core.c revision d8d28c8f00e84a72e8bee39a85835635417bee49
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#include <linux/compiler.h> 77 78#include <asm/switch_to.h> 79#include <asm/tlb.h> 80#include <asm/irq_regs.h> 81#include <asm/mutex.h> 82#ifdef CONFIG_PARAVIRT 83#include <asm/paravirt.h> 84#endif 85 86#include "sched.h" 87#include "../workqueue_internal.h" 88#include "../smpboot.h" 89 90#define CREATE_TRACE_POINTS 91#include <trace/events/sched.h> 92 93#ifdef smp_mb__before_atomic 94void __smp_mb__before_atomic(void) 95{ 96 smp_mb__before_atomic(); 97} 98EXPORT_SYMBOL(__smp_mb__before_atomic); 99#endif 100 101#ifdef smp_mb__after_atomic 102void __smp_mb__after_atomic(void) 103{ 104 smp_mb__after_atomic(); 105} 106EXPORT_SYMBOL(__smp_mb__after_atomic); 107#endif 108 109void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 110{ 111 unsigned long delta; 112 ktime_t soft, hard, now; 113 114 for (;;) { 115 if (hrtimer_active(period_timer)) 116 break; 117 118 now = hrtimer_cb_get_time(period_timer); 119 hrtimer_forward(period_timer, now, period); 120 121 soft = hrtimer_get_softexpires(period_timer); 122 hard = hrtimer_get_expires(period_timer); 123 delta = ktime_to_ns(ktime_sub(hard, soft)); 124 __hrtimer_start_range_ns(period_timer, soft, delta, 125 HRTIMER_MODE_ABS_PINNED, 0); 126 } 127} 128 129DEFINE_MUTEX(sched_domains_mutex); 130DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 131 132static void update_rq_clock_task(struct rq *rq, s64 delta); 133 134void update_rq_clock(struct rq *rq) 135{ 136 s64 delta; 137 138 if (rq->skip_clock_update > 0) 139 return; 140 141 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 142 rq->clock += delta; 143 update_rq_clock_task(rq, delta); 144} 145 146/* 147 * Debugging: various feature bits 148 */ 149 150#define SCHED_FEAT(name, enabled) \ 151 (1UL << __SCHED_FEAT_##name) * enabled | 152 153const_debug unsigned int sysctl_sched_features = 154#include "features.h" 155 0; 156 157#undef SCHED_FEAT 158 159#ifdef CONFIG_SCHED_DEBUG 160#define SCHED_FEAT(name, enabled) \ 161 #name , 162 163static const char * const sched_feat_names[] = { 164#include "features.h" 165}; 166 167#undef SCHED_FEAT 168 169static int sched_feat_show(struct seq_file *m, void *v) 170{ 171 int i; 172 173 for (i = 0; i < __SCHED_FEAT_NR; i++) { 174 if (!(sysctl_sched_features & (1UL << i))) 175 seq_puts(m, "NO_"); 176 seq_printf(m, "%s ", sched_feat_names[i]); 177 } 178 seq_puts(m, "\n"); 179 180 return 0; 181} 182 183#ifdef HAVE_JUMP_LABEL 184 185#define jump_label_key__true STATIC_KEY_INIT_TRUE 186#define jump_label_key__false STATIC_KEY_INIT_FALSE 187 188#define SCHED_FEAT(name, enabled) \ 189 jump_label_key__##enabled , 190 191struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 192#include "features.h" 193}; 194 195#undef SCHED_FEAT 196 197static void sched_feat_disable(int i) 198{ 199 if (static_key_enabled(&sched_feat_keys[i])) 200 static_key_slow_dec(&sched_feat_keys[i]); 201} 202 203static void sched_feat_enable(int i) 204{ 205 if (!static_key_enabled(&sched_feat_keys[i])) 206 static_key_slow_inc(&sched_feat_keys[i]); 207} 208#else 209static void sched_feat_disable(int i) { }; 210static void sched_feat_enable(int i) { }; 211#endif /* HAVE_JUMP_LABEL */ 212 213static int sched_feat_set(char *cmp) 214{ 215 int i; 216 int neg = 0; 217 218 if (strncmp(cmp, "NO_", 3) == 0) { 219 neg = 1; 220 cmp += 3; 221 } 222 223 for (i = 0; i < __SCHED_FEAT_NR; i++) { 224 if (strcmp(cmp, sched_feat_names[i]) == 0) { 225 if (neg) { 226 sysctl_sched_features &= ~(1UL << i); 227 sched_feat_disable(i); 228 } else { 229 sysctl_sched_features |= (1UL << i); 230 sched_feat_enable(i); 231 } 232 break; 233 } 234 } 235 236 return i; 237} 238 239static ssize_t 240sched_feat_write(struct file *filp, const char __user *ubuf, 241 size_t cnt, loff_t *ppos) 242{ 243 char buf[64]; 244 char *cmp; 245 int i; 246 247 if (cnt > 63) 248 cnt = 63; 249 250 if (copy_from_user(&buf, ubuf, cnt)) 251 return -EFAULT; 252 253 buf[cnt] = 0; 254 cmp = strstrip(buf); 255 256 i = sched_feat_set(cmp); 257 if (i == __SCHED_FEAT_NR) 258 return -EINVAL; 259 260 *ppos += cnt; 261 262 return cnt; 263} 264 265static int sched_feat_open(struct inode *inode, struct file *filp) 266{ 267 return single_open(filp, sched_feat_show, NULL); 268} 269 270static const struct file_operations sched_feat_fops = { 271 .open = sched_feat_open, 272 .write = sched_feat_write, 273 .read = seq_read, 274 .llseek = seq_lseek, 275 .release = single_release, 276}; 277 278static __init int sched_init_debug(void) 279{ 280 debugfs_create_file("sched_features", 0644, NULL, NULL, 281 &sched_feat_fops); 282 283 return 0; 284} 285late_initcall(sched_init_debug); 286#endif /* CONFIG_SCHED_DEBUG */ 287 288/* 289 * Number of tasks to iterate in a single balance run. 290 * Limited because this is done with IRQs disabled. 291 */ 292const_debug unsigned int sysctl_sched_nr_migrate = 32; 293 294/* 295 * period over which we average the RT time consumption, measured 296 * in ms. 297 * 298 * default: 1s 299 */ 300const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 301 302/* 303 * period over which we measure -rt task cpu usage in us. 304 * default: 1s 305 */ 306unsigned int sysctl_sched_rt_period = 1000000; 307 308__read_mostly int scheduler_running; 309 310/* 311 * part of the period that we allow rt tasks to run in us. 312 * default: 0.95s 313 */ 314int sysctl_sched_rt_runtime = 950000; 315 316/* 317 * __task_rq_lock - lock the rq @p resides on. 318 */ 319static inline struct rq *__task_rq_lock(struct task_struct *p) 320 __acquires(rq->lock) 321{ 322 struct rq *rq; 323 324 lockdep_assert_held(&p->pi_lock); 325 326 for (;;) { 327 rq = task_rq(p); 328 raw_spin_lock(&rq->lock); 329 if (likely(rq == task_rq(p))) 330 return rq; 331 raw_spin_unlock(&rq->lock); 332 } 333} 334 335/* 336 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 337 */ 338static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 339 __acquires(p->pi_lock) 340 __acquires(rq->lock) 341{ 342 struct rq *rq; 343 344 for (;;) { 345 raw_spin_lock_irqsave(&p->pi_lock, *flags); 346 rq = task_rq(p); 347 raw_spin_lock(&rq->lock); 348 if (likely(rq == task_rq(p))) 349 return rq; 350 raw_spin_unlock(&rq->lock); 351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 352 } 353} 354 355static void __task_rq_unlock(struct rq *rq) 356 __releases(rq->lock) 357{ 358 raw_spin_unlock(&rq->lock); 359} 360 361static inline void 362task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 363 __releases(rq->lock) 364 __releases(p->pi_lock) 365{ 366 raw_spin_unlock(&rq->lock); 367 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 368} 369 370/* 371 * this_rq_lock - lock this runqueue and disable interrupts. 372 */ 373static struct rq *this_rq_lock(void) 374 __acquires(rq->lock) 375{ 376 struct rq *rq; 377 378 local_irq_disable(); 379 rq = this_rq(); 380 raw_spin_lock(&rq->lock); 381 382 return rq; 383} 384 385#ifdef CONFIG_SCHED_HRTICK 386/* 387 * Use HR-timers to deliver accurate preemption points. 388 */ 389 390static void hrtick_clear(struct rq *rq) 391{ 392 if (hrtimer_active(&rq->hrtick_timer)) 393 hrtimer_cancel(&rq->hrtick_timer); 394} 395 396/* 397 * High-resolution timer tick. 398 * Runs from hardirq context with interrupts disabled. 399 */ 400static enum hrtimer_restart hrtick(struct hrtimer *timer) 401{ 402 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 403 404 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 405 406 raw_spin_lock(&rq->lock); 407 update_rq_clock(rq); 408 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 409 raw_spin_unlock(&rq->lock); 410 411 return HRTIMER_NORESTART; 412} 413 414#ifdef CONFIG_SMP 415 416static int __hrtick_restart(struct rq *rq) 417{ 418 struct hrtimer *timer = &rq->hrtick_timer; 419 ktime_t time = hrtimer_get_softexpires(timer); 420 421 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 422} 423 424/* 425 * called from hardirq (IPI) context 426 */ 427static void __hrtick_start(void *arg) 428{ 429 struct rq *rq = arg; 430 431 raw_spin_lock(&rq->lock); 432 __hrtick_restart(rq); 433 rq->hrtick_csd_pending = 0; 434 raw_spin_unlock(&rq->lock); 435} 436 437/* 438 * Called to set the hrtick timer state. 439 * 440 * called with rq->lock held and irqs disabled 441 */ 442void hrtick_start(struct rq *rq, u64 delay) 443{ 444 struct hrtimer *timer = &rq->hrtick_timer; 445 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 446 447 hrtimer_set_expires(timer, time); 448 449 if (rq == this_rq()) { 450 __hrtick_restart(rq); 451 } else if (!rq->hrtick_csd_pending) { 452 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 453 rq->hrtick_csd_pending = 1; 454 } 455} 456 457static int 458hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 459{ 460 int cpu = (int)(long)hcpu; 461 462 switch (action) { 463 case CPU_UP_CANCELED: 464 case CPU_UP_CANCELED_FROZEN: 465 case CPU_DOWN_PREPARE: 466 case CPU_DOWN_PREPARE_FROZEN: 467 case CPU_DEAD: 468 case CPU_DEAD_FROZEN: 469 hrtick_clear(cpu_rq(cpu)); 470 return NOTIFY_OK; 471 } 472 473 return NOTIFY_DONE; 474} 475 476static __init void init_hrtick(void) 477{ 478 hotcpu_notifier(hotplug_hrtick, 0); 479} 480#else 481/* 482 * Called to set the hrtick timer state. 483 * 484 * called with rq->lock held and irqs disabled 485 */ 486void hrtick_start(struct rq *rq, u64 delay) 487{ 488 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 489 HRTIMER_MODE_REL_PINNED, 0); 490} 491 492static inline void init_hrtick(void) 493{ 494} 495#endif /* CONFIG_SMP */ 496 497static void init_rq_hrtick(struct rq *rq) 498{ 499#ifdef CONFIG_SMP 500 rq->hrtick_csd_pending = 0; 501 502 rq->hrtick_csd.flags = 0; 503 rq->hrtick_csd.func = __hrtick_start; 504 rq->hrtick_csd.info = rq; 505#endif 506 507 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 508 rq->hrtick_timer.function = hrtick; 509} 510#else /* CONFIG_SCHED_HRTICK */ 511static inline void hrtick_clear(struct rq *rq) 512{ 513} 514 515static inline void init_rq_hrtick(struct rq *rq) 516{ 517} 518 519static inline void init_hrtick(void) 520{ 521} 522#endif /* CONFIG_SCHED_HRTICK */ 523 524/* 525 * cmpxchg based fetch_or, macro so it works for different integer types 526 */ 527#define fetch_or(ptr, val) \ 528({ typeof(*(ptr)) __old, __val = *(ptr); \ 529 for (;;) { \ 530 __old = cmpxchg((ptr), __val, __val | (val)); \ 531 if (__old == __val) \ 532 break; \ 533 __val = __old; \ 534 } \ 535 __old; \ 536}) 537 538#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) 539/* 540 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 541 * this avoids any races wrt polling state changes and thereby avoids 542 * spurious IPIs. 543 */ 544static bool set_nr_and_not_polling(struct task_struct *p) 545{ 546 struct thread_info *ti = task_thread_info(p); 547 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 548} 549 550/* 551 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. 552 * 553 * If this returns true, then the idle task promises to call 554 * sched_ttwu_pending() and reschedule soon. 555 */ 556static bool set_nr_if_polling(struct task_struct *p) 557{ 558 struct thread_info *ti = task_thread_info(p); 559 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags); 560 561 for (;;) { 562 if (!(val & _TIF_POLLING_NRFLAG)) 563 return false; 564 if (val & _TIF_NEED_RESCHED) 565 return true; 566 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); 567 if (old == val) 568 break; 569 val = old; 570 } 571 return true; 572} 573 574#else 575static bool set_nr_and_not_polling(struct task_struct *p) 576{ 577 set_tsk_need_resched(p); 578 return true; 579} 580 581#ifdef CONFIG_SMP 582static bool set_nr_if_polling(struct task_struct *p) 583{ 584 return false; 585} 586#endif 587#endif 588 589/* 590 * resched_task - mark a task 'to be rescheduled now'. 591 * 592 * On UP this means the setting of the need_resched flag, on SMP it 593 * might also involve a cross-CPU call to trigger the scheduler on 594 * the target CPU. 595 */ 596void resched_task(struct task_struct *p) 597{ 598 int cpu; 599 600 lockdep_assert_held(&task_rq(p)->lock); 601 602 if (test_tsk_need_resched(p)) 603 return; 604 605 cpu = task_cpu(p); 606 607 if (cpu == smp_processor_id()) { 608 set_tsk_need_resched(p); 609 set_preempt_need_resched(); 610 return; 611 } 612 613 if (set_nr_and_not_polling(p)) 614 smp_send_reschedule(cpu); 615 else 616 trace_sched_wake_idle_without_ipi(cpu); 617} 618 619void resched_cpu(int cpu) 620{ 621 struct rq *rq = cpu_rq(cpu); 622 unsigned long flags; 623 624 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 625 return; 626 resched_task(cpu_curr(cpu)); 627 raw_spin_unlock_irqrestore(&rq->lock, flags); 628} 629 630#ifdef CONFIG_SMP 631#ifdef CONFIG_NO_HZ_COMMON 632/* 633 * In the semi idle case, use the nearest busy cpu for migrating timers 634 * from an idle cpu. This is good for power-savings. 635 * 636 * We don't do similar optimization for completely idle system, as 637 * selecting an idle cpu will add more delays to the timers than intended 638 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 639 */ 640int get_nohz_timer_target(int pinned) 641{ 642 int cpu = smp_processor_id(); 643 int i; 644 struct sched_domain *sd; 645 646 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 647 return cpu; 648 649 rcu_read_lock(); 650 for_each_domain(cpu, sd) { 651 for_each_cpu(i, sched_domain_span(sd)) { 652 if (!idle_cpu(i)) { 653 cpu = i; 654 goto unlock; 655 } 656 } 657 } 658unlock: 659 rcu_read_unlock(); 660 return cpu; 661} 662/* 663 * When add_timer_on() enqueues a timer into the timer wheel of an 664 * idle CPU then this timer might expire before the next timer event 665 * which is scheduled to wake up that CPU. In case of a completely 666 * idle system the next event might even be infinite time into the 667 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 668 * leaves the inner idle loop so the newly added timer is taken into 669 * account when the CPU goes back to idle and evaluates the timer 670 * wheel for the next timer event. 671 */ 672static void wake_up_idle_cpu(int cpu) 673{ 674 struct rq *rq = cpu_rq(cpu); 675 676 if (cpu == smp_processor_id()) 677 return; 678 679 if (set_nr_and_not_polling(rq->idle)) 680 smp_send_reschedule(cpu); 681 else 682 trace_sched_wake_idle_without_ipi(cpu); 683} 684 685static bool wake_up_full_nohz_cpu(int cpu) 686{ 687 if (tick_nohz_full_cpu(cpu)) { 688 if (cpu != smp_processor_id() || 689 tick_nohz_tick_stopped()) 690 smp_send_reschedule(cpu); 691 return true; 692 } 693 694 return false; 695} 696 697void wake_up_nohz_cpu(int cpu) 698{ 699 if (!wake_up_full_nohz_cpu(cpu)) 700 wake_up_idle_cpu(cpu); 701} 702 703static inline bool got_nohz_idle_kick(void) 704{ 705 int cpu = smp_processor_id(); 706 707 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 708 return false; 709 710 if (idle_cpu(cpu) && !need_resched()) 711 return true; 712 713 /* 714 * We can't run Idle Load Balance on this CPU for this time so we 715 * cancel it and clear NOHZ_BALANCE_KICK 716 */ 717 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 718 return false; 719} 720 721#else /* CONFIG_NO_HZ_COMMON */ 722 723static inline bool got_nohz_idle_kick(void) 724{ 725 return false; 726} 727 728#endif /* CONFIG_NO_HZ_COMMON */ 729 730#ifdef CONFIG_NO_HZ_FULL 731bool sched_can_stop_tick(void) 732{ 733 struct rq *rq; 734 735 rq = this_rq(); 736 737 /* Make sure rq->nr_running update is visible after the IPI */ 738 smp_rmb(); 739 740 /* More than one running task need preemption */ 741 if (rq->nr_running > 1) 742 return false; 743 744 return true; 745} 746#endif /* CONFIG_NO_HZ_FULL */ 747 748void sched_avg_update(struct rq *rq) 749{ 750 s64 period = sched_avg_period(); 751 752 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 753 /* 754 * Inline assembly required to prevent the compiler 755 * optimising this loop into a divmod call. 756 * See __iter_div_u64_rem() for another example of this. 757 */ 758 asm("" : "+rm" (rq->age_stamp)); 759 rq->age_stamp += period; 760 rq->rt_avg /= 2; 761 } 762} 763 764#endif /* CONFIG_SMP */ 765 766#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 767 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 768/* 769 * Iterate task_group tree rooted at *from, calling @down when first entering a 770 * node and @up when leaving it for the final time. 771 * 772 * Caller must hold rcu_lock or sufficient equivalent. 773 */ 774int walk_tg_tree_from(struct task_group *from, 775 tg_visitor down, tg_visitor up, void *data) 776{ 777 struct task_group *parent, *child; 778 int ret; 779 780 parent = from; 781 782down: 783 ret = (*down)(parent, data); 784 if (ret) 785 goto out; 786 list_for_each_entry_rcu(child, &parent->children, siblings) { 787 parent = child; 788 goto down; 789 790up: 791 continue; 792 } 793 ret = (*up)(parent, data); 794 if (ret || parent == from) 795 goto out; 796 797 child = parent; 798 parent = parent->parent; 799 if (parent) 800 goto up; 801out: 802 return ret; 803} 804 805int tg_nop(struct task_group *tg, void *data) 806{ 807 return 0; 808} 809#endif 810 811static void set_load_weight(struct task_struct *p) 812{ 813 int prio = p->static_prio - MAX_RT_PRIO; 814 struct load_weight *load = &p->se.load; 815 816 /* 817 * SCHED_IDLE tasks get minimal weight: 818 */ 819 if (p->policy == SCHED_IDLE) { 820 load->weight = scale_load(WEIGHT_IDLEPRIO); 821 load->inv_weight = WMULT_IDLEPRIO; 822 return; 823 } 824 825 load->weight = scale_load(prio_to_weight[prio]); 826 load->inv_weight = prio_to_wmult[prio]; 827} 828 829static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 830{ 831 update_rq_clock(rq); 832 sched_info_queued(rq, p); 833 p->sched_class->enqueue_task(rq, p, flags); 834} 835 836static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 837{ 838 update_rq_clock(rq); 839 sched_info_dequeued(rq, p); 840 p->sched_class->dequeue_task(rq, p, flags); 841} 842 843void activate_task(struct rq *rq, struct task_struct *p, int flags) 844{ 845 if (task_contributes_to_load(p)) 846 rq->nr_uninterruptible--; 847 848 enqueue_task(rq, p, flags); 849} 850 851void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 852{ 853 if (task_contributes_to_load(p)) 854 rq->nr_uninterruptible++; 855 856 dequeue_task(rq, p, flags); 857} 858 859static void update_rq_clock_task(struct rq *rq, s64 delta) 860{ 861/* 862 * In theory, the compile should just see 0 here, and optimize out the call 863 * to sched_rt_avg_update. But I don't trust it... 864 */ 865#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 866 s64 steal = 0, irq_delta = 0; 867#endif 868#ifdef CONFIG_IRQ_TIME_ACCOUNTING 869 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 870 871 /* 872 * Since irq_time is only updated on {soft,}irq_exit, we might run into 873 * this case when a previous update_rq_clock() happened inside a 874 * {soft,}irq region. 875 * 876 * When this happens, we stop ->clock_task and only update the 877 * prev_irq_time stamp to account for the part that fit, so that a next 878 * update will consume the rest. This ensures ->clock_task is 879 * monotonic. 880 * 881 * It does however cause some slight miss-attribution of {soft,}irq 882 * time, a more accurate solution would be to update the irq_time using 883 * the current rq->clock timestamp, except that would require using 884 * atomic ops. 885 */ 886 if (irq_delta > delta) 887 irq_delta = delta; 888 889 rq->prev_irq_time += irq_delta; 890 delta -= irq_delta; 891#endif 892#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 893 if (static_key_false((¶virt_steal_rq_enabled))) { 894 steal = paravirt_steal_clock(cpu_of(rq)); 895 steal -= rq->prev_steal_time_rq; 896 897 if (unlikely(steal > delta)) 898 steal = delta; 899 900 rq->prev_steal_time_rq += steal; 901 delta -= steal; 902 } 903#endif 904 905 rq->clock_task += delta; 906 907#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 908 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 909 sched_rt_avg_update(rq, irq_delta + steal); 910#endif 911} 912 913void sched_set_stop_task(int cpu, struct task_struct *stop) 914{ 915 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 916 struct task_struct *old_stop = cpu_rq(cpu)->stop; 917 918 if (stop) { 919 /* 920 * Make it appear like a SCHED_FIFO task, its something 921 * userspace knows about and won't get confused about. 922 * 923 * Also, it will make PI more or less work without too 924 * much confusion -- but then, stop work should not 925 * rely on PI working anyway. 926 */ 927 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 928 929 stop->sched_class = &stop_sched_class; 930 } 931 932 cpu_rq(cpu)->stop = stop; 933 934 if (old_stop) { 935 /* 936 * Reset it back to a normal scheduling class so that 937 * it can die in pieces. 938 */ 939 old_stop->sched_class = &rt_sched_class; 940 } 941} 942 943/* 944 * __normal_prio - return the priority that is based on the static prio 945 */ 946static inline int __normal_prio(struct task_struct *p) 947{ 948 return p->static_prio; 949} 950 951/* 952 * Calculate the expected normal priority: i.e. priority 953 * without taking RT-inheritance into account. Might be 954 * boosted by interactivity modifiers. Changes upon fork, 955 * setprio syscalls, and whenever the interactivity 956 * estimator recalculates. 957 */ 958static inline int normal_prio(struct task_struct *p) 959{ 960 int prio; 961 962 if (task_has_dl_policy(p)) 963 prio = MAX_DL_PRIO-1; 964 else if (task_has_rt_policy(p)) 965 prio = MAX_RT_PRIO-1 - p->rt_priority; 966 else 967 prio = __normal_prio(p); 968 return prio; 969} 970 971/* 972 * Calculate the current priority, i.e. the priority 973 * taken into account by the scheduler. This value might 974 * be boosted by RT tasks, or might be boosted by 975 * interactivity modifiers. Will be RT if the task got 976 * RT-boosted. If not then it returns p->normal_prio. 977 */ 978static int effective_prio(struct task_struct *p) 979{ 980 p->normal_prio = normal_prio(p); 981 /* 982 * If we are RT tasks or we were boosted to RT priority, 983 * keep the priority unchanged. Otherwise, update priority 984 * to the normal priority: 985 */ 986 if (!rt_prio(p->prio)) 987 return p->normal_prio; 988 return p->prio; 989} 990 991/** 992 * task_curr - is this task currently executing on a CPU? 993 * @p: the task in question. 994 * 995 * Return: 1 if the task is currently executing. 0 otherwise. 996 */ 997inline int task_curr(const struct task_struct *p) 998{ 999 return cpu_curr(task_cpu(p)) == p; 1000} 1001 1002static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1003 const struct sched_class *prev_class, 1004 int oldprio) 1005{ 1006 if (prev_class != p->sched_class) { 1007 if (prev_class->switched_from) 1008 prev_class->switched_from(rq, p); 1009 p->sched_class->switched_to(rq, p); 1010 } else if (oldprio != p->prio || dl_task(p)) 1011 p->sched_class->prio_changed(rq, p, oldprio); 1012} 1013 1014void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1015{ 1016 const struct sched_class *class; 1017 1018 if (p->sched_class == rq->curr->sched_class) { 1019 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1020 } else { 1021 for_each_class(class) { 1022 if (class == rq->curr->sched_class) 1023 break; 1024 if (class == p->sched_class) { 1025 resched_task(rq->curr); 1026 break; 1027 } 1028 } 1029 } 1030 1031 /* 1032 * A queue event has occurred, and we're going to schedule. In 1033 * this case, we can save a useless back to back clock update. 1034 */ 1035 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1036 rq->skip_clock_update = 1; 1037} 1038 1039#ifdef CONFIG_SMP 1040void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1041{ 1042#ifdef CONFIG_SCHED_DEBUG 1043 /* 1044 * We should never call set_task_cpu() on a blocked task, 1045 * ttwu() will sort out the placement. 1046 */ 1047 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1048 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 1049 1050#ifdef CONFIG_LOCKDEP 1051 /* 1052 * The caller should hold either p->pi_lock or rq->lock, when changing 1053 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1054 * 1055 * sched_move_task() holds both and thus holding either pins the cgroup, 1056 * see task_group(). 1057 * 1058 * Furthermore, all task_rq users should acquire both locks, see 1059 * task_rq_lock(). 1060 */ 1061 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1062 lockdep_is_held(&task_rq(p)->lock))); 1063#endif 1064#endif 1065 1066 trace_sched_migrate_task(p, new_cpu); 1067 1068 if (task_cpu(p) != new_cpu) { 1069 if (p->sched_class->migrate_task_rq) 1070 p->sched_class->migrate_task_rq(p, new_cpu); 1071 p->se.nr_migrations++; 1072 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1073 } 1074 1075 __set_task_cpu(p, new_cpu); 1076} 1077 1078static void __migrate_swap_task(struct task_struct *p, int cpu) 1079{ 1080 if (p->on_rq) { 1081 struct rq *src_rq, *dst_rq; 1082 1083 src_rq = task_rq(p); 1084 dst_rq = cpu_rq(cpu); 1085 1086 deactivate_task(src_rq, p, 0); 1087 set_task_cpu(p, cpu); 1088 activate_task(dst_rq, p, 0); 1089 check_preempt_curr(dst_rq, p, 0); 1090 } else { 1091 /* 1092 * Task isn't running anymore; make it appear like we migrated 1093 * it before it went to sleep. This means on wakeup we make the 1094 * previous cpu our targer instead of where it really is. 1095 */ 1096 p->wake_cpu = cpu; 1097 } 1098} 1099 1100struct migration_swap_arg { 1101 struct task_struct *src_task, *dst_task; 1102 int src_cpu, dst_cpu; 1103}; 1104 1105static int migrate_swap_stop(void *data) 1106{ 1107 struct migration_swap_arg *arg = data; 1108 struct rq *src_rq, *dst_rq; 1109 int ret = -EAGAIN; 1110 1111 src_rq = cpu_rq(arg->src_cpu); 1112 dst_rq = cpu_rq(arg->dst_cpu); 1113 1114 double_raw_lock(&arg->src_task->pi_lock, 1115 &arg->dst_task->pi_lock); 1116 double_rq_lock(src_rq, dst_rq); 1117 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1118 goto unlock; 1119 1120 if (task_cpu(arg->src_task) != arg->src_cpu) 1121 goto unlock; 1122 1123 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1124 goto unlock; 1125 1126 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1127 goto unlock; 1128 1129 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1130 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1131 1132 ret = 0; 1133 1134unlock: 1135 double_rq_unlock(src_rq, dst_rq); 1136 raw_spin_unlock(&arg->dst_task->pi_lock); 1137 raw_spin_unlock(&arg->src_task->pi_lock); 1138 1139 return ret; 1140} 1141 1142/* 1143 * Cross migrate two tasks 1144 */ 1145int migrate_swap(struct task_struct *cur, struct task_struct *p) 1146{ 1147 struct migration_swap_arg arg; 1148 int ret = -EINVAL; 1149 1150 arg = (struct migration_swap_arg){ 1151 .src_task = cur, 1152 .src_cpu = task_cpu(cur), 1153 .dst_task = p, 1154 .dst_cpu = task_cpu(p), 1155 }; 1156 1157 if (arg.src_cpu == arg.dst_cpu) 1158 goto out; 1159 1160 /* 1161 * These three tests are all lockless; this is OK since all of them 1162 * will be re-checked with proper locks held further down the line. 1163 */ 1164 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1165 goto out; 1166 1167 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1168 goto out; 1169 1170 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1171 goto out; 1172 1173 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1174 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1175 1176out: 1177 return ret; 1178} 1179 1180struct migration_arg { 1181 struct task_struct *task; 1182 int dest_cpu; 1183}; 1184 1185static int migration_cpu_stop(void *data); 1186 1187/* 1188 * wait_task_inactive - wait for a thread to unschedule. 