p->prio)) { 1842 put_cpu(); 1843 return -EAGAIN; 1844 } else if (rt_prio(p->prio)) { 1845 p->sched_class = &rt_sched_class; 1846 } else { 1847 p->sched_class = &fair_sched_class; 1848 } 1849 1850 if (p->sched_class->task_fork) 1851 p->sched_class->task_fork(p); 1852 1853 /* 1854 * The child is not yet in the pid-hash so no cgroup attach races, 1855 * and the cgroup is pinned to this child due to cgroup_fork() 1856 * is ran before sched_fork(). 1857 * 1858 * Silence PROVE_RCU. 1859 */ 1860 raw_spin_lock_irqsave(&p->pi_lock, flags); 1861 set_task_cpu(p, cpu); 1862 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1863 1864#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1865 if (likely(sched_info_on())) 1866 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1867#endif 1868#if defined(CONFIG_SMP) 1869 p->on_cpu = 0; 1870#endif 1871 init_task_preempt_count(p); 1872#ifdef CONFIG_SMP 1873 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1874 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1875#endif 1876 1877 put_cpu(); 1878 return 0; 1879} 1880 1881unsigned long to_ratio(u64 period, u64 runtime) 1882{ 1883 if (runtime == RUNTIME_INF) 1884 return 1ULL << 20; 1885 1886 /* 1887 * Doing this here saves a lot of checks in all 1888 * the calling paths, and returning zero seems 1889 * safe for them anyway. 1890 */ 1891 if (period == 0) 1892 return 0; 1893 1894 return div64_u64(runtime << 20, period); 1895} 1896 1897#ifdef CONFIG_SMP 1898inline struct dl_bw *dl_bw_of(int i) 1899{ 1900 return &cpu_rq(i)->rd->dl_bw; 1901} 1902 1903static inline int dl_bw_cpus(int i) 1904{ 1905 struct root_domain *rd = cpu_rq(i)->rd; 1906 int cpus = 0; 1907 1908 for_each_cpu_and(i, rd->span, cpu_active_mask) 1909 cpus++; 1910 1911 return cpus; 1912} 1913#else 1914inline struct dl_bw *dl_bw_of(int i) 1915{ 1916 return &cpu_rq(i)->dl.dl_bw; 1917} 1918 1919static inline int dl_bw_cpus(int i) 1920{ 1921 return 1; 1922} 1923#endif 1924 1925static inline 1926void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1927{ 1928 dl_b->total_bw -= tsk_bw; 1929} 1930 1931static inline 1932void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1933{ 1934 dl_b->total_bw += tsk_bw; 1935} 1936 1937static inline 1938bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1939{ 1940 return dl_b->bw != -1 && 1941 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1942} 1943 1944/* 1945 * We must be sure that accepting a new task (or allowing changing the 1946 * parameters of an existing one) is consistent with the bandwidth 1947 * constraints. If yes, this function also accordingly updates the currently 1948 * allocated bandwidth to reflect the new situation. 1949 * 1950 * This function is called while holding p's rq->lock. 1951 */ 1952static int dl_overflow(struct task_struct *p, int policy, 1953 const struct sched_attr *attr) 1954{ 1955 1956 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1957 u64 period = attr->sched_period; 1958 u64 runtime = attr->sched_runtime; 1959 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1960 int cpus, err = -1; 1961 1962 if (new_bw == p->dl.dl_bw) 1963 return 0; 1964 1965 /* 1966 * Either if a task, enters, leave, or stays -deadline but changes 1967 * its parameters, we may need to update accordingly the total 1968 * allocated bandwidth of the container. 1969 */ 1970 raw_spin_lock(&dl_b->lock); 1971 cpus = dl_bw_cpus(task_cpu(p)); 1972 if (dl_policy(policy) && !task_has_dl_policy(p) && 1973 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1974 __dl_add(dl_b, new_bw); 1975 err = 0; 1976 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1977 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1978 __dl_clear(dl_b, p->dl.dl_bw); 1979 __dl_add(dl_b, new_bw); 1980 err = 0; 1981 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1982 __dl_clear(dl_b, p->dl.dl_bw); 1983 err = 0; 1984 } 1985 raw_spin_unlock(&dl_b->lock); 1986 1987 return err; 1988} 1989 1990extern void init_dl_bw(struct dl_bw *dl_b); 1991 1992/* 1993 * wake_up_new_task - wake up a newly created task for the first time. 1994 * 1995 * This function will do some initial scheduler statistics housekeeping 1996 * that must be done for every newly created context, then puts the task 1997 * on the runqueue and wakes it. 1998 */ 1999void wake_up_new_task(struct task_struct *p) 2000{ 2001 unsigned long flags; 2002 struct rq *rq; 2003 2004 raw_spin_lock_irqsave(&p->pi_lock, flags); 2005#ifdef CONFIG_SMP 2006 /* 2007 * Fork balancing, do it here and not earlier because: 2008 * - cpus_allowed can change in the fork path 2009 * - any previously selected cpu might disappear through hotplug 2010 */ 2011 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2012#endif 2013 2014 /* Initialize new task's runnable average */ 2015 init_task_runnable_average(p); 2016 rq = __task_rq_lock(p); 2017 activate_task(rq, p, 0); 2018 p->on_rq = 1; 2019 trace_sched_wakeup_new(p, true); 2020 check_preempt_curr(rq, p, WF_FORK); 2021#ifdef CONFIG_SMP 2022 if (p->sched_class->task_woken) 2023 p->sched_class->task_woken(rq, p); 2024#endif 2025 task_rq_unlock(rq, p, &flags); 2026} 2027 2028#ifdef CONFIG_PREEMPT_NOTIFIERS 2029 2030/** 2031 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2032 * @notifier: notifier struct to register 2033 */ 2034void preempt_notifier_register(struct preempt_notifier *notifier) 2035{ 2036 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2037} 2038EXPORT_SYMBOL_GPL(preempt_notifier_register); 2039 2040/** 2041 * preempt_notifier_unregister - no longer interested in preemption notifications 2042 * @notifier: notifier struct to unregister 2043 * 2044 * This is safe to call from within a preemption notifier. 2045 */ 2046void preempt_notifier_unregister(struct preempt_notifier *notifier) 2047{ 2048 hlist_del(¬ifier->link); 2049} 2050EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2051 2052static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2053{ 2054 struct preempt_notifier *notifier; 2055 2056 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2057 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2058} 2059 2060static void 2061fire_sched_out_preempt_notifiers(struct task_struct *curr, 2062 struct task_struct *next) 2063{ 2064 struct preempt_notifier *notifier; 2065 2066 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2067 notifier->ops->sched_out(notifier, next); 2068} 2069 2070#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2071 2072static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2073{ 2074} 2075 2076static void 2077fire_sched_out_preempt_notifiers(struct task_struct *curr, 2078 struct task_struct *next) 2079{ 2080} 2081 2082#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2083 2084/** 2085 * prepare_task_switch - prepare to switch tasks 2086 * @rq: the runqueue preparing to switch 2087 * @prev: the current task that is being switched out 2088 * @next: the task we are going to switch to. 2089 * 2090 * This is called with the rq lock held and interrupts off. It must 2091 * be paired with a subsequent finish_task_switch after the context 2092 * switch. 2093 * 2094 * prepare_task_switch sets up locking and calls architecture specific 2095 * hooks. 2096 */ 2097static inline void 2098prepare_task_switch(struct rq *rq, struct task_struct *prev, 2099 struct task_struct *next) 2100{ 2101 trace_sched_switch(prev, next); 2102 sched_info_switch(rq, prev, next); 2103 perf_event_task_sched_out(prev, next); 2104 fire_sched_out_preempt_notifiers(prev, next); 2105 prepare_lock_switch(rq, next); 2106 prepare_arch_switch(next); 2107} 2108 2109/** 2110 * finish_task_switch - clean up after a task-switch 2111 * @rq: runqueue associated with task-switch 2112 * @prev: the thread we just switched away from. 2113 * 2114 * finish_task_switch must be called after the context switch, paired 2115 * with a prepare_task_switch call before the context switch. 2116 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2117 * and do any other architecture-specific cleanup actions. 2118 * 2119 * Note that we may have delayed dropping an mm in context_switch(). If 2120 * so, we finish that here outside of the runqueue lock. (Doing it 2121 * with the lock held can cause deadlocks; see schedule() for 2122 * details.) 2123 */ 2124static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2125 __releases(rq->lock) 2126{ 2127 struct mm_struct *mm = rq->prev_mm; 2128 long prev_state; 2129 2130 rq->prev_mm = NULL; 2131 2132 /* 2133 * A task struct has one reference for the use as "current". 2134 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2135 * schedule one last time. The schedule call will never return, and 2136 * the scheduled task must drop that reference. 2137 * The test for TASK_DEAD must occur while the runqueue locks are 2138 * still held, otherwise prev could be scheduled on another cpu, die 2139 * there before we look at prev->state, and then the reference would 2140 * be dropped twice. 2141 * Manfred Spraul <manfred@colorfullife.com> 2142 */ 2143 prev_state = prev->state; 2144 vtime_task_switch(prev); 2145 finish_arch_switch(prev); 2146 perf_event_task_sched_in(prev, current); 2147 finish_lock_switch(rq, prev); 2148 finish_arch_post_lock_switch(); 2149 2150 fire_sched_in_preempt_notifiers(current); 2151 if (mm) 2152 mmdrop(mm); 2153 if (unlikely(prev_state == TASK_DEAD)) { 2154 task_numa_free(prev); 2155 2156 if (prev->sched_class->task_dead) 2157 prev->sched_class->task_dead(prev); 2158 2159 /* 2160 * Remove function-return probe instances associated with this 2161 * task and put them back on the free list. 2162 */ 2163 kprobe_flush_task(prev); 2164 put_task_struct(prev); 2165 } 2166 2167 tick_nohz_task_switch(current); 2168} 2169 2170#ifdef CONFIG_SMP 2171 2172/* rq->lock is NOT held, but preemption is disabled */ 2173static inline void post_schedule(struct rq *rq) 2174{ 2175 if (rq->post_schedule) { 2176 unsigned long flags; 2177 2178 raw_spin_lock_irqsave(&rq->lock, flags); 2179 if (rq->curr->sched_class->post_schedule) 2180 rq->curr->sched_class->post_schedule(rq); 2181 raw_spin_unlock_irqrestore(&rq->lock, flags); 2182 2183 rq->post_schedule = 0; 2184 } 2185} 2186 2187#else 2188 2189static inline void post_schedule(struct rq *rq) 2190{ 2191} 2192 2193#endif 2194 2195/** 2196 * schedule_tail - first thing a freshly forked thread must call. 2197 * @prev: the thread we just switched away from. 2198 */ 2199asmlinkage void schedule_tail(struct task_struct *prev) 2200 __releases(rq->lock) 2201{ 2202 struct rq *rq = this_rq(); 2203 2204 finish_task_switch(rq, prev); 2205 2206 /* 2207 * FIXME: do we need to worry about rq being invalidated by the 2208 * task_switch? 2209 */ 2210 post_schedule(rq); 2211 2212#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2213 /* In this case, finish_task_switch does not reenable preemption */ 2214 preempt_enable(); 2215#endif 2216 if (current->set_child_tid) 2217 put_user(task_pid_vnr(current), current->set_child_tid); 2218} 2219 2220/* 2221 * context_switch - switch to the new MM and the new 2222 * thread's register state. 2223 */ 2224static inline void 2225context_switch(struct rq *rq, struct task_struct *prev, 2226 struct task_struct *next) 2227{ 2228 struct mm_struct *mm, *oldmm; 2229 2230 prepare_task_switch(rq, prev, next); 2231 2232 mm = next->mm; 2233 oldmm = prev->active_mm; 2234 /* 2235 * For paravirt, this is coupled with an exit in switch_to to 2236 * combine the page table reload and the switch backend into 2237 * one hypercall. 2238 */ 2239 arch_start_context_switch(prev); 2240 2241 if (!mm) { 2242 next->active_mm = oldmm; 2243 atomic_inc(&oldmm->mm_count); 2244 enter_lazy_tlb(oldmm, next); 2245 } else 2246 switch_mm(oldmm, mm, next); 2247 2248 if (!prev->mm) { 2249 prev->active_mm = NULL; 2250 rq->prev_mm = oldmm; 2251 } 2252 /* 2253 * Since the runqueue lock will be released by the next 2254 * task (which is an invalid locking op but in the case 2255 * of the scheduler it's an obvious special-case), so we 2256 * do an early lockdep release here: 2257 */ 2258#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2259 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2260#endif 2261 2262 context_tracking_task_switch(prev, next); 2263 /* Here we just switch the register state and the stack. */ 2264 switch_to(prev, next, prev); 2265 2266 barrier(); 2267 /* 2268 * this_rq must be evaluated again because prev may have moved 2269 * CPUs since it called schedule(), thus the 'rq' on its stack 2270 * frame will be invalid. 2271 */ 2272 finish_task_switch(this_rq(), prev); 2273} 2274 2275/* 2276 * nr_running and nr_context_switches: 2277 * 2278 * externally visible scheduler statistics: current number of runnable 2279 * threads, total number of context switches performed since bootup. 2280 */ 2281unsigned long nr_running(void) 2282{ 2283 unsigned long i, sum = 0; 2284 2285 for_each_online_cpu(i) 2286 sum += cpu_rq(i)->nr_running; 2287 2288 return sum; 2289} 2290 2291unsigned long long nr_context_switches(void) 2292{ 2293 int i; 2294 unsigned long long sum = 0; 2295 2296 for_each_possible_cpu(i) 2297 sum += cpu_rq(i)->nr_switches; 2298 2299 return sum; 2300} 2301 2302unsigned long nr_iowait(void) 2303{ 2304 unsigned long i, sum = 0; 2305 2306 for_each_possible_cpu(i) 2307 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2308 2309 return sum; 2310} 2311 2312unsigned long nr_iowait_cpu(int cpu) 2313{ 2314 struct rq *this = cpu_rq(cpu); 2315 return atomic_read(&this->nr_iowait); 2316} 2317 2318#ifdef CONFIG_SMP 2319 2320/* 2321 * sched_exec - execve() is a valuable balancing opportunity, because at 2322 * this point the task has the smallest effective memory and cache footprint. 2323 */ 2324void sched_exec(void) 2325{ 2326 struct task_struct *p = current; 2327 unsigned long flags; 2328 int dest_cpu; 2329 2330 raw_spin_lock_irqsave(&p->pi_lock, flags); 2331 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2332 if (dest_cpu == smp_processor_id()) 2333 goto unlock; 2334 2335 if (likely(cpu_active(dest_cpu))) { 2336 struct migration_arg arg = { p, dest_cpu }; 2337 2338 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2339 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2340 return; 2341 } 2342unlock: 2343 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2344} 2345 2346#endif 2347 2348DEFINE_PER_CPU(struct kernel_stat, kstat); 2349DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2350 2351EXPORT_PER_CPU_SYMBOL(kstat); 2352EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2353 2354/* 2355 * Return any ns on the sched_clock that have not yet been accounted in 2356 * @p in case that task is currently running. 2357 * 2358 * Called with task_rq_lock() held on @rq. 2359 */ 2360static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2361{ 2362 u64 ns = 0; 2363 2364 if (task_current(rq, p)) { 2365 update_rq_clock(rq); 2366 ns = rq_clock_task(rq) - p->se.exec_start; 2367 if ((s64)ns < 0) 2368 ns = 0; 2369 } 2370 2371 return ns; 2372} 2373 2374unsigned long long task_delta_exec(struct task_struct *p) 2375{ 2376 unsigned long flags; 2377 struct rq *rq; 2378 u64 ns = 0; 2379 2380 rq = task_rq_lock(p, &flags); 2381 ns = do_task_delta_exec(p, rq); 2382 task_rq_unlock(rq, p, &flags); 2383 2384 return ns; 2385} 2386 2387/* 2388 * Return accounted runtime for the task. 2389 * In case the task is currently running, return the runtime plus current's 2390 * pending runtime that have not been accounted yet. 2391 */ 2392unsigned long long task_sched_runtime(struct task_struct *p) 2393{ 2394 unsigned long flags; 2395 struct rq *rq; 2396 u64 ns = 0; 2397 2398#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2399 /* 2400 * 64-bit doesn't need locks to atomically read a 64bit value. 2401 * So we have a optimization chance when the task's delta_exec is 0. 2402 * Reading ->on_cpu is racy, but this is ok. 2403 * 2404 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2405 * If we race with it entering cpu, unaccounted time is 0. This is 2406 * indistinguishable from the read occurring a few cycles earlier. 2407 */ 2408 if (!p->on_cpu) 2409 return p->se.sum_exec_runtime; 2410#endif 2411 2412 rq = task_rq_lock(p, &flags); 2413 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2414 task_rq_unlock(rq, p, &flags); 2415 2416 return ns; 2417} 2418 2419/* 2420 * This function gets called by the timer code, with HZ frequency. 2421 * We call it with interrupts disabled. 2422 */ 2423void scheduler_tick(void) 2424{ 2425 int cpu = smp_processor_id(); 2426 struct rq *rq = cpu_rq(cpu); 2427 struct task_struct *curr = rq->curr; 2428 2429 sched_clock_tick(); 2430 2431 raw_spin_lock(&rq->lock); 2432 update_rq_clock(rq); 2433 curr->sched_class->task_tick(rq, curr, 0); 2434 update_cpu_load_active(rq); 2435 raw_spin_unlock(&rq->lock); 2436 2437 perf_event_task_tick(); 2438 2439#ifdef CONFIG_SMP 2440 rq->idle_balance = idle_cpu(cpu); 2441 trigger_load_balance(rq); 2442#endif 2443 rq_last_tick_reset(rq); 2444} 2445 2446#ifdef CONFIG_NO_HZ_FULL 2447/** 2448 * scheduler_tick_max_deferment 2449 * 2450 * Keep at least one tick per second when a single 2451 * active task is running because the scheduler doesn't 2452 * yet completely support full dynticks environment. 2453 * 2454 * This makes sure that uptime, CFS vruntime, load 2455 * balancing, etc... continue to move forward, even 2456 * with a very low granularity. 2457 * 2458 * Return: Maximum deferment in nanoseconds. 2459 */ 2460u64 scheduler_tick_max_deferment(void) 2461{ 2462 struct rq *rq = this_rq(); 2463 unsigned long next, now = ACCESS_ONCE(jiffies); 2464 2465 next = rq->last_sched_tick + HZ; 2466 2467 if (time_before_eq(next, now)) 2468 return 0; 2469 2470 return jiffies_to_nsecs(next - now); 2471} 2472#endif 2473 2474notrace unsigned long get_parent_ip(unsigned long addr) 2475{ 2476 if (in_lock_functions(addr)) { 2477 addr = CALLER_ADDR2; 2478 if (in_lock_functions(addr)) 2479 addr = CALLER_ADDR3; 2480 } 2481 return addr; 2482} 2483 2484#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2485 defined(CONFIG_PREEMPT_TRACER)) 2486 2487void __kprobes preempt_count_add(int val) 2488{ 2489#ifdef CONFIG_DEBUG_PREEMPT 2490 /* 2491 * Underflow? 2492 */ 2493 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2494 return; 2495#endif 2496 __preempt_count_add(val); 2497#ifdef CONFIG_DEBUG_PREEMPT 2498 /* 2499 * Spinlock count overflowing soon? 2500 */ 2501 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2502 PREEMPT_MASK - 10); 2503#endif 2504 if (preempt_count() == val) { 2505 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2506#ifdef CONFIG_DEBUG_PREEMPT 2507 current->preempt_disable_ip = ip; 2508#endif 2509 trace_preempt_off(CALLER_ADDR0, ip); 2510 } 2511} 2512EXPORT_SYMBOL(preempt_count_add); 2513 2514void __kprobes preempt_count_sub(int val) 2515{ 2516#ifdef CONFIG_DEBUG_PREEMPT 2517 /* 2518 * Underflow? 2519 */ 2520 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2521 return; 2522 /* 2523 * Is the spinlock portion underflowing? 2524 */ 2525 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2526 !(preempt_count() & PREEMPT_MASK))) 2527 return; 2528#endif 2529 2530 if (preempt_count() == val) 2531 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2532 __preempt_count_sub(val); 2533} 2534EXPORT_SYMBOL(preempt_count_sub); 2535 2536#endif 2537 2538/* 2539 * Print scheduling while atomic bug: 2540 */ 2541static noinline void __schedule_bug(struct task_struct *prev) 2542{ 2543 if (oops_in_progress) 2544 return; 2545 2546 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2547 prev->comm, prev->pid, preempt_count()); 2548 2549 debug_show_held_locks(prev); 2550 print_modules(); 2551 if (irqs_disabled()) 2552 print_irqtrace_events(prev); 2553#ifdef CONFIG_DEBUG_PREEMPT 2554 if (in_atomic_preempt_off()) { 2555 pr_err("Preemption disabled at:"); 2556 print_ip_sym(current->preempt_disable_ip); 2557 pr_cont("\n"); 2558 } 2559#endif 2560 dump_stack(); 2561 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2562} 2563 2564/* 2565 * Various schedule()-time debugging checks and statistics: 2566 */ 2567static inline void schedule_debug(struct task_struct *prev) 2568{ 2569 /* 2570 * Test if we are atomic. Since do_exit() needs to call into 2571 * schedule() atomically, we ignore that path. Otherwise whine 2572 * if we are scheduling when we should not. 2573 */ 2574 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2575 __schedule_bug(prev); 2576 rcu_sleep_check(); 2577 2578 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2579 2580 schedstat_inc(this_rq(), sched_count); 2581} 2582 2583/* 2584 * Pick up the highest-prio task: 2585 */ 2586static inline struct task_struct * 2587pick_next_task(struct rq *rq, struct task_struct *prev) 2588{ 2589 const struct sched_class *class = &fair_sched_class; 2590 struct task_struct *p; 2591 2592 /* 2593 * Optimization: we know that if all tasks are in 2594 * the fair class we can call that function directly: 2595 */ 2596 if (likely(prev->sched_class == class && 2597 rq->nr_running == rq->cfs.h_nr_running)) { 2598 p = fair_sched_class.pick_next_task(rq, prev); 2599 if (likely(p && p != RETRY_TASK)) 2600 return p; 2601 } 2602 2603again: 2604 for_each_class(class) { 2605 p = class->pick_next_task(rq, prev); 2606 if (p) { 2607 if (unlikely(p == RETRY_TASK)) 2608 goto again; 2609 return p; 2610 } 2611 } 2612 2613 BUG(); /* the idle class will always have a runnable task */ 2614} 2615 2616/* 2617 * __schedule() is the main scheduler function. 