1 /* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29 #include <linux/mm.h> 30 #include <linux/module.h> 31 #include <linux/nmi.h> 32 #include <linux/init.h> 33 #include <linux/uaccess.h> 34 #include <linux/highmem.h> 35 #include <asm/mmu_context.h> 36 #include <linux/interrupt.h> 37 #include <linux/capability.h> 38 #include <linux/completion.h> 39 #include <linux/kernel_stat.h> 40 #include <linux/debug_locks.h> 41 #include <linux/perf_event.h> 42 #include <linux/security.h> 43 #include <linux/notifier.h> 44 #include <linux/profile.h> 45 #include <linux/freezer.h> 46 #include <linux/vmalloc.h> 47 #include <linux/blkdev.h> 48 #include <linux/delay.h> 49 #include <linux/pid_namespace.h> 50 #include <linux/smp.h> 51 #include <linux/threads.h> 52 #include <linux/timer.h> 53 #include <linux/rcupdate.h> 54 #include <linux/cpu.h> 55 #include <linux/cpuset.h> 56 #include <linux/percpu.h> 57 #include <linux/proc_fs.h> 58 #include <linux/seq_file.h> 59 #include <linux/sysctl.h> 60 #include <linux/syscalls.h> 61 #include <linux/times.h> 62 #include <linux/tsacct_kern.h> 63 #include <linux/kprobes.h> 64 #include <linux/delayacct.h> 65 #include <linux/unistd.h> 66 #include <linux/pagemap.h> 67 #include <linux/hrtimer.h> 68 #include <linux/tick.h> 69 #include <linux/debugfs.h> 70 #include <linux/ctype.h> 71 #include <linux/ftrace.h> 72 #include <linux/slab.h> 73 #include <linux/init_task.h> 74 #include <linux/binfmts.h> 75 #include <linux/context_tracking.h> 76 #include <linux/compiler.h> 77 78 #include <asm/switch_to.h> 79 #include <asm/tlb.h> 80 #include <asm/irq_regs.h> 81 #include <asm/mutex.h> 82 #ifdef CONFIG_PARAVIRT 83 #include <asm/paravirt.h> 84 #endif 85 86 #include "sched.h" 87 #include "../workqueue_internal.h" 88 #include "../smpboot.h" 89 90 #define CREATE_TRACE_POINTS 91 #include <trace/events/sched.h> 92 93 DEFINE_MUTEX(sched_domains_mutex); 94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 95 96 static void update_rq_clock_task(struct rq *rq, s64 delta); 97 98 void update_rq_clock(struct rq *rq) 99 { 100 s64 delta; 101 102 lockdep_assert_held(&rq->lock); 103 104 if (rq->clock_skip_update & RQCF_ACT_SKIP) 105 return; 106 107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 108 if (delta < 0) 109 return; 110 rq->clock += delta; 111 update_rq_clock_task(rq, delta); 112 } 113 114 /* 115 * Debugging: various feature bits 116 */ 117 118 #define SCHED_FEAT(name, enabled) \ 119 (1UL << __SCHED_FEAT_##name) * enabled | 120 121 const_debug unsigned int sysctl_sched_features = 122 #include "features.h" 123 0; 124 125 #undef SCHED_FEAT 126 127 #ifdef CONFIG_SCHED_DEBUG 128 #define SCHED_FEAT(name, enabled) \ 129 #name , 130 131 static const char * const sched_feat_names[] = { 132 #include "features.h" 133 }; 134 135 #undef SCHED_FEAT 136 137 static int sched_feat_show(struct seq_file *m, void *v) 138 { 139 int i; 140 141 for (i = 0; i < __SCHED_FEAT_NR; i++) { 142 if (!(sysctl_sched_features & (1UL << i))) 143 seq_puts(m, "NO_"); 144 seq_printf(m, "%s ", sched_feat_names[i]); 145 } 146 seq_puts(m, "\n"); 147 148 return 0; 149 } 150 151 #ifdef HAVE_JUMP_LABEL 152 153 #define jump_label_key__true STATIC_KEY_INIT_TRUE 154 #define jump_label_key__false STATIC_KEY_INIT_FALSE 155 156 #define SCHED_FEAT(name, enabled) \ 157 jump_label_key__##enabled , 158 159 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 160 #include "features.h" 161 }; 162 163 #undef SCHED_FEAT 164 165 static void sched_feat_disable(int i) 166 { 167 static_key_disable(&sched_feat_keys[i]); 168 } 169 170 static void sched_feat_enable(int i) 171 { 172 static_key_enable(&sched_feat_keys[i]); 173 } 174 #else 175 static void sched_feat_disable(int i) { }; 176 static void sched_feat_enable(int i) { }; 177 #endif /* HAVE_JUMP_LABEL */ 178 179 static int sched_feat_set(char *cmp) 180 { 181 int i; 182 int neg = 0; 183 184 if (strncmp(cmp, "NO_", 3) == 0) { 185 neg = 1; 186 cmp += 3; 187 } 188 189 for (i = 0; i < __SCHED_FEAT_NR; i++) { 190 if (strcmp(cmp, sched_feat_names[i]) == 0) { 191 if (neg) { 192 sysctl_sched_features &= ~(1UL << i); 193 sched_feat_disable(i); 194 } else { 195 sysctl_sched_features |= (1UL << i); 196 sched_feat_enable(i); 197 } 198 break; 199 } 200 } 201 202 return i; 203 } 204 205 static ssize_t 206 sched_feat_write(struct file *filp, const char __user *ubuf, 207 size_t cnt, loff_t *ppos) 208 { 209 char buf[64]; 210 char *cmp; 211 int i; 212 struct inode *inode; 213 214 if (cnt > 63) 215 cnt = 63; 216 217 if (copy_from_user(&buf, ubuf, cnt)) 218 return -EFAULT; 219 220 buf[cnt] = 0; 221 cmp = strstrip(buf); 222 223 /* Ensure the static_key remains in a consistent state */ 224 inode = file_inode(filp); 225 mutex_lock(&inode->i_mutex); 226 i = sched_feat_set(cmp); 227 mutex_unlock(&inode->i_mutex); 228 if (i == __SCHED_FEAT_NR) 229 return -EINVAL; 230 231 *ppos += cnt; 232 233 return cnt; 234 } 235 236 static int sched_feat_open(struct inode *inode, struct file *filp) 237 { 238 return single_open(filp, sched_feat_show, NULL); 239 } 240 241 static const struct file_operations sched_feat_fops = { 242 .open = sched_feat_open, 243 .write = sched_feat_write, 244 .read = seq_read, 245 .llseek = seq_lseek, 246 .release = single_release, 247 }; 248 249 static __init int sched_init_debug(void) 250 { 251 debugfs_create_file("sched_features", 0644, NULL, NULL, 252 &sched_feat_fops); 253 254 return 0; 255 } 256 late_initcall(sched_init_debug); 257 #endif /* CONFIG_SCHED_DEBUG */ 258 259 /* 260 * Number of tasks to iterate in a single balance run. 261 * Limited because this is done with IRQs disabled. 262 */ 263 const_debug unsigned int sysctl_sched_nr_migrate = 32; 264 265 /* 266 * period over which we average the RT time consumption, measured 267 * in ms. 268 * 269 * default: 1s 270 */ 271 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 272 273 /* 274 * period over which we measure -rt task cpu usage in us. 275 * default: 1s 276 */ 277 unsigned int sysctl_sched_rt_period = 1000000; 278 279 __read_mostly int scheduler_running; 280 281 /* 282 * part of the period that we allow rt tasks to run in us. 283 * default: 0.95s 284 */ 285 int sysctl_sched_rt_runtime = 950000; 286 287 /* cpus with isolated domains */ 288 cpumask_var_t cpu_isolated_map; 289 290 /* 291 * this_rq_lock - lock this runqueue and disable interrupts. 292 */ 293 static struct rq *this_rq_lock(void) 294 __acquires(rq->lock) 295 { 296 struct rq *rq; 297 298 local_irq_disable(); 299 rq = this_rq(); 300 raw_spin_lock(&rq->lock); 301 302 return rq; 303 } 304 305 #ifdef CONFIG_SCHED_HRTICK 306 /* 307 * Use HR-timers to deliver accurate preemption points. 308 */ 309 310 static void hrtick_clear(struct rq *rq) 311 { 312 if (hrtimer_active(&rq->hrtick_timer)) 313 hrtimer_cancel(&rq->hrtick_timer); 314 } 315 316 /* 317 * High-resolution timer tick. 318 * Runs from hardirq context with interrupts disabled. 319 */ 320 static enum hrtimer_restart hrtick(struct hrtimer *timer) 321 { 322 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 323 324 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 325 326 raw_spin_lock(&rq->lock); 327 update_rq_clock(rq); 328 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 329 raw_spin_unlock(&rq->lock); 330 331 return HRTIMER_NORESTART; 332 } 333 334 #ifdef CONFIG_SMP 335 336 static void __hrtick_restart(struct rq *rq) 337 { 338 struct hrtimer *timer = &rq->hrtick_timer; 339 340 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED); 341 } 342 343 /* 344 * called from hardirq (IPI) context 345 */ 346 static void __hrtick_start(void *arg) 347 { 348 struct rq *rq = arg; 349 350 raw_spin_lock(&rq->lock); 351 __hrtick_restart(rq); 352 rq->hrtick_csd_pending = 0; 353 raw_spin_unlock(&rq->lock); 354 } 355 356 /* 357 * Called to set the hrtick timer state. 358 * 359 * called with rq->lock held and irqs disabled 360 */ 361 void hrtick_start(struct rq *rq, u64 delay) 362 { 363 struct hrtimer *timer = &rq->hrtick_timer; 364 ktime_t time; 365 s64 delta; 366 367 /* 368 * Don't schedule slices shorter than 10000ns, that just 369 * doesn't make sense and can cause timer DoS. 370 */ 371 delta = max_t(s64, delay, 10000LL); 372 time = ktime_add_ns(timer->base->get_time(), delta); 373 374 hrtimer_set_expires(timer, time); 375 376 if (rq == this_rq()) { 377 __hrtick_restart(rq); 378 } else if (!rq->hrtick_csd_pending) { 379 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 380 rq->hrtick_csd_pending = 1; 381 } 382 } 383 384 static int 385 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 386 { 387 int cpu = (int)(long)hcpu; 388 389 switch (action) { 390 case CPU_UP_CANCELED: 391 case CPU_UP_CANCELED_FROZEN: 392 case CPU_DOWN_PREPARE: 393 case CPU_DOWN_PREPARE_FROZEN: 394 case CPU_DEAD: 395 case CPU_DEAD_FROZEN: 396 hrtick_clear(cpu_rq(cpu)); 397 return NOTIFY_OK; 398 } 399 400 return NOTIFY_DONE; 401 } 402 403 static __init void init_hrtick(void) 404 { 405 hotcpu_notifier(hotplug_hrtick, 0); 406 } 407 #else 408 /* 409 * Called to set the hrtick timer state. 410 * 411 * called with rq->lock held and irqs disabled 412 */ 413 void hrtick_start(struct rq *rq, u64 delay) 414 { 415 /* 416 * Don't schedule slices shorter than 10000ns, that just 417 * doesn't make sense. Rely on vruntime for fairness. 418 */ 419 delay = max_t(u64, delay, 10000LL); 420 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), 421 HRTIMER_MODE_REL_PINNED); 422 } 423 424 static inline void init_hrtick(void) 425 { 426 } 427 #endif /* CONFIG_SMP */ 428 429 static void init_rq_hrtick(struct rq *rq) 430 { 431 #ifdef CONFIG_SMP 432 rq->hrtick_csd_pending = 0; 433 434 rq->hrtick_csd.flags = 0; 435 rq->hrtick_csd.func = __hrtick_start; 436 rq->hrtick_csd.info = rq; 437 #endif 438 439 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 440 rq->hrtick_timer.function = hrtick; 441 } 442 #else /* CONFIG_SCHED_HRTICK */ 443 static inline void hrtick_clear(struct rq *rq) 444 { 445 } 446 447 static inline void init_rq_hrtick(struct rq *rq) 448 { 449 } 450 451 static inline void init_hrtick(void) 452 { 453 } 454 #endif /* CONFIG_SCHED_HRTICK */ 455 456 /* 457 * cmpxchg based fetch_or, macro so it works for different integer types 458 */ 459 #define fetch_or(ptr, val) \ 460 ({ typeof(*(ptr)) __old, __val = *(ptr); \ 461 for (;;) { \ 462 __old = cmpxchg((ptr), __val, __val | (val)); \ 463 if (__old == __val) \ 464 break; \ 465 __val = __old; \ 466 } \ 467 __old; \ 468 }) 469 470 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) 471 /* 472 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 473 * this avoids any races wrt polling state changes and thereby avoids 474 * spurious IPIs. 475 */ 476 static bool set_nr_and_not_polling(struct task_struct *p) 477 { 478 struct thread_info *ti = task_thread_info(p); 479 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 480 } 481 482 /* 483 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. 484 * 485 * If this returns true, then the idle task promises to call 486 * sched_ttwu_pending() and reschedule soon. 487 */ 488 static bool set_nr_if_polling(struct task_struct *p) 489 { 490 struct thread_info *ti = task_thread_info(p); 491 typeof(ti->flags) old, val = READ_ONCE(ti->flags); 492 493 for (;;) { 494 if (!(val & _TIF_POLLING_NRFLAG)) 495 return false; 496 if (val & _TIF_NEED_RESCHED) 497 return true; 498 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); 499 if (old == val) 500 break; 501 val = old; 502 } 503 return true; 504 } 505 506 #else 507 static bool set_nr_and_not_polling(struct task_struct *p) 508 { 509 set_tsk_need_resched(p); 510 return true; 511 } 512 513 #ifdef CONFIG_SMP 514 static bool set_nr_if_polling(struct task_struct *p) 515 { 516 return false; 517 } 518 #endif 519 #endif 520 521 void wake_q_add(struct wake_q_head *head, struct task_struct *task) 522 { 523 struct wake_q_node *node = &task->wake_q; 524 525 /* 526 * Atomically grab the task, if ->wake_q is !nil already it means 527 * its already queued (either by us or someone else) and will get the 528 * wakeup due to that. 529 * 530 * This cmpxchg() implies a full barrier, which pairs with the write 531 * barrier implied by the wakeup in wake_up_list(). 532 */ 533 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL)) 534 return; 535 536 get_task_struct(task); 537 538 /* 539 * The head is context local, there can be no concurrency. 540 */ 541 *head->lastp = node; 542 head->lastp = &node->next; 543 } 544 545 void wake_up_q(struct wake_q_head *head) 546 { 547 struct wake_q_node *node = head->first; 548 549 while (node != WAKE_Q_TAIL) { 550 struct task_struct *task; 551 552 task = container_of(node, struct task_struct, wake_q); 553 BUG_ON(!task); 554 /* task can safely be re-inserted now */ 555 node = node->next; 556 task->wake_q.next = NULL; 557 558 /* 559 * wake_up_process() implies a wmb() to pair with the queueing 560 * in wake_q_add() so as not to miss wakeups. 561 */ 562 wake_up_process(task); 563 put_task_struct(task); 564 } 565 } 566 567 /* 568 * resched_curr - mark rq's current task 'to be rescheduled now'. 569 * 570 * On UP this means the setting of the need_resched flag, on SMP it 571 * might also involve a cross-CPU call to trigger the scheduler on 572 * the target CPU. 573 */ 574 void resched_curr(struct rq *rq) 575 { 576 struct task_struct *curr = rq->curr; 577 int cpu; 578 579 lockdep_assert_held(&rq->lock); 580 581 if (test_tsk_need_resched(curr)) 582 return; 583 584 cpu = cpu_of(rq); 585 586 if (cpu == smp_processor_id()) { 587 set_tsk_need_resched(curr); 588 set_preempt_need_resched(); 589 return; 590 } 591 592 if (set_nr_and_not_polling(curr)) 593 smp_send_reschedule(cpu); 594 else 595 trace_sched_wake_idle_without_ipi(cpu); 596 } 597 598 void resched_cpu(int cpu) 599 { 600 struct rq *rq = cpu_rq(cpu); 601 unsigned long flags; 602 603 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 604 return; 605 resched_curr(rq); 606 raw_spin_unlock_irqrestore(&rq->lock, flags); 607 } 608 609 #ifdef CONFIG_SMP 610 #ifdef CONFIG_NO_HZ_COMMON 611 /* 612 * In the semi idle case, use the nearest busy cpu for migrating timers 613 * from an idle cpu. This is good for power-savings. 614 * 615 * We don't do similar optimization for completely idle system, as 616 * selecting an idle cpu will add more delays to the timers than intended 617 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 618 */ 619 int get_nohz_timer_target(void) 620 { 621 int i, cpu = smp_processor_id(); 622 struct sched_domain *sd; 623 624 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu)) 625 return cpu; 626 627 rcu_read_lock(); 628 for_each_domain(cpu, sd) { 629 for_each_cpu(i, sched_domain_span(sd)) { 630 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) { 631 cpu = i; 632 goto unlock; 633 } 634 } 635 } 636 637 if (!is_housekeeping_cpu(cpu)) 638 cpu = housekeeping_any_cpu(); 639 unlock: 640 rcu_read_unlock(); 641 return cpu; 642 } 643 /* 644 * When add_timer_on() enqueues a timer into the timer wheel of an 645 * idle CPU then this timer might expire before the next timer event 646 * which is scheduled to wake up that CPU. In case of a completely 647 * idle system the next event might even be infinite time into the 648 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 649 * leaves the inner idle loop so the newly added timer is taken into 650 * account when the CPU goes back to idle and evaluates the timer 651 * wheel for the next timer event. 652 */ 653 static void wake_up_idle_cpu(int cpu) 654 { 655 struct rq *rq = cpu_rq(cpu); 656 657 if (cpu == smp_processor_id()) 658 return; 659 660 if (set_nr_and_not_polling(rq->idle)) 661 smp_send_reschedule(cpu); 662 else 663 trace_sched_wake_idle_without_ipi(cpu); 664 } 665 666 static bool wake_up_full_nohz_cpu(int cpu) 667 { 668 /* 669 * We just need the target to call irq_exit() and re-evaluate 670 * the next tick. The nohz full kick at least implies that. 671 * If needed we can still optimize that later with an 672 * empty IRQ. 673 */ 674 if (tick_nohz_full_cpu(cpu)) { 675 if (cpu != smp_processor_id() || 676 tick_nohz_tick_stopped()) 677 tick_nohz_full_kick_cpu(cpu); 678 return true; 679 } 680 681 return false; 682 } 683 684 void wake_up_nohz_cpu(int cpu) 685 { 686 if (!wake_up_full_nohz_cpu(cpu)) 687 wake_up_idle_cpu(cpu); 688 } 689 690 static inline bool got_nohz_idle_kick(void) 691 { 692 int cpu = smp_processor_id(); 693 694 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 695 return false; 696 697 if (idle_cpu(cpu) && !need_resched()) 698 return true; 699 700 /* 701 * We can't run Idle Load Balance on this CPU for this time so we 702 * cancel it and clear NOHZ_BALANCE_KICK 703 */ 704 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 705 return false; 706 } 707 708 #else /* CONFIG_NO_HZ_COMMON */ 709 710 static inline bool got_nohz_idle_kick(void) 711 { 712 return false; 713 } 714 715 #endif /* CONFIG_NO_HZ_COMMON */ 716 717 #ifdef CONFIG_NO_HZ_FULL 718 bool sched_can_stop_tick(void) 719 { 720 /* 721 * FIFO realtime policy runs the highest priority task. Other runnable 722 * tasks are of a lower priority. The scheduler tick does nothing. 723 */ 724 if (current->policy == SCHED_FIFO) 725 return true; 726 727 /* 728 * Round-robin realtime tasks time slice with other tasks at the same 729 * realtime priority. Is this task the only one at this priority? 730 */ 731 if (current->policy == SCHED_RR) { 732 struct sched_rt_entity *rt_se = ¤t->rt; 733 734 return rt_se->run_list.prev == rt_se->run_list.next; 735 } 736 737 /* 738 * More than one running task need preemption. 739 * nr_running update is assumed to be visible 740 * after IPI is sent from wakers. 741 */ 742 if (this_rq()->nr_running > 1) 743 return false; 744 745 return true; 746 } 747 #endif /* CONFIG_NO_HZ_FULL */ 748 749 void sched_avg_update(struct rq *rq) 750 { 751 s64 period = sched_avg_period(); 752 753 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 754 /* 755 * Inline assembly required to prevent the compiler 756 * optimising this loop into a divmod call. 757 * See __iter_div_u64_rem() for another example of this. 758 */ 759 asm("" : "+rm" (rq->age_stamp)); 760 rq->age_stamp += period; 761 rq->rt_avg /= 2; 762 } 763 } 764 765 #endif /* CONFIG_SMP */ 766 767 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 768 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 769 /* 770 * Iterate task_group tree rooted at *from, calling @down when first entering a 771 * node and @up when leaving it for the final time. 772 * 773 * Caller must hold rcu_lock or sufficient equivalent. 774 */ 775 int walk_tg_tree_from(struct task_group *from, 776 tg_visitor down, tg_visitor up, void *data) 777 { 778 struct task_group *parent, *child; 779 int ret; 780 781 parent = from; 782 783 down: 784 ret = (*down)(parent, data); 785 if (ret) 786 goto out; 787 list_for_each_entry_rcu(child, &parent->children, siblings) { 788 parent = child; 789 goto down; 790 791 up: 792 continue; 793 } 794 ret = (*up)(parent, data); 795 if (ret || parent == from) 796 goto out; 797 798 child = parent; 799 parent = parent->parent; 800 if (parent) 801 goto up; 802 out: 803 return ret; 804 } 805 806 int tg_nop(struct task_group *tg, void *data) 807 { 808 return 0; 809 } 810 #endif 811 812 static void set_load_weight(struct task_struct *p) 813 { 814 int prio = p->static_prio - MAX_RT_PRIO; 815 struct load_weight *load = &p->se.load; 816 817 /* 818 * SCHED_IDLE tasks get minimal weight: 819 */ 820 if (p->policy == SCHED_IDLE) { 821 load->weight = scale_load(WEIGHT_IDLEPRIO); 822 load->inv_weight = WMULT_IDLEPRIO; 823 return; 824 } 825 826 load->weight = scale_load(prio_to_weight[prio]); 827 load->inv_weight = prio_to_wmult[prio]; 828 } 829 830 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 831 { 832 update_rq_clock(rq); 833 sched_info_queued(rq, p); 834 p->sched_class->enqueue_task(rq, p, flags); 835 } 836 837 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 838 { 839 update_rq_clock(rq); 840 sched_info_dequeued(rq, p); 841 p->sched_class->dequeue_task(rq, p, flags); 842 } 843 844 void activate_task(struct rq *rq, struct task_struct *p, int flags) 845 { 846 if (task_contributes_to_load(p)) 847 rq->nr_uninterruptible--; 848 849 enqueue_task(rq, p, flags); 850 } 851 852 void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 853 { 854 if (task_contributes_to_load(p)) 855 rq->nr_uninterruptible++; 856 857 dequeue_task(rq, p, flags); 858 } 859 860 static void update_rq_clock_task(struct rq *rq, s64 delta) 861 { 862 /* 863 * In theory, the compile should just see 0 here, and optimize out the call 864 * to sched_rt_avg_update. But I don't trust it... 865 */ 866 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 867 s64 steal = 0, irq_delta = 0; 868 #endif 869 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 870 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 871 872 /* 873 * Since irq_time is only updated on {soft,}irq_exit, we might run into 874 * this case when a previous update_rq_clock() happened inside a 875 * {soft,}irq region. 876 * 877 * When this happens, we stop ->clock_task and only update the 878 * prev_irq_time stamp to account for the part that fit, so that a next 879 * update will consume the rest. This ensures ->clock_task is 880 * monotonic. 881 * 882 * It does however cause some slight miss-attribution of {soft,}irq 883 * time, a more accurate solution would be to update the irq_time using 884 * the current rq->clock timestamp, except that would require using 885 * atomic ops. 886 */ 887 if (irq_delta > delta) 888 irq_delta = delta; 889 890 rq->prev_irq_time += irq_delta; 891 delta -= irq_delta; 892 #endif 893 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 894 if (static_key_false((¶virt_steal_rq_enabled))) { 895 steal = paravirt_steal_clock(cpu_of(rq)); 896 steal -= rq->prev_steal_time_rq; 897 898 if (unlikely(steal > delta)) 899 steal = delta; 900 901 rq->prev_steal_time_rq += steal; 902 delta -= steal; 903 } 904 #endif 905 906 rq->clock_task += delta; 907 908 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 909 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 910 sched_rt_avg_update(rq, irq_delta + steal); 911 #endif 912 } 913 914 void sched_set_stop_task(int cpu, struct task_struct *stop) 915 { 916 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 917 struct task_struct *old_stop = cpu_rq(cpu)->stop; 918 919 if (stop) { 920 /* 921 * Make it appear like a SCHED_FIFO task, its something 922 * userspace knows about and won't get confused about. 923 * 924 * Also, it will make PI more or less work without too 925 * much confusion -- but then, stop work should not 926 * rely on PI working anyway. 927 */ 928 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 929 930 stop->sched_class = &stop_sched_class; 931 } 932 933 cpu_rq(cpu)->stop = stop; 934 935 if (old_stop) { 936 /* 937 * Reset it back to a normal scheduling class so that 938 * it can die in pieces. 939 */ 940 old_stop->sched_class = &rt_sched_class; 941 } 942 } 943 944 /* 945 * __normal_prio - return the priority that is based on the static prio 946 */ 947 static inline int __normal_prio(struct task_struct *p) 948 { 949 return p->static_prio; 950 } 951 952 /* 953 * Calculate the expected normal priority: i.e. priority 954 * without taking RT-inheritance into account. Might be 955 * boosted by interactivity modifiers. Changes upon fork, 956 * setprio syscalls, and whenever the interactivity 957 * estimator recalculates. 958 */ 959 static inline int normal_prio(struct task_struct *p) 960 { 961 int prio; 962 963 if (task_has_dl_policy(p)) 964 prio = MAX_DL_PRIO-1; 965 else if (task_has_rt_policy(p)) 966 prio = MAX_RT_PRIO-1 - p->rt_priority; 967 else 968 prio = __normal_prio(p); 969 return prio; 970 } 971 972 /* 973 * Calculate the current priority, i.e. the priority 974 * taken into account by the scheduler. This value might 975 * be boosted by RT tasks, or might be boosted by 976 * interactivity modifiers. Will be RT if the task got 977 * RT-boosted. If not then it returns p->normal_prio. 978 */ 979 static int effective_prio(struct task_struct *p) 980 { 981 p->normal_prio = normal_prio(p); 982 /* 983 * If we are RT tasks or we were boosted to RT priority, 984 * keep the priority unchanged. Otherwise, update priority 985 * to the normal priority: 986 */ 987 if (!rt_prio(p->prio)) 988 return p->normal_prio; 989 return p->prio; 990 } 991 992 /** 993 * task_curr - is this task currently executing on a CPU? 994 * @p: the task in question. 995 * 996 * Return: 1 if the task is currently executing. 0 otherwise. 997 */ 998 inline int task_curr(const struct task_struct *p) 999 { 1000 return cpu_curr(task_cpu(p)) == p; 1001 } 1002 1003 /* 1004 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock, 1005 * use the balance_callback list if you want balancing. 1006 * 1007 * this means any call to check_class_changed() must be followed by a call to 1008 * balance_callback(). 1009 */ 1010 static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1011 const struct sched_class *prev_class, 1012 int oldprio) 1013 { 1014 if (prev_class != p->sched_class) { 1015 if (prev_class->switched_from) 1016 prev_class->switched_from(rq, p); 1017 1018 p->sched_class->switched_to(rq, p); 1019 } else if (oldprio != p->prio || dl_task(p)) 1020 p->sched_class->prio_changed(rq, p, oldprio); 1021 } 1022 1023 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1024 { 1025 const struct sched_class *class; 1026 1027 if (p->sched_class == rq->curr->sched_class) { 1028 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1029 } else { 1030 for_each_class(class) { 1031 if (class == rq->curr->sched_class) 1032 break; 1033 if (class == p->sched_class) { 1034 resched_curr(rq); 1035 break; 1036 } 1037 } 1038 } 1039 1040 /* 1041 * A queue event has occurred, and we're going to schedule. In 1042 * this case, we can save a useless back to back clock update. 1043 */ 1044 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) 1045 rq_clock_skip_update(rq, true); 1046 } 1047 1048 #ifdef CONFIG_SMP 1049 /* 1050 * This is how migration works: 1051 * 1052 * 1) we invoke migration_cpu_stop() on the target CPU using 1053 * stop_one_cpu(). 1054 * 2) stopper starts to run (implicitly forcing the migrated thread 1055 * off the CPU) 1056 * 3) it checks whether the migrated task is still in the wrong runqueue. 1057 * 4) if it's in the wrong runqueue then the migration thread removes 1058 * it and puts it into the right queue. 1059 * 5) stopper completes and stop_one_cpu() returns and the migration 1060 * is done. 1061 */ 1062 1063 /* 1064 * move_queued_task - move a queued task to new rq. 1065 * 1066 * Returns (locked) new rq. Old rq's lock is released. 1067 */ 1068 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu) 1069 { 1070 lockdep_assert_held(&rq->lock); 1071 1072 dequeue_task(rq, p, 0); 1073 p->on_rq = TASK_ON_RQ_MIGRATING; 1074 set_task_cpu(p, new_cpu); 1075 raw_spin_unlock(&rq->lock); 1076 1077 rq = cpu_rq(new_cpu); 1078 1079 raw_spin_lock(&rq->lock); 1080 BUG_ON(task_cpu(p) != new_cpu); 1081 p->on_rq = TASK_ON_RQ_QUEUED; 1082 enqueue_task(rq, p, 0); 1083 check_preempt_curr(rq, p, 0); 1084 1085 return rq; 1086 } 1087 1088 struct migration_arg { 1089 struct task_struct *task; 1090 int dest_cpu; 1091 }; 1092 1093 /* 1094 * Move (not current) task off this cpu, onto dest cpu. We're doing 1095 * this because either it can't run here any more (set_cpus_allowed() 1096 * away from this CPU, or CPU going down), or because we're 1097 * attempting to rebalance this task on exec (sched_exec). 1098 * 1099 * So we race with normal scheduler movements, but that's OK, as long 1100 * as the task is no longer on this CPU. 1101 */ 1102 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu) 1103 { 1104 if (unlikely(!cpu_active(dest_cpu))) 1105 return rq; 1106 1107 /* Affinity changed (again). */ 1108 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1109 return rq; 1110 1111 rq = move_queued_task(rq, p, dest_cpu); 1112 1113 return rq; 1114 } 1115 1116 /* 1117 * migration_cpu_stop - this will be executed by a highprio stopper thread 1118 * and performs thread migration by bumping thread off CPU then 1119 * 'pushing' onto another runqueue. 1120 */ 1121 static int migration_cpu_stop(void *data) 1122 { 1123 struct migration_arg *arg = data; 1124 struct task_struct *p = arg->task; 1125 struct rq *rq = this_rq(); 1126 1127 /* 1128 * The original target cpu might have gone down and we might 1129 * be on another cpu but it doesn't matter. 1130 */ 1131 local_irq_disable(); 1132 /* 1133 * We need to explicitly wake pending tasks before running 1134 * __migrate_task() such that we will not miss enforcing cpus_allowed 1135 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. 1136 */ 1137 sched_ttwu_pending(); 1138 1139 raw_spin_lock(&p->pi_lock); 1140 raw_spin_lock(&rq->lock); 1141 /* 1142 * If task_rq(p) != rq, it cannot be migrated here, because we're 1143 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because 1144 * we're holding p->pi_lock. 1145 */ 1146 if (task_rq(p) == rq && task_on_rq_queued(p)) 1147 rq = __migrate_task(rq, p, arg->dest_cpu); 1148 raw_spin_unlock(&rq->lock); 1149 raw_spin_unlock(&p->pi_lock); 1150 1151 local_irq_enable(); 1152 return 0; 1153 } 1154 1155 /* 1156 * sched_class::set_cpus_allowed must do the below, but is not required to 1157 * actually call this function. 