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 * __task_rq_lock - lock the rq @p resides on. 301 */ 302 static inline struct rq *__task_rq_lock(struct task_struct *p) 303 __acquires(rq->lock) 304 { 305 struct rq *rq; 306 307 lockdep_assert_held(&p->pi_lock); 308 309 for (;;) { 310 rq = task_rq(p); 311 raw_spin_lock(&rq->lock); 312 if (likely(rq == task_rq(p))) 313 return rq; 314 raw_spin_unlock(&rq->lock); 315 } 316 } 317 318 /* 319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 320 */ 321 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 322 __acquires(p->pi_lock) 323 __acquires(rq->lock) 324 { 325 struct rq *rq; 326 327 for (;;) { 328 raw_spin_lock_irqsave(&p->pi_lock, *flags); 329 rq = task_rq(p); 330 raw_spin_lock(&rq->lock); 331 if (likely(rq == task_rq(p))) 332 return rq; 333 raw_spin_unlock(&rq->lock); 334 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 335 } 336 } 337 338 static void __task_rq_unlock(struct rq *rq) 339 __releases(rq->lock) 340 { 341 raw_spin_unlock(&rq->lock); 342 } 343 344 static inline void 345 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 346 __releases(rq->lock) 347 __releases(p->pi_lock) 348 { 349 raw_spin_unlock(&rq->lock); 350 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 351 } 352 353 /* 354 * this_rq_lock - lock this runqueue and disable interrupts. 355 */ 356 static struct rq *this_rq_lock(void) 357 __acquires(rq->lock) 358 { 359 struct rq *rq; 360 361 local_irq_disable(); 362 rq = this_rq(); 363 raw_spin_lock(&rq->lock); 364 365 return rq; 366 } 367 368 #ifdef CONFIG_SCHED_HRTICK 369 /* 370 * Use HR-timers to deliver accurate preemption points. 371 */ 372 373 static void hrtick_clear(struct rq *rq) 374 { 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377 } 378 379 /* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383 static enum hrtimer_restart hrtick(struct hrtimer *timer) 384 { 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395 } 396 397 #ifdef CONFIG_SMP 398 399 static int __hrtick_restart(struct rq *rq) 400 { 401 struct hrtimer *timer = &rq->hrtick_timer; 402 ktime_t time = hrtimer_get_softexpires(timer); 403 404 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 405 } 406 407 /* 408 * called from hardirq (IPI) context 409 */ 410 static void __hrtick_start(void *arg) 411 { 412 struct rq *rq = arg; 413 414 raw_spin_lock(&rq->lock); 415 __hrtick_restart(rq); 416 rq->hrtick_csd_pending = 0; 417 raw_spin_unlock(&rq->lock); 418 } 419 420 /* 421 * Called to set the hrtick timer state. 422 * 423 * called with rq->lock held and irqs disabled 424 */ 425 void hrtick_start(struct rq *rq, u64 delay) 426 { 427 struct hrtimer *timer = &rq->hrtick_timer; 428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 429 430 hrtimer_set_expires(timer, time); 431 432 if (rq == this_rq()) { 433 __hrtick_restart(rq); 434 } else if (!rq->hrtick_csd_pending) { 435 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 436 rq->hrtick_csd_pending = 1; 437 } 438 } 439 440 static int 441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 442 { 443 int cpu = (int)(long)hcpu; 444 445 switch (action) { 446 case CPU_UP_CANCELED: 447 case CPU_UP_CANCELED_FROZEN: 448 case CPU_DOWN_PREPARE: 449 case CPU_DOWN_PREPARE_FROZEN: 450 case CPU_DEAD: 451 case CPU_DEAD_FROZEN: 452 hrtick_clear(cpu_rq(cpu)); 453 return NOTIFY_OK; 454 } 455 456 return NOTIFY_DONE; 457 } 458 459 static __init void init_hrtick(void) 460 { 461 hotcpu_notifier(hotplug_hrtick, 0); 462 } 463 #else 464 /* 465 * Called to set the hrtick timer state. 466 * 467 * called with rq->lock held and irqs disabled 468 */ 469 void hrtick_start(struct rq *rq, u64 delay) 470 { 471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 472 HRTIMER_MODE_REL_PINNED, 0); 473 } 474 475 static inline void init_hrtick(void) 476 { 477 } 478 #endif /* CONFIG_SMP */ 479 480 static void init_rq_hrtick(struct rq *rq) 481 { 482 #ifdef CONFIG_SMP 483 rq->hrtick_csd_pending = 0; 484 485 rq->hrtick_csd.flags = 0; 486 rq->hrtick_csd.func = __hrtick_start; 487 rq->hrtick_csd.info = rq; 488 #endif 489 490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 491 rq->hrtick_timer.function = hrtick; 492 } 493 #else /* CONFIG_SCHED_HRTICK */ 494 static inline void hrtick_clear(struct rq *rq) 495 { 496 } 497 498 static inline void init_rq_hrtick(struct rq *rq) 499 { 500 } 501 502 static inline void init_hrtick(void) 503 { 504 } 505 #endif /* CONFIG_SCHED_HRTICK */ 506 507 /* 508 * resched_task - mark a task 'to be rescheduled now'. 509 * 510 * On UP this means the setting of the need_resched flag, on SMP it 511 * might also involve a cross-CPU call to trigger the scheduler on 512 * the target CPU. 513 */ 514 void resched_task(struct task_struct *p) 515 { 516 int cpu; 517 518 lockdep_assert_held(&task_rq(p)->lock); 519 520 if (test_tsk_need_resched(p)) 521 return; 522 523 set_tsk_need_resched(p); 524 525 cpu = task_cpu(p); 526 if (cpu == smp_processor_id()) { 527 set_preempt_need_resched(); 528 return; 529 } 530 531 /* NEED_RESCHED must be visible before we test polling */ 532 smp_mb(); 533 if (!tsk_is_polling(p)) 534 smp_send_reschedule(cpu); 535 } 536 537 void resched_cpu(int cpu) 538 { 539 struct rq *rq = cpu_rq(cpu); 540 unsigned long flags; 541 542 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 543 return; 544 resched_task(cpu_curr(cpu)); 545 raw_spin_unlock_irqrestore(&rq->lock, flags); 546 } 547 548 #ifdef CONFIG_SMP 549 #ifdef CONFIG_NO_HZ_COMMON 550 /* 551 * In the semi idle case, use the nearest busy cpu for migrating timers 552 * from an idle cpu. This is good for power-savings. 553 * 554 * We don't do similar optimization for completely idle system, as 555 * selecting an idle cpu will add more delays to the timers than intended 556 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 557 */ 558 int get_nohz_timer_target(int pinned) 559 { 560 int cpu = smp_processor_id(); 561 int i; 562 struct sched_domain *sd; 563 564 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 565 return cpu; 566 567 rcu_read_lock(); 568 for_each_domain(cpu, sd) { 569 for_each_cpu(i, sched_domain_span(sd)) { 570 if (!idle_cpu(i)) { 571 cpu = i; 572 goto unlock; 573 } 574 } 575 } 576 unlock: 577 rcu_read_unlock(); 578 return cpu; 579 } 580 /* 581 * When add_timer_on() enqueues a timer into the timer wheel of an 582 * idle CPU then this timer might expire before the next timer event 583 * which is scheduled to wake up that CPU. In case of a completely 584 * idle system the next event might even be infinite time into the 585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 586 * leaves the inner idle loop so the newly added timer is taken into 587 * account when the CPU goes back to idle and evaluates the timer 588 * wheel for the next timer event. 589 */ 590 static void wake_up_idle_cpu(int cpu) 591 { 592 struct rq *rq = cpu_rq(cpu); 593 594 if (cpu == smp_processor_id()) 595 return; 596 597 /* 598 * This is safe, as this function is called with the timer 599 * wheel base lock of (cpu) held. When the CPU is on the way 600 * to idle and has not yet set rq->curr to idle then it will 601 * be serialized on the timer wheel base lock and take the new 602 * timer into account automatically. 603 */ 604 if (rq->curr != rq->idle) 605 return; 606 607 /* 608 * We can set TIF_RESCHED on the idle task of the other CPU 609 * lockless. The worst case is that the other CPU runs the 610 * idle task through an additional NOOP schedule() 611 */ 612 set_tsk_need_resched(rq->idle); 613 614 /* NEED_RESCHED must be visible before we test polling */ 615 smp_mb(); 616 if (!tsk_is_polling(rq->idle)) 617 smp_send_reschedule(cpu); 618 } 619 620 static bool wake_up_full_nohz_cpu(int cpu) 621 { 622 if (tick_nohz_full_cpu(cpu)) { 623 if (cpu != smp_processor_id() || 624 tick_nohz_tick_stopped()) 625 smp_send_reschedule(cpu); 626 return true; 627 } 628 629 return false; 630 } 631 632 void wake_up_nohz_cpu(int cpu) 633 { 634 if (!wake_up_full_nohz_cpu(cpu)) 635 wake_up_idle_cpu(cpu); 636 } 637 638 static inline bool got_nohz_idle_kick(void) 639 { 640 int cpu = smp_processor_id(); 641 642 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 643 return false; 644 645 if (idle_cpu(cpu) && !need_resched()) 646 return true; 647 648 /* 649 * We can't run Idle Load Balance on this CPU for this time so we 650 * cancel it and clear NOHZ_BALANCE_KICK 651 */ 652 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 653 return false; 654 } 655 656 #else /* CONFIG_NO_HZ_COMMON */ 657 658 static inline bool got_nohz_idle_kick(void) 659 { 660 return false; 661 } 662 663 #endif /* CONFIG_NO_HZ_COMMON */ 664 665 #ifdef CONFIG_NO_HZ_FULL 666 bool sched_can_stop_tick(void) 667 { 668 struct rq *rq; 669 670 rq = this_rq(); 671 672 /* Make sure rq->nr_running update is visible after the IPI */ 673 smp_rmb(); 674 675 /* More than one running task need preemption */ 676 if (rq->nr_running > 1) 677 return false; 678 679 return true; 680 } 681 #endif /* CONFIG_NO_HZ_FULL */ 682 683 void sched_avg_update(struct rq *rq) 684 { 685 s64 period = sched_avg_period(); 686 687 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 688 /* 689 * Inline assembly required to prevent the compiler 690 * optimising this loop into a divmod call. 691 * See __iter_div_u64_rem() for another example of this. 692 */ 693 asm("" : "+rm" (rq->age_stamp)); 694 rq->age_stamp += period; 695 rq->rt_avg /= 2; 696 } 697 } 698 699 #endif /* CONFIG_SMP */ 700 701 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 702 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 703 /* 704 * Iterate task_group tree rooted at *from, calling @down when first entering a 705 * node and @up when leaving it for the final time. 706 * 707 * Caller must hold rcu_lock or sufficient equivalent. 708 */ 709 int walk_tg_tree_from(struct task_group *from, 710 tg_visitor down, tg_visitor up, void *data) 711 { 712 struct task_group *parent, *child; 713 int ret; 714 715 parent = from; 716 717 down: 718 ret = (*down)(parent, data); 719 if (ret) 720 goto out; 721 list_for_each_entry_rcu(child, &parent->children, siblings) { 722 parent = child; 723 goto down; 724 725 up: 726 continue; 727 } 728 ret = (*up)(parent, data); 729 if (ret || parent == from) 730 goto out; 731 732 child = parent; 733 parent = parent->parent; 734 if (parent) 735 goto up; 736 out: 737 return ret; 738 } 739 740 int tg_nop(struct task_group *tg, void *data) 741 { 742 return 0; 743 } 744 #endif 745 746 static void set_load_weight(struct task_struct *p) 747 { 748 int prio = p->static_prio - MAX_RT_PRIO; 749 struct load_weight *load = &p->se.load; 750 751 /* 752 * SCHED_IDLE tasks get minimal weight: 753 */ 754 if (p->policy == SCHED_IDLE) { 755 load->weight = scale_load(WEIGHT_IDLEPRIO); 756 load->inv_weight = WMULT_IDLEPRIO; 757 return; 758 } 759 760 load->weight = scale_load(prio_to_weight[prio]); 761 load->inv_weight = prio_to_wmult[prio]; 762 } 763 764 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 765 { 766 update_rq_clock(rq); 767 sched_info_queued(rq, p); 768 p->sched_class->enqueue_task(rq, p, flags); 769 } 770 771 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 772 { 773 update_rq_clock(rq); 774 sched_info_dequeued(rq, p); 775 p->sched_class->dequeue_task(rq, p, flags); 776 } 777 778 void activate_task(struct rq *rq, struct task_struct *p, int flags) 779 { 780 if (task_contributes_to_load(p)) 781 rq->nr_uninterruptible--; 782 783 enqueue_task(rq, p, flags); 784 } 785 786 void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 787 { 788 if (task_contributes_to_load(p)) 789 rq->nr_uninterruptible++; 790 791 dequeue_task(rq, p, flags); 792 } 793 794 static void update_rq_clock_task(struct rq *rq, s64 delta) 795 { 796 /* 797 * In theory, the compile should just see 0 here, and optimize out the call 798 * to sched_rt_avg_update. But I don't trust it... 799 */ 800 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 801 s64 steal = 0, irq_delta = 0; 802 #endif 803 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 804 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 805 806 /* 807 * Since irq_time is only updated on {soft,}irq_exit, we might run into 808 * this case when a previous update_rq_clock() happened inside a 809 * {soft,}irq region. 810 * 811 * When this happens, we stop ->clock_task and only update the 812 * prev_irq_time stamp to account for the part that fit, so that a next 813 * update will consume the rest. This ensures ->clock_task is 814 * monotonic. 815 * 816 * It does however cause some slight miss-attribution of {soft,}irq 817 * time, a more accurate solution would be to update the irq_time using 818 * the current rq->clock timestamp, except that would require using 819 * atomic ops. 820 */ 821 if (irq_delta > delta) 822 irq_delta = delta; 823 824 rq->prev_irq_time += irq_delta; 825 delta -= irq_delta; 826 #endif 827 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 828 if (static_key_false((¶virt_steal_rq_enabled))) { 829 steal = paravirt_steal_clock(cpu_of(rq)); 830 steal -= rq->prev_steal_time_rq; 831 832 if (unlikely(steal > delta)) 833 steal = delta; 834 835 rq->prev_steal_time_rq += steal; 836 delta -= steal; 837 } 838 #endif 839 840 rq->clock_task += delta; 841 842 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 843 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 844 sched_rt_avg_update(rq, irq_delta + steal); 845 #endif 846 } 847 848 void sched_set_stop_task(int cpu, struct task_struct *stop) 849 { 850 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 851 struct task_struct *old_stop = cpu_rq(cpu)->stop; 852 853 if (stop) { 854 /* 855 * Make it appear like a SCHED_FIFO task, its something 856 * userspace knows about and won't get confused about. 857 * 858 * Also, it will make PI more or less work without too 859 * much confusion -- but then, stop work should not 860 * rely on PI working anyway. 861 */ 862 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 863 864 stop->sched_class = &stop_sched_class; 865 } 866 867 cpu_rq(cpu)->stop = stop; 868 869 if (old_stop) { 870 /* 871 * Reset it back to a normal scheduling class so that 872 * it can die in pieces. 873 */ 874 old_stop->sched_class = &rt_sched_class; 875 } 876 } 877 878 /* 879 * __normal_prio - return the priority that is based on the static prio 880 */ 881 static inline int __normal_prio(struct task_struct *p) 882 { 883 return p->static_prio; 884 } 885 886 /* 887 * Calculate the expected normal priority: i.e. priority 888 * without taking RT-inheritance into account. Might be 889 * boosted by interactivity modifiers. Changes upon fork, 890 * setprio syscalls, and whenever the interactivity 891 * estimator recalculates. 892 */ 893 static inline int normal_prio(struct task_struct *p) 894 { 895 int prio; 896 897 if (task_has_dl_policy(p)) 898 prio = MAX_DL_PRIO-1; 899 else if (task_has_rt_policy(p)) 900 prio = MAX_RT_PRIO-1 - p->rt_priority; 901 else 902 prio = __normal_prio(p); 903 return prio; 904 } 905 906 /* 907 * Calculate the current priority, i.e. the priority 908 * taken into account by the scheduler. This value might 909 * be boosted by RT tasks, or might be boosted by 910 * interactivity modifiers. Will be RT if the task got 911 * RT-boosted. If not then it returns p->normal_prio. 912 */ 913 static int effective_prio(struct task_struct *p) 914 { 915 p->normal_prio = normal_prio(p); 916 /* 917 * If we are RT tasks or we were boosted to RT priority, 918 * keep the priority unchanged. Otherwise, update priority 919 * to the normal priority: 920 */ 921 if (!rt_prio(p->prio)) 922 return p->normal_prio; 923 return p->prio; 924 } 925 926 /** 927 * task_curr - is this task currently executing on a CPU? 928 * @p: the task in question. 929 * 930 * Return: 1 if the task is currently executing. 0 otherwise. 931 */ 932 inline int task_curr(const struct task_struct *p) 933 { 934 return cpu_curr(task_cpu(p)) == p; 935 } 936 937 static inline void check_class_changed(struct rq *rq, struct task_struct *p, 938 const struct sched_class *prev_class, 939 int oldprio) 940 { 941 if (prev_class != p->sched_class) { 942 if (prev_class->switched_from) 943 prev_class->switched_from(rq, p); 944 p->sched_class->switched_to(rq, p); 945 } else if (oldprio != p->prio || dl_task(p)) 946 p->sched_class->prio_changed(rq, p, oldprio); 947 } 948 949 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 950 { 951 const struct sched_class *class; 952 953 if (p->sched_class == rq->curr->sched_class) { 954 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 955 } else { 956 for_each_class(class) { 957 if (class == rq->curr->sched_class) 958 break; 959 if (class == p->sched_class) { 960 resched_task(rq->curr); 961 break; 962 } 963 } 964 } 965 966 /* 967 * A queue event has occurred, and we're going to schedule. In 968 * this case, we can save a useless back to back clock update. 969 */ 970 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 971 rq->skip_clock_update = 1; 972 } 973 974 #ifdef CONFIG_SMP 975 void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 976 { 977 #ifdef CONFIG_SCHED_DEBUG 978 /* 979 * We should never call set_task_cpu() on a blocked task, 980 * ttwu() will sort out the placement. 981 */ 982 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 983 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 984 985 #ifdef CONFIG_LOCKDEP 986 /* 987 * The caller should hold either p->pi_lock or rq->lock, when changing 988 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 989 * 990 * sched_move_task() holds both and thus holding either pins the cgroup, 991 * see task_group(). 992 * 993 * Furthermore, all task_rq users should acquire both locks, see 994 * task_rq_lock(). 995 */ 996 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 997 lockdep_is_held(&task_rq(p)->lock))); 998 #endif 999 #endif 1000 1001 trace_sched_migrate_task(p, new_cpu); 1002 1003 if (task_cpu(p) != new_cpu) { 1004 if (p->sched_class->migrate_task_rq) 1005 p->sched_class->migrate_task_rq(p, new_cpu); 1006 p->se.nr_migrations++; 1007 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1008 } 1009 1010 __set_task_cpu(p, new_cpu); 1011 } 1012 1013 static void __migrate_swap_task(struct task_struct *p, int cpu) 1014 { 1015 if (p->on_rq) { 1016 struct rq *src_rq, *dst_rq; 1017 1018 src_rq = task_rq(p); 1019 dst_rq = cpu_rq(cpu); 1020 1021 deactivate_task(src_rq, p, 0); 1022 set_task_cpu(p, cpu); 1023 activate_task(dst_rq, p, 0); 1024 check_preempt_curr(dst_rq, p, 0); 1025 } else { 1026 /* 1027 * Task isn't running anymore; make it appear like we migrated 1028 * it before it went to sleep. This means on wakeup we make the 1029 * previous cpu our targer instead of where it really is. 1030 */ 1031 p->wake_cpu = cpu; 1032 } 1033 } 1034 1035 struct migration_swap_arg { 1036 struct task_struct *src_task, *dst_task; 1037 int src_cpu, dst_cpu; 1038 }; 1039 1040 static int migrate_swap_stop(void *data) 1041 { 1042 struct migration_swap_arg *arg = data; 1043 struct rq *src_rq, *dst_rq; 1044 int ret = -EAGAIN; 1045 1046 src_rq = cpu_rq(arg->src_cpu); 1047 dst_rq = cpu_rq(arg->dst_cpu); 1048 1049 double_raw_lock(&arg->src_task->pi_lock, 1050 &arg->dst_task->pi_lock); 1051 double_rq_lock(src_rq, dst_rq); 1052 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1053 goto unlock; 1054 1055 if (task_cpu(arg->src_task) != arg->src_cpu) 1056 goto unlock; 1057 1058 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1059 goto unlock; 1060 1061 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1062 goto unlock; 1063 1064 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1065 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1066 1067 ret = 0; 1068 1069 unlock: 1070 double_rq_unlock(src_rq, dst_rq); 1071 raw_spin_unlock(&arg->dst_task->pi_lock); 1072 raw_spin_unlock(&arg->src_task->pi_lock); 1073 1074 return ret; 1075 } 1076 1077 /* 1078 * Cross migrate two tasks 1079 */ 1080 int migrate_swap(struct task_struct *cur, struct task_struct *p) 1081 { 1082 struct migration_swap_arg arg; 1083 int ret = -EINVAL; 1084 1085 arg = (struct migration_swap_arg){ 1086 .src_task = cur, 1087 .src_cpu = task_cpu(cur), 1088 .dst_task = p, 1089 .dst_cpu = task_cpu(p), 1090 }; 1091 1092 if (arg.src_cpu == arg.dst_cpu) 1093 goto out; 1094 1095 /* 1096 * These three tests are all lockless; this is OK since all of them 1097 * will be re-checked with proper locks held further down the line. 1098 */ 1099 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1100 goto out; 1101 1102 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1103 goto out; 1104 1105 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1106 goto out; 1107 1108 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1109 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1110 1111 out: 1112 return ret; 1113 } 1114 1115 struct migration_arg { 1116 struct task_struct *task; 1117 int dest_cpu; 1118 }; 1119 1120 static int migration_cpu_stop(void *data); 1121 1122 /* 1123 * wait_task_inactive - wait for a thread to unschedule. 1124 * 1125 * If @match_state is nonzero, it's the @p->state value just checked and 1126 * not expected to change. If it changes, i.e. @p might have woken up, 1127 * then return zero. When we succeed in waiting for @p to be off its CPU, 1128 * we return a positive number (its total switch count). If a second call 1129 * a short while later returns the same number, the caller can be sure that 1130 * @p has remained unscheduled the whole time. 1131 * 1132 * The caller must ensure that the task *will* unschedule sometime soon, 1133 * else this function might spin for a *long* time. This function can't 1134 * be called with interrupts off, or it may introduce deadlock with 1135 * smp_call_function() if an IPI is sent by the same process we are 1136 * waiting to become inactive. 1137 */ 1138 unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1139 { 1140 unsigned long flags; 1141 int running, on_rq; 1142 unsigned long ncsw; 1143 struct rq *rq; 1144 1145 for (;;) { 1146 /* 1147 * We do the initial early heuristics without holding 1148 * any task-queue locks at all. We'll only try to get 1149 * the runqueue lock when things look like they will 1150 * work out! 1151 */ 1152 rq = task_rq(p); 1153 1154 /* 1155 * If the task is actively running on another CPU 1156 * still, just relax and busy-wait without holding 1157 * any locks. 1158 * 1159 * NOTE! Since we don't hold any locks, it's not 1160 * even sure that "rq" stays as the right runqueue! 1161 * But we don't care, since "task_running()" will 1162 * return false if the runqueue has changed and p 1163 * is actually now running somewhere else! 1164 */ 1165 while (task_running(rq, p)) { 1166 if (match_state && unlikely(p->state != match_state)) 1167 return 0; 1168 cpu_relax(); 1169 } 1170 1171 /* 1172 * Ok, time to look more closely! We need the rq 1173 * lock now, to be *sure*. If we're wrong, we'll 1174 * just go back and repeat. 1175 */ 1176 rq = task_rq_lock(p, &flags); 1177 trace_sched_wait_task(p); 1178 running = task_running(rq, p); 1179 on_rq = p->on_rq; 1180 ncsw = 0; 1181 if (!match_state || p->state == match_state) 1182 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1183 task_rq_unlock(rq, p, &flags); 1184 1185 /* 1186 * If it changed from the expected state, bail out now. 1187 */ 1188 if (unlikely(!ncsw)) 1189 break; 1190 1191 /* 1192 * Was it really running after all now that we 1193 * checked with the proper locks actually held? 1194 * 1195 * Oops. Go back and try again.. 1196 */ 1197 if (unlikely(running)) { 1198 cpu_relax(); 1199 continue; 1200 } 1201 1202 /* 1203 * It's not enough that it's not actively running, 1204 * it must be off the runqueue _entirely_, and not 1205 * preempted! 1206 * 1207 * So if it was still runnable (but just not actively 1208 * running right now), it's preempted, and we should 1209 * yield - it could be a while. 1210 */ 1211 if (unlikely(on_rq)) { 1212 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1213 1214 set_current_state(TASK_UNINTERRUPTIBLE); 1215 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1216 continue; 1217 } 1218 1219 /* 1220 * Ahh, all good. It wasn't running, and it wasn't 1221 * runnable, which means that it will never become 1222 * running in the future either. We're all done! 1223 */ 1224 break; 1225 } 1226 1227 return ncsw; 1228 } 1229 1230 /*** 1231 * kick_process - kick a running thread to enter/exit the kernel 1232 * @p: the to-be-kicked thread 1233 * 1234 * Cause a process which is running on another CPU to enter 1235 * kernel-mode, without any delay. (to get signals handled.) 1236 * 1237 * NOTE: this function doesn't have to take the runqueue lock, 1238 * because all it wants to ensure is that the remote task enters 1239 * the kernel. If the IPI races and the task has been migrated 1240 * to another CPU then no harm is done and the purpose has been 1241 * achieved as well. 1242 */ 1243 void kick_process(struct task_struct *p) 1244 { 1245 int cpu; 1246 1247 preempt_disable(); 1248 cpu = task_cpu(p); 1249 if ((cpu != smp_processor_id()) && task_curr(p)) 1250 smp_send_reschedule(cpu); 1251 preempt_enable(); 1252 } 1253 EXPORT_SYMBOL_GPL(kick_process); 1254 #endif /* CONFIG_SMP */ 1255 1256 #ifdef CONFIG_SMP 1257 /* 1258 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1259 */ 1260 static int select_fallback_rq(int cpu, struct task_struct *p) 1261 { 1262 int nid = cpu_to_node(cpu); 1263 const struct cpumask *nodemask = NULL; 1264 enum { cpuset, possible, fail } state = cpuset; 1265 int dest_cpu; 1266 1267 /* 1268 * If the node that the cpu is on has been offlined, cpu_to_node() 1269 * will return -1. There is no cpu on the node, and we should 1270 * select the cpu on the other node. 1271 */ 1272 if (nid != -1) { 1273 nodemask = cpumask_of_node(nid); 1274 1275 /* Look for allowed, online CPU in same node. */ 1276 for_each_cpu(dest_cpu, nodemask) { 1277 if (!cpu_online(dest_cpu)) 1278 continue; 1279 if (!cpu_active(dest_cpu)) 1280 continue; 1281 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1282 return dest_cpu; 1283 } 1284 } 1285 1286 for (;;) { 1287 /* Any allowed, online CPU? */ 1288 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1289 if (!cpu_online(dest_cpu)) 1290 continue; 1291 if (!cpu_active(dest_cpu)) 1292 continue; 1293 goto out; 1294 } 1295 1296 switch (state) { 1297 case cpuset: 1298 /* No more Mr. Nice Guy. */ 1299 cpuset_cpus_allowed_fallback(p); 1300 state = possible; 1301 break; 1302 1303 case possible: 1304 do_set_cpus_allowed(p, cpu_possible_mask); 1305 state = fail; 1306 break; 1307 1308 case fail: 1309 BUG(); 1310 break; 1311 } 1312 } 1313 1314 out: 1315 if (state != cpuset) { 1316 /* 1317 * Don't tell them about moving exiting tasks or 1318 * kernel threads (both mm NULL), since they never 1319 * leave kernel. 