2618 * 2619 * The main means of driving the scheduler and thus entering this function are: 2620 * 2621 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2622 * 2623 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2624 * paths. For example, see arch/x86/entry_64.S. 2625 * 2626 * To drive preemption between tasks, the scheduler sets the flag in timer 2627 * interrupt handler scheduler_tick(). 2628 * 2629 * 3. Wakeups don't really cause entry into schedule(). They add a 2630 * task to the run-queue and that's it. 2631 * 2632 * Now, if the new task added to the run-queue preempts the current 2633 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2634 * called on the nearest possible occasion: 2635 * 2636 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2637 * 2638 * - in syscall or exception context, at the next outmost 2639 * preempt_enable(). (this might be as soon as the wake_up()'s 2640 * spin_unlock()!) 2641 * 2642 * - in IRQ context, return from interrupt-handler to 2643 * preemptible context 2644 * 2645 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2646 * then at the next: 2647 * 2648 * - cond_resched() call 2649 * - explicit schedule() call 2650 * - return from syscall or exception to user-space 2651 * - return from interrupt-handler to user-space 2652 */ 2653static void __sched __schedule(void) 2654{ 2655 struct task_struct *prev, *next; 2656 unsigned long *switch_count; 2657 struct rq *rq; 2658 int cpu; 2659 2660need_resched: 2661 preempt_disable(); 2662 cpu = smp_processor_id(); 2663 rq = cpu_rq(cpu); 2664 rcu_note_context_switch(cpu); 2665 prev = rq->curr; 2666 2667 schedule_debug(prev); 2668 2669 if (sched_feat(HRTICK)) 2670 hrtick_clear(rq); 2671 2672 /* 2673 * Make sure that signal_pending_state()->signal_pending() below 2674 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2675 * done by the caller to avoid the race with signal_wake_up(). 2676 */ 2677 smp_mb__before_spinlock(); 2678 raw_spin_lock_irq(&rq->lock); 2679 2680 switch_count = &prev->nivcsw; 2681 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2682 if (unlikely(signal_pending_state(prev->state, prev))) { 2683 prev->state = TASK_RUNNING; 2684 } else { 2685 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2686 prev->on_rq = 0; 2687 2688 /* 2689 * If a worker went to sleep, notify and ask workqueue 2690 * whether it wants to wake up a task to maintain 2691 * concurrency. 2692 */ 2693 if (prev->flags & PF_WQ_WORKER) { 2694 struct task_struct *to_wakeup; 2695 2696 to_wakeup = wq_worker_sleeping(prev, cpu); 2697 if (to_wakeup) 2698 try_to_wake_up_local(to_wakeup); 2699 } 2700 } 2701 switch_count = &prev->nvcsw; 2702 } 2703 2704 if (prev->on_rq || rq->skip_clock_update < 0) 2705 update_rq_clock(rq); 2706 2707 next = pick_next_task(rq, prev); 2708 clear_tsk_need_resched(prev); 2709 clear_preempt_need_resched(); 2710 rq->skip_clock_update = 0; 2711 2712 if (likely(prev != next)) { 2713 rq->nr_switches++; 2714 rq->curr = next; 2715 ++*switch_count; 2716 2717 context_switch(rq, prev, next); /* unlocks the rq */ 2718 /* 2719 * The context switch have flipped the stack from under us 2720 * and restored the local variables which were saved when 2721 * this task called schedule() in the past. prev == current 2722 * is still correct, but it can be moved to another cpu/rq. 2723 */ 2724 cpu = smp_processor_id(); 2725 rq = cpu_rq(cpu); 2726 } else 2727 raw_spin_unlock_irq(&rq->lock); 2728 2729 post_schedule(rq); 2730 2731 sched_preempt_enable_no_resched(); 2732 if (need_resched()) 2733 goto need_resched; 2734} 2735 2736static inline void sched_submit_work(struct task_struct *tsk) 2737{ 2738 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2739 return; 2740 /* 2741 * If we are going to sleep and we have plugged IO queued, 2742 * make sure to submit it to avoid deadlocks. 2743 */ 2744 if (blk_needs_flush_plug(tsk)) 2745 blk_schedule_flush_plug(tsk); 2746} 2747 2748asmlinkage void __sched schedule(void) 2749{ 2750 struct task_struct *tsk = current; 2751 2752 sched_submit_work(tsk); 2753 __schedule(); 2754} 2755EXPORT_SYMBOL(schedule); 2756 2757#ifdef CONFIG_CONTEXT_TRACKING 2758asmlinkage void __sched schedule_user(void) 2759{ 2760 /* 2761 * If we come here after a random call to set_need_resched(), 2762 * or we have been woken up remotely but the IPI has not yet arrived, 2763 * we haven't yet exited the RCU idle mode. Do it here manually until 2764 * we find a better solution. 2765 */ 2766 user_exit(); 2767 schedule(); 2768 user_enter(); 2769} 2770#endif 2771 2772/** 2773 * schedule_preempt_disabled - called with preemption disabled 2774 * 2775 * Returns with preemption disabled. Note: preempt_count must be 1 2776 */ 2777void __sched schedule_preempt_disabled(void) 2778{ 2779 sched_preempt_enable_no_resched(); 2780 schedule(); 2781 preempt_disable(); 2782} 2783 2784#ifdef CONFIG_PREEMPT 2785/* 2786 * this is the entry point to schedule() from in-kernel preemption 2787 * off of preempt_enable. Kernel preemptions off return from interrupt 2788 * occur there and call schedule directly. 2789 */ 2790asmlinkage void __sched notrace preempt_schedule(void) 2791{ 2792 /* 2793 * If there is a non-zero preempt_count or interrupts are disabled, 2794 * we do not want to preempt the current task. Just return.. 2795 */ 2796 if (likely(!preemptible())) 2797 return; 2798 2799 do { 2800 __preempt_count_add(PREEMPT_ACTIVE); 2801 __schedule(); 2802 __preempt_count_sub(PREEMPT_ACTIVE); 2803 2804 /* 2805 * Check again in case we missed a preemption opportunity 2806 * between schedule and now. 2807 */ 2808 barrier(); 2809 } while (need_resched()); 2810} 2811EXPORT_SYMBOL(preempt_schedule); 2812#endif /* CONFIG_PREEMPT */ 2813 2814/* 2815 * this is the entry point to schedule() from kernel preemption 2816 * off of irq context. 2817 * Note, that this is called and return with irqs disabled. This will 2818 * protect us against recursive calling from irq. 2819 */ 2820asmlinkage void __sched preempt_schedule_irq(void) 2821{ 2822 enum ctx_state prev_state; 2823 2824 /* Catch callers which need to be fixed */ 2825 BUG_ON(preempt_count() || !irqs_disabled()); 2826 2827 prev_state = exception_enter(); 2828 2829 do { 2830 __preempt_count_add(PREEMPT_ACTIVE); 2831 local_irq_enable(); 2832 __schedule(); 2833 local_irq_disable(); 2834 __preempt_count_sub(PREEMPT_ACTIVE); 2835 2836 /* 2837 * Check again in case we missed a preemption opportunity 2838 * between schedule and now. 2839 */ 2840 barrier(); 2841 } while (need_resched()); 2842 2843 exception_exit(prev_state); 2844} 2845 2846int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2847 void *key) 2848{ 2849 return try_to_wake_up(curr->private, mode, wake_flags); 2850} 2851EXPORT_SYMBOL(default_wake_function); 2852 2853static long __sched 2854sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2855{ 2856 unsigned long flags; 2857 wait_queue_t wait; 2858 2859 init_waitqueue_entry(&wait, current); 2860 2861 __set_current_state(state); 2862 2863 spin_lock_irqsave(&q->lock, flags); 2864 __add_wait_queue(q, &wait); 2865 spin_unlock(&q->lock); 2866 timeout = schedule_timeout(timeout); 2867 spin_lock_irq(&q->lock); 2868 __remove_wait_queue(q, &wait); 2869 spin_unlock_irqrestore(&q->lock, flags); 2870 2871 return timeout; 2872} 2873 2874void __sched interruptible_sleep_on(wait_queue_head_t *q) 2875{ 2876 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2877} 2878EXPORT_SYMBOL(interruptible_sleep_on); 2879 2880long __sched 2881interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2882{ 2883 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2884} 2885EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2886 2887void __sched sleep_on(wait_queue_head_t *q) 2888{ 2889 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2890} 2891EXPORT_SYMBOL(sleep_on); 2892 2893long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2894{ 2895 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2896} 2897EXPORT_SYMBOL(sleep_on_timeout); 2898 2899#ifdef CONFIG_RT_MUTEXES 2900 2901/* 2902 * rt_mutex_setprio - set the current priority of a task 2903 * @p: task 2904 * @prio: prio value (kernel-internal form) 2905 * 2906 * This function changes the 'effective' priority of a task. It does 2907 * not touch ->normal_prio like __setscheduler(). 2908 * 2909 * Used by the rt_mutex code to implement priority inheritance 2910 * logic. Call site only calls if the priority of the task changed. 2911 */ 2912void rt_mutex_setprio(struct task_struct *p, int prio) 2913{ 2914 int oldprio, on_rq, running, enqueue_flag = 0; 2915 struct rq *rq; 2916 const struct sched_class *prev_class; 2917 2918 BUG_ON(prio > MAX_PRIO); 2919 2920 rq = __task_rq_lock(p); 2921 2922 /* 2923 * Idle task boosting is a nono in general. There is one 2924 * exception, when PREEMPT_RT and NOHZ is active: 2925 * 2926 * The idle task calls get_next_timer_interrupt() and holds 2927 * the timer wheel base->lock on the CPU and another CPU wants 2928 * to access the timer (probably to cancel it). We can safely 2929 * ignore the boosting request, as the idle CPU runs this code 2930 * with interrupts disabled and will complete the lock 2931 * protected section without being interrupted. So there is no 2932 * real need to boost. 2933 */ 2934 if (unlikely(p == rq->idle)) { 2935 WARN_ON(p != rq->curr); 2936 WARN_ON(p->pi_blocked_on); 2937 goto out_unlock; 2938 } 2939 2940 trace_sched_pi_setprio(p, prio); 2941 p->pi_top_task = rt_mutex_get_top_task(p); 2942 oldprio = p->prio; 2943 prev_class = p->sched_class; 2944 on_rq = p->on_rq; 2945 running = task_current(rq, p); 2946 if (on_rq) 2947 dequeue_task(rq, p, 0); 2948 if (running) 2949 p->sched_class->put_prev_task(rq, p); 2950 2951 /* 2952 * Boosting condition are: 2953 * 1. -rt task is running and holds mutex A 2954 * --> -dl task blocks on mutex A 2955 * 2956 * 2. -dl task is running and holds mutex A 2957 * --> -dl task blocks on mutex A and could preempt the 2958 * running task 2959 */ 2960 if (dl_prio(prio)) { 2961 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2962 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2963 p->dl.dl_boosted = 1; 2964 p->dl.dl_throttled = 0; 2965 enqueue_flag = ENQUEUE_REPLENISH; 2966 } else 2967 p->dl.dl_boosted = 0; 2968 p->sched_class = &dl_sched_class; 2969 } else if (rt_prio(prio)) { 2970 if (dl_prio(oldprio)) 2971 p->dl.dl_boosted = 0; 2972 if (oldprio < prio) 2973 enqueue_flag = ENQUEUE_HEAD; 2974 p->sched_class = &rt_sched_class; 2975 } else { 2976 if (dl_prio(oldprio)) 2977 p->dl.dl_boosted = 0; 2978 p->sched_class = &fair_sched_class; 2979 } 2980 2981 p->prio = prio; 2982 2983 if (running) 2984 p->sched_class->set_curr_task(rq); 2985 if (on_rq) 2986 enqueue_task(rq, p, enqueue_flag); 2987 2988 check_class_changed(rq, p, prev_class, oldprio); 2989out_unlock: 2990 __task_rq_unlock(rq); 2991} 2992#endif 2993 2994void set_user_nice(struct task_struct *p, long nice) 2995{ 2996 int old_prio, delta, on_rq; 2997 unsigned long flags; 2998 struct rq *rq; 2999 3000 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 3001 return; 3002 /* 3003 * We have to be careful, if called from sys_setpriority(), 3004 * the task might be in the middle of scheduling on another CPU. 3005 */ 3006 rq = task_rq_lock(p, &flags); 3007 /* 3008 * The RT priorities are set via sched_setscheduler(), but we still 3009 * allow the 'normal' nice value to be set - but as expected 3010 * it wont have any effect on scheduling until the task is 3011 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3012 */ 3013 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3014 p->static_prio = NICE_TO_PRIO(nice); 3015 goto out_unlock; 3016 } 3017 on_rq = p->on_rq; 3018 if (on_rq) 3019 dequeue_task(rq, p, 0); 3020 3021 p->static_prio = NICE_TO_PRIO(nice); 3022 set_load_weight(p); 3023 old_prio = p->prio; 3024 p->prio = effective_prio(p); 3025 delta = p->prio - old_prio; 3026 3027 if (on_rq) { 3028 enqueue_task(rq, p, 0); 3029 /* 3030 * If the task increased its priority or is running and 3031 * lowered its priority, then reschedule its CPU: 3032 */ 3033 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3034 resched_task(rq->curr); 3035 } 3036out_unlock: 3037 task_rq_unlock(rq, p, &flags); 3038} 3039EXPORT_SYMBOL(set_user_nice); 3040 3041/* 3042 * can_nice - check if a task can reduce its nice value 3043 * @p: task 3044 * @nice: nice value 3045 */ 3046int can_nice(const struct task_struct *p, const int nice) 3047{ 3048 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3049 int nice_rlim = 20 - nice; 3050 3051 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3052 capable(CAP_SYS_NICE)); 3053} 3054 3055#ifdef __ARCH_WANT_SYS_NICE 3056 3057/* 3058 * sys_nice - change the priority of the current process. 3059 * @increment: priority increment 3060 * 3061 * sys_setpriority is a more generic, but much slower function that 3062 * does similar things. 3063 */ 3064SYSCALL_DEFINE1(nice, int, increment) 3065{ 3066 long nice, retval; 3067 3068 /* 3069 * Setpriority might change our priority at the same moment. 3070 * We don't have to worry. Conceptually one call occurs first 3071 * and we have a single winner. 3072 */ 3073 if (increment < -40) 3074 increment = -40; 3075 if (increment > 40) 3076 increment = 40; 3077 3078 nice = task_nice(current) + increment; 3079 if (nice < MIN_NICE) 3080 nice = MIN_NICE; 3081 if (nice > MAX_NICE) 3082 nice = MAX_NICE; 3083 3084 if (increment < 0 && !can_nice(current, nice)) 3085 return -EPERM; 3086 3087 retval = security_task_setnice(current, nice); 3088 if (retval) 3089 return retval; 3090 3091 set_user_nice(current, nice); 3092 return 0; 3093} 3094 3095#endif 3096 3097/** 3098 * task_prio - return the priority value of a given task. 3099 * @p: the task in question. 3100 * 3101 * Return: The priority value as seen by users in /proc. 3102 * RT tasks are offset by -200. Normal tasks are centered 3103 * around 0, value goes from -16 to +15. 3104 */ 3105int task_prio(const struct task_struct *p) 3106{ 3107 return p->prio - MAX_RT_PRIO; 3108} 3109 3110/** 3111 * idle_cpu - is a given cpu idle currently? 3112 * @cpu: the processor in question. 3113 * 3114 * Return: 1 if the CPU is currently idle. 0 otherwise. 3115 */ 3116int idle_cpu(int cpu) 3117{ 3118 struct rq *rq = cpu_rq(cpu); 3119 3120 if (rq->curr != rq->idle) 3121 return 0; 3122 3123 if (rq->nr_running) 3124 return 0; 3125 3126#ifdef CONFIG_SMP 3127 if (!llist_empty(&rq->wake_list)) 3128 return 0; 3129#endif 3130 3131 return 1; 3132} 3133 3134/** 3135 * idle_task - return the idle task for a given cpu. 3136 * @cpu: the processor in question. 3137 * 3138 * Return: The idle task for the cpu @cpu. 3139 */ 3140struct task_struct *idle_task(int cpu) 3141{ 3142 return cpu_rq(cpu)->idle; 3143} 3144 3145/** 3146 * find_process_by_pid - find a process with a matching PID value. 3147 * @pid: the pid in question. 3148 * 3149 * The task of @pid, if found. %NULL otherwise. 3150 */ 3151static struct task_struct *find_process_by_pid(pid_t pid) 3152{ 3153 return pid ? find_task_by_vpid(pid) : current; 3154} 3155 3156/* 3157 * This function initializes the sched_dl_entity of a newly becoming 3158 * SCHED_DEADLINE task. 3159 * 3160 * Only the static values are considered here, the actual runtime and the 3161 * absolute deadline will be properly calculated when the task is enqueued 3162 * for the first time with its new policy. 3163 */ 3164static void 3165__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3166{ 3167 struct sched_dl_entity *dl_se = &p->dl; 3168 3169 init_dl_task_timer(dl_se); 3170 dl_se->dl_runtime = attr->sched_runtime; 3171 dl_se->dl_deadline = attr->sched_deadline; 3172 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3173 dl_se->flags = attr->sched_flags; 3174 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3175 dl_se->dl_throttled = 0; 3176 dl_se->dl_new = 1; 3177} 3178 3179static void __setscheduler_params(struct task_struct *p, 3180 const struct sched_attr *attr) 3181{ 3182 int policy = attr->sched_policy; 3183 3184 if (policy == -1) /* setparam */ 3185 policy = p->policy; 3186 3187 p->policy = policy; 3188 3189 if (dl_policy(policy)) 3190 __setparam_dl(p, attr); 3191 else if (fair_policy(policy)) 3192 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3193 3194 /* 3195 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3196 * !rt_policy. Always setting this ensures that things like 3197 * getparam()/getattr() don't report silly values for !rt tasks. 3198 */ 3199 p->rt_priority = attr->sched_priority; 3200 set_load_weight(p); 3201} 3202 3203/* Actually do priority change: must hold pi & rq lock. */ 3204static void __setscheduler(struct rq *rq, struct task_struct *p, 3205 const struct sched_attr *attr) 3206{ 3207 __setscheduler_params(p, attr); 3208 3209 if (dl_prio(p->prio)) 3210 p->sched_class = &dl_sched_class; 3211 else if (rt_prio(p->prio)) 3212 p->sched_class = &rt_sched_class; 3213 else 3214 p->sched_class = &fair_sched_class; 3215} 3216 3217static void 3218__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3219{ 3220 struct sched_dl_entity *dl_se = &p->dl; 3221 3222 attr->sched_priority = p->rt_priority; 3223 attr->sched_runtime = dl_se->dl_runtime; 3224 attr->sched_deadline = dl_se->dl_deadline; 3225 attr->sched_period = dl_se->dl_period; 3226 attr->sched_flags = dl_se->flags; 3227} 3228 3229/* 3230 * This function validates the new parameters of a -deadline task. 3231 * We ask for the deadline not being zero, and greater or equal 3232 * than the runtime, as well as the period of being zero or 3233 * greater than deadline. Furthermore, we have to be sure that 3234 * user parameters are above the internal resolution (1us); we 3235 * check sched_runtime only since it is always the smaller one. 3236 */ 3237static bool 3238__checkparam_dl(const struct sched_attr *attr) 3239{ 3240 return attr && attr->sched_deadline != 0 && 3241 (attr->sched_period == 0 || 3242 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3243 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3244 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3245} 3246 3247/* 3248 * check the target process has a UID that matches the current process's 3249 */ 3250static bool check_same_owner(struct task_struct *p) 3251{ 3252 const struct cred *cred = current_cred(), *pcred; 3253 bool match; 3254 3255 rcu_read_lock(); 3256 pcred = __task_cred(p); 3257 match = (uid_eq(cred->euid, pcred->euid) || 3258 uid_eq(cred->euid, pcred->uid)); 3259 rcu_read_unlock(); 3260 return match; 3261} 3262 3263static int __sched_setscheduler(struct task_struct *p, 3264 const struct sched_attr *attr, 3265 bool user) 3266{ 3267 int newprio = MAX_RT_PRIO - 1 - attr->sched_priority; 3268 int retval, oldprio, oldpolicy = -1, on_rq, running; 3269 int policy = attr->sched_policy; 3270 unsigned long flags; 3271 const struct sched_class *prev_class; 3272 struct rq *rq; 3273 int reset_on_fork; 3274 3275 /* may grab non-irq protected spin_locks */ 3276 BUG_ON(in_interrupt()); 3277recheck: 3278 /* double check policy once rq lock held */ 3279 if (policy < 0) { 3280 reset_on_fork = p->sched_reset_on_fork; 3281 policy = oldpolicy = p->policy; 3282 } else { 3283 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3284 3285 if (policy != SCHED_DEADLINE && 3286 policy != SCHED_FIFO && policy != SCHED_RR && 3287 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3288 policy != SCHED_IDLE) 3289 return -EINVAL; 3290 } 3291 3292 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3293 return -EINVAL; 3294 3295 /* 3296 * Valid priorities for SCHED_FIFO and SCHED_RR are 3297 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3298 * SCHED_BATCH and SCHED_IDLE is 0. 