1158 */ 1159 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask) 1160 { 1161 cpumask_copy(&p->cpus_allowed, new_mask); 1162 p->nr_cpus_allowed = cpumask_weight(new_mask); 1163 } 1164 1165 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 1166 { 1167 struct rq *rq = task_rq(p); 1168 bool queued, running; 1169 1170 lockdep_assert_held(&p->pi_lock); 1171 1172 queued = task_on_rq_queued(p); 1173 running = task_current(rq, p); 1174 1175 if (queued) { 1176 /* 1177 * Because __kthread_bind() calls this on blocked tasks without 1178 * holding rq->lock. 1179 */ 1180 lockdep_assert_held(&rq->lock); 1181 dequeue_task(rq, p, 0); 1182 } 1183 if (running) 1184 put_prev_task(rq, p); 1185 1186 p->sched_class->set_cpus_allowed(p, new_mask); 1187 1188 if (running) 1189 p->sched_class->set_curr_task(rq); 1190 if (queued) 1191 enqueue_task(rq, p, 0); 1192 } 1193 1194 /* 1195 * Change a given task's CPU affinity. Migrate the thread to a 1196 * proper CPU and schedule it away if the CPU it's executing on 1197 * is removed from the allowed bitmask. 1198 * 1199 * NOTE: the caller must have a valid reference to the task, the 1200 * task must not exit() & deallocate itself prematurely. The 1201 * call is not atomic; no spinlocks may be held. 1202 */ 1203 static int __set_cpus_allowed_ptr(struct task_struct *p, 1204 const struct cpumask *new_mask, bool check) 1205 { 1206 unsigned long flags; 1207 struct rq *rq; 1208 unsigned int dest_cpu; 1209 int ret = 0; 1210 1211 rq = task_rq_lock(p, &flags); 1212 1213 /* 1214 * Must re-check here, to close a race against __kthread_bind(), 1215 * sched_setaffinity() is not guaranteed to observe the flag. 1216 */ 1217 if (check && (p->flags & PF_NO_SETAFFINITY)) { 1218 ret = -EINVAL; 1219 goto out; 1220 } 1221 1222 if (cpumask_equal(&p->cpus_allowed, new_mask)) 1223 goto out; 1224 1225 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 1226 ret = -EINVAL; 1227 goto out; 1228 } 1229 1230 do_set_cpus_allowed(p, new_mask); 1231 1232 /* Can the task run on the task's current CPU? If so, we're done */ 1233 if (cpumask_test_cpu(task_cpu(p), new_mask)) 1234 goto out; 1235 1236 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 1237 if (task_running(rq, p) || p->state == TASK_WAKING) { 1238 struct migration_arg arg = { p, dest_cpu }; 1239 /* Need help from migration thread: drop lock and wait. */ 1240 task_rq_unlock(rq, p, &flags); 1241 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 1242 tlb_migrate_finish(p->mm); 1243 return 0; 1244 } else if (task_on_rq_queued(p)) { 1245 /* 1246 * OK, since we're going to drop the lock immediately 1247 * afterwards anyway. 1248 */ 1249 lockdep_unpin_lock(&rq->lock); 1250 rq = move_queued_task(rq, p, dest_cpu); 1251 lockdep_pin_lock(&rq->lock); 1252 } 1253 out: 1254 task_rq_unlock(rq, p, &flags); 1255 1256 return ret; 1257 } 1258 1259 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 1260 { 1261 return __set_cpus_allowed_ptr(p, new_mask, false); 1262 } 1263 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 1264 1265 void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1266 { 1267 #ifdef CONFIG_SCHED_DEBUG 1268 /* 1269 * We should never call set_task_cpu() on a blocked task, 1270 * ttwu() will sort out the placement. 1271 */ 1272 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1273 !p->on_rq); 1274 1275 #ifdef CONFIG_LOCKDEP 1276 /* 1277 * The caller should hold either p->pi_lock or rq->lock, when changing 1278 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1279 * 1280 * sched_move_task() holds both and thus holding either pins the cgroup, 1281 * see task_group(). 1282 * 1283 * Furthermore, all task_rq users should acquire both locks, see 1284 * task_rq_lock(). 1285 */ 1286 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1287 lockdep_is_held(&task_rq(p)->lock))); 1288 #endif 1289 #endif 1290 1291 trace_sched_migrate_task(p, new_cpu); 1292 1293 if (task_cpu(p) != new_cpu) { 1294 if (p->sched_class->migrate_task_rq) 1295 p->sched_class->migrate_task_rq(p, new_cpu); 1296 p->se.nr_migrations++; 1297 perf_event_task_migrate(p); 1298 } 1299 1300 __set_task_cpu(p, new_cpu); 1301 } 1302 1303 static void __migrate_swap_task(struct task_struct *p, int cpu) 1304 { 1305 if (task_on_rq_queued(p)) { 1306 struct rq *src_rq, *dst_rq; 1307 1308 src_rq = task_rq(p); 1309 dst_rq = cpu_rq(cpu); 1310 1311 deactivate_task(src_rq, p, 0); 1312 set_task_cpu(p, cpu); 1313 activate_task(dst_rq, p, 0); 1314 check_preempt_curr(dst_rq, p, 0); 1315 } else { 1316 /* 1317 * Task isn't running anymore; make it appear like we migrated 1318 * it before it went to sleep. This means on wakeup we make the 1319 * previous cpu our targer instead of where it really is. 1320 */ 1321 p->wake_cpu = cpu; 1322 } 1323 } 1324 1325 struct migration_swap_arg { 1326 struct task_struct *src_task, *dst_task; 1327 int src_cpu, dst_cpu; 1328 }; 1329 1330 static int migrate_swap_stop(void *data) 1331 { 1332 struct migration_swap_arg *arg = data; 1333 struct rq *src_rq, *dst_rq; 1334 int ret = -EAGAIN; 1335 1336 src_rq = cpu_rq(arg->src_cpu); 1337 dst_rq = cpu_rq(arg->dst_cpu); 1338 1339 double_raw_lock(&arg->src_task->pi_lock, 1340 &arg->dst_task->pi_lock); 1341 double_rq_lock(src_rq, dst_rq); 1342 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1343 goto unlock; 1344 1345 if (task_cpu(arg->src_task) != arg->src_cpu) 1346 goto unlock; 1347 1348 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1349 goto unlock; 1350 1351 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1352 goto unlock; 1353 1354 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1355 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1356 1357 ret = 0; 1358 1359 unlock: 1360 double_rq_unlock(src_rq, dst_rq); 1361 raw_spin_unlock(&arg->dst_task->pi_lock); 1362 raw_spin_unlock(&arg->src_task->pi_lock); 1363 1364 return ret; 1365 } 1366 1367 /* 1368 * Cross migrate two tasks 1369 */ 1370 int migrate_swap(struct task_struct *cur, struct task_struct *p) 1371 { 1372 struct migration_swap_arg arg; 1373 int ret = -EINVAL; 1374 1375 arg = (struct migration_swap_arg){ 1376 .src_task = cur, 1377 .src_cpu = task_cpu(cur), 1378 .dst_task = p, 1379 .dst_cpu = task_cpu(p), 1380 }; 1381 1382 if (arg.src_cpu == arg.dst_cpu) 1383 goto out; 1384 1385 /* 1386 * These three tests are all lockless; this is OK since all of them 1387 * will be re-checked with proper locks held further down the line. 1388 */ 1389 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1390 goto out; 1391 1392 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1393 goto out; 1394 1395 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1396 goto out; 1397 1398 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1399 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1400 1401 out: 1402 return ret; 1403 } 1404 1405 /* 1406 * wait_task_inactive - wait for a thread to unschedule. 1407 * 1408 * If @match_state is nonzero, it's the @p->state value just checked and 1409 * not expected to change. If it changes, i.e. @p might have woken up, 1410 * then return zero. When we succeed in waiting for @p to be off its CPU, 1411 * we return a positive number (its total switch count). If a second call 1412 * a short while later returns the same number, the caller can be sure that 1413 * @p has remained unscheduled the whole time. 1414 * 1415 * The caller must ensure that the task *will* unschedule sometime soon, 1416 * else this function might spin for a *long* time. This function can't 1417 * be called with interrupts off, or it may introduce deadlock with 1418 * smp_call_function() if an IPI is sent by the same process we are 1419 * waiting to become inactive. 1420 */ 1421 unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1422 { 1423 unsigned long flags; 1424 int running, queued; 1425 unsigned long ncsw; 1426 struct rq *rq; 1427 1428 for (;;) { 1429 /* 1430 * We do the initial early heuristics without holding 1431 * any task-queue locks at all. We'll only try to get 1432 * the runqueue lock when things look like they will 1433 * work out! 1434 */ 1435 rq = task_rq(p); 1436 1437 /* 1438 * If the task is actively running on another CPU 1439 * still, just relax and busy-wait without holding 1440 * any locks. 1441 * 1442 * NOTE! Since we don't hold any locks, it's not 1443 * even sure that "rq" stays as the right runqueue! 1444 * But we don't care, since "task_running()" will 1445 * return false if the runqueue has changed and p 1446 * is actually now running somewhere else! 1447 */ 1448 while (task_running(rq, p)) { 1449 if (match_state && unlikely(p->state != match_state)) 1450 return 0; 1451 cpu_relax(); 1452 } 1453 1454 /* 1455 * Ok, time to look more closely! We need the rq 1456 * lock now, to be *sure*. If we're wrong, we'll 1457 * just go back and repeat. 1458 */ 1459 rq = task_rq_lock(p, &flags); 1460 trace_sched_wait_task(p); 1461 running = task_running(rq, p); 1462 queued = task_on_rq_queued(p); 1463 ncsw = 0; 1464 if (!match_state || p->state == match_state) 1465 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1466 task_rq_unlock(rq, p, &flags); 1467 1468 /* 1469 * If it changed from the expected state, bail out now. 1470 */ 1471 if (unlikely(!ncsw)) 1472 break; 1473 1474 /* 1475 * Was it really running after all now that we 1476 * checked with the proper locks actually held? 1477 * 1478 * Oops. Go back and try again.. 1479 */ 1480 if (unlikely(running)) { 1481 cpu_relax(); 1482 continue; 1483 } 1484 1485 /* 1486 * It's not enough that it's not actively running, 1487 * it must be off the runqueue _entirely_, and not 1488 * preempted! 1489 * 1490 * So if it was still runnable (but just not actively 1491 * running right now), it's preempted, and we should 1492 * yield - it could be a while. 1493 */ 1494 if (unlikely(queued)) { 1495 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1496 1497 set_current_state(TASK_UNINTERRUPTIBLE); 1498 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1499 continue; 1500 } 1501 1502 /* 1503 * Ahh, all good. It wasn't running, and it wasn't 1504 * runnable, which means that it will never become 1505 * running in the future either. We're all done! 1506 */ 1507 break; 1508 } 1509 1510 return ncsw; 1511 } 1512 1513 /*** 1514 * kick_process - kick a running thread to enter/exit the kernel 1515 * @p: the to-be-kicked thread 1516 * 1517 * Cause a process which is running on another CPU to enter 1518 * kernel-mode, without any delay. (to get signals handled.) 1519 * 1520 * NOTE: this function doesn't have to take the runqueue lock, 1521 * because all it wants to ensure is that the remote task enters 1522 * the kernel. If the IPI races and the task has been migrated 1523 * to another CPU then no harm is done and the purpose has been 1524 * achieved as well. 1525 */ 1526 void kick_process(struct task_struct *p) 1527 { 1528 int cpu; 1529 1530 preempt_disable(); 1531 cpu = task_cpu(p); 1532 if ((cpu != smp_processor_id()) && task_curr(p)) 1533 smp_send_reschedule(cpu); 1534 preempt_enable(); 1535 } 1536 EXPORT_SYMBOL_GPL(kick_process); 1537 1538 /* 1539 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1540 */ 1541 static int select_fallback_rq(int cpu, struct task_struct *p) 1542 { 1543 int nid = cpu_to_node(cpu); 1544 const struct cpumask *nodemask = NULL; 1545 enum { cpuset, possible, fail } state = cpuset; 1546 int dest_cpu; 1547 1548 /* 1549 * If the node that the cpu is on has been offlined, cpu_to_node() 1550 * will return -1. There is no cpu on the node, and we should 1551 * select the cpu on the other node. 1552 */ 1553 if (nid != -1) { 1554 nodemask = cpumask_of_node(nid); 1555 1556 /* Look for allowed, online CPU in same node. */ 1557 for_each_cpu(dest_cpu, nodemask) { 1558 if (!cpu_online(dest_cpu)) 1559 continue; 1560 if (!cpu_active(dest_cpu)) 1561 continue; 1562 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1563 return dest_cpu; 1564 } 1565 } 1566 1567 for (;;) { 1568 /* Any allowed, online CPU? */ 1569 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1570 if (!cpu_online(dest_cpu)) 1571 continue; 1572 if (!cpu_active(dest_cpu)) 1573 continue; 1574 goto out; 1575 } 1576 1577 switch (state) { 1578 case cpuset: 1579 /* No more Mr. Nice Guy. */ 1580 cpuset_cpus_allowed_fallback(p); 1581 state = possible; 1582 break; 1583 1584 case possible: 1585 do_set_cpus_allowed(p, cpu_possible_mask); 1586 state = fail; 1587 break; 1588 1589 case fail: 1590 BUG(); 1591 break; 1592 } 1593 } 1594 1595 out: 1596 if (state != cpuset) { 1597 /* 1598 * Don't tell them about moving exiting tasks or 1599 * kernel threads (both mm NULL), since they never 1600 * leave kernel. 1601 */ 1602 if (p->mm && printk_ratelimit()) { 1603 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 1604 task_pid_nr(p), p->comm, cpu); 1605 } 1606 } 1607 1608 return dest_cpu; 1609 } 1610 1611 /* 1612 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1613 */ 1614 static inline 1615 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1616 { 1617 lockdep_assert_held(&p->pi_lock); 1618 1619 if (p->nr_cpus_allowed > 1) 1620 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1621 1622 /* 1623 * In order not to call set_task_cpu() on a blocking task we need 1624 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1625 * cpu. 1626 * 1627 * Since this is common to all placement strategies, this lives here. 1628 * 1629 * [ this allows ->select_task() to simply return task_cpu(p) and 1630 * not worry about this generic constraint ] 1631 */ 1632 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1633 !cpu_online(cpu))) 1634 cpu = select_fallback_rq(task_cpu(p), p); 1635 1636 return cpu; 1637 } 1638 1639 static void update_avg(u64 *avg, u64 sample) 1640 { 1641 s64 diff = sample - *avg; 1642 *avg += diff >> 3; 1643 } 1644 1645 #else 1646 1647 static inline int __set_cpus_allowed_ptr(struct task_struct *p, 1648 const struct cpumask *new_mask, bool check) 1649 { 1650 return set_cpus_allowed_ptr(p, new_mask); 1651 } 1652 1653 #endif /* CONFIG_SMP */ 1654 1655 static void 1656 ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1657 { 1658 #ifdef CONFIG_SCHEDSTATS 1659 struct rq *rq = this_rq(); 1660 1661 #ifdef CONFIG_SMP 1662 int this_cpu = smp_processor_id(); 1663 1664 if (cpu == this_cpu) { 1665 schedstat_inc(rq, ttwu_local); 1666 schedstat_inc(p, se.statistics.nr_wakeups_local); 1667 } else { 1668 struct sched_domain *sd; 1669 1670 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1671 rcu_read_lock(); 1672 for_each_domain(this_cpu, sd) { 1673 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1674 schedstat_inc(sd, ttwu_wake_remote); 1675 break; 1676 } 1677 } 1678 rcu_read_unlock(); 1679 } 1680 1681 if (wake_flags & WF_MIGRATED) 1682 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1683 1684 #endif /* CONFIG_SMP */ 1685 1686 schedstat_inc(rq, ttwu_count); 1687 schedstat_inc(p, se.statistics.nr_wakeups); 1688 1689 if (wake_flags & WF_SYNC) 1690 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1691 1692 #endif /* CONFIG_SCHEDSTATS */ 1693 } 1694 1695 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1696 { 1697 activate_task(rq, p, en_flags); 1698 p->on_rq = TASK_ON_RQ_QUEUED; 1699 1700 /* if a worker is waking up, notify workqueue */ 1701 if (p->flags & PF_WQ_WORKER) 1702 wq_worker_waking_up(p, cpu_of(rq)); 1703 } 1704 1705 /* 1706 * Mark the task runnable and perform wakeup-preemption. 1707 */ 1708 static void 1709 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1710 { 1711 check_preempt_curr(rq, p, wake_flags); 1712 p->state = TASK_RUNNING; 1713 trace_sched_wakeup(p); 1714 1715 #ifdef CONFIG_SMP 1716 if (p->sched_class->task_woken) { 1717 /* 1718 * Our task @p is fully woken up and running; so its safe to 1719 * drop the rq->lock, hereafter rq is only used for statistics. 1720 */ 1721 lockdep_unpin_lock(&rq->lock); 1722 p->sched_class->task_woken(rq, p); 1723 lockdep_pin_lock(&rq->lock); 1724 } 1725 1726 if (rq->idle_stamp) { 1727 u64 delta = rq_clock(rq) - rq->idle_stamp; 1728 u64 max = 2*rq->max_idle_balance_cost; 1729 1730 update_avg(&rq->avg_idle, delta); 1731 1732 if (rq->avg_idle > max) 1733 rq->avg_idle = max; 1734 1735 rq->idle_stamp = 0; 1736 } 1737 #endif 1738 } 1739 1740 static void 1741 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1742 { 1743 lockdep_assert_held(&rq->lock); 1744 1745 #ifdef CONFIG_SMP 1746 if (p->sched_contributes_to_load) 1747 rq->nr_uninterruptible--; 1748 #endif 1749 1750 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1751 ttwu_do_wakeup(rq, p, wake_flags); 1752 } 1753 1754 /* 1755 * Called in case the task @p isn't fully descheduled from its runqueue, 1756 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1757 * since all we need to do is flip p->state to TASK_RUNNING, since 1758 * the task is still ->on_rq. 1759 */ 1760 static int ttwu_remote(struct task_struct *p, int wake_flags) 1761 { 1762 struct rq *rq; 1763 int ret = 0; 1764 1765 rq = __task_rq_lock(p); 1766 if (task_on_rq_queued(p)) { 1767 /* check_preempt_curr() may use rq clock */ 1768 update_rq_clock(rq); 1769 ttwu_do_wakeup(rq, p, wake_flags); 1770 ret = 1; 1771 } 1772 __task_rq_unlock(rq); 1773 1774 return ret; 1775 } 1776 1777 #ifdef CONFIG_SMP 1778 void sched_ttwu_pending(void) 1779 { 1780 struct rq *rq = this_rq(); 1781 struct llist_node *llist = llist_del_all(&rq->wake_list); 1782 struct task_struct *p; 1783 unsigned long flags; 1784 1785 if (!llist) 1786 return; 1787 1788 raw_spin_lock_irqsave(&rq->lock, flags); 1789 lockdep_pin_lock(&rq->lock); 1790 1791 while (llist) { 1792 p = llist_entry(llist, struct task_struct, wake_entry); 1793 llist = llist_next(llist); 1794 ttwu_do_activate(rq, p, 0); 1795 } 1796 1797 lockdep_unpin_lock(&rq->lock); 1798 raw_spin_unlock_irqrestore(&rq->lock, flags); 1799 } 1800 1801 void scheduler_ipi(void) 1802 { 1803 /* 1804 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1805 * TIF_NEED_RESCHED remotely (for the first time) will also send 1806 * this IPI. 1807 */ 1808 preempt_fold_need_resched(); 1809 1810 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1811 return; 1812 1813 /* 1814 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1815 * traditionally all their work was done from the interrupt return 1816 * path. Now that we actually do some work, we need to make sure 1817 * we do call them. 1818 * 1819 * Some archs already do call them, luckily irq_enter/exit nest 1820 * properly. 1821 * 1822 * Arguably we should visit all archs and update all handlers, 1823 * however a fair share of IPIs are still resched only so this would 1824 * somewhat pessimize the simple resched case. 1825 */ 1826 irq_enter(); 1827 sched_ttwu_pending(); 1828 1829 /* 1830 * Check if someone kicked us for doing the nohz idle load balance. 1831 */ 1832 if (unlikely(got_nohz_idle_kick())) { 1833 this_rq()->idle_balance = 1; 1834 raise_softirq_irqoff(SCHED_SOFTIRQ); 1835 } 1836 irq_exit(); 1837 } 1838 1839 static void ttwu_queue_remote(struct task_struct *p, int cpu) 1840 { 1841 struct rq *rq = cpu_rq(cpu); 1842 1843 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { 1844 if (!set_nr_if_polling(rq->idle)) 1845 smp_send_reschedule(cpu); 1846 else 1847 trace_sched_wake_idle_without_ipi(cpu); 1848 } 1849 } 1850 1851 void wake_up_if_idle(int cpu) 1852 { 1853 struct rq *rq = cpu_rq(cpu); 1854 unsigned long flags; 1855 1856 rcu_read_lock(); 1857 1858 if (!is_idle_task(rcu_dereference(rq->curr))) 1859 goto out; 1860 1861 if (set_nr_if_polling(rq->idle)) { 1862 trace_sched_wake_idle_without_ipi(cpu); 1863 } else { 1864 raw_spin_lock_irqsave(&rq->lock, flags); 1865 if (is_idle_task(rq->curr)) 1866 smp_send_reschedule(cpu); 1867 /* Else cpu is not in idle, do nothing here */ 1868 raw_spin_unlock_irqrestore(&rq->lock, flags); 1869 } 1870 1871 out: 1872 rcu_read_unlock(); 1873 } 1874 1875 bool cpus_share_cache(int this_cpu, int that_cpu) 1876 { 1877 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1878 } 1879 #endif /* CONFIG_SMP */ 1880 1881 static void ttwu_queue(struct task_struct *p, int cpu) 1882 { 1883 struct rq *rq = cpu_rq(cpu); 1884 1885 #if defined(CONFIG_SMP) 1886 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1887 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1888 ttwu_queue_remote(p, cpu); 1889 return; 1890 } 1891 #endif 1892 1893 raw_spin_lock(&rq->lock); 1894 lockdep_pin_lock(&rq->lock); 1895 ttwu_do_activate(rq, p, 0); 1896 lockdep_unpin_lock(&rq->lock); 1897 raw_spin_unlock(&rq->lock); 1898 } 1899 1900 /** 1901 * try_to_wake_up - wake up a thread 1902 * @p: the thread to be awakened 1903 * @state: the mask of task states that can be woken 1904 * @wake_flags: wake modifier flags (WF_*) 1905 * 1906 * Put it on the run-queue if it's not already there. The "current" 1907 * thread is always on the run-queue (except when the actual 1908 * re-schedule is in progress), and as such you're allowed to do 1909 * the simpler "current->state = TASK_RUNNING" to mark yourself 1910 * runnable without the overhead of this. 1911 * 1912 * Return: %true if @p was woken up, %false if it was already running. 1913 * or @state didn't match @p's state. 1914 */ 1915 static int 1916 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1917 { 1918 unsigned long flags; 1919 int cpu, success = 0; 1920 1921 /* 1922 * If we are going to wake up a thread waiting for CONDITION we 1923 * need to ensure that CONDITION=1 done by the caller can not be 1924 * reordered with p->state check below. This pairs with mb() in 1925 * set_current_state() the waiting thread does. 1926 */ 1927 smp_mb__before_spinlock(); 1928 raw_spin_lock_irqsave(&p->pi_lock, flags); 1929 if (!(p->state & state)) 1930 goto out; 1931 1932 trace_sched_waking(p); 1933 1934 success = 1; /* we're going to change ->state */ 1935 cpu = task_cpu(p); 1936 1937 if (p->on_rq && ttwu_remote(p, wake_flags)) 1938 goto stat; 1939 1940 #ifdef CONFIG_SMP 1941 /* 1942 * If the owning (remote) cpu is still in the middle of schedule() with 1943 * this task as prev, wait until its done referencing the task. 1944 */ 1945 while (p->on_cpu) 1946 cpu_relax(); 1947 /* 1948 * Pairs with the smp_wmb() in finish_lock_switch(). 1949 */ 1950 smp_rmb(); 1951 1952 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1953 p->state = TASK_WAKING; 1954 1955 if (p->sched_class->task_waking) 1956 p->sched_class->task_waking(p); 1957 1958 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1959 if (task_cpu(p) != cpu) { 1960 wake_flags |= WF_MIGRATED; 1961 set_task_cpu(p, cpu); 1962 } 1963 #endif /* CONFIG_SMP */ 1964 1965 ttwu_queue(p, cpu); 1966 stat: 1967 ttwu_stat(p, cpu, wake_flags); 1968 out: 1969 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1970 1971 return success; 1972 } 1973 1974 /** 1975 * try_to_wake_up_local - try to wake up a local task with rq lock held 1976 * @p: the thread to be awakened 1977 * 1978 * Put @p on the run-queue if it's not already there. The caller must 1979 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1980 * the current task. 1981 */ 1982 static void try_to_wake_up_local(struct task_struct *p) 1983 { 1984 struct rq *rq = task_rq(p); 1985 1986 if (WARN_ON_ONCE(rq != this_rq()) || 1987 WARN_ON_ONCE(p == current)) 1988 return; 1989 1990 lockdep_assert_held(&rq->lock); 1991 1992 if (!raw_spin_trylock(&p->pi_lock)) { 1993 /* 1994 * This is OK, because current is on_cpu, which avoids it being 1995 * picked for load-balance and preemption/IRQs are still 1996 * disabled avoiding further scheduler activity on it and we've 1997 * not yet picked a replacement task. 1998 */ 1999 lockdep_unpin_lock(&rq->lock); 2000 raw_spin_unlock(&rq->lock); 2001 raw_spin_lock(&p->pi_lock); 2002 raw_spin_lock(&rq->lock); 2003 lockdep_pin_lock(&rq->lock); 2004 } 2005 2006 if (!(p->state & TASK_NORMAL)) 2007 goto out; 2008 2009 trace_sched_waking(p); 2010 2011 if (!task_on_rq_queued(p)) 2012 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 2013 2014 ttwu_do_wakeup(rq, p, 0); 2015 ttwu_stat(p, smp_processor_id(), 0); 2016 out: 2017 raw_spin_unlock(&p->pi_lock); 2018 } 2019 2020 /** 2021 * wake_up_process - Wake up a specific process 2022 * @p: The process to be woken up. 2023 * 2024 * Attempt to wake up the nominated process and move it to the set of runnable 2025 * processes. 2026 * 2027 * Return: 1 if the process was woken up, 0 if it was already running. 2028 * 2029 * It may be assumed that this function implies a write memory barrier before 2030 * changing the task state if and only if any tasks are woken up. 2031 */ 2032 int wake_up_process(struct task_struct *p) 2033 { 2034 WARN_ON(task_is_stopped_or_traced(p)); 2035 return try_to_wake_up(p, TASK_NORMAL, 0); 2036 } 2037 EXPORT_SYMBOL(wake_up_process); 2038 2039 int wake_up_state(struct task_struct *p, unsigned int state) 2040 { 2041 return try_to_wake_up(p, state, 0); 2042 } 2043 2044 /* 2045 * This function clears the sched_dl_entity static params. 2046 */ 2047 void __dl_clear_params(struct task_struct *p) 2048 { 2049 struct sched_dl_entity *dl_se = &p->dl; 2050 2051 dl_se->dl_runtime = 0; 2052 dl_se->dl_deadline = 0; 2053 dl_se->dl_period = 0; 2054 dl_se->flags = 0; 2055 dl_se->dl_bw = 0; 2056 2057 dl_se->dl_throttled = 0; 2058 dl_se->dl_new = 1; 2059 dl_se->dl_yielded = 0; 2060 } 2061 2062 /* 2063 * Perform scheduler related setup for a newly forked process p. 2064 * p is forked by current. 2065 * 2066 * __sched_fork() is basic setup used by init_idle() too: 2067 */ 2068 static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 2069 { 2070 p->on_rq = 0; 2071 2072 p->se.on_rq = 0; 2073 p->se.exec_start = 0; 2074 p->se.sum_exec_runtime = 0; 2075 p->se.prev_sum_exec_runtime = 0; 2076 p->se.nr_migrations = 0; 2077 p->se.vruntime = 0; 2078 INIT_LIST_HEAD(&p->se.group_node); 2079 2080 #ifdef CONFIG_SCHEDSTATS 2081 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 2082 #endif 2083 2084 RB_CLEAR_NODE(&p->dl.rb_node); 2085 init_dl_task_timer(&p->dl); 2086 __dl_clear_params(p); 2087 2088 INIT_LIST_HEAD(&p->rt.run_list); 2089 2090 #ifdef CONFIG_PREEMPT_NOTIFIERS 2091 INIT_HLIST_HEAD(&p->preempt_notifiers); 2092 #endif 2093 2094 #ifdef CONFIG_NUMA_BALANCING 2095 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 2096 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 2097 p->mm->numa_scan_seq = 0; 2098 } 2099 2100 if (clone_flags & CLONE_VM) 2101 p->numa_preferred_nid = current->numa_preferred_nid; 2102 else 2103 p->numa_preferred_nid = -1; 2104 2105 p->node_stamp = 0ULL; 2106 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 2107 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 2108 p->numa_work.next = &p->numa_work; 2109 p->numa_faults = NULL; 2110 p->last_task_numa_placement = 0; 2111 p->last_sum_exec_runtime = 0; 2112 2113 p->numa_group = NULL; 2114 #endif /* CONFIG_NUMA_BALANCING */ 2115 } 2116 2117 #ifdef CONFIG_NUMA_BALANCING 2118 #ifdef CONFIG_SCHED_DEBUG 2119 void set_numabalancing_state(bool enabled) 2120 { 2121 if (enabled) 2122 sched_feat_set("NUMA"); 2123 else 2124 sched_feat_set("NO_NUMA"); 2125 } 2126 #else 2127 __read_mostly bool numabalancing_enabled; 2128 2129 void set_numabalancing_state(bool enabled) 2130 { 2131 numabalancing_enabled = enabled; 2132 } 2133 #endif /* CONFIG_SCHED_DEBUG */ 2134 2135 #ifdef CONFIG_PROC_SYSCTL 2136 int sysctl_numa_balancing(struct ctl_table *table, int write, 2137 void __user *buffer, size_t *lenp, loff_t *ppos) 2138 { 2139 struct ctl_table t; 2140 int err; 2141 int state = numabalancing_enabled; 2142 2143 if (write && !capable(CAP_SYS_ADMIN)) 2144 return -EPERM; 2145 2146 t = *table; 2147 t.data = &state; 2148 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 2149 if (err < 0) 2150 return err; 2151 if (write) 2152 set_numabalancing_state(state); 2153 return err; 2154 } 2155 #endif 2156 #endif 2157 2158 /* 2159 * fork()/clone()-time setup: 2160 */ 2161 int sched_fork(unsigned long clone_flags, struct task_struct *p) 2162 { 2163 unsigned long flags; 2164 int cpu = get_cpu(); 2165 2166 __sched_fork(clone_flags, p); 2167 /* 2168 * We mark the process as running here. This guarantees that 2169 * nobody will actually run it, and a signal or other external 2170 * event cannot wake it up and insert it on the runqueue either. 2171 */ 2172 p->state = TASK_RUNNING; 2173 2174 /* 2175 * Make sure we do not leak PI boosting priority to the child. 2176 */ 2177 p->prio = current->normal_prio; 2178 2179 /* 2180 * Revert to default priority/policy on fork if requested. 2181 */ 2182 if (unlikely(p->sched_reset_on_fork)) { 2183 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 2184 p->policy = SCHED_NORMAL; 2185 p->static_prio = NICE_TO_PRIO(0); 2186 p->rt_priority = 0; 2187 } else if (PRIO_TO_NICE(p->static_prio) < 0) 2188 p->static_prio = NICE_TO_PRIO(0); 2189 2190 p->prio = p->normal_prio = __normal_prio(p); 2191 set_load_weight(p); 2192 2193 /* 2194 * We don't need the reset flag anymore after the fork. It has 2195 * fulfilled its duty: 2196 */ 2197 p->sched_reset_on_fork = 0; 2198 } 2199 2200 if (dl_prio(p->prio)) { 2201 put_cpu(); 2202 return -EAGAIN; 2203 } else if (rt_prio(p->prio)) { 2204 p->sched_class = &rt_sched_class; 2205 } else { 2206 p->sched_class = &fair_sched_class; 2207 } 2208 2209 if (p->sched_class->task_fork) 2210 p->sched_class->task_fork(p); 2211 2212 /* 2213 * The child is not yet in the pid-hash so no cgroup attach races, 2214 * and the cgroup is pinned to this child due to cgroup_fork() 2215 * is ran before sched_fork(). 