| 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 | |
| 76 | #include <asm/switch_to.h> |
| 77 | #include <asm/tlb.h> |
| 78 | #include <asm/irq_regs.h> |
| 79 | #include <asm/mutex.h> |
| 80 | #ifdef CONFIG_PARAVIRT |
| 81 | #include <asm/paravirt.h> |
| 82 | #endif |
| 83 | |
| 84 | #include "sched.h" |
| 85 | #include "../workqueue_sched.h" |
| 86 | #include "../smpboot.h" |
| 87 | |
| 88 | #define CREATE_TRACE_POINTS |
| 89 | #include <trace/events/sched.h> |
| 90 | |
| 91 | void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) |
| 92 | { |
| 93 | unsigned long delta; |
| 94 | ktime_t soft, hard, now; |
| 95 | |
| 96 | for (;;) { |
| 97 | if (hrtimer_active(period_timer)) |
| 98 | break; |
| 99 | |
| 100 | now = hrtimer_cb_get_time(period_timer); |
| 101 | hrtimer_forward(period_timer, now, period); |
| 102 | |
| 103 | soft = hrtimer_get_softexpires(period_timer); |
| 104 | hard = hrtimer_get_expires(period_timer); |
| 105 | delta = ktime_to_ns(ktime_sub(hard, soft)); |
| 106 | __hrtimer_start_range_ns(period_timer, soft, delta, |
| 107 | HRTIMER_MODE_ABS_PINNED, 0); |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | DEFINE_MUTEX(sched_domains_mutex); |
| 112 | DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); |
| 113 | |
| 114 | static void update_rq_clock_task(struct rq *rq, s64 delta); |
| 115 | |
| 116 | void update_rq_clock(struct rq *rq) |
| 117 | { |
| 118 | s64 delta; |
| 119 | |
| 120 | if (rq->skip_clock_update > 0) |
| 121 | return; |
| 122 | |
| 123 | delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; |
| 124 | rq->clock += delta; |
| 125 | update_rq_clock_task(rq, delta); |
| 126 | } |
| 127 | |
| 128 | /* |
| 129 | * Debugging: various feature bits |
| 130 | */ |
| 131 | |
| 132 | #define SCHED_FEAT(name, enabled) \ |
| 133 | (1UL << __SCHED_FEAT_##name) * enabled | |
| 134 | |
| 135 | const_debug unsigned int sysctl_sched_features = |
| 136 | #include "features.h" |
| 137 | 0; |
| 138 | |
| 139 | #undef SCHED_FEAT |
| 140 | |
| 141 | #ifdef CONFIG_SCHED_DEBUG |
| 142 | #define SCHED_FEAT(name, enabled) \ |
| 143 | #name , |
| 144 | |
| 145 | static const char * const sched_feat_names[] = { |
| 146 | #include "features.h" |
| 147 | }; |
| 148 | |
| 149 | #undef SCHED_FEAT |
| 150 | |
| 151 | static int sched_feat_show(struct seq_file *m, void *v) |
| 152 | { |
| 153 | int i; |
| 154 | |
| 155 | for (i = 0; i < __SCHED_FEAT_NR; i++) { |
| 156 | if (!(sysctl_sched_features & (1UL << i))) |
| 157 | seq_puts(m, "NO_"); |
| 158 | seq_printf(m, "%s ", sched_feat_names[i]); |
| 159 | } |
| 160 | seq_puts(m, "\n"); |
| 161 | |
| 162 | return 0; |
| 163 | } |
| 164 | |
| 165 | #ifdef HAVE_JUMP_LABEL |
| 166 | |
| 167 | #define jump_label_key__true STATIC_KEY_INIT_TRUE |
| 168 | #define jump_label_key__false STATIC_KEY_INIT_FALSE |
| 169 | |
| 170 | #define SCHED_FEAT(name, enabled) \ |
| 171 | jump_label_key__##enabled , |
| 172 | |
| 173 | struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { |
| 174 | #include "features.h" |
| 175 | }; |
| 176 | |
| 177 | #undef SCHED_FEAT |
| 178 | |
| 179 | static void sched_feat_disable(int i) |
| 180 | { |
| 181 | if (static_key_enabled(&sched_feat_keys[i])) |
| 182 | static_key_slow_dec(&sched_feat_keys[i]); |
| 183 | } |
| 184 | |
| 185 | static void sched_feat_enable(int i) |
| 186 | { |
| 187 | if (!static_key_enabled(&sched_feat_keys[i])) |
| 188 | static_key_slow_inc(&sched_feat_keys[i]); |
| 189 | } |
| 190 | #else |
| 191 | static void sched_feat_disable(int i) { }; |
| 192 | static void sched_feat_enable(int i) { }; |
| 193 | #endif /* HAVE_JUMP_LABEL */ |
| 194 | |
| 195 | static ssize_t |
| 196 | sched_feat_write(struct file *filp, const char __user *ubuf, |
| 197 | size_t cnt, loff_t *ppos) |
| 198 | { |
| 199 | char buf[64]; |
| 200 | char *cmp; |
| 201 | int neg = 0; |
| 202 | int i; |
| 203 | |
| 204 | if (cnt > 63) |
| 205 | cnt = 63; |
| 206 | |
| 207 | if (copy_from_user(&buf, ubuf, cnt)) |
| 208 | return -EFAULT; |
| 209 | |
| 210 | buf[cnt] = 0; |
| 211 | cmp = strstrip(buf); |
| 212 | |
| 213 | if (strncmp(cmp, "NO_", 3) == 0) { |
| 214 | neg = 1; |
| 215 | cmp += 3; |
| 216 | } |
| 217 | |
| 218 | for (i = 0; i < __SCHED_FEAT_NR; i++) { |
| 219 | if (strcmp(cmp, sched_feat_names[i]) == 0) { |
| 220 | if (neg) { |
| 221 | sysctl_sched_features &= ~(1UL << i); |
| 222 | sched_feat_disable(i); |
| 223 | } else { |
| 224 | sysctl_sched_features |= (1UL << i); |
| 225 | sched_feat_enable(i); |
| 226 | } |
| 227 | break; |
| 228 | } |
| 229 | } |
| 230 | |
| 231 | if (i == __SCHED_FEAT_NR) |
| 232 | return -EINVAL; |
| 233 | |
| 234 | *ppos += cnt; |
| 235 | |
| 236 | return cnt; |
| 237 | } |
| 238 | |
| 239 | static int sched_feat_open(struct inode *inode, struct file *filp) |
| 240 | { |
| 241 | return single_open(filp, sched_feat_show, NULL); |
| 242 | } |
| 243 | |
| 244 | static const struct file_operations sched_feat_fops = { |
| 245 | .open = sched_feat_open, |
| 246 | .write = sched_feat_write, |
| 247 | .read = seq_read, |
| 248 | .llseek = seq_lseek, |
| 249 | .release = single_release, |
| 250 | }; |
| 251 | |
| 252 | static __init int sched_init_debug(void) |
| 253 | { |
| 254 | debugfs_create_file("sched_features", 0644, NULL, NULL, |
| 255 | &sched_feat_fops); |
| 256 | |
| 257 | return 0; |
| 258 | } |
| 259 | late_initcall(sched_init_debug); |
| 260 | #endif /* CONFIG_SCHED_DEBUG */ |
| 261 | |
| 262 | /* |
| 263 | * Number of tasks to iterate in a single balance run. |
| 264 | * Limited because this is done with IRQs disabled. |
| 265 | */ |
| 266 | const_debug unsigned int sysctl_sched_nr_migrate = 32; |
| 267 | |
| 268 | /* |
| 269 | * period over which we average the RT time consumption, measured |
| 270 | * in ms. |
| 271 | * |
| 272 | * default: 1s |
| 273 | */ |
| 274 | const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; |
| 275 | |
| 276 | /* |
| 277 | * period over which we measure -rt task cpu usage in us. |
| 278 | * default: 1s |
| 279 | */ |
| 280 | unsigned int sysctl_sched_rt_period = 1000000; |
| 281 | |
| 282 | __read_mostly int scheduler_running; |
| 283 | |
| 284 | /* |
| 285 | * part of the period that we allow rt tasks to run in us. |
| 286 | * default: 0.95s |
| 287 | */ |
| 288 | int sysctl_sched_rt_runtime = 950000; |
| 289 | |
| 290 | |
| 291 | |
| 292 | /* |
| 293 | * __task_rq_lock - lock the rq @p resides on. |
| 294 | */ |
| 295 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
| 296 | __acquires(rq->lock) |
| 297 | { |
| 298 | struct rq *rq; |
| 299 | |
| 300 | lockdep_assert_held(&p->pi_lock); |
| 301 | |
| 302 | for (;;) { |
| 303 | rq = task_rq(p); |
| 304 | raw_spin_lock(&rq->lock); |
| 305 | if (likely(rq == task_rq(p))) |
| 306 | return rq; |
| 307 | raw_spin_unlock(&rq->lock); |
| 308 | } |
| 309 | } |
| 310 | |
| 311 | /* |
| 312 | * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. |
| 313 | */ |
| 314 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
| 315 | __acquires(p->pi_lock) |
| 316 | __acquires(rq->lock) |
| 317 | { |
| 318 | struct rq *rq; |
| 319 | |
| 320 | for (;;) { |
| 321 | raw_spin_lock_irqsave(&p->pi_lock, *flags); |
| 322 | rq = task_rq(p); |
| 323 | raw_spin_lock(&rq->lock); |
| 324 | if (likely(rq == task_rq(p))) |
| 325 | return rq; |
| 326 | raw_spin_unlock(&rq->lock); |
| 327 | raw_spin_unlock_irqrestore(&p->pi_lock, *flags); |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | static void __task_rq_unlock(struct rq *rq) |
| 332 | __releases(rq->lock) |
| 333 | { |
| 334 | raw_spin_unlock(&rq->lock); |
| 335 | } |
| 336 | |
| 337 | static inline void |
| 338 | task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) |
| 339 | __releases(rq->lock) |
| 340 | __releases(p->pi_lock) |
| 341 | { |
| 342 | raw_spin_unlock(&rq->lock); |
| 343 | raw_spin_unlock_irqrestore(&p->pi_lock, *flags); |
| 344 | } |
| 345 | |
| 346 | /* |
| 347 | * this_rq_lock - lock this runqueue and disable interrupts. |
| 348 | */ |
| 349 | static struct rq *this_rq_lock(void) |
| 350 | __acquires(rq->lock) |
| 351 | { |
| 352 | struct rq *rq; |
| 353 | |
| 354 | local_irq_disable(); |
| 355 | rq = this_rq(); |
| 356 | raw_spin_lock(&rq->lock); |
| 357 | |
| 358 | return rq; |
| 359 | } |
| 360 | |
| 361 | #ifdef CONFIG_SCHED_HRTICK |
| 362 | /* |
| 363 | * Use HR-timers to deliver accurate preemption points. |
| 364 | * |
| 365 | * Its all a bit involved since we cannot program an hrt while holding the |
| 366 | * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a |
| 367 | * reschedule event. |
| 368 | * |
| 369 | * When we get rescheduled we reprogram the hrtick_timer outside of the |
| 370 | * rq->lock. |
| 371 | */ |
| 372 | |
| 373 | static void hrtick_clear(struct rq *rq) |
| 374 | { |
| 375 | if (hrtimer_active(&rq->hrtick_timer)) |
| 376 | hrtimer_cancel(&rq->hrtick_timer); |
| 377 | } |
| 378 | |
| 379 | /* |
| 380 | * High-resolution timer tick. |
| 381 | * Runs from hardirq context with interrupts disabled. |
| 382 | */ |
| 383 | static enum hrtimer_restart hrtick(struct hrtimer *timer) |
| 384 | { |
| 385 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); |
| 386 | |
| 387 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); |
| 388 | |
| 389 | raw_spin_lock(&rq->lock); |
| 390 | update_rq_clock(rq); |
| 391 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); |
| 392 | raw_spin_unlock(&rq->lock); |
| 393 | |
| 394 | return HRTIMER_NORESTART; |
| 395 | } |
| 396 | |
| 397 | #ifdef CONFIG_SMP |
| 398 | /* |
| 399 | * called from hardirq (IPI) context |
| 400 | */ |
| 401 | static void __hrtick_start(void *arg) |
| 402 | { |
| 403 | struct rq *rq = arg; |
| 404 | |
| 405 | raw_spin_lock(&rq->lock); |
| 406 | hrtimer_restart(&rq->hrtick_timer); |
| 407 | rq->hrtick_csd_pending = 0; |
| 408 | raw_spin_unlock(&rq->lock); |
| 409 | } |
| 410 | |
| 411 | /* |
| 412 | * Called to set the hrtick timer state. |
| 413 | * |
| 414 | * called with rq->lock held and irqs disabled |
| 415 | */ |
| 416 | void hrtick_start(struct rq *rq, u64 delay) |
| 417 | { |
| 418 | struct hrtimer *timer = &rq->hrtick_timer; |
| 419 | ktime_t time = ktime_add_ns(timer->base->get_time(), delay); |
| 420 | |
| 421 | hrtimer_set_expires(timer, time); |
| 422 | |
| 423 | if (rq == this_rq()) { |
| 424 | hrtimer_restart(timer); |
| 425 | } else if (!rq->hrtick_csd_pending) { |
| 426 | __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); |
| 427 | rq->hrtick_csd_pending = 1; |
| 428 | } |
| 429 | } |
| 430 | |
| 431 | static int |
| 432 | hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) |
| 433 | { |
| 434 | int cpu = (int)(long)hcpu; |
| 435 | |
| 436 | switch (action) { |
| 437 | case CPU_UP_CANCELED: |
| 438 | case CPU_UP_CANCELED_FROZEN: |
| 439 | case CPU_DOWN_PREPARE: |
| 440 | case CPU_DOWN_PREPARE_FROZEN: |
| 441 | case CPU_DEAD: |
| 442 | case CPU_DEAD_FROZEN: |
| 443 | hrtick_clear(cpu_rq(cpu)); |
| 444 | return NOTIFY_OK; |
| 445 | } |
| 446 | |
| 447 | return NOTIFY_DONE; |
| 448 | } |
| 449 | |
| 450 | static __init void init_hrtick(void) |
| 451 | { |
| 452 | hotcpu_notifier(hotplug_hrtick, 0); |
| 453 | } |
| 454 | #else |
| 455 | /* |
| 456 | * Called to set the hrtick timer state. |
| 457 | * |
| 458 | * called with rq->lock held and irqs disabled |
| 459 | */ |
| 460 | void hrtick_start(struct rq *rq, u64 delay) |
| 461 | { |
| 462 | __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, |
| 463 | HRTIMER_MODE_REL_PINNED, 0); |
| 464 | } |
| 465 | |
| 466 | static inline void init_hrtick(void) |
| 467 | { |
| 468 | } |
| 469 | #endif /* CONFIG_SMP */ |
| 470 | |
| 471 | static void init_rq_hrtick(struct rq *rq) |
| 472 | { |
| 473 | #ifdef CONFIG_SMP |
| 474 | rq->hrtick_csd_pending = 0; |
| 475 | |
| 476 | rq->hrtick_csd.flags = 0; |
| 477 | rq->hrtick_csd.func = __hrtick_start; |
| 478 | rq->hrtick_csd.info = rq; |
| 479 | #endif |
| 480 | |
| 481 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
| 482 | rq->hrtick_timer.function = hrtick; |
| 483 | } |
| 484 | #else /* CONFIG_SCHED_HRTICK */ |
| 485 | static inline void hrtick_clear(struct rq *rq) |
| 486 | { |
| 487 | } |
| 488 | |
| 489 | static inline void init_rq_hrtick(struct rq *rq) |
| 490 | { |
| 491 | } |
| 492 | |
| 493 | static inline void init_hrtick(void) |
| 494 | { |
| 495 | } |
| 496 | #endif /* CONFIG_SCHED_HRTICK */ |
| 497 | |
| 498 | /* |
| 499 | * resched_task - mark a task 'to be rescheduled now'. |
| 500 | * |
| 501 | * On UP this means the setting of the need_resched flag, on SMP it |
| 502 | * might also involve a cross-CPU call to trigger the scheduler on |
| 503 | * the target CPU. |
| 504 | */ |
| 505 | #ifdef CONFIG_SMP |
| 506 | |
| 507 | #ifndef tsk_is_polling |
| 508 | #define tsk_is_polling(t) 0 |
| 509 | #endif |
| 510 | |
| 511 | void resched_task(struct task_struct *p) |
| 512 | { |
| 513 | int cpu; |
| 514 | |
| 515 | assert_raw_spin_locked(&task_rq(p)->lock); |
| 516 | |
| 517 | if (test_tsk_need_resched(p)) |
| 518 | return; |
| 519 | |
| 520 | set_tsk_need_resched(p); |
| 521 | |
| 522 | cpu = task_cpu(p); |
| 523 | if (cpu == smp_processor_id()) |
| 524 | return; |
| 525 | |
| 526 | /* NEED_RESCHED must be visible before we test polling */ |
| 527 | smp_mb(); |
| 528 | if (!tsk_is_polling(p)) |
| 529 | smp_send_reschedule(cpu); |
| 530 | } |
| 531 | |
| 532 | void resched_cpu(int cpu) |
| 533 | { |
| 534 | struct rq *rq = cpu_rq(cpu); |
| 535 | unsigned long flags; |
| 536 | |
| 537 | if (!raw_spin_trylock_irqsave(&rq->lock, flags)) |
| 538 | return; |
| 539 | resched_task(cpu_curr(cpu)); |
| 540 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
| 541 | } |
| 542 | |
| 543 | #ifdef CONFIG_NO_HZ |
| 544 | /* |
| 545 | * In the semi idle case, use the nearest busy cpu for migrating timers |
| 546 | * from an idle cpu. This is good for power-savings. |
| 547 | * |
| 548 | * We don't do similar optimization for completely idle system, as |
| 549 | * selecting an idle cpu will add more delays to the timers than intended |
| 550 | * (as that cpu's timer base may not be uptodate wrt jiffies etc). |
| 551 | */ |
| 552 | int get_nohz_timer_target(void) |
| 553 | { |
| 554 | int cpu = smp_processor_id(); |
| 555 | int i; |
| 556 | struct sched_domain *sd; |
| 557 | |
| 558 | rcu_read_lock(); |
| 559 | for_each_domain(cpu, sd) { |
| 560 | for_each_cpu(i, sched_domain_span(sd)) { |
| 561 | if (!idle_cpu(i)) { |
| 562 | cpu = i; |
| 563 | goto unlock; |
| 564 | } |
| 565 | } |
| 566 | } |
| 567 | unlock: |
| 568 | rcu_read_unlock(); |
| 569 | return cpu; |
| 570 | } |
| 571 | /* |
| 572 | * When add_timer_on() enqueues a timer into the timer wheel of an |
| 573 | * idle CPU then this timer might expire before the next timer event |
| 574 | * which is scheduled to wake up that CPU. In case of a completely |
| 575 | * idle system the next event might even be infinite time into the |
| 576 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and |
| 577 | * leaves the inner idle loop so the newly added timer is taken into |
| 578 | * account when the CPU goes back to idle and evaluates the timer |
| 579 | * wheel for the next timer event. |
| 580 | */ |
| 581 | void wake_up_idle_cpu(int cpu) |
| 582 | { |
| 583 | struct rq *rq = cpu_rq(cpu); |
| 584 | |
| 585 | if (cpu == smp_processor_id()) |
| 586 | return; |
| 587 | |
| 588 | /* |
| 589 | * This is safe, as this function is called with the timer |
| 590 | * wheel base lock of (cpu) held. When the CPU is on the way |
| 591 | * to idle and has not yet set rq->curr to idle then it will |
| 592 | * be serialized on the timer wheel base lock and take the new |
| 593 | * timer into account automatically. |
| 594 | */ |
| 595 | if (rq->curr != rq->idle) |
| 596 | return; |
| 597 | |
| 598 | /* |
| 599 | * We can set TIF_RESCHED on the idle task of the other CPU |
| 600 | * lockless. The worst case is that the other CPU runs the |
| 601 | * idle task through an additional NOOP schedule() |
| 602 | */ |
| 603 | set_tsk_need_resched(rq->idle); |
| 604 | |
| 605 | /* NEED_RESCHED must be visible before we test polling */ |
| 606 | smp_mb(); |
| 607 | if (!tsk_is_polling(rq->idle)) |
| 608 | smp_send_reschedule(cpu); |
| 609 | } |
| 610 | |
| 611 | static inline bool got_nohz_idle_kick(void) |
| 612 | { |
| 613 | int cpu = smp_processor_id(); |
| 614 | return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); |
| 615 | } |
| 616 | |
| 617 | #else /* CONFIG_NO_HZ */ |
| 618 | |
| 619 | static inline bool got_nohz_idle_kick(void) |
| 620 | { |
| 621 | return false; |
| 622 | } |
| 623 | |
| 624 | #endif /* CONFIG_NO_HZ */ |
| 625 | |
| 626 | void sched_avg_update(struct rq *rq) |
| 627 | { |
| 628 | s64 period = sched_avg_period(); |
| 629 | |
| 630 | while ((s64)(rq->clock - rq->age_stamp) > period) { |
| 631 | /* |
| 632 | * Inline assembly required to prevent the compiler |
| 633 | * optimising this loop into a divmod call. |
| 634 | * See __iter_div_u64_rem() for another example of this. |
| 635 | */ |
| 636 | asm("" : "+rm" (rq->age_stamp)); |
| 637 | rq->age_stamp += period; |
| 638 | rq->rt_avg /= 2; |
| 639 | } |
| 640 | } |
| 641 | |
| 642 | #else /* !CONFIG_SMP */ |
| 643 | void resched_task(struct task_struct *p) |
| 644 | { |
| 645 | assert_raw_spin_locked(&task_rq(p)->lock); |
| 646 | set_tsk_need_resched(p); |
| 647 | } |
| 648 | #endif /* CONFIG_SMP */ |
| 649 | |
| 650 | #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ |
| 651 | (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) |
| 652 | /* |
| 653 | * Iterate task_group tree rooted at *from, calling @down when first entering a |
| 654 | * node and @up when leaving it for the final time. |
| 655 | * |
| 656 | * Caller must hold rcu_lock or sufficient equivalent. |
| 657 | */ |
| 658 | int walk_tg_tree_from(struct task_group *from, |
| 659 | tg_visitor down, tg_visitor up, void *data) |
| 660 | { |
| 661 | struct task_group *parent, *child; |
| 662 | int ret; |
| 663 | |
| 664 | parent = from; |
| 665 | |
| 666 | down: |
| 667 | ret = (*down)(parent, data); |
| 668 | if (ret) |
| 669 | goto out; |
| 670 | list_for_each_entry_rcu(child, &parent->children, siblings) { |
| 671 | parent = child; |
| 672 | goto down; |
| 673 | |
| 674 | up: |
| 675 | continue; |
| 676 | } |
| 677 | ret = (*up)(parent, data); |
| 678 | if (ret || parent == from) |
| 679 | goto out; |
| 680 | |
| 681 | child = parent; |
| 682 | parent = parent->parent; |
| 683 | if (parent) |
| 684 | goto up; |
| 685 | out: |
| 686 | return ret; |
| 687 | } |
| 688 | |
| 689 | int tg_nop(struct task_group *tg, void *data) |
| 690 | { |
| 691 | return 0; |
| 692 | } |
| 693 | #endif |
| 694 | |
| 695 | static void set_load_weight(struct task_struct *p) |
| 696 | { |
| 697 | int prio = p->static_prio - MAX_RT_PRIO; |
| 698 | struct load_weight *load = &p->se.load; |
| 699 | |
| 700 | /* |
| 701 | * SCHED_IDLE tasks get minimal weight: |
| 702 | */ |
| 703 | if (p->policy == SCHED_IDLE) { |
| 704 | load->weight = scale_load(WEIGHT_IDLEPRIO); |
| 705 | load->inv_weight = WMULT_IDLEPRIO; |
| 706 | return; |
| 707 | } |
| 708 | |
| 709 | load->weight = scale_load(prio_to_weight[prio]); |
| 710 | load->inv_weight = prio_to_wmult[prio]; |
| 711 | } |
| 712 | |
| 713 | static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) |
| 714 | { |
| 715 | update_rq_clock(rq); |
| 716 | sched_info_queued(p); |
| 717 | p->sched_class->enqueue_task(rq, p, flags); |
| 718 | } |
| 719 | |
| 720 | static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) |
| 721 | { |
| 722 | update_rq_clock(rq); |
| 723 | sched_info_dequeued(p); |
| 724 | p->sched_class->dequeue_task(rq, p, flags); |
| 725 | } |
| 726 | |
| 727 | void activate_task(struct rq *rq, struct task_struct *p, int flags) |
| 728 | { |
| 729 | if (task_contributes_to_load(p)) |
| 730 | rq->nr_uninterruptible--; |
| 731 | |
| 732 | enqueue_task(rq, p, flags); |
| 733 | } |
| 734 | |
| 735 | void deactivate_task(struct rq *rq, struct task_struct *p, int flags) |
| 736 | { |
| 737 | if (task_contributes_to_load(p)) |
| 738 | rq->nr_uninterruptible++; |
| 739 | |
| 740 | dequeue_task(rq, p, flags); |
| 741 | } |
| 742 | |
| 743 | static void update_rq_clock_task(struct rq *rq, s64 delta) |
| 744 | { |
| 745 | /* |
| 746 | * In theory, the compile should just see 0 here, and optimize out the call |
| 747 | * to sched_rt_avg_update. But I don't trust it... |
| 748 | */ |
| 749 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) |
| 750 | s64 steal = 0, irq_delta = 0; |
| 751 | #endif |
| 752 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING |
| 753 | irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; |
| 754 | |
| 755 | /* |
| 756 | * Since irq_time is only updated on {soft,}irq_exit, we might run into |
| 757 | * this case when a previous update_rq_clock() happened inside a |
| 758 | * {soft,}irq region. |
| 759 | * |
| 760 | * When this happens, we stop ->clock_task and only update the |
| 761 | * prev_irq_time stamp to account for the part that fit, so that a next |
| 762 | * update will consume the rest. This ensures ->clock_task is |
| 763 | * monotonic. |
| 764 | * |
| 765 | * It does however cause some slight miss-attribution of {soft,}irq |
| 766 | * time, a more accurate solution would be to update the irq_time using |
| 767 | * the current rq->clock timestamp, except that would require using |
| 768 | * atomic ops. |
| 769 | */ |
| 770 | if (irq_delta > delta) |
| 771 | irq_delta = delta; |
| 772 | |
| 773 | rq->prev_irq_time += irq_delta; |
| 774 | delta -= irq_delta; |
| 775 | #endif |
| 776 | #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING |
| 777 | if (static_key_false((¶virt_steal_rq_enabled))) { |
| 778 | u64 st; |
| 779 | |
| 780 | steal = paravirt_steal_clock(cpu_of(rq)); |
| 781 | steal -= rq->prev_steal_time_rq; |
| 782 | |
| 783 | if (unlikely(steal > delta)) |
| 784 | steal = delta; |
| 785 | |
| 786 | st = steal_ticks(steal); |
| 787 | steal = st * TICK_NSEC; |
| 788 | |
| 789 | rq->prev_steal_time_rq += steal; |
| 790 | |
| 791 | delta -= steal; |
| 792 | } |
| 793 | #endif |
| 794 | |
| 795 | rq->clock_task += delta; |
| 796 | |
| 797 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) |
| 798 | if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) |
| 799 | sched_rt_avg_update(rq, irq_delta + steal); |
| 800 | #endif |
| 801 | } |
| 802 | |
| 803 | void sched_set_stop_task(int cpu, struct task_struct *stop) |
| 804 | { |
| 805 | struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; |
| 806 | struct task_struct *old_stop = cpu_rq(cpu)->stop; |
| 807 | |
| 808 | if (stop) { |
| 809 | /* |
| 810 | * Make it appear like a SCHED_FIFO task, its something |
| 811 | * userspace knows about and won't get confused about. |
| 812 | * |
| 813 | * Also, it will make PI more or less work without too |
| 814 | * much confusion -- but then, stop work should not |
| 815 | * rely on PI working anyway. |
| 816 | */ |
| 817 | sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); |
| 818 | |
| 819 | stop->sched_class = &stop_sched_class; |
| 820 | } |
| 821 | |
| 822 | cpu_rq(cpu)->stop = stop; |
| 823 | |
| 824 | if (old_stop) { |
| 825 | /* |
| 826 | * Reset it back to a normal scheduling class so that |
| 827 | * it can die in pieces. |
| 828 | */ |
| 829 | old_stop->sched_class = &rt_sched_class; |
| 830 | } |
| 831 | } |
| 832 | |
| 833 | /* |
| 834 | * __normal_prio - return the priority that is based on the static prio |
| 835 | */ |
| 836 | static inline int __normal_prio(struct task_struct *p) |
| 837 | { |
| 838 | return p->static_prio; |
| 839 | } |
| 840 | |
| 841 | /* |
| 842 | * Calculate the expected normal priority: i.e. priority |
| 843 | * without taking RT-inheritance into account. Might be |
| 844 | * boosted by interactivity modifiers. Changes upon fork, |
| 845 | * setprio syscalls, and whenever the interactivity |
| 846 | * estimator recalculates. |
| 847 | */ |
| 848 | static inline int normal_prio(struct task_struct *p) |
| 849 | { |
| 850 | int prio; |
| 851 | |
| 852 | if (task_has_rt_policy(p)) |
| 853 | prio = MAX_RT_PRIO-1 - p->rt_priority; |
| 854 | else |
| 855 | prio = __normal_prio(p); |
| 856 | return prio; |
| 857 | } |
| 858 | |
| 859 | /* |
| 860 | * Calculate the current priority, i.e. the priority |
| 861 | * taken into account by the scheduler. This value might |
| 862 | * be boosted by RT tasks, or might be boosted by |
| 863 | * interactivity modifiers. Will be RT if the task got |
| 864 | * RT-boosted. If not then it returns p->normal_prio. |
| 865 | */ |
| 866 | static int effective_prio(struct task_struct *p) |
| 867 | { |
| 868 | p->normal_prio = normal_prio(p); |
| 869 | /* |
| 870 | * If we are RT tasks or we were boosted to RT priority, |
| 871 | * keep the priority unchanged. Otherwise, update priority |
| 872 | * to the normal priority: |
| 873 | */ |
| 874 | if (!rt_prio(p->prio)) |
| 875 | return p->normal_prio; |
| 876 | return p->prio; |
| 877 | } |
| 878 | |
| 879 | /** |
| 880 | * task_curr - is this task currently executing on a CPU? |
| 881 | * @p: the task in question. |
| 882 | */ |
| 883 | inline int task_curr(const struct task_struct *p) |
| 884 | { |
| 885 | return cpu_curr(task_cpu(p)) == p; |
| 886 | } |
| 887 | |
| 888 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, |
| 889 | const struct sched_class *prev_class, |
| 890 | int oldprio) |
| 891 | { |
| 892 | if (prev_class != p->sched_class) { |
| 893 | if (prev_class->switched_from) |
| 894 | prev_class->switched_from(rq, p); |
| 895 | p->sched_class->switched_to(rq, p); |
| 896 | } else if (oldprio != p->prio) |
| 897 | p->sched_class->prio_changed(rq, p, oldprio); |
| 898 | } |
| 899 | |
| 900 | void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) |
| 901 | { |
| 902 | const struct sched_class *class; |
| 903 | |
| 904 | if (p->sched_class == rq->curr->sched_class) { |
| 905 | rq->curr->sched_class->check_preempt_curr(rq, p, flags); |
| 906 | } else { |
| 907 | for_each_class(class) { |
| 908 | if (class == rq->curr->sched_class) |
| 909 | break; |
| 910 | if (class == p->sched_class) { |
| 911 | resched_task(rq->curr); |
| 912 | break; |
| 913 | } |
| 914 | } |
| 915 | } |
| 916 | |
| 917 | /* |
| 918 | * A queue event has occurred, and we're going to schedule. In |
| 919 | * this case, we can save a useless back to back clock update. |
| 920 | */ |
| 921 | if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) |
| 922 | rq->skip_clock_update = 1; |
| 923 | } |
| 924 | |
| 925 | #ifdef CONFIG_SMP |
| 926 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
| 927 | { |
| 928 | #ifdef CONFIG_SCHED_DEBUG |
| 929 | /* |
| 930 | * We should never call set_task_cpu() on a blocked task, |
| 931 | * ttwu() will sort out the placement. |
| 932 | */ |
| 933 | WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && |
| 934 | !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); |
| 935 | |
| 936 | #ifdef CONFIG_LOCKDEP |
| 937 | /* |
| 938 | * The caller should hold either p->pi_lock or rq->lock, when changing |
| 939 | * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. |
| 940 | * |
| 941 | * sched_move_task() holds both and thus holding either pins the cgroup, |
| 942 | * see task_group(). |
| 943 | * |
| 944 | * Furthermore, all task_rq users should acquire both locks, see |
| 945 | * task_rq_lock(). |
| 946 | */ |
| 947 | WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || |
| 948 | lockdep_is_held(&task_rq(p)->lock))); |
| 949 | #endif |
| 950 | #endif |
| 951 | |
| 952 | trace_sched_migrate_task(p, new_cpu); |
| 953 | |
| 954 | if (task_cpu(p) != new_cpu) { |
| 955 | p->se.nr_migrations++; |
| 956 | perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); |
| 957 | } |
| 958 | |
| 959 | __set_task_cpu(p, new_cpu); |
| 960 | } |
| 961 | |
| 962 | struct migration_arg { |
| 963 | struct task_struct *task; |
| 964 | int dest_cpu; |
| 965 | }; |
| 966 | |
| 967 | static int migration_cpu_stop(void *data); |
| 968 | |
| 969 | /* |
| 970 | * wait_task_inactive - wait for a thread to unschedule. |
| 971 | * |
| 972 | * If @match_state is nonzero, it's the @p->state value just checked and |
| 973 | * not expected to change. If it changes, i.e. @p might have woken up, |
| 974 | * then return zero. When we succeed in waiting for @p to be off its CPU, |
| 975 | * we return a positive number (its total switch count). If a second call |
| 976 | * a short while later returns the same number, the caller can be sure that |
| 977 | * @p has remained unscheduled the whole time. |
| 978 | * |
| 979 | * The caller must ensure that the task *will* unschedule sometime soon, |
| 980 | * else this function might spin for a *long* time. This function can't |
| 981 | * be called with interrupts off, or it may introduce deadlock with |
| 982 | * smp_call_function() if an IPI is sent by the same process we are |
| 983 | * waiting to become inactive. |
| 984 | */ |
| 985 | unsigned long wait_task_inactive(struct task_struct *p, long match_state) |
| 986 | { |
| 987 | unsigned long flags; |
| 988 | int running, on_rq; |
| 989 | unsigned long ncsw; |
| 990 | struct rq *rq; |
| 991 | |
| 992 | for (;;) { |
| 993 | /* |
| 994 | * We do the initial early heuristics without holding |
| 995 | * any task-queue locks at all. We'll only try to get |
| 996 | * the runqueue lock when things look like they will |
| 997 | * work out! |
| 998 | */ |
| 999 | rq = task_rq(p); |
| 1000 | |
| 1001 | /* |
| 1002 | * If the task is actively running on another CPU |
| 1003 | * still, just relax and busy-wait without holding |
| 1004 | * any locks. |
| 1005 | * |
| 1006 | * NOTE! Since we don't hold any locks, it's not |
| 1007 | * even sure that "rq" stays as the right runqueue! |
| 1008 | * But we don't care, since "task_running()" will |
| 1009 | * return false if the runqueue has changed and p |
| 1010 | * is actually now running somewhere else! |
| 1011 | */ |
| 1012 | while (task_running(rq, p)) { |
| 1013 | if (match_state && unlikely(p->state != match_state)) |
| 1014 | return 0; |
| 1015 | cpu_relax(); |
| 1016 | } |
| 1017 | |
| 1018 | /* |
| 1019 | * Ok, time to look more closely! We need the rq |
| 1020 | * lock now, to be *sure*. If we're wrong, we'll |
| 1021 | * just go back and repeat. |
| 1022 | */ |
| 1023 | rq = task_rq_lock(p, &flags); |
| 1024 | trace_sched_wait_task(p); |
| 1025 | running = task_running(rq, p); |
| 1026 | on_rq = p->on_rq; |
| 1027 | ncsw = 0; |
| 1028 | if (!match_state || p->state == match_state) |
| 1029 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ |
| 1030 | task_rq_unlock(rq, p, &flags); |
| 1031 | |
| 1032 | /* |
| 1033 | * If it changed from the expected state, bail out now. |
| 1034 | */ |
| 1035 | if (unlikely(!ncsw)) |
| 1036 | break; |
| 1037 | |
| 1038 | /* |
| 1039 | * Was it really running after all now that we |
| 1040 | * checked with the proper locks actually held? |
| 1041 | * |
| 1042 | * Oops. Go back and try again.. |
| 1043 | */ |
| 1044 | if (unlikely(running)) { |
| 1045 | cpu_relax(); |
| 1046 | continue; |
| 1047 | } |
| 1048 | |
| 1049 | /* |
| 1050 | * It's not enough that it's not actively running, |
| 1051 | * it must be off the runqueue _entirely_, and not |
| 1052 | * preempted! |
| 1053 | * |
| 1054 | * So if it was still runnable (but just not actively |
| 1055 | * running right now), it's preempted, and we should |
| 1056 | * yield - it could be a while. |
| 1057 | */ |
| 1058 | if (unlikely(on_rq)) { |
| 1059 | ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); |
| 1060 | |
| 1061 | set_current_state(TASK_UNINTERRUPTIBLE); |
| 1062 | schedule_hrtimeout(&to, HRTIMER_MODE_REL); |
| 1063 | continue; |
| 1064 | } |
| 1065 | |
| 1066 | /* |
| 1067 | * Ahh, all good. It wasn't running, and it wasn't |
| 1068 | * runnable, which means that it will never become |
| 1069 | * running in the future either. We're all done! |
| 1070 | */ |
| 1071 | break; |
| 1072 | } |
| 1073 | |
| 1074 | return ncsw; |
| 1075 | } |
| 1076 | |
| 1077 | /*** |
| 1078 | * kick_process - kick a running thread to enter/exit the kernel |
| 1079 | * @p: the to-be-kicked thread |
| 1080 | * |
| 1081 | * Cause a process which is running on another CPU to enter |
| 1082 | * kernel-mode, without any delay. (to get signals handled.) |
| 1083 | * |
| 1084 | * NOTE: this function doesn't have to take the runqueue lock, |
| 1085 | * because all it wants to ensure is that the remote task enters |
| 1086 | * the kernel. If the IPI races and the task has been migrated |
| 1087 | * to another CPU then no harm is done and the purpose has been |
| 1088 | * achieved as well. |
| 1089 | */ |
| 1090 | void kick_process(struct task_struct *p) |
| 1091 | { |
| 1092 | int cpu; |
| 1093 | |
| 1094 | preempt_disable(); |
| 1095 | cpu = task_cpu(p); |
| 1096 | if ((cpu != smp_processor_id()) && task_curr(p)) |
| 1097 | smp_send_reschedule(cpu); |
| 1098 | preempt_enable(); |
| 1099 | } |
| 1100 | EXPORT_SYMBOL_GPL(kick_process); |
| 1101 | #endif /* CONFIG_SMP */ |
| 1102 | |
| 1103 | #ifdef CONFIG_SMP |
| 1104 | /* |
| 1105 | * ->cpus_allowed is protected by both rq->lock and p->pi_lock |
| 1106 | */ |
| 1107 | static int select_fallback_rq(int cpu, struct task_struct *p) |
| 1108 | { |
| 1109 | const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); |
| 1110 | enum { cpuset, possible, fail } state = cpuset; |
| 1111 | int dest_cpu; |
| 1112 | |
| 1113 | /* Look for allowed, online CPU in same node. */ |
| 1114 | for_each_cpu(dest_cpu, nodemask) { |
| 1115 | if (!cpu_online(dest_cpu)) |
| 1116 | continue; |
| 1117 | if (!cpu_active(dest_cpu)) |
| 1118 | continue; |
| 1119 | if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) |
| 1120 | return dest_cpu; |
| 1121 | } |
| 1122 | |
| 1123 | for (;;) { |
| 1124 | /* Any allowed, online CPU? */ |
| 1125 | for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { |
| 1126 | if (!cpu_online(dest_cpu)) |
| 1127 | continue; |