3299 */ 3300 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3301 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3302 return -EINVAL; 3303 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3304 (rt_policy(policy) != (attr->sched_priority != 0))) 3305 return -EINVAL; 3306 3307 /* 3308 * Allow unprivileged RT tasks to decrease priority: 3309 */ 3310 if (user && !capable(CAP_SYS_NICE)) { 3311 if (fair_policy(policy)) { 3312 if (attr->sched_nice < task_nice(p) && 3313 !can_nice(p, attr->sched_nice)) 3314 return -EPERM; 3315 } 3316 3317 if (rt_policy(policy)) { 3318 unsigned long rlim_rtprio = 3319 task_rlimit(p, RLIMIT_RTPRIO); 3320 3321 /* can't set/change the rt policy */ 3322 if (policy != p->policy && !rlim_rtprio) 3323 return -EPERM; 3324 3325 /* can't increase priority */ 3326 if (attr->sched_priority > p->rt_priority && 3327 attr->sched_priority > rlim_rtprio) 3328 return -EPERM; 3329 } 3330 3331 /* 3332 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3333 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3334 */ 3335 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3336 if (!can_nice(p, task_nice(p))) 3337 return -EPERM; 3338 } 3339 3340 /* can't change other user's priorities */ 3341 if (!check_same_owner(p)) 3342 return -EPERM; 3343 3344 /* Normal users shall not reset the sched_reset_on_fork flag */ 3345 if (p->sched_reset_on_fork && !reset_on_fork) 3346 return -EPERM; 3347 } 3348 3349 if (user) { 3350 retval = security_task_setscheduler(p); 3351 if (retval) 3352 return retval; 3353 } 3354 3355 /* 3356 * make sure no PI-waiters arrive (or leave) while we are 3357 * changing the priority of the task: 3358 * 3359 * To be able to change p->policy safely, the appropriate 3360 * runqueue lock must be held. 3361 */ 3362 rq = task_rq_lock(p, &flags); 3363 3364 /* 3365 * Changing the policy of the stop threads its a very bad idea 3366 */ 3367 if (p == rq->stop) { 3368 task_rq_unlock(rq, p, &flags); 3369 return -EINVAL; 3370 } 3371 3372 /* 3373 * If not changing anything there's no need to proceed further, 3374 * but store a possible modification of reset_on_fork. 3375 */ 3376 if (unlikely(policy == p->policy)) { 3377 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3378 goto change; 3379 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3380 goto change; 3381 if (dl_policy(policy)) 3382 goto change; 3383 3384 p->sched_reset_on_fork = reset_on_fork; 3385 task_rq_unlock(rq, p, &flags); 3386 return 0; 3387 } 3388change: 3389 3390 if (user) { 3391#ifdef CONFIG_RT_GROUP_SCHED 3392 /* 3393 * Do not allow realtime tasks into groups that have no runtime 3394 * assigned. 3395 */ 3396 if (rt_bandwidth_enabled() && rt_policy(policy) && 3397 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3398 !task_group_is_autogroup(task_group(p))) { 3399 task_rq_unlock(rq, p, &flags); 3400 return -EPERM; 3401 } 3402#endif 3403#ifdef CONFIG_SMP 3404 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3405 cpumask_t *span = rq->rd->span; 3406 3407 /* 3408 * Don't allow tasks with an affinity mask smaller than 3409 * the entire root_domain to become SCHED_DEADLINE. We 3410 * will also fail if there's no bandwidth available. 3411 */ 3412 if (!cpumask_subset(span, &p->cpus_allowed) || 3413 rq->rd->dl_bw.bw == 0) { 3414 task_rq_unlock(rq, p, &flags); 3415 return -EPERM; 3416 } 3417 } 3418#endif 3419 } 3420 3421 /* recheck policy now with rq lock held */ 3422 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3423 policy = oldpolicy = -1; 3424 task_rq_unlock(rq, p, &flags); 3425 goto recheck; 3426 } 3427 3428 /* 3429 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3430 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3431 * is available. 3432 */ 3433 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3434 task_rq_unlock(rq, p, &flags); 3435 return -EBUSY; 3436 } 3437 3438 p->sched_reset_on_fork = reset_on_fork; 3439 oldprio = p->prio; 3440 3441 /* 3442 * Special case for priority boosted tasks. 3443 * 3444 * If the new priority is lower or equal (user space view) 3445 * than the current (boosted) priority, we just store the new 3446 * normal parameters and do not touch the scheduler class and 3447 * the runqueue. This will be done when the task deboost 3448 * itself. 3449 */ 3450 if (rt_mutex_check_prio(p, newprio)) { 3451 __setscheduler_params(p, attr); 3452 task_rq_unlock(rq, p, &flags); 3453 return 0; 3454 } 3455 3456 on_rq = p->on_rq; 3457 running = task_current(rq, p); 3458 if (on_rq) 3459 dequeue_task(rq, p, 0); 3460 if (running) 3461 p->sched_class->put_prev_task(rq, p); 3462 3463 prev_class = p->sched_class; 3464 __setscheduler(rq, p, attr); 3465 3466 if (running) 3467 p->sched_class->set_curr_task(rq); 3468 if (on_rq) { 3469 /* 3470 * We enqueue to tail when the priority of a task is 3471 * increased (user space view). 3472 */ 3473 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3474 } 3475 3476 check_class_changed(rq, p, prev_class, oldprio); 3477 task_rq_unlock(rq, p, &flags); 3478 3479 rt_mutex_adjust_pi(p); 3480 3481 return 0; 3482} 3483 3484static int _sched_setscheduler(struct task_struct *p, int policy, 3485 const struct sched_param *param, bool check) 3486{ 3487 struct sched_attr attr = { 3488 .sched_policy = policy, 3489 .sched_priority = param->sched_priority, 3490 .sched_nice = PRIO_TO_NICE(p->static_prio), 3491 }; 3492 3493 /* 3494 * Fixup the legacy SCHED_RESET_ON_FORK hack 3495 */ 3496 if (policy & SCHED_RESET_ON_FORK) { 3497 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3498 policy &= ~SCHED_RESET_ON_FORK; 3499 attr.sched_policy = policy; 3500 } 3501 3502 return __sched_setscheduler(p, &attr, check); 3503} 3504/** 3505 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3506 * @p: the task in question. 3507 * @policy: new policy. 3508 * @param: structure containing the new RT priority. 3509 * 3510 * Return: 0 on success. An error code otherwise. 3511 * 3512 * NOTE that the task may be already dead. 3513 */ 3514int sched_setscheduler(struct task_struct *p, int policy, 3515 const struct sched_param *param) 3516{ 3517 return _sched_setscheduler(p, policy, param, true); 3518} 3519EXPORT_SYMBOL_GPL(sched_setscheduler); 3520 3521int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3522{ 3523 return __sched_setscheduler(p, attr, true); 3524} 3525EXPORT_SYMBOL_GPL(sched_setattr); 3526 3527/** 3528 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3529 * @p: the task in question. 3530 * @policy: new policy. 3531 * @param: structure containing the new RT priority. 3532 * 3533 * Just like sched_setscheduler, only don't bother checking if the 3534 * current context has permission. For example, this is needed in 3535 * stop_machine(): we create temporary high priority worker threads, 3536 * but our caller might not have that capability. 3537 * 3538 * Return: 0 on success. An error code otherwise. 3539 */ 3540int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3541 const struct sched_param *param) 3542{ 3543 return _sched_setscheduler(p, policy, param, false); 3544} 3545 3546static int 3547do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3548{ 3549 struct sched_param lparam; 3550 struct task_struct *p; 3551 int retval; 3552 3553 if (!param || pid < 0) 3554 return -EINVAL; 3555 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3556 return -EFAULT; 3557 3558 rcu_read_lock(); 3559 retval = -ESRCH; 3560 p = find_process_by_pid(pid); 3561 if (p != NULL) 3562 retval = sched_setscheduler(p, policy, &lparam); 3563 rcu_read_unlock(); 3564 3565 return retval; 3566} 3567 3568/* 3569 * Mimics kernel/events/core.c perf_copy_attr(). 3570 */ 3571static int sched_copy_attr(struct sched_attr __user *uattr, 3572 struct sched_attr *attr) 3573{ 3574 u32 size; 3575 int ret; 3576 3577 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3578 return -EFAULT; 3579 3580 /* 3581 * zero the full structure, so that a short copy will be nice. 3582 */ 3583 memset(attr, 0, sizeof(*attr)); 3584 3585 ret = get_user(size, &uattr->size); 3586 if (ret) 3587 return ret; 3588 3589 if (size > PAGE_SIZE) /* silly large */ 3590 goto err_size; 3591 3592 if (!size) /* abi compat */ 3593 size = SCHED_ATTR_SIZE_VER0; 3594 3595 if (size < SCHED_ATTR_SIZE_VER0) 3596 goto err_size; 3597 3598 /* 3599 * If we're handed a bigger struct than we know of, 3600 * ensure all the unknown bits are 0 - i.e. new 3601 * user-space does not rely on any kernel feature 3602 * extensions we dont know about yet. 3603 */ 3604 if (size > sizeof(*attr)) { 3605 unsigned char __user *addr; 3606 unsigned char __user *end; 3607 unsigned char val; 3608 3609 addr = (void __user *)uattr + sizeof(*attr); 3610 end = (void __user *)uattr + size; 3611 3612 for (; addr < end; addr++) { 3613 ret = get_user(val, addr); 3614 if (ret) 3615 return ret; 3616 if (val) 3617 goto err_size; 3618 } 3619 size = sizeof(*attr); 3620 } 3621 3622 ret = copy_from_user(attr, uattr, size); 3623 if (ret) 3624 return -EFAULT; 3625 3626 /* 3627 * XXX: do we want to be lenient like existing syscalls; or do we want 3628 * to be strict and return an error on out-of-bounds values? 3629 */ 3630 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3631 3632out: 3633 return ret; 3634 3635err_size: 3636 put_user(sizeof(*attr), &uattr->size); 3637 ret = -E2BIG; 3638 goto out; 3639} 3640 3641/** 3642 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3643 * @pid: the pid in question. 3644 * @policy: new policy. 3645 * @param: structure containing the new RT priority. 3646 * 3647 * Return: 0 on success. An error code otherwise. 3648 */ 3649SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3650 struct sched_param __user *, param) 3651{ 3652 /* negative values for policy are not valid */ 3653 if (policy < 0) 3654 return -EINVAL; 3655 3656 return do_sched_setscheduler(pid, policy, param); 3657} 3658 3659/** 3660 * sys_sched_setparam - set/change the RT priority of a thread 3661 * @pid: the pid in question. 3662 * @param: structure containing the new RT priority. 3663 * 3664 * Return: 0 on success. An error code otherwise. 3665 */ 3666SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3667{ 3668 return do_sched_setscheduler(pid, -1, param); 3669} 3670 3671/** 3672 * sys_sched_setattr - same as above, but with extended sched_attr 3673 * @pid: the pid in question. 3674 * @uattr: structure containing the extended parameters. 3675 */ 3676SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr) 3677{ 3678 struct sched_attr attr; 3679 struct task_struct *p; 3680 int retval; 3681 3682 if (!uattr || pid < 0) 3683 return -EINVAL; 3684 3685 if (sched_copy_attr(uattr, &attr)) 3686 return -EFAULT; 3687 3688 rcu_read_lock(); 3689 retval = -ESRCH; 3690 p = find_process_by_pid(pid); 3691 if (p != NULL) 3692 retval = sched_setattr(p, &attr); 3693 rcu_read_unlock(); 3694 3695 return retval; 3696} 3697 3698/** 3699 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3700 * @pid: the pid in question. 3701 * 3702 * Return: On success, the policy of the thread. Otherwise, a negative error 3703 * code. 3704 */ 3705SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3706{ 3707 struct task_struct *p; 3708 int retval; 3709 3710 if (pid < 0) 3711 return -EINVAL; 3712 3713 retval = -ESRCH; 3714 rcu_read_lock(); 3715 p = find_process_by_pid(pid); 3716 if (p) { 3717 retval = security_task_getscheduler(p); 3718 if (!retval) 3719 retval = p->policy 3720 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3721 } 3722 rcu_read_unlock(); 3723 return retval; 3724} 3725 3726/** 3727 * sys_sched_getparam - get the RT priority of a thread 3728 * @pid: the pid in question. 3729 * @param: structure containing the RT priority. 3730 * 3731 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3732 * code. 3733 */ 3734SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3735{ 3736 struct sched_param lp; 3737 struct task_struct *p; 3738 int retval; 3739 3740 if (!param || pid < 0) 3741 return -EINVAL; 3742 3743 rcu_read_lock(); 3744 p = find_process_by_pid(pid); 3745 retval = -ESRCH; 3746 if (!p) 3747 goto out_unlock; 3748 3749 retval = security_task_getscheduler(p); 3750 if (retval) 3751 goto out_unlock; 3752 3753 if (task_has_dl_policy(p)) { 3754 retval = -EINVAL; 3755 goto out_unlock; 3756 } 3757 lp.sched_priority = p->rt_priority; 3758 rcu_read_unlock(); 3759 3760 /* 3761 * This one might sleep, we cannot do it with a spinlock held ... 3762 */ 3763 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3764 3765 return retval; 3766 3767out_unlock: 3768 rcu_read_unlock(); 3769 return retval; 3770} 3771 3772static int sched_read_attr(struct sched_attr __user *uattr, 3773 struct sched_attr *attr, 3774 unsigned int usize) 3775{ 3776 int ret; 3777 3778 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3779 return -EFAULT; 3780 3781 /* 3782 * If we're handed a smaller struct than we know of, 3783 * ensure all the unknown bits are 0 - i.e. old 3784 * user-space does not get uncomplete information. 3785 */ 3786 if (usize < sizeof(*attr)) { 3787 unsigned char *addr; 3788 unsigned char *end; 3789 3790 addr = (void *)attr + usize; 3791 end = (void *)attr + sizeof(*attr); 3792 3793 for (; addr < end; addr++) { 3794 if (*addr) 3795 goto err_size; 3796 } 3797 3798 attr->size = usize; 3799 } 3800 3801 ret = copy_to_user(uattr, attr, usize); 3802 if (ret) 3803 return -EFAULT; 3804 3805out: 3806 return ret; 3807 3808err_size: 3809 ret = -E2BIG; 3810 goto out; 3811} 3812 3813/** 3814 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3815 * @pid: the pid in question. 3816 * @uattr: structure containing the extended parameters. 3817 * @size: sizeof(attr) for fwd/bwd comp. 3818 */ 3819SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3820 unsigned int, size) 3821{ 3822 struct sched_attr attr = { 3823 .size = sizeof(struct sched_attr), 3824 }; 3825 struct task_struct *p; 3826 int retval; 3827 3828 if (!uattr || pid < 0 || size > PAGE_SIZE || 3829 size < SCHED_ATTR_SIZE_VER0) 3830 return -EINVAL; 3831 3832 rcu_read_lock(); 3833 p = find_process_by_pid(pid); 3834 retval = -ESRCH; 3835 if (!p) 3836 goto out_unlock; 3837 3838 retval = security_task_getscheduler(p); 3839 if (retval) 3840 goto out_unlock; 3841 3842 attr.sched_policy = p->policy; 3843 if (p->sched_reset_on_fork) 3844 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3845 if (task_has_dl_policy(p)) 3846 __getparam_dl(p, &attr); 3847 else if (task_has_rt_policy(p)) 3848 attr.sched_priority = p->rt_priority; 3849 else 3850 attr.sched_nice = task_nice(p); 3851 3852 rcu_read_unlock(); 3853 3854 retval = sched_read_attr(uattr, &attr, size); 3855 return retval; 3856 3857out_unlock: 3858 rcu_read_unlock(); 3859 return retval; 3860} 3861 3862long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3863{ 3864 cpumask_var_t cpus_allowed, new_mask; 3865 struct task_struct *p; 3866 int retval; 3867 3868 rcu_read_lock(); 3869 3870 p = find_process_by_pid(pid); 3871 if (!p) { 3872 rcu_read_unlock(); 3873 return -ESRCH; 3874 } 3875 3876 /* Prevent p going away */ 3877 get_task_struct(p); 3878 rcu_read_unlock(); 3879 3880 if (p->flags & PF_NO_SETAFFINITY) { 3881 retval = -EINVAL; 3882 goto out_put_task; 3883 } 3884 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3885 retval = -ENOMEM; 3886 goto out_put_task; 3887 } 3888 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3889 retval = -ENOMEM; 3890 goto out_free_cpus_allowed; 3891 } 3892 retval = -EPERM; 3893 if (!check_same_owner(p)) { 3894 rcu_read_lock(); 3895 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3896 rcu_read_unlock(); 3897 goto out_unlock; 3898 } 3899 rcu_read_unlock(); 3900 } 3901 3902 retval = security_task_setscheduler(p); 3903 if (retval) 3904 goto out_unlock; 3905 3906 3907 cpuset_cpus_allowed(p, cpus_allowed); 3908 cpumask_and(new_mask, in_mask, cpus_allowed); 3909 3910 /* 3911 * Since bandwidth control happens on root_domain basis, 3912 * if admission test is enabled, we only admit -deadline 3913 * tasks allowed to run on all the CPUs in the task's 3914 * root_domain. 3915 */ 3916#ifdef CONFIG_SMP 3917 if (task_has_dl_policy(p)) { 3918 const struct cpumask *span = task_rq(p)->rd->span; 3919 3920 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3921 retval = -EBUSY; 3922 goto out_unlock; 3923 } 3924 } 3925#endif 3926again: 3927 retval = set_cpus_allowed_ptr(p, new_mask); 3928 3929 if (!retval) { 3930 cpuset_cpus_allowed(p, cpus_allowed); 3931 if (!cpumask_subset(new_mask, cpus_allowed)) { 3932 /* 3933 * We must have raced with a concurrent cpuset 3934 * update. Just reset the cpus_allowed to the 3935 * cpuset's cpus_allowed 3936 */ 3937 cpumask_copy(new_mask, cpus_allowed); 3938 goto again; 3939 } 3940 } 3941out_unlock: 3942 free_cpumask_var(new_mask); 3943out_free_cpus_allowed: 3944 free_cpumask_var(cpus_allowed); 3945out_put_task: 3946 put_task_struct(p); 3947 return retval; 3948} 3949 3950static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3951 struct cpumask *new_mask) 3952{ 3953 if (len < cpumask_size()) 3954 cpumask_clear(new_mask); 3955 else if (len > cpumask_size()) 3956 len = cpumask_size(); 3957 3958 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3959} 3960 3961/** 3962 * sys_sched_setaffinity - set the cpu affinity of a process 3963 * @pid: pid of the process 3964 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3965 * @user_mask_ptr: user-space pointer to the new cpu mask 3966 * 3967 * Return: 0 on success. An error code otherwise. 3968 */ 3969SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3970 unsigned long __user *, user_mask_ptr) 3971{ 3972 cpumask_var_t new_mask; 3973 int retval; 3974 3975 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3976 return -ENOMEM; 3977 3978 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3979 if (retval == 0) 3980 retval = sched_setaffinity(pid, new_mask); 3981 free_cpumask_var(new_mask); 3982 return retval; 3983} 3984 3985long sched_getaffinity(pid_t pid, struct cpumask *mask) 3986{ 3987 struct task_struct *p; 3988 unsigned long flags; 3989 int retval; 3990 3991 rcu_read_lock(); 3992 3993 retval = -ESRCH; 3994 p = find_process_by_pid(pid); 3995 if (!p) 3996 goto out_unlock; 3997 3998 retval = security_task_getscheduler(p); 3999 if (retval) 4000 goto out_unlock; 4001 4002 raw_spin_lock_irqsave(&p->pi_lock, flags); 4003 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4004 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4005 4006out_unlock: 4007 rcu_read_unlock(); 4008 4009 return retval; 4010} 4011 4012/** 4013 * sys_sched_getaffinity - get the cpu affinity of a process 4014 * @pid: pid of the process 4015 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4016 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4017 * 4018 * Return: 0 on success. An error code otherwise. 4019 */ 4020SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4021 unsigned long __user *, user_mask_ptr) 4022{ 4023 int ret; 4024 cpumask_var_t mask; 4025 4026 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4027 return -EINVAL; 4028 if (len & (sizeof(unsigned long)-1)) 4029 return -EINVAL; 4030 4031 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4032 return -ENOMEM; 4033 4034 ret = sched_getaffinity(pid, mask); 4035 if (ret == 0) { 4036 size_t retlen = min_t(size_t, len, cpumask_size()); 4037 4038 if (copy_to_user(user_mask_ptr, mask, retlen)) 4039 ret = -EFAULT; 4040 else 4041 ret = retlen; 4042 } 4043 free_cpumask_var(mask); 4044 4045 return ret; 4046} 4047 4048/** 4049 * sys_sched_yield - yield the current processor to other threads. 4050 * 4051 * This function yields the current CPU to other tasks. If there are no 4052 * other threads running on this CPU then this function will return. 4053 * 4054 * Return: 0. 4055 */ 4056SYSCALL_DEFINE0(sched_yield) 4057{ 4058 struct rq *rq = this_rq_lock(); 4059 4060 schedstat_inc(rq, yld_count); 4061 current->sched_class->yield_task(rq); 4062 4063 /* 4064 * Since we are going to call schedule() anyway, there's 4065 * no need to preempt or enable interrupts: 4066 */ 4067 __release(rq->lock); 4068 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4069 do_raw_spin_unlock(&rq->lock); 4070 sched_preempt_enable_no_resched(); 4071 4072 schedule(); 4073 4074 return 0; 4075} 4076 4077static void __cond_resched(void) 4078{ 4079 __preempt_count_add(PREEMPT_ACTIVE); 4080 __schedule(); 4081 __preempt_count_sub(PREEMPT_ACTIVE); 4082} 4083 4084int __sched _cond_resched(void) 4085{ 4086 if (should_resched()) { 4087 __cond_resched(); 4088 return 1; 4089 } 4090 return 0; 4091} 4092EXPORT_SYMBOL(_cond_resched); 4093 4094/* 4095 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4096 * call schedule, and on return reacquire the lock. 4097 * 4098 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4099 * operations here to prevent schedule() from being called twice (once via 4100 * spin_unlock(), once by hand). 4101 */ 4102int __cond_resched_lock(spinlock_t *lock) 4103{ 4104 int resched = should_resched(); 4105 int ret = 0; 4106 4107 lockdep_assert_held(lock); 4108 4109 if (spin_needbreak(lock) || resched) { 4110 spin_unlock(lock); 4111 if (resched) 4112 __cond_resched(); 4113 else 4114 cpu_relax(); 4115 ret = 1; 4116 spin_lock(lock); 4117 } 4118 return ret; 4119} 4120EXPORT_SYMBOL(__cond_resched_lock); 4121 4122int __sched __cond_resched_softirq(void) 4123{ 4124 BUG_ON(!in_softirq()); 4125 4126 if (should_resched()) { 4127 local_bh_enable(); 4128 __cond_resched(); 4129 local_bh_disable(); 4130 return 1; 4131 } 4132 return 0; 4133} 4134EXPORT_SYMBOL(__cond_resched_softirq); 4135 4136/** 4137 * yield - yield the current processor to other threads. 4138 * 4139 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4140 * 4141 * The scheduler is at all times free to pick the calling task as the most 4142 * eligible task to run, if removing the yield() call from your code breaks 4143 * it, its already broken. 4144 * 4145 * Typical broken usage is: 4146 * 4147 * while (!event) 4148 * yield(); 4149 * 4150 * where one assumes that yield() will let 'the other' process run that will 4151 * make event true. If the current task is a SCHED_FIFO task that will never 4152 * happen. Never use yield() as a progress guarantee!! 4153 * 4154 * If you want to use yield() to wait for something, use wait_event(). 