1320 */ 1321 if (p->mm && printk_ratelimit()) { 1322 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1323 task_pid_nr(p), p->comm, cpu); 1324 } 1325 } 1326 1327 return dest_cpu; 1328 } 1329 1330 /* 1331 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1332 */ 1333 static inline 1334 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1335 { 1336 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1337 1338 /* 1339 * In order not to call set_task_cpu() on a blocking task we need 1340 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1341 * cpu. 1342 * 1343 * Since this is common to all placement strategies, this lives here. 1344 * 1345 * [ this allows ->select_task() to simply return task_cpu(p) and 1346 * not worry about this generic constraint ] 1347 */ 1348 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1349 !cpu_online(cpu))) 1350 cpu = select_fallback_rq(task_cpu(p), p); 1351 1352 return cpu; 1353 } 1354 1355 static void update_avg(u64 *avg, u64 sample) 1356 { 1357 s64 diff = sample - *avg; 1358 *avg += diff >> 3; 1359 } 1360 #endif 1361 1362 static void 1363 ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1364 { 1365 #ifdef CONFIG_SCHEDSTATS 1366 struct rq *rq = this_rq(); 1367 1368 #ifdef CONFIG_SMP 1369 int this_cpu = smp_processor_id(); 1370 1371 if (cpu == this_cpu) { 1372 schedstat_inc(rq, ttwu_local); 1373 schedstat_inc(p, se.statistics.nr_wakeups_local); 1374 } else { 1375 struct sched_domain *sd; 1376 1377 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1378 rcu_read_lock(); 1379 for_each_domain(this_cpu, sd) { 1380 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1381 schedstat_inc(sd, ttwu_wake_remote); 1382 break; 1383 } 1384 } 1385 rcu_read_unlock(); 1386 } 1387 1388 if (wake_flags & WF_MIGRATED) 1389 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1390 1391 #endif /* CONFIG_SMP */ 1392 1393 schedstat_inc(rq, ttwu_count); 1394 schedstat_inc(p, se.statistics.nr_wakeups); 1395 1396 if (wake_flags & WF_SYNC) 1397 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1398 1399 #endif /* CONFIG_SCHEDSTATS */ 1400 } 1401 1402 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1403 { 1404 activate_task(rq, p, en_flags); 1405 p->on_rq = 1; 1406 1407 /* if a worker is waking up, notify workqueue */ 1408 if (p->flags & PF_WQ_WORKER) 1409 wq_worker_waking_up(p, cpu_of(rq)); 1410 } 1411 1412 /* 1413 * Mark the task runnable and perform wakeup-preemption. 1414 */ 1415 static void 1416 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1417 { 1418 check_preempt_curr(rq, p, wake_flags); 1419 trace_sched_wakeup(p, true); 1420 1421 p->state = TASK_RUNNING; 1422 #ifdef CONFIG_SMP 1423 if (p->sched_class->task_woken) 1424 p->sched_class->task_woken(rq, p); 1425 1426 if (rq->idle_stamp) { 1427 u64 delta = rq_clock(rq) - rq->idle_stamp; 1428 u64 max = 2*rq->max_idle_balance_cost; 1429 1430 update_avg(&rq->avg_idle, delta); 1431 1432 if (rq->avg_idle > max) 1433 rq->avg_idle = max; 1434 1435 rq->idle_stamp = 0; 1436 } 1437 #endif 1438 } 1439 1440 static void 1441 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1442 { 1443 #ifdef CONFIG_SMP 1444 if (p->sched_contributes_to_load) 1445 rq->nr_uninterruptible--; 1446 #endif 1447 1448 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1449 ttwu_do_wakeup(rq, p, wake_flags); 1450 } 1451 1452 /* 1453 * Called in case the task @p isn't fully descheduled from its runqueue, 1454 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1455 * since all we need to do is flip p->state to TASK_RUNNING, since 1456 * the task is still ->on_rq. 1457 */ 1458 static int ttwu_remote(struct task_struct *p, int wake_flags) 1459 { 1460 struct rq *rq; 1461 int ret = 0; 1462 1463 rq = __task_rq_lock(p); 1464 if (p->on_rq) { 1465 /* check_preempt_curr() may use rq clock */ 1466 update_rq_clock(rq); 1467 ttwu_do_wakeup(rq, p, wake_flags); 1468 ret = 1; 1469 } 1470 __task_rq_unlock(rq); 1471 1472 return ret; 1473 } 1474 1475 #ifdef CONFIG_SMP 1476 static void sched_ttwu_pending(void) 1477 { 1478 struct rq *rq = this_rq(); 1479 struct llist_node *llist = llist_del_all(&rq->wake_list); 1480 struct task_struct *p; 1481 1482 raw_spin_lock(&rq->lock); 1483 1484 while (llist) { 1485 p = llist_entry(llist, struct task_struct, wake_entry); 1486 llist = llist_next(llist); 1487 ttwu_do_activate(rq, p, 0); 1488 } 1489 1490 raw_spin_unlock(&rq->lock); 1491 } 1492 1493 void scheduler_ipi(void) 1494 { 1495 /* 1496 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1497 * TIF_NEED_RESCHED remotely (for the first time) will also send 1498 * this IPI. 1499 */ 1500 preempt_fold_need_resched(); 1501 1502 if (llist_empty(&this_rq()->wake_list) 1503 && !tick_nohz_full_cpu(smp_processor_id()) 1504 && !got_nohz_idle_kick()) 1505 return; 1506 1507 /* 1508 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1509 * traditionally all their work was done from the interrupt return 1510 * path. Now that we actually do some work, we need to make sure 1511 * we do call them. 1512 * 1513 * Some archs already do call them, luckily irq_enter/exit nest 1514 * properly. 1515 * 1516 * Arguably we should visit all archs and update all handlers, 1517 * however a fair share of IPIs are still resched only so this would 1518 * somewhat pessimize the simple resched case. 1519 */ 1520 irq_enter(); 1521 tick_nohz_full_check(); 1522 sched_ttwu_pending(); 1523 1524 /* 1525 * Check if someone kicked us for doing the nohz idle load balance. 1526 */ 1527 if (unlikely(got_nohz_idle_kick())) { 1528 this_rq()->idle_balance = 1; 1529 raise_softirq_irqoff(SCHED_SOFTIRQ); 1530 } 1531 irq_exit(); 1532 } 1533 1534 static void ttwu_queue_remote(struct task_struct *p, int cpu) 1535 { 1536 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1537 smp_send_reschedule(cpu); 1538 } 1539 1540 bool cpus_share_cache(int this_cpu, int that_cpu) 1541 { 1542 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1543 } 1544 #endif /* CONFIG_SMP */ 1545 1546 static void ttwu_queue(struct task_struct *p, int cpu) 1547 { 1548 struct rq *rq = cpu_rq(cpu); 1549 1550 #if defined(CONFIG_SMP) 1551 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1552 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1553 ttwu_queue_remote(p, cpu); 1554 return; 1555 } 1556 #endif 1557 1558 raw_spin_lock(&rq->lock); 1559 ttwu_do_activate(rq, p, 0); 1560 raw_spin_unlock(&rq->lock); 1561 } 1562 1563 /** 1564 * try_to_wake_up - wake up a thread 1565 * @p: the thread to be awakened 1566 * @state: the mask of task states that can be woken 1567 * @wake_flags: wake modifier flags (WF_*) 1568 * 1569 * Put it on the run-queue if it's not already there. The "current" 1570 * thread is always on the run-queue (except when the actual 1571 * re-schedule is in progress), and as such you're allowed to do 1572 * the simpler "current->state = TASK_RUNNING" to mark yourself 1573 * runnable without the overhead of this. 1574 * 1575 * Return: %true if @p was woken up, %false if it was already running. 1576 * or @state didn't match @p's state. 1577 */ 1578 static int 1579 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1580 { 1581 unsigned long flags; 1582 int cpu, success = 0; 1583 1584 /* 1585 * If we are going to wake up a thread waiting for CONDITION we 1586 * need to ensure that CONDITION=1 done by the caller can not be 1587 * reordered with p->state check below. This pairs with mb() in 1588 * set_current_state() the waiting thread does. 1589 */ 1590 smp_mb__before_spinlock(); 1591 raw_spin_lock_irqsave(&p->pi_lock, flags); 1592 if (!(p->state & state)) 1593 goto out; 1594 1595 success = 1; /* we're going to change ->state */ 1596 cpu = task_cpu(p); 1597 1598 if (p->on_rq && ttwu_remote(p, wake_flags)) 1599 goto stat; 1600 1601 #ifdef CONFIG_SMP 1602 /* 1603 * If the owning (remote) cpu is still in the middle of schedule() with 1604 * this task as prev, wait until its done referencing the task. 1605 */ 1606 while (p->on_cpu) 1607 cpu_relax(); 1608 /* 1609 * Pairs with the smp_wmb() in finish_lock_switch(). 1610 */ 1611 smp_rmb(); 1612 1613 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1614 p->state = TASK_WAKING; 1615 1616 if (p->sched_class->task_waking) 1617 p->sched_class->task_waking(p); 1618 1619 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1620 if (task_cpu(p) != cpu) { 1621 wake_flags |= WF_MIGRATED; 1622 set_task_cpu(p, cpu); 1623 } 1624 #endif /* CONFIG_SMP */ 1625 1626 ttwu_queue(p, cpu); 1627 stat: 1628 ttwu_stat(p, cpu, wake_flags); 1629 out: 1630 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1631 1632 return success; 1633 } 1634 1635 /** 1636 * try_to_wake_up_local - try to wake up a local task with rq lock held 1637 * @p: the thread to be awakened 1638 * 1639 * Put @p on the run-queue if it's not already there. The caller must 1640 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1641 * the current task. 1642 */ 1643 static void try_to_wake_up_local(struct task_struct *p) 1644 { 1645 struct rq *rq = task_rq(p); 1646 1647 if (WARN_ON_ONCE(rq != this_rq()) || 1648 WARN_ON_ONCE(p == current)) 1649 return; 1650 1651 lockdep_assert_held(&rq->lock); 1652 1653 if (!raw_spin_trylock(&p->pi_lock)) { 1654 raw_spin_unlock(&rq->lock); 1655 raw_spin_lock(&p->pi_lock); 1656 raw_spin_lock(&rq->lock); 1657 } 1658 1659 if (!(p->state & TASK_NORMAL)) 1660 goto out; 1661 1662 if (!p->on_rq) 1663 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1664 1665 ttwu_do_wakeup(rq, p, 0); 1666 ttwu_stat(p, smp_processor_id(), 0); 1667 out: 1668 raw_spin_unlock(&p->pi_lock); 1669 } 1670 1671 /** 1672 * wake_up_process - Wake up a specific process 1673 * @p: The process to be woken up. 1674 * 1675 * Attempt to wake up the nominated process and move it to the set of runnable 1676 * processes. 1677 * 1678 * Return: 1 if the process was woken up, 0 if it was already running. 1679 * 1680 * It may be assumed that this function implies a write memory barrier before 1681 * changing the task state if and only if any tasks are woken up. 1682 */ 1683 int wake_up_process(struct task_struct *p) 1684 { 1685 WARN_ON(task_is_stopped_or_traced(p)); 1686 return try_to_wake_up(p, TASK_NORMAL, 0); 1687 } 1688 EXPORT_SYMBOL(wake_up_process); 1689 1690 int wake_up_state(struct task_struct *p, unsigned int state) 1691 { 1692 return try_to_wake_up(p, state, 0); 1693 } 1694 1695 /* 1696 * Perform scheduler related setup for a newly forked process p. 1697 * p is forked by current. 1698 * 1699 * __sched_fork() is basic setup used by init_idle() too: 1700 */ 1701 static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1702 { 1703 p->on_rq = 0; 1704 1705 p->se.on_rq = 0; 1706 p->se.exec_start = 0; 1707 p->se.sum_exec_runtime = 0; 1708 p->se.prev_sum_exec_runtime = 0; 1709 p->se.nr_migrations = 0; 1710 p->se.vruntime = 0; 1711 INIT_LIST_HEAD(&p->se.group_node); 1712 1713 #ifdef CONFIG_SCHEDSTATS 1714 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1715 #endif 1716 1717 RB_CLEAR_NODE(&p->dl.rb_node); 1718 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1719 p->dl.dl_runtime = p->dl.runtime = 0; 1720 p->dl.dl_deadline = p->dl.deadline = 0; 1721 p->dl.dl_period = 0; 1722 p->dl.flags = 0; 1723 1724 INIT_LIST_HEAD(&p->rt.run_list); 1725 1726 #ifdef CONFIG_PREEMPT_NOTIFIERS 1727 INIT_HLIST_HEAD(&p->preempt_notifiers); 1728 #endif 1729 1730 #ifdef CONFIG_NUMA_BALANCING 1731 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1732 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1733 p->mm->numa_scan_seq = 0; 1734 } 1735 1736 if (clone_flags & CLONE_VM) 1737 p->numa_preferred_nid = current->numa_preferred_nid; 1738 else 1739 p->numa_preferred_nid = -1; 1740 1741 p->node_stamp = 0ULL; 1742 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1743 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1744 p->numa_work.next = &p->numa_work; 1745 p->numa_faults_memory = NULL; 1746 p->numa_faults_buffer_memory = NULL; 1747 p->last_task_numa_placement = 0; 1748 p->last_sum_exec_runtime = 0; 1749 1750 INIT_LIST_HEAD(&p->numa_entry); 1751 p->numa_group = NULL; 1752 #endif /* CONFIG_NUMA_BALANCING */ 1753 } 1754 1755 #ifdef CONFIG_NUMA_BALANCING 1756 #ifdef CONFIG_SCHED_DEBUG 1757 void set_numabalancing_state(bool enabled) 1758 { 1759 if (enabled) 1760 sched_feat_set("NUMA"); 1761 else 1762 sched_feat_set("NO_NUMA"); 1763 } 1764 #else 1765 __read_mostly bool numabalancing_enabled; 1766 1767 void set_numabalancing_state(bool enabled) 1768 { 1769 numabalancing_enabled = enabled; 1770 } 1771 #endif /* CONFIG_SCHED_DEBUG */ 1772 1773 #ifdef CONFIG_PROC_SYSCTL 1774 int sysctl_numa_balancing(struct ctl_table *table, int write, 1775 void __user *buffer, size_t *lenp, loff_t *ppos) 1776 { 1777 struct ctl_table t; 1778 int err; 1779 int state = numabalancing_enabled; 1780 1781 if (write && !capable(CAP_SYS_ADMIN)) 1782 return -EPERM; 1783 1784 t = *table; 1785 t.data = &state; 1786 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1787 if (err < 0) 1788 return err; 1789 if (write) 1790 set_numabalancing_state(state); 1791 return err; 1792 } 1793 #endif 1794 #endif 1795 1796 /* 1797 * fork()/clone()-time setup: 1798 */ 1799 int sched_fork(unsigned long clone_flags, struct task_struct *p) 1800 { 1801 unsigned long flags; 1802 int cpu = get_cpu(); 1803 1804 __sched_fork(clone_flags, p); 1805 /* 1806 * We mark the process as running here. This guarantees that 1807 * nobody will actually run it, and a signal or other external 1808 * event cannot wake it up and insert it on the runqueue either. 1809 */ 1810 p->state = TASK_RUNNING; 1811 1812 /* 1813 * Make sure we do not leak PI boosting priority to the child. 1814 */ 1815 p->prio = current->normal_prio; 1816 1817 /* 1818 * Revert to default priority/policy on fork if requested. 1819 */ 1820 if (unlikely(p->sched_reset_on_fork)) { 1821 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1822 p->policy = SCHED_NORMAL; 1823 p->static_prio = NICE_TO_PRIO(0); 1824 p->rt_priority = 0; 1825 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1826 p->static_prio = NICE_TO_PRIO(0); 1827 1828 p->prio = p->normal_prio = __normal_prio(p); 1829 set_load_weight(p); 1830 1831 /* 1832 * We don't need the reset flag anymore after the fork. It has 1833 * fulfilled its duty: 1834 */ 1835 p->sched_reset_on_fork = 0; 1836 } 1837 1838 if (dl_prio(p->prio)) { 1839 put_cpu(); 1840 return -EAGAIN; 1841 } else if (rt_prio(p->prio)) { 1842 p->sched_class = &rt_sched_class; 1843 } else { 1844 p->sched_class = &fair_sched_class; 1845 } 1846 1847 if (p->sched_class->task_fork) 1848 p->sched_class->task_fork(p); 1849 1850 /* 1851 * The child is not yet in the pid-hash so no cgroup attach races, 1852 * and the cgroup is pinned to this child due to cgroup_fork() 1853 * is ran before sched_fork(). 1854 * 1855 * Silence PROVE_RCU. 1856 */ 1857 raw_spin_lock_irqsave(&p->pi_lock, flags); 1858 set_task_cpu(p, cpu); 1859 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1860 1861 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1862 if (likely(sched_info_on())) 1863 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1864 #endif 1865 #if defined(CONFIG_SMP) 1866 p->on_cpu = 0; 1867 #endif 1868 init_task_preempt_count(p); 1869 #ifdef CONFIG_SMP 1870 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1871 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1872 #endif 1873 1874 put_cpu(); 1875 return 0; 1876 } 1877 1878 unsigned long to_ratio(u64 period, u64 runtime) 1879 { 1880 if (runtime == RUNTIME_INF) 1881 return 1ULL << 20; 1882 1883 /* 1884 * Doing this here saves a lot of checks in all 1885 * the calling paths, and returning zero seems 1886 * safe for them anyway. 1887 */ 1888 if (period == 0) 1889 return 0; 1890 1891 return div64_u64(runtime << 20, period); 1892 } 1893 1894 #ifdef CONFIG_SMP 1895 inline struct dl_bw *dl_bw_of(int i) 1896 { 1897 return &cpu_rq(i)->rd->dl_bw; 1898 } 1899 1900 static inline int dl_bw_cpus(int i) 1901 { 1902 struct root_domain *rd = cpu_rq(i)->rd; 1903 int cpus = 0; 1904 1905 for_each_cpu_and(i, rd->span, cpu_active_mask) 1906 cpus++; 1907 1908 return cpus; 1909 } 1910 #else 1911 inline struct dl_bw *dl_bw_of(int i) 1912 { 1913 return &cpu_rq(i)->dl.dl_bw; 1914 } 1915 1916 static inline int dl_bw_cpus(int i) 1917 { 1918 return 1; 1919 } 1920 #endif 1921 1922 static inline 1923 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1924 { 1925 dl_b->total_bw -= tsk_bw; 1926 } 1927 1928 static inline 1929 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1930 { 1931 dl_b->total_bw += tsk_bw; 1932 } 1933 1934 static inline 1935 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1936 { 1937 return dl_b->bw != -1 && 1938 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1939 } 1940 1941 /* 1942 * We must be sure that accepting a new task (or allowing changing the 1943 * parameters of an existing one) is consistent with the bandwidth 1944 * constraints. If yes, this function also accordingly updates the currently 1945 * allocated bandwidth to reflect the new situation. 1946 * 1947 * This function is called while holding p's rq->lock. 1948 */ 1949 static int dl_overflow(struct task_struct *p, int policy, 1950 const struct sched_attr *attr) 1951 { 1952 1953 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1954 u64 period = attr->sched_period ?: attr->sched_deadline; 1955 u64 runtime = attr->sched_runtime; 1956 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1957 int cpus, err = -1; 1958 1959 if (new_bw == p->dl.dl_bw) 1960 return 0; 1961 1962 /* 1963 * Either if a task, enters, leave, or stays -deadline but changes 1964 * its parameters, we may need to update accordingly the total 1965 * allocated bandwidth of the container. 1966 */ 1967 raw_spin_lock(&dl_b->lock); 1968 cpus = dl_bw_cpus(task_cpu(p)); 1969 if (dl_policy(policy) && !task_has_dl_policy(p) && 1970 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1971 __dl_add(dl_b, new_bw); 1972 err = 0; 1973 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1974 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1975 __dl_clear(dl_b, p->dl.dl_bw); 1976 __dl_add(dl_b, new_bw); 1977 err = 0; 1978 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1979 __dl_clear(dl_b, p->dl.dl_bw); 1980 err = 0; 1981 } 1982 raw_spin_unlock(&dl_b->lock); 1983 1984 return err; 1985 } 1986 1987 extern void init_dl_bw(struct dl_bw *dl_b); 1988 1989 /* 1990 * wake_up_new_task - wake up a newly created task for the first time. 1991 * 1992 * This function will do some initial scheduler statistics housekeeping 1993 * that must be done for every newly created context, then puts the task 1994 * on the runqueue and wakes it. 1995 */ 1996 void wake_up_new_task(struct task_struct *p) 1997 { 1998 unsigned long flags; 1999 struct rq *rq; 2000 2001 raw_spin_lock_irqsave(&p->pi_lock, flags); 2002 #ifdef CONFIG_SMP 2003 /* 2004 * Fork balancing, do it here and not earlier because: 2005 * - cpus_allowed can change in the fork path 2006 * - any previously selected cpu might disappear through hotplug 2007 */ 2008 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2009 #endif 2010 2011 /* Initialize new task's runnable average */ 2012 init_task_runnable_average(p); 2013 rq = __task_rq_lock(p); 2014 activate_task(rq, p, 0); 2015 p->on_rq = 1; 2016 trace_sched_wakeup_new(p, true); 2017 check_preempt_curr(rq, p, WF_FORK); 2018 #ifdef CONFIG_SMP 2019 if (p->sched_class->task_woken) 2020 p->sched_class->task_woken(rq, p); 2021 #endif 2022 task_rq_unlock(rq, p, &flags); 2023 } 2024 2025 #ifdef CONFIG_PREEMPT_NOTIFIERS 2026 2027 /** 2028 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2029 * @notifier: notifier struct to register 2030 */ 2031 void preempt_notifier_register(struct preempt_notifier *notifier) 2032 { 2033 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2034 } 2035 EXPORT_SYMBOL_GPL(preempt_notifier_register); 2036 2037 /** 2038 * preempt_notifier_unregister - no longer interested in preemption notifications 2039 * @notifier: notifier struct to unregister 2040 * 2041 * This is safe to call from within a preemption notifier. 2042 */ 2043 void preempt_notifier_unregister(struct preempt_notifier *notifier) 2044 { 2045 hlist_del(¬ifier->link); 2046 } 2047 EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2048 2049 static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2050 { 2051 struct preempt_notifier *notifier; 2052 2053 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2054 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2055 } 2056 2057 static void 2058 fire_sched_out_preempt_notifiers(struct task_struct *curr, 2059 struct task_struct *next) 2060 { 2061 struct preempt_notifier *notifier; 2062 2063 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2064 notifier->ops->sched_out(notifier, next); 2065 } 2066 2067 #else /* !CONFIG_PREEMPT_NOTIFIERS */ 2068 2069 static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2070 { 2071 } 2072 2073 static void 2074 fire_sched_out_preempt_notifiers(struct task_struct *curr, 2075 struct task_struct *next) 2076 { 2077 } 2078 2079 #endif /* CONFIG_PREEMPT_NOTIFIERS */ 2080 2081 /** 2082 * prepare_task_switch - prepare to switch tasks 2083 * @rq: the runqueue preparing to switch 2084 * @prev: the current task that is being switched out 2085 * @next: the task we are going to switch to. 2086 * 2087 * This is called with the rq lock held and interrupts off. It must 2088 * be paired with a subsequent finish_task_switch after the context 2089 * switch. 2090 * 2091 * prepare_task_switch sets up locking and calls architecture specific 2092 * hooks. 2093 */ 2094 static inline void 2095 prepare_task_switch(struct rq *rq, struct task_struct *prev, 2096 struct task_struct *next) 2097 { 2098 trace_sched_switch(prev, next); 2099 sched_info_switch(rq, prev, next); 2100 perf_event_task_sched_out(prev, next); 2101 fire_sched_out_preempt_notifiers(prev, next); 2102 prepare_lock_switch(rq, next); 2103 prepare_arch_switch(next); 2104 } 2105 2106 /** 2107 * finish_task_switch - clean up after a task-switch 2108 * @rq: runqueue associated with task-switch 2109 * @prev: the thread we just switched away from. 2110 * 2111 * finish_task_switch must be called after the context switch, paired 2112 * with a prepare_task_switch call before the context switch. 2113 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2114 * and do any other architecture-specific cleanup actions. 2115 * 2116 * Note that we may have delayed dropping an mm in context_switch(). If 2117 * so, we finish that here outside of the runqueue lock. (Doing it 2118 * with the lock held can cause deadlocks; see schedule() for 2119 * details.) 2120 */ 2121 static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2122 __releases(rq->lock) 2123 { 2124 struct mm_struct *mm = rq->prev_mm; 2125 long prev_state; 2126 2127 rq->prev_mm = NULL; 2128 2129 /* 2130 * A task struct has one reference for the use as "current". 2131 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2132 * schedule one last time. The schedule call will never return, and 2133 * the scheduled task must drop that reference. 2134 * The test for TASK_DEAD must occur while the runqueue locks are 2135 * still held, otherwise prev could be scheduled on another cpu, die 2136 * there before we look at prev->state, and then the reference would 2137 * be dropped twice. 2138 * Manfred Spraul <manfred@colorfullife.com> 2139 */ 2140 prev_state = prev->state; 2141 vtime_task_switch(prev); 2142 finish_arch_switch(prev); 2143 perf_event_task_sched_in(prev, current); 2144 finish_lock_switch(rq, prev); 2145 finish_arch_post_lock_switch(); 2146 2147 fire_sched_in_preempt_notifiers(current); 2148 if (mm) 2149 mmdrop(mm); 2150 if (unlikely(prev_state == TASK_DEAD)) { 2151 if (prev->sched_class->task_dead) 2152 prev->sched_class->task_dead(prev); 2153 2154 /* 2155 * Remove function-return probe instances associated with this 2156 * task and put them back on the free list. 2157 */ 2158 kprobe_flush_task(prev); 2159 put_task_struct(prev); 2160 } 2161 2162 tick_nohz_task_switch(current); 2163 } 2164 2165 #ifdef CONFIG_SMP 2166 2167 /* rq->lock is NOT held, but preemption is disabled */ 2168 static inline void post_schedule(struct rq *rq) 2169 { 2170 if (rq->post_schedule) { 2171 unsigned long flags; 2172 2173 raw_spin_lock_irqsave(&rq->lock, flags); 2174 if (rq->curr->sched_class->post_schedule) 2175 rq->curr->sched_class->post_schedule(rq); 2176 raw_spin_unlock_irqrestore(&rq->lock, flags); 2177 2178 rq->post_schedule = 0; 2179 } 2180 } 2181 2182 #else 2183 2184 static inline void post_schedule(struct rq *rq) 2185 { 2186 } 2187 2188 #endif 2189 2190 /** 2191 * schedule_tail - first thing a freshly forked thread must call. 2192 * @prev: the thread we just switched away from. 2193 */ 2194 asmlinkage void schedule_tail(struct task_struct *prev) 2195 __releases(rq->lock) 2196 { 2197 struct rq *rq = this_rq(); 2198 2199 finish_task_switch(rq, prev); 2200 2201 /* 2202 * FIXME: do we need to worry about rq being invalidated by the 2203 * task_switch? 2204 */ 2205 post_schedule(rq); 2206 2207 #ifdef __ARCH_WANT_UNLOCKED_CTXSW 2208 /* In this case, finish_task_switch does not reenable preemption */ 2209 preempt_enable(); 2210 #endif 2211 if (current->set_child_tid) 2212 put_user(task_pid_vnr(current), current->set_child_tid); 2213 } 2214 2215 /* 2216 * context_switch - switch to the new MM and the new 2217 * thread's register state. 2218 */ 2219 static inline void 2220 context_switch(struct rq *rq, struct task_struct *prev, 2221 struct task_struct *next) 2222 { 2223 struct mm_struct *mm, *oldmm; 2224 2225 prepare_task_switch(rq, prev, next); 2226 2227 mm = next->mm; 2228 oldmm = prev->active_mm; 2229 /* 2230 * For paravirt, this is coupled with an exit in switch_to to 2231 * combine the page table reload and the switch backend into 2232 * one hypercall. 