| 1128 | if (!cpu_active(dest_cpu)) |
| 1129 | continue; |
| 1130 | goto out; |
| 1131 | } |
| 1132 | |
| 1133 | switch (state) { |
| 1134 | case cpuset: |
| 1135 | /* No more Mr. Nice Guy. */ |
| 1136 | cpuset_cpus_allowed_fallback(p); |
| 1137 | state = possible; |
| 1138 | break; |
| 1139 | |
| 1140 | case possible: |
| 1141 | do_set_cpus_allowed(p, cpu_possible_mask); |
| 1142 | state = fail; |
| 1143 | break; |
| 1144 | |
| 1145 | case fail: |
| 1146 | BUG(); |
| 1147 | break; |
| 1148 | } |
| 1149 | } |
| 1150 | |
| 1151 | out: |
| 1152 | if (state != cpuset) { |
| 1153 | /* |
| 1154 | * Don't tell them about moving exiting tasks or |
| 1155 | * kernel threads (both mm NULL), since they never |
| 1156 | * leave kernel. |
| 1157 | */ |
| 1158 | if (p->mm && printk_ratelimit()) { |
| 1159 | printk_sched("process %d (%s) no longer affine to cpu%d\n", |
| 1160 | task_pid_nr(p), p->comm, cpu); |
| 1161 | } |
| 1162 | } |
| 1163 | |
| 1164 | return dest_cpu; |
| 1165 | } |
| 1166 | |
| 1167 | /* |
| 1168 | * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. |
| 1169 | */ |
| 1170 | static inline |
| 1171 | int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) |
| 1172 | { |
| 1173 | int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); |
| 1174 | |
| 1175 | /* |
| 1176 | * In order not to call set_task_cpu() on a blocking task we need |
| 1177 | * to rely on ttwu() to place the task on a valid ->cpus_allowed |
| 1178 | * cpu. |
| 1179 | * |
| 1180 | * Since this is common to all placement strategies, this lives here. |
| 1181 | * |
| 1182 | * [ this allows ->select_task() to simply return task_cpu(p) and |
| 1183 | * not worry about this generic constraint ] |
| 1184 | */ |
| 1185 | if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || |
| 1186 | !cpu_online(cpu))) |
| 1187 | cpu = select_fallback_rq(task_cpu(p), p); |
| 1188 | |
| 1189 | return cpu; |
| 1190 | } |
| 1191 | |
| 1192 | static void update_avg(u64 *avg, u64 sample) |
| 1193 | { |
| 1194 | s64 diff = sample - *avg; |
| 1195 | *avg += diff >> 3; |
| 1196 | } |
| 1197 | #endif |
| 1198 | |
| 1199 | static void |
| 1200 | ttwu_stat(struct task_struct *p, int cpu, int wake_flags) |
| 1201 | { |
| 1202 | #ifdef CONFIG_SCHEDSTATS |
| 1203 | struct rq *rq = this_rq(); |
| 1204 | |
| 1205 | #ifdef CONFIG_SMP |
| 1206 | int this_cpu = smp_processor_id(); |
| 1207 | |
| 1208 | if (cpu == this_cpu) { |
| 1209 | schedstat_inc(rq, ttwu_local); |
| 1210 | schedstat_inc(p, se.statistics.nr_wakeups_local); |
| 1211 | } else { |
| 1212 | struct sched_domain *sd; |
| 1213 | |
| 1214 | schedstat_inc(p, se.statistics.nr_wakeups_remote); |
| 1215 | rcu_read_lock(); |
| 1216 | for_each_domain(this_cpu, sd) { |
| 1217 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
| 1218 | schedstat_inc(sd, ttwu_wake_remote); |
| 1219 | break; |
| 1220 | } |
| 1221 | } |
| 1222 | rcu_read_unlock(); |
| 1223 | } |
| 1224 | |
| 1225 | if (wake_flags & WF_MIGRATED) |
| 1226 | schedstat_inc(p, se.statistics.nr_wakeups_migrate); |
| 1227 | |
| 1228 | #endif /* CONFIG_SMP */ |
| 1229 | |
| 1230 | schedstat_inc(rq, ttwu_count); |
| 1231 | schedstat_inc(p, se.statistics.nr_wakeups); |
| 1232 | |
| 1233 | if (wake_flags & WF_SYNC) |
| 1234 | schedstat_inc(p, se.statistics.nr_wakeups_sync); |
| 1235 | |
| 1236 | #endif /* CONFIG_SCHEDSTATS */ |
| 1237 | } |
| 1238 | |
| 1239 | static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) |
| 1240 | { |
| 1241 | activate_task(rq, p, en_flags); |
| 1242 | p->on_rq = 1; |
| 1243 | |
| 1244 | /* if a worker is waking up, notify workqueue */ |
| 1245 | if (p->flags & PF_WQ_WORKER) |
| 1246 | wq_worker_waking_up(p, cpu_of(rq)); |
| 1247 | } |
| 1248 | |
| 1249 | /* |
| 1250 | * Mark the task runnable and perform wakeup-preemption. |
| 1251 | */ |
| 1252 | static void |
| 1253 | ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
| 1254 | { |
| 1255 | trace_sched_wakeup(p, true); |
| 1256 | check_preempt_curr(rq, p, wake_flags); |
| 1257 | |
| 1258 | p->state = TASK_RUNNING; |
| 1259 | #ifdef CONFIG_SMP |
| 1260 | if (p->sched_class->task_woken) |
| 1261 | p->sched_class->task_woken(rq, p); |
| 1262 | |
| 1263 | if (rq->idle_stamp) { |
| 1264 | u64 delta = rq->clock - rq->idle_stamp; |
| 1265 | u64 max = 2*sysctl_sched_migration_cost; |
| 1266 | |
| 1267 | if (delta > max) |
| 1268 | rq->avg_idle = max; |
| 1269 | else |
| 1270 | update_avg(&rq->avg_idle, delta); |
| 1271 | rq->idle_stamp = 0; |
| 1272 | } |
| 1273 | #endif |
| 1274 | } |
| 1275 | |
| 1276 | static void |
| 1277 | ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) |
| 1278 | { |
| 1279 | #ifdef CONFIG_SMP |
| 1280 | if (p->sched_contributes_to_load) |
| 1281 | rq->nr_uninterruptible--; |
| 1282 | #endif |
| 1283 | |
| 1284 | ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); |
| 1285 | ttwu_do_wakeup(rq, p, wake_flags); |
| 1286 | } |
| 1287 | |
| 1288 | /* |
| 1289 | * Called in case the task @p isn't fully descheduled from its runqueue, |
| 1290 | * in this case we must do a remote wakeup. Its a 'light' wakeup though, |
| 1291 | * since all we need to do is flip p->state to TASK_RUNNING, since |
| 1292 | * the task is still ->on_rq. |
| 1293 | */ |
| 1294 | static int ttwu_remote(struct task_struct *p, int wake_flags) |
| 1295 | { |
| 1296 | struct rq *rq; |
| 1297 | int ret = 0; |
| 1298 | |
| 1299 | rq = __task_rq_lock(p); |
| 1300 | if (p->on_rq) { |
| 1301 | ttwu_do_wakeup(rq, p, wake_flags); |
| 1302 | ret = 1; |
| 1303 | } |
| 1304 | __task_rq_unlock(rq); |
| 1305 | |
| 1306 | return ret; |
| 1307 | } |
| 1308 | |
| 1309 | #ifdef CONFIG_SMP |
| 1310 | static void sched_ttwu_pending(void) |
| 1311 | { |
| 1312 | struct rq *rq = this_rq(); |
| 1313 | struct llist_node *llist = llist_del_all(&rq->wake_list); |
| 1314 | struct task_struct *p; |
| 1315 | |
| 1316 | raw_spin_lock(&rq->lock); |
| 1317 | |
| 1318 | while (llist) { |
| 1319 | p = llist_entry(llist, struct task_struct, wake_entry); |
| 1320 | llist = llist_next(llist); |
| 1321 | ttwu_do_activate(rq, p, 0); |
| 1322 | } |
| 1323 | |
| 1324 | raw_spin_unlock(&rq->lock); |
| 1325 | } |
| 1326 | |
| 1327 | void scheduler_ipi(void) |
| 1328 | { |
| 1329 | if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) |
| 1330 | return; |
| 1331 | |
| 1332 | /* |
| 1333 | * Not all reschedule IPI handlers call irq_enter/irq_exit, since |
| 1334 | * traditionally all their work was done from the interrupt return |
| 1335 | * path. Now that we actually do some work, we need to make sure |
| 1336 | * we do call them. |
| 1337 | * |
| 1338 | * Some archs already do call them, luckily irq_enter/exit nest |
| 1339 | * properly. |
| 1340 | * |
| 1341 | * Arguably we should visit all archs and update all handlers, |
| 1342 | * however a fair share of IPIs are still resched only so this would |
| 1343 | * somewhat pessimize the simple resched case. |
| 1344 | */ |
| 1345 | irq_enter(); |
| 1346 | sched_ttwu_pending(); |
| 1347 | |
| 1348 | /* |
| 1349 | * Check if someone kicked us for doing the nohz idle load balance. |
| 1350 | */ |
| 1351 | if (unlikely(got_nohz_idle_kick() && !need_resched())) { |
| 1352 | this_rq()->idle_balance = 1; |
| 1353 | raise_softirq_irqoff(SCHED_SOFTIRQ); |
| 1354 | } |
| 1355 | irq_exit(); |
| 1356 | } |
| 1357 | |
| 1358 | static void ttwu_queue_remote(struct task_struct *p, int cpu) |
| 1359 | { |
| 1360 | if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) |
| 1361 | smp_send_reschedule(cpu); |
| 1362 | } |
| 1363 | |
| 1364 | bool cpus_share_cache(int this_cpu, int that_cpu) |
| 1365 | { |
| 1366 | return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); |
| 1367 | } |
| 1368 | #endif /* CONFIG_SMP */ |
| 1369 | |
| 1370 | static void ttwu_queue(struct task_struct *p, int cpu) |
| 1371 | { |
| 1372 | struct rq *rq = cpu_rq(cpu); |
| 1373 | |
| 1374 | #if defined(CONFIG_SMP) |
| 1375 | if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { |
| 1376 | sched_clock_cpu(cpu); /* sync clocks x-cpu */ |
| 1377 | ttwu_queue_remote(p, cpu); |
| 1378 | return; |
| 1379 | } |
| 1380 | #endif |
| 1381 | |
| 1382 | raw_spin_lock(&rq->lock); |
| 1383 | ttwu_do_activate(rq, p, 0); |
| 1384 | raw_spin_unlock(&rq->lock); |
| 1385 | } |
| 1386 | |
| 1387 | /** |
| 1388 | * try_to_wake_up - wake up a thread |
| 1389 | * @p: the thread to be awakened |
| 1390 | * @state: the mask of task states that can be woken |
| 1391 | * @wake_flags: wake modifier flags (WF_*) |
| 1392 | * |
| 1393 | * Put it on the run-queue if it's not already there. The "current" |
| 1394 | * thread is always on the run-queue (except when the actual |
| 1395 | * re-schedule is in progress), and as such you're allowed to do |
| 1396 | * the simpler "current->state = TASK_RUNNING" to mark yourself |
| 1397 | * runnable without the overhead of this. |
| 1398 | * |
| 1399 | * Returns %true if @p was woken up, %false if it was already running |
| 1400 | * or @state didn't match @p's state. |
| 1401 | */ |
| 1402 | static int |
| 1403 | try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) |
| 1404 | { |
| 1405 | unsigned long flags; |
| 1406 | int cpu, success = 0; |
| 1407 | |
| 1408 | smp_wmb(); |
| 1409 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
| 1410 | if (!(p->state & state)) |
| 1411 | goto out; |
| 1412 | |
| 1413 | success = 1; /* we're going to change ->state */ |
| 1414 | cpu = task_cpu(p); |
| 1415 | |
| 1416 | if (p->on_rq && ttwu_remote(p, wake_flags)) |
| 1417 | goto stat; |
| 1418 | |
| 1419 | #ifdef CONFIG_SMP |
| 1420 | /* |
| 1421 | * If the owning (remote) cpu is still in the middle of schedule() with |
| 1422 | * this task as prev, wait until its done referencing the task. |
| 1423 | */ |
| 1424 | while (p->on_cpu) |
| 1425 | cpu_relax(); |
| 1426 | /* |
| 1427 | * Pairs with the smp_wmb() in finish_lock_switch(). |
| 1428 | */ |
| 1429 | smp_rmb(); |
| 1430 | |
| 1431 | p->sched_contributes_to_load = !!task_contributes_to_load(p); |
| 1432 | p->state = TASK_WAKING; |
| 1433 | |
| 1434 | if (p->sched_class->task_waking) |
| 1435 | p->sched_class->task_waking(p); |
| 1436 | |
| 1437 | cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); |
| 1438 | if (task_cpu(p) != cpu) { |
| 1439 | wake_flags |= WF_MIGRATED; |
| 1440 | set_task_cpu(p, cpu); |
| 1441 | } |
| 1442 | #endif /* CONFIG_SMP */ |
| 1443 | |
| 1444 | ttwu_queue(p, cpu); |
| 1445 | stat: |
| 1446 | ttwu_stat(p, cpu, wake_flags); |
| 1447 | out: |
| 1448 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
| 1449 | |
| 1450 | return success; |
| 1451 | } |
| 1452 | |
| 1453 | /** |
| 1454 | * try_to_wake_up_local - try to wake up a local task with rq lock held |
| 1455 | * @p: the thread to be awakened |
| 1456 | * |
| 1457 | * Put @p on the run-queue if it's not already there. The caller must |
| 1458 | * ensure that this_rq() is locked, @p is bound to this_rq() and not |
| 1459 | * the current task. |
| 1460 | */ |
| 1461 | static void try_to_wake_up_local(struct task_struct *p) |
| 1462 | { |
| 1463 | struct rq *rq = task_rq(p); |
| 1464 | |
| 1465 | BUG_ON(rq != this_rq()); |
| 1466 | BUG_ON(p == current); |
| 1467 | lockdep_assert_held(&rq->lock); |
| 1468 | |
| 1469 | if (!raw_spin_trylock(&p->pi_lock)) { |
| 1470 | raw_spin_unlock(&rq->lock); |
| 1471 | raw_spin_lock(&p->pi_lock); |
| 1472 | raw_spin_lock(&rq->lock); |
| 1473 | } |
| 1474 | |
| 1475 | if (!(p->state & TASK_NORMAL)) |
| 1476 | goto out; |
| 1477 | |
| 1478 | if (!p->on_rq) |
| 1479 | ttwu_activate(rq, p, ENQUEUE_WAKEUP); |
| 1480 | |
| 1481 | ttwu_do_wakeup(rq, p, 0); |
| 1482 | ttwu_stat(p, smp_processor_id(), 0); |
| 1483 | out: |
| 1484 | raw_spin_unlock(&p->pi_lock); |
| 1485 | } |
| 1486 | |
| 1487 | /** |
| 1488 | * wake_up_process - Wake up a specific process |
| 1489 | * @p: The process to be woken up. |
| 1490 | * |
| 1491 | * Attempt to wake up the nominated process and move it to the set of runnable |
| 1492 | * processes. Returns 1 if the process was woken up, 0 if it was already |
| 1493 | * running. |
| 1494 | * |
| 1495 | * It may be assumed that this function implies a write memory barrier before |
| 1496 | * changing the task state if and only if any tasks are woken up. |
| 1497 | */ |
| 1498 | int wake_up_process(struct task_struct *p) |
| 1499 | { |
| 1500 | return try_to_wake_up(p, TASK_ALL, 0); |
| 1501 | } |
| 1502 | EXPORT_SYMBOL(wake_up_process); |
| 1503 | |
| 1504 | int wake_up_state(struct task_struct *p, unsigned int state) |
| 1505 | { |
| 1506 | return try_to_wake_up(p, state, 0); |
| 1507 | } |
| 1508 | |
| 1509 | /* |
| 1510 | * Perform scheduler related setup for a newly forked process p. |
| 1511 | * p is forked by current. |
| 1512 | * |
| 1513 | * __sched_fork() is basic setup used by init_idle() too: |
| 1514 | */ |
| 1515 | static void __sched_fork(struct task_struct *p) |
| 1516 | { |
| 1517 | p->on_rq = 0; |
| 1518 | |
| 1519 | p->se.on_rq = 0; |
| 1520 | p->se.exec_start = 0; |
| 1521 | p->se.sum_exec_runtime = 0; |
| 1522 | p->se.prev_sum_exec_runtime = 0; |
| 1523 | p->se.nr_migrations = 0; |
| 1524 | p->se.vruntime = 0; |
| 1525 | INIT_LIST_HEAD(&p->se.group_node); |
| 1526 | |
| 1527 | #ifdef CONFIG_SCHEDSTATS |
| 1528 | memset(&p->se.statistics, 0, sizeof(p->se.statistics)); |
| 1529 | #endif |
| 1530 | |
| 1531 | INIT_LIST_HEAD(&p->rt.run_list); |
| 1532 | |
| 1533 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 1534 | INIT_HLIST_HEAD(&p->preempt_notifiers); |
| 1535 | #endif |
| 1536 | |
| 1537 | #ifdef CONFIG_NUMA_BALANCING |
| 1538 | if (p->mm && atomic_read(&p->mm->mm_users) == 1) { |
| 1539 | p->mm->numa_next_scan = jiffies; |
| 1540 | p->mm->numa_scan_seq = 0; |
| 1541 | } |
| 1542 | |
| 1543 | p->node_stamp = 0ULL; |
| 1544 | p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; |
| 1545 | p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; |
| 1546 | p->numa_scan_period = sysctl_numa_balancing_scan_period_min; |
| 1547 | p->numa_work.next = &p->numa_work; |
| 1548 | #endif /* CONFIG_NUMA_BALANCING */ |
| 1549 | } |
| 1550 | |
| 1551 | /* |
| 1552 | * fork()/clone()-time setup: |
| 1553 | */ |
| 1554 | void sched_fork(struct task_struct *p) |
| 1555 | { |
| 1556 | unsigned long flags; |
| 1557 | int cpu = get_cpu(); |
| 1558 | |
| 1559 | __sched_fork(p); |
| 1560 | /* |
| 1561 | * We mark the process as running here. This guarantees that |
| 1562 | * nobody will actually run it, and a signal or other external |
| 1563 | * event cannot wake it up and insert it on the runqueue either. |
| 1564 | */ |
| 1565 | p->state = TASK_RUNNING; |
| 1566 | |
| 1567 | /* |
| 1568 | * Make sure we do not leak PI boosting priority to the child. |
| 1569 | */ |
| 1570 | p->prio = current->normal_prio; |
| 1571 | |
| 1572 | /* |
| 1573 | * Revert to default priority/policy on fork if requested. |
| 1574 | */ |
| 1575 | if (unlikely(p->sched_reset_on_fork)) { |
| 1576 | if (task_has_rt_policy(p)) { |
| 1577 | p->policy = SCHED_NORMAL; |
| 1578 | p->static_prio = NICE_TO_PRIO(0); |
| 1579 | p->rt_priority = 0; |
| 1580 | } else if (PRIO_TO_NICE(p->static_prio) < 0) |
| 1581 | p->static_prio = NICE_TO_PRIO(0); |
| 1582 | |
| 1583 | p->prio = p->normal_prio = __normal_prio(p); |
| 1584 | set_load_weight(p); |
| 1585 | |
| 1586 | /* |
| 1587 | * We don't need the reset flag anymore after the fork. It has |
| 1588 | * fulfilled its duty: |
| 1589 | */ |
| 1590 | p->sched_reset_on_fork = 0; |
| 1591 | } |
| 1592 | |
| 1593 | if (!rt_prio(p->prio)) |
| 1594 | p->sched_class = &fair_sched_class; |
| 1595 | |
| 1596 | if (p->sched_class->task_fork) |
| 1597 | p->sched_class->task_fork(p); |
| 1598 | |
| 1599 | /* |
| 1600 | * The child is not yet in the pid-hash so no cgroup attach races, |
| 1601 | * and the cgroup is pinned to this child due to cgroup_fork() |
| 1602 | * is ran before sched_fork(). |
| 1603 | * |
| 1604 | * Silence PROVE_RCU. |
| 1605 | */ |
| 1606 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
| 1607 | set_task_cpu(p, cpu); |
| 1608 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
| 1609 | |
| 1610 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
| 1611 | if (likely(sched_info_on())) |
| 1612 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
| 1613 | #endif |
| 1614 | #if defined(CONFIG_SMP) |
| 1615 | p->on_cpu = 0; |
| 1616 | #endif |
| 1617 | #ifdef CONFIG_PREEMPT_COUNT |
| 1618 | /* Want to start with kernel preemption disabled. */ |
| 1619 | task_thread_info(p)->preempt_count = 1; |
| 1620 | #endif |
| 1621 | #ifdef CONFIG_SMP |
| 1622 | plist_node_init(&p->pushable_tasks, MAX_PRIO); |
| 1623 | #endif |
| 1624 | |
| 1625 | put_cpu(); |
| 1626 | } |
| 1627 | |
| 1628 | /* |
| 1629 | * wake_up_new_task - wake up a newly created task for the first time. |
| 1630 | * |
| 1631 | * This function will do some initial scheduler statistics housekeeping |
| 1632 | * that must be done for every newly created context, then puts the task |
| 1633 | * on the runqueue and wakes it. |
| 1634 | */ |
| 1635 | void wake_up_new_task(struct task_struct *p) |
| 1636 | { |
| 1637 | unsigned long flags; |
| 1638 | struct rq *rq; |
| 1639 | |
| 1640 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
| 1641 | #ifdef CONFIG_SMP |
| 1642 | /* |
| 1643 | * Fork balancing, do it here and not earlier because: |
| 1644 | * - cpus_allowed can change in the fork path |
| 1645 | * - any previously selected cpu might disappear through hotplug |
| 1646 | */ |
| 1647 | set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); |
| 1648 | #endif |
| 1649 | |
| 1650 | rq = __task_rq_lock(p); |
| 1651 | activate_task(rq, p, 0); |
| 1652 | p->on_rq = 1; |
| 1653 | trace_sched_wakeup_new(p, true); |
| 1654 | check_preempt_curr(rq, p, WF_FORK); |
| 1655 | #ifdef CONFIG_SMP |
| 1656 | if (p->sched_class->task_woken) |
| 1657 | p->sched_class->task_woken(rq, p); |
| 1658 | #endif |
| 1659 | task_rq_unlock(rq, p, &flags); |
| 1660 | } |
| 1661 | |
| 1662 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 1663 | |
| 1664 | /** |
| 1665 | * preempt_notifier_register - tell me when current is being preempted & rescheduled |
| 1666 | * @notifier: notifier struct to register |
| 1667 | */ |
| 1668 | void preempt_notifier_register(struct preempt_notifier *notifier) |
| 1669 | { |
| 1670 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); |
| 1671 | } |
| 1672 | EXPORT_SYMBOL_GPL(preempt_notifier_register); |
| 1673 | |
| 1674 | /** |
| 1675 | * preempt_notifier_unregister - no longer interested in preemption notifications |
| 1676 | * @notifier: notifier struct to unregister |
| 1677 | * |
| 1678 | * This is safe to call from within a preemption notifier. |
| 1679 | */ |
| 1680 | void preempt_notifier_unregister(struct preempt_notifier *notifier) |
| 1681 | { |
| 1682 | hlist_del(¬ifier->link); |
| 1683 | } |
| 1684 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); |
| 1685 | |
| 1686 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
| 1687 | { |
| 1688 | struct preempt_notifier *notifier; |
| 1689 | struct hlist_node *node; |
| 1690 | |
| 1691 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) |
| 1692 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); |
| 1693 | } |
| 1694 | |
| 1695 | static void |
| 1696 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
| 1697 | struct task_struct *next) |
| 1698 | { |
| 1699 | struct preempt_notifier *notifier; |
| 1700 | struct hlist_node *node; |
| 1701 | |
| 1702 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) |
| 1703 | notifier->ops->sched_out(notifier, next); |
| 1704 | } |
| 1705 | |
| 1706 | #else /* !CONFIG_PREEMPT_NOTIFIERS */ |
| 1707 | |
| 1708 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
| 1709 | { |
| 1710 | } |
| 1711 | |
| 1712 | static void |
| 1713 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
| 1714 | struct task_struct *next) |
| 1715 | { |
| 1716 | } |
| 1717 | |
| 1718 | #endif /* CONFIG_PREEMPT_NOTIFIERS */ |
| 1719 | |
| 1720 | /** |
| 1721 | * prepare_task_switch - prepare to switch tasks |
| 1722 | * @rq: the runqueue preparing to switch |
| 1723 | * @prev: the current task that is being switched out |
| 1724 | * @next: the task we are going to switch to. |
| 1725 | * |
| 1726 | * This is called with the rq lock held and interrupts off. It must |
| 1727 | * be paired with a subsequent finish_task_switch after the context |
| 1728 | * switch. |
| 1729 | * |
| 1730 | * prepare_task_switch sets up locking and calls architecture specific |
| 1731 | * hooks. |
| 1732 | */ |
| 1733 | static inline void |
| 1734 | prepare_task_switch(struct rq *rq, struct task_struct *prev, |
| 1735 | struct task_struct *next) |
| 1736 | { |
| 1737 | trace_sched_switch(prev, next); |
| 1738 | sched_info_switch(prev, next); |
| 1739 | perf_event_task_sched_out(prev, next); |
| 1740 | fire_sched_out_preempt_notifiers(prev, next); |
| 1741 | prepare_lock_switch(rq, next); |
| 1742 | prepare_arch_switch(next); |
| 1743 | } |
| 1744 | |
| 1745 | /** |
| 1746 | * finish_task_switch - clean up after a task-switch |
| 1747 | * @rq: runqueue associated with task-switch |
| 1748 | * @prev: the thread we just switched away from. |
| 1749 | * |
| 1750 | * finish_task_switch must be called after the context switch, paired |
| 1751 | * with a prepare_task_switch call before the context switch. |
| 1752 | * finish_task_switch will reconcile locking set up by prepare_task_switch, |
| 1753 | * and do any other architecture-specific cleanup actions. |
| 1754 | * |
| 1755 | * Note that we may have delayed dropping an mm in context_switch(). If |
| 1756 | * so, we finish that here outside of the runqueue lock. (Doing it |
| 1757 | * with the lock held can cause deadlocks; see schedule() for |
| 1758 | * details.) |
| 1759 | */ |
| 1760 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) |
| 1761 | __releases(rq->lock) |
| 1762 | { |
| 1763 | struct mm_struct *mm = rq->prev_mm; |
| 1764 | long prev_state; |
| 1765 | |
| 1766 | rq->prev_mm = NULL; |
| 1767 | |
| 1768 | /* |
| 1769 | * A task struct has one reference for the use as "current". |
| 1770 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
| 1771 | * schedule one last time. The schedule call will never return, and |
| 1772 | * the scheduled task must drop that reference. |
| 1773 | * The test for TASK_DEAD must occur while the runqueue locks are |
| 1774 | * still held, otherwise prev could be scheduled on another cpu, die |
| 1775 | * there before we look at prev->state, and then the reference would |
| 1776 | * be dropped twice. |
| 1777 | * Manfred Spraul <manfred@colorfullife.com> |
| 1778 | */ |
| 1779 | prev_state = prev->state; |
| 1780 | vtime_task_switch(prev); |
| 1781 | finish_arch_switch(prev); |
| 1782 | perf_event_task_sched_in(prev, current); |
| 1783 | finish_lock_switch(rq, prev); |
| 1784 | finish_arch_post_lock_switch(); |
| 1785 | |
| 1786 | fire_sched_in_preempt_notifiers(current); |
| 1787 | if (mm) |
| 1788 | mmdrop(mm); |
| 1789 | if (unlikely(prev_state == TASK_DEAD)) { |
| 1790 | /* |
| 1791 | * Remove function-return probe instances associated with this |
| 1792 | * task and put them back on the free list. |
| 1793 | */ |
| 1794 | kprobe_flush_task(prev); |
| 1795 | put_task_struct(prev); |
| 1796 | } |
| 1797 | } |
| 1798 | |
| 1799 | #ifdef CONFIG_SMP |
| 1800 | |
| 1801 | /* assumes rq->lock is held */ |
| 1802 | static inline void pre_schedule(struct rq *rq, struct task_struct *prev) |
| 1803 | { |
| 1804 | if (prev->sched_class->pre_schedule) |
| 1805 | prev->sched_class->pre_schedule(rq, prev); |
| 1806 | } |
| 1807 | |
| 1808 | /* rq->lock is NOT held, but preemption is disabled */ |
| 1809 | static inline void post_schedule(struct rq *rq) |
| 1810 | { |
| 1811 | if (rq->post_schedule) { |
| 1812 | unsigned long flags; |
| 1813 | |
| 1814 | raw_spin_lock_irqsave(&rq->lock, flags); |
| 1815 | if (rq->curr->sched_class->post_schedule) |
| 1816 | rq->curr->sched_class->post_schedule(rq); |
| 1817 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
| 1818 | |
| 1819 | rq->post_schedule = 0; |
| 1820 | } |
| 1821 | } |
| 1822 | |
| 1823 | #else |
| 1824 | |
| 1825 | static inline void pre_schedule(struct rq *rq, struct task_struct *p) |
| 1826 | { |
| 1827 | } |
| 1828 | |
| 1829 | static inline void post_schedule(struct rq *rq) |
| 1830 | { |
| 1831 | } |
| 1832 | |
| 1833 | #endif |
| 1834 | |
| 1835 | /** |
| 1836 | * schedule_tail - first thing a freshly forked thread must call. |
| 1837 | * @prev: the thread we just switched away from. |
| 1838 | */ |
| 1839 | asmlinkage void schedule_tail(struct task_struct *prev) |
| 1840 | __releases(rq->lock) |
| 1841 | { |
| 1842 | struct rq *rq = this_rq(); |
| 1843 | |
| 1844 | finish_task_switch(rq, prev); |
| 1845 | |
| 1846 | /* |
| 1847 | * FIXME: do we need to worry about rq being invalidated by the |
| 1848 | * task_switch? |
| 1849 | */ |
| 1850 | post_schedule(rq); |
| 1851 | |
| 1852 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW |
| 1853 | /* In this case, finish_task_switch does not reenable preemption */ |
| 1854 | preempt_enable(); |
| 1855 | #endif |
| 1856 | if (current->set_child_tid) |
| 1857 | put_user(task_pid_vnr(current), current->set_child_tid); |
| 1858 | } |
| 1859 | |
| 1860 | /* |
| 1861 | * context_switch - switch to the new MM and the new |
| 1862 | * thread's register state. |
| 1863 | */ |
| 1864 | static inline void |
| 1865 | context_switch(struct rq *rq, struct task_struct *prev, |
| 1866 | struct task_struct *next) |
| 1867 | { |
| 1868 | struct mm_struct *mm, *oldmm; |
| 1869 | |
| 1870 | prepare_task_switch(rq, prev, next); |
| 1871 | |
| 1872 | mm = next->mm; |
| 1873 | oldmm = prev->active_mm; |
| 1874 | /* |
| 1875 | * For paravirt, this is coupled with an exit in switch_to to |
| 1876 | * combine the page table reload and the switch backend into |
| 1877 | * one hypercall. |
| 1878 | */ |
| 1879 | arch_start_context_switch(prev); |
| 1880 | |
| 1881 | if (!mm) { |
| 1882 | next->active_mm = oldmm; |
| 1883 | atomic_inc(&oldmm->mm_count); |
| 1884 | enter_lazy_tlb(oldmm, next); |
| 1885 | } else |
| 1886 | switch_mm(oldmm, mm, next); |
| 1887 | |
| 1888 | if (!prev->mm) { |
| 1889 | prev->active_mm = NULL; |
| 1890 | rq->prev_mm = oldmm; |
| 1891 | } |
| 1892 | /* |
| 1893 | * Since the runqueue lock will be released by the next |
| 1894 | * task (which is an invalid locking op but in the case |
| 1895 | * of the scheduler it's an obvious special-case), so we |
| 1896 | * do an early lockdep release here: |
| 1897 | */ |
| 1898 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
| 1899 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
| 1900 | #endif |
| 1901 | |
| 1902 | /* Here we just switch the register state and the stack. */ |
| 1903 | rcu_switch(prev, next); |
| 1904 | switch_to(prev, next, prev); |
| 1905 | |
| 1906 | barrier(); |
| 1907 | /* |
| 1908 | * this_rq must be evaluated again because prev may have moved |
| 1909 | * CPUs since it called schedule(), thus the 'rq' on its stack |
| 1910 | * frame will be invalid. |
| 1911 | */ |
| 1912 | finish_task_switch(this_rq(), prev); |
| 1913 | } |
| 1914 | |
| 1915 | /* |
| 1916 | * nr_running, nr_uninterruptible and nr_context_switches: |
| 1917 | * |
| 1918 | * externally visible scheduler statistics: current number of runnable |
| 1919 | * threads, current number of uninterruptible-sleeping threads, total |
| 1920 | * number of context switches performed since bootup. |
| 1921 | */ |
| 1922 | unsigned long nr_running(void) |
| 1923 | { |
| 1924 | unsigned long i, sum = 0; |
| 1925 | |
| 1926 | for_each_online_cpu(i) |
| 1927 | sum += cpu_rq(i)->nr_running; |
| 1928 | |
| 1929 | return sum; |
| 1930 | } |
| 1931 | |
| 1932 | unsigned long nr_uninterruptible(void) |
| 1933 | { |
| 1934 | unsigned long i, sum = 0; |
| 1935 | |
| 1936 | for_each_possible_cpu(i) |
| 1937 | sum += cpu_rq(i)->nr_uninterruptible; |
| 1938 | |
| 1939 | /* |
| 1940 | * Since we read the counters lockless, it might be slightly |
| 1941 | * inaccurate. Do not allow it to go below zero though: |
| 1942 | */ |
| 1943 | if (unlikely((long)sum < 0)) |
| 1944 | sum = 0; |
| 1945 | |
| 1946 | return sum; |
| 1947 | } |
| 1948 | |
| 1949 | unsigned long long nr_context_switches(void) |
| 1950 | { |
| 1951 | int i; |
| 1952 | unsigned long long sum = 0; |
| 1953 | |
| 1954 | for_each_possible_cpu(i) |
| 1955 | sum += cpu_rq(i)->nr_switches; |
| 1956 | |
| 1957 | return sum; |
| 1958 | } |
| 1959 | |
| 1960 | unsigned long nr_iowait(void) |
| 1961 | { |
| 1962 | unsigned long i, sum = 0; |
| 1963 | |
| 1964 | for_each_possible_cpu(i) |
| 1965 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
| 1966 | |
| 1967 | return sum; |
| 1968 | } |
| 1969 | |
| 1970 | unsigned long nr_iowait_cpu(int cpu) |
| 1971 | { |
| 1972 | struct rq *this = cpu_rq(cpu); |
| 1973 | return atomic_read(&this->nr_iowait); |
| 1974 | } |
| 1975 | |
| 1976 | unsigned long this_cpu_load(void) |
| 1977 | { |
| 1978 | struct rq *this = this_rq(); |
| 1979 | return this->cpu_load[0]; |
| 1980 | } |
| 1981 | |
| 1982 | |
| 1983 | /* |
| 1984 | * Global load-average calculations |
| 1985 | * |
| 1986 | * We take a distributed and async approach to calculating the global load-avg |
| 1987 | * in order to minimize overhead. |
| 1988 | * |
| 1989 | * The global load average is an exponentially decaying average of nr_running + |
| 1990 | * nr_uninterruptible. |
| 1991 | * |
| 1992 | * Once every LOAD_FREQ: |
| 1993 | * |
| 1994 | * nr_active = 0; |
| 1995 | * for_each_possible_cpu(cpu) |
| 1996 | * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; |
| 1997 | * |
| 1998 | * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) |
| 1999 | * |
| 2000 | * Due to a number of reasons the above turns in the mess below: |
| 2001 | * |
| 2002 | * - for_each_possible_cpu() is prohibitively expensive on machines with |
| 2003 | * serious number of cpus, therefore we need to take a distributed approach |
| 2004 | * to calculating nr_active. |
| 2005 | * |
| 2006 | * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 |
| 2007 | * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } |
| 2008 | * |
| 2009 | * So assuming nr_active := 0 when we start out -- true per definition, we |
| 2010 | * can simply take per-cpu deltas and fold those into a global accumulate |
| 2011 | * to obtain the same result. See calc_load_fold_active(). |
| 2012 | * |
| 2013 | * Furthermore, in order to avoid synchronizing all per-cpu delta folding |
| 2014 | * across the machine, we assume 10 ticks is sufficient time for every |
| 2015 | * cpu to have completed this task. |
| 2016 | * |
| 2017 | * This places an upper-bound on the IRQ-off latency of the machine. Then |
| 2018 | * again, being late doesn't loose the delta, just wrecks the sample. |
| 2019 | * |
| 2020 | * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because |
| 2021 | * this would add another cross-cpu cacheline miss and atomic operation |
| 2022 | * to the wakeup path. Instead we increment on whatever cpu the task ran |
| 2023 | * when it went into uninterruptible state and decrement on whatever cpu |
| 2024 | * did the wakeup. This means that only the sum of nr_uninterruptible over |
| 2025 | * all cpus yields the correct result. |
| 2026 | * |
| 2027 | * This covers the NO_HZ=n code, for extra head-aches, see the comment below. |
| 2028 | */ |
| 2029 | |
| 2030 | /* Variables and functions for calc_load */ |
| 2031 | static atomic_long_t calc_load_tasks; |
| 2032 | static unsigned long calc_load_update; |
| 2033 | unsigned long avenrun[3]; |
| 2034 | EXPORT_SYMBOL(avenrun); /* should be removed */ |
| 2035 | |
| 2036 | /** |
| 2037 | * get_avenrun - get the load average array |
| 2038 | * @loads: pointer to dest load array |
| 2039 | * @offset: offset to add |
| 2040 | * @shift: shift count to shift the result left |
| 2041 | * |
| 2042 | * These values are estimates at best, so no need for locking. |
| 2043 | */ |
| 2044 | void get_avenrun(unsigned long *loads, unsigned long offset, int shift) |
| 2045 | { |
| 2046 | loads[0] = (avenrun[0] + offset) << shift; |
| 2047 | loads[1] = (avenrun[1] + offset) << shift; |
| 2048 | loads[2] = (avenrun[2] + offset) << shift; |
| 2049 | } |
| 2050 | |
| 2051 | static long calc_load_fold_active(struct rq *this_rq) |
| 2052 | { |
| 2053 | long nr_active, delta = 0; |
| 2054 | |
| 2055 | nr_active = this_rq->nr_running; |
| 2056 | nr_active += (long) this_rq->nr_uninterruptible; |
| 2057 | |
| 2058 | if (nr_active != this_rq->calc_load_active) { |
| 2059 | delta = nr_active - this_rq->calc_load_active; |
| 2060 | this_rq->calc_load_active = nr_active; |
| 2061 | } |
| 2062 | |
| 2063 | return delta; |
| 2064 | } |
| 2065 | |
| 2066 | /* |
| 2067 | * a1 = a0 * e + a * (1 - e) |
| 2068 | */ |
| 2069 | static unsigned long |
| 2070 | calc_load(unsigned long load, unsigned long exp, unsigned long active) |
| 2071 | { |
| 2072 | load *= exp; |
| 2073 | load += active * (FIXED_1 - exp); |
| 2074 | load += 1UL << (FSHIFT - 1); |
| 2075 | return load >> FSHIFT; |
| 2076 | } |
| 2077 | |
| 2078 | #ifdef CONFIG_NO_HZ |
| 2079 | /* |
| 2080 | * Handle NO_HZ for the global load-average. |
| 2081 | * |
| 2082 | * Since the above described distributed algorithm to compute the global |
| 2083 | * load-average relies on per-cpu sampling from the tick, it is affected by |
| 2084 | * NO_HZ. |
| 2085 | * |
| 2086 | * The basic idea is to fold the nr_active delta into a global idle-delta upon |
| 2087 | * entering NO_HZ state such that we can include this as an 'extra' cpu delta |
| 2088 | * when we read the global state. |
| 2089 | * |
| 2090 | * Obviously reality has to ruin such a delightfully simple scheme: |
| 2091 | * |
| 2092 | * - When we go NO_HZ idle during the window, we can negate our sample |
| 2093 | * contribution, causing under-accounting. |
| 2094 | * |
| 2095 | * We avoid this by keeping two idle-delta counters and flipping them |
