| 1 | /* |
| 2 | * kernel/sched.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 <linux/smp_lock.h> |
| 36 | #include <asm/mmu_context.h> |
| 37 | #include <linux/interrupt.h> |
| 38 | #include <linux/capability.h> |
| 39 | #include <linux/completion.h> |
| 40 | #include <linux/kernel_stat.h> |
| 41 | #include <linux/debug_locks.h> |
| 42 | #include <linux/perf_counter.h> |
| 43 | #include <linux/security.h> |
| 44 | #include <linux/notifier.h> |
| 45 | #include <linux/profile.h> |
| 46 | #include <linux/freezer.h> |
| 47 | #include <linux/vmalloc.h> |
| 48 | #include <linux/blkdev.h> |
| 49 | #include <linux/delay.h> |
| 50 | #include <linux/pid_namespace.h> |
| 51 | #include <linux/smp.h> |
| 52 | #include <linux/threads.h> |
| 53 | #include <linux/timer.h> |
| 54 | #include <linux/rcupdate.h> |
| 55 | #include <linux/cpu.h> |
| 56 | #include <linux/cpuset.h> |
| 57 | #include <linux/percpu.h> |
| 58 | #include <linux/kthread.h> |
| 59 | #include <linux/proc_fs.h> |
| 60 | #include <linux/seq_file.h> |
| 61 | #include <linux/sysctl.h> |
| 62 | #include <linux/syscalls.h> |
| 63 | #include <linux/times.h> |
| 64 | #include <linux/tsacct_kern.h> |
| 65 | #include <linux/kprobes.h> |
| 66 | #include <linux/delayacct.h> |
| 67 | #include <linux/unistd.h> |
| 68 | #include <linux/pagemap.h> |
| 69 | #include <linux/hrtimer.h> |
| 70 | #include <linux/tick.h> |
| 71 | #include <linux/debugfs.h> |
| 72 | #include <linux/ctype.h> |
| 73 | #include <linux/ftrace.h> |
| 74 | |
| 75 | #include <asm/tlb.h> |
| 76 | #include <asm/irq_regs.h> |
| 77 | |
| 78 | #include "sched_cpupri.h" |
| 79 | |
| 80 | #define CREATE_TRACE_POINTS |
| 81 | #include <trace/events/sched.h> |
| 82 | |
| 83 | /* |
| 84 | * Convert user-nice values [ -20 ... 0 ... 19 ] |
| 85 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], |
| 86 | * and back. |
| 87 | */ |
| 88 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) |
| 89 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) |
| 90 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) |
| 91 | |
| 92 | /* |
| 93 | * 'User priority' is the nice value converted to something we |
| 94 | * can work with better when scaling various scheduler parameters, |
| 95 | * it's a [ 0 ... 39 ] range. |
| 96 | */ |
| 97 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) |
| 98 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) |
| 99 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) |
| 100 | |
| 101 | /* |
| 102 | * Helpers for converting nanosecond timing to jiffy resolution |
| 103 | */ |
| 104 | #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) |
| 105 | |
| 106 | #define NICE_0_LOAD SCHED_LOAD_SCALE |
| 107 | #define NICE_0_SHIFT SCHED_LOAD_SHIFT |
| 108 | |
| 109 | /* |
| 110 | * These are the 'tuning knobs' of the scheduler: |
| 111 | * |
| 112 | * default timeslice is 100 msecs (used only for SCHED_RR tasks). |
| 113 | * Timeslices get refilled after they expire. |
| 114 | */ |
| 115 | #define DEF_TIMESLICE (100 * HZ / 1000) |
| 116 | |
| 117 | /* |
| 118 | * single value that denotes runtime == period, ie unlimited time. |
| 119 | */ |
| 120 | #define RUNTIME_INF ((u64)~0ULL) |
| 121 | |
| 122 | static inline int rt_policy(int policy) |
| 123 | { |
| 124 | if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR)) |
| 125 | return 1; |
| 126 | return 0; |
| 127 | } |
| 128 | |
| 129 | static inline int task_has_rt_policy(struct task_struct *p) |
| 130 | { |
| 131 | return rt_policy(p->policy); |
| 132 | } |
| 133 | |
| 134 | /* |
| 135 | * This is the priority-queue data structure of the RT scheduling class: |
| 136 | */ |
| 137 | struct rt_prio_array { |
| 138 | DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ |
| 139 | struct list_head queue[MAX_RT_PRIO]; |
| 140 | }; |
| 141 | |
| 142 | struct rt_bandwidth { |
| 143 | /* nests inside the rq lock: */ |
| 144 | spinlock_t rt_runtime_lock; |
| 145 | ktime_t rt_period; |
| 146 | u64 rt_runtime; |
| 147 | struct hrtimer rt_period_timer; |
| 148 | }; |
| 149 | |
| 150 | static struct rt_bandwidth def_rt_bandwidth; |
| 151 | |
| 152 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); |
| 153 | |
| 154 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) |
| 155 | { |
| 156 | struct rt_bandwidth *rt_b = |
| 157 | container_of(timer, struct rt_bandwidth, rt_period_timer); |
| 158 | ktime_t now; |
| 159 | int overrun; |
| 160 | int idle = 0; |
| 161 | |
| 162 | for (;;) { |
| 163 | now = hrtimer_cb_get_time(timer); |
| 164 | overrun = hrtimer_forward(timer, now, rt_b->rt_period); |
| 165 | |
| 166 | if (!overrun) |
| 167 | break; |
| 168 | |
| 169 | idle = do_sched_rt_period_timer(rt_b, overrun); |
| 170 | } |
| 171 | |
| 172 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; |
| 173 | } |
| 174 | |