1189 * 1190 * If @match_state is nonzero, it's the @p->state value just checked and 1191 * not expected to change. If it changes, i.e. @p might have woken up, 1192 * then return zero. When we succeed in waiting for @p to be off its CPU, 1193 * we return a positive number (its total switch count). If a second call 1194 * a short while later returns the same number, the caller can be sure that 1195 * @p has remained unscheduled the whole time. 1196 * 1197 * The caller must ensure that the task *will* unschedule sometime soon, 1198 * else this function might spin for a *long* time. This function can't 1199 * be called with interrupts off, or it may introduce deadlock with 1200 * smp_call_function() if an IPI is sent by the same process we are 1201 * waiting to become inactive. 1202 */ 1203unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1204{ 1205 unsigned long flags; 1206 int running, on_rq; 1207 unsigned long ncsw; 1208 struct rq *rq; 1209 1210 for (;;) { 1211 /* 1212 * We do the initial early heuristics without holding 1213 * any task-queue locks at all. We'll only try to get 1214 * the runqueue lock when things look like they will 1215 * work out! 1216 */ 1217 rq = task_rq(p); 1218 1219 /* 1220 * If the task is actively running on another CPU 1221 * still, just relax and busy-wait without holding 1222 * any locks. 1223 * 1224 * NOTE! Since we don't hold any locks, it's not 1225 * even sure that "rq" stays as the right runqueue! 1226 * But we don't care, since "task_running()" will 1227 * return false if the runqueue has changed and p 1228 * is actually now running somewhere else! 1229 */ 1230 while (task_running(rq, p)) { 1231 if (match_state && unlikely(p->state != match_state)) 1232 return 0; 1233 cpu_relax(); 1234 } 1235 1236 /* 1237 * Ok, time to look more closely! We need the rq 1238 * lock now, to be *sure*. If we're wrong, we'll 1239 * just go back and repeat. 1240 */ 1241 rq = task_rq_lock(p, &flags); 1242 trace_sched_wait_task(p); 1243 running = task_running(rq, p); 1244 on_rq = p->on_rq; 1245 ncsw = 0; 1246 if (!match_state || p->state == match_state) 1247 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1248 task_rq_unlock(rq, p, &flags); 1249 1250 /* 1251 * If it changed from the expected state, bail out now. 1252 */ 1253 if (unlikely(!ncsw)) 1254 break; 1255 1256 /* 1257 * Was it really running after all now that we 1258 * checked with the proper locks actually held? 1259 * 1260 * Oops. Go back and try again.. 1261 */ 1262 if (unlikely(running)) { 1263 cpu_relax(); 1264 continue; 1265 } 1266 1267 /* 1268 * It's not enough that it's not actively running, 1269 * it must be off the runqueue _entirely_, and not 1270 * preempted! 1271 * 1272 * So if it was still runnable (but just not actively 1273 * running right now), it's preempted, and we should 1274 * yield - it could be a while. 1275 */ 1276 if (unlikely(on_rq)) { 1277 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1278 1279 set_current_state(TASK_UNINTERRUPTIBLE); 1280 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1281 continue; 1282 } 1283 1284 /* 1285 * Ahh, all good. It wasn't running, and it wasn't 1286 * runnable, which means that it will never become 1287 * running in the future either. We're all done! 1288 */ 1289 break; 1290 } 1291 1292 return ncsw; 1293} 1294 1295/*** 1296 * kick_process - kick a running thread to enter/exit the kernel 1297 * @p: the to-be-kicked thread 1298 * 1299 * Cause a process which is running on another CPU to enter 1300 * kernel-mode, without any delay. (to get signals handled.) 1301 * 1302 * NOTE: this function doesn't have to take the runqueue lock, 1303 * because all it wants to ensure is that the remote task enters 1304 * the kernel. If the IPI races and the task has been migrated 1305 * to another CPU then no harm is done and the purpose has been 1306 * achieved as well. 1307 */ 1308void kick_process(struct task_struct *p) 1309{ 1310 int cpu; 1311 1312 preempt_disable(); 1313 cpu = task_cpu(p); 1314 if ((cpu != smp_processor_id()) && task_curr(p)) 1315 smp_send_reschedule(cpu); 1316 preempt_enable(); 1317} 1318EXPORT_SYMBOL_GPL(kick_process); 1319#endif /* CONFIG_SMP */ 1320 1321#ifdef CONFIG_SMP 1322/* 1323 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1324 */ 1325static int select_fallback_rq(int cpu, struct task_struct *p) 1326{ 1327 int nid = cpu_to_node(cpu); 1328 const struct cpumask *nodemask = NULL; 1329 enum { cpuset, possible, fail } state = cpuset; 1330 int dest_cpu; 1331 1332 /* 1333 * If the node that the cpu is on has been offlined, cpu_to_node() 1334 * will return -1. There is no cpu on the node, and we should 1335 * select the cpu on the other node. 1336 */ 1337 if (nid != -1) { 1338 nodemask = cpumask_of_node(nid); 1339 1340 /* Look for allowed, online CPU in same node. */ 1341 for_each_cpu(dest_cpu, nodemask) { 1342 if (!cpu_online(dest_cpu)) 1343 continue; 1344 if (!cpu_active(dest_cpu)) 1345 continue; 1346 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1347 return dest_cpu; 1348 } 1349 } 1350 1351 for (;;) { 1352 /* Any allowed, online CPU? */ 1353 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1354 if (!cpu_online(dest_cpu)) 1355 continue; 1356 if (!cpu_active(dest_cpu)) 1357 continue; 1358 goto out; 1359 } 1360 1361 switch (state) { 1362 case cpuset: 1363 /* No more Mr. Nice Guy. */ 1364 cpuset_cpus_allowed_fallback(p); 1365 state = possible; 1366 break; 1367 1368 case possible: 1369 do_set_cpus_allowed(p, cpu_possible_mask); 1370 state = fail; 1371 break; 1372 1373 case fail: 1374 BUG(); 1375 break; 1376 } 1377 } 1378 1379out: 1380 if (state != cpuset) { 1381 /* 1382 * Don't tell them about moving exiting tasks or 1383 * kernel threads (both mm NULL), since they never 1384 * leave kernel. 1385 */ 1386 if (p->mm && printk_ratelimit()) { 1387 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 1388 task_pid_nr(p), p->comm, cpu); 1389 } 1390 } 1391 1392 return dest_cpu; 1393} 1394 1395/* 1396 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1397 */ 1398static inline 1399int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1400{ 1401 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1402 1403 /* 1404 * In order not to call set_task_cpu() on a blocking task we need 1405 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1406 * cpu. 1407 * 1408 * Since this is common to all placement strategies, this lives here. 1409 * 1410 * [ this allows ->select_task() to simply return task_cpu(p) and 1411 * not worry about this generic constraint ] 1412 */ 1413 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1414 !cpu_online(cpu))) 1415 cpu = select_fallback_rq(task_cpu(p), p); 1416 1417 return cpu; 1418} 1419 1420static void update_avg(u64 *avg, u64 sample) 1421{ 1422 s64 diff = sample - *avg; 1423 *avg += diff >> 3; 1424} 1425#endif 1426 1427static void 1428ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1429{ 1430#ifdef CONFIG_SCHEDSTATS 1431 struct rq *rq = this_rq(); 1432 1433#ifdef CONFIG_SMP 1434 int this_cpu = smp_processor_id(); 1435 1436 if (cpu == this_cpu) { 1437 schedstat_inc(rq, ttwu_local); 1438 schedstat_inc(p, se.statistics.nr_wakeups_local); 1439 } else { 1440 struct sched_domain *sd; 1441 1442 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1443 rcu_read_lock(); 1444 for_each_domain(this_cpu, sd) { 1445 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1446 schedstat_inc(sd, ttwu_wake_remote); 1447 break; 1448 } 1449 } 1450 rcu_read_unlock(); 1451 } 1452 1453 if (wake_flags & WF_MIGRATED) 1454 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1455 1456#endif /* CONFIG_SMP */ 1457 1458 schedstat_inc(rq, ttwu_count); 1459 schedstat_inc(p, se.statistics.nr_wakeups); 1460 1461 if (wake_flags & WF_SYNC) 1462 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1463 1464#endif /* CONFIG_SCHEDSTATS */ 1465} 1466 1467static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1468{ 1469 activate_task(rq, p, en_flags); 1470 p->on_rq = 1; 1471 1472 /* if a worker is waking up, notify workqueue */ 1473 if (p->flags & PF_WQ_WORKER) 1474 wq_worker_waking_up(p, cpu_of(rq)); 1475} 1476 1477/* 1478 * Mark the task runnable and perform wakeup-preemption. 1479 */ 1480static void 1481ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1482{ 1483 check_preempt_curr(rq, p, wake_flags); 1484 trace_sched_wakeup(p, true); 1485 1486 p->state = TASK_RUNNING; 1487#ifdef CONFIG_SMP 1488 if (p->sched_class->task_woken) 1489 p->sched_class->task_woken(rq, p); 1490 1491 if (rq->idle_stamp) { 1492 u64 delta = rq_clock(rq) - rq->idle_stamp; 1493 u64 max = 2*rq->max_idle_balance_cost; 1494 1495 update_avg(&rq->avg_idle, delta); 1496 1497 if (rq->avg_idle > max) 1498 rq->avg_idle = max; 1499 1500 rq->idle_stamp = 0; 1501 } 1502#endif 1503} 1504 1505static void 1506ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1507{ 1508#ifdef CONFIG_SMP 1509 if (p->sched_contributes_to_load) 1510 rq->nr_uninterruptible--; 1511#endif 1512 1513 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1514 ttwu_do_wakeup(rq, p, wake_flags); 1515} 1516 1517/* 1518 * Called in case the task @p isn't fully descheduled from its runqueue, 1519 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1520 * since all we need to do is flip p->state to TASK_RUNNING, since 1521 * the task is still ->on_rq. 1522 */ 1523static int ttwu_remote(struct task_struct *p, int wake_flags) 1524{ 1525 struct rq *rq; 1526 int ret = 0; 1527 1528 rq = __task_rq_lock(p); 1529 if (p->on_rq) { 1530 /* check_preempt_curr() may use rq clock */ 1531 update_rq_clock(rq); 1532 ttwu_do_wakeup(rq, p, wake_flags); 1533 ret = 1; 1534 } 1535 __task_rq_unlock(rq); 1536 1537 return ret; 1538} 1539 1540#ifdef CONFIG_SMP 1541void sched_ttwu_pending(void) 1542{ 1543 struct rq *rq = this_rq(); 1544 struct llist_node *llist = llist_del_all(&rq->wake_list); 1545 struct task_struct *p; 1546 unsigned long flags; 1547 1548 if (!llist) 1549 return; 1550 1551 raw_spin_lock_irqsave(&rq->lock, flags); 1552 1553 while (llist) { 1554 p = llist_entry(llist, struct task_struct, wake_entry); 1555 llist = llist_next(llist); 1556 ttwu_do_activate(rq, p, 0); 1557 } 1558 1559 raw_spin_unlock_irqrestore(&rq->lock, flags); 1560} 1561 1562void scheduler_ipi(void) 1563{ 1564 /* 1565 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1566 * TIF_NEED_RESCHED remotely (for the first time) will also send 1567 * this IPI. 1568 */ 1569 preempt_fold_need_resched(); 1570 1571 if (llist_empty(&this_rq()->wake_list) 1572 && !tick_nohz_full_cpu(smp_processor_id()) 1573 && !got_nohz_idle_kick()) 1574 return; 1575 1576 /* 1577 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1578 * traditionally all their work was done from the interrupt return 1579 * path. Now that we actually do some work, we need to make sure 1580 * we do call them. 1581 * 1582 * Some archs already do call them, luckily irq_enter/exit nest 1583 * properly. 1584 * 1585 * Arguably we should visit all archs and update all handlers, 1586 * however a fair share of IPIs are still resched only so this would 1587 * somewhat pessimize the simple resched case. 1588 */ 1589 irq_enter(); 1590 tick_nohz_full_check(); 1591 sched_ttwu_pending(); 1592 1593 /* 1594 * Check if someone kicked us for doing the nohz idle load balance. 1595 */ 1596 if (unlikely(got_nohz_idle_kick())) { 1597 this_rq()->idle_balance = 1; 1598 raise_softirq_irqoff(SCHED_SOFTIRQ); 1599 } 1600 irq_exit(); 1601} 1602 1603static void ttwu_queue_remote(struct task_struct *p, int cpu) 1604{ 1605 struct rq *rq = cpu_rq(cpu); 1606 1607 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { 1608 if (!set_nr_if_polling(rq->idle)) 1609 smp_send_reschedule(cpu); 1610 else 1611 trace_sched_wake_idle_without_ipi(cpu); 1612 } 1613} 1614 1615bool cpus_share_cache(int this_cpu, int that_cpu) 1616{ 1617 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1618} 1619#endif /* CONFIG_SMP */ 1620 1621static void ttwu_queue(struct task_struct *p, int cpu) 1622{ 1623 struct rq *rq = cpu_rq(cpu); 1624 1625#if defined(CONFIG_SMP) 1626 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1627 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1628 ttwu_queue_remote(p, cpu); 1629 return; 1630 } 1631#endif 1632 1633 raw_spin_lock(&rq->lock); 1634 ttwu_do_activate(rq, p, 0); 1635 raw_spin_unlock(&rq->lock); 1636} 1637 1638/** 1639 * try_to_wake_up - wake up a thread 1640 * @p: the thread to be awakened 1641 * @state: the mask of task states that can be woken 1642 * @wake_flags: wake modifier flags (WF_*) 1643 * 1644 * Put it on the run-queue if it's not already there. The "current" 1645 * thread is always on the run-queue (except when the actual 1646 * re-schedule is in progress), and as such you're allowed to do 1647 * the simpler "current->state = TASK_RUNNING" to mark yourself 1648 * runnable without the overhead of this. 1649 * 1650 * Return: %true if @p was woken up, %false if it was already running. 1651 * or @state didn't match @p's state. 1652 */ 1653static int 1654try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1655{ 1656 unsigned long flags; 1657 int cpu, success = 0; 1658 1659 /* 1660 * If we are going to wake up a thread waiting for CONDITION we 1661 * need to ensure that CONDITION=1 done by the caller can not be 1662 * reordered with p->state check below. This pairs with mb() in 1663 * set_current_state() the waiting thread does. 1664 */ 1665 smp_mb__before_spinlock(); 1666 raw_spin_lock_irqsave(&p->pi_lock, flags); 1667 if (!(p->state & state)) 1668 goto out; 1669 1670 success = 1; /* we're going to change ->state */ 1671 cpu = task_cpu(p); 1672 1673 if (p->on_rq && ttwu_remote(p, wake_flags)) 1674 goto stat; 1675 1676#ifdef CONFIG_SMP 1677 /* 1678 * If the owning (remote) cpu is still in the middle of schedule() with 1679 * this task as prev, wait until its done referencing the task. 1680 */ 1681 while (p->on_cpu) 1682 cpu_relax(); 1683 /* 1684 * Pairs with the smp_wmb() in finish_lock_switch(). 1685 */ 1686 smp_rmb(); 1687 1688 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1689 p->state = TASK_WAKING; 1690 1691 if (p->sched_class->task_waking) 1692 p->sched_class->task_waking(p); 1693 1694 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1695 if (task_cpu(p) != cpu) { 1696 wake_flags |= WF_MIGRATED; 1697 set_task_cpu(p, cpu); 1698 } 1699#endif /* CONFIG_SMP */ 1700 1701 ttwu_queue(p, cpu); 1702stat: 1703 ttwu_stat(p, cpu, wake_flags); 1704out: 1705 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1706 1707 return success; 1708} 1709 1710/** 1711 * try_to_wake_up_local - try to wake up a local task with rq lock held 1712 * @p: the thread to be awakened 1713 * 1714 * Put @p on the run-queue if it's not already there. The caller must 1715 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1716 * the current task. 1717 */ 1718static void try_to_wake_up_local(struct task_struct *p) 1719{ 1720 struct rq *rq = task_rq(p); 1721 1722 if (WARN_ON_ONCE(rq != this_rq()) || 1723 WARN_ON_ONCE(p == current)) 1724 return; 1725 1726 lockdep_assert_held(&rq->lock); 1727 1728 if (!raw_spin_trylock(&p->pi_lock)) { 1729 raw_spin_unlock(&rq->lock); 1730 raw_spin_lock(&p->pi_lock); 1731 raw_spin_lock(&rq->lock); 1732 } 1733 1734 if (!(p->state & TASK_NORMAL)) 1735 goto out; 1736 1737 if (!p->on_rq) 1738 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1739 1740 ttwu_do_wakeup(rq, p, 0); 1741 ttwu_stat(p, smp_processor_id(), 0); 1742out: 1743 raw_spin_unlock(&p->pi_lock); 1744} 1745 1746/** 1747 * wake_up_process - Wake up a specific process 1748 * @p: The process to be woken up. 1749 * 1750 * Attempt to wake up the nominated process and move it to the set of runnable 1751 * processes. 1752 * 1753 * Return: 1 if the process was woken up, 0 if it was already running. 1754 * 1755 * It may be assumed that this function implies a write memory barrier before 1756 * changing the task state if and only if any tasks are woken up. 1757 */ 1758int wake_up_process(struct task_struct *p) 1759{ 1760 WARN_ON(task_is_stopped_or_traced(p)); 1761 return try_to_wake_up(p, TASK_NORMAL, 0); 1762} 1763EXPORT_SYMBOL(wake_up_process); 1764 1765int wake_up_state(struct task_struct *p, unsigned int state) 1766{ 1767 return try_to_wake_up(p, state, 0); 1768} 1769 1770/* 1771 * Perform scheduler related setup for a newly forked process p. 1772 * p is forked by current. 1773 * 1774 * __sched_fork() is basic setup used by init_idle() too: 1775 */ 1776static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1777{ 1778 p->on_rq = 0; 1779 1780 p->se.on_rq = 0; 1781 p->se.exec_start = 0; 1782 p->se.sum_exec_runtime = 0; 1783 p->se.prev_sum_exec_runtime = 0; 1784 p->se.nr_migrations = 0; 1785 p->se.vruntime = 0; 1786 INIT_LIST_HEAD(&p->se.group_node); 1787 1788#ifdef CONFIG_SCHEDSTATS 1789 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1790#endif 1791 1792 RB_CLEAR_NODE(&p->dl.rb_node); 1793 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1794 p->dl.dl_runtime = p->dl.runtime = 0; 1795 p->dl.dl_deadline = p->dl.deadline = 0; 1796 p->dl.dl_period = 0; 1797 p->dl.flags = 0; 1798 1799 INIT_LIST_HEAD(&p->rt.run_list); 1800 1801#ifdef CONFIG_PREEMPT_NOTIFIERS 1802 INIT_HLIST_HEAD(&p->preempt_notifiers); 1803#endif 1804 1805#ifdef CONFIG_NUMA_BALANCING 1806 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1807 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1808 p->mm->numa_scan_seq = 0; 1809 } 1810 1811 if (clone_flags & CLONE_VM) 1812 p->numa_preferred_nid = current->numa_preferred_nid; 1813 else 1814 p->numa_preferred_nid = -1; 1815 1816 p->node_stamp = 0ULL; 1817 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1818 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1819 p->numa_work.next = &p->numa_work; 1820 p->numa_faults_memory = NULL; 1821 p->numa_faults_buffer_memory = NULL; 1822 p->last_task_numa_placement = 0; 1823 p->last_sum_exec_runtime = 0; 1824 1825 INIT_LIST_HEAD(&p->numa_entry); 1826 p->numa_group = NULL; 1827#endif /* CONFIG_NUMA_BALANCING */ 1828} 1829 1830#ifdef CONFIG_NUMA_BALANCING 1831#ifdef CONFIG_SCHED_DEBUG 1832void set_numabalancing_state(bool enabled) 1833{ 1834 if (enabled) 1835 sched_feat_set("NUMA"); 1836 else 1837 sched_feat_set("NO_NUMA"); 1838} 1839#else 1840__read_mostly bool numabalancing_enabled; 1841 1842void set_numabalancing_state(bool enabled) 1843{ 1844 numabalancing_enabled = enabled; 1845} 1846#endif /* CONFIG_SCHED_DEBUG */ 1847 1848#ifdef CONFIG_PROC_SYSCTL 1849int sysctl_numa_balancing(struct ctl_table *table, int write, 1850 void __user *buffer, size_t *lenp, loff_t *ppos) 1851{ 1852 struct ctl_table t; 1853 int err; 1854 int state = numabalancing_enabled; 1855 1856 if (write && !capable(CAP_SYS_ADMIN)) 1857 return -EPERM; 1858 1859 t = *table; 1860 t.data = &state; 1861 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1862 if (err < 0) 1863 return err; 1864 if (write) 1865 set_numabalancing_state(state); 1866 return err; 1867} 1868#endif 1869#endif 1870 1871/* 1872 * fork()/clone()-time setup: 1873 */ 1874int sched_fork(unsigned long clone_flags, struct task_struct *p) 1875{ 1876 unsigned long flags; 1877 int cpu = get_cpu(); 1878 1879 __sched_fork(clone_flags, p); 1880 /* 1881 * We mark the process as running here. This guarantees that 1882 * nobody will actually run it, and a signal or other external 1883 * event cannot wake it up and insert it on the runqueue either. 1884 */ 1885 p->state = TASK_RUNNING; 1886 1887 /* 1888 * Make sure we do not leak PI boosting priority to the child. 1889 */ 1890 p->prio = current->normal_prio; 1891 1892 /* 1893 * Revert to default priority/policy on fork if requested. 1894 */ 1895 if (unlikely(p->sched_reset_on_fork)) { 1896 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1897 p->policy = SCHED_NORMAL; 1898 p->static_prio = NICE_TO_PRIO(0); 1899 p->rt_priority = 0; 1900 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1901 p->static_prio = NICE_TO_PRIO(0); 1902 1903 p->prio = p->normal_prio = __normal_prio(p); 1904 set_load_weight(p); 1905 1906 /* 1907 * We don't need the reset flag anymore after the fork. It has 1908 * fulfilled its duty: 1909 */ 1910 p->sched_reset_on_fork = 0; 1911 } 1912 1913 if (dl_prio(p->prio)) { 1914 put_cpu(); 1915 return -EAGAIN; 1916 } else if (rt_prio(p->prio)) { 1917 p->sched_class = &rt_sched_class; 1918 } else { 1919 p->sched_class = &fair_sched_class; 1920 } 1921 1922 if (p->sched_class->task_fork) 1923 p->sched_class->task_fork(p); 1924 1925 /* 1926 * The child is not yet in the pid-hash so no cgroup attach races, 1927 * and the cgroup is pinned to this child due to cgroup_fork() 1928 * is ran before sched_fork(). 1929 * 1930 * Silence PROVE_RCU. 1931 */ 1932 raw_spin_lock_irqsave(&p->pi_lock, flags); 1933 set_task_cpu(p, cpu); 1934 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1935 1936#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1937 if (likely(sched_info_on())) 1938 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1939#endif 1940#if defined(CONFIG_SMP) 1941 p->on_cpu = 0; 1942#endif 1943 init_task_preempt_count(p); 1944#ifdef CONFIG_SMP 1945 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1946 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1947#endif 1948 1949 put_cpu(); 1950 return 0; 1951} 1952 1953unsigned long to_ratio(u64 period, u64 runtime) 1954{ 1955 if (runtime == RUNTIME_INF) 1956 return 1ULL << 20; 1957 1958 /* 1959 * Doing this here saves a lot of checks in all 1960 * the calling paths, and returning zero seems 1961 * safe for them anyway. 1962 */ 1963 if (period == 0) 1964 return 0; 1965 1966 return div64_u64(runtime << 20, period); 1967} 1968 1969#ifdef CONFIG_SMP 1970inline struct dl_bw *dl_bw_of(int i) 1971{ 1972 return &cpu_rq(i)->rd->dl_bw; 1973} 1974 1975static inline int dl_bw_cpus(int i) 1976{ 1977 struct root_domain *rd = cpu_rq(i)->rd; 1978 int cpus = 0; 1979 1980 for_each_cpu_and(i, rd->span, cpu_active_mask) 1981 cpus++; 1982 1983 return cpus; 1984} 1985#else 1986inline struct dl_bw *dl_bw_of(int i) 1987{ 1988 return &cpu_rq(i)->dl.dl_bw; 1989} 1990 1991static inline int dl_bw_cpus(int i) 1992{ 1993 return 1; 1994} 1995#endif 1996 1997static inline 1998void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1999{ 2000 dl_b->total_bw -= tsk_bw; 2001} 2002 2003static inline 2004void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 2005{ 2006 dl_b->total_bw += tsk_bw; 2007} 2008 2009static inline 2010bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 2011{ 2012 return dl_b->bw != -1 && 2013 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 2014} 2015 2016/* 2017 * We must be sure that accepting a new task (or allowing changing the 2018 * parameters of an existing one) is consistent with the bandwidth 2019 * constraints. If yes, this function also accordingly updates the currently 2020 * allocated bandwidth to reflect the new situation. 2021 * 2022 * This function is called while holding p's rq->lock. 2023 */ 2024static int dl_overflow(struct task_struct *p, int policy, 2025 const struct sched_attr *attr) 2026{ 2027 2028 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 2029 u64 period = attr->sched_period ?: attr->sched_deadline; 2030 u64 runtime = attr->sched_runtime; 2031 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 2032 int cpus, err = -1; 2033 2034 if (new_bw == p->dl.dl_bw) 2035 return 0; 2036 2037 /* 2038 * Either if a task, enters, leave, or stays -deadline but changes 2039 * its parameters, we may need to update accordingly the total 2040 * allocated bandwidth of the container. 2041 */ 2042 raw_spin_lock(&dl_b->lock); 2043 cpus = dl_bw_cpus(task_cpu(p)); 2044 if (dl_policy(policy) && !task_has_dl_policy(p) && 2045 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2046 __dl_add(dl_b, new_bw); 2047 err = 0; 2048 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2049 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2050 __dl_clear(dl_b, p->dl.dl_bw); 2051 __dl_add(dl_b, new_bw); 2052 err = 0; 2053 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2054 __dl_clear(dl_b, p->dl.dl_bw); 2055 err = 0; 2056 } 2057 raw_spin_unlock(&dl_b->lock); 2058 2059 return err; 2060} 2061 2062extern void init_dl_bw(struct dl_bw *dl_b); 2063 2064/* 2065 * wake_up_new_task - wake up a newly created task for the first time. 2066 * 2067 * This function will do some initial scheduler statistics housekeeping 2068 * that must be done for every newly created context, then puts the task 2069 * on the runqueue and wakes it. 2070 */ 2071void wake_up_new_task(struct task_struct *p) 2072{ 2073 unsigned long flags; 2074 struct rq *rq; 2075 2076 raw_spin_lock_irqsave(&p->pi_lock, flags); 2077#ifdef CONFIG_SMP 2078 /* 2079 * Fork balancing, do it here and not earlier because: 2080 * - cpus_allowed can change in the fork path 2081 * - any previously selected cpu might disappear through hotplug 2082 */ 2083 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2084#endif 2085 2086 /* Initialize new task's runnable average */ 2087 init_task_runnable_average(p); 2088 rq = __task_rq_lock(p); 2089 activate_task(rq, p, 0); 2090 p->on_rq = 1; 2091 trace_sched_wakeup_new(p, true); 2092 check_preempt_curr(rq, p, WF_FORK); 2093#ifdef CONFIG_SMP 2094 if (p->sched_class->task_woken) 2095 p->sched_class->task_woken(rq, p); 2096#endif 2097 task_rq_unlock(rq, p, &flags); 2098} 2099 2100#ifdef CONFIG_PREEMPT_NOTIFIERS 2101 2102/** 2103 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2104 * @notifier: notifier struct to register 2105 */ 2106void preempt_notifier_register(struct preempt_notifier *notifier) 2107{ 2108 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2109} 2110EXPORT_SYMBOL_GPL(preempt_notifier_register); 2111 2112/** 2113 * preempt_notifier_unregister - no longer interested in preemption notifications 2114 * @notifier: notifier struct to unregister 2115 * 2116 * This is safe to call from within a preemption notifier. 2117 */ 2118void preempt_notifier_unregister(struct preempt_notifier *notifier) 2119{ 2120 hlist_del(¬ifier->link); 2121} 2122EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2123 2124static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2125{ 2126 struct preempt_notifier *notifier; 2127 2128 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2129 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2130} 2131 2132static void 2133fire_sched_out_preempt_notifiers(struct task_struct *curr, 2134 struct task_struct *next) 2135{ 2136 struct preempt_notifier *notifier; 2137 2138 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2139 notifier->ops->sched_out(notifier, next); 2140} 2141 2142#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2143 2144static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2145{ 2146} 2147 2148static void 2149fire_sched_out_preempt_notifiers(struct task_struct *curr, 2150 struct task_struct *next) 2151{ 2152} 2153 2154#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2155 2156/** 2157 * prepare_task_switch - prepare to switch tasks 2158 * @rq: the runqueue preparing to switch 2159 * @prev: the current task that is being switched out 2160 * @next: the task we are going to switch to. 2161 * 2162 * This is called with the rq lock held and interrupts off. It must 2163 * be paired with a subsequent finish_task_switch after the context 2164 * switch. 