| 2096 | * when the window starts, thus separating old and new NO_HZ load. |
| 2097 | * |
| 2098 | * The only trick is the slight shift in index flip for read vs write. |
| 2099 | * |
| 2100 | * 0s 5s 10s 15s |
| 2101 | * +10 +10 +10 +10 |
| 2102 | * |-|-----------|-|-----------|-|-----------|-| |
| 2103 | * r:0 0 1 1 0 0 1 1 0 |
| 2104 | * w:0 1 1 0 0 1 1 0 0 |
| 2105 | * |
| 2106 | * This ensures we'll fold the old idle contribution in this window while |
| 2107 | * accumlating the new one. |
| 2108 | * |
| 2109 | * - When we wake up from NO_HZ idle during the window, we push up our |
| 2110 | * contribution, since we effectively move our sample point to a known |
| 2111 | * busy state. |
| 2112 | * |
| 2113 | * This is solved by pushing the window forward, and thus skipping the |
| 2114 | * sample, for this cpu (effectively using the idle-delta for this cpu which |
| 2115 | * was in effect at the time the window opened). This also solves the issue |
| 2116 | * of having to deal with a cpu having been in NOHZ idle for multiple |
| 2117 | * LOAD_FREQ intervals. |
| 2118 | * |
| 2119 | * When making the ILB scale, we should try to pull this in as well. |
| 2120 | */ |
| 2121 | static atomic_long_t calc_load_idle[2]; |
| 2122 | static int calc_load_idx; |
| 2123 | |
| 2124 | static inline int calc_load_write_idx(void) |
| 2125 | { |
| 2126 | int idx = calc_load_idx; |
| 2127 | |
| 2128 | /* |
| 2129 | * See calc_global_nohz(), if we observe the new index, we also |
| 2130 | * need to observe the new update time. |
| 2131 | */ |
| 2132 | smp_rmb(); |
| 2133 | |
| 2134 | /* |
| 2135 | * If the folding window started, make sure we start writing in the |
| 2136 | * next idle-delta. |
| 2137 | */ |
| 2138 | if (!time_before(jiffies, calc_load_update)) |
| 2139 | idx++; |
| 2140 | |
| 2141 | return idx & 1; |
| 2142 | } |
| 2143 | |
| 2144 | static inline int calc_load_read_idx(void) |
| 2145 | { |
| 2146 | return calc_load_idx & 1; |
| 2147 | } |
| 2148 | |
| 2149 | void calc_load_enter_idle(void) |
| 2150 | { |
| 2151 | struct rq *this_rq = this_rq(); |
| 2152 | long delta; |
| 2153 | |
| 2154 | /* |
| 2155 | * We're going into NOHZ mode, if there's any pending delta, fold it |
| 2156 | * into the pending idle delta. |
| 2157 | */ |
| 2158 | delta = calc_load_fold_active(this_rq); |
| 2159 | if (delta) { |
| 2160 | int idx = calc_load_write_idx(); |
| 2161 | atomic_long_add(delta, &calc_load_idle[idx]); |
| 2162 | } |
| 2163 | } |
| 2164 | |
| 2165 | void calc_load_exit_idle(void) |
| 2166 | { |
| 2167 | struct rq *this_rq = this_rq(); |
| 2168 | |
| 2169 | /* |
| 2170 | * If we're still before the sample window, we're done. |
| 2171 | */ |
| 2172 | if (time_before(jiffies, this_rq->calc_load_update)) |
| 2173 | return; |
| 2174 | |
| 2175 | /* |
| 2176 | * We woke inside or after the sample window, this means we're already |
| 2177 | * accounted through the nohz accounting, so skip the entire deal and |
| 2178 | * sync up for the next window. |
| 2179 | */ |
| 2180 | this_rq->calc_load_update = calc_load_update; |
| 2181 | if (time_before(jiffies, this_rq->calc_load_update + 10)) |
| 2182 | this_rq->calc_load_update += LOAD_FREQ; |
| 2183 | } |
| 2184 | |
| 2185 | static long calc_load_fold_idle(void) |
| 2186 | { |
| 2187 | int idx = calc_load_read_idx(); |
| 2188 | long delta = 0; |
| 2189 | |
| 2190 | if (atomic_long_read(&calc_load_idle[idx])) |
| 2191 | delta = atomic_long_xchg(&calc_load_idle[idx], 0); |
| 2192 | |
| 2193 | return delta; |
| 2194 | } |
| 2195 | |
| 2196 | /** |
| 2197 | * fixed_power_int - compute: x^n, in O(log n) time |
| 2198 | * |
| 2199 | * @x: base of the power |
| 2200 | * @frac_bits: fractional bits of @x |
| 2201 | * @n: power to raise @x to. |
| 2202 | * |
| 2203 | * By exploiting the relation between the definition of the natural power |
| 2204 | * function: x^n := x*x*...*x (x multiplied by itself for n times), and |
| 2205 | * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, |
| 2206 | * (where: n_i \elem {0, 1}, the binary vector representing n), |
| 2207 | * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is |
| 2208 | * of course trivially computable in O(log_2 n), the length of our binary |
| 2209 | * vector. |
| 2210 | */ |
| 2211 | static unsigned long |
| 2212 | fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) |
| 2213 | { |
| 2214 | unsigned long result = 1UL << frac_bits; |
| 2215 | |
| 2216 | if (n) for (;;) { |
| 2217 | if (n & 1) { |
| 2218 | result *= x; |
| 2219 | result += 1UL << (frac_bits - 1); |
| 2220 | result >>= frac_bits; |
| 2221 | } |
| 2222 | n >>= 1; |
| 2223 | if (!n) |
| 2224 | break; |
| 2225 | x *= x; |
| 2226 | x += 1UL << (frac_bits - 1); |
| 2227 | x >>= frac_bits; |
| 2228 | } |
| 2229 | |
| 2230 | return result; |
| 2231 | } |
| 2232 | |
| 2233 | /* |
| 2234 | * a1 = a0 * e + a * (1 - e) |
| 2235 | * |
| 2236 | * a2 = a1 * e + a * (1 - e) |
| 2237 | * = (a0 * e + a * (1 - e)) * e + a * (1 - e) |
| 2238 | * = a0 * e^2 + a * (1 - e) * (1 + e) |
| 2239 | * |
| 2240 | * a3 = a2 * e + a * (1 - e) |
| 2241 | * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) |
| 2242 | * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) |
| 2243 | * |
| 2244 | * ... |
| 2245 | * |
| 2246 | * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] |
| 2247 | * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) |
| 2248 | * = a0 * e^n + a * (1 - e^n) |
| 2249 | * |
| 2250 | * [1] application of the geometric series: |
| 2251 | * |
| 2252 | * n 1 - x^(n+1) |
| 2253 | * S_n := \Sum x^i = ------------- |
| 2254 | * i=0 1 - x |
| 2255 | */ |
| 2256 | static unsigned long |
| 2257 | calc_load_n(unsigned long load, unsigned long exp, |
| 2258 | unsigned long active, unsigned int n) |
| 2259 | { |
| 2260 | |
| 2261 | return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); |
| 2262 | } |
| 2263 | |
| 2264 | /* |
| 2265 | * NO_HZ can leave us missing all per-cpu ticks calling |
| 2266 | * calc_load_account_active(), but since an idle CPU folds its delta into |
| 2267 | * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold |
| 2268 | * in the pending idle delta if our idle period crossed a load cycle boundary. |
| 2269 | * |
| 2270 | * Once we've updated the global active value, we need to apply the exponential |
| 2271 | * weights adjusted to the number of cycles missed. |
| 2272 | */ |
| 2273 | static void calc_global_nohz(void) |
| 2274 | { |
| 2275 | long delta, active, n; |
| 2276 | |
| 2277 | if (!time_before(jiffies, calc_load_update + 10)) { |
| 2278 | /* |
| 2279 | * Catch-up, fold however many we are behind still |
| 2280 | */ |
| 2281 | delta = jiffies - calc_load_update - 10; |
| 2282 | n = 1 + (delta / LOAD_FREQ); |
| 2283 | |
| 2284 | active = atomic_long_read(&calc_load_tasks); |
| 2285 | active = active > 0 ? active * FIXED_1 : 0; |
| 2286 | |
| 2287 | avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); |
| 2288 | avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); |
| 2289 | avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); |
| 2290 | |
| 2291 | calc_load_update += n * LOAD_FREQ; |
| 2292 | } |
| 2293 | |
| 2294 | /* |
| 2295 | * Flip the idle index... |
| 2296 | * |
| 2297 | * Make sure we first write the new time then flip the index, so that |
| 2298 | * calc_load_write_idx() will see the new time when it reads the new |
| 2299 | * index, this avoids a double flip messing things up. |
| 2300 | */ |
| 2301 | smp_wmb(); |
| 2302 | calc_load_idx++; |
| 2303 | } |
| 2304 | #else /* !CONFIG_NO_HZ */ |
| 2305 | |
| 2306 | static inline long calc_load_fold_idle(void) { return 0; } |
| 2307 | static inline void calc_global_nohz(void) { } |
| 2308 | |
| 2309 | #endif /* CONFIG_NO_HZ */ |
| 2310 | |
| 2311 | /* |
| 2312 | * calc_load - update the avenrun load estimates 10 ticks after the |
| 2313 | * CPUs have updated calc_load_tasks. |
| 2314 | */ |
| 2315 | void calc_global_load(unsigned long ticks) |
| 2316 | { |
| 2317 | long active, delta; |
| 2318 | |
| 2319 | if (time_before(jiffies, calc_load_update + 10)) |
| 2320 | return; |
| 2321 | |
| 2322 | /* |
| 2323 | * Fold the 'old' idle-delta to include all NO_HZ cpus. |
| 2324 | */ |
| 2325 | delta = calc_load_fold_idle(); |
| 2326 | if (delta) |
| 2327 | atomic_long_add(delta, &calc_load_tasks); |
| 2328 | |
| 2329 | active = atomic_long_read(&calc_load_tasks); |
| 2330 | active = active > 0 ? active * FIXED_1 : 0; |
| 2331 | |
| 2332 | avenrun[0] = calc_load(avenrun[0], EXP_1, active); |
| 2333 | avenrun[1] = calc_load(avenrun[1], EXP_5, active); |
| 2334 | avenrun[2] = calc_load(avenrun[2], EXP_15, active); |
| 2335 | |
| 2336 | calc_load_update += LOAD_FREQ; |
| 2337 | |
| 2338 | /* |
| 2339 | * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk. |
| 2340 | */ |
| 2341 | calc_global_nohz(); |
| 2342 | } |
| 2343 | |
| 2344 | /* |
| 2345 | * Called from update_cpu_load() to periodically update this CPU's |
| 2346 | * active count. |
| 2347 | */ |
| 2348 | static void calc_load_account_active(struct rq *this_rq) |
| 2349 | { |
| 2350 | long delta; |
| 2351 | |
| 2352 | if (time_before(jiffies, this_rq->calc_load_update)) |
| 2353 | return; |
| 2354 | |
| 2355 | delta = calc_load_fold_active(this_rq); |
| 2356 | if (delta) |
| 2357 | atomic_long_add(delta, &calc_load_tasks); |
| 2358 | |
| 2359 | this_rq->calc_load_update += LOAD_FREQ; |
| 2360 | } |
| 2361 | |
| 2362 | /* |
| 2363 | * End of global load-average stuff |
| 2364 | */ |
| 2365 | |
| 2366 | /* |
| 2367 | * The exact cpuload at various idx values, calculated at every tick would be |
| 2368 | * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load |
| 2369 | * |
| 2370 | * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called |
| 2371 | * on nth tick when cpu may be busy, then we have: |
| 2372 | * load = ((2^idx - 1) / 2^idx)^(n-1) * load |
| 2373 | * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load |
| 2374 | * |
| 2375 | * decay_load_missed() below does efficient calculation of |
| 2376 | * load = ((2^idx - 1) / 2^idx)^(n-1) * load |
| 2377 | * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load |
| 2378 | * |
| 2379 | * The calculation is approximated on a 128 point scale. |
| 2380 | * degrade_zero_ticks is the number of ticks after which load at any |
| 2381 | * particular idx is approximated to be zero. |
| 2382 | * degrade_factor is a precomputed table, a row for each load idx. |
| 2383 | * Each column corresponds to degradation factor for a power of two ticks, |
| 2384 | * based on 128 point scale. |
| 2385 | * Example: |
| 2386 | * row 2, col 3 (=12) says that the degradation at load idx 2 after |
| 2387 | * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). |
| 2388 | * |
| 2389 | * With this power of 2 load factors, we can degrade the load n times |
| 2390 | * by looking at 1 bits in n and doing as many mult/shift instead of |
| 2391 | * n mult/shifts needed by the exact degradation. |
| 2392 | */ |
| 2393 | #define DEGRADE_SHIFT 7 |
| 2394 | static const unsigned char |
| 2395 | degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; |
| 2396 | static const unsigned char |
| 2397 | degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { |
| 2398 | {0, 0, 0, 0, 0, 0, 0, 0}, |
| 2399 | {64, 32, 8, 0, 0, 0, 0, 0}, |
| 2400 | {96, 72, 40, 12, 1, 0, 0}, |
| 2401 | {112, 98, 75, 43, 15, 1, 0}, |
| 2402 | {120, 112, 98, 76, 45, 16, 2} }; |
| 2403 | |
| 2404 | /* |
| 2405 | * Update cpu_load for any missed ticks, due to tickless idle. The backlog |
| 2406 | * would be when CPU is idle and so we just decay the old load without |
| 2407 | * adding any new load. |
| 2408 | */ |
| 2409 | static unsigned long |
| 2410 | decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) |
| 2411 | { |
| 2412 | int j = 0; |
| 2413 | |
| 2414 | if (!missed_updates) |
| 2415 | return load; |
| 2416 | |
| 2417 | if (missed_updates >= degrade_zero_ticks[idx]) |
| 2418 | return 0; |
| 2419 | |
| 2420 | if (idx == 1) |
| 2421 | return load >> missed_updates; |
| 2422 | |
| 2423 | while (missed_updates) { |
| 2424 | if (missed_updates % 2) |
| 2425 | load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; |
| 2426 | |
| 2427 | missed_updates >>= 1; |
| 2428 | j++; |
| 2429 | } |
| 2430 | return load; |
| 2431 | } |
| 2432 | |
| 2433 | /* |
| 2434 | * Update rq->cpu_load[] statistics. This function is usually called every |
| 2435 | * scheduler tick (TICK_NSEC). With tickless idle this will not be called |
| 2436 | * every tick. We fix it up based on jiffies. |
| 2437 | */ |
| 2438 | static void __update_cpu_load(struct rq *this_rq, unsigned long this_load, |
| 2439 | unsigned long pending_updates) |
| 2440 | { |
| 2441 | int i, scale; |
| 2442 | |
| 2443 | this_rq->nr_load_updates++; |
| 2444 | |
| 2445 | /* Update our load: */ |
| 2446 | this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ |
| 2447 | for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { |
| 2448 | unsigned long old_load, new_load; |
| 2449 | |
| 2450 | /* scale is effectively 1 << i now, and >> i divides by scale */ |
| 2451 | |
| 2452 | old_load = this_rq->cpu_load[i]; |
| 2453 | old_load = decay_load_missed(old_load, pending_updates - 1, i); |
| 2454 | new_load = this_load; |
| 2455 | /* |
| 2456 | * Round up the averaging division if load is increasing. This |
| 2457 | * prevents us from getting stuck on 9 if the load is 10, for |
| 2458 | * example. |
| 2459 | */ |
| 2460 | if (new_load > old_load) |
| 2461 | new_load += scale - 1; |
| 2462 | |
| 2463 | this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; |
| 2464 | } |
| 2465 | |
| 2466 | sched_avg_update(this_rq); |
| 2467 | } |
| 2468 | |
| 2469 | #ifdef CONFIG_NO_HZ |
| 2470 | /* |
| 2471 | * There is no sane way to deal with nohz on smp when using jiffies because the |
| 2472 | * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading |
| 2473 | * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. |
| 2474 | * |
| 2475 | * Therefore we cannot use the delta approach from the regular tick since that |
| 2476 | * would seriously skew the load calculation. However we'll make do for those |
| 2477 | * updates happening while idle (nohz_idle_balance) or coming out of idle |
| 2478 | * (tick_nohz_idle_exit). |
| 2479 | * |
| 2480 | * This means we might still be one tick off for nohz periods. |
| 2481 | */ |
| 2482 | |
| 2483 | /* |
| 2484 | * Called from nohz_idle_balance() to update the load ratings before doing the |
| 2485 | * idle balance. |
| 2486 | */ |
| 2487 | void update_idle_cpu_load(struct rq *this_rq) |
| 2488 | { |
| 2489 | unsigned long curr_jiffies = ACCESS_ONCE(jiffies); |
| 2490 | unsigned long load = this_rq->load.weight; |
| 2491 | unsigned long pending_updates; |
| 2492 | |
| 2493 | /* |
| 2494 | * bail if there's load or we're actually up-to-date. |
| 2495 | */ |
| 2496 | if (load || curr_jiffies == this_rq->last_load_update_tick) |
| 2497 | return; |
| 2498 | |
| 2499 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; |
| 2500 | this_rq->last_load_update_tick = curr_jiffies; |
| 2501 | |
| 2502 | __update_cpu_load(this_rq, load, pending_updates); |
| 2503 | } |
| 2504 | |
| 2505 | /* |
| 2506 | * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed. |
| 2507 | */ |
| 2508 | void update_cpu_load_nohz(void) |
| 2509 | { |
| 2510 | struct rq *this_rq = this_rq(); |
| 2511 | unsigned long curr_jiffies = ACCESS_ONCE(jiffies); |
| 2512 | unsigned long pending_updates; |
| 2513 | |
| 2514 | if (curr_jiffies == this_rq->last_load_update_tick) |
| 2515 | return; |
| 2516 | |
| 2517 | raw_spin_lock(&this_rq->lock); |
| 2518 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; |
| 2519 | if (pending_updates) { |
| 2520 | this_rq->last_load_update_tick = curr_jiffies; |
| 2521 | /* |
| 2522 | * We were idle, this means load 0, the current load might be |
| 2523 | * !0 due to remote wakeups and the sort. |
| 2524 | */ |
| 2525 | __update_cpu_load(this_rq, 0, pending_updates); |
| 2526 | } |
| 2527 | raw_spin_unlock(&this_rq->lock); |
| 2528 | } |
| 2529 | #endif /* CONFIG_NO_HZ */ |
| 2530 | |
| 2531 | /* |
| 2532 | * Called from scheduler_tick() |
| 2533 | */ |
| 2534 | static void update_cpu_load_active(struct rq *this_rq) |
| 2535 | { |
| 2536 | /* |
| 2537 | * See the mess around update_idle_cpu_load() / update_cpu_load_nohz(). |
| 2538 | */ |
| 2539 | this_rq->last_load_update_tick = jiffies; |
| 2540 | __update_cpu_load(this_rq, this_rq->load.weight, 1); |
| 2541 | |
| 2542 | calc_load_account_active(this_rq); |
| 2543 | } |
| 2544 | |
| 2545 | #ifdef CONFIG_SMP |
| 2546 | |
| 2547 | /* |
| 2548 | * sched_exec - execve() is a valuable balancing opportunity, because at |
| 2549 | * this point the task has the smallest effective memory and cache footprint. |
| 2550 | */ |
| 2551 | void sched_exec(void) |
| 2552 | { |
| 2553 | struct task_struct *p = current; |
| 2554 | unsigned long flags; |
| 2555 | int dest_cpu; |
| 2556 | |
| 2557 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
| 2558 | dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); |
| 2559 | if (dest_cpu == smp_processor_id()) |
| 2560 | goto unlock; |
| 2561 | |
| 2562 | if (likely(cpu_active(dest_cpu))) { |
| 2563 | struct migration_arg arg = { p, dest_cpu }; |
| 2564 | |
| 2565 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
| 2566 | stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); |
| 2567 | return; |
| 2568 | } |
| 2569 | unlock: |
| 2570 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
| 2571 | } |
| 2572 | |
| 2573 | #endif |
| 2574 | |
| 2575 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
| 2576 | DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); |
| 2577 | |
| 2578 | EXPORT_PER_CPU_SYMBOL(kstat); |
| 2579 | EXPORT_PER_CPU_SYMBOL(kernel_cpustat); |
| 2580 | |
| 2581 | /* |
| 2582 | * Return any ns on the sched_clock that have not yet been accounted in |
| 2583 | * @p in case that task is currently running. |
| 2584 | * |
| 2585 | * Called with task_rq_lock() held on @rq. |
| 2586 | */ |
| 2587 | static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) |
| 2588 | { |
| 2589 | u64 ns = 0; |
| 2590 | |
| 2591 | if (task_current(rq, p)) { |
| 2592 | update_rq_clock(rq); |
| 2593 | ns = rq->clock_task - p->se.exec_start; |
| 2594 | if ((s64)ns < 0) |
| 2595 | ns = 0; |
| 2596 | } |
| 2597 | |
| 2598 | return ns; |
| 2599 | } |
| 2600 | |
| 2601 | unsigned long long task_delta_exec(struct task_struct *p) |
| 2602 | { |
| 2603 | unsigned long flags; |
| 2604 | struct rq *rq; |
| 2605 | u64 ns = 0; |
| 2606 | |
| 2607 | rq = task_rq_lock(p, &flags); |
| 2608 | ns = do_task_delta_exec(p, rq); |
| 2609 | task_rq_unlock(rq, p, &flags); |
| 2610 | |
| 2611 | return ns; |
| 2612 | } |
| 2613 | |
| 2614 | /* |
| 2615 | * Return accounted runtime for the task. |
| 2616 | * In case the task is currently running, return the runtime plus current's |
| 2617 | * pending runtime that have not been accounted yet. |
| 2618 | */ |
| 2619 | unsigned long long task_sched_runtime(struct task_struct *p) |
| 2620 | { |
| 2621 | unsigned long flags; |
| 2622 | struct rq *rq; |
| 2623 | u64 ns = 0; |
| 2624 | |
| 2625 | rq = task_rq_lock(p, &flags); |
| 2626 | ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); |
| 2627 | task_rq_unlock(rq, p, &flags); |
| 2628 | |
| 2629 | return ns; |
| 2630 | } |
| 2631 | |
| 2632 | /* |
| 2633 | * This function gets called by the timer code, with HZ frequency. |
| 2634 | * We call it with interrupts disabled. |
| 2635 | */ |
| 2636 | void scheduler_tick(void) |
| 2637 | { |
| 2638 | int cpu = smp_processor_id(); |
| 2639 | struct rq *rq = cpu_rq(cpu); |
| 2640 | struct task_struct *curr = rq->curr; |
| 2641 | |
| 2642 | sched_clock_tick(); |
| 2643 | |
| 2644 | raw_spin_lock(&rq->lock); |
| 2645 | update_rq_clock(rq); |
| 2646 | update_cpu_load_active(rq); |
| 2647 | curr->sched_class->task_tick(rq, curr, 0); |
| 2648 | raw_spin_unlock(&rq->lock); |
| 2649 | |
| 2650 | perf_event_task_tick(); |
| 2651 | |
| 2652 | #ifdef CONFIG_SMP |
| 2653 | rq->idle_balance = idle_cpu(cpu); |
| 2654 | trigger_load_balance(rq, cpu); |
| 2655 | #endif |
| 2656 | } |
| 2657 | |
| 2658 | notrace unsigned long get_parent_ip(unsigned long addr) |
| 2659 | { |
| 2660 | if (in_lock_functions(addr)) { |
| 2661 | addr = CALLER_ADDR2; |
| 2662 | if (in_lock_functions(addr)) |
| 2663 | addr = CALLER_ADDR3; |
| 2664 | } |
| 2665 | return addr; |
| 2666 | } |
| 2667 | |
| 2668 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ |
| 2669 | defined(CONFIG_PREEMPT_TRACER)) |
| 2670 | |
| 2671 | void __kprobes add_preempt_count(int val) |
| 2672 | { |
| 2673 | #ifdef CONFIG_DEBUG_PREEMPT |
| 2674 | /* |
| 2675 | * Underflow? |
| 2676 | */ |
| 2677 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
| 2678 | return; |
| 2679 | #endif |
| 2680 | preempt_count() += val; |
| 2681 | #ifdef CONFIG_DEBUG_PREEMPT |
| 2682 | /* |
| 2683 | * Spinlock count overflowing soon? |
| 2684 | */ |
| 2685 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
| 2686 | PREEMPT_MASK - 10); |
| 2687 | #endif |
| 2688 | if (preempt_count() == val) |
| 2689 | trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); |
| 2690 | } |
| 2691 | EXPORT_SYMBOL(add_preempt_count); |
| 2692 | |
| 2693 | void __kprobes sub_preempt_count(int val) |
| 2694 | { |
| 2695 | #ifdef CONFIG_DEBUG_PREEMPT |
| 2696 | /* |
| 2697 | * Underflow? |
| 2698 | */ |
| 2699 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
| 2700 | return; |
| 2701 | /* |
| 2702 | * Is the spinlock portion underflowing? |
| 2703 | */ |
| 2704 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
| 2705 | !(preempt_count() & PREEMPT_MASK))) |
| 2706 | return; |
| 2707 | #endif |
| 2708 | |
| 2709 | if (preempt_count() == val) |
| 2710 | trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); |
| 2711 | preempt_count() -= val; |
| 2712 | } |
| 2713 | EXPORT_SYMBOL(sub_preempt_count); |
| 2714 | |
| 2715 | #endif |
| 2716 | |
| 2717 | /* |
| 2718 | * Print scheduling while atomic bug: |
| 2719 | */ |
| 2720 | static noinline void __schedule_bug(struct task_struct *prev) |
| 2721 | { |
| 2722 | if (oops_in_progress) |
| 2723 | return; |
| 2724 | |
| 2725 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", |
| 2726 | prev->comm, prev->pid, preempt_count()); |
| 2727 | |
| 2728 | debug_show_held_locks(prev); |
| 2729 | print_modules(); |
| 2730 | if (irqs_disabled()) |
| 2731 | print_irqtrace_events(prev); |
| 2732 | dump_stack(); |
| 2733 | add_taint(TAINT_WARN); |
| 2734 | } |
| 2735 | |
| 2736 | /* |
| 2737 | * Various schedule()-time debugging checks and statistics: |
| 2738 | */ |
| 2739 | static inline void schedule_debug(struct task_struct *prev) |
| 2740 | { |
| 2741 | /* |
| 2742 | * Test if we are atomic. Since do_exit() needs to call into |
| 2743 | * schedule() atomically, we ignore that path for now. |
| 2744 | * Otherwise, whine if we are scheduling when we should not be. |
| 2745 | */ |
| 2746 | if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) |
| 2747 | __schedule_bug(prev); |
| 2748 | rcu_sleep_check(); |
| 2749 | |
| 2750 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
| 2751 | |
| 2752 | schedstat_inc(this_rq(), sched_count); |
| 2753 | } |
| 2754 | |
| 2755 | static void put_prev_task(struct rq *rq, struct task_struct *prev) |
| 2756 | { |
| 2757 | if (prev->on_rq || rq->skip_clock_update < 0) |
| 2758 | update_rq_clock(rq); |
| 2759 | prev->sched_class->put_prev_task(rq, prev); |
| 2760 | } |
| 2761 | |
| 2762 | /* |
| 2763 | * Pick up the highest-prio task: |
| 2764 | */ |
| 2765 | static inline struct task_struct * |
| 2766 | pick_next_task(struct rq *rq) |
| 2767 | { |
| 2768 | const struct sched_class *class; |
| 2769 | struct task_struct *p; |
| 2770 | |
| 2771 | /* |
| 2772 | * Optimization: we know that if all tasks are in |
| 2773 | * the fair class we can call that function directly: |
| 2774 | */ |
| 2775 | if (likely(rq->nr_running == rq->cfs.h_nr_running)) { |
| 2776 | p = fair_sched_class.pick_next_task(rq); |
| 2777 | if (likely(p)) |
| 2778 | return p; |
| 2779 | } |
| 2780 | |
| 2781 | for_each_class(class) { |
| 2782 | p = class->pick_next_task(rq); |
| 2783 | if (p) |
| 2784 | return p; |
| 2785 | } |
| 2786 | |
| 2787 | BUG(); /* the idle class will always have a runnable task */ |
| 2788 | } |
| 2789 | |
| 2790 | /* |
| 2791 | * __schedule() is the main scheduler function. |
| 2792 | * |
| 2793 | * The main means of driving the scheduler and thus entering this function are: |
| 2794 | * |
| 2795 | * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. |
| 2796 | * |
| 2797 | * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return |
| 2798 | * paths. For example, see arch/x86/entry_64.S. |
| 2799 | * |
| 2800 | * To drive preemption between tasks, the scheduler sets the flag in timer |
| 2801 | * interrupt handler scheduler_tick(). |
| 2802 | * |
| 2803 | * 3. Wakeups don't really cause entry into schedule(). They add a |
| 2804 | * task to the run-queue and that's it. |
| 2805 | * |
| 2806 | * Now, if the new task added to the run-queue preempts the current |
| 2807 | * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets |
| 2808 | * called on the nearest possible occasion: |
| 2809 | * |
| 2810 | * - If the kernel is preemptible (CONFIG_PREEMPT=y): |
| 2811 | * |
| 2812 | * - in syscall or exception context, at the next outmost |
| 2813 | * preempt_enable(). (this might be as soon as the wake_up()'s |
| 2814 | * spin_unlock()!) |
| 2815 | * |
| 2816 | * - in IRQ context, return from interrupt-handler to |
| 2817 | * preemptible context |
| 2818 | * |
| 2819 | * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) |
| 2820 | * then at the next: |
| 2821 | * |
| 2822 | * - cond_resched() call |
| 2823 | * - explicit schedule() call |
| 2824 | * - return from syscall or exception to user-space |
| 2825 | * - return from interrupt-handler to user-space |
| 2826 | */ |
| 2827 | static void __sched __schedule(void) |
| 2828 | { |
| 2829 | struct task_struct *prev, *next; |
| 2830 | unsigned long *switch_count; |
| 2831 | struct rq *rq; |
| 2832 | int cpu; |
| 2833 | |
| 2834 | need_resched: |
| 2835 | preempt_disable(); |
| 2836 | cpu = smp_processor_id(); |
| 2837 | rq = cpu_rq(cpu); |
| 2838 | rcu_note_context_switch(cpu); |
| 2839 | prev = rq->curr; |
| 2840 | |
| 2841 | schedule_debug(prev); |
| 2842 | |
| 2843 | if (sched_feat(HRTICK)) |
| 2844 | hrtick_clear(rq); |
| 2845 | |
| 2846 | raw_spin_lock_irq(&rq->lock); |
| 2847 | |
| 2848 | switch_count = &prev->nivcsw; |
| 2849 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
| 2850 | if (unlikely(signal_pending_state(prev->state, prev))) { |
| 2851 | prev->state = TASK_RUNNING; |
| 2852 | } else { |
| 2853 | deactivate_task(rq, prev, DEQUEUE_SLEEP); |
| 2854 | prev->on_rq = 0; |
| 2855 | |
| 2856 | /* |
| 2857 | * If a worker went to sleep, notify and ask workqueue |
| 2858 | * whether it wants to wake up a task to maintain |
| 2859 | * concurrency. |
| 2860 | */ |
| 2861 | if (prev->flags & PF_WQ_WORKER) { |
| 2862 | struct task_struct *to_wakeup; |
| 2863 | |
| 2864 | to_wakeup = wq_worker_sleeping(prev, cpu); |
| 2865 | if (to_wakeup) |
| 2866 | try_to_wake_up_local(to_wakeup); |
| 2867 | } |
| 2868 | } |
| 2869 | switch_count = &prev->nvcsw; |
| 2870 | } |
| 2871 | |
| 2872 | pre_schedule(rq, prev); |
| 2873 | |
| 2874 | if (unlikely(!rq->nr_running)) |
| 2875 | idle_balance(cpu, rq); |
| 2876 | |
| 2877 | put_prev_task(rq, prev); |
| 2878 | next = pick_next_task(rq); |
| 2879 | clear_tsk_need_resched(prev); |
| 2880 | rq->skip_clock_update = 0; |
| 2881 | |
| 2882 | if (likely(prev != next)) { |
| 2883 | rq->nr_switches++; |
| 2884 | rq->curr = next; |
| 2885 | ++*switch_count; |
| 2886 | |
| 2887 | context_switch(rq, prev, next); /* unlocks the rq */ |
| 2888 | /* |
| 2889 | * The context switch have flipped the stack from under us |
| 2890 | * and restored the local variables which were saved when |
| 2891 | * this task called schedule() in the past. prev == current |
| 2892 | * is still correct, but it can be moved to another cpu/rq. |
| 2893 | */ |
| 2894 | cpu = smp_processor_id(); |
| 2895 | rq = cpu_rq(cpu); |
| 2896 | } else |
| 2897 | raw_spin_unlock_irq(&rq->lock); |
| 2898 | |
| 2899 | post_schedule(rq); |
| 2900 | |
| 2901 | sched_preempt_enable_no_resched(); |
| 2902 | if (need_resched()) |
| 2903 | goto need_resched; |
| 2904 | } |
| 2905 | |
| 2906 | static inline void sched_submit_work(struct task_struct *tsk) |
| 2907 | { |
| 2908 | if (!tsk->state || tsk_is_pi_blocked(tsk)) |
| 2909 | return; |
| 2910 | /* |
| 2911 | * If we are going to sleep and we have plugged IO queued, |
| 2912 | * make sure to submit it to avoid deadlocks. |
| 2913 | */ |
| 2914 | if (blk_needs_flush_plug(tsk)) |
| 2915 | blk_schedule_flush_plug(tsk); |
| 2916 | } |
| 2917 | |
| 2918 | asmlinkage void __sched schedule(void) |
| 2919 | { |
| 2920 | struct task_struct *tsk = current; |
| 2921 | |
| 2922 | sched_submit_work(tsk); |
| 2923 | __schedule(); |
| 2924 | } |
| 2925 | EXPORT_SYMBOL(schedule); |
| 2926 | |
| 2927 | #ifdef CONFIG_RCU_USER_QS |
| 2928 | asmlinkage void __sched schedule_user(void) |
| 2929 | { |
| 2930 | /* |
| 2931 | * If we come here after a random call to set_need_resched(), |
| 2932 | * or we have been woken up remotely but the IPI has not yet arrived, |
| 2933 | * we haven't yet exited the RCU idle mode. Do it here manually until |
| 2934 | * we find a better solution. |
| 2935 | */ |
| 2936 | rcu_user_exit(); |
| 2937 | schedule(); |
| 2938 | rcu_user_enter(); |
| 2939 | } |
| 2940 | #endif |
| 2941 | |
| 2942 | /** |
| 2943 | * schedule_preempt_disabled - called with preemption disabled |
| 2944 | * |
| 2945 | * Returns with preemption disabled. Note: preempt_count must be 1 |
| 2946 | */ |
| 2947 | void __sched schedule_preempt_disabled(void) |
| 2948 | { |
| 2949 | sched_preempt_enable_no_resched(); |
| 2950 | schedule(); |
| 2951 | preempt_disable(); |
| 2952 | } |
| 2953 | |
| 2954 | #ifdef CONFIG_MUTEX_SPIN_ON_OWNER |
| 2955 | |
| 2956 | static inline bool owner_running(struct mutex *lock, struct task_struct *owner) |
| 2957 | { |
| 2958 | if (lock->owner != owner) |
| 2959 | return false; |
| 2960 | |
| 2961 | /* |
| 2962 | * Ensure we emit the owner->on_cpu, dereference _after_ checking |
| 2963 | * lock->owner still matches owner, if that fails, owner might |
| 2964 | * point to free()d memory, if it still matches, the rcu_read_lock() |
| 2965 | * ensures the memory stays valid. |
| 2966 | */ |
| 2967 | barrier(); |
| 2968 | |
| 2969 | return owner->on_cpu; |
| 2970 | } |
| 2971 | |
| 2972 | /* |
| 2973 | * Look out! "owner" is an entirely speculative pointer |
| 2974 | * access and not reliable. |
| 2975 | */ |
| 2976 | int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) |
| 2977 | { |
| 2978 | if (!sched_feat(OWNER_SPIN)) |
| 2979 | return 0; |
| 2980 | |
| 2981 | rcu_read_lock(); |
| 2982 | while (owner_running(lock, owner)) { |
| 2983 | if (need_resched()) |
| 2984 | break; |
| 2985 | |
| 2986 | arch_mutex_cpu_relax(); |
| 2987 | } |
| 2988 | rcu_read_unlock(); |
| 2989 | |
| 2990 | /* |
| 2991 | * We break out the loop above on need_resched() and when the |
| 2992 | * owner changed, which is a sign for heavy contention. Return |
| 2993 | * success only when lock->owner is NULL. |
| 2994 | */ |
| 2995 | return lock->owner == NULL; |
| 2996 | } |
| 2997 | #endif |
| 2998 | |
| 2999 | #ifdef CONFIG_PREEMPT |
| 3000 | /* |
| 3001 | * this is the entry point to schedule() from in-kernel preemption |
| 3002 | * off of preempt_enable. Kernel preemptions off return from interrupt |
| 3003 | * occur there and call schedule directly. |
| 3004 | */ |
| 3005 | asmlinkage void __sched notrace preempt_schedule(void) |
| 3006 | { |
| 3007 | struct thread_info *ti = current_thread_info(); |
| 3008 | |
| 3009 | /* |
| 3010 | * If there is a non-zero preempt_count or interrupts are disabled, |
| 3011 | * we do not want to preempt the current task. Just return.. |
| 3012 | */ |
| 3013 | if (likely(ti->preempt_count || irqs_disabled())) |
| 3014 | return; |
| 3015 | |
| 3016 | do { |
| 3017 | add_preempt_count_notrace(PREEMPT_ACTIVE); |
| 3018 | __schedule(); |
| 3019 | sub_preempt_count_notrace(PREEMPT_ACTIVE); |
| 3020 | |
| 3021 | /* |
| 3022 | * Check again in case we missed a preemption opportunity |
| 3023 | * between schedule and now. |
| 3024 | */ |
| 3025 | barrier(); |
| 3026 | } while (need_resched()); |
| 3027 | } |
| 3028 | EXPORT_SYMBOL(preempt_schedule); |
| 3029 | |
| 3030 | /* |
| 3031 | * this is the entry point to schedule() from kernel preemption |
| 3032 | * off of irq context. |
| 3033 | * Note, that this is called and return with irqs disabled. This will |
| 3034 | * protect us against recursive calling from irq. |
| 3035 | */ |
| 3036 | asmlinkage void __sched preempt_schedule_irq(void) |
| 3037 | { |
| 3038 | struct thread_info *ti = current_thread_info(); |
| 3039 | |
| 3040 | /* Catch callers which need to be fixed */ |
| 3041 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
| 3042 | |
| 3043 | rcu_user_exit(); |
| 3044 | do { |
| 3045 | add_preempt_count(PREEMPT_ACTIVE); |
| 3046 | local_irq_enable(); |
| 3047 | __schedule(); |
| 3048 | local_irq_disable(); |
| 3049 | sub_preempt_count(PREEMPT_ACTIVE); |
| 3050 | |
| 3051 | /* |
| 3052 | * Check again in case we missed a preemption opportunity |
| 3053 | * between schedule and now. |
| 3054 | */ |
| 3055 | barrier(); |
| 3056 | } while (need_resched()); |
| 3057 | } |
| 3058 | |
| 3059 | #endif /* CONFIG_PREEMPT */ |
| 3060 | |
| 3061 | int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, |
| 3062 | void *key) |
| 3063 | { |
| 3064 | return try_to_wake_up(curr->private, mode, wake_flags); |
| 3065 | } |
| 3066 | EXPORT_SYMBOL(default_wake_function); |
| 3067 | |