| 175 | static |
| 176 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) |
| 177 | { |
| 178 | rt_b->rt_period = ns_to_ktime(period); |
| 179 | rt_b->rt_runtime = runtime; |
| 180 | |
| 181 | spin_lock_init(&rt_b->rt_runtime_lock); |
| 182 | |
| 183 | hrtimer_init(&rt_b->rt_period_timer, |
| 184 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
| 185 | rt_b->rt_period_timer.function = sched_rt_period_timer; |
| 186 | } |
| 187 | |
| 188 | static inline int rt_bandwidth_enabled(void) |
| 189 | { |
| 190 | return sysctl_sched_rt_runtime >= 0; |
| 191 | } |
| 192 | |
| 193 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) |
| 194 | { |
| 195 | ktime_t now; |
| 196 | |
| 197 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) |
| 198 | return; |
| 199 | |
| 200 | if (hrtimer_active(&rt_b->rt_period_timer)) |
| 201 | return; |
| 202 | |
| 203 | spin_lock(&rt_b->rt_runtime_lock); |
| 204 | for (;;) { |
| 205 | unsigned long delta; |
| 206 | ktime_t soft, hard; |
| 207 | |
| 208 | if (hrtimer_active(&rt_b->rt_period_timer)) |
| 209 | break; |
| 210 | |
| 211 | now = hrtimer_cb_get_time(&rt_b->rt_period_timer); |
| 212 | hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period); |
| 213 | |
| 214 | soft = hrtimer_get_softexpires(&rt_b->rt_period_timer); |
| 215 | hard = hrtimer_get_expires(&rt_b->rt_period_timer); |
| 216 | delta = ktime_to_ns(ktime_sub(hard, soft)); |
| 217 | __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta, |
| 218 | HRTIMER_MODE_ABS_PINNED, 0); |
| 219 | } |
| 220 | spin_unlock(&rt_b->rt_runtime_lock); |
| 221 | } |
| 222 | |
| 223 | #ifdef CONFIG_RT_GROUP_SCHED |
| 224 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) |
| 225 | { |
| 226 | hrtimer_cancel(&rt_b->rt_period_timer); |
| 227 | } |
| 228 | #endif |
| 229 | |
| 230 | /* |
| 231 | * sched_domains_mutex serializes calls to arch_init_sched_domains, |
| 232 | * detach_destroy_domains and partition_sched_domains. |
| 233 | */ |
| 234 | static DEFINE_MUTEX(sched_domains_mutex); |
| 235 | |
| 236 | #ifdef CONFIG_GROUP_SCHED |
| 237 | |
| 238 | #include <linux/cgroup.h> |
| 239 | |
| 240 | struct cfs_rq; |
| 241 | |
| 242 | static LIST_HEAD(task_groups); |
| 243 | |
| 244 | /* task group related information */ |
| 245 | struct task_group { |
| 246 | #ifdef CONFIG_CGROUP_SCHED |
| 247 | struct cgroup_subsys_state css; |
| 248 | #endif |
| 249 | |
| 250 | #ifdef CONFIG_USER_SCHED |
| 251 | uid_t uid; |
| 252 | #endif |
| 253 | |
| 254 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 255 | /* schedulable entities of this group on each cpu */ |
| 256 | struct sched_entity **se; |
| 257 | /* runqueue "owned" by this group on each cpu */ |
| 258 | struct cfs_rq **cfs_rq; |
| 259 | unsigned long shares; |
| 260 | #endif |
| 261 | |
| 262 | #ifdef CONFIG_RT_GROUP_SCHED |
| 263 | struct sched_rt_entity **rt_se; |
| 264 | struct rt_rq **rt_rq; |
| 265 | |
| 266 | struct rt_bandwidth rt_bandwidth; |
| 267 | #endif |
| 268 | |
| 269 | struct rcu_head rcu; |
| 270 | struct list_head list; |
| 271 | |
| 272 | struct task_group *parent; |
| 273 | struct list_head siblings; |
| 274 | struct list_head children; |
| 275 | }; |
| 276 | |
| 277 | #ifdef CONFIG_USER_SCHED |
| 278 | |
| 279 | /* Helper function to pass uid information to create_sched_user() */ |
| 280 | void set_tg_uid(struct user_struct *user) |
| 281 | { |
| 282 | user->tg->uid = user->uid; |
| 283 | } |
| 284 | |
| 285 | /* |
| 286 | * Root task group. |
| 287 | * Every UID task group (including init_task_group aka UID-0) will |
| 288 | * be a child to this group. |
| 289 | */ |
| 290 | struct task_group root_task_group; |
| 291 | |
| 292 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 293 | /* Default task group's sched entity on each cpu */ |
| 294 | static DEFINE_PER_CPU(struct sched_entity, init_sched_entity); |
| 295 | /* Default task group's cfs_rq on each cpu */ |
| 296 | static DEFINE_PER_CPU(struct cfs_rq, init_tg_cfs_rq) ____cacheline_aligned_in_smp; |
| 297 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 298 | |
| 299 | #ifdef CONFIG_RT_GROUP_SCHED |
| 300 | static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity); |
| 301 | static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp; |
| 302 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 303 | #else /* !CONFIG_USER_SCHED */ |
| 304 | #define root_task_group init_task_group |
| 305 | #endif /* CONFIG_USER_SCHED */ |
| 306 | |
| 307 | /* task_group_lock serializes add/remove of task groups and also changes to |
| 308 | * a task group's cpu shares. |
| 309 | */ |
| 310 | static DEFINE_SPINLOCK(task_group_lock); |
| 311 | |
| 312 | #ifdef CONFIG_SMP |
| 313 | static int root_task_group_empty(void) |
| 314 | { |
| 315 | return list_empty(&root_task_group.children); |
| 316 | } |