2233 */ 2234 arch_start_context_switch(prev); 2235 2236 if (!mm) { 2237 next->active_mm = oldmm; 2238 atomic_inc(&oldmm->mm_count); 2239 enter_lazy_tlb(oldmm, next); 2240 } else 2241 switch_mm(oldmm, mm, next); 2242 2243 if (!prev->mm) { 2244 prev->active_mm = NULL; 2245 rq->prev_mm = oldmm; 2246 } 2247 /* 2248 * Since the runqueue lock will be released by the next 2249 * task (which is an invalid locking op but in the case 2250 * of the scheduler it's an obvious special-case), so we 2251 * do an early lockdep release here: 2252 */ 2253 #ifndef __ARCH_WANT_UNLOCKED_CTXSW 2254 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2255 #endif 2256 2257 context_tracking_task_switch(prev, next); 2258 /* Here we just switch the register state and the stack. */ 2259 switch_to(prev, next, prev); 2260 2261 barrier(); 2262 /* 2263 * this_rq must be evaluated again because prev may have moved 2264 * CPUs since it called schedule(), thus the 'rq' on its stack 2265 * frame will be invalid. 2266 */ 2267 finish_task_switch(this_rq(), prev); 2268 } 2269 2270 /* 2271 * nr_running and nr_context_switches: 2272 * 2273 * externally visible scheduler statistics: current number of runnable 2274 * threads, total number of context switches performed since bootup. 2275 */ 2276 unsigned long nr_running(void) 2277 { 2278 unsigned long i, sum = 0; 2279 2280 for_each_online_cpu(i) 2281 sum += cpu_rq(i)->nr_running; 2282 2283 return sum; 2284 } 2285 2286 unsigned long long nr_context_switches(void) 2287 { 2288 int i; 2289 unsigned long long sum = 0; 2290 2291 for_each_possible_cpu(i) 2292 sum += cpu_rq(i)->nr_switches; 2293 2294 return sum; 2295 } 2296 2297 unsigned long nr_iowait(void) 2298 { 2299 unsigned long i, sum = 0; 2300 2301 for_each_possible_cpu(i) 2302 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2303 2304 return sum; 2305 } 2306 2307 unsigned long nr_iowait_cpu(int cpu) 2308 { 2309 struct rq *this = cpu_rq(cpu); 2310 return atomic_read(&this->nr_iowait); 2311 } 2312 2313 #ifdef CONFIG_SMP 2314 2315 /* 2316 * sched_exec - execve() is a valuable balancing opportunity, because at 2317 * this point the task has the smallest effective memory and cache footprint. 2318 */ 2319 void sched_exec(void) 2320 { 2321 struct task_struct *p = current; 2322 unsigned long flags; 2323 int dest_cpu; 2324 2325 raw_spin_lock_irqsave(&p->pi_lock, flags); 2326 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2327 if (dest_cpu == smp_processor_id()) 2328 goto unlock; 2329 2330 if (likely(cpu_active(dest_cpu))) { 2331 struct migration_arg arg = { p, dest_cpu }; 2332 2333 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2334 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2335 return; 2336 } 2337 unlock: 2338 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2339 } 2340 2341 #endif 2342 2343 DEFINE_PER_CPU(struct kernel_stat, kstat); 2344 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2345 2346 EXPORT_PER_CPU_SYMBOL(kstat); 2347 EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2348 2349 /* 2350 * Return any ns on the sched_clock that have not yet been accounted in 2351 * @p in case that task is currently running. 2352 * 2353 * Called with task_rq_lock() held on @rq. 2354 */ 2355 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2356 { 2357 u64 ns = 0; 2358 2359 if (task_current(rq, p)) { 2360 update_rq_clock(rq); 2361 ns = rq_clock_task(rq) - p->se.exec_start; 2362 if ((s64)ns < 0) 2363 ns = 0; 2364 } 2365 2366 return ns; 2367 } 2368 2369 unsigned long long task_delta_exec(struct task_struct *p) 2370 { 2371 unsigned long flags; 2372 struct rq *rq; 2373 u64 ns = 0; 2374 2375 rq = task_rq_lock(p, &flags); 2376 ns = do_task_delta_exec(p, rq); 2377 task_rq_unlock(rq, p, &flags); 2378 2379 return ns; 2380 } 2381 2382 /* 2383 * Return accounted runtime for the task. 2384 * In case the task is currently running, return the runtime plus current's 2385 * pending runtime that have not been accounted yet. 2386 */ 2387 unsigned long long task_sched_runtime(struct task_struct *p) 2388 { 2389 unsigned long flags; 2390 struct rq *rq; 2391 u64 ns = 0; 2392 2393 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2394 /* 2395 * 64-bit doesn't need locks to atomically read a 64bit value. 2396 * So we have a optimization chance when the task's delta_exec is 0. 2397 * Reading ->on_cpu is racy, but this is ok. 2398 * 2399 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2400 * If we race with it entering cpu, unaccounted time is 0. This is 2401 * indistinguishable from the read occurring a few cycles earlier. 2402 */ 2403 if (!p->on_cpu) 2404 return p->se.sum_exec_runtime; 2405 #endif 2406 2407 rq = task_rq_lock(p, &flags); 2408 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2409 task_rq_unlock(rq, p, &flags); 2410 2411 return ns; 2412 } 2413 2414 /* 2415 * This function gets called by the timer code, with HZ frequency. 2416 * We call it with interrupts disabled. 2417 */ 2418 void scheduler_tick(void) 2419 { 2420 int cpu = smp_processor_id(); 2421 struct rq *rq = cpu_rq(cpu); 2422 struct task_struct *curr = rq->curr; 2423 2424 sched_clock_tick(); 2425 2426 raw_spin_lock(&rq->lock); 2427 update_rq_clock(rq); 2428 curr->sched_class->task_tick(rq, curr, 0); 2429 update_cpu_load_active(rq); 2430 raw_spin_unlock(&rq->lock); 2431 2432 perf_event_task_tick(); 2433 2434 #ifdef CONFIG_SMP 2435 rq->idle_balance = idle_cpu(cpu); 2436 trigger_load_balance(rq); 2437 #endif 2438 rq_last_tick_reset(rq); 2439 } 2440 2441 #ifdef CONFIG_NO_HZ_FULL 2442 /** 2443 * scheduler_tick_max_deferment 2444 * 2445 * Keep at least one tick per second when a single 2446 * active task is running because the scheduler doesn't 2447 * yet completely support full dynticks environment. 2448 * 2449 * This makes sure that uptime, CFS vruntime, load 2450 * balancing, etc... continue to move forward, even 2451 * with a very low granularity. 2452 * 2453 * Return: Maximum deferment in nanoseconds. 2454 */ 2455 u64 scheduler_tick_max_deferment(void) 2456 { 2457 struct rq *rq = this_rq(); 2458 unsigned long next, now = ACCESS_ONCE(jiffies); 2459 2460 next = rq->last_sched_tick + HZ; 2461 2462 if (time_before_eq(next, now)) 2463 return 0; 2464 2465 return jiffies_to_nsecs(next - now); 2466 } 2467 #endif 2468 2469 notrace unsigned long get_parent_ip(unsigned long addr) 2470 { 2471 if (in_lock_functions(addr)) { 2472 addr = CALLER_ADDR2; 2473 if (in_lock_functions(addr)) 2474 addr = CALLER_ADDR3; 2475 } 2476 return addr; 2477 } 2478 2479 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2480 defined(CONFIG_PREEMPT_TRACER)) 2481 2482 void __kprobes preempt_count_add(int val) 2483 { 2484 #ifdef CONFIG_DEBUG_PREEMPT 2485 /* 2486 * Underflow? 2487 */ 2488 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2489 return; 2490 #endif 2491 __preempt_count_add(val); 2492 #ifdef CONFIG_DEBUG_PREEMPT 2493 /* 2494 * Spinlock count overflowing soon? 2495 */ 2496 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2497 PREEMPT_MASK - 10); 2498 #endif 2499 if (preempt_count() == val) { 2500 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2501 #ifdef CONFIG_DEBUG_PREEMPT 2502 current->preempt_disable_ip = ip; 2503 #endif 2504 trace_preempt_off(CALLER_ADDR0, ip); 2505 } 2506 } 2507 EXPORT_SYMBOL(preempt_count_add); 2508 2509 void __kprobes preempt_count_sub(int val) 2510 { 2511 #ifdef CONFIG_DEBUG_PREEMPT 2512 /* 2513 * Underflow? 2514 */ 2515 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2516 return; 2517 /* 2518 * Is the spinlock portion underflowing? 2519 */ 2520 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2521 !(preempt_count() & PREEMPT_MASK))) 2522 return; 2523 #endif 2524 2525 if (preempt_count() == val) 2526 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2527 __preempt_count_sub(val); 2528 } 2529 EXPORT_SYMBOL(preempt_count_sub); 2530 2531 #endif 2532 2533 /* 2534 * Print scheduling while atomic bug: 2535 */ 2536 static noinline void __schedule_bug(struct task_struct *prev) 2537 { 2538 if (oops_in_progress) 2539 return; 2540 2541 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2542 prev->comm, prev->pid, preempt_count()); 2543 2544 debug_show_held_locks(prev); 2545 print_modules(); 2546 if (irqs_disabled()) 2547 print_irqtrace_events(prev); 2548 #ifdef CONFIG_DEBUG_PREEMPT 2549 if (in_atomic_preempt_off()) { 2550 pr_err("Preemption disabled at:"); 2551 print_ip_sym(current->preempt_disable_ip); 2552 pr_cont("\n"); 2553 } 2554 #endif 2555 dump_stack(); 2556 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2557 } 2558 2559 /* 2560 * Various schedule()-time debugging checks and statistics: 2561 */ 2562 static inline void schedule_debug(struct task_struct *prev) 2563 { 2564 /* 2565 * Test if we are atomic. Since do_exit() needs to call into 2566 * schedule() atomically, we ignore that path. Otherwise whine 2567 * if we are scheduling when we should not. 2568 */ 2569 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2570 __schedule_bug(prev); 2571 rcu_sleep_check(); 2572 2573 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2574 2575 schedstat_inc(this_rq(), sched_count); 2576 } 2577 2578 /* 2579 * Pick up the highest-prio task: 2580 */ 2581 static inline struct task_struct * 2582 pick_next_task(struct rq *rq, struct task_struct *prev) 2583 { 2584 const struct sched_class *class = &fair_sched_class; 2585 struct task_struct *p; 2586 2587 /* 2588 * Optimization: we know that if all tasks are in 2589 * the fair class we can call that function directly: 2590 */ 2591 if (likely(prev->sched_class == class && 2592 rq->nr_running == rq->cfs.h_nr_running)) { 2593 p = fair_sched_class.pick_next_task(rq, prev); 2594 if (likely(p && p != RETRY_TASK)) 2595 return p; 2596 } 2597 2598 again: 2599 for_each_class(class) { 2600 p = class->pick_next_task(rq, prev); 2601 if (p) { 2602 if (unlikely(p == RETRY_TASK)) 2603 goto again; 2604 return p; 2605 } 2606 } 2607 2608 BUG(); /* the idle class will always have a runnable task */ 2609 } 2610 2611 /* 2612 * __schedule() is the main scheduler function. 2613 * 2614 * The main means of driving the scheduler and thus entering this function are: 2615 * 2616 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2617 * 2618 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2619 * paths. For example, see arch/x86/entry_64.S. 2620 * 2621 * To drive preemption between tasks, the scheduler sets the flag in timer 2622 * interrupt handler scheduler_tick(). 2623 * 2624 * 3. Wakeups don't really cause entry into schedule(). They add a 2625 * task to the run-queue and that's it. 2626 * 2627 * Now, if the new task added to the run-queue preempts the current 2628 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2629 * called on the nearest possible occasion: 2630 * 2631 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2632 * 2633 * - in syscall or exception context, at the next outmost 2634 * preempt_enable(). (this might be as soon as the wake_up()'s 2635 * spin_unlock()!) 2636 * 2637 * - in IRQ context, return from interrupt-handler to 2638 * preemptible context 2639 * 2640 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2641 * then at the next: 2642 * 2643 * - cond_resched() call 2644 * - explicit schedule() call 2645 * - return from syscall or exception to user-space 2646 * - return from interrupt-handler to user-space 2647 */ 2648 static void __sched __schedule(void) 2649 { 2650 struct task_struct *prev, *next; 2651 unsigned long *switch_count; 2652 struct rq *rq; 2653 int cpu; 2654 2655 need_resched: 2656 preempt_disable(); 2657 cpu = smp_processor_id(); 2658 rq = cpu_rq(cpu); 2659 rcu_note_context_switch(cpu); 2660 prev = rq->curr; 2661 2662 schedule_debug(prev); 2663 2664 if (sched_feat(HRTICK)) 2665 hrtick_clear(rq); 2666 2667 /* 2668 * Make sure that signal_pending_state()->signal_pending() below 2669 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2670 * done by the caller to avoid the race with signal_wake_up(). 2671 */ 2672 smp_mb__before_spinlock(); 2673 raw_spin_lock_irq(&rq->lock); 2674 2675 switch_count = &prev->nivcsw; 2676 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2677 if (unlikely(signal_pending_state(prev->state, prev))) { 2678 prev->state = TASK_RUNNING; 2679 } else { 2680 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2681 prev->on_rq = 0; 2682 2683 /* 2684 * If a worker went to sleep, notify and ask workqueue 2685 * whether it wants to wake up a task to maintain 2686 * concurrency. 2687 */ 2688 if (prev->flags & PF_WQ_WORKER) { 2689 struct task_struct *to_wakeup; 2690 2691 to_wakeup = wq_worker_sleeping(prev, cpu); 2692 if (to_wakeup) 2693 try_to_wake_up_local(to_wakeup); 2694 } 2695 } 2696 switch_count = &prev->nvcsw; 2697 } 2698 2699 if (prev->on_rq || rq->skip_clock_update < 0) 2700 update_rq_clock(rq); 2701 2702 next = pick_next_task(rq, prev); 2703 clear_tsk_need_resched(prev); 2704 clear_preempt_need_resched(); 2705 rq->skip_clock_update = 0; 2706 2707 if (likely(prev != next)) { 2708 rq->nr_switches++; 2709 rq->curr = next; 2710 ++*switch_count; 2711 2712 context_switch(rq, prev, next); /* unlocks the rq */ 2713 /* 2714 * The context switch have flipped the stack from under us 2715 * and restored the local variables which were saved when 2716 * this task called schedule() in the past. prev == current 2717 * is still correct, but it can be moved to another cpu/rq. 2718 */ 2719 cpu = smp_processor_id(); 2720 rq = cpu_rq(cpu); 2721 } else 2722 raw_spin_unlock_irq(&rq->lock); 2723 2724 post_schedule(rq); 2725 2726 sched_preempt_enable_no_resched(); 2727 if (need_resched()) 2728 goto need_resched; 2729 } 2730 2731 static inline void sched_submit_work(struct task_struct *tsk) 2732 { 2733 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2734 return; 2735 /* 2736 * If we are going to sleep and we have plugged IO queued, 2737 * make sure to submit it to avoid deadlocks. 2738 */ 2739 if (blk_needs_flush_plug(tsk)) 2740 blk_schedule_flush_plug(tsk); 2741 } 2742 2743 asmlinkage void __sched schedule(void) 2744 { 2745 struct task_struct *tsk = current; 2746 2747 sched_submit_work(tsk); 2748 __schedule(); 2749 } 2750 EXPORT_SYMBOL(schedule); 2751 2752 #ifdef CONFIG_CONTEXT_TRACKING 2753 asmlinkage void __sched schedule_user(void) 2754 { 2755 /* 2756 * If we come here after a random call to set_need_resched(), 2757 * or we have been woken up remotely but the IPI has not yet arrived, 2758 * we haven't yet exited the RCU idle mode. Do it here manually until 2759 * we find a better solution. 2760 */ 2761 user_exit(); 2762 schedule(); 2763 user_enter(); 2764 } 2765 #endif 2766 2767 /** 2768 * schedule_preempt_disabled - called with preemption disabled 2769 * 2770 * Returns with preemption disabled. Note: preempt_count must be 1 2771 */ 2772 void __sched schedule_preempt_disabled(void) 2773 { 2774 sched_preempt_enable_no_resched(); 2775 schedule(); 2776 preempt_disable(); 2777 } 2778 2779 #ifdef CONFIG_PREEMPT 2780 /* 2781 * this is the entry point to schedule() from in-kernel preemption 2782 * off of preempt_enable. Kernel preemptions off return from interrupt 2783 * occur there and call schedule directly. 2784 */ 2785 asmlinkage void __sched notrace preempt_schedule(void) 2786 { 2787 /* 2788 * If there is a non-zero preempt_count or interrupts are disabled, 2789 * we do not want to preempt the current task. Just return.. 2790 */ 2791 if (likely(!preemptible())) 2792 return; 2793 2794 do { 2795 __preempt_count_add(PREEMPT_ACTIVE); 2796 __schedule(); 2797 __preempt_count_sub(PREEMPT_ACTIVE); 2798 2799 /* 2800 * Check again in case we missed a preemption opportunity 2801 * between schedule and now. 2802 */ 2803 barrier(); 2804 } while (need_resched()); 2805 } 2806 EXPORT_SYMBOL(preempt_schedule); 2807 #endif /* CONFIG_PREEMPT */ 2808 2809 /* 2810 * this is the entry point to schedule() from kernel preemption 2811 * off of irq context. 2812 * Note, that this is called and return with irqs disabled. This will 2813 * protect us against recursive calling from irq. 2814 */ 2815 asmlinkage void __sched preempt_schedule_irq(void) 2816 { 2817 enum ctx_state prev_state; 2818 2819 /* Catch callers which need to be fixed */ 2820 BUG_ON(preempt_count() || !irqs_disabled()); 2821 2822 prev_state = exception_enter(); 2823 2824 do { 2825 __preempt_count_add(PREEMPT_ACTIVE); 2826 local_irq_enable(); 2827 __schedule(); 2828 local_irq_disable(); 2829 __preempt_count_sub(PREEMPT_ACTIVE); 2830 2831 /* 2832 * Check again in case we missed a preemption opportunity 2833 * between schedule and now. 2834 */ 2835 barrier(); 2836 } while (need_resched()); 2837 2838 exception_exit(prev_state); 2839 } 2840 2841 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2842 void *key) 2843 { 2844 return try_to_wake_up(curr->private, mode, wake_flags); 2845 } 2846 EXPORT_SYMBOL(default_wake_function); 2847 2848 static long __sched 2849 sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2850 { 2851 unsigned long flags; 2852 wait_queue_t wait; 2853 2854 init_waitqueue_entry(&wait, current); 2855 2856 __set_current_state(state); 2857 2858 spin_lock_irqsave(&q->lock, flags); 2859 __add_wait_queue(q, &wait); 2860 spin_unlock(&q->lock); 2861 timeout = schedule_timeout(timeout); 2862 spin_lock_irq(&q->lock); 2863 __remove_wait_queue(q, &wait); 2864 spin_unlock_irqrestore(&q->lock, flags); 2865 2866 return timeout; 2867 } 2868 2869 void __sched interruptible_sleep_on(wait_queue_head_t *q) 2870 { 2871 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2872 } 2873 EXPORT_SYMBOL(interruptible_sleep_on); 2874 2875 long __sched 2876 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2877 { 2878 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2879 } 2880 EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2881 2882 void __sched sleep_on(wait_queue_head_t *q) 2883 { 2884 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2885 } 2886 EXPORT_SYMBOL(sleep_on); 2887 2888 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2889 { 2890 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2891 } 2892 EXPORT_SYMBOL(sleep_on_timeout); 2893 2894 #ifdef CONFIG_RT_MUTEXES 2895 2896 /* 2897 * rt_mutex_setprio - set the current priority of a task 2898 * @p: task 2899 * @prio: prio value (kernel-internal form) 2900 * 2901 * This function changes the 'effective' priority of a task. It does 2902 * not touch ->normal_prio like __setscheduler(). 2903 * 2904 * Used by the rt_mutex code to implement priority inheritance 2905 * logic. Call site only calls if the priority of the task changed. 2906 */ 2907 void rt_mutex_setprio(struct task_struct *p, int prio) 2908 { 2909 int oldprio, on_rq, running, enqueue_flag = 0; 2910 struct rq *rq; 2911 const struct sched_class *prev_class; 2912 2913 BUG_ON(prio > MAX_PRIO); 2914 2915 rq = __task_rq_lock(p); 2916 2917 /* 2918 * Idle task boosting is a nono in general. There is one 2919 * exception, when PREEMPT_RT and NOHZ is active: 2920 * 2921 * The idle task calls get_next_timer_interrupt() and holds 2922 * the timer wheel base->lock on the CPU and another CPU wants 2923 * to access the timer (probably to cancel it). We can safely 2924 * ignore the boosting request, as the idle CPU runs this code 2925 * with interrupts disabled and will complete the lock 2926 * protected section without being interrupted. So there is no 2927 * real need to boost. 2928 */ 2929 if (unlikely(p == rq->idle)) { 2930 WARN_ON(p != rq->curr); 2931 WARN_ON(p->pi_blocked_on); 2932 goto out_unlock; 2933 } 2934 2935 trace_sched_pi_setprio(p, prio); 2936 p->pi_top_task = rt_mutex_get_top_task(p); 2937 oldprio = p->prio; 2938 prev_class = p->sched_class; 2939 on_rq = p->on_rq; 2940 running = task_current(rq, p); 2941 if (on_rq) 2942 dequeue_task(rq, p, 0); 2943 if (running) 2944 p->sched_class->put_prev_task(rq, p); 2945 2946 /* 2947 * Boosting condition are: 2948 * 1. -rt task is running and holds mutex A 2949 * --> -dl task blocks on mutex A 2950 * 2951 * 2. -dl task is running and holds mutex A 2952 * --> -dl task blocks on mutex A and could preempt the 2953 * running task 2954 */ 2955 if (dl_prio(prio)) { 2956 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2957 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2958 p->dl.dl_boosted = 1; 2959 p->dl.dl_throttled = 0; 2960 enqueue_flag = ENQUEUE_REPLENISH; 2961 } else 2962 p->dl.dl_boosted = 0; 2963 p->sched_class = &dl_sched_class; 2964 } else if (rt_prio(prio)) { 2965 if (dl_prio(oldprio)) 2966 p->dl.dl_boosted = 0; 2967 if (oldprio < prio) 2968 enqueue_flag = ENQUEUE_HEAD; 2969 p->sched_class = &rt_sched_class; 2970 } else { 2971 if (dl_prio(oldprio)) 2972 p->dl.dl_boosted = 0; 2973 p->sched_class = &fair_sched_class; 2974 } 2975 2976 p->prio = prio; 2977 2978 if (running) 2979 p->sched_class->set_curr_task(rq); 2980 if (on_rq) 2981 enqueue_task(rq, p, enqueue_flag); 2982 2983 check_class_changed(rq, p, prev_class, oldprio); 2984 out_unlock: 2985 __task_rq_unlock(rq); 2986 } 2987 #endif 2988 2989 void set_user_nice(struct task_struct *p, long nice) 2990 { 2991 int old_prio, delta, on_rq; 2992 unsigned long flags; 2993 struct rq *rq; 2994 2995 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 2996 return; 2997 /* 2998 * We have to be careful, if called from sys_setpriority(), 2999 * the task might be in the middle of scheduling on another CPU. 3000 */ 3001 rq = task_rq_lock(p, &flags); 3002 /* 3003 * The RT priorities are set via sched_setscheduler(), but we still 3004 * allow the 'normal' nice value to be set - but as expected 3005 * it wont have any effect on scheduling until the task is 3006 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3007 */ 3008 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3009 p->static_prio = NICE_TO_PRIO(nice); 3010 goto out_unlock; 3011 } 3012 on_rq = p->on_rq; 3013 if (on_rq) 3014 dequeue_task(rq, p, 0); 3015 3016 p->static_prio = NICE_TO_PRIO(nice); 3017 set_load_weight(p); 3018 old_prio = p->prio; 3019 p->prio = effective_prio(p); 3020 delta = p->prio - old_prio; 3021 3022 if (on_rq) { 3023 enqueue_task(rq, p, 0); 3024 /* 3025 * If the task increased its priority or is running and 3026 * lowered its priority, then reschedule its CPU: 3027 */ 3028 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3029 resched_task(rq->curr); 3030 } 3031 out_unlock: 3032 task_rq_unlock(rq, p, &flags); 3033 } 3034 EXPORT_SYMBOL(set_user_nice); 3035 3036 /* 3037 * can_nice - check if a task can reduce its nice value 3038 * @p: task 3039 * @nice: nice value 3040 */ 3041 int can_nice(const struct task_struct *p, const int nice) 3042 { 3043 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3044 int nice_rlim = 20 - nice; 3045 3046 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3047 capable(CAP_SYS_NICE)); 3048 } 3049 3050 #ifdef __ARCH_WANT_SYS_NICE 3051 3052 /* 3053 * sys_nice - change the priority of the current process. 3054 * @increment: priority increment 3055 * 3056 * sys_setpriority is a more generic, but much slower function that 3057 * does similar things. 3058 */ 3059 SYSCALL_DEFINE1(nice, int, increment) 3060 { 3061 long nice, retval; 3062 3063 /* 3064 * Setpriority might change our priority at the same moment. 3065 * We don't have to worry. Conceptually one call occurs first 3066 * and we have a single winner. 3067 */ 3068 if (increment < -40) 3069 increment = -40; 3070 if (increment > 40) 3071 increment = 40; 3072 3073 nice = task_nice(current) + increment; 3074 if (nice < MIN_NICE) 3075 nice = MIN_NICE; 3076 if (nice > MAX_NICE) 3077 nice = MAX_NICE; 3078 3079 if (increment < 0 && !can_nice(current, nice)) 3080 return -EPERM; 3081 3082 retval = security_task_setnice(current, nice); 3083 if (retval) 3084 return retval; 3085 3086 set_user_nice(current, nice); 3087 return 0; 3088 } 3089 3090 #endif 3091 3092 /** 3093 * task_prio - return the priority value of a given task. 3094 * @p: the task in question. 3095 * 3096 * Return: The priority value as seen by users in /proc. 3097 * RT tasks are offset by -200. Normal tasks are centered 3098 * around 0, value goes from -16 to +15. 3099 */ 3100 int task_prio(const struct task_struct *p) 3101 { 3102 return p->prio - MAX_RT_PRIO; 3103 } 3104 3105 /** 3106 * idle_cpu - is a given cpu idle currently? 3107 * @cpu: the processor in question. 3108 * 3109 * Return: 1 if the CPU is currently idle. 0 otherwise. 3110 */ 3111 int idle_cpu(int cpu) 3112 { 3113 struct rq *rq = cpu_rq(cpu); 3114 3115 if (rq->curr != rq->idle) 3116 return 0; 3117 3118 if (rq->nr_running) 3119 return 0; 3120 3121 #ifdef CONFIG_SMP 3122 if (!llist_empty(&rq->wake_list)) 3123 return 0; 3124 #endif 3125 3126 return 1; 3127 } 3128 3129 /** 3130 * idle_task - return the idle task for a given cpu. 3131 * @cpu: the processor in question. 3132 * 3133 * Return: The idle task for the cpu @cpu. 3134 */ 3135 struct task_struct *idle_task(int cpu) 3136 { 3137 return cpu_rq(cpu)->idle; 3138 } 3139 3140 /** 3141 * find_process_by_pid - find a process with a matching PID value. 3142 * @pid: the pid in question. 3143 * 3144 * The task of @pid, if found. %NULL otherwise. 3145 */ 3146 static struct task_struct *find_process_by_pid(pid_t pid) 3147 { 3148 return pid ? find_task_by_vpid(pid) : current; 3149 } 3150 3151 /* 3152 * This function initializes the sched_dl_entity of a newly becoming 3153 * SCHED_DEADLINE task. 3154 * 3155 * Only the static values are considered here, the actual runtime and the 3156 * absolute deadline will be properly calculated when the task is enqueued 3157 * for the first time with its new policy. 3158 */ 3159 static void 3160 __setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3161 { 3162 struct sched_dl_entity *dl_se = &p->dl; 3163 3164 init_dl_task_timer(dl_se); 3165 dl_se->dl_runtime = attr->sched_runtime; 3166 dl_se->dl_deadline = attr->sched_deadline; 3167 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3168 dl_se->flags = attr->sched_flags; 3169 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3170 dl_se->dl_throttled = 0; 3171 dl_se->dl_new = 1; 3172 } 3173 3174 static void __setscheduler_params(struct task_struct *p, 3175 const struct sched_attr *attr) 3176 { 3177 int policy = attr->sched_policy; 3178 3179 if (policy == -1) /* setparam */ 3180 policy = p->policy; 3181 3182 p->policy = policy; 3183 3184 if (dl_policy(policy)) 3185 __setparam_dl(p, attr); 3186 else if (fair_policy(policy)) 3187 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3188 3189 /* 3190 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3191 * !rt_policy. Always setting this ensures that things like 3192 * getparam()/getattr() don't report silly values for !rt tasks. 