2216 * 2217 * Silence PROVE_RCU. 2218 */ 2219 raw_spin_lock_irqsave(&p->pi_lock, flags); 2220 set_task_cpu(p, cpu); 2221 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2222 2223 #ifdef CONFIG_SCHED_INFO 2224 if (likely(sched_info_on())) 2225 memset(&p->sched_info, 0, sizeof(p->sched_info)); 2226 #endif 2227 #if defined(CONFIG_SMP) 2228 p->on_cpu = 0; 2229 #endif 2230 init_task_preempt_count(p); 2231 #ifdef CONFIG_SMP 2232 plist_node_init(&p->pushable_tasks, MAX_PRIO); 2233 RB_CLEAR_NODE(&p->pushable_dl_tasks); 2234 #endif 2235 2236 put_cpu(); 2237 return 0; 2238 } 2239 2240 unsigned long to_ratio(u64 period, u64 runtime) 2241 { 2242 if (runtime == RUNTIME_INF) 2243 return 1ULL << 20; 2244 2245 /* 2246 * Doing this here saves a lot of checks in all 2247 * the calling paths, and returning zero seems 2248 * safe for them anyway. 2249 */ 2250 if (period == 0) 2251 return 0; 2252 2253 return div64_u64(runtime << 20, period); 2254 } 2255 2256 #ifdef CONFIG_SMP 2257 inline struct dl_bw *dl_bw_of(int i) 2258 { 2259 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), 2260 "sched RCU must be held"); 2261 return &cpu_rq(i)->rd->dl_bw; 2262 } 2263 2264 static inline int dl_bw_cpus(int i) 2265 { 2266 struct root_domain *rd = cpu_rq(i)->rd; 2267 int cpus = 0; 2268 2269 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), 2270 "sched RCU must be held"); 2271 for_each_cpu_and(i, rd->span, cpu_active_mask) 2272 cpus++; 2273 2274 return cpus; 2275 } 2276 #else 2277 inline struct dl_bw *dl_bw_of(int i) 2278 { 2279 return &cpu_rq(i)->dl.dl_bw; 2280 } 2281 2282 static inline int dl_bw_cpus(int i) 2283 { 2284 return 1; 2285 } 2286 #endif 2287 2288 /* 2289 * We must be sure that accepting a new task (or allowing changing the 2290 * parameters of an existing one) is consistent with the bandwidth 2291 * constraints. If yes, this function also accordingly updates the currently 2292 * allocated bandwidth to reflect the new situation. 2293 * 2294 * This function is called while holding p's rq->lock. 2295 * 2296 * XXX we should delay bw change until the task's 0-lag point, see 2297 * __setparam_dl(). 2298 */ 2299 static int dl_overflow(struct task_struct *p, int policy, 2300 const struct sched_attr *attr) 2301 { 2302 2303 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 2304 u64 period = attr->sched_period ?: attr->sched_deadline; 2305 u64 runtime = attr->sched_runtime; 2306 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 2307 int cpus, err = -1; 2308 2309 if (new_bw == p->dl.dl_bw) 2310 return 0; 2311 2312 /* 2313 * Either if a task, enters, leave, or stays -deadline but changes 2314 * its parameters, we may need to update accordingly the total 2315 * allocated bandwidth of the container. 2316 */ 2317 raw_spin_lock(&dl_b->lock); 2318 cpus = dl_bw_cpus(task_cpu(p)); 2319 if (dl_policy(policy) && !task_has_dl_policy(p) && 2320 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2321 __dl_add(dl_b, new_bw); 2322 err = 0; 2323 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2324 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2325 __dl_clear(dl_b, p->dl.dl_bw); 2326 __dl_add(dl_b, new_bw); 2327 err = 0; 2328 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2329 __dl_clear(dl_b, p->dl.dl_bw); 2330 err = 0; 2331 } 2332 raw_spin_unlock(&dl_b->lock); 2333 2334 return err; 2335 } 2336 2337 extern void init_dl_bw(struct dl_bw *dl_b); 2338 2339 /* 2340 * wake_up_new_task - wake up a newly created task for the first time. 2341 * 2342 * This function will do some initial scheduler statistics housekeeping 2343 * that must be done for every newly created context, then puts the task 2344 * on the runqueue and wakes it. 2345 */ 2346 void wake_up_new_task(struct task_struct *p) 2347 { 2348 unsigned long flags; 2349 struct rq *rq; 2350 2351 raw_spin_lock_irqsave(&p->pi_lock, flags); 2352 #ifdef CONFIG_SMP 2353 /* 2354 * Fork balancing, do it here and not earlier because: 2355 * - cpus_allowed can change in the fork path 2356 * - any previously selected cpu might disappear through hotplug 2357 */ 2358 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2359 #endif 2360 2361 /* Initialize new task's runnable average */ 2362 init_entity_runnable_average(&p->se); 2363 rq = __task_rq_lock(p); 2364 activate_task(rq, p, 0); 2365 p->on_rq = TASK_ON_RQ_QUEUED; 2366 trace_sched_wakeup_new(p); 2367 check_preempt_curr(rq, p, WF_FORK); 2368 #ifdef CONFIG_SMP 2369 if (p->sched_class->task_woken) 2370 p->sched_class->task_woken(rq, p); 2371 #endif 2372 task_rq_unlock(rq, p, &flags); 2373 } 2374 2375 #ifdef CONFIG_PREEMPT_NOTIFIERS 2376 2377 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE; 2378 2379 void preempt_notifier_inc(void) 2380 { 2381 static_key_slow_inc(&preempt_notifier_key); 2382 } 2383 EXPORT_SYMBOL_GPL(preempt_notifier_inc); 2384 2385 void preempt_notifier_dec(void) 2386 { 2387 static_key_slow_dec(&preempt_notifier_key); 2388 } 2389 EXPORT_SYMBOL_GPL(preempt_notifier_dec); 2390 2391 /** 2392 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2393 * @notifier: notifier struct to register 2394 */ 2395 void preempt_notifier_register(struct preempt_notifier *notifier) 2396 { 2397 if (!static_key_false(&preempt_notifier_key)) 2398 WARN(1, "registering preempt_notifier while notifiers disabled\n"); 2399 2400 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2401 } 2402 EXPORT_SYMBOL_GPL(preempt_notifier_register); 2403 2404 /** 2405 * preempt_notifier_unregister - no longer interested in preemption notifications 2406 * @notifier: notifier struct to unregister 2407 * 2408 * This is *not* safe to call from within a preemption notifier. 2409 */ 2410 void preempt_notifier_unregister(struct preempt_notifier *notifier) 2411 { 2412 hlist_del(¬ifier->link); 2413 } 2414 EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2415 2416 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr) 2417 { 2418 struct preempt_notifier *notifier; 2419 2420 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2421 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2422 } 2423 2424 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2425 { 2426 if (static_key_false(&preempt_notifier_key)) 2427 __fire_sched_in_preempt_notifiers(curr); 2428 } 2429 2430 static void 2431 __fire_sched_out_preempt_notifiers(struct task_struct *curr, 2432 struct task_struct *next) 2433 { 2434 struct preempt_notifier *notifier; 2435 2436 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2437 notifier->ops->sched_out(notifier, next); 2438 } 2439 2440 static __always_inline void 2441 fire_sched_out_preempt_notifiers(struct task_struct *curr, 2442 struct task_struct *next) 2443 { 2444 if (static_key_false(&preempt_notifier_key)) 2445 __fire_sched_out_preempt_notifiers(curr, next); 2446 } 2447 2448 #else /* !CONFIG_PREEMPT_NOTIFIERS */ 2449 2450 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2451 { 2452 } 2453 2454 static inline void 2455 fire_sched_out_preempt_notifiers(struct task_struct *curr, 2456 struct task_struct *next) 2457 { 2458 } 2459 2460 #endif /* CONFIG_PREEMPT_NOTIFIERS */ 2461 2462 /** 2463 * prepare_task_switch - prepare to switch tasks 2464 * @rq: the runqueue preparing to switch 2465 * @prev: the current task that is being switched out 2466 * @next: the task we are going to switch to. 2467 * 2468 * This is called with the rq lock held and interrupts off. It must 2469 * be paired with a subsequent finish_task_switch after the context 2470 * switch. 2471 * 2472 * prepare_task_switch sets up locking and calls architecture specific 2473 * hooks. 2474 */ 2475 static inline void 2476 prepare_task_switch(struct rq *rq, struct task_struct *prev, 2477 struct task_struct *next) 2478 { 2479 trace_sched_switch(prev, next); 2480 sched_info_switch(rq, prev, next); 2481 perf_event_task_sched_out(prev, next); 2482 fire_sched_out_preempt_notifiers(prev, next); 2483 prepare_lock_switch(rq, next); 2484 prepare_arch_switch(next); 2485 } 2486 2487 /** 2488 * finish_task_switch - clean up after a task-switch 2489 * @prev: the thread we just switched away from. 2490 * 2491 * finish_task_switch must be called after the context switch, paired 2492 * with a prepare_task_switch call before the context switch. 2493 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2494 * and do any other architecture-specific cleanup actions. 2495 * 2496 * Note that we may have delayed dropping an mm in context_switch(). If 2497 * so, we finish that here outside of the runqueue lock. (Doing it 2498 * with the lock held can cause deadlocks; see schedule() for 2499 * details.) 2500 * 2501 * The context switch have flipped the stack from under us and restored the 2502 * local variables which were saved when this task called schedule() in the 2503 * past. prev == current is still correct but we need to recalculate this_rq 2504 * because prev may have moved to another CPU. 2505 */ 2506 static struct rq *finish_task_switch(struct task_struct *prev) 2507 __releases(rq->lock) 2508 { 2509 struct rq *rq = this_rq(); 2510 struct mm_struct *mm = rq->prev_mm; 2511 long prev_state; 2512 2513 rq->prev_mm = NULL; 2514 2515 /* 2516 * A task struct has one reference for the use as "current". 2517 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2518 * schedule one last time. The schedule call will never return, and 2519 * the scheduled task must drop that reference. 2520 * The test for TASK_DEAD must occur while the runqueue locks are 2521 * still held, otherwise prev could be scheduled on another cpu, die 2522 * there before we look at prev->state, and then the reference would 2523 * be dropped twice. 2524 * Manfred Spraul <manfred@colorfullife.com> 2525 */ 2526 prev_state = prev->state; 2527 vtime_task_switch(prev); 2528 perf_event_task_sched_in(prev, current); 2529 finish_lock_switch(rq, prev); 2530 finish_arch_post_lock_switch(); 2531 2532 fire_sched_in_preempt_notifiers(current); 2533 if (mm) 2534 mmdrop(mm); 2535 if (unlikely(prev_state == TASK_DEAD)) { 2536 if (prev->sched_class->task_dead) 2537 prev->sched_class->task_dead(prev); 2538 2539 /* 2540 * Remove function-return probe instances associated with this 2541 * task and put them back on the free list. 2542 */ 2543 kprobe_flush_task(prev); 2544 put_task_struct(prev); 2545 } 2546 2547 tick_nohz_task_switch(); 2548 return rq; 2549 } 2550 2551 #ifdef CONFIG_SMP 2552 2553 /* rq->lock is NOT held, but preemption is disabled */ 2554 static void __balance_callback(struct rq *rq) 2555 { 2556 struct callback_head *head, *next; 2557 void (*func)(struct rq *rq); 2558 unsigned long flags; 2559 2560 raw_spin_lock_irqsave(&rq->lock, flags); 2561 head = rq->balance_callback; 2562 rq->balance_callback = NULL; 2563 while (head) { 2564 func = (void (*)(struct rq *))head->func; 2565 next = head->next; 2566 head->next = NULL; 2567 head = next; 2568 2569 func(rq); 2570 } 2571 raw_spin_unlock_irqrestore(&rq->lock, flags); 2572 } 2573 2574 static inline void balance_callback(struct rq *rq) 2575 { 2576 if (unlikely(rq->balance_callback)) 2577 __balance_callback(rq); 2578 } 2579 2580 #else 2581 2582 static inline void balance_callback(struct rq *rq) 2583 { 2584 } 2585 2586 #endif 2587 2588 /** 2589 * schedule_tail - first thing a freshly forked thread must call. 2590 * @prev: the thread we just switched away from. 2591 */ 2592 asmlinkage __visible void schedule_tail(struct task_struct *prev) 2593 __releases(rq->lock) 2594 { 2595 struct rq *rq; 2596 2597 /* finish_task_switch() drops rq->lock and enables preemtion */ 2598 preempt_disable(); 2599 rq = finish_task_switch(prev); 2600 balance_callback(rq); 2601 preempt_enable(); 2602 2603 if (current->set_child_tid) 2604 put_user(task_pid_vnr(current), current->set_child_tid); 2605 } 2606 2607 /* 2608 * context_switch - switch to the new MM and the new thread's register state. 2609 */ 2610 static inline struct rq * 2611 context_switch(struct rq *rq, struct task_struct *prev, 2612 struct task_struct *next) 2613 { 2614 struct mm_struct *mm, *oldmm; 2615 2616 prepare_task_switch(rq, prev, next); 2617 2618 mm = next->mm; 2619 oldmm = prev->active_mm; 2620 /* 2621 * For paravirt, this is coupled with an exit in switch_to to 2622 * combine the page table reload and the switch backend into 2623 * one hypercall. 2624 */ 2625 arch_start_context_switch(prev); 2626 2627 if (!mm) { 2628 next->active_mm = oldmm; 2629 atomic_inc(&oldmm->mm_count); 2630 enter_lazy_tlb(oldmm, next); 2631 } else 2632 switch_mm(oldmm, mm, next); 2633 2634 if (!prev->mm) { 2635 prev->active_mm = NULL; 2636 rq->prev_mm = oldmm; 2637 } 2638 /* 2639 * Since the runqueue lock will be released by the next 2640 * task (which is an invalid locking op but in the case 2641 * of the scheduler it's an obvious special-case), so we 2642 * do an early lockdep release here: 2643 */ 2644 lockdep_unpin_lock(&rq->lock); 2645 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2646 2647 /* Here we just switch the register state and the stack. */ 2648 switch_to(prev, next, prev); 2649 barrier(); 2650 2651 return finish_task_switch(prev); 2652 } 2653 2654 /* 2655 * nr_running and nr_context_switches: 2656 * 2657 * externally visible scheduler statistics: current number of runnable 2658 * threads, total number of context switches performed since bootup. 2659 */ 2660 unsigned long nr_running(void) 2661 { 2662 unsigned long i, sum = 0; 2663 2664 for_each_online_cpu(i) 2665 sum += cpu_rq(i)->nr_running; 2666 2667 return sum; 2668 } 2669 2670 /* 2671 * Check if only the current task is running on the cpu. 2672 * 2673 * Caution: this function does not check that the caller has disabled 2674 * preemption, thus the result might have a time-of-check-to-time-of-use 2675 * race. The caller is responsible to use it correctly, for example: 2676 * 2677 * - from a non-preemptable section (of course) 2678 * 2679 * - from a thread that is bound to a single CPU 2680 * 2681 * - in a loop with very short iterations (e.g. a polling loop) 2682 */ 2683 bool single_task_running(void) 2684 { 2685 return raw_rq()->nr_running == 1; 2686 } 2687 EXPORT_SYMBOL(single_task_running); 2688 2689 unsigned long long nr_context_switches(void) 2690 { 2691 int i; 2692 unsigned long long sum = 0; 2693 2694 for_each_possible_cpu(i) 2695 sum += cpu_rq(i)->nr_switches; 2696 2697 return sum; 2698 } 2699 2700 unsigned long nr_iowait(void) 2701 { 2702 unsigned long i, sum = 0; 2703 2704 for_each_possible_cpu(i) 2705 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2706 2707 return sum; 2708 } 2709 2710 unsigned long nr_iowait_cpu(int cpu) 2711 { 2712 struct rq *this = cpu_rq(cpu); 2713 return atomic_read(&this->nr_iowait); 2714 } 2715 2716 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load) 2717 { 2718 struct rq *rq = this_rq(); 2719 *nr_waiters = atomic_read(&rq->nr_iowait); 2720 *load = rq->load.weight; 2721 } 2722 2723 #ifdef CONFIG_SMP 2724 2725 /* 2726 * sched_exec - execve() is a valuable balancing opportunity, because at 2727 * this point the task has the smallest effective memory and cache footprint. 2728 */ 2729 void sched_exec(void) 2730 { 2731 struct task_struct *p = current; 2732 unsigned long flags; 2733 int dest_cpu; 2734 2735 raw_spin_lock_irqsave(&p->pi_lock, flags); 2736 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2737 if (dest_cpu == smp_processor_id()) 2738 goto unlock; 2739 2740 if (likely(cpu_active(dest_cpu))) { 2741 struct migration_arg arg = { p, dest_cpu }; 2742 2743 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2744 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2745 return; 2746 } 2747 unlock: 2748 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2749 } 2750 2751 #endif 2752 2753 DEFINE_PER_CPU(struct kernel_stat, kstat); 2754 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2755 2756 EXPORT_PER_CPU_SYMBOL(kstat); 2757 EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2758 2759 /* 2760 * Return accounted runtime for the task. 2761 * In case the task is currently running, return the runtime plus current's 2762 * pending runtime that have not been accounted yet. 2763 */ 2764 unsigned long long task_sched_runtime(struct task_struct *p) 2765 { 2766 unsigned long flags; 2767 struct rq *rq; 2768 u64 ns; 2769 2770 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2771 /* 2772 * 64-bit doesn't need locks to atomically read a 64bit value. 2773 * So we have a optimization chance when the task's delta_exec is 0. 2774 * Reading ->on_cpu is racy, but this is ok. 2775 * 2776 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2777 * If we race with it entering cpu, unaccounted time is 0. This is 2778 * indistinguishable from the read occurring a few cycles earlier. 2779 * If we see ->on_cpu without ->on_rq, the task is leaving, and has 2780 * been accounted, so we're correct here as well. 2781 */ 2782 if (!p->on_cpu || !task_on_rq_queued(p)) 2783 return p->se.sum_exec_runtime; 2784 #endif 2785 2786 rq = task_rq_lock(p, &flags); 2787 /* 2788 * Must be ->curr _and_ ->on_rq. If dequeued, we would 2789 * project cycles that may never be accounted to this 2790 * thread, breaking clock_gettime(). 2791 */ 2792 if (task_current(rq, p) && task_on_rq_queued(p)) { 2793 update_rq_clock(rq); 2794 p->sched_class->update_curr(rq); 2795 } 2796 ns = p->se.sum_exec_runtime; 2797 task_rq_unlock(rq, p, &flags); 2798 2799 return ns; 2800 } 2801 2802 /* 2803 * This function gets called by the timer code, with HZ frequency. 2804 * We call it with interrupts disabled. 2805 */ 2806 void scheduler_tick(void) 2807 { 2808 int cpu = smp_processor_id(); 2809 struct rq *rq = cpu_rq(cpu); 2810 struct task_struct *curr = rq->curr; 2811 2812 sched_clock_tick(); 2813 2814 raw_spin_lock(&rq->lock); 2815 update_rq_clock(rq); 2816 curr->sched_class->task_tick(rq, curr, 0); 2817 update_cpu_load_active(rq); 2818 calc_global_load_tick(rq); 2819 raw_spin_unlock(&rq->lock); 2820 2821 perf_event_task_tick(); 2822 2823 #ifdef CONFIG_SMP 2824 rq->idle_balance = idle_cpu(cpu); 2825 trigger_load_balance(rq); 2826 #endif 2827 rq_last_tick_reset(rq); 2828 } 2829 2830 #ifdef CONFIG_NO_HZ_FULL 2831 /** 2832 * scheduler_tick_max_deferment 2833 * 2834 * Keep at least one tick per second when a single 2835 * active task is running because the scheduler doesn't 2836 * yet completely support full dynticks environment. 2837 * 2838 * This makes sure that uptime, CFS vruntime, load 2839 * balancing, etc... continue to move forward, even 2840 * with a very low granularity. 2841 * 2842 * Return: Maximum deferment in nanoseconds. 2843 */ 2844 u64 scheduler_tick_max_deferment(void) 2845 { 2846 struct rq *rq = this_rq(); 2847 unsigned long next, now = READ_ONCE(jiffies); 2848 2849 next = rq->last_sched_tick + HZ; 2850 2851 if (time_before_eq(next, now)) 2852 return 0; 2853 2854 return jiffies_to_nsecs(next - now); 2855 } 2856 #endif 2857 2858 notrace unsigned long get_parent_ip(unsigned long addr) 2859 { 2860 if (in_lock_functions(addr)) { 2861 addr = CALLER_ADDR2; 2862 if (in_lock_functions(addr)) 2863 addr = CALLER_ADDR3; 2864 } 2865 return addr; 2866 } 2867 2868 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2869 defined(CONFIG_PREEMPT_TRACER)) 2870 2871 void preempt_count_add(int val) 2872 { 2873 #ifdef CONFIG_DEBUG_PREEMPT 2874 /* 2875 * Underflow? 2876 */ 2877 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2878 return; 2879 #endif 2880 __preempt_count_add(val); 2881 #ifdef CONFIG_DEBUG_PREEMPT 2882 /* 2883 * Spinlock count overflowing soon? 2884 */ 2885 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2886 PREEMPT_MASK - 10); 2887 #endif 2888 if (preempt_count() == val) { 2889 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2890 #ifdef CONFIG_DEBUG_PREEMPT 2891 current->preempt_disable_ip = ip; 2892 #endif 2893 trace_preempt_off(CALLER_ADDR0, ip); 2894 } 2895 } 2896 EXPORT_SYMBOL(preempt_count_add); 2897 NOKPROBE_SYMBOL(preempt_count_add); 2898 2899 void preempt_count_sub(int val) 2900 { 2901 #ifdef CONFIG_DEBUG_PREEMPT 2902 /* 2903 * Underflow? 2904 */ 2905 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2906 return; 2907 /* 2908 * Is the spinlock portion underflowing? 2909 */ 2910 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2911 !(preempt_count() & PREEMPT_MASK))) 2912 return; 2913 #endif 2914 2915 if (preempt_count() == val) 2916 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2917 __preempt_count_sub(val); 2918 } 2919 EXPORT_SYMBOL(preempt_count_sub); 2920 NOKPROBE_SYMBOL(preempt_count_sub); 2921 2922 #endif 2923 2924 /* 2925 * Print scheduling while atomic bug: 2926 */ 2927 static noinline void __schedule_bug(struct task_struct *prev) 2928 { 2929 if (oops_in_progress) 2930 return; 2931 2932 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2933 prev->comm, prev->pid, preempt_count()); 2934 2935 debug_show_held_locks(prev); 2936 print_modules(); 2937 if (irqs_disabled()) 2938 print_irqtrace_events(prev); 2939 #ifdef CONFIG_DEBUG_PREEMPT 2940 if (in_atomic_preempt_off()) { 2941 pr_err("Preemption disabled at:"); 2942 print_ip_sym(current->preempt_disable_ip); 2943 pr_cont("\n"); 2944 } 2945 #endif 2946 dump_stack(); 2947 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2948 } 2949 2950 /* 2951 * Various schedule()-time debugging checks and statistics: 2952 */ 2953 static inline void schedule_debug(struct task_struct *prev) 2954 { 2955 #ifdef CONFIG_SCHED_STACK_END_CHECK 2956 BUG_ON(unlikely(task_stack_end_corrupted(prev))); 2957 #endif 2958 /* 2959 * Test if we are atomic. Since do_exit() needs to call into 2960 * schedule() atomically, we ignore that path. Otherwise whine 2961 * if we are scheduling when we should not. 2962 */ 2963 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2964 __schedule_bug(prev); 2965 rcu_sleep_check(); 2966 2967 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2968 2969 schedstat_inc(this_rq(), sched_count); 2970 } 2971 2972 /* 2973 * Pick up the highest-prio task: 2974 */ 2975 static inline struct task_struct * 2976 pick_next_task(struct rq *rq, struct task_struct *prev) 2977 { 2978 const struct sched_class *class = &fair_sched_class; 2979 struct task_struct *p; 2980 2981 /* 2982 * Optimization: we know that if all tasks are in 2983 * the fair class we can call that function directly: 2984 */ 2985 if (likely(prev->sched_class == class && 2986 rq->nr_running == rq->cfs.h_nr_running)) { 2987 p = fair_sched_class.pick_next_task(rq, prev); 2988 if (unlikely(p == RETRY_TASK)) 2989 goto again; 2990 2991 /* assumes fair_sched_class->next == idle_sched_class */ 2992 if (unlikely(!p)) 2993 p = idle_sched_class.pick_next_task(rq, prev); 2994 2995 return p; 2996 } 2997 2998 again: 2999 for_each_class(class) { 3000 p = class->pick_next_task(rq, prev); 3001 if (p) { 3002 if (unlikely(p == RETRY_TASK)) 3003 goto again; 3004 return p; 3005 } 3006 } 3007 3008 BUG(); /* the idle class will always have a runnable task */ 3009 } 3010 3011 /* 3012 * __schedule() is the main scheduler function. 3013 * 3014 * The main means of driving the scheduler and thus entering this function are: 3015 * 3016 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 3017 * 3018 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 3019 * paths. For example, see arch/x86/entry_64.S. 3020 * 3021 * To drive preemption between tasks, the scheduler sets the flag in timer 3022 * interrupt handler scheduler_tick(). 3023 * 3024 * 3. Wakeups don't really cause entry into schedule(). They add a 3025 * task to the run-queue and that's it. 3026 * 3027 * Now, if the new task added to the run-queue preempts the current 3028 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 3029 * called on the nearest possible occasion: 3030 * 3031 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 3032 * 3033 * - in syscall or exception context, at the next outmost 3034 * preempt_enable(). (this might be as soon as the wake_up()'s 3035 * spin_unlock()!) 3036 * 3037 * - in IRQ context, return from interrupt-handler to 3038 * preemptible context 3039 * 3040 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 3041 * then at the next: 3042 * 3043 * - cond_resched() call 3044 * - explicit schedule() call 3045 * - return from syscall or exception to user-space 3046 * - return from interrupt-handler to user-space 3047 * 3048 * WARNING: must be called with preemption disabled! 3049 */ 3050 static void __sched __schedule(void) 3051 { 3052 struct task_struct *prev, *next; 3053 unsigned long *switch_count; 3054 struct rq *rq; 3055 int cpu; 3056 3057 cpu = smp_processor_id(); 3058 rq = cpu_rq(cpu); 3059 rcu_note_context_switch(); 3060 prev = rq->curr; 3061 3062 schedule_debug(prev); 3063 3064 if (sched_feat(HRTICK)) 3065 hrtick_clear(rq); 3066 3067 /* 3068 * Make sure that signal_pending_state()->signal_pending() below 3069 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 3070 * done by the caller to avoid the race with signal_wake_up(). 3071 */ 3072 smp_mb__before_spinlock(); 3073 raw_spin_lock_irq(&rq->lock); 3074 lockdep_pin_lock(&rq->lock); 3075 3076 rq->clock_skip_update <<= 1; /* promote REQ to ACT */ 3077 3078 switch_count = &prev->nivcsw; 3079 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 3080 if (unlikely(signal_pending_state(prev->state, prev))) { 3081 prev->state = TASK_RUNNING; 3082 } else { 3083 deactivate_task(rq, prev, DEQUEUE_SLEEP); 3084 prev->on_rq = 0; 3085 3086 /* 3087 * If a worker went to sleep, notify and ask workqueue 3088 * whether it wants to wake up a task to maintain 3089 * concurrency. 3090 */ 3091 if (prev->flags & PF_WQ_WORKER) { 3092 struct task_struct *to_wakeup; 3093 3094 to_wakeup = wq_worker_sleeping(prev, cpu); 3095 if (to_wakeup) 3096 try_to_wake_up_local(to_wakeup); 3097 } 3098 } 3099 switch_count = &prev->nvcsw; 3100 } 3101 3102 if (task_on_rq_queued(prev)) 3103 update_rq_clock(rq); 3104 3105 next = pick_next_task(rq, prev); 3106 clear_tsk_need_resched(prev); 3107 clear_preempt_need_resched(); 3108 rq->clock_skip_update = 0; 3109 3110 if (likely(prev != next)) { 3111 rq->nr_switches++; 3112 rq->curr = next; 3113 ++*switch_count; 3114 3115 rq = context_switch(rq, prev, next); /* unlocks the rq */ 3116 cpu = cpu_of(rq); 3117 } else { 3118 lockdep_unpin_lock(&rq->lock); 3119 raw_spin_unlock_irq(&rq->lock); 3120 } 3121 3122 balance_callback(rq); 3123 } 3124 3125 static inline void sched_submit_work(struct task_struct *tsk) 3126 { 3127 if (!tsk->state || tsk_is_pi_blocked(tsk)) 3128 return; 3129 /* 3130 * If we are going to sleep and we have plugged IO queued, 3131 * make sure to submit it to avoid deadlocks. 3132 */ 3133 if (blk_needs_flush_plug(tsk)) 3134 blk_schedule_flush_plug(tsk); 3135 } 3136 3137 asmlinkage __visible void __sched schedule(void) 3138 { 3139 struct task_struct *tsk = current; 3140 3141 sched_submit_work(tsk); 3142 do { 3143 preempt_disable(); 3144 __schedule(); 3145 sched_preempt_enable_no_resched(); 3146 } while (need_resched()); 3147 } 3148 EXPORT_SYMBOL(schedule); 3149 3150 #ifdef CONFIG_CONTEXT_TRACKING 3151 asmlinkage __visible void __sched schedule_user(void) 3152 { 3153 /* 3154 * If we come here after a random call to set_need_resched(), 3155 * or we have been woken up remotely but the IPI has not yet arrived, 3156 * we haven't yet exited the RCU idle mode. Do it here manually until 3157 * we find a better solution. 3158 * 3159 * NB: There are buggy callers of this function. Ideally we 3160 * should warn if prev_state != CONTEXT_USER, but that will trigger 3161 * too frequently to make sense yet. 3162 */ 3163 enum ctx_state prev_state = exception_enter(); 3164 schedule(); 3165 exception_exit(prev_state); 3166 } 3167 #endif 3168 3169 /** 3170 * schedule_preempt_disabled - called with preemption disabled 3171 * 3172 * Returns with preemption disabled. Note: preempt_count must be 1 3173 */ 3174 void __sched schedule_preempt_disabled(void) 3175 { 3176 sched_preempt_enable_no_resched(); 3177 schedule(); 3178 preempt_disable(); 3179 } 3180 3181 static void __sched notrace preempt_schedule_common(void) 3182 { 3183 do { 3184 preempt_active_enter(); 3185 __schedule(); 3186 preempt_active_exit(); 3187 3188 /* 3189 * Check again in case we missed a preemption opportunity 3190 * between schedule and now. 3191 */ 3192 } while (need_resched()); 3193 } 3194 3195 #ifdef CONFIG_PREEMPT 3196 /* 3197 * this is the entry point to schedule() from in-kernel preemption 3198 * off of preempt_enable. Kernel preemptions off return from interrupt 3199 * occur there and call schedule directly. 3200 */ 3201 asmlinkage __visible void __sched notrace preempt_schedule(void) 3202 { 3203 /* 3204 * If there is a non-zero preempt_count or interrupts are disabled, 3205 * we do not want to preempt the current task. Just return.. 3206 */ 3207 if (likely(!preemptible())) 3208 return; 3209 3210 preempt_schedule_common(); 3211 } 3212 NOKPROBE_SYMBOL(preempt_schedule); 3213 EXPORT_SYMBOL(preempt_schedule); 3214 3215 /** 3216 * preempt_schedule_notrace - preempt_schedule called by tracing 3217 * 3218 * The tracing infrastructure uses preempt_enable_notrace to prevent 3219 * recursion and tracing preempt enabling caused by the tracing 3220 * infrastructure itself. But as tracing can happen in areas coming 3221 * from userspace or just about to enter userspace, a preempt enable 3222 * can occur before user_exit() is called. This will cause the scheduler 3223 * to be called when the system is still in usermode. 3224 * 3225 * To prevent this, the preempt_enable_notrace will use this function 3226 * instead of preempt_schedule() to exit user context if needed before 3227 * calling the scheduler. 3228 */ 3229 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void) 3230 { 3231 enum ctx_state prev_ctx; 3232 3233 if (likely(!preemptible())) 3234 return; 3235 3236 do { 3237 /* 3238 * Use raw __prempt_count() ops that don't call function. 3239 * We can't call functions before disabling preemption which 3240 * disarm preemption tracing recursions. 