| 317 | #endif |
| 318 | |
| 319 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 320 | #ifdef CONFIG_USER_SCHED |
| 321 | # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD) |
| 322 | #else /* !CONFIG_USER_SCHED */ |
| 323 | # define INIT_TASK_GROUP_LOAD NICE_0_LOAD |
| 324 | #endif /* CONFIG_USER_SCHED */ |
| 325 | |
| 326 | /* |
| 327 | * A weight of 0 or 1 can cause arithmetics problems. |
| 328 | * A weight of a cfs_rq is the sum of weights of which entities |
| 329 | * are queued on this cfs_rq, so a weight of a entity should not be |
| 330 | * too large, so as the shares value of a task group. |
| 331 | * (The default weight is 1024 - so there's no practical |
| 332 | * limitation from this.) |
| 333 | */ |
| 334 | #define MIN_SHARES 2 |
| 335 | #define MAX_SHARES (1UL << 18) |
| 336 | |
| 337 | static int init_task_group_load = INIT_TASK_GROUP_LOAD; |
| 338 | #endif |
| 339 | |
| 340 | /* Default task group. |
| 341 | * Every task in system belong to this group at bootup. |
| 342 | */ |
| 343 | struct task_group init_task_group; |
| 344 | |
| 345 | /* return group to which a task belongs */ |
| 346 | static inline struct task_group *task_group(struct task_struct *p) |
| 347 | { |
| 348 | struct task_group *tg; |
| 349 | |
| 350 | #ifdef CONFIG_USER_SCHED |
| 351 | rcu_read_lock(); |
| 352 | tg = __task_cred(p)->user->tg; |
| 353 | rcu_read_unlock(); |
| 354 | #elif defined(CONFIG_CGROUP_SCHED) |
| 355 | tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id), |
| 356 | struct task_group, css); |
| 357 | #else |
| 358 | tg = &init_task_group; |
| 359 | #endif |
| 360 | return tg; |
| 361 | } |
| 362 | |
| 363 | /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ |
| 364 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) |
| 365 | { |
| 366 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 367 | p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; |
| 368 | p->se.parent = task_group(p)->se[cpu]; |
| 369 | #endif |
| 370 | |
| 371 | #ifdef CONFIG_RT_GROUP_SCHED |
| 372 | p->rt.rt_rq = task_group(p)->rt_rq[cpu]; |
| 373 | p->rt.parent = task_group(p)->rt_se[cpu]; |
| 374 | #endif |
| 375 | } |
| 376 | |
| 377 | #else |
| 378 | |
| 379 | #ifdef CONFIG_SMP |
| 380 | static int root_task_group_empty(void) |
| 381 | { |
| 382 | return 1; |
| 383 | } |
| 384 | #endif |
| 385 | |
| 386 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } |
| 387 | static inline struct task_group *task_group(struct task_struct *p) |
| 388 | { |
| 389 | return NULL; |
| 390 | } |
| 391 | |
| 392 | #endif /* CONFIG_GROUP_SCHED */ |
| 393 | |
| 394 | /* CFS-related fields in a runqueue */ |
| 395 | struct cfs_rq { |
| 396 | struct load_weight load; |
| 397 | unsigned long nr_running; |
| 398 | |
| 399 | u64 exec_clock; |
| 400 | u64 min_vruntime; |
| 401 | |
| 402 | struct rb_root tasks_timeline; |
| 403 | struct rb_node *rb_leftmost; |
| 404 | |
| 405 | struct list_head tasks; |
| 406 | struct list_head *balance_iterator; |
| 407 | |
| 408 | /* |
| 409 | * 'curr' points to currently running entity on this cfs_rq. |
| 410 | * It is set to NULL otherwise (i.e when none are currently running). |
| 411 | */ |
| 412 | struct sched_entity *curr, *next, *last; |
| 413 | |
| 414 | unsigned int nr_spread_over; |
| 415 | |
| 416 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 417 | struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ |
| 418 | |
| 419 | /* |
| 420 | * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in |
| 421 | * a hierarchy). Non-leaf lrqs hold other higher schedulable entities |
| 422 | * (like users, containers etc.) |
| 423 | * |
| 424 | * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This |
| 425 | * list is used during load balance. |
| 426 | */ |
| 427 | struct list_head leaf_cfs_rq_list; |
| 428 | struct task_group *tg; /* group that "owns" this runqueue */ |
| 429 | |
| 430 | #ifdef CONFIG_SMP |
| 431 | /* |
| 432 | * the part of load.weight contributed by tasks |
| 433 | */ |
| 434 | unsigned long task_weight; |
| 435 | |
| 436 | /* |
| 437 | * h_load = weight * f(tg) |
| 438 | * |
| 439 | * Where f(tg) is the recursive weight fraction assigned to |
| 440 | * this group. |
| 441 | */ |
| 442 | unsigned long h_load; |
| 443 | |
| 444 | /* |
| 445 | * this cpu's part of tg->shares |
| 446 | */ |
| 447 | unsigned long shares; |
| 448 | |
| 449 | /* |
| 450 | * load.weight at the time we set shares |
| 451 | */ |
| 452 | unsigned long rq_weight; |
| 453 | #endif |
| 454 | #endif |
| 455 | }; |
| 456 | |
| 457 | /* Real-Time classes' related field in a runqueue: */ |
| 458 | struct rt_rq { |
| 459 | struct rt_prio_array active; |
| 460 | unsigned long rt_nr_running; |
| 461 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
| 462 | struct { |
| 463 | int curr; /* highest queued rt task prio */ |