| 3068 | /* |
| 3069 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
| 3070 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve |
| 3071 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
| 3072 | * |
| 3073 | * There are circumstances in which we can try to wake a task which has already |
| 3074 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
| 3075 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
| 3076 | */ |
| 3077 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, |
| 3078 | int nr_exclusive, int wake_flags, void *key) |
| 3079 | { |
| 3080 | wait_queue_t *curr, *next; |
| 3081 | |
| 3082 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { |
| 3083 | unsigned flags = curr->flags; |
| 3084 | |
| 3085 | if (curr->func(curr, mode, wake_flags, key) && |
| 3086 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) |
| 3087 | break; |
| 3088 | } |
| 3089 | } |
| 3090 | |
| 3091 | /** |
| 3092 | * __wake_up - wake up threads blocked on a waitqueue. |
| 3093 | * @q: the waitqueue |
| 3094 | * @mode: which threads |
| 3095 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
| 3096 | * @key: is directly passed to the wakeup function |
| 3097 | * |
| 3098 | * It may be assumed that this function implies a write memory barrier before |
| 3099 | * changing the task state if and only if any tasks are woken up. |
| 3100 | */ |
| 3101 | void __wake_up(wait_queue_head_t *q, unsigned int mode, |
| 3102 | int nr_exclusive, void *key) |
| 3103 | { |
| 3104 | unsigned long flags; |
| 3105 | |
| 3106 | spin_lock_irqsave(&q->lock, flags); |
| 3107 | __wake_up_common(q, mode, nr_exclusive, 0, key); |
| 3108 | spin_unlock_irqrestore(&q->lock, flags); |
| 3109 | } |
| 3110 | EXPORT_SYMBOL(__wake_up); |
| 3111 | |
| 3112 | /* |
| 3113 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. |
| 3114 | */ |
| 3115 | void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) |
| 3116 | { |
| 3117 | __wake_up_common(q, mode, nr, 0, NULL); |
| 3118 | } |
| 3119 | EXPORT_SYMBOL_GPL(__wake_up_locked); |
| 3120 | |
| 3121 | void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) |
| 3122 | { |
| 3123 | __wake_up_common(q, mode, 1, 0, key); |
| 3124 | } |
| 3125 | EXPORT_SYMBOL_GPL(__wake_up_locked_key); |
| 3126 | |
| 3127 | /** |
| 3128 | * __wake_up_sync_key - wake up threads blocked on a waitqueue. |
| 3129 | * @q: the waitqueue |
| 3130 | * @mode: which threads |
| 3131 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
| 3132 | * @key: opaque value to be passed to wakeup targets |
| 3133 | * |
| 3134 | * The sync wakeup differs that the waker knows that it will schedule |
| 3135 | * away soon, so while the target thread will be woken up, it will not |
| 3136 | * be migrated to another CPU - ie. the two threads are 'synchronized' |
| 3137 | * with each other. This can prevent needless bouncing between CPUs. |
| 3138 | * |
| 3139 | * On UP it can prevent extra preemption. |
| 3140 | * |
| 3141 | * It may be assumed that this function implies a write memory barrier before |
| 3142 | * changing the task state if and only if any tasks are woken up. |
| 3143 | */ |
| 3144 | void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, |
| 3145 | int nr_exclusive, void *key) |
| 3146 | { |
| 3147 | unsigned long flags; |
| 3148 | int wake_flags = WF_SYNC; |
| 3149 | |
| 3150 | if (unlikely(!q)) |
| 3151 | return; |
| 3152 | |
| 3153 | if (unlikely(!nr_exclusive)) |
| 3154 | wake_flags = 0; |
| 3155 | |
| 3156 | spin_lock_irqsave(&q->lock, flags); |
| 3157 | __wake_up_common(q, mode, nr_exclusive, wake_flags, key); |
| 3158 | spin_unlock_irqrestore(&q->lock, flags); |
| 3159 | } |
| 3160 | EXPORT_SYMBOL_GPL(__wake_up_sync_key); |
| 3161 | |
| 3162 | /* |
| 3163 | * __wake_up_sync - see __wake_up_sync_key() |
| 3164 | */ |
| 3165 | void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
| 3166 | { |
| 3167 | __wake_up_sync_key(q, mode, nr_exclusive, NULL); |
| 3168 | } |
| 3169 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ |
| 3170 | |
| 3171 | /** |
| 3172 | * complete: - signals a single thread waiting on this completion |
| 3173 | * @x: holds the state of this particular completion |
| 3174 | * |
| 3175 | * This will wake up a single thread waiting on this completion. Threads will be |
| 3176 | * awakened in the same order in which they were queued. |
| 3177 | * |
| 3178 | * See also complete_all(), wait_for_completion() and related routines. |
| 3179 | * |
| 3180 | * It may be assumed that this function implies a write memory barrier before |
| 3181 | * changing the task state if and only if any tasks are woken up. |
| 3182 | */ |
| 3183 | void complete(struct completion *x) |
| 3184 | { |
| 3185 | unsigned long flags; |
| 3186 | |
| 3187 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3188 | x->done++; |
| 3189 | __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); |
| 3190 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3191 | } |
| 3192 | EXPORT_SYMBOL(complete); |
| 3193 | |
| 3194 | /** |
| 3195 | * complete_all: - signals all threads waiting on this completion |
| 3196 | * @x: holds the state of this particular completion |
| 3197 | * |
| 3198 | * This will wake up all threads waiting on this particular completion event. |
| 3199 | * |
| 3200 | * It may be assumed that this function implies a write memory barrier before |
| 3201 | * changing the task state if and only if any tasks are woken up. |
| 3202 | */ |
| 3203 | void complete_all(struct completion *x) |
| 3204 | { |
| 3205 | unsigned long flags; |
| 3206 | |
| 3207 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3208 | x->done += UINT_MAX/2; |
| 3209 | __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); |
| 3210 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3211 | } |
| 3212 | EXPORT_SYMBOL(complete_all); |
| 3213 | |
| 3214 | static inline long __sched |
| 3215 | do_wait_for_common(struct completion *x, long timeout, int state) |
| 3216 | { |
| 3217 | if (!x->done) { |
| 3218 | DECLARE_WAITQUEUE(wait, current); |
| 3219 | |
| 3220 | __add_wait_queue_tail_exclusive(&x->wait, &wait); |
| 3221 | do { |
| 3222 | if (signal_pending_state(state, current)) { |
| 3223 | timeout = -ERESTARTSYS; |
| 3224 | break; |
| 3225 | } |
| 3226 | __set_current_state(state); |
| 3227 | spin_unlock_irq(&x->wait.lock); |
| 3228 | timeout = schedule_timeout(timeout); |
| 3229 | spin_lock_irq(&x->wait.lock); |
| 3230 | } while (!x->done && timeout); |
| 3231 | __remove_wait_queue(&x->wait, &wait); |
| 3232 | if (!x->done) |
| 3233 | return timeout; |
| 3234 | } |
| 3235 | x->done--; |
| 3236 | return timeout ?: 1; |
| 3237 | } |
| 3238 | |
| 3239 | static long __sched |
| 3240 | wait_for_common(struct completion *x, long timeout, int state) |
| 3241 | { |
| 3242 | might_sleep(); |
| 3243 | |
| 3244 | spin_lock_irq(&x->wait.lock); |
| 3245 | timeout = do_wait_for_common(x, timeout, state); |
| 3246 | spin_unlock_irq(&x->wait.lock); |
| 3247 | return timeout; |
| 3248 | } |
| 3249 | |
| 3250 | /** |
| 3251 | * wait_for_completion: - waits for completion of a task |
| 3252 | * @x: holds the state of this particular completion |
| 3253 | * |
| 3254 | * This waits to be signaled for completion of a specific task. It is NOT |
| 3255 | * interruptible and there is no timeout. |
| 3256 | * |
| 3257 | * See also similar routines (i.e. wait_for_completion_timeout()) with timeout |
| 3258 | * and interrupt capability. Also see complete(). |
| 3259 | */ |
| 3260 | void __sched wait_for_completion(struct completion *x) |
| 3261 | { |
| 3262 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); |
| 3263 | } |
| 3264 | EXPORT_SYMBOL(wait_for_completion); |
| 3265 | |
| 3266 | /** |
| 3267 | * wait_for_completion_timeout: - waits for completion of a task (w/timeout) |
| 3268 | * @x: holds the state of this particular completion |
| 3269 | * @timeout: timeout value in jiffies |
| 3270 | * |
| 3271 | * This waits for either a completion of a specific task to be signaled or for a |
| 3272 | * specified timeout to expire. The timeout is in jiffies. It is not |
| 3273 | * interruptible. |
| 3274 | * |
| 3275 | * The return value is 0 if timed out, and positive (at least 1, or number of |
| 3276 | * jiffies left till timeout) if completed. |
| 3277 | */ |
| 3278 | unsigned long __sched |
| 3279 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
| 3280 | { |
| 3281 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); |
| 3282 | } |
| 3283 | EXPORT_SYMBOL(wait_for_completion_timeout); |
| 3284 | |
| 3285 | /** |
| 3286 | * wait_for_completion_interruptible: - waits for completion of a task (w/intr) |
| 3287 | * @x: holds the state of this particular completion |
| 3288 | * |
| 3289 | * This waits for completion of a specific task to be signaled. It is |
| 3290 | * interruptible. |
| 3291 | * |
| 3292 | * The return value is -ERESTARTSYS if interrupted, 0 if completed. |
| 3293 | */ |
| 3294 | int __sched wait_for_completion_interruptible(struct completion *x) |
| 3295 | { |
| 3296 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); |
| 3297 | if (t == -ERESTARTSYS) |
| 3298 | return t; |
| 3299 | return 0; |
| 3300 | } |
| 3301 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
| 3302 | |
| 3303 | /** |
| 3304 | * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) |
| 3305 | * @x: holds the state of this particular completion |
| 3306 | * @timeout: timeout value in jiffies |
| 3307 | * |
| 3308 | * This waits for either a completion of a specific task to be signaled or for a |
| 3309 | * specified timeout to expire. It is interruptible. The timeout is in jiffies. |
| 3310 | * |
| 3311 | * The return value is -ERESTARTSYS if interrupted, 0 if timed out, |
| 3312 | * positive (at least 1, or number of jiffies left till timeout) if completed. |
| 3313 | */ |
| 3314 | long __sched |
| 3315 | wait_for_completion_interruptible_timeout(struct completion *x, |
| 3316 | unsigned long timeout) |
| 3317 | { |
| 3318 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); |
| 3319 | } |
| 3320 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
| 3321 | |
| 3322 | /** |
| 3323 | * wait_for_completion_killable: - waits for completion of a task (killable) |
| 3324 | * @x: holds the state of this particular completion |
| 3325 | * |
| 3326 | * This waits to be signaled for completion of a specific task. It can be |
| 3327 | * interrupted by a kill signal. |
| 3328 | * |
| 3329 | * The return value is -ERESTARTSYS if interrupted, 0 if completed. |
| 3330 | */ |
| 3331 | int __sched wait_for_completion_killable(struct completion *x) |
| 3332 | { |
| 3333 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); |
| 3334 | if (t == -ERESTARTSYS) |
| 3335 | return t; |
| 3336 | return 0; |
| 3337 | } |
| 3338 | EXPORT_SYMBOL(wait_for_completion_killable); |
| 3339 | |
| 3340 | /** |
| 3341 | * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) |
| 3342 | * @x: holds the state of this particular completion |
| 3343 | * @timeout: timeout value in jiffies |
| 3344 | * |
| 3345 | * This waits for either a completion of a specific task to be |
| 3346 | * signaled or for a specified timeout to expire. It can be |
| 3347 | * interrupted by a kill signal. The timeout is in jiffies. |
| 3348 | * |
| 3349 | * The return value is -ERESTARTSYS if interrupted, 0 if timed out, |
| 3350 | * positive (at least 1, or number of jiffies left till timeout) if completed. |
| 3351 | */ |
| 3352 | long __sched |
| 3353 | wait_for_completion_killable_timeout(struct completion *x, |
| 3354 | unsigned long timeout) |
| 3355 | { |
| 3356 | return wait_for_common(x, timeout, TASK_KILLABLE); |
| 3357 | } |
| 3358 | EXPORT_SYMBOL(wait_for_completion_killable_timeout); |
| 3359 | |
| 3360 | /** |
| 3361 | * try_wait_for_completion - try to decrement a completion without blocking |
| 3362 | * @x: completion structure |
| 3363 | * |
| 3364 | * Returns: 0 if a decrement cannot be done without blocking |
| 3365 | * 1 if a decrement succeeded. |
| 3366 | * |
| 3367 | * If a completion is being used as a counting completion, |
| 3368 | * attempt to decrement the counter without blocking. This |
| 3369 | * enables us to avoid waiting if the resource the completion |
| 3370 | * is protecting is not available. |
| 3371 | */ |
| 3372 | bool try_wait_for_completion(struct completion *x) |
| 3373 | { |
| 3374 | unsigned long flags; |
| 3375 | int ret = 1; |
| 3376 | |
| 3377 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3378 | if (!x->done) |
| 3379 | ret = 0; |
| 3380 | else |
| 3381 | x->done--; |
| 3382 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3383 | return ret; |
| 3384 | } |
| 3385 | EXPORT_SYMBOL(try_wait_for_completion); |
| 3386 | |
| 3387 | /** |
| 3388 | * completion_done - Test to see if a completion has any waiters |
| 3389 | * @x: completion structure |
| 3390 | * |
| 3391 | * Returns: 0 if there are waiters (wait_for_completion() in progress) |
| 3392 | * 1 if there are no waiters. |
| 3393 | * |
| 3394 | */ |
| 3395 | bool completion_done(struct completion *x) |
| 3396 | { |
| 3397 | unsigned long flags; |
| 3398 | int ret = 1; |
| 3399 | |
| 3400 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3401 | if (!x->done) |
| 3402 | ret = 0; |
| 3403 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3404 | return ret; |
| 3405 | } |
| 3406 | EXPORT_SYMBOL(completion_done); |
| 3407 | |
| 3408 | static long __sched |
| 3409 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) |
| 3410 | { |
| 3411 | unsigned long flags; |
| 3412 | wait_queue_t wait; |
| 3413 | |
| 3414 | init_waitqueue_entry(&wait, current); |
| 3415 | |
| 3416 | __set_current_state(state); |
| 3417 | |
| 3418 | spin_lock_irqsave(&q->lock, flags); |
| 3419 | __add_wait_queue(q, &wait); |
| 3420 | spin_unlock(&q->lock); |
| 3421 | timeout = schedule_timeout(timeout); |
| 3422 | spin_lock_irq(&q->lock); |
| 3423 | __remove_wait_queue(q, &wait); |
| 3424 | spin_unlock_irqrestore(&q->lock, flags); |
| 3425 | |
| 3426 | return timeout; |
| 3427 | } |
| 3428 | |
| 3429 | void __sched interruptible_sleep_on(wait_queue_head_t *q) |
| 3430 | { |
| 3431 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
| 3432 | } |
| 3433 | EXPORT_SYMBOL(interruptible_sleep_on); |
| 3434 | |
| 3435 | long __sched |
| 3436 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
| 3437 | { |
| 3438 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); |
| 3439 | } |
| 3440 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
| 3441 | |
| 3442 | void __sched sleep_on(wait_queue_head_t *q) |
| 3443 | { |
| 3444 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
| 3445 | } |
| 3446 | EXPORT_SYMBOL(sleep_on); |
| 3447 | |
| 3448 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
| 3449 | { |
| 3450 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); |
| 3451 | } |
| 3452 | EXPORT_SYMBOL(sleep_on_timeout); |
| 3453 | |
| 3454 | #ifdef CONFIG_RT_MUTEXES |
| 3455 | |
| 3456 | /* |
| 3457 | * rt_mutex_setprio - set the current priority of a task |
| 3458 | * @p: task |
| 3459 | * @prio: prio value (kernel-internal form) |
| 3460 | * |
| 3461 | * This function changes the 'effective' priority of a task. It does |
| 3462 | * not touch ->normal_prio like __setscheduler(). |
| 3463 | * |
| 3464 | * Used by the rt_mutex code to implement priority inheritance logic. |
| 3465 | */ |
| 3466 | void rt_mutex_setprio(struct task_struct *p, int prio) |
| 3467 | { |
| 3468 | int oldprio, on_rq, running; |
| 3469 | struct rq *rq; |
| 3470 | const struct sched_class *prev_class; |
| 3471 | |
| 3472 | BUG_ON(prio < 0 || prio > MAX_PRIO); |
| 3473 | |
| 3474 | rq = __task_rq_lock(p); |
| 3475 | |
| 3476 | /* |
| 3477 | * Idle task boosting is a nono in general. There is one |
| 3478 | * exception, when PREEMPT_RT and NOHZ is active: |
| 3479 | * |
| 3480 | * The idle task calls get_next_timer_interrupt() and holds |
| 3481 | * the timer wheel base->lock on the CPU and another CPU wants |
| 3482 | * to access the timer (probably to cancel it). We can safely |
| 3483 | * ignore the boosting request, as the idle CPU runs this code |
| 3484 | * with interrupts disabled and will complete the lock |
| 3485 | * protected section without being interrupted. So there is no |
| 3486 | * real need to boost. |
| 3487 | */ |
| 3488 | if (unlikely(p == rq->idle)) { |
| 3489 | WARN_ON(p != rq->curr); |
| 3490 | WARN_ON(p->pi_blocked_on); |
| 3491 | goto out_unlock; |
| 3492 | } |
| 3493 | |
| 3494 | trace_sched_pi_setprio(p, prio); |
| 3495 | oldprio = p->prio; |
| 3496 | prev_class = p->sched_class; |
| 3497 | on_rq = p->on_rq; |
| 3498 | running = task_current(rq, p); |
| 3499 | if (on_rq) |
| 3500 | dequeue_task(rq, p, 0); |
| 3501 | if (running) |
| 3502 | p->sched_class->put_prev_task(rq, p); |
| 3503 | |
| 3504 | if (rt_prio(prio)) |
| 3505 | p->sched_class = &rt_sched_class; |
| 3506 | else |
| 3507 | p->sched_class = &fair_sched_class; |
| 3508 | |
| 3509 | p->prio = prio; |
| 3510 | |
| 3511 | if (running) |
| 3512 | p->sched_class->set_curr_task(rq); |
| 3513 | if (on_rq) |
| 3514 | enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); |
| 3515 | |
| 3516 | check_class_changed(rq, p, prev_class, oldprio); |
| 3517 | out_unlock: |
| 3518 | __task_rq_unlock(rq); |
| 3519 | } |
| 3520 | #endif |
| 3521 | void set_user_nice(struct task_struct *p, long nice) |
| 3522 | { |
| 3523 | int old_prio, delta, on_rq; |
| 3524 | unsigned long flags; |
| 3525 | struct rq *rq; |
| 3526 | |
| 3527 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) |
| 3528 | return; |
| 3529 | /* |
| 3530 | * We have to be careful, if called from sys_setpriority(), |
| 3531 | * the task might be in the middle of scheduling on another CPU. |
| 3532 | */ |
| 3533 | rq = task_rq_lock(p, &flags); |
| 3534 | /* |
| 3535 | * The RT priorities are set via sched_setscheduler(), but we still |
| 3536 | * allow the 'normal' nice value to be set - but as expected |
| 3537 | * it wont have any effect on scheduling until the task is |
| 3538 | * SCHED_FIFO/SCHED_RR: |
| 3539 | */ |
| 3540 | if (task_has_rt_policy(p)) { |
| 3541 | p->static_prio = NICE_TO_PRIO(nice); |
| 3542 | goto out_unlock; |
| 3543 | } |
| 3544 | on_rq = p->on_rq; |
| 3545 | if (on_rq) |
| 3546 | dequeue_task(rq, p, 0); |
| 3547 | |
| 3548 | p->static_prio = NICE_TO_PRIO(nice); |
| 3549 | set_load_weight(p); |
| 3550 | old_prio = p->prio; |
| 3551 | p->prio = effective_prio(p); |
| 3552 | delta = p->prio - old_prio; |
| 3553 | |
| 3554 | if (on_rq) { |
| 3555 | enqueue_task(rq, p, 0); |
| 3556 | /* |
| 3557 | * If the task increased its priority or is running and |
| 3558 | * lowered its priority, then reschedule its CPU: |
| 3559 | */ |
| 3560 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
| 3561 | resched_task(rq->curr); |
| 3562 | } |
| 3563 | out_unlock: |
| 3564 | task_rq_unlock(rq, p, &flags); |
| 3565 | } |
| 3566 | EXPORT_SYMBOL(set_user_nice); |
| 3567 | |
| 3568 | /* |
| 3569 | * can_nice - check if a task can reduce its nice value |
| 3570 | * @p: task |
| 3571 | * @nice: nice value |
| 3572 | */ |
| 3573 | int can_nice(const struct task_struct *p, const int nice) |
| 3574 | { |
| 3575 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
| 3576 | int nice_rlim = 20 - nice; |
| 3577 | |
| 3578 | return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || |
| 3579 | capable(CAP_SYS_NICE)); |
| 3580 | } |
| 3581 | |
| 3582 | #ifdef __ARCH_WANT_SYS_NICE |
| 3583 | |
| 3584 | /* |
| 3585 | * sys_nice - change the priority of the current process. |
| 3586 | * @increment: priority increment |
| 3587 | * |
| 3588 | * sys_setpriority is a more generic, but much slower function that |
| 3589 | * does similar things. |
| 3590 | */ |
| 3591 | SYSCALL_DEFINE1(nice, int, increment) |
| 3592 | { |
| 3593 | long nice, retval; |
| 3594 | |
| 3595 | /* |
| 3596 | * Setpriority might change our priority at the same moment. |
| 3597 | * We don't have to worry. Conceptually one call occurs first |
| 3598 | * and we have a single winner. |
| 3599 | */ |
| 3600 | if (increment < -40) |
| 3601 | increment = -40; |
| 3602 | if (increment > 40) |
| 3603 | increment = 40; |
| 3604 | |
| 3605 | nice = TASK_NICE(current) + increment; |
| 3606 | if (nice < -20) |
| 3607 | nice = -20; |
| 3608 | if (nice > 19) |
| 3609 | nice = 19; |
| 3610 | |
| 3611 | if (increment < 0 && !can_nice(current, nice)) |
| 3612 | return -EPERM; |
| 3613 | |
| 3614 | retval = security_task_setnice(current, nice); |
| 3615 | if (retval) |
| 3616 | return retval; |
| 3617 | |
| 3618 | set_user_nice(current, nice); |
| 3619 | return 0; |
| 3620 | } |
| 3621 | |
| 3622 | #endif |
| 3623 | |
| 3624 | /** |
| 3625 | * task_prio - return the priority value of a given task. |
| 3626 | * @p: the task in question. |
| 3627 | * |
| 3628 | * This is the priority value as seen by users in /proc. |
| 3629 | * RT tasks are offset by -200. Normal tasks are centered |
| 3630 | * around 0, value goes from -16 to +15. |
| 3631 | */ |
| 3632 | int task_prio(const struct task_struct *p) |
| 3633 | { |
| 3634 | return p->prio - MAX_RT_PRIO; |
| 3635 | } |
| 3636 | |
| 3637 | /** |
| 3638 | * task_nice - return the nice value of a given task. |
| 3639 | * @p: the task in question. |
| 3640 | */ |
| 3641 | int task_nice(const struct task_struct *p) |
| 3642 | { |
| 3643 | return TASK_NICE(p); |
| 3644 | } |
| 3645 | EXPORT_SYMBOL(task_nice); |
| 3646 | |
| 3647 | /** |
| 3648 | * idle_cpu - is a given cpu idle currently? |
| 3649 | * @cpu: the processor in question. |
| 3650 | */ |
| 3651 | int idle_cpu(int cpu) |
| 3652 | { |
| 3653 | struct rq *rq = cpu_rq(cpu); |
| 3654 | |
| 3655 | if (rq->curr != rq->idle) |
| 3656 | return 0; |
| 3657 | |
| 3658 | if (rq->nr_running) |
| 3659 | return 0; |
| 3660 | |
| 3661 | #ifdef CONFIG_SMP |
| 3662 | if (!llist_empty(&rq->wake_list)) |
| 3663 | return 0; |
| 3664 | #endif |
| 3665 | |
| 3666 | return 1; |
| 3667 | } |
| 3668 | |
| 3669 | /** |
| 3670 | * idle_task - return the idle task for a given cpu. |
| 3671 | * @cpu: the processor in question. |
| 3672 | */ |
| 3673 | struct task_struct *idle_task(int cpu) |
| 3674 | { |
| 3675 | return cpu_rq(cpu)->idle; |
| 3676 | } |
| 3677 | |
| 3678 | /** |
| 3679 | * find_process_by_pid - find a process with a matching PID value. |
| 3680 | * @pid: the pid in question. |
| 3681 | */ |
| 3682 | static struct task_struct *find_process_by_pid(pid_t pid) |
| 3683 | { |
| 3684 | return pid ? find_task_by_vpid(pid) : current; |
| 3685 | } |
| 3686 | |
| 3687 | /* Actually do priority change: must hold rq lock. */ |
| 3688 | static void |
| 3689 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) |
| 3690 | { |
| 3691 | p->policy = policy; |
| 3692 | p->rt_priority = prio; |
| 3693 | p->normal_prio = normal_prio(p); |
| 3694 | /* we are holding p->pi_lock already */ |
| 3695 | p->prio = rt_mutex_getprio(p); |
| 3696 | if (rt_prio(p->prio)) |
| 3697 | p->sched_class = &rt_sched_class; |
| 3698 | else |
| 3699 | p->sched_class = &fair_sched_class; |
| 3700 | set_load_weight(p); |
| 3701 | } |
| 3702 | |
| 3703 | /* |
| 3704 | * check the target process has a UID that matches the current process's |
| 3705 | */ |
| 3706 | static bool check_same_owner(struct task_struct *p) |
| 3707 | { |
| 3708 | const struct cred *cred = current_cred(), *pcred; |
| 3709 | bool match; |
| 3710 | |
| 3711 | rcu_read_lock(); |
| 3712 | pcred = __task_cred(p); |
| 3713 | match = (uid_eq(cred->euid, pcred->euid) || |
| 3714 | uid_eq(cred->euid, pcred->uid)); |
| 3715 | rcu_read_unlock(); |
| 3716 | return match; |
| 3717 | } |
| 3718 | |
| 3719 | static int __sched_setscheduler(struct task_struct *p, int policy, |
| 3720 | const struct sched_param *param, bool user) |
| 3721 | { |
| 3722 | int retval, oldprio, oldpolicy = -1, on_rq, running; |
| 3723 | unsigned long flags; |
| 3724 | const struct sched_class *prev_class; |
| 3725 | struct rq *rq; |
| 3726 | int reset_on_fork; |
| 3727 | |
| 3728 | /* may grab non-irq protected spin_locks */ |
| 3729 | BUG_ON(in_interrupt()); |
| 3730 | recheck: |
| 3731 | /* double check policy once rq lock held */ |
| 3732 | if (policy < 0) { |
| 3733 | reset_on_fork = p->sched_reset_on_fork; |
| 3734 | policy = oldpolicy = p->policy; |
| 3735 | } else { |
| 3736 | reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); |
| 3737 | policy &= ~SCHED_RESET_ON_FORK; |
| 3738 | |
| 3739 | if (policy != SCHED_FIFO && policy != SCHED_RR && |
| 3740 | policy != SCHED_NORMAL && policy != SCHED_BATCH && |
| 3741 | policy != SCHED_IDLE) |
| 3742 | return -EINVAL; |
| 3743 | } |
| 3744 | |
| 3745 | /* |
| 3746 | * Valid priorities for SCHED_FIFO and SCHED_RR are |
| 3747 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, |
| 3748 | * SCHED_BATCH and SCHED_IDLE is 0. |
| 3749 | */ |
| 3750 | if (param->sched_priority < 0 || |
| 3751 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
| 3752 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
| 3753 | return -EINVAL; |
| 3754 | if (rt_policy(policy) != (param->sched_priority != 0)) |
| 3755 | return -EINVAL; |
| 3756 | |
| 3757 | /* |
| 3758 | * Allow unprivileged RT tasks to decrease priority: |
| 3759 | */ |
| 3760 | if (user && !capable(CAP_SYS_NICE)) { |
| 3761 | if (rt_policy(policy)) { |
| 3762 | unsigned long rlim_rtprio = |
| 3763 | task_rlimit(p, RLIMIT_RTPRIO); |
| 3764 | |
| 3765 | /* can't set/change the rt policy */ |
| 3766 | if (policy != p->policy && !rlim_rtprio) |
| 3767 | return -EPERM; |
| 3768 | |
| 3769 | /* can't increase priority */ |
| 3770 | if (param->sched_priority > p->rt_priority && |
| 3771 | param->sched_priority > rlim_rtprio) |
| 3772 | return -EPERM; |
| 3773 | } |
| 3774 | |
| 3775 | /* |
| 3776 | * Treat SCHED_IDLE as nice 20. Only allow a switch to |
| 3777 | * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. |
| 3778 | */ |
| 3779 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { |
| 3780 | if (!can_nice(p, TASK_NICE(p))) |
| 3781 | return -EPERM; |
| 3782 | } |
| 3783 | |
| 3784 | /* can't change other user's priorities */ |
| 3785 | if (!check_same_owner(p)) |
| 3786 | return -EPERM; |
| 3787 | |
| 3788 | /* Normal users shall not reset the sched_reset_on_fork flag */ |
| 3789 | if (p->sched_reset_on_fork && !reset_on_fork) |
| 3790 | return -EPERM; |
| 3791 | } |
| 3792 | |
| 3793 | if (user) { |
| 3794 | retval = security_task_setscheduler(p); |
| 3795 | if (retval) |
| 3796 | return retval; |
| 3797 | } |
| 3798 | |
| 3799 | /* |
| 3800 | * make sure no PI-waiters arrive (or leave) while we are |
| 3801 | * changing the priority of the task: |
| 3802 | * |
| 3803 | * To be able to change p->policy safely, the appropriate |
| 3804 | * runqueue lock must be held. |
| 3805 | */ |
| 3806 | rq = task_rq_lock(p, &flags); |
| 3807 | |
| 3808 | /* |
| 3809 | * Changing the policy of the stop threads its a very bad idea |
| 3810 | */ |
| 3811 | if (p == rq->stop) { |
| 3812 | task_rq_unlock(rq, p, &flags); |
| 3813 | return -EINVAL; |
| 3814 | } |
| 3815 | |
| 3816 | /* |
| 3817 | * If not changing anything there's no need to proceed further: |
| 3818 | */ |
| 3819 | if (unlikely(policy == p->policy && (!rt_policy(policy) || |
| 3820 | param->sched_priority == p->rt_priority))) { |
| 3821 | task_rq_unlock(rq, p, &flags); |
| 3822 | return 0; |
| 3823 | } |
| 3824 | |
| 3825 | #ifdef CONFIG_RT_GROUP_SCHED |
| 3826 | if (user) { |
| 3827 | /* |
| 3828 | * Do not allow realtime tasks into groups that have no runtime |
| 3829 | * assigned. |
| 3830 | */ |
| 3831 | if (rt_bandwidth_enabled() && rt_policy(policy) && |
| 3832 | task_group(p)->rt_bandwidth.rt_runtime == 0 && |
| 3833 | !task_group_is_autogroup(task_group(p))) { |
| 3834 | task_rq_unlock(rq, p, &flags); |
| 3835 | return -EPERM; |
| 3836 | } |
| 3837 | } |
| 3838 | #endif |
| 3839 | |
| 3840 | /* recheck policy now with rq lock held */ |
| 3841 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { |
| 3842 | policy = oldpolicy = -1; |
| 3843 | task_rq_unlock(rq, p, &flags); |
| 3844 | goto recheck; |
| 3845 | } |
| 3846 | on_rq = p->on_rq; |
| 3847 | running = task_current(rq, p); |
| 3848 | if (on_rq) |
| 3849 | dequeue_task(rq, p, 0); |
| 3850 | if (running) |
| 3851 | p->sched_class->put_prev_task(rq, p); |
| 3852 | |
| 3853 | p->sched_reset_on_fork = reset_on_fork; |
| 3854 | |
| 3855 | oldprio = p->prio; |
| 3856 | prev_class = p->sched_class; |
| 3857 | __setscheduler(rq, p, policy, param->sched_priority); |
| 3858 | |
| 3859 | if (running) |
| 3860 | p->sched_class->set_curr_task(rq); |
| 3861 | if (on_rq) |
| 3862 | enqueue_task(rq, p, 0); |
| 3863 | |
| 3864 | check_class_changed(rq, p, prev_class, oldprio); |
| 3865 | task_rq_unlock(rq, p, &flags); |
| 3866 | |
| 3867 | rt_mutex_adjust_pi(p); |
| 3868 | |
| 3869 | return 0; |
| 3870 | } |
| 3871 | |
| 3872 | /** |
| 3873 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. |
| 3874 | * @p: the task in question. |
| 3875 | * @policy: new policy. |
| 3876 | * @param: structure containing the new RT priority. |
| 3877 | * |
| 3878 | * NOTE that the task may be already dead. |
| 3879 | */ |
| 3880 | int sched_setscheduler(struct task_struct *p, int policy, |
| 3881 | const struct sched_param *param) |
| 3882 | { |
| 3883 | return __sched_setscheduler(p, policy, param, true); |
| 3884 | } |
| 3885 | EXPORT_SYMBOL_GPL(sched_setscheduler); |
| 3886 | |
| 3887 | /** |
| 3888 | * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. |
| 3889 | * @p: the task in question. |
| 3890 | * @policy: new policy. |
| 3891 | * @param: structure containing the new RT priority. |
| 3892 | * |
| 3893 | * Just like sched_setscheduler, only don't bother checking if the |
| 3894 | * current context has permission. For example, this is needed in |
| 3895 | * stop_machine(): we create temporary high priority worker threads, |
| 3896 | * but our caller might not have that capability. |
| 3897 | */ |
| 3898 | int sched_setscheduler_nocheck(struct task_struct *p, int policy, |
| 3899 | const struct sched_param *param) |
| 3900 | { |
| 3901 | return __sched_setscheduler(p, policy, param, false); |
| 3902 | } |
| 3903 | |
| 3904 | static int |
| 3905 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) |
| 3906 | { |
| 3907 | struct sched_param lparam; |
| 3908 | struct task_struct *p; |
| 3909 | int retval; |
| 3910 | |
| 3911 | if (!param || pid < 0) |
| 3912 | return -EINVAL; |
| 3913 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) |
| 3914 | return -EFAULT; |
| 3915 | |
| 3916 | rcu_read_lock(); |
| 3917 | retval = -ESRCH; |
| 3918 | p = find_process_by_pid(pid); |
| 3919 | if (p != NULL) |
| 3920 | retval = sched_setscheduler(p, policy, &lparam); |
| 3921 | rcu_read_unlock(); |
| 3922 | |
| 3923 | return retval; |
| 3924 | } |
| 3925 | |
| 3926 | /** |
| 3927 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority |
| 3928 | * @pid: the pid in question. |
| 3929 | * @policy: new policy. |
| 3930 | * @param: structure containing the new RT priority. |
| 3931 | */ |
| 3932 | SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, |
| 3933 | struct sched_param __user *, param) |
| 3934 | { |
| 3935 | /* negative values for policy are not valid */ |
| 3936 | if (policy < 0) |
| 3937 | return -EINVAL; |
| 3938 | |
| 3939 | return do_sched_setscheduler(pid, policy, param); |
| 3940 | } |
| 3941 | |
| 3942 | /** |
| 3943 | * sys_sched_setparam - set/change the RT priority of a thread |
| 3944 | * @pid: the pid in question. |
| 3945 | * @param: structure containing the new RT priority. |
| 3946 | */ |
| 3947 | SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) |
| 3948 | { |
| 3949 | return do_sched_setscheduler(pid, -1, param); |
| 3950 | } |
| 3951 | |
| 3952 | /** |
| 3953 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread |
| 3954 | * @pid: the pid in question. |
| 3955 | */ |
| 3956 | SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) |
| 3957 | { |
| 3958 | struct task_struct *p; |
| 3959 | int retval; |
| 3960 | |
| 3961 | if (pid < 0) |
| 3962 | return -EINVAL; |
| 3963 | |
| 3964 | retval = -ESRCH; |
| 3965 | rcu_read_lock(); |
| 3966 | p = find_process_by_pid(pid); |
| 3967 | if (p) { |
| 3968 | retval = security_task_getscheduler(p); |
| 3969 | if (!retval) |
| 3970 | retval = p->policy |
| 3971 | | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); |
| 3972 | } |
| 3973 | rcu_read_unlock(); |
| 3974 | return retval; |
| 3975 | } |
| 3976 | |
| 3977 | /** |
| 3978 | * sys_sched_getparam - get the RT priority of a thread |
| 3979 | * @pid: the pid in question. |
| 3980 | * @param: structure containing the RT priority. |
| 3981 | */ |
| 3982 | SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) |
| 3983 | { |
| 3984 | struct sched_param lp; |
| 3985 | struct task_struct *p; |
| 3986 | int retval; |
| 3987 | |
| 3988 | if (!param || pid < 0) |
| 3989 | return -EINVAL; |
| 3990 | |
| 3991 | rcu_read_lock(); |
| 3992 | p = find_process_by_pid(pid); |
| 3993 | retval = -ESRCH; |
| 3994 | if (!p) |
| 3995 | goto out_unlock; |
| 3996 | |
| 3997 | retval = security_task_getscheduler(p); |
| 3998 | if (retval) |
| 3999 | goto out_unlock; |
| 4000 | |
| 4001 | lp.sched_priority = p->rt_priority; |
| 4002 | rcu_read_unlock(); |
| 4003 | |
| 4004 | /* |
| 4005 | * This one might sleep, we cannot do it with a spinlock held ... |
| 4006 | */ |
| 4007 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; |
| 4008 | |
| 4009 | return retval; |
| 4010 | |
| 4011 | out_unlock: |
| 4012 | rcu_read_unlock(); |
| 4013 | return retval; |
| 4014 | } |
| 4015 | |
| 4016 | long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) |
| 4017 | { |
| 4018 | cpumask_var_t cpus_allowed, new_mask; |
| 4019 | struct task_struct *p; |
| 4020 | int retval; |
| 4021 | |
| 4022 | get_online_cpus(); |
| 4023 | rcu_read_lock(); |
| 4024 | |
| 4025 | p = find_process_by_pid(pid); |
| 4026 | if (!p) { |