3193 */ 3194 p->rt_priority = attr->sched_priority; 3195 p->normal_prio = normal_prio(p); 3196 set_load_weight(p); 3197 } 3198 3199 /* Actually do priority change: must hold pi & rq lock. */ 3200 static void __setscheduler(struct rq *rq, struct task_struct *p, 3201 const struct sched_attr *attr) 3202 { 3203 __setscheduler_params(p, attr); 3204 3205 /* 3206 * If we get here, there was no pi waiters boosting the 3207 * task. It is safe to use the normal prio. 3208 */ 3209 p->prio = normal_prio(p); 3210 3211 if (dl_prio(p->prio)) 3212 p->sched_class = &dl_sched_class; 3213 else if (rt_prio(p->prio)) 3214 p->sched_class = &rt_sched_class; 3215 else 3216 p->sched_class = &fair_sched_class; 3217 } 3218 3219 static void 3220 __getparam_dl(struct task_struct *p, struct sched_attr *attr) 3221 { 3222 struct sched_dl_entity *dl_se = &p->dl; 3223 3224 attr->sched_priority = p->rt_priority; 3225 attr->sched_runtime = dl_se->dl_runtime; 3226 attr->sched_deadline = dl_se->dl_deadline; 3227 attr->sched_period = dl_se->dl_period; 3228 attr->sched_flags = dl_se->flags; 3229 } 3230 3231 /* 3232 * This function validates the new parameters of a -deadline task. 3233 * We ask for the deadline not being zero, and greater or equal 3234 * than the runtime, as well as the period of being zero or 3235 * greater than deadline. Furthermore, we have to be sure that 3236 * user parameters are above the internal resolution (1us); we 3237 * check sched_runtime only since it is always the smaller one. 3238 */ 3239 static bool 3240 __checkparam_dl(const struct sched_attr *attr) 3241 { 3242 return attr && attr->sched_deadline != 0 && 3243 (attr->sched_period == 0 || 3244 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3245 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3246 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3247 } 3248 3249 /* 3250 * check the target process has a UID that matches the current process's 3251 */ 3252 static bool check_same_owner(struct task_struct *p) 3253 { 3254 const struct cred *cred = current_cred(), *pcred; 3255 bool match; 3256 3257 rcu_read_lock(); 3258 pcred = __task_cred(p); 3259 match = (uid_eq(cred->euid, pcred->euid) || 3260 uid_eq(cred->euid, pcred->uid)); 3261 rcu_read_unlock(); 3262 return match; 3263 } 3264 3265 static int __sched_setscheduler(struct task_struct *p, 3266 const struct sched_attr *attr, 3267 bool user) 3268 { 3269 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3270 MAX_RT_PRIO - 1 - attr->sched_priority; 3271 int retval, oldprio, oldpolicy = -1, on_rq, running; 3272 int policy = attr->sched_policy; 3273 unsigned long flags; 3274 const struct sched_class *prev_class; 3275 struct rq *rq; 3276 int reset_on_fork; 3277 3278 /* may grab non-irq protected spin_locks */ 3279 BUG_ON(in_interrupt()); 3280 recheck: 3281 /* double check policy once rq lock held */ 3282 if (policy < 0) { 3283 reset_on_fork = p->sched_reset_on_fork; 3284 policy = oldpolicy = p->policy; 3285 } else { 3286 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3287 3288 if (policy != SCHED_DEADLINE && 3289 policy != SCHED_FIFO && policy != SCHED_RR && 3290 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3291 policy != SCHED_IDLE) 3292 return -EINVAL; 3293 } 3294 3295 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3296 return -EINVAL; 3297 3298 /* 3299 * Valid priorities for SCHED_FIFO and SCHED_RR are 3300 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3301 * SCHED_BATCH and SCHED_IDLE is 0. 3302 */ 3303 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3304 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3305 return -EINVAL; 3306 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3307 (rt_policy(policy) != (attr->sched_priority != 0))) 3308 return -EINVAL; 3309 3310 /* 3311 * Allow unprivileged RT tasks to decrease priority: 3312 */ 3313 if (user && !capable(CAP_SYS_NICE)) { 3314 if (fair_policy(policy)) { 3315 if (attr->sched_nice < task_nice(p) && 3316 !can_nice(p, attr->sched_nice)) 3317 return -EPERM; 3318 } 3319 3320 if (rt_policy(policy)) { 3321 unsigned long rlim_rtprio = 3322 task_rlimit(p, RLIMIT_RTPRIO); 3323 3324 /* can't set/change the rt policy */ 3325 if (policy != p->policy && !rlim_rtprio) 3326 return -EPERM; 3327 3328 /* can't increase priority */ 3329 if (attr->sched_priority > p->rt_priority && 3330 attr->sched_priority > rlim_rtprio) 3331 return -EPERM; 3332 } 3333 3334 /* 3335 * Can't set/change SCHED_DEADLINE policy at all for now 3336 * (safest behavior); in the future we would like to allow 3337 * unprivileged DL tasks to increase their relative deadline 3338 * or reduce their runtime (both ways reducing utilization) 3339 */ 3340 if (dl_policy(policy)) 3341 return -EPERM; 3342 3343 /* 3344 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3345 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3346 */ 3347 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3348 if (!can_nice(p, task_nice(p))) 3349 return -EPERM; 3350 } 3351 3352 /* can't change other user's priorities */ 3353 if (!check_same_owner(p)) 3354 return -EPERM; 3355 3356 /* Normal users shall not reset the sched_reset_on_fork flag */ 3357 if (p->sched_reset_on_fork && !reset_on_fork) 3358 return -EPERM; 3359 } 3360 3361 if (user) { 3362 retval = security_task_setscheduler(p); 3363 if (retval) 3364 return retval; 3365 } 3366 3367 /* 3368 * make sure no PI-waiters arrive (or leave) while we are 3369 * changing the priority of the task: 3370 * 3371 * To be able to change p->policy safely, the appropriate 3372 * runqueue lock must be held. 3373 */ 3374 rq = task_rq_lock(p, &flags); 3375 3376 /* 3377 * Changing the policy of the stop threads its a very bad idea 3378 */ 3379 if (p == rq->stop) { 3380 task_rq_unlock(rq, p, &flags); 3381 return -EINVAL; 3382 } 3383 3384 /* 3385 * If not changing anything there's no need to proceed further, 3386 * but store a possible modification of reset_on_fork. 3387 */ 3388 if (unlikely(policy == p->policy)) { 3389 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3390 goto change; 3391 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3392 goto change; 3393 if (dl_policy(policy)) 3394 goto change; 3395 3396 p->sched_reset_on_fork = reset_on_fork; 3397 task_rq_unlock(rq, p, &flags); 3398 return 0; 3399 } 3400 change: 3401 3402 if (user) { 3403 #ifdef CONFIG_RT_GROUP_SCHED 3404 /* 3405 * Do not allow realtime tasks into groups that have no runtime 3406 * assigned. 3407 */ 3408 if (rt_bandwidth_enabled() && rt_policy(policy) && 3409 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3410 !task_group_is_autogroup(task_group(p))) { 3411 task_rq_unlock(rq, p, &flags); 3412 return -EPERM; 3413 } 3414 #endif 3415 #ifdef CONFIG_SMP 3416 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3417 cpumask_t *span = rq->rd->span; 3418 3419 /* 3420 * Don't allow tasks with an affinity mask smaller than 3421 * the entire root_domain to become SCHED_DEADLINE. We 3422 * will also fail if there's no bandwidth available. 3423 */ 3424 if (!cpumask_subset(span, &p->cpus_allowed) || 3425 rq->rd->dl_bw.bw == 0) { 3426 task_rq_unlock(rq, p, &flags); 3427 return -EPERM; 3428 } 3429 } 3430 #endif 3431 } 3432 3433 /* recheck policy now with rq lock held */ 3434 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3435 policy = oldpolicy = -1; 3436 task_rq_unlock(rq, p, &flags); 3437 goto recheck; 3438 } 3439 3440 /* 3441 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3442 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3443 * is available. 3444 */ 3445 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3446 task_rq_unlock(rq, p, &flags); 3447 return -EBUSY; 3448 } 3449 3450 p->sched_reset_on_fork = reset_on_fork; 3451 oldprio = p->prio; 3452 3453 /* 3454 * Special case for priority boosted tasks. 3455 * 3456 * If the new priority is lower or equal (user space view) 3457 * than the current (boosted) priority, we just store the new 3458 * normal parameters and do not touch the scheduler class and 3459 * the runqueue. This will be done when the task deboost 3460 * itself. 3461 */ 3462 if (rt_mutex_check_prio(p, newprio)) { 3463 __setscheduler_params(p, attr); 3464 task_rq_unlock(rq, p, &flags); 3465 return 0; 3466 } 3467 3468 on_rq = p->on_rq; 3469 running = task_current(rq, p); 3470 if (on_rq) 3471 dequeue_task(rq, p, 0); 3472 if (running) 3473 p->sched_class->put_prev_task(rq, p); 3474 3475 prev_class = p->sched_class; 3476 __setscheduler(rq, p, attr); 3477 3478 if (running) 3479 p->sched_class->set_curr_task(rq); 3480 if (on_rq) { 3481 /* 3482 * We enqueue to tail when the priority of a task is 3483 * increased (user space view). 3484 */ 3485 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3486 } 3487 3488 check_class_changed(rq, p, prev_class, oldprio); 3489 task_rq_unlock(rq, p, &flags); 3490 3491 rt_mutex_adjust_pi(p); 3492 3493 return 0; 3494 } 3495 3496 static int _sched_setscheduler(struct task_struct *p, int policy, 3497 const struct sched_param *param, bool check) 3498 { 3499 struct sched_attr attr = { 3500 .sched_policy = policy, 3501 .sched_priority = param->sched_priority, 3502 .sched_nice = PRIO_TO_NICE(p->static_prio), 3503 }; 3504 3505 /* 3506 * Fixup the legacy SCHED_RESET_ON_FORK hack 3507 */ 3508 if (policy & SCHED_RESET_ON_FORK) { 3509 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3510 policy &= ~SCHED_RESET_ON_FORK; 3511 attr.sched_policy = policy; 3512 } 3513 3514 return __sched_setscheduler(p, &attr, check); 3515 } 3516 /** 3517 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3518 * @p: the task in question. 3519 * @policy: new policy. 3520 * @param: structure containing the new RT priority. 3521 * 3522 * Return: 0 on success. An error code otherwise. 3523 * 3524 * NOTE that the task may be already dead. 3525 */ 3526 int sched_setscheduler(struct task_struct *p, int policy, 3527 const struct sched_param *param) 3528 { 3529 return _sched_setscheduler(p, policy, param, true); 3530 } 3531 EXPORT_SYMBOL_GPL(sched_setscheduler); 3532 3533 int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3534 { 3535 return __sched_setscheduler(p, attr, true); 3536 } 3537 EXPORT_SYMBOL_GPL(sched_setattr); 3538 3539 /** 3540 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3541 * @p: the task in question. 3542 * @policy: new policy. 3543 * @param: structure containing the new RT priority. 3544 * 3545 * Just like sched_setscheduler, only don't bother checking if the 3546 * current context has permission. For example, this is needed in 3547 * stop_machine(): we create temporary high priority worker threads, 3548 * but our caller might not have that capability. 3549 * 3550 * Return: 0 on success. An error code otherwise. 3551 */ 3552 int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3553 const struct sched_param *param) 3554 { 3555 return _sched_setscheduler(p, policy, param, false); 3556 } 3557 3558 static int 3559 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3560 { 3561 struct sched_param lparam; 3562 struct task_struct *p; 3563 int retval; 3564 3565 if (!param || pid < 0) 3566 return -EINVAL; 3567 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3568 return -EFAULT; 3569 3570 rcu_read_lock(); 3571 retval = -ESRCH; 3572 p = find_process_by_pid(pid); 3573 if (p != NULL) 3574 retval = sched_setscheduler(p, policy, &lparam); 3575 rcu_read_unlock(); 3576 3577 return retval; 3578 } 3579 3580 /* 3581 * Mimics kernel/events/core.c perf_copy_attr(). 3582 */ 3583 static int sched_copy_attr(struct sched_attr __user *uattr, 3584 struct sched_attr *attr) 3585 { 3586 u32 size; 3587 int ret; 3588 3589 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3590 return -EFAULT; 3591 3592 /* 3593 * zero the full structure, so that a short copy will be nice. 3594 */ 3595 memset(attr, 0, sizeof(*attr)); 3596 3597 ret = get_user(size, &uattr->size); 3598 if (ret) 3599 return ret; 3600 3601 if (size > PAGE_SIZE) /* silly large */ 3602 goto err_size; 3603 3604 if (!size) /* abi compat */ 3605 size = SCHED_ATTR_SIZE_VER0; 3606 3607 if (size < SCHED_ATTR_SIZE_VER0) 3608 goto err_size; 3609 3610 /* 3611 * If we're handed a bigger struct than we know of, 3612 * ensure all the unknown bits are 0 - i.e. new 3613 * user-space does not rely on any kernel feature 3614 * extensions we dont know about yet. 3615 */ 3616 if (size > sizeof(*attr)) { 3617 unsigned char __user *addr; 3618 unsigned char __user *end; 3619 unsigned char val; 3620 3621 addr = (void __user *)uattr + sizeof(*attr); 3622 end = (void __user *)uattr + size; 3623 3624 for (; addr < end; addr++) { 3625 ret = get_user(val, addr); 3626 if (ret) 3627 return ret; 3628 if (val) 3629 goto err_size; 3630 } 3631 size = sizeof(*attr); 3632 } 3633 3634 ret = copy_from_user(attr, uattr, size); 3635 if (ret) 3636 return -EFAULT; 3637 3638 /* 3639 * XXX: do we want to be lenient like existing syscalls; or do we want 3640 * to be strict and return an error on out-of-bounds values? 3641 */ 3642 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3643 3644 out: 3645 return ret; 3646 3647 err_size: 3648 put_user(sizeof(*attr), &uattr->size); 3649 ret = -E2BIG; 3650 goto out; 3651 } 3652 3653 /** 3654 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3655 * @pid: the pid in question. 3656 * @policy: new policy. 3657 * @param: structure containing the new RT priority. 3658 * 3659 * Return: 0 on success. An error code otherwise. 3660 */ 3661 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3662 struct sched_param __user *, param) 3663 { 3664 /* negative values for policy are not valid */ 3665 if (policy < 0) 3666 return -EINVAL; 3667 3668 return do_sched_setscheduler(pid, policy, param); 3669 } 3670 3671 /** 3672 * sys_sched_setparam - set/change the RT priority of a thread 3673 * @pid: the pid in question. 3674 * @param: structure containing the new RT priority. 3675 * 3676 * Return: 0 on success. An error code otherwise. 3677 */ 3678 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3679 { 3680 return do_sched_setscheduler(pid, -1, param); 3681 } 3682 3683 /** 3684 * sys_sched_setattr - same as above, but with extended sched_attr 3685 * @pid: the pid in question. 3686 * @uattr: structure containing the extended parameters. 3687 */ 3688 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3689 unsigned int, flags) 3690 { 3691 struct sched_attr attr; 3692 struct task_struct *p; 3693 int retval; 3694 3695 if (!uattr || pid < 0 || flags) 3696 return -EINVAL; 3697 3698 if (sched_copy_attr(uattr, &attr)) 3699 return -EFAULT; 3700 3701 rcu_read_lock(); 3702 retval = -ESRCH; 3703 p = find_process_by_pid(pid); 3704 if (p != NULL) 3705 retval = sched_setattr(p, &attr); 3706 rcu_read_unlock(); 3707 3708 return retval; 3709 } 3710 3711 /** 3712 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3713 * @pid: the pid in question. 3714 * 3715 * Return: On success, the policy of the thread. Otherwise, a negative error 3716 * code. 3717 */ 3718 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3719 { 3720 struct task_struct *p; 3721 int retval; 3722 3723 if (pid < 0) 3724 return -EINVAL; 3725 3726 retval = -ESRCH; 3727 rcu_read_lock(); 3728 p = find_process_by_pid(pid); 3729 if (p) { 3730 retval = security_task_getscheduler(p); 3731 if (!retval) 3732 retval = p->policy 3733 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3734 } 3735 rcu_read_unlock(); 3736 return retval; 3737 } 3738 3739 /** 3740 * sys_sched_getparam - get the RT priority of a thread 3741 * @pid: the pid in question. 3742 * @param: structure containing the RT priority. 3743 * 3744 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3745 * code. 3746 */ 3747 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3748 { 3749 struct sched_param lp; 3750 struct task_struct *p; 3751 int retval; 3752 3753 if (!param || pid < 0) 3754 return -EINVAL; 3755 3756 rcu_read_lock(); 3757 p = find_process_by_pid(pid); 3758 retval = -ESRCH; 3759 if (!p) 3760 goto out_unlock; 3761 3762 retval = security_task_getscheduler(p); 3763 if (retval) 3764 goto out_unlock; 3765 3766 if (task_has_dl_policy(p)) { 3767 retval = -EINVAL; 3768 goto out_unlock; 3769 } 3770 lp.sched_priority = p->rt_priority; 3771 rcu_read_unlock(); 3772 3773 /* 3774 * This one might sleep, we cannot do it with a spinlock held ... 3775 */ 3776 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3777 3778 return retval; 3779 3780 out_unlock: 3781 rcu_read_unlock(); 3782 return retval; 3783 } 3784 3785 static int sched_read_attr(struct sched_attr __user *uattr, 3786 struct sched_attr *attr, 3787 unsigned int usize) 3788 { 3789 int ret; 3790 3791 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3792 return -EFAULT; 3793 3794 /* 3795 * If we're handed a smaller struct than we know of, 3796 * ensure all the unknown bits are 0 - i.e. old 3797 * user-space does not get uncomplete information. 3798 */ 3799 if (usize < sizeof(*attr)) { 3800 unsigned char *addr; 3801 unsigned char *end; 3802 3803 addr = (void *)attr + usize; 3804 end = (void *)attr + sizeof(*attr); 3805 3806 for (; addr < end; addr++) { 3807 if (*addr) 3808 goto err_size; 3809 } 3810 3811 attr->size = usize; 3812 } 3813 3814 ret = copy_to_user(uattr, attr, attr->size); 3815 if (ret) 3816 return -EFAULT; 3817 3818 out: 3819 return ret; 3820 3821 err_size: 3822 ret = -E2BIG; 3823 goto out; 3824 } 3825 3826 /** 3827 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3828 * @pid: the pid in question. 3829 * @uattr: structure containing the extended parameters. 3830 * @size: sizeof(attr) for fwd/bwd comp. 3831 */ 3832 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3833 unsigned int, size, unsigned int, flags) 3834 { 3835 struct sched_attr attr = { 3836 .size = sizeof(struct sched_attr), 3837 }; 3838 struct task_struct *p; 3839 int retval; 3840 3841 if (!uattr || pid < 0 || size > PAGE_SIZE || 3842 size < SCHED_ATTR_SIZE_VER0 || flags) 3843 return -EINVAL; 3844 3845 rcu_read_lock(); 3846 p = find_process_by_pid(pid); 3847 retval = -ESRCH; 3848 if (!p) 3849 goto out_unlock; 3850 3851 retval = security_task_getscheduler(p); 3852 if (retval) 3853 goto out_unlock; 3854 3855 attr.sched_policy = p->policy; 3856 if (p->sched_reset_on_fork) 3857 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3858 if (task_has_dl_policy(p)) 3859 __getparam_dl(p, &attr); 3860 else if (task_has_rt_policy(p)) 3861 attr.sched_priority = p->rt_priority; 3862 else 3863 attr.sched_nice = task_nice(p); 3864 3865 rcu_read_unlock(); 3866 3867 retval = sched_read_attr(uattr, &attr, size); 3868 return retval; 3869 3870 out_unlock: 3871 rcu_read_unlock(); 3872 return retval; 3873 } 3874 3875 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3876 { 3877 cpumask_var_t cpus_allowed, new_mask; 3878 struct task_struct *p; 3879 int retval; 3880 3881 rcu_read_lock(); 3882 3883 p = find_process_by_pid(pid); 3884 if (!p) { 3885 rcu_read_unlock(); 3886 return -ESRCH; 3887 } 3888 3889 /* Prevent p going away */ 3890 get_task_struct(p); 3891 rcu_read_unlock(); 3892 3893 if (p->flags & PF_NO_SETAFFINITY) { 3894 retval = -EINVAL; 3895 goto out_put_task; 3896 } 3897 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3898 retval = -ENOMEM; 3899 goto out_put_task; 3900 } 3901 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3902 retval = -ENOMEM; 3903 goto out_free_cpus_allowed; 3904 } 3905 retval = -EPERM; 3906 if (!check_same_owner(p)) { 3907 rcu_read_lock(); 3908 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3909 rcu_read_unlock(); 3910 goto out_unlock; 3911 } 3912 rcu_read_unlock(); 3913 } 3914 3915 retval = security_task_setscheduler(p); 3916 if (retval) 3917 goto out_unlock; 3918 3919 3920 cpuset_cpus_allowed(p, cpus_allowed); 3921 cpumask_and(new_mask, in_mask, cpus_allowed); 3922 3923 /* 3924 * Since bandwidth control happens on root_domain basis, 3925 * if admission test is enabled, we only admit -deadline 3926 * tasks allowed to run on all the CPUs in the task's 3927 * root_domain. 3928 */ 3929 #ifdef CONFIG_SMP 3930 if (task_has_dl_policy(p)) { 3931 const struct cpumask *span = task_rq(p)->rd->span; 3932 3933 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3934 retval = -EBUSY; 3935 goto out_unlock; 3936 } 3937 } 3938 #endif 3939 again: 3940 retval = set_cpus_allowed_ptr(p, new_mask); 3941 3942 if (!retval) { 3943 cpuset_cpus_allowed(p, cpus_allowed); 3944 if (!cpumask_subset(new_mask, cpus_allowed)) { 3945 /* 3946 * We must have raced with a concurrent cpuset 3947 * update. Just reset the cpus_allowed to the 3948 * cpuset's cpus_allowed 3949 */ 3950 cpumask_copy(new_mask, cpus_allowed); 3951 goto again; 3952 } 3953 } 3954 out_unlock: 3955 free_cpumask_var(new_mask); 3956 out_free_cpus_allowed: 3957 free_cpumask_var(cpus_allowed); 3958 out_put_task: 3959 put_task_struct(p); 3960 return retval; 3961 } 3962 3963 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3964 struct cpumask *new_mask) 3965 { 3966 if (len < cpumask_size()) 3967 cpumask_clear(new_mask); 3968 else if (len > cpumask_size()) 3969 len = cpumask_size(); 3970 3971 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3972 } 3973 3974 /** 3975 * sys_sched_setaffinity - set the cpu affinity of a process 3976 * @pid: pid of the process 3977 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3978 * @user_mask_ptr: user-space pointer to the new cpu mask 3979 * 3980 * Return: 0 on success. An error code otherwise. 3981 */ 3982 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3983 unsigned long __user *, user_mask_ptr) 3984 { 3985 cpumask_var_t new_mask; 3986 int retval; 3987 3988 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3989 return -ENOMEM; 3990 3991 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3992 if (retval == 0) 3993 retval = sched_setaffinity(pid, new_mask); 3994 free_cpumask_var(new_mask); 3995 return retval; 3996 } 3997 3998 long sched_getaffinity(pid_t pid, struct cpumask *mask) 3999 { 4000 struct task_struct *p; 4001 unsigned long flags; 4002 int retval; 4003 4004 rcu_read_lock(); 4005 4006 retval = -ESRCH; 4007 p = find_process_by_pid(pid); 4008 if (!p) 4009 goto out_unlock; 4010 4011 retval = security_task_getscheduler(p); 4012 if (retval) 4013 goto out_unlock; 4014 4015 raw_spin_lock_irqsave(&p->pi_lock, flags); 4016 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4017 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4018 4019 out_unlock: 4020 rcu_read_unlock(); 4021 4022 return retval; 4023 } 4024 4025 /** 4026 * sys_sched_getaffinity - get the cpu affinity of a process 4027 * @pid: pid of the process 4028 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4029 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4030 * 4031 * Return: 0 on success. An error code otherwise. 4032 */ 4033 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4034 unsigned long __user *, user_mask_ptr) 4035 { 4036 int ret; 4037 cpumask_var_t mask; 4038 4039 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4040 return -EINVAL; 4041 if (len & (sizeof(unsigned long)-1)) 4042 return -EINVAL; 4043 4044 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4045 return -ENOMEM; 4046 4047 ret = sched_getaffinity(pid, mask); 4048 if (ret == 0) { 4049 size_t retlen = min_t(size_t, len, cpumask_size()); 4050 4051 if (copy_to_user(user_mask_ptr, mask, retlen)) 4052 ret = -EFAULT; 4053 else 4054 ret = retlen; 4055 } 4056 free_cpumask_var(mask); 4057 4058 return ret; 4059 } 4060 4061 /** 4062 * sys_sched_yield - yield the current processor to other threads. 4063 * 4064 * This function yields the current CPU to other tasks. If there are no 4065 * other threads running on this CPU then this function will return. 4066 * 4067 * Return: 0. 4068 */ 4069 SYSCALL_DEFINE0(sched_yield) 4070 { 4071 struct rq *rq = this_rq_lock(); 4072 4073 schedstat_inc(rq, yld_count); 4074 current->sched_class->yield_task(rq); 4075 4076 /* 4077 * Since we are going to call schedule() anyway, there's 4078 * no need to preempt or enable interrupts: 4079 */ 4080 __release(rq->lock); 4081 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4082 do_raw_spin_unlock(&rq->lock); 4083 sched_preempt_enable_no_resched(); 4084 4085 schedule(); 4086 4087 return 0; 4088 } 4089 4090 static void __cond_resched(void) 4091 { 4092 __preempt_count_add(PREEMPT_ACTIVE); 4093 __schedule(); 4094 __preempt_count_sub(PREEMPT_ACTIVE); 4095 } 4096 4097 int __sched _cond_resched(void) 4098 { 4099 if (should_resched()) { 4100 __cond_resched(); 4101 return 1; 4102 } 4103 return 0; 4104 } 4105 EXPORT_SYMBOL(_cond_resched); 4106 4107 /* 4108 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4109 * call schedule, and on return reacquire the lock. 4110 * 4111 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4112 * operations here to prevent schedule() from being called twice (once via 4113 * spin_unlock(), once by hand). 4114 */ 4115 int __cond_resched_lock(spinlock_t *lock) 4116 { 4117 int resched = should_resched(); 4118 int ret = 0; 4119 4120 lockdep_assert_held(lock); 4121 4122 if (spin_needbreak(lock) || resched) { 4123 spin_unlock(lock); 4124 if (resched) 4125 __cond_resched(); 4126 else 4127 cpu_relax(); 4128 ret = 1; 4129 spin_lock(lock); 4130 } 4131 return ret; 4132 } 4133 EXPORT_SYMBOL(__cond_resched_lock); 4134 4135 int __sched __cond_resched_softirq(void) 4136 { 4137 BUG_ON(!in_softirq()); 4138 4139 if (should_resched()) { 4140 local_bh_enable(); 4141 __cond_resched(); 4142 local_bh_disable(); 4143 return 1; 4144 } 4145 return 0; 4146 } 4147 EXPORT_SYMBOL(__cond_resched_softirq); 4148 4149 /** 4150 * yield - yield the current processor to other threads. 4151 * 4152 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4153 * 4154 * The scheduler is at all times free to pick the calling task as the most 4155 * eligible task to run, if removing the yield() call from your code breaks 4156 * it, its already broken. 4157 * 4158 * Typical broken usage is: 4159 * 4160 * while (!event) 4161 * yield(); 4162 * 4163 * where one assumes that yield() will let 'the other' process run that will 4164 * make event true. If the current task is a SCHED_FIFO task that will never 4165 * happen. Never use yield() as a progress guarantee!! 4166 * 4167 * If you want to use yield() to wait for something, use wait_event(). 4168 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4169 * If you still want to use yield(), do not! 4170 */ 4171 void __sched yield(void) 4172 { 4173 set_current_state(TASK_RUNNING); 4174 sys_sched_yield(); 4175 } 4176 EXPORT_SYMBOL(yield); 4177 4178 /** 4179 * yield_to - yield the current processor to another thread in 4180 * your thread group, or accelerate that thread toward the 4181 * processor it's on. 4182 * @p: target task 4183 * @preempt: whether task preemption is allowed or not 4184 * 4185 * It's the caller's job to ensure that the target task struct 4186 * can't go away on us before we can do any checks. 