3241 */ 3242 __preempt_count_add(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET); 3243 barrier(); 3244 /* 3245 * Needs preempt disabled in case user_exit() is traced 3246 * and the tracer calls preempt_enable_notrace() causing 3247 * an infinite recursion. 3248 */ 3249 prev_ctx = exception_enter(); 3250 __schedule(); 3251 exception_exit(prev_ctx); 3252 3253 barrier(); 3254 __preempt_count_sub(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET); 3255 } while (need_resched()); 3256 } 3257 EXPORT_SYMBOL_GPL(preempt_schedule_notrace); 3258 3259 #endif /* CONFIG_PREEMPT */ 3260 3261 /* 3262 * this is the entry point to schedule() from kernel preemption 3263 * off of irq context. 3264 * Note, that this is called and return with irqs disabled. This will 3265 * protect us against recursive calling from irq. 3266 */ 3267 asmlinkage __visible void __sched preempt_schedule_irq(void) 3268 { 3269 enum ctx_state prev_state; 3270 3271 /* Catch callers which need to be fixed */ 3272 BUG_ON(preempt_count() || !irqs_disabled()); 3273 3274 prev_state = exception_enter(); 3275 3276 do { 3277 preempt_active_enter(); 3278 local_irq_enable(); 3279 __schedule(); 3280 local_irq_disable(); 3281 preempt_active_exit(); 3282 } while (need_resched()); 3283 3284 exception_exit(prev_state); 3285 } 3286 3287 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3288 void *key) 3289 { 3290 return try_to_wake_up(curr->private, mode, wake_flags); 3291 } 3292 EXPORT_SYMBOL(default_wake_function); 3293 3294 #ifdef CONFIG_RT_MUTEXES 3295 3296 /* 3297 * rt_mutex_setprio - set the current priority of a task 3298 * @p: task 3299 * @prio: prio value (kernel-internal form) 3300 * 3301 * This function changes the 'effective' priority of a task. It does 3302 * not touch ->normal_prio like __setscheduler(). 3303 * 3304 * Used by the rt_mutex code to implement priority inheritance 3305 * logic. Call site only calls if the priority of the task changed. 3306 */ 3307 void rt_mutex_setprio(struct task_struct *p, int prio) 3308 { 3309 int oldprio, queued, running, enqueue_flag = 0; 3310 struct rq *rq; 3311 const struct sched_class *prev_class; 3312 3313 BUG_ON(prio > MAX_PRIO); 3314 3315 rq = __task_rq_lock(p); 3316 3317 /* 3318 * Idle task boosting is a nono in general. There is one 3319 * exception, when PREEMPT_RT and NOHZ is active: 3320 * 3321 * The idle task calls get_next_timer_interrupt() and holds 3322 * the timer wheel base->lock on the CPU and another CPU wants 3323 * to access the timer (probably to cancel it). We can safely 3324 * ignore the boosting request, as the idle CPU runs this code 3325 * with interrupts disabled and will complete the lock 3326 * protected section without being interrupted. So there is no 3327 * real need to boost. 3328 */ 3329 if (unlikely(p == rq->idle)) { 3330 WARN_ON(p != rq->curr); 3331 WARN_ON(p->pi_blocked_on); 3332 goto out_unlock; 3333 } 3334 3335 trace_sched_pi_setprio(p, prio); 3336 oldprio = p->prio; 3337 prev_class = p->sched_class; 3338 queued = task_on_rq_queued(p); 3339 running = task_current(rq, p); 3340 if (queued) 3341 dequeue_task(rq, p, 0); 3342 if (running) 3343 put_prev_task(rq, p); 3344 3345 /* 3346 * Boosting condition are: 3347 * 1. -rt task is running and holds mutex A 3348 * --> -dl task blocks on mutex A 3349 * 3350 * 2. -dl task is running and holds mutex A 3351 * --> -dl task blocks on mutex A and could preempt the 3352 * running task 3353 */ 3354 if (dl_prio(prio)) { 3355 struct task_struct *pi_task = rt_mutex_get_top_task(p); 3356 if (!dl_prio(p->normal_prio) || 3357 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) { 3358 p->dl.dl_boosted = 1; 3359 enqueue_flag = ENQUEUE_REPLENISH; 3360 } else 3361 p->dl.dl_boosted = 0; 3362 p->sched_class = &dl_sched_class; 3363 } else if (rt_prio(prio)) { 3364 if (dl_prio(oldprio)) 3365 p->dl.dl_boosted = 0; 3366 if (oldprio < prio) 3367 enqueue_flag = ENQUEUE_HEAD; 3368 p->sched_class = &rt_sched_class; 3369 } else { 3370 if (dl_prio(oldprio)) 3371 p->dl.dl_boosted = 0; 3372 if (rt_prio(oldprio)) 3373 p->rt.timeout = 0; 3374 p->sched_class = &fair_sched_class; 3375 } 3376 3377 p->prio = prio; 3378 3379 if (running) 3380 p->sched_class->set_curr_task(rq); 3381 if (queued) 3382 enqueue_task(rq, p, enqueue_flag); 3383 3384 check_class_changed(rq, p, prev_class, oldprio); 3385 out_unlock: 3386 preempt_disable(); /* avoid rq from going away on us */ 3387 __task_rq_unlock(rq); 3388 3389 balance_callback(rq); 3390 preempt_enable(); 3391 } 3392 #endif 3393 3394 void set_user_nice(struct task_struct *p, long nice) 3395 { 3396 int old_prio, delta, queued; 3397 unsigned long flags; 3398 struct rq *rq; 3399 3400 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 3401 return; 3402 /* 3403 * We have to be careful, if called from sys_setpriority(), 3404 * the task might be in the middle of scheduling on another CPU. 3405 */ 3406 rq = task_rq_lock(p, &flags); 3407 /* 3408 * The RT priorities are set via sched_setscheduler(), but we still 3409 * allow the 'normal' nice value to be set - but as expected 3410 * it wont have any effect on scheduling until the task is 3411 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3412 */ 3413 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3414 p->static_prio = NICE_TO_PRIO(nice); 3415 goto out_unlock; 3416 } 3417 queued = task_on_rq_queued(p); 3418 if (queued) 3419 dequeue_task(rq, p, 0); 3420 3421 p->static_prio = NICE_TO_PRIO(nice); 3422 set_load_weight(p); 3423 old_prio = p->prio; 3424 p->prio = effective_prio(p); 3425 delta = p->prio - old_prio; 3426 3427 if (queued) { 3428 enqueue_task(rq, p, 0); 3429 /* 3430 * If the task increased its priority or is running and 3431 * lowered its priority, then reschedule its CPU: 3432 */ 3433 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3434 resched_curr(rq); 3435 } 3436 out_unlock: 3437 task_rq_unlock(rq, p, &flags); 3438 } 3439 EXPORT_SYMBOL(set_user_nice); 3440 3441 /* 3442 * can_nice - check if a task can reduce its nice value 3443 * @p: task 3444 * @nice: nice value 3445 */ 3446 int can_nice(const struct task_struct *p, const int nice) 3447 { 3448 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3449 int nice_rlim = nice_to_rlimit(nice); 3450 3451 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3452 capable(CAP_SYS_NICE)); 3453 } 3454 3455 #ifdef __ARCH_WANT_SYS_NICE 3456 3457 /* 3458 * sys_nice - change the priority of the current process. 3459 * @increment: priority increment 3460 * 3461 * sys_setpriority is a more generic, but much slower function that 3462 * does similar things. 3463 */ 3464 SYSCALL_DEFINE1(nice, int, increment) 3465 { 3466 long nice, retval; 3467 3468 /* 3469 * Setpriority might change our priority at the same moment. 3470 * We don't have to worry. Conceptually one call occurs first 3471 * and we have a single winner. 3472 */ 3473 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 3474 nice = task_nice(current) + increment; 3475 3476 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 3477 if (increment < 0 && !can_nice(current, nice)) 3478 return -EPERM; 3479 3480 retval = security_task_setnice(current, nice); 3481 if (retval) 3482 return retval; 3483 3484 set_user_nice(current, nice); 3485 return 0; 3486 } 3487 3488 #endif 3489 3490 /** 3491 * task_prio - return the priority value of a given task. 3492 * @p: the task in question. 3493 * 3494 * Return: The priority value as seen by users in /proc. 3495 * RT tasks are offset by -200. Normal tasks are centered 3496 * around 0, value goes from -16 to +15. 3497 */ 3498 int task_prio(const struct task_struct *p) 3499 { 3500 return p->prio - MAX_RT_PRIO; 3501 } 3502 3503 /** 3504 * idle_cpu - is a given cpu idle currently? 3505 * @cpu: the processor in question. 3506 * 3507 * Return: 1 if the CPU is currently idle. 0 otherwise. 3508 */ 3509 int idle_cpu(int cpu) 3510 { 3511 struct rq *rq = cpu_rq(cpu); 3512 3513 if (rq->curr != rq->idle) 3514 return 0; 3515 3516 if (rq->nr_running) 3517 return 0; 3518 3519 #ifdef CONFIG_SMP 3520 if (!llist_empty(&rq->wake_list)) 3521 return 0; 3522 #endif 3523 3524 return 1; 3525 } 3526 3527 /** 3528 * idle_task - return the idle task for a given cpu. 3529 * @cpu: the processor in question. 3530 * 3531 * Return: The idle task for the cpu @cpu. 3532 */ 3533 struct task_struct *idle_task(int cpu) 3534 { 3535 return cpu_rq(cpu)->idle; 3536 } 3537 3538 /** 3539 * find_process_by_pid - find a process with a matching PID value. 3540 * @pid: the pid in question. 3541 * 3542 * The task of @pid, if found. %NULL otherwise. 3543 */ 3544 static struct task_struct *find_process_by_pid(pid_t pid) 3545 { 3546 return pid ? find_task_by_vpid(pid) : current; 3547 } 3548 3549 /* 3550 * This function initializes the sched_dl_entity of a newly becoming 3551 * SCHED_DEADLINE task. 3552 * 3553 * Only the static values are considered here, the actual runtime and the 3554 * absolute deadline will be properly calculated when the task is enqueued 3555 * for the first time with its new policy. 3556 */ 3557 static void 3558 __setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3559 { 3560 struct sched_dl_entity *dl_se = &p->dl; 3561 3562 dl_se->dl_runtime = attr->sched_runtime; 3563 dl_se->dl_deadline = attr->sched_deadline; 3564 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3565 dl_se->flags = attr->sched_flags; 3566 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3567 3568 /* 3569 * Changing the parameters of a task is 'tricky' and we're not doing 3570 * the correct thing -- also see task_dead_dl() and switched_from_dl(). 3571 * 3572 * What we SHOULD do is delay the bandwidth release until the 0-lag 3573 * point. This would include retaining the task_struct until that time 3574 * and change dl_overflow() to not immediately decrement the current 3575 * amount. 3576 * 3577 * Instead we retain the current runtime/deadline and let the new 3578 * parameters take effect after the current reservation period lapses. 3579 * This is safe (albeit pessimistic) because the 0-lag point is always 3580 * before the current scheduling deadline. 3581 * 3582 * We can still have temporary overloads because we do not delay the 3583 * change in bandwidth until that time; so admission control is 3584 * not on the safe side. It does however guarantee tasks will never 3585 * consume more than promised. 3586 */ 3587 } 3588 3589 /* 3590 * sched_setparam() passes in -1 for its policy, to let the functions 3591 * it calls know not to change it. 3592 */ 3593 #define SETPARAM_POLICY -1 3594 3595 static void __setscheduler_params(struct task_struct *p, 3596 const struct sched_attr *attr) 3597 { 3598 int policy = attr->sched_policy; 3599 3600 if (policy == SETPARAM_POLICY) 3601 policy = p->policy; 3602 3603 p->policy = policy; 3604 3605 if (dl_policy(policy)) 3606 __setparam_dl(p, attr); 3607 else if (fair_policy(policy)) 3608 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3609 3610 /* 3611 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3612 * !rt_policy. Always setting this ensures that things like 3613 * getparam()/getattr() don't report silly values for !rt tasks. 3614 */ 3615 p->rt_priority = attr->sched_priority; 3616 p->normal_prio = normal_prio(p); 3617 set_load_weight(p); 3618 } 3619 3620 /* Actually do priority change: must hold pi & rq lock. */ 3621 static void __setscheduler(struct rq *rq, struct task_struct *p, 3622 const struct sched_attr *attr, bool keep_boost) 3623 { 3624 __setscheduler_params(p, attr); 3625 3626 /* 3627 * Keep a potential priority boosting if called from 3628 * sched_setscheduler(). 3629 */ 3630 if (keep_boost) 3631 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p)); 3632 else 3633 p->prio = normal_prio(p); 3634 3635 if (dl_prio(p->prio)) 3636 p->sched_class = &dl_sched_class; 3637 else if (rt_prio(p->prio)) 3638 p->sched_class = &rt_sched_class; 3639 else 3640 p->sched_class = &fair_sched_class; 3641 } 3642 3643 static void 3644 __getparam_dl(struct task_struct *p, struct sched_attr *attr) 3645 { 3646 struct sched_dl_entity *dl_se = &p->dl; 3647 3648 attr->sched_priority = p->rt_priority; 3649 attr->sched_runtime = dl_se->dl_runtime; 3650 attr->sched_deadline = dl_se->dl_deadline; 3651 attr->sched_period = dl_se->dl_period; 3652 attr->sched_flags = dl_se->flags; 3653 } 3654 3655 /* 3656 * This function validates the new parameters of a -deadline task. 3657 * We ask for the deadline not being zero, and greater or equal 3658 * than the runtime, as well as the period of being zero or 3659 * greater than deadline. Furthermore, we have to be sure that 3660 * user parameters are above the internal resolution of 1us (we 3661 * check sched_runtime only since it is always the smaller one) and 3662 * below 2^63 ns (we have to check both sched_deadline and 3663 * sched_period, as the latter can be zero). 3664 */ 3665 static bool 3666 __checkparam_dl(const struct sched_attr *attr) 3667 { 3668 /* deadline != 0 */ 3669 if (attr->sched_deadline == 0) 3670 return false; 3671 3672 /* 3673 * Since we truncate DL_SCALE bits, make sure we're at least 3674 * that big. 3675 */ 3676 if (attr->sched_runtime < (1ULL << DL_SCALE)) 3677 return false; 3678 3679 /* 3680 * Since we use the MSB for wrap-around and sign issues, make 3681 * sure it's not set (mind that period can be equal to zero). 3682 */ 3683 if (attr->sched_deadline & (1ULL << 63) || 3684 attr->sched_period & (1ULL << 63)) 3685 return false; 3686 3687 /* runtime <= deadline <= period (if period != 0) */ 3688 if ((attr->sched_period != 0 && 3689 attr->sched_period < attr->sched_deadline) || 3690 attr->sched_deadline < attr->sched_runtime) 3691 return false; 3692 3693 return true; 3694 } 3695 3696 /* 3697 * check the target process has a UID that matches the current process's 3698 */ 3699 static bool check_same_owner(struct task_struct *p) 3700 { 3701 const struct cred *cred = current_cred(), *pcred; 3702 bool match; 3703 3704 rcu_read_lock(); 3705 pcred = __task_cred(p); 3706 match = (uid_eq(cred->euid, pcred->euid) || 3707 uid_eq(cred->euid, pcred->uid)); 3708 rcu_read_unlock(); 3709 return match; 3710 } 3711 3712 static bool dl_param_changed(struct task_struct *p, 3713 const struct sched_attr *attr) 3714 { 3715 struct sched_dl_entity *dl_se = &p->dl; 3716 3717 if (dl_se->dl_runtime != attr->sched_runtime || 3718 dl_se->dl_deadline != attr->sched_deadline || 3719 dl_se->dl_period != attr->sched_period || 3720 dl_se->flags != attr->sched_flags) 3721 return true; 3722 3723 return false; 3724 } 3725 3726 static int __sched_setscheduler(struct task_struct *p, 3727 const struct sched_attr *attr, 3728 bool user, bool pi) 3729 { 3730 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3731 MAX_RT_PRIO - 1 - attr->sched_priority; 3732 int retval, oldprio, oldpolicy = -1, queued, running; 3733 int new_effective_prio, policy = attr->sched_policy; 3734 unsigned long flags; 3735 const struct sched_class *prev_class; 3736 struct rq *rq; 3737 int reset_on_fork; 3738 3739 /* may grab non-irq protected spin_locks */ 3740 BUG_ON(in_interrupt()); 3741 recheck: 3742 /* double check policy once rq lock held */ 3743 if (policy < 0) { 3744 reset_on_fork = p->sched_reset_on_fork; 3745 policy = oldpolicy = p->policy; 3746 } else { 3747 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3748 3749 if (policy != SCHED_DEADLINE && 3750 policy != SCHED_FIFO && policy != SCHED_RR && 3751 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3752 policy != SCHED_IDLE) 3753 return -EINVAL; 3754 } 3755 3756 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3757 return -EINVAL; 3758 3759 /* 3760 * Valid priorities for SCHED_FIFO and SCHED_RR are 3761 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3762 * SCHED_BATCH and SCHED_IDLE is 0. 3763 */ 3764 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3765 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3766 return -EINVAL; 3767 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3768 (rt_policy(policy) != (attr->sched_priority != 0))) 3769 return -EINVAL; 3770 3771 /* 3772 * Allow unprivileged RT tasks to decrease priority: 3773 */ 3774 if (user && !capable(CAP_SYS_NICE)) { 3775 if (fair_policy(policy)) { 3776 if (attr->sched_nice < task_nice(p) && 3777 !can_nice(p, attr->sched_nice)) 3778 return -EPERM; 3779 } 3780 3781 if (rt_policy(policy)) { 3782 unsigned long rlim_rtprio = 3783 task_rlimit(p, RLIMIT_RTPRIO); 3784 3785 /* can't set/change the rt policy */ 3786 if (policy != p->policy && !rlim_rtprio) 3787 return -EPERM; 3788 3789 /* can't increase priority */ 3790 if (attr->sched_priority > p->rt_priority && 3791 attr->sched_priority > rlim_rtprio) 3792 return -EPERM; 3793 } 3794 3795 /* 3796 * Can't set/change SCHED_DEADLINE policy at all for now 3797 * (safest behavior); in the future we would like to allow 3798 * unprivileged DL tasks to increase their relative deadline 3799 * or reduce their runtime (both ways reducing utilization) 3800 */ 3801 if (dl_policy(policy)) 3802 return -EPERM; 3803 3804 /* 3805 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3806 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3807 */ 3808 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3809 if (!can_nice(p, task_nice(p))) 3810 return -EPERM; 3811 } 3812 3813 /* can't change other user's priorities */ 3814 if (!check_same_owner(p)) 3815 return -EPERM; 3816 3817 /* Normal users shall not reset the sched_reset_on_fork flag */ 3818 if (p->sched_reset_on_fork && !reset_on_fork) 3819 return -EPERM; 3820 } 3821 3822 if (user) { 3823 retval = security_task_setscheduler(p); 3824 if (retval) 3825 return retval; 3826 } 3827 3828 /* 3829 * make sure no PI-waiters arrive (or leave) while we are 3830 * changing the priority of the task: 3831 * 3832 * To be able to change p->policy safely, the appropriate 3833 * runqueue lock must be held. 3834 */ 3835 rq = task_rq_lock(p, &flags); 3836 3837 /* 3838 * Changing the policy of the stop threads its a very bad idea 3839 */ 3840 if (p == rq->stop) { 3841 task_rq_unlock(rq, p, &flags); 3842 return -EINVAL; 3843 } 3844 3845 /* 3846 * If not changing anything there's no need to proceed further, 3847 * but store a possible modification of reset_on_fork. 3848 */ 3849 if (unlikely(policy == p->policy)) { 3850 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3851 goto change; 3852 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3853 goto change; 3854 if (dl_policy(policy) && dl_param_changed(p, attr)) 3855 goto change; 3856 3857 p->sched_reset_on_fork = reset_on_fork; 3858 task_rq_unlock(rq, p, &flags); 3859 return 0; 3860 } 3861 change: 3862 3863 if (user) { 3864 #ifdef CONFIG_RT_GROUP_SCHED 3865 /* 3866 * Do not allow realtime tasks into groups that have no runtime 3867 * assigned. 3868 */ 3869 if (rt_bandwidth_enabled() && rt_policy(policy) && 3870 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3871 !task_group_is_autogroup(task_group(p))) { 3872 task_rq_unlock(rq, p, &flags); 3873 return -EPERM; 3874 } 3875 #endif 3876 #ifdef CONFIG_SMP 3877 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3878 cpumask_t *span = rq->rd->span; 3879 3880 /* 3881 * Don't allow tasks with an affinity mask smaller than 3882 * the entire root_domain to become SCHED_DEADLINE. We 3883 * will also fail if there's no bandwidth available. 3884 */ 3885 if (!cpumask_subset(span, &p->cpus_allowed) || 3886 rq->rd->dl_bw.bw == 0) { 3887 task_rq_unlock(rq, p, &flags); 3888 return -EPERM; 3889 } 3890 } 3891 #endif 3892 } 3893 3894 /* recheck policy now with rq lock held */ 3895 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3896 policy = oldpolicy = -1; 3897 task_rq_unlock(rq, p, &flags); 3898 goto recheck; 3899 } 3900 3901 /* 3902 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3903 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3904 * is available. 3905 */ 3906 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3907 task_rq_unlock(rq, p, &flags); 3908 return -EBUSY; 3909 } 3910 3911 p->sched_reset_on_fork = reset_on_fork; 3912 oldprio = p->prio; 3913 3914 if (pi) { 3915 /* 3916 * Take priority boosted tasks into account. If the new 3917 * effective priority is unchanged, we just store the new 3918 * normal parameters and do not touch the scheduler class and 3919 * the runqueue. This will be done when the task deboost 3920 * itself. 3921 */ 3922 new_effective_prio = rt_mutex_get_effective_prio(p, newprio); 3923 if (new_effective_prio == oldprio) { 3924 __setscheduler_params(p, attr); 3925 task_rq_unlock(rq, p, &flags); 3926 return 0; 3927 } 3928 } 3929 3930 queued = task_on_rq_queued(p); 3931 running = task_current(rq, p); 3932 if (queued) 3933 dequeue_task(rq, p, 0); 3934 if (running) 3935 put_prev_task(rq, p); 3936 3937 prev_class = p->sched_class; 3938 __setscheduler(rq, p, attr, pi); 3939 3940 if (running) 3941 p->sched_class->set_curr_task(rq); 3942 if (queued) { 3943 /* 3944 * We enqueue to tail when the priority of a task is 3945 * increased (user space view). 3946 */ 3947 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3948 } 3949 3950 check_class_changed(rq, p, prev_class, oldprio); 3951 preempt_disable(); /* avoid rq from going away on us */ 3952 task_rq_unlock(rq, p, &flags); 3953 3954 if (pi) 3955 rt_mutex_adjust_pi(p); 3956 3957 /* 3958 * Run balance callbacks after we've adjusted the PI chain. 3959 */ 3960 balance_callback(rq); 3961 preempt_enable(); 3962 3963 return 0; 3964 } 3965 3966 static int _sched_setscheduler(struct task_struct *p, int policy, 3967 const struct sched_param *param, bool check) 3968 { 3969 struct sched_attr attr = { 3970 .sched_policy = policy, 3971 .sched_priority = param->sched_priority, 3972 .sched_nice = PRIO_TO_NICE(p->static_prio), 3973 }; 3974 3975 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ 3976 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { 3977 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3978 policy &= ~SCHED_RESET_ON_FORK; 3979 attr.sched_policy = policy; 3980 } 3981 3982 return __sched_setscheduler(p, &attr, check, true); 3983 } 3984 /** 3985 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3986 * @p: the task in question. 3987 * @policy: new policy. 3988 * @param: structure containing the new RT priority. 3989 * 3990 * Return: 0 on success. An error code otherwise. 3991 * 3992 * NOTE that the task may be already dead. 3993 */ 3994 int sched_setscheduler(struct task_struct *p, int policy, 3995 const struct sched_param *param) 3996 { 3997 return _sched_setscheduler(p, policy, param, true); 3998 } 3999 EXPORT_SYMBOL_GPL(sched_setscheduler); 4000 4001 int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 4002 { 4003 return __sched_setscheduler(p, attr, true, true); 4004 } 4005 EXPORT_SYMBOL_GPL(sched_setattr); 4006 4007 /** 4008 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 4009 * @p: the task in question. 4010 * @policy: new policy. 4011 * @param: structure containing the new RT priority. 4012 * 4013 * Just like sched_setscheduler, only don't bother checking if the 4014 * current context has permission. For example, this is needed in 4015 * stop_machine(): we create temporary high priority worker threads, 4016 * but our caller might not have that capability. 4017 * 4018 * Return: 0 on success. An error code otherwise. 4019 */ 4020 int sched_setscheduler_nocheck(struct task_struct *p, int policy, 4021 const struct sched_param *param) 4022 { 4023 return _sched_setscheduler(p, policy, param, false); 4024 } 4025 4026 static int 4027 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 4028 { 4029 struct sched_param lparam; 4030 struct task_struct *p; 4031 int retval; 4032 4033 if (!param || pid < 0) 4034 return -EINVAL; 4035 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 4036 return -EFAULT; 4037 4038 rcu_read_lock(); 4039 retval = -ESRCH; 4040 p = find_process_by_pid(pid); 4041 if (p != NULL) 4042 retval = sched_setscheduler(p, policy, &lparam); 4043 rcu_read_unlock(); 4044 4045 return retval; 4046 } 4047 4048 /* 4049 * Mimics kernel/events/core.c perf_copy_attr(). 4050 */ 4051 static int sched_copy_attr(struct sched_attr __user *uattr, 4052 struct sched_attr *attr) 4053 { 4054 u32 size; 4055 int ret; 4056 4057 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 4058 return -EFAULT; 4059 4060 /* 4061 * zero the full structure, so that a short copy will be nice. 4062 */ 4063 memset(attr, 0, sizeof(*attr)); 4064 4065 ret = get_user(size, &uattr->size); 4066 if (ret) 4067 return ret; 4068 4069 if (size > PAGE_SIZE) /* silly large */ 4070 goto err_size; 4071 4072 if (!size) /* abi compat */ 4073 size = SCHED_ATTR_SIZE_VER0; 4074 4075 if (size < SCHED_ATTR_SIZE_VER0) 4076 goto err_size; 4077 4078 /* 4079 * If we're handed a bigger struct than we know of, 4080 * ensure all the unknown bits are 0 - i.e. new 4081 * user-space does not rely on any kernel feature 4082 * extensions we dont know about yet. 4083 */ 4084 if (size > sizeof(*attr)) { 4085 unsigned char __user *addr; 4086 unsigned char __user *end; 4087 unsigned char val; 4088 4089 addr = (void __user *)uattr + sizeof(*attr); 4090 end = (void __user *)uattr + size; 4091 4092 for (; addr < end; addr++) { 4093 ret = get_user(val, addr); 4094 if (ret) 4095 return ret; 4096 if (val) 4097 goto err_size; 4098 } 4099 size = sizeof(*attr); 4100 } 4101 4102 ret = copy_from_user(attr, uattr, size); 4103 if (ret) 4104 return -EFAULT; 4105 4106 /* 4107 * XXX: do we want to be lenient like existing syscalls; or do we want 4108 * to be strict and return an error on out-of-bounds values? 4109 */ 4110 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 4111 4112 return 0; 4113 4114 err_size: 4115 put_user(sizeof(*attr), &uattr->size); 4116 return -E2BIG; 4117 } 4118 4119 /** 4120 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 4121 * @pid: the pid in question. 4122 * @policy: new policy. 4123 * @param: structure containing the new RT priority. 4124 * 4125 * Return: 0 on success. An error code otherwise. 4126 */ 4127 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 4128 struct sched_param __user *, param) 4129 { 4130 /* negative values for policy are not valid */ 4131 if (policy < 0) 4132 return -EINVAL; 4133 4134 return do_sched_setscheduler(pid, policy, param); 4135 } 4136 4137 /** 4138 * sys_sched_setparam - set/change the RT priority of a thread 4139 * @pid: the pid in question. 4140 * @param: structure containing the new RT priority. 4141 * 4142 * Return: 0 on success. An error code otherwise. 4143 */ 4144 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 4145 { 4146 return do_sched_setscheduler(pid, SETPARAM_POLICY, param); 4147 } 4148 4149 /** 4150 * sys_sched_setattr - same as above, but with extended sched_attr 4151 * @pid: the pid in question. 4152 * @uattr: structure containing the extended parameters. 4153 * @flags: for future extension. 4154 */ 4155 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 4156 unsigned int, flags) 4157 { 4158 struct sched_attr attr; 4159 struct task_struct *p; 4160 int retval; 4161 4162 if (!uattr || pid < 0 || flags) 4163 return -EINVAL; 4164 4165 retval = sched_copy_attr(uattr, &attr); 4166 if (retval) 4167 return retval; 4168 4169 if ((int)attr.sched_policy < 0) 4170 return -EINVAL; 4171 4172 rcu_read_lock(); 4173 retval = -ESRCH; 4174 p = find_process_by_pid(pid); 4175 if (p != NULL) 4176 retval = sched_setattr(p, &attr); 4177 rcu_read_unlock(); 4178 4179 return retval; 4180 } 4181 4182 /** 4183 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 4184 * @pid: the pid in question. 4185 * 4186 * Return: On success, the policy of the thread. Otherwise, a negative error 4187 * code. 4188 */ 4189 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 4190 { 4191 struct task_struct *p; 4192 int retval; 4193 4194 if (pid < 0) 4195 return -EINVAL; 4196 4197 retval = -ESRCH; 4198 rcu_read_lock(); 4199 p = find_process_by_pid(pid); 4200 if (p) { 4201 retval = security_task_getscheduler(p); 4202 if (!retval) 4203 retval = p->policy 4204 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4205 } 4206 rcu_read_unlock(); 4207 return retval; 4208 } 4209 4210 /** 4211 * sys_sched_getparam - get the RT priority of a thread 4212 * @pid: the pid in question. 4213 * @param: structure containing the RT priority. 4214 * 4215 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 4216 * code. 4217 */ 4218 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4219 { 4220 struct sched_param lp = { .sched_priority = 0 }; 4221 struct task_struct *p; 4222 int retval; 4223 4224 if (!param || pid < 0) 4225 return -EINVAL; 4226 4227 rcu_read_lock(); 4228 p = find_process_by_pid(pid); 4229 retval = -ESRCH; 4230 if (!p) 4231 goto out_unlock; 4232 4233 retval = security_task_getscheduler(p); 4234 if (retval) 4235 goto out_unlock; 4236 4237 if (task_has_rt_policy(p)) 4238 lp.sched_priority = p->rt_priority; 4239 rcu_read_unlock(); 4240 4241 /* 4242 * This one might sleep, we cannot do it with a spinlock held ... 4243 */ 4244 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4245 4246 return retval; 4247 4248 out_unlock: 4249 rcu_read_unlock(); 4250 return retval; 4251 } 4252 4253 static int sched_read_attr(struct sched_attr __user *uattr, 4254 struct sched_attr *attr, 4255 unsigned int usize) 4256 { 4257 int ret; 4258 4259 if (!access_ok(VERIFY_WRITE, uattr, usize)) 4260 return -EFAULT; 4261 4262 /* 4263 * If we're handed a smaller struct than we know of, 4264 * ensure all the unknown bits are 0 - i.e. old 4265 * user-space does not get uncomplete information. 