| 4027 | rcu_read_unlock(); |
| 4028 | put_online_cpus(); |
| 4029 | return -ESRCH; |
| 4030 | } |
| 4031 | |
| 4032 | /* Prevent p going away */ |
| 4033 | get_task_struct(p); |
| 4034 | rcu_read_unlock(); |
| 4035 | |
| 4036 | if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { |
| 4037 | retval = -ENOMEM; |
| 4038 | goto out_put_task; |
| 4039 | } |
| 4040 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { |
| 4041 | retval = -ENOMEM; |
| 4042 | goto out_free_cpus_allowed; |
| 4043 | } |
| 4044 | retval = -EPERM; |
| 4045 | if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE)) |
| 4046 | goto out_unlock; |
| 4047 | |
| 4048 | retval = security_task_setscheduler(p); |
| 4049 | if (retval) |
| 4050 | goto out_unlock; |
| 4051 | |
| 4052 | cpuset_cpus_allowed(p, cpus_allowed); |
| 4053 | cpumask_and(new_mask, in_mask, cpus_allowed); |
| 4054 | again: |
| 4055 | retval = set_cpus_allowed_ptr(p, new_mask); |
| 4056 | |
| 4057 | if (!retval) { |
| 4058 | cpuset_cpus_allowed(p, cpus_allowed); |
| 4059 | if (!cpumask_subset(new_mask, cpus_allowed)) { |
| 4060 | /* |
| 4061 | * We must have raced with a concurrent cpuset |
| 4062 | * update. Just reset the cpus_allowed to the |
| 4063 | * cpuset's cpus_allowed |
| 4064 | */ |
| 4065 | cpumask_copy(new_mask, cpus_allowed); |
| 4066 | goto again; |
| 4067 | } |
| 4068 | } |
| 4069 | out_unlock: |
| 4070 | free_cpumask_var(new_mask); |
| 4071 | out_free_cpus_allowed: |
| 4072 | free_cpumask_var(cpus_allowed); |
| 4073 | out_put_task: |
| 4074 | put_task_struct(p); |
| 4075 | put_online_cpus(); |
| 4076 | return retval; |
| 4077 | } |
| 4078 | |
| 4079 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, |
| 4080 | struct cpumask *new_mask) |
| 4081 | { |
| 4082 | if (len < cpumask_size()) |
| 4083 | cpumask_clear(new_mask); |
| 4084 | else if (len > cpumask_size()) |
| 4085 | len = cpumask_size(); |
| 4086 | |
| 4087 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; |
| 4088 | } |
| 4089 | |
| 4090 | /** |
| 4091 | * sys_sched_setaffinity - set the cpu affinity of a process |
| 4092 | * @pid: pid of the process |
| 4093 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 4094 | * @user_mask_ptr: user-space pointer to the new cpu mask |
| 4095 | */ |
| 4096 | SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, |
| 4097 | unsigned long __user *, user_mask_ptr) |
| 4098 | { |
| 4099 | cpumask_var_t new_mask; |
| 4100 | int retval; |
| 4101 | |
| 4102 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) |
| 4103 | return -ENOMEM; |
| 4104 | |
| 4105 | retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); |
| 4106 | if (retval == 0) |
| 4107 | retval = sched_setaffinity(pid, new_mask); |
| 4108 | free_cpumask_var(new_mask); |
| 4109 | return retval; |
| 4110 | } |
| 4111 | |
| 4112 | long sched_getaffinity(pid_t pid, struct cpumask *mask) |
| 4113 | { |
| 4114 | struct task_struct *p; |
| 4115 | unsigned long flags; |
| 4116 | int retval; |
| 4117 | |
| 4118 | get_online_cpus(); |
| 4119 | rcu_read_lock(); |
| 4120 | |
| 4121 | retval = -ESRCH; |
| 4122 | p = find_process_by_pid(pid); |
| 4123 | if (!p) |
| 4124 | goto out_unlock; |
| 4125 | |
| 4126 | retval = security_task_getscheduler(p); |
| 4127 | if (retval) |
| 4128 | goto out_unlock; |
| 4129 | |
| 4130 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
| 4131 | cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); |
| 4132 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
| 4133 | |
| 4134 | out_unlock: |
| 4135 | rcu_read_unlock(); |
| 4136 | put_online_cpus(); |
| 4137 | |
| 4138 | return retval; |
| 4139 | } |
| 4140 | |
| 4141 | /** |
| 4142 | * sys_sched_getaffinity - get the cpu affinity of a process |
| 4143 | * @pid: pid of the process |
| 4144 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 4145 | * @user_mask_ptr: user-space pointer to hold the current cpu mask |
| 4146 | */ |
| 4147 | SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, |
| 4148 | unsigned long __user *, user_mask_ptr) |
| 4149 | { |
| 4150 | int ret; |
| 4151 | cpumask_var_t mask; |
| 4152 | |
| 4153 | if ((len * BITS_PER_BYTE) < nr_cpu_ids) |
| 4154 | return -EINVAL; |
| 4155 | if (len & (sizeof(unsigned long)-1)) |
| 4156 | return -EINVAL; |
| 4157 | |
| 4158 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) |
| 4159 | return -ENOMEM; |
| 4160 | |
| 4161 | ret = sched_getaffinity(pid, mask); |
| 4162 | if (ret == 0) { |
| 4163 | size_t retlen = min_t(size_t, len, cpumask_size()); |
| 4164 | |
| 4165 | if (copy_to_user(user_mask_ptr, mask, retlen)) |
| 4166 | ret = -EFAULT; |
| 4167 | else |
| 4168 | ret = retlen; |
| 4169 | } |
| 4170 | free_cpumask_var(mask); |
| 4171 | |
| 4172 | return ret; |
| 4173 | } |
| 4174 | |
| 4175 | /** |
| 4176 | * sys_sched_yield - yield the current processor to other threads. |
| 4177 | * |
| 4178 | * This function yields the current CPU to other tasks. If there are no |
| 4179 | * other threads running on this CPU then this function will return. |
| 4180 | */ |
| 4181 | SYSCALL_DEFINE0(sched_yield) |
| 4182 | { |
| 4183 | struct rq *rq = this_rq_lock(); |
| 4184 | |
| 4185 | schedstat_inc(rq, yld_count); |
| 4186 | current->sched_class->yield_task(rq); |
| 4187 | |
| 4188 | /* |
| 4189 | * Since we are going to call schedule() anyway, there's |
| 4190 | * no need to preempt or enable interrupts: |
| 4191 | */ |
| 4192 | __release(rq->lock); |
| 4193 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
| 4194 | do_raw_spin_unlock(&rq->lock); |
| 4195 | sched_preempt_enable_no_resched(); |
| 4196 | |
| 4197 | schedule(); |
| 4198 | |
| 4199 | return 0; |
| 4200 | } |
| 4201 | |
| 4202 | static inline int should_resched(void) |
| 4203 | { |
| 4204 | return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); |
| 4205 | } |
| 4206 | |
| 4207 | static void __cond_resched(void) |
| 4208 | { |
| 4209 | add_preempt_count(PREEMPT_ACTIVE); |
| 4210 | __schedule(); |
| 4211 | sub_preempt_count(PREEMPT_ACTIVE); |
| 4212 | } |
| 4213 | |
| 4214 | int __sched _cond_resched(void) |
| 4215 | { |
| 4216 | if (should_resched()) { |
| 4217 | __cond_resched(); |
| 4218 | return 1; |
| 4219 | } |
| 4220 | return 0; |
| 4221 | } |
| 4222 | EXPORT_SYMBOL(_cond_resched); |
| 4223 | |
| 4224 | /* |
| 4225 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, |
| 4226 | * call schedule, and on return reacquire the lock. |
| 4227 | * |
| 4228 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
| 4229 | * operations here to prevent schedule() from being called twice (once via |
| 4230 | * spin_unlock(), once by hand). |
| 4231 | */ |
| 4232 | int __cond_resched_lock(spinlock_t *lock) |
| 4233 | { |
| 4234 | int resched = should_resched(); |
| 4235 | int ret = 0; |
| 4236 | |
| 4237 | lockdep_assert_held(lock); |
| 4238 | |
| 4239 | if (spin_needbreak(lock) || resched) { |
| 4240 | spin_unlock(lock); |
| 4241 | if (resched) |
| 4242 | __cond_resched(); |
| 4243 | else |
| 4244 | cpu_relax(); |
| 4245 | ret = 1; |
| 4246 | spin_lock(lock); |
| 4247 | } |
| 4248 | return ret; |
| 4249 | } |
| 4250 | EXPORT_SYMBOL(__cond_resched_lock); |
| 4251 | |
| 4252 | int __sched __cond_resched_softirq(void) |
| 4253 | { |
| 4254 | BUG_ON(!in_softirq()); |
| 4255 | |
| 4256 | if (should_resched()) { |
| 4257 | local_bh_enable(); |
| 4258 | __cond_resched(); |
| 4259 | local_bh_disable(); |
| 4260 | return 1; |
| 4261 | } |
| 4262 | return 0; |
| 4263 | } |
| 4264 | EXPORT_SYMBOL(__cond_resched_softirq); |
| 4265 | |
| 4266 | /** |
| 4267 | * yield - yield the current processor to other threads. |
| 4268 | * |
| 4269 | * Do not ever use this function, there's a 99% chance you're doing it wrong. |
| 4270 | * |
| 4271 | * The scheduler is at all times free to pick the calling task as the most |
| 4272 | * eligible task to run, if removing the yield() call from your code breaks |
| 4273 | * it, its already broken. |
| 4274 | * |
| 4275 | * Typical broken usage is: |
| 4276 | * |
| 4277 | * while (!event) |
| 4278 | * yield(); |
| 4279 | * |
| 4280 | * where one assumes that yield() will let 'the other' process run that will |
| 4281 | * make event true. If the current task is a SCHED_FIFO task that will never |
| 4282 | * happen. Never use yield() as a progress guarantee!! |
| 4283 | * |
| 4284 | * If you want to use yield() to wait for something, use wait_event(). |
| 4285 | * If you want to use yield() to be 'nice' for others, use cond_resched(). |
| 4286 | * If you still want to use yield(), do not! |
| 4287 | */ |
| 4288 | void __sched yield(void) |
| 4289 | { |
| 4290 | set_current_state(TASK_RUNNING); |
| 4291 | sys_sched_yield(); |
| 4292 | } |
| 4293 | EXPORT_SYMBOL(yield); |
| 4294 | |
| 4295 | /** |
| 4296 | * yield_to - yield the current processor to another thread in |
| 4297 | * your thread group, or accelerate that thread toward the |
| 4298 | * processor it's on. |
| 4299 | * @p: target task |
| 4300 | * @preempt: whether task preemption is allowed or not |
| 4301 | * |
| 4302 | * It's the caller's job to ensure that the target task struct |
| 4303 | * can't go away on us before we can do any checks. |
| 4304 | * |
| 4305 | * Returns true if we indeed boosted the target task. |
| 4306 | */ |
| 4307 | bool __sched yield_to(struct task_struct *p, bool preempt) |
| 4308 | { |
| 4309 | struct task_struct *curr = current; |
| 4310 | struct rq *rq, *p_rq; |
| 4311 | unsigned long flags; |
| 4312 | bool yielded = 0; |
| 4313 | |
| 4314 | local_irq_save(flags); |
| 4315 | rq = this_rq(); |
| 4316 | |
| 4317 | again: |
| 4318 | p_rq = task_rq(p); |
| 4319 | double_rq_lock(rq, p_rq); |
| 4320 | while (task_rq(p) != p_rq) { |
| 4321 | double_rq_unlock(rq, p_rq); |
| 4322 | goto again; |
| 4323 | } |
| 4324 | |
| 4325 | if (!curr->sched_class->yield_to_task) |
| 4326 | goto out; |
| 4327 | |
| 4328 | if (curr->sched_class != p->sched_class) |
| 4329 | goto out; |
| 4330 | |
| 4331 | if (task_running(p_rq, p) || p->state) |
| 4332 | goto out; |
| 4333 | |
| 4334 | yielded = curr->sched_class->yield_to_task(rq, p, preempt); |
| 4335 | if (yielded) { |
| 4336 | schedstat_inc(rq, yld_count); |
| 4337 | /* |
| 4338 | * Make p's CPU reschedule; pick_next_entity takes care of |
| 4339 | * fairness. |
| 4340 | */ |
| 4341 | if (preempt && rq != p_rq) |
| 4342 | resched_task(p_rq->curr); |
| 4343 | } |
| 4344 | |
| 4345 | out: |
| 4346 | double_rq_unlock(rq, p_rq); |
| 4347 | local_irq_restore(flags); |
| 4348 | |
| 4349 | if (yielded) |
| 4350 | schedule(); |
| 4351 | |
| 4352 | return yielded; |
| 4353 | } |
| 4354 | EXPORT_SYMBOL_GPL(yield_to); |
| 4355 | |
| 4356 | /* |
| 4357 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
| 4358 | * that process accounting knows that this is a task in IO wait state. |
| 4359 | */ |
| 4360 | void __sched io_schedule(void) |
| 4361 | { |
| 4362 | struct rq *rq = raw_rq(); |
| 4363 | |
| 4364 | delayacct_blkio_start(); |
| 4365 | atomic_inc(&rq->nr_iowait); |
| 4366 | blk_flush_plug(current); |
| 4367 | current->in_iowait = 1; |
| 4368 | schedule(); |
| 4369 | current->in_iowait = 0; |
| 4370 | atomic_dec(&rq->nr_iowait); |
| 4371 | delayacct_blkio_end(); |
| 4372 | } |
| 4373 | EXPORT_SYMBOL(io_schedule); |
| 4374 | |
| 4375 | long __sched io_schedule_timeout(long timeout) |
| 4376 | { |
| 4377 | struct rq *rq = raw_rq(); |
| 4378 | long ret; |
| 4379 | |
| 4380 | delayacct_blkio_start(); |
| 4381 | atomic_inc(&rq->nr_iowait); |
| 4382 | blk_flush_plug(current); |
| 4383 | current->in_iowait = 1; |
| 4384 | ret = schedule_timeout(timeout); |
| 4385 | current->in_iowait = 0; |
| 4386 | atomic_dec(&rq->nr_iowait); |
| 4387 | delayacct_blkio_end(); |
| 4388 | return ret; |
| 4389 | } |
| 4390 | |
| 4391 | /** |
| 4392 | * sys_sched_get_priority_max - return maximum RT priority. |
| 4393 | * @policy: scheduling class. |
| 4394 | * |
| 4395 | * this syscall returns the maximum rt_priority that can be used |
| 4396 | * by a given scheduling class. |
| 4397 | */ |
| 4398 | SYSCALL_DEFINE1(sched_get_priority_max, int, policy) |
| 4399 | { |
| 4400 | int ret = -EINVAL; |
| 4401 | |
| 4402 | switch (policy) { |
| 4403 | case SCHED_FIFO: |
| 4404 | case SCHED_RR: |
| 4405 | ret = MAX_USER_RT_PRIO-1; |
| 4406 | break; |
| 4407 | case SCHED_NORMAL: |
| 4408 | case SCHED_BATCH: |
| 4409 | case SCHED_IDLE: |
| 4410 | ret = 0; |
| 4411 | break; |
| 4412 | } |
| 4413 | return ret; |
| 4414 | } |
| 4415 | |
| 4416 | /** |
| 4417 | * sys_sched_get_priority_min - return minimum RT priority. |
| 4418 | * @policy: scheduling class. |
| 4419 | * |
| 4420 | * this syscall returns the minimum rt_priority that can be used |
| 4421 | * by a given scheduling class. |
| 4422 | */ |
| 4423 | SYSCALL_DEFINE1(sched_get_priority_min, int, policy) |
| 4424 | { |
| 4425 | int ret = -EINVAL; |
| 4426 | |
| 4427 | switch (policy) { |
| 4428 | case SCHED_FIFO: |
| 4429 | case SCHED_RR: |
| 4430 | ret = 1; |
| 4431 | break; |
| 4432 | case SCHED_NORMAL: |
| 4433 | case SCHED_BATCH: |
| 4434 | case SCHED_IDLE: |
| 4435 | ret = 0; |
| 4436 | } |
| 4437 | return ret; |
| 4438 | } |
| 4439 | |
| 4440 | /** |
| 4441 | * sys_sched_rr_get_interval - return the default timeslice of a process. |
| 4442 | * @pid: pid of the process. |
| 4443 | * @interval: userspace pointer to the timeslice value. |
| 4444 | * |
| 4445 | * this syscall writes the default timeslice value of a given process |
| 4446 | * into the user-space timespec buffer. A value of '0' means infinity. |
| 4447 | */ |
| 4448 | SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, |
| 4449 | struct timespec __user *, interval) |
| 4450 | { |
| 4451 | struct task_struct *p; |
| 4452 | unsigned int time_slice; |
| 4453 | unsigned long flags; |
| 4454 | struct rq *rq; |
| 4455 | int retval; |
| 4456 | struct timespec t; |
| 4457 | |
| 4458 | if (pid < 0) |
| 4459 | return -EINVAL; |
| 4460 | |
| 4461 | retval = -ESRCH; |
| 4462 | rcu_read_lock(); |
| 4463 | p = find_process_by_pid(pid); |
| 4464 | if (!p) |
| 4465 | goto out_unlock; |
| 4466 | |
| 4467 | retval = security_task_getscheduler(p); |
| 4468 | if (retval) |
| 4469 | goto out_unlock; |
| 4470 | |
| 4471 | rq = task_rq_lock(p, &flags); |
| 4472 | time_slice = p->sched_class->get_rr_interval(rq, p); |
| 4473 | task_rq_unlock(rq, p, &flags); |
| 4474 | |
| 4475 | rcu_read_unlock(); |
| 4476 | jiffies_to_timespec(time_slice, &t); |
| 4477 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
| 4478 | return retval; |
| 4479 | |
| 4480 | out_unlock: |
| 4481 | rcu_read_unlock(); |
| 4482 | return retval; |
| 4483 | } |
| 4484 | |
| 4485 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; |
| 4486 | |
| 4487 | void sched_show_task(struct task_struct *p) |
| 4488 | { |
| 4489 | unsigned long free = 0; |
| 4490 | unsigned state; |
| 4491 | |
| 4492 | state = p->state ? __ffs(p->state) + 1 : 0; |
| 4493 | printk(KERN_INFO "%-15.15s %c", p->comm, |
| 4494 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); |
| 4495 | #if BITS_PER_LONG == 32 |
| 4496 | if (state == TASK_RUNNING) |
| 4497 | printk(KERN_CONT " running "); |
| 4498 | else |
| 4499 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); |
| 4500 | #else |
| 4501 | if (state == TASK_RUNNING) |
| 4502 | printk(KERN_CONT " running task "); |
| 4503 | else |
| 4504 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); |
| 4505 | #endif |
| 4506 | #ifdef CONFIG_DEBUG_STACK_USAGE |
| 4507 | free = stack_not_used(p); |
| 4508 | #endif |
| 4509 | printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, |
| 4510 | task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)), |
| 4511 | (unsigned long)task_thread_info(p)->flags); |
| 4512 | |
| 4513 | show_stack(p, NULL); |
| 4514 | } |
| 4515 | |
| 4516 | void show_state_filter(unsigned long state_filter) |
| 4517 | { |
| 4518 | struct task_struct *g, *p; |
| 4519 | |
| 4520 | #if BITS_PER_LONG == 32 |
| 4521 | printk(KERN_INFO |
| 4522 | " task PC stack pid father\n"); |
| 4523 | #else |
| 4524 | printk(KERN_INFO |
| 4525 | " task PC stack pid father\n"); |
| 4526 | #endif |
| 4527 | rcu_read_lock(); |
| 4528 | do_each_thread(g, p) { |
| 4529 | /* |
| 4530 | * reset the NMI-timeout, listing all files on a slow |
| 4531 | * console might take a lot of time: |
| 4532 | */ |
| 4533 | touch_nmi_watchdog(); |
| 4534 | if (!state_filter || (p->state & state_filter)) |
| 4535 | sched_show_task(p); |
| 4536 | } while_each_thread(g, p); |
| 4537 | |
| 4538 | touch_all_softlockup_watchdogs(); |
| 4539 | |
| 4540 | #ifdef CONFIG_SCHED_DEBUG |
| 4541 | sysrq_sched_debug_show(); |
| 4542 | #endif |
| 4543 | rcu_read_unlock(); |
| 4544 | /* |
| 4545 | * Only show locks if all tasks are dumped: |
| 4546 | */ |
| 4547 | if (!state_filter) |
| 4548 | debug_show_all_locks(); |
| 4549 | } |
| 4550 | |
| 4551 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) |
| 4552 | { |
| 4553 | idle->sched_class = &idle_sched_class; |
| 4554 | } |
| 4555 | |
| 4556 | /** |
| 4557 | * init_idle - set up an idle thread for a given CPU |
| 4558 | * @idle: task in question |
| 4559 | * @cpu: cpu the idle task belongs to |
| 4560 | * |
| 4561 | * NOTE: this function does not set the idle thread's NEED_RESCHED |
| 4562 | * flag, to make booting more robust. |
| 4563 | */ |
| 4564 | void __cpuinit init_idle(struct task_struct *idle, int cpu) |
| 4565 | { |
| 4566 | struct rq *rq = cpu_rq(cpu); |
| 4567 | unsigned long flags; |
| 4568 | |
| 4569 | raw_spin_lock_irqsave(&rq->lock, flags); |
| 4570 | |
| 4571 | __sched_fork(idle); |
| 4572 | idle->state = TASK_RUNNING; |
| 4573 | idle->se.exec_start = sched_clock(); |
| 4574 | |
| 4575 | do_set_cpus_allowed(idle, cpumask_of(cpu)); |
| 4576 | /* |
| 4577 | * We're having a chicken and egg problem, even though we are |
| 4578 | * holding rq->lock, the cpu isn't yet set to this cpu so the |
| 4579 | * lockdep check in task_group() will fail. |
| 4580 | * |
| 4581 | * Similar case to sched_fork(). / Alternatively we could |
| 4582 | * use task_rq_lock() here and obtain the other rq->lock. |
| 4583 | * |
| 4584 | * Silence PROVE_RCU |
| 4585 | */ |
| 4586 | rcu_read_lock(); |
| 4587 | __set_task_cpu(idle, cpu); |
| 4588 | rcu_read_unlock(); |
| 4589 | |
| 4590 | rq->curr = rq->idle = idle; |
| 4591 | #if defined(CONFIG_SMP) |
| 4592 | idle->on_cpu = 1; |
| 4593 | #endif |
| 4594 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
| 4595 | |
| 4596 | /* Set the preempt count _outside_ the spinlocks! */ |
| 4597 | task_thread_info(idle)->preempt_count = 0; |
| 4598 | |
| 4599 | /* |
| 4600 | * The idle tasks have their own, simple scheduling class: |
| 4601 | */ |
| 4602 | idle->sched_class = &idle_sched_class; |
| 4603 | ftrace_graph_init_idle_task(idle, cpu); |
| 4604 | #if defined(CONFIG_SMP) |
| 4605 | sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); |
| 4606 | #endif |
| 4607 | } |
| 4608 | |
| 4609 | #ifdef CONFIG_SMP |
| 4610 | void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) |
| 4611 | { |
| 4612 | if (p->sched_class && p->sched_class->set_cpus_allowed) |
| 4613 | p->sched_class->set_cpus_allowed(p, new_mask); |
| 4614 | |
| 4615 | cpumask_copy(&p->cpus_allowed, new_mask); |
| 4616 | p->nr_cpus_allowed = cpumask_weight(new_mask); |
| 4617 | } |
| 4618 | |
| 4619 | /* |
| 4620 | * This is how migration works: |
| 4621 | * |
| 4622 | * 1) we invoke migration_cpu_stop() on the target CPU using |
| 4623 | * stop_one_cpu(). |
| 4624 | * 2) stopper starts to run (implicitly forcing the migrated thread |
| 4625 | * off the CPU) |
| 4626 | * 3) it checks whether the migrated task is still in the wrong runqueue. |
| 4627 | * 4) if it's in the wrong runqueue then the migration thread removes |
| 4628 | * it and puts it into the right queue. |
| 4629 | * 5) stopper completes and stop_one_cpu() returns and the migration |
| 4630 | * is done. |
| 4631 | */ |
| 4632 | |
| 4633 | /* |
| 4634 | * Change a given task's CPU affinity. Migrate the thread to a |
| 4635 | * proper CPU and schedule it away if the CPU it's executing on |
| 4636 | * is removed from the allowed bitmask. |
| 4637 | * |
| 4638 | * NOTE: the caller must have a valid reference to the task, the |
| 4639 | * task must not exit() & deallocate itself prematurely. The |
| 4640 | * call is not atomic; no spinlocks may be held. |
| 4641 | */ |
| 4642 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) |
| 4643 | { |
| 4644 | unsigned long flags; |
| 4645 | struct rq *rq; |
| 4646 | unsigned int dest_cpu; |
| 4647 | int ret = 0; |
| 4648 | |
| 4649 | rq = task_rq_lock(p, &flags); |
| 4650 | |
| 4651 | if (cpumask_equal(&p->cpus_allowed, new_mask)) |
| 4652 | goto out; |
| 4653 | |
| 4654 | if (!cpumask_intersects(new_mask, cpu_active_mask)) { |
| 4655 | ret = -EINVAL; |
| 4656 | goto out; |
| 4657 | } |
| 4658 | |
| 4659 | if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { |
| 4660 | ret = -EINVAL; |
| 4661 | goto out; |
| 4662 | } |
| 4663 | |
| 4664 | do_set_cpus_allowed(p, new_mask); |
| 4665 | |
| 4666 | /* Can the task run on the task's current CPU? If so, we're done */ |
| 4667 | if (cpumask_test_cpu(task_cpu(p), new_mask)) |
| 4668 | goto out; |
| 4669 | |
| 4670 | dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); |
| 4671 | if (p->on_rq) { |
| 4672 | struct migration_arg arg = { p, dest_cpu }; |
| 4673 | /* Need help from migration thread: drop lock and wait. */ |
| 4674 | task_rq_unlock(rq, p, &flags); |
| 4675 | stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); |
| 4676 | tlb_migrate_finish(p->mm); |
| 4677 | return 0; |
| 4678 | } |
| 4679 | out: |
| 4680 | task_rq_unlock(rq, p, &flags); |
| 4681 | |
| 4682 | return ret; |
| 4683 | } |
| 4684 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); |
| 4685 | |
| 4686 | /* |
| 4687 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
| 4688 | * this because either it can't run here any more (set_cpus_allowed() |
| 4689 | * away from this CPU, or CPU going down), or because we're |
| 4690 | * attempting to rebalance this task on exec (sched_exec). |
| 4691 | * |
| 4692 | * So we race with normal scheduler movements, but that's OK, as long |
| 4693 | * as the task is no longer on this CPU. |
| 4694 | * |
| 4695 | * Returns non-zero if task was successfully migrated. |
| 4696 | */ |
| 4697 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
| 4698 | { |
| 4699 | struct rq *rq_dest, *rq_src; |
| 4700 | int ret = 0; |
| 4701 | |
| 4702 | if (unlikely(!cpu_active(dest_cpu))) |
| 4703 | return ret; |
| 4704 | |
| 4705 | rq_src = cpu_rq(src_cpu); |
| 4706 | rq_dest = cpu_rq(dest_cpu); |
| 4707 | |
| 4708 | raw_spin_lock(&p->pi_lock); |
| 4709 | double_rq_lock(rq_src, rq_dest); |
| 4710 | /* Already moved. */ |
| 4711 | if (task_cpu(p) != src_cpu) |
| 4712 | goto done; |
| 4713 | /* Affinity changed (again). */ |
| 4714 | if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) |
| 4715 | goto fail; |
| 4716 | |
| 4717 | /* |
| 4718 | * If we're not on a rq, the next wake-up will ensure we're |
| 4719 | * placed properly. |
| 4720 | */ |
| 4721 | if (p->on_rq) { |
| 4722 | dequeue_task(rq_src, p, 0); |
| 4723 | set_task_cpu(p, dest_cpu); |
| 4724 | enqueue_task(rq_dest, p, 0); |
| 4725 | check_preempt_curr(rq_dest, p, 0); |
| 4726 | } |
| 4727 | done: |
| 4728 | ret = 1; |
| 4729 | fail: |
| 4730 | double_rq_unlock(rq_src, rq_dest); |
| 4731 | raw_spin_unlock(&p->pi_lock); |
| 4732 | return ret; |
| 4733 | } |
| 4734 | |
| 4735 | /* |
| 4736 | * migration_cpu_stop - this will be executed by a highprio stopper thread |
| 4737 | * and performs thread migration by bumping thread off CPU then |
| 4738 | * 'pushing' onto another runqueue. |
| 4739 | */ |
| 4740 | static int migration_cpu_stop(void *data) |
| 4741 | { |
| 4742 | struct migration_arg *arg = data; |
| 4743 | |
| 4744 | /* |
| 4745 | * The original target cpu might have gone down and we might |
| 4746 | * be on another cpu but it doesn't matter. |
| 4747 | */ |
| 4748 | local_irq_disable(); |
| 4749 | __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); |
| 4750 | local_irq_enable(); |
| 4751 | return 0; |
| 4752 | } |
| 4753 | |
| 4754 | #ifdef CONFIG_HOTPLUG_CPU |
| 4755 | |
| 4756 | /* |
| 4757 | * Ensures that the idle task is using init_mm right before its cpu goes |
| 4758 | * offline. |
| 4759 | */ |
| 4760 | void idle_task_exit(void) |
| 4761 | { |
| 4762 | struct mm_struct *mm = current->active_mm; |
| 4763 | |
| 4764 | BUG_ON(cpu_online(smp_processor_id())); |
| 4765 | |
| 4766 | if (mm != &init_mm) |
| 4767 | switch_mm(mm, &init_mm, current); |
| 4768 | mmdrop(mm); |
| 4769 | } |
| 4770 | |
| 4771 | /* |
| 4772 | * Since this CPU is going 'away' for a while, fold any nr_active delta |
| 4773 | * we might have. Assumes we're called after migrate_tasks() so that the |
| 4774 | * nr_active count is stable. |
| 4775 | * |
| 4776 | * Also see the comment "Global load-average calculations". |
| 4777 | */ |
| 4778 | static void calc_load_migrate(struct rq *rq) |
| 4779 | { |
| 4780 | long delta = calc_load_fold_active(rq); |
| 4781 | if (delta) |
| 4782 | atomic_long_add(delta, &calc_load_tasks); |
| 4783 | } |
| 4784 | |
| 4785 | /* |
| 4786 | * Migrate all tasks from the rq, sleeping tasks will be migrated by |
| 4787 | * try_to_wake_up()->select_task_rq(). |
| 4788 | * |
| 4789 | * Called with rq->lock held even though we'er in stop_machine() and |
| 4790 | * there's no concurrency possible, we hold the required locks anyway |
| 4791 | * because of lock validation efforts. |
| 4792 | */ |
| 4793 | static void migrate_tasks(unsigned int dead_cpu) |
| 4794 | { |
| 4795 | struct rq *rq = cpu_rq(dead_cpu); |
| 4796 | struct task_struct *next, *stop = rq->stop; |
| 4797 | int dest_cpu; |
| 4798 | |
| 4799 | /* |
| 4800 | * Fudge the rq selection such that the below task selection loop |
| 4801 | * doesn't get stuck on the currently eligible stop task. |
| 4802 | * |
| 4803 | * We're currently inside stop_machine() and the rq is either stuck |
| 4804 | * in the stop_machine_cpu_stop() loop, or we're executing this code, |
| 4805 | * either way we should never end up calling schedule() until we're |
| 4806 | * done here. |
| 4807 | */ |
| 4808 | rq->stop = NULL; |
| 4809 | |
| 4810 | for ( ; ; ) { |
| 4811 | /* |
| 4812 | * There's this thread running, bail when that's the only |
| 4813 | * remaining thread. |
| 4814 | */ |
| 4815 | if (rq->nr_running == 1) |
| 4816 | break; |
| 4817 | |
| 4818 | next = pick_next_task(rq); |
| 4819 | BUG_ON(!next); |
| 4820 | next->sched_class->put_prev_task(rq, next); |
| 4821 | |
| 4822 | /* Find suitable destination for @next, with force if needed. */ |
| 4823 | dest_cpu = select_fallback_rq(dead_cpu, next); |
| 4824 | raw_spin_unlock(&rq->lock); |
| 4825 | |
| 4826 | __migrate_task(next, dead_cpu, dest_cpu); |
| 4827 | |
| 4828 | raw_spin_lock(&rq->lock); |
| 4829 | } |
| 4830 | |
| 4831 | rq->stop = stop; |
| 4832 | } |
| 4833 | |
| 4834 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 4835 | |
| 4836 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) |
| 4837 | |
| 4838 | static struct ctl_table sd_ctl_dir[] = { |
| 4839 | { |
| 4840 | .procname = "sched_domain", |
| 4841 | .mode = 0555, |
| 4842 | }, |
| 4843 | {} |
| 4844 | }; |
| 4845 | |
| 4846 | static struct ctl_table sd_ctl_root[] = { |
| 4847 | { |
| 4848 | .procname = "kernel", |
| 4849 | .mode = 0555, |
| 4850 | .child = sd_ctl_dir, |
| 4851 | }, |
| 4852 | {} |
| 4853 | }; |
| 4854 | |
| 4855 | static struct ctl_table *sd_alloc_ctl_entry(int n) |
| 4856 | { |
| 4857 | struct ctl_table *entry = |
| 4858 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); |
| 4859 | |
| 4860 | return entry; |
| 4861 | } |
| 4862 | |
| 4863 | static void sd_free_ctl_entry(struct ctl_table **tablep) |
| 4864 | { |
| 4865 | struct ctl_table *entry; |
| 4866 | |
| 4867 | /* |
| 4868 | * In the intermediate directories, both the child directory and |
| 4869 | * procname are dynamically allocated and could fail but the mode |
| 4870 | * will always be set. In the lowest directory the names are |
| 4871 | * static strings and all have proc handlers. |
| 4872 | */ |
| 4873 | for (entry = *tablep; entry->mode; entry++) { |
| 4874 | if (entry->child) |
| 4875 | sd_free_ctl_entry(&entry->child); |
| 4876 | if (entry->proc_handler == NULL) |
| 4877 | kfree(entry->procname); |
| 4878 | } |
| 4879 | |
| 4880 | kfree(*tablep); |
| 4881 | *tablep = NULL; |
| 4882 | } |
| 4883 | |
| 4884 | static int min_load_idx = 0; |
| 4885 | static int max_load_idx = CPU_LOAD_IDX_MAX; |
| 4886 | |
| 4887 | static void |
| 4888 | set_table_entry(struct ctl_table *entry, |
| 4889 | const char *procname, void *data, int maxlen, |
| 4890 | umode_t mode, proc_handler *proc_handler, |
| 4891 | bool load_idx) |
| 4892 | { |
| 4893 | entry->procname = procname; |
| 4894 | entry->data = data; |
| 4895 | entry->maxlen = maxlen; |
| 4896 | entry->mode = mode; |
| 4897 | entry->proc_handler = proc_handler; |
| 4898 | |
| 4899 | if (load_idx) { |
| 4900 | entry->extra1 = &min_load_idx; |
| 4901 | entry->extra2 = &max_load_idx; |
| 4902 | } |
| 4903 | } |
| 4904 | |
| 4905 | static struct ctl_table * |
| 4906 | sd_alloc_ctl_domain_table(struct sched_domain *sd) |
| 4907 | { |
| 4908 | struct ctl_table *table = sd_alloc_ctl_entry(13); |
| 4909 | |
| 4910 | if (table == NULL) |
| 4911 | return NULL; |
| 4912 | |
| 4913 | set_table_entry(&table[0], "min_interval", &sd->min_interval, |
| 4914 | sizeof(long), 0644, proc_doulongvec_minmax, false); |
| 4915 | set_table_entry(&table[1], "max_interval", &sd->max_interval, |
| 4916 | sizeof(long), 0644, proc_doulongvec_minmax, false); |
| 4917 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, |
| 4918 | sizeof(int), 0644, proc_dointvec_minmax, true); |
| 4919 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, |
| 4920 | sizeof(int), 0644, proc_dointvec_minmax, true); |
| 4921 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, |
| 4922 | sizeof(int), 0644, proc_dointvec_minmax, true); |
| 4923 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, |
| 4924 | sizeof(int), 0644, proc_dointvec_minmax, true); |
| 4925 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, |
| 4926 | sizeof(int), 0644, proc_dointvec_minmax, true); |
| 4927 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, |
| 4928 | sizeof(int), 0644, proc_dointvec_minmax, false); |
| 4929 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, |
| 4930 | sizeof(int), 0644, proc_dointvec_minmax, false); |
| 4931 | set_table_entry(&table[9], "cache_nice_tries", |
| 4932 | &sd->cache_nice_tries, |
| 4933 | sizeof(int), 0644, proc_dointvec_minmax, false); |
| 4934 | set_table_entry(&table[10], "flags", &sd->flags, |
| 4935 | sizeof(int), 0644, proc_dointvec_minmax, false); |
| 4936 | set_table_entry(&table[11], "name", sd->name, |
| 4937 | CORENAME_MAX_SIZE, 0444, proc_dostring, false); |
| 4938 | /* &table[12] is terminator */ |
| 4939 | |
| 4940 | return table; |
| 4941 | } |
| 4942 | |
| 4943 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) |
| 4944 | { |
| 4945 | struct ctl_table *entry, *table; |
| 4946 | struct sched_domain *sd; |
| 4947 | int domain_num = 0, i; |
| 4948 | char buf[32]; |
| 4949 | |
| 4950 | for_each_domain(cpu, sd) |
| 4951 | domain_num++; |
| 4952 | entry = table = sd_alloc_ctl_entry(domain_num + 1); |
| 4953 | if (table == NULL) |
| 4954 | return NULL; |
| 4955 | |
| 4956 | i = 0; |
| 4957 | for_each_domain(cpu, sd) { |
| 4958 | snprintf(buf, 32, "domain%d", i); |
| 4959 | entry->procname = kstrdup(buf, GFP_KERNEL); |
| 4960 | entry->mode = 0555; |
| 4961 | entry->child = sd_alloc_ctl_domain_table(sd); |
| 4962 | entry++; |
| 4963 | i++; |
| 4964 | } |
| 4965 | return table; |
| 4966 | } |
| 4967 | |
| 4968 | static struct ctl_table_header *sd_sysctl_header; |
| 4969 | static void register_sched_domain_sysctl(void) |
| 4970 | { |
| 4971 | int i, cpu_num = num_possible_cpus(); |
| 4972 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); |
| 4973 | char buf[32]; |
| 4974 | |
| 4975 | WARN_ON(sd_ctl_dir[0].child); |
| 4976 | sd_ctl_dir[0].child = entry; |
| 4977 | |
| 4978 | if (entry == NULL) |
| 4979 | return; |
| 4980 | |
| 4981 | for_each_possible_cpu(i) { |
| 4982 | snprintf(buf, 32, "cpu%d", i); |
| 4983 | entry->procname = kstrdup(buf, GFP_KERNEL); |
| 4984 | entry->mode = 0555; |
| 4985 | entry->child = sd_alloc_ctl_cpu_table(i); |
| 4986 | entry++; |
| 4987 | } |
| 4988 | |
| 4989 | WARN_ON(sd_sysctl_header); |
| 4990 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); |
| 4991 | } |
| 4992 | |
| 4993 | /* may be called multiple times per register */ |
| 4994 | static void unregister_sched_domain_sysctl(void) |
| 4995 | { |
| 4996 | if (sd_sysctl_header) |
| 4997 | unregister_sysctl_table(sd_sysctl_header); |
| 4998 | sd_sysctl_header = NULL; |
| 4999 | if (sd_ctl_dir[0].child) |
| 5000 | sd_free_ctl_entry(&sd_ctl_dir[0].child); |
| 5001 | } |
| 5002 | #else |
| 5003 | static void register_sched_domain_sysctl(void) |
| 5004 | { |
| 5005 | } |
| 5006 | static void unregister_sched_domain_sysctl(void) |
| 5007 | { |
| 5008 | } |
| 5009 | #endif |
| 5010 | |
| 5011 | static void set_rq_online(struct rq *rq) |
| 5012 | { |
| 5013 | if (!rq->online) { |
| 5014 | const struct sched_class *class; |
| 5015 | |
| 5016 | cpumask_set_cpu(rq->cpu, rq->rd->online); |
| 5017 | rq->online = 1; |
| 5018 | |
| 5019 | for_each_class(class) { |
| 5020 | if (class->rq_online) |
| 5021 | class->rq_online(rq); |
| 5022 | } |
| 5023 | } |
| 5024 | } |
| 5025 | |
| 5026 | static void set_rq_offline(struct rq *rq) |
| 5027 | { |
| 5028 | if (rq->online) { |
| 5029 | const struct sched_class *class; |
| 5030 | |
| 5031 | for_each_class(class) { |
| 5032 | if (class->rq_offline) |
| 5033 | class->rq_offline(rq); |
| 5034 | } |
| 5035 | |
| 5036 | cpumask_clear_cpu(rq->cpu, rq->rd->online); |
| 5037 | rq->online = 0; |
| 5038 | } |
| 5039 | } |
| 5040 | |
| 5041 | /* |