4155 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4156 * If you still want to use yield(), do not! 4157 */ 4158void __sched yield(void) 4159{ 4160 set_current_state(TASK_RUNNING); 4161 sys_sched_yield(); 4162} 4163EXPORT_SYMBOL(yield); 4164 4165/** 4166 * yield_to - yield the current processor to another thread in 4167 * your thread group, or accelerate that thread toward the 4168 * processor it's on. 4169 * @p: target task 4170 * @preempt: whether task preemption is allowed or not 4171 * 4172 * It's the caller's job to ensure that the target task struct 4173 * can't go away on us before we can do any checks. 4174 * 4175 * Return: 4176 * true (>0) if we indeed boosted the target task. 4177 * false (0) if we failed to boost the target. 4178 * -ESRCH if there's no task to yield to. 4179 */ 4180bool __sched yield_to(struct task_struct *p, bool preempt) 4181{ 4182 struct task_struct *curr = current; 4183 struct rq *rq, *p_rq; 4184 unsigned long flags; 4185 int yielded = 0; 4186 4187 local_irq_save(flags); 4188 rq = this_rq(); 4189 4190again: 4191 p_rq = task_rq(p); 4192 /* 4193 * If we're the only runnable task on the rq and target rq also 4194 * has only one task, there's absolutely no point in yielding. 4195 */ 4196 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4197 yielded = -ESRCH; 4198 goto out_irq; 4199 } 4200 4201 double_rq_lock(rq, p_rq); 4202 if (task_rq(p) != p_rq) { 4203 double_rq_unlock(rq, p_rq); 4204 goto again; 4205 } 4206 4207 if (!curr->sched_class->yield_to_task) 4208 goto out_unlock; 4209 4210 if (curr->sched_class != p->sched_class) 4211 goto out_unlock; 4212 4213 if (task_running(p_rq, p) || p->state) 4214 goto out_unlock; 4215 4216 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4217 if (yielded) { 4218 schedstat_inc(rq, yld_count); 4219 /* 4220 * Make p's CPU reschedule; pick_next_entity takes care of 4221 * fairness. 4222 */ 4223 if (preempt && rq != p_rq) 4224 resched_task(p_rq->curr); 4225 } 4226 4227out_unlock: 4228 double_rq_unlock(rq, p_rq); 4229out_irq: 4230 local_irq_restore(flags); 4231 4232 if (yielded > 0) 4233 schedule(); 4234 4235 return yielded; 4236} 4237EXPORT_SYMBOL_GPL(yield_to); 4238 4239/* 4240 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4241 * that process accounting knows that this is a task in IO wait state. 4242 */ 4243void __sched io_schedule(void) 4244{ 4245 struct rq *rq = raw_rq(); 4246 4247 delayacct_blkio_start(); 4248 atomic_inc(&rq->nr_iowait); 4249 blk_flush_plug(current); 4250 current->in_iowait = 1; 4251 schedule(); 4252 current->in_iowait = 0; 4253 atomic_dec(&rq->nr_iowait); 4254 delayacct_blkio_end(); 4255} 4256EXPORT_SYMBOL(io_schedule); 4257 4258long __sched io_schedule_timeout(long timeout) 4259{ 4260 struct rq *rq = raw_rq(); 4261 long ret; 4262 4263 delayacct_blkio_start(); 4264 atomic_inc(&rq->nr_iowait); 4265 blk_flush_plug(current); 4266 current->in_iowait = 1; 4267 ret = schedule_timeout(timeout); 4268 current->in_iowait = 0; 4269 atomic_dec(&rq->nr_iowait); 4270 delayacct_blkio_end(); 4271 return ret; 4272} 4273 4274/** 4275 * sys_sched_get_priority_max - return maximum RT priority. 4276 * @policy: scheduling class. 4277 * 4278 * Return: On success, this syscall returns the maximum 4279 * rt_priority that can be used by a given scheduling class. 4280 * On failure, a negative error code is returned. 4281 */ 4282SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4283{ 4284 int ret = -EINVAL; 4285 4286 switch (policy) { 4287 case SCHED_FIFO: 4288 case SCHED_RR: 4289 ret = MAX_USER_RT_PRIO-1; 4290 break; 4291 case SCHED_DEADLINE: 4292 case SCHED_NORMAL: 4293 case SCHED_BATCH: 4294 case SCHED_IDLE: 4295 ret = 0; 4296 break; 4297 } 4298 return ret; 4299} 4300 4301/** 4302 * sys_sched_get_priority_min - return minimum RT priority. 4303 * @policy: scheduling class. 4304 * 4305 * Return: On success, this syscall returns the minimum 4306 * rt_priority that can be used by a given scheduling class. 4307 * On failure, a negative error code is returned. 4308 */ 4309SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4310{ 4311 int ret = -EINVAL; 4312 4313 switch (policy) { 4314 case SCHED_FIFO: 4315 case SCHED_RR: 4316 ret = 1; 4317 break; 4318 case SCHED_DEADLINE: 4319 case SCHED_NORMAL: 4320 case SCHED_BATCH: 4321 case SCHED_IDLE: 4322 ret = 0; 4323 } 4324 return ret; 4325} 4326 4327/** 4328 * sys_sched_rr_get_interval - return the default timeslice of a process. 4329 * @pid: pid of the process. 4330 * @interval: userspace pointer to the timeslice value. 4331 * 4332 * this syscall writes the default timeslice value of a given process 4333 * into the user-space timespec buffer. A value of '0' means infinity. 4334 * 4335 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4336 * an error code. 4337 */ 4338SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4339 struct timespec __user *, interval) 4340{ 4341 struct task_struct *p; 4342 unsigned int time_slice; 4343 unsigned long flags; 4344 struct rq *rq; 4345 int retval; 4346 struct timespec t; 4347 4348 if (pid < 0) 4349 return -EINVAL; 4350 4351 retval = -ESRCH; 4352 rcu_read_lock(); 4353 p = find_process_by_pid(pid); 4354 if (!p) 4355 goto out_unlock; 4356 4357 retval = security_task_getscheduler(p); 4358 if (retval) 4359 goto out_unlock; 4360 4361 rq = task_rq_lock(p, &flags); 4362 time_slice = 0; 4363 if (p->sched_class->get_rr_interval) 4364 time_slice = p->sched_class->get_rr_interval(rq, p); 4365 task_rq_unlock(rq, p, &flags); 4366 4367 rcu_read_unlock(); 4368 jiffies_to_timespec(time_slice, &t); 4369 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4370 return retval; 4371 4372out_unlock: 4373 rcu_read_unlock(); 4374 return retval; 4375} 4376 4377static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4378 4379void sched_show_task(struct task_struct *p) 4380{ 4381 unsigned long free = 0; 4382 int ppid; 4383 unsigned state; 4384 4385 state = p->state ? __ffs(p->state) + 1 : 0; 4386 printk(KERN_INFO "%-15.15s %c", p->comm, 4387 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4388#if BITS_PER_LONG == 32 4389 if (state == TASK_RUNNING) 4390 printk(KERN_CONT " running "); 4391 else 4392 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4393#else 4394 if (state == TASK_RUNNING) 4395 printk(KERN_CONT " running task "); 4396 else 4397 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4398#endif 4399#ifdef CONFIG_DEBUG_STACK_USAGE 4400 free = stack_not_used(p); 4401#endif 4402 rcu_read_lock(); 4403 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4404 rcu_read_unlock(); 4405 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4406 task_pid_nr(p), ppid, 4407 (unsigned long)task_thread_info(p)->flags); 4408 4409 print_worker_info(KERN_INFO, p); 4410 show_stack(p, NULL); 4411} 4412 4413void show_state_filter(unsigned long state_filter) 4414{ 4415 struct task_struct *g, *p; 4416 4417#if BITS_PER_LONG == 32 4418 printk(KERN_INFO 4419 " task PC stack pid father\n"); 4420#else 4421 printk(KERN_INFO 4422 " task PC stack pid father\n"); 4423#endif 4424 rcu_read_lock(); 4425 do_each_thread(g, p) { 4426 /* 4427 * reset the NMI-timeout, listing all files on a slow 4428 * console might take a lot of time: 4429 */ 4430 touch_nmi_watchdog(); 4431 if (!state_filter || (p->state & state_filter)) 4432 sched_show_task(p); 4433 } while_each_thread(g, p); 4434 4435 touch_all_softlockup_watchdogs(); 4436 4437#ifdef CONFIG_SCHED_DEBUG 4438 sysrq_sched_debug_show(); 4439#endif 4440 rcu_read_unlock(); 4441 /* 4442 * Only show locks if all tasks are dumped: 4443 */ 4444 if (!state_filter) 4445 debug_show_all_locks(); 4446} 4447 4448void init_idle_bootup_task(struct task_struct *idle) 4449{ 4450 idle->sched_class = &idle_sched_class; 4451} 4452 4453/** 4454 * init_idle - set up an idle thread for a given CPU 4455 * @idle: task in question 4456 * @cpu: cpu the idle task belongs to 4457 * 4458 * NOTE: this function does not set the idle thread's NEED_RESCHED 4459 * flag, to make booting more robust. 4460 */ 4461void init_idle(struct task_struct *idle, int cpu) 4462{ 4463 struct rq *rq = cpu_rq(cpu); 4464 unsigned long flags; 4465 4466 raw_spin_lock_irqsave(&rq->lock, flags); 4467 4468 __sched_fork(0, idle); 4469 idle->state = TASK_RUNNING; 4470 idle->se.exec_start = sched_clock(); 4471 4472 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4473 /* 4474 * We're having a chicken and egg problem, even though we are 4475 * holding rq->lock, the cpu isn't yet set to this cpu so the 4476 * lockdep check in task_group() will fail. 4477 * 4478 * Similar case to sched_fork(). / Alternatively we could 4479 * use task_rq_lock() here and obtain the other rq->lock. 4480 * 4481 * Silence PROVE_RCU 4482 */ 4483 rcu_read_lock(); 4484 __set_task_cpu(idle, cpu); 4485 rcu_read_unlock(); 4486 4487 rq->curr = rq->idle = idle; 4488 idle->on_rq = 1; 4489#if defined(CONFIG_SMP) 4490 idle->on_cpu = 1; 4491#endif 4492 raw_spin_unlock_irqrestore(&rq->lock, flags); 4493 4494 /* Set the preempt count _outside_ the spinlocks! */ 4495 init_idle_preempt_count(idle, cpu); 4496 4497 /* 4498 * The idle tasks have their own, simple scheduling class: 4499 */ 4500 idle->sched_class = &idle_sched_class; 4501 ftrace_graph_init_idle_task(idle, cpu); 4502 vtime_init_idle(idle, cpu); 4503#if defined(CONFIG_SMP) 4504 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4505#endif 4506} 4507 4508#ifdef CONFIG_SMP 4509void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4510{ 4511 if (p->sched_class && p->sched_class->set_cpus_allowed) 4512 p->sched_class->set_cpus_allowed(p, new_mask); 4513 4514 cpumask_copy(&p->cpus_allowed, new_mask); 4515 p->nr_cpus_allowed = cpumask_weight(new_mask); 4516} 4517 4518/* 4519 * This is how migration works: 4520 * 4521 * 1) we invoke migration_cpu_stop() on the target CPU using 4522 * stop_one_cpu(). 4523 * 2) stopper starts to run (implicitly forcing the migrated thread 4524 * off the CPU) 4525 * 3) it checks whether the migrated task is still in the wrong runqueue. 4526 * 4) if it's in the wrong runqueue then the migration thread removes 4527 * it and puts it into the right queue. 4528 * 5) stopper completes and stop_one_cpu() returns and the migration 4529 * is done. 4530 */ 4531 4532/* 4533 * Change a given task's CPU affinity. Migrate the thread to a 4534 * proper CPU and schedule it away if the CPU it's executing on 4535 * is removed from the allowed bitmask. 4536 * 4537 * NOTE: the caller must have a valid reference to the task, the 4538 * task must not exit() & deallocate itself prematurely. The 4539 * call is not atomic; no spinlocks may be held. 4540 */ 4541int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4542{ 4543 unsigned long flags; 4544 struct rq *rq; 4545 unsigned int dest_cpu; 4546 int ret = 0; 4547 4548 rq = task_rq_lock(p, &flags); 4549 4550 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4551 goto out; 4552 4553 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4554 ret = -EINVAL; 4555 goto out; 4556 } 4557 4558 do_set_cpus_allowed(p, new_mask); 4559 4560 /* Can the task run on the task's current CPU? If so, we're done */ 4561 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4562 goto out; 4563 4564 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4565 if (p->on_rq) { 4566 struct migration_arg arg = { p, dest_cpu }; 4567 /* Need help from migration thread: drop lock and wait. */ 4568 task_rq_unlock(rq, p, &flags); 4569 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4570 tlb_migrate_finish(p->mm); 4571 return 0; 4572 } 4573out: 4574 task_rq_unlock(rq, p, &flags); 4575 4576 return ret; 4577} 4578EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4579 4580/* 4581 * Move (not current) task off this cpu, onto dest cpu. We're doing 4582 * this because either it can't run here any more (set_cpus_allowed() 4583 * away from this CPU, or CPU going down), or because we're 4584 * attempting to rebalance this task on exec (sched_exec). 4585 * 4586 * So we race with normal scheduler movements, but that's OK, as long 4587 * as the task is no longer on this CPU. 4588 * 4589 * Returns non-zero if task was successfully migrated. 4590 */ 4591static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4592{ 4593 struct rq *rq_dest, *rq_src; 4594 int ret = 0; 4595 4596 if (unlikely(!cpu_active(dest_cpu))) 4597 return ret; 4598 4599 rq_src = cpu_rq(src_cpu); 4600 rq_dest = cpu_rq(dest_cpu); 4601 4602 raw_spin_lock(&p->pi_lock); 4603 double_rq_lock(rq_src, rq_dest); 4604 /* Already moved. */ 4605 if (task_cpu(p) != src_cpu) 4606 goto done; 4607 /* Affinity changed (again). */ 4608 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4609 goto fail; 4610 4611 /* 4612 * If we're not on a rq, the next wake-up will ensure we're 4613 * placed properly. 4614 */ 4615 if (p->on_rq) { 4616 dequeue_task(rq_src, p, 0); 4617 set_task_cpu(p, dest_cpu); 4618 enqueue_task(rq_dest, p, 0); 4619 check_preempt_curr(rq_dest, p, 0); 4620 } 4621done: 4622 ret = 1; 4623fail: 4624 double_rq_unlock(rq_src, rq_dest); 4625 raw_spin_unlock(&p->pi_lock); 4626 return ret; 4627} 4628 4629#ifdef CONFIG_NUMA_BALANCING 4630/* Migrate current task p to target_cpu */ 4631int migrate_task_to(struct task_struct *p, int target_cpu) 4632{ 4633 struct migration_arg arg = { p, target_cpu }; 4634 int curr_cpu = task_cpu(p); 4635 4636 if (curr_cpu == target_cpu) 4637 return 0; 4638 4639 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4640 return -EINVAL; 4641 4642 /* TODO: This is not properly updating schedstats */ 4643 4644 trace_sched_move_numa(p, curr_cpu, target_cpu); 4645 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4646} 4647 4648/* 4649 * Requeue a task on a given node and accurately track the number of NUMA 4650 * tasks on the runqueues 4651 */ 4652void sched_setnuma(struct task_struct *p, int nid) 4653{ 4654 struct rq *rq; 4655 unsigned long flags; 4656 bool on_rq, running; 4657 4658 rq = task_rq_lock(p, &flags); 4659 on_rq = p->on_rq; 4660 running = task_current(rq, p); 4661 4662 if (on_rq) 4663 dequeue_task(rq, p, 0); 4664 if (running) 4665 p->sched_class->put_prev_task(rq, p); 4666 4667 p->numa_preferred_nid = nid; 4668 4669 if (running) 4670 p->sched_class->set_curr_task(rq); 4671 if (on_rq) 4672 enqueue_task(rq, p, 0); 4673 task_rq_unlock(rq, p, &flags); 4674} 4675#endif 4676 4677/* 4678 * migration_cpu_stop - this will be executed by a highprio stopper thread 4679 * and performs thread migration by bumping thread off CPU then 4680 * 'pushing' onto another runqueue. 4681 */ 4682static int migration_cpu_stop(void *data) 4683{ 4684 struct migration_arg *arg = data; 4685 4686 /* 4687 * The original target cpu might have gone down and we might 4688 * be on another cpu but it doesn't matter. 4689 */ 4690 local_irq_disable(); 4691 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4692 local_irq_enable(); 4693 return 0; 4694} 4695 4696#ifdef CONFIG_HOTPLUG_CPU 4697 4698/* 4699 * Ensures that the idle task is using init_mm right before its cpu goes 4700 * offline. 4701 */ 4702void idle_task_exit(void) 4703{ 4704 struct mm_struct *mm = current->active_mm; 4705 4706 BUG_ON(cpu_online(smp_processor_id())); 4707 4708 if (mm != &init_mm) 4709 switch_mm(mm, &init_mm, current); 4710 mmdrop(mm); 4711} 4712 4713/* 4714 * Since this CPU is going 'away' for a while, fold any nr_active delta 4715 * we might have. Assumes we're called after migrate_tasks() so that the 4716 * nr_active count is stable. 4717 * 4718 * Also see the comment "Global load-average calculations". 4719 */ 4720static void calc_load_migrate(struct rq *rq) 4721{ 4722 long delta = calc_load_fold_active(rq); 4723 if (delta) 4724 atomic_long_add(delta, &calc_load_tasks); 4725} 4726 4727static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4728{ 4729} 4730 4731static const struct sched_class fake_sched_class = { 4732 .put_prev_task = put_prev_task_fake, 4733}; 4734 4735static struct task_struct fake_task = { 4736 /* 4737 * Avoid pull_{rt,dl}_task() 4738 */ 4739 .prio = MAX_PRIO + 1, 4740 .sched_class = &fake_sched_class, 4741}; 4742 4743/* 4744 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4745 * try_to_wake_up()->select_task_rq(). 4746 * 4747 * Called with rq->lock held even though we'er in stop_machine() and 4748 * there's no concurrency possible, we hold the required locks anyway 4749 * because of lock validation efforts. 4750 */ 4751static void migrate_tasks(unsigned int dead_cpu) 4752{ 4753 struct rq *rq = cpu_rq(dead_cpu); 4754 struct task_struct *next, *stop = rq->stop; 4755 int dest_cpu; 4756 4757 /* 4758 * Fudge the rq selection such that the below task selection loop 4759 * doesn't get stuck on the currently eligible stop task. 4760 * 4761 * We're currently inside stop_machine() and the rq is either stuck 4762 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4763 * either way we should never end up calling schedule() until we're 4764 * done here. 4765 */ 4766 rq->stop = NULL; 4767 4768 /* 4769 * put_prev_task() and pick_next_task() sched 4770 * class method both need to have an up-to-date 4771 * value of rq->clock[_task] 4772 */ 4773 update_rq_clock(rq); 4774 4775 for ( ; ; ) { 4776 /* 4777 * There's this thread running, bail when that's the only 4778 * remaining thread. 4779 */ 4780 if (rq->nr_running == 1) 4781 break; 4782 4783 next = pick_next_task(rq, &fake_task); 4784 BUG_ON(!next); 4785 next->sched_class->put_prev_task(rq, next); 4786 4787 /* Find suitable destination for @next, with force if needed. */ 4788 dest_cpu = select_fallback_rq(dead_cpu, next); 4789 raw_spin_unlock(&rq->lock); 4790 4791 __migrate_task(next, dead_cpu, dest_cpu); 4792 4793 raw_spin_lock(&rq->lock); 4794 } 4795 4796 rq->stop = stop; 4797} 4798 4799#endif /* CONFIG_HOTPLUG_CPU */ 4800 4801#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4802 4803static struct ctl_table sd_ctl_dir[] = { 4804 { 4805 .procname = "sched_domain", 4806 .mode = 0555, 4807 }, 4808 {} 4809}; 4810 4811static struct ctl_table sd_ctl_root[] = { 4812 { 4813 .procname = "kernel", 4814 .mode = 0555, 4815 .child = sd_ctl_dir, 4816 }, 4817 {} 4818}; 4819 4820static struct ctl_table *sd_alloc_ctl_entry(int n) 4821{ 4822 struct ctl_table *entry = 4823 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4824 4825 return entry; 4826} 4827 4828static void sd_free_ctl_entry(struct ctl_table **tablep) 4829{ 4830 struct ctl_table *entry; 4831 4832 /* 4833 * In the intermediate directories, both the child directory and 4834 * procname are dynamically allocated and could fail but the mode 4835 * will always be set. In the lowest directory the names are 4836 * static strings and all have proc handlers. 