| 464 | #ifdef CONFIG_SMP |
| 465 | int next; /* next highest */ |
| 466 | #endif |
| 467 | } highest_prio; |
| 468 | #endif |
| 469 | #ifdef CONFIG_SMP |
| 470 | unsigned long rt_nr_migratory; |
| 471 | unsigned long rt_nr_total; |
| 472 | int overloaded; |
| 473 | struct plist_head pushable_tasks; |
| 474 | #endif |
| 475 | int rt_throttled; |
| 476 | u64 rt_time; |
| 477 | u64 rt_runtime; |
| 478 | /* Nests inside the rq lock: */ |
| 479 | spinlock_t rt_runtime_lock; |
| 480 | |
| 481 | #ifdef CONFIG_RT_GROUP_SCHED |
| 482 | unsigned long rt_nr_boosted; |
| 483 | |
| 484 | struct rq *rq; |
| 485 | struct list_head leaf_rt_rq_list; |
| 486 | struct task_group *tg; |
| 487 | struct sched_rt_entity *rt_se; |
| 488 | #endif |
| 489 | }; |
| 490 | |
| 491 | #ifdef CONFIG_SMP |
| 492 | |
| 493 | /* |
| 494 | * We add the notion of a root-domain which will be used to define per-domain |
| 495 | * variables. Each exclusive cpuset essentially defines an island domain by |
| 496 | * fully partitioning the member cpus from any other cpuset. Whenever a new |
| 497 | * exclusive cpuset is created, we also create and attach a new root-domain |
| 498 | * object. |
| 499 | * |
| 500 | */ |
| 501 | struct root_domain { |
| 502 | atomic_t refcount; |
| 503 | cpumask_var_t span; |
| 504 | cpumask_var_t online; |
| 505 | |
| 506 | /* |
| 507 | * The "RT overload" flag: it gets set if a CPU has more than |
| 508 | * one runnable RT task. |
| 509 | */ |
| 510 | cpumask_var_t rto_mask; |
| 511 | atomic_t rto_count; |
| 512 | #ifdef CONFIG_SMP |
| 513 | struct cpupri cpupri; |
| 514 | #endif |
| 515 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
| 516 | /* |
| 517 | * Preferred wake up cpu nominated by sched_mc balance that will be |
| 518 | * used when most cpus are idle in the system indicating overall very |
| 519 | * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2) |
| 520 | */ |
| 521 | unsigned int sched_mc_preferred_wakeup_cpu; |
| 522 | #endif |
| 523 | }; |
| 524 | |
| 525 | /* |
| 526 | * By default the system creates a single root-domain with all cpus as |
| 527 | * members (mimicking the global state we have today). |
| 528 | */ |
| 529 | static struct root_domain def_root_domain; |
| 530 | |
| 531 | #endif |
| 532 | |
| 533 | /* |
| 534 | * This is the main, per-CPU runqueue data structure. |
| 535 | * |
| 536 | * Locking rule: those places that want to lock multiple runqueues |
| 537 | * (such as the load balancing or the thread migration code), lock |
| 538 | * acquire operations must be ordered by ascending &runqueue. |
| 539 | */ |
| 540 | struct rq { |
| 541 | /* runqueue lock: */ |
| 542 | spinlock_t lock; |
| 543 | |
| 544 | /* |
| 545 | * nr_running and cpu_load should be in the same cacheline because |
| 546 | * remote CPUs use both these fields when doing load calculation. |
| 547 | */ |
| 548 | unsigned long nr_running; |
| 549 | #define CPU_LOAD_IDX_MAX 5 |
| 550 | unsigned long cpu_load[CPU_LOAD_IDX_MAX]; |
| 551 | #ifdef CONFIG_NO_HZ |
| 552 | unsigned long last_tick_seen; |
| 553 | unsigned char in_nohz_recently; |
| 554 | #endif |
| 555 | /* capture load from *all* tasks on this cpu: */ |
| 556 | struct load_weight load; |
| 557 | unsigned long nr_load_updates; |
| 558 | u64 nr_switches; |
| 559 | u64 nr_migrations_in; |
| 560 | |
| 561 | struct cfs_rq cfs; |
| 562 | struct rt_rq rt; |
| 563 | |
| 564 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 565 | /* list of leaf cfs_rq on this cpu: */ |
| 566 | struct list_head leaf_cfs_rq_list; |
| 567 | #endif |
| 568 | #ifdef CONFIG_RT_GROUP_SCHED |
| 569 | struct list_head leaf_rt_rq_list; |
| 570 | #endif |
| 571 | |
| 572 | /* |
| 573 | * This is part of a global counter where only the total sum |
| 574 | * over all CPUs matters. A task can increase this counter on |
| 575 | * one CPU and if it got migrated afterwards it may decrease |
| 576 | * it on another CPU. Always updated under the runqueue lock: |
| 577 | */ |
| 578 | unsigned long nr_uninterruptible; |
| 579 | |
| 580 | struct task_struct *curr, *idle; |
| 581 | unsigned long next_balance; |
| 582 | struct mm_struct *prev_mm; |
| 583 | |
| 584 | u64 clock; |
| 585 | |
| 586 | atomic_t nr_iowait; |
| 587 | |
| 588 | #ifdef CONFIG_SMP |
| 589 | struct root_domain *rd; |
| 590 | struct sched_domain *sd; |
| 591 | |
| 592 | unsigned char idle_at_tick; |
| 593 | /* For active balancing */ |
| 594 | int post_schedule; |
| 595 | int active_balance; |
| 596 | int push_cpu; |
| 597 | /* cpu of this runqueue: */ |
| 598 | int cpu; |
| 599 | int online; |
| 600 | |
| 601 | unsigned long avg_load_per_task; |
| 602 | |
| 603 | struct task_struct *migration_thread; |
| 604 | struct list_head migration_queue; |
| 605 | |
| 606 | u64 rt_avg; |
| 607 | u64 age_stamp; |