2165 * 2166 * prepare_task_switch sets up locking and calls architecture specific 2167 * hooks. 2168 */ 2169static inline void 2170prepare_task_switch(struct rq *rq, struct task_struct *prev, 2171 struct task_struct *next) 2172{ 2173 trace_sched_switch(prev, next); 2174 sched_info_switch(rq, prev, next); 2175 perf_event_task_sched_out(prev, next); 2176 fire_sched_out_preempt_notifiers(prev, next); 2177 prepare_lock_switch(rq, next); 2178 prepare_arch_switch(next); 2179} 2180 2181/** 2182 * finish_task_switch - clean up after a task-switch 2183 * @rq: runqueue associated with task-switch 2184 * @prev: the thread we just switched away from. 2185 * 2186 * finish_task_switch must be called after the context switch, paired 2187 * with a prepare_task_switch call before the context switch. 2188 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2189 * and do any other architecture-specific cleanup actions. 2190 * 2191 * Note that we may have delayed dropping an mm in context_switch(). If 2192 * so, we finish that here outside of the runqueue lock. (Doing it 2193 * with the lock held can cause deadlocks; see schedule() for 2194 * details.) 2195 */ 2196static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2197 __releases(rq->lock) 2198{ 2199 struct mm_struct *mm = rq->prev_mm; 2200 long prev_state; 2201 2202 rq->prev_mm = NULL; 2203 2204 /* 2205 * A task struct has one reference for the use as "current". 2206 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2207 * schedule one last time. The schedule call will never return, and 2208 * the scheduled task must drop that reference. 2209 * The test for TASK_DEAD must occur while the runqueue locks are 2210 * still held, otherwise prev could be scheduled on another cpu, die 2211 * there before we look at prev->state, and then the reference would 2212 * be dropped twice. 2213 * Manfred Spraul <manfred@colorfullife.com> 2214 */ 2215 prev_state = prev->state; 2216 vtime_task_switch(prev); 2217 finish_arch_switch(prev); 2218 perf_event_task_sched_in(prev, current); 2219 finish_lock_switch(rq, prev); 2220 finish_arch_post_lock_switch(); 2221 2222 fire_sched_in_preempt_notifiers(current); 2223 if (mm) 2224 mmdrop(mm); 2225 if (unlikely(prev_state == TASK_DEAD)) { 2226 if (prev->sched_class->task_dead) 2227 prev->sched_class->task_dead(prev); 2228 2229 /* 2230 * Remove function-return probe instances associated with this 2231 * task and put them back on the free list. 2232 */ 2233 kprobe_flush_task(prev); 2234 put_task_struct(prev); 2235 } 2236 2237 tick_nohz_task_switch(current); 2238} 2239 2240#ifdef CONFIG_SMP 2241 2242/* rq->lock is NOT held, but preemption is disabled */ 2243static inline void post_schedule(struct rq *rq) 2244{ 2245 if (rq->post_schedule) { 2246 unsigned long flags; 2247 2248 raw_spin_lock_irqsave(&rq->lock, flags); 2249 if (rq->curr->sched_class->post_schedule) 2250 rq->curr->sched_class->post_schedule(rq); 2251 raw_spin_unlock_irqrestore(&rq->lock, flags); 2252 2253 rq->post_schedule = 0; 2254 } 2255} 2256 2257#else 2258 2259static inline void post_schedule(struct rq *rq) 2260{ 2261} 2262 2263#endif 2264 2265/** 2266 * schedule_tail - first thing a freshly forked thread must call. 2267 * @prev: the thread we just switched away from. 2268 */ 2269asmlinkage __visible void schedule_tail(struct task_struct *prev) 2270 __releases(rq->lock) 2271{ 2272 struct rq *rq = this_rq(); 2273 2274 finish_task_switch(rq, prev); 2275 2276 /* 2277 * FIXME: do we need to worry about rq being invalidated by the 2278 * task_switch? 2279 */ 2280 post_schedule(rq); 2281 2282#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2283 /* In this case, finish_task_switch does not reenable preemption */ 2284 preempt_enable(); 2285#endif 2286 if (current->set_child_tid) 2287 put_user(task_pid_vnr(current), current->set_child_tid); 2288} 2289 2290/* 2291 * context_switch - switch to the new MM and the new 2292 * thread's register state. 2293 */ 2294static inline void 2295context_switch(struct rq *rq, struct task_struct *prev, 2296 struct task_struct *next) 2297{ 2298 struct mm_struct *mm, *oldmm; 2299 2300 prepare_task_switch(rq, prev, next); 2301 2302 mm = next->mm; 2303 oldmm = prev->active_mm; 2304 /* 2305 * For paravirt, this is coupled with an exit in switch_to to 2306 * combine the page table reload and the switch backend into 2307 * one hypercall. 2308 */ 2309 arch_start_context_switch(prev); 2310 2311 if (!mm) { 2312 next->active_mm = oldmm; 2313 atomic_inc(&oldmm->mm_count); 2314 enter_lazy_tlb(oldmm, next); 2315 } else 2316 switch_mm(oldmm, mm, next); 2317 2318 if (!prev->mm) { 2319 prev->active_mm = NULL; 2320 rq->prev_mm = oldmm; 2321 } 2322 /* 2323 * Since the runqueue lock will be released by the next 2324 * task (which is an invalid locking op but in the case 2325 * of the scheduler it's an obvious special-case), so we 2326 * do an early lockdep release here: 2327 */ 2328#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2329 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2330#endif 2331 2332 context_tracking_task_switch(prev, next); 2333 /* Here we just switch the register state and the stack. */ 2334 switch_to(prev, next, prev); 2335 2336 barrier(); 2337 /* 2338 * this_rq must be evaluated again because prev may have moved 2339 * CPUs since it called schedule(), thus the 'rq' on its stack 2340 * frame will be invalid. 2341 */ 2342 finish_task_switch(this_rq(), prev); 2343} 2344 2345/* 2346 * nr_running and nr_context_switches: 2347 * 2348 * externally visible scheduler statistics: current number of runnable 2349 * threads, total number of context switches performed since bootup. 2350 */ 2351unsigned long nr_running(void) 2352{ 2353 unsigned long i, sum = 0; 2354 2355 for_each_online_cpu(i) 2356 sum += cpu_rq(i)->nr_running; 2357 2358 return sum; 2359} 2360 2361unsigned long long nr_context_switches(void) 2362{ 2363 int i; 2364 unsigned long long sum = 0; 2365 2366 for_each_possible_cpu(i) 2367 sum += cpu_rq(i)->nr_switches; 2368 2369 return sum; 2370} 2371 2372unsigned long nr_iowait(void) 2373{ 2374 unsigned long i, sum = 0; 2375 2376 for_each_possible_cpu(i) 2377 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2378 2379 return sum; 2380} 2381 2382unsigned long nr_iowait_cpu(int cpu) 2383{ 2384 struct rq *this = cpu_rq(cpu); 2385 return atomic_read(&this->nr_iowait); 2386} 2387 2388#ifdef CONFIG_SMP 2389 2390/* 2391 * sched_exec - execve() is a valuable balancing opportunity, because at 2392 * this point the task has the smallest effective memory and cache footprint. 2393 */ 2394void sched_exec(void) 2395{ 2396 struct task_struct *p = current; 2397 unsigned long flags; 2398 int dest_cpu; 2399 2400 raw_spin_lock_irqsave(&p->pi_lock, flags); 2401 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2402 if (dest_cpu == smp_processor_id()) 2403 goto unlock; 2404 2405 if (likely(cpu_active(dest_cpu))) { 2406 struct migration_arg arg = { p, dest_cpu }; 2407 2408 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2409 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2410 return; 2411 } 2412unlock: 2413 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2414} 2415 2416#endif 2417 2418DEFINE_PER_CPU(struct kernel_stat, kstat); 2419DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2420 2421EXPORT_PER_CPU_SYMBOL(kstat); 2422EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2423 2424/* 2425 * Return any ns on the sched_clock that have not yet been accounted in 2426 * @p in case that task is currently running. 2427 * 2428 * Called with task_rq_lock() held on @rq. 2429 */ 2430static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2431{ 2432 u64 ns = 0; 2433 2434 if (task_current(rq, p)) { 2435 update_rq_clock(rq); 2436 ns = rq_clock_task(rq) - p->se.exec_start; 2437 if ((s64)ns < 0) 2438 ns = 0; 2439 } 2440 2441 return ns; 2442} 2443 2444unsigned long long task_delta_exec(struct task_struct *p) 2445{ 2446 unsigned long flags; 2447 struct rq *rq; 2448 u64 ns = 0; 2449 2450 rq = task_rq_lock(p, &flags); 2451 ns = do_task_delta_exec(p, rq); 2452 task_rq_unlock(rq, p, &flags); 2453 2454 return ns; 2455} 2456 2457/* 2458 * Return accounted runtime for the task. 2459 * In case the task is currently running, return the runtime plus current's 2460 * pending runtime that have not been accounted yet. 2461 */ 2462unsigned long long task_sched_runtime(struct task_struct *p) 2463{ 2464 unsigned long flags; 2465 struct rq *rq; 2466 u64 ns = 0; 2467 2468#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2469 /* 2470 * 64-bit doesn't need locks to atomically read a 64bit value. 2471 * So we have a optimization chance when the task's delta_exec is 0. 2472 * Reading ->on_cpu is racy, but this is ok. 2473 * 2474 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2475 * If we race with it entering cpu, unaccounted time is 0. This is 2476 * indistinguishable from the read occurring a few cycles earlier. 2477 */ 2478 if (!p->on_cpu) 2479 return p->se.sum_exec_runtime; 2480#endif 2481 2482 rq = task_rq_lock(p, &flags); 2483 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2484 task_rq_unlock(rq, p, &flags); 2485 2486 return ns; 2487} 2488 2489/* 2490 * This function gets called by the timer code, with HZ frequency. 2491 * We call it with interrupts disabled. 2492 */ 2493void scheduler_tick(void) 2494{ 2495 int cpu = smp_processor_id(); 2496 struct rq *rq = cpu_rq(cpu); 2497 struct task_struct *curr = rq->curr; 2498 2499 sched_clock_tick(); 2500 2501 raw_spin_lock(&rq->lock); 2502 update_rq_clock(rq); 2503 curr->sched_class->task_tick(rq, curr, 0); 2504 update_cpu_load_active(rq); 2505 raw_spin_unlock(&rq->lock); 2506 2507 perf_event_task_tick(); 2508 2509#ifdef CONFIG_SMP 2510 rq->idle_balance = idle_cpu(cpu); 2511 trigger_load_balance(rq); 2512#endif 2513 rq_last_tick_reset(rq); 2514} 2515 2516#ifdef CONFIG_NO_HZ_FULL 2517/** 2518 * scheduler_tick_max_deferment 2519 * 2520 * Keep at least one tick per second when a single 2521 * active task is running because the scheduler doesn't 2522 * yet completely support full dynticks environment. 2523 * 2524 * This makes sure that uptime, CFS vruntime, load 2525 * balancing, etc... continue to move forward, even 2526 * with a very low granularity. 2527 * 2528 * Return: Maximum deferment in nanoseconds. 2529 */ 2530u64 scheduler_tick_max_deferment(void) 2531{ 2532 struct rq *rq = this_rq(); 2533 unsigned long next, now = ACCESS_ONCE(jiffies); 2534 2535 next = rq->last_sched_tick + HZ; 2536 2537 if (time_before_eq(next, now)) 2538 return 0; 2539 2540 return jiffies_to_nsecs(next - now); 2541} 2542#endif 2543 2544notrace unsigned long get_parent_ip(unsigned long addr) 2545{ 2546 if (in_lock_functions(addr)) { 2547 addr = CALLER_ADDR2; 2548 if (in_lock_functions(addr)) 2549 addr = CALLER_ADDR3; 2550 } 2551 return addr; 2552} 2553 2554#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2555 defined(CONFIG_PREEMPT_TRACER)) 2556 2557void preempt_count_add(int val) 2558{ 2559#ifdef CONFIG_DEBUG_PREEMPT 2560 /* 2561 * Underflow? 2562 */ 2563 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2564 return; 2565#endif 2566 __preempt_count_add(val); 2567#ifdef CONFIG_DEBUG_PREEMPT 2568 /* 2569 * Spinlock count overflowing soon? 2570 */ 2571 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2572 PREEMPT_MASK - 10); 2573#endif 2574 if (preempt_count() == val) { 2575 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2576#ifdef CONFIG_DEBUG_PREEMPT 2577 current->preempt_disable_ip = ip; 2578#endif 2579 trace_preempt_off(CALLER_ADDR0, ip); 2580 } 2581} 2582EXPORT_SYMBOL(preempt_count_add); 2583NOKPROBE_SYMBOL(preempt_count_add); 2584 2585void preempt_count_sub(int val) 2586{ 2587#ifdef CONFIG_DEBUG_PREEMPT 2588 /* 2589 * Underflow? 2590 */ 2591 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2592 return; 2593 /* 2594 * Is the spinlock portion underflowing? 2595 */ 2596 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2597 !(preempt_count() & PREEMPT_MASK))) 2598 return; 2599#endif 2600 2601 if (preempt_count() == val) 2602 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2603 __preempt_count_sub(val); 2604} 2605EXPORT_SYMBOL(preempt_count_sub); 2606NOKPROBE_SYMBOL(preempt_count_sub); 2607 2608#endif 2609 2610/* 2611 * Print scheduling while atomic bug: 2612 */ 2613static noinline void __schedule_bug(struct task_struct *prev) 2614{ 2615 if (oops_in_progress) 2616 return; 2617 2618 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2619 prev->comm, prev->pid, preempt_count()); 2620 2621 debug_show_held_locks(prev); 2622 print_modules(); 2623 if (irqs_disabled()) 2624 print_irqtrace_events(prev); 2625#ifdef CONFIG_DEBUG_PREEMPT 2626 if (in_atomic_preempt_off()) { 2627 pr_err("Preemption disabled at:"); 2628 print_ip_sym(current->preempt_disable_ip); 2629 pr_cont("\n"); 2630 } 2631#endif 2632 dump_stack(); 2633 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2634} 2635 2636/* 2637 * Various schedule()-time debugging checks and statistics: 2638 */ 2639static inline void schedule_debug(struct task_struct *prev) 2640{ 2641 /* 2642 * Test if we are atomic. Since do_exit() needs to call into 2643 * schedule() atomically, we ignore that path. Otherwise whine 2644 * if we are scheduling when we should not. 2645 */ 2646 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2647 __schedule_bug(prev); 2648 rcu_sleep_check(); 2649 2650 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2651 2652 schedstat_inc(this_rq(), sched_count); 2653} 2654 2655/* 2656 * Pick up the highest-prio task: 2657 */ 2658static inline struct task_struct * 2659pick_next_task(struct rq *rq, struct task_struct *prev) 2660{ 2661 const struct sched_class *class = &fair_sched_class; 2662 struct task_struct *p; 2663 2664 /* 2665 * Optimization: we know that if all tasks are in 2666 * the fair class we can call that function directly: 2667 */ 2668 if (likely(prev->sched_class == class && 2669 rq->nr_running == rq->cfs.h_nr_running)) { 2670 p = fair_sched_class.pick_next_task(rq, prev); 2671 if (unlikely(p == RETRY_TASK)) 2672 goto again; 2673 2674 /* assumes fair_sched_class->next == idle_sched_class */ 2675 if (unlikely(!p)) 2676 p = idle_sched_class.pick_next_task(rq, prev); 2677 2678 return p; 2679 } 2680 2681again: 2682 for_each_class(class) { 2683 p = class->pick_next_task(rq, prev); 2684 if (p) { 2685 if (unlikely(p == RETRY_TASK)) 2686 goto again; 2687 return p; 2688 } 2689 } 2690 2691 BUG(); /* the idle class will always have a runnable task */ 2692} 2693 2694/* 2695 * __schedule() is the main scheduler function. 2696 * 2697 * The main means of driving the scheduler and thus entering this function are: 2698 * 2699 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2700 * 2701 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2702 * paths. For example, see arch/x86/entry_64.S. 2703 * 2704 * To drive preemption between tasks, the scheduler sets the flag in timer 2705 * interrupt handler scheduler_tick(). 2706 * 2707 * 3. Wakeups don't really cause entry into schedule(). They add a 2708 * task to the run-queue and that's it. 2709 * 2710 * Now, if the new task added to the run-queue preempts the current 2711 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2712 * called on the nearest possible occasion: 2713 * 2714 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2715 * 2716 * - in syscall or exception context, at the next outmost 2717 * preempt_enable(). (this might be as soon as the wake_up()'s 2718 * spin_unlock()!) 2719 * 2720 * - in IRQ context, return from interrupt-handler to 2721 * preemptible context 2722 * 2723 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2724 * then at the next: 2725 * 2726 * - cond_resched() call 2727 * - explicit schedule() call 2728 * - return from syscall or exception to user-space 2729 * - return from interrupt-handler to user-space 2730 */ 2731static void __sched __schedule(void) 2732{ 2733 struct task_struct *prev, *next; 2734 unsigned long *switch_count; 2735 struct rq *rq; 2736 int cpu; 2737 2738need_resched: 2739 preempt_disable(); 2740 cpu = smp_processor_id(); 2741 rq = cpu_rq(cpu); 2742 rcu_note_context_switch(cpu); 2743 prev = rq->curr; 2744 2745 schedule_debug(prev); 2746 2747 if (sched_feat(HRTICK)) 2748 hrtick_clear(rq); 2749 2750 /* 2751 * Make sure that signal_pending_state()->signal_pending() below 2752 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2753 * done by the caller to avoid the race with signal_wake_up(). 2754 */ 2755 smp_mb__before_spinlock(); 2756 raw_spin_lock_irq(&rq->lock); 2757 2758 switch_count = &prev->nivcsw; 2759 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2760 if (unlikely(signal_pending_state(prev->state, prev))) { 2761 prev->state = TASK_RUNNING; 2762 } else { 2763 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2764 prev->on_rq = 0; 2765 2766 /* 2767 * If a worker went to sleep, notify and ask workqueue 2768 * whether it wants to wake up a task to maintain 2769 * concurrency. 2770 */ 2771 if (prev->flags & PF_WQ_WORKER) { 2772 struct task_struct *to_wakeup; 2773 2774 to_wakeup = wq_worker_sleeping(prev, cpu); 2775 if (to_wakeup) 2776 try_to_wake_up_local(to_wakeup); 2777 } 2778 } 2779 switch_count = &prev->nvcsw; 2780 } 2781 2782 if (prev->on_rq || rq->skip_clock_update < 0) 2783 update_rq_clock(rq); 2784 2785 next = pick_next_task(rq, prev); 2786 clear_tsk_need_resched(prev); 2787 clear_preempt_need_resched(); 2788 rq->skip_clock_update = 0; 2789 2790 if (likely(prev != next)) { 2791 rq->nr_switches++; 2792 rq->curr = next; 2793 ++*switch_count; 2794 2795 context_switch(rq, prev, next); /* unlocks the rq */ 2796 /* 2797 * The context switch have flipped the stack from under us 2798 * and restored the local variables which were saved when 2799 * this task called schedule() in the past. prev == current 2800 * is still correct, but it can be moved to another cpu/rq. 2801 */ 2802 cpu = smp_processor_id(); 2803 rq = cpu_rq(cpu); 2804 } else 2805 raw_spin_unlock_irq(&rq->lock); 2806 2807 post_schedule(rq); 2808 2809 sched_preempt_enable_no_resched(); 2810 if (need_resched()) 2811 goto need_resched; 2812} 2813 2814static inline void sched_submit_work(struct task_struct *tsk) 2815{ 2816 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2817 return; 2818 /* 2819 * If we are going to sleep and we have plugged IO queued, 2820 * make sure to submit it to avoid deadlocks. 2821 */ 2822 if (blk_needs_flush_plug(tsk)) 2823 blk_schedule_flush_plug(tsk); 2824} 2825 2826asmlinkage __visible void __sched schedule(void) 2827{ 2828 struct task_struct *tsk = current; 2829 2830 sched_submit_work(tsk); 2831 __schedule(); 2832} 2833EXPORT_SYMBOL(schedule); 2834 2835#ifdef CONFIG_CONTEXT_TRACKING 2836asmlinkage __visible void __sched schedule_user(void) 2837{ 2838 /* 2839 * If we come here after a random call to set_need_resched(), 2840 * or we have been woken up remotely but the IPI has not yet arrived, 2841 * we haven't yet exited the RCU idle mode. Do it here manually until 2842 * we find a better solution. 2843 */ 2844 user_exit(); 2845 schedule(); 2846 user_enter(); 2847} 2848#endif 2849 2850/** 2851 * schedule_preempt_disabled - called with preemption disabled 2852 * 2853 * Returns with preemption disabled. Note: preempt_count must be 1 2854 */ 2855void __sched schedule_preempt_disabled(void) 2856{ 2857 sched_preempt_enable_no_resched(); 2858 schedule(); 2859 preempt_disable(); 2860} 2861 2862#ifdef CONFIG_PREEMPT 2863/* 2864 * this is the entry point to schedule() from in-kernel preemption 2865 * off of preempt_enable. Kernel preemptions off return from interrupt 2866 * occur there and call schedule directly. 2867 */ 2868asmlinkage __visible void __sched notrace preempt_schedule(void) 2869{ 2870 /* 2871 * If there is a non-zero preempt_count or interrupts are disabled, 2872 * we do not want to preempt the current task. Just return.. 2873 */ 2874 if (likely(!preemptible())) 2875 return; 2876 2877 do { 2878 __preempt_count_add(PREEMPT_ACTIVE); 2879 __schedule(); 2880 __preempt_count_sub(PREEMPT_ACTIVE); 2881 2882 /* 2883 * Check again in case we missed a preemption opportunity 2884 * between schedule and now. 2885 */ 2886 barrier(); 2887 } while (need_resched()); 2888} 2889NOKPROBE_SYMBOL(preempt_schedule); 2890EXPORT_SYMBOL(preempt_schedule); 2891#endif /* CONFIG_PREEMPT */ 2892 2893/* 2894 * this is the entry point to schedule() from kernel preemption 2895 * off of irq context. 2896 * Note, that this is called and return with irqs disabled. This will 2897 * protect us against recursive calling from irq. 2898 */ 2899asmlinkage __visible void __sched preempt_schedule_irq(void) 2900{ 2901 enum ctx_state prev_state; 2902 2903 /* Catch callers which need to be fixed */ 2904 BUG_ON(preempt_count() || !irqs_disabled()); 2905 2906 prev_state = exception_enter(); 2907 2908 do { 2909 __preempt_count_add(PREEMPT_ACTIVE); 2910 local_irq_enable(); 2911 __schedule(); 2912 local_irq_disable(); 2913 __preempt_count_sub(PREEMPT_ACTIVE); 2914 2915 /* 2916 * Check again in case we missed a preemption opportunity 2917 * between schedule and now. 2918 */ 2919 barrier(); 2920 } while (need_resched()); 2921 2922 exception_exit(prev_state); 2923} 2924 2925int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2926 void *key) 2927{ 2928 return try_to_wake_up(curr->private, mode, wake_flags); 2929} 2930EXPORT_SYMBOL(default_wake_function); 2931 2932#ifdef CONFIG_RT_MUTEXES 2933 2934/* 2935 * rt_mutex_setprio - set the current priority of a task 2936 * @p: task 2937 * @prio: prio value (kernel-internal form) 2938 * 2939 * This function changes the 'effective' priority of a task. It does 2940 * not touch ->normal_prio like __setscheduler(). 2941 * 2942 * Used by the rt_mutex code to implement priority inheritance 2943 * logic. Call site only calls if the priority of the task changed. 2944 */ 2945void rt_mutex_setprio(struct task_struct *p, int prio) 2946{ 2947 int oldprio, on_rq, running, enqueue_flag = 0; 2948 struct rq *rq; 2949 const struct sched_class *prev_class; 2950 2951 BUG_ON(prio > MAX_PRIO); 2952 2953 rq = __task_rq_lock(p); 2954 2955 /* 2956 * Idle task boosting is a nono in general. There is one 2957 * exception, when PREEMPT_RT and NOHZ is active: 2958 * 2959 * The idle task calls get_next_timer_interrupt() and holds 2960 * the timer wheel base->lock on the CPU and another CPU wants 2961 * to access the timer (probably to cancel it). We can safely 2962 * ignore the boosting request, as the idle CPU runs this code 2963 * with interrupts disabled and will complete the lock 2964 * protected section without being interrupted. So there is no 2965 * real need to boost. 2966 */ 2967 if (unlikely(p == rq->idle)) { 2968 WARN_ON(p != rq->curr); 2969 WARN_ON(p->pi_blocked_on); 2970 goto out_unlock; 2971 } 2972 2973 trace_sched_pi_setprio(p, prio); 2974 p->pi_top_task = rt_mutex_get_top_task(p); 2975 oldprio = p->prio; 2976 prev_class = p->sched_class; 2977 on_rq = p->on_rq; 2978 running = task_current(rq, p); 2979 if (on_rq) 2980 dequeue_task(rq, p, 0); 2981 if (running) 2982 p->sched_class->put_prev_task(rq, p); 2983 2984 /* 2985 * Boosting condition are: 2986 * 1. -rt task is running and holds mutex A 2987 * --> -dl task blocks on mutex A 2988 * 2989 * 2. -dl task is running and holds mutex A 2990 * --> -dl task blocks on mutex A and could preempt the 2991 * running task 2992 */ 2993 if (dl_prio(prio)) { 2994 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2995 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2996 p->dl.dl_boosted = 1; 2997 p->dl.dl_throttled = 0; 2998 enqueue_flag = ENQUEUE_REPLENISH; 2999 } else 3000 p->dl.dl_boosted = 0; 3001 p->sched_class = &dl_sched_class; 3002 } else if (rt_prio(prio)) { 3003 if (dl_prio(oldprio)) 3004 p->dl.dl_boosted = 0; 3005 if (oldprio < prio) 3006 enqueue_flag = ENQUEUE_HEAD; 3007 p->sched_class = &rt_sched_class; 3008 } else { 3009 if (dl_prio(oldprio)) 3010 p->dl.dl_boosted = 0; 3011 p->sched_class = &fair_sched_class; 3012 } 3013 3014 p->prio = prio; 3015 3016 if (running) 3017 p->sched_class->set_curr_task(rq); 3018 if (on_rq) 3019 enqueue_task(rq, p, enqueue_flag); 3020 3021 check_class_changed(rq, p, prev_class, oldprio); 3022out_unlock: 3023 __task_rq_unlock(rq); 3024} 3025#endif 3026 3027void set_user_nice(struct task_struct *p, long nice) 3028{ 3029 int old_prio, delta, on_rq; 3030 unsigned long flags; 3031 struct rq *rq; 3032 3033 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 3034 return; 3035 /* 3036 * We have to be careful, if called from sys_setpriority(), 3037 * the task might be in the middle of scheduling on another CPU. 3038 */ 3039 rq = task_rq_lock(p, &flags); 3040 /* 3041 * The RT priorities are set via sched_setscheduler(), but we still 3042 * allow the 'normal' nice value to be set - but as expected 3043 * it wont have any effect on scheduling until the task is 3044 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3045 */ 3046 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3047 p->static_prio = NICE_TO_PRIO(nice); 3048 goto out_unlock; 3049 } 3050 on_rq = p->on_rq; 3051 if (on_rq) 3052 dequeue_task(rq, p, 0); 3053 3054 p->static_prio = NICE_TO_PRIO(nice); 3055 set_load_weight(p); 3056 old_prio = p->prio; 3057 p->prio = effective_prio(p); 3058 delta = p->prio - old_prio; 3059 3060 if (on_rq) { 3061 enqueue_task(rq, p, 0); 3062 /* 3063 * If the task increased its priority or is running and 3064 * lowered its priority, then reschedule its CPU: 3065 */ 3066 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3067 resched_task(rq->curr); 3068 } 3069out_unlock: 3070 task_rq_unlock(rq, p, &flags); 3071} 3072EXPORT_SYMBOL(set_user_nice); 3073 3074/* 3075 * can_nice - check if a task can reduce its nice value 3076 * @p: task 3077 * @nice: nice value 3078 */ 3079int can_nice(const struct task_struct *p, const int nice) 3080{ 3081 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3082 int nice_rlim = nice_to_rlimit(nice); 3083 3084 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3085 capable(CAP_SYS_NICE)); 3086} 3087 3088#ifdef __ARCH_WANT_SYS_NICE 3089 3090/* 3091 * sys_nice - change the priority of the current process. 3092 * @increment: priority increment 3093 * 3094 * sys_setpriority is a more generic, but much slower function that 3095 * does similar things. 3096 */ 3097SYSCALL_DEFINE1(nice, int, increment) 3098{ 3099 long nice, retval; 3100 3101 /* 3102 * Setpriority might change our priority at the same moment. 3103 * We don't have to worry. Conceptually one call occurs first 3104 * and we have a single winner. 