4187 * 4188 * Return: 4189 * true (>0) if we indeed boosted the target task. 4190 * false (0) if we failed to boost the target. 4191 * -ESRCH if there's no task to yield to. 4192 */ 4193 bool __sched yield_to(struct task_struct *p, bool preempt) 4194 { 4195 struct task_struct *curr = current; 4196 struct rq *rq, *p_rq; 4197 unsigned long flags; 4198 int yielded = 0; 4199 4200 local_irq_save(flags); 4201 rq = this_rq(); 4202 4203 again: 4204 p_rq = task_rq(p); 4205 /* 4206 * If we're the only runnable task on the rq and target rq also 4207 * has only one task, there's absolutely no point in yielding. 4208 */ 4209 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4210 yielded = -ESRCH; 4211 goto out_irq; 4212 } 4213 4214 double_rq_lock(rq, p_rq); 4215 if (task_rq(p) != p_rq) { 4216 double_rq_unlock(rq, p_rq); 4217 goto again; 4218 } 4219 4220 if (!curr->sched_class->yield_to_task) 4221 goto out_unlock; 4222 4223 if (curr->sched_class != p->sched_class) 4224 goto out_unlock; 4225 4226 if (task_running(p_rq, p) || p->state) 4227 goto out_unlock; 4228 4229 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4230 if (yielded) { 4231 schedstat_inc(rq, yld_count); 4232 /* 4233 * Make p's CPU reschedule; pick_next_entity takes care of 4234 * fairness. 4235 */ 4236 if (preempt && rq != p_rq) 4237 resched_task(p_rq->curr); 4238 } 4239 4240 out_unlock: 4241 double_rq_unlock(rq, p_rq); 4242 out_irq: 4243 local_irq_restore(flags); 4244 4245 if (yielded > 0) 4246 schedule(); 4247 4248 return yielded; 4249 } 4250 EXPORT_SYMBOL_GPL(yield_to); 4251 4252 /* 4253 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4254 * that process accounting knows that this is a task in IO wait state. 4255 */ 4256 void __sched io_schedule(void) 4257 { 4258 struct rq *rq = raw_rq(); 4259 4260 delayacct_blkio_start(); 4261 atomic_inc(&rq->nr_iowait); 4262 blk_flush_plug(current); 4263 current->in_iowait = 1; 4264 schedule(); 4265 current->in_iowait = 0; 4266 atomic_dec(&rq->nr_iowait); 4267 delayacct_blkio_end(); 4268 } 4269 EXPORT_SYMBOL(io_schedule); 4270 4271 long __sched io_schedule_timeout(long timeout) 4272 { 4273 struct rq *rq = raw_rq(); 4274 long ret; 4275 4276 delayacct_blkio_start(); 4277 atomic_inc(&rq->nr_iowait); 4278 blk_flush_plug(current); 4279 current->in_iowait = 1; 4280 ret = schedule_timeout(timeout); 4281 current->in_iowait = 0; 4282 atomic_dec(&rq->nr_iowait); 4283 delayacct_blkio_end(); 4284 return ret; 4285 } 4286 4287 /** 4288 * sys_sched_get_priority_max - return maximum RT priority. 4289 * @policy: scheduling class. 4290 * 4291 * Return: On success, this syscall returns the maximum 4292 * rt_priority that can be used by a given scheduling class. 4293 * On failure, a negative error code is returned. 4294 */ 4295 SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4296 { 4297 int ret = -EINVAL; 4298 4299 switch (policy) { 4300 case SCHED_FIFO: 4301 case SCHED_RR: 4302 ret = MAX_USER_RT_PRIO-1; 4303 break; 4304 case SCHED_DEADLINE: 4305 case SCHED_NORMAL: 4306 case SCHED_BATCH: 4307 case SCHED_IDLE: 4308 ret = 0; 4309 break; 4310 } 4311 return ret; 4312 } 4313 4314 /** 4315 * sys_sched_get_priority_min - return minimum RT priority. 4316 * @policy: scheduling class. 4317 * 4318 * Return: On success, this syscall returns the minimum 4319 * rt_priority that can be used by a given scheduling class. 4320 * On failure, a negative error code is returned. 4321 */ 4322 SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4323 { 4324 int ret = -EINVAL; 4325 4326 switch (policy) { 4327 case SCHED_FIFO: 4328 case SCHED_RR: 4329 ret = 1; 4330 break; 4331 case SCHED_DEADLINE: 4332 case SCHED_NORMAL: 4333 case SCHED_BATCH: 4334 case SCHED_IDLE: 4335 ret = 0; 4336 } 4337 return ret; 4338 } 4339 4340 /** 4341 * sys_sched_rr_get_interval - return the default timeslice of a process. 4342 * @pid: pid of the process. 4343 * @interval: userspace pointer to the timeslice value. 4344 * 4345 * this syscall writes the default timeslice value of a given process 4346 * into the user-space timespec buffer. A value of '0' means infinity. 4347 * 4348 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4349 * an error code. 4350 */ 4351 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4352 struct timespec __user *, interval) 4353 { 4354 struct task_struct *p; 4355 unsigned int time_slice; 4356 unsigned long flags; 4357 struct rq *rq; 4358 int retval; 4359 struct timespec t; 4360 4361 if (pid < 0) 4362 return -EINVAL; 4363 4364 retval = -ESRCH; 4365 rcu_read_lock(); 4366 p = find_process_by_pid(pid); 4367 if (!p) 4368 goto out_unlock; 4369 4370 retval = security_task_getscheduler(p); 4371 if (retval) 4372 goto out_unlock; 4373 4374 rq = task_rq_lock(p, &flags); 4375 time_slice = 0; 4376 if (p->sched_class->get_rr_interval) 4377 time_slice = p->sched_class->get_rr_interval(rq, p); 4378 task_rq_unlock(rq, p, &flags); 4379 4380 rcu_read_unlock(); 4381 jiffies_to_timespec(time_slice, &t); 4382 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4383 return retval; 4384 4385 out_unlock: 4386 rcu_read_unlock(); 4387 return retval; 4388 } 4389 4390 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4391 4392 void sched_show_task(struct task_struct *p) 4393 { 4394 unsigned long free = 0; 4395 int ppid; 4396 unsigned state; 4397 4398 state = p->state ? __ffs(p->state) + 1 : 0; 4399 printk(KERN_INFO "%-15.15s %c", p->comm, 4400 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4401 #if BITS_PER_LONG == 32 4402 if (state == TASK_RUNNING) 4403 printk(KERN_CONT " running "); 4404 else 4405 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4406 #else 4407 if (state == TASK_RUNNING) 4408 printk(KERN_CONT " running task "); 4409 else 4410 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4411 #endif 4412 #ifdef CONFIG_DEBUG_STACK_USAGE 4413 free = stack_not_used(p); 4414 #endif 4415 rcu_read_lock(); 4416 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4417 rcu_read_unlock(); 4418 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4419 task_pid_nr(p), ppid, 4420 (unsigned long)task_thread_info(p)->flags); 4421 4422 print_worker_info(KERN_INFO, p); 4423 show_stack(p, NULL); 4424 } 4425 4426 void show_state_filter(unsigned long state_filter) 4427 { 4428 struct task_struct *g, *p; 4429 4430 #if BITS_PER_LONG == 32 4431 printk(KERN_INFO 4432 " task PC stack pid father\n"); 4433 #else 4434 printk(KERN_INFO 4435 " task PC stack pid father\n"); 4436 #endif 4437 rcu_read_lock(); 4438 do_each_thread(g, p) { 4439 /* 4440 * reset the NMI-timeout, listing all files on a slow 4441 * console might take a lot of time: 4442 */ 4443 touch_nmi_watchdog(); 4444 if (!state_filter || (p->state & state_filter)) 4445 sched_show_task(p); 4446 } while_each_thread(g, p); 4447 4448 touch_all_softlockup_watchdogs(); 4449 4450 #ifdef CONFIG_SCHED_DEBUG 4451 sysrq_sched_debug_show(); 4452 #endif 4453 rcu_read_unlock(); 4454 /* 4455 * Only show locks if all tasks are dumped: 4456 */ 4457 if (!state_filter) 4458 debug_show_all_locks(); 4459 } 4460 4461 void init_idle_bootup_task(struct task_struct *idle) 4462 { 4463 idle->sched_class = &idle_sched_class; 4464 } 4465 4466 /** 4467 * init_idle - set up an idle thread for a given CPU 4468 * @idle: task in question 4469 * @cpu: cpu the idle task belongs to 4470 * 4471 * NOTE: this function does not set the idle thread's NEED_RESCHED 4472 * flag, to make booting more robust. 4473 */ 4474 void init_idle(struct task_struct *idle, int cpu) 4475 { 4476 struct rq *rq = cpu_rq(cpu); 4477 unsigned long flags; 4478 4479 raw_spin_lock_irqsave(&rq->lock, flags); 4480 4481 __sched_fork(0, idle); 4482 idle->state = TASK_RUNNING; 4483 idle->se.exec_start = sched_clock(); 4484 4485 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4486 /* 4487 * We're having a chicken and egg problem, even though we are 4488 * holding rq->lock, the cpu isn't yet set to this cpu so the 4489 * lockdep check in task_group() will fail. 4490 * 4491 * Similar case to sched_fork(). / Alternatively we could 4492 * use task_rq_lock() here and obtain the other rq->lock. 4493 * 4494 * Silence PROVE_RCU 4495 */ 4496 rcu_read_lock(); 4497 __set_task_cpu(idle, cpu); 4498 rcu_read_unlock(); 4499 4500 rq->curr = rq->idle = idle; 4501 idle->on_rq = 1; 4502 #if defined(CONFIG_SMP) 4503 idle->on_cpu = 1; 4504 #endif 4505 raw_spin_unlock_irqrestore(&rq->lock, flags); 4506 4507 /* Set the preempt count _outside_ the spinlocks! */ 4508 init_idle_preempt_count(idle, cpu); 4509 4510 /* 4511 * The idle tasks have their own, simple scheduling class: 4512 */ 4513 idle->sched_class = &idle_sched_class; 4514 ftrace_graph_init_idle_task(idle, cpu); 4515 vtime_init_idle(idle, cpu); 4516 #if defined(CONFIG_SMP) 4517 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4518 #endif 4519 } 4520 4521 #ifdef CONFIG_SMP 4522 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4523 { 4524 if (p->sched_class && p->sched_class->set_cpus_allowed) 4525 p->sched_class->set_cpus_allowed(p, new_mask); 4526 4527 cpumask_copy(&p->cpus_allowed, new_mask); 4528 p->nr_cpus_allowed = cpumask_weight(new_mask); 4529 } 4530 4531 /* 4532 * This is how migration works: 4533 * 4534 * 1) we invoke migration_cpu_stop() on the target CPU using 4535 * stop_one_cpu(). 4536 * 2) stopper starts to run (implicitly forcing the migrated thread 4537 * off the CPU) 4538 * 3) it checks whether the migrated task is still in the wrong runqueue. 4539 * 4) if it's in the wrong runqueue then the migration thread removes 4540 * it and puts it into the right queue. 4541 * 5) stopper completes and stop_one_cpu() returns and the migration 4542 * is done. 4543 */ 4544 4545 /* 4546 * Change a given task's CPU affinity. Migrate the thread to a 4547 * proper CPU and schedule it away if the CPU it's executing on 4548 * is removed from the allowed bitmask. 4549 * 4550 * NOTE: the caller must have a valid reference to the task, the 4551 * task must not exit() & deallocate itself prematurely. The 4552 * call is not atomic; no spinlocks may be held. 4553 */ 4554 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4555 { 4556 unsigned long flags; 4557 struct rq *rq; 4558 unsigned int dest_cpu; 4559 int ret = 0; 4560 4561 rq = task_rq_lock(p, &flags); 4562 4563 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4564 goto out; 4565 4566 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4567 ret = -EINVAL; 4568 goto out; 4569 } 4570 4571 do_set_cpus_allowed(p, new_mask); 4572 4573 /* Can the task run on the task's current CPU? If so, we're done */ 4574 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4575 goto out; 4576 4577 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4578 if (p->on_rq) { 4579 struct migration_arg arg = { p, dest_cpu }; 4580 /* Need help from migration thread: drop lock and wait. */ 4581 task_rq_unlock(rq, p, &flags); 4582 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4583 tlb_migrate_finish(p->mm); 4584 return 0; 4585 } 4586 out: 4587 task_rq_unlock(rq, p, &flags); 4588 4589 return ret; 4590 } 4591 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4592 4593 /* 4594 * Move (not current) task off this cpu, onto dest cpu. We're doing 4595 * this because either it can't run here any more (set_cpus_allowed() 4596 * away from this CPU, or CPU going down), or because we're 4597 * attempting to rebalance this task on exec (sched_exec). 4598 * 4599 * So we race with normal scheduler movements, but that's OK, as long 4600 * as the task is no longer on this CPU. 4601 * 4602 * Returns non-zero if task was successfully migrated. 4603 */ 4604 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4605 { 4606 struct rq *rq_dest, *rq_src; 4607 int ret = 0; 4608 4609 if (unlikely(!cpu_active(dest_cpu))) 4610 return ret; 4611 4612 rq_src = cpu_rq(src_cpu); 4613 rq_dest = cpu_rq(dest_cpu); 4614 4615 raw_spin_lock(&p->pi_lock); 4616 double_rq_lock(rq_src, rq_dest); 4617 /* Already moved. */ 4618 if (task_cpu(p) != src_cpu) 4619 goto done; 4620 /* Affinity changed (again). */ 4621 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4622 goto fail; 4623 4624 /* 4625 * If we're not on a rq, the next wake-up will ensure we're 4626 * placed properly. 4627 */ 4628 if (p->on_rq) { 4629 dequeue_task(rq_src, p, 0); 4630 set_task_cpu(p, dest_cpu); 4631 enqueue_task(rq_dest, p, 0); 4632 check_preempt_curr(rq_dest, p, 0); 4633 } 4634 done: 4635 ret = 1; 4636 fail: 4637 double_rq_unlock(rq_src, rq_dest); 4638 raw_spin_unlock(&p->pi_lock); 4639 return ret; 4640 } 4641 4642 #ifdef CONFIG_NUMA_BALANCING 4643 /* Migrate current task p to target_cpu */ 4644 int migrate_task_to(struct task_struct *p, int target_cpu) 4645 { 4646 struct migration_arg arg = { p, target_cpu }; 4647 int curr_cpu = task_cpu(p); 4648 4649 if (curr_cpu == target_cpu) 4650 return 0; 4651 4652 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4653 return -EINVAL; 4654 4655 /* TODO: This is not properly updating schedstats */ 4656 4657 trace_sched_move_numa(p, curr_cpu, target_cpu); 4658 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4659 } 4660 4661 /* 4662 * Requeue a task on a given node and accurately track the number of NUMA 4663 * tasks on the runqueues 4664 */ 4665 void sched_setnuma(struct task_struct *p, int nid) 4666 { 4667 struct rq *rq; 4668 unsigned long flags; 4669 bool on_rq, running; 4670 4671 rq = task_rq_lock(p, &flags); 4672 on_rq = p->on_rq; 4673 running = task_current(rq, p); 4674 4675 if (on_rq) 4676 dequeue_task(rq, p, 0); 4677 if (running) 4678 p->sched_class->put_prev_task(rq, p); 4679 4680 p->numa_preferred_nid = nid; 4681 4682 if (running) 4683 p->sched_class->set_curr_task(rq); 4684 if (on_rq) 4685 enqueue_task(rq, p, 0); 4686 task_rq_unlock(rq, p, &flags); 4687 } 4688 #endif 4689 4690 /* 4691 * migration_cpu_stop - this will be executed by a highprio stopper thread 4692 * and performs thread migration by bumping thread off CPU then 4693 * 'pushing' onto another runqueue. 4694 */ 4695 static int migration_cpu_stop(void *data) 4696 { 4697 struct migration_arg *arg = data; 4698 4699 /* 4700 * The original target cpu might have gone down and we might 4701 * be on another cpu but it doesn't matter. 4702 */ 4703 local_irq_disable(); 4704 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4705 local_irq_enable(); 4706 return 0; 4707 } 4708 4709 #ifdef CONFIG_HOTPLUG_CPU 4710 4711 /* 4712 * Ensures that the idle task is using init_mm right before its cpu goes 4713 * offline. 4714 */ 4715 void idle_task_exit(void) 4716 { 4717 struct mm_struct *mm = current->active_mm; 4718 4719 BUG_ON(cpu_online(smp_processor_id())); 4720 4721 if (mm != &init_mm) { 4722 switch_mm(mm, &init_mm, current); 4723 finish_arch_post_lock_switch(); 4724 } 4725 mmdrop(mm); 4726 } 4727 4728 /* 4729 * Since this CPU is going 'away' for a while, fold any nr_active delta 4730 * we might have. Assumes we're called after migrate_tasks() so that the 4731 * nr_active count is stable. 4732 * 4733 * Also see the comment "Global load-average calculations". 4734 */ 4735 static void calc_load_migrate(struct rq *rq) 4736 { 4737 long delta = calc_load_fold_active(rq); 4738 if (delta) 4739 atomic_long_add(delta, &calc_load_tasks); 4740 } 4741 4742 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4743 { 4744 } 4745 4746 static const struct sched_class fake_sched_class = { 4747 .put_prev_task = put_prev_task_fake, 4748 }; 4749 4750 static struct task_struct fake_task = { 4751 /* 4752 * Avoid pull_{rt,dl}_task() 4753 */ 4754 .prio = MAX_PRIO + 1, 4755 .sched_class = &fake_sched_class, 4756 }; 4757 4758 /* 4759 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4760 * try_to_wake_up()->select_task_rq(). 4761 * 4762 * Called with rq->lock held even though we'er in stop_machine() and 4763 * there's no concurrency possible, we hold the required locks anyway 4764 * because of lock validation efforts. 4765 */ 4766 static void migrate_tasks(unsigned int dead_cpu) 4767 { 4768 struct rq *rq = cpu_rq(dead_cpu); 4769 struct task_struct *next, *stop = rq->stop; 4770 int dest_cpu; 4771 4772 /* 4773 * Fudge the rq selection such that the below task selection loop 4774 * doesn't get stuck on the currently eligible stop task. 4775 * 4776 * We're currently inside stop_machine() and the rq is either stuck 4777 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4778 * either way we should never end up calling schedule() until we're 4779 * done here. 4780 */ 4781 rq->stop = NULL; 4782 4783 /* 4784 * put_prev_task() and pick_next_task() sched 4785 * class method both need to have an up-to-date 4786 * value of rq->clock[_task] 4787 */ 4788 update_rq_clock(rq); 4789 4790 for ( ; ; ) { 4791 /* 4792 * There's this thread running, bail when that's the only 4793 * remaining thread. 4794 */ 4795 if (rq->nr_running == 1) 4796 break; 4797 4798 next = pick_next_task(rq, &fake_task); 4799 BUG_ON(!next); 4800 next->sched_class->put_prev_task(rq, next); 4801 4802 /* Find suitable destination for @next, with force if needed. */ 4803 dest_cpu = select_fallback_rq(dead_cpu, next); 4804 raw_spin_unlock(&rq->lock); 4805 4806 __migrate_task(next, dead_cpu, dest_cpu); 4807 4808 raw_spin_lock(&rq->lock); 4809 } 4810 4811 rq->stop = stop; 4812 } 4813 4814 #endif /* CONFIG_HOTPLUG_CPU */ 4815 4816 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4817 4818 static struct ctl_table sd_ctl_dir[] = { 4819 { 4820 .procname = "sched_domain", 4821 .mode = 0555, 4822 }, 4823 {} 4824 }; 4825 4826 static struct ctl_table sd_ctl_root[] = { 4827 { 4828 .procname = "kernel", 4829 .mode = 0555, 4830 .child = sd_ctl_dir, 4831 }, 4832 {} 4833 }; 4834 4835 static struct ctl_table *sd_alloc_ctl_entry(int n) 4836 { 4837 struct ctl_table *entry = 4838 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4839 4840 return entry; 4841 } 4842 4843 static void sd_free_ctl_entry(struct ctl_table **tablep) 4844 { 4845 struct ctl_table *entry; 4846 4847 /* 4848 * In the intermediate directories, both the child directory and 4849 * procname are dynamically allocated and could fail but the mode 4850 * will always be set. In the lowest directory the names are 4851 * static strings and all have proc handlers. 4852 */ 4853 for (entry = *tablep; entry->mode; entry++) { 4854 if (entry->child) 4855 sd_free_ctl_entry(&entry->child); 4856 if (entry->proc_handler == NULL) 4857 kfree(entry->procname); 4858 } 4859 4860 kfree(*tablep); 4861 *tablep = NULL; 4862 } 4863 4864 static int min_load_idx = 0; 4865 static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4866 4867 static void 4868 set_table_entry(struct ctl_table *entry, 4869 const char *procname, void *data, int maxlen, 4870 umode_t mode, proc_handler *proc_handler, 4871 bool load_idx) 4872 { 4873 entry->procname = procname; 4874 entry->data = data; 4875 entry->maxlen = maxlen; 4876 entry->mode = mode; 4877 entry->proc_handler = proc_handler; 4878 4879 if (load_idx) { 4880 entry->extra1 = &min_load_idx; 4881 entry->extra2 = &max_load_idx; 4882 } 4883 } 4884 4885 static struct ctl_table * 4886 sd_alloc_ctl_domain_table(struct sched_domain *sd) 4887 { 4888 struct ctl_table *table = sd_alloc_ctl_entry(14); 4889 4890 if (table == NULL) 4891 return NULL; 4892 4893 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4894 sizeof(long), 0644, proc_doulongvec_minmax, false); 4895 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4896 sizeof(long), 0644, proc_doulongvec_minmax, false); 4897 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4898 sizeof(int), 0644, proc_dointvec_minmax, true); 4899 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4900 sizeof(int), 0644, proc_dointvec_minmax, true); 4901 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4902 sizeof(int), 0644, proc_dointvec_minmax, true); 4903 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4904 sizeof(int), 0644, proc_dointvec_minmax, true); 4905 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4906 sizeof(int), 0644, proc_dointvec_minmax, true); 4907 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4908 sizeof(int), 0644, proc_dointvec_minmax, false); 4909 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4910 sizeof(int), 0644, proc_dointvec_minmax, false); 4911 set_table_entry(&table[9], "cache_nice_tries", 4912 &sd->cache_nice_tries, 4913 sizeof(int), 0644, proc_dointvec_minmax, false); 4914 set_table_entry(&table[10], "flags", &sd->flags, 4915 sizeof(int), 0644, proc_dointvec_minmax, false); 4916 set_table_entry(&table[11], "max_newidle_lb_cost", 4917 &sd->max_newidle_lb_cost, 4918 sizeof(long), 0644, proc_doulongvec_minmax, false); 4919 set_table_entry(&table[12], "name", sd->name, 4920 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4921 /* &table[13] is terminator */ 4922 4923 return table; 4924 } 4925 4926 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4927 { 4928 struct ctl_table *entry, *table; 4929 struct sched_domain *sd; 4930 int domain_num = 0, i; 4931 char buf[32]; 4932 4933 for_each_domain(cpu, sd) 4934 domain_num++; 4935 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4936 if (table == NULL) 4937 return NULL; 4938 4939 i = 0; 4940 for_each_domain(cpu, sd) { 4941 snprintf(buf, 32, "domain%d", i); 4942 entry->procname = kstrdup(buf, GFP_KERNEL); 4943 entry->mode = 0555; 4944 entry->child = sd_alloc_ctl_domain_table(sd); 4945 entry++; 4946 i++; 4947 } 4948 return table; 4949 } 4950 4951 static struct ctl_table_header *sd_sysctl_header; 4952 static void register_sched_domain_sysctl(void) 4953 { 4954 int i, cpu_num = num_possible_cpus(); 4955 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4956 char buf[32]; 4957 4958 WARN_ON(sd_ctl_dir[0].child); 4959 sd_ctl_dir[0].child = entry; 4960 4961 if (entry == NULL) 4962 return; 4963 4964 for_each_possible_cpu(i) { 4965 snprintf(buf, 32, "cpu%d", i); 4966 entry->procname = kstrdup(buf, GFP_KERNEL); 4967 entry->mode = 0555; 4968 entry->child = sd_alloc_ctl_cpu_table(i); 4969 entry++; 4970 } 4971 4972 WARN_ON(sd_sysctl_header); 4973 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4974 } 4975 4976 /* may be called multiple times per register */ 4977 static void unregister_sched_domain_sysctl(void) 4978 { 4979 if (sd_sysctl_header) 4980 unregister_sysctl_table(sd_sysctl_header); 4981 sd_sysctl_header = NULL; 4982 if (sd_ctl_dir[0].child) 4983 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4984 } 4985 #else 4986 static void register_sched_domain_sysctl(void) 4987 { 4988 } 4989 static void unregister_sched_domain_sysctl(void) 4990 { 4991 } 4992 #endif 4993 4994 static void set_rq_online(struct rq *rq) 4995 { 4996 if (!rq->online) { 4997 const struct sched_class *class; 4998 4999 cpumask_set_cpu(rq->cpu, rq->rd->online); 5000 rq->online = 1; 5001 5002 for_each_class(class) { 5003 if (class->rq_online) 5004 class->rq_online(rq); 5005 } 5006 } 5007 } 5008 5009 static void set_rq_offline(struct rq *rq) 5010 { 5011 if (rq->online) { 5012 const struct sched_class *class; 5013 5014 for_each_class(class) { 5015 if (class->rq_offline) 5016 class->rq_offline(rq); 5017 } 5018 5019 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5020 rq->online = 0; 5021 } 5022 } 5023 5024 /* 5025 * migration_call - callback that gets triggered when a CPU is added. 5026 * Here we can start up the necessary migration thread for the new CPU. 5027 */ 5028 static int 5029 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5030 { 5031 int cpu = (long)hcpu; 5032 unsigned long flags; 5033 struct rq *rq = cpu_rq(cpu); 5034 5035 switch (action & ~CPU_TASKS_FROZEN) { 5036 5037 case CPU_UP_PREPARE: 5038 rq->calc_load_update = calc_load_update; 5039 break; 5040 5041 case CPU_ONLINE: 5042 /* Update our root-domain */ 5043 raw_spin_lock_irqsave(&rq->lock, flags); 5044 if (rq->rd) { 5045 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5046 5047 set_rq_online(rq); 5048 } 5049 raw_spin_unlock_irqrestore(&rq->lock, flags); 5050 break; 5051 5052 #ifdef CONFIG_HOTPLUG_CPU 5053 case CPU_DYING: 5054 sched_ttwu_pending(); 5055 /* Update our root-domain */ 5056 raw_spin_lock_irqsave(&rq->lock, flags); 5057 if (rq->rd) { 5058 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5059 set_rq_offline(rq); 5060 } 5061 migrate_tasks(cpu); 5062 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5063 raw_spin_unlock_irqrestore(&rq->lock, flags); 5064 break; 5065 5066 case CPU_DEAD: 5067 calc_load_migrate(rq); 5068 break; 5069 #endif 5070 } 5071 5072 update_max_interval(); 5073 5074 return NOTIFY_OK; 5075 } 5076 5077 /* 5078 * Register at high priority so that task migration (migrate_all_tasks) 5079 * happens before everything else. This has to be lower priority than 5080 * the notifier in the perf_event subsystem, though. 5081 */ 5082 static struct notifier_block migration_notifier = { 5083 .notifier_call = migration_call, 5084 .priority = CPU_PRI_MIGRATION, 5085 }; 5086 5087 static int sched_cpu_active(struct notifier_block *nfb, 5088 unsigned long action, void *hcpu) 5089 { 5090 switch (action & ~CPU_TASKS_FROZEN) { 5091 case CPU_STARTING: 5092 case CPU_DOWN_FAILED: 5093 set_cpu_active((long)hcpu, true); 5094 return NOTIFY_OK; 5095 default: 5096 return NOTIFY_DONE; 5097 } 5098 } 5099 5100 static int sched_cpu_inactive(struct notifier_block *nfb, 5101 unsigned long action, void *hcpu) 5102 { 5103 unsigned long flags; 5104 long cpu = (long)hcpu; 5105 5106 switch (action & ~CPU_TASKS_FROZEN) { 5107 case CPU_DOWN_PREPARE: 5108 set_cpu_active(cpu, false); 5109 5110 /* explicitly allow suspend */ 5111 if (!