4266 */ 4267 if (usize < sizeof(*attr)) { 4268 unsigned char *addr; 4269 unsigned char *end; 4270 4271 addr = (void *)attr + usize; 4272 end = (void *)attr + sizeof(*attr); 4273 4274 for (; addr < end; addr++) { 4275 if (*addr) 4276 return -EFBIG; 4277 } 4278 4279 attr->size = usize; 4280 } 4281 4282 ret = copy_to_user(uattr, attr, attr->size); 4283 if (ret) 4284 return -EFAULT; 4285 4286 return 0; 4287 } 4288 4289 /** 4290 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 4291 * @pid: the pid in question. 4292 * @uattr: structure containing the extended parameters. 4293 * @size: sizeof(attr) for fwd/bwd comp. 4294 * @flags: for future extension. 4295 */ 4296 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 4297 unsigned int, size, unsigned int, flags) 4298 { 4299 struct sched_attr attr = { 4300 .size = sizeof(struct sched_attr), 4301 }; 4302 struct task_struct *p; 4303 int retval; 4304 4305 if (!uattr || pid < 0 || size > PAGE_SIZE || 4306 size < SCHED_ATTR_SIZE_VER0 || flags) 4307 return -EINVAL; 4308 4309 rcu_read_lock(); 4310 p = find_process_by_pid(pid); 4311 retval = -ESRCH; 4312 if (!p) 4313 goto out_unlock; 4314 4315 retval = security_task_getscheduler(p); 4316 if (retval) 4317 goto out_unlock; 4318 4319 attr.sched_policy = p->policy; 4320 if (p->sched_reset_on_fork) 4321 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 4322 if (task_has_dl_policy(p)) 4323 __getparam_dl(p, &attr); 4324 else if (task_has_rt_policy(p)) 4325 attr.sched_priority = p->rt_priority; 4326 else 4327 attr.sched_nice = task_nice(p); 4328 4329 rcu_read_unlock(); 4330 4331 retval = sched_read_attr(uattr, &attr, size); 4332 return retval; 4333 4334 out_unlock: 4335 rcu_read_unlock(); 4336 return retval; 4337 } 4338 4339 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4340 { 4341 cpumask_var_t cpus_allowed, new_mask; 4342 struct task_struct *p; 4343 int retval; 4344 4345 rcu_read_lock(); 4346 4347 p = find_process_by_pid(pid); 4348 if (!p) { 4349 rcu_read_unlock(); 4350 return -ESRCH; 4351 } 4352 4353 /* Prevent p going away */ 4354 get_task_struct(p); 4355 rcu_read_unlock(); 4356 4357 if (p->flags & PF_NO_SETAFFINITY) { 4358 retval = -EINVAL; 4359 goto out_put_task; 4360 } 4361 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4362 retval = -ENOMEM; 4363 goto out_put_task; 4364 } 4365 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4366 retval = -ENOMEM; 4367 goto out_free_cpus_allowed; 4368 } 4369 retval = -EPERM; 4370 if (!check_same_owner(p)) { 4371 rcu_read_lock(); 4372 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 4373 rcu_read_unlock(); 4374 goto out_free_new_mask; 4375 } 4376 rcu_read_unlock(); 4377 } 4378 4379 retval = security_task_setscheduler(p); 4380 if (retval) 4381 goto out_free_new_mask; 4382 4383 4384 cpuset_cpus_allowed(p, cpus_allowed); 4385 cpumask_and(new_mask, in_mask, cpus_allowed); 4386 4387 /* 4388 * Since bandwidth control happens on root_domain basis, 4389 * if admission test is enabled, we only admit -deadline 4390 * tasks allowed to run on all the CPUs in the task's 4391 * root_domain. 4392 */ 4393 #ifdef CONFIG_SMP 4394 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { 4395 rcu_read_lock(); 4396 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) { 4397 retval = -EBUSY; 4398 rcu_read_unlock(); 4399 goto out_free_new_mask; 4400 } 4401 rcu_read_unlock(); 4402 } 4403 #endif 4404 again: 4405 retval = __set_cpus_allowed_ptr(p, new_mask, true); 4406 4407 if (!retval) { 4408 cpuset_cpus_allowed(p, cpus_allowed); 4409 if (!cpumask_subset(new_mask, cpus_allowed)) { 4410 /* 4411 * We must have raced with a concurrent cpuset 4412 * update. Just reset the cpus_allowed to the 4413 * cpuset's cpus_allowed 4414 */ 4415 cpumask_copy(new_mask, cpus_allowed); 4416 goto again; 4417 } 4418 } 4419 out_free_new_mask: 4420 free_cpumask_var(new_mask); 4421 out_free_cpus_allowed: 4422 free_cpumask_var(cpus_allowed); 4423 out_put_task: 4424 put_task_struct(p); 4425 return retval; 4426 } 4427 4428 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4429 struct cpumask *new_mask) 4430 { 4431 if (len < cpumask_size()) 4432 cpumask_clear(new_mask); 4433 else if (len > cpumask_size()) 4434 len = cpumask_size(); 4435 4436 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4437 } 4438 4439 /** 4440 * sys_sched_setaffinity - set the cpu affinity of a process 4441 * @pid: pid of the process 4442 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4443 * @user_mask_ptr: user-space pointer to the new cpu mask 4444 * 4445 * Return: 0 on success. An error code otherwise. 4446 */ 4447 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4448 unsigned long __user *, user_mask_ptr) 4449 { 4450 cpumask_var_t new_mask; 4451 int retval; 4452 4453 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4454 return -ENOMEM; 4455 4456 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4457 if (retval == 0) 4458 retval = sched_setaffinity(pid, new_mask); 4459 free_cpumask_var(new_mask); 4460 return retval; 4461 } 4462 4463 long sched_getaffinity(pid_t pid, struct cpumask *mask) 4464 { 4465 struct task_struct *p; 4466 unsigned long flags; 4467 int retval; 4468 4469 rcu_read_lock(); 4470 4471 retval = -ESRCH; 4472 p = find_process_by_pid(pid); 4473 if (!p) 4474 goto out_unlock; 4475 4476 retval = security_task_getscheduler(p); 4477 if (retval) 4478 goto out_unlock; 4479 4480 raw_spin_lock_irqsave(&p->pi_lock, flags); 4481 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4482 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4483 4484 out_unlock: 4485 rcu_read_unlock(); 4486 4487 return retval; 4488 } 4489 4490 /** 4491 * sys_sched_getaffinity - get the cpu affinity of a process 4492 * @pid: pid of the process 4493 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4494 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4495 * 4496 * Return: 0 on success. An error code otherwise. 4497 */ 4498 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4499 unsigned long __user *, user_mask_ptr) 4500 { 4501 int ret; 4502 cpumask_var_t mask; 4503 4504 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4505 return -EINVAL; 4506 if (len & (sizeof(unsigned long)-1)) 4507 return -EINVAL; 4508 4509 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4510 return -ENOMEM; 4511 4512 ret = sched_getaffinity(pid, mask); 4513 if (ret == 0) { 4514 size_t retlen = min_t(size_t, len, cpumask_size()); 4515 4516 if (copy_to_user(user_mask_ptr, mask, retlen)) 4517 ret = -EFAULT; 4518 else 4519 ret = retlen; 4520 } 4521 free_cpumask_var(mask); 4522 4523 return ret; 4524 } 4525 4526 /** 4527 * sys_sched_yield - yield the current processor to other threads. 4528 * 4529 * This function yields the current CPU to other tasks. If there are no 4530 * other threads running on this CPU then this function will return. 4531 * 4532 * Return: 0. 4533 */ 4534 SYSCALL_DEFINE0(sched_yield) 4535 { 4536 struct rq *rq = this_rq_lock(); 4537 4538 schedstat_inc(rq, yld_count); 4539 current->sched_class->yield_task(rq); 4540 4541 /* 4542 * Since we are going to call schedule() anyway, there's 4543 * no need to preempt or enable interrupts: 4544 */ 4545 __release(rq->lock); 4546 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4547 do_raw_spin_unlock(&rq->lock); 4548 sched_preempt_enable_no_resched(); 4549 4550 schedule(); 4551 4552 return 0; 4553 } 4554 4555 int __sched _cond_resched(void) 4556 { 4557 if (should_resched(0)) { 4558 preempt_schedule_common(); 4559 return 1; 4560 } 4561 return 0; 4562 } 4563 EXPORT_SYMBOL(_cond_resched); 4564 4565 /* 4566 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4567 * call schedule, and on return reacquire the lock. 4568 * 4569 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4570 * operations here to prevent schedule() from being called twice (once via 4571 * spin_unlock(), once by hand). 4572 */ 4573 int __cond_resched_lock(spinlock_t *lock) 4574 { 4575 int resched = should_resched(PREEMPT_LOCK_OFFSET); 4576 int ret = 0; 4577 4578 lockdep_assert_held(lock); 4579 4580 if (spin_needbreak(lock) || resched) { 4581 spin_unlock(lock); 4582 if (resched) 4583 preempt_schedule_common(); 4584 else 4585 cpu_relax(); 4586 ret = 1; 4587 spin_lock(lock); 4588 } 4589 return ret; 4590 } 4591 EXPORT_SYMBOL(__cond_resched_lock); 4592 4593 int __sched __cond_resched_softirq(void) 4594 { 4595 BUG_ON(!in_softirq()); 4596 4597 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) { 4598 local_bh_enable(); 4599 preempt_schedule_common(); 4600 local_bh_disable(); 4601 return 1; 4602 } 4603 return 0; 4604 } 4605 EXPORT_SYMBOL(__cond_resched_softirq); 4606 4607 /** 4608 * yield - yield the current processor to other threads. 4609 * 4610 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4611 * 4612 * The scheduler is at all times free to pick the calling task as the most 4613 * eligible task to run, if removing the yield() call from your code breaks 4614 * it, its already broken. 4615 * 4616 * Typical broken usage is: 4617 * 4618 * while (!event) 4619 * yield(); 4620 * 4621 * where one assumes that yield() will let 'the other' process run that will 4622 * make event true. If the current task is a SCHED_FIFO task that will never 4623 * happen. Never use yield() as a progress guarantee!! 4624 * 4625 * If you want to use yield() to wait for something, use wait_event(). 4626 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4627 * If you still want to use yield(), do not! 4628 */ 4629 void __sched yield(void) 4630 { 4631 set_current_state(TASK_RUNNING); 4632 sys_sched_yield(); 4633 } 4634 EXPORT_SYMBOL(yield); 4635 4636 /** 4637 * yield_to - yield the current processor to another thread in 4638 * your thread group, or accelerate that thread toward the 4639 * processor it's on. 4640 * @p: target task 4641 * @preempt: whether task preemption is allowed or not 4642 * 4643 * It's the caller's job to ensure that the target task struct 4644 * can't go away on us before we can do any checks. 4645 * 4646 * Return: 4647 * true (>0) if we indeed boosted the target task. 4648 * false (0) if we failed to boost the target. 4649 * -ESRCH if there's no task to yield to. 4650 */ 4651 int __sched yield_to(struct task_struct *p, bool preempt) 4652 { 4653 struct task_struct *curr = current; 4654 struct rq *rq, *p_rq; 4655 unsigned long flags; 4656 int yielded = 0; 4657 4658 local_irq_save(flags); 4659 rq = this_rq(); 4660 4661 again: 4662 p_rq = task_rq(p); 4663 /* 4664 * If we're the only runnable task on the rq and target rq also 4665 * has only one task, there's absolutely no point in yielding. 4666 */ 4667 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4668 yielded = -ESRCH; 4669 goto out_irq; 4670 } 4671 4672 double_rq_lock(rq, p_rq); 4673 if (task_rq(p) != p_rq) { 4674 double_rq_unlock(rq, p_rq); 4675 goto again; 4676 } 4677 4678 if (!curr->sched_class->yield_to_task) 4679 goto out_unlock; 4680 4681 if (curr->sched_class != p->sched_class) 4682 goto out_unlock; 4683 4684 if (task_running(p_rq, p) || p->state) 4685 goto out_unlock; 4686 4687 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4688 if (yielded) { 4689 schedstat_inc(rq, yld_count); 4690 /* 4691 * Make p's CPU reschedule; pick_next_entity takes care of 4692 * fairness. 4693 */ 4694 if (preempt && rq != p_rq) 4695 resched_curr(p_rq); 4696 } 4697 4698 out_unlock: 4699 double_rq_unlock(rq, p_rq); 4700 out_irq: 4701 local_irq_restore(flags); 4702 4703 if (yielded > 0) 4704 schedule(); 4705 4706 return yielded; 4707 } 4708 EXPORT_SYMBOL_GPL(yield_to); 4709 4710 /* 4711 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4712 * that process accounting knows that this is a task in IO wait state. 4713 */ 4714 long __sched io_schedule_timeout(long timeout) 4715 { 4716 int old_iowait = current->in_iowait; 4717 struct rq *rq; 4718 long ret; 4719 4720 current->in_iowait = 1; 4721 blk_schedule_flush_plug(current); 4722 4723 delayacct_blkio_start(); 4724 rq = raw_rq(); 4725 atomic_inc(&rq->nr_iowait); 4726 ret = schedule_timeout(timeout); 4727 current->in_iowait = old_iowait; 4728 atomic_dec(&rq->nr_iowait); 4729 delayacct_blkio_end(); 4730 4731 return ret; 4732 } 4733 EXPORT_SYMBOL(io_schedule_timeout); 4734 4735 /** 4736 * sys_sched_get_priority_max - return maximum RT priority. 4737 * @policy: scheduling class. 4738 * 4739 * Return: On success, this syscall returns the maximum 4740 * rt_priority that can be used by a given scheduling class. 4741 * On failure, a negative error code is returned. 4742 */ 4743 SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4744 { 4745 int ret = -EINVAL; 4746 4747 switch (policy) { 4748 case SCHED_FIFO: 4749 case SCHED_RR: 4750 ret = MAX_USER_RT_PRIO-1; 4751 break; 4752 case SCHED_DEADLINE: 4753 case SCHED_NORMAL: 4754 case SCHED_BATCH: 4755 case SCHED_IDLE: 4756 ret = 0; 4757 break; 4758 } 4759 return ret; 4760 } 4761 4762 /** 4763 * sys_sched_get_priority_min - return minimum RT priority. 4764 * @policy: scheduling class. 4765 * 4766 * Return: On success, this syscall returns the minimum 4767 * rt_priority that can be used by a given scheduling class. 4768 * On failure, a negative error code is returned. 4769 */ 4770 SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4771 { 4772 int ret = -EINVAL; 4773 4774 switch (policy) { 4775 case SCHED_FIFO: 4776 case SCHED_RR: 4777 ret = 1; 4778 break; 4779 case SCHED_DEADLINE: 4780 case SCHED_NORMAL: 4781 case SCHED_BATCH: 4782 case SCHED_IDLE: 4783 ret = 0; 4784 } 4785 return ret; 4786 } 4787 4788 /** 4789 * sys_sched_rr_get_interval - return the default timeslice of a process. 4790 * @pid: pid of the process. 4791 * @interval: userspace pointer to the timeslice value. 4792 * 4793 * this syscall writes the default timeslice value of a given process 4794 * into the user-space timespec buffer. A value of '0' means infinity. 4795 * 4796 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4797 * an error code. 4798 */ 4799 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4800 struct timespec __user *, interval) 4801 { 4802 struct task_struct *p; 4803 unsigned int time_slice; 4804 unsigned long flags; 4805 struct rq *rq; 4806 int retval; 4807 struct timespec t; 4808 4809 if (pid < 0) 4810 return -EINVAL; 4811 4812 retval = -ESRCH; 4813 rcu_read_lock(); 4814 p = find_process_by_pid(pid); 4815 if (!p) 4816 goto out_unlock; 4817 4818 retval = security_task_getscheduler(p); 4819 if (retval) 4820 goto out_unlock; 4821 4822 rq = task_rq_lock(p, &flags); 4823 time_slice = 0; 4824 if (p->sched_class->get_rr_interval) 4825 time_slice = p->sched_class->get_rr_interval(rq, p); 4826 task_rq_unlock(rq, p, &flags); 4827 4828 rcu_read_unlock(); 4829 jiffies_to_timespec(time_slice, &t); 4830 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4831 return retval; 4832 4833 out_unlock: 4834 rcu_read_unlock(); 4835 return retval; 4836 } 4837 4838 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4839 4840 void sched_show_task(struct task_struct *p) 4841 { 4842 unsigned long free = 0; 4843 int ppid; 4844 unsigned long state = p->state; 4845 4846 if (state) 4847 state = __ffs(state) + 1; 4848 printk(KERN_INFO "%-15.15s %c", p->comm, 4849 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4850 #if BITS_PER_LONG == 32 4851 if (state == TASK_RUNNING) 4852 printk(KERN_CONT " running "); 4853 else 4854 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4855 #else 4856 if (state == TASK_RUNNING) 4857 printk(KERN_CONT " running task "); 4858 else 4859 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4860 #endif 4861 #ifdef CONFIG_DEBUG_STACK_USAGE 4862 free = stack_not_used(p); 4863 #endif 4864 ppid = 0; 4865 rcu_read_lock(); 4866 if (pid_alive(p)) 4867 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4868 rcu_read_unlock(); 4869 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4870 task_pid_nr(p), ppid, 4871 (unsigned long)task_thread_info(p)->flags); 4872 4873 print_worker_info(KERN_INFO, p); 4874 show_stack(p, NULL); 4875 } 4876 4877 void show_state_filter(unsigned long state_filter) 4878 { 4879 struct task_struct *g, *p; 4880 4881 #if BITS_PER_LONG == 32 4882 printk(KERN_INFO 4883 " task PC stack pid father\n"); 4884 #else 4885 printk(KERN_INFO 4886 " task PC stack pid father\n"); 4887 #endif 4888 rcu_read_lock(); 4889 for_each_process_thread(g, p) { 4890 /* 4891 * reset the NMI-timeout, listing all files on a slow 4892 * console might take a lot of time: 4893 */ 4894 touch_nmi_watchdog(); 4895 if (!state_filter || (p->state & state_filter)) 4896 sched_show_task(p); 4897 } 4898 4899 touch_all_softlockup_watchdogs(); 4900 4901 #ifdef CONFIG_SCHED_DEBUG 4902 sysrq_sched_debug_show(); 4903 #endif 4904 rcu_read_unlock(); 4905 /* 4906 * Only show locks if all tasks are dumped: 4907 */ 4908 if (!state_filter) 4909 debug_show_all_locks(); 4910 } 4911 4912 void init_idle_bootup_task(struct task_struct *idle) 4913 { 4914 idle->sched_class = &idle_sched_class; 4915 } 4916 4917 /** 4918 * init_idle - set up an idle thread for a given CPU 4919 * @idle: task in question 4920 * @cpu: cpu the idle task belongs to 4921 * 4922 * NOTE: this function does not set the idle thread's NEED_RESCHED 4923 * flag, to make booting more robust. 4924 */ 4925 void init_idle(struct task_struct *idle, int cpu) 4926 { 4927 struct rq *rq = cpu_rq(cpu); 4928 unsigned long flags; 4929 4930 raw_spin_lock_irqsave(&idle->pi_lock, flags); 4931 raw_spin_lock(&rq->lock); 4932 4933 __sched_fork(0, idle); 4934 idle->state = TASK_RUNNING; 4935 idle->se.exec_start = sched_clock(); 4936 4937 #ifdef CONFIG_SMP 4938 /* 4939 * Its possible that init_idle() gets called multiple times on a task, 4940 * in that case do_set_cpus_allowed() will not do the right thing. 4941 * 4942 * And since this is boot we can forgo the serialization. 4943 */ 4944 set_cpus_allowed_common(idle, cpumask_of(cpu)); 4945 #endif 4946 /* 4947 * We're having a chicken and egg problem, even though we are 4948 * holding rq->lock, the cpu isn't yet set to this cpu so the 4949 * lockdep check in task_group() will fail. 4950 * 4951 * Similar case to sched_fork(). / Alternatively we could 4952 * use task_rq_lock() here and obtain the other rq->lock. 4953 * 4954 * Silence PROVE_RCU 4955 */ 4956 rcu_read_lock(); 4957 __set_task_cpu(idle, cpu); 4958 rcu_read_unlock(); 4959 4960 rq->curr = rq->idle = idle; 4961 idle->on_rq = TASK_ON_RQ_QUEUED; 4962 #ifdef CONFIG_SMP 4963 idle->on_cpu = 1; 4964 #endif 4965 raw_spin_unlock(&rq->lock); 4966 raw_spin_unlock_irqrestore(&idle->pi_lock, flags); 4967 4968 /* Set the preempt count _outside_ the spinlocks! */ 4969 init_idle_preempt_count(idle, cpu); 4970 4971 /* 4972 * The idle tasks have their own, simple scheduling class: 4973 */ 4974 idle->sched_class = &idle_sched_class; 4975 ftrace_graph_init_idle_task(idle, cpu); 4976 vtime_init_idle(idle, cpu); 4977 #ifdef CONFIG_SMP 4978 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4979 #endif 4980 } 4981 4982 int cpuset_cpumask_can_shrink(const struct cpumask *cur, 4983 const struct cpumask *trial) 4984 { 4985 int ret = 1, trial_cpus; 4986 struct dl_bw *cur_dl_b; 4987 unsigned long flags; 4988 4989 if (!cpumask_weight(cur)) 4990 return ret; 4991 4992 rcu_read_lock_sched(); 4993 cur_dl_b = dl_bw_of(cpumask_any(cur)); 4994 trial_cpus = cpumask_weight(trial); 4995 4996 raw_spin_lock_irqsave(&cur_dl_b->lock, flags); 4997 if (cur_dl_b->bw != -1 && 4998 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw) 4999 ret = 0; 5000 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags); 5001 rcu_read_unlock_sched(); 5002 5003 return ret; 5004 } 5005 5006 int task_can_attach(struct task_struct *p, 5007 const struct cpumask *cs_cpus_allowed) 5008 { 5009 int ret = 0; 5010 5011 /* 5012 * Kthreads which disallow setaffinity shouldn't be moved 5013 * to a new cpuset; we don't want to change their cpu 5014 * affinity and isolating such threads by their set of 5015 * allowed nodes is unnecessary. Thus, cpusets are not 5016 * applicable for such threads. This prevents checking for 5017 * success of set_cpus_allowed_ptr() on all attached tasks 5018 * before cpus_allowed may be changed. 5019 */ 5020 if (p->flags & PF_NO_SETAFFINITY) { 5021 ret = -EINVAL; 5022 goto out; 5023 } 5024 5025 #ifdef CONFIG_SMP 5026 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span, 5027 cs_cpus_allowed)) { 5028 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask, 5029 cs_cpus_allowed); 5030 struct dl_bw *dl_b; 5031 bool overflow; 5032 int cpus; 5033 unsigned long flags; 5034 5035 rcu_read_lock_sched(); 5036 dl_b = dl_bw_of(dest_cpu); 5037 raw_spin_lock_irqsave(&dl_b->lock, flags); 5038 cpus = dl_bw_cpus(dest_cpu); 5039 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw); 5040 if (overflow) 5041 ret = -EBUSY; 5042 else { 5043 /* 5044 * We reserve space for this task in the destination 5045 * root_domain, as we can't fail after this point. 5046 * We will free resources in the source root_domain 5047 * later on (see set_cpus_allowed_dl()). 5048 */ 5049 __dl_add(dl_b, p->dl.dl_bw); 5050 } 5051 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5052 rcu_read_unlock_sched(); 5053 5054 } 5055 #endif 5056 out: 5057 return ret; 5058 } 5059 5060 #ifdef CONFIG_SMP 5061 5062 #ifdef CONFIG_NUMA_BALANCING 5063 /* Migrate current task p to target_cpu */ 5064 int migrate_task_to(struct task_struct *p, int target_cpu) 5065 { 5066 struct migration_arg arg = { p, target_cpu }; 5067 int curr_cpu = task_cpu(p); 5068 5069 if (curr_cpu == target_cpu) 5070 return 0; 5071 5072 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 5073 return -EINVAL; 5074 5075 /* TODO: This is not properly updating schedstats */ 5076 5077 trace_sched_move_numa(p, curr_cpu, target_cpu); 5078 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 5079 } 5080 5081 /* 5082 * Requeue a task on a given node and accurately track the number of NUMA 5083 * tasks on the runqueues 5084 */ 5085 void sched_setnuma(struct task_struct *p, int nid) 5086 { 5087 struct rq *rq; 5088 unsigned long flags; 5089 bool queued, running; 5090 5091 rq = task_rq_lock(p, &flags); 5092 queued = task_on_rq_queued(p); 5093 running = task_current(rq, p); 5094 5095 if (queued) 5096 dequeue_task(rq, p, 0); 5097 if (running) 5098 put_prev_task(rq, p); 5099 5100 p->numa_preferred_nid = nid; 5101 5102 if (running) 5103 p->sched_class->set_curr_task(rq); 5104 if (queued) 5105 enqueue_task(rq, p, 0); 5106 task_rq_unlock(rq, p, &flags); 5107 } 5108 #endif /* CONFIG_NUMA_BALANCING */ 5109 5110 #ifdef CONFIG_HOTPLUG_CPU 5111 /* 5112 * Ensures that the idle task is using init_mm right before its cpu goes 5113 * offline. 5114 */ 5115 void idle_task_exit(void) 5116 { 5117 struct mm_struct *mm = current->active_mm; 5118 5119 BUG_ON(cpu_online(smp_processor_id())); 5120 5121 if (mm != &init_mm) { 5122 switch_mm(mm, &init_mm, current); 5123 finish_arch_post_lock_switch(); 5124 } 5125 mmdrop(mm); 5126 } 5127 5128 /* 5129 * Since this CPU is going 'away' for a while, fold any nr_active delta 5130 * we might have. Assumes we're called after migrate_tasks() so that the 5131 * nr_active count is stable. 5132 * 5133 * Also see the comment "Global load-average calculations". 5134 */ 5135 static void calc_load_migrate(struct rq *rq) 5136 { 5137 long delta = calc_load_fold_active(rq); 5138 if (delta) 5139 atomic_long_add(delta, &calc_load_tasks); 5140 } 5141 5142 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 5143 { 5144 } 5145 5146 static const struct sched_class fake_sched_class = { 5147 .put_prev_task = put_prev_task_fake, 5148 }; 5149 5150 static struct task_struct fake_task = { 5151 /* 5152 * Avoid pull_{rt,dl}_task() 5153 */ 5154 .prio = MAX_PRIO + 1, 5155 .sched_class = &fake_sched_class, 5156 }; 5157 5158 /* 5159 * Migrate all tasks from the rq, sleeping tasks will be migrated by 5160 * try_to_wake_up()->select_task_rq(). 5161 * 5162 * Called with rq->lock held even though we'er in stop_machine() and 5163 * there's no concurrency possible, we hold the required locks anyway 5164 * because of lock validation efforts. 5165 */ 5166 static void migrate_tasks(struct rq *dead_rq) 5167 { 5168 struct rq *rq = dead_rq; 5169 struct task_struct *next, *stop = rq->stop; 5170 int dest_cpu; 5171 5172 /* 5173 * Fudge the rq selection such that the below task selection loop 5174 * doesn't get stuck on the currently eligible stop task. 5175 * 5176 * We're currently inside stop_machine() and the rq is either stuck 5177 * in the stop_machine_cpu_stop() loop, or we're executing this code, 5178 * either way we should never end up calling schedule() until we're 5179 * done here. 5180 */ 5181 rq->stop = NULL; 5182 5183 /* 5184 * put_prev_task() and pick_next_task() sched 5185 * class method both need to have an up-to-date 5186 * value of rq->clock[_task] 5187 */ 5188 update_rq_clock(rq); 5189 5190 for (;;) { 5191 /* 5192 * There's this thread running, bail when that's the only 5193 * remaining thread. 5194 */ 5195 if (rq->nr_running == 1) 5196 break; 5197 5198 /* 5199 * pick_next_task assumes pinned rq->lock. 5200 */ 5201 lockdep_pin_lock(&rq->lock); 5202 next = pick_next_task(rq, &fake_task); 5203 BUG_ON(!next); 5204 next->sched_class->put_prev_task(rq, next); 5205 5206 /* 5207 * Rules for changing task_struct::cpus_allowed are holding 5208 * both pi_lock and rq->lock, such that holding either 5209 * stabilizes the mask. 5210 * 5211 * Drop rq->lock is not quite as disastrous as it usually is 5212 * because !cpu_active at this point, which means load-balance 5213 * will not interfere. Also, stop-machine. 5214 */ 5215 lockdep_unpin_lock(&rq->lock); 5216 raw_spin_unlock(&rq->lock); 5217 raw_spin_lock(&next->pi_lock); 5218 raw_spin_lock(&rq->lock); 5219 5220 /* 5221 * Since we're inside stop-machine, _nothing_ should have 5222 * changed the task, WARN if weird stuff happened, because in 5223 * that case the above rq->lock drop is a fail too. 5224 */ 5225 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) { 5226 raw_spin_unlock(&next->pi_lock); 5227 continue; 5228 } 5229 5230 /* Find suitable destination for @next, with force if needed. */ 5231 dest_cpu = select_fallback_rq(dead_rq->cpu, next); 5232 5233 rq = __migrate_task(rq, next, dest_cpu); 5234 if (rq != dead_rq) { 5235 raw_spin_unlock(&rq->lock); 5236 rq = dead_rq; 5237 raw_spin_lock(&rq->lock); 5238 } 5239 raw_spin_unlock(&next->pi_lock); 5240 } 5241 5242 rq->stop = stop; 5243 } 5244 #endif /* CONFIG_HOTPLUG_CPU */ 5245 5246 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 5247 5248 static struct ctl_table sd_ctl_dir[] = { 5249 { 5250 .procname = "sched_domain", 5251 .mode = 0555, 5252 }, 5253 {} 5254 }; 5255 5256 static struct ctl_table sd_ctl_root[] = { 5257 { 5258 .procname = "kernel", 5259 .mode = 0555, 5260 .child = sd_ctl_dir, 5261 }, 5262 {} 5263 }; 5264 5265 static struct ctl_table *sd_alloc_ctl_entry(int n) 5266 { 5267 struct ctl_table *entry = 5268 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 5269 5270 return entry; 5271 } 5272 5273 static void sd_free_ctl_entry(struct ctl_table **tablep) 5274 { 5275 struct ctl_table *entry; 5276 5277 /* 5278 * In the intermediate directories, both the child directory and 5279 * procname are dynamically allocated and could fail but the mode 5280 * will always be set. In the lowest directory the names are 5281 * static strings and all have proc handlers. 