| 5042 | * migration_call - callback that gets triggered when a CPU is added. |
| 5043 | * Here we can start up the necessary migration thread for the new CPU. |
| 5044 | */ |
| 5045 | static int __cpuinit |
| 5046 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) |
| 5047 | { |
| 5048 | int cpu = (long)hcpu; |
| 5049 | unsigned long flags; |
| 5050 | struct rq *rq = cpu_rq(cpu); |
| 5051 | |
| 5052 | switch (action & ~CPU_TASKS_FROZEN) { |
| 5053 | |
| 5054 | case CPU_UP_PREPARE: |
| 5055 | rq->calc_load_update = calc_load_update; |
| 5056 | break; |
| 5057 | |
| 5058 | case CPU_ONLINE: |
| 5059 | /* Update our root-domain */ |
| 5060 | raw_spin_lock_irqsave(&rq->lock, flags); |
| 5061 | if (rq->rd) { |
| 5062 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
| 5063 | |
| 5064 | set_rq_online(rq); |
| 5065 | } |
| 5066 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
| 5067 | break; |
| 5068 | |
| 5069 | #ifdef CONFIG_HOTPLUG_CPU |
| 5070 | case CPU_DYING: |
| 5071 | sched_ttwu_pending(); |
| 5072 | /* Update our root-domain */ |
| 5073 | raw_spin_lock_irqsave(&rq->lock, flags); |
| 5074 | if (rq->rd) { |
| 5075 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
| 5076 | set_rq_offline(rq); |
| 5077 | } |
| 5078 | migrate_tasks(cpu); |
| 5079 | BUG_ON(rq->nr_running != 1); /* the migration thread */ |
| 5080 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
| 5081 | break; |
| 5082 | |
| 5083 | case CPU_DEAD: |
| 5084 | calc_load_migrate(rq); |
| 5085 | break; |
| 5086 | #endif |
| 5087 | } |
| 5088 | |
| 5089 | update_max_interval(); |
| 5090 | |
| 5091 | return NOTIFY_OK; |
| 5092 | } |
| 5093 | |
| 5094 | /* |
| 5095 | * Register at high priority so that task migration (migrate_all_tasks) |
| 5096 | * happens before everything else. This has to be lower priority than |
| 5097 | * the notifier in the perf_event subsystem, though. |
| 5098 | */ |
| 5099 | static struct notifier_block __cpuinitdata migration_notifier = { |
| 5100 | .notifier_call = migration_call, |
| 5101 | .priority = CPU_PRI_MIGRATION, |
| 5102 | }; |
| 5103 | |
| 5104 | static int __cpuinit sched_cpu_active(struct notifier_block *nfb, |
| 5105 | unsigned long action, void *hcpu) |
| 5106 | { |
| 5107 | switch (action & ~CPU_TASKS_FROZEN) { |
| 5108 | case CPU_STARTING: |
| 5109 | case CPU_DOWN_FAILED: |
| 5110 | set_cpu_active((long)hcpu, true); |
| 5111 | return NOTIFY_OK; |
| 5112 | default: |
| 5113 | return NOTIFY_DONE; |
| 5114 | } |
| 5115 | } |
| 5116 | |
| 5117 | static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, |
| 5118 | unsigned long action, void *hcpu) |
| 5119 | { |
| 5120 | switch (action & ~CPU_TASKS_FROZEN) { |
| 5121 | case CPU_DOWN_PREPARE: |
| 5122 | set_cpu_active((long)hcpu, false); |
| 5123 | return NOTIFY_OK; |
| 5124 | default: |
| 5125 | return NOTIFY_DONE; |
| 5126 | } |
| 5127 | } |
| 5128 | |
| 5129 | static int __init migration_init(void) |
| 5130 | { |
| 5131 | void *cpu = (void *)(long)smp_processor_id(); |
| 5132 | int err; |
| 5133 | |
| 5134 | /* Initialize migration for the boot CPU */ |
| 5135 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
| 5136 | BUG_ON(err == NOTIFY_BAD); |
| 5137 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
| 5138 | register_cpu_notifier(&migration_notifier); |
| 5139 | |
| 5140 | /* Register cpu active notifiers */ |
| 5141 | cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); |
| 5142 | cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); |
| 5143 | |
| 5144 | return 0; |
| 5145 | } |
| 5146 | early_initcall(migration_init); |
| 5147 | #endif |
| 5148 | |
| 5149 | #ifdef CONFIG_SMP |
| 5150 | |
| 5151 | static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ |
| 5152 | |
| 5153 | #ifdef CONFIG_SCHED_DEBUG |
| 5154 | |
| 5155 | static __read_mostly int sched_debug_enabled; |
| 5156 | |
| 5157 | static int __init sched_debug_setup(char *str) |
| 5158 | { |
| 5159 | sched_debug_enabled = 1; |
| 5160 | |
| 5161 | return 0; |
| 5162 | } |
| 5163 | early_param("sched_debug", sched_debug_setup); |
| 5164 | |
| 5165 | static inline bool sched_debug(void) |
| 5166 | { |
| 5167 | return sched_debug_enabled; |
| 5168 | } |
| 5169 | |
| 5170 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, |
| 5171 | struct cpumask *groupmask) |
| 5172 | { |
| 5173 | struct sched_group *group = sd->groups; |
| 5174 | char str[256]; |
| 5175 | |
| 5176 | cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); |
| 5177 | cpumask_clear(groupmask); |
| 5178 | |
| 5179 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); |
| 5180 | |
| 5181 | if (!(sd->flags & SD_LOAD_BALANCE)) { |
| 5182 | printk("does not load-balance\n"); |
| 5183 | if (sd->parent) |
| 5184 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" |
| 5185 | " has parent"); |
| 5186 | return -1; |
| 5187 | } |
| 5188 | |
| 5189 | printk(KERN_CONT "span %s level %s\n", str, sd->name); |
| 5190 | |
| 5191 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
| 5192 | printk(KERN_ERR "ERROR: domain->span does not contain " |
| 5193 | "CPU%d\n", cpu); |
| 5194 | } |
| 5195 | if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { |
| 5196 | printk(KERN_ERR "ERROR: domain->groups does not contain" |
| 5197 | " CPU%d\n", cpu); |
| 5198 | } |
| 5199 | |
| 5200 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); |
| 5201 | do { |
| 5202 | if (!group) { |
| 5203 | printk("\n"); |
| 5204 | printk(KERN_ERR "ERROR: group is NULL\n"); |
| 5205 | break; |
| 5206 | } |
| 5207 | |
| 5208 | /* |
| 5209 | * Even though we initialize ->power to something semi-sane, |
| 5210 | * we leave power_orig unset. This allows us to detect if |
| 5211 | * domain iteration is still funny without causing /0 traps. |
| 5212 | */ |
| 5213 | if (!group->sgp->power_orig) { |
| 5214 | printk(KERN_CONT "\n"); |
| 5215 | printk(KERN_ERR "ERROR: domain->cpu_power not " |
| 5216 | "set\n"); |
| 5217 | break; |
| 5218 | } |
| 5219 | |
| 5220 | if (!cpumask_weight(sched_group_cpus(group))) { |
| 5221 | printk(KERN_CONT "\n"); |
| 5222 | printk(KERN_ERR "ERROR: empty group\n"); |
| 5223 | break; |
| 5224 | } |
| 5225 | |
| 5226 | if (!(sd->flags & SD_OVERLAP) && |
| 5227 | cpumask_intersects(groupmask, sched_group_cpus(group))) { |
| 5228 | printk(KERN_CONT "\n"); |
| 5229 | printk(KERN_ERR "ERROR: repeated CPUs\n"); |
| 5230 | break; |
| 5231 | } |
| 5232 | |
| 5233 | cpumask_or(groupmask, groupmask, sched_group_cpus(group)); |
| 5234 | |
| 5235 | cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); |
| 5236 | |
| 5237 | printk(KERN_CONT " %s", str); |
| 5238 | if (group->sgp->power != SCHED_POWER_SCALE) { |
| 5239 | printk(KERN_CONT " (cpu_power = %d)", |
| 5240 | group->sgp->power); |
| 5241 | } |
| 5242 | |
| 5243 | group = group->next; |
| 5244 | } while (group != sd->groups); |
| 5245 | printk(KERN_CONT "\n"); |
| 5246 | |
| 5247 | if (!cpumask_equal(sched_domain_span(sd), groupmask)) |
| 5248 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
| 5249 | |
| 5250 | if (sd->parent && |
| 5251 | !cpumask_subset(groupmask, sched_domain_span(sd->parent))) |
| 5252 | printk(KERN_ERR "ERROR: parent span is not a superset " |
| 5253 | "of domain->span\n"); |
| 5254 | return 0; |
| 5255 | } |
| 5256 | |
| 5257 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
| 5258 | { |
| 5259 | int level = 0; |
| 5260 | |
| 5261 | if (!sched_debug_enabled) |
| 5262 | return; |
| 5263 | |
| 5264 | if (!sd) { |
| 5265 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); |
| 5266 | return; |
| 5267 | } |
| 5268 | |
| 5269 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
| 5270 | |
| 5271 | for (;;) { |
| 5272 | if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) |
| 5273 | break; |
| 5274 | level++; |
| 5275 | sd = sd->parent; |
| 5276 | if (!sd) |
| 5277 | break; |
| 5278 | } |
| 5279 | } |
| 5280 | #else /* !CONFIG_SCHED_DEBUG */ |
| 5281 | # define sched_domain_debug(sd, cpu) do { } while (0) |
| 5282 | static inline bool sched_debug(void) |
| 5283 | { |
| 5284 | return false; |
| 5285 | } |
| 5286 | #endif /* CONFIG_SCHED_DEBUG */ |
| 5287 | |
| 5288 | static int sd_degenerate(struct sched_domain *sd) |
| 5289 | { |
| 5290 | if (cpumask_weight(sched_domain_span(sd)) == 1) |
| 5291 | return 1; |
| 5292 | |
| 5293 | /* Following flags need at least 2 groups */ |
| 5294 | if (sd->flags & (SD_LOAD_BALANCE | |
| 5295 | SD_BALANCE_NEWIDLE | |
| 5296 | SD_BALANCE_FORK | |
| 5297 | SD_BALANCE_EXEC | |
| 5298 | SD_SHARE_CPUPOWER | |
| 5299 | SD_SHARE_PKG_RESOURCES)) { |
| 5300 | if (sd->groups != sd->groups->next) |
| 5301 | return 0; |
| 5302 | } |
| 5303 | |
| 5304 | /* Following flags don't use groups */ |
| 5305 | if (sd->flags & (SD_WAKE_AFFINE)) |
| 5306 | return 0; |
| 5307 | |
| 5308 | return 1; |
| 5309 | } |
| 5310 | |
| 5311 | static int |
| 5312 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) |
| 5313 | { |
| 5314 | unsigned long cflags = sd->flags, pflags = parent->flags; |
| 5315 | |
| 5316 | if (sd_degenerate(parent)) |
| 5317 | return 1; |
| 5318 | |
| 5319 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) |
| 5320 | return 0; |
| 5321 | |
| 5322 | /* Flags needing groups don't count if only 1 group in parent */ |
| 5323 | if (parent->groups == parent->groups->next) { |
| 5324 | pflags &= ~(SD_LOAD_BALANCE | |
| 5325 | SD_BALANCE_NEWIDLE | |
| 5326 | SD_BALANCE_FORK | |
| 5327 | SD_BALANCE_EXEC | |
| 5328 | SD_SHARE_CPUPOWER | |
| 5329 | SD_SHARE_PKG_RESOURCES); |
| 5330 | if (nr_node_ids == 1) |
| 5331 | pflags &= ~SD_SERIALIZE; |
| 5332 | } |
| 5333 | if (~cflags & pflags) |
| 5334 | return 0; |
| 5335 | |
| 5336 | return 1; |
| 5337 | } |
| 5338 | |
| 5339 | static void free_rootdomain(struct rcu_head *rcu) |
| 5340 | { |
| 5341 | struct root_domain *rd = container_of(rcu, struct root_domain, rcu); |
| 5342 | |
| 5343 | cpupri_cleanup(&rd->cpupri); |
| 5344 | free_cpumask_var(rd->rto_mask); |
| 5345 | free_cpumask_var(rd->online); |
| 5346 | free_cpumask_var(rd->span); |
| 5347 | kfree(rd); |
| 5348 | } |
| 5349 | |
| 5350 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) |
| 5351 | { |
| 5352 | struct root_domain *old_rd = NULL; |
| 5353 | unsigned long flags; |
| 5354 | |
| 5355 | raw_spin_lock_irqsave(&rq->lock, flags); |
| 5356 | |
| 5357 | if (rq->rd) { |
| 5358 | old_rd = rq->rd; |
| 5359 | |
| 5360 | if (cpumask_test_cpu(rq->cpu, old_rd->online)) |
| 5361 | set_rq_offline(rq); |
| 5362 | |
| 5363 | cpumask_clear_cpu(rq->cpu, old_rd->span); |
| 5364 | |
| 5365 | /* |
| 5366 | * If we dont want to free the old_rt yet then |
| 5367 | * set old_rd to NULL to skip the freeing later |
| 5368 | * in this function: |
| 5369 | */ |
| 5370 | if (!atomic_dec_and_test(&old_rd->refcount)) |
| 5371 | old_rd = NULL; |
| 5372 | } |
| 5373 | |
| 5374 | atomic_inc(&rd->refcount); |
| 5375 | rq->rd = rd; |
| 5376 | |
| 5377 | cpumask_set_cpu(rq->cpu, rd->span); |
| 5378 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) |
| 5379 | set_rq_online(rq); |
| 5380 | |
| 5381 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
| 5382 | |
| 5383 | if (old_rd) |
| 5384 | call_rcu_sched(&old_rd->rcu, free_rootdomain); |
| 5385 | } |
| 5386 | |
| 5387 | static int init_rootdomain(struct root_domain *rd) |
| 5388 | { |
| 5389 | memset(rd, 0, sizeof(*rd)); |
| 5390 | |
| 5391 | if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) |
| 5392 | goto out; |
| 5393 | if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) |
| 5394 | goto free_span; |
| 5395 | if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) |
| 5396 | goto free_online; |
| 5397 | |
| 5398 | if (cpupri_init(&rd->cpupri) != 0) |
| 5399 | goto free_rto_mask; |
| 5400 | return 0; |
| 5401 | |
| 5402 | free_rto_mask: |
| 5403 | free_cpumask_var(rd->rto_mask); |
| 5404 | free_online: |
| 5405 | free_cpumask_var(rd->online); |
| 5406 | free_span: |
| 5407 | free_cpumask_var(rd->span); |
| 5408 | out: |
| 5409 | return -ENOMEM; |
| 5410 | } |
| 5411 | |
| 5412 | /* |
| 5413 | * By default the system creates a single root-domain with all cpus as |
| 5414 | * members (mimicking the global state we have today). |
| 5415 | */ |
| 5416 | struct root_domain def_root_domain; |
| 5417 | |
| 5418 | static void init_defrootdomain(void) |
| 5419 | { |
| 5420 | init_rootdomain(&def_root_domain); |
| 5421 | |
| 5422 | atomic_set(&def_root_domain.refcount, 1); |
| 5423 | } |
| 5424 | |
| 5425 | static struct root_domain *alloc_rootdomain(void) |
| 5426 | { |
| 5427 | struct root_domain *rd; |
| 5428 | |
| 5429 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); |
| 5430 | if (!rd) |
| 5431 | return NULL; |
| 5432 | |
| 5433 | if (init_rootdomain(rd) != 0) { |
| 5434 | kfree(rd); |
| 5435 | return NULL; |
| 5436 | } |
| 5437 | |
| 5438 | return rd; |
| 5439 | } |
| 5440 | |
| 5441 | static void free_sched_groups(struct sched_group *sg, int free_sgp) |
| 5442 | { |
| 5443 | struct sched_group *tmp, *first; |
| 5444 | |
| 5445 | if (!sg) |
| 5446 | return; |
| 5447 | |
| 5448 | first = sg; |
| 5449 | do { |
| 5450 | tmp = sg->next; |
| 5451 | |
| 5452 | if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) |
| 5453 | kfree(sg->sgp); |
| 5454 | |
| 5455 | kfree(sg); |
| 5456 | sg = tmp; |
| 5457 | } while (sg != first); |
| 5458 | } |
| 5459 | |
| 5460 | static void free_sched_domain(struct rcu_head *rcu) |
| 5461 | { |
| 5462 | struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); |
| 5463 | |
| 5464 | /* |
| 5465 | * If its an overlapping domain it has private groups, iterate and |
| 5466 | * nuke them all. |
| 5467 | */ |
| 5468 | if (sd->flags & SD_OVERLAP) { |
| 5469 | free_sched_groups(sd->groups, 1); |
| 5470 | } else if (atomic_dec_and_test(&sd->groups->ref)) { |
| 5471 | kfree(sd->groups->sgp); |
| 5472 | kfree(sd->groups); |
| 5473 | } |
| 5474 | kfree(sd); |
| 5475 | } |
| 5476 | |
| 5477 | static void destroy_sched_domain(struct sched_domain *sd, int cpu) |
| 5478 | { |
| 5479 | call_rcu(&sd->rcu, free_sched_domain); |
| 5480 | } |
| 5481 | |
| 5482 | static void destroy_sched_domains(struct sched_domain *sd, int cpu) |
| 5483 | { |
| 5484 | for (; sd; sd = sd->parent) |
| 5485 | destroy_sched_domain(sd, cpu); |
| 5486 | } |
| 5487 | |
| 5488 | /* |
| 5489 | * Keep a special pointer to the highest sched_domain that has |
| 5490 | * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this |
| 5491 | * allows us to avoid some pointer chasing select_idle_sibling(). |
| 5492 | * |
| 5493 | * Also keep a unique ID per domain (we use the first cpu number in |
| 5494 | * the cpumask of the domain), this allows us to quickly tell if |
| 5495 | * two cpus are in the same cache domain, see cpus_share_cache(). |
| 5496 | */ |
| 5497 | DEFINE_PER_CPU(struct sched_domain *, sd_llc); |
| 5498 | DEFINE_PER_CPU(int, sd_llc_id); |
| 5499 | |
| 5500 | static void update_top_cache_domain(int cpu) |
| 5501 | { |
| 5502 | struct sched_domain *sd; |
| 5503 | int id = cpu; |
| 5504 | |
| 5505 | sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); |
| 5506 | if (sd) |
| 5507 | id = cpumask_first(sched_domain_span(sd)); |
| 5508 | |
| 5509 | rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); |
| 5510 | per_cpu(sd_llc_id, cpu) = id; |
| 5511 | } |
| 5512 | |
| 5513 | /* |
| 5514 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
| 5515 | * hold the hotplug lock. |
| 5516 | */ |
| 5517 | static void |
| 5518 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) |
| 5519 | { |
| 5520 | struct rq *rq = cpu_rq(cpu); |
| 5521 | struct sched_domain *tmp; |
| 5522 | |
| 5523 | /* Remove the sched domains which do not contribute to scheduling. */ |
| 5524 | for (tmp = sd; tmp; ) { |
| 5525 | struct sched_domain *parent = tmp->parent; |
| 5526 | if (!parent) |
| 5527 | break; |
| 5528 | |
| 5529 | if (sd_parent_degenerate(tmp, parent)) { |
| 5530 | tmp->parent = parent->parent; |
| 5531 | if (parent->parent) |
| 5532 | parent->parent->child = tmp; |
| 5533 | destroy_sched_domain(parent, cpu); |
| 5534 | } else |
| 5535 | tmp = tmp->parent; |
| 5536 | } |
| 5537 | |
| 5538 | if (sd && sd_degenerate(sd)) { |
| 5539 | tmp = sd; |
| 5540 | sd = sd->parent; |
| 5541 | destroy_sched_domain(tmp, cpu); |
| 5542 | if (sd) |
| 5543 | sd->child = NULL; |
| 5544 | } |
| 5545 | |
| 5546 | sched_domain_debug(sd, cpu); |
| 5547 | |
| 5548 | rq_attach_root(rq, rd); |
| 5549 | tmp = rq->sd; |
| 5550 | rcu_assign_pointer(rq->sd, sd); |
| 5551 | destroy_sched_domains(tmp, cpu); |
| 5552 | |
| 5553 | update_top_cache_domain(cpu); |
| 5554 | } |
| 5555 | |
| 5556 | /* cpus with isolated domains */ |
| 5557 | static cpumask_var_t cpu_isolated_map; |
| 5558 | |
| 5559 | /* Setup the mask of cpus configured for isolated domains */ |
| 5560 | static int __init isolated_cpu_setup(char *str) |
| 5561 | { |
| 5562 | alloc_bootmem_cpumask_var(&cpu_isolated_map); |
| 5563 | cpulist_parse(str, cpu_isolated_map); |
| 5564 | return 1; |
| 5565 | } |
| 5566 | |
| 5567 | __setup("isolcpus=", isolated_cpu_setup); |
| 5568 | |
| 5569 | static const struct cpumask *cpu_cpu_mask(int cpu) |
| 5570 | { |
| 5571 | return cpumask_of_node(cpu_to_node(cpu)); |
| 5572 | } |
| 5573 | |
| 5574 | struct sd_data { |
| 5575 | struct sched_domain **__percpu sd; |
| 5576 | struct sched_group **__percpu sg; |
| 5577 | struct sched_group_power **__percpu sgp; |
| 5578 | }; |
| 5579 | |
| 5580 | struct s_data { |
| 5581 | struct sched_domain ** __percpu sd; |
| 5582 | struct root_domain *rd; |
| 5583 | }; |
| 5584 | |
| 5585 | enum s_alloc { |
| 5586 | sa_rootdomain, |
| 5587 | sa_sd, |
| 5588 | sa_sd_storage, |
| 5589 | sa_none, |
| 5590 | }; |
| 5591 | |
| 5592 | struct sched_domain_topology_level; |
| 5593 | |
| 5594 | typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); |
| 5595 | typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); |
| 5596 | |
| 5597 | #define SDTL_OVERLAP 0x01 |
| 5598 | |
| 5599 | struct sched_domain_topology_level { |
| 5600 | sched_domain_init_f init; |
| 5601 | sched_domain_mask_f mask; |
| 5602 | int flags; |
| 5603 | int numa_level; |
| 5604 | struct sd_data data; |
| 5605 | }; |
| 5606 | |
| 5607 | /* |
| 5608 | * Build an iteration mask that can exclude certain CPUs from the upwards |
| 5609 | * domain traversal. |
| 5610 | * |
| 5611 | * Asymmetric node setups can result in situations where the domain tree is of |
| 5612 | * unequal depth, make sure to skip domains that already cover the entire |
| 5613 | * range. |
| 5614 | * |
| 5615 | * In that case build_sched_domains() will have terminated the iteration early |
| 5616 | * and our sibling sd spans will be empty. Domains should always include the |
| 5617 | * cpu they're built on, so check that. |
| 5618 | * |
| 5619 | */ |
| 5620 | static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) |
| 5621 | { |
| 5622 | const struct cpumask *span = sched_domain_span(sd); |
| 5623 | struct sd_data *sdd = sd->private; |
| 5624 | struct sched_domain *sibling; |
| 5625 | int i; |
| 5626 | |
| 5627 | for_each_cpu(i, span) { |
| 5628 | sibling = *per_cpu_ptr(sdd->sd, i); |
| 5629 | if (!cpumask_test_cpu(i, sched_domain_span(sibling))) |
| 5630 | continue; |
| 5631 | |
| 5632 | cpumask_set_cpu(i, sched_group_mask(sg)); |
| 5633 | } |
| 5634 | } |
| 5635 | |
| 5636 | /* |
| 5637 | * Return the canonical balance cpu for this group, this is the first cpu |
| 5638 | * of this group that's also in the iteration mask. |
| 5639 | */ |
| 5640 | int group_balance_cpu(struct sched_group *sg) |
| 5641 | { |
| 5642 | return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); |
| 5643 | } |
| 5644 | |
| 5645 | static int |
| 5646 | build_overlap_sched_groups(struct sched_domain *sd, int cpu) |
| 5647 | { |
| 5648 | struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; |
| 5649 | const struct cpumask *span = sched_domain_span(sd); |
| 5650 | struct cpumask *covered = sched_domains_tmpmask; |
| 5651 | struct sd_data *sdd = sd->private; |
| 5652 | struct sched_domain *child; |
| 5653 | int i; |
| 5654 | |
| 5655 | cpumask_clear(covered); |
| 5656 | |
| 5657 | for_each_cpu(i, span) { |
| 5658 | struct cpumask *sg_span; |
| 5659 | |
| 5660 | if (cpumask_test_cpu(i, covered)) |
| 5661 | continue; |
| 5662 | |
| 5663 | child = *per_cpu_ptr(sdd->sd, i); |
| 5664 | |
| 5665 | /* See the comment near build_group_mask(). */ |
| 5666 | if (!cpumask_test_cpu(i, sched_domain_span(child))) |
| 5667 | continue; |
| 5668 | |
| 5669 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), |
| 5670 | GFP_KERNEL, cpu_to_node(cpu)); |
| 5671 | |
| 5672 | if (!sg) |
| 5673 | goto fail; |
| 5674 | |
| 5675 | sg_span = sched_group_cpus(sg); |
| 5676 | if (child->child) { |
| 5677 | child = child->child; |
| 5678 | cpumask_copy(sg_span, sched_domain_span(child)); |
| 5679 | } else |
| 5680 | cpumask_set_cpu(i, sg_span); |
| 5681 | |
| 5682 | cpumask_or(covered, covered, sg_span); |
| 5683 | |
| 5684 | sg->sgp = *per_cpu_ptr(sdd->sgp, i); |
| 5685 | if (atomic_inc_return(&sg->sgp->ref) == 1) |
| 5686 | build_group_mask(sd, sg); |
| 5687 | |
| 5688 | /* |
| 5689 | * Initialize sgp->power such that even if we mess up the |
| 5690 | * domains and no possible iteration will get us here, we won't |
| 5691 | * die on a /0 trap. |
| 5692 | */ |
| 5693 | sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); |
| 5694 | |
| 5695 | /* |
| 5696 | * Make sure the first group of this domain contains the |
| 5697 | * canonical balance cpu. Otherwise the sched_domain iteration |
| 5698 | * breaks. See update_sg_lb_stats(). |
| 5699 | */ |
| 5700 | if ((!groups && cpumask_test_cpu(cpu, sg_span)) || |
| 5701 | group_balance_cpu(sg) == cpu) |
| 5702 | groups = sg; |
| 5703 | |
| 5704 | if (!first) |
| 5705 | first = sg; |
| 5706 | if (last) |
| 5707 | last->next = sg; |
| 5708 | last = sg; |
| 5709 | last->next = first; |
| 5710 | } |
| 5711 | sd->groups = groups; |
| 5712 | |
| 5713 | return 0; |
| 5714 | |
| 5715 | fail: |
| 5716 | free_sched_groups(first, 0); |
| 5717 | |
| 5718 | return -ENOMEM; |
| 5719 | } |
| 5720 | |
| 5721 | static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) |
| 5722 | { |
| 5723 | struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); |
| 5724 | struct sched_domain *child = sd->child; |
| 5725 | |
| 5726 | if (child) |
| 5727 | cpu = cpumask_first(sched_domain_span(child)); |
| 5728 | |
| 5729 | if (sg) { |
| 5730 | *sg = *per_cpu_ptr(sdd->sg, cpu); |
| 5731 | (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); |
| 5732 | atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ |
| 5733 | } |
| 5734 | |
| 5735 | return cpu; |
| 5736 | } |
| 5737 | |
| 5738 | /* |
| 5739 | * build_sched_groups will build a circular linked list of the groups |
| 5740 | * covered by the given span, and will set each group's ->cpumask correctly, |
| 5741 | * and ->cpu_power to 0. |
| 5742 | * |
| 5743 | * Assumes the sched_domain tree is fully constructed |
| 5744 | */ |
| 5745 | static int |
| 5746 | build_sched_groups(struct sched_domain *sd, int cpu) |
| 5747 | { |
| 5748 | struct sched_group *first = NULL, *last = NULL; |
| 5749 | struct sd_data *sdd = sd->private; |
| 5750 | const struct cpumask *span = sched_domain_span(sd); |
| 5751 | struct cpumask *covered; |
| 5752 | int i; |
| 5753 | |
| 5754 | get_group(cpu, sdd, &sd->groups); |
| 5755 | atomic_inc(&sd->groups->ref); |
| 5756 | |
| 5757 | if (cpu != cpumask_first(sched_domain_span(sd))) |
| 5758 | return 0; |
| 5759 | |
| 5760 | lockdep_assert_held(&sched_domains_mutex); |
| 5761 | covered = sched_domains_tmpmask; |
| 5762 | |
| 5763 | cpumask_clear(covered); |
| 5764 | |
| 5765 | for_each_cpu(i, span) { |
| 5766 | struct sched_group *sg; |
| 5767 | int group = get_group(i, sdd, &sg); |
| 5768 | int j; |
| 5769 | |
| 5770 | if (cpumask_test_cpu(i, covered)) |
| 5771 | continue; |
| 5772 | |
| 5773 | cpumask_clear(sched_group_cpus(sg)); |
| 5774 | sg->sgp->power = 0; |
| 5775 | cpumask_setall(sched_group_mask(sg)); |
| 5776 | |
| 5777 | for_each_cpu(j, span) { |
| 5778 | if (get_group(j, sdd, NULL) != group) |
| 5779 | continue; |
| 5780 | |
| 5781 | cpumask_set_cpu(j, covered); |
| 5782 | cpumask_set_cpu(j, sched_group_cpus(sg)); |
| 5783 | } |
| 5784 | |
| 5785 | if (!first) |
| 5786 | first = sg; |
| 5787 | if (last) |
| 5788 | last->next = sg; |
| 5789 | last = sg; |
| 5790 | } |
| 5791 | last->next = first; |
| 5792 | |
| 5793 | return 0; |
| 5794 | } |
| 5795 | |
| 5796 | /* |
| 5797 | * Initialize sched groups cpu_power. |
| 5798 | * |
| 5799 | * cpu_power indicates the capacity of sched group, which is used while |
| 5800 | * distributing the load between different sched groups in a sched domain. |
| 5801 | * Typically cpu_power for all the groups in a sched domain will be same unless |
| 5802 | * there are asymmetries in the topology. If there are asymmetries, group |
| 5803 | * having more cpu_power will pickup more load compared to the group having |
| 5804 | * less cpu_power. |
| 5805 | */ |
| 5806 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) |
| 5807 | { |
| 5808 | struct sched_group *sg = sd->groups; |
| 5809 | |
| 5810 | WARN_ON(!sd || !sg); |
| 5811 | |
| 5812 | do { |
| 5813 | sg->group_weight = cpumask_weight(sched_group_cpus(sg)); |
| 5814 | sg = sg->next; |
| 5815 | } while (sg != sd->groups); |
| 5816 | |
| 5817 | if (cpu != group_balance_cpu(sg)) |
| 5818 | return; |
| 5819 | |
| 5820 | update_group_power(sd, cpu); |
| 5821 | atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); |
| 5822 | } |
| 5823 | |
| 5824 | int __weak arch_sd_sibling_asym_packing(void) |
| 5825 | { |
| 5826 | return 0*SD_ASYM_PACKING; |
| 5827 | } |
| 5828 | |
| 5829 | /* |
| 5830 | * Initializers for schedule domains |
| 5831 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() |
| 5832 | */ |
| 5833 | |
| 5834 | #ifdef CONFIG_SCHED_DEBUG |
| 5835 | # define SD_INIT_NAME(sd, type) sd->name = #type |
| 5836 | #else |
| 5837 | # define SD_INIT_NAME(sd, type) do { } while (0) |
| 5838 | #endif |
| 5839 | |
| 5840 | #define SD_INIT_FUNC(type) \ |
| 5841 | static noinline struct sched_domain * \ |
| 5842 | sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ |
| 5843 | { \ |
| 5844 | struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ |
| 5845 | *sd = SD_##type##_INIT; \ |
| 5846 | SD_INIT_NAME(sd, type); \ |
| 5847 | sd->private = &tl->data; \ |
| 5848 | return sd; \ |
| 5849 | } |
| 5850 | |
| 5851 | SD_INIT_FUNC(CPU) |
| 5852 | #ifdef CONFIG_SCHED_SMT |
| 5853 | SD_INIT_FUNC(SIBLING) |
| 5854 | #endif |
| 5855 | #ifdef CONFIG_SCHED_MC |
| 5856 | SD_INIT_FUNC(MC) |
| 5857 | #endif |
| 5858 | #ifdef CONFIG_SCHED_BOOK |
| 5859 | SD_INIT_FUNC(BOOK) |
| 5860 | #endif |
| 5861 | |
| 5862 | static int default_relax_domain_level = -1; |
| 5863 | int sched_domain_level_max; |
| 5864 | |
| 5865 | static int __init setup_relax_domain_level(char *str) |
| 5866 | { |
| 5867 | if (kstrtoint(str, 0, &default_relax_domain_level)) |
| 5868 | pr_warn("Unable to set relax_domain_level\n"); |
| 5869 | |
| 5870 | return 1; |
| 5871 | } |
| 5872 | __setup("relax_domain_level=", setup_relax_domain_level); |
| 5873 | |
| 5874 | static void set_domain_attribute(struct sched_domain *sd, |
| 5875 | struct sched_domain_attr *attr) |
| 5876 | { |
| 5877 | int request; |
| 5878 | |
| 5879 | if (!attr || attr->relax_domain_level < 0) { |
| 5880 | if (default_relax_domain_level < 0) |
| 5881 | return; |
| 5882 | else |
| 5883 | request = default_relax_domain_level; |
| 5884 | } else |
| 5885 | request = attr->relax_domain_level; |
| 5886 | if (request < sd->level) { |
| 5887 | /* turn off idle balance on this domain */ |
| 5888 | sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); |
| 5889 | } else { |
| 5890 | /* turn on idle balance on this domain */ |
| 5891 | sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); |
| 5892 | } |
| 5893 | } |
| 5894 | |
| 5895 | static void __sdt_free(const struct cpumask *cpu_map); |
| 5896 | static int __sdt_alloc(const struct cpumask *cpu_map); |
| 5897 | |
| 5898 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what, |
| 5899 | const struct cpumask *cpu_map) |
| 5900 | { |
| 5901 | switch (what) { |
| 5902 | case sa_rootdomain: |
| 5903 | if (!atomic_read(&d->rd->refcount)) |
| 5904 | free_rootdomain(&d->rd->rcu); /* fall through */ |
| 5905 | case sa_sd: |
| 5906 | free_percpu(d->sd); /* fall through */ |
| 5907 | case sa_sd_storage: |
| 5908 | __sdt_free(cpu_map); /* fall through */ |
| 5909 | case sa_none: |
| 5910 | break; |
| 5911 | } |
| 5912 | } |
| 5913 | |
| 5914 | static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, |
| 5915 | const struct cpumask *cpu_map) |
| 5916 | { |
| 5917 | memset(d, 0, sizeof(*d)); |
| 5918 | |
| 5919 | if (__sdt_alloc(cpu_map)) |
| 5920 | return sa_sd_storage; |
| 5921 | d->sd = alloc_percpu(struct sched_domain *); |
| 5922 | if (!d->sd) |
| 5923 | return sa_sd_storage; |
| 5924 | d->rd = alloc_rootdomain(); |
| 5925 | if (!d->rd) |
| 5926 | return sa_sd; |
| 5927 | return sa_rootdomain; |
| 5928 | } |
| 5929 | |
| 5930 | /* |
| 5931 | * NULL the sd_data elements we've used to build the sched_domain and |
| 5932 | * sched_group structure so that the subsequent __free_domain_allocs() |
| 5933 | * will not free the data we're using. |
| 5934 | */ |
| 5935 | static void claim_allocations(int cpu, struct sched_domain *sd) |
| 5936 | { |
| 5937 | struct sd_data *sdd = sd->private; |
| 5938 | |
| 5939 | WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); |
| 5940 | *per_cpu_ptr(sdd->sd, cpu) = NULL; |
| 5941 | |
| 5942 | if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) |
| 5943 | *per_cpu_ptr(sdd->sg, cpu) = NULL; |
| 5944 | |
| 5945 | if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) |
| 5946 | *per_cpu_ptr(sdd->sgp, cpu) = NULL; |
| 5947 | } |
| 5948 | |
| 5949 | #ifdef CONFIG_SCHED_SMT |
| 5950 | static const struct cpumask *cpu_smt_mask(int cpu) |
| 5951 | { |
| 5952 | return topology_thread_cpumask(cpu); |
| 5953 | } |
| 5954 | #endif |
| 5955 | |
| 5956 | /* |
| 5957 | * Topology list, bottom-up. |
| 5958 | */ |
| 5959 | static struct sched_domain_topology_level default_topology[] = { |
| 5960 | #ifdef CONFIG_SCHED_SMT |
| 5961 | { sd_init_SIBLING, cpu_smt_mask, }, |
| 5962 | #endif |
| 5963 | #ifdef CONFIG_SCHED_MC |
| 5964 | { sd_init_MC, cpu_coregroup_mask, }, |
| 5965 | #endif |
| 5966 | #ifdef CONFIG_SCHED_BOOK |
| 5967 | { sd_init_BOOK, cpu_book_mask, }, |
| 5968 | #endif |
| 5969 | { sd_init_CPU, cpu_cpu_mask, }, |
| 5970 | { NULL, }, |
| 5971 | }; |
| 5972 | |
| 5973 | static struct sched_domain_topology_level *sched_domain_topology = default_topology; |
| 5974 | |
| 5975 | #ifdef CONFIG_NUMA |
| 5976 | |
| 5977 | static int sched_domains_numa_levels; |
| 5978 | static int *sched_domains_numa_distance; |
| 5979 | static struct cpumask ***sched_domains_numa_masks; |
| 5980 | static int sched_domains_curr_level; |
| 5981 | |
| 5982 | static inline int sd_local_flags(int level) |
| 5983 | { |
| 5984 | if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) |
| 5985 | return 0; |
| 5986 | |
| 5987 | return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; |
| 5988 | } |
| 5989 | |
| 5990 | static struct sched_domain * |
| 5991 | sd_numa_init(struct sched_domain_topology_level *tl, int cpu) |
| 5992 | { |
| 5993 | struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); |
| 5994 | int level = tl->numa_level; |
| 5995 | int sd_weight = cpumask_weight( |
| 5996 | sched_domains_numa_masks[level][cpu_to_node(cpu)]); |
| 5997 | |
| 5998 | *sd = (struct sched_domain){ |
| 5999 | .min_interval = sd_weight, |
| 6000 | .max_interval = 2*sd_weight, |
| 6001 | .busy_factor = 32, |
| 6002 | .imbalance_pct = 125, |
| 6003 | .cache_nice_tries = 2, |
| 6004 | .busy_idx = 3, |
| 6005 | .idle_idx = 2, |
| 6006 | .newidle_idx = 0, |
| 6007 | .wake_idx = 0, |
| 6008 | .forkexec_idx = 0, |
| 6009 | |
| 6010 | .flags = 1*SD_LOAD_BALANCE |
| 6011 | | 1*SD_BALANCE_NEWIDLE |
| 6012 | | 0*SD_BALANCE_EXEC |
| 6013 | | 0*SD_BALANCE_FORK |
| 6014 | | 0*SD_BALANCE_WAKE |
| 6015 | | 0*SD_WAKE_AFFINE |
| 6016 | | 0*SD_SHARE_CPUPOWER |
| 6017 | | 0*SD_SHARE_PKG_RESOURCES |
| 6018 | | 1*SD_SERIALIZE |
| 6019 | | 0*SD_PREFER_SIBLING |
| 6020 | | sd_local_flags(level) |
| 6021 | , |
| 6022 | .last_balance = jiffies, |
| 6023 | .balance_interval = sd_weight, |
| 6024 | }; |
| 6025 | SD_INIT_NAME(sd, NUMA); |
| 6026 | sd->private = &tl->data; |
| 6027 | |
| 6028 | /* |
| 6029 | * Ugly hack to pass state to sd_numa_mask()... |
| 6030 | */ |
| 6031 | sched_domains_curr_level = tl->numa_level; |
| 6032 | |
| 6033 | return sd; |
| 6034 | } |
| 6035 | |
| 6036 | static const struct cpumask *sd_numa_mask(int cpu) |
| 6037 | { |
| 6038 | return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; |
| 6039 | } |
| 6040 | |
| 6041 | static void sched_numa_warn(const char *str) |
| 6042 | { |
| 6043 | static int done = false; |
| 6044 | int i,j; |
| 6045 | |
| 6046 | if (done) |
| 6047 | return; |
| 6048 | |
| 6049 | done = true; |
| 6050 | |
| 6051 | printk(KERN_WARNING "ERROR: %s\n\n", str); |
| 6052 | |
| 6053 | for (i = 0; i < nr_node_ids; i++) { |
| 6054 | printk(KERN_WARNING " "); |
| 6055 | for (j = 0; j < nr_node_ids; j++) |
| 6056 | printk(KERN_CONT "%02d ", node_distance(i,j)); |
| 6057 | printk(KERN_CONT "\n"); |
| 6058 | } |
| 6059 | printk(KERN_WARNING "\n"); |
| 6060 | } |
| 6061 | |
| 6062 | static bool find_numa_distance(int distance) |
| 6063 | { |
| 6064 | int i; |
| 6065 | |
| 6066 | if (distance == node_distance(0, 0)) |
| 6067 | return true; |
| 6068 | |
| 6069 | for (i = 0; i < sched_domains_numa_levels; i++) { |
| 6070 | if (sched_domains_numa_distance[i] == distance) |
| 6071 | return true; |
| 6072 | } |
| 6073 | |
| 6074 | return false; |
| 6075 | } |
| 6076 | |
| 6077 | static void sched_init_numa(void) |
| 6078 | { |
| 6079 | int next_distance, curr_distance = node_distance(0, 0); |
| 6080 | struct sched_domain_topology_level *tl; |