(action & CPU_TASKS_FROZEN)) { 5112 struct dl_bw *dl_b = dl_bw_of(cpu); 5113 bool overflow; 5114 int cpus; 5115 5116 raw_spin_lock_irqsave(&dl_b->lock, flags); 5117 cpus = dl_bw_cpus(cpu); 5118 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5119 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5120 5121 if (overflow) 5122 return notifier_from_errno(-EBUSY); 5123 } 5124 return NOTIFY_OK; 5125 } 5126 5127 return NOTIFY_DONE; 5128 } 5129 5130 static int __init migration_init(void) 5131 { 5132 void *cpu = (void *)(long)smp_processor_id(); 5133 int err; 5134 5135 /* Initialize migration for the boot CPU */ 5136 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5137 BUG_ON(err == NOTIFY_BAD); 5138 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5139 register_cpu_notifier(&migration_notifier); 5140 5141 /* Register cpu active notifiers */ 5142 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5143 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5144 5145 return 0; 5146 } 5147 early_initcall(migration_init); 5148 #endif 5149 5150 #ifdef CONFIG_SMP 5151 5152 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5153 5154 #ifdef CONFIG_SCHED_DEBUG 5155 5156 static __read_mostly int sched_debug_enabled; 5157 5158 static int __init sched_debug_setup(char *str) 5159 { 5160 sched_debug_enabled = 1; 5161 5162 return 0; 5163 } 5164 early_param("sched_debug", sched_debug_setup); 5165 5166 static inline bool sched_debug(void) 5167 { 5168 return sched_debug_enabled; 5169 } 5170 5171 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5172 struct cpumask *groupmask) 5173 { 5174 struct sched_group *group = sd->groups; 5175 char str[256]; 5176 5177 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5178 cpumask_clear(groupmask); 5179 5180 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5181 5182 if (!(sd->flags & SD_LOAD_BALANCE)) { 5183 printk("does not load-balance\n"); 5184 if (sd->parent) 5185 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5186 " has parent"); 5187 return -1; 5188 } 5189 5190 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5191 5192 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5193 printk(KERN_ERR "ERROR: domain->span does not contain " 5194 "CPU%d\n", cpu); 5195 } 5196 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5197 printk(KERN_ERR "ERROR: domain->groups does not contain" 5198 " CPU%d\n", cpu); 5199 } 5200 5201 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5202 do { 5203 if (!group) { 5204 printk("\n"); 5205 printk(KERN_ERR "ERROR: group is NULL\n"); 5206 break; 5207 } 5208 5209 /* 5210 * Even though we initialize ->power to something semi-sane, 5211 * we leave power_orig unset. This allows us to detect if 5212 * domain iteration is still funny without causing /0 traps. 5213 */ 5214 if (!group->sgp->power_orig) { 5215 printk(KERN_CONT "\n"); 5216 printk(KERN_ERR "ERROR: domain->cpu_power not " 5217 "set\n"); 5218 break; 5219 } 5220 5221 if (!cpumask_weight(sched_group_cpus(group))) { 5222 printk(KERN_CONT "\n"); 5223 printk(KERN_ERR "ERROR: empty group\n"); 5224 break; 5225 } 5226 5227 if (!(sd->flags & SD_OVERLAP) && 5228 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5229 printk(KERN_CONT "\n"); 5230 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5231 break; 5232 } 5233 5234 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5235 5236 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5237 5238 printk(KERN_CONT " %s", str); 5239 if (group->sgp->power != SCHED_POWER_SCALE) { 5240 printk(KERN_CONT " (cpu_power = %d)", 5241 group->sgp->power); 5242 } 5243 5244 group = group->next; 5245 } while (group != sd->groups); 5246 printk(KERN_CONT "\n"); 5247 5248 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5249 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5250 5251 if (sd->parent && 5252 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5253 printk(KERN_ERR "ERROR: parent span is not a superset " 5254 "of domain->span\n"); 5255 return 0; 5256 } 5257 5258 static void sched_domain_debug(struct sched_domain *sd, int cpu) 5259 { 5260 int level = 0; 5261 5262 if (!sched_debug_enabled) 5263 return; 5264 5265 if (!sd) { 5266 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5267 return; 5268 } 5269 5270 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5271 5272 for (;;) { 5273 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5274 break; 5275 level++; 5276 sd = sd->parent; 5277 if (!sd) 5278 break; 5279 } 5280 } 5281 #else /* !CONFIG_SCHED_DEBUG */ 5282 # define sched_domain_debug(sd, cpu) do { } while (0) 5283 static inline bool sched_debug(void) 5284 { 5285 return false; 5286 } 5287 #endif /* CONFIG_SCHED_DEBUG */ 5288 5289 static int sd_degenerate(struct sched_domain *sd) 5290 { 5291 if (cpumask_weight(sched_domain_span(sd)) == 1) 5292 return 1; 5293 5294 /* Following flags need at least 2 groups */ 5295 if (sd->flags & (SD_LOAD_BALANCE | 5296 SD_BALANCE_NEWIDLE | 5297 SD_BALANCE_FORK | 5298 SD_BALANCE_EXEC | 5299 SD_SHARE_CPUPOWER | 5300 SD_SHARE_PKG_RESOURCES)) { 5301 if (sd->groups != sd->groups->next) 5302 return 0; 5303 } 5304 5305 /* Following flags don't use groups */ 5306 if (sd->flags & (SD_WAKE_AFFINE)) 5307 return 0; 5308 5309 return 1; 5310 } 5311 5312 static int 5313 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5314 { 5315 unsigned long cflags = sd->flags, pflags = parent->flags; 5316 5317 if (sd_degenerate(parent)) 5318 return 1; 5319 5320 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5321 return 0; 5322 5323 /* Flags needing groups don't count if only 1 group in parent */ 5324 if (parent->groups == parent->groups->next) { 5325 pflags &= ~(SD_LOAD_BALANCE | 5326 SD_BALANCE_NEWIDLE | 5327 SD_BALANCE_FORK | 5328 SD_BALANCE_EXEC | 5329 SD_SHARE_CPUPOWER | 5330 SD_SHARE_PKG_RESOURCES | 5331 SD_PREFER_SIBLING); 5332 if (nr_node_ids == 1) 5333 pflags &= ~SD_SERIALIZE; 5334 } 5335 if (~cflags & pflags) 5336 return 0; 5337 5338 return 1; 5339 } 5340 5341 static void free_rootdomain(struct rcu_head *rcu) 5342 { 5343 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5344 5345 cpupri_cleanup(&rd->cpupri); 5346 cpudl_cleanup(&rd->cpudl); 5347 free_cpumask_var(rd->dlo_mask); 5348 free_cpumask_var(rd->rto_mask); 5349 free_cpumask_var(rd->online); 5350 free_cpumask_var(rd->span); 5351 kfree(rd); 5352 } 5353 5354 static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5355 { 5356 struct root_domain *old_rd = NULL; 5357 unsigned long flags; 5358 5359 raw_spin_lock_irqsave(&rq->lock, flags); 5360 5361 if (rq->rd) { 5362 old_rd = rq->rd; 5363 5364 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5365 set_rq_offline(rq); 5366 5367 cpumask_clear_cpu(rq->cpu, old_rd->span); 5368 5369 /* 5370 * If we dont want to free the old_rd yet then 5371 * set old_rd to NULL to skip the freeing later 5372 * in this function: 5373 */ 5374 if (!atomic_dec_and_test(&old_rd->refcount)) 5375 old_rd = NULL; 5376 } 5377 5378 atomic_inc(&rd->refcount); 5379 rq->rd = rd; 5380 5381 cpumask_set_cpu(rq->cpu, rd->span); 5382 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5383 set_rq_online(rq); 5384 5385 raw_spin_unlock_irqrestore(&rq->lock, flags); 5386 5387 if (old_rd) 5388 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5389 } 5390 5391 static int init_rootdomain(struct root_domain *rd) 5392 { 5393 memset(rd, 0, sizeof(*rd)); 5394 5395 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5396 goto out; 5397 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5398 goto free_span; 5399 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5400 goto free_online; 5401 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5402 goto free_dlo_mask; 5403 5404 init_dl_bw(&rd->dl_bw); 5405 if (cpudl_init(&rd->cpudl) != 0) 5406 goto free_dlo_mask; 5407 5408 if (cpupri_init(&rd->cpupri) != 0) 5409 goto free_rto_mask; 5410 return 0; 5411 5412 free_rto_mask: 5413 free_cpumask_var(rd->rto_mask); 5414 free_dlo_mask: 5415 free_cpumask_var(rd->dlo_mask); 5416 free_online: 5417 free_cpumask_var(rd->online); 5418 free_span: 5419 free_cpumask_var(rd->span); 5420 out: 5421 return -ENOMEM; 5422 } 5423 5424 /* 5425 * By default the system creates a single root-domain with all cpus as 5426 * members (mimicking the global state we have today). 5427 */ 5428 struct root_domain def_root_domain; 5429 5430 static void init_defrootdomain(void) 5431 { 5432 init_rootdomain(&def_root_domain); 5433 5434 atomic_set(&def_root_domain.refcount, 1); 5435 } 5436 5437 static struct root_domain *alloc_rootdomain(void) 5438 { 5439 struct root_domain *rd; 5440 5441 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5442 if (!rd) 5443 return NULL; 5444 5445 if (init_rootdomain(rd) != 0) { 5446 kfree(rd); 5447 return NULL; 5448 } 5449 5450 return rd; 5451 } 5452 5453 static void free_sched_groups(struct sched_group *sg, int free_sgp) 5454 { 5455 struct sched_group *tmp, *first; 5456 5457 if (!sg) 5458 return; 5459 5460 first = sg; 5461 do { 5462 tmp = sg->next; 5463 5464 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5465 kfree(sg->sgp); 5466 5467 kfree(sg); 5468 sg = tmp; 5469 } while (sg != first); 5470 } 5471 5472 static void free_sched_domain(struct rcu_head *rcu) 5473 { 5474 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5475 5476 /* 5477 * If its an overlapping domain it has private groups, iterate and 5478 * nuke them all. 5479 */ 5480 if (sd->flags & SD_OVERLAP) { 5481 free_sched_groups(sd->groups, 1); 5482 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5483 kfree(sd->groups->sgp); 5484 kfree(sd->groups); 5485 } 5486 kfree(sd); 5487 } 5488 5489 static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5490 { 5491 call_rcu(&sd->rcu, free_sched_domain); 5492 } 5493 5494 static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5495 { 5496 for (; sd; sd = sd->parent) 5497 destroy_sched_domain(sd, cpu); 5498 } 5499 5500 /* 5501 * Keep a special pointer to the highest sched_domain that has 5502 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5503 * allows us to avoid some pointer chasing select_idle_sibling(). 5504 * 5505 * Also keep a unique ID per domain (we use the first cpu number in 5506 * the cpumask of the domain), this allows us to quickly tell if 5507 * two cpus are in the same cache domain, see cpus_share_cache(). 5508 */ 5509 DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5510 DEFINE_PER_CPU(int, sd_llc_size); 5511 DEFINE_PER_CPU(int, sd_llc_id); 5512 DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5513 DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5514 DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5515 5516 static void update_top_cache_domain(int cpu) 5517 { 5518 struct sched_domain *sd; 5519 struct sched_domain *busy_sd = NULL; 5520 int id = cpu; 5521 int size = 1; 5522 5523 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5524 if (sd) { 5525 id = cpumask_first(sched_domain_span(sd)); 5526 size = cpumask_weight(sched_domain_span(sd)); 5527 busy_sd = sd->parent; /* sd_busy */ 5528 } 5529 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5530 5531 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5532 per_cpu(sd_llc_size, cpu) = size; 5533 per_cpu(sd_llc_id, cpu) = id; 5534 5535 sd = lowest_flag_domain(cpu, SD_NUMA); 5536 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5537 5538 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5539 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5540 } 5541 5542 /* 5543 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5544 * hold the hotplug lock. 5545 */ 5546 static void 5547 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5548 { 5549 struct rq *rq = cpu_rq(cpu); 5550 struct sched_domain *tmp; 5551 5552 /* Remove the sched domains which do not contribute to scheduling. */ 5553 for (tmp = sd; tmp; ) { 5554 struct sched_domain *parent = tmp->parent; 5555 if (!parent) 5556 break; 5557 5558 if (sd_parent_degenerate(tmp, parent)) { 5559 tmp->parent = parent->parent; 5560 if (parent->parent) 5561 parent->parent->child = tmp; 5562 /* 5563 * Transfer SD_PREFER_SIBLING down in case of a 5564 * degenerate parent; the spans match for this 5565 * so the property transfers. 5566 */ 5567 if (parent->flags & SD_PREFER_SIBLING) 5568 tmp->flags |= SD_PREFER_SIBLING; 5569 destroy_sched_domain(parent, cpu); 5570 } else 5571 tmp = tmp->parent; 5572 } 5573 5574 if (sd && sd_degenerate(sd)) { 5575 tmp = sd; 5576 sd = sd->parent; 5577 destroy_sched_domain(tmp, cpu); 5578 if (sd) 5579 sd->child = NULL; 5580 } 5581 5582 sched_domain_debug(sd, cpu); 5583 5584 rq_attach_root(rq, rd); 5585 tmp = rq->sd; 5586 rcu_assign_pointer(rq->sd, sd); 5587 destroy_sched_domains(tmp, cpu); 5588 5589 update_top_cache_domain(cpu); 5590 } 5591 5592 /* cpus with isolated domains */ 5593 static cpumask_var_t cpu_isolated_map; 5594 5595 /* Setup the mask of cpus configured for isolated domains */ 5596 static int __init isolated_cpu_setup(char *str) 5597 { 5598 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5599 cpulist_parse(str, cpu_isolated_map); 5600 return 1; 5601 } 5602 5603 __setup("isolcpus=", isolated_cpu_setup); 5604 5605 static const struct cpumask *cpu_cpu_mask(int cpu) 5606 { 5607 return cpumask_of_node(cpu_to_node(cpu)); 5608 } 5609 5610 struct sd_data { 5611 struct sched_domain **__percpu sd; 5612 struct sched_group **__percpu sg; 5613 struct sched_group_power **__percpu sgp; 5614 }; 5615 5616 struct s_data { 5617 struct sched_domain ** __percpu sd; 5618 struct root_domain *rd; 5619 }; 5620 5621 enum s_alloc { 5622 sa_rootdomain, 5623 sa_sd, 5624 sa_sd_storage, 5625 sa_none, 5626 }; 5627 5628 struct sched_domain_topology_level; 5629 5630 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5631 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5632 5633 #define SDTL_OVERLAP 0x01 5634 5635 struct sched_domain_topology_level { 5636 sched_domain_init_f init; 5637 sched_domain_mask_f mask; 5638 int flags; 5639 int numa_level; 5640 struct sd_data data; 5641 }; 5642 5643 /* 5644 * Build an iteration mask that can exclude certain CPUs from the upwards 5645 * domain traversal. 5646 * 5647 * Asymmetric node setups can result in situations where the domain tree is of 5648 * unequal depth, make sure to skip domains that already cover the entire 5649 * range. 5650 * 5651 * In that case build_sched_domains() will have terminated the iteration early 5652 * and our sibling sd spans will be empty. Domains should always include the 5653 * cpu they're built on, so check that. 5654 * 5655 */ 5656 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5657 { 5658 const struct cpumask *span = sched_domain_span(sd); 5659 struct sd_data *sdd = sd->private; 5660 struct sched_domain *sibling; 5661 int i; 5662 5663 for_each_cpu(i, span) { 5664 sibling = *per_cpu_ptr(sdd->sd, i); 5665 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5666 continue; 5667 5668 cpumask_set_cpu(i, sched_group_mask(sg)); 5669 } 5670 } 5671 5672 /* 5673 * Return the canonical balance cpu for this group, this is the first cpu 5674 * of this group that's also in the iteration mask. 5675 */ 5676 int group_balance_cpu(struct sched_group *sg) 5677 { 5678 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5679 } 5680 5681 static int 5682 build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5683 { 5684 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5685 const struct cpumask *span = sched_domain_span(sd); 5686 struct cpumask *covered = sched_domains_tmpmask; 5687 struct sd_data *sdd = sd->private; 5688 struct sched_domain *child; 5689 int i; 5690 5691 cpumask_clear(covered); 5692 5693 for_each_cpu(i, span) { 5694 struct cpumask *sg_span; 5695 5696 if (cpumask_test_cpu(i, covered)) 5697 continue; 5698 5699 child = *per_cpu_ptr(sdd->sd, i); 5700 5701 /* See the comment near build_group_mask(). */ 5702 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5703 continue; 5704 5705 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5706 GFP_KERNEL, cpu_to_node(cpu)); 5707 5708 if (!sg) 5709 goto fail; 5710 5711 sg_span = sched_group_cpus(sg); 5712 if (child->child) { 5713 child = child->child; 5714 cpumask_copy(sg_span, sched_domain_span(child)); 5715 } else 5716 cpumask_set_cpu(i, sg_span); 5717 5718 cpumask_or(covered, covered, sg_span); 5719 5720 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5721 if (atomic_inc_return(&sg->sgp->ref) == 1) 5722 build_group_mask(sd, sg); 5723 5724 /* 5725 * Initialize sgp->power such that even if we mess up the 5726 * domains and no possible iteration will get us here, we won't 5727 * die on a /0 trap. 5728 */ 5729 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5730 sg->sgp->power_orig = sg->sgp->power; 5731 5732 /* 5733 * Make sure the first group of this domain contains the 5734 * canonical balance cpu. Otherwise the sched_domain iteration 5735 * breaks. See update_sg_lb_stats(). 5736 */ 5737 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5738 group_balance_cpu(sg) == cpu) 5739 groups = sg; 5740 5741 if (!first) 5742 first = sg; 5743 if (last) 5744 last->next = sg; 5745 last = sg; 5746 last->next = first; 5747 } 5748 sd->groups = groups; 5749 5750 return 0; 5751 5752 fail: 5753 free_sched_groups(first, 0); 5754 5755 return -ENOMEM; 5756 } 5757 5758 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5759 { 5760 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5761 struct sched_domain *child = sd->child; 5762 5763 if (child) 5764 cpu = cpumask_first(sched_domain_span(child)); 5765 5766 if (sg) { 5767 *sg = *per_cpu_ptr(sdd->sg, cpu); 5768 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5769 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5770 } 5771 5772 return cpu; 5773 } 5774 5775 /* 5776 * build_sched_groups will build a circular linked list of the groups 5777 * covered by the given span, and will set each group's ->cpumask correctly, 5778 * and ->cpu_power to 0. 5779 * 5780 * Assumes the sched_domain tree is fully constructed 5781 */ 5782 static int 5783 build_sched_groups(struct sched_domain *sd, int cpu) 5784 { 5785 struct sched_group *first = NULL, *last = NULL; 5786 struct sd_data *sdd = sd->private; 5787 const struct cpumask *span = sched_domain_span(sd); 5788 struct cpumask *covered; 5789 int i; 5790 5791 get_group(cpu, sdd, &sd->groups); 5792 atomic_inc(&sd->groups->ref); 5793 5794 if (cpu != cpumask_first(span)) 5795 return 0; 5796 5797 lockdep_assert_held(&sched_domains_mutex); 5798 covered = sched_domains_tmpmask; 5799 5800 cpumask_clear(covered); 5801 5802 for_each_cpu(i, span) { 5803 struct sched_group *sg; 5804 int group, j; 5805 5806 if (cpumask_test_cpu(i, covered)) 5807 continue; 5808 5809 group = get_group(i, sdd, &sg); 5810 cpumask_clear(sched_group_cpus(sg)); 5811 sg->sgp->power = 0; 5812 cpumask_setall(sched_group_mask(sg)); 5813 5814 for_each_cpu(j, span) { 5815 if (get_group(j, sdd, NULL) != group) 5816 continue; 5817 5818 cpumask_set_cpu(j, covered); 5819 cpumask_set_cpu(j, sched_group_cpus(sg)); 5820 } 5821 5822 if (!first) 5823 first = sg; 5824 if (last) 5825 last->next = sg; 5826 last = sg; 5827 } 5828 last->next = first; 5829 5830 return 0; 5831 } 5832 5833 /* 5834 * Initialize sched groups cpu_power. 5835 * 5836 * cpu_power indicates the capacity of sched group, which is used while 5837 * distributing the load between different sched groups in a sched domain. 5838 * Typically cpu_power for all the groups in a sched domain will be same unless 5839 * there are asymmetries in the topology. If there are asymmetries, group 5840 * having more cpu_power will pickup more load compared to the group having 5841 * less cpu_power. 5842 */ 5843 static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5844 { 5845 struct sched_group *sg = sd->groups; 5846 5847 WARN_ON(!sg); 5848 5849 do { 5850 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5851 sg = sg->next; 5852 } while (sg != sd->groups); 5853 5854 if (cpu != group_balance_cpu(sg)) 5855 return; 5856 5857 update_group_power(sd, cpu); 5858 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5859 } 5860 5861 int __weak arch_sd_sibling_asym_packing(void) 5862 { 5863 return 0*SD_ASYM_PACKING; 5864 } 5865 5866 /* 5867 * Initializers for schedule domains 5868 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5869 */ 5870 5871 #ifdef CONFIG_SCHED_DEBUG 5872 # define SD_INIT_NAME(sd, type) sd->name = #type 5873 #else 5874 # define SD_INIT_NAME(sd, type) do { } while (0) 5875 #endif 5876 5877 #define SD_INIT_FUNC(type) \ 5878 static noinline struct sched_domain * \ 5879 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5880 { \ 5881 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5882 *sd = SD_##type##_INIT; \ 5883 SD_INIT_NAME(sd, type); \ 5884 sd->private = &tl->data; \ 5885 return sd; \ 5886 } 5887 5888 SD_INIT_FUNC(CPU) 5889 #ifdef CONFIG_SCHED_SMT 5890 SD_INIT_FUNC(SIBLING) 5891 #endif 5892 #ifdef CONFIG_SCHED_MC 5893 SD_INIT_FUNC(MC) 5894 #endif 5895 #ifdef CONFIG_SCHED_BOOK 5896 SD_INIT_FUNC(BOOK) 5897 #endif 5898 5899 static int default_relax_domain_level = -1; 5900 int sched_domain_level_max; 5901 5902 static int __init setup_relax_domain_level(char *str) 5903 { 5904 if (kstrtoint(str, 0, &default_relax_domain_level)) 5905 pr_warn("Unable to set relax_domain_level\n"); 5906 5907 return 1; 5908 } 5909 __setup("relax_domain_level=", setup_relax_domain_level); 5910 5911 static void set_domain_attribute(struct sched_domain *sd, 5912 struct sched_domain_attr *attr) 5913 { 5914 int request; 5915 5916 if (!attr || attr->relax_domain_level < 0) { 5917 if (default_relax_domain_level < 0) 5918 return; 5919 else 5920 request = default_relax_domain_level; 5921 } else 5922 request = attr->relax_domain_level; 5923 if (request < sd->level) { 5924 /* turn off idle balance on this domain */ 5925 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5926 } else { 5927 /* turn on idle balance on this domain */ 5928 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5929 } 5930 } 5931 5932 static void __sdt_free(const struct cpumask *cpu_map); 5933 static int __sdt_alloc(const struct cpumask *cpu_map); 5934 5935 static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5936 const struct cpumask *cpu_map) 5937 { 5938 switch (what) { 5939 case sa_rootdomain: 5940 if (!atomic_read(&d->rd->refcount)) 5941 free_rootdomain(&d->rd->rcu); /* fall through */ 5942 case sa_sd: 5943 free_percpu(d->sd); /* fall through */ 5944 case sa_sd_storage: 5945 __sdt_free(cpu_map); /* fall through */ 5946 case sa_none: 5947 break; 5948 } 5949 } 5950 5951 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5952 const struct cpumask *cpu_map) 5953 { 5954 memset(d, 0, sizeof(*d)); 5955 5956 if (__sdt_alloc(cpu_map)) 5957 return sa_sd_storage; 5958 d->sd = alloc_percpu(struct sched_domain *); 5959 if (!d->sd) 5960 return sa_sd_storage; 5961 d->rd = alloc_rootdomain(); 5962 if (!d->rd) 5963 return sa_sd; 5964 return sa_rootdomain; 5965 } 5966 5967 /* 5968 * NULL the sd_data elements we've used to build the sched_domain and 5969 * sched_group structure so that the subsequent __free_domain_allocs() 5970 * will not free the data we're using. 5971 */ 5972 static void claim_allocations(int cpu, struct sched_domain *sd) 5973 { 5974 struct sd_data *sdd = sd->private; 5975 5976 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5977 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5978 5979 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5980 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5981 5982 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5983 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5984 } 5985 5986 #ifdef CONFIG_SCHED_SMT 5987 static const struct cpumask *cpu_smt_mask(int cpu) 5988 { 5989 return topology_thread_cpumask(cpu); 5990 } 5991 #endif 5992 5993 /* 5994 * Topology list, bottom-up. 5995 */ 5996 static struct sched_domain_topology_level default_topology[] = { 5997 #ifdef CONFIG_SCHED_SMT 5998 { sd_init_SIBLING, cpu_smt_mask, }, 5999 #endif 6000 #ifdef CONFIG_SCHED_MC 6001 { sd_init_MC, cpu_coregroup_mask, }, 6002 #endif 6003 #ifdef CONFIG_SCHED_BOOK 6004 { sd_init_BOOK, cpu_book_mask, }, 6005 #endif 6006 { sd_init_CPU, cpu_cpu_mask, }, 6007 { NULL, }, 6008 }; 6009 6010 static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6011 6012 #define for_each_sd_topology(tl) \ 6013 for (tl = sched_domain_topology; tl->init; tl++) 6014 6015 #ifdef CONFIG_NUMA 6016 6017 static int sched_domains_numa_levels; 6018 static int *sched_domains_numa_distance; 6019 static struct cpumask ***sched_domains_numa_masks; 6020 static int sched_domains_curr_level; 6021 6022 static inline int sd_local_flags(int level) 6023 { 6024 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 6025 return 0; 6026 6027 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 6028 } 6029 6030 static struct sched_domain * 6031 sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 6032 { 6033 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6034 int level = tl->numa_level; 6035 int sd_weight = cpumask_weight( 6036 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 6037 6038 *sd = (struct sched_domain){ 6039 .min_interval = sd_weight, 6040 .max_interval = 2*sd_weight, 6041 .busy_factor = 32, 6042 .imbalance_pct = 125, 6043 .cache_nice_tries = 2, 6044 .busy_idx = 3, 6045 .idle_idx = 2, 6046 .newidle_idx = 0, 6047 .wake_idx = 0, 6048 .forkexec_idx = 0, 6049 6050 .flags = 1*SD_LOAD_BALANCE 6051 | 1*SD_BALANCE_NEWIDLE 6052 | 0*SD_BALANCE_EXEC 6053 | 0*SD_BALANCE_FORK 6054 | 0*SD_BALANCE_WAKE 6055 | 0*SD_WAKE_AFFINE 6056 | 0*SD_SHARE_CPUPOWER 6057 | 0*SD_SHARE_PKG_RESOURCES 6058 | 1*SD_SERIALIZE 6059 | 0*SD_PREFER_SIBLING 6060 | 1*SD_NUMA 6061 | sd_local_flags(level) 6062 , 6063 .last_balance = jiffies, 6064 .balance_interval = sd_weight, 6065 }; 6066 SD_INIT_NAME(sd, NUMA); 6067 sd->private = &tl->data; 6068 6069 /* 6070 * Ugly hack to pass state to sd_numa_mask()... 6071 */ 6072 sched_domains_curr_level = tl->numa_level; 6073 6074 return sd; 6075 } 6076 6077 static const struct cpumask *sd_numa_mask(int cpu) 6078 { 6079 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6080 } 6081 6082 static void sched_numa_warn(const char *str) 6083 { 6084 static int done = false; 6085 int i,j; 6086 6087 if (done) 6088 return; 6089 6090 done = true; 6091 6092 printk(KERN_WARNING "ERROR: %s\n\n", str); 6093 6094 for (i = 0; i < nr_node_ids; i++) { 6095 printk(KERN_WARNING " "); 6096 for (j = 0; j < nr_node_ids; j++) 6097 printk(KERN_CONT "%02d ", node_distance(i,j)); 6098 printk(KERN_CONT "\n"); 6099 } 6100 printk(KERN_WARNING "\n"); 6101 } 6102 6103 static bool find_numa_distance(int distance) 6104 { 6105 int i; 6106 6107 if (distance == node_distance(0, 0)) 6108 return true; 6109 6110 for (i = 0; i < sched_domains_numa_levels; i++) { 6111 if (sched_domains_numa_distance[i] == distance) 6112 return true; 6113 } 6114 6115 return false; 6116 } 6117 6118 static void sched_init_numa(void) 6119 { 6120 int next_distance, curr_distance = node_distance(0, 0); 6121 struct sched_domain_topology_level *tl; 6122 int level = 0; 6123 int i, j, k; 6124 6125 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6126 if (!sched_domains_numa_distance) 6127 return; 6128 6129 /* 6130 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6131 * unique distances in the node_distance() table. 