| 608 | #endif |
| 609 | |
| 610 | /* calc_load related fields */ |
| 611 | unsigned long calc_load_update; |
| 612 | long calc_load_active; |
| 613 | |
| 614 | #ifdef CONFIG_SCHED_HRTICK |
| 615 | #ifdef CONFIG_SMP |
| 616 | int hrtick_csd_pending; |
| 617 | struct call_single_data hrtick_csd; |
| 618 | #endif |
| 619 | struct hrtimer hrtick_timer; |
| 620 | #endif |
| 621 | |
| 622 | #ifdef CONFIG_SCHEDSTATS |
| 623 | /* latency stats */ |
| 624 | struct sched_info rq_sched_info; |
| 625 | unsigned long long rq_cpu_time; |
| 626 | /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ |
| 627 | |
| 628 | /* sys_sched_yield() stats */ |
| 629 | unsigned int yld_count; |
| 630 | |
| 631 | /* schedule() stats */ |
| 632 | unsigned int sched_switch; |
| 633 | unsigned int sched_count; |
| 634 | unsigned int sched_goidle; |
| 635 | |
| 636 | /* try_to_wake_up() stats */ |
| 637 | unsigned int ttwu_count; |
| 638 | unsigned int ttwu_local; |
| 639 | |
| 640 | /* BKL stats */ |
| 641 | unsigned int bkl_count; |
| 642 | #endif |
| 643 | }; |
| 644 | |
| 645 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); |
| 646 | |
| 647 | static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync) |
| 648 | { |
| 649 | rq->curr->sched_class->check_preempt_curr(rq, p, sync); |
| 650 | } |
| 651 | |
| 652 | static inline int cpu_of(struct rq *rq) |
| 653 | { |
| 654 | #ifdef CONFIG_SMP |
| 655 | return rq->cpu; |
| 656 | #else |
| 657 | return 0; |
| 658 | #endif |
| 659 | } |
| 660 | |
| 661 | /* |
| 662 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. |
| 663 | * See detach_destroy_domains: synchronize_sched for details. |
| 664 | * |
| 665 | * The domain tree of any CPU may only be accessed from within |
| 666 | * preempt-disabled sections. |
| 667 | */ |
| 668 | #define for_each_domain(cpu, __sd) \ |
| 669 | for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) |
| 670 | |
| 671 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) |
| 672 | #define this_rq() (&__get_cpu_var(runqueues)) |
| 673 | #define task_rq(p) cpu_rq(task_cpu(p)) |
| 674 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) |
| 675 | #define raw_rq() (&__raw_get_cpu_var(runqueues)) |
| 676 | |
| 677 | inline void update_rq_clock(struct rq *rq) |
| 678 | { |
| 679 | rq->clock = sched_clock_cpu(cpu_of(rq)); |
| 680 | } |
| 681 | |
| 682 | /* |
| 683 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: |
| 684 | */ |
| 685 | #ifdef CONFIG_SCHED_DEBUG |
| 686 | # define const_debug __read_mostly |
| 687 | #else |
| 688 | # define const_debug static const |
| 689 | #endif |
| 690 | |
| 691 | /** |
| 692 | * runqueue_is_locked |
| 693 | * |
| 694 | * Returns true if the current cpu runqueue is locked. |
| 695 | * This interface allows printk to be called with the runqueue lock |
| 696 | * held and know whether or not it is OK to wake up the klogd. |
| 697 | */ |
| 698 | int runqueue_is_locked(void) |
| 699 | { |
| 700 | int cpu = get_cpu(); |
| 701 | struct rq *rq = cpu_rq(cpu); |
| 702 | int ret; |
| 703 | |
| 704 | ret = spin_is_locked(&rq->lock); |
| 705 | put_cpu(); |
| 706 | return ret; |
| 707 | } |
| 708 | |
| 709 | /* |
| 710 | * Debugging: various feature bits |
| 711 | */ |
| 712 | |
| 713 | #define SCHED_FEAT(name, enabled) \ |
| 714 | __SCHED_FEAT_##name , |
| 715 | |
| 716 | enum { |
| 717 | #include "sched_features.h" |
| 718 | }; |
| 719 | |
| 720 | #undef SCHED_FEAT |
| 721 | |
| 722 | #define SCHED_FEAT(name, enabled) \ |
| 723 | (1UL << __SCHED_FEAT_##name) * enabled | |
| 724 | |
| 725 | const_debug unsigned int sysctl_sched_features = |
| 726 | #include "sched_features.h" |
| 727 | 0; |
| 728 | |
| 729 | #undef SCHED_FEAT |
| 730 | |
| 731 | #ifdef CONFIG_SCHED_DEBUG |
| 732 | #define SCHED_FEAT(name, enabled) \ |
| 733 | #name , |
| 734 | |
| 735 | static __read_mostly char *sched_feat_names[] = { |
| 736 | #include "sched_features.h" |
| 737 | NULL |
| 738 | }; |
| 739 | |
| 740 | #undef SCHED_FEAT |
| 741 | |
| 742 | static int sched_feat_show(struct seq_file *m, void *v) |
| 743 | { |
| 744 | int i; |
| 745 | |
| 746 | for (i = 0; sched_feat_names[i]; i++) { |
| 747 | if (!(sysctl_sched_features & (1UL << i))) |
| 748 | seq_puts(m, "NO_"); |
| 749 | seq_printf(m, "%s ", sched_feat_names[i]); |
| 750 | } |
| 751 | seq_puts(m, "\n"); |
| 752 | |
| 753 | return 0; |
| 754 | } |
| 755 | |
| 756 | static ssize_t |
| 757 | sched_feat_write(struct file *filp, const char __user *ubuf, |
| 758 | size_t cnt, loff_t *ppos) |
| 759 | { |
| 760 | char buf[64]; |
| 761 | char *cmp = buf; |
| 762 | int neg = 0; |
| 763 | int i; |
| 764 | |
| 765 | if (cnt > 63) |
| 766 | cnt = 63; |
| 767 | |
| 768 | if (copy_from_user(&buf, ubuf, cnt)) |
| 769 | return -EFAULT; |
| 770 | |
| 771 | buf[cnt] = 0; |
| 772 | |
| 773 | if (strncmp(buf, "NO_", 3) == 0) { |
