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