6132 * 6133 * Assumes node_distance(0,j) includes all distances in 6134 * node_distance(i,j) in order to avoid cubic time. 6135 */ 6136 next_distance = curr_distance; 6137 for (i = 0; i < nr_node_ids; i++) { 6138 for (j = 0; j < nr_node_ids; j++) { 6139 for (k = 0; k < nr_node_ids; k++) { 6140 int distance = node_distance(i, k); 6141 6142 if (distance > curr_distance && 6143 (distance < next_distance || 6144 next_distance == curr_distance)) 6145 next_distance = distance; 6146 6147 /* 6148 * While not a strong assumption it would be nice to know 6149 * about cases where if node A is connected to B, B is not 6150 * equally connected to A. 6151 */ 6152 if (sched_debug() && node_distance(k, i) != distance) 6153 sched_numa_warn("Node-distance not symmetric"); 6154 6155 if (sched_debug() && i && !find_numa_distance(distance)) 6156 sched_numa_warn("Node-0 not representative"); 6157 } 6158 if (next_distance != curr_distance) { 6159 sched_domains_numa_distance[level++] = next_distance; 6160 sched_domains_numa_levels = level; 6161 curr_distance = next_distance; 6162 } else break; 6163 } 6164 6165 /* 6166 * In case of sched_debug() we verify the above assumption. 6167 */ 6168 if (!sched_debug()) 6169 break; 6170 } 6171 /* 6172 * 'level' contains the number of unique distances, excluding the 6173 * identity distance node_distance(i,i). 6174 * 6175 * The sched_domains_numa_distance[] array includes the actual distance 6176 * numbers. 6177 */ 6178 6179 /* 6180 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6181 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6182 * the array will contain less then 'level' members. This could be 6183 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6184 * in other functions. 6185 * 6186 * We reset it to 'level' at the end of this function. 6187 */ 6188 sched_domains_numa_levels = 0; 6189 6190 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6191 if (!sched_domains_numa_masks) 6192 return; 6193 6194 /* 6195 * Now for each level, construct a mask per node which contains all 6196 * cpus of nodes that are that many hops away from us. 6197 */ 6198 for (i = 0; i < level; i++) { 6199 sched_domains_numa_masks[i] = 6200 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6201 if (!sched_domains_numa_masks[i]) 6202 return; 6203 6204 for (j = 0; j < nr_node_ids; j++) { 6205 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6206 if (!mask) 6207 return; 6208 6209 sched_domains_numa_masks[i][j] = mask; 6210 6211 for (k = 0; k < nr_node_ids; k++) { 6212 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6213 continue; 6214 6215 cpumask_or(mask, mask, cpumask_of_node(k)); 6216 } 6217 } 6218 } 6219 6220 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6221 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6222 if (!tl) 6223 return; 6224 6225 /* 6226 * Copy the default topology bits.. 6227 */ 6228 for (i = 0; default_topology[i].init; i++) 6229 tl[i] = default_topology[i]; 6230 6231 /* 6232 * .. and append 'j' levels of NUMA goodness. 6233 */ 6234 for (j = 0; j < level; i++, j++) { 6235 tl[i] = (struct sched_domain_topology_level){ 6236 .init = sd_numa_init, 6237 .mask = sd_numa_mask, 6238 .flags = SDTL_OVERLAP, 6239 .numa_level = j, 6240 }; 6241 } 6242 6243 sched_domain_topology = tl; 6244 6245 sched_domains_numa_levels = level; 6246 } 6247 6248 static void sched_domains_numa_masks_set(int cpu) 6249 { 6250 int i, j; 6251 int node = cpu_to_node(cpu); 6252 6253 for (i = 0; i < sched_domains_numa_levels; i++) { 6254 for (j = 0; j < nr_node_ids; j++) { 6255 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6256 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6257 } 6258 } 6259 } 6260 6261 static void sched_domains_numa_masks_clear(int cpu) 6262 { 6263 int i, j; 6264 for (i = 0; i < sched_domains_numa_levels; i++) { 6265 for (j = 0; j < nr_node_ids; j++) 6266 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6267 } 6268 } 6269 6270 /* 6271 * Update sched_domains_numa_masks[level][node] array when new cpus 6272 * are onlined. 6273 */ 6274 static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6275 unsigned long action, 6276 void *hcpu) 6277 { 6278 int cpu = (long)hcpu; 6279 6280 switch (action & ~CPU_TASKS_FROZEN) { 6281 case CPU_ONLINE: 6282 sched_domains_numa_masks_set(cpu); 6283 break; 6284 6285 case CPU_DEAD: 6286 sched_domains_numa_masks_clear(cpu); 6287 break; 6288 6289 default: 6290 return NOTIFY_DONE; 6291 } 6292 6293 return NOTIFY_OK; 6294 } 6295 #else 6296 static inline void sched_init_numa(void) 6297 { 6298 } 6299 6300 static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6301 unsigned long action, 6302 void *hcpu) 6303 { 6304 return 0; 6305 } 6306 #endif /* CONFIG_NUMA */ 6307 6308 static int __sdt_alloc(const struct cpumask *cpu_map) 6309 { 6310 struct sched_domain_topology_level *tl; 6311 int j; 6312 6313 for_each_sd_topology(tl) { 6314 struct sd_data *sdd = &tl->data; 6315 6316 sdd->sd = alloc_percpu(struct sched_domain *); 6317 if (!sdd->sd) 6318 return -ENOMEM; 6319 6320 sdd->sg = alloc_percpu(struct sched_group *); 6321 if (!sdd->sg) 6322 return -ENOMEM; 6323 6324 sdd->sgp = alloc_percpu(struct sched_group_power *); 6325 if (!sdd->sgp) 6326 return -ENOMEM; 6327 6328 for_each_cpu(j, cpu_map) { 6329 struct sched_domain *sd; 6330 struct sched_group *sg; 6331 struct sched_group_power *sgp; 6332 6333 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6334 GFP_KERNEL, cpu_to_node(j)); 6335 if (!sd) 6336 return -ENOMEM; 6337 6338 *per_cpu_ptr(sdd->sd, j) = sd; 6339 6340 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6341 GFP_KERNEL, cpu_to_node(j)); 6342 if (!sg) 6343 return -ENOMEM; 6344 6345 sg->next = sg; 6346 6347 *per_cpu_ptr(sdd->sg, j) = sg; 6348 6349 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6350 GFP_KERNEL, cpu_to_node(j)); 6351 if (!sgp) 6352 return -ENOMEM; 6353 6354 *per_cpu_ptr(sdd->sgp, j) = sgp; 6355 } 6356 } 6357 6358 return 0; 6359 } 6360 6361 static void __sdt_free(const struct cpumask *cpu_map) 6362 { 6363 struct sched_domain_topology_level *tl; 6364 int j; 6365 6366 for_each_sd_topology(tl) { 6367 struct sd_data *sdd = &tl->data; 6368 6369 for_each_cpu(j, cpu_map) { 6370 struct sched_domain *sd; 6371 6372 if (sdd->sd) { 6373 sd = *per_cpu_ptr(sdd->sd, j); 6374 if (sd && (sd->flags & SD_OVERLAP)) 6375 free_sched_groups(sd->groups, 0); 6376 kfree(*per_cpu_ptr(sdd->sd, j)); 6377 } 6378 6379 if (sdd->sg) 6380 kfree(*per_cpu_ptr(sdd->sg, j)); 6381 if (sdd->sgp) 6382 kfree(*per_cpu_ptr(sdd->sgp, j)); 6383 } 6384 free_percpu(sdd->sd); 6385 sdd->sd = NULL; 6386 free_percpu(sdd->sg); 6387 sdd->sg = NULL; 6388 free_percpu(sdd->sgp); 6389 sdd->sgp = NULL; 6390 } 6391 } 6392 6393 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6394 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6395 struct sched_domain *child, int cpu) 6396 { 6397 struct sched_domain *sd = tl->init(tl, cpu); 6398 if (!sd) 6399 return child; 6400 6401 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6402 if (child) { 6403 sd->level = child->level + 1; 6404 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6405 child->parent = sd; 6406 sd->child = child; 6407 } 6408 set_domain_attribute(sd, attr); 6409 6410 return sd; 6411 } 6412 6413 /* 6414 * Build sched domains for a given set of cpus and attach the sched domains 6415 * to the individual cpus 6416 */ 6417 static int build_sched_domains(const struct cpumask *cpu_map, 6418 struct sched_domain_attr *attr) 6419 { 6420 enum s_alloc alloc_state; 6421 struct sched_domain *sd; 6422 struct s_data d; 6423 int i, ret = -ENOMEM; 6424 6425 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6426 if (alloc_state != sa_rootdomain) 6427 goto error; 6428 6429 /* Set up domains for cpus specified by the cpu_map. */ 6430 for_each_cpu(i, cpu_map) { 6431 struct sched_domain_topology_level *tl; 6432 6433 sd = NULL; 6434 for_each_sd_topology(tl) { 6435 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6436 if (tl == sched_domain_topology) 6437 *per_cpu_ptr(d.sd, i) = sd; 6438 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6439 sd->flags |= SD_OVERLAP; 6440 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6441 break; 6442 } 6443 } 6444 6445 /* Build the groups for the domains */ 6446 for_each_cpu(i, cpu_map) { 6447 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6448 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6449 if (sd->flags & SD_OVERLAP) { 6450 if (build_overlap_sched_groups(sd, i)) 6451 goto error; 6452 } else { 6453 if (build_sched_groups(sd, i)) 6454 goto error; 6455 } 6456 } 6457 } 6458 6459 /* Calculate CPU power for physical packages and nodes */ 6460 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6461 if (!cpumask_test_cpu(i, cpu_map)) 6462 continue; 6463 6464 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6465 claim_allocations(i, sd); 6466 init_sched_groups_power(i, sd); 6467 } 6468 } 6469 6470 /* Attach the domains */ 6471 rcu_read_lock(); 6472 for_each_cpu(i, cpu_map) { 6473 sd = *per_cpu_ptr(d.sd, i); 6474 cpu_attach_domain(sd, d.rd, i); 6475 } 6476 rcu_read_unlock(); 6477 6478 ret = 0; 6479 error: 6480 __free_domain_allocs(&d, alloc_state, cpu_map); 6481 return ret; 6482 } 6483 6484 static cpumask_var_t *doms_cur; /* current sched domains */ 6485 static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6486 static struct sched_domain_attr *dattr_cur; 6487 /* attribues of custom domains in 'doms_cur' */ 6488 6489 /* 6490 * Special case: If a kmalloc of a doms_cur partition (array of 6491 * cpumask) fails, then fallback to a single sched domain, 6492 * as determined by the single cpumask fallback_doms. 6493 */ 6494 static cpumask_var_t fallback_doms; 6495 6496 /* 6497 * arch_update_cpu_topology lets virtualized architectures update the 6498 * cpu core maps. It is supposed to return 1 if the topology changed 6499 * or 0 if it stayed the same. 6500 */ 6501 int __attribute__((weak)) arch_update_cpu_topology(void) 6502 { 6503 return 0; 6504 } 6505 6506 cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6507 { 6508 int i; 6509 cpumask_var_t *doms; 6510 6511 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6512 if (!doms) 6513 return NULL; 6514 for (i = 0; i < ndoms; i++) { 6515 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6516 free_sched_domains(doms, i); 6517 return NULL; 6518 } 6519 } 6520 return doms; 6521 } 6522 6523 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6524 { 6525 unsigned int i; 6526 for (i = 0; i < ndoms; i++) 6527 free_cpumask_var(doms[i]); 6528 kfree(doms); 6529 } 6530 6531 /* 6532 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6533 * For now this just excludes isolated cpus, but could be used to 6534 * exclude other special cases in the future. 6535 */ 6536 static int init_sched_domains(const struct cpumask *cpu_map) 6537 { 6538 int err; 6539 6540 arch_update_cpu_topology(); 6541 ndoms_cur = 1; 6542 doms_cur = alloc_sched_domains(ndoms_cur); 6543 if (!doms_cur) 6544 doms_cur = &fallback_doms; 6545 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6546 err = build_sched_domains(doms_cur[0], NULL); 6547 register_sched_domain_sysctl(); 6548 6549 return err; 6550 } 6551 6552 /* 6553 * Detach sched domains from a group of cpus specified in cpu_map 6554 * These cpus will now be attached to the NULL domain 6555 */ 6556 static void detach_destroy_domains(const struct cpumask *cpu_map) 6557 { 6558 int i; 6559 6560 rcu_read_lock(); 6561 for_each_cpu(i, cpu_map) 6562 cpu_attach_domain(NULL, &def_root_domain, i); 6563 rcu_read_unlock(); 6564 } 6565 6566 /* handle null as "default" */ 6567 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6568 struct sched_domain_attr *new, int idx_new) 6569 { 6570 struct sched_domain_attr tmp; 6571 6572 /* fast path */ 6573 if (!new && !cur) 6574 return 1; 6575 6576 tmp = SD_ATTR_INIT; 6577 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6578 new ? (new + idx_new) : &tmp, 6579 sizeof(struct sched_domain_attr)); 6580 } 6581 6582 /* 6583 * Partition sched domains as specified by the 'ndoms_new' 6584 * cpumasks in the array doms_new[] of cpumasks. This compares 6585 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6586 * It destroys each deleted domain and builds each new domain. 6587 * 6588 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6589 * The masks don't intersect (don't overlap.) We should setup one 6590 * sched domain for each mask. CPUs not in any of the cpumasks will 6591 * not be load balanced. If the same cpumask appears both in the 6592 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6593 * it as it is. 6594 * 6595 * The passed in 'doms_new' should be allocated using 6596 * alloc_sched_domains. This routine takes ownership of it and will 6597 * free_sched_domains it when done with it. If the caller failed the 6598 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6599 * and partition_sched_domains() will fallback to the single partition 6600 * 'fallback_doms', it also forces the domains to be rebuilt. 6601 * 6602 * If doms_new == NULL it will be replaced with cpu_online_mask. 6603 * ndoms_new == 0 is a special case for destroying existing domains, 6604 * and it will not create the default domain. 6605 * 6606 * Call with hotplug lock held 6607 */ 6608 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6609 struct sched_domain_attr *dattr_new) 6610 { 6611 int i, j, n; 6612 int new_topology; 6613 6614 mutex_lock(&sched_domains_mutex); 6615 6616 /* always unregister in case we don't destroy any domains */ 6617 unregister_sched_domain_sysctl(); 6618 6619 /* Let architecture update cpu core mappings. */ 6620 new_topology = arch_update_cpu_topology(); 6621 6622 n = doms_new ? ndoms_new : 0; 6623 6624 /* Destroy deleted domains */ 6625 for (i = 0; i < ndoms_cur; i++) { 6626 for (j = 0; j < n && !new_topology; j++) { 6627 if (cpumask_equal(doms_cur[i], doms_new[j]) 6628 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6629 goto match1; 6630 } 6631 /* no match - a current sched domain not in new doms_new[] */ 6632 detach_destroy_domains(doms_cur[i]); 6633 match1: 6634 ; 6635 } 6636 6637 n = ndoms_cur; 6638 if (doms_new == NULL) { 6639 n = 0; 6640 doms_new = &fallback_doms; 6641 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6642 WARN_ON_ONCE(dattr_new); 6643 } 6644 6645 /* Build new domains */ 6646 for (i = 0; i < ndoms_new; i++) { 6647 for (j = 0; j < n && !new_topology; j++) { 6648 if (cpumask_equal(doms_new[i], doms_cur[j]) 6649 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6650 goto match2; 6651 } 6652 /* no match - add a new doms_new */ 6653 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6654 match2: 6655 ; 6656 } 6657 6658 /* Remember the new sched domains */ 6659 if (doms_cur != &fallback_doms) 6660 free_sched_domains(doms_cur, ndoms_cur); 6661 kfree(dattr_cur); /* kfree(NULL) is safe */ 6662 doms_cur = doms_new; 6663 dattr_cur = dattr_new; 6664 ndoms_cur = ndoms_new; 6665 6666 register_sched_domain_sysctl(); 6667 6668 mutex_unlock(&sched_domains_mutex); 6669 } 6670 6671 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6672 6673 /* 6674 * Update cpusets according to cpu_active mask. If cpusets are 6675 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6676 * around partition_sched_domains(). 6677 * 6678 * If we come here as part of a suspend/resume, don't touch cpusets because we 6679 * want to restore it back to its original state upon resume anyway. 6680 */ 6681 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6682 void *hcpu) 6683 { 6684 switch (action) { 6685 case CPU_ONLINE_FROZEN: 6686 case CPU_DOWN_FAILED_FROZEN: 6687 6688 /* 6689 * num_cpus_frozen tracks how many CPUs are involved in suspend 6690 * resume sequence. As long as this is not the last online 6691 * operation in the resume sequence, just build a single sched 6692 * domain, ignoring cpusets. 6693 */ 6694 num_cpus_frozen--; 6695 if (likely(num_cpus_frozen)) { 6696 partition_sched_domains(1, NULL, NULL); 6697 break; 6698 } 6699 6700 /* 6701 * This is the last CPU online operation. So fall through and 6702 * restore the original sched domains by considering the 6703 * cpuset configurations. 6704 */ 6705 6706 case CPU_ONLINE: 6707 case CPU_DOWN_FAILED: 6708 cpuset_update_active_cpus(true); 6709 break; 6710 default: 6711 return NOTIFY_DONE; 6712 } 6713 return NOTIFY_OK; 6714 } 6715 6716 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6717 void *hcpu) 6718 { 6719 switch (action) { 6720 case CPU_DOWN_PREPARE: 6721 cpuset_update_active_cpus(false); 6722 break; 6723 case CPU_DOWN_PREPARE_FROZEN: 6724 num_cpus_frozen++; 6725 partition_sched_domains(1, NULL, NULL); 6726 break; 6727 default: 6728 return NOTIFY_DONE; 6729 } 6730 return NOTIFY_OK; 6731 } 6732 6733 void __init sched_init_smp(void) 6734 { 6735 cpumask_var_t non_isolated_cpus; 6736 6737 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6738 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6739 6740 sched_init_numa(); 6741 6742 /* 6743 * There's no userspace yet to cause hotplug operations; hence all the 6744 * cpu masks are stable and all blatant races in the below code cannot 6745 * happen. 6746 */ 6747 mutex_lock(&sched_domains_mutex); 6748 init_sched_domains(cpu_active_mask); 6749 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6750 if (cpumask_empty(non_isolated_cpus)) 6751 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6752 mutex_unlock(&sched_domains_mutex); 6753 6754 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6755 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6756 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6757 6758 init_hrtick(); 6759 6760 /* Move init over to a non-isolated CPU */ 6761 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6762 BUG(); 6763 sched_init_granularity(); 6764 free_cpumask_var(non_isolated_cpus); 6765 6766 init_sched_rt_class(); 6767 init_sched_dl_class(); 6768 } 6769 #else 6770 void __init sched_init_smp(void) 6771 { 6772 sched_init_granularity(); 6773 } 6774 #endif /* CONFIG_SMP */ 6775 6776 const_debug unsigned int sysctl_timer_migration = 1; 6777 6778 int in_sched_functions(unsigned long addr) 6779 { 6780 return in_lock_functions(addr) || 6781 (addr >= (unsigned long)__sched_text_start 6782 && addr < (unsigned long)__sched_text_end); 6783 } 6784 6785 #ifdef CONFIG_CGROUP_SCHED 6786 /* 6787 * Default task group. 6788 * Every task in system belongs to this group at bootup. 6789 */ 6790 struct task_group root_task_group; 6791 LIST_HEAD(task_groups); 6792 #endif 6793 6794 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6795 6796 void __init sched_init(void) 6797 { 6798 int i, j; 6799 unsigned long alloc_size = 0, ptr; 6800 6801 #ifdef CONFIG_FAIR_GROUP_SCHED 6802 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6803 #endif 6804 #ifdef CONFIG_RT_GROUP_SCHED 6805 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6806 #endif 6807 #ifdef CONFIG_CPUMASK_OFFSTACK 6808 alloc_size += num_possible_cpus() * cpumask_size(); 6809 #endif 6810 if (alloc_size) { 6811 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6812 6813 #ifdef CONFIG_FAIR_GROUP_SCHED 6814 root_task_group.se = (struct sched_entity **)ptr; 6815 ptr += nr_cpu_ids * sizeof(void **); 6816 6817 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6818 ptr += nr_cpu_ids * sizeof(void **); 6819 6820 #endif /* CONFIG_FAIR_GROUP_SCHED */ 6821 #ifdef CONFIG_RT_GROUP_SCHED 6822 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6823 ptr += nr_cpu_ids * sizeof(void **); 6824 6825 root_task_group.rt_rq = (struct rt_rq **)ptr; 6826 ptr += nr_cpu_ids * sizeof(void **); 6827 6828 #endif /* CONFIG_RT_GROUP_SCHED */ 6829 #ifdef CONFIG_CPUMASK_OFFSTACK 6830 for_each_possible_cpu(i) { 6831 per_cpu(load_balance_mask, i) = (void *)ptr; 6832 ptr += cpumask_size(); 6833 } 6834 #endif /* CONFIG_CPUMASK_OFFSTACK */ 6835 } 6836 6837 init_rt_bandwidth(&def_rt_bandwidth, 6838 global_rt_period(), global_rt_runtime()); 6839 init_dl_bandwidth(&def_dl_bandwidth, 6840 global_rt_period(), global_rt_runtime()); 6841 6842 #ifdef CONFIG_SMP 6843 init_defrootdomain(); 6844 #endif 6845 6846 #ifdef CONFIG_RT_GROUP_SCHED 6847 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6848 global_rt_period(), global_rt_runtime()); 6849 #endif /* CONFIG_RT_GROUP_SCHED */ 6850 6851 #ifdef CONFIG_CGROUP_SCHED 6852 list_add(&root_task_group.list, &task_groups); 6853 INIT_LIST_HEAD(&root_task_group.children); 6854 INIT_LIST_HEAD(&root_task_group.siblings); 6855 autogroup_init(&init_task); 6856 6857 #endif /* CONFIG_CGROUP_SCHED */ 6858 6859 for_each_possible_cpu(i) { 6860 struct rq *rq; 6861 6862 rq = cpu_rq(i); 6863 raw_spin_lock_init(&rq->lock); 6864 rq->nr_running = 0; 6865 rq->calc_load_active = 0; 6866 rq->calc_load_update = jiffies + LOAD_FREQ; 6867 init_cfs_rq(&rq->cfs); 6868 init_rt_rq(&rq->rt, rq); 6869 init_dl_rq(&rq->dl, rq); 6870 #ifdef CONFIG_FAIR_GROUP_SCHED 6871 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6872 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6873 /* 6874 * How much cpu bandwidth does root_task_group get? 6875 * 6876 * In case of task-groups formed thr' the cgroup filesystem, it 6877 * gets 100% of the cpu resources in the system. This overall 6878 * system cpu resource is divided among the tasks of 6879 * root_task_group and its child task-groups in a fair manner, 6880 * based on each entity's (task or task-group's) weight 6881 * (se->load.weight). 6882 * 6883 * In other words, if root_task_group has 10 tasks of weight 6884 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6885 * then A0's share of the cpu resource is: 6886 * 6887 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6888 * 6889 * We achieve this by letting root_task_group's tasks sit 6890 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6891 */ 6892 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6893 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6894 #endif /* CONFIG_FAIR_GROUP_SCHED */ 6895 6896 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6897 #ifdef CONFIG_RT_GROUP_SCHED 6898 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6899 #endif 6900 6901 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6902 rq->cpu_load[j] = 0; 6903 6904 rq->last_load_update_tick = jiffies; 6905 6906 #ifdef CONFIG_SMP 6907 rq->sd = NULL; 6908 rq->rd = NULL; 6909 rq->cpu_power = SCHED_POWER_SCALE; 6910 rq->post_schedule = 0; 6911 rq->active_balance = 0; 6912 rq->next_balance = jiffies; 6913 rq->push_cpu = 0; 6914 rq->cpu = i; 6915 rq->online = 0; 6916 rq->idle_stamp = 0; 6917 rq->avg_idle = 2*sysctl_sched_migration_cost; 6918 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6919 6920 INIT_LIST_HEAD(&rq->cfs_tasks); 6921 6922 rq_attach_root(rq, &def_root_domain); 6923 #ifdef CONFIG_NO_HZ_COMMON 6924 rq->nohz_flags = 0; 6925 #endif 6926 #ifdef CONFIG_NO_HZ_FULL 6927 rq->last_sched_tick = 0; 6928 #endif 6929 #endif 6930 init_rq_hrtick(rq); 6931 atomic_set(&rq->nr_iowait, 0); 6932 } 6933 6934 set_load_weight(&init_task); 6935 6936 #ifdef CONFIG_PREEMPT_NOTIFIERS 6937 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6938 #endif 6939 6940 /* 6941 * The boot idle thread does lazy MMU switching as well: 6942 */ 6943 atomic_inc(&init_mm.mm_count); 6944 enter_lazy_tlb(&init_mm, current); 6945 6946 /* 6947 * Make us the idle thread. Technically, schedule() should not be 6948 * called from this thread, however somewhere below it might be, 6949 * but because we are the idle thread, we just pick up running again 6950 * when this runqueue becomes "idle". 6951 */ 6952 init_idle(current, smp_processor_id()); 6953 6954 calc_load_update = jiffies + LOAD_FREQ; 6955 6956 /* 6957 * During early bootup we pretend to be a normal task: 6958 */ 6959 current->sched_class = &fair_sched_class; 6960 6961 #ifdef CONFIG_SMP 6962 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6963 /* May be allocated at isolcpus cmdline parse time */ 6964 if (cpu_isolated_map == NULL) 6965 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6966 idle_thread_set_boot_cpu(); 6967 #endif 6968 init_sched_fair_class(); 6969 6970 scheduler_running = 1; 6971 } 6972 6973 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6974 static inline int preempt_count_equals(int preempt_offset) 6975 { 6976 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6977 6978 return (nested == preempt_offset); 6979 } 6980 6981 void __might_sleep(const char *file, int line, int preempt_offset) 6982 { 6983 static unsigned long prev_jiffy; /* ratelimiting */ 6984 6985 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6986 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 6987 !is_idle_task(current)) || 6988 system_state != SYSTEM_RUNNING || oops_in_progress) 6989 return; 6990 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6991 return; 6992 prev_jiffy = jiffies; 6993 6994 printk(KERN_ERR 6995 "BUG: sleeping function called from invalid context at %s:%d\n", 6996 file, line); 6997 printk(KERN_ERR 6998 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6999 in_atomic(), irqs_disabled(), 7000 current->pid, current->comm); 7001 7002 debug_show_held_locks(current); 7003 if (irqs_disabled()) 7004 print_irqtrace_events(current); 7005 #ifdef CONFIG_DEBUG_PREEMPT 7006 if (!preempt_count_equals(preempt_offset)) { 7007 pr_err("Preemption disabled at:"); 7008 print_ip_sym(current->preempt_disable_ip); 7009 pr_cont("\n"); 7010 } 7011 #endif 7012 dump_stack(); 7013 } 7014 EXPORT_SYMBOL(__might_sleep); 7015 #endif 7016 7017 #ifdef CONFIG_MAGIC_SYSRQ 7018 static void normalize_task(struct rq *rq, struct task_struct *p) 7019 { 7020 const struct sched_class *prev_class = p->sched_class; 7021 struct sched_attr attr = { 7022 .sched_policy = SCHED_NORMAL, 7023 }; 7024 int old_prio = p->prio; 7025 int on_rq; 7026 7027 on_rq = p->on_rq; 7028 if (on_rq) 7029 dequeue_task(rq, p, 0); 7030 __setscheduler(rq, p, &attr); 7031 if (on_rq) { 7032 enqueue_task(rq, p, 0); 7033 resched_task(rq->curr); 7034 } 7035 7036 check_class_changed(rq, p, prev_class, old_prio); 7037 } 7038 7039 void normalize_rt_tasks(void) 7040 { 7041 struct task_struct *g, *p; 7042 unsigned long flags; 7043 struct rq *rq; 7044 7045 read_lock_irqsave(&tasklist_lock, flags); 7046 do_each_thread(g, p) { 7047 /* 7048 * Only normalize user tasks: 7049 */ 7050 if (!p->mm) 7051 continue; 7052 7053 p->se.exec_start = 0; 7054 #ifdef CONFIG_SCHEDSTATS 7055 p->se.statistics.wait_start = 0; 7056 p->se.statistics.sleep_start = 0; 7057 p->se.statistics.block_start = 0; 7058 #endif 7059 7060 if (!dl_task(p) && !rt_task(p)) { 7061 /* 7062 * Renice negative nice level userspace 7063 * tasks back to 0: 7064 */ 7065 if (task_nice(p) < 0 && p->mm) 7066 set_user_nice(p, 0); 7067 continue; 7068 } 7069 7070 raw_spin_lock(&p->pi_lock); 7071 rq = __task_rq_lock(p); 7072 7073 normalize_task(rq, p); 7074 7075 __task_rq_unlock(rq); 7076 raw_spin_unlock(&p->pi_lock); 7077 } while_each_thread(g, p); 7078 7079 read_unlock_irqrestore(&tasklist_lock, flags); 7080 } 7081 7082 #endif /* CONFIG_MAGIC_SYSRQ */ 7083 7084 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7085 /* 7086 * These functions are only useful for the IA64 MCA handling, or kdb. 7087 * 7088 * They can only be called when the whole system has been 7089 * stopped - every CPU needs to be quiescent, and no scheduling 7090 * activity can take place. Using them for anything else would 7091 * be a serious bug, and as a result, they aren't even visible 7092 * under any other configuration. 