| 774 | neg = 1; |
| 775 | cmp += 3; |
| 776 | } |
| 777 | |
| 778 | for (i = 0; sched_feat_names[i]; i++) { |
| 779 | int len = strlen(sched_feat_names[i]); |
| 780 | |
| 781 | if (strncmp(cmp, sched_feat_names[i], len) == 0) { |
| 782 | if (neg) |
| 783 | sysctl_sched_features &= ~(1UL << i); |
| 784 | else |
| 785 | sysctl_sched_features |= (1UL << i); |
| 786 | break; |
| 787 | } |
| 788 | } |
| 789 | |
| 790 | if (!sched_feat_names[i]) |
| 791 | return -EINVAL; |
| 792 | |
| 793 | filp->f_pos += cnt; |
| 794 | |
| 795 | return cnt; |
| 796 | } |
| 797 | |
| 798 | static int sched_feat_open(struct inode *inode, struct file *filp) |
| 799 | { |
| 800 | return single_open(filp, sched_feat_show, NULL); |
| 801 | } |
| 802 | |
| 803 | static struct file_operations sched_feat_fops = { |
| 804 | .open = sched_feat_open, |
| 805 | .write = sched_feat_write, |
| 806 | .read = seq_read, |
| 807 | .llseek = seq_lseek, |
| 808 | .release = single_release, |
| 809 | }; |
| 810 | |
| 811 | static __init int sched_init_debug(void) |
| 812 | { |
| 813 | debugfs_create_file("sched_features", 0644, NULL, NULL, |
| 814 | &sched_feat_fops); |
| 815 | |
| 816 | return 0; |
| 817 | } |
| 818 | late_initcall(sched_init_debug); |
| 819 | |
| 820 | #endif |
| 821 | |
| 822 | #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) |
| 823 | |
| 824 | /* |
| 825 | * Number of tasks to iterate in a single balance run. |
| 826 | * Limited because this is done with IRQs disabled. |
| 827 | */ |
| 828 | const_debug unsigned int sysctl_sched_nr_migrate = 32; |
| 829 | |
| 830 | /* |
| 831 | * ratelimit for updating the group shares. |
| 832 | * default: 0.25ms |
| 833 | */ |
| 834 | unsigned int sysctl_sched_shares_ratelimit = 250000; |
| 835 | |
| 836 | /* |
| 837 | * Inject some fuzzyness into changing the per-cpu group shares |
| 838 | * this avoids remote rq-locks at the expense of fairness. |
| 839 | * default: 4 |
| 840 | */ |
| 841 | unsigned int sysctl_sched_shares_thresh = 4; |
| 842 | |
| 843 | /* |
| 844 | * period over which we average the RT time consumption, measured |
| 845 | * in ms. |
| 846 | * |
| 847 | * default: 1s |
| 848 | */ |
| 849 | const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; |
| 850 | |
| 851 | /* |
| 852 | * period over which we measure -rt task cpu usage in us. |
| 853 | * default: 1s |
| 854 | */ |
| 855 | unsigned int sysctl_sched_rt_period = 1000000; |
| 856 | |
| 857 | static __read_mostly int scheduler_running; |
| 858 | |
| 859 | /* |
| 860 | * part of the period that we allow rt tasks to run in us. |
| 861 | * default: 0.95s |
| 862 | */ |
| 863 | int sysctl_sched_rt_runtime = 950000; |
| 864 | |
| 865 | static inline u64 global_rt_period(void) |
| 866 | { |
| 867 | return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; |
| 868 | } |
| 869 | |
| 870 | static inline u64 global_rt_runtime(void) |
| 871 | { |
| 872 | if (sysctl_sched_rt_runtime < 0) |
| 873 | return RUNTIME_INF; |
| 874 | |
| 875 | return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; |
| 876 | } |
| 877 | |
| 878 | #ifndef prepare_arch_switch |
| 879 | # define prepare_arch_switch(next) do { } while (0) |
| 880 | #endif |
| 881 | #ifndef finish_arch_switch |
| 882 | # define finish_arch_switch(prev) do { } while (0) |
| 883 | #endif |
| 884 | |
| 885 | static inline int task_current(struct rq *rq, struct task_struct *p) |
| 886 | { |
| 887 | return rq->curr == p; |
| 888 | } |
| 889 | |
| 890 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
| 891 | static inline int task_running(struct rq *rq, struct task_struct *p) |
| 892 | { |
| 893 | return task_current(rq, p); |
| 894 | } |
| 895 | |
| 896 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
| 897 | { |
| 898 | } |
| 899 | |
| 900 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
| 901 | { |
| 902 | #ifdef CONFIG_DEBUG_SPINLOCK |
| 903 | /* this is a valid case when another task releases the spinlock */ |
| 904 | rq->lock.owner = current; |
| 905 | #endif |
| 906 | /* |
| 907 | * If we are tracking spinlock dependencies then we have to |
| 908 | * fix up the runqueue lock - which gets 'carried over' from |
| 909 | * prev into current: |
| 910 | */ |
| 911 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); |
| 912 | |
| 913 | spin_unlock_irq(&rq->lock); |
| 914 | } |
| 915 | |
| 916 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ |
| 917 | static inline int task_running(struct rq *rq, struct task_struct *p) |
| 918 | { |
| 919 | #ifdef CONFIG_SMP |
| 920 | return p->oncpu; |
| 921 | #else |
| 922 | return task_current(rq, p); |
| 923 | #endif |
| 924 | } |
| 925 | |
| 926 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
| 927 | { |
| 928 | #ifdef CONFIG_SMP |
| 929 | /* |
| 930 | * We can optimise this out completely for !SMP, because the |