3105 */ 3106 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 3107 nice = task_nice(current) + increment; 3108 3109 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 3110 if (increment < 0 && !can_nice(current, nice)) 3111 return -EPERM; 3112 3113 retval = security_task_setnice(current, nice); 3114 if (retval) 3115 return retval; 3116 3117 set_user_nice(current, nice); 3118 return 0; 3119} 3120 3121#endif 3122 3123/** 3124 * task_prio - return the priority value of a given task. 3125 * @p: the task in question. 3126 * 3127 * Return: The priority value as seen by users in /proc. 3128 * RT tasks are offset by -200. Normal tasks are centered 3129 * around 0, value goes from -16 to +15. 3130 */ 3131int task_prio(const struct task_struct *p) 3132{ 3133 return p->prio - MAX_RT_PRIO; 3134} 3135 3136/** 3137 * idle_cpu - is a given cpu idle currently? 3138 * @cpu: the processor in question. 3139 * 3140 * Return: 1 if the CPU is currently idle. 0 otherwise. 3141 */ 3142int idle_cpu(int cpu) 3143{ 3144 struct rq *rq = cpu_rq(cpu); 3145 3146 if (rq->curr != rq->idle) 3147 return 0; 3148 3149 if (rq->nr_running) 3150 return 0; 3151 3152#ifdef CONFIG_SMP 3153 if (!llist_empty(&rq->wake_list)) 3154 return 0; 3155#endif 3156 3157 return 1; 3158} 3159 3160/** 3161 * idle_task - return the idle task for a given cpu. 3162 * @cpu: the processor in question. 3163 * 3164 * Return: The idle task for the cpu @cpu. 3165 */ 3166struct task_struct *idle_task(int cpu) 3167{ 3168 return cpu_rq(cpu)->idle; 3169} 3170 3171/** 3172 * find_process_by_pid - find a process with a matching PID value. 3173 * @pid: the pid in question. 3174 * 3175 * The task of @pid, if found. %NULL otherwise. 3176 */ 3177static struct task_struct *find_process_by_pid(pid_t pid) 3178{ 3179 return pid ? find_task_by_vpid(pid) : current; 3180} 3181 3182/* 3183 * This function initializes the sched_dl_entity of a newly becoming 3184 * SCHED_DEADLINE task. 3185 * 3186 * Only the static values are considered here, the actual runtime and the 3187 * absolute deadline will be properly calculated when the task is enqueued 3188 * for the first time with its new policy. 3189 */ 3190static void 3191__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3192{ 3193 struct sched_dl_entity *dl_se = &p->dl; 3194 3195 init_dl_task_timer(dl_se); 3196 dl_se->dl_runtime = attr->sched_runtime; 3197 dl_se->dl_deadline = attr->sched_deadline; 3198 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3199 dl_se->flags = attr->sched_flags; 3200 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3201 dl_se->dl_throttled = 0; 3202 dl_se->dl_new = 1; 3203 dl_se->dl_yielded = 0; 3204} 3205 3206static void __setscheduler_params(struct task_struct *p, 3207 const struct sched_attr *attr) 3208{ 3209 int policy = attr->sched_policy; 3210 3211 if (policy == -1) /* setparam */ 3212 policy = p->policy; 3213 3214 p->policy = policy; 3215 3216 if (dl_policy(policy)) 3217 __setparam_dl(p, attr); 3218 else if (fair_policy(policy)) 3219 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3220 3221 /* 3222 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3223 * !rt_policy. Always setting this ensures that things like 3224 * getparam()/getattr() don't report silly values for !rt tasks. 3225 */ 3226 p->rt_priority = attr->sched_priority; 3227 p->normal_prio = normal_prio(p); 3228 set_load_weight(p); 3229} 3230 3231/* Actually do priority change: must hold pi & rq lock. */ 3232static void __setscheduler(struct rq *rq, struct task_struct *p, 3233 const struct sched_attr *attr) 3234{ 3235 __setscheduler_params(p, attr); 3236 3237 /* 3238 * If we get here, there was no pi waiters boosting the 3239 * task. It is safe to use the normal prio. 3240 */ 3241 p->prio = normal_prio(p); 3242 3243 if (dl_prio(p->prio)) 3244 p->sched_class = &dl_sched_class; 3245 else if (rt_prio(p->prio)) 3246 p->sched_class = &rt_sched_class; 3247 else 3248 p->sched_class = &fair_sched_class; 3249} 3250 3251static void 3252__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3253{ 3254 struct sched_dl_entity *dl_se = &p->dl; 3255 3256 attr->sched_priority = p->rt_priority; 3257 attr->sched_runtime = dl_se->dl_runtime; 3258 attr->sched_deadline = dl_se->dl_deadline; 3259 attr->sched_period = dl_se->dl_period; 3260 attr->sched_flags = dl_se->flags; 3261} 3262 3263/* 3264 * This function validates the new parameters of a -deadline task. 3265 * We ask for the deadline not being zero, and greater or equal 3266 * than the runtime, as well as the period of being zero or 3267 * greater than deadline. Furthermore, we have to be sure that 3268 * user parameters are above the internal resolution of 1us (we 3269 * check sched_runtime only since it is always the smaller one) and 3270 * below 2^63 ns (we have to check both sched_deadline and 3271 * sched_period, as the latter can be zero). 3272 */ 3273static bool 3274__checkparam_dl(const struct sched_attr *attr) 3275{ 3276 /* deadline != 0 */ 3277 if (attr->sched_deadline == 0) 3278 return false; 3279 3280 /* 3281 * Since we truncate DL_SCALE bits, make sure we're at least 3282 * that big. 3283 */ 3284 if (attr->sched_runtime < (1ULL << DL_SCALE)) 3285 return false; 3286 3287 /* 3288 * Since we use the MSB for wrap-around and sign issues, make 3289 * sure it's not set (mind that period can be equal to zero). 3290 */ 3291 if (attr->sched_deadline & (1ULL << 63) || 3292 attr->sched_period & (1ULL << 63)) 3293 return false; 3294 3295 /* runtime <= deadline <= period (if period != 0) */ 3296 if ((attr->sched_period != 0 && 3297 attr->sched_period < attr->sched_deadline) || 3298 attr->sched_deadline < attr->sched_runtime) 3299 return false; 3300 3301 return true; 3302} 3303 3304/* 3305 * check the target process has a UID that matches the current process's 3306 */ 3307static bool check_same_owner(struct task_struct *p) 3308{ 3309 const struct cred *cred = current_cred(), *pcred; 3310 bool match; 3311 3312 rcu_read_lock(); 3313 pcred = __task_cred(p); 3314 match = (uid_eq(cred->euid, pcred->euid) || 3315 uid_eq(cred->euid, pcred->uid)); 3316 rcu_read_unlock(); 3317 return match; 3318} 3319 3320static int __sched_setscheduler(struct task_struct *p, 3321 const struct sched_attr *attr, 3322 bool user) 3323{ 3324 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3325 MAX_RT_PRIO - 1 - attr->sched_priority; 3326 int retval, oldprio, oldpolicy = -1, on_rq, running; 3327 int policy = attr->sched_policy; 3328 unsigned long flags; 3329 const struct sched_class *prev_class; 3330 struct rq *rq; 3331 int reset_on_fork; 3332 3333 /* may grab non-irq protected spin_locks */ 3334 BUG_ON(in_interrupt()); 3335recheck: 3336 /* double check policy once rq lock held */ 3337 if (policy < 0) { 3338 reset_on_fork = p->sched_reset_on_fork; 3339 policy = oldpolicy = p->policy; 3340 } else { 3341 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3342 3343 if (policy != SCHED_DEADLINE && 3344 policy != SCHED_FIFO && policy != SCHED_RR && 3345 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3346 policy != SCHED_IDLE) 3347 return -EINVAL; 3348 } 3349 3350 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3351 return -EINVAL; 3352 3353 /* 3354 * Valid priorities for SCHED_FIFO and SCHED_RR are 3355 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3356 * SCHED_BATCH and SCHED_IDLE is 0. 3357 */ 3358 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3359 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3360 return -EINVAL; 3361 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3362 (rt_policy(policy) != (attr->sched_priority != 0))) 3363 return -EINVAL; 3364 3365 /* 3366 * Allow unprivileged RT tasks to decrease priority: 3367 */ 3368 if (user && !capable(CAP_SYS_NICE)) { 3369 if (fair_policy(policy)) { 3370 if (attr->sched_nice < task_nice(p) && 3371 !can_nice(p, attr->sched_nice)) 3372 return -EPERM; 3373 } 3374 3375 if (rt_policy(policy)) { 3376 unsigned long rlim_rtprio = 3377 task_rlimit(p, RLIMIT_RTPRIO); 3378 3379 /* can't set/change the rt policy */ 3380 if (policy != p->policy && !rlim_rtprio) 3381 return -EPERM; 3382 3383 /* can't increase priority */ 3384 if (attr->sched_priority > p->rt_priority && 3385 attr->sched_priority > rlim_rtprio) 3386 return -EPERM; 3387 } 3388 3389 /* 3390 * Can't set/change SCHED_DEADLINE policy at all for now 3391 * (safest behavior); in the future we would like to allow 3392 * unprivileged DL tasks to increase their relative deadline 3393 * or reduce their runtime (both ways reducing utilization) 3394 */ 3395 if (dl_policy(policy)) 3396 return -EPERM; 3397 3398 /* 3399 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3400 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3401 */ 3402 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3403 if (!can_nice(p, task_nice(p))) 3404 return -EPERM; 3405 } 3406 3407 /* can't change other user's priorities */ 3408 if (!check_same_owner(p)) 3409 return -EPERM; 3410 3411 /* Normal users shall not reset the sched_reset_on_fork flag */ 3412 if (p->sched_reset_on_fork && !reset_on_fork) 3413 return -EPERM; 3414 } 3415 3416 if (user) { 3417 retval = security_task_setscheduler(p); 3418 if (retval) 3419 return retval; 3420 } 3421 3422 /* 3423 * make sure no PI-waiters arrive (or leave) while we are 3424 * changing the priority of the task: 3425 * 3426 * To be able to change p->policy safely, the appropriate 3427 * runqueue lock must be held. 3428 */ 3429 rq = task_rq_lock(p, &flags); 3430 3431 /* 3432 * Changing the policy of the stop threads its a very bad idea 3433 */ 3434 if (p == rq->stop) { 3435 task_rq_unlock(rq, p, &flags); 3436 return -EINVAL; 3437 } 3438 3439 /* 3440 * If not changing anything there's no need to proceed further, 3441 * but store a possible modification of reset_on_fork. 3442 */ 3443 if (unlikely(policy == p->policy)) { 3444 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3445 goto change; 3446 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3447 goto change; 3448 if (dl_policy(policy)) 3449 goto change; 3450 3451 p->sched_reset_on_fork = reset_on_fork; 3452 task_rq_unlock(rq, p, &flags); 3453 return 0; 3454 } 3455change: 3456 3457 if (user) { 3458#ifdef CONFIG_RT_GROUP_SCHED 3459 /* 3460 * Do not allow realtime tasks into groups that have no runtime 3461 * assigned. 3462 */ 3463 if (rt_bandwidth_enabled() && rt_policy(policy) && 3464 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3465 !task_group_is_autogroup(task_group(p))) { 3466 task_rq_unlock(rq, p, &flags); 3467 return -EPERM; 3468 } 3469#endif 3470#ifdef CONFIG_SMP 3471 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3472 cpumask_t *span = rq->rd->span; 3473 3474 /* 3475 * Don't allow tasks with an affinity mask smaller than 3476 * the entire root_domain to become SCHED_DEADLINE. We 3477 * will also fail if there's no bandwidth available. 3478 */ 3479 if (!cpumask_subset(span, &p->cpus_allowed) || 3480 rq->rd->dl_bw.bw == 0) { 3481 task_rq_unlock(rq, p, &flags); 3482 return -EPERM; 3483 } 3484 } 3485#endif 3486 } 3487 3488 /* recheck policy now with rq lock held */ 3489 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3490 policy = oldpolicy = -1; 3491 task_rq_unlock(rq, p, &flags); 3492 goto recheck; 3493 } 3494 3495 /* 3496 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3497 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3498 * is available. 3499 */ 3500 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3501 task_rq_unlock(rq, p, &flags); 3502 return -EBUSY; 3503 } 3504 3505 p->sched_reset_on_fork = reset_on_fork; 3506 oldprio = p->prio; 3507 3508 /* 3509 * Special case for priority boosted tasks. 3510 * 3511 * If the new priority is lower or equal (user space view) 3512 * than the current (boosted) priority, we just store the new 3513 * normal parameters and do not touch the scheduler class and 3514 * the runqueue. This will be done when the task deboost 3515 * itself. 3516 */ 3517 if (rt_mutex_check_prio(p, newprio)) { 3518 __setscheduler_params(p, attr); 3519 task_rq_unlock(rq, p, &flags); 3520 return 0; 3521 } 3522 3523 on_rq = p->on_rq; 3524 running = task_current(rq, p); 3525 if (on_rq) 3526 dequeue_task(rq, p, 0); 3527 if (running) 3528 p->sched_class->put_prev_task(rq, p); 3529 3530 prev_class = p->sched_class; 3531 __setscheduler(rq, p, attr); 3532 3533 if (running) 3534 p->sched_class->set_curr_task(rq); 3535 if (on_rq) { 3536 /* 3537 * We enqueue to tail when the priority of a task is 3538 * increased (user space view). 3539 */ 3540 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3541 } 3542 3543 check_class_changed(rq, p, prev_class, oldprio); 3544 task_rq_unlock(rq, p, &flags); 3545 3546 rt_mutex_adjust_pi(p); 3547 3548 return 0; 3549} 3550 3551static int _sched_setscheduler(struct task_struct *p, int policy, 3552 const struct sched_param *param, bool check) 3553{ 3554 struct sched_attr attr = { 3555 .sched_policy = policy, 3556 .sched_priority = param->sched_priority, 3557 .sched_nice = PRIO_TO_NICE(p->static_prio), 3558 }; 3559 3560 /* 3561 * Fixup the legacy SCHED_RESET_ON_FORK hack, except if 3562 * the policy=-1 was passed by sched_setparam(). 3563 */ 3564 if ((policy != -1) && (policy & SCHED_RESET_ON_FORK)) { 3565 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3566 policy &= ~SCHED_RESET_ON_FORK; 3567 attr.sched_policy = policy; 3568 } 3569 3570 return __sched_setscheduler(p, &attr, check); 3571} 3572/** 3573 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3574 * @p: the task in question. 3575 * @policy: new policy. 3576 * @param: structure containing the new RT priority. 3577 * 3578 * Return: 0 on success. An error code otherwise. 3579 * 3580 * NOTE that the task may be already dead. 3581 */ 3582int sched_setscheduler(struct task_struct *p, int policy, 3583 const struct sched_param *param) 3584{ 3585 return _sched_setscheduler(p, policy, param, true); 3586} 3587EXPORT_SYMBOL_GPL(sched_setscheduler); 3588 3589int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3590{ 3591 return __sched_setscheduler(p, attr, true); 3592} 3593EXPORT_SYMBOL_GPL(sched_setattr); 3594 3595/** 3596 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3597 * @p: the task in question. 3598 * @policy: new policy. 3599 * @param: structure containing the new RT priority. 3600 * 3601 * Just like sched_setscheduler, only don't bother checking if the 3602 * current context has permission. For example, this is needed in 3603 * stop_machine(): we create temporary high priority worker threads, 3604 * but our caller might not have that capability. 3605 * 3606 * Return: 0 on success. An error code otherwise. 3607 */ 3608int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3609 const struct sched_param *param) 3610{ 3611 return _sched_setscheduler(p, policy, param, false); 3612} 3613 3614static int 3615do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3616{ 3617 struct sched_param lparam; 3618 struct task_struct *p; 3619 int retval; 3620 3621 if (!param || pid < 0) 3622 return -EINVAL; 3623 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3624 return -EFAULT; 3625 3626 rcu_read_lock(); 3627 retval = -ESRCH; 3628 p = find_process_by_pid(pid); 3629 if (p != NULL) 3630 retval = sched_setscheduler(p, policy, &lparam); 3631 rcu_read_unlock(); 3632 3633 return retval; 3634} 3635 3636/* 3637 * Mimics kernel/events/core.c perf_copy_attr(). 3638 */ 3639static int sched_copy_attr(struct sched_attr __user *uattr, 3640 struct sched_attr *attr) 3641{ 3642 u32 size; 3643 int ret; 3644 3645 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3646 return -EFAULT; 3647 3648 /* 3649 * zero the full structure, so that a short copy will be nice. 3650 */ 3651 memset(attr, 0, sizeof(*attr)); 3652 3653 ret = get_user(size, &uattr->size); 3654 if (ret) 3655 return ret; 3656 3657 if (size > PAGE_SIZE) /* silly large */ 3658 goto err_size; 3659 3660 if (!size) /* abi compat */ 3661 size = SCHED_ATTR_SIZE_VER0; 3662 3663 if (size < SCHED_ATTR_SIZE_VER0) 3664 goto err_size; 3665 3666 /* 3667 * If we're handed a bigger struct than we know of, 3668 * ensure all the unknown bits are 0 - i.e. new 3669 * user-space does not rely on any kernel feature 3670 * extensions we dont know about yet. 3671 */ 3672 if (size > sizeof(*attr)) { 3673 unsigned char __user *addr; 3674 unsigned char __user *end; 3675 unsigned char val; 3676 3677 addr = (void __user *)uattr + sizeof(*attr); 3678 end = (void __user *)uattr + size; 3679 3680 for (; addr < end; addr++) { 3681 ret = get_user(val, addr); 3682 if (ret) 3683 return ret; 3684 if (val) 3685 goto err_size; 3686 } 3687 size = sizeof(*attr); 3688 } 3689 3690 ret = copy_from_user(attr, uattr, size); 3691 if (ret) 3692 return -EFAULT; 3693 3694 /* 3695 * XXX: do we want to be lenient like existing syscalls; or do we want 3696 * to be strict and return an error on out-of-bounds values? 3697 */ 3698 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3699 3700 return 0; 3701 3702err_size: 3703 put_user(sizeof(*attr), &uattr->size); 3704 return -E2BIG; 3705} 3706 3707/** 3708 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3709 * @pid: the pid in question. 3710 * @policy: new policy. 3711 * @param: structure containing the new RT priority. 3712 * 3713 * Return: 0 on success. An error code otherwise. 3714 */ 3715SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3716 struct sched_param __user *, param) 3717{ 3718 /* negative values for policy are not valid */ 3719 if (policy < 0) 3720 return -EINVAL; 3721 3722 return do_sched_setscheduler(pid, policy, param); 3723} 3724 3725/** 3726 * sys_sched_setparam - set/change the RT priority of a thread 3727 * @pid: the pid in question. 3728 * @param: structure containing the new RT priority. 3729 * 3730 * Return: 0 on success. An error code otherwise. 3731 */ 3732SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3733{ 3734 return do_sched_setscheduler(pid, -1, param); 3735} 3736 3737/** 3738 * sys_sched_setattr - same as above, but with extended sched_attr 3739 * @pid: the pid in question. 3740 * @uattr: structure containing the extended parameters. 3741 * @flags: for future extension. 3742 */ 3743SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3744 unsigned int, flags) 3745{ 3746 struct sched_attr attr; 3747 struct task_struct *p; 3748 int retval; 3749 3750 if (!uattr || pid < 0 || flags) 3751 return -EINVAL; 3752 3753 retval = sched_copy_attr(uattr, &attr); 3754 if (retval) 3755 return retval; 3756 3757 if ((int)attr.sched_policy < 0) 3758 return -EINVAL; 3759 3760 rcu_read_lock(); 3761 retval = -ESRCH; 3762 p = find_process_by_pid(pid); 3763 if (p != NULL) 3764 retval = sched_setattr(p, &attr); 3765 rcu_read_unlock(); 3766 3767 return retval; 3768} 3769 3770/** 3771 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3772 * @pid: the pid in question. 3773 * 3774 * Return: On success, the policy of the thread. Otherwise, a negative error 3775 * code. 3776 */ 3777SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3778{ 3779 struct task_struct *p; 3780 int retval; 3781 3782 if (pid < 0) 3783 return -EINVAL; 3784 3785 retval = -ESRCH; 3786 rcu_read_lock(); 3787 p = find_process_by_pid(pid); 3788 if (p) { 3789 retval = security_task_getscheduler(p); 3790 if (!retval) 3791 retval = p->policy 3792 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3793 } 3794 rcu_read_unlock(); 3795 return retval; 3796} 3797 3798/** 3799 * sys_sched_getparam - get the RT priority of a thread 3800 * @pid: the pid in question. 3801 * @param: structure containing the RT priority. 3802 * 3803 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3804 * code. 3805 */ 3806SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3807{ 3808 struct sched_param lp = { .sched_priority = 0 }; 3809 struct task_struct *p; 3810 int retval; 3811 3812 if (!param || pid < 0) 3813 return -EINVAL; 3814 3815 rcu_read_lock(); 3816 p = find_process_by_pid(pid); 3817 retval = -ESRCH; 3818 if (!p) 3819 goto out_unlock; 3820 3821 retval = security_task_getscheduler(p); 3822 if (retval) 3823 goto out_unlock; 3824 3825 if (task_has_rt_policy(p)) 3826 lp.sched_priority = p->rt_priority; 3827 rcu_read_unlock(); 3828 3829 /* 3830 * This one might sleep, we cannot do it with a spinlock held ... 3831 */ 3832 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3833 3834 return retval; 3835 3836out_unlock: 3837 rcu_read_unlock(); 3838 return retval; 3839} 3840 3841static int sched_read_attr(struct sched_attr __user *uattr, 3842 struct sched_attr *attr, 3843 unsigned int usize) 3844{ 3845 int ret; 3846 3847 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3848 return -EFAULT; 3849 3850 /* 3851 * If we're handed a smaller struct than we know of, 3852 * ensure all the unknown bits are 0 - i.e. old 3853 * user-space does not get uncomplete information. 3854 */ 3855 if (usize < sizeof(*attr)) { 3856 unsigned char *addr; 3857 unsigned char *end; 3858 3859 addr = (void *)attr + usize; 3860 end = (void *)attr + sizeof(*attr); 3861 3862 for (; addr < end; addr++) { 3863 if (*addr) 3864 return -EFBIG; 3865 } 3866 3867 attr->size = usize; 3868 } 3869 3870 ret = copy_to_user(uattr, attr, attr->size); 3871 if (ret) 3872 return -EFAULT; 3873 3874 return 0; 3875} 3876 3877/** 3878 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3879 * @pid: the pid in question. 3880 * @uattr: structure containing the extended parameters. 3881 * @size: sizeof(attr) for fwd/bwd comp. 3882 * @flags: for future extension. 3883 */ 3884SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3885 unsigned int, size, unsigned int, flags) 3886{ 3887 struct sched_attr attr = { 3888 .size = sizeof(struct sched_attr), 3889 }; 3890 struct task_struct *p; 3891 int retval; 3892 3893 if (!uattr || pid < 0 || size > PAGE_SIZE || 3894 size < SCHED_ATTR_SIZE_VER0 || flags) 3895 return -EINVAL; 3896 3897 rcu_read_lock(); 3898 p = find_process_by_pid(pid); 3899 retval = -ESRCH; 3900 if (!p) 3901 goto out_unlock; 3902 3903 retval = security_task_getscheduler(p); 3904 if (retval) 3905 goto out_unlock; 3906 3907 attr.sched_policy = p->policy; 3908 if (p->sched_reset_on_fork) 3909 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3910 if (task_has_dl_policy(p)) 3911 __getparam_dl(p, &attr); 3912 else if (task_has_rt_policy(p)) 3913 attr.sched_priority = p->rt_priority; 3914 else 3915 attr.sched_nice = task_nice(p); 3916 3917 rcu_read_unlock(); 3918 3919 retval = sched_read_attr(uattr, &attr, size); 3920 return retval; 3921 3922out_unlock: 3923 rcu_read_unlock(); 3924 return retval; 3925} 3926 3927long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3928{ 3929 cpumask_var_t cpus_allowed, new_mask; 3930 struct task_struct *p; 3931 int retval; 3932 3933 rcu_read_lock(); 3934 3935 p = find_process_by_pid(pid); 3936 if (!p) { 3937 rcu_read_unlock(); 3938 return -ESRCH; 3939 } 3940 3941 /* Prevent p going away */ 3942 get_task_struct(p); 3943 rcu_read_unlock(); 3944 3945 if (p->flags & PF_NO_SETAFFINITY) { 3946 retval = -EINVAL; 3947 goto out_put_task; 3948 } 3949 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3950 retval = -ENOMEM; 3951 goto out_put_task; 3952 } 3953 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3954 retval = -ENOMEM; 3955 goto out_free_cpus_allowed; 3956 } 3957 retval = -EPERM; 3958 if (!check_same_owner(p)) { 3959 rcu_read_lock(); 3960 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3961 rcu_read_unlock(); 3962 goto out_unlock; 3963 } 3964 rcu_read_unlock(); 3965 } 3966 3967 retval = security_task_setscheduler(p); 3968 if (retval) 3969 goto out_unlock; 3970 3971 3972 cpuset_cpus_allowed(p, cpus_allowed); 3973 cpumask_and(new_mask, in_mask, cpus_allowed); 3974 3975 /* 3976 * Since bandwidth control happens on root_domain basis, 3977 * if admission test is enabled, we only admit -deadline 3978 * tasks allowed to run on all the CPUs in the task's 3979 * root_domain. 3980 */ 3981#ifdef CONFIG_SMP 3982 if (task_has_dl_policy(p)) { 3983 const struct cpumask *span = task_rq(p)->rd->span; 3984 3985 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3986 retval = -EBUSY; 3987 goto out_unlock; 3988 } 3989 } 3990#endif 3991again: 3992 retval = set_cpus_allowed_ptr(p, new_mask); 3993 3994 if (!retval) { 3995 cpuset_cpus_allowed(p, cpus_allowed); 3996 if (!cpumask_subset(new_mask, cpus_allowed)) { 3997 /* 3998 * We must have raced with a concurrent cpuset 3999 * update. Just reset the cpus_allowed to the 4000 * cpuset's cpus_allowed 4001 */ 4002 cpumask_copy(new_mask, cpus_allowed); 4003 goto again; 4004 } 4005 } 4006out_unlock: 4007 free_cpumask_var(new_mask); 4008out_free_cpus_allowed: 4009 free_cpumask_var(cpus_allowed); 4010out_put_task: 4011 put_task_struct(p); 4012 return retval; 4013} 4014 4015static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4016 struct cpumask *new_mask) 4017{ 4018 if (len < cpumask_size()) 4019 cpumask_clear(new_mask); 4020 else if (len > cpumask_size()) 4021 len = cpumask_size(); 4022 4023 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4024} 4025 4026/** 4027 * sys_sched_setaffinity - set the cpu affinity of a process 4028 * @pid: pid of the process 4029 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4030 * @user_mask_ptr: user-space pointer to the new cpu mask 4031 * 4032 * Return: 0 on success. An error code otherwise. 4033 */ 4034SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4035 unsigned long __user *, user_mask_ptr) 4036{ 4037 cpumask_var_t new_mask; 4038 int retval; 4039 4040 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4041 return -ENOMEM; 4042 4043 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4044 if (retval == 0) 4045 retval = sched_setaffinity(pid, new_mask); 4046 free_cpumask_var(new_mask); 4047 return retval; 4048} 4049 4050long sched_getaffinity(pid_t pid, struct cpumask *mask) 4051{ 4052 struct task_struct *p; 4053 unsigned long flags; 4054 int retval; 4055 4056 rcu_read_lock(); 4057 4058 retval = -ESRCH; 4059 p = find_process_by_pid(pid); 4060 if (!p) 4061 goto out_unlock; 4062 4063 retval = security_task_getscheduler(p); 4064 if (retval) 4065 goto out_unlock; 4066 4067 raw_spin_lock_irqsave(&p->pi_lock, flags); 4068 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4069 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4070 4071out_unlock: 4072 rcu_read_unlock(); 4073 4074 return retval; 4075} 4076 4077/** 4078 * sys_sched_getaffinity - get the cpu affinity of a process 4079 * @pid: pid of the process 4080 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4081 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4082 * 4083 * Return: 0 on success. An error code otherwise. 4084 */ 4085SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4086 unsigned long __user *, user_mask_ptr) 4087{ 4088 int ret; 4089 cpumask_var_t mask; 4090 4091 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4092 return -EINVAL; 4093 if (len & (sizeof(unsigned long)-1)) 4094 return -EINVAL; 4095 4096 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4097 return -ENOMEM; 4098 4099 ret = sched_getaffinity(pid, mask); 4100 if (ret == 0) { 4101 size_t retlen = min_t(size_t, len, cpumask_size()); 4102 4103 if (copy_to_user(user_mask_ptr, mask, retlen)) 4104 ret = -EFAULT; 4105 else 4106 ret = retlen; 4107 } 4108 free_cpumask_var(mask); 4109 4110 return ret; 4111} 4112 4113/** 4114 * sys_sched_yield - yield the current processor to other threads. 