5282 */ 5283 for (entry = *tablep; entry->mode; entry++) { 5284 if (entry->child) 5285 sd_free_ctl_entry(&entry->child); 5286 if (entry->proc_handler == NULL) 5287 kfree(entry->procname); 5288 } 5289 5290 kfree(*tablep); 5291 *tablep = NULL; 5292 } 5293 5294 static int min_load_idx = 0; 5295 static int max_load_idx = CPU_LOAD_IDX_MAX-1; 5296 5297 static void 5298 set_table_entry(struct ctl_table *entry, 5299 const char *procname, void *data, int maxlen, 5300 umode_t mode, proc_handler *proc_handler, 5301 bool load_idx) 5302 { 5303 entry->procname = procname; 5304 entry->data = data; 5305 entry->maxlen = maxlen; 5306 entry->mode = mode; 5307 entry->proc_handler = proc_handler; 5308 5309 if (load_idx) { 5310 entry->extra1 = &min_load_idx; 5311 entry->extra2 = &max_load_idx; 5312 } 5313 } 5314 5315 static struct ctl_table * 5316 sd_alloc_ctl_domain_table(struct sched_domain *sd) 5317 { 5318 struct ctl_table *table = sd_alloc_ctl_entry(14); 5319 5320 if (table == NULL) 5321 return NULL; 5322 5323 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5324 sizeof(long), 0644, proc_doulongvec_minmax, false); 5325 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5326 sizeof(long), 0644, proc_doulongvec_minmax, false); 5327 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5328 sizeof(int), 0644, proc_dointvec_minmax, true); 5329 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5330 sizeof(int), 0644, proc_dointvec_minmax, true); 5331 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5332 sizeof(int), 0644, proc_dointvec_minmax, true); 5333 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5334 sizeof(int), 0644, proc_dointvec_minmax, true); 5335 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5336 sizeof(int), 0644, proc_dointvec_minmax, true); 5337 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5338 sizeof(int), 0644, proc_dointvec_minmax, false); 5339 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5340 sizeof(int), 0644, proc_dointvec_minmax, false); 5341 set_table_entry(&table[9], "cache_nice_tries", 5342 &sd->cache_nice_tries, 5343 sizeof(int), 0644, proc_dointvec_minmax, false); 5344 set_table_entry(&table[10], "flags", &sd->flags, 5345 sizeof(int), 0644, proc_dointvec_minmax, false); 5346 set_table_entry(&table[11], "max_newidle_lb_cost", 5347 &sd->max_newidle_lb_cost, 5348 sizeof(long), 0644, proc_doulongvec_minmax, false); 5349 set_table_entry(&table[12], "name", sd->name, 5350 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 5351 /* &table[13] is terminator */ 5352 5353 return table; 5354 } 5355 5356 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5357 { 5358 struct ctl_table *entry, *table; 5359 struct sched_domain *sd; 5360 int domain_num = 0, i; 5361 char buf[32]; 5362 5363 for_each_domain(cpu, sd) 5364 domain_num++; 5365 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5366 if (table == NULL) 5367 return NULL; 5368 5369 i = 0; 5370 for_each_domain(cpu, sd) { 5371 snprintf(buf, 32, "domain%d", i); 5372 entry->procname = kstrdup(buf, GFP_KERNEL); 5373 entry->mode = 0555; 5374 entry->child = sd_alloc_ctl_domain_table(sd); 5375 entry++; 5376 i++; 5377 } 5378 return table; 5379 } 5380 5381 static struct ctl_table_header *sd_sysctl_header; 5382 static void register_sched_domain_sysctl(void) 5383 { 5384 int i, cpu_num = num_possible_cpus(); 5385 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5386 char buf[32]; 5387 5388 WARN_ON(sd_ctl_dir[0].child); 5389 sd_ctl_dir[0].child = entry; 5390 5391 if (entry == NULL) 5392 return; 5393 5394 for_each_possible_cpu(i) { 5395 snprintf(buf, 32, "cpu%d", i); 5396 entry->procname = kstrdup(buf, GFP_KERNEL); 5397 entry->mode = 0555; 5398 entry->child = sd_alloc_ctl_cpu_table(i); 5399 entry++; 5400 } 5401 5402 WARN_ON(sd_sysctl_header); 5403 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5404 } 5405 5406 /* may be called multiple times per register */ 5407 static void unregister_sched_domain_sysctl(void) 5408 { 5409 unregister_sysctl_table(sd_sysctl_header); 5410 sd_sysctl_header = NULL; 5411 if (sd_ctl_dir[0].child) 5412 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5413 } 5414 #else 5415 static void register_sched_domain_sysctl(void) 5416 { 5417 } 5418 static void unregister_sched_domain_sysctl(void) 5419 { 5420 } 5421 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */ 5422 5423 static void set_rq_online(struct rq *rq) 5424 { 5425 if (!rq->online) { 5426 const struct sched_class *class; 5427 5428 cpumask_set_cpu(rq->cpu, rq->rd->online); 5429 rq->online = 1; 5430 5431 for_each_class(class) { 5432 if (class->rq_online) 5433 class->rq_online(rq); 5434 } 5435 } 5436 } 5437 5438 static void set_rq_offline(struct rq *rq) 5439 { 5440 if (rq->online) { 5441 const struct sched_class *class; 5442 5443 for_each_class(class) { 5444 if (class->rq_offline) 5445 class->rq_offline(rq); 5446 } 5447 5448 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5449 rq->online = 0; 5450 } 5451 } 5452 5453 /* 5454 * migration_call - callback that gets triggered when a CPU is added. 5455 * Here we can start up the necessary migration thread for the new CPU. 5456 */ 5457 static int 5458 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5459 { 5460 int cpu = (long)hcpu; 5461 unsigned long flags; 5462 struct rq *rq = cpu_rq(cpu); 5463 5464 switch (action & ~CPU_TASKS_FROZEN) { 5465 5466 case CPU_UP_PREPARE: 5467 rq->calc_load_update = calc_load_update; 5468 break; 5469 5470 case CPU_ONLINE: 5471 /* Update our root-domain */ 5472 raw_spin_lock_irqsave(&rq->lock, flags); 5473 if (rq->rd) { 5474 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5475 5476 set_rq_online(rq); 5477 } 5478 raw_spin_unlock_irqrestore(&rq->lock, flags); 5479 break; 5480 5481 #ifdef CONFIG_HOTPLUG_CPU 5482 case CPU_DYING: 5483 sched_ttwu_pending(); 5484 /* Update our root-domain */ 5485 raw_spin_lock_irqsave(&rq->lock, flags); 5486 if (rq->rd) { 5487 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5488 set_rq_offline(rq); 5489 } 5490 migrate_tasks(rq); 5491 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5492 raw_spin_unlock_irqrestore(&rq->lock, flags); 5493 break; 5494 5495 case CPU_DEAD: 5496 calc_load_migrate(rq); 5497 break; 5498 #endif 5499 } 5500 5501 update_max_interval(); 5502 5503 return NOTIFY_OK; 5504 } 5505 5506 /* 5507 * Register at high priority so that task migration (migrate_all_tasks) 5508 * happens before everything else. This has to be lower priority than 5509 * the notifier in the perf_event subsystem, though. 5510 */ 5511 static struct notifier_block migration_notifier = { 5512 .notifier_call = migration_call, 5513 .priority = CPU_PRI_MIGRATION, 5514 }; 5515 5516 static void set_cpu_rq_start_time(void) 5517 { 5518 int cpu = smp_processor_id(); 5519 struct rq *rq = cpu_rq(cpu); 5520 rq->age_stamp = sched_clock_cpu(cpu); 5521 } 5522 5523 static int sched_cpu_active(struct notifier_block *nfb, 5524 unsigned long action, void *hcpu) 5525 { 5526 switch (action & ~CPU_TASKS_FROZEN) { 5527 case CPU_STARTING: 5528 set_cpu_rq_start_time(); 5529 return NOTIFY_OK; 5530 case CPU_ONLINE: 5531 /* 5532 * At this point a starting CPU has marked itself as online via 5533 * set_cpu_online(). But it might not yet have marked itself 5534 * as active, which is essential from here on. 5535 * 5536 * Thus, fall-through and help the starting CPU along. 5537 */ 5538 case CPU_DOWN_FAILED: 5539 set_cpu_active((long)hcpu, true); 5540 return NOTIFY_OK; 5541 default: 5542 return NOTIFY_DONE; 5543 } 5544 } 5545 5546 static int sched_cpu_inactive(struct notifier_block *nfb, 5547 unsigned long action, void *hcpu) 5548 { 5549 switch (action & ~CPU_TASKS_FROZEN) { 5550 case CPU_DOWN_PREPARE: 5551 set_cpu_active((long)hcpu, false); 5552 return NOTIFY_OK; 5553 default: 5554 return NOTIFY_DONE; 5555 } 5556 } 5557 5558 static int __init migration_init(void) 5559 { 5560 void *cpu = (void *)(long)smp_processor_id(); 5561 int err; 5562 5563 /* Initialize migration for the boot CPU */ 5564 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5565 BUG_ON(err == NOTIFY_BAD); 5566 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5567 register_cpu_notifier(&migration_notifier); 5568 5569 /* Register cpu active notifiers */ 5570 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5571 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5572 5573 return 0; 5574 } 5575 early_initcall(migration_init); 5576 5577 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5578 5579 #ifdef CONFIG_SCHED_DEBUG 5580 5581 static __read_mostly int sched_debug_enabled; 5582 5583 static int __init sched_debug_setup(char *str) 5584 { 5585 sched_debug_enabled = 1; 5586 5587 return 0; 5588 } 5589 early_param("sched_debug", sched_debug_setup); 5590 5591 static inline bool sched_debug(void) 5592 { 5593 return sched_debug_enabled; 5594 } 5595 5596 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5597 struct cpumask *groupmask) 5598 { 5599 struct sched_group *group = sd->groups; 5600 5601 cpumask_clear(groupmask); 5602 5603 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5604 5605 if (!(sd->flags & SD_LOAD_BALANCE)) { 5606 printk("does not load-balance\n"); 5607 if (sd->parent) 5608 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5609 " has parent"); 5610 return -1; 5611 } 5612 5613 printk(KERN_CONT "span %*pbl level %s\n", 5614 cpumask_pr_args(sched_domain_span(sd)), sd->name); 5615 5616 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5617 printk(KERN_ERR "ERROR: domain->span does not contain " 5618 "CPU%d\n", cpu); 5619 } 5620 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5621 printk(KERN_ERR "ERROR: domain->groups does not contain" 5622 " CPU%d\n", cpu); 5623 } 5624 5625 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5626 do { 5627 if (!group) { 5628 printk("\n"); 5629 printk(KERN_ERR "ERROR: group is NULL\n"); 5630 break; 5631 } 5632 5633 if (!cpumask_weight(sched_group_cpus(group))) { 5634 printk(KERN_CONT "\n"); 5635 printk(KERN_ERR "ERROR: empty group\n"); 5636 break; 5637 } 5638 5639 if (!(sd->flags & SD_OVERLAP) && 5640 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5641 printk(KERN_CONT "\n"); 5642 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5643 break; 5644 } 5645 5646 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5647 5648 printk(KERN_CONT " %*pbl", 5649 cpumask_pr_args(sched_group_cpus(group))); 5650 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { 5651 printk(KERN_CONT " (cpu_capacity = %d)", 5652 group->sgc->capacity); 5653 } 5654 5655 group = group->next; 5656 } while (group != sd->groups); 5657 printk(KERN_CONT "\n"); 5658 5659 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5660 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5661 5662 if (sd->parent && 5663 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5664 printk(KERN_ERR "ERROR: parent span is not a superset " 5665 "of domain->span\n"); 5666 return 0; 5667 } 5668 5669 static void sched_domain_debug(struct sched_domain *sd, int cpu) 5670 { 5671 int level = 0; 5672 5673 if (!sched_debug_enabled) 5674 return; 5675 5676 if (!sd) { 5677 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5678 return; 5679 } 5680 5681 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5682 5683 for (;;) { 5684 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5685 break; 5686 level++; 5687 sd = sd->parent; 5688 if (!sd) 5689 break; 5690 } 5691 } 5692 #else /* !CONFIG_SCHED_DEBUG */ 5693 # define sched_domain_debug(sd, cpu) do { } while (0) 5694 static inline bool sched_debug(void) 5695 { 5696 return false; 5697 } 5698 #endif /* CONFIG_SCHED_DEBUG */ 5699 5700 static int sd_degenerate(struct sched_domain *sd) 5701 { 5702 if (cpumask_weight(sched_domain_span(sd)) == 1) 5703 return 1; 5704 5705 /* Following flags need at least 2 groups */ 5706 if (sd->flags & (SD_LOAD_BALANCE | 5707 SD_BALANCE_NEWIDLE | 5708 SD_BALANCE_FORK | 5709 SD_BALANCE_EXEC | 5710 SD_SHARE_CPUCAPACITY | 5711 SD_SHARE_PKG_RESOURCES | 5712 SD_SHARE_POWERDOMAIN)) { 5713 if (sd->groups != sd->groups->next) 5714 return 0; 5715 } 5716 5717 /* Following flags don't use groups */ 5718 if (sd->flags & (SD_WAKE_AFFINE)) 5719 return 0; 5720 5721 return 1; 5722 } 5723 5724 static int 5725 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5726 { 5727 unsigned long cflags = sd->flags, pflags = parent->flags; 5728 5729 if (sd_degenerate(parent)) 5730 return 1; 5731 5732 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5733 return 0; 5734 5735 /* Flags needing groups don't count if only 1 group in parent */ 5736 if (parent->groups == parent->groups->next) { 5737 pflags &= ~(SD_LOAD_BALANCE | 5738 SD_BALANCE_NEWIDLE | 5739 SD_BALANCE_FORK | 5740 SD_BALANCE_EXEC | 5741 SD_SHARE_CPUCAPACITY | 5742 SD_SHARE_PKG_RESOURCES | 5743 SD_PREFER_SIBLING | 5744 SD_SHARE_POWERDOMAIN); 5745 if (nr_node_ids == 1) 5746 pflags &= ~SD_SERIALIZE; 5747 } 5748 if (~cflags & pflags) 5749 return 0; 5750 5751 return 1; 5752 } 5753 5754 static void free_rootdomain(struct rcu_head *rcu) 5755 { 5756 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5757 5758 cpupri_cleanup(&rd->cpupri); 5759 cpudl_cleanup(&rd->cpudl); 5760 free_cpumask_var(rd->dlo_mask); 5761 free_cpumask_var(rd->rto_mask); 5762 free_cpumask_var(rd->online); 5763 free_cpumask_var(rd->span); 5764 kfree(rd); 5765 } 5766 5767 static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5768 { 5769 struct root_domain *old_rd = NULL; 5770 unsigned long flags; 5771 5772 raw_spin_lock_irqsave(&rq->lock, flags); 5773 5774 if (rq->rd) { 5775 old_rd = rq->rd; 5776 5777 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5778 set_rq_offline(rq); 5779 5780 cpumask_clear_cpu(rq->cpu, old_rd->span); 5781 5782 /* 5783 * If we dont want to free the old_rd yet then 5784 * set old_rd to NULL to skip the freeing later 5785 * in this function: 5786 */ 5787 if (!atomic_dec_and_test(&old_rd->refcount)) 5788 old_rd = NULL; 5789 } 5790 5791 atomic_inc(&rd->refcount); 5792 rq->rd = rd; 5793 5794 cpumask_set_cpu(rq->cpu, rd->span); 5795 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5796 set_rq_online(rq); 5797 5798 raw_spin_unlock_irqrestore(&rq->lock, flags); 5799 5800 if (old_rd) 5801 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5802 } 5803 5804 static int init_rootdomain(struct root_domain *rd) 5805 { 5806 memset(rd, 0, sizeof(*rd)); 5807 5808 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5809 goto out; 5810 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5811 goto free_span; 5812 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5813 goto free_online; 5814 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5815 goto free_dlo_mask; 5816 5817 init_dl_bw(&rd->dl_bw); 5818 if (cpudl_init(&rd->cpudl) != 0) 5819 goto free_dlo_mask; 5820 5821 if (cpupri_init(&rd->cpupri) != 0) 5822 goto free_rto_mask; 5823 return 0; 5824 5825 free_rto_mask: 5826 free_cpumask_var(rd->rto_mask); 5827 free_dlo_mask: 5828 free_cpumask_var(rd->dlo_mask); 5829 free_online: 5830 free_cpumask_var(rd->online); 5831 free_span: 5832 free_cpumask_var(rd->span); 5833 out: 5834 return -ENOMEM; 5835 } 5836 5837 /* 5838 * By default the system creates a single root-domain with all cpus as 5839 * members (mimicking the global state we have today). 5840 */ 5841 struct root_domain def_root_domain; 5842 5843 static void init_defrootdomain(void) 5844 { 5845 init_rootdomain(&def_root_domain); 5846 5847 atomic_set(&def_root_domain.refcount, 1); 5848 } 5849 5850 static struct root_domain *alloc_rootdomain(void) 5851 { 5852 struct root_domain *rd; 5853 5854 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5855 if (!rd) 5856 return NULL; 5857 5858 if (init_rootdomain(rd) != 0) { 5859 kfree(rd); 5860 return NULL; 5861 } 5862 5863 return rd; 5864 } 5865 5866 static void free_sched_groups(struct sched_group *sg, int free_sgc) 5867 { 5868 struct sched_group *tmp, *first; 5869 5870 if (!sg) 5871 return; 5872 5873 first = sg; 5874 do { 5875 tmp = sg->next; 5876 5877 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) 5878 kfree(sg->sgc); 5879 5880 kfree(sg); 5881 sg = tmp; 5882 } while (sg != first); 5883 } 5884 5885 static void free_sched_domain(struct rcu_head *rcu) 5886 { 5887 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5888 5889 /* 5890 * If its an overlapping domain it has private groups, iterate and 5891 * nuke them all. 5892 */ 5893 if (sd->flags & SD_OVERLAP) { 5894 free_sched_groups(sd->groups, 1); 5895 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5896 kfree(sd->groups->sgc); 5897 kfree(sd->groups); 5898 } 5899 kfree(sd); 5900 } 5901 5902 static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5903 { 5904 call_rcu(&sd->rcu, free_sched_domain); 5905 } 5906 5907 static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5908 { 5909 for (; sd; sd = sd->parent) 5910 destroy_sched_domain(sd, cpu); 5911 } 5912 5913 /* 5914 * Keep a special pointer to the highest sched_domain that has 5915 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5916 * allows us to avoid some pointer chasing select_idle_sibling(). 5917 * 5918 * Also keep a unique ID per domain (we use the first cpu number in 5919 * the cpumask of the domain), this allows us to quickly tell if 5920 * two cpus are in the same cache domain, see cpus_share_cache(). 5921 */ 5922 DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5923 DEFINE_PER_CPU(int, sd_llc_size); 5924 DEFINE_PER_CPU(int, sd_llc_id); 5925 DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5926 DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5927 DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5928 5929 static void update_top_cache_domain(int cpu) 5930 { 5931 struct sched_domain *sd; 5932 struct sched_domain *busy_sd = NULL; 5933 int id = cpu; 5934 int size = 1; 5935 5936 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5937 if (sd) { 5938 id = cpumask_first(sched_domain_span(sd)); 5939 size = cpumask_weight(sched_domain_span(sd)); 5940 busy_sd = sd->parent; /* sd_busy */ 5941 } 5942 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5943 5944 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5945 per_cpu(sd_llc_size, cpu) = size; 5946 per_cpu(sd_llc_id, cpu) = id; 5947 5948 sd = lowest_flag_domain(cpu, SD_NUMA); 5949 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5950 5951 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5952 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5953 } 5954 5955 /* 5956 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5957 * hold the hotplug lock. 5958 */ 5959 static void 5960 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5961 { 5962 struct rq *rq = cpu_rq(cpu); 5963 struct sched_domain *tmp; 5964 5965 /* Remove the sched domains which do not contribute to scheduling. */ 5966 for (tmp = sd; tmp; ) { 5967 struct sched_domain *parent = tmp->parent; 5968 if (!parent) 5969 break; 5970 5971 if (sd_parent_degenerate(tmp, parent)) { 5972 tmp->parent = parent->parent; 5973 if (parent->parent) 5974 parent->parent->child = tmp; 5975 /* 5976 * Transfer SD_PREFER_SIBLING down in case of a 5977 * degenerate parent; the spans match for this 5978 * so the property transfers. 5979 */ 5980 if (parent->flags & SD_PREFER_SIBLING) 5981 tmp->flags |= SD_PREFER_SIBLING; 5982 destroy_sched_domain(parent, cpu); 5983 } else 5984 tmp = tmp->parent; 5985 } 5986 5987 if (sd && sd_degenerate(sd)) { 5988 tmp = sd; 5989 sd = sd->parent; 5990 destroy_sched_domain(tmp, cpu); 5991 if (sd) 5992 sd->child = NULL; 5993 } 5994 5995 sched_domain_debug(sd, cpu); 5996 5997 rq_attach_root(rq, rd); 5998 tmp = rq->sd; 5999 rcu_assign_pointer(rq->sd, sd); 6000 destroy_sched_domains(tmp, cpu); 6001 6002 update_top_cache_domain(cpu); 6003 } 6004 6005 /* Setup the mask of cpus configured for isolated domains */ 6006 static int __init isolated_cpu_setup(char *str) 6007 { 6008 alloc_bootmem_cpumask_var(&cpu_isolated_map); 6009 cpulist_parse(str, cpu_isolated_map); 6010 return 1; 6011 } 6012 6013 __setup("isolcpus=", isolated_cpu_setup); 6014 6015 struct s_data { 6016 struct sched_domain ** __percpu sd; 6017 struct root_domain *rd; 6018 }; 6019 6020 enum s_alloc { 6021 sa_rootdomain, 6022 sa_sd, 6023 sa_sd_storage, 6024 sa_none, 6025 }; 6026 6027 /* 6028 * Build an iteration mask that can exclude certain CPUs from the upwards 6029 * domain traversal. 6030 * 6031 * Asymmetric node setups can result in situations where the domain tree is of 6032 * unequal depth, make sure to skip domains that already cover the entire 6033 * range. 6034 * 6035 * In that case build_sched_domains() will have terminated the iteration early 6036 * and our sibling sd spans will be empty. Domains should always include the 6037 * cpu they're built on, so check that. 6038 * 6039 */ 6040 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 6041 { 6042 const struct cpumask *span = sched_domain_span(sd); 6043 struct sd_data *sdd = sd->private; 6044 struct sched_domain *sibling; 6045 int i; 6046 6047 for_each_cpu(i, span) { 6048 sibling = *per_cpu_ptr(sdd->sd, i); 6049 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 6050 continue; 6051 6052 cpumask_set_cpu(i, sched_group_mask(sg)); 6053 } 6054 } 6055 6056 /* 6057 * Return the canonical balance cpu for this group, this is the first cpu 6058 * of this group that's also in the iteration mask. 6059 */ 6060 int group_balance_cpu(struct sched_group *sg) 6061 { 6062 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 6063 } 6064 6065 static int 6066 build_overlap_sched_groups(struct sched_domain *sd, int cpu) 6067 { 6068 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 6069 const struct cpumask *span = sched_domain_span(sd); 6070 struct cpumask *covered = sched_domains_tmpmask; 6071 struct sd_data *sdd = sd->private; 6072 struct sched_domain *sibling; 6073 int i; 6074 6075 cpumask_clear(covered); 6076 6077 for_each_cpu(i, span) { 6078 struct cpumask *sg_span; 6079 6080 if (cpumask_test_cpu(i, covered)) 6081 continue; 6082 6083 sibling = *per_cpu_ptr(sdd->sd, i); 6084 6085 /* See the comment near build_group_mask(). */ 6086 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 6087 continue; 6088 6089 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6090 GFP_KERNEL, cpu_to_node(cpu)); 6091 6092 if (!sg) 6093 goto fail; 6094 6095 sg_span = sched_group_cpus(sg); 6096 if (sibling->child) 6097 cpumask_copy(sg_span, sched_domain_span(sibling->child)); 6098 else 6099 cpumask_set_cpu(i, sg_span); 6100 6101 cpumask_or(covered, covered, sg_span); 6102 6103 sg->sgc = *per_cpu_ptr(sdd->sgc, i); 6104 if (atomic_inc_return(&sg->sgc->ref) == 1) 6105 build_group_mask(sd, sg); 6106 6107 /* 6108 * Initialize sgc->capacity such that even if we mess up the 6109 * domains and no possible iteration will get us here, we won't 6110 * die on a /0 trap. 6111 */ 6112 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); 6113 6114 /* 6115 * Make sure the first group of this domain contains the 6116 * canonical balance cpu. Otherwise the sched_domain iteration 6117 * breaks. See update_sg_lb_stats(). 6118 */ 6119 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 6120 group_balance_cpu(sg) == cpu) 6121 groups = sg; 6122 6123 if (!first) 6124 first = sg; 6125 if (last) 6126 last->next = sg; 6127 last = sg; 6128 last->next = first; 6129 } 6130 sd->groups = groups; 6131 6132 return 0; 6133 6134 fail: 6135 free_sched_groups(first, 0); 6136 6137 return -ENOMEM; 6138 } 6139 6140 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 6141 { 6142 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 6143 struct sched_domain *child = sd->child; 6144 6145 if (child) 6146 cpu = cpumask_first(sched_domain_span(child)); 6147 6148 if (sg) { 6149 *sg = *per_cpu_ptr(sdd->sg, cpu); 6150 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); 6151 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ 6152 } 6153 6154 return cpu; 6155 } 6156 6157 /* 6158 * build_sched_groups will build a circular linked list of the groups 6159 * covered by the given span, and will set each group's ->cpumask correctly, 6160 * and ->cpu_capacity to 0. 6161 * 6162 * Assumes the sched_domain tree is fully constructed 6163 */ 6164 static int 6165 build_sched_groups(struct sched_domain *sd, int cpu) 6166 { 6167 struct sched_group *first = NULL, *last = NULL; 6168 struct sd_data *sdd = sd->private; 6169 const struct cpumask *span = sched_domain_span(sd); 6170 struct cpumask *covered; 6171 int i; 6172 6173 get_group(cpu, sdd, &sd->groups); 6174 atomic_inc(&sd->groups->ref); 6175 6176 if (cpu != cpumask_first(span)) 6177 return 0; 6178 6179 lockdep_assert_held(&sched_domains_mutex); 6180 covered = sched_domains_tmpmask; 6181 6182 cpumask_clear(covered); 6183 6184 for_each_cpu(i, span) { 6185 struct sched_group *sg; 6186 int group, j; 6187 6188 if (cpumask_test_cpu(i, covered)) 6189 continue; 6190 6191 group = get_group(i, sdd, &sg); 6192 cpumask_setall(sched_group_mask(sg)); 6193 6194 for_each_cpu(j, span) { 6195 if (get_group(j, sdd, NULL) != group) 6196 continue; 6197 6198 cpumask_set_cpu(j, covered); 6199 cpumask_set_cpu(j, sched_group_cpus(sg)); 6200 } 6201 6202 if (!first) 6203 first = sg; 6204 if (last) 6205 last->next = sg; 6206 last = sg; 6207 } 6208 last->next = first; 6209 6210 return 0; 6211 } 6212 6213 /* 6214 * Initialize sched groups cpu_capacity. 6215 * 6216 * cpu_capacity indicates the capacity of sched group, which is used while 6217 * distributing the load between different sched groups in a sched domain. 6218 * Typically cpu_capacity for all the groups in a sched domain will be same 6219 * unless there are asymmetries in the topology. If there are asymmetries, 6220 * group having more cpu_capacity will pickup more load compared to the 6221 * group having less cpu_capacity. 6222 */ 6223 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) 6224 { 6225 struct sched_group *sg = sd->groups; 6226 6227 WARN_ON(!sg); 6228 6229 do { 6230 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 6231 sg = sg->next; 6232 } while (sg != sd->groups); 6233 6234 if (cpu != group_balance_cpu(sg)) 6235 return; 6236 6237 update_group_capacity(sd, cpu); 6238 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); 6239 } 6240 6241 /* 6242 * Initializers for schedule domains 6243 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 6244 */ 6245 6246 static int default_relax_domain_level = -1; 6247 int sched_domain_level_max; 6248 6249 static int __init setup_relax_domain_level(char *str) 6250 { 6251 if (kstrtoint(str, 0, &default_relax_domain_level)) 6252 pr_warn("Unable to set relax_domain_level\n"); 6253 6254 return 1; 6255 } 6256 __setup("relax_domain_level=", setup_relax_domain_level); 6257 6258 static void set_domain_attribute(struct sched_domain *sd, 6259 struct sched_domain_attr *attr) 6260 { 6261 int request; 6262 6263 if (!attr || attr->relax_domain_level < 0) { 6264 if (default_relax_domain_level < 0) 6265 return; 6266 else 6267 request = default_relax_domain_level; 6268 } else 6269 request = attr->relax_domain_level; 6270 if (request < sd->level) { 6271 /* turn off idle balance on this domain */ 6272 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6273 } else { 6274 /* turn on idle balance on this domain */ 6275 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6276 } 6277 } 6278 6279 static void __sdt_free(const struct cpumask *cpu_map); 6280 static int __sdt_alloc(const struct cpumask *cpu_map); 6281 6282 static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6283 const struct cpumask *cpu_map) 6284 { 6285 switch (what) { 6286 case sa_rootdomain: 6287 if (!atomic_read(&d->rd->refcount)) 6288 free_rootdomain(&d->rd->rcu); /* fall through */ 6289 case sa_sd: 6290 free_percpu(d->sd); /* fall through */ 6291 case sa_sd_storage: 6292 __sdt_free(cpu_map); /* fall through */ 6293 case sa_none: 6294 break; 6295 } 6296 } 6297 6298 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6299 const struct cpumask *cpu_map) 6300 { 6301 memset(d, 0, sizeof(*d)); 6302 6303 if (__sdt_alloc(cpu_map)) 6304 return sa_sd_storage; 6305 d->sd = alloc_percpu(struct sched_domain *); 6306 if (!d->sd) 6307 return sa_sd_storage; 6308 d->rd = alloc_rootdomain(); 6309 if (!d->rd) 6310 return sa_sd; 6311 return sa_rootdomain; 6312 } 6313 6314 /* 6315 * NULL the sd_data elements we've used to build the sched_domain and 6316 * sched_group structure so that the subsequent __free_domain_allocs() 6317 * will not free the data we're using. 6318 */ 6319 static void claim_allocations(int cpu, struct sched_domain *sd) 6320 { 6321 struct sd_data *sdd = sd->private; 6322 6323 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6324 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6325 6326 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6327 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6328 6329 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) 6330 *per_cpu_ptr(sdd->sgc, cpu) = NULL; 6331 } 6332 6333 #ifdef CONFIG_NUMA 6334 static int sched_domains_numa_levels; 6335 enum numa_topology_type sched_numa_topology_type; 6336 static int *sched_domains_numa_distance; 6337 int sched_max_numa_distance; 6338 static struct cpumask ***sched_domains_numa_masks; 6339 static int sched_domains_curr_level; 6340 #endif 6341 6342 /* 6343 * SD_flags allowed in topology descriptions. 6344 * 6345 * SD_SHARE_CPUCAPACITY - describes SMT topologies 6346 * SD_SHARE_PKG_RESOURCES - describes shared caches 6347 * SD_NUMA - describes NUMA topologies 6348 * SD_SHARE_POWERDOMAIN - describes shared power domain 6349 * 6350 * Odd one out: 6351 * SD_ASYM_PACKING - describes SMT quirks 6352 */ 6353 #define TOPOLOGY_SD_FLAGS \ 6354 (SD_SHARE_CPUCAPACITY | \ 6355 SD_SHARE_PKG_RESOURCES | \ 6356 SD_NUMA | \ 6357 SD_ASYM_PACKING | \ 6358 SD_SHARE_POWERDOMAIN) 6359 6360 static struct sched_domain * 6361 sd_init(struct sched_domain_topology_level *tl, int cpu) 6362 { 6363 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6364 int sd_weight, sd_flags = 0; 6365 6366 #ifdef CONFIG_NUMA 6367 /* 6368 * Ugly hack to pass state to sd_numa_mask()... 6369 */ 6370 sched_domains_curr_level = tl->numa_level; 6371 #endif 6372 6373 sd_weight = cpumask_weight(tl->mask(cpu)); 6374 6375 if (tl->sd_flags) 6376 sd_flags = (*tl->sd_flags)(); 6377 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 6378 "wrong sd_flags in topology description\n")) 6379 sd_flags &= ~TOPOLOGY_SD_FLAGS; 6380 6381 *sd = (struct sched_domain){ 6382 .min_interval = sd_weight, 6383 .max_interval = 2*sd_weight, 6384 .busy_factor = 32, 6385 .imbalance_pct = 125, 6386 6387 .cache_nice_tries = 0, 6388 .busy_idx = 0, 6389 .idle_idx = 0, 6390 .newidle_idx = 0, 6391 .wake_idx = 0, 6392 .forkexec_idx = 0, 6393 6394 .flags = 1*SD_LOAD_BALANCE 6395 | 1*SD_BALANCE_NEWIDLE 6396 | 1*SD_BALANCE_EXEC 6397 | 1*SD_BALANCE_FORK 6398 | 0*SD_BALANCE_WAKE 6399 | 1*SD_WAKE_AFFINE 6400 | 0*SD_SHARE_CPUCAPACITY 6401 | 0*SD_SHARE_PKG_RESOURCES 6402 | 0*SD_SERIALIZE 6403 | 0*SD_PREFER_SIBLING 6404 | 0*SD_NUMA 6405 | sd_flags 6406 , 6407 6408 .last_balance = jiffies, 6409 .balance_interval = sd_weight, 6410 .smt_gain = 0, 6411 .max_newidle_lb_cost = 0, 6412 .next_decay_max_lb_cost = jiffies, 6413 #ifdef CONFIG_SCHED_DEBUG 6414 .name = tl->name, 6415 #endif 6416 }; 6417 6418 /* 6419 * Convert topological properties into behaviour. 