7093 */ 7094 7095 /** 7096 * curr_task - return the current task for a given cpu. 7097 * @cpu: the processor in question. 7098 * 7099 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7100 * 7101 * Return: The current task for @cpu. 7102 */ 7103 struct task_struct *curr_task(int cpu) 7104 { 7105 return cpu_curr(cpu); 7106 } 7107 7108 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7109 7110 #ifdef CONFIG_IA64 7111 /** 7112 * set_curr_task - set the current task for a given cpu. 7113 * @cpu: the processor in question. 7114 * @p: the task pointer to set. 7115 * 7116 * Description: This function must only be used when non-maskable interrupts 7117 * are serviced on a separate stack. It allows the architecture to switch the 7118 * notion of the current task on a cpu in a non-blocking manner. This function 7119 * must be called with all CPU's synchronized, and interrupts disabled, the 7120 * and caller must save the original value of the current task (see 7121 * curr_task() above) and restore that value before reenabling interrupts and 7122 * re-starting the system. 7123 * 7124 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7125 */ 7126 void set_curr_task(int cpu, struct task_struct *p) 7127 { 7128 cpu_curr(cpu) = p; 7129 } 7130 7131 #endif 7132 7133 #ifdef CONFIG_CGROUP_SCHED 7134 /* task_group_lock serializes the addition/removal of task groups */ 7135 static DEFINE_SPINLOCK(task_group_lock); 7136 7137 static void free_sched_group(struct task_group *tg) 7138 { 7139 free_fair_sched_group(tg); 7140 free_rt_sched_group(tg); 7141 autogroup_free(tg); 7142 kfree(tg); 7143 } 7144 7145 /* allocate runqueue etc for a new task group */ 7146 struct task_group *sched_create_group(struct task_group *parent) 7147 { 7148 struct task_group *tg; 7149 7150 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7151 if (!tg) 7152 return ERR_PTR(-ENOMEM); 7153 7154 if (!alloc_fair_sched_group(tg, parent)) 7155 goto err; 7156 7157 if (!alloc_rt_sched_group(tg, parent)) 7158 goto err; 7159 7160 return tg; 7161 7162 err: 7163 free_sched_group(tg); 7164 return ERR_PTR(-ENOMEM); 7165 } 7166 7167 void sched_online_group(struct task_group *tg, struct task_group *parent) 7168 { 7169 unsigned long flags; 7170 7171 spin_lock_irqsave(&task_group_lock, flags); 7172 list_add_rcu(&tg->list, &task_groups); 7173 7174 WARN_ON(!parent); /* root should already exist */ 7175 7176 tg->parent = parent; 7177 INIT_LIST_HEAD(&tg->children); 7178 list_add_rcu(&tg->siblings, &parent->children); 7179 spin_unlock_irqrestore(&task_group_lock, flags); 7180 } 7181 7182 /* rcu callback to free various structures associated with a task group */ 7183 static void free_sched_group_rcu(struct rcu_head *rhp) 7184 { 7185 /* now it should be safe to free those cfs_rqs */ 7186 free_sched_group(container_of(rhp, struct task_group, rcu)); 7187 } 7188 7189 /* Destroy runqueue etc associated with a task group */ 7190 void sched_destroy_group(struct task_group *tg) 7191 { 7192 /* wait for possible concurrent references to cfs_rqs complete */ 7193 call_rcu(&tg->rcu, free_sched_group_rcu); 7194 } 7195 7196 void sched_offline_group(struct task_group *tg) 7197 { 7198 unsigned long flags; 7199 int i; 7200 7201 /* end participation in shares distribution */ 7202 for_each_possible_cpu(i) 7203 unregister_fair_sched_group(tg, i); 7204 7205 spin_lock_irqsave(&task_group_lock, flags); 7206 list_del_rcu(&tg->list); 7207 list_del_rcu(&tg->siblings); 7208 spin_unlock_irqrestore(&task_group_lock, flags); 7209 } 7210 7211 /* change task's runqueue when it moves between groups. 7212 * The caller of this function should have put the task in its new group 7213 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7214 * reflect its new group. 7215 */ 7216 void sched_move_task(struct task_struct *tsk) 7217 { 7218 struct task_group *tg; 7219 int on_rq, running; 7220 unsigned long flags; 7221 struct rq *rq; 7222 7223 rq = task_rq_lock(tsk, &flags); 7224 7225 running = task_current(rq, tsk); 7226 on_rq = tsk->on_rq; 7227 7228 if (on_rq) 7229 dequeue_task(rq, tsk, 0); 7230 if (unlikely(running)) 7231 tsk->sched_class->put_prev_task(rq, tsk); 7232 7233 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 7234 lockdep_is_held(&tsk->sighand->siglock)), 7235 struct task_group, css); 7236 tg = autogroup_task_group(tsk, tg); 7237 tsk->sched_task_group = tg; 7238 7239 #ifdef CONFIG_FAIR_GROUP_SCHED 7240 if (tsk->sched_class->task_move_group) 7241 tsk->sched_class->task_move_group(tsk, on_rq); 7242 else 7243 #endif 7244 set_task_rq(tsk, task_cpu(tsk)); 7245 7246 if (unlikely(running)) 7247 tsk->sched_class->set_curr_task(rq); 7248 if (on_rq) 7249 enqueue_task(rq, tsk, 0); 7250 7251 task_rq_unlock(rq, tsk, &flags); 7252 } 7253 #endif /* CONFIG_CGROUP_SCHED */ 7254 7255 #ifdef CONFIG_RT_GROUP_SCHED 7256 /* 7257 * Ensure that the real time constraints are schedulable. 7258 */ 7259 static DEFINE_MUTEX(rt_constraints_mutex); 7260 7261 /* Must be called with tasklist_lock held */ 7262 static inline int tg_has_rt_tasks(struct task_group *tg) 7263 { 7264 struct task_struct *g, *p; 7265 7266 do_each_thread(g, p) { 7267 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7268 return 1; 7269 } while_each_thread(g, p); 7270 7271 return 0; 7272 } 7273 7274 struct rt_schedulable_data { 7275 struct task_group *tg; 7276 u64 rt_period; 7277 u64 rt_runtime; 7278 }; 7279 7280 static int tg_rt_schedulable(struct task_group *tg, void *data) 7281 { 7282 struct rt_schedulable_data *d = data; 7283 struct task_group *child; 7284 unsigned long total, sum = 0; 7285 u64 period, runtime; 7286 7287 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7288 runtime = tg->rt_bandwidth.rt_runtime; 7289 7290 if (tg == d->tg) { 7291 period = d->rt_period; 7292 runtime = d->rt_runtime; 7293 } 7294 7295 /* 7296 * Cannot have more runtime than the period. 7297 */ 7298 if (runtime > period && runtime != RUNTIME_INF) 7299 return -EINVAL; 7300 7301 /* 7302 * Ensure we don't starve existing RT tasks. 7303 */ 7304 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7305 return -EBUSY; 7306 7307 total = to_ratio(period, runtime); 7308 7309 /* 7310 * Nobody can have more than the global setting allows. 7311 */ 7312 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7313 return -EINVAL; 7314 7315 /* 7316 * The sum of our children's runtime should not exceed our own. 7317 */ 7318 list_for_each_entry_rcu(child, &tg->children, siblings) { 7319 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7320 runtime = child->rt_bandwidth.rt_runtime; 7321 7322 if (child == d->tg) { 7323 period = d->rt_period; 7324 runtime = d->rt_runtime; 7325 } 7326 7327 sum += to_ratio(period, runtime); 7328 } 7329 7330 if (sum > total) 7331 return -EINVAL; 7332 7333 return 0; 7334 } 7335 7336 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7337 { 7338 int ret; 7339 7340 struct rt_schedulable_data data = { 7341 .tg = tg, 7342 .rt_period = period, 7343 .rt_runtime = runtime, 7344 }; 7345 7346 rcu_read_lock(); 7347 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7348 rcu_read_unlock(); 7349 7350 return ret; 7351 } 7352 7353 static int tg_set_rt_bandwidth(struct task_group *tg, 7354 u64 rt_period, u64 rt_runtime) 7355 { 7356 int i, err = 0; 7357 7358 mutex_lock(&rt_constraints_mutex); 7359 read_lock(&tasklist_lock); 7360 err = __rt_schedulable(tg, rt_period, rt_runtime); 7361 if (err) 7362 goto unlock; 7363 7364 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7365 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7366 tg->rt_bandwidth.rt_runtime = rt_runtime; 7367 7368 for_each_possible_cpu(i) { 7369 struct rt_rq *rt_rq = tg->rt_rq[i]; 7370 7371 raw_spin_lock(&rt_rq->rt_runtime_lock); 7372 rt_rq->rt_runtime = rt_runtime; 7373 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7374 } 7375 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7376 unlock: 7377 read_unlock(&tasklist_lock); 7378 mutex_unlock(&rt_constraints_mutex); 7379 7380 return err; 7381 } 7382 7383 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7384 { 7385 u64 rt_runtime, rt_period; 7386 7387 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7388 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7389 if (rt_runtime_us < 0) 7390 rt_runtime = RUNTIME_INF; 7391 7392 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7393 } 7394 7395 static long sched_group_rt_runtime(struct task_group *tg) 7396 { 7397 u64 rt_runtime_us; 7398 7399 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7400 return -1; 7401 7402 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7403 do_div(rt_runtime_us, NSEC_PER_USEC); 7404 return rt_runtime_us; 7405 } 7406 7407 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7408 { 7409 u64 rt_runtime, rt_period; 7410 7411 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7412 rt_runtime = tg->rt_bandwidth.rt_runtime; 7413 7414 if (rt_period == 0) 7415 return -EINVAL; 7416 7417 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7418 } 7419 7420 static long sched_group_rt_period(struct task_group *tg) 7421 { 7422 u64 rt_period_us; 7423 7424 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7425 do_div(rt_period_us, NSEC_PER_USEC); 7426 return rt_period_us; 7427 } 7428 #endif /* CONFIG_RT_GROUP_SCHED */ 7429 7430 #ifdef CONFIG_RT_GROUP_SCHED 7431 static int sched_rt_global_constraints(void) 7432 { 7433 int ret = 0; 7434 7435 mutex_lock(&rt_constraints_mutex); 7436 read_lock(&tasklist_lock); 7437 ret = __rt_schedulable(NULL, 0, 0); 7438 read_unlock(&tasklist_lock); 7439 mutex_unlock(&rt_constraints_mutex); 7440 7441 return ret; 7442 } 7443 7444 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7445 { 7446 /* Don't accept realtime tasks when there is no way for them to run */ 7447 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7448 return 0; 7449 7450 return 1; 7451 } 7452 7453 #else /* !CONFIG_RT_GROUP_SCHED */ 7454 static int sched_rt_global_constraints(void) 7455 { 7456 unsigned long flags; 7457 int i, ret = 0; 7458 7459 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7460 for_each_possible_cpu(i) { 7461 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7462 7463 raw_spin_lock(&rt_rq->rt_runtime_lock); 7464 rt_rq->rt_runtime = global_rt_runtime(); 7465 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7466 } 7467 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7468 7469 return ret; 7470 } 7471 #endif /* CONFIG_RT_GROUP_SCHED */ 7472 7473 static int sched_dl_global_constraints(void) 7474 { 7475 u64 runtime = global_rt_runtime(); 7476 u64 period = global_rt_period(); 7477 u64 new_bw = to_ratio(period, runtime); 7478 int cpu, ret = 0; 7479 unsigned long flags; 7480 7481 /* 7482 * Here we want to check the bandwidth not being set to some 7483 * value smaller than the currently allocated bandwidth in 7484 * any of the root_domains. 7485 * 7486 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7487 * cycling on root_domains... Discussion on different/better 7488 * solutions is welcome! 7489 */ 7490 for_each_possible_cpu(cpu) { 7491 struct dl_bw *dl_b = dl_bw_of(cpu); 7492 7493 raw_spin_lock_irqsave(&dl_b->lock, flags); 7494 if (new_bw < dl_b->total_bw) 7495 ret = -EBUSY; 7496 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7497 7498 if (ret) 7499 break; 7500 } 7501 7502 return ret; 7503 } 7504 7505 static void sched_dl_do_global(void) 7506 { 7507 u64 new_bw = -1; 7508 int cpu; 7509 unsigned long flags; 7510 7511 def_dl_bandwidth.dl_period = global_rt_period(); 7512 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7513 7514 if (global_rt_runtime() != RUNTIME_INF) 7515 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7516 7517 /* 7518 * FIXME: As above... 7519 */ 7520 for_each_possible_cpu(cpu) { 7521 struct dl_bw *dl_b = dl_bw_of(cpu); 7522 7523 raw_spin_lock_irqsave(&dl_b->lock, flags); 7524 dl_b->bw = new_bw; 7525 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7526 } 7527 } 7528 7529 static int sched_rt_global_validate(void) 7530 { 7531 if (sysctl_sched_rt_period <= 0) 7532 return -EINVAL; 7533 7534 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7535 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7536 return -EINVAL; 7537 7538 return 0; 7539 } 7540 7541 static void sched_rt_do_global(void) 7542 { 7543 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7544 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7545 } 7546 7547 int sched_rt_handler(struct ctl_table *table, int write, 7548 void __user *buffer, size_t *lenp, 7549 loff_t *ppos) 7550 { 7551 int old_period, old_runtime; 7552 static DEFINE_MUTEX(mutex); 7553 int ret; 7554 7555 mutex_lock(&mutex); 7556 old_period = sysctl_sched_rt_period; 7557 old_runtime = sysctl_sched_rt_runtime; 7558 7559 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7560 7561 if (!ret && write) { 7562 ret = sched_rt_global_validate(); 7563 if (ret) 7564 goto undo; 7565 7566 ret = sched_rt_global_constraints(); 7567 if (ret) 7568 goto undo; 7569 7570 ret = sched_dl_global_constraints(); 7571 if (ret) 7572 goto undo; 7573 7574 sched_rt_do_global(); 7575 sched_dl_do_global(); 7576 } 7577 if (0) { 7578 undo: 7579 sysctl_sched_rt_period = old_period; 7580 sysctl_sched_rt_runtime = old_runtime; 7581 } 7582 mutex_unlock(&mutex); 7583 7584 return ret; 7585 } 7586 7587 int sched_rr_handler(struct ctl_table *table, int write, 7588 void __user *buffer, size_t *lenp, 7589 loff_t *ppos) 7590 { 7591 int ret; 7592 static DEFINE_MUTEX(mutex); 7593 7594 mutex_lock(&mutex); 7595 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7596 /* make sure that internally we keep jiffies */ 7597 /* also, writing zero resets timeslice to default */ 7598 if (!ret && write) { 7599 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7600 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7601 } 7602 mutex_unlock(&mutex); 7603 return ret; 7604 } 7605 7606 #ifdef CONFIG_CGROUP_SCHED 7607 7608 static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7609 { 7610 return css ? container_of(css, struct task_group, css) : NULL; 7611 } 7612 7613 static struct cgroup_subsys_state * 7614 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7615 { 7616 struct task_group *parent = css_tg(parent_css); 7617 struct task_group *tg; 7618 7619 if (!parent) { 7620 /* This is early initialization for the top cgroup */ 7621 return &root_task_group.css; 7622 } 7623 7624 tg = sched_create_group(parent); 7625 if (IS_ERR(tg)) 7626 return ERR_PTR(-ENOMEM); 7627 7628 return &tg->css; 7629 } 7630 7631 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7632 { 7633 struct task_group *tg = css_tg(css); 7634 struct task_group *parent = css_tg(css_parent(css)); 7635 7636 if (parent) 7637 sched_online_group(tg, parent); 7638 return 0; 7639 } 7640 7641 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7642 { 7643 struct task_group *tg = css_tg(css); 7644 7645 sched_destroy_group(tg); 7646 } 7647 7648 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7649 { 7650 struct task_group *tg = css_tg(css); 7651 7652 sched_offline_group(tg); 7653 } 7654 7655 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7656 struct cgroup_taskset *tset) 7657 { 7658 struct task_struct *task; 7659 7660 cgroup_taskset_for_each(task, css, tset) { 7661 #ifdef CONFIG_RT_GROUP_SCHED 7662 if (!sched_rt_can_attach(css_tg(css), task)) 7663 return -EINVAL; 7664 #else 7665 /* We don't support RT-tasks being in separate groups */ 7666 if (task->sched_class != &fair_sched_class) 7667 return -EINVAL; 7668 #endif 7669 } 7670 return 0; 7671 } 7672 7673 static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7674 struct cgroup_taskset *tset) 7675 { 7676 struct task_struct *task; 7677 7678 cgroup_taskset_for_each(task, css, tset) 7679 sched_move_task(task); 7680 } 7681 7682 static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7683 struct cgroup_subsys_state *old_css, 7684 struct task_struct *task) 7685 { 7686 /* 7687 * cgroup_exit() is called in the copy_process() failure path. 7688 * Ignore this case since the task hasn't ran yet, this avoids 7689 * trying to poke a half freed task state from generic code. 7690 */ 7691 if (!(task->flags & PF_EXITING)) 7692 return; 7693 7694 sched_move_task(task); 7695 } 7696 7697 #ifdef CONFIG_FAIR_GROUP_SCHED 7698 static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7699 struct cftype *cftype, u64 shareval) 7700 { 7701 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7702 } 7703 7704 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7705 struct cftype *cft) 7706 { 7707 struct task_group *tg = css_tg(css); 7708 7709 return (u64) scale_load_down(tg->shares); 7710 } 7711 7712 #ifdef CONFIG_CFS_BANDWIDTH 7713 static DEFINE_MUTEX(cfs_constraints_mutex); 7714 7715 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7716 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7717 7718 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7719 7720 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7721 { 7722 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7723 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7724 7725 if (tg == &root_task_group) 7726 return -EINVAL; 7727 7728 /* 7729 * Ensure we have at some amount of bandwidth every period. This is 7730 * to prevent reaching a state of large arrears when throttled via 7731 * entity_tick() resulting in prolonged exit starvation. 7732 */ 7733 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7734 return -EINVAL; 7735 7736 /* 7737 * Likewise, bound things on the otherside by preventing insane quota 7738 * periods. This also allows us to normalize in computing quota 7739 * feasibility. 7740 */ 7741 if (period > max_cfs_quota_period) 7742 return -EINVAL; 7743 7744 mutex_lock(&cfs_constraints_mutex); 7745 ret = __cfs_schedulable(tg, period, quota); 7746 if (ret) 7747 goto out_unlock; 7748 7749 runtime_enabled = quota != RUNTIME_INF; 7750 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7751 /* 7752 * If we need to toggle cfs_bandwidth_used, off->on must occur 7753 * before making related changes, and on->off must occur afterwards 7754 */ 7755 if (runtime_enabled && !runtime_was_enabled) 7756 cfs_bandwidth_usage_inc(); 7757 raw_spin_lock_irq(&cfs_b->lock); 7758 cfs_b->period = ns_to_ktime(period); 7759 cfs_b->quota = quota; 7760 7761 __refill_cfs_bandwidth_runtime(cfs_b); 7762 /* restart the period timer (if active) to handle new period expiry */ 7763 if (runtime_enabled && cfs_b->timer_active) { 7764 /* force a reprogram */ 7765 cfs_b->timer_active = 0; 7766 __start_cfs_bandwidth(cfs_b); 7767 } 7768 raw_spin_unlock_irq(&cfs_b->lock); 7769 7770 for_each_possible_cpu(i) { 7771 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7772 struct rq *rq = cfs_rq->rq; 7773 7774 raw_spin_lock_irq(&rq->lock); 7775 cfs_rq->runtime_enabled = runtime_enabled; 7776 cfs_rq->runtime_remaining = 0; 7777 7778 if (cfs_rq->throttled) 7779 unthrottle_cfs_rq(cfs_rq); 7780 raw_spin_unlock_irq(&rq->lock); 7781 } 7782 if (runtime_was_enabled && !runtime_enabled) 7783 cfs_bandwidth_usage_dec(); 7784 out_unlock: 7785 mutex_unlock(&cfs_constraints_mutex); 7786 7787 return ret; 7788 } 7789 7790 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7791 { 7792 u64 quota, period; 7793 7794 period = ktime_to_ns(tg->cfs_bandwidth.period); 7795 if (cfs_quota_us < 0) 7796 quota = RUNTIME_INF; 7797 else 7798 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7799 7800 return tg_set_cfs_bandwidth(tg, period, quota); 7801 } 7802 7803 long tg_get_cfs_quota(struct task_group *tg) 7804 { 7805 u64 quota_us; 7806 7807 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7808 return -1; 7809 7810 quota_us = tg->cfs_bandwidth.quota; 7811 do_div(quota_us, NSEC_PER_USEC); 7812 7813 return quota_us; 7814 } 7815 7816 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7817 { 7818 u64 quota, period; 7819 7820 period = (u64)cfs_period_us * NSEC_PER_USEC; 7821 quota = tg->cfs_bandwidth.quota; 7822 7823 return tg_set_cfs_bandwidth(tg, period, quota); 7824 } 7825 7826 long tg_get_cfs_period(struct task_group *tg) 7827 { 7828 u64 cfs_period_us; 7829 7830 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7831 do_div(cfs_period_us, NSEC_PER_USEC); 7832 7833 return cfs_period_us; 7834 } 7835 7836 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7837 struct cftype *cft) 7838 { 7839 return tg_get_cfs_quota(css_tg(css)); 7840 } 7841 7842 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7843 struct cftype *cftype, s64 cfs_quota_us) 7844 { 7845 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7846 } 7847 7848 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7849 struct cftype *cft) 7850 { 7851 return tg_get_cfs_period(css_tg(css)); 7852 } 7853 7854 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7855 struct cftype *cftype, u64 cfs_period_us) 7856 { 7857 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7858 } 7859 7860 struct cfs_schedulable_data { 7861 struct task_group *tg; 7862 u64 period, quota; 7863 }; 7864 7865 /* 7866 * normalize group quota/period to be quota/max_period 7867 * note: units are usecs 7868 */ 7869 static u64 normalize_cfs_quota(struct task_group *tg, 7870 struct cfs_schedulable_data *d) 7871 { 7872 u64 quota, period; 7873 7874 if (tg == d->tg) { 7875 period = d->period; 7876 quota = d->quota; 7877 } else { 7878 period = tg_get_cfs_period(tg); 7879 quota = tg_get_cfs_quota(tg); 7880 } 7881 7882 /* note: these should typically be equivalent */ 7883 if (quota == RUNTIME_INF || quota == -1) 7884 return RUNTIME_INF; 7885 7886 return to_ratio(period, quota); 7887 } 7888 7889 static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7890 { 7891 struct cfs_schedulable_data *d = data; 7892 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7893 s64 quota = 0, parent_quota = -1; 7894 7895 if (!tg->parent) { 7896 quota = RUNTIME_INF; 7897 } else { 7898 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7899 7900 quota = normalize_cfs_quota(tg, d); 7901 parent_quota = parent_b->hierarchal_quota; 7902 7903 /* 7904 * ensure max(child_quota) <= parent_quota, inherit when no 7905 * limit is set 7906 */ 7907 if (quota == RUNTIME_INF) 7908 quota = parent_quota; 7909 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7910 return -EINVAL; 7911 } 7912 cfs_b->hierarchal_quota = quota; 7913 7914 return 0; 7915 } 7916 7917 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7918 { 7919 int ret; 7920 struct cfs_schedulable_data data = { 7921 .tg = tg, 7922 .period = period, 7923 .quota = quota, 7924 }; 7925 7926 if (quota != RUNTIME_INF) { 7927 do_div(data.period, NSEC_PER_USEC); 7928 do_div(data.quota, NSEC_PER_USEC); 7929 } 7930 7931 rcu_read_lock(); 7932 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7933 rcu_read_unlock(); 7934 7935 return ret; 7936 } 7937 7938 static int cpu_stats_show(struct seq_file *sf, void *v) 7939 { 7940 struct task_group *tg = css_tg(seq_css(sf)); 7941 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7942 7943 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7944 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7945 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7946 7947 return 0; 7948 } 7949 #endif /* CONFIG_CFS_BANDWIDTH */ 7950 #endif /* CONFIG_FAIR_GROUP_SCHED */ 7951 7952 #ifdef CONFIG_RT_GROUP_SCHED 7953 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7954 struct cftype *cft, s64 val) 7955 { 7956 return sched_group_set_rt_runtime(css_tg(css), val); 7957 } 7958 7959 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7960 struct cftype *cft) 7961 { 7962 return sched_group_rt_runtime(css_tg(css)); 7963 } 7964 7965 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7966 struct cftype *cftype, u64 rt_period_us) 7967 { 7968 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7969 } 7970 7971 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7972 struct cftype *cft) 7973 { 7974 return sched_group_rt_period(css_tg(css)); 7975 } 7976 #endif /* CONFIG_RT_GROUP_SCHED */ 7977 7978 static struct cftype cpu_files[] = { 7979 #ifdef CONFIG_FAIR_GROUP_SCHED 7980 { 7981 .name = "shares", 7982 .read_u64 = cpu_shares_read_u64, 7983 .write_u64 = cpu_shares_write_u64, 7984 }, 7985 #endif 7986 #ifdef CONFIG_CFS_BANDWIDTH 7987 { 7988 .name = "cfs_quota_us", 7989 .read_s64 = cpu_cfs_quota_read_s64, 7990 .write_s64 = cpu_cfs_quota_write_s64, 7991 }, 7992 { 7993 .name = "cfs_period_us", 7994 .read_u64 = cpu_cfs_period_read_u64, 7995 .write_u64 = cpu_cfs_period_write_u64, 7996 }, 7997 { 7998 .name = "stat", 7999 .seq_show = cpu_stats_show, 8000 }, 8001 #endif 8002 #ifdef CONFIG_RT_GROUP_SCHED 8003 { 8004 .name = "rt_runtime_us", 8005 .read_s64 = cpu_rt_runtime_read, 8006 .write_s64 = cpu_rt_runtime_write, 8007 }, 8008 { 8009 .name = "rt_period_us", 8010 .read_u64 = cpu_rt_period_read_uint, 8011 .write_u64 = cpu_rt_period_write_uint, 8012 }, 8013 #endif 8014 { } /* terminate */ 8015 }; 8016 8017 struct cgroup_subsys cpu_cgroup_subsys = { 8018 .name = "cpu", 8019 .css_alloc = cpu_cgroup_css_alloc, 8020 .css_free = cpu_cgroup_css_free, 8021 .css_online = cpu_cgroup_css_online, 8022 .css_offline = cpu_cgroup_css_offline, 8023 .can_attach = cpu_cgroup_can_attach, 8024 .attach = cpu_cgroup_attach, 8025 .exit = cpu_cgroup_exit, 8026 .subsys_id = cpu_cgroup_subsys_id, 8027 .base_cftypes = cpu_files, 8028 .early_init = 1, 8029 }; 8030 8031 #endif /* CONFIG_CGROUP_SCHED */ 8032 8033 void dump_cpu_task(int cpu) 8034 { 8035 pr_info("Task dump for CPU %d:\n", cpu); 8036 sched_show_task(cpu_curr(cpu)); 8037 } 8038