4115 * 4116 * This function yields the current CPU to other tasks. If there are no 4117 * other threads running on this CPU then this function will return. 4118 * 4119 * Return: 0. 4120 */ 4121SYSCALL_DEFINE0(sched_yield) 4122{ 4123 struct rq *rq = this_rq_lock(); 4124 4125 schedstat_inc(rq, yld_count); 4126 current->sched_class->yield_task(rq); 4127 4128 /* 4129 * Since we are going to call schedule() anyway, there's 4130 * no need to preempt or enable interrupts: 4131 */ 4132 __release(rq->lock); 4133 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4134 do_raw_spin_unlock(&rq->lock); 4135 sched_preempt_enable_no_resched(); 4136 4137 schedule(); 4138 4139 return 0; 4140} 4141 4142static void __cond_resched(void) 4143{ 4144 __preempt_count_add(PREEMPT_ACTIVE); 4145 __schedule(); 4146 __preempt_count_sub(PREEMPT_ACTIVE); 4147} 4148 4149int __sched _cond_resched(void) 4150{ 4151 if (should_resched()) { 4152 __cond_resched(); 4153 return 1; 4154 } 4155 return 0; 4156} 4157EXPORT_SYMBOL(_cond_resched); 4158 4159/* 4160 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4161 * call schedule, and on return reacquire the lock. 4162 * 4163 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4164 * operations here to prevent schedule() from being called twice (once via 4165 * spin_unlock(), once by hand). 4166 */ 4167int __cond_resched_lock(spinlock_t *lock) 4168{ 4169 int resched = should_resched(); 4170 int ret = 0; 4171 4172 lockdep_assert_held(lock); 4173 4174 if (spin_needbreak(lock) || resched) { 4175 spin_unlock(lock); 4176 if (resched) 4177 __cond_resched(); 4178 else 4179 cpu_relax(); 4180 ret = 1; 4181 spin_lock(lock); 4182 } 4183 return ret; 4184} 4185EXPORT_SYMBOL(__cond_resched_lock); 4186 4187int __sched __cond_resched_softirq(void) 4188{ 4189 BUG_ON(!in_softirq()); 4190 4191 if (should_resched()) { 4192 local_bh_enable(); 4193 __cond_resched(); 4194 local_bh_disable(); 4195 return 1; 4196 } 4197 return 0; 4198} 4199EXPORT_SYMBOL(__cond_resched_softirq); 4200 4201/** 4202 * yield - yield the current processor to other threads. 4203 * 4204 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4205 * 4206 * The scheduler is at all times free to pick the calling task as the most 4207 * eligible task to run, if removing the yield() call from your code breaks 4208 * it, its already broken. 4209 * 4210 * Typical broken usage is: 4211 * 4212 * while (!event) 4213 * yield(); 4214 * 4215 * where one assumes that yield() will let 'the other' process run that will 4216 * make event true. If the current task is a SCHED_FIFO task that will never 4217 * happen. Never use yield() as a progress guarantee!! 4218 * 4219 * If you want to use yield() to wait for something, use wait_event(). 4220 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4221 * If you still want to use yield(), do not! 4222 */ 4223void __sched yield(void) 4224{ 4225 set_current_state(TASK_RUNNING); 4226 sys_sched_yield(); 4227} 4228EXPORT_SYMBOL(yield); 4229 4230/** 4231 * yield_to - yield the current processor to another thread in 4232 * your thread group, or accelerate that thread toward the 4233 * processor it's on. 4234 * @p: target task 4235 * @preempt: whether task preemption is allowed or not 4236 * 4237 * It's the caller's job to ensure that the target task struct 4238 * can't go away on us before we can do any checks. 4239 * 4240 * Return: 4241 * true (>0) if we indeed boosted the target task. 4242 * false (0) if we failed to boost the target. 4243 * -ESRCH if there's no task to yield to. 4244 */ 4245int __sched yield_to(struct task_struct *p, bool preempt) 4246{ 4247 struct task_struct *curr = current; 4248 struct rq *rq, *p_rq; 4249 unsigned long flags; 4250 int yielded = 0; 4251 4252 local_irq_save(flags); 4253 rq = this_rq(); 4254 4255again: 4256 p_rq = task_rq(p); 4257 /* 4258 * If we're the only runnable task on the rq and target rq also 4259 * has only one task, there's absolutely no point in yielding. 4260 */ 4261 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4262 yielded = -ESRCH; 4263 goto out_irq; 4264 } 4265 4266 double_rq_lock(rq, p_rq); 4267 if (task_rq(p) != p_rq) { 4268 double_rq_unlock(rq, p_rq); 4269 goto again; 4270 } 4271 4272 if (!curr->sched_class->yield_to_task) 4273 goto out_unlock; 4274 4275 if (curr->sched_class != p->sched_class) 4276 goto out_unlock; 4277 4278 if (task_running(p_rq, p) || p->state) 4279 goto out_unlock; 4280 4281 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4282 if (yielded) { 4283 schedstat_inc(rq, yld_count); 4284 /* 4285 * Make p's CPU reschedule; pick_next_entity takes care of 4286 * fairness. 4287 */ 4288 if (preempt && rq != p_rq) 4289 resched_task(p_rq->curr); 4290 } 4291 4292out_unlock: 4293 double_rq_unlock(rq, p_rq); 4294out_irq: 4295 local_irq_restore(flags); 4296 4297 if (yielded > 0) 4298 schedule(); 4299 4300 return yielded; 4301} 4302EXPORT_SYMBOL_GPL(yield_to); 4303 4304/* 4305 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4306 * that process accounting knows that this is a task in IO wait state. 4307 */ 4308void __sched io_schedule(void) 4309{ 4310 struct rq *rq = raw_rq(); 4311 4312 delayacct_blkio_start(); 4313 atomic_inc(&rq->nr_iowait); 4314 blk_flush_plug(current); 4315 current->in_iowait = 1; 4316 schedule(); 4317 current->in_iowait = 0; 4318 atomic_dec(&rq->nr_iowait); 4319 delayacct_blkio_end(); 4320} 4321EXPORT_SYMBOL(io_schedule); 4322 4323long __sched io_schedule_timeout(long timeout) 4324{ 4325 struct rq *rq = raw_rq(); 4326 long ret; 4327 4328 delayacct_blkio_start(); 4329 atomic_inc(&rq->nr_iowait); 4330 blk_flush_plug(current); 4331 current->in_iowait = 1; 4332 ret = schedule_timeout(timeout); 4333 current->in_iowait = 0; 4334 atomic_dec(&rq->nr_iowait); 4335 delayacct_blkio_end(); 4336 return ret; 4337} 4338 4339/** 4340 * sys_sched_get_priority_max - return maximum RT priority. 4341 * @policy: scheduling class. 4342 * 4343 * Return: On success, this syscall returns the maximum 4344 * rt_priority that can be used by a given scheduling class. 4345 * On failure, a negative error code is returned. 4346 */ 4347SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4348{ 4349 int ret = -EINVAL; 4350 4351 switch (policy) { 4352 case SCHED_FIFO: 4353 case SCHED_RR: 4354 ret = MAX_USER_RT_PRIO-1; 4355 break; 4356 case SCHED_DEADLINE: 4357 case SCHED_NORMAL: 4358 case SCHED_BATCH: 4359 case SCHED_IDLE: 4360 ret = 0; 4361 break; 4362 } 4363 return ret; 4364} 4365 4366/** 4367 * sys_sched_get_priority_min - return minimum RT priority. 4368 * @policy: scheduling class. 4369 * 4370 * Return: On success, this syscall returns the minimum 4371 * rt_priority that can be used by a given scheduling class. 4372 * On failure, a negative error code is returned. 4373 */ 4374SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4375{ 4376 int ret = -EINVAL; 4377 4378 switch (policy) { 4379 case SCHED_FIFO: 4380 case SCHED_RR: 4381 ret = 1; 4382 break; 4383 case SCHED_DEADLINE: 4384 case SCHED_NORMAL: 4385 case SCHED_BATCH: 4386 case SCHED_IDLE: 4387 ret = 0; 4388 } 4389 return ret; 4390} 4391 4392/** 4393 * sys_sched_rr_get_interval - return the default timeslice of a process. 4394 * @pid: pid of the process. 4395 * @interval: userspace pointer to the timeslice value. 4396 * 4397 * this syscall writes the default timeslice value of a given process 4398 * into the user-space timespec buffer. A value of '0' means infinity. 4399 * 4400 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4401 * an error code. 4402 */ 4403SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4404 struct timespec __user *, interval) 4405{ 4406 struct task_struct *p; 4407 unsigned int time_slice; 4408 unsigned long flags; 4409 struct rq *rq; 4410 int retval; 4411 struct timespec t; 4412 4413 if (pid < 0) 4414 return -EINVAL; 4415 4416 retval = -ESRCH; 4417 rcu_read_lock(); 4418 p = find_process_by_pid(pid); 4419 if (!p) 4420 goto out_unlock; 4421 4422 retval = security_task_getscheduler(p); 4423 if (retval) 4424 goto out_unlock; 4425 4426 rq = task_rq_lock(p, &flags); 4427 time_slice = 0; 4428 if (p->sched_class->get_rr_interval) 4429 time_slice = p->sched_class->get_rr_interval(rq, p); 4430 task_rq_unlock(rq, p, &flags); 4431 4432 rcu_read_unlock(); 4433 jiffies_to_timespec(time_slice, &t); 4434 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4435 return retval; 4436 4437out_unlock: 4438 rcu_read_unlock(); 4439 return retval; 4440} 4441 4442static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4443 4444void sched_show_task(struct task_struct *p) 4445{ 4446 unsigned long free = 0; 4447 int ppid; 4448 unsigned state; 4449 4450 state = p->state ? __ffs(p->state) + 1 : 0; 4451 printk(KERN_INFO "%-15.15s %c", p->comm, 4452 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4453#if BITS_PER_LONG == 32 4454 if (state == TASK_RUNNING) 4455 printk(KERN_CONT " running "); 4456 else 4457 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4458#else 4459 if (state == TASK_RUNNING) 4460 printk(KERN_CONT " running task "); 4461 else 4462 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4463#endif 4464#ifdef CONFIG_DEBUG_STACK_USAGE 4465 free = stack_not_used(p); 4466#endif 4467 rcu_read_lock(); 4468 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4469 rcu_read_unlock(); 4470 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4471 task_pid_nr(p), ppid, 4472 (unsigned long)task_thread_info(p)->flags); 4473 4474 print_worker_info(KERN_INFO, p); 4475 show_stack(p, NULL); 4476} 4477 4478void show_state_filter(unsigned long state_filter) 4479{ 4480 struct task_struct *g, *p; 4481 4482#if BITS_PER_LONG == 32 4483 printk(KERN_INFO 4484 " task PC stack pid father\n"); 4485#else 4486 printk(KERN_INFO 4487 " task PC stack pid father\n"); 4488#endif 4489 rcu_read_lock(); 4490 do_each_thread(g, p) { 4491 /* 4492 * reset the NMI-timeout, listing all files on a slow 4493 * console might take a lot of time: 4494 */ 4495 touch_nmi_watchdog(); 4496 if (!state_filter || (p->state & state_filter)) 4497 sched_show_task(p); 4498 } while_each_thread(g, p); 4499 4500 touch_all_softlockup_watchdogs(); 4501 4502#ifdef CONFIG_SCHED_DEBUG 4503 sysrq_sched_debug_show(); 4504#endif 4505 rcu_read_unlock(); 4506 /* 4507 * Only show locks if all tasks are dumped: 4508 */ 4509 if (!state_filter) 4510 debug_show_all_locks(); 4511} 4512 4513void init_idle_bootup_task(struct task_struct *idle) 4514{ 4515 idle->sched_class = &idle_sched_class; 4516} 4517 4518/** 4519 * init_idle - set up an idle thread for a given CPU 4520 * @idle: task in question 4521 * @cpu: cpu the idle task belongs to 4522 * 4523 * NOTE: this function does not set the idle thread's NEED_RESCHED 4524 * flag, to make booting more robust. 4525 */ 4526void init_idle(struct task_struct *idle, int cpu) 4527{ 4528 struct rq *rq = cpu_rq(cpu); 4529 unsigned long flags; 4530 4531 raw_spin_lock_irqsave(&rq->lock, flags); 4532 4533 __sched_fork(0, idle); 4534 idle->state = TASK_RUNNING; 4535 idle->se.exec_start = sched_clock(); 4536 4537 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4538 /* 4539 * We're having a chicken and egg problem, even though we are 4540 * holding rq->lock, the cpu isn't yet set to this cpu so the 4541 * lockdep check in task_group() will fail. 4542 * 4543 * Similar case to sched_fork(). / Alternatively we could 4544 * use task_rq_lock() here and obtain the other rq->lock. 4545 * 4546 * Silence PROVE_RCU 4547 */ 4548 rcu_read_lock(); 4549 __set_task_cpu(idle, cpu); 4550 rcu_read_unlock(); 4551 4552 rq->curr = rq->idle = idle; 4553 idle->on_rq = 1; 4554#if defined(CONFIG_SMP) 4555 idle->on_cpu = 1; 4556#endif 4557 raw_spin_unlock_irqrestore(&rq->lock, flags); 4558 4559 /* Set the preempt count _outside_ the spinlocks! */ 4560 init_idle_preempt_count(idle, cpu); 4561 4562 /* 4563 * The idle tasks have their own, simple scheduling class: 4564 */ 4565 idle->sched_class = &idle_sched_class; 4566 ftrace_graph_init_idle_task(idle, cpu); 4567 vtime_init_idle(idle, cpu); 4568#if defined(CONFIG_SMP) 4569 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4570#endif 4571} 4572 4573#ifdef CONFIG_SMP 4574void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4575{ 4576 if (p->sched_class && p->sched_class->set_cpus_allowed) 4577 p->sched_class->set_cpus_allowed(p, new_mask); 4578 4579 cpumask_copy(&p->cpus_allowed, new_mask); 4580 p->nr_cpus_allowed = cpumask_weight(new_mask); 4581} 4582 4583/* 4584 * This is how migration works: 4585 * 4586 * 1) we invoke migration_cpu_stop() on the target CPU using 4587 * stop_one_cpu(). 4588 * 2) stopper starts to run (implicitly forcing the migrated thread 4589 * off the CPU) 4590 * 3) it checks whether the migrated task is still in the wrong runqueue. 4591 * 4) if it's in the wrong runqueue then the migration thread removes 4592 * it and puts it into the right queue. 4593 * 5) stopper completes and stop_one_cpu() returns and the migration 4594 * is done. 4595 */ 4596 4597/* 4598 * Change a given task's CPU affinity. Migrate the thread to a 4599 * proper CPU and schedule it away if the CPU it's executing on 4600 * is removed from the allowed bitmask. 4601 * 4602 * NOTE: the caller must have a valid reference to the task, the 4603 * task must not exit() & deallocate itself prematurely. The 4604 * call is not atomic; no spinlocks may be held. 4605 */ 4606int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4607{ 4608 unsigned long flags; 4609 struct rq *rq; 4610 unsigned int dest_cpu; 4611 int ret = 0; 4612 4613 rq = task_rq_lock(p, &flags); 4614 4615 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4616 goto out; 4617 4618 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4619 ret = -EINVAL; 4620 goto out; 4621 } 4622 4623 do_set_cpus_allowed(p, new_mask); 4624 4625 /* Can the task run on the task's current CPU? If so, we're done */ 4626 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4627 goto out; 4628 4629 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4630 if (p->on_rq) { 4631 struct migration_arg arg = { p, dest_cpu }; 4632 /* Need help from migration thread: drop lock and wait. */ 4633 task_rq_unlock(rq, p, &flags); 4634 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4635 tlb_migrate_finish(p->mm); 4636 return 0; 4637 } 4638out: 4639 task_rq_unlock(rq, p, &flags); 4640 4641 return ret; 4642} 4643EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4644 4645/* 4646 * Move (not current) task off this cpu, onto dest cpu. We're doing 4647 * this because either it can't run here any more (set_cpus_allowed() 4648 * away from this CPU, or CPU going down), or because we're 4649 * attempting to rebalance this task on exec (sched_exec). 4650 * 4651 * So we race with normal scheduler movements, but that's OK, as long 4652 * as the task is no longer on this CPU. 4653 * 4654 * Returns non-zero if task was successfully migrated. 4655 */ 4656static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4657{ 4658 struct rq *rq_dest, *rq_src; 4659 int ret = 0; 4660 4661 if (unlikely(!cpu_active(dest_cpu))) 4662 return ret; 4663 4664 rq_src = cpu_rq(src_cpu); 4665 rq_dest = cpu_rq(dest_cpu); 4666 4667 raw_spin_lock(&p->pi_lock); 4668 double_rq_lock(rq_src, rq_dest); 4669 /* Already moved. */ 4670 if (task_cpu(p) != src_cpu) 4671 goto done; 4672 /* Affinity changed (again). */ 4673 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4674 goto fail; 4675 4676 /* 4677 * If we're not on a rq, the next wake-up will ensure we're 4678 * placed properly. 4679 */ 4680 if (p->on_rq) { 4681 dequeue_task(rq_src, p, 0); 4682 set_task_cpu(p, dest_cpu); 4683 enqueue_task(rq_dest, p, 0); 4684 check_preempt_curr(rq_dest, p, 0); 4685 } 4686done: 4687 ret = 1; 4688fail: 4689 double_rq_unlock(rq_src, rq_dest); 4690 raw_spin_unlock(&p->pi_lock); 4691 return ret; 4692} 4693 4694#ifdef CONFIG_NUMA_BALANCING 4695/* Migrate current task p to target_cpu */ 4696int migrate_task_to(struct task_struct *p, int target_cpu) 4697{ 4698 struct migration_arg arg = { p, target_cpu }; 4699 int curr_cpu = task_cpu(p); 4700 4701 if (curr_cpu == target_cpu) 4702 return 0; 4703 4704 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4705 return -EINVAL; 4706 4707 /* TODO: This is not properly updating schedstats */ 4708 4709 trace_sched_move_numa(p, curr_cpu, target_cpu); 4710 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4711} 4712 4713/* 4714 * Requeue a task on a given node and accurately track the number of NUMA 4715 * tasks on the runqueues 4716 */ 4717void sched_setnuma(struct task_struct *p, int nid) 4718{ 4719 struct rq *rq; 4720 unsigned long flags; 4721 bool on_rq, running; 4722 4723 rq = task_rq_lock(p, &flags); 4724 on_rq = p->on_rq; 4725 running = task_current(rq, p); 4726 4727 if (on_rq) 4728 dequeue_task(rq, p, 0); 4729 if (running) 4730 p->sched_class->put_prev_task(rq, p); 4731 4732 p->numa_preferred_nid = nid; 4733 4734 if (running) 4735 p->sched_class->set_curr_task(rq); 4736 if (on_rq) 4737 enqueue_task(rq, p, 0); 4738 task_rq_unlock(rq, p, &flags); 4739} 4740#endif 4741 4742/* 4743 * migration_cpu_stop - this will be executed by a highprio stopper thread 4744 * and performs thread migration by bumping thread off CPU then 4745 * 'pushing' onto another runqueue. 4746 */ 4747static int migration_cpu_stop(void *data) 4748{ 4749 struct migration_arg *arg = data; 4750 4751 /* 4752 * The original target cpu might have gone down and we might 4753 * be on another cpu but it doesn't matter. 4754 */ 4755 local_irq_disable(); 4756 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4757 local_irq_enable(); 4758 return 0; 4759} 4760 4761#ifdef CONFIG_HOTPLUG_CPU 4762 4763/* 4764 * Ensures that the idle task is using init_mm right before its cpu goes 4765 * offline. 4766 */ 4767void idle_task_exit(void) 4768{ 4769 struct mm_struct *mm = current->active_mm; 4770 4771 BUG_ON(cpu_online(smp_processor_id())); 4772 4773 if (mm != &init_mm) { 4774 switch_mm(mm, &init_mm, current); 4775 finish_arch_post_lock_switch(); 4776 } 4777 mmdrop(mm); 4778} 4779 4780/* 4781 * Since this CPU is going 'away' for a while, fold any nr_active delta 4782 * we might have. Assumes we're called after migrate_tasks() so that the 4783 * nr_active count is stable. 4784 * 4785 * Also see the comment "Global load-average calculations". 4786 */ 4787static void calc_load_migrate(struct rq *rq) 4788{ 4789 long delta = calc_load_fold_active(rq); 4790 if (delta) 4791 atomic_long_add(delta, &calc_load_tasks); 4792} 4793 4794static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4795{ 4796} 4797 4798static const struct sched_class fake_sched_class = { 4799 .put_prev_task = put_prev_task_fake, 4800}; 4801 4802static struct task_struct fake_task = { 4803 /* 4804 * Avoid pull_{rt,dl}_task() 4805 */ 4806 .prio = MAX_PRIO + 1, 4807 .sched_class = &fake_sched_class, 4808}; 4809 4810/* 4811 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4812 * try_to_wake_up()->select_task_rq(). 4813 * 4814 * Called with rq->lock held even though we'er in stop_machine() and 4815 * there's no concurrency possible, we hold the required locks anyway 4816 * because of lock validation efforts. 4817 */ 4818static void migrate_tasks(unsigned int dead_cpu) 4819{ 4820 struct rq *rq = cpu_rq(dead_cpu); 4821 struct task_struct *next, *stop = rq->stop; 4822 int dest_cpu; 4823 4824 /* 4825 * Fudge the rq selection such that the below task selection loop 4826 * doesn't get stuck on the currently eligible stop task. 4827 * 4828 * We're currently inside stop_machine() and the rq is either stuck 4829 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4830 * either way we should never end up calling schedule() until we're 4831 * done here. 4832 */ 4833 rq->stop = NULL; 4834 4835 /* 4836 * put_prev_task() and pick_next_task() sched 4837 * class method both need to have an up-to-date 4838 * value of rq->clock[_task] 4839 */ 4840 update_rq_clock(rq); 4841 4842 for ( ; ; ) { 4843 /* 4844 * There's this thread running, bail when that's the only 4845 * remaining thread. 4846 */ 4847 if (rq->nr_running == 1) 4848 break; 4849 4850 next = pick_next_task(rq, &fake_task); 4851 BUG_ON(!next); 4852 next->sched_class->put_prev_task(rq, next); 4853 4854 /* Find suitable destination for @next, with force if needed. */ 4855 dest_cpu = select_fallback_rq(dead_cpu, next); 4856 raw_spin_unlock(&rq->lock); 4857 4858 __migrate_task(next, dead_cpu, dest_cpu); 4859 4860 raw_spin_lock(&rq->lock); 4861 } 4862 4863 rq->stop = stop; 4864} 4865 4866#endif /* CONFIG_HOTPLUG_CPU */ 4867 4868#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4869 4870static struct ctl_table sd_ctl_dir[] = { 4871 { 4872 .procname = "sched_domain", 4873 .mode = 0555, 4874 }, 4875 {} 4876}; 4877 4878static struct ctl_table sd_ctl_root[] = { 4879 { 4880 .procname = "kernel", 4881 .mode = 0555, 4882 .child = sd_ctl_dir, 4883 }, 4884 {} 4885}; 4886 4887static struct ctl_table *sd_alloc_ctl_entry(int n) 4888{ 4889 struct ctl_table *entry = 4890 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4891 4892 return entry; 4893} 4894 4895static void sd_free_ctl_entry(struct ctl_table **tablep) 4896{ 4897 struct ctl_table *entry; 4898 4899 /* 4900 * In the intermediate directories, both the child directory and 4901 * procname are dynamically allocated and could fail but the mode 4902 * will always be set. In the lowest directory the names are 4903 * static strings and all have proc handlers. 4904 */ 4905 for (entry = *tablep; entry->mode; entry++) { 4906 if (entry->child) 4907 sd_free_ctl_entry(&entry->child); 4908 if (entry->proc_handler == NULL) 4909 kfree(entry->procname); 4910 } 4911 4912 kfree(*tablep); 4913 *tablep = NULL; 4914} 4915 4916static int min_load_idx = 0; 4917static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4918 4919static void 4920set_table_entry(struct ctl_table *entry, 4921 const char *procname, void *data, int maxlen, 4922 umode_t mode, proc_handler *proc_handler, 4923 bool load_idx) 4924{ 4925 entry->procname = procname; 4926 entry->data = data; 4927 entry->maxlen = maxlen; 4928 entry->mode = mode; 4929 entry->proc_handler = proc_handler; 4930 4931 if (load_idx) { 4932 entry->extra1 = &min_load_idx; 4933 entry->extra2 = &max_load_idx; 4934 } 4935} 4936 4937static struct ctl_table * 4938sd_alloc_ctl_domain_table(struct sched_domain *sd) 4939{ 4940 struct ctl_table *table = sd_alloc_ctl_entry(14); 4941 4942 if (table == NULL) 4943 return NULL; 4944 4945 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4946 sizeof(long), 0644, proc_doulongvec_minmax, false); 4947 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4948 sizeof(long), 0644, proc_doulongvec_minmax, false); 4949 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4950 sizeof(int), 0644, proc_dointvec_minmax, true); 4951 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4952 sizeof(int), 0644, proc_dointvec_minmax, true); 4953 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4954 sizeof(int), 0644, proc_dointvec_minmax, true); 4955 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4956 sizeof(int), 0644, proc_dointvec_minmax, true); 4957 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4958 sizeof(int), 0644, proc_dointvec_minmax, true); 4959 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4960 sizeof(int), 0644, proc_dointvec_minmax, false); 4961 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4962 sizeof(int), 0644, proc_dointvec_minmax, false); 4963 set_table_entry(&table[9], "cache_nice_tries", 4964 &sd->cache_nice_tries, 4965 sizeof(int), 0644, proc_dointvec_minmax, false); 4966 set_table_entry(&table[10], "flags", &sd->flags, 4967 sizeof(int), 0644, proc_dointvec_minmax, false); 4968 set_table_entry(&table[11], "max_newidle_lb_cost", 4969 &sd->max_newidle_lb_cost, 4970 sizeof(long), 0644, proc_doulongvec_minmax, false); 4971 set_table_entry(&table[12], "name", sd->name, 4972 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4973 /* &table[13] is terminator */ 4974 4975 return table; 4976} 4977 4978static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4979{ 4980 struct ctl_table *entry, *table; 4981 struct sched_domain *sd; 4982 int domain_num = 0, i; 4983 char buf[32]; 4984 4985 for_each_domain(cpu, sd) 4986 domain_num++; 4987 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4988 if (table == NULL) 4989 return NULL; 4990 4991 i = 0; 4992 for_each_domain(cpu, sd) { 4993 snprintf(buf, 32, "domain%d", i); 4994 entry->procname = kstrdup(buf, GFP_KERNEL); 4995 entry->mode = 0555; 4996 entry->child = sd_alloc_ctl_domain_table(sd); 4997 entry++; 4998 i++; 4999 } 5000 return table; 5001} 5002 5003static struct ctl_table_header *sd_sysctl_header; 5004static void register_sched_domain_sysctl(void) 5005{ 5006 int i, cpu_num = num_possible_cpus(); 5007 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5008 char buf[32]; 5009 5010 WARN_ON(sd_ctl_dir[0].child); 5011 sd_ctl_dir[0].child = entry; 5012 5013 if (entry == NULL) 5014 return; 5015 5016 for_each_possible_cpu(i) { 5017 snprintf(buf, 32, "cpu%d", i); 5018 entry->procname = kstrdup(buf, GFP_KERNEL); 5019 entry->mode = 0555; 5020 entry->child = sd_alloc_ctl_cpu_table(i); 5021 entry++; 5022 } 5023 5024 WARN_ON(sd_sysctl_header); 5025 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5026} 5027 5028/* may be called multiple times per register */ 5029static void unregister_sched_domain_sysctl(void) 5030{ 5031 if (sd_sysctl_header) 5032 unregister_sysctl_table(sd_sysctl_header); 5033 sd_sysctl_header = NULL; 5034 if (sd_ctl_dir[0].child) 5035 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5036} 5037#else 5038static void register_sched_domain_sysctl(void) 5039{ 5040} 5041static void unregister_sched_domain_sysctl(void) 5042{ 5043} 5044#endif 5045 5046static void set_rq_online(struct rq *rq) 5047{ 5048 if (!rq->online) { 5049 const struct sched_class *class; 5050 5051 cpumask_set_cpu(rq->cpu, rq->rd->online); 5052 rq->online = 1; 5053 5054 for_each_class(class) { 5055 if (class->rq_online) 5056 class->rq_online(rq); 5057 } 5058 } 5059} 5060 5061static void set_rq_offline(struct rq *rq) 5062{ 5063 if (rq->online) { 5064 const struct sched_class *class; 5065 5066 for_each_class(class) { 5067 if (class->rq_offline) 5068 class->rq_offline(rq); 5069 } 5070 5071 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5072 rq->online = 0; 5073 } 5074} 5075 5076/* 5077 * migration_call - callback that gets triggered when a CPU is added. 5078 * Here we can start up the necessary migration thread for the new CPU. 5079 */ 5080static int 5081migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5082{ 5083 int cpu = (long)hcpu; 5084 unsigned long flags; 5085 struct rq *rq = cpu_rq(cpu); 5086 5087 switch (action & ~CPU_TASKS_FROZEN) { 5088 5089 case CPU_UP_PREPARE: 5090 rq->calc_load_update = calc_load_update; 5091 break; 5092 5093 case CPU_ONLINE: 5094 /* Update our root-domain */ 5095 raw_spin_lock_irqsave(&rq->lock, flags); 5096 if (rq->rd) { 5097 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5098 5099 set_rq_online(rq); 5100 } 5101 raw_spin_unlock_irqrestore(&rq->lock, flags); 5102 break; 5103 5104#ifdef CONFIG_HOTPLUG_CPU 5105 case CPU_DYING: 5106 sched_ttwu_pending(); 5107 /* Update our root-domain */ 5108 raw_spin_lock_irqsave(&rq->lock, flags); 5109 if (rq->rd) { 5110 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5111 set_rq_offline(rq); 5112 } 5113 migrate_tasks(cpu); 5114 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5115 raw_spin_unlock_irqrestore(&rq->lock, flags); 5116 break; 5117 5118 case CPU_DEAD: 5119 calc_load_migrate(rq); 5120 break; 5121#endif 5122 } 5123 5124 update_max_interval(); 5125 5126 return NOTIFY_OK; 5127} 5128 5129/* 5130 * Register at high priority so that task migration (migrate_all_tasks) 5131 * happens before everything else. This has to be lower priority than 5132 * the notifier in the perf_event subsystem, though. 