4837 */ 4838 for (entry = *tablep; entry->mode; entry++) { 4839 if (entry->child) 4840 sd_free_ctl_entry(&entry->child); 4841 if (entry->proc_handler == NULL) 4842 kfree(entry->procname); 4843 } 4844 4845 kfree(*tablep); 4846 *tablep = NULL; 4847} 4848 4849static int min_load_idx = 0; 4850static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4851 4852static void 4853set_table_entry(struct ctl_table *entry, 4854 const char *procname, void *data, int maxlen, 4855 umode_t mode, proc_handler *proc_handler, 4856 bool load_idx) 4857{ 4858 entry->procname = procname; 4859 entry->data = data; 4860 entry->maxlen = maxlen; 4861 entry->mode = mode; 4862 entry->proc_handler = proc_handler; 4863 4864 if (load_idx) { 4865 entry->extra1 = &min_load_idx; 4866 entry->extra2 = &max_load_idx; 4867 } 4868} 4869 4870static struct ctl_table * 4871sd_alloc_ctl_domain_table(struct sched_domain *sd) 4872{ 4873 struct ctl_table *table = sd_alloc_ctl_entry(14); 4874 4875 if (table == NULL) 4876 return NULL; 4877 4878 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4879 sizeof(long), 0644, proc_doulongvec_minmax, false); 4880 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4881 sizeof(long), 0644, proc_doulongvec_minmax, false); 4882 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4883 sizeof(int), 0644, proc_dointvec_minmax, true); 4884 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4885 sizeof(int), 0644, proc_dointvec_minmax, true); 4886 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4887 sizeof(int), 0644, proc_dointvec_minmax, true); 4888 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4889 sizeof(int), 0644, proc_dointvec_minmax, true); 4890 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4891 sizeof(int), 0644, proc_dointvec_minmax, true); 4892 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4893 sizeof(int), 0644, proc_dointvec_minmax, false); 4894 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4895 sizeof(int), 0644, proc_dointvec_minmax, false); 4896 set_table_entry(&table[9], "cache_nice_tries", 4897 &sd->cache_nice_tries, 4898 sizeof(int), 0644, proc_dointvec_minmax, false); 4899 set_table_entry(&table[10], "flags", &sd->flags, 4900 sizeof(int), 0644, proc_dointvec_minmax, false); 4901 set_table_entry(&table[11], "max_newidle_lb_cost", 4902 &sd->max_newidle_lb_cost, 4903 sizeof(long), 0644, proc_doulongvec_minmax, false); 4904 set_table_entry(&table[12], "name", sd->name, 4905 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4906 /* &table[13] is terminator */ 4907 4908 return table; 4909} 4910 4911static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4912{ 4913 struct ctl_table *entry, *table; 4914 struct sched_domain *sd; 4915 int domain_num = 0, i; 4916 char buf[32]; 4917 4918 for_each_domain(cpu, sd) 4919 domain_num++; 4920 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4921 if (table == NULL) 4922 return NULL; 4923 4924 i = 0; 4925 for_each_domain(cpu, sd) { 4926 snprintf(buf, 32, "domain%d", i); 4927 entry->procname = kstrdup(buf, GFP_KERNEL); 4928 entry->mode = 0555; 4929 entry->child = sd_alloc_ctl_domain_table(sd); 4930 entry++; 4931 i++; 4932 } 4933 return table; 4934} 4935 4936static struct ctl_table_header *sd_sysctl_header; 4937static void register_sched_domain_sysctl(void) 4938{ 4939 int i, cpu_num = num_possible_cpus(); 4940 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4941 char buf[32]; 4942 4943 WARN_ON(sd_ctl_dir[0].child); 4944 sd_ctl_dir[0].child = entry; 4945 4946 if (entry == NULL) 4947 return; 4948 4949 for_each_possible_cpu(i) { 4950 snprintf(buf, 32, "cpu%d", i); 4951 entry->procname = kstrdup(buf, GFP_KERNEL); 4952 entry->mode = 0555; 4953 entry->child = sd_alloc_ctl_cpu_table(i); 4954 entry++; 4955 } 4956 4957 WARN_ON(sd_sysctl_header); 4958 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4959} 4960 4961/* may be called multiple times per register */ 4962static void unregister_sched_domain_sysctl(void) 4963{ 4964 if (sd_sysctl_header) 4965 unregister_sysctl_table(sd_sysctl_header); 4966 sd_sysctl_header = NULL; 4967 if (sd_ctl_dir[0].child) 4968 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4969} 4970#else 4971static void register_sched_domain_sysctl(void) 4972{ 4973} 4974static void unregister_sched_domain_sysctl(void) 4975{ 4976} 4977#endif 4978 4979static void set_rq_online(struct rq *rq) 4980{ 4981 if (!rq->online) { 4982 const struct sched_class *class; 4983 4984 cpumask_set_cpu(rq->cpu, rq->rd->online); 4985 rq->online = 1; 4986 4987 for_each_class(class) { 4988 if (class->rq_online) 4989 class->rq_online(rq); 4990 } 4991 } 4992} 4993 4994static void set_rq_offline(struct rq *rq) 4995{ 4996 if (rq->online) { 4997 const struct sched_class *class; 4998 4999 for_each_class(class) { 5000 if (class->rq_offline) 5001 class->rq_offline(rq); 5002 } 5003 5004 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5005 rq->online = 0; 5006 } 5007} 5008 5009/* 5010 * migration_call - callback that gets triggered when a CPU is added. 5011 * Here we can start up the necessary migration thread for the new CPU. 5012 */ 5013static int 5014migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5015{ 5016 int cpu = (long)hcpu; 5017 unsigned long flags; 5018 struct rq *rq = cpu_rq(cpu); 5019 5020 switch (action & ~CPU_TASKS_FROZEN) { 5021 5022 case CPU_UP_PREPARE: 5023 rq->calc_load_update = calc_load_update; 5024 break; 5025 5026 case CPU_ONLINE: 5027 /* Update our root-domain */ 5028 raw_spin_lock_irqsave(&rq->lock, flags); 5029 if (rq->rd) { 5030 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5031 5032 set_rq_online(rq); 5033 } 5034 raw_spin_unlock_irqrestore(&rq->lock, flags); 5035 break; 5036 5037#ifdef CONFIG_HOTPLUG_CPU 5038 case CPU_DYING: 5039 sched_ttwu_pending(); 5040 /* Update our root-domain */ 5041 raw_spin_lock_irqsave(&rq->lock, flags); 5042 if (rq->rd) { 5043 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5044 set_rq_offline(rq); 5045 } 5046 migrate_tasks(cpu); 5047 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5048 raw_spin_unlock_irqrestore(&rq->lock, flags); 5049 break; 5050 5051 case CPU_DEAD: 5052 calc_load_migrate(rq); 5053 break; 5054#endif 5055 } 5056 5057 update_max_interval(); 5058 5059 return NOTIFY_OK; 5060} 5061 5062/* 5063 * Register at high priority so that task migration (migrate_all_tasks) 5064 * happens before everything else. This has to be lower priority than 5065 * the notifier in the perf_event subsystem, though. 5066 */ 5067static struct notifier_block migration_notifier = { 5068 .notifier_call = migration_call, 5069 .priority = CPU_PRI_MIGRATION, 5070}; 5071 5072static int sched_cpu_active(struct notifier_block *nfb, 5073 unsigned long action, void *hcpu) 5074{ 5075 switch (action & ~CPU_TASKS_FROZEN) { 5076 case CPU_STARTING: 5077 case CPU_DOWN_FAILED: 5078 set_cpu_active((long)hcpu, true); 5079 return NOTIFY_OK; 5080 default: 5081 return NOTIFY_DONE; 5082 } 5083} 5084 5085static int sched_cpu_inactive(struct notifier_block *nfb, 5086 unsigned long action, void *hcpu) 5087{ 5088 unsigned long flags; 5089 long cpu = (long)hcpu; 5090 5091 switch (action & ~CPU_TASKS_FROZEN) { 5092 case CPU_DOWN_PREPARE: 5093 set_cpu_active(cpu, false); 5094 5095 /* explicitly allow suspend */ 5096 if (!(action & CPU_TASKS_FROZEN)) { 5097 struct dl_bw *dl_b = dl_bw_of(cpu); 5098 bool overflow; 5099 int cpus; 5100 5101 raw_spin_lock_irqsave(&dl_b->lock, flags); 5102 cpus = dl_bw_cpus(cpu); 5103 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5104 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5105 5106 if (overflow) 5107 return notifier_from_errno(-EBUSY); 5108 } 5109 return NOTIFY_OK; 5110 } 5111 5112 return NOTIFY_DONE; 5113} 5114 5115static int __init migration_init(void) 5116{ 5117 void *cpu = (void *)(long)smp_processor_id(); 5118 int err; 5119 5120 /* Initialize migration for the boot CPU */ 5121 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5122 BUG_ON(err == NOTIFY_BAD); 5123 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5124 register_cpu_notifier(&migration_notifier); 5125 5126 /* Register cpu active notifiers */ 5127 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5128 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5129 5130 return 0; 5131} 5132early_initcall(migration_init); 5133#endif 5134 5135#ifdef CONFIG_SMP 5136 5137static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5138 5139#ifdef CONFIG_SCHED_DEBUG 5140 5141static __read_mostly int sched_debug_enabled; 5142 5143static int __init sched_debug_setup(char *str) 5144{ 5145 sched_debug_enabled = 1; 5146 5147 return 0; 5148} 5149early_param("sched_debug", sched_debug_setup); 5150 5151static inline bool sched_debug(void) 5152{ 5153 return sched_debug_enabled; 5154} 5155 5156static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5157 struct cpumask *groupmask) 5158{ 5159 struct sched_group *group = sd->groups; 5160 char str[256]; 5161 5162 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5163 cpumask_clear(groupmask); 5164 5165 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5166 5167 if (!(sd->flags & SD_LOAD_BALANCE)) { 5168 printk("does not load-balance\n"); 5169 if (sd->parent) 5170 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5171 " has parent"); 5172 return -1; 5173 } 5174 5175 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5176 5177 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5178 printk(KERN_ERR "ERROR: domain->span does not contain " 5179 "CPU%d\n", cpu); 5180 } 5181 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5182 printk(KERN_ERR "ERROR: domain->groups does not contain" 5183 " CPU%d\n", cpu); 5184 } 5185 5186 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5187 do { 5188 if (!group) { 5189 printk("\n"); 5190 printk(KERN_ERR "ERROR: group is NULL\n"); 5191 break; 5192 } 5193 5194 /* 5195 * Even though we initialize ->power to something semi-sane, 5196 * we leave power_orig unset. This allows us to detect if 5197 * domain iteration is still funny without causing /0 traps. 5198 */ 5199 if (!group->sgp->power_orig) { 5200 printk(KERN_CONT "\n"); 5201 printk(KERN_ERR "ERROR: domain->cpu_power not " 5202 "set\n"); 5203 break; 5204 } 5205 5206 if (!cpumask_weight(sched_group_cpus(group))) { 5207 printk(KERN_CONT "\n"); 5208 printk(KERN_ERR "ERROR: empty group\n"); 5209 break; 5210 } 5211 5212 if (!(sd->flags & SD_OVERLAP) && 5213 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5214 printk(KERN_CONT "\n"); 5215 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5216 break; 5217 } 5218 5219 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5220 5221 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5222 5223 printk(KERN_CONT " %s", str); 5224 if (group->sgp->power != SCHED_POWER_SCALE) { 5225 printk(KERN_CONT " (cpu_power = %d)", 5226 group->sgp->power); 5227 } 5228 5229 group = group->next; 5230 } while (group != sd->groups); 5231 printk(KERN_CONT "\n"); 5232 5233 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5234 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5235 5236 if (sd->parent && 5237 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5238 printk(KERN_ERR "ERROR: parent span is not a superset " 5239 "of domain->span\n"); 5240 return 0; 5241} 5242 5243static void sched_domain_debug(struct sched_domain *sd, int cpu) 5244{ 5245 int level = 0; 5246 5247 if (!sched_debug_enabled) 5248 return; 5249 5250 if (!sd) { 5251 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5252 return; 5253 } 5254 5255 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5256 5257 for (;;) { 5258 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5259 break; 5260 level++; 5261 sd = sd->parent; 5262 if (!sd) 5263 break; 5264 } 5265} 5266#else /* !CONFIG_SCHED_DEBUG */ 5267# define sched_domain_debug(sd, cpu) do { } while (0) 5268static inline bool sched_debug(void) 5269{ 5270 return false; 5271} 5272#endif /* CONFIG_SCHED_DEBUG */ 5273 5274static int sd_degenerate(struct sched_domain *sd) 5275{ 5276 if (cpumask_weight(sched_domain_span(sd)) == 1) 5277 return 1; 5278 5279 /* Following flags need at least 2 groups */ 5280 if (sd->flags & (SD_LOAD_BALANCE | 5281 SD_BALANCE_NEWIDLE | 5282 SD_BALANCE_FORK | 5283 SD_BALANCE_EXEC | 5284 SD_SHARE_CPUPOWER | 5285 SD_SHARE_PKG_RESOURCES)) { 5286 if (sd->groups != sd->groups->next) 5287 return 0; 5288 } 5289 5290 /* Following flags don't use groups */ 5291 if (sd->flags & (SD_WAKE_AFFINE)) 5292 return 0; 5293 5294 return 1; 5295} 5296 5297static int 5298sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5299{ 5300 unsigned long cflags = sd->flags, pflags = parent->flags; 5301 5302 if (sd_degenerate(parent)) 5303 return 1; 5304 5305 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5306 return 0; 5307 5308 /* Flags needing groups don't count if only 1 group in parent */ 5309 if (parent->groups == parent->groups->next) { 5310 pflags &= ~(SD_LOAD_BALANCE | 5311 SD_BALANCE_NEWIDLE | 5312 SD_BALANCE_FORK | 5313 SD_BALANCE_EXEC | 5314 SD_SHARE_CPUPOWER | 5315 SD_SHARE_PKG_RESOURCES | 5316 SD_PREFER_SIBLING); 5317 if (nr_node_ids == 1) 5318 pflags &= ~SD_SERIALIZE; 5319 } 5320 if (~cflags & pflags) 5321 return 0; 5322 5323 return 1; 5324} 5325 5326static void free_rootdomain(struct rcu_head *rcu) 5327{ 5328 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5329 5330 cpupri_cleanup(&rd->cpupri); 5331 cpudl_cleanup(&rd->cpudl); 5332 free_cpumask_var(rd->dlo_mask); 5333 free_cpumask_var(rd->rto_mask); 5334 free_cpumask_var(rd->online); 5335 free_cpumask_var(rd->span); 5336 kfree(rd); 5337} 5338 5339static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5340{ 5341 struct root_domain *old_rd = NULL; 5342 unsigned long flags; 5343 5344 raw_spin_lock_irqsave(&rq->lock, flags); 5345 5346 if (rq->rd) { 5347 old_rd = rq->rd; 5348 5349 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5350 set_rq_offline(rq); 5351 5352 cpumask_clear_cpu(rq->cpu, old_rd->span); 5353 5354 /* 5355 * If we dont want to free the old_rd yet then 5356 * set old_rd to NULL to skip the freeing later 5357 * in this function: 5358 */ 5359 if (!atomic_dec_and_test(&old_rd->refcount)) 5360 old_rd = NULL; 5361 } 5362 5363 atomic_inc(&rd->refcount); 5364 rq->rd = rd; 5365 5366 cpumask_set_cpu(rq->cpu, rd->span); 5367 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5368 set_rq_online(rq); 5369 5370 raw_spin_unlock_irqrestore(&rq->lock, flags); 5371 5372 if (old_rd) 5373 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5374} 5375 5376static int init_rootdomain(struct root_domain *rd) 5377{ 5378 memset(rd, 0, sizeof(*rd)); 5379 5380 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5381 goto out; 5382 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5383 goto free_span; 5384 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5385 goto free_online; 5386 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5387 goto free_dlo_mask; 5388 5389 init_dl_bw(&rd->dl_bw); 5390 if (cpudl_init(&rd->cpudl) != 0) 5391 goto free_dlo_mask; 5392 5393 if (cpupri_init(&rd->cpupri) != 0) 5394 goto free_rto_mask; 5395 return 0; 5396 5397free_rto_mask: 5398 free_cpumask_var(rd->rto_mask); 5399free_dlo_mask: 5400 free_cpumask_var(rd->dlo_mask); 5401free_online: 5402 free_cpumask_var(rd->online); 5403free_span: 5404 free_cpumask_var(rd->span); 5405out: 5406 return -ENOMEM; 5407} 5408 5409/* 5410 * By default the system creates a single root-domain with all cpus as 5411 * members (mimicking the global state we have today). 5412 */ 5413struct root_domain def_root_domain; 5414 5415static void init_defrootdomain(void) 5416{ 5417 init_rootdomain(&def_root_domain); 5418 5419 atomic_set(&def_root_domain.refcount, 1); 5420} 5421 5422static struct root_domain *alloc_rootdomain(void) 5423{ 5424 struct root_domain *rd; 5425 5426 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5427 if (!rd) 5428 return NULL; 5429 5430 if (init_rootdomain(rd) != 0) { 5431 kfree(rd); 5432 return NULL; 5433 } 5434 5435 return rd; 5436} 5437 5438static void free_sched_groups(struct sched_group *sg, int free_sgp) 5439{ 5440 struct sched_group *tmp, *first; 5441 5442 if (!sg) 5443 return; 5444 5445 first = sg; 5446 do { 5447 tmp = sg->next; 5448 5449 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5450 kfree(sg->sgp); 5451 5452 kfree(sg); 5453 sg = tmp; 5454 } while (sg != first); 5455} 5456 5457static void free_sched_domain(struct rcu_head *rcu) 5458{ 5459 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5460 5461 /* 5462 * If its an overlapping domain it has private groups, iterate and 5463 * nuke them all. 5464 */ 5465 if (sd->flags & SD_OVERLAP) { 5466 free_sched_groups(sd->groups, 1); 5467 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5468 kfree(sd->groups->sgp); 5469 kfree(sd->groups); 5470 } 5471 kfree(sd); 5472} 5473 5474static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5475{ 5476 call_rcu(&sd->rcu, free_sched_domain); 5477} 5478 5479static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5480{ 5481 for (; sd; sd = sd->parent) 5482 destroy_sched_domain(sd, cpu); 5483} 5484 5485/* 5486 * Keep a special pointer to the highest sched_domain that has 5487 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5488 * allows us to avoid some pointer chasing select_idle_sibling(). 5489 * 5490 * Also keep a unique ID per domain (we use the first cpu number in 5491 * the cpumask of the domain), this allows us to quickly tell if 5492 * two cpus are in the same cache domain, see cpus_share_cache(). 5493 */ 5494DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5495DEFINE_PER_CPU(int, sd_llc_size); 5496DEFINE_PER_CPU(int, sd_llc_id); 5497DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5498DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5499DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5500 5501static void update_top_cache_domain(int cpu) 5502{ 5503 struct sched_domain *sd; 5504 struct sched_domain *busy_sd = NULL; 5505 int id = cpu; 5506 int size = 1; 5507 5508 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5509 if (sd) { 5510 id = cpumask_first(sched_domain_span(sd)); 5511 size = cpumask_weight(sched_domain_span(sd)); 5512 busy_sd = sd->parent; /* sd_busy */ 5513 } 5514 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5515 5516 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5517 per_cpu(sd_llc_size, cpu) = size; 5518 per_cpu(sd_llc_id, cpu) = id; 5519 5520 sd = lowest_flag_domain(cpu, SD_NUMA); 5521 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5522 5523 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5524 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5525} 5526 5527/* 5528 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5529 * hold the hotplug lock. 5530 */ 5531static void 5532cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5533{ 5534 struct rq *rq = cpu_rq(cpu); 5535 struct sched_domain *tmp; 5536 5537 /* Remove the sched domains which do not contribute to scheduling. */ 5538 for (tmp = sd; tmp; ) { 5539 struct sched_domain *parent = tmp->parent; 5540 if (!parent) 5541 break; 5542 5543 if (sd_parent_degenerate(tmp, parent)) { 5544 tmp->parent = parent->parent; 5545 if (parent->parent) 5546 parent->parent->child = tmp; 5547 /* 5548 * Transfer SD_PREFER_SIBLING down in case of a 5549 * degenerate parent; the spans match for this 5550 * so the property transfers. 5551 */ 5552 if (parent->flags & SD_PREFER_SIBLING) 5553 tmp->flags |= SD_PREFER_SIBLING; 5554 destroy_sched_domain(parent, cpu); 5555 } else 5556 tmp = tmp->parent; 5557 } 5558 5559 if (sd && sd_degenerate(sd)) { 5560 tmp = sd; 5561 sd = sd->parent; 5562 destroy_sched_domain(tmp, cpu); 5563 if (sd) 5564 sd->child = NULL; 5565 } 5566 5567 sched_domain_debug(sd, cpu); 5568 5569 rq_attach_root(rq, rd); 5570 tmp = rq->sd; 5571 rcu_assign_pointer(rq->sd, sd); 5572 destroy_sched_domains(tmp, cpu); 5573 5574 update_top_cache_domain(cpu); 5575} 5576 5577/* cpus with isolated domains */ 5578static cpumask_var_t cpu_isolated_map; 5579 5580/* Setup the mask of cpus configured for isolated domains */ 5581static int __init isolated_cpu_setup(char *str) 5582{ 5583 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5584 cpulist_parse(str, cpu_isolated_map); 5585 return 1; 5586} 5587 5588__setup("isolcpus=", isolated_cpu_setup); 5589 5590static const struct cpumask *cpu_cpu_mask(int cpu) 5591{ 5592 return cpumask_of_node(cpu_to_node(cpu)); 5593} 5594 5595struct sd_data { 5596 struct sched_domain **__percpu sd; 5597 struct sched_group **__percpu sg; 5598 struct sched_group_power **__percpu sgp; 5599}; 5600 5601struct s_data { 5602 struct sched_domain ** __percpu sd; 5603 struct root_domain *rd; 5604}; 5605 5606enum s_alloc { 5607 sa_rootdomain, 5608 sa_sd, 5609 sa_sd_storage, 5610 sa_none, 5611}; 5612 5613struct sched_domain_topology_level; 5614 5615typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5616typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5617 5618#define SDTL_OVERLAP 0x01 5619 5620struct sched_domain_topology_level { 5621 sched_domain_init_f init; 5622 sched_domain_mask_f mask; 5623 int flags; 5624 int numa_level; 5625 struct sd_data data; 5626}; 5627 5628/* 5629 * Build an iteration mask that can exclude certain CPUs from the upwards 5630 * domain traversal. 5631 * 5632 * Asymmetric node setups can result in situations where the domain tree is of 5633 * unequal depth, make sure to skip domains that already cover the entire 5634 * range. 5635 * 5636 * In that case build_sched_domains() will have terminated the iteration early 5637 * and our sibling sd spans will be empty. Domains should always include the 5638 * cpu they're built on, so check that. 5639 * 5640 */ 5641static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5642{ 5643 const struct cpumask *span = sched_domain_span(sd); 5644 struct sd_data *sdd = sd->private; 5645 struct sched_domain *sibling; 5646 int i; 5647 5648 for_each_cpu(i, span) { 5649 sibling = *per_cpu_ptr(sdd->sd, i); 5650 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5651 continue; 5652 5653 cpumask_set_cpu(i, sched_group_mask(sg)); 5654 } 5655} 5656 5657/* 5658 * Return the canonical balance cpu for this group, this is the first cpu 5659 * of this group that's also in the iteration mask. 5660 */ 5661int group_balance_cpu(struct sched_group *sg) 5662{ 5663 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5664} 5665 5666static int 5667build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5668{ 5669 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5670 const struct cpumask *span = sched_domain_span(sd); 5671 struct cpumask *covered = sched_domains_tmpmask; 5672 struct sd_data *sdd = sd->private; 5673 struct sched_domain *child; 5674 int i; 5675 5676 cpumask_clear(covered); 5677 5678 for_each_cpu(i, span) { 5679 struct cpumask *sg_span; 5680 5681 if (cpumask_test_cpu(i, covered)) 5682 continue; 5683 5684 child = *per_cpu_ptr(sdd->sd, i); 5685 5686 /* See the comment near build_group_mask(). */ 5687 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5688 continue; 5689 5690 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5691 GFP_KERNEL, cpu_to_node(cpu)); 5692 5693 if (!sg) 5694 goto fail; 5695 5696 sg_span = sched_group_cpus(sg); 5697 if (child->child) { 5698 child = child->child; 5699 cpumask_copy(sg_span, sched_domain_span(child)); 5700 } else 5701 cpumask_set_cpu(i, sg_span); 5702 5703 cpumask_or(covered, covered, sg_span); 5704 5705 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5706 if (atomic_inc_return(&sg->sgp->ref) == 1) 5707 build_group_mask(sd, sg); 5708 5709 /* 5710 * Initialize sgp->power such that even if we mess up the 5711 * domains and no possible iteration will get us here, we won't 5712 * die on a /0 trap. 5713 */ 5714 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5715 sg->sgp->power_orig = sg->sgp->power; 5716 5717 /* 5718 * Make sure the first group of this domain contains the 5719 * canonical balance cpu. Otherwise the sched_domain iteration 5720 * breaks. See update_sg_lb_stats(). 