| 931 | * SMP rebalancing from interrupt is the only thing that cares |
| 932 | * here. |
| 933 | */ |
| 934 | next->oncpu = 1; |
| 935 | #endif |
| 936 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
| 937 | spin_unlock_irq(&rq->lock); |
| 938 | #else |
| 939 | spin_unlock(&rq->lock); |
| 940 | #endif |
| 941 | } |
| 942 | |
| 943 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
| 944 | { |
| 945 | #ifdef CONFIG_SMP |
| 946 | /* |
| 947 | * After ->oncpu is cleared, the task can be moved to a different CPU. |
| 948 | * We must ensure this doesn't happen until the switch is completely |
| 949 | * finished. |
| 950 | */ |
| 951 | smp_wmb(); |
| 952 | prev->oncpu = 0; |
| 953 | #endif |
| 954 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
| 955 | local_irq_enable(); |
| 956 | #endif |
| 957 | } |
| 958 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ |
| 959 | |
| 960 | /* |
| 961 | * __task_rq_lock - lock the runqueue a given task resides on. |
| 962 | * Must be called interrupts disabled. |
| 963 | */ |
| 964 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
| 965 | __acquires(rq->lock) |
| 966 | { |
| 967 | for (;;) { |
| 968 | struct rq *rq = task_rq(p); |
| 969 | spin_lock(&rq->lock); |
| 970 | if (likely(rq == task_rq(p))) |
| 971 | return rq; |
| 972 | spin_unlock(&rq->lock); |
| 973 | } |
| 974 | } |
| 975 | |
| 976 | /* |
| 977 | * task_rq_lock - lock the runqueue a given task resides on and disable |
| 978 | * interrupts. Note the ordering: we can safely lookup the task_rq without |
| 979 | * explicitly disabling preemption. |
| 980 | */ |
| 981 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
| 982 | __acquires(rq->lock) |
| 983 | { |
| 984 | struct rq *rq; |
| 985 | |
| 986 | for (;;) { |
| 987 | local_irq_save(*flags); |
| 988 | rq = task_rq(p); |
| 989 | spin_lock(&rq->lock); |
| 990 | if (likely(rq == task_rq(p))) |
| 991 | return rq; |
| 992 | spin_unlock_irqrestore(&rq->lock, *flags); |
| 993 | } |
| 994 | } |
| 995 | |
| 996 | void task_rq_unlock_wait(struct task_struct *p) |
| 997 | { |
| 998 | struct rq *rq = task_rq(p); |
| 999 | |
| 1000 | smp_mb(); /* spin-unlock-wait is not a full memory barrier */ |
| 1001 | spin_unlock_wait(&rq->lock); |
| 1002 | } |
| 1003 | |
| 1004 | static void __task_rq_unlock(struct rq *rq) |
| 1005 | __releases(rq->lock) |
| 1006 | { |
| 1007 | spin_unlock(&rq->lock); |
| 1008 | } |
| 1009 | |
| 1010 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) |
| 1011 | __releases(rq->lock) |
| 1012 | { |
| 1013 | spin_unlock_irqrestore(&rq->lock, *flags); |
| 1014 | } |
| 1015 | |
| 1016 | /* |
| 1017 | * this_rq_lock - lock this runqueue and disable interrupts. |
| 1018 | */ |
| 1019 | static struct rq *this_rq_lock(void) |
| 1020 | __acquires(rq->lock) |
| 1021 | { |
| 1022 | struct rq *rq; |
| 1023 | |
| 1024 | local_irq_disable(); |
| 1025 | rq = this_rq(); |
| 1026 | spin_lock(&rq->lock); |
| 1027 | |
| 1028 | return rq; |
| 1029 | } |
| 1030 | |
| 1031 | #ifdef CONFIG_SCHED_HRTICK |
| 1032 | /* |
| 1033 | * Use HR-timers to deliver accurate preemption points. |
| 1034 | * |
| 1035 | * Its all a bit involved since we cannot program an hrt while holding the |
| 1036 | * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a |
| 1037 | * reschedule event. |
| 1038 | * |
| 1039 | * When we get rescheduled we reprogram the hrtick_timer outside of the |
| 1040 | * rq->lock. |
| 1041 | */ |
| 1042 | |
| 1043 | /* |
| 1044 | * Use hrtick when: |
| 1045 | * - enabled by features |
| 1046 | * - hrtimer is actually high res |
| 1047 | */ |
| 1048 | static inline int hrtick_enabled(struct rq *rq) |
| 1049 | { |
| 1050 | if (!sched_feat(HRTICK)) |
| 1051 | return 0; |
| 1052 | if (!cpu_active(cpu_of(rq))) |
| 1053 | return 0; |
| 1054 | return hrtimer_is_hres_active(&rq->hrtick_timer); |
| 1055 | } |
| 1056 | |
| 1057 | static void hrtick_clear(struct rq *rq) |
| 1058 | { |
| 1059 | if (hrtimer_active(&rq->hrtick_timer)) |
| 1060 | hrtimer_cancel(&rq->hrtick_timer); |
| 1061 | } |
| 1062 | |
| 1063 | /* |
| 1064 | * High-resolution timer tick. |
| 1065 | * Runs from hardirq context with interrupts disabled. |
| 1066 | */ |
| 1067 | static enum hrtimer_restart hrtick(struct hrtimer *timer) |
| 1068 | { |
| 1069 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); |
| 1070 | |
| 1071 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); |
| 1072 | |
| 1073 | spin_lock(&rq->lock); |
| 1074 | update_rq_clock(rq); |
| 1075 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); |
| 1076 | spin_unlock(&rq->lock); |
| 1077 | |
| 1078 | return HRTIMER_NORESTART; |
| 1079 | } |
| 1080 | |
| 1081 | #ifdef CONFIG_SMP |
| 1082 | /* |
| 1083 | * called from hardirq (IPI) context |
| 1084 | */ |
| 1085 | static void __hrtick_start(void *arg) |