5133 */ 5134static struct notifier_block migration_notifier = { 5135 .notifier_call = migration_call, 5136 .priority = CPU_PRI_MIGRATION, 5137}; 5138 5139static void __cpuinit set_cpu_rq_start_time(void) 5140{ 5141 int cpu = smp_processor_id(); 5142 struct rq *rq = cpu_rq(cpu); 5143 rq->age_stamp = sched_clock_cpu(cpu); 5144} 5145 5146static int sched_cpu_active(struct notifier_block *nfb, 5147 unsigned long action, void *hcpu) 5148{ 5149 switch (action & ~CPU_TASKS_FROZEN) { 5150 case CPU_STARTING: 5151 set_cpu_rq_start_time(); 5152 return NOTIFY_OK; 5153 case CPU_DOWN_FAILED: 5154 set_cpu_active((long)hcpu, true); 5155 return NOTIFY_OK; 5156 default: 5157 return NOTIFY_DONE; 5158 } 5159} 5160 5161static int sched_cpu_inactive(struct notifier_block *nfb, 5162 unsigned long action, void *hcpu) 5163{ 5164 unsigned long flags; 5165 long cpu = (long)hcpu; 5166 5167 switch (action & ~CPU_TASKS_FROZEN) { 5168 case CPU_DOWN_PREPARE: 5169 set_cpu_active(cpu, false); 5170 5171 /* explicitly allow suspend */ 5172 if (!(action & CPU_TASKS_FROZEN)) { 5173 struct dl_bw *dl_b = dl_bw_of(cpu); 5174 bool overflow; 5175 int cpus; 5176 5177 raw_spin_lock_irqsave(&dl_b->lock, flags); 5178 cpus = dl_bw_cpus(cpu); 5179 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5180 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5181 5182 if (overflow) 5183 return notifier_from_errno(-EBUSY); 5184 } 5185 return NOTIFY_OK; 5186 } 5187 5188 return NOTIFY_DONE; 5189} 5190 5191static int __init migration_init(void) 5192{ 5193 void *cpu = (void *)(long)smp_processor_id(); 5194 int err; 5195 5196 /* Initialize migration for the boot CPU */ 5197 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5198 BUG_ON(err == NOTIFY_BAD); 5199 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5200 register_cpu_notifier(&migration_notifier); 5201 5202 /* Register cpu active notifiers */ 5203 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5204 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5205 5206 return 0; 5207} 5208early_initcall(migration_init); 5209#endif 5210 5211#ifdef CONFIG_SMP 5212 5213static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5214 5215#ifdef CONFIG_SCHED_DEBUG 5216 5217static __read_mostly int sched_debug_enabled; 5218 5219static int __init sched_debug_setup(char *str) 5220{ 5221 sched_debug_enabled = 1; 5222 5223 return 0; 5224} 5225early_param("sched_debug", sched_debug_setup); 5226 5227static inline bool sched_debug(void) 5228{ 5229 return sched_debug_enabled; 5230} 5231 5232static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5233 struct cpumask *groupmask) 5234{ 5235 struct sched_group *group = sd->groups; 5236 char str[256]; 5237 5238 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5239 cpumask_clear(groupmask); 5240 5241 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5242 5243 if (!(sd->flags & SD_LOAD_BALANCE)) { 5244 printk("does not load-balance\n"); 5245 if (sd->parent) 5246 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5247 " has parent"); 5248 return -1; 5249 } 5250 5251 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5252 5253 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5254 printk(KERN_ERR "ERROR: domain->span does not contain " 5255 "CPU%d\n", cpu); 5256 } 5257 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5258 printk(KERN_ERR "ERROR: domain->groups does not contain" 5259 " CPU%d\n", cpu); 5260 } 5261 5262 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5263 do { 5264 if (!group) { 5265 printk("\n"); 5266 printk(KERN_ERR "ERROR: group is NULL\n"); 5267 break; 5268 } 5269 5270 /* 5271 * Even though we initialize ->capacity to something semi-sane, 5272 * we leave capacity_orig unset. This allows us to detect if 5273 * domain iteration is still funny without causing /0 traps. 5274 */ 5275 if (!group->sgc->capacity_orig) { 5276 printk(KERN_CONT "\n"); 5277 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n"); 5278 break; 5279 } 5280 5281 if (!cpumask_weight(sched_group_cpus(group))) { 5282 printk(KERN_CONT "\n"); 5283 printk(KERN_ERR "ERROR: empty group\n"); 5284 break; 5285 } 5286 5287 if (!(sd->flags & SD_OVERLAP) && 5288 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5289 printk(KERN_CONT "\n"); 5290 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5291 break; 5292 } 5293 5294 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5295 5296 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5297 5298 printk(KERN_CONT " %s", str); 5299 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { 5300 printk(KERN_CONT " (cpu_capacity = %d)", 5301 group->sgc->capacity); 5302 } 5303 5304 group = group->next; 5305 } while (group != sd->groups); 5306 printk(KERN_CONT "\n"); 5307 5308 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5309 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5310 5311 if (sd->parent && 5312 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5313 printk(KERN_ERR "ERROR: parent span is not a superset " 5314 "of domain->span\n"); 5315 return 0; 5316} 5317 5318static void sched_domain_debug(struct sched_domain *sd, int cpu) 5319{ 5320 int level = 0; 5321 5322 if (!sched_debug_enabled) 5323 return; 5324 5325 if (!sd) { 5326 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5327 return; 5328 } 5329 5330 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5331 5332 for (;;) { 5333 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5334 break; 5335 level++; 5336 sd = sd->parent; 5337 if (!sd) 5338 break; 5339 } 5340} 5341#else /* !CONFIG_SCHED_DEBUG */ 5342# define sched_domain_debug(sd, cpu) do { } while (0) 5343static inline bool sched_debug(void) 5344{ 5345 return false; 5346} 5347#endif /* CONFIG_SCHED_DEBUG */ 5348 5349static int sd_degenerate(struct sched_domain *sd) 5350{ 5351 if (cpumask_weight(sched_domain_span(sd)) == 1) 5352 return 1; 5353 5354 /* Following flags need at least 2 groups */ 5355 if (sd->flags & (SD_LOAD_BALANCE | 5356 SD_BALANCE_NEWIDLE | 5357 SD_BALANCE_FORK | 5358 SD_BALANCE_EXEC | 5359 SD_SHARE_CPUCAPACITY | 5360 SD_SHARE_PKG_RESOURCES | 5361 SD_SHARE_POWERDOMAIN)) { 5362 if (sd->groups != sd->groups->next) 5363 return 0; 5364 } 5365 5366 /* Following flags don't use groups */ 5367 if (sd->flags & (SD_WAKE_AFFINE)) 5368 return 0; 5369 5370 return 1; 5371} 5372 5373static int 5374sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5375{ 5376 unsigned long cflags = sd->flags, pflags = parent->flags; 5377 5378 if (sd_degenerate(parent)) 5379 return 1; 5380 5381 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5382 return 0; 5383 5384 /* Flags needing groups don't count if only 1 group in parent */ 5385 if (parent->groups == parent->groups->next) { 5386 pflags &= ~(SD_LOAD_BALANCE | 5387 SD_BALANCE_NEWIDLE | 5388 SD_BALANCE_FORK | 5389 SD_BALANCE_EXEC | 5390 SD_SHARE_CPUCAPACITY | 5391 SD_SHARE_PKG_RESOURCES | 5392 SD_PREFER_SIBLING | 5393 SD_SHARE_POWERDOMAIN); 5394 if (nr_node_ids == 1) 5395 pflags &= ~SD_SERIALIZE; 5396 } 5397 if (~cflags & pflags) 5398 return 0; 5399 5400 return 1; 5401} 5402 5403static void free_rootdomain(struct rcu_head *rcu) 5404{ 5405 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5406 5407 cpupri_cleanup(&rd->cpupri); 5408 cpudl_cleanup(&rd->cpudl); 5409 free_cpumask_var(rd->dlo_mask); 5410 free_cpumask_var(rd->rto_mask); 5411 free_cpumask_var(rd->online); 5412 free_cpumask_var(rd->span); 5413 kfree(rd); 5414} 5415 5416static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5417{ 5418 struct root_domain *old_rd = NULL; 5419 unsigned long flags; 5420 5421 raw_spin_lock_irqsave(&rq->lock, flags); 5422 5423 if (rq->rd) { 5424 old_rd = rq->rd; 5425 5426 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5427 set_rq_offline(rq); 5428 5429 cpumask_clear_cpu(rq->cpu, old_rd->span); 5430 5431 /* 5432 * If we dont want to free the old_rd yet then 5433 * set old_rd to NULL to skip the freeing later 5434 * in this function: 5435 */ 5436 if (!atomic_dec_and_test(&old_rd->refcount)) 5437 old_rd = NULL; 5438 } 5439 5440 atomic_inc(&rd->refcount); 5441 rq->rd = rd; 5442 5443 cpumask_set_cpu(rq->cpu, rd->span); 5444 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5445 set_rq_online(rq); 5446 5447 raw_spin_unlock_irqrestore(&rq->lock, flags); 5448 5449 if (old_rd) 5450 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5451} 5452 5453static int init_rootdomain(struct root_domain *rd) 5454{ 5455 memset(rd, 0, sizeof(*rd)); 5456 5457 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5458 goto out; 5459 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5460 goto free_span; 5461 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5462 goto free_online; 5463 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5464 goto free_dlo_mask; 5465 5466 init_dl_bw(&rd->dl_bw); 5467 if (cpudl_init(&rd->cpudl) != 0) 5468 goto free_dlo_mask; 5469 5470 if (cpupri_init(&rd->cpupri) != 0) 5471 goto free_rto_mask; 5472 return 0; 5473 5474free_rto_mask: 5475 free_cpumask_var(rd->rto_mask); 5476free_dlo_mask: 5477 free_cpumask_var(rd->dlo_mask); 5478free_online: 5479 free_cpumask_var(rd->online); 5480free_span: 5481 free_cpumask_var(rd->span); 5482out: 5483 return -ENOMEM; 5484} 5485 5486/* 5487 * By default the system creates a single root-domain with all cpus as 5488 * members (mimicking the global state we have today). 5489 */ 5490struct root_domain def_root_domain; 5491 5492static void init_defrootdomain(void) 5493{ 5494 init_rootdomain(&def_root_domain); 5495 5496 atomic_set(&def_root_domain.refcount, 1); 5497} 5498 5499static struct root_domain *alloc_rootdomain(void) 5500{ 5501 struct root_domain *rd; 5502 5503 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5504 if (!rd) 5505 return NULL; 5506 5507 if (init_rootdomain(rd) != 0) { 5508 kfree(rd); 5509 return NULL; 5510 } 5511 5512 return rd; 5513} 5514 5515static void free_sched_groups(struct sched_group *sg, int free_sgc) 5516{ 5517 struct sched_group *tmp, *first; 5518 5519 if (!sg) 5520 return; 5521 5522 first = sg; 5523 do { 5524 tmp = sg->next; 5525 5526 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) 5527 kfree(sg->sgc); 5528 5529 kfree(sg); 5530 sg = tmp; 5531 } while (sg != first); 5532} 5533 5534static void free_sched_domain(struct rcu_head *rcu) 5535{ 5536 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5537 5538 /* 5539 * If its an overlapping domain it has private groups, iterate and 5540 * nuke them all. 5541 */ 5542 if (sd->flags & SD_OVERLAP) { 5543 free_sched_groups(sd->groups, 1); 5544 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5545 kfree(sd->groups->sgc); 5546 kfree(sd->groups); 5547 } 5548 kfree(sd); 5549} 5550 5551static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5552{ 5553 call_rcu(&sd->rcu, free_sched_domain); 5554} 5555 5556static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5557{ 5558 for (; sd; sd = sd->parent) 5559 destroy_sched_domain(sd, cpu); 5560} 5561 5562/* 5563 * Keep a special pointer to the highest sched_domain that has 5564 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5565 * allows us to avoid some pointer chasing select_idle_sibling(). 5566 * 5567 * Also keep a unique ID per domain (we use the first cpu number in 5568 * the cpumask of the domain), this allows us to quickly tell if 5569 * two cpus are in the same cache domain, see cpus_share_cache(). 5570 */ 5571DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5572DEFINE_PER_CPU(int, sd_llc_size); 5573DEFINE_PER_CPU(int, sd_llc_id); 5574DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5575DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5576DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5577 5578static void update_top_cache_domain(int cpu) 5579{ 5580 struct sched_domain *sd; 5581 struct sched_domain *busy_sd = NULL; 5582 int id = cpu; 5583 int size = 1; 5584 5585 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5586 if (sd) { 5587 id = cpumask_first(sched_domain_span(sd)); 5588 size = cpumask_weight(sched_domain_span(sd)); 5589 busy_sd = sd->parent; /* sd_busy */ 5590 } 5591 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5592 5593 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5594 per_cpu(sd_llc_size, cpu) = size; 5595 per_cpu(sd_llc_id, cpu) = id; 5596 5597 sd = lowest_flag_domain(cpu, SD_NUMA); 5598 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5599 5600 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5601 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5602} 5603 5604/* 5605 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5606 * hold the hotplug lock. 5607 */ 5608static void 5609cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5610{ 5611 struct rq *rq = cpu_rq(cpu); 5612 struct sched_domain *tmp; 5613 5614 /* Remove the sched domains which do not contribute to scheduling. */ 5615 for (tmp = sd; tmp; ) { 5616 struct sched_domain *parent = tmp->parent; 5617 if (!parent) 5618 break; 5619 5620 if (sd_parent_degenerate(tmp, parent)) { 5621 tmp->parent = parent->parent; 5622 if (parent->parent) 5623 parent->parent->child = tmp; 5624 /* 5625 * Transfer SD_PREFER_SIBLING down in case of a 5626 * degenerate parent; the spans match for this 5627 * so the property transfers. 5628 */ 5629 if (parent->flags & SD_PREFER_SIBLING) 5630 tmp->flags |= SD_PREFER_SIBLING; 5631 destroy_sched_domain(parent, cpu); 5632 } else 5633 tmp = tmp->parent; 5634 } 5635 5636 if (sd && sd_degenerate(sd)) { 5637 tmp = sd; 5638 sd = sd->parent; 5639 destroy_sched_domain(tmp, cpu); 5640 if (sd) 5641 sd->child = NULL; 5642 } 5643 5644 sched_domain_debug(sd, cpu); 5645 5646 rq_attach_root(rq, rd); 5647 tmp = rq->sd; 5648 rcu_assign_pointer(rq->sd, sd); 5649 destroy_sched_domains(tmp, cpu); 5650 5651 update_top_cache_domain(cpu); 5652} 5653 5654/* cpus with isolated domains */ 5655static cpumask_var_t cpu_isolated_map; 5656 5657/* Setup the mask of cpus configured for isolated domains */ 5658static int __init isolated_cpu_setup(char *str) 5659{ 5660 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5661 cpulist_parse(str, cpu_isolated_map); 5662 return 1; 5663} 5664 5665__setup("isolcpus=", isolated_cpu_setup); 5666 5667struct s_data { 5668 struct sched_domain ** __percpu sd; 5669 struct root_domain *rd; 5670}; 5671 5672enum s_alloc { 5673 sa_rootdomain, 5674 sa_sd, 5675 sa_sd_storage, 5676 sa_none, 5677}; 5678 5679/* 5680 * Build an iteration mask that can exclude certain CPUs from the upwards 5681 * domain traversal. 5682 * 5683 * Asymmetric node setups can result in situations where the domain tree is of 5684 * unequal depth, make sure to skip domains that already cover the entire 5685 * range. 5686 * 5687 * In that case build_sched_domains() will have terminated the iteration early 5688 * and our sibling sd spans will be empty. Domains should always include the 5689 * cpu they're built on, so check that. 5690 * 5691 */ 5692static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5693{ 5694 const struct cpumask *span = sched_domain_span(sd); 5695 struct sd_data *sdd = sd->private; 5696 struct sched_domain *sibling; 5697 int i; 5698 5699 for_each_cpu(i, span) { 5700 sibling = *per_cpu_ptr(sdd->sd, i); 5701 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5702 continue; 5703 5704 cpumask_set_cpu(i, sched_group_mask(sg)); 5705 } 5706} 5707 5708/* 5709 * Return the canonical balance cpu for this group, this is the first cpu 5710 * of this group that's also in the iteration mask. 5711 */ 5712int group_balance_cpu(struct sched_group *sg) 5713{ 5714 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5715} 5716 5717static int 5718build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5719{ 5720 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5721 const struct cpumask *span = sched_domain_span(sd); 5722 struct cpumask *covered = sched_domains_tmpmask; 5723 struct sd_data *sdd = sd->private; 5724 struct sched_domain *child; 5725 int i; 5726 5727 cpumask_clear(covered); 5728 5729 for_each_cpu(i, span) { 5730 struct cpumask *sg_span; 5731 5732 if (cpumask_test_cpu(i, covered)) 5733 continue; 5734 5735 child = *per_cpu_ptr(sdd->sd, i); 5736 5737 /* See the comment near build_group_mask(). */ 5738 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5739 continue; 5740 5741 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5742 GFP_KERNEL, cpu_to_node(cpu)); 5743 5744 if (!sg) 5745 goto fail; 5746 5747 sg_span = sched_group_cpus(sg); 5748 if (child->child) { 5749 child = child->child; 5750 cpumask_copy(sg_span, sched_domain_span(child)); 5751 } else 5752 cpumask_set_cpu(i, sg_span); 5753 5754 cpumask_or(covered, covered, sg_span); 5755 5756 sg->sgc = *per_cpu_ptr(sdd->sgc, i); 5757 if (atomic_inc_return(&sg->sgc->ref) == 1) 5758 build_group_mask(sd, sg); 5759 5760 /* 5761 * Initialize sgc->capacity such that even if we mess up the 5762 * domains and no possible iteration will get us here, we won't 5763 * die on a /0 trap. 5764 */ 5765 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); 5766 sg->sgc->capacity_orig = sg->sgc->capacity; 5767 5768 /* 5769 * Make sure the first group of this domain contains the 5770 * canonical balance cpu. Otherwise the sched_domain iteration 5771 * breaks. See update_sg_lb_stats(). 5772 */ 5773 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5774 group_balance_cpu(sg) == cpu) 5775 groups = sg; 5776 5777 if (!first) 5778 first = sg; 5779 if (last) 5780 last->next = sg; 5781 last = sg; 5782 last->next = first; 5783 } 5784 sd->groups = groups; 5785 5786 return 0; 5787 5788fail: 5789 free_sched_groups(first, 0); 5790 5791 return -ENOMEM; 5792} 5793 5794static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5795{ 5796 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5797 struct sched_domain *child = sd->child; 5798 5799 if (child) 5800 cpu = cpumask_first(sched_domain_span(child)); 5801 5802 if (sg) { 5803 *sg = *per_cpu_ptr(sdd->sg, cpu); 5804 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); 5805 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ 5806 } 5807 5808 return cpu; 5809} 5810 5811/* 5812 * build_sched_groups will build a circular linked list of the groups 5813 * covered by the given span, and will set each group's ->cpumask correctly, 5814 * and ->cpu_capacity to 0. 5815 * 5816 * Assumes the sched_domain tree is fully constructed 5817 */ 5818static int 5819build_sched_groups(struct sched_domain *sd, int cpu) 5820{ 5821 struct sched_group *first = NULL, *last = NULL; 5822 struct sd_data *sdd = sd->private; 5823 const struct cpumask *span = sched_domain_span(sd); 5824 struct cpumask *covered; 5825 int i; 5826 5827 get_group(cpu, sdd, &sd->groups); 5828 atomic_inc(&sd->groups->ref); 5829 5830 if (cpu != cpumask_first(span)) 5831 return 0; 5832 5833 lockdep_assert_held(&sched_domains_mutex); 5834 covered = sched_domains_tmpmask; 5835 5836 cpumask_clear(covered); 5837 5838 for_each_cpu(i, span) { 5839 struct sched_group *sg; 5840 int group, j; 5841 5842 if (cpumask_test_cpu(i, covered)) 5843 continue; 5844 5845 group = get_group(i, sdd, &sg); 5846 cpumask_setall(sched_group_mask(sg)); 5847 5848 for_each_cpu(j, span) { 5849 if (get_group(j, sdd, NULL) != group) 5850 continue; 5851 5852 cpumask_set_cpu(j, covered); 5853 cpumask_set_cpu(j, sched_group_cpus(sg)); 5854 } 5855 5856 if (!first) 5857 first = sg; 5858 if (last) 5859 last->next = sg; 5860 last = sg; 5861 } 5862 last->next = first; 5863 5864 return 0; 5865} 5866 5867/* 5868 * Initialize sched groups cpu_capacity. 5869 * 5870 * cpu_capacity indicates the capacity of sched group, which is used while 5871 * distributing the load between different sched groups in a sched domain. 5872 * Typically cpu_capacity for all the groups in a sched domain will be same 5873 * unless there are asymmetries in the topology. If there are asymmetries, 5874 * group having more cpu_capacity will pickup more load compared to the 5875 * group having less cpu_capacity. 5876 */ 5877static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) 5878{ 5879 struct sched_group *sg = sd->groups; 5880 5881 WARN_ON(!sg); 5882 5883 do { 5884 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5885 sg = sg->next; 5886 } while (sg != sd->groups); 5887 5888 if (cpu != group_balance_cpu(sg)) 5889 return; 5890 5891 update_group_capacity(sd, cpu); 5892 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); 5893} 5894 5895/* 5896 * Initializers for schedule domains 5897 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5898 */ 5899 5900static int default_relax_domain_level = -1; 5901int sched_domain_level_max; 5902 5903static int __init setup_relax_domain_level(char *str) 5904{ 5905 if (kstrtoint(str, 0, &default_relax_domain_level)) 5906 pr_warn("Unable to set relax_domain_level\n"); 5907 5908 return 1; 5909} 5910__setup("relax_domain_level=", setup_relax_domain_level); 5911 5912static void set_domain_attribute(struct sched_domain *sd, 5913 struct sched_domain_attr *attr) 5914{ 5915 int request; 5916 5917 if (!attr || attr->relax_domain_level < 0) { 5918 if (default_relax_domain_level < 0) 5919 return; 5920 else 5921 request = default_relax_domain_level; 5922 } else 5923 request = attr->relax_domain_level; 5924 if (request < sd->level) { 5925 /* turn off idle balance on this domain */ 5926 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5927 } else { 5928 /* turn on idle balance on this domain */ 5929 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5930 } 5931} 5932 5933static void __sdt_free(const struct cpumask *cpu_map); 5934static int __sdt_alloc(const struct cpumask *cpu_map); 5935 5936static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5937 const struct cpumask *cpu_map) 5938{ 5939 switch (what) { 5940 case sa_rootdomain: 5941 if (!atomic_read(&d->rd->refcount)) 5942 free_rootdomain(&d->rd->rcu); /* fall through */ 5943 case sa_sd: 5944 free_percpu(d->sd); /* fall through */ 5945 case sa_sd_storage: 5946 __sdt_free(cpu_map); /* fall through */ 5947 case sa_none: 5948 break; 5949 } 5950} 5951 5952static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5953 const struct cpumask *cpu_map) 5954{ 5955 memset(d, 0, sizeof(*d)); 5956 5957 if (__sdt_alloc(cpu_map)) 5958 return sa_sd_storage; 5959 d->sd = alloc_percpu(struct sched_domain *); 5960 if (!d->sd) 5961 return sa_sd_storage; 5962 d->rd = alloc_rootdomain(); 5963 if (!d->rd) 5964 return sa_sd; 5965 return sa_rootdomain; 5966} 5967 5968/* 5969 * NULL the sd_data elements we've used to build the sched_domain and 5970 * sched_group structure so that the subsequent __free_domain_allocs() 5971 * will not free the data we're using. 5972 */ 5973static void claim_allocations(int cpu, struct sched_domain *sd) 5974{ 5975 struct sd_data *sdd = sd->private; 5976 5977 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5978 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5979 5980 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5981 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5982 5983 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) 5984 *per_cpu_ptr(sdd->sgc, cpu) = NULL; 5985} 5986 5987#ifdef CONFIG_NUMA 5988static int sched_domains_numa_levels; 5989static int *sched_domains_numa_distance; 5990static struct cpumask ***sched_domains_numa_masks; 5991static int sched_domains_curr_level; 5992#endif 5993 5994/* 5995 * SD_flags allowed in topology descriptions. 5996 * 5997 * SD_SHARE_CPUCAPACITY - describes SMT topologies 5998 * SD_SHARE_PKG_RESOURCES - describes shared caches 5999 * SD_NUMA - describes NUMA topologies 6000 * SD_SHARE_POWERDOMAIN - describes shared power domain 6001 * 6002 * Odd one out: 6003 * SD_ASYM_PACKING - describes SMT quirks 6004 */ 6005#define TOPOLOGY_SD_FLAGS \ 6006 (SD_SHARE_CPUCAPACITY | \ 6007 SD_SHARE_PKG_RESOURCES | \ 6008 SD_NUMA | \ 6009 SD_ASYM_PACKING | \ 6010 SD_SHARE_POWERDOMAIN) 6011 6012static struct sched_domain * 6013sd_init(struct sched_domain_topology_level *tl, int cpu) 6014{ 6015 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6016 int sd_weight, sd_flags = 0; 6017 6018#ifdef CONFIG_NUMA 6019 /* 6020 * Ugly hack to pass state to sd_numa_mask()... 6021 */ 6022 sched_domains_curr_level = tl->numa_level; 6023#endif 6024 6025 sd_weight = cpumask_weight(tl->mask(cpu)); 6026 6027 if (tl->sd_flags) 6028 sd_flags = (*tl->sd_flags)(); 6029 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 6030 "wrong sd_flags in topology description\n")) 6031 sd_flags &= ~TOPOLOGY_SD_FLAGS; 6032 6033 *sd = (struct sched_domain){ 6034 .min_interval = sd_weight, 6035 .max_interval = 2*sd_weight, 6036 .busy_factor = 32, 6037 .imbalance_pct = 125, 6038 6039 .cache_nice_tries = 0, 6040 .busy_idx = 0, 6041 .idle_idx = 0, 6042 .newidle_idx = 0, 6043 .wake_idx = 0, 6044 .forkexec_idx = 0, 6045 6046 .flags = 1*SD_LOAD_BALANCE 6047 | 1*SD_BALANCE_NEWIDLE 6048 | 1*SD_BALANCE_EXEC 6049 | 1*SD_BALANCE_FORK 6050 | 0*SD_BALANCE_WAKE 6051 | 1*SD_WAKE_AFFINE 6052 | 0*SD_SHARE_CPUCAPACITY 6053 | 0*SD_SHARE_PKG_RESOURCES 6054 | 0*SD_SERIALIZE 6055 | 0*SD_PREFER_SIBLING 6056 | 0*SD_NUMA 6057 | sd_flags 6058 , 6059 6060 .last_balance = jiffies, 6061 .balance_interval = sd_weight, 6062 .smt_gain = 0, 6063 .max_newidle_lb_cost = 0, 6064 .next_decay_max_lb_cost = jiffies, 6065#ifdef CONFIG_SCHED_DEBUG 6066 .name = tl->name, 6067#endif 6068 }; 6069 6070 /* 6071 * Convert topological properties into behaviour. 