5721 */ 5722 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5723 group_balance_cpu(sg) == cpu) 5724 groups = sg; 5725 5726 if (!first) 5727 first = sg; 5728 if (last) 5729 last->next = sg; 5730 last = sg; 5731 last->next = first; 5732 } 5733 sd->groups = groups; 5734 5735 return 0; 5736 5737fail: 5738 free_sched_groups(first, 0); 5739 5740 return -ENOMEM; 5741} 5742 5743static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5744{ 5745 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5746 struct sched_domain *child = sd->child; 5747 5748 if (child) 5749 cpu = cpumask_first(sched_domain_span(child)); 5750 5751 if (sg) { 5752 *sg = *per_cpu_ptr(sdd->sg, cpu); 5753 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5754 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5755 } 5756 5757 return cpu; 5758} 5759 5760/* 5761 * build_sched_groups will build a circular linked list of the groups 5762 * covered by the given span, and will set each group's ->cpumask correctly, 5763 * and ->cpu_power to 0. 5764 * 5765 * Assumes the sched_domain tree is fully constructed 5766 */ 5767static int 5768build_sched_groups(struct sched_domain *sd, int cpu) 5769{ 5770 struct sched_group *first = NULL, *last = NULL; 5771 struct sd_data *sdd = sd->private; 5772 const struct cpumask *span = sched_domain_span(sd); 5773 struct cpumask *covered; 5774 int i; 5775 5776 get_group(cpu, sdd, &sd->groups); 5777 atomic_inc(&sd->groups->ref); 5778 5779 if (cpu != cpumask_first(span)) 5780 return 0; 5781 5782 lockdep_assert_held(&sched_domains_mutex); 5783 covered = sched_domains_tmpmask; 5784 5785 cpumask_clear(covered); 5786 5787 for_each_cpu(i, span) { 5788 struct sched_group *sg; 5789 int group, j; 5790 5791 if (cpumask_test_cpu(i, covered)) 5792 continue; 5793 5794 group = get_group(i, sdd, &sg); 5795 cpumask_clear(sched_group_cpus(sg)); 5796 sg->sgp->power = 0; 5797 cpumask_setall(sched_group_mask(sg)); 5798 5799 for_each_cpu(j, span) { 5800 if (get_group(j, sdd, NULL) != group) 5801 continue; 5802 5803 cpumask_set_cpu(j, covered); 5804 cpumask_set_cpu(j, sched_group_cpus(sg)); 5805 } 5806 5807 if (!first) 5808 first = sg; 5809 if (last) 5810 last->next = sg; 5811 last = sg; 5812 } 5813 last->next = first; 5814 5815 return 0; 5816} 5817 5818/* 5819 * Initialize sched groups cpu_power. 5820 * 5821 * cpu_power indicates the capacity of sched group, which is used while 5822 * distributing the load between different sched groups in a sched domain. 5823 * Typically cpu_power for all the groups in a sched domain will be same unless 5824 * there are asymmetries in the topology. If there are asymmetries, group 5825 * having more cpu_power will pickup more load compared to the group having 5826 * less cpu_power. 5827 */ 5828static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5829{ 5830 struct sched_group *sg = sd->groups; 5831 5832 WARN_ON(!sg); 5833 5834 do { 5835 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5836 sg = sg->next; 5837 } while (sg != sd->groups); 5838 5839 if (cpu != group_balance_cpu(sg)) 5840 return; 5841 5842 update_group_power(sd, cpu); 5843 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5844} 5845 5846int __weak arch_sd_sibling_asym_packing(void) 5847{ 5848 return 0*SD_ASYM_PACKING; 5849} 5850 5851/* 5852 * Initializers for schedule domains 5853 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5854 */ 5855 5856#ifdef CONFIG_SCHED_DEBUG 5857# define SD_INIT_NAME(sd, type) sd->name = #type 5858#else 5859# define SD_INIT_NAME(sd, type) do { } while (0) 5860#endif 5861 5862#define SD_INIT_FUNC(type) \ 5863static noinline struct sched_domain * \ 5864sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5865{ \ 5866 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5867 *sd = SD_##type##_INIT; \ 5868 SD_INIT_NAME(sd, type); \ 5869 sd->private = &tl->data; \ 5870 return sd; \ 5871} 5872 5873SD_INIT_FUNC(CPU) 5874#ifdef CONFIG_SCHED_SMT 5875 SD_INIT_FUNC(SIBLING) 5876#endif 5877#ifdef CONFIG_SCHED_MC 5878 SD_INIT_FUNC(MC) 5879#endif 5880#ifdef CONFIG_SCHED_BOOK 5881 SD_INIT_FUNC(BOOK) 5882#endif 5883 5884static int default_relax_domain_level = -1; 5885int sched_domain_level_max; 5886 5887static int __init setup_relax_domain_level(char *str) 5888{ 5889 if (kstrtoint(str, 0, &default_relax_domain_level)) 5890 pr_warn("Unable to set relax_domain_level\n"); 5891 5892 return 1; 5893} 5894__setup("relax_domain_level=", setup_relax_domain_level); 5895 5896static void set_domain_attribute(struct sched_domain *sd, 5897 struct sched_domain_attr *attr) 5898{ 5899 int request; 5900 5901 if (!attr || attr->relax_domain_level < 0) { 5902 if (default_relax_domain_level < 0) 5903 return; 5904 else 5905 request = default_relax_domain_level; 5906 } else 5907 request = attr->relax_domain_level; 5908 if (request < sd->level) { 5909 /* turn off idle balance on this domain */ 5910 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5911 } else { 5912 /* turn on idle balance on this domain */ 5913 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5914 } 5915} 5916 5917static void __sdt_free(const struct cpumask *cpu_map); 5918static int __sdt_alloc(const struct cpumask *cpu_map); 5919 5920static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5921 const struct cpumask *cpu_map) 5922{ 5923 switch (what) { 5924 case sa_rootdomain: 5925 if (!atomic_read(&d->rd->refcount)) 5926 free_rootdomain(&d->rd->rcu); /* fall through */ 5927 case sa_sd: 5928 free_percpu(d->sd); /* fall through */ 5929 case sa_sd_storage: 5930 __sdt_free(cpu_map); /* fall through */ 5931 case sa_none: 5932 break; 5933 } 5934} 5935 5936static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5937 const struct cpumask *cpu_map) 5938{ 5939 memset(d, 0, sizeof(*d)); 5940 5941 if (__sdt_alloc(cpu_map)) 5942 return sa_sd_storage; 5943 d->sd = alloc_percpu(struct sched_domain *); 5944 if (!d->sd) 5945 return sa_sd_storage; 5946 d->rd = alloc_rootdomain(); 5947 if (!d->rd) 5948 return sa_sd; 5949 return sa_rootdomain; 5950} 5951 5952/* 5953 * NULL the sd_data elements we've used to build the sched_domain and 5954 * sched_group structure so that the subsequent __free_domain_allocs() 5955 * will not free the data we're using. 5956 */ 5957static void claim_allocations(int cpu, struct sched_domain *sd) 5958{ 5959 struct sd_data *sdd = sd->private; 5960 5961 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5962 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5963 5964 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5965 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5966 5967 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5968 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5969} 5970 5971#ifdef CONFIG_SCHED_SMT 5972static const struct cpumask *cpu_smt_mask(int cpu) 5973{ 5974 return topology_thread_cpumask(cpu); 5975} 5976#endif 5977 5978/* 5979 * Topology list, bottom-up. 5980 */ 5981static struct sched_domain_topology_level default_topology[] = { 5982#ifdef CONFIG_SCHED_SMT 5983 { sd_init_SIBLING, cpu_smt_mask, }, 5984#endif 5985#ifdef CONFIG_SCHED_MC 5986 { sd_init_MC, cpu_coregroup_mask, }, 5987#endif 5988#ifdef CONFIG_SCHED_BOOK 5989 { sd_init_BOOK, cpu_book_mask, }, 5990#endif 5991 { sd_init_CPU, cpu_cpu_mask, }, 5992 { NULL, }, 5993}; 5994 5995static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5996 5997#define for_each_sd_topology(tl) \ 5998 for (tl = sched_domain_topology; tl->init; tl++) 5999 6000#ifdef CONFIG_NUMA 6001 6002static int sched_domains_numa_levels; 6003static int *sched_domains_numa_distance; 6004static struct cpumask ***sched_domains_numa_masks; 6005static int sched_domains_curr_level; 6006 6007static inline int sd_local_flags(int level) 6008{ 6009 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 6010 return 0; 6011 6012 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 6013} 6014 6015static struct sched_domain * 6016sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 6017{ 6018 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6019 int level = tl->numa_level; 6020 int sd_weight = cpumask_weight( 6021 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 6022 6023 *sd = (struct sched_domain){ 6024 .min_interval = sd_weight, 6025 .max_interval = 2*sd_weight, 6026 .busy_factor = 32, 6027 .imbalance_pct = 125, 6028 .cache_nice_tries = 2, 6029 .busy_idx = 3, 6030 .idle_idx = 2, 6031 .newidle_idx = 0, 6032 .wake_idx = 0, 6033 .forkexec_idx = 0, 6034 6035 .flags = 1*SD_LOAD_BALANCE 6036 | 1*SD_BALANCE_NEWIDLE 6037 | 0*SD_BALANCE_EXEC 6038 | 0*SD_BALANCE_FORK 6039 | 0*SD_BALANCE_WAKE 6040 | 0*SD_WAKE_AFFINE 6041 | 0*SD_SHARE_CPUPOWER 6042 | 0*SD_SHARE_PKG_RESOURCES 6043 | 1*SD_SERIALIZE 6044 | 0*SD_PREFER_SIBLING 6045 | 1*SD_NUMA 6046 | sd_local_flags(level) 6047 , 6048 .last_balance = jiffies, 6049 .balance_interval = sd_weight, 6050 }; 6051 SD_INIT_NAME(sd, NUMA); 6052 sd->private = &tl->data; 6053 6054 /* 6055 * Ugly hack to pass state to sd_numa_mask()... 6056 */ 6057 sched_domains_curr_level = tl->numa_level; 6058 6059 return sd; 6060} 6061 6062static const struct cpumask *sd_numa_mask(int cpu) 6063{ 6064 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6065} 6066 6067static void sched_numa_warn(const char *str) 6068{ 6069 static int done = false; 6070 int i,j; 6071 6072 if (done) 6073 return; 6074 6075 done = true; 6076 6077 printk(KERN_WARNING "ERROR: %s\n\n", str); 6078 6079 for (i = 0; i < nr_node_ids; i++) { 6080 printk(KERN_WARNING " "); 6081 for (j = 0; j < nr_node_ids; j++) 6082 printk(KERN_CONT "%02d ", node_distance(i,j)); 6083 printk(KERN_CONT "\n"); 6084 } 6085 printk(KERN_WARNING "\n"); 6086} 6087 6088static bool find_numa_distance(int distance) 6089{ 6090 int i; 6091 6092 if (distance == node_distance(0, 0)) 6093 return true; 6094 6095 for (i = 0; i < sched_domains_numa_levels; i++) { 6096 if (sched_domains_numa_distance[i] == distance) 6097 return true; 6098 } 6099 6100 return false; 6101} 6102 6103static void sched_init_numa(void) 6104{ 6105 int next_distance, curr_distance = node_distance(0, 0); 6106 struct sched_domain_topology_level *tl; 6107 int level = 0; 6108 int i, j, k; 6109 6110 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6111 if (!sched_domains_numa_distance) 6112 return; 6113 6114 /* 6115 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6116 * unique distances in the node_distance() table. 6117 * 6118 * Assumes node_distance(0,j) includes all distances in 6119 * node_distance(i,j) in order to avoid cubic time. 6120 */ 6121 next_distance = curr_distance; 6122 for (i = 0; i < nr_node_ids; i++) { 6123 for (j = 0; j < nr_node_ids; j++) { 6124 for (k = 0; k < nr_node_ids; k++) { 6125 int distance = node_distance(i, k); 6126 6127 if (distance > curr_distance && 6128 (distance < next_distance || 6129 next_distance == curr_distance)) 6130 next_distance = distance; 6131 6132 /* 6133 * While not a strong assumption it would be nice to know 6134 * about cases where if node A is connected to B, B is not 6135 * equally connected to A. 6136 */ 6137 if (sched_debug() && node_distance(k, i) != distance) 6138 sched_numa_warn("Node-distance not symmetric"); 6139 6140 if (sched_debug() && i && !find_numa_distance(distance)) 6141 sched_numa_warn("Node-0 not representative"); 6142 } 6143 if (next_distance != curr_distance) { 6144 sched_domains_numa_distance[level++] = next_distance; 6145 sched_domains_numa_levels = level; 6146 curr_distance = next_distance; 6147 } else break; 6148 } 6149 6150 /* 6151 * In case of sched_debug() we verify the above assumption. 6152 */ 6153 if (!sched_debug()) 6154 break; 6155 } 6156 /* 6157 * 'level' contains the number of unique distances, excluding the 6158 * identity distance node_distance(i,i). 6159 * 6160 * The sched_domains_numa_distance[] array includes the actual distance 6161 * numbers. 6162 */ 6163 6164 /* 6165 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6166 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6167 * the array will contain less then 'level' members. This could be 6168 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6169 * in other functions. 6170 * 6171 * We reset it to 'level' at the end of this function. 6172 */ 6173 sched_domains_numa_levels = 0; 6174 6175 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6176 if (!sched_domains_numa_masks) 6177 return; 6178 6179 /* 6180 * Now for each level, construct a mask per node which contains all 6181 * cpus of nodes that are that many hops away from us. 6182 */ 6183 for (i = 0; i < level; i++) { 6184 sched_domains_numa_masks[i] = 6185 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6186 if (!sched_domains_numa_masks[i]) 6187 return; 6188 6189 for (j = 0; j < nr_node_ids; j++) { 6190 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6191 if (!mask) 6192 return; 6193 6194 sched_domains_numa_masks[i][j] = mask; 6195 6196 for (k = 0; k < nr_node_ids; k++) { 6197 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6198 continue; 6199 6200 cpumask_or(mask, mask, cpumask_of_node(k)); 6201 } 6202 } 6203 } 6204 6205 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6206 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6207 if (!tl) 6208 return; 6209 6210 /* 6211 * Copy the default topology bits.. 6212 */ 6213 for (i = 0; default_topology[i].init; i++) 6214 tl[i] = default_topology[i]; 6215 6216 /* 6217 * .. and append 'j' levels of NUMA goodness. 6218 */ 6219 for (j = 0; j < level; i++, j++) { 6220 tl[i] = (struct sched_domain_topology_level){ 6221 .init = sd_numa_init, 6222 .mask = sd_numa_mask, 6223 .flags = SDTL_OVERLAP, 6224 .numa_level = j, 6225 }; 6226 } 6227 6228 sched_domain_topology = tl; 6229 6230 sched_domains_numa_levels = level; 6231} 6232 6233static void sched_domains_numa_masks_set(int cpu) 6234{ 6235 int i, j; 6236 int node = cpu_to_node(cpu); 6237 6238 for (i = 0; i < sched_domains_numa_levels; i++) { 6239 for (j = 0; j < nr_node_ids; j++) { 6240 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6241 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6242 } 6243 } 6244} 6245 6246static void sched_domains_numa_masks_clear(int cpu) 6247{ 6248 int i, j; 6249 for (i = 0; i < sched_domains_numa_levels; i++) { 6250 for (j = 0; j < nr_node_ids; j++) 6251 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6252 } 6253} 6254 6255/* 6256 * Update sched_domains_numa_masks[level][node] array when new cpus 6257 * are onlined. 6258 */ 6259static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6260 unsigned long action, 6261 void *hcpu) 6262{ 6263 int cpu = (long)hcpu; 6264 6265 switch (action & ~CPU_TASKS_FROZEN) { 6266 case CPU_ONLINE: 6267 sched_domains_numa_masks_set(cpu); 6268 break; 6269 6270 case CPU_DEAD: 6271 sched_domains_numa_masks_clear(cpu); 6272 break; 6273 6274 default: 6275 return NOTIFY_DONE; 6276 } 6277 6278 return NOTIFY_OK; 6279} 6280#else 6281static inline void sched_init_numa(void) 6282{ 6283} 6284 6285static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6286 unsigned long action, 6287 void *hcpu) 6288{ 6289 return 0; 6290} 6291#endif /* CONFIG_NUMA */ 6292 6293static int __sdt_alloc(const struct cpumask *cpu_map) 6294{ 6295 struct sched_domain_topology_level *tl; 6296 int j; 6297 6298 for_each_sd_topology(tl) { 6299 struct sd_data *sdd = &tl->data; 6300 6301 sdd->sd = alloc_percpu(struct sched_domain *); 6302 if (!sdd->sd) 6303 return -ENOMEM; 6304 6305 sdd->sg = alloc_percpu(struct sched_group *); 6306 if (!sdd->sg) 6307 return -ENOMEM; 6308 6309 sdd->sgp = alloc_percpu(struct sched_group_power *); 6310 if (!sdd->sgp) 6311 return -ENOMEM; 6312 6313 for_each_cpu(j, cpu_map) { 6314 struct sched_domain *sd; 6315 struct sched_group *sg; 6316 struct sched_group_power *sgp; 6317 6318 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6319 GFP_KERNEL, cpu_to_node(j)); 6320 if (!sd) 6321 return -ENOMEM; 6322 6323 *per_cpu_ptr(sdd->sd, j) = sd; 6324 6325 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6326 GFP_KERNEL, cpu_to_node(j)); 6327 if (!sg) 6328 return -ENOMEM; 6329 6330 sg->next = sg; 6331 6332 *per_cpu_ptr(sdd->sg, j) = sg; 6333 6334 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6335 GFP_KERNEL, cpu_to_node(j)); 6336 if (!sgp) 6337 return -ENOMEM; 6338 6339 *per_cpu_ptr(sdd->sgp, j) = sgp; 6340 } 6341 } 6342 6343 return 0; 6344} 6345 6346static void __sdt_free(const struct cpumask *cpu_map) 6347{ 6348 struct sched_domain_topology_level *tl; 6349 int j; 6350 6351 for_each_sd_topology(tl) { 6352 struct sd_data *sdd = &tl->data; 6353 6354 for_each_cpu(j, cpu_map) { 6355 struct sched_domain *sd; 6356 6357 if (sdd->sd) { 6358 sd = *per_cpu_ptr(sdd->sd, j); 6359 if (sd && (sd->flags & SD_OVERLAP)) 6360 free_sched_groups(sd->groups, 0); 6361 kfree(*per_cpu_ptr(sdd->sd, j)); 6362 } 6363 6364 if (sdd->sg) 6365 kfree(*per_cpu_ptr(sdd->sg, j)); 6366 if (sdd->sgp) 6367 kfree(*per_cpu_ptr(sdd->sgp, j)); 6368 } 6369 free_percpu(sdd->sd); 6370 sdd->sd = NULL; 6371 free_percpu(sdd->sg); 6372 sdd->sg = NULL; 6373 free_percpu(sdd->sgp); 6374 sdd->sgp = NULL; 6375 } 6376} 6377 6378struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6379 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6380 struct sched_domain *child, int cpu) 6381{ 6382 struct sched_domain *sd = tl->init(tl, cpu); 6383 if (!sd) 6384 return child; 6385 6386 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6387 if (child) { 6388 sd->level = child->level + 1; 6389 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6390 child->parent = sd; 6391 sd->child = child; 6392 } 6393 set_domain_attribute(sd, attr); 6394 6395 return sd; 6396} 6397 6398/* 6399 * Build sched domains for a given set of cpus and attach the sched domains 6400 * to the individual cpus 6401 */ 6402static int build_sched_domains(const struct cpumask *cpu_map, 6403 struct sched_domain_attr *attr) 6404{ 6405 enum s_alloc alloc_state; 6406 struct sched_domain *sd; 6407 struct s_data d; 6408 int i, ret = -ENOMEM; 6409 6410 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6411 if (alloc_state != sa_rootdomain) 6412 goto error; 6413 6414 /* Set up domains for cpus specified by the cpu_map. */ 6415 for_each_cpu(i, cpu_map) { 6416 struct sched_domain_topology_level *tl; 6417 6418 sd = NULL; 6419 for_each_sd_topology(tl) { 6420 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6421 if (tl == sched_domain_topology) 6422 *per_cpu_ptr(d.sd, i) = sd; 6423 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6424 sd->flags |= SD_OVERLAP; 6425 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6426 break; 6427 } 6428 } 6429 6430 /* Build the groups for the domains */ 6431 for_each_cpu(i, cpu_map) { 6432 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6433 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6434 if (sd->flags & SD_OVERLAP) { 6435 if (build_overlap_sched_groups(sd, i)) 6436 goto error; 6437 } else { 6438 if (build_sched_groups(sd, i)) 6439 goto error; 6440 } 6441 } 6442 } 6443 6444 /* Calculate CPU power for physical packages and nodes */ 6445 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6446 if (!cpumask_test_cpu(i, cpu_map)) 6447 continue; 6448 6449 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6450 claim_allocations(i, sd); 6451 init_sched_groups_power(i, sd); 6452 } 6453 } 6454 6455 /* Attach the domains */ 6456 rcu_read_lock(); 6457 for_each_cpu(i, cpu_map) { 6458 sd = *per_cpu_ptr(d.sd, i); 6459 cpu_attach_domain(sd, d.rd, i); 6460 } 6461 rcu_read_unlock(); 6462 6463 ret = 0; 6464error: 6465 __free_domain_allocs(&d, alloc_state, cpu_map); 6466 return ret; 6467} 6468 6469static cpumask_var_t *doms_cur; /* current sched domains */ 6470static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6471static struct sched_domain_attr *dattr_cur; 6472 /* attribues of custom domains in 'doms_cur' */ 6473 6474/* 6475 * Special case: If a kmalloc of a doms_cur partition (array of 6476 * cpumask) fails, then fallback to a single sched domain, 6477 * as determined by the single cpumask fallback_doms. 6478 */ 6479static cpumask_var_t fallback_doms; 6480 6481/* 6482 * arch_update_cpu_topology lets virtualized architectures update the 6483 * cpu core maps. It is supposed to return 1 if the topology changed 6484 * or 0 if it stayed the same. 6485 */ 6486int __attribute__((weak)) arch_update_cpu_topology(void) 6487{ 6488 return 0; 6489} 6490 6491cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6492{ 6493 int i; 6494 cpumask_var_t *doms; 6495 6496 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6497 if (!doms) 6498 return NULL; 6499 for (i = 0; i < ndoms; i++) { 6500 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6501 free_sched_domains(doms, i); 6502 return NULL; 6503 } 6504 } 6505 return doms; 6506} 6507 6508void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6509{ 6510 unsigned int i; 6511 for (i = 0; i < ndoms; i++) 6512 free_cpumask_var(doms[i]); 6513 kfree(doms); 6514} 6515 6516/* 6517 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6518 * For now this just excludes isolated cpus, but could be used to 6519 * exclude other special cases in the future. 