6420 */ 6421 6422 if (sd->flags & SD_SHARE_CPUCAPACITY) { 6423 sd->flags |= SD_PREFER_SIBLING; 6424 sd->imbalance_pct = 110; 6425 sd->smt_gain = 1178; /* ~15% */ 6426 6427 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6428 sd->imbalance_pct = 117; 6429 sd->cache_nice_tries = 1; 6430 sd->busy_idx = 2; 6431 6432 #ifdef CONFIG_NUMA 6433 } else if (sd->flags & SD_NUMA) { 6434 sd->cache_nice_tries = 2; 6435 sd->busy_idx = 3; 6436 sd->idle_idx = 2; 6437 6438 sd->flags |= SD_SERIALIZE; 6439 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6440 sd->flags &= ~(SD_BALANCE_EXEC | 6441 SD_BALANCE_FORK | 6442 SD_WAKE_AFFINE); 6443 } 6444 6445 #endif 6446 } else { 6447 sd->flags |= SD_PREFER_SIBLING; 6448 sd->cache_nice_tries = 1; 6449 sd->busy_idx = 2; 6450 sd->idle_idx = 1; 6451 } 6452 6453 sd->private = &tl->data; 6454 6455 return sd; 6456 } 6457 6458 /* 6459 * Topology list, bottom-up. 6460 */ 6461 static struct sched_domain_topology_level default_topology[] = { 6462 #ifdef CONFIG_SCHED_SMT 6463 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6464 #endif 6465 #ifdef CONFIG_SCHED_MC 6466 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6467 #endif 6468 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6469 { NULL, }, 6470 }; 6471 6472 struct sched_domain_topology_level *sched_domain_topology = default_topology; 6473 6474 #define for_each_sd_topology(tl) \ 6475 for (tl = sched_domain_topology; tl->mask; tl++) 6476 6477 void set_sched_topology(struct sched_domain_topology_level *tl) 6478 { 6479 sched_domain_topology = tl; 6480 } 6481 6482 #ifdef CONFIG_NUMA 6483 6484 static const struct cpumask *sd_numa_mask(int cpu) 6485 { 6486 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6487 } 6488 6489 static void sched_numa_warn(const char *str) 6490 { 6491 static int done = false; 6492 int i,j; 6493 6494 if (done) 6495 return; 6496 6497 done = true; 6498 6499 printk(KERN_WARNING "ERROR: %s\n\n", str); 6500 6501 for (i = 0; i < nr_node_ids; i++) { 6502 printk(KERN_WARNING " "); 6503 for (j = 0; j < nr_node_ids; j++) 6504 printk(KERN_CONT "%02d ", node_distance(i,j)); 6505 printk(KERN_CONT "\n"); 6506 } 6507 printk(KERN_WARNING "\n"); 6508 } 6509 6510 bool find_numa_distance(int distance) 6511 { 6512 int i; 6513 6514 if (distance == node_distance(0, 0)) 6515 return true; 6516 6517 for (i = 0; i < sched_domains_numa_levels; i++) { 6518 if (sched_domains_numa_distance[i] == distance) 6519 return true; 6520 } 6521 6522 return false; 6523 } 6524 6525 /* 6526 * A system can have three types of NUMA topology: 6527 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system 6528 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes 6529 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane 6530 * 6531 * The difference between a glueless mesh topology and a backplane 6532 * topology lies in whether communication between not directly 6533 * connected nodes goes through intermediary nodes (where programs 6534 * could run), or through backplane controllers. This affects 6535 * placement of programs. 6536 * 6537 * The type of topology can be discerned with the following tests: 6538 * - If the maximum distance between any nodes is 1 hop, the system 6539 * is directly connected. 6540 * - If for two nodes A and B, located N > 1 hops away from each other, 6541 * there is an intermediary node C, which is < N hops away from both 6542 * nodes A and B, the system is a glueless mesh. 6543 */ 6544 static void init_numa_topology_type(void) 6545 { 6546 int a, b, c, n; 6547 6548 n = sched_max_numa_distance; 6549 6550 if (sched_domains_numa_levels <= 1) { 6551 sched_numa_topology_type = NUMA_DIRECT; 6552 return; 6553 } 6554 6555 for_each_online_node(a) { 6556 for_each_online_node(b) { 6557 /* Find two nodes furthest removed from each other. */ 6558 if (node_distance(a, b) < n) 6559 continue; 6560 6561 /* Is there an intermediary node between a and b? */ 6562 for_each_online_node(c) { 6563 if (node_distance(a, c) < n && 6564 node_distance(b, c) < n) { 6565 sched_numa_topology_type = 6566 NUMA_GLUELESS_MESH; 6567 return; 6568 } 6569 } 6570 6571 sched_numa_topology_type = NUMA_BACKPLANE; 6572 return; 6573 } 6574 } 6575 } 6576 6577 static void sched_init_numa(void) 6578 { 6579 int next_distance, curr_distance = node_distance(0, 0); 6580 struct sched_domain_topology_level *tl; 6581 int level = 0; 6582 int i, j, k; 6583 6584 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6585 if (!sched_domains_numa_distance) 6586 return; 6587 6588 /* 6589 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6590 * unique distances in the node_distance() table. 6591 * 6592 * Assumes node_distance(0,j) includes all distances in 6593 * node_distance(i,j) in order to avoid cubic time. 6594 */ 6595 next_distance = curr_distance; 6596 for (i = 0; i < nr_node_ids; i++) { 6597 for (j = 0; j < nr_node_ids; j++) { 6598 for (k = 0; k < nr_node_ids; k++) { 6599 int distance = node_distance(i, k); 6600 6601 if (distance > curr_distance && 6602 (distance < next_distance || 6603 next_distance == curr_distance)) 6604 next_distance = distance; 6605 6606 /* 6607 * While not a strong assumption it would be nice to know 6608 * about cases where if node A is connected to B, B is not 6609 * equally connected to A. 6610 */ 6611 if (sched_debug() && node_distance(k, i) != distance) 6612 sched_numa_warn("Node-distance not symmetric"); 6613 6614 if (sched_debug() && i && !find_numa_distance(distance)) 6615 sched_numa_warn("Node-0 not representative"); 6616 } 6617 if (next_distance != curr_distance) { 6618 sched_domains_numa_distance[level++] = next_distance; 6619 sched_domains_numa_levels = level; 6620 curr_distance = next_distance; 6621 } else break; 6622 } 6623 6624 /* 6625 * In case of sched_debug() we verify the above assumption. 6626 */ 6627 if (!sched_debug()) 6628 break; 6629 } 6630 6631 if (!level) 6632 return; 6633 6634 /* 6635 * 'level' contains the number of unique distances, excluding the 6636 * identity distance node_distance(i,i). 6637 * 6638 * The sched_domains_numa_distance[] array includes the actual distance 6639 * numbers. 6640 */ 6641 6642 /* 6643 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6644 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6645 * the array will contain less then 'level' members. This could be 6646 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6647 * in other functions. 6648 * 6649 * We reset it to 'level' at the end of this function. 6650 */ 6651 sched_domains_numa_levels = 0; 6652 6653 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6654 if (!sched_domains_numa_masks) 6655 return; 6656 6657 /* 6658 * Now for each level, construct a mask per node which contains all 6659 * cpus of nodes that are that many hops away from us. 6660 */ 6661 for (i = 0; i < level; i++) { 6662 sched_domains_numa_masks[i] = 6663 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6664 if (!sched_domains_numa_masks[i]) 6665 return; 6666 6667 for (j = 0; j < nr_node_ids; j++) { 6668 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6669 if (!mask) 6670 return; 6671 6672 sched_domains_numa_masks[i][j] = mask; 6673 6674 for (k = 0; k < nr_node_ids; k++) { 6675 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6676 continue; 6677 6678 cpumask_or(mask, mask, cpumask_of_node(k)); 6679 } 6680 } 6681 } 6682 6683 /* Compute default topology size */ 6684 for (i = 0; sched_domain_topology[i].mask; i++); 6685 6686 tl = kzalloc((i + level + 1) * 6687 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6688 if (!tl) 6689 return; 6690 6691 /* 6692 * Copy the default topology bits.. 6693 */ 6694 for (i = 0; sched_domain_topology[i].mask; i++) 6695 tl[i] = sched_domain_topology[i]; 6696 6697 /* 6698 * .. and append 'j' levels of NUMA goodness. 6699 */ 6700 for (j = 0; j < level; i++, j++) { 6701 tl[i] = (struct sched_domain_topology_level){ 6702 .mask = sd_numa_mask, 6703 .sd_flags = cpu_numa_flags, 6704 .flags = SDTL_OVERLAP, 6705 .numa_level = j, 6706 SD_INIT_NAME(NUMA) 6707 }; 6708 } 6709 6710 sched_domain_topology = tl; 6711 6712 sched_domains_numa_levels = level; 6713 sched_max_numa_distance = sched_domains_numa_distance[level - 1]; 6714 6715 init_numa_topology_type(); 6716 } 6717 6718 static void sched_domains_numa_masks_set(int cpu) 6719 { 6720 int i, j; 6721 int node = cpu_to_node(cpu); 6722 6723 for (i = 0; i < sched_domains_numa_levels; i++) { 6724 for (j = 0; j < nr_node_ids; j++) { 6725 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6726 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6727 } 6728 } 6729 } 6730 6731 static void sched_domains_numa_masks_clear(int cpu) 6732 { 6733 int i, j; 6734 for (i = 0; i < sched_domains_numa_levels; i++) { 6735 for (j = 0; j < nr_node_ids; j++) 6736 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6737 } 6738 } 6739 6740 /* 6741 * Update sched_domains_numa_masks[level][node] array when new cpus 6742 * are onlined. 6743 */ 6744 static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6745 unsigned long action, 6746 void *hcpu) 6747 { 6748 int cpu = (long)hcpu; 6749 6750 switch (action & ~CPU_TASKS_FROZEN) { 6751 case CPU_ONLINE: 6752 sched_domains_numa_masks_set(cpu); 6753 break; 6754 6755 case CPU_DEAD: 6756 sched_domains_numa_masks_clear(cpu); 6757 break; 6758 6759 default: 6760 return NOTIFY_DONE; 6761 } 6762 6763 return NOTIFY_OK; 6764 } 6765 #else 6766 static inline void sched_init_numa(void) 6767 { 6768 } 6769 6770 static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6771 unsigned long action, 6772 void *hcpu) 6773 { 6774 return 0; 6775 } 6776 #endif /* CONFIG_NUMA */ 6777 6778 static int __sdt_alloc(const struct cpumask *cpu_map) 6779 { 6780 struct sched_domain_topology_level *tl; 6781 int j; 6782 6783 for_each_sd_topology(tl) { 6784 struct sd_data *sdd = &tl->data; 6785 6786 sdd->sd = alloc_percpu(struct sched_domain *); 6787 if (!sdd->sd) 6788 return -ENOMEM; 6789 6790 sdd->sg = alloc_percpu(struct sched_group *); 6791 if (!sdd->sg) 6792 return -ENOMEM; 6793 6794 sdd->sgc = alloc_percpu(struct sched_group_capacity *); 6795 if (!sdd->sgc) 6796 return -ENOMEM; 6797 6798 for_each_cpu(j, cpu_map) { 6799 struct sched_domain *sd; 6800 struct sched_group *sg; 6801 struct sched_group_capacity *sgc; 6802 6803 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6804 GFP_KERNEL, cpu_to_node(j)); 6805 if (!sd) 6806 return -ENOMEM; 6807 6808 *per_cpu_ptr(sdd->sd, j) = sd; 6809 6810 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6811 GFP_KERNEL, cpu_to_node(j)); 6812 if (!sg) 6813 return -ENOMEM; 6814 6815 sg->next = sg; 6816 6817 *per_cpu_ptr(sdd->sg, j) = sg; 6818 6819 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), 6820 GFP_KERNEL, cpu_to_node(j)); 6821 if (!sgc) 6822 return -ENOMEM; 6823 6824 *per_cpu_ptr(sdd->sgc, j) = sgc; 6825 } 6826 } 6827 6828 return 0; 6829 } 6830 6831 static void __sdt_free(const struct cpumask *cpu_map) 6832 { 6833 struct sched_domain_topology_level *tl; 6834 int j; 6835 6836 for_each_sd_topology(tl) { 6837 struct sd_data *sdd = &tl->data; 6838 6839 for_each_cpu(j, cpu_map) { 6840 struct sched_domain *sd; 6841 6842 if (sdd->sd) { 6843 sd = *per_cpu_ptr(sdd->sd, j); 6844 if (sd && (sd->flags & SD_OVERLAP)) 6845 free_sched_groups(sd->groups, 0); 6846 kfree(*per_cpu_ptr(sdd->sd, j)); 6847 } 6848 6849 if (sdd->sg) 6850 kfree(*per_cpu_ptr(sdd->sg, j)); 6851 if (sdd->sgc) 6852 kfree(*per_cpu_ptr(sdd->sgc, j)); 6853 } 6854 free_percpu(sdd->sd); 6855 sdd->sd = NULL; 6856 free_percpu(sdd->sg); 6857 sdd->sg = NULL; 6858 free_percpu(sdd->sgc); 6859 sdd->sgc = NULL; 6860 } 6861 } 6862 6863 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6864 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6865 struct sched_domain *child, int cpu) 6866 { 6867 struct sched_domain *sd = sd_init(tl, cpu); 6868 if (!sd) 6869 return child; 6870 6871 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6872 if (child) { 6873 sd->level = child->level + 1; 6874 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6875 child->parent = sd; 6876 sd->child = child; 6877 6878 if (!cpumask_subset(sched_domain_span(child), 6879 sched_domain_span(sd))) { 6880 pr_err("BUG: arch topology borken\n"); 6881 #ifdef CONFIG_SCHED_DEBUG 6882 pr_err(" the %s domain not a subset of the %s domain\n", 6883 child->name, sd->name); 6884 #endif 6885 /* Fixup, ensure @sd has at least @child cpus. */ 6886 cpumask_or(sched_domain_span(sd), 6887 sched_domain_span(sd), 6888 sched_domain_span(child)); 6889 } 6890 6891 } 6892 set_domain_attribute(sd, attr); 6893 6894 return sd; 6895 } 6896 6897 /* 6898 * Build sched domains for a given set of cpus and attach the sched domains 6899 * to the individual cpus 6900 */ 6901 static int build_sched_domains(const struct cpumask *cpu_map, 6902 struct sched_domain_attr *attr) 6903 { 6904 enum s_alloc alloc_state; 6905 struct sched_domain *sd; 6906 struct s_data d; 6907 int i, ret = -ENOMEM; 6908 6909 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6910 if (alloc_state != sa_rootdomain) 6911 goto error; 6912 6913 /* Set up domains for cpus specified by the cpu_map. */ 6914 for_each_cpu(i, cpu_map) { 6915 struct sched_domain_topology_level *tl; 6916 6917 sd = NULL; 6918 for_each_sd_topology(tl) { 6919 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6920 if (tl == sched_domain_topology) 6921 *per_cpu_ptr(d.sd, i) = sd; 6922 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6923 sd->flags |= SD_OVERLAP; 6924 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6925 break; 6926 } 6927 } 6928 6929 /* Build the groups for the domains */ 6930 for_each_cpu(i, cpu_map) { 6931 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6932 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6933 if (sd->flags & SD_OVERLAP) { 6934 if (build_overlap_sched_groups(sd, i)) 6935 goto error; 6936 } else { 6937 if (build_sched_groups(sd, i)) 6938 goto error; 6939 } 6940 } 6941 } 6942 6943 /* Calculate CPU capacity for physical packages and nodes */ 6944 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6945 if (!cpumask_test_cpu(i, cpu_map)) 6946 continue; 6947 6948 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6949 claim_allocations(i, sd); 6950 init_sched_groups_capacity(i, sd); 6951 } 6952 } 6953 6954 /* Attach the domains */ 6955 rcu_read_lock(); 6956 for_each_cpu(i, cpu_map) { 6957 sd = *per_cpu_ptr(d.sd, i); 6958 cpu_attach_domain(sd, d.rd, i); 6959 } 6960 rcu_read_unlock(); 6961 6962 ret = 0; 6963 error: 6964 __free_domain_allocs(&d, alloc_state, cpu_map); 6965 return ret; 6966 } 6967 6968 static cpumask_var_t *doms_cur; /* current sched domains */ 6969 static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6970 static struct sched_domain_attr *dattr_cur; 6971 /* attribues of custom domains in 'doms_cur' */ 6972 6973 /* 6974 * Special case: If a kmalloc of a doms_cur partition (array of 6975 * cpumask) fails, then fallback to a single sched domain, 6976 * as determined by the single cpumask fallback_doms. 6977 */ 6978 static cpumask_var_t fallback_doms; 6979 6980 /* 6981 * arch_update_cpu_topology lets virtualized architectures update the 6982 * cpu core maps. It is supposed to return 1 if the topology changed 6983 * or 0 if it stayed the same. 6984 */ 6985 int __weak arch_update_cpu_topology(void) 6986 { 6987 return 0; 6988 } 6989 6990 cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6991 { 6992 int i; 6993 cpumask_var_t *doms; 6994 6995 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6996 if (!doms) 6997 return NULL; 6998 for (i = 0; i < ndoms; i++) { 6999 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 7000 free_sched_domains(doms, i); 7001 return NULL; 7002 } 7003 } 7004 return doms; 7005 } 7006 7007 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 7008 { 7009 unsigned int i; 7010 for (i = 0; i < ndoms; i++) 7011 free_cpumask_var(doms[i]); 7012 kfree(doms); 7013 } 7014 7015 /* 7016 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 7017 * For now this just excludes isolated cpus, but could be used to 7018 * exclude other special cases in the future. 7019 */ 7020 static int init_sched_domains(const struct cpumask *cpu_map) 7021 { 7022 int err; 7023 7024 arch_update_cpu_topology(); 7025 ndoms_cur = 1; 7026 doms_cur = alloc_sched_domains(ndoms_cur); 7027 if (!doms_cur) 7028 doms_cur = &fallback_doms; 7029 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 7030 err = build_sched_domains(doms_cur[0], NULL); 7031 register_sched_domain_sysctl(); 7032 7033 return err; 7034 } 7035 7036 /* 7037 * Detach sched domains from a group of cpus specified in cpu_map 7038 * These cpus will now be attached to the NULL domain 7039 */ 7040 static void detach_destroy_domains(const struct cpumask *cpu_map) 7041 { 7042 int i; 7043 7044 rcu_read_lock(); 7045 for_each_cpu(i, cpu_map) 7046 cpu_attach_domain(NULL, &def_root_domain, i); 7047 rcu_read_unlock(); 7048 } 7049 7050 /* handle null as "default" */ 7051 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 7052 struct sched_domain_attr *new, int idx_new) 7053 { 7054 struct sched_domain_attr tmp; 7055 7056 /* fast path */ 7057 if (!new && !cur) 7058 return 1; 7059 7060 tmp = SD_ATTR_INIT; 7061 return !memcmp(cur ? (cur + idx_cur) : &tmp, 7062 new ? (new + idx_new) : &tmp, 7063 sizeof(struct sched_domain_attr)); 7064 } 7065 7066 /* 7067 * Partition sched domains as specified by the 'ndoms_new' 7068 * cpumasks in the array doms_new[] of cpumasks. This compares 7069 * doms_new[] to the current sched domain partitioning, doms_cur[]. 7070 * It destroys each deleted domain and builds each new domain. 7071 * 7072 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 7073 * The masks don't intersect (don't overlap.) We should setup one 7074 * sched domain for each mask. CPUs not in any of the cpumasks will 7075 * not be load balanced. If the same cpumask appears both in the 7076 * current 'doms_cur' domains and in the new 'doms_new', we can leave 7077 * it as it is. 7078 * 7079 * The passed in 'doms_new' should be allocated using 7080 * alloc_sched_domains. This routine takes ownership of it and will 7081 * free_sched_domains it when done with it. If the caller failed the 7082 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 7083 * and partition_sched_domains() will fallback to the single partition 7084 * 'fallback_doms', it also forces the domains to be rebuilt. 7085 * 7086 * If doms_new == NULL it will be replaced with cpu_online_mask. 7087 * ndoms_new == 0 is a special case for destroying existing domains, 7088 * and it will not create the default domain. 7089 * 7090 * Call with hotplug lock held 7091 */ 7092 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 7093 struct sched_domain_attr *dattr_new) 7094 { 7095 int i, j, n; 7096 int new_topology; 7097 7098 mutex_lock(&sched_domains_mutex); 7099 7100 /* always unregister in case we don't destroy any domains */ 7101 unregister_sched_domain_sysctl(); 7102 7103 /* Let architecture update cpu core mappings. */ 7104 new_topology = arch_update_cpu_topology(); 7105 7106 n = doms_new ? ndoms_new : 0; 7107 7108 /* Destroy deleted domains */ 7109 for (i = 0; i < ndoms_cur; i++) { 7110 for (j = 0; j < n && !new_topology; j++) { 7111 if (cpumask_equal(doms_cur[i], doms_new[j]) 7112 && dattrs_equal(dattr_cur, i, dattr_new, j)) 7113 goto match1; 7114 } 7115 /* no match - a current sched domain not in new doms_new[] */ 7116 detach_destroy_domains(doms_cur[i]); 7117 match1: 7118 ; 7119 } 7120 7121 n = ndoms_cur; 7122 if (doms_new == NULL) { 7123 n = 0; 7124 doms_new = &fallback_doms; 7125 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 7126 WARN_ON_ONCE(dattr_new); 7127 } 7128 7129 /* Build new domains */ 7130 for (i = 0; i < ndoms_new; i++) { 7131 for (j = 0; j < n && !new_topology; j++) { 7132 if (cpumask_equal(doms_new[i], doms_cur[j]) 7133 && dattrs_equal(dattr_new, i, dattr_cur, j)) 7134 goto match2; 7135 } 7136 /* no match - add a new doms_new */ 7137 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 7138 match2: 7139 ; 7140 } 7141 7142 /* Remember the new sched domains */ 7143 if (doms_cur != &fallback_doms) 7144 free_sched_domains(doms_cur, ndoms_cur); 7145 kfree(dattr_cur); /* kfree(NULL) is safe */ 7146 doms_cur = doms_new; 7147 dattr_cur = dattr_new; 7148 ndoms_cur = ndoms_new; 7149 7150 register_sched_domain_sysctl(); 7151 7152 mutex_unlock(&sched_domains_mutex); 7153 } 7154 7155 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 7156 7157 /* 7158 * Update cpusets according to cpu_active mask. If cpusets are 7159 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 7160 * around partition_sched_domains(). 7161 * 7162 * If we come here as part of a suspend/resume, don't touch cpusets because we 7163 * want to restore it back to its original state upon resume anyway. 7164 */ 7165 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 7166 void *hcpu) 7167 { 7168 switch (action) { 7169 case CPU_ONLINE_FROZEN: 7170 case CPU_DOWN_FAILED_FROZEN: 7171 7172 /* 7173 * num_cpus_frozen tracks how many CPUs are involved in suspend 7174 * resume sequence. As long as this is not the last online 7175 * operation in the resume sequence, just build a single sched 7176 * domain, ignoring cpusets. 7177 */ 7178 num_cpus_frozen--; 7179 if (likely(num_cpus_frozen)) { 7180 partition_sched_domains(1, NULL, NULL); 7181 break; 7182 } 7183 7184 /* 7185 * This is the last CPU online operation. So fall through and 7186 * restore the original sched domains by considering the 7187 * cpuset configurations. 7188 */ 7189 7190 case CPU_ONLINE: 7191 cpuset_update_active_cpus(true); 7192 break; 7193 default: 7194 return NOTIFY_DONE; 7195 } 7196 return NOTIFY_OK; 7197 } 7198 7199 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 7200 void *hcpu) 7201 { 7202 unsigned long flags; 7203 long cpu = (long)hcpu; 7204 struct dl_bw *dl_b; 7205 bool overflow; 7206 int cpus; 7207 7208 switch (action) { 7209 case CPU_DOWN_PREPARE: 7210 rcu_read_lock_sched(); 7211 dl_b = dl_bw_of(cpu); 7212 7213 raw_spin_lock_irqsave(&dl_b->lock, flags); 7214 cpus = dl_bw_cpus(cpu); 7215 overflow = __dl_overflow(dl_b, cpus, 0, 0); 7216 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7217 7218 rcu_read_unlock_sched(); 7219 7220 if (overflow) 7221 return notifier_from_errno(-EBUSY); 7222 cpuset_update_active_cpus(false); 7223 break; 7224 case CPU_DOWN_PREPARE_FROZEN: 7225 num_cpus_frozen++; 7226 partition_sched_domains(1, NULL, NULL); 7227 break; 7228 default: 7229 return NOTIFY_DONE; 7230 } 7231 return NOTIFY_OK; 7232 } 7233 7234 void __init sched_init_smp(void) 7235 { 7236 cpumask_var_t non_isolated_cpus; 7237 7238 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 7239 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 7240 7241 /* nohz_full won't take effect without isolating the cpus. */ 7242 tick_nohz_full_add_cpus_to(cpu_isolated_map); 7243 7244 sched_init_numa(); 7245 7246 /* 7247 * There's no userspace yet to cause hotplug operations; hence all the 7248 * cpu masks are stable and all blatant races in the below code cannot 7249 * happen. 7250 */ 7251 mutex_lock(&sched_domains_mutex); 7252 init_sched_domains(cpu_active_mask); 7253 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 7254 if (cpumask_empty(non_isolated_cpus)) 7255 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 7256 mutex_unlock(&sched_domains_mutex); 7257 7258 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 7259 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 7260 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 7261 7262 init_hrtick(); 7263 7264 /* Move init over to a non-isolated CPU */ 7265 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 7266 BUG(); 7267 sched_init_granularity(); 7268 free_cpumask_var(non_isolated_cpus); 7269 7270 init_sched_rt_class(); 7271 init_sched_dl_class(); 7272 } 7273 #else 7274 void __init sched_init_smp(void) 7275 { 7276 sched_init_granularity(); 7277 } 7278 #endif /* CONFIG_SMP */ 7279 7280 int in_sched_functions(unsigned long addr) 7281 { 7282 return in_lock_functions(addr) || 7283 (addr >= (unsigned long)__sched_text_start 7284 && addr < (unsigned long)__sched_text_end); 7285 } 7286 7287 #ifdef CONFIG_CGROUP_SCHED 7288 /* 7289 * Default task group. 7290 * Every task in system belongs to this group at bootup. 7291 */ 7292 struct task_group root_task_group; 7293 LIST_HEAD(task_groups); 7294 #endif 7295 7296 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 7297 7298 void __init sched_init(void) 7299 { 7300 int i, j; 7301 unsigned long alloc_size = 0, ptr; 7302 7303 #ifdef CONFIG_FAIR_GROUP_SCHED 7304 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 7305 #endif 7306 #ifdef CONFIG_RT_GROUP_SCHED 7307 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 7308 #endif 7309 if (alloc_size) { 7310 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 7311 7312 #ifdef CONFIG_FAIR_GROUP_SCHED 7313 root_task_group.se = (struct sched_entity **)ptr; 7314 ptr += nr_cpu_ids * sizeof(void **); 7315 7316 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 7317 ptr += nr_cpu_ids * sizeof(void **); 7318 7319 #endif /* CONFIG_FAIR_GROUP_SCHED */ 7320 #ifdef CONFIG_RT_GROUP_SCHED 7321 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 7322 ptr += nr_cpu_ids * sizeof(void **); 7323 7324 root_task_group.rt_rq = (struct rt_rq **)ptr; 7325 ptr += nr_cpu_ids * sizeof(void **); 7326 7327 #endif /* CONFIG_RT_GROUP_SCHED */ 7328 } 7329 #ifdef CONFIG_CPUMASK_OFFSTACK 7330 for_each_possible_cpu(i) { 7331 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node( 7332 cpumask_size(), GFP_KERNEL, cpu_to_node(i)); 7333 } 7334 #endif /* CONFIG_CPUMASK_OFFSTACK */ 7335 7336 init_rt_bandwidth(&def_rt_bandwidth, 7337 global_rt_period(), global_rt_runtime()); 7338 init_dl_bandwidth(&def_dl_bandwidth, 7339 global_rt_period(), global_rt_runtime()); 7340 7341 #ifdef CONFIG_SMP 7342 init_defrootdomain(); 7343 #endif 7344 7345 #ifdef CONFIG_RT_GROUP_SCHED 7346 init_rt_bandwidth(&root_task_group.rt_bandwidth, 7347 global_rt_period(), global_rt_runtime()); 7348 #endif /* CONFIG_RT_GROUP_SCHED */ 7349 7350 #ifdef CONFIG_CGROUP_SCHED 7351 list_add(&root_task_group.list, &task_groups); 7352 INIT_LIST_HEAD(&root_task_group.children); 7353 INIT_LIST_HEAD(&root_task_group.siblings); 7354 autogroup_init(&init_task); 7355 7356 #endif /* CONFIG_CGROUP_SCHED */ 7357 7358 for_each_possible_cpu(i) { 7359 struct rq *rq; 7360 7361 rq = cpu_rq(i); 7362 raw_spin_lock_init(&rq->lock); 7363 rq->nr_running = 0; 7364 rq->calc_load_active = 0; 7365 rq->calc_load_update = jiffies + LOAD_FREQ; 7366 init_cfs_rq(&rq->cfs); 7367 init_rt_rq(&rq->rt); 7368 init_dl_rq(&rq->dl); 7369 #ifdef CONFIG_FAIR_GROUP_SCHED 7370 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 7371 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 7372 /* 7373 * How much cpu bandwidth does root_task_group get? 7374 * 7375 * In case of task-groups formed thr' the cgroup filesystem, it 7376 * gets 100% of the cpu resources in the system. This overall 7377 * system cpu resource is divided among the tasks of 7378 * root_task_group and its child task-groups in a fair manner, 7379 * based on each entity's (task or task-group's) weight 7380 * (se->load.weight). 7381 * 7382 * In other words, if root_task_group has 10 tasks of weight 7383 * 1024) and two child groups A0 and A1 (of weight 1024 each), 7384 * then A0's share of the cpu resource is: 7385 * 7386 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 7387 * 7388 * We achieve this by letting root_task_group's tasks sit 7389 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 7390 */ 7391 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 7392 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 7393 #endif /* CONFIG_FAIR_GROUP_SCHED */ 7394 7395 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 7396 #ifdef CONFIG_RT_GROUP_SCHED 7397 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 7398 #endif 7399 7400 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 7401 rq->cpu_load[j] = 0; 7402 7403 rq->last_load_update_tick = jiffies; 7404 7405 #ifdef CONFIG_SMP 7406 rq->sd = NULL; 7407 rq->rd = NULL; 7408 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE; 7409 rq->balance_callback = NULL; 7410 rq->active_balance = 0; 7411 rq->next_balance = jiffies; 7412 rq->push_cpu = 0; 7413 rq->cpu = i; 7414 rq->online = 0; 7415 rq->idle_stamp = 0; 7416 rq->avg_idle = 2*sysctl_sched_migration_cost; 7417 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 7418 7419 INIT_LIST_HEAD(&rq->cfs_tasks); 7420 7421 rq_attach_root(rq, &def_root_domain); 7422 #ifdef CONFIG_NO_HZ_COMMON 7423 rq->nohz_flags = 0; 7424 #endif 7425 #ifdef CONFIG_NO_HZ_FULL 7426 rq->last_sched_tick = 0; 7427 #endif 7428 #endif 7429 init_rq_hrtick(rq); 7430 atomic_set(&rq->nr_iowait, 0); 7431 } 7432 7433 set_load_weight(&init_task); 7434 7435 #ifdef CONFIG_PREEMPT_NOTIFIERS 7436 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7437 #endif 7438 7439 /* 7440 * The boot idle thread does lazy MMU switching as well: 7441 */ 7442 atomic_inc(&init_mm.mm_count); 7443 enter_lazy_tlb(&init_mm, current); 7444 7445 /* 7446 * During early bootup we pretend to be a normal task: 7447 */ 7448 current->sched_class = &fair_sched_class; 7449 7450 /* 7451 * Make us the idle thread. Technically, schedule() should not be 7452 * called from this thread, however somewhere below it might be, 7453 * but because we are the idle thread, we just pick up running again 7454 * when this runqueue becomes "idle". 7455 */ 7456 init_idle(current, smp_processor_id()); 7457 7458 calc_load_update = jiffies + LOAD_FREQ; 7459 7460 #ifdef CONFIG_SMP 7461 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7462 /* May be allocated at isolcpus cmdline parse time */ 7463 if (cpu_isolated_map == NULL) 7464 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7465 idle_thread_set_boot_cpu(); 7466 set_cpu_rq_start_time(); 7467 #endif 7468 init_sched_fair_class(); 7469 7470 scheduler_running = 1; 7471 } 7472 7473 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7474 static inline int preempt_count_equals(int preempt_offset) 7475 { 7476 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7477 7478 return (nested == preempt_offset); 7479 } 7480 7481 void __might_sleep(const char *file, int line, int preempt_offset) 7482 { 7483 /* 7484 * Blocking primitives will set (and therefore destroy) current->state, 7485 * since we will exit with TASK_RUNNING make sure we enter with it, 7486 * otherwise we will destroy state. 