6072 */ 6073 6074 if (sd->flags & SD_SHARE_CPUCAPACITY) { 6075 sd->imbalance_pct = 110; 6076 sd->smt_gain = 1178; /* ~15% */ 6077 6078 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6079 sd->imbalance_pct = 117; 6080 sd->cache_nice_tries = 1; 6081 sd->busy_idx = 2; 6082 6083#ifdef CONFIG_NUMA 6084 } else if (sd->flags & SD_NUMA) { 6085 sd->cache_nice_tries = 2; 6086 sd->busy_idx = 3; 6087 sd->idle_idx = 2; 6088 6089 sd->flags |= SD_SERIALIZE; 6090 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6091 sd->flags &= ~(SD_BALANCE_EXEC | 6092 SD_BALANCE_FORK | 6093 SD_WAKE_AFFINE); 6094 } 6095 6096#endif 6097 } else { 6098 sd->flags |= SD_PREFER_SIBLING; 6099 sd->cache_nice_tries = 1; 6100 sd->busy_idx = 2; 6101 sd->idle_idx = 1; 6102 } 6103 6104 sd->private = &tl->data; 6105 6106 return sd; 6107} 6108 6109/* 6110 * Topology list, bottom-up. 6111 */ 6112static struct sched_domain_topology_level default_topology[] = { 6113#ifdef CONFIG_SCHED_SMT 6114 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6115#endif 6116#ifdef CONFIG_SCHED_MC 6117 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6118#endif 6119 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6120 { NULL, }, 6121}; 6122 6123struct sched_domain_topology_level *sched_domain_topology = default_topology; 6124 6125#define for_each_sd_topology(tl) \ 6126 for (tl = sched_domain_topology; tl->mask; tl++) 6127 6128void set_sched_topology(struct sched_domain_topology_level *tl) 6129{ 6130 sched_domain_topology = tl; 6131} 6132 6133#ifdef CONFIG_NUMA 6134 6135static const struct cpumask *sd_numa_mask(int cpu) 6136{ 6137 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6138} 6139 6140static void sched_numa_warn(const char *str) 6141{ 6142 static int done = false; 6143 int i,j; 6144 6145 if (done) 6146 return; 6147 6148 done = true; 6149 6150 printk(KERN_WARNING "ERROR: %s\n\n", str); 6151 6152 for (i = 0; i < nr_node_ids; i++) { 6153 printk(KERN_WARNING " "); 6154 for (j = 0; j < nr_node_ids; j++) 6155 printk(KERN_CONT "%02d ", node_distance(i,j)); 6156 printk(KERN_CONT "\n"); 6157 } 6158 printk(KERN_WARNING "\n"); 6159} 6160 6161static bool find_numa_distance(int distance) 6162{ 6163 int i; 6164 6165 if (distance == node_distance(0, 0)) 6166 return true; 6167 6168 for (i = 0; i < sched_domains_numa_levels; i++) { 6169 if (sched_domains_numa_distance[i] == distance) 6170 return true; 6171 } 6172 6173 return false; 6174} 6175 6176static void sched_init_numa(void) 6177{ 6178 int next_distance, curr_distance = node_distance(0, 0); 6179 struct sched_domain_topology_level *tl; 6180 int level = 0; 6181 int i, j, k; 6182 6183 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6184 if (!sched_domains_numa_distance) 6185 return; 6186 6187 /* 6188 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6189 * unique distances in the node_distance() table. 6190 * 6191 * Assumes node_distance(0,j) includes all distances in 6192 * node_distance(i,j) in order to avoid cubic time. 6193 */ 6194 next_distance = curr_distance; 6195 for (i = 0; i < nr_node_ids; i++) { 6196 for (j = 0; j < nr_node_ids; j++) { 6197 for (k = 0; k < nr_node_ids; k++) { 6198 int distance = node_distance(i, k); 6199 6200 if (distance > curr_distance && 6201 (distance < next_distance || 6202 next_distance == curr_distance)) 6203 next_distance = distance; 6204 6205 /* 6206 * While not a strong assumption it would be nice to know 6207 * about cases where if node A is connected to B, B is not 6208 * equally connected to A. 6209 */ 6210 if (sched_debug() && node_distance(k, i) != distance) 6211 sched_numa_warn("Node-distance not symmetric"); 6212 6213 if (sched_debug() && i && !find_numa_distance(distance)) 6214 sched_numa_warn("Node-0 not representative"); 6215 } 6216 if (next_distance != curr_distance) { 6217 sched_domains_numa_distance[level++] = next_distance; 6218 sched_domains_numa_levels = level; 6219 curr_distance = next_distance; 6220 } else break; 6221 } 6222 6223 /* 6224 * In case of sched_debug() we verify the above assumption. 6225 */ 6226 if (!sched_debug()) 6227 break; 6228 } 6229 /* 6230 * 'level' contains the number of unique distances, excluding the 6231 * identity distance node_distance(i,i). 6232 * 6233 * The sched_domains_numa_distance[] array includes the actual distance 6234 * numbers. 6235 */ 6236 6237 /* 6238 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6239 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6240 * the array will contain less then 'level' members. This could be 6241 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6242 * in other functions. 6243 * 6244 * We reset it to 'level' at the end of this function. 6245 */ 6246 sched_domains_numa_levels = 0; 6247 6248 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6249 if (!sched_domains_numa_masks) 6250 return; 6251 6252 /* 6253 * Now for each level, construct a mask per node which contains all 6254 * cpus of nodes that are that many hops away from us. 6255 */ 6256 for (i = 0; i < level; i++) { 6257 sched_domains_numa_masks[i] = 6258 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6259 if (!sched_domains_numa_masks[i]) 6260 return; 6261 6262 for (j = 0; j < nr_node_ids; j++) { 6263 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6264 if (!mask) 6265 return; 6266 6267 sched_domains_numa_masks[i][j] = mask; 6268 6269 for (k = 0; k < nr_node_ids; k++) { 6270 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6271 continue; 6272 6273 cpumask_or(mask, mask, cpumask_of_node(k)); 6274 } 6275 } 6276 } 6277 6278 /* Compute default topology size */ 6279 for (i = 0; sched_domain_topology[i].mask; i++); 6280 6281 tl = kzalloc((i + level + 1) * 6282 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6283 if (!tl) 6284 return; 6285 6286 /* 6287 * Copy the default topology bits.. 6288 */ 6289 for (i = 0; sched_domain_topology[i].mask; i++) 6290 tl[i] = sched_domain_topology[i]; 6291 6292 /* 6293 * .. and append 'j' levels of NUMA goodness. 6294 */ 6295 for (j = 0; j < level; i++, j++) { 6296 tl[i] = (struct sched_domain_topology_level){ 6297 .mask = sd_numa_mask, 6298 .sd_flags = cpu_numa_flags, 6299 .flags = SDTL_OVERLAP, 6300 .numa_level = j, 6301 SD_INIT_NAME(NUMA) 6302 }; 6303 } 6304 6305 sched_domain_topology = tl; 6306 6307 sched_domains_numa_levels = level; 6308} 6309 6310static void sched_domains_numa_masks_set(int cpu) 6311{ 6312 int i, j; 6313 int node = cpu_to_node(cpu); 6314 6315 for (i = 0; i < sched_domains_numa_levels; i++) { 6316 for (j = 0; j < nr_node_ids; j++) { 6317 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6318 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6319 } 6320 } 6321} 6322 6323static void sched_domains_numa_masks_clear(int cpu) 6324{ 6325 int i, j; 6326 for (i = 0; i < sched_domains_numa_levels; i++) { 6327 for (j = 0; j < nr_node_ids; j++) 6328 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6329 } 6330} 6331 6332/* 6333 * Update sched_domains_numa_masks[level][node] array when new cpus 6334 * are onlined. 6335 */ 6336static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6337 unsigned long action, 6338 void *hcpu) 6339{ 6340 int cpu = (long)hcpu; 6341 6342 switch (action & ~CPU_TASKS_FROZEN) { 6343 case CPU_ONLINE: 6344 sched_domains_numa_masks_set(cpu); 6345 break; 6346 6347 case CPU_DEAD: 6348 sched_domains_numa_masks_clear(cpu); 6349 break; 6350 6351 default: 6352 return NOTIFY_DONE; 6353 } 6354 6355 return NOTIFY_OK; 6356} 6357#else 6358static inline void sched_init_numa(void) 6359{ 6360} 6361 6362static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6363 unsigned long action, 6364 void *hcpu) 6365{ 6366 return 0; 6367} 6368#endif /* CONFIG_NUMA */ 6369 6370static int __sdt_alloc(const struct cpumask *cpu_map) 6371{ 6372 struct sched_domain_topology_level *tl; 6373 int j; 6374 6375 for_each_sd_topology(tl) { 6376 struct sd_data *sdd = &tl->data; 6377 6378 sdd->sd = alloc_percpu(struct sched_domain *); 6379 if (!sdd->sd) 6380 return -ENOMEM; 6381 6382 sdd->sg = alloc_percpu(struct sched_group *); 6383 if (!sdd->sg) 6384 return -ENOMEM; 6385 6386 sdd->sgc = alloc_percpu(struct sched_group_capacity *); 6387 if (!sdd->sgc) 6388 return -ENOMEM; 6389 6390 for_each_cpu(j, cpu_map) { 6391 struct sched_domain *sd; 6392 struct sched_group *sg; 6393 struct sched_group_capacity *sgc; 6394 6395 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6396 GFP_KERNEL, cpu_to_node(j)); 6397 if (!sd) 6398 return -ENOMEM; 6399 6400 *per_cpu_ptr(sdd->sd, j) = sd; 6401 6402 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6403 GFP_KERNEL, cpu_to_node(j)); 6404 if (!sg) 6405 return -ENOMEM; 6406 6407 sg->next = sg; 6408 6409 *per_cpu_ptr(sdd->sg, j) = sg; 6410 6411 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), 6412 GFP_KERNEL, cpu_to_node(j)); 6413 if (!sgc) 6414 return -ENOMEM; 6415 6416 *per_cpu_ptr(sdd->sgc, j) = sgc; 6417 } 6418 } 6419 6420 return 0; 6421} 6422 6423static void __sdt_free(const struct cpumask *cpu_map) 6424{ 6425 struct sched_domain_topology_level *tl; 6426 int j; 6427 6428 for_each_sd_topology(tl) { 6429 struct sd_data *sdd = &tl->data; 6430 6431 for_each_cpu(j, cpu_map) { 6432 struct sched_domain *sd; 6433 6434 if (sdd->sd) { 6435 sd = *per_cpu_ptr(sdd->sd, j); 6436 if (sd && (sd->flags & SD_OVERLAP)) 6437 free_sched_groups(sd->groups, 0); 6438 kfree(*per_cpu_ptr(sdd->sd, j)); 6439 } 6440 6441 if (sdd->sg) 6442 kfree(*per_cpu_ptr(sdd->sg, j)); 6443 if (sdd->sgc) 6444 kfree(*per_cpu_ptr(sdd->sgc, j)); 6445 } 6446 free_percpu(sdd->sd); 6447 sdd->sd = NULL; 6448 free_percpu(sdd->sg); 6449 sdd->sg = NULL; 6450 free_percpu(sdd->sgc); 6451 sdd->sgc = NULL; 6452 } 6453} 6454 6455struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6456 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6457 struct sched_domain *child, int cpu) 6458{ 6459 struct sched_domain *sd = sd_init(tl, cpu); 6460 if (!sd) 6461 return child; 6462 6463 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6464 if (child) { 6465 sd->level = child->level + 1; 6466 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6467 child->parent = sd; 6468 sd->child = child; 6469 } 6470 set_domain_attribute(sd, attr); 6471 6472 return sd; 6473} 6474 6475/* 6476 * Build sched domains for a given set of cpus and attach the sched domains 6477 * to the individual cpus 6478 */ 6479static int build_sched_domains(const struct cpumask *cpu_map, 6480 struct sched_domain_attr *attr) 6481{ 6482 enum s_alloc alloc_state; 6483 struct sched_domain *sd; 6484 struct s_data d; 6485 int i, ret = -ENOMEM; 6486 6487 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6488 if (alloc_state != sa_rootdomain) 6489 goto error; 6490 6491 /* Set up domains for cpus specified by the cpu_map. */ 6492 for_each_cpu(i, cpu_map) { 6493 struct sched_domain_topology_level *tl; 6494 6495 sd = NULL; 6496 for_each_sd_topology(tl) { 6497 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6498 if (tl == sched_domain_topology) 6499 *per_cpu_ptr(d.sd, i) = sd; 6500 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6501 sd->flags |= SD_OVERLAP; 6502 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6503 break; 6504 } 6505 } 6506 6507 /* Build the groups for the domains */ 6508 for_each_cpu(i, cpu_map) { 6509 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6510 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6511 if (sd->flags & SD_OVERLAP) { 6512 if (build_overlap_sched_groups(sd, i)) 6513 goto error; 6514 } else { 6515 if (build_sched_groups(sd, i)) 6516 goto error; 6517 } 6518 } 6519 } 6520 6521 /* Calculate CPU capacity for physical packages and nodes */ 6522 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6523 if (!cpumask_test_cpu(i, cpu_map)) 6524 continue; 6525 6526 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6527 claim_allocations(i, sd); 6528 init_sched_groups_capacity(i, sd); 6529 } 6530 } 6531 6532 /* Attach the domains */ 6533 rcu_read_lock(); 6534 for_each_cpu(i, cpu_map) { 6535 sd = *per_cpu_ptr(d.sd, i); 6536 cpu_attach_domain(sd, d.rd, i); 6537 } 6538 rcu_read_unlock(); 6539 6540 ret = 0; 6541error: 6542 __free_domain_allocs(&d, alloc_state, cpu_map); 6543 return ret; 6544} 6545 6546static cpumask_var_t *doms_cur; /* current sched domains */ 6547static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6548static struct sched_domain_attr *dattr_cur; 6549 /* attribues of custom domains in 'doms_cur' */ 6550 6551/* 6552 * Special case: If a kmalloc of a doms_cur partition (array of 6553 * cpumask) fails, then fallback to a single sched domain, 6554 * as determined by the single cpumask fallback_doms. 6555 */ 6556static cpumask_var_t fallback_doms; 6557 6558/* 6559 * arch_update_cpu_topology lets virtualized architectures update the 6560 * cpu core maps. It is supposed to return 1 if the topology changed 6561 * or 0 if it stayed the same. 6562 */ 6563int __weak arch_update_cpu_topology(void) 6564{ 6565 return 0; 6566} 6567 6568cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6569{ 6570 int i; 6571 cpumask_var_t *doms; 6572 6573 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6574 if (!doms) 6575 return NULL; 6576 for (i = 0; i < ndoms; i++) { 6577 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6578 free_sched_domains(doms, i); 6579 return NULL; 6580 } 6581 } 6582 return doms; 6583} 6584 6585void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6586{ 6587 unsigned int i; 6588 for (i = 0; i < ndoms; i++) 6589 free_cpumask_var(doms[i]); 6590 kfree(doms); 6591} 6592 6593/* 6594 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6595 * For now this just excludes isolated cpus, but could be used to 6596 * exclude other special cases in the future. 6597 */ 6598static int init_sched_domains(const struct cpumask *cpu_map) 6599{ 6600 int err; 6601 6602 arch_update_cpu_topology(); 6603 ndoms_cur = 1; 6604 doms_cur = alloc_sched_domains(ndoms_cur); 6605 if (!doms_cur) 6606 doms_cur = &fallback_doms; 6607 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6608 err = build_sched_domains(doms_cur[0], NULL); 6609 register_sched_domain_sysctl(); 6610 6611 return err; 6612} 6613 6614/* 6615 * Detach sched domains from a group of cpus specified in cpu_map 6616 * These cpus will now be attached to the NULL domain 6617 */ 6618static void detach_destroy_domains(const struct cpumask *cpu_map) 6619{ 6620 int i; 6621 6622 rcu_read_lock(); 6623 for_each_cpu(i, cpu_map) 6624 cpu_attach_domain(NULL, &def_root_domain, i); 6625 rcu_read_unlock(); 6626} 6627 6628/* handle null as "default" */ 6629static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6630 struct sched_domain_attr *new, int idx_new) 6631{ 6632 struct sched_domain_attr tmp; 6633 6634 /* fast path */ 6635 if (!new && !cur) 6636 return 1; 6637 6638 tmp = SD_ATTR_INIT; 6639 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6640 new ? (new + idx_new) : &tmp, 6641 sizeof(struct sched_domain_attr)); 6642} 6643 6644/* 6645 * Partition sched domains as specified by the 'ndoms_new' 6646 * cpumasks in the array doms_new[] of cpumasks. This compares 6647 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6648 * It destroys each deleted domain and builds each new domain. 6649 * 6650 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6651 * The masks don't intersect (don't overlap.) We should setup one 6652 * sched domain for each mask. CPUs not in any of the cpumasks will 6653 * not be load balanced. If the same cpumask appears both in the 6654 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6655 * it as it is. 6656 * 6657 * The passed in 'doms_new' should be allocated using 6658 * alloc_sched_domains. This routine takes ownership of it and will 6659 * free_sched_domains it when done with it. If the caller failed the 6660 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6661 * and partition_sched_domains() will fallback to the single partition 6662 * 'fallback_doms', it also forces the domains to be rebuilt. 6663 * 6664 * If doms_new == NULL it will be replaced with cpu_online_mask. 6665 * ndoms_new == 0 is a special case for destroying existing domains, 6666 * and it will not create the default domain. 6667 * 6668 * Call with hotplug lock held 6669 */ 6670void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6671 struct sched_domain_attr *dattr_new) 6672{ 6673 int i, j, n; 6674 int new_topology; 6675 6676 mutex_lock(&sched_domains_mutex); 6677 6678 /* always unregister in case we don't destroy any domains */ 6679 unregister_sched_domain_sysctl(); 6680 6681 /* Let architecture update cpu core mappings. */ 6682 new_topology = arch_update_cpu_topology(); 6683 6684 n = doms_new ? ndoms_new : 0; 6685 6686 /* Destroy deleted domains */ 6687 for (i = 0; i < ndoms_cur; i++) { 6688 for (j = 0; j < n && !new_topology; j++) { 6689 if (cpumask_equal(doms_cur[i], doms_new[j]) 6690 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6691 goto match1; 6692 } 6693 /* no match - a current sched domain not in new doms_new[] */ 6694 detach_destroy_domains(doms_cur[i]); 6695match1: 6696 ; 6697 } 6698 6699 n = ndoms_cur; 6700 if (doms_new == NULL) { 6701 n = 0; 6702 doms_new = &fallback_doms; 6703 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6704 WARN_ON_ONCE(dattr_new); 6705 } 6706 6707 /* Build new domains */ 6708 for (i = 0; i < ndoms_new; i++) { 6709 for (j = 0; j < n && !new_topology; j++) { 6710 if (cpumask_equal(doms_new[i], doms_cur[j]) 6711 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6712 goto match2; 6713 } 6714 /* no match - add a new doms_new */ 6715 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6716match2: 6717 ; 6718 } 6719 6720 /* Remember the new sched domains */ 6721 if (doms_cur != &fallback_doms) 6722 free_sched_domains(doms_cur, ndoms_cur); 6723 kfree(dattr_cur); /* kfree(NULL) is safe */ 6724 doms_cur = doms_new; 6725 dattr_cur = dattr_new; 6726 ndoms_cur = ndoms_new; 6727 6728 register_sched_domain_sysctl(); 6729 6730 mutex_unlock(&sched_domains_mutex); 6731} 6732 6733static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6734 6735/* 6736 * Update cpusets according to cpu_active mask. If cpusets are 6737 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6738 * around partition_sched_domains(). 6739 * 6740 * If we come here as part of a suspend/resume, don't touch cpusets because we 6741 * want to restore it back to its original state upon resume anyway. 6742 */ 6743static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6744 void *hcpu) 6745{ 6746 switch (action) { 6747 case CPU_ONLINE_FROZEN: 6748 case CPU_DOWN_FAILED_FROZEN: 6749 6750 /* 6751 * num_cpus_frozen tracks how many CPUs are involved in suspend 6752 * resume sequence. As long as this is not the last online 6753 * operation in the resume sequence, just build a single sched 6754 * domain, ignoring cpusets. 6755 */ 6756 num_cpus_frozen--; 6757 if (likely(num_cpus_frozen)) { 6758 partition_sched_domains(1, NULL, NULL); 6759 break; 6760 } 6761 6762 /* 6763 * This is the last CPU online operation. So fall through and 6764 * restore the original sched domains by considering the 6765 * cpuset configurations. 6766 */ 6767 6768 case CPU_ONLINE: 6769 case CPU_DOWN_FAILED: 6770 cpuset_update_active_cpus(true); 6771 break; 6772 default: 6773 return NOTIFY_DONE; 6774 } 6775 return NOTIFY_OK; 6776} 6777 6778static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6779 void *hcpu) 6780{ 6781 switch (action) { 6782 case CPU_DOWN_PREPARE: 6783 cpuset_update_active_cpus(false); 6784 break; 6785 case CPU_DOWN_PREPARE_FROZEN: 6786 num_cpus_frozen++; 6787 partition_sched_domains(1, NULL, NULL); 6788 break; 6789 default: 6790 return NOTIFY_DONE; 6791 } 6792 return NOTIFY_OK; 6793} 6794 6795void __init sched_init_smp(void) 6796{ 6797 cpumask_var_t non_isolated_cpus; 6798 6799 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6800 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6801 6802 sched_init_numa(); 6803 6804 /* 6805 * There's no userspace yet to cause hotplug operations; hence all the 6806 * cpu masks are stable and all blatant races in the below code cannot 6807 * happen. 6808 */ 6809 mutex_lock(&sched_domains_mutex); 6810 init_sched_domains(cpu_active_mask); 6811 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6812 if (cpumask_empty(non_isolated_cpus)) 6813 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6814 mutex_unlock(&sched_domains_mutex); 6815 6816 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6817 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6818 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6819 6820 init_hrtick(); 6821 6822 /* Move init over to a non-isolated CPU */ 6823 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6824 BUG(); 6825 sched_init_granularity(); 6826 free_cpumask_var(non_isolated_cpus); 6827 6828 init_sched_rt_class(); 6829 init_sched_dl_class(); 6830} 6831#else 6832void __init sched_init_smp(void) 6833{ 6834 sched_init_granularity(); 6835} 6836#endif /* CONFIG_SMP */ 6837 6838const_debug unsigned int sysctl_timer_migration = 1; 6839 6840int in_sched_functions(unsigned long addr) 6841{ 6842 return in_lock_functions(addr) || 6843 (addr >= (unsigned long)__sched_text_start 6844 && addr < (unsigned long)__sched_text_end); 6845} 6846 6847#ifdef CONFIG_CGROUP_SCHED 6848/* 6849 * Default task group. 6850 * Every task in system belongs to this group at bootup. 6851 */ 6852struct task_group root_task_group; 6853LIST_HEAD(task_groups); 6854#endif 6855 6856DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6857 6858void __init sched_init(void) 6859{ 6860 int i, j; 6861 unsigned long alloc_size = 0, ptr; 6862 6863#ifdef CONFIG_FAIR_GROUP_SCHED 6864 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6865#endif 6866#ifdef CONFIG_RT_GROUP_SCHED 6867 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6868#endif 6869#ifdef CONFIG_CPUMASK_OFFSTACK 6870 alloc_size += num_possible_cpus() * cpumask_size(); 6871#endif 6872 if (alloc_size) { 6873 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6874 6875#ifdef CONFIG_FAIR_GROUP_SCHED 6876 root_task_group.se = (struct sched_entity **)ptr; 6877 ptr += nr_cpu_ids * sizeof(void **); 6878 6879 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6880 ptr += nr_cpu_ids * sizeof(void **); 6881 6882#endif /* CONFIG_FAIR_GROUP_SCHED */ 6883#ifdef CONFIG_RT_GROUP_SCHED 6884 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6885 ptr += nr_cpu_ids * sizeof(void **); 6886 6887 root_task_group.rt_rq = (struct rt_rq **)ptr; 6888 ptr += nr_cpu_ids * sizeof(void **); 6889 6890#endif /* CONFIG_RT_GROUP_SCHED */ 6891#ifdef CONFIG_CPUMASK_OFFSTACK 6892 for_each_possible_cpu(i) { 6893 per_cpu(load_balance_mask, i) = (void *)ptr; 6894 ptr += cpumask_size(); 6895 } 6896#endif /* CONFIG_CPUMASK_OFFSTACK */ 6897 } 6898 6899 init_rt_bandwidth(&def_rt_bandwidth, 6900 global_rt_period(), global_rt_runtime()); 6901 init_dl_bandwidth(&def_dl_bandwidth, 6902 global_rt_period(), global_rt_runtime()); 6903 6904#ifdef CONFIG_SMP 6905 init_defrootdomain(); 6906#endif 6907 6908#ifdef CONFIG_RT_GROUP_SCHED 6909 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6910 global_rt_period(), global_rt_runtime()); 6911#endif /* CONFIG_RT_GROUP_SCHED */ 6912 6913#ifdef CONFIG_CGROUP_SCHED 6914 list_add(&root_task_group.list, &task_groups); 6915 INIT_LIST_HEAD(&root_task_group.children); 6916 INIT_LIST_HEAD(&root_task_group.siblings); 6917 autogroup_init(&init_task); 6918 6919#endif /* CONFIG_CGROUP_SCHED */ 6920 6921 for_each_possible_cpu(i) { 6922 struct rq *rq; 6923 6924 rq = cpu_rq(i); 6925 raw_spin_lock_init(&rq->lock); 6926 rq->nr_running = 0; 6927 rq->calc_load_active = 0; 6928 rq->calc_load_update = jiffies + LOAD_FREQ; 6929 init_cfs_rq(&rq->cfs); 6930 init_rt_rq(&rq->rt, rq); 6931 init_dl_rq(&rq->dl, rq); 6932#ifdef CONFIG_FAIR_GROUP_SCHED 6933 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6934 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6935 /* 6936 * How much cpu bandwidth does root_task_group get? 6937 * 6938 * In case of task-groups formed thr' the cgroup filesystem, it 6939 * gets 100% of the cpu resources in the system. This overall 6940 * system cpu resource is divided among the tasks of 6941 * root_task_group and its child task-groups in a fair manner, 6942 * based on each entity's (task or task-group's) weight 6943 * (se->load.weight). 6944 * 6945 * In other words, if root_task_group has 10 tasks of weight 6946 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6947 * then A0's share of the cpu resource is: 6948 * 6949 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6950 * 6951 * We achieve this by letting root_task_group's tasks sit 6952 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6953 */ 6954 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6955 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6956#endif /* CONFIG_FAIR_GROUP_SCHED */ 6957 6958 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6959#ifdef CONFIG_RT_GROUP_SCHED 6960 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6961#endif 6962 6963 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6964 rq->cpu_load[j] = 0; 6965 6966 rq->last_load_update_tick = jiffies; 6967 6968#ifdef CONFIG_SMP 6969 rq->sd = NULL; 6970 rq->rd = NULL; 6971 rq->cpu_capacity = SCHED_CAPACITY_SCALE; 6972 rq->post_schedule = 0; 6973 rq->active_balance = 0; 6974 rq->next_balance = jiffies; 6975 rq->push_cpu = 0; 6976 rq->cpu = i; 6977 rq->online = 0; 6978 rq->idle_stamp = 0; 6979 rq->avg_idle = 2*sysctl_sched_migration_cost; 6980 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6981 6982 INIT_LIST_HEAD(&rq->cfs_tasks); 6983 6984 rq_attach_root(rq, &def_root_domain); 6985#ifdef CONFIG_NO_HZ_COMMON 6986 rq->nohz_flags = 0; 6987#endif 6988#ifdef CONFIG_NO_HZ_FULL 6989 rq->last_sched_tick = 0; 6990#endif 6991#endif 6992 init_rq_hrtick(rq); 6993 atomic_set(&rq->nr_iowait, 0); 6994 } 6995 6996 set_load_weight(&init_task); 6997 6998#ifdef CONFIG_PREEMPT_NOTIFIERS 6999 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7000#endif 7001 7002 /* 7003 * The boot idle thread does lazy MMU switching as well: 7004 */ 7005 atomic_inc(&init_mm.mm_count); 7006 enter_lazy_tlb(&init_mm, current); 7007 7008 /* 7009 * Make us the idle thread. Technically, schedule() should not be 7010 * called from this thread, however somewhere below it might be, 7011 * but because we are the idle thread, we just pick up running again 7012 * when this runqueue becomes "idle". 7013 */ 7014 init_idle(current, smp_processor_id()); 7015 7016 calc_load_update = jiffies + LOAD_FREQ; 7017 7018 /* 7019 * During early bootup we pretend to be a normal task: 7020 */ 7021 current->sched_class = &fair_sched_class; 7022 7023#ifdef CONFIG_SMP 7024 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7025 /* May be allocated at isolcpus cmdline parse time */ 7026 if (cpu_isolated_map == NULL) 7027 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7028 idle_thread_set_boot_cpu(); 7029 set_cpu_rq_start_time(); 7030#endif 7031 init_sched_fair_class(); 7032 7033 scheduler_running = 1; 7034} 7035 7036#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7037static inline int preempt_count_equals(int preempt_offset) 7038{ 7039 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7040 7041 return (nested == preempt_offset); 7042} 7043 7044void __might_sleep(const char *file, int line, int preempt_offset) 7045{ 7046 static unsigned long prev_jiffy; /* ratelimiting */ 7047 7048 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7049 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 7050 !is_idle_task(current)) || 7051 system_state != SYSTEM_RUNNING || oops_in_progress) 7052 return; 7053 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7054 return; 7055 prev_jiffy = jiffies; 7056 7057 printk(KERN_ERR 7058 "BUG: sleeping function called from invalid context at %s:%d\n", 7059 file, line); 7060 printk(KERN_ERR 7061 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7062 in_atomic(), irqs_disabled(), 7063 current->pid, current->comm); 7064 7065 debug_show_held_locks(current); 7066 if (irqs_disabled()) 7067 print_irqtrace_events(current); 7068#ifdef CONFIG_DEBUG_PREEMPT 7069 if (!preempt_count_equals(preempt_offset)) { 7070 pr_err("Preemption disabled at:"); 7071 print_ip_sym(current->preempt_disable_ip); 7072 pr_cont("\n"); 7073 } 7074#endif 7075 dump_stack(); 7076} 7077EXPORT_SYMBOL(__might_sleep); 7078#endif 7079 7080#ifdef CONFIG_MAGIC_SYSRQ 7081static void normalize_task(struct rq *rq, struct task_struct *p) 7082{ 7083 const struct sched_class *prev_class = p->sched_class; 7084 struct sched_attr attr = { 7085 .sched_policy = SCHED_NORMAL, 7086 }; 7087 int old_prio = p->prio; 7088 int on_rq; 7089 7090 on_rq = p->on_rq; 7091 if (on_rq) 7092 dequeue_task(rq, p, 0); 7093 __setscheduler(rq, p, &attr); 7094 if (on_rq) { 7095 enqueue_task(rq, p, 0); 7096 resched_task(rq->curr); 7097 } 7098 7099 check_class_changed(rq, p, prev_class, old_prio); 7100} 7101 7102void normalize_rt_tasks(void) 7103{ 7104 struct task_struct *g, *p; 7105 unsigned long flags; 7106 struct rq *rq; 7107 7108 read_lock_irqsave(&tasklist_lock, flags); 7109 do_each_thread(g, p) { 7110 /* 7111 * Only normalize user tasks: 7112 */ 7113 if (!p->mm) 7114 continue; 7115 7116 p->se.exec_start = 0; 7117#ifdef CONFIG_SCHEDSTATS 7118 p->se.statistics.wait_start = 0; 7119 p->se.statistics.sleep_start = 0; 7120 p->se.statistics.block_start = 0; 7121#endif 7122 7123 if (!dl_task(p) && !rt_task(p)) { 7124 /* 7125 * Renice negative nice level userspace 7126 * tasks back to 0: 7127 */ 7128 if (task_nice(p) < 0 && p->mm) 7129 set_user_nice(p, 0); 7130 continue; 7131 } 7132 7133 raw_spin_lock(&p->pi_lock); 7134 rq = __task_rq_lock(p); 7135 7136 normalize_task(rq, p); 7137 7138 __task_rq_unlock(rq); 7139 raw_spin_unlock(&p->pi_lock); 7140 } while_each_thread(g, p); 7141 7142 read_unlock_irqrestore(&tasklist_lock, flags); 7143} 7144 7145#endif /* CONFIG_MAGIC_SYSRQ */ 7146 7147#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7148/* 7149 * These functions are only useful for the IA64 MCA handling, or kdb. 7150 * 7151 * They can only be called when the whole system has been 7152 * stopped - every CPU needs to be quiescent, and no scheduling 7153 * activity can take place. Using them for anything else would 7154 * be a serious bug, and as a result, they aren't even visible 7155 * under any other configuration. 