6520 */ 6521static int init_sched_domains(const struct cpumask *cpu_map) 6522{ 6523 int err; 6524 6525 arch_update_cpu_topology(); 6526 ndoms_cur = 1; 6527 doms_cur = alloc_sched_domains(ndoms_cur); 6528 if (!doms_cur) 6529 doms_cur = &fallback_doms; 6530 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6531 err = build_sched_domains(doms_cur[0], NULL); 6532 register_sched_domain_sysctl(); 6533 6534 return err; 6535} 6536 6537/* 6538 * Detach sched domains from a group of cpus specified in cpu_map 6539 * These cpus will now be attached to the NULL domain 6540 */ 6541static void detach_destroy_domains(const struct cpumask *cpu_map) 6542{ 6543 int i; 6544 6545 rcu_read_lock(); 6546 for_each_cpu(i, cpu_map) 6547 cpu_attach_domain(NULL, &def_root_domain, i); 6548 rcu_read_unlock(); 6549} 6550 6551/* handle null as "default" */ 6552static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6553 struct sched_domain_attr *new, int idx_new) 6554{ 6555 struct sched_domain_attr tmp; 6556 6557 /* fast path */ 6558 if (!new && !cur) 6559 return 1; 6560 6561 tmp = SD_ATTR_INIT; 6562 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6563 new ? (new + idx_new) : &tmp, 6564 sizeof(struct sched_domain_attr)); 6565} 6566 6567/* 6568 * Partition sched domains as specified by the 'ndoms_new' 6569 * cpumasks in the array doms_new[] of cpumasks. This compares 6570 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6571 * It destroys each deleted domain and builds each new domain. 6572 * 6573 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6574 * The masks don't intersect (don't overlap.) We should setup one 6575 * sched domain for each mask. CPUs not in any of the cpumasks will 6576 * not be load balanced. If the same cpumask appears both in the 6577 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6578 * it as it is. 6579 * 6580 * The passed in 'doms_new' should be allocated using 6581 * alloc_sched_domains. This routine takes ownership of it and will 6582 * free_sched_domains it when done with it. If the caller failed the 6583 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6584 * and partition_sched_domains() will fallback to the single partition 6585 * 'fallback_doms', it also forces the domains to be rebuilt. 6586 * 6587 * If doms_new == NULL it will be replaced with cpu_online_mask. 6588 * ndoms_new == 0 is a special case for destroying existing domains, 6589 * and it will not create the default domain. 6590 * 6591 * Call with hotplug lock held 6592 */ 6593void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6594 struct sched_domain_attr *dattr_new) 6595{ 6596 int i, j, n; 6597 int new_topology; 6598 6599 mutex_lock(&sched_domains_mutex); 6600 6601 /* always unregister in case we don't destroy any domains */ 6602 unregister_sched_domain_sysctl(); 6603 6604 /* Let architecture update cpu core mappings. */ 6605 new_topology = arch_update_cpu_topology(); 6606 6607 n = doms_new ? ndoms_new : 0; 6608 6609 /* Destroy deleted domains */ 6610 for (i = 0; i < ndoms_cur; i++) { 6611 for (j = 0; j < n && !new_topology; j++) { 6612 if (cpumask_equal(doms_cur[i], doms_new[j]) 6613 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6614 goto match1; 6615 } 6616 /* no match - a current sched domain not in new doms_new[] */ 6617 detach_destroy_domains(doms_cur[i]); 6618match1: 6619 ; 6620 } 6621 6622 n = ndoms_cur; 6623 if (doms_new == NULL) { 6624 n = 0; 6625 doms_new = &fallback_doms; 6626 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6627 WARN_ON_ONCE(dattr_new); 6628 } 6629 6630 /* Build new domains */ 6631 for (i = 0; i < ndoms_new; i++) { 6632 for (j = 0; j < n && !new_topology; j++) { 6633 if (cpumask_equal(doms_new[i], doms_cur[j]) 6634 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6635 goto match2; 6636 } 6637 /* no match - add a new doms_new */ 6638 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6639match2: 6640 ; 6641 } 6642 6643 /* Remember the new sched domains */ 6644 if (doms_cur != &fallback_doms) 6645 free_sched_domains(doms_cur, ndoms_cur); 6646 kfree(dattr_cur); /* kfree(NULL) is safe */ 6647 doms_cur = doms_new; 6648 dattr_cur = dattr_new; 6649 ndoms_cur = ndoms_new; 6650 6651 register_sched_domain_sysctl(); 6652 6653 mutex_unlock(&sched_domains_mutex); 6654} 6655 6656static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6657 6658/* 6659 * Update cpusets according to cpu_active mask. If cpusets are 6660 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6661 * around partition_sched_domains(). 6662 * 6663 * If we come here as part of a suspend/resume, don't touch cpusets because we 6664 * want to restore it back to its original state upon resume anyway. 6665 */ 6666static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6667 void *hcpu) 6668{ 6669 switch (action) { 6670 case CPU_ONLINE_FROZEN: 6671 case CPU_DOWN_FAILED_FROZEN: 6672 6673 /* 6674 * num_cpus_frozen tracks how many CPUs are involved in suspend 6675 * resume sequence. As long as this is not the last online 6676 * operation in the resume sequence, just build a single sched 6677 * domain, ignoring cpusets. 6678 */ 6679 num_cpus_frozen--; 6680 if (likely(num_cpus_frozen)) { 6681 partition_sched_domains(1, NULL, NULL); 6682 break; 6683 } 6684 6685 /* 6686 * This is the last CPU online operation. So fall through and 6687 * restore the original sched domains by considering the 6688 * cpuset configurations. 6689 */ 6690 6691 case CPU_ONLINE: 6692 case CPU_DOWN_FAILED: 6693 cpuset_update_active_cpus(true); 6694 break; 6695 default: 6696 return NOTIFY_DONE; 6697 } 6698 return NOTIFY_OK; 6699} 6700 6701static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6702 void *hcpu) 6703{ 6704 switch (action) { 6705 case CPU_DOWN_PREPARE: 6706 cpuset_update_active_cpus(false); 6707 break; 6708 case CPU_DOWN_PREPARE_FROZEN: 6709 num_cpus_frozen++; 6710 partition_sched_domains(1, NULL, NULL); 6711 break; 6712 default: 6713 return NOTIFY_DONE; 6714 } 6715 return NOTIFY_OK; 6716} 6717 6718void __init sched_init_smp(void) 6719{ 6720 cpumask_var_t non_isolated_cpus; 6721 6722 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6723 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6724 6725 sched_init_numa(); 6726 6727 /* 6728 * There's no userspace yet to cause hotplug operations; hence all the 6729 * cpu masks are stable and all blatant races in the below code cannot 6730 * happen. 6731 */ 6732 mutex_lock(&sched_domains_mutex); 6733 init_sched_domains(cpu_active_mask); 6734 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6735 if (cpumask_empty(non_isolated_cpus)) 6736 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6737 mutex_unlock(&sched_domains_mutex); 6738 6739 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6740 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6741 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6742 6743 init_hrtick(); 6744 6745 /* Move init over to a non-isolated CPU */ 6746 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6747 BUG(); 6748 sched_init_granularity(); 6749 free_cpumask_var(non_isolated_cpus); 6750 6751 init_sched_rt_class(); 6752 init_sched_dl_class(); 6753} 6754#else 6755void __init sched_init_smp(void) 6756{ 6757 sched_init_granularity(); 6758} 6759#endif /* CONFIG_SMP */ 6760 6761const_debug unsigned int sysctl_timer_migration = 1; 6762 6763int in_sched_functions(unsigned long addr) 6764{ 6765 return in_lock_functions(addr) || 6766 (addr >= (unsigned long)__sched_text_start 6767 && addr < (unsigned long)__sched_text_end); 6768} 6769 6770#ifdef CONFIG_CGROUP_SCHED 6771/* 6772 * Default task group. 6773 * Every task in system belongs to this group at bootup. 6774 */ 6775struct task_group root_task_group; 6776LIST_HEAD(task_groups); 6777#endif 6778 6779DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6780 6781void __init sched_init(void) 6782{ 6783 int i, j; 6784 unsigned long alloc_size = 0, ptr; 6785 6786#ifdef CONFIG_FAIR_GROUP_SCHED 6787 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6788#endif 6789#ifdef CONFIG_RT_GROUP_SCHED 6790 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6791#endif 6792#ifdef CONFIG_CPUMASK_OFFSTACK 6793 alloc_size += num_possible_cpus() * cpumask_size(); 6794#endif 6795 if (alloc_size) { 6796 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6797 6798#ifdef CONFIG_FAIR_GROUP_SCHED 6799 root_task_group.se = (struct sched_entity **)ptr; 6800 ptr += nr_cpu_ids * sizeof(void **); 6801 6802 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6803 ptr += nr_cpu_ids * sizeof(void **); 6804 6805#endif /* CONFIG_FAIR_GROUP_SCHED */ 6806#ifdef CONFIG_RT_GROUP_SCHED 6807 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6808 ptr += nr_cpu_ids * sizeof(void **); 6809 6810 root_task_group.rt_rq = (struct rt_rq **)ptr; 6811 ptr += nr_cpu_ids * sizeof(void **); 6812 6813#endif /* CONFIG_RT_GROUP_SCHED */ 6814#ifdef CONFIG_CPUMASK_OFFSTACK 6815 for_each_possible_cpu(i) { 6816 per_cpu(load_balance_mask, i) = (void *)ptr; 6817 ptr += cpumask_size(); 6818 } 6819#endif /* CONFIG_CPUMASK_OFFSTACK */ 6820 } 6821 6822 init_rt_bandwidth(&def_rt_bandwidth, 6823 global_rt_period(), global_rt_runtime()); 6824 init_dl_bandwidth(&def_dl_bandwidth, 6825 global_rt_period(), global_rt_runtime()); 6826 6827#ifdef CONFIG_SMP 6828 init_defrootdomain(); 6829#endif 6830 6831#ifdef CONFIG_RT_GROUP_SCHED 6832 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6833 global_rt_period(), global_rt_runtime()); 6834#endif /* CONFIG_RT_GROUP_SCHED */ 6835 6836#ifdef CONFIG_CGROUP_SCHED 6837 list_add(&root_task_group.list, &task_groups); 6838 INIT_LIST_HEAD(&root_task_group.children); 6839 INIT_LIST_HEAD(&root_task_group.siblings); 6840 autogroup_init(&init_task); 6841 6842#endif /* CONFIG_CGROUP_SCHED */ 6843 6844 for_each_possible_cpu(i) { 6845 struct rq *rq; 6846 6847 rq = cpu_rq(i); 6848 raw_spin_lock_init(&rq->lock); 6849 rq->nr_running = 0; 6850 rq->calc_load_active = 0; 6851 rq->calc_load_update = jiffies + LOAD_FREQ; 6852 init_cfs_rq(&rq->cfs); 6853 init_rt_rq(&rq->rt, rq); 6854 init_dl_rq(&rq->dl, rq); 6855#ifdef CONFIG_FAIR_GROUP_SCHED 6856 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6857 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6858 /* 6859 * How much cpu bandwidth does root_task_group get? 6860 * 6861 * In case of task-groups formed thr' the cgroup filesystem, it 6862 * gets 100% of the cpu resources in the system. This overall 6863 * system cpu resource is divided among the tasks of 6864 * root_task_group and its child task-groups in a fair manner, 6865 * based on each entity's (task or task-group's) weight 6866 * (se->load.weight). 6867 * 6868 * In other words, if root_task_group has 10 tasks of weight 6869 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6870 * then A0's share of the cpu resource is: 6871 * 6872 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6873 * 6874 * We achieve this by letting root_task_group's tasks sit 6875 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6876 */ 6877 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6878 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6879#endif /* CONFIG_FAIR_GROUP_SCHED */ 6880 6881 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6882#ifdef CONFIG_RT_GROUP_SCHED 6883 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6884#endif 6885 6886 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6887 rq->cpu_load[j] = 0; 6888 6889 rq->last_load_update_tick = jiffies; 6890 6891#ifdef CONFIG_SMP 6892 rq->sd = NULL; 6893 rq->rd = NULL; 6894 rq->cpu_power = SCHED_POWER_SCALE; 6895 rq->post_schedule = 0; 6896 rq->active_balance = 0; 6897 rq->next_balance = jiffies; 6898 rq->push_cpu = 0; 6899 rq->cpu = i; 6900 rq->online = 0; 6901 rq->idle_stamp = 0; 6902 rq->avg_idle = 2*sysctl_sched_migration_cost; 6903 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6904 6905 INIT_LIST_HEAD(&rq->cfs_tasks); 6906 6907 rq_attach_root(rq, &def_root_domain); 6908#ifdef CONFIG_NO_HZ_COMMON 6909 rq->nohz_flags = 0; 6910#endif 6911#ifdef CONFIG_NO_HZ_FULL 6912 rq->last_sched_tick = 0; 6913#endif 6914#endif 6915 init_rq_hrtick(rq); 6916 atomic_set(&rq->nr_iowait, 0); 6917 } 6918 6919 set_load_weight(&init_task); 6920 6921#ifdef CONFIG_PREEMPT_NOTIFIERS 6922 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6923#endif 6924 6925 /* 6926 * The boot idle thread does lazy MMU switching as well: 6927 */ 6928 atomic_inc(&init_mm.mm_count); 6929 enter_lazy_tlb(&init_mm, current); 6930 6931 /* 6932 * Make us the idle thread. Technically, schedule() should not be 6933 * called from this thread, however somewhere below it might be, 6934 * but because we are the idle thread, we just pick up running again 6935 * when this runqueue becomes "idle". 6936 */ 6937 init_idle(current, smp_processor_id()); 6938 6939 calc_load_update = jiffies + LOAD_FREQ; 6940 6941 /* 6942 * During early bootup we pretend to be a normal task: 6943 */ 6944 current->sched_class = &fair_sched_class; 6945 6946#ifdef CONFIG_SMP 6947 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6948 /* May be allocated at isolcpus cmdline parse time */ 6949 if (cpu_isolated_map == NULL) 6950 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6951 idle_thread_set_boot_cpu(); 6952#endif 6953 init_sched_fair_class(); 6954 6955 scheduler_running = 1; 6956} 6957 6958#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6959static inline int preempt_count_equals(int preempt_offset) 6960{ 6961 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6962 6963 return (nested == preempt_offset); 6964} 6965 6966void __might_sleep(const char *file, int line, int preempt_offset) 6967{ 6968 static unsigned long prev_jiffy; /* ratelimiting */ 6969 6970 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6971 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 6972 !is_idle_task(current)) || 6973 system_state != SYSTEM_RUNNING || oops_in_progress) 6974 return; 6975 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6976 return; 6977 prev_jiffy = jiffies; 6978 6979 printk(KERN_ERR 6980 "BUG: sleeping function called from invalid context at %s:%d\n", 6981 file, line); 6982 printk(KERN_ERR 6983 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6984 in_atomic(), irqs_disabled(), 6985 current->pid, current->comm); 6986 6987 debug_show_held_locks(current); 6988 if (irqs_disabled()) 6989 print_irqtrace_events(current); 6990#ifdef CONFIG_DEBUG_PREEMPT 6991 if (!preempt_count_equals(preempt_offset)) { 6992 pr_err("Preemption disabled at:"); 6993 print_ip_sym(current->preempt_disable_ip); 6994 pr_cont("\n"); 6995 } 6996#endif 6997 dump_stack(); 6998} 6999EXPORT_SYMBOL(__might_sleep); 7000#endif 7001 7002#ifdef CONFIG_MAGIC_SYSRQ 7003static void normalize_task(struct rq *rq, struct task_struct *p) 7004{ 7005 const struct sched_class *prev_class = p->sched_class; 7006 struct sched_attr attr = { 7007 .sched_policy = SCHED_NORMAL, 7008 }; 7009 int old_prio = p->prio; 7010 int on_rq; 7011 7012 on_rq = p->on_rq; 7013 if (on_rq) 7014 dequeue_task(rq, p, 0); 7015 __setscheduler(rq, p, &attr); 7016 if (on_rq) { 7017 enqueue_task(rq, p, 0); 7018 resched_task(rq->curr); 7019 } 7020 7021 check_class_changed(rq, p, prev_class, old_prio); 7022} 7023 7024void normalize_rt_tasks(void) 7025{ 7026 struct task_struct *g, *p; 7027 unsigned long flags; 7028 struct rq *rq; 7029 7030 read_lock_irqsave(&tasklist_lock, flags); 7031 do_each_thread(g, p) { 7032 /* 7033 * Only normalize user tasks: 7034 */ 7035 if (!p->mm) 7036 continue; 7037 7038 p->se.exec_start = 0; 7039#ifdef CONFIG_SCHEDSTATS 7040 p->se.statistics.wait_start = 0; 7041 p->se.statistics.sleep_start = 0; 7042 p->se.statistics.block_start = 0; 7043#endif 7044 7045 if (!dl_task(p) && !rt_task(p)) { 7046 /* 7047 * Renice negative nice level userspace 7048 * tasks back to 0: 7049 */ 7050 if (task_nice(p) < 0 && p->mm) 7051 set_user_nice(p, 0); 7052 continue; 7053 } 7054 7055 raw_spin_lock(&p->pi_lock); 7056 rq = __task_rq_lock(p); 7057 7058 normalize_task(rq, p); 7059 7060 __task_rq_unlock(rq); 7061 raw_spin_unlock(&p->pi_lock); 7062 } while_each_thread(g, p); 7063 7064 read_unlock_irqrestore(&tasklist_lock, flags); 7065} 7066 7067#endif /* CONFIG_MAGIC_SYSRQ */ 7068 7069#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7070/* 7071 * These functions are only useful for the IA64 MCA handling, or kdb. 7072 * 7073 * They can only be called when the whole system has been 7074 * stopped - every CPU needs to be quiescent, and no scheduling 7075 * activity can take place. Using them for anything else would 7076 * be a serious bug, and as a result, they aren't even visible 7077 * under any other configuration. 7078 */ 7079 7080/** 7081 * curr_task - return the current task for a given cpu. 7082 * @cpu: the processor in question. 7083 * 7084 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7085 * 7086 * Return: The current task for @cpu. 7087 */ 7088struct task_struct *curr_task(int cpu) 7089{ 7090 return cpu_curr(cpu); 7091} 7092 7093#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7094 7095#ifdef CONFIG_IA64 7096/** 7097 * set_curr_task - set the current task for a given cpu. 7098 * @cpu: the processor in question. 7099 * @p: the task pointer to set. 7100 * 7101 * Description: This function must only be used when non-maskable interrupts 7102 * are serviced on a separate stack. It allows the architecture to switch the 7103 * notion of the current task on a cpu in a non-blocking manner. This function 7104 * must be called with all CPU's synchronized, and interrupts disabled, the 7105 * and caller must save the original value of the current task (see 7106 * curr_task() above) and restore that value before reenabling interrupts and 7107 * re-starting the system. 7108 * 7109 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7110 */ 7111void set_curr_task(int cpu, struct task_struct *p) 7112{ 7113 cpu_curr(cpu) = p; 7114} 7115 7116#endif 7117 7118#ifdef CONFIG_CGROUP_SCHED 7119/* task_group_lock serializes the addition/removal of task groups */ 7120static DEFINE_SPINLOCK(task_group_lock); 7121 7122static void free_sched_group(struct task_group *tg) 7123{ 7124 free_fair_sched_group(tg); 7125 free_rt_sched_group(tg); 7126 autogroup_free(tg); 7127 kfree(tg); 7128} 7129 7130/* allocate runqueue etc for a new task group */ 7131struct task_group *sched_create_group(struct task_group *parent) 7132{ 7133 struct task_group *tg; 7134 7135 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7136 if (!tg) 7137 return ERR_PTR(-ENOMEM); 7138 7139 if (!alloc_fair_sched_group(tg, parent)) 7140 goto err; 7141 7142 if (!alloc_rt_sched_group(tg, parent)) 7143 goto err; 7144 7145 return tg; 7146 7147err: 7148 free_sched_group(tg); 7149 return ERR_PTR(-ENOMEM); 7150} 7151 7152void sched_online_group(struct task_group *tg, struct task_group *parent) 7153{ 7154 unsigned long flags; 7155 7156 spin_lock_irqsave(&task_group_lock, flags); 7157 list_add_rcu(&tg->list, &task_groups); 7158 7159 WARN_ON(!parent); /* root should already exist */ 7160 7161 tg->parent = parent; 7162 INIT_LIST_HEAD(&tg->children); 7163 list_add_rcu(&tg->siblings, &parent->children); 7164 spin_unlock_irqrestore(&task_group_lock, flags); 7165} 7166 7167/* rcu callback to free various structures associated with a task group */ 7168static void free_sched_group_rcu(struct rcu_head *rhp) 7169{ 7170 /* now it should be safe to free those cfs_rqs */ 7171 free_sched_group(container_of(rhp, struct task_group, rcu)); 7172} 7173 7174/* Destroy runqueue etc associated with a task group */ 7175void sched_destroy_group(struct task_group *tg) 7176{ 7177 /* wait for possible concurrent references to cfs_rqs complete */ 7178 call_rcu(&tg->rcu, free_sched_group_rcu); 7179} 7180 7181void sched_offline_group(struct task_group *tg) 7182{ 7183 unsigned long flags; 7184 int i; 7185 7186 /* end participation in shares distribution */ 7187 for_each_possible_cpu(i) 7188 unregister_fair_sched_group(tg, i); 7189 7190 spin_lock_irqsave(&task_group_lock, flags); 7191 list_del_rcu(&tg->list); 7192 list_del_rcu(&tg->siblings); 7193 spin_unlock_irqrestore(&task_group_lock, flags); 7194} 7195 7196/* change task's runqueue when it moves between groups. 7197 * The caller of this function should have put the task in its new group 7198 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7199 * reflect its new group. 