| 6081 | int level = 0; |
| 6082 | int i, j, k; |
| 6083 | |
| 6084 | sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); |
| 6085 | if (!sched_domains_numa_distance) |
| 6086 | return; |
| 6087 | |
| 6088 | /* |
| 6089 | * O(nr_nodes^2) deduplicating selection sort -- in order to find the |
| 6090 | * unique distances in the node_distance() table. |
| 6091 | * |
| 6092 | * Assumes node_distance(0,j) includes all distances in |
| 6093 | * node_distance(i,j) in order to avoid cubic time. |
| 6094 | */ |
| 6095 | next_distance = curr_distance; |
| 6096 | for (i = 0; i < nr_node_ids; i++) { |
| 6097 | for (j = 0; j < nr_node_ids; j++) { |
| 6098 | for (k = 0; k < nr_node_ids; k++) { |
| 6099 | int distance = node_distance(i, k); |
| 6100 | |
| 6101 | if (distance > curr_distance && |
| 6102 | (distance < next_distance || |
| 6103 | next_distance == curr_distance)) |
| 6104 | next_distance = distance; |
| 6105 | |
| 6106 | /* |
| 6107 | * While not a strong assumption it would be nice to know |
| 6108 | * about cases where if node A is connected to B, B is not |
| 6109 | * equally connected to A. |
| 6110 | */ |
| 6111 | if (sched_debug() && node_distance(k, i) != distance) |
| 6112 | sched_numa_warn("Node-distance not symmetric"); |
| 6113 | |
| 6114 | if (sched_debug() && i && !find_numa_distance(distance)) |
| 6115 | sched_numa_warn("Node-0 not representative"); |
| 6116 | } |
| 6117 | if (next_distance != curr_distance) { |
| 6118 | sched_domains_numa_distance[level++] = next_distance; |
| 6119 | sched_domains_numa_levels = level; |
| 6120 | curr_distance = next_distance; |
| 6121 | } else break; |
| 6122 | } |
| 6123 | |
| 6124 | /* |
| 6125 | * In case of sched_debug() we verify the above assumption. |
| 6126 | */ |
| 6127 | if (!sched_debug()) |
| 6128 | break; |
| 6129 | } |
| 6130 | /* |
| 6131 | * 'level' contains the number of unique distances, excluding the |
| 6132 | * identity distance node_distance(i,i). |
| 6133 | * |
| 6134 | * The sched_domains_nume_distance[] array includes the actual distance |
| 6135 | * numbers. |
| 6136 | */ |
| 6137 | |
| 6138 | /* |
| 6139 | * Here, we should temporarily reset sched_domains_numa_levels to 0. |
| 6140 | * If it fails to allocate memory for array sched_domains_numa_masks[][], |
| 6141 | * the array will contain less then 'level' members. This could be |
| 6142 | * dangerous when we use it to iterate array sched_domains_numa_masks[][] |
| 6143 | * in other functions. |
| 6144 | * |
| 6145 | * We reset it to 'level' at the end of this function. |
| 6146 | */ |
| 6147 | sched_domains_numa_levels = 0; |
| 6148 | |
| 6149 | sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); |
| 6150 | if (!sched_domains_numa_masks) |
| 6151 | return; |
| 6152 | |
| 6153 | /* |
| 6154 | * Now for each level, construct a mask per node which contains all |
| 6155 | * cpus of nodes that are that many hops away from us. |
| 6156 | */ |
| 6157 | for (i = 0; i < level; i++) { |
| 6158 | sched_domains_numa_masks[i] = |
| 6159 | kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); |
| 6160 | if (!sched_domains_numa_masks[i]) |
| 6161 | return; |
| 6162 | |
| 6163 | for (j = 0; j < nr_node_ids; j++) { |
| 6164 | struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); |
| 6165 | if (!mask) |
| 6166 | return; |
| 6167 | |
| 6168 | sched_domains_numa_masks[i][j] = mask; |
| 6169 | |
| 6170 | for (k = 0; k < nr_node_ids; k++) { |
| 6171 | if (node_distance(j, k) > sched_domains_numa_distance[i]) |
| 6172 | continue; |
| 6173 | |
| 6174 | cpumask_or(mask, mask, cpumask_of_node(k)); |
| 6175 | } |
| 6176 | } |
| 6177 | } |
| 6178 | |
| 6179 | tl = kzalloc((ARRAY_SIZE(default_topology) + level) * |
| 6180 | sizeof(struct sched_domain_topology_level), GFP_KERNEL); |
| 6181 | if (!tl) |
| 6182 | return; |
| 6183 | |
| 6184 | /* |
| 6185 | * Copy the default topology bits.. |
| 6186 | */ |
| 6187 | for (i = 0; default_topology[i].init; i++) |
| 6188 | tl[i] = default_topology[i]; |
| 6189 | |
| 6190 | /* |
| 6191 | * .. and append 'j' levels of NUMA goodness. |
| 6192 | */ |
| 6193 | for (j = 0; j < level; i++, j++) { |
| 6194 | tl[i] = (struct sched_domain_topology_level){ |
| 6195 | .init = sd_numa_init, |
| 6196 | .mask = sd_numa_mask, |
| 6197 | .flags = SDTL_OVERLAP, |
| 6198 | .numa_level = j, |
| 6199 | }; |
| 6200 | } |
| 6201 | |
| 6202 | sched_domain_topology = tl; |
| 6203 | |
| 6204 | sched_domains_numa_levels = level; |
| 6205 | } |
| 6206 | |
| 6207 | static void sched_domains_numa_masks_set(int cpu) |
| 6208 | { |
| 6209 | int i, j; |
| 6210 | int node = cpu_to_node(cpu); |
| 6211 | |
| 6212 | for (i = 0; i < sched_domains_numa_levels; i++) { |
| 6213 | for (j = 0; j < nr_node_ids; j++) { |
| 6214 | if (node_distance(j, node) <= sched_domains_numa_distance[i]) |
| 6215 | cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); |
| 6216 | } |
| 6217 | } |
| 6218 | } |
| 6219 | |
| 6220 | static void sched_domains_numa_masks_clear(int cpu) |
| 6221 | { |
| 6222 | int i, j; |
| 6223 | for (i = 0; i < sched_domains_numa_levels; i++) { |
| 6224 | for (j = 0; j < nr_node_ids; j++) |
| 6225 | cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); |
| 6226 | } |
| 6227 | } |
| 6228 | |
| 6229 | /* |
| 6230 | * Update sched_domains_numa_masks[level][node] array when new cpus |
| 6231 | * are onlined. |
| 6232 | */ |
| 6233 | static int sched_domains_numa_masks_update(struct notifier_block *nfb, |
| 6234 | unsigned long action, |
| 6235 | void *hcpu) |
| 6236 | { |
| 6237 | int cpu = (long)hcpu; |
| 6238 | |
| 6239 | switch (action & ~CPU_TASKS_FROZEN) { |
| 6240 | case CPU_ONLINE: |
| 6241 | sched_domains_numa_masks_set(cpu); |
| 6242 | break; |
| 6243 | |
| 6244 | case CPU_DEAD: |
| 6245 | sched_domains_numa_masks_clear(cpu); |
| 6246 | break; |
| 6247 | |
| 6248 | default: |
| 6249 | return NOTIFY_DONE; |
| 6250 | } |
| 6251 | |
| 6252 | return NOTIFY_OK; |
| 6253 | } |
| 6254 | #else |
| 6255 | static inline void sched_init_numa(void) |
| 6256 | { |
| 6257 | } |
| 6258 | |
| 6259 | static int sched_domains_numa_masks_update(struct notifier_block *nfb, |
| 6260 | unsigned long action, |
| 6261 | void *hcpu) |
| 6262 | { |
| 6263 | return 0; |
| 6264 | } |
| 6265 | #endif /* CONFIG_NUMA */ |
| 6266 | |
| 6267 | static int __sdt_alloc(const struct cpumask *cpu_map) |
| 6268 | { |
| 6269 | struct sched_domain_topology_level *tl; |
| 6270 | int j; |
| 6271 | |
| 6272 | for (tl = sched_domain_topology; tl->init; tl++) { |
| 6273 | struct sd_data *sdd = &tl->data; |
| 6274 | |
| 6275 | sdd->sd = alloc_percpu(struct sched_domain *); |
| 6276 | if (!sdd->sd) |
| 6277 | return -ENOMEM; |
| 6278 | |
| 6279 | sdd->sg = alloc_percpu(struct sched_group *); |
| 6280 | if (!sdd->sg) |
| 6281 | return -ENOMEM; |
| 6282 | |
| 6283 | sdd->sgp = alloc_percpu(struct sched_group_power *); |
| 6284 | if (!sdd->sgp) |
| 6285 | return -ENOMEM; |
| 6286 | |
| 6287 | for_each_cpu(j, cpu_map) { |
| 6288 | struct sched_domain *sd; |
| 6289 | struct sched_group *sg; |
| 6290 | struct sched_group_power *sgp; |
| 6291 | |
| 6292 | sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), |
| 6293 | GFP_KERNEL, cpu_to_node(j)); |
| 6294 | if (!sd) |
| 6295 | return -ENOMEM; |
| 6296 | |
| 6297 | *per_cpu_ptr(sdd->sd, j) = sd; |
| 6298 | |
| 6299 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), |
| 6300 | GFP_KERNEL, cpu_to_node(j)); |
| 6301 | if (!sg) |
| 6302 | return -ENOMEM; |
| 6303 | |
| 6304 | sg->next = sg; |
| 6305 | |
| 6306 | *per_cpu_ptr(sdd->sg, j) = sg; |
| 6307 | |
| 6308 | sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), |
| 6309 | GFP_KERNEL, cpu_to_node(j)); |
| 6310 | if (!sgp) |
| 6311 | return -ENOMEM; |
| 6312 | |
| 6313 | *per_cpu_ptr(sdd->sgp, j) = sgp; |
| 6314 | } |
| 6315 | } |
| 6316 | |
| 6317 | return 0; |
| 6318 | } |
| 6319 | |
| 6320 | static void __sdt_free(const struct cpumask *cpu_map) |
| 6321 | { |
| 6322 | struct sched_domain_topology_level *tl; |
| 6323 | int j; |
| 6324 | |
| 6325 | for (tl = sched_domain_topology; tl->init; tl++) { |
| 6326 | struct sd_data *sdd = &tl->data; |
| 6327 | |
| 6328 | for_each_cpu(j, cpu_map) { |
| 6329 | struct sched_domain *sd; |
| 6330 | |
| 6331 | if (sdd->sd) { |
| 6332 | sd = *per_cpu_ptr(sdd->sd, j); |
| 6333 | if (sd && (sd->flags & SD_OVERLAP)) |
| 6334 | free_sched_groups(sd->groups, 0); |
| 6335 | kfree(*per_cpu_ptr(sdd->sd, j)); |
| 6336 | } |
| 6337 | |
| 6338 | if (sdd->sg) |
| 6339 | kfree(*per_cpu_ptr(sdd->sg, j)); |
| 6340 | if (sdd->sgp) |
| 6341 | kfree(*per_cpu_ptr(sdd->sgp, j)); |
| 6342 | } |
| 6343 | free_percpu(sdd->sd); |
| 6344 | sdd->sd = NULL; |
| 6345 | free_percpu(sdd->sg); |
| 6346 | sdd->sg = NULL; |
| 6347 | free_percpu(sdd->sgp); |
| 6348 | sdd->sgp = NULL; |
| 6349 | } |
| 6350 | } |
| 6351 | |
| 6352 | struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, |
| 6353 | struct s_data *d, const struct cpumask *cpu_map, |
| 6354 | struct sched_domain_attr *attr, struct sched_domain *child, |
| 6355 | int cpu) |
| 6356 | { |
| 6357 | struct sched_domain *sd = tl->init(tl, cpu); |
| 6358 | if (!sd) |
| 6359 | return child; |
| 6360 | |
| 6361 | cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); |
| 6362 | if (child) { |
| 6363 | sd->level = child->level + 1; |
| 6364 | sched_domain_level_max = max(sched_domain_level_max, sd->level); |
| 6365 | child->parent = sd; |
| 6366 | } |
| 6367 | sd->child = child; |
| 6368 | set_domain_attribute(sd, attr); |
| 6369 | |
| 6370 | return sd; |
| 6371 | } |
| 6372 | |
| 6373 | /* |
| 6374 | * Build sched domains for a given set of cpus and attach the sched domains |
| 6375 | * to the individual cpus |
| 6376 | */ |
| 6377 | static int build_sched_domains(const struct cpumask *cpu_map, |
| 6378 | struct sched_domain_attr *attr) |
| 6379 | { |
| 6380 | enum s_alloc alloc_state = sa_none; |
| 6381 | struct sched_domain *sd; |
| 6382 | struct s_data d; |
| 6383 | int i, ret = -ENOMEM; |
| 6384 | |
| 6385 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map); |
| 6386 | if (alloc_state != sa_rootdomain) |
| 6387 | goto error; |
| 6388 | |
| 6389 | /* Set up domains for cpus specified by the cpu_map. */ |
| 6390 | for_each_cpu(i, cpu_map) { |
| 6391 | struct sched_domain_topology_level *tl; |
| 6392 | |
| 6393 | sd = NULL; |
| 6394 | for (tl = sched_domain_topology; tl->init; tl++) { |
| 6395 | sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); |
| 6396 | if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) |
| 6397 | sd->flags |= SD_OVERLAP; |
| 6398 | if (cpumask_equal(cpu_map, sched_domain_span(sd))) |
| 6399 | break; |
| 6400 | } |
| 6401 | |
| 6402 | while (sd->child) |
| 6403 | sd = sd->child; |
| 6404 | |
| 6405 | *per_cpu_ptr(d.sd, i) = sd; |
| 6406 | } |
| 6407 | |
| 6408 | /* Build the groups for the domains */ |
| 6409 | for_each_cpu(i, cpu_map) { |
| 6410 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { |
| 6411 | sd->span_weight = cpumask_weight(sched_domain_span(sd)); |
| 6412 | if (sd->flags & SD_OVERLAP) { |
| 6413 | if (build_overlap_sched_groups(sd, i)) |
| 6414 | goto error; |
| 6415 | } else { |
| 6416 | if (build_sched_groups(sd, i)) |
| 6417 | goto error; |
| 6418 | } |
| 6419 | } |
| 6420 | } |
| 6421 | |
| 6422 | /* Calculate CPU power for physical packages and nodes */ |
| 6423 | for (i = nr_cpumask_bits-1; i >= 0; i--) { |
| 6424 | if (!cpumask_test_cpu(i, cpu_map)) |
| 6425 | continue; |
| 6426 | |
| 6427 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { |
| 6428 | claim_allocations(i, sd); |
| 6429 | init_sched_groups_power(i, sd); |
| 6430 | } |
| 6431 | } |
| 6432 | |
| 6433 | /* Attach the domains */ |
| 6434 | rcu_read_lock(); |
| 6435 | for_each_cpu(i, cpu_map) { |
| 6436 | sd = *per_cpu_ptr(d.sd, i); |
| 6437 | cpu_attach_domain(sd, d.rd, i); |
| 6438 | } |
| 6439 | rcu_read_unlock(); |
| 6440 | |
| 6441 | ret = 0; |
| 6442 | error: |
| 6443 | __free_domain_allocs(&d, alloc_state, cpu_map); |
| 6444 | return ret; |
| 6445 | } |
| 6446 | |
| 6447 | static cpumask_var_t *doms_cur; /* current sched domains */ |
| 6448 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ |
| 6449 | static struct sched_domain_attr *dattr_cur; |
| 6450 | /* attribues of custom domains in 'doms_cur' */ |
| 6451 | |
| 6452 | /* |
| 6453 | * Special case: If a kmalloc of a doms_cur partition (array of |
| 6454 | * cpumask) fails, then fallback to a single sched domain, |
| 6455 | * as determined by the single cpumask fallback_doms. |
| 6456 | */ |
| 6457 | static cpumask_var_t fallback_doms; |
| 6458 | |
| 6459 | /* |
| 6460 | * arch_update_cpu_topology lets virtualized architectures update the |
| 6461 | * cpu core maps. It is supposed to return 1 if the topology changed |
| 6462 | * or 0 if it stayed the same. |
| 6463 | */ |
| 6464 | int __attribute__((weak)) arch_update_cpu_topology(void) |
| 6465 | { |
| 6466 | return 0; |
| 6467 | } |
| 6468 | |
| 6469 | cpumask_var_t *alloc_sched_domains(unsigned int ndoms) |
| 6470 | { |
| 6471 | int i; |
| 6472 | cpumask_var_t *doms; |
| 6473 | |
| 6474 | doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); |
| 6475 | if (!doms) |
| 6476 | return NULL; |
| 6477 | for (i = 0; i < ndoms; i++) { |
| 6478 | if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { |
| 6479 | free_sched_domains(doms, i); |
| 6480 | return NULL; |
| 6481 | } |
| 6482 | } |
| 6483 | return doms; |
| 6484 | } |
| 6485 | |
| 6486 | void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) |
| 6487 | { |
| 6488 | unsigned int i; |
| 6489 | for (i = 0; i < ndoms; i++) |
| 6490 | free_cpumask_var(doms[i]); |
| 6491 | kfree(doms); |
| 6492 | } |
| 6493 | |
| 6494 | /* |
| 6495 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
| 6496 | * For now this just excludes isolated cpus, but could be used to |
| 6497 | * exclude other special cases in the future. |
| 6498 | */ |
| 6499 | static int init_sched_domains(const struct cpumask *cpu_map) |
| 6500 | { |
| 6501 | int err; |
| 6502 | |
| 6503 | arch_update_cpu_topology(); |
| 6504 | ndoms_cur = 1; |
| 6505 | doms_cur = alloc_sched_domains(ndoms_cur); |
| 6506 | if (!doms_cur) |
| 6507 | doms_cur = &fallback_doms; |
| 6508 | cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); |
| 6509 | err = build_sched_domains(doms_cur[0], NULL); |
| 6510 | register_sched_domain_sysctl(); |
| 6511 | |
| 6512 | return err; |
| 6513 | } |
| 6514 | |
| 6515 | /* |
| 6516 | * Detach sched domains from a group of cpus specified in cpu_map |
| 6517 | * These cpus will now be attached to the NULL domain |
| 6518 | */ |
| 6519 | static void detach_destroy_domains(const struct cpumask *cpu_map) |
| 6520 | { |
| 6521 | int i; |
| 6522 | |
| 6523 | rcu_read_lock(); |
| 6524 | for_each_cpu(i, cpu_map) |
| 6525 | cpu_attach_domain(NULL, &def_root_domain, i); |
| 6526 | rcu_read_unlock(); |
| 6527 | } |
| 6528 | |
| 6529 | /* handle null as "default" */ |
| 6530 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, |
| 6531 | struct sched_domain_attr *new, int idx_new) |
| 6532 | { |
| 6533 | struct sched_domain_attr tmp; |
| 6534 | |
| 6535 | /* fast path */ |
| 6536 | if (!new && !cur) |
| 6537 | return 1; |
| 6538 | |
| 6539 | tmp = SD_ATTR_INIT; |
| 6540 | return !memcmp(cur ? (cur + idx_cur) : &tmp, |
| 6541 | new ? (new + idx_new) : &tmp, |
| 6542 | sizeof(struct sched_domain_attr)); |
| 6543 | } |
| 6544 | |
| 6545 | /* |
| 6546 | * Partition sched domains as specified by the 'ndoms_new' |
| 6547 | * cpumasks in the array doms_new[] of cpumasks. This compares |
| 6548 | * doms_new[] to the current sched domain partitioning, doms_cur[]. |
| 6549 | * It destroys each deleted domain and builds each new domain. |
| 6550 | * |
| 6551 | * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. |
| 6552 | * The masks don't intersect (don't overlap.) We should setup one |
| 6553 | * sched domain for each mask. CPUs not in any of the cpumasks will |
| 6554 | * not be load balanced. If the same cpumask appears both in the |
| 6555 | * current 'doms_cur' domains and in the new 'doms_new', we can leave |
| 6556 | * it as it is. |
| 6557 | * |
| 6558 | * The passed in 'doms_new' should be allocated using |
| 6559 | * alloc_sched_domains. This routine takes ownership of it and will |
| 6560 | * free_sched_domains it when done with it. If the caller failed the |
| 6561 | * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, |
| 6562 | * and partition_sched_domains() will fallback to the single partition |
| 6563 | * 'fallback_doms', it also forces the domains to be rebuilt. |
| 6564 | * |
| 6565 | * If doms_new == NULL it will be replaced with cpu_online_mask. |
| 6566 | * ndoms_new == 0 is a special case for destroying existing domains, |
| 6567 | * and it will not create the default domain. |
| 6568 | * |
| 6569 | * Call with hotplug lock held |
| 6570 | */ |
| 6571 | void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], |
| 6572 | struct sched_domain_attr *dattr_new) |
| 6573 | { |
| 6574 | int i, j, n; |
| 6575 | int new_topology; |
| 6576 | |
| 6577 | mutex_lock(&sched_domains_mutex); |
| 6578 | |
| 6579 | /* always unregister in case we don't destroy any domains */ |
| 6580 | unregister_sched_domain_sysctl(); |
| 6581 | |
| 6582 | /* Let architecture update cpu core mappings. */ |
| 6583 | new_topology = arch_update_cpu_topology(); |
| 6584 | |
| 6585 | n = doms_new ? ndoms_new : 0; |
| 6586 | |
| 6587 | /* Destroy deleted domains */ |
| 6588 | for (i = 0; i < ndoms_cur; i++) { |
| 6589 | for (j = 0; j < n && !new_topology; j++) { |
| 6590 | if (cpumask_equal(doms_cur[i], doms_new[j]) |
| 6591 | && dattrs_equal(dattr_cur, i, dattr_new, j)) |
| 6592 | goto match1; |
| 6593 | } |
| 6594 | /* no match - a current sched domain not in new doms_new[] */ |
| 6595 | detach_destroy_domains(doms_cur[i]); |
| 6596 | match1: |
| 6597 | ; |
| 6598 | } |
| 6599 | |
| 6600 | if (doms_new == NULL) { |
| 6601 | ndoms_cur = 0; |
| 6602 | doms_new = &fallback_doms; |
| 6603 | cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); |
| 6604 | WARN_ON_ONCE(dattr_new); |
| 6605 | } |
| 6606 | |
| 6607 | /* Build new domains */ |
| 6608 | for (i = 0; i < ndoms_new; i++) { |
| 6609 | for (j = 0; j < ndoms_cur && !new_topology; j++) { |
| 6610 | if (cpumask_equal(doms_new[i], doms_cur[j]) |
| 6611 | && dattrs_equal(dattr_new, i, dattr_cur, j)) |
| 6612 | goto match2; |
| 6613 | } |
| 6614 | /* no match - add a new doms_new */ |
| 6615 | build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); |
| 6616 | match2: |
| 6617 | ; |
| 6618 | } |
| 6619 | |
| 6620 | /* Remember the new sched domains */ |
| 6621 | if (doms_cur != &fallback_doms) |
| 6622 | free_sched_domains(doms_cur, ndoms_cur); |
| 6623 | kfree(dattr_cur); /* kfree(NULL) is safe */ |
| 6624 | doms_cur = doms_new; |
| 6625 | dattr_cur = dattr_new; |
| 6626 | ndoms_cur = ndoms_new; |
| 6627 | |
| 6628 | register_sched_domain_sysctl(); |
| 6629 | |
| 6630 | mutex_unlock(&sched_domains_mutex); |
| 6631 | } |
| 6632 | |
| 6633 | static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ |
| 6634 | |
| 6635 | /* |
| 6636 | * Update cpusets according to cpu_active mask. If cpusets are |
| 6637 | * disabled, cpuset_update_active_cpus() becomes a simple wrapper |
| 6638 | * around partition_sched_domains(). |
| 6639 | * |
| 6640 | * If we come here as part of a suspend/resume, don't touch cpusets because we |
| 6641 | * want to restore it back to its original state upon resume anyway. |
| 6642 | */ |
| 6643 | static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, |
| 6644 | void *hcpu) |
| 6645 | { |
| 6646 | switch (action) { |
| 6647 | case CPU_ONLINE_FROZEN: |
| 6648 | case CPU_DOWN_FAILED_FROZEN: |
| 6649 | |
| 6650 | /* |
| 6651 | * num_cpus_frozen tracks how many CPUs are involved in suspend |
| 6652 | * resume sequence. As long as this is not the last online |
| 6653 | * operation in the resume sequence, just build a single sched |
| 6654 | * domain, ignoring cpusets. |
| 6655 | */ |
| 6656 | num_cpus_frozen--; |
| 6657 | if (likely(num_cpus_frozen)) { |
| 6658 | partition_sched_domains(1, NULL, NULL); |
| 6659 | break; |
| 6660 | } |
| 6661 | |
| 6662 | /* |
| 6663 | * This is the last CPU online operation. So fall through and |
| 6664 | * restore the original sched domains by considering the |
| 6665 | * cpuset configurations. |
| 6666 | */ |
| 6667 | |
| 6668 | case CPU_ONLINE: |
| 6669 | case CPU_DOWN_FAILED: |
| 6670 | cpuset_update_active_cpus(true); |
| 6671 | break; |
| 6672 | default: |
| 6673 | return NOTIFY_DONE; |
| 6674 | } |
| 6675 | return NOTIFY_OK; |
| 6676 | } |
| 6677 | |
| 6678 | static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, |
| 6679 | void *hcpu) |
| 6680 | { |
| 6681 | switch (action) { |
| 6682 | case CPU_DOWN_PREPARE: |
| 6683 | cpuset_update_active_cpus(false); |
| 6684 | break; |
| 6685 | case CPU_DOWN_PREPARE_FROZEN: |
| 6686 | num_cpus_frozen++; |
| 6687 | partition_sched_domains(1, NULL, NULL); |
| 6688 | break; |
| 6689 | default: |
| 6690 | return NOTIFY_DONE; |
| 6691 | } |
| 6692 | return NOTIFY_OK; |
| 6693 | } |
| 6694 | |
| 6695 | void __init sched_init_smp(void) |
| 6696 | { |
| 6697 | cpumask_var_t non_isolated_cpus; |
| 6698 | |
| 6699 | alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); |
| 6700 | alloc_cpumask_var(&fallback_doms, GFP_KERNEL); |
| 6701 | |
| 6702 | sched_init_numa(); |
| 6703 | |
| 6704 | get_online_cpus(); |
| 6705 | mutex_lock(&sched_domains_mutex); |
| 6706 | init_sched_domains(cpu_active_mask); |
| 6707 | cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); |
| 6708 | if (cpumask_empty(non_isolated_cpus)) |
| 6709 | cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); |
| 6710 | mutex_unlock(&sched_domains_mutex); |
| 6711 | put_online_cpus(); |
| 6712 | |
| 6713 | hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); |
| 6714 | hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); |
| 6715 | hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); |
| 6716 | |
| 6717 | /* RT runtime code needs to handle some hotplug events */ |
| 6718 | hotcpu_notifier(update_runtime, 0); |
| 6719 | |
| 6720 | init_hrtick(); |
| 6721 | |
| 6722 | /* Move init over to a non-isolated CPU */ |
| 6723 | if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) |
| 6724 | BUG(); |
| 6725 | sched_init_granularity(); |
| 6726 | free_cpumask_var(non_isolated_cpus); |
| 6727 | |
| 6728 | init_sched_rt_class(); |
| 6729 | } |
| 6730 | #else |
| 6731 | void __init sched_init_smp(void) |
| 6732 | { |
| 6733 | sched_init_granularity(); |
| 6734 | } |
| 6735 | #endif /* CONFIG_SMP */ |
| 6736 | |
| 6737 | const_debug unsigned int sysctl_timer_migration = 1; |
| 6738 | |
| 6739 | int in_sched_functions(unsigned long addr) |
| 6740 | { |
| 6741 | return in_lock_functions(addr) || |
| 6742 | (addr >= (unsigned long)__sched_text_start |
| 6743 | && addr < (unsigned long)__sched_text_end); |
| 6744 | } |
| 6745 | |
| 6746 | #ifdef CONFIG_CGROUP_SCHED |
| 6747 | struct task_group root_task_group; |
| 6748 | LIST_HEAD(task_groups); |
| 6749 | #endif |
| 6750 | |
| 6751 | DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); |
| 6752 | |
| 6753 | void __init sched_init(void) |
| 6754 | { |
| 6755 | int i, j; |
| 6756 | unsigned long alloc_size = 0, ptr; |
| 6757 | |
| 6758 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 6759 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); |
| 6760 | #endif |
| 6761 | #ifdef CONFIG_RT_GROUP_SCHED |
| 6762 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); |
| 6763 | #endif |
| 6764 | #ifdef CONFIG_CPUMASK_OFFSTACK |
| 6765 | alloc_size += num_possible_cpus() * cpumask_size(); |
| 6766 | #endif |
| 6767 | if (alloc_size) { |
| 6768 | ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); |
| 6769 | |
| 6770 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 6771 | root_task_group.se = (struct sched_entity **)ptr; |
| 6772 | ptr += nr_cpu_ids * sizeof(void **); |
| 6773 | |
| 6774 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; |
| 6775 | ptr += nr_cpu_ids * sizeof(void **); |
| 6776 | |
| 6777 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 6778 | #ifdef CONFIG_RT_GROUP_SCHED |
| 6779 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; |
| 6780 | ptr += nr_cpu_ids * sizeof(void **); |
| 6781 | |
| 6782 | root_task_group.rt_rq = (struct rt_rq **)ptr; |
| 6783 | ptr += nr_cpu_ids * sizeof(void **); |
| 6784 | |
| 6785 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 6786 | #ifdef CONFIG_CPUMASK_OFFSTACK |
| 6787 | for_each_possible_cpu(i) { |
| 6788 | per_cpu(load_balance_tmpmask, i) = (void *)ptr; |
| 6789 | ptr += cpumask_size(); |
| 6790 | } |
| 6791 | #endif /* CONFIG_CPUMASK_OFFSTACK */ |
| 6792 | } |
| 6793 | |
| 6794 | #ifdef CONFIG_SMP |
| 6795 | init_defrootdomain(); |
| 6796 | #endif |
| 6797 | |
| 6798 | init_rt_bandwidth(&def_rt_bandwidth, |
| 6799 | global_rt_period(), global_rt_runtime()); |
| 6800 | |
| 6801 | #ifdef CONFIG_RT_GROUP_SCHED |
| 6802 | init_rt_bandwidth(&root_task_group.rt_bandwidth, |
| 6803 | global_rt_period(), global_rt_runtime()); |
| 6804 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 6805 | |
| 6806 | #ifdef CONFIG_CGROUP_SCHED |
| 6807 | list_add(&root_task_group.list, &task_groups); |
| 6808 | INIT_LIST_HEAD(&root_task_group.children); |
| 6809 | INIT_LIST_HEAD(&root_task_group.siblings); |
| 6810 | autogroup_init(&init_task); |
| 6811 | |
| 6812 | #endif /* CONFIG_CGROUP_SCHED */ |
| 6813 | |
| 6814 | #ifdef CONFIG_CGROUP_CPUACCT |
| 6815 | root_cpuacct.cpustat = &kernel_cpustat; |
| 6816 | root_cpuacct.cpuusage = alloc_percpu(u64); |
| 6817 | /* Too early, not expected to fail */ |
| 6818 | BUG_ON(!root_cpuacct.cpuusage); |
| 6819 | #endif |
| 6820 | for_each_possible_cpu(i) { |
| 6821 | struct rq *rq; |
| 6822 | |
| 6823 | rq = cpu_rq(i); |
| 6824 | raw_spin_lock_init(&rq->lock); |
| 6825 | rq->nr_running = 0; |
| 6826 | rq->calc_load_active = 0; |
| 6827 | rq->calc_load_update = jiffies + LOAD_FREQ; |
| 6828 | init_cfs_rq(&rq->cfs); |
| 6829 | init_rt_rq(&rq->rt, rq); |
| 6830 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 6831 | root_task_group.shares = ROOT_TASK_GROUP_LOAD; |
| 6832 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
| 6833 | /* |
| 6834 | * How much cpu bandwidth does root_task_group get? |
| 6835 | * |
| 6836 | * In case of task-groups formed thr' the cgroup filesystem, it |
| 6837 | * gets 100% of the cpu resources in the system. This overall |
| 6838 | * system cpu resource is divided among the tasks of |
| 6839 | * root_task_group and its child task-groups in a fair manner, |
| 6840 | * based on each entity's (task or task-group's) weight |
| 6841 | * (se->load.weight). |
| 6842 | * |
| 6843 | * In other words, if root_task_group has 10 tasks of weight |
| 6844 | * 1024) and two child groups A0 and A1 (of weight 1024 each), |
| 6845 | * then A0's share of the cpu resource is: |
| 6846 | * |
| 6847 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% |
| 6848 | * |
| 6849 | * We achieve this by letting root_task_group's tasks sit |
| 6850 | * directly in rq->cfs (i.e root_task_group->se[] = NULL). |
| 6851 | */ |
| 6852 | init_cfs_bandwidth(&root_task_group.cfs_bandwidth); |
| 6853 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); |
| 6854 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 6855 | |
| 6856 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; |
| 6857 | #ifdef CONFIG_RT_GROUP_SCHED |
| 6858 | INIT_LIST_HEAD(&rq->leaf_rt_rq_list); |
| 6859 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); |
| 6860 | #endif |
| 6861 | |
| 6862 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) |
| 6863 | rq->cpu_load[j] = 0; |
| 6864 | |
| 6865 | rq->last_load_update_tick = jiffies; |
| 6866 | |
| 6867 | #ifdef CONFIG_SMP |
| 6868 | rq->sd = NULL; |
| 6869 | rq->rd = NULL; |
| 6870 | rq->cpu_power = SCHED_POWER_SCALE; |
| 6871 | rq->post_schedule = 0; |
| 6872 | rq->active_balance = 0; |
| 6873 | rq->next_balance = jiffies; |
| 6874 | rq->push_cpu = 0; |
| 6875 | rq->cpu = i; |
| 6876 | rq->online = 0; |
| 6877 | rq->idle_stamp = 0; |
| 6878 | rq->avg_idle = 2*sysctl_sched_migration_cost; |
| 6879 | |
| 6880 | INIT_LIST_HEAD(&rq->cfs_tasks); |
| 6881 | |
| 6882 | rq_attach_root(rq, &def_root_domain); |
| 6883 | #ifdef CONFIG_NO_HZ |
| 6884 | rq->nohz_flags = 0; |
| 6885 | #endif |
| 6886 | #endif |
| 6887 | init_rq_hrtick(rq); |
| 6888 | atomic_set(&rq->nr_iowait, 0); |
| 6889 | } |
| 6890 | |
| 6891 | set_load_weight(&init_task); |
| 6892 | |
| 6893 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 6894 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); |
| 6895 | #endif |
| 6896 | |
| 6897 | #ifdef CONFIG_RT_MUTEXES |
| 6898 | plist_head_init(&init_task.pi_waiters); |
| 6899 | #endif |
| 6900 | |
| 6901 | /* |
| 6902 | * The boot idle thread does lazy MMU switching as well: |
| 6903 | */ |
| 6904 | atomic_inc(&init_mm.mm_count); |
| 6905 | enter_lazy_tlb(&init_mm, current); |
| 6906 | |
| 6907 | /* |
| 6908 | * Make us the idle thread. Technically, schedule() should not be |
| 6909 | * called from this thread, however somewhere below it might be, |
| 6910 | * but because we are the idle thread, we just pick up running again |
| 6911 | * when this runqueue becomes "idle". |
| 6912 | */ |
| 6913 | init_idle(current, smp_processor_id()); |
| 6914 | |
| 6915 | calc_load_update = jiffies + LOAD_FREQ; |
| 6916 | |
| 6917 | /* |
| 6918 | * During early bootup we pretend to be a normal task: |
| 6919 | */ |
| 6920 | current->sched_class = &fair_sched_class; |
| 6921 | |
| 6922 | #ifdef CONFIG_SMP |
| 6923 | zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); |
| 6924 | /* May be allocated at isolcpus cmdline parse time */ |
| 6925 | if (cpu_isolated_map == NULL) |
| 6926 | zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); |
| 6927 | idle_thread_set_boot_cpu(); |
| 6928 | #endif |
| 6929 | init_sched_fair_class(); |
| 6930 | |
| 6931 | scheduler_running = 1; |
| 6932 | } |
| 6933 | |
| 6934 | #ifdef CONFIG_DEBUG_ATOMIC_SLEEP |
| 6935 | static inline int preempt_count_equals(int preempt_offset) |
| 6936 | { |
| 6937 | int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); |
| 6938 | |
| 6939 | return (nested == preempt_offset); |
| 6940 | } |
| 6941 | |
| 6942 | void __might_sleep(const char *file, int line, int preempt_offset) |
| 6943 | { |
| 6944 | static unsigned long prev_jiffy; /* ratelimiting */ |
| 6945 | |
| 6946 | rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ |
| 6947 | if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || |
| 6948 | system_state != SYSTEM_RUNNING || oops_in_progress) |
| 6949 | return; |
| 6950 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
| 6951 | return; |
| 6952 | prev_jiffy = jiffies; |
| 6953 | |
| 6954 | printk(KERN_ERR |
| 6955 | "BUG: sleeping function called from invalid context at %s:%d\n", |
| 6956 | file, line); |
| 6957 | printk(KERN_ERR |
| 6958 | "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", |
| 6959 | in_atomic(), irqs_disabled(), |
| 6960 | current->pid, current->comm); |
| 6961 | |
| 6962 | debug_show_held_locks(current); |
| 6963 | if (irqs_disabled()) |
| 6964 | print_irqtrace_events(current); |
| 6965 | dump_stack(); |
| 6966 | } |
| 6967 | EXPORT_SYMBOL(__might_sleep); |
| 6968 | #endif |
| 6969 | |
| 6970 | #ifdef CONFIG_MAGIC_SYSRQ |
| 6971 | static void normalize_task(struct rq *rq, struct task_struct *p) |
| 6972 | { |
| 6973 | const struct sched_class *prev_class = p->sched_class; |
| 6974 | int old_prio = p->prio; |
| 6975 | int on_rq; |
| 6976 | |
| 6977 | on_rq = p->on_rq; |
| 6978 | if (on_rq) |
| 6979 | dequeue_task(rq, p, 0); |
| 6980 | __setscheduler(rq, p, SCHED_NORMAL, 0); |
| 6981 | if (on_rq) { |
| 6982 | enqueue_task(rq, p, 0); |
| 6983 | resched_task(rq->curr); |
| 6984 | } |
| 6985 | |
| 6986 | check_class_changed(rq, p, prev_class, old_prio); |
| 6987 | } |
| 6988 | |
| 6989 | void normalize_rt_tasks(void) |
| 6990 | { |
| 6991 | struct task_struct *g, *p; |
| 6992 | unsigned long flags; |
| 6993 | struct rq *rq; |
| 6994 | |
| 6995 | read_lock_irqsave(&tasklist_lock, flags); |
| 6996 | do_each_thread(g, p) { |
| 6997 | /* |
| 6998 | * Only normalize user tasks: |
| 6999 | */ |
| 7000 | if (!p->mm) |
| 7001 | continue; |
| 7002 | |
| 7003 | p->se.exec_start = 0; |
| 7004 | #ifdef CONFIG_SCHEDSTATS |
| 7005 | p->se.statistics.wait_start = 0; |
| 7006 | p->se.statistics.sleep_start = 0; |
| 7007 | p->se.statistics.block_start = 0; |
| 7008 | #endif |
| 7009 | |
| 7010 | if (!rt_task(p)) { |
| 7011 | /* |
| 7012 | * Renice negative nice level userspace |
| 7013 | * tasks back to 0: |
| 7014 | */ |
| 7015 | if (TASK_NICE(p) < 0 && p->mm) |
| 7016 | set_user_nice(p, 0); |
| 7017 | continue; |
| 7018 | } |
| 7019 | |
| 7020 | raw_spin_lock(&p->pi_lock); |
| 7021 | rq = __task_rq_lock(p); |
| 7022 | |
| 7023 | normalize_task(rq, p); |
| 7024 | |
| 7025 | __task_rq_unlock(rq); |
| 7026 | raw_spin_unlock(&p->pi_lock); |
| 7027 | } while_each_thread(g, p); |
| 7028 | |
| 7029 | read_unlock_irqrestore(&tasklist_lock, flags); |
| 7030 | } |
| 7031 | |
| 7032 | #endif /* CONFIG_MAGIC_SYSRQ */ |
| 7033 | |
| 7034 | #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) |
| 7035 | /* |
| 7036 | * These functions are only useful for the IA64 MCA handling, or kdb. |
| 7037 | * |
| 7038 | * They can only be called when the whole system has been |
| 7039 | * stopped - every CPU needs to be quiescent, and no scheduling |
| 7040 | * activity can take place. Using them for anything else would |
| 7041 | * be a serious bug, and as a result, they aren't even visible |
| 7042 | * under any other configuration. |
| 7043 | */ |
| 7044 | |
| 7045 | /** |
| 7046 | * curr_task - return the current task for a given cpu. |
| 7047 | * @cpu: the processor in question. |
| 7048 | * |
| 7049 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 7050 | */ |