7156 */ 7157 7158/** 7159 * curr_task - return the current task for a given cpu. 7160 * @cpu: the processor in question. 7161 * 7162 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7163 * 7164 * Return: The current task for @cpu. 7165 */ 7166struct task_struct *curr_task(int cpu) 7167{ 7168 return cpu_curr(cpu); 7169} 7170 7171#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7172 7173#ifdef CONFIG_IA64 7174/** 7175 * set_curr_task - set the current task for a given cpu. 7176 * @cpu: the processor in question. 7177 * @p: the task pointer to set. 7178 * 7179 * Description: This function must only be used when non-maskable interrupts 7180 * are serviced on a separate stack. It allows the architecture to switch the 7181 * notion of the current task on a cpu in a non-blocking manner. This function 7182 * must be called with all CPU's synchronized, and interrupts disabled, the 7183 * and caller must save the original value of the current task (see 7184 * curr_task() above) and restore that value before reenabling interrupts and 7185 * re-starting the system. 7186 * 7187 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7188 */ 7189void set_curr_task(int cpu, struct task_struct *p) 7190{ 7191 cpu_curr(cpu) = p; 7192} 7193 7194#endif 7195 7196#ifdef CONFIG_CGROUP_SCHED 7197/* task_group_lock serializes the addition/removal of task groups */ 7198static DEFINE_SPINLOCK(task_group_lock); 7199 7200static void free_sched_group(struct task_group *tg) 7201{ 7202 free_fair_sched_group(tg); 7203 free_rt_sched_group(tg); 7204 autogroup_free(tg); 7205 kfree(tg); 7206} 7207 7208/* allocate runqueue etc for a new task group */ 7209struct task_group *sched_create_group(struct task_group *parent) 7210{ 7211 struct task_group *tg; 7212 7213 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7214 if (!tg) 7215 return ERR_PTR(-ENOMEM); 7216 7217 if (!alloc_fair_sched_group(tg, parent)) 7218 goto err; 7219 7220 if (!alloc_rt_sched_group(tg, parent)) 7221 goto err; 7222 7223 return tg; 7224 7225err: 7226 free_sched_group(tg); 7227 return ERR_PTR(-ENOMEM); 7228} 7229 7230void sched_online_group(struct task_group *tg, struct task_group *parent) 7231{ 7232 unsigned long flags; 7233 7234 spin_lock_irqsave(&task_group_lock, flags); 7235 list_add_rcu(&tg->list, &task_groups); 7236 7237 WARN_ON(!parent); /* root should already exist */ 7238 7239 tg->parent = parent; 7240 INIT_LIST_HEAD(&tg->children); 7241 list_add_rcu(&tg->siblings, &parent->children); 7242 spin_unlock_irqrestore(&task_group_lock, flags); 7243} 7244 7245/* rcu callback to free various structures associated with a task group */ 7246static void free_sched_group_rcu(struct rcu_head *rhp) 7247{ 7248 /* now it should be safe to free those cfs_rqs */ 7249 free_sched_group(container_of(rhp, struct task_group, rcu)); 7250} 7251 7252/* Destroy runqueue etc associated with a task group */ 7253void sched_destroy_group(struct task_group *tg) 7254{ 7255 /* wait for possible concurrent references to cfs_rqs complete */ 7256 call_rcu(&tg->rcu, free_sched_group_rcu); 7257} 7258 7259void sched_offline_group(struct task_group *tg) 7260{ 7261 unsigned long flags; 7262 int i; 7263 7264 /* end participation in shares distribution */ 7265 for_each_possible_cpu(i) 7266 unregister_fair_sched_group(tg, i); 7267 7268 spin_lock_irqsave(&task_group_lock, flags); 7269 list_del_rcu(&tg->list); 7270 list_del_rcu(&tg->siblings); 7271 spin_unlock_irqrestore(&task_group_lock, flags); 7272} 7273 7274/* change task's runqueue when it moves between groups. 7275 * The caller of this function should have put the task in its new group 7276 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7277 * reflect its new group. 7278 */ 7279void sched_move_task(struct task_struct *tsk) 7280{ 7281 struct task_group *tg; 7282 int on_rq, running; 7283 unsigned long flags; 7284 struct rq *rq; 7285 7286 rq = task_rq_lock(tsk, &flags); 7287 7288 running = task_current(rq, tsk); 7289 on_rq = tsk->on_rq; 7290 7291 if (on_rq) 7292 dequeue_task(rq, tsk, 0); 7293 if (unlikely(running)) 7294 tsk->sched_class->put_prev_task(rq, tsk); 7295 7296 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7297 lockdep_is_held(&tsk->sighand->siglock)), 7298 struct task_group, css); 7299 tg = autogroup_task_group(tsk, tg); 7300 tsk->sched_task_group = tg; 7301 7302#ifdef CONFIG_FAIR_GROUP_SCHED 7303 if (tsk->sched_class->task_move_group) 7304 tsk->sched_class->task_move_group(tsk, on_rq); 7305 else 7306#endif 7307 set_task_rq(tsk, task_cpu(tsk)); 7308 7309 if (unlikely(running)) 7310 tsk->sched_class->set_curr_task(rq); 7311 if (on_rq) 7312 enqueue_task(rq, tsk, 0); 7313 7314 task_rq_unlock(rq, tsk, &flags); 7315} 7316#endif /* CONFIG_CGROUP_SCHED */ 7317 7318#ifdef CONFIG_RT_GROUP_SCHED 7319/* 7320 * Ensure that the real time constraints are schedulable. 7321 */ 7322static DEFINE_MUTEX(rt_constraints_mutex); 7323 7324/* Must be called with tasklist_lock held */ 7325static inline int tg_has_rt_tasks(struct task_group *tg) 7326{ 7327 struct task_struct *g, *p; 7328 7329 do_each_thread(g, p) { 7330 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7331 return 1; 7332 } while_each_thread(g, p); 7333 7334 return 0; 7335} 7336 7337struct rt_schedulable_data { 7338 struct task_group *tg; 7339 u64 rt_period; 7340 u64 rt_runtime; 7341}; 7342 7343static int tg_rt_schedulable(struct task_group *tg, void *data) 7344{ 7345 struct rt_schedulable_data *d = data; 7346 struct task_group *child; 7347 unsigned long total, sum = 0; 7348 u64 period, runtime; 7349 7350 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7351 runtime = tg->rt_bandwidth.rt_runtime; 7352 7353 if (tg == d->tg) { 7354 period = d->rt_period; 7355 runtime = d->rt_runtime; 7356 } 7357 7358 /* 7359 * Cannot have more runtime than the period. 7360 */ 7361 if (runtime > period && runtime != RUNTIME_INF) 7362 return -EINVAL; 7363 7364 /* 7365 * Ensure we don't starve existing RT tasks. 7366 */ 7367 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7368 return -EBUSY; 7369 7370 total = to_ratio(period, runtime); 7371 7372 /* 7373 * Nobody can have more than the global setting allows. 7374 */ 7375 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7376 return -EINVAL; 7377 7378 /* 7379 * The sum of our children's runtime should not exceed our own. 7380 */ 7381 list_for_each_entry_rcu(child, &tg->children, siblings) { 7382 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7383 runtime = child->rt_bandwidth.rt_runtime; 7384 7385 if (child == d->tg) { 7386 period = d->rt_period; 7387 runtime = d->rt_runtime; 7388 } 7389 7390 sum += to_ratio(period, runtime); 7391 } 7392 7393 if (sum > total) 7394 return -EINVAL; 7395 7396 return 0; 7397} 7398 7399static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7400{ 7401 int ret; 7402 7403 struct rt_schedulable_data data = { 7404 .tg = tg, 7405 .rt_period = period, 7406 .rt_runtime = runtime, 7407 }; 7408 7409 rcu_read_lock(); 7410 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7411 rcu_read_unlock(); 7412 7413 return ret; 7414} 7415 7416static int tg_set_rt_bandwidth(struct task_group *tg, 7417 u64 rt_period, u64 rt_runtime) 7418{ 7419 int i, err = 0; 7420 7421 mutex_lock(&rt_constraints_mutex); 7422 read_lock(&tasklist_lock); 7423 err = __rt_schedulable(tg, rt_period, rt_runtime); 7424 if (err) 7425 goto unlock; 7426 7427 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7428 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7429 tg->rt_bandwidth.rt_runtime = rt_runtime; 7430 7431 for_each_possible_cpu(i) { 7432 struct rt_rq *rt_rq = tg->rt_rq[i]; 7433 7434 raw_spin_lock(&rt_rq->rt_runtime_lock); 7435 rt_rq->rt_runtime = rt_runtime; 7436 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7437 } 7438 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7439unlock: 7440 read_unlock(&tasklist_lock); 7441 mutex_unlock(&rt_constraints_mutex); 7442 7443 return err; 7444} 7445 7446static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7447{ 7448 u64 rt_runtime, rt_period; 7449 7450 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7451 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7452 if (rt_runtime_us < 0) 7453 rt_runtime = RUNTIME_INF; 7454 7455 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7456} 7457 7458static long sched_group_rt_runtime(struct task_group *tg) 7459{ 7460 u64 rt_runtime_us; 7461 7462 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7463 return -1; 7464 7465 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7466 do_div(rt_runtime_us, NSEC_PER_USEC); 7467 return rt_runtime_us; 7468} 7469 7470static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7471{ 7472 u64 rt_runtime, rt_period; 7473 7474 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7475 rt_runtime = tg->rt_bandwidth.rt_runtime; 7476 7477 if (rt_period == 0) 7478 return -EINVAL; 7479 7480 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7481} 7482 7483static long sched_group_rt_period(struct task_group *tg) 7484{ 7485 u64 rt_period_us; 7486 7487 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7488 do_div(rt_period_us, NSEC_PER_USEC); 7489 return rt_period_us; 7490} 7491#endif /* CONFIG_RT_GROUP_SCHED */ 7492 7493#ifdef CONFIG_RT_GROUP_SCHED 7494static int sched_rt_global_constraints(void) 7495{ 7496 int ret = 0; 7497 7498 mutex_lock(&rt_constraints_mutex); 7499 read_lock(&tasklist_lock); 7500 ret = __rt_schedulable(NULL, 0, 0); 7501 read_unlock(&tasklist_lock); 7502 mutex_unlock(&rt_constraints_mutex); 7503 7504 return ret; 7505} 7506 7507static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7508{ 7509 /* Don't accept realtime tasks when there is no way for them to run */ 7510 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7511 return 0; 7512 7513 return 1; 7514} 7515 7516#else /* !CONFIG_RT_GROUP_SCHED */ 7517static int sched_rt_global_constraints(void) 7518{ 7519 unsigned long flags; 7520 int i, ret = 0; 7521 7522 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7523 for_each_possible_cpu(i) { 7524 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7525 7526 raw_spin_lock(&rt_rq->rt_runtime_lock); 7527 rt_rq->rt_runtime = global_rt_runtime(); 7528 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7529 } 7530 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7531 7532 return ret; 7533} 7534#endif /* CONFIG_RT_GROUP_SCHED */ 7535 7536static int sched_dl_global_constraints(void) 7537{ 7538 u64 runtime = global_rt_runtime(); 7539 u64 period = global_rt_period(); 7540 u64 new_bw = to_ratio(period, runtime); 7541 int cpu, ret = 0; 7542 unsigned long flags; 7543 7544 /* 7545 * Here we want to check the bandwidth not being set to some 7546 * value smaller than the currently allocated bandwidth in 7547 * any of the root_domains. 7548 * 7549 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7550 * cycling on root_domains... Discussion on different/better 7551 * solutions is welcome! 7552 */ 7553 for_each_possible_cpu(cpu) { 7554 struct dl_bw *dl_b = dl_bw_of(cpu); 7555 7556 raw_spin_lock_irqsave(&dl_b->lock, flags); 7557 if (new_bw < dl_b->total_bw) 7558 ret = -EBUSY; 7559 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7560 7561 if (ret) 7562 break; 7563 } 7564 7565 return ret; 7566} 7567 7568static void sched_dl_do_global(void) 7569{ 7570 u64 new_bw = -1; 7571 int cpu; 7572 unsigned long flags; 7573 7574 def_dl_bandwidth.dl_period = global_rt_period(); 7575 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7576 7577 if (global_rt_runtime() != RUNTIME_INF) 7578 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7579 7580 /* 7581 * FIXME: As above... 7582 */ 7583 for_each_possible_cpu(cpu) { 7584 struct dl_bw *dl_b = dl_bw_of(cpu); 7585 7586 raw_spin_lock_irqsave(&dl_b->lock, flags); 7587 dl_b->bw = new_bw; 7588 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7589 } 7590} 7591 7592static int sched_rt_global_validate(void) 7593{ 7594 if (sysctl_sched_rt_period <= 0) 7595 return -EINVAL; 7596 7597 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7598 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7599 return -EINVAL; 7600 7601 return 0; 7602} 7603 7604static void sched_rt_do_global(void) 7605{ 7606 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7607 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7608} 7609 7610int sched_rt_handler(struct ctl_table *table, int write, 7611 void __user *buffer, size_t *lenp, 7612 loff_t *ppos) 7613{ 7614 int old_period, old_runtime; 7615 static DEFINE_MUTEX(mutex); 7616 int ret; 7617 7618 mutex_lock(&mutex); 7619 old_period = sysctl_sched_rt_period; 7620 old_runtime = sysctl_sched_rt_runtime; 7621 7622 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7623 7624 if (!ret && write) { 7625 ret = sched_rt_global_validate(); 7626 if (ret) 7627 goto undo; 7628 7629 ret = sched_rt_global_constraints(); 7630 if (ret) 7631 goto undo; 7632 7633 ret = sched_dl_global_constraints(); 7634 if (ret) 7635 goto undo; 7636 7637 sched_rt_do_global(); 7638 sched_dl_do_global(); 7639 } 7640 if (0) { 7641undo: 7642 sysctl_sched_rt_period = old_period; 7643 sysctl_sched_rt_runtime = old_runtime; 7644 } 7645 mutex_unlock(&mutex); 7646 7647 return ret; 7648} 7649 7650int sched_rr_handler(struct ctl_table *table, int write, 7651 void __user *buffer, size_t *lenp, 7652 loff_t *ppos) 7653{ 7654 int ret; 7655 static DEFINE_MUTEX(mutex); 7656 7657 mutex_lock(&mutex); 7658 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7659 /* make sure that internally we keep jiffies */ 7660 /* also, writing zero resets timeslice to default */ 7661 if (!ret && write) { 7662 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7663 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7664 } 7665 mutex_unlock(&mutex); 7666 return ret; 7667} 7668 7669#ifdef CONFIG_CGROUP_SCHED 7670 7671static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7672{ 7673 return css ? container_of(css, struct task_group, css) : NULL; 7674} 7675 7676static struct cgroup_subsys_state * 7677cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7678{ 7679 struct task_group *parent = css_tg(parent_css); 7680 struct task_group *tg; 7681 7682 if (!parent) { 7683 /* This is early initialization for the top cgroup */ 7684 return &root_task_group.css; 7685 } 7686 7687 tg = sched_create_group(parent); 7688 if (IS_ERR(tg)) 7689 return ERR_PTR(-ENOMEM); 7690 7691 return &tg->css; 7692} 7693 7694static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7695{ 7696 struct task_group *tg = css_tg(css); 7697 struct task_group *parent = css_tg(css->parent); 7698 7699 if (parent) 7700 sched_online_group(tg, parent); 7701 return 0; 7702} 7703 7704static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7705{ 7706 struct task_group *tg = css_tg(css); 7707 7708 sched_destroy_group(tg); 7709} 7710 7711static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7712{ 7713 struct task_group *tg = css_tg(css); 7714 7715 sched_offline_group(tg); 7716} 7717 7718static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7719 struct cgroup_taskset *tset) 7720{ 7721 struct task_struct *task; 7722 7723 cgroup_taskset_for_each(task, tset) { 7724#ifdef CONFIG_RT_GROUP_SCHED 7725 if (!sched_rt_can_attach(css_tg(css), task)) 7726 return -EINVAL; 7727#else 7728 /* We don't support RT-tasks being in separate groups */ 7729 if (task->sched_class != &fair_sched_class) 7730 return -EINVAL; 7731#endif 7732 } 7733 return 0; 7734} 7735 7736static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7737 struct cgroup_taskset *tset) 7738{ 7739 struct task_struct *task; 7740 7741 cgroup_taskset_for_each(task, tset) 7742 sched_move_task(task); 7743} 7744 7745static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7746 struct cgroup_subsys_state *old_css, 7747 struct task_struct *task) 7748{ 7749 /* 7750 * cgroup_exit() is called in the copy_process() failure path. 7751 * Ignore this case since the task hasn't ran yet, this avoids 7752 * trying to poke a half freed task state from generic code. 7753 */ 7754 if (!(task->flags & PF_EXITING)) 7755 return; 7756 7757 sched_move_task(task); 7758} 7759 7760#ifdef CONFIG_FAIR_GROUP_SCHED 7761static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7762 struct cftype *cftype, u64 shareval) 7763{ 7764 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7765} 7766 7767static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7768 struct cftype *cft) 7769{ 7770 struct task_group *tg = css_tg(css); 7771 7772 return (u64) scale_load_down(tg->shares); 7773} 7774 7775#ifdef CONFIG_CFS_BANDWIDTH 7776static DEFINE_MUTEX(cfs_constraints_mutex); 7777 7778const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7779const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7780 7781static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7782 7783static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7784{ 7785 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7786 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7787 7788 if (tg == &root_task_group) 7789 return -EINVAL; 7790 7791 /* 7792 * Ensure we have at some amount of bandwidth every period. This is 7793 * to prevent reaching a state of large arrears when throttled via 7794 * entity_tick() resulting in prolonged exit starvation. 7795 */ 7796 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7797 return -EINVAL; 7798 7799 /* 7800 * Likewise, bound things on the otherside by preventing insane quota 7801 * periods. This also allows us to normalize in computing quota 7802 * feasibility. 7803 */ 7804 if (period > max_cfs_quota_period) 7805 return -EINVAL; 7806 7807 mutex_lock(&cfs_constraints_mutex); 7808 ret = __cfs_schedulable(tg, period, quota); 7809 if (ret) 7810 goto out_unlock; 7811 7812 runtime_enabled = quota != RUNTIME_INF; 7813 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7814 /* 7815 * If we need to toggle cfs_bandwidth_used, off->on must occur 7816 * before making related changes, and on->off must occur afterwards 7817 */ 7818 if (runtime_enabled && !runtime_was_enabled) 7819 cfs_bandwidth_usage_inc(); 7820 raw_spin_lock_irq(&cfs_b->lock); 7821 cfs_b->period = ns_to_ktime(period); 7822 cfs_b->quota = quota; 7823 7824 __refill_cfs_bandwidth_runtime(cfs_b); 7825 /* restart the period timer (if active) to handle new period expiry */ 7826 if (runtime_enabled && cfs_b->timer_active) { 7827 /* force a reprogram */ 7828 __start_cfs_bandwidth(cfs_b, true); 7829 } 7830 raw_spin_unlock_irq(&cfs_b->lock); 7831 7832 for_each_possible_cpu(i) { 7833 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7834 struct rq *rq = cfs_rq->rq; 7835 7836 raw_spin_lock_irq(&rq->lock); 7837 cfs_rq->runtime_enabled = runtime_enabled; 7838 cfs_rq->runtime_remaining = 0; 7839 7840 if (cfs_rq->throttled) 7841 unthrottle_cfs_rq(cfs_rq); 7842 raw_spin_unlock_irq(&rq->lock); 7843 } 7844 if (runtime_was_enabled && !runtime_enabled) 7845 cfs_bandwidth_usage_dec(); 7846out_unlock: 7847 mutex_unlock(&cfs_constraints_mutex); 7848 7849 return ret; 7850} 7851 7852int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7853{ 7854 u64 quota, period; 7855 7856 period = ktime_to_ns(tg->cfs_bandwidth.period); 7857 if (cfs_quota_us < 0) 7858 quota = RUNTIME_INF; 7859 else 7860 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7861 7862 return tg_set_cfs_bandwidth(tg, period, quota); 7863} 7864 7865long tg_get_cfs_quota(struct task_group *tg) 7866{ 7867 u64 quota_us; 7868 7869 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7870 return -1; 7871 7872 quota_us = tg->cfs_bandwidth.quota; 7873 do_div(quota_us, NSEC_PER_USEC); 7874 7875 return quota_us; 7876} 7877 7878int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7879{ 7880 u64 quota, period; 7881 7882 period = (u64)cfs_period_us * NSEC_PER_USEC; 7883 quota = tg->cfs_bandwidth.quota; 7884 7885 return tg_set_cfs_bandwidth(tg, period, quota); 7886} 7887 7888long tg_get_cfs_period(struct task_group *tg) 7889{ 7890 u64 cfs_period_us; 7891 7892 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7893 do_div(cfs_period_us, NSEC_PER_USEC); 7894 7895 return cfs_period_us; 7896} 7897 7898static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7899 struct cftype *cft) 7900{ 7901 return tg_get_cfs_quota(css_tg(css)); 7902} 7903 7904static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7905 struct cftype *cftype, s64 cfs_quota_us) 7906{ 7907 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7908} 7909 7910static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7911 struct cftype *cft) 7912{ 7913 return tg_get_cfs_period(css_tg(css)); 7914} 7915 7916static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7917 struct cftype *cftype, u64 cfs_period_us) 7918{ 7919 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7920} 7921 7922struct cfs_schedulable_data { 7923 struct task_group *tg; 7924 u64 period, quota; 7925}; 7926 7927/* 7928 * normalize group quota/period to be quota/max_period 7929 * note: units are usecs 7930 */ 7931static u64 normalize_cfs_quota(struct task_group *tg, 7932 struct cfs_schedulable_data *d) 7933{ 7934 u64 quota, period; 7935 7936 if (tg == d->tg) { 7937 period = d->period; 7938 quota = d->quota; 7939 } else { 7940 period = tg_get_cfs_period(tg); 7941 quota = tg_get_cfs_quota(tg); 7942 } 7943 7944 /* note: these should typically be equivalent */ 7945 if (quota == RUNTIME_INF || quota == -1) 7946 return RUNTIME_INF; 7947 7948 return to_ratio(period, quota); 7949} 7950 7951static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7952{ 7953 struct cfs_schedulable_data *d = data; 7954 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7955 s64 quota = 0, parent_quota = -1; 7956 7957 if (!tg->parent) { 7958 quota = RUNTIME_INF; 7959 } else { 7960 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7961 7962 quota = normalize_cfs_quota(tg, d); 7963 parent_quota = parent_b->hierarchal_quota; 7964 7965 /* 7966 * ensure max(child_quota) <= parent_quota, inherit when no 7967 * limit is set 7968 */ 7969 if (quota == RUNTIME_INF) 7970 quota = parent_quota; 7971 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7972 return -EINVAL; 7973 } 7974 cfs_b->hierarchal_quota = quota; 7975 7976 return 0; 7977} 7978 7979static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7980{ 7981 int ret; 7982 struct cfs_schedulable_data data = { 7983 .tg = tg, 7984 .period = period, 7985 .quota = quota, 7986 }; 7987 7988 if (quota != RUNTIME_INF) { 7989 do_div(data.period, NSEC_PER_USEC); 7990 do_div(data.quota, NSEC_PER_USEC); 7991 } 7992 7993 rcu_read_lock(); 7994 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7995 rcu_read_unlock(); 7996 7997 return ret; 7998} 7999 8000static int cpu_stats_show(struct seq_file *sf, void *v) 8001{ 8002 struct task_group *tg = css_tg(seq_css(sf)); 8003 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8004 8005 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 8006 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 8007 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 8008 8009 return 0; 8010} 8011#endif /* CONFIG_CFS_BANDWIDTH */ 8012#endif /* CONFIG_FAIR_GROUP_SCHED */ 8013 8014#ifdef CONFIG_RT_GROUP_SCHED 8015static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 8016 struct cftype *cft, s64 val) 8017{ 8018 return sched_group_set_rt_runtime(css_tg(css), val); 8019} 8020 8021static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 8022 struct cftype *cft) 8023{ 8024 return sched_group_rt_runtime(css_tg(css)); 8025} 8026 8027static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 8028 struct cftype *cftype, u64 rt_period_us) 8029{ 8030 return sched_group_set_rt_period(css_tg(css), rt_period_us); 8031} 8032 8033static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 8034 struct cftype *cft) 8035{ 8036 return sched_group_rt_period(css_tg(css)); 8037} 8038#endif /* CONFIG_RT_GROUP_SCHED */ 8039 8040static struct cftype cpu_files[] = { 8041#ifdef CONFIG_FAIR_GROUP_SCHED 8042 { 8043 .name = "shares", 8044 .read_u64 = cpu_shares_read_u64, 8045 .write_u64 = cpu_shares_write_u64, 8046 }, 8047#endif 8048#ifdef CONFIG_CFS_BANDWIDTH 8049 { 8050 .name = "cfs_quota_us", 8051 .read_s64 = cpu_cfs_quota_read_s64, 8052 .write_s64 = cpu_cfs_quota_write_s64, 8053 }, 8054 { 8055 .name = "cfs_period_us", 8056 .read_u64 = cpu_cfs_period_read_u64, 8057 .write_u64 = cpu_cfs_period_write_u64, 8058 }, 8059 { 8060 .name = "stat", 8061 .seq_show = cpu_stats_show, 8062 }, 8063#endif 8064#ifdef CONFIG_RT_GROUP_SCHED 8065 { 8066 .name = "rt_runtime_us", 8067 .read_s64 = cpu_rt_runtime_read, 8068 .write_s64 = cpu_rt_runtime_write, 8069 }, 8070 { 8071 .name = "rt_period_us", 8072 .read_u64 = cpu_rt_period_read_uint, 8073 .write_u64 = cpu_rt_period_write_uint, 8074 }, 8075#endif 8076 { } /* terminate */ 8077}; 8078 8079struct cgroup_subsys cpu_cgrp_subsys = { 8080 .css_alloc = cpu_cgroup_css_alloc, 8081 .css_free = cpu_cgroup_css_free, 8082 .css_online = cpu_cgroup_css_online, 8083 .css_offline = cpu_cgroup_css_offline, 8084 .can_attach = cpu_cgroup_can_attach, 8085 .attach = cpu_cgroup_attach, 8086 .exit = cpu_cgroup_exit, 8087 .base_cftypes = cpu_files, 8088 .early_init = 1, 8089}; 8090 8091#endif /* CONFIG_CGROUP_SCHED */ 8092 8093void dump_cpu_task(int cpu) 8094{ 8095 pr_info("Task dump for CPU %d:\n", cpu); 8096 sched_show_task(cpu_curr(cpu)); 8097} 8098