7200 */ 7201void sched_move_task(struct task_struct *tsk) 7202{ 7203 struct task_group *tg; 7204 int on_rq, running; 7205 unsigned long flags; 7206 struct rq *rq; 7207 7208 rq = task_rq_lock(tsk, &flags); 7209 7210 running = task_current(rq, tsk); 7211 on_rq = tsk->on_rq; 7212 7213 if (on_rq) 7214 dequeue_task(rq, tsk, 0); 7215 if (unlikely(running)) 7216 tsk->sched_class->put_prev_task(rq, tsk); 7217 7218 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 7219 lockdep_is_held(&tsk->sighand->siglock)), 7220 struct task_group, css); 7221 tg = autogroup_task_group(tsk, tg); 7222 tsk->sched_task_group = tg; 7223 7224#ifdef CONFIG_FAIR_GROUP_SCHED 7225 if (tsk->sched_class->task_move_group) 7226 tsk->sched_class->task_move_group(tsk, on_rq); 7227 else 7228#endif 7229 set_task_rq(tsk, task_cpu(tsk)); 7230 7231 if (unlikely(running)) 7232 tsk->sched_class->set_curr_task(rq); 7233 if (on_rq) 7234 enqueue_task(rq, tsk, 0); 7235 7236 task_rq_unlock(rq, tsk, &flags); 7237} 7238#endif /* CONFIG_CGROUP_SCHED */ 7239 7240#ifdef CONFIG_RT_GROUP_SCHED 7241/* 7242 * Ensure that the real time constraints are schedulable. 7243 */ 7244static DEFINE_MUTEX(rt_constraints_mutex); 7245 7246/* Must be called with tasklist_lock held */ 7247static inline int tg_has_rt_tasks(struct task_group *tg) 7248{ 7249 struct task_struct *g, *p; 7250 7251 do_each_thread(g, p) { 7252 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7253 return 1; 7254 } while_each_thread(g, p); 7255 7256 return 0; 7257} 7258 7259struct rt_schedulable_data { 7260 struct task_group *tg; 7261 u64 rt_period; 7262 u64 rt_runtime; 7263}; 7264 7265static int tg_rt_schedulable(struct task_group *tg, void *data) 7266{ 7267 struct rt_schedulable_data *d = data; 7268 struct task_group *child; 7269 unsigned long total, sum = 0; 7270 u64 period, runtime; 7271 7272 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7273 runtime = tg->rt_bandwidth.rt_runtime; 7274 7275 if (tg == d->tg) { 7276 period = d->rt_period; 7277 runtime = d->rt_runtime; 7278 } 7279 7280 /* 7281 * Cannot have more runtime than the period. 7282 */ 7283 if (runtime > period && runtime != RUNTIME_INF) 7284 return -EINVAL; 7285 7286 /* 7287 * Ensure we don't starve existing RT tasks. 7288 */ 7289 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7290 return -EBUSY; 7291 7292 total = to_ratio(period, runtime); 7293 7294 /* 7295 * Nobody can have more than the global setting allows. 7296 */ 7297 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7298 return -EINVAL; 7299 7300 /* 7301 * The sum of our children's runtime should not exceed our own. 7302 */ 7303 list_for_each_entry_rcu(child, &tg->children, siblings) { 7304 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7305 runtime = child->rt_bandwidth.rt_runtime; 7306 7307 if (child == d->tg) { 7308 period = d->rt_period; 7309 runtime = d->rt_runtime; 7310 } 7311 7312 sum += to_ratio(period, runtime); 7313 } 7314 7315 if (sum > total) 7316 return -EINVAL; 7317 7318 return 0; 7319} 7320 7321static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7322{ 7323 int ret; 7324 7325 struct rt_schedulable_data data = { 7326 .tg = tg, 7327 .rt_period = period, 7328 .rt_runtime = runtime, 7329 }; 7330 7331 rcu_read_lock(); 7332 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7333 rcu_read_unlock(); 7334 7335 return ret; 7336} 7337 7338static int tg_set_rt_bandwidth(struct task_group *tg, 7339 u64 rt_period, u64 rt_runtime) 7340{ 7341 int i, err = 0; 7342 7343 mutex_lock(&rt_constraints_mutex); 7344 read_lock(&tasklist_lock); 7345 err = __rt_schedulable(tg, rt_period, rt_runtime); 7346 if (err) 7347 goto unlock; 7348 7349 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7350 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7351 tg->rt_bandwidth.rt_runtime = rt_runtime; 7352 7353 for_each_possible_cpu(i) { 7354 struct rt_rq *rt_rq = tg->rt_rq[i]; 7355 7356 raw_spin_lock(&rt_rq->rt_runtime_lock); 7357 rt_rq->rt_runtime = rt_runtime; 7358 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7359 } 7360 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7361unlock: 7362 read_unlock(&tasklist_lock); 7363 mutex_unlock(&rt_constraints_mutex); 7364 7365 return err; 7366} 7367 7368static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7369{ 7370 u64 rt_runtime, rt_period; 7371 7372 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7373 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7374 if (rt_runtime_us < 0) 7375 rt_runtime = RUNTIME_INF; 7376 7377 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7378} 7379 7380static long sched_group_rt_runtime(struct task_group *tg) 7381{ 7382 u64 rt_runtime_us; 7383 7384 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7385 return -1; 7386 7387 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7388 do_div(rt_runtime_us, NSEC_PER_USEC); 7389 return rt_runtime_us; 7390} 7391 7392static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7393{ 7394 u64 rt_runtime, rt_period; 7395 7396 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7397 rt_runtime = tg->rt_bandwidth.rt_runtime; 7398 7399 if (rt_period == 0) 7400 return -EINVAL; 7401 7402 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7403} 7404 7405static long sched_group_rt_period(struct task_group *tg) 7406{ 7407 u64 rt_period_us; 7408 7409 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7410 do_div(rt_period_us, NSEC_PER_USEC); 7411 return rt_period_us; 7412} 7413#endif /* CONFIG_RT_GROUP_SCHED */ 7414 7415#ifdef CONFIG_RT_GROUP_SCHED 7416static int sched_rt_global_constraints(void) 7417{ 7418 int ret = 0; 7419 7420 mutex_lock(&rt_constraints_mutex); 7421 read_lock(&tasklist_lock); 7422 ret = __rt_schedulable(NULL, 0, 0); 7423 read_unlock(&tasklist_lock); 7424 mutex_unlock(&rt_constraints_mutex); 7425 7426 return ret; 7427} 7428 7429static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7430{ 7431 /* Don't accept realtime tasks when there is no way for them to run */ 7432 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7433 return 0; 7434 7435 return 1; 7436} 7437 7438#else /* !CONFIG_RT_GROUP_SCHED */ 7439static int sched_rt_global_constraints(void) 7440{ 7441 unsigned long flags; 7442 int i, ret = 0; 7443 7444 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7445 for_each_possible_cpu(i) { 7446 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7447 7448 raw_spin_lock(&rt_rq->rt_runtime_lock); 7449 rt_rq->rt_runtime = global_rt_runtime(); 7450 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7451 } 7452 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7453 7454 return ret; 7455} 7456#endif /* CONFIG_RT_GROUP_SCHED */ 7457 7458static int sched_dl_global_constraints(void) 7459{ 7460 u64 runtime = global_rt_runtime(); 7461 u64 period = global_rt_period(); 7462 u64 new_bw = to_ratio(period, runtime); 7463 int cpu, ret = 0; 7464 7465 /* 7466 * Here we want to check the bandwidth not being set to some 7467 * value smaller than the currently allocated bandwidth in 7468 * any of the root_domains. 7469 * 7470 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7471 * cycling on root_domains... Discussion on different/better 7472 * solutions is welcome! 7473 */ 7474 for_each_possible_cpu(cpu) { 7475 struct dl_bw *dl_b = dl_bw_of(cpu); 7476 7477 raw_spin_lock(&dl_b->lock); 7478 if (new_bw < dl_b->total_bw) 7479 ret = -EBUSY; 7480 raw_spin_unlock(&dl_b->lock); 7481 7482 if (ret) 7483 break; 7484 } 7485 7486 return ret; 7487} 7488 7489static void sched_dl_do_global(void) 7490{ 7491 u64 new_bw = -1; 7492 int cpu; 7493 7494 def_dl_bandwidth.dl_period = global_rt_period(); 7495 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7496 7497 if (global_rt_runtime() != RUNTIME_INF) 7498 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7499 7500 /* 7501 * FIXME: As above... 7502 */ 7503 for_each_possible_cpu(cpu) { 7504 struct dl_bw *dl_b = dl_bw_of(cpu); 7505 7506 raw_spin_lock(&dl_b->lock); 7507 dl_b->bw = new_bw; 7508 raw_spin_unlock(&dl_b->lock); 7509 } 7510} 7511 7512static int sched_rt_global_validate(void) 7513{ 7514 if (sysctl_sched_rt_period <= 0) 7515 return -EINVAL; 7516 7517 if (sysctl_sched_rt_runtime > sysctl_sched_rt_period) 7518 return -EINVAL; 7519 7520 return 0; 7521} 7522 7523static void sched_rt_do_global(void) 7524{ 7525 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7526 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7527} 7528 7529int sched_rt_handler(struct ctl_table *table, int write, 7530 void __user *buffer, size_t *lenp, 7531 loff_t *ppos) 7532{ 7533 int old_period, old_runtime; 7534 static DEFINE_MUTEX(mutex); 7535 int ret; 7536 7537 mutex_lock(&mutex); 7538 old_period = sysctl_sched_rt_period; 7539 old_runtime = sysctl_sched_rt_runtime; 7540 7541 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7542 7543 if (!ret && write) { 7544 ret = sched_rt_global_validate(); 7545 if (ret) 7546 goto undo; 7547 7548 ret = sched_rt_global_constraints(); 7549 if (ret) 7550 goto undo; 7551 7552 ret = sched_dl_global_constraints(); 7553 if (ret) 7554 goto undo; 7555 7556 sched_rt_do_global(); 7557 sched_dl_do_global(); 7558 } 7559 if (0) { 7560undo: 7561 sysctl_sched_rt_period = old_period; 7562 sysctl_sched_rt_runtime = old_runtime; 7563 } 7564 mutex_unlock(&mutex); 7565 7566 return ret; 7567} 7568 7569int sched_rr_handler(struct ctl_table *table, int write, 7570 void __user *buffer, size_t *lenp, 7571 loff_t *ppos) 7572{ 7573 int ret; 7574 static DEFINE_MUTEX(mutex); 7575 7576 mutex_lock(&mutex); 7577 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7578 /* make sure that internally we keep jiffies */ 7579 /* also, writing zero resets timeslice to default */ 7580 if (!ret && write) { 7581 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7582 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7583 } 7584 mutex_unlock(&mutex); 7585 return ret; 7586} 7587 7588#ifdef CONFIG_CGROUP_SCHED 7589 7590static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7591{ 7592 return css ? container_of(css, struct task_group, css) : NULL; 7593} 7594 7595static struct cgroup_subsys_state * 7596cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7597{ 7598 struct task_group *parent = css_tg(parent_css); 7599 struct task_group *tg; 7600 7601 if (!parent) { 7602 /* This is early initialization for the top cgroup */ 7603 return &root_task_group.css; 7604 } 7605 7606 tg = sched_create_group(parent); 7607 if (IS_ERR(tg)) 7608 return ERR_PTR(-ENOMEM); 7609 7610 return &tg->css; 7611} 7612 7613static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7614{ 7615 struct task_group *tg = css_tg(css); 7616 struct task_group *parent = css_tg(css_parent(css)); 7617 7618 if (parent) 7619 sched_online_group(tg, parent); 7620 return 0; 7621} 7622 7623static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7624{ 7625 struct task_group *tg = css_tg(css); 7626 7627 sched_destroy_group(tg); 7628} 7629 7630static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7631{ 7632 struct task_group *tg = css_tg(css); 7633 7634 sched_offline_group(tg); 7635} 7636 7637static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7638 struct cgroup_taskset *tset) 7639{ 7640 struct task_struct *task; 7641 7642 cgroup_taskset_for_each(task, css, tset) { 7643#ifdef CONFIG_RT_GROUP_SCHED 7644 if (!sched_rt_can_attach(css_tg(css), task)) 7645 return -EINVAL; 7646#else 7647 /* We don't support RT-tasks being in separate groups */ 7648 if (task->sched_class != &fair_sched_class) 7649 return -EINVAL; 7650#endif 7651 } 7652 return 0; 7653} 7654 7655static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7656 struct cgroup_taskset *tset) 7657{ 7658 struct task_struct *task; 7659 7660 cgroup_taskset_for_each(task, css, tset) 7661 sched_move_task(task); 7662} 7663 7664static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7665 struct cgroup_subsys_state *old_css, 7666 struct task_struct *task) 7667{ 7668 /* 7669 * cgroup_exit() is called in the copy_process() failure path. 7670 * Ignore this case since the task hasn't ran yet, this avoids 7671 * trying to poke a half freed task state from generic code. 7672 */ 7673 if (!(task->flags & PF_EXITING)) 7674 return; 7675 7676 sched_move_task(task); 7677} 7678 7679#ifdef CONFIG_FAIR_GROUP_SCHED 7680static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7681 struct cftype *cftype, u64 shareval) 7682{ 7683 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7684} 7685 7686static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7687 struct cftype *cft) 7688{ 7689 struct task_group *tg = css_tg(css); 7690 7691 return (u64) scale_load_down(tg->shares); 7692} 7693 7694#ifdef CONFIG_CFS_BANDWIDTH 7695static DEFINE_MUTEX(cfs_constraints_mutex); 7696 7697const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7698const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7699 7700static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7701 7702static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7703{ 7704 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7705 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7706 7707 if (tg == &root_task_group) 7708 return -EINVAL; 7709 7710 /* 7711 * Ensure we have at some amount of bandwidth every period. This is 7712 * to prevent reaching a state of large arrears when throttled via 7713 * entity_tick() resulting in prolonged exit starvation. 7714 */ 7715 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7716 return -EINVAL; 7717 7718 /* 7719 * Likewise, bound things on the otherside by preventing insane quota 7720 * periods. This also allows us to normalize in computing quota 7721 * feasibility. 7722 */ 7723 if (period > max_cfs_quota_period) 7724 return -EINVAL; 7725 7726 mutex_lock(&cfs_constraints_mutex); 7727 ret = __cfs_schedulable(tg, period, quota); 7728 if (ret) 7729 goto out_unlock; 7730 7731 runtime_enabled = quota != RUNTIME_INF; 7732 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7733 /* 7734 * If we need to toggle cfs_bandwidth_used, off->on must occur 7735 * before making related changes, and on->off must occur afterwards 7736 */ 7737 if (runtime_enabled && !runtime_was_enabled) 7738 cfs_bandwidth_usage_inc(); 7739 raw_spin_lock_irq(&cfs_b->lock); 7740 cfs_b->period = ns_to_ktime(period); 7741 cfs_b->quota = quota; 7742 7743 __refill_cfs_bandwidth_runtime(cfs_b); 7744 /* restart the period timer (if active) to handle new period expiry */ 7745 if (runtime_enabled && cfs_b->timer_active) { 7746 /* force a reprogram */ 7747 cfs_b->timer_active = 0; 7748 __start_cfs_bandwidth(cfs_b); 7749 } 7750 raw_spin_unlock_irq(&cfs_b->lock); 7751 7752 for_each_possible_cpu(i) { 7753 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7754 struct rq *rq = cfs_rq->rq; 7755 7756 raw_spin_lock_irq(&rq->lock); 7757 cfs_rq->runtime_enabled = runtime_enabled; 7758 cfs_rq->runtime_remaining = 0; 7759 7760 if (cfs_rq->throttled) 7761 unthrottle_cfs_rq(cfs_rq); 7762 raw_spin_unlock_irq(&rq->lock); 7763 } 7764 if (runtime_was_enabled && !runtime_enabled) 7765 cfs_bandwidth_usage_dec(); 7766out_unlock: 7767 mutex_unlock(&cfs_constraints_mutex); 7768 7769 return ret; 7770} 7771 7772int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7773{ 7774 u64 quota, period; 7775 7776 period = ktime_to_ns(tg->cfs_bandwidth.period); 7777 if (cfs_quota_us < 0) 7778 quota = RUNTIME_INF; 7779 else 7780 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7781 7782 return tg_set_cfs_bandwidth(tg, period, quota); 7783} 7784 7785long tg_get_cfs_quota(struct task_group *tg) 7786{ 7787 u64 quota_us; 7788 7789 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7790 return -1; 7791 7792 quota_us = tg->cfs_bandwidth.quota; 7793 do_div(quota_us, NSEC_PER_USEC); 7794 7795 return quota_us; 7796} 7797 7798int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7799{ 7800 u64 quota, period; 7801 7802 period = (u64)cfs_period_us * NSEC_PER_USEC; 7803 quota = tg->cfs_bandwidth.quota; 7804 7805 return tg_set_cfs_bandwidth(tg, period, quota); 7806} 7807 7808long tg_get_cfs_period(struct task_group *tg) 7809{ 7810 u64 cfs_period_us; 7811 7812 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7813 do_div(cfs_period_us, NSEC_PER_USEC); 7814 7815 return cfs_period_us; 7816} 7817 7818static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7819 struct cftype *cft) 7820{ 7821 return tg_get_cfs_quota(css_tg(css)); 7822} 7823 7824static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7825 struct cftype *cftype, s64 cfs_quota_us) 7826{ 7827 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7828} 7829 7830static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7831 struct cftype *cft) 7832{ 7833 return tg_get_cfs_period(css_tg(css)); 7834} 7835 7836static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7837 struct cftype *cftype, u64 cfs_period_us) 7838{ 7839 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7840} 7841 7842struct cfs_schedulable_data { 7843 struct task_group *tg; 7844 u64 period, quota; 7845}; 7846 7847/* 7848 * normalize group quota/period to be quota/max_period 7849 * note: units are usecs 7850 */ 7851static u64 normalize_cfs_quota(struct task_group *tg, 7852 struct cfs_schedulable_data *d) 7853{ 7854 u64 quota, period; 7855 7856 if (tg == d->tg) { 7857 period = d->period; 7858 quota = d->quota; 7859 } else { 7860 period = tg_get_cfs_period(tg); 7861 quota = tg_get_cfs_quota(tg); 7862 } 7863 7864 /* note: these should typically be equivalent */ 7865 if (quota == RUNTIME_INF || quota == -1) 7866 return RUNTIME_INF; 7867 7868 return to_ratio(period, quota); 7869} 7870 7871static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7872{ 7873 struct cfs_schedulable_data *d = data; 7874 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7875 s64 quota = 0, parent_quota = -1; 7876 7877 if (!tg->parent) { 7878 quota = RUNTIME_INF; 7879 } else { 7880 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7881 7882 quota = normalize_cfs_quota(tg, d); 7883 parent_quota = parent_b->hierarchal_quota; 7884 7885 /* 7886 * ensure max(child_quota) <= parent_quota, inherit when no 7887 * limit is set 7888 */ 7889 if (quota == RUNTIME_INF) 7890 quota = parent_quota; 7891 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7892 return -EINVAL; 7893 } 7894 cfs_b->hierarchal_quota = quota; 7895 7896 return 0; 7897} 7898 7899static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7900{ 7901 int ret; 7902 struct cfs_schedulable_data data = { 7903 .tg = tg, 7904 .period = period, 7905 .quota = quota, 7906 }; 7907 7908 if (quota != RUNTIME_INF) { 7909 do_div(data.period, NSEC_PER_USEC); 7910 do_div(data.quota, NSEC_PER_USEC); 7911 } 7912 7913 rcu_read_lock(); 7914 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7915 rcu_read_unlock(); 7916 7917 return ret; 7918} 7919 7920static int cpu_stats_show(struct seq_file *sf, void *v) 7921{ 7922 struct task_group *tg = css_tg(seq_css(sf)); 7923 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7924 7925 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7926 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7927 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7928 7929 return 0; 7930} 7931#endif /* CONFIG_CFS_BANDWIDTH */ 7932#endif /* CONFIG_FAIR_GROUP_SCHED */ 7933 7934#ifdef CONFIG_RT_GROUP_SCHED 7935static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7936 struct cftype *cft, s64 val) 7937{ 7938 return sched_group_set_rt_runtime(css_tg(css), val); 7939} 7940 7941static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7942 struct cftype *cft) 7943{ 7944 return sched_group_rt_runtime(css_tg(css)); 7945} 7946 7947static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7948 struct cftype *cftype, u64 rt_period_us) 7949{ 7950 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7951} 7952 7953static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7954 struct cftype *cft) 7955{ 7956 return sched_group_rt_period(css_tg(css)); 7957} 7958#endif /* CONFIG_RT_GROUP_SCHED */ 7959 7960static struct cftype cpu_files[] = { 7961#ifdef CONFIG_FAIR_GROUP_SCHED 7962 { 7963 .name = "shares", 7964 .read_u64 = cpu_shares_read_u64, 7965 .write_u64 = cpu_shares_write_u64, 7966 }, 7967#endif 7968#ifdef CONFIG_CFS_BANDWIDTH 7969 { 7970 .name = "cfs_quota_us", 7971 .read_s64 = cpu_cfs_quota_read_s64, 7972 .write_s64 = cpu_cfs_quota_write_s64, 7973 }, 7974 { 7975 .name = "cfs_period_us", 7976 .read_u64 = cpu_cfs_period_read_u64, 7977 .write_u64 = cpu_cfs_period_write_u64, 7978 }, 7979 { 7980 .name = "stat", 7981 .seq_show = cpu_stats_show, 7982 }, 7983#endif 7984#ifdef CONFIG_RT_GROUP_SCHED 7985 { 7986 .name = "rt_runtime_us", 7987 .read_s64 = cpu_rt_runtime_read, 7988 .write_s64 = cpu_rt_runtime_write, 7989 }, 7990 { 7991 .name = "rt_period_us", 7992 .read_u64 = cpu_rt_period_read_uint, 7993 .write_u64 = cpu_rt_period_write_uint, 7994 }, 7995#endif 7996 { } /* terminate */ 7997}; 7998 7999struct cgroup_subsys cpu_cgroup_subsys = { 8000 .name = "cpu", 8001 .css_alloc = cpu_cgroup_css_alloc, 8002 .css_free = cpu_cgroup_css_free, 8003 .css_online = cpu_cgroup_css_online, 8004 .css_offline = cpu_cgroup_css_offline, 8005 .can_attach = cpu_cgroup_can_attach, 8006 .attach = cpu_cgroup_attach, 8007 .exit = cpu_cgroup_exit, 8008 .subsys_id = cpu_cgroup_subsys_id, 8009 .base_cftypes = cpu_files, 8010 .early_init = 1, 8011}; 8012 8013#endif /* CONFIG_CGROUP_SCHED */ 8014 8015void dump_cpu_task(int cpu) 8016{ 8017 pr_info("Task dump for CPU %d:\n", cpu); 8018 sched_show_task(cpu_curr(cpu)); 8019} 8020