| 7051 | struct task_struct *curr_task(int cpu) |
| 7052 | { |
| 7053 | return cpu_curr(cpu); |
| 7054 | } |
| 7055 | |
| 7056 | #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ |
| 7057 | |
| 7058 | #ifdef CONFIG_IA64 |
| 7059 | /** |
| 7060 | * set_curr_task - set the current task for a given cpu. |
| 7061 | * @cpu: the processor in question. |
| 7062 | * @p: the task pointer to set. |
| 7063 | * |
| 7064 | * Description: This function must only be used when non-maskable interrupts |
| 7065 | * are serviced on a separate stack. It allows the architecture to switch the |
| 7066 | * notion of the current task on a cpu in a non-blocking manner. This function |
| 7067 | * must be called with all CPU's synchronized, and interrupts disabled, the |
| 7068 | * and caller must save the original value of the current task (see |
| 7069 | * curr_task() above) and restore that value before reenabling interrupts and |
| 7070 | * re-starting the system. |
| 7071 | * |
| 7072 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 7073 | */ |
| 7074 | void set_curr_task(int cpu, struct task_struct *p) |
| 7075 | { |
| 7076 | cpu_curr(cpu) = p; |
| 7077 | } |
| 7078 | |
| 7079 | #endif |
| 7080 | |
| 7081 | #ifdef CONFIG_CGROUP_SCHED |
| 7082 | /* task_group_lock serializes the addition/removal of task groups */ |
| 7083 | static DEFINE_SPINLOCK(task_group_lock); |
| 7084 | |
| 7085 | static void free_sched_group(struct task_group *tg) |
| 7086 | { |
| 7087 | free_fair_sched_group(tg); |
| 7088 | free_rt_sched_group(tg); |
| 7089 | autogroup_free(tg); |
| 7090 | kfree(tg); |
| 7091 | } |
| 7092 | |
| 7093 | /* allocate runqueue etc for a new task group */ |
| 7094 | struct task_group *sched_create_group(struct task_group *parent) |
| 7095 | { |
| 7096 | struct task_group *tg; |
| 7097 | unsigned long flags; |
| 7098 | |
| 7099 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); |
| 7100 | if (!tg) |
| 7101 | return ERR_PTR(-ENOMEM); |
| 7102 | |
| 7103 | if (!alloc_fair_sched_group(tg, parent)) |
| 7104 | goto err; |
| 7105 | |
| 7106 | if (!alloc_rt_sched_group(tg, parent)) |
| 7107 | goto err; |
| 7108 | |
| 7109 | spin_lock_irqsave(&task_group_lock, flags); |
| 7110 | list_add_rcu(&tg->list, &task_groups); |
| 7111 | |
| 7112 | WARN_ON(!parent); /* root should already exist */ |
| 7113 | |
| 7114 | tg->parent = parent; |
| 7115 | INIT_LIST_HEAD(&tg->children); |
| 7116 | list_add_rcu(&tg->siblings, &parent->children); |
| 7117 | spin_unlock_irqrestore(&task_group_lock, flags); |
| 7118 | |
| 7119 | return tg; |
| 7120 | |
| 7121 | err: |
| 7122 | free_sched_group(tg); |
| 7123 | return ERR_PTR(-ENOMEM); |
| 7124 | } |
| 7125 | |
| 7126 | /* rcu callback to free various structures associated with a task group */ |
| 7127 | static void free_sched_group_rcu(struct rcu_head *rhp) |
| 7128 | { |
| 7129 | /* now it should be safe to free those cfs_rqs */ |
| 7130 | free_sched_group(container_of(rhp, struct task_group, rcu)); |
| 7131 | } |
| 7132 | |
| 7133 | /* Destroy runqueue etc associated with a task group */ |
| 7134 | void sched_destroy_group(struct task_group *tg) |
| 7135 | { |
| 7136 | unsigned long flags; |
| 7137 | int i; |
| 7138 | |
| 7139 | /* end participation in shares distribution */ |
| 7140 | for_each_possible_cpu(i) |
| 7141 | unregister_fair_sched_group(tg, i); |
| 7142 | |
| 7143 | spin_lock_irqsave(&task_group_lock, flags); |
| 7144 | list_del_rcu(&tg->list); |
| 7145 | list_del_rcu(&tg->siblings); |
| 7146 | spin_unlock_irqrestore(&task_group_lock, flags); |
| 7147 | |
| 7148 | /* wait for possible concurrent references to cfs_rqs complete */ |
| 7149 | call_rcu(&tg->rcu, free_sched_group_rcu); |
| 7150 | } |
| 7151 | |
| 7152 | /* change task's runqueue when it moves between groups. |
| 7153 | * The caller of this function should have put the task in its new group |
| 7154 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to |
| 7155 | * reflect its new group. |
| 7156 | */ |
| 7157 | void sched_move_task(struct task_struct *tsk) |
| 7158 | { |
| 7159 | struct task_group *tg; |
| 7160 | int on_rq, running; |
| 7161 | unsigned long flags; |
| 7162 | struct rq *rq; |
| 7163 | |
| 7164 | rq = task_rq_lock(tsk, &flags); |
| 7165 | |
| 7166 | running = task_current(rq, tsk); |
| 7167 | on_rq = tsk->on_rq; |
| 7168 | |
| 7169 | if (on_rq) |
| 7170 | dequeue_task(rq, tsk, 0); |
| 7171 | if (unlikely(running)) |
| 7172 | tsk->sched_class->put_prev_task(rq, tsk); |
| 7173 | |
| 7174 | tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, |
| 7175 | lockdep_is_held(&tsk->sighand->siglock)), |
| 7176 | struct task_group, css); |
| 7177 | tg = autogroup_task_group(tsk, tg); |
| 7178 | tsk->sched_task_group = tg; |
| 7179 | |
| 7180 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 7181 | if (tsk->sched_class->task_move_group) |
| 7182 | tsk->sched_class->task_move_group(tsk, on_rq); |
| 7183 | else |
| 7184 | #endif |
| 7185 | set_task_rq(tsk, task_cpu(tsk)); |
| 7186 | |
| 7187 | if (unlikely(running)) |
| 7188 | tsk->sched_class->set_curr_task(rq); |
| 7189 | if (on_rq) |
| 7190 | enqueue_task(rq, tsk, 0); |
| 7191 | |
| 7192 | task_rq_unlock(rq, tsk, &flags); |
| 7193 | } |
| 7194 | #endif /* CONFIG_CGROUP_SCHED */ |
| 7195 | |
| 7196 | #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) |
| 7197 | static unsigned long to_ratio(u64 period, u64 runtime) |
| 7198 | { |
| 7199 | if (runtime == RUNTIME_INF) |
| 7200 | return 1ULL << 20; |
| 7201 | |
| 7202 | return div64_u64(runtime << 20, period); |
| 7203 | } |
| 7204 | #endif |
| 7205 | |
| 7206 | #ifdef CONFIG_RT_GROUP_SCHED |
| 7207 | /* |
| 7208 | * Ensure that the real time constraints are schedulable. |
| 7209 | */ |
| 7210 | static DEFINE_MUTEX(rt_constraints_mutex); |
| 7211 | |
| 7212 | /* Must be called with tasklist_lock held */ |
| 7213 | static inline int tg_has_rt_tasks(struct task_group *tg) |
| 7214 | { |
| 7215 | struct task_struct *g, *p; |
| 7216 | |
| 7217 | do_each_thread(g, p) { |
| 7218 | if (rt_task(p) && task_rq(p)->rt.tg == tg) |
| 7219 | return 1; |
| 7220 | } while_each_thread(g, p); |
| 7221 | |
| 7222 | return 0; |
| 7223 | } |
| 7224 | |
| 7225 | struct rt_schedulable_data { |
| 7226 | struct task_group *tg; |
| 7227 | u64 rt_period; |
| 7228 | u64 rt_runtime; |
| 7229 | }; |
| 7230 | |
| 7231 | static int tg_rt_schedulable(struct task_group *tg, void *data) |
| 7232 | { |
| 7233 | struct rt_schedulable_data *d = data; |
| 7234 | struct task_group *child; |
| 7235 | unsigned long total, sum = 0; |
| 7236 | u64 period, runtime; |
| 7237 | |
| 7238 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); |
| 7239 | runtime = tg->rt_bandwidth.rt_runtime; |
| 7240 | |
| 7241 | if (tg == d->tg) { |
| 7242 | period = d->rt_period; |
| 7243 | runtime = d->rt_runtime; |
| 7244 | } |
| 7245 | |
| 7246 | /* |
| 7247 | * Cannot have more runtime than the period. |
| 7248 | */ |
| 7249 | if (runtime > period && runtime != RUNTIME_INF) |
| 7250 | return -EINVAL; |
| 7251 | |
| 7252 | /* |
| 7253 | * Ensure we don't starve existing RT tasks. |
| 7254 | */ |
| 7255 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) |
| 7256 | return -EBUSY; |
| 7257 | |
| 7258 | total = to_ratio(period, runtime); |
| 7259 | |
| 7260 | /* |
| 7261 | * Nobody can have more than the global setting allows. |
| 7262 | */ |
| 7263 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) |
| 7264 | return -EINVAL; |
| 7265 | |
| 7266 | /* |
| 7267 | * The sum of our children's runtime should not exceed our own. |
| 7268 | */ |
| 7269 | list_for_each_entry_rcu(child, &tg->children, siblings) { |
| 7270 | period = ktime_to_ns(child->rt_bandwidth.rt_period); |
| 7271 | runtime = child->rt_bandwidth.rt_runtime; |
| 7272 | |
| 7273 | if (child == d->tg) { |
| 7274 | period = d->rt_period; |
| 7275 | runtime = d->rt_runtime; |
| 7276 | } |
| 7277 | |
| 7278 | sum += to_ratio(period, runtime); |
| 7279 | } |
| 7280 | |
| 7281 | if (sum > total) |
| 7282 | return -EINVAL; |
| 7283 | |
| 7284 | return 0; |
| 7285 | } |
| 7286 | |
| 7287 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) |
| 7288 | { |
| 7289 | int ret; |
| 7290 | |
| 7291 | struct rt_schedulable_data data = { |
| 7292 | .tg = tg, |
| 7293 | .rt_period = period, |
| 7294 | .rt_runtime = runtime, |
| 7295 | }; |
| 7296 | |
| 7297 | rcu_read_lock(); |
| 7298 | ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); |
| 7299 | rcu_read_unlock(); |
| 7300 | |
| 7301 | return ret; |
| 7302 | } |
| 7303 | |
| 7304 | static int tg_set_rt_bandwidth(struct task_group *tg, |
| 7305 | u64 rt_period, u64 rt_runtime) |
| 7306 | { |
| 7307 | int i, err = 0; |
| 7308 | |
| 7309 | mutex_lock(&rt_constraints_mutex); |
| 7310 | read_lock(&tasklist_lock); |
| 7311 | err = __rt_schedulable(tg, rt_period, rt_runtime); |
| 7312 | if (err) |
| 7313 | goto unlock; |
| 7314 | |
| 7315 | raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
| 7316 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); |
| 7317 | tg->rt_bandwidth.rt_runtime = rt_runtime; |
| 7318 | |
| 7319 | for_each_possible_cpu(i) { |
| 7320 | struct rt_rq *rt_rq = tg->rt_rq[i]; |
| 7321 | |
| 7322 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
| 7323 | rt_rq->rt_runtime = rt_runtime; |
| 7324 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
| 7325 | } |
| 7326 | raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
| 7327 | unlock: |
| 7328 | read_unlock(&tasklist_lock); |
| 7329 | mutex_unlock(&rt_constraints_mutex); |
| 7330 | |
| 7331 | return err; |
| 7332 | } |
| 7333 | |
| 7334 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) |
| 7335 | { |
| 7336 | u64 rt_runtime, rt_period; |
| 7337 | |
| 7338 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); |
| 7339 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; |
| 7340 | if (rt_runtime_us < 0) |
| 7341 | rt_runtime = RUNTIME_INF; |
| 7342 | |
| 7343 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); |
| 7344 | } |
| 7345 | |
| 7346 | long sched_group_rt_runtime(struct task_group *tg) |
| 7347 | { |
| 7348 | u64 rt_runtime_us; |
| 7349 | |
| 7350 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) |
| 7351 | return -1; |
| 7352 | |
| 7353 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; |
| 7354 | do_div(rt_runtime_us, NSEC_PER_USEC); |
| 7355 | return rt_runtime_us; |
| 7356 | } |
| 7357 | |
| 7358 | int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) |
| 7359 | { |
| 7360 | u64 rt_runtime, rt_period; |
| 7361 | |
| 7362 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; |
| 7363 | rt_runtime = tg->rt_bandwidth.rt_runtime; |
| 7364 | |
| 7365 | if (rt_period == 0) |
| 7366 | return -EINVAL; |
| 7367 | |
| 7368 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); |
| 7369 | } |
| 7370 | |
| 7371 | long sched_group_rt_period(struct task_group *tg) |
| 7372 | { |
| 7373 | u64 rt_period_us; |
| 7374 | |
| 7375 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); |
| 7376 | do_div(rt_period_us, NSEC_PER_USEC); |
| 7377 | return rt_period_us; |
| 7378 | } |
| 7379 | |
| 7380 | static int sched_rt_global_constraints(void) |
| 7381 | { |
| 7382 | u64 runtime, period; |
| 7383 | int ret = 0; |
| 7384 | |
| 7385 | if (sysctl_sched_rt_period <= 0) |
| 7386 | return -EINVAL; |
| 7387 | |
| 7388 | runtime = global_rt_runtime(); |
| 7389 | period = global_rt_period(); |
| 7390 | |
| 7391 | /* |
| 7392 | * Sanity check on the sysctl variables. |
| 7393 | */ |
| 7394 | if (runtime > period && runtime != RUNTIME_INF) |
| 7395 | return -EINVAL; |
| 7396 | |
| 7397 | mutex_lock(&rt_constraints_mutex); |
| 7398 | read_lock(&tasklist_lock); |
| 7399 | ret = __rt_schedulable(NULL, 0, 0); |
| 7400 | read_unlock(&tasklist_lock); |
| 7401 | mutex_unlock(&rt_constraints_mutex); |
| 7402 | |
| 7403 | return ret; |
| 7404 | } |
| 7405 | |
| 7406 | int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) |
| 7407 | { |
| 7408 | /* Don't accept realtime tasks when there is no way for them to run */ |
| 7409 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) |
| 7410 | return 0; |
| 7411 | |
| 7412 | return 1; |
| 7413 | } |
| 7414 | |
| 7415 | #else /* !CONFIG_RT_GROUP_SCHED */ |
| 7416 | static int sched_rt_global_constraints(void) |
| 7417 | { |
| 7418 | unsigned long flags; |
| 7419 | int i; |
| 7420 | |
| 7421 | if (sysctl_sched_rt_period <= 0) |
| 7422 | return -EINVAL; |
| 7423 | |
| 7424 | /* |
| 7425 | * There's always some RT tasks in the root group |
| 7426 | * -- migration, kstopmachine etc.. |
| 7427 | */ |
| 7428 | if (sysctl_sched_rt_runtime == 0) |
| 7429 | return -EBUSY; |
| 7430 | |
| 7431 | raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); |
| 7432 | for_each_possible_cpu(i) { |
| 7433 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; |
| 7434 | |
| 7435 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
| 7436 | rt_rq->rt_runtime = global_rt_runtime(); |
| 7437 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
| 7438 | } |
| 7439 | raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); |
| 7440 | |
| 7441 | return 0; |
| 7442 | } |
| 7443 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 7444 | |
| 7445 | int sched_rt_handler(struct ctl_table *table, int write, |
| 7446 | void __user *buffer, size_t *lenp, |
| 7447 | loff_t *ppos) |
| 7448 | { |
| 7449 | int ret; |
| 7450 | int old_period, old_runtime; |
| 7451 | static DEFINE_MUTEX(mutex); |
| 7452 | |
| 7453 | mutex_lock(&mutex); |
| 7454 | old_period = sysctl_sched_rt_period; |
| 7455 | old_runtime = sysctl_sched_rt_runtime; |
| 7456 | |
| 7457 | ret = proc_dointvec(table, write, buffer, lenp, ppos); |
| 7458 | |
| 7459 | if (!ret && write) { |
| 7460 | ret = sched_rt_global_constraints(); |
| 7461 | if (ret) { |
| 7462 | sysctl_sched_rt_period = old_period; |
| 7463 | sysctl_sched_rt_runtime = old_runtime; |
| 7464 | } else { |
| 7465 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); |
| 7466 | def_rt_bandwidth.rt_period = |
| 7467 | ns_to_ktime(global_rt_period()); |
| 7468 | } |
| 7469 | } |
| 7470 | mutex_unlock(&mutex); |
| 7471 | |
| 7472 | return ret; |
| 7473 | } |
| 7474 | |
| 7475 | #ifdef CONFIG_CGROUP_SCHED |
| 7476 | |
| 7477 | /* return corresponding task_group object of a cgroup */ |
| 7478 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) |
| 7479 | { |
| 7480 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), |
| 7481 | struct task_group, css); |
| 7482 | } |
| 7483 | |
| 7484 | static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp) |
| 7485 | { |
| 7486 | struct task_group *tg, *parent; |
| 7487 | |
| 7488 | if (!cgrp->parent) { |
| 7489 | /* This is early initialization for the top cgroup */ |
| 7490 | return &root_task_group.css; |
| 7491 | } |
| 7492 | |
| 7493 | parent = cgroup_tg(cgrp->parent); |
| 7494 | tg = sched_create_group(parent); |
| 7495 | if (IS_ERR(tg)) |
| 7496 | return ERR_PTR(-ENOMEM); |
| 7497 | |
| 7498 | return &tg->css; |
| 7499 | } |
| 7500 | |
| 7501 | static void cpu_cgroup_destroy(struct cgroup *cgrp) |
| 7502 | { |
| 7503 | struct task_group *tg = cgroup_tg(cgrp); |
| 7504 | |
| 7505 | sched_destroy_group(tg); |
| 7506 | } |
| 7507 | |
| 7508 | static int cpu_cgroup_can_attach(struct cgroup *cgrp, |
| 7509 | struct cgroup_taskset *tset) |
| 7510 | { |
| 7511 | struct task_struct *task; |
| 7512 | |
| 7513 | cgroup_taskset_for_each(task, cgrp, tset) { |
| 7514 | #ifdef CONFIG_RT_GROUP_SCHED |
| 7515 | if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) |
| 7516 | return -EINVAL; |
| 7517 | #else |
| 7518 | /* We don't support RT-tasks being in separate groups */ |
| 7519 | if (task->sched_class != &fair_sched_class) |
| 7520 | return -EINVAL; |
| 7521 | #endif |
| 7522 | } |
| 7523 | return 0; |
| 7524 | } |
| 7525 | |
| 7526 | static void cpu_cgroup_attach(struct cgroup *cgrp, |
| 7527 | struct cgroup_taskset *tset) |
| 7528 | { |
| 7529 | struct task_struct *task; |
| 7530 | |
| 7531 | cgroup_taskset_for_each(task, cgrp, tset) |
| 7532 | sched_move_task(task); |
| 7533 | } |
| 7534 | |
| 7535 | static void |
| 7536 | cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, |
| 7537 | struct task_struct *task) |
| 7538 | { |
| 7539 | /* |
| 7540 | * cgroup_exit() is called in the copy_process() failure path. |
| 7541 | * Ignore this case since the task hasn't ran yet, this avoids |
| 7542 | * trying to poke a half freed task state from generic code. |
| 7543 | */ |
| 7544 | if (!(task->flags & PF_EXITING)) |
| 7545 | return; |
| 7546 | |
| 7547 | sched_move_task(task); |
| 7548 | } |
| 7549 | |
| 7550 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 7551 | static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, |
| 7552 | u64 shareval) |
| 7553 | { |
| 7554 | return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); |
| 7555 | } |
| 7556 | |
| 7557 | static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) |
| 7558 | { |
| 7559 | struct task_group *tg = cgroup_tg(cgrp); |
| 7560 | |
| 7561 | return (u64) scale_load_down(tg->shares); |
| 7562 | } |
| 7563 | |
| 7564 | #ifdef CONFIG_CFS_BANDWIDTH |
| 7565 | static DEFINE_MUTEX(cfs_constraints_mutex); |
| 7566 | |
| 7567 | const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ |
| 7568 | const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ |
| 7569 | |
| 7570 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); |
| 7571 | |
| 7572 | static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) |
| 7573 | { |
| 7574 | int i, ret = 0, runtime_enabled, runtime_was_enabled; |
| 7575 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
| 7576 | |
| 7577 | if (tg == &root_task_group) |
| 7578 | return -EINVAL; |
| 7579 | |
| 7580 | /* |
| 7581 | * Ensure we have at some amount of bandwidth every period. This is |
| 7582 | * to prevent reaching a state of large arrears when throttled via |
| 7583 | * entity_tick() resulting in prolonged exit starvation. |
| 7584 | */ |
| 7585 | if (quota < min_cfs_quota_period || period < min_cfs_quota_period) |
| 7586 | return -EINVAL; |
| 7587 | |
| 7588 | /* |
| 7589 | * Likewise, bound things on the otherside by preventing insane quota |
| 7590 | * periods. This also allows us to normalize in computing quota |
| 7591 | * feasibility. |
| 7592 | */ |
| 7593 | if (period > max_cfs_quota_period) |
| 7594 | return -EINVAL; |
| 7595 | |
| 7596 | mutex_lock(&cfs_constraints_mutex); |
| 7597 | ret = __cfs_schedulable(tg, period, quota); |
| 7598 | if (ret) |
| 7599 | goto out_unlock; |
| 7600 | |
| 7601 | runtime_enabled = quota != RUNTIME_INF; |
| 7602 | runtime_was_enabled = cfs_b->quota != RUNTIME_INF; |
| 7603 | account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); |
| 7604 | raw_spin_lock_irq(&cfs_b->lock); |
| 7605 | cfs_b->period = ns_to_ktime(period); |
| 7606 | cfs_b->quota = quota; |
| 7607 | |
| 7608 | __refill_cfs_bandwidth_runtime(cfs_b); |
| 7609 | /* restart the period timer (if active) to handle new period expiry */ |
| 7610 | if (runtime_enabled && cfs_b->timer_active) { |
| 7611 | /* force a reprogram */ |
| 7612 | cfs_b->timer_active = 0; |
| 7613 | __start_cfs_bandwidth(cfs_b); |
| 7614 | } |
| 7615 | raw_spin_unlock_irq(&cfs_b->lock); |
| 7616 | |
| 7617 | for_each_possible_cpu(i) { |
| 7618 | struct cfs_rq *cfs_rq = tg->cfs_rq[i]; |
| 7619 | struct rq *rq = cfs_rq->rq; |
| 7620 | |
| 7621 | raw_spin_lock_irq(&rq->lock); |
| 7622 | cfs_rq->runtime_enabled = runtime_enabled; |
| 7623 | cfs_rq->runtime_remaining = 0; |
| 7624 | |
| 7625 | if (cfs_rq->throttled) |
| 7626 | unthrottle_cfs_rq(cfs_rq); |
| 7627 | raw_spin_unlock_irq(&rq->lock); |
| 7628 | } |
| 7629 | out_unlock: |
| 7630 | mutex_unlock(&cfs_constraints_mutex); |
| 7631 | |
| 7632 | return ret; |
| 7633 | } |
| 7634 | |
| 7635 | int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) |
| 7636 | { |
| 7637 | u64 quota, period; |
| 7638 | |
| 7639 | period = ktime_to_ns(tg->cfs_bandwidth.period); |
| 7640 | if (cfs_quota_us < 0) |
| 7641 | quota = RUNTIME_INF; |
| 7642 | else |
| 7643 | quota = (u64)cfs_quota_us * NSEC_PER_USEC; |
| 7644 | |
| 7645 | return tg_set_cfs_bandwidth(tg, period, quota); |
| 7646 | } |
| 7647 | |
| 7648 | long tg_get_cfs_quota(struct task_group *tg) |
| 7649 | { |
| 7650 | u64 quota_us; |
| 7651 | |
| 7652 | if (tg->cfs_bandwidth.quota == RUNTIME_INF) |
| 7653 | return -1; |
| 7654 | |
| 7655 | quota_us = tg->cfs_bandwidth.quota; |
| 7656 | do_div(quota_us, NSEC_PER_USEC); |
| 7657 | |
| 7658 | return quota_us; |
| 7659 | } |
| 7660 | |
| 7661 | int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) |
| 7662 | { |
| 7663 | u64 quota, period; |
| 7664 | |
| 7665 | period = (u64)cfs_period_us * NSEC_PER_USEC; |
| 7666 | quota = tg->cfs_bandwidth.quota; |
| 7667 | |
| 7668 | return tg_set_cfs_bandwidth(tg, period, quota); |
| 7669 | } |
| 7670 | |
| 7671 | long tg_get_cfs_period(struct task_group *tg) |
| 7672 | { |
| 7673 | u64 cfs_period_us; |
| 7674 | |
| 7675 | cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); |
| 7676 | do_div(cfs_period_us, NSEC_PER_USEC); |
| 7677 | |
| 7678 | return cfs_period_us; |
| 7679 | } |
| 7680 | |
| 7681 | static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) |
| 7682 | { |
| 7683 | return tg_get_cfs_quota(cgroup_tg(cgrp)); |
| 7684 | } |
| 7685 | |
| 7686 | static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, |
| 7687 | s64 cfs_quota_us) |
| 7688 | { |
| 7689 | return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); |
| 7690 | } |
| 7691 | |
| 7692 | static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) |
| 7693 | { |
| 7694 | return tg_get_cfs_period(cgroup_tg(cgrp)); |
| 7695 | } |
| 7696 | |
| 7697 | static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, |
| 7698 | u64 cfs_period_us) |
| 7699 | { |
| 7700 | return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); |
| 7701 | } |
| 7702 | |
| 7703 | struct cfs_schedulable_data { |
| 7704 | struct task_group *tg; |
| 7705 | u64 period, quota; |
| 7706 | }; |
| 7707 | |
| 7708 | /* |
| 7709 | * normalize group quota/period to be quota/max_period |
| 7710 | * note: units are usecs |
| 7711 | */ |
| 7712 | static u64 normalize_cfs_quota(struct task_group *tg, |
| 7713 | struct cfs_schedulable_data *d) |
| 7714 | { |
| 7715 | u64 quota, period; |
| 7716 | |
| 7717 | if (tg == d->tg) { |
| 7718 | period = d->period; |
| 7719 | quota = d->quota; |
| 7720 | } else { |
| 7721 | period = tg_get_cfs_period(tg); |
| 7722 | quota = tg_get_cfs_quota(tg); |
| 7723 | } |
| 7724 | |
| 7725 | /* note: these should typically be equivalent */ |
| 7726 | if (quota == RUNTIME_INF || quota == -1) |
| 7727 | return RUNTIME_INF; |
| 7728 | |
| 7729 | return to_ratio(period, quota); |
| 7730 | } |
| 7731 | |
| 7732 | static int tg_cfs_schedulable_down(struct task_group *tg, void *data) |
| 7733 | { |
| 7734 | struct cfs_schedulable_data *d = data; |
| 7735 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
| 7736 | s64 quota = 0, parent_quota = -1; |
| 7737 | |
| 7738 | if (!tg->parent) { |
| 7739 | quota = RUNTIME_INF; |
| 7740 | } else { |
| 7741 | struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; |
| 7742 | |
| 7743 | quota = normalize_cfs_quota(tg, d); |
| 7744 | parent_quota = parent_b->hierarchal_quota; |
| 7745 | |
| 7746 | /* |
| 7747 | * ensure max(child_quota) <= parent_quota, inherit when no |
| 7748 | * limit is set |
| 7749 | */ |
| 7750 | if (quota == RUNTIME_INF) |
| 7751 | quota = parent_quota; |
| 7752 | else if (parent_quota != RUNTIME_INF && quota > parent_quota) |
| 7753 | return -EINVAL; |
| 7754 | } |
| 7755 | cfs_b->hierarchal_quota = quota; |
| 7756 | |
| 7757 | return 0; |
| 7758 | } |
| 7759 | |
| 7760 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) |
| 7761 | { |
| 7762 | int ret; |
| 7763 | struct cfs_schedulable_data data = { |
| 7764 | .tg = tg, |
| 7765 | .period = period, |
| 7766 | .quota = quota, |
| 7767 | }; |
| 7768 | |
| 7769 | if (quota != RUNTIME_INF) { |
| 7770 | do_div(data.period, NSEC_PER_USEC); |
| 7771 | do_div(data.quota, NSEC_PER_USEC); |
| 7772 | } |
| 7773 | |
| 7774 | rcu_read_lock(); |
| 7775 | ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); |
| 7776 | rcu_read_unlock(); |
| 7777 | |
| 7778 | return ret; |
| 7779 | } |
| 7780 | |
| 7781 | static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, |
| 7782 | struct cgroup_map_cb *cb) |
| 7783 | { |
| 7784 | struct task_group *tg = cgroup_tg(cgrp); |
| 7785 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
| 7786 | |
| 7787 | cb->fill(cb, "nr_periods", cfs_b->nr_periods); |
| 7788 | cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); |
| 7789 | cb->fill(cb, "throttled_time", cfs_b->throttled_time); |
| 7790 | |
| 7791 | return 0; |
| 7792 | } |
| 7793 | #endif /* CONFIG_CFS_BANDWIDTH */ |
| 7794 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 7795 | |
| 7796 | #ifdef CONFIG_RT_GROUP_SCHED |
| 7797 | static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, |
| 7798 | s64 val) |
| 7799 | { |
| 7800 | return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); |
| 7801 | } |
| 7802 | |
| 7803 | static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) |
| 7804 | { |
| 7805 | return sched_group_rt_runtime(cgroup_tg(cgrp)); |
| 7806 | } |
| 7807 | |
| 7808 | static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, |
| 7809 | u64 rt_period_us) |
| 7810 | { |
| 7811 | return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); |
| 7812 | } |
| 7813 | |
| 7814 | static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) |
| 7815 | { |
| 7816 | return sched_group_rt_period(cgroup_tg(cgrp)); |
| 7817 | } |
| 7818 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 7819 | |
| 7820 | static struct cftype cpu_files[] = { |
| 7821 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 7822 | { |
| 7823 | .name = "shares", |
| 7824 | .read_u64 = cpu_shares_read_u64, |
| 7825 | .write_u64 = cpu_shares_write_u64, |
| 7826 | }, |
| 7827 | #endif |
| 7828 | #ifdef CONFIG_CFS_BANDWIDTH |
| 7829 | { |
| 7830 | .name = "cfs_quota_us", |
| 7831 | .read_s64 = cpu_cfs_quota_read_s64, |
| 7832 | .write_s64 = cpu_cfs_quota_write_s64, |
| 7833 | }, |
| 7834 | { |
| 7835 | .name = "cfs_period_us", |
| 7836 | .read_u64 = cpu_cfs_period_read_u64, |
| 7837 | .write_u64 = cpu_cfs_period_write_u64, |
| 7838 | }, |
| 7839 | { |
| 7840 | .name = "stat", |
| 7841 | .read_map = cpu_stats_show, |
| 7842 | }, |
| 7843 | #endif |
| 7844 | #ifdef CONFIG_RT_GROUP_SCHED |
| 7845 | { |
| 7846 | .name = "rt_runtime_us", |
| 7847 | .read_s64 = cpu_rt_runtime_read, |
| 7848 | .write_s64 = cpu_rt_runtime_write, |
| 7849 | }, |
| 7850 | { |
| 7851 | .name = "rt_period_us", |
| 7852 | .read_u64 = cpu_rt_period_read_uint, |
| 7853 | .write_u64 = cpu_rt_period_write_uint, |
| 7854 | }, |
| 7855 | #endif |
| 7856 | { } /* terminate */ |
| 7857 | }; |
| 7858 | |
| 7859 | struct cgroup_subsys cpu_cgroup_subsys = { |
| 7860 | .name = "cpu", |
| 7861 | .create = cpu_cgroup_create, |
| 7862 | .destroy = cpu_cgroup_destroy, |
| 7863 | .can_attach = cpu_cgroup_can_attach, |
| 7864 | .attach = cpu_cgroup_attach, |
| 7865 | .exit = cpu_cgroup_exit, |
| 7866 | .subsys_id = cpu_cgroup_subsys_id, |
| 7867 | .base_cftypes = cpu_files, |
| 7868 | .early_init = 1, |
| 7869 | }; |
| 7870 | |
| 7871 | #endif /* CONFIG_CGROUP_SCHED */ |
| 7872 | |
| 7873 | #ifdef CONFIG_CGROUP_CPUACCT |
| 7874 | |
| 7875 | /* |
| 7876 | * CPU accounting code for task groups. |
| 7877 | * |
| 7878 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh |
| 7879 | * (balbir@in.ibm.com). |
| 7880 | */ |
| 7881 | |
| 7882 | struct cpuacct root_cpuacct; |
| 7883 | |
| 7884 | /* create a new cpu accounting group */ |
| 7885 | static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp) |
| 7886 | { |
| 7887 | struct cpuacct *ca; |
| 7888 | |
| 7889 | if (!cgrp->parent) |
| 7890 | return &root_cpuacct.css; |
| 7891 | |
| 7892 | ca = kzalloc(sizeof(*ca), GFP_KERNEL); |
| 7893 | if (!ca) |
| 7894 | goto out; |
| 7895 | |
| 7896 | ca->cpuusage = alloc_percpu(u64); |
| 7897 | if (!ca->cpuusage) |
| 7898 | goto out_free_ca; |
| 7899 | |
| 7900 | ca->cpustat = alloc_percpu(struct kernel_cpustat); |
| 7901 | if (!ca->cpustat) |
| 7902 | goto out_free_cpuusage; |
| 7903 | |
| 7904 | return &ca->css; |
| 7905 | |
| 7906 | out_free_cpuusage: |
| 7907 | free_percpu(ca->cpuusage); |
| 7908 | out_free_ca: |
| 7909 | kfree(ca); |
| 7910 | out: |
| 7911 | return ERR_PTR(-ENOMEM); |
| 7912 | } |
| 7913 | |
| 7914 | /* destroy an existing cpu accounting group */ |
| 7915 | static void cpuacct_destroy(struct cgroup *cgrp) |
| 7916 | { |
| 7917 | struct cpuacct *ca = cgroup_ca(cgrp); |
| 7918 | |
| 7919 | free_percpu(ca->cpustat); |
| 7920 | free_percpu(ca->cpuusage); |
| 7921 | kfree(ca); |
| 7922 | } |
| 7923 | |
| 7924 | static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) |
| 7925 | { |
| 7926 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); |
| 7927 | u64 data; |
| 7928 | |
| 7929 | #ifndef CONFIG_64BIT |
| 7930 | /* |
| 7931 | * Take rq->lock to make 64-bit read safe on 32-bit platforms. |
| 7932 | */ |
| 7933 | raw_spin_lock_irq(&cpu_rq(cpu)->lock); |
| 7934 | data = *cpuusage; |
| 7935 | raw_spin_unlock_irq(&cpu_rq(cpu)->lock); |
| 7936 | #else |
| 7937 | data = *cpuusage; |
| 7938 | #endif |
| 7939 | |
| 7940 | return data; |
| 7941 | } |
| 7942 | |
| 7943 | static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) |
| 7944 | { |
| 7945 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); |
| 7946 | |
| 7947 | #ifndef CONFIG_64BIT |
| 7948 | /* |
| 7949 | * Take rq->lock to make 64-bit write safe on 32-bit platforms. |
| 7950 | */ |
| 7951 | raw_spin_lock_irq(&cpu_rq(cpu)->lock); |
| 7952 | *cpuusage = val; |
| 7953 | raw_spin_unlock_irq(&cpu_rq(cpu)->lock); |
| 7954 | #else |
| 7955 | *cpuusage = val; |
| 7956 | #endif |
| 7957 | } |
| 7958 | |
| 7959 | /* return total cpu usage (in nanoseconds) of a group */ |
| 7960 | static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) |
| 7961 | { |
| 7962 | struct cpuacct *ca = cgroup_ca(cgrp); |
| 7963 | u64 totalcpuusage = 0; |
| 7964 | int i; |
| 7965 | |
| 7966 | for_each_present_cpu(i) |
| 7967 | totalcpuusage += cpuacct_cpuusage_read(ca, i); |
| 7968 | |
| 7969 | return totalcpuusage; |
| 7970 | } |
| 7971 | |
| 7972 | static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, |
| 7973 | u64 reset) |
| 7974 | { |
| 7975 | struct cpuacct *ca = cgroup_ca(cgrp); |
| 7976 | int err = 0; |
| 7977 | int i; |
| 7978 | |
| 7979 | if (reset) { |
| 7980 | err = -EINVAL; |
| 7981 | goto out; |
| 7982 | } |
| 7983 | |
| 7984 | for_each_present_cpu(i) |
| 7985 | cpuacct_cpuusage_write(ca, i, 0); |
| 7986 | |
| 7987 | out: |
| 7988 | return err; |
| 7989 | } |
| 7990 | |
| 7991 | static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, |
| 7992 | struct seq_file *m) |
| 7993 | { |
| 7994 | struct cpuacct *ca = cgroup_ca(cgroup); |
| 7995 | u64 percpu; |
| 7996 | int i; |
| 7997 | |
| 7998 | for_each_present_cpu(i) { |
| 7999 | percpu = cpuacct_cpuusage_read(ca, i); |
| 8000 | seq_printf(m, "%llu ", (unsigned long long) percpu); |
| 8001 | } |
| 8002 | seq_printf(m, "\n"); |
| 8003 | return 0; |
| 8004 | } |
| 8005 | |
| 8006 | static const char *cpuacct_stat_desc[] = { |
| 8007 | [CPUACCT_STAT_USER] = "user", |
| 8008 | [CPUACCT_STAT_SYSTEM] = "system", |
| 8009 | }; |
| 8010 | |
| 8011 | static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, |
| 8012 | struct cgroup_map_cb *cb) |
| 8013 | { |
| 8014 | struct cpuacct *ca = cgroup_ca(cgrp); |
| 8015 | int cpu; |
| 8016 | s64 val = 0; |
| 8017 | |
| 8018 | for_each_online_cpu(cpu) { |
| 8019 | struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); |
| 8020 | val += kcpustat->cpustat[CPUTIME_USER]; |
| 8021 | val += kcpustat->cpustat[CPUTIME_NICE]; |
| 8022 | } |
| 8023 | val = cputime64_to_clock_t(val); |
| 8024 | cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); |
| 8025 | |
| 8026 | val = 0; |
| 8027 | for_each_online_cpu(cpu) { |
| 8028 | struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); |
| 8029 | val += kcpustat->cpustat[CPUTIME_SYSTEM]; |
| 8030 | val += kcpustat->cpustat[CPUTIME_IRQ]; |
| 8031 | val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; |
| 8032 | } |
| 8033 | |
| 8034 | val = cputime64_to_clock_t(val); |
| 8035 | cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); |
| 8036 | |
| 8037 | return 0; |
| 8038 | } |
| 8039 | |
| 8040 | static struct cftype files[] = { |
| 8041 | { |
| 8042 | .name = "usage", |
| 8043 | .read_u64 = cpuusage_read, |
| 8044 | .write_u64 = cpuusage_write, |
| 8045 | }, |
| 8046 | { |
| 8047 | .name = "usage_percpu", |
| 8048 | .read_seq_string = cpuacct_percpu_seq_read, |
| 8049 | }, |
| 8050 | { |
| 8051 | .name = "stat", |
| 8052 | .read_map = cpuacct_stats_show, |
| 8053 | }, |
| 8054 | { } /* terminate */ |
| 8055 | }; |
| 8056 | |
| 8057 | /* |
| 8058 | * charge this task's execution time to its accounting group. |
| 8059 | * |
| 8060 | * called with rq->lock held. |
| 8061 | */ |
| 8062 | void cpuacct_charge(struct task_struct *tsk, u64 cputime) |
| 8063 | { |
| 8064 | struct cpuacct *ca; |
| 8065 | int cpu; |
| 8066 | |
| 8067 | if (unlikely(!cpuacct_subsys.active)) |
| 8068 | return; |
| 8069 | |
| 8070 | cpu = task_cpu(tsk); |
| 8071 | |
| 8072 | rcu_read_lock(); |
| 8073 | |
| 8074 | ca = task_ca(tsk); |
| 8075 | |
| 8076 | for (; ca; ca = parent_ca(ca)) { |
| 8077 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); |
| 8078 | *cpuusage += cputime; |
| 8079 | } |
| 8080 | |
| 8081 | rcu_read_unlock(); |
| 8082 | } |
| 8083 | |
| 8084 | struct cgroup_subsys cpuacct_subsys = { |
| 8085 | .name = "cpuacct", |
| 8086 | .create = cpuacct_create, |
| 8087 | .destroy = cpuacct_destroy, |
| 8088 | .subsys_id = cpuacct_subsys_id, |
| 8089 | .base_cftypes = files, |
| 8090 | }; |
| 8091 | #endif /* CONFIG_CGROUP_CPUACCT */ |