7487 */ 7488 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change, 7489 "do not call blocking ops when !TASK_RUNNING; " 7490 "state=%lx set at [<%p>] %pS\n", 7491 current->state, 7492 (void *)current->task_state_change, 7493 (void *)current->task_state_change); 7494 7495 ___might_sleep(file, line, preempt_offset); 7496 } 7497 EXPORT_SYMBOL(__might_sleep); 7498 7499 void ___might_sleep(const char *file, int line, int preempt_offset) 7500 { 7501 static unsigned long prev_jiffy; /* ratelimiting */ 7502 7503 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7504 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 7505 !is_idle_task(current)) || 7506 system_state != SYSTEM_RUNNING || oops_in_progress) 7507 return; 7508 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7509 return; 7510 prev_jiffy = jiffies; 7511 7512 printk(KERN_ERR 7513 "BUG: sleeping function called from invalid context at %s:%d\n", 7514 file, line); 7515 printk(KERN_ERR 7516 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7517 in_atomic(), irqs_disabled(), 7518 current->pid, current->comm); 7519 7520 if (task_stack_end_corrupted(current)) 7521 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); 7522 7523 debug_show_held_locks(current); 7524 if (irqs_disabled()) 7525 print_irqtrace_events(current); 7526 #ifdef CONFIG_DEBUG_PREEMPT 7527 if (!preempt_count_equals(preempt_offset)) { 7528 pr_err("Preemption disabled at:"); 7529 print_ip_sym(current->preempt_disable_ip); 7530 pr_cont("\n"); 7531 } 7532 #endif 7533 dump_stack(); 7534 } 7535 EXPORT_SYMBOL(___might_sleep); 7536 #endif 7537 7538 #ifdef CONFIG_MAGIC_SYSRQ 7539 void normalize_rt_tasks(void) 7540 { 7541 struct task_struct *g, *p; 7542 struct sched_attr attr = { 7543 .sched_policy = SCHED_NORMAL, 7544 }; 7545 7546 read_lock(&tasklist_lock); 7547 for_each_process_thread(g, p) { 7548 /* 7549 * Only normalize user tasks: 7550 */ 7551 if (p->flags & PF_KTHREAD) 7552 continue; 7553 7554 p->se.exec_start = 0; 7555 #ifdef CONFIG_SCHEDSTATS 7556 p->se.statistics.wait_start = 0; 7557 p->se.statistics.sleep_start = 0; 7558 p->se.statistics.block_start = 0; 7559 #endif 7560 7561 if (!dl_task(p) && !rt_task(p)) { 7562 /* 7563 * Renice negative nice level userspace 7564 * tasks back to 0: 7565 */ 7566 if (task_nice(p) < 0) 7567 set_user_nice(p, 0); 7568 continue; 7569 } 7570 7571 __sched_setscheduler(p, &attr, false, false); 7572 } 7573 read_unlock(&tasklist_lock); 7574 } 7575 7576 #endif /* CONFIG_MAGIC_SYSRQ */ 7577 7578 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7579 /* 7580 * These functions are only useful for the IA64 MCA handling, or kdb. 7581 * 7582 * They can only be called when the whole system has been 7583 * stopped - every CPU needs to be quiescent, and no scheduling 7584 * activity can take place. Using them for anything else would 7585 * be a serious bug, and as a result, they aren't even visible 7586 * under any other configuration. 7587 */ 7588 7589 /** 7590 * curr_task - return the current task for a given cpu. 7591 * @cpu: the processor in question. 7592 * 7593 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7594 * 7595 * Return: The current task for @cpu. 7596 */ 7597 struct task_struct *curr_task(int cpu) 7598 { 7599 return cpu_curr(cpu); 7600 } 7601 7602 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7603 7604 #ifdef CONFIG_IA64 7605 /** 7606 * set_curr_task - set the current task for a given cpu. 7607 * @cpu: the processor in question. 7608 * @p: the task pointer to set. 7609 * 7610 * Description: This function must only be used when non-maskable interrupts 7611 * are serviced on a separate stack. It allows the architecture to switch the 7612 * notion of the current task on a cpu in a non-blocking manner. This function 7613 * must be called with all CPU's synchronized, and interrupts disabled, the 7614 * and caller must save the original value of the current task (see 7615 * curr_task() above) and restore that value before reenabling interrupts and 7616 * re-starting the system. 7617 * 7618 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7619 */ 7620 void set_curr_task(int cpu, struct task_struct *p) 7621 { 7622 cpu_curr(cpu) = p; 7623 } 7624 7625 #endif 7626 7627 #ifdef CONFIG_CGROUP_SCHED 7628 /* task_group_lock serializes the addition/removal of task groups */ 7629 static DEFINE_SPINLOCK(task_group_lock); 7630 7631 static void free_sched_group(struct task_group *tg) 7632 { 7633 free_fair_sched_group(tg); 7634 free_rt_sched_group(tg); 7635 autogroup_free(tg); 7636 kfree(tg); 7637 } 7638 7639 /* allocate runqueue etc for a new task group */ 7640 struct task_group *sched_create_group(struct task_group *parent) 7641 { 7642 struct task_group *tg; 7643 7644 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7645 if (!tg) 7646 return ERR_PTR(-ENOMEM); 7647 7648 if (!alloc_fair_sched_group(tg, parent)) 7649 goto err; 7650 7651 if (!alloc_rt_sched_group(tg, parent)) 7652 goto err; 7653 7654 return tg; 7655 7656 err: 7657 free_sched_group(tg); 7658 return ERR_PTR(-ENOMEM); 7659 } 7660 7661 void sched_online_group(struct task_group *tg, struct task_group *parent) 7662 { 7663 unsigned long flags; 7664 7665 spin_lock_irqsave(&task_group_lock, flags); 7666 list_add_rcu(&tg->list, &task_groups); 7667 7668 WARN_ON(!parent); /* root should already exist */ 7669 7670 tg->parent = parent; 7671 INIT_LIST_HEAD(&tg->children); 7672 list_add_rcu(&tg->siblings, &parent->children); 7673 spin_unlock_irqrestore(&task_group_lock, flags); 7674 } 7675 7676 /* rcu callback to free various structures associated with a task group */ 7677 static void free_sched_group_rcu(struct rcu_head *rhp) 7678 { 7679 /* now it should be safe to free those cfs_rqs */ 7680 free_sched_group(container_of(rhp, struct task_group, rcu)); 7681 } 7682 7683 /* Destroy runqueue etc associated with a task group */ 7684 void sched_destroy_group(struct task_group *tg) 7685 { 7686 /* wait for possible concurrent references to cfs_rqs complete */ 7687 call_rcu(&tg->rcu, free_sched_group_rcu); 7688 } 7689 7690 void sched_offline_group(struct task_group *tg) 7691 { 7692 unsigned long flags; 7693 int i; 7694 7695 /* end participation in shares distribution */ 7696 for_each_possible_cpu(i) 7697 unregister_fair_sched_group(tg, i); 7698 7699 spin_lock_irqsave(&task_group_lock, flags); 7700 list_del_rcu(&tg->list); 7701 list_del_rcu(&tg->siblings); 7702 spin_unlock_irqrestore(&task_group_lock, flags); 7703 } 7704 7705 /* change task's runqueue when it moves between groups. 7706 * The caller of this function should have put the task in its new group 7707 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7708 * reflect its new group. 7709 */ 7710 void sched_move_task(struct task_struct *tsk) 7711 { 7712 struct task_group *tg; 7713 int queued, running; 7714 unsigned long flags; 7715 struct rq *rq; 7716 7717 rq = task_rq_lock(tsk, &flags); 7718 7719 running = task_current(rq, tsk); 7720 queued = task_on_rq_queued(tsk); 7721 7722 if (queued) 7723 dequeue_task(rq, tsk, 0); 7724 if (unlikely(running)) 7725 put_prev_task(rq, tsk); 7726 7727 /* 7728 * All callers are synchronized by task_rq_lock(); we do not use RCU 7729 * which is pointless here. Thus, we pass "true" to task_css_check() 7730 * to prevent lockdep warnings. 7731 */ 7732 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true), 7733 struct task_group, css); 7734 tg = autogroup_task_group(tsk, tg); 7735 tsk->sched_task_group = tg; 7736 7737 #ifdef CONFIG_FAIR_GROUP_SCHED 7738 if (tsk->sched_class->task_move_group) 7739 tsk->sched_class->task_move_group(tsk, queued); 7740 else 7741 #endif 7742 set_task_rq(tsk, task_cpu(tsk)); 7743 7744 if (unlikely(running)) 7745 tsk->sched_class->set_curr_task(rq); 7746 if (queued) 7747 enqueue_task(rq, tsk, 0); 7748 7749 task_rq_unlock(rq, tsk, &flags); 7750 } 7751 #endif /* CONFIG_CGROUP_SCHED */ 7752 7753 #ifdef CONFIG_RT_GROUP_SCHED 7754 /* 7755 * Ensure that the real time constraints are schedulable. 7756 */ 7757 static DEFINE_MUTEX(rt_constraints_mutex); 7758 7759 /* Must be called with tasklist_lock held */ 7760 static inline int tg_has_rt_tasks(struct task_group *tg) 7761 { 7762 struct task_struct *g, *p; 7763 7764 /* 7765 * Autogroups do not have RT tasks; see autogroup_create(). 7766 */ 7767 if (task_group_is_autogroup(tg)) 7768 return 0; 7769 7770 for_each_process_thread(g, p) { 7771 if (rt_task(p) && task_group(p) == tg) 7772 return 1; 7773 } 7774 7775 return 0; 7776 } 7777 7778 struct rt_schedulable_data { 7779 struct task_group *tg; 7780 u64 rt_period; 7781 u64 rt_runtime; 7782 }; 7783 7784 static int tg_rt_schedulable(struct task_group *tg, void *data) 7785 { 7786 struct rt_schedulable_data *d = data; 7787 struct task_group *child; 7788 unsigned long total, sum = 0; 7789 u64 period, runtime; 7790 7791 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7792 runtime = tg->rt_bandwidth.rt_runtime; 7793 7794 if (tg == d->tg) { 7795 period = d->rt_period; 7796 runtime = d->rt_runtime; 7797 } 7798 7799 /* 7800 * Cannot have more runtime than the period. 7801 */ 7802 if (runtime > period && runtime != RUNTIME_INF) 7803 return -EINVAL; 7804 7805 /* 7806 * Ensure we don't starve existing RT tasks. 7807 */ 7808 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7809 return -EBUSY; 7810 7811 total = to_ratio(period, runtime); 7812 7813 /* 7814 * Nobody can have more than the global setting allows. 7815 */ 7816 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7817 return -EINVAL; 7818 7819 /* 7820 * The sum of our children's runtime should not exceed our own. 7821 */ 7822 list_for_each_entry_rcu(child, &tg->children, siblings) { 7823 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7824 runtime = child->rt_bandwidth.rt_runtime; 7825 7826 if (child == d->tg) { 7827 period = d->rt_period; 7828 runtime = d->rt_runtime; 7829 } 7830 7831 sum += to_ratio(period, runtime); 7832 } 7833 7834 if (sum > total) 7835 return -EINVAL; 7836 7837 return 0; 7838 } 7839 7840 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7841 { 7842 int ret; 7843 7844 struct rt_schedulable_data data = { 7845 .tg = tg, 7846 .rt_period = period, 7847 .rt_runtime = runtime, 7848 }; 7849 7850 rcu_read_lock(); 7851 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7852 rcu_read_unlock(); 7853 7854 return ret; 7855 } 7856 7857 static int tg_set_rt_bandwidth(struct task_group *tg, 7858 u64 rt_period, u64 rt_runtime) 7859 { 7860 int i, err = 0; 7861 7862 /* 7863 * Disallowing the root group RT runtime is BAD, it would disallow the 7864 * kernel creating (and or operating) RT threads. 7865 */ 7866 if (tg == &root_task_group && rt_runtime == 0) 7867 return -EINVAL; 7868 7869 /* No period doesn't make any sense. */ 7870 if (rt_period == 0) 7871 return -EINVAL; 7872 7873 mutex_lock(&rt_constraints_mutex); 7874 read_lock(&tasklist_lock); 7875 err = __rt_schedulable(tg, rt_period, rt_runtime); 7876 if (err) 7877 goto unlock; 7878 7879 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7880 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7881 tg->rt_bandwidth.rt_runtime = rt_runtime; 7882 7883 for_each_possible_cpu(i) { 7884 struct rt_rq *rt_rq = tg->rt_rq[i]; 7885 7886 raw_spin_lock(&rt_rq->rt_runtime_lock); 7887 rt_rq->rt_runtime = rt_runtime; 7888 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7889 } 7890 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7891 unlock: 7892 read_unlock(&tasklist_lock); 7893 mutex_unlock(&rt_constraints_mutex); 7894 7895 return err; 7896 } 7897 7898 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7899 { 7900 u64 rt_runtime, rt_period; 7901 7902 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7903 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7904 if (rt_runtime_us < 0) 7905 rt_runtime = RUNTIME_INF; 7906 7907 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7908 } 7909 7910 static long sched_group_rt_runtime(struct task_group *tg) 7911 { 7912 u64 rt_runtime_us; 7913 7914 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7915 return -1; 7916 7917 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7918 do_div(rt_runtime_us, NSEC_PER_USEC); 7919 return rt_runtime_us; 7920 } 7921 7922 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us) 7923 { 7924 u64 rt_runtime, rt_period; 7925 7926 rt_period = rt_period_us * NSEC_PER_USEC; 7927 rt_runtime = tg->rt_bandwidth.rt_runtime; 7928 7929 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7930 } 7931 7932 static long sched_group_rt_period(struct task_group *tg) 7933 { 7934 u64 rt_period_us; 7935 7936 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7937 do_div(rt_period_us, NSEC_PER_USEC); 7938 return rt_period_us; 7939 } 7940 #endif /* CONFIG_RT_GROUP_SCHED */ 7941 7942 #ifdef CONFIG_RT_GROUP_SCHED 7943 static int sched_rt_global_constraints(void) 7944 { 7945 int ret = 0; 7946 7947 mutex_lock(&rt_constraints_mutex); 7948 read_lock(&tasklist_lock); 7949 ret = __rt_schedulable(NULL, 0, 0); 7950 read_unlock(&tasklist_lock); 7951 mutex_unlock(&rt_constraints_mutex); 7952 7953 return ret; 7954 } 7955 7956 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7957 { 7958 /* Don't accept realtime tasks when there is no way for them to run */ 7959 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7960 return 0; 7961 7962 return 1; 7963 } 7964 7965 #else /* !CONFIG_RT_GROUP_SCHED */ 7966 static int sched_rt_global_constraints(void) 7967 { 7968 unsigned long flags; 7969 int i, ret = 0; 7970 7971 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7972 for_each_possible_cpu(i) { 7973 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7974 7975 raw_spin_lock(&rt_rq->rt_runtime_lock); 7976 rt_rq->rt_runtime = global_rt_runtime(); 7977 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7978 } 7979 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7980 7981 return ret; 7982 } 7983 #endif /* CONFIG_RT_GROUP_SCHED */ 7984 7985 static int sched_dl_global_validate(void) 7986 { 7987 u64 runtime = global_rt_runtime(); 7988 u64 period = global_rt_period(); 7989 u64 new_bw = to_ratio(period, runtime); 7990 struct dl_bw *dl_b; 7991 int cpu, ret = 0; 7992 unsigned long flags; 7993 7994 /* 7995 * Here we want to check the bandwidth not being set to some 7996 * value smaller than the currently allocated bandwidth in 7997 * any of the root_domains. 7998 * 7999 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 8000 * cycling on root_domains... Discussion on different/better 8001 * solutions is welcome! 8002 */ 8003 for_each_possible_cpu(cpu) { 8004 rcu_read_lock_sched(); 8005 dl_b = dl_bw_of(cpu); 8006 8007 raw_spin_lock_irqsave(&dl_b->lock, flags); 8008 if (new_bw < dl_b->total_bw) 8009 ret = -EBUSY; 8010 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 8011 8012 rcu_read_unlock_sched(); 8013 8014 if (ret) 8015 break; 8016 } 8017 8018 return ret; 8019 } 8020 8021 static void sched_dl_do_global(void) 8022 { 8023 u64 new_bw = -1; 8024 struct dl_bw *dl_b; 8025 int cpu; 8026 unsigned long flags; 8027 8028 def_dl_bandwidth.dl_period = global_rt_period(); 8029 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 8030 8031 if (global_rt_runtime() != RUNTIME_INF) 8032 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 8033 8034 /* 8035 * FIXME: As above... 8036 */ 8037 for_each_possible_cpu(cpu) { 8038 rcu_read_lock_sched(); 8039 dl_b = dl_bw_of(cpu); 8040 8041 raw_spin_lock_irqsave(&dl_b->lock, flags); 8042 dl_b->bw = new_bw; 8043 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 8044 8045 rcu_read_unlock_sched(); 8046 } 8047 } 8048 8049 static int sched_rt_global_validate(void) 8050 { 8051 if (sysctl_sched_rt_period <= 0) 8052 return -EINVAL; 8053 8054 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 8055 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 8056 return -EINVAL; 8057 8058 return 0; 8059 } 8060 8061 static void sched_rt_do_global(void) 8062 { 8063 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 8064 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 8065 } 8066 8067 int sched_rt_handler(struct ctl_table *table, int write, 8068 void __user *buffer, size_t *lenp, 8069 loff_t *ppos) 8070 { 8071 int old_period, old_runtime; 8072 static DEFINE_MUTEX(mutex); 8073 int ret; 8074 8075 mutex_lock(&mutex); 8076 old_period = sysctl_sched_rt_period; 8077 old_runtime = sysctl_sched_rt_runtime; 8078 8079 ret = proc_dointvec(table, write, buffer, lenp, ppos); 8080 8081 if (!ret && write) { 8082 ret = sched_rt_global_validate(); 8083 if (ret) 8084 goto undo; 8085 8086 ret = sched_dl_global_validate(); 8087 if (ret) 8088 goto undo; 8089 8090 ret = sched_rt_global_constraints(); 8091 if (ret) 8092 goto undo; 8093 8094 sched_rt_do_global(); 8095 sched_dl_do_global(); 8096 } 8097 if (0) { 8098 undo: 8099 sysctl_sched_rt_period = old_period; 8100 sysctl_sched_rt_runtime = old_runtime; 8101 } 8102 mutex_unlock(&mutex); 8103 8104 return ret; 8105 } 8106 8107 int sched_rr_handler(struct ctl_table *table, int write, 8108 void __user *buffer, size_t *lenp, 8109 loff_t *ppos) 8110 { 8111 int ret; 8112 static DEFINE_MUTEX(mutex); 8113 8114 mutex_lock(&mutex); 8115 ret = proc_dointvec(table, write, buffer, lenp, ppos); 8116 /* make sure that internally we keep jiffies */ 8117 /* also, writing zero resets timeslice to default */ 8118 if (!ret && write) { 8119 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 8120 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 8121 } 8122 mutex_unlock(&mutex); 8123 return ret; 8124 } 8125 8126 #ifdef CONFIG_CGROUP_SCHED 8127 8128 static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 8129 { 8130 return css ? container_of(css, struct task_group, css) : NULL; 8131 } 8132 8133 static struct cgroup_subsys_state * 8134 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 8135 { 8136 struct task_group *parent = css_tg(parent_css); 8137 struct task_group *tg; 8138 8139 if (!parent) { 8140 /* This is early initialization for the top cgroup */ 8141 return &root_task_group.css; 8142 } 8143 8144 tg = sched_create_group(parent); 8145 if (IS_ERR(tg)) 8146 return ERR_PTR(-ENOMEM); 8147 8148 return &tg->css; 8149 } 8150 8151 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 8152 { 8153 struct task_group *tg = css_tg(css); 8154 struct task_group *parent = css_tg(css->parent); 8155 8156 if (parent) 8157 sched_online_group(tg, parent); 8158 return 0; 8159 } 8160 8161 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 8162 { 8163 struct task_group *tg = css_tg(css); 8164 8165 sched_destroy_group(tg); 8166 } 8167 8168 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 8169 { 8170 struct task_group *tg = css_tg(css); 8171 8172 sched_offline_group(tg); 8173 } 8174 8175 static void cpu_cgroup_fork(struct task_struct *task, void *private) 8176 { 8177 sched_move_task(task); 8178 } 8179 8180 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 8181 struct cgroup_taskset *tset) 8182 { 8183 struct task_struct *task; 8184 8185 cgroup_taskset_for_each(task, tset) { 8186 #ifdef CONFIG_RT_GROUP_SCHED 8187 if (!sched_rt_can_attach(css_tg(css), task)) 8188 return -EINVAL; 8189 #else 8190 /* We don't support RT-tasks being in separate groups */ 8191 if (task->sched_class != &fair_sched_class) 8192 return -EINVAL; 8193 #endif 8194 } 8195 return 0; 8196 } 8197 8198 static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 8199 struct cgroup_taskset *tset) 8200 { 8201 struct task_struct *task; 8202 8203 cgroup_taskset_for_each(task, tset) 8204 sched_move_task(task); 8205 } 8206 8207 static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 8208 struct cgroup_subsys_state *old_css, 8209 struct task_struct *task) 8210 { 8211 /* 8212 * cgroup_exit() is called in the copy_process() failure path. 8213 * Ignore this case since the task hasn't ran yet, this avoids 8214 * trying to poke a half freed task state from generic code. 8215 */ 8216 if (!(task->flags & PF_EXITING)) 8217 return; 8218 8219 sched_move_task(task); 8220 } 8221 8222 #ifdef CONFIG_FAIR_GROUP_SCHED 8223 static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 8224 struct cftype *cftype, u64 shareval) 8225 { 8226 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 8227 } 8228 8229 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 8230 struct cftype *cft) 8231 { 8232 struct task_group *tg = css_tg(css); 8233 8234 return (u64) scale_load_down(tg->shares); 8235 } 8236 8237 #ifdef CONFIG_CFS_BANDWIDTH 8238 static DEFINE_MUTEX(cfs_constraints_mutex); 8239 8240 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 8241 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 8242 8243 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 8244 8245 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 8246 { 8247 int i, ret = 0, runtime_enabled, runtime_was_enabled; 8248 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8249 8250 if (tg == &root_task_group) 8251 return -EINVAL; 8252 8253 /* 8254 * Ensure we have at some amount of bandwidth every period. This is 8255 * to prevent reaching a state of large arrears when throttled via 8256 * entity_tick() resulting in prolonged exit starvation. 8257 */ 8258 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 8259 return -EINVAL; 8260 8261 /* 8262 * Likewise, bound things on the otherside by preventing insane quota 8263 * periods. This also allows us to normalize in computing quota 8264 * feasibility. 8265 */ 8266 if (period > max_cfs_quota_period) 8267 return -EINVAL; 8268 8269 /* 8270 * Prevent race between setting of cfs_rq->runtime_enabled and 8271 * unthrottle_offline_cfs_rqs(). 8272 */ 8273 get_online_cpus(); 8274 mutex_lock(&cfs_constraints_mutex); 8275 ret = __cfs_schedulable(tg, period, quota); 8276 if (ret) 8277 goto out_unlock; 8278 8279 runtime_enabled = quota != RUNTIME_INF; 8280 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 8281 /* 8282 * If we need to toggle cfs_bandwidth_used, off->on must occur 8283 * before making related changes, and on->off must occur afterwards 8284 */ 8285 if (runtime_enabled && !runtime_was_enabled) 8286 cfs_bandwidth_usage_inc(); 8287 raw_spin_lock_irq(&cfs_b->lock); 8288 cfs_b->period = ns_to_ktime(period); 8289 cfs_b->quota = quota; 8290 8291 __refill_cfs_bandwidth_runtime(cfs_b); 8292 /* restart the period timer (if active) to handle new period expiry */ 8293 if (runtime_enabled) 8294 start_cfs_bandwidth(cfs_b); 8295 raw_spin_unlock_irq(&cfs_b->lock); 8296 8297 for_each_online_cpu(i) { 8298 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 8299 struct rq *rq = cfs_rq->rq; 8300 8301 raw_spin_lock_irq(&rq->lock); 8302 cfs_rq->runtime_enabled = runtime_enabled; 8303 cfs_rq->runtime_remaining = 0; 8304 8305 if (cfs_rq->throttled) 8306 unthrottle_cfs_rq(cfs_rq); 8307 raw_spin_unlock_irq(&rq->lock); 8308 } 8309 if (runtime_was_enabled && !runtime_enabled) 8310 cfs_bandwidth_usage_dec(); 8311 out_unlock: 8312 mutex_unlock(&cfs_constraints_mutex); 8313 put_online_cpus(); 8314 8315 return ret; 8316 } 8317 8318 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 8319 { 8320 u64 quota, period; 8321 8322 period = ktime_to_ns(tg->cfs_bandwidth.period); 8323 if (cfs_quota_us < 0) 8324 quota = RUNTIME_INF; 8325 else 8326 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 8327 8328 return tg_set_cfs_bandwidth(tg, period, quota); 8329 } 8330 8331 long tg_get_cfs_quota(struct task_group *tg) 8332 { 8333 u64 quota_us; 8334 8335 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 8336 return -1; 8337 8338 quota_us = tg->cfs_bandwidth.quota; 8339 do_div(quota_us, NSEC_PER_USEC); 8340 8341 return quota_us; 8342 } 8343 8344 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 8345 { 8346 u64 quota, period; 8347 8348 period = (u64)cfs_period_us * NSEC_PER_USEC; 8349 quota = tg->cfs_bandwidth.quota; 8350 8351 return tg_set_cfs_bandwidth(tg, period, quota); 8352 } 8353 8354 long tg_get_cfs_period(struct task_group *tg) 8355 { 8356 u64 cfs_period_us; 8357 8358 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 8359 do_div(cfs_period_us, NSEC_PER_USEC); 8360 8361 return cfs_period_us; 8362 } 8363 8364 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 8365 struct cftype *cft) 8366 { 8367 return tg_get_cfs_quota(css_tg(css)); 8368 } 8369 8370 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 8371 struct cftype *cftype, s64 cfs_quota_us) 8372 { 8373 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 8374 } 8375 8376 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 8377 struct cftype *cft) 8378 { 8379 return tg_get_cfs_period(css_tg(css)); 8380 } 8381 8382 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 8383 struct cftype *cftype, u64 cfs_period_us) 8384 { 8385 return tg_set_cfs_period(css_tg(css), cfs_period_us); 8386 } 8387 8388 struct cfs_schedulable_data { 8389 struct task_group *tg; 8390 u64 period, quota; 8391 }; 8392 8393 /* 8394 * normalize group quota/period to be quota/max_period 8395 * note: units are usecs 8396 */ 8397 static u64 normalize_cfs_quota(struct task_group *tg, 8398 struct cfs_schedulable_data *d) 8399 { 8400 u64 quota, period; 8401 8402 if (tg == d->tg) { 8403 period = d->period; 8404 quota = d->quota; 8405 } else { 8406 period = tg_get_cfs_period(tg); 8407 quota = tg_get_cfs_quota(tg); 8408 } 8409 8410 /* note: these should typically be equivalent */ 8411 if (quota == RUNTIME_INF || quota == -1) 8412 return RUNTIME_INF; 8413 8414 return to_ratio(period, quota); 8415 } 8416 8417 static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 8418 { 8419 struct cfs_schedulable_data *d = data; 8420 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8421 s64 quota = 0, parent_quota = -1; 8422 8423 if (!tg->parent) { 8424 quota = RUNTIME_INF; 8425 } else { 8426 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 8427 8428 quota = normalize_cfs_quota(tg, d); 8429 parent_quota = parent_b->hierarchical_quota; 8430 8431 /* 8432 * ensure max(child_quota) <= parent_quota, inherit when no 8433 * limit is set 8434 */ 8435 if (quota == RUNTIME_INF) 8436 quota = parent_quota; 8437 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 8438 return -EINVAL; 8439 } 8440 cfs_b->hierarchical_quota = quota; 8441 8442 return 0; 8443 } 8444 8445 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 8446 { 8447 int ret; 8448 struct cfs_schedulable_data data = { 8449 .tg = tg, 8450 .period = period, 8451 .quota = quota, 8452 }; 8453 8454 if (quota != RUNTIME_INF) { 8455 do_div(data.period, NSEC_PER_USEC); 8456 do_div(data.quota, NSEC_PER_USEC); 8457 } 8458 8459 rcu_read_lock(); 8460 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 8461 rcu_read_unlock(); 8462 8463 return ret; 8464 } 8465 8466 static int cpu_stats_show(struct seq_file *sf, void *v) 8467 { 8468 struct task_group *tg = css_tg(seq_css(sf)); 8469 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8470 8471 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 8472 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 8473 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 8474 8475 return 0; 8476 } 8477 #endif /* CONFIG_CFS_BANDWIDTH */ 8478 #endif /* CONFIG_FAIR_GROUP_SCHED */ 8479 8480 #ifdef CONFIG_RT_GROUP_SCHED 8481 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 8482 struct cftype *cft, s64 val) 8483 { 8484 return sched_group_set_rt_runtime(css_tg(css), val); 8485 } 8486 8487 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 8488 struct cftype *cft) 8489 { 8490 return sched_group_rt_runtime(css_tg(css)); 8491 } 8492 8493 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 8494 struct cftype *cftype, u64 rt_period_us) 8495 { 8496 return sched_group_set_rt_period(css_tg(css), rt_period_us); 8497 } 8498 8499 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 8500 struct cftype *cft) 8501 { 8502 return sched_group_rt_period(css_tg(css)); 8503 } 8504 #endif /* CONFIG_RT_GROUP_SCHED */ 8505 8506 static struct cftype cpu_files[] = { 8507 #ifdef CONFIG_FAIR_GROUP_SCHED 8508 { 8509 .name = "shares", 8510 .read_u64 = cpu_shares_read_u64, 8511 .write_u64 = cpu_shares_write_u64, 8512 }, 8513 #endif 8514 #ifdef CONFIG_CFS_BANDWIDTH 8515 { 8516 .name = "cfs_quota_us", 8517 .read_s64 = cpu_cfs_quota_read_s64, 8518 .write_s64 = cpu_cfs_quota_write_s64, 8519 }, 8520 { 8521 .name = "cfs_period_us", 8522 .read_u64 = cpu_cfs_period_read_u64, 8523 .write_u64 = cpu_cfs_period_write_u64, 8524 }, 8525 { 8526 .name = "stat", 8527 .seq_show = cpu_stats_show, 8528 }, 8529 #endif 8530 #ifdef CONFIG_RT_GROUP_SCHED 8531 { 8532 .name = "rt_runtime_us", 8533 .read_s64 = cpu_rt_runtime_read, 8534 .write_s64 = cpu_rt_runtime_write, 8535 }, 8536 { 8537 .name = "rt_period_us", 8538 .read_u64 = cpu_rt_period_read_uint, 8539 .write_u64 = cpu_rt_period_write_uint, 8540 }, 8541 #endif 8542 { } /* terminate */ 8543 }; 8544 8545 struct cgroup_subsys cpu_cgrp_subsys = { 8546 .css_alloc = cpu_cgroup_css_alloc, 8547 .css_free = cpu_cgroup_css_free, 8548 .css_online = cpu_cgroup_css_online, 8549 .css_offline = cpu_cgroup_css_offline, 8550 .fork = cpu_cgroup_fork, 8551 .can_attach = cpu_cgroup_can_attach, 8552 .attach = cpu_cgroup_attach, 8553 .exit = cpu_cgroup_exit, 8554 .legacy_cftypes = cpu_files, 8555 .early_init = 1, 8556 }; 8557 8558 #endif /* CONFIG_CGROUP_SCHED */ 8559 8560 void dump_cpu_task(int cpu) 8561 { 8562 pr_info("Task dump for CPU %d:\n", cpu); 8563 sched_show_task(cpu_curr(cpu)); 8564 } 8565