| 1086 | { |
| 1087 | struct rq *rq = arg; |
| 1088 | |
| 1089 | spin_lock(&rq->lock); |
| 1090 | hrtimer_restart(&rq->hrtick_timer); |
| 1091 | rq->hrtick_csd_pending = 0; |
| 1092 | spin_unlock(&rq->lock); |
| 1093 | } |
| 1094 | |
| 1095 | /* |
| 1096 | * Called to set the hrtick timer state. |
| 1097 | * |
| 1098 | * called with rq->lock held and irqs disabled |
| 1099 | */ |
| 1100 | static void hrtick_start(struct rq *rq, u64 delay) |
| 1101 | { |
| 1102 | struct hrtimer *timer = &rq->hrtick_timer; |
| 1103 | ktime_t time = ktime_add_ns(timer->base->get_time(), delay); |
| 1104 | |
| 1105 | hrtimer_set_expires(timer, time); |
| 1106 | |
| 1107 | if (rq == this_rq()) { |
| 1108 | hrtimer_restart(timer); |
| 1109 | } else if (!rq->hrtick_csd_pending) { |
| 1110 | __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); |
| 1111 | rq->hrtick_csd_pending = 1; |
| 1112 | } |
| 1113 | } |
| 1114 | |
| 1115 | static int |
| 1116 | hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) |
| 1117 | { |
| 1118 | int cpu = (int)(long)hcpu; |
| 1119 | |
| 1120 | switch (action) { |
| 1121 | case CPU_UP_CANCELED: |
| 1122 | case CPU_UP_CANCELED_FROZEN: |
| 1123 | case CPU_DOWN_PREPARE: |
| 1124 | case CPU_DOWN_PREPARE_FROZEN: |
| 1125 | case CPU_DEAD: |
| 1126 | case CPU_DEAD_FROZEN: |
| 1127 | hrtick_clear(cpu_rq(cpu)); |
| 1128 | return NOTIFY_OK; |
| 1129 | } |
| 1130 | |
| 1131 | return NOTIFY_DONE; |
| 1132 | } |
| 1133 | |
| 1134 | static __init void init_hrtick(void) |
| 1135 | { |
| 1136 | hotcpu_notifier(hotplug_hrtick, 0); |
| 1137 | } |
| 1138 | #else |
| 1139 | /* |
| 1140 | * Called to set the hrtick timer state. |
| 1141 | * |
| 1142 | * called with rq->lock held and irqs disabled |
| 1143 | */ |
| 1144 | static void hrtick_start(struct rq *rq, u64 delay) |
| 1145 | { |
| 1146 | __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, |
| 1147 | HRTIMER_MODE_REL_PINNED, 0); |
| 1148 | } |
| 1149 | |
| 1150 | static inline void init_hrtick(void) |
| 1151 | { |
| 1152 | } |
| 1153 | #endif /* CONFIG_SMP */ |
| 1154 | |
| 1155 | static void init_rq_hrtick(struct rq *rq) |
| 1156 | { |
| 1157 | #ifdef CONFIG_SMP |
| 1158 | rq->hrtick_csd_pending = 0; |
| 1159 | |
| 1160 | rq->hrtick_csd.flags = 0; |
| 1161 | rq->hrtick_csd.func = __hrtick_start; |
| 1162 | rq->hrtick_csd.info = rq; |
| 1163 | #endif |
| 1164 | |
| 1165 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
| 1166 | rq->hrtick_timer.function = hrtick; |
| 1167 | } |
| 1168 | #else /* CONFIG_SCHED_HRTICK */ |
| 1169 | static inline void hrtick_clear(struct rq *rq) |
| 1170 | { |
| 1171 | } |
| 1172 | |
| 1173 | static inline void init_rq_hrtick(struct rq *rq) |
| 1174 | { |
| 1175 | } |
| 1176 | |
| 1177 | static inline void init_hrtick(void) |
| 1178 | { |
| 1179 | } |
| 1180 | #endif /* CONFIG_SCHED_HRTICK */ |
| 1181 | |
| 1182 | /* |
| 1183 | * resched_task - mark a task 'to be rescheduled now'. |
| 1184 | * |
| 1185 | * On UP this means the setting of the need_resched flag, on SMP it |
| 1186 | * might also involve a cross-CPU call to trigger the scheduler on |
| 1187 | * the target CPU. |
| 1188 | */ |
| 1189 | #ifdef CONFIG_SMP |
| 1190 | |
| 1191 | #ifndef tsk_is_polling |
| 1192 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) |
| 1193 | #endif |
| 1194 | |
| 1195 | static void resched_task(struct task_struct *p) |
| 1196 | { |
| 1197 | int cpu; |
| 1198 | |
| 1199 | assert_spin_locked(&task_rq(p)->lock); |
| 1200 | |
| 1201 | if (test_tsk_need_resched(p)) |
| 1202 | return; |
| 1203 | |
| 1204 | set_tsk_need_resched(p); |
| 1205 | |
| 1206 | cpu = task_cpu(p); |
| 1207 | if (cpu == smp_processor_id()) |
| 1208 | return; |
| 1209 | |
| 1210 | /* NEED_RESCHED must be visible before we test polling */ |
| 1211 | smp_mb(); |
| 1212 | if (!tsk_is_polling(p)) |
| 1213 | smp_send_reschedule(cpu); |
| 1214 | } |
| 1215 | |
| 1216 | static void resched_cpu(int cpu) |
| 1217 | { |
| 1218 | struct rq *rq = cpu_rq(cpu); |
| 1219 | unsigned long flags; |
| 1220 | |
| 1221 | if (!spin_trylock_irqsave(&rq->lock, flags)) |
| 1222 | return; |
| 1223 | resched_task(cpu_curr(cpu)); |
| 1224 | spin_unlock_irqrestore(&rq->lock, flags); |
| 1225 | } |
| 1226 | |
| 1227 | #ifdef CONFIG_NO_HZ |
| 1228 | /* |
| 1229 | * When add_timer_on() enqueues a timer into the timer wheel of an |
| 1230 | * idle CPU then this timer might expire before the next timer event |
| 1231 | * which is scheduled to wake up that CPU. In case of a completely |
| 1232 | * idle system the next event might even be infinite time into the |
| 1233 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and |
| 1234 | * leaves the inner idle loop so the newly added timer is taken into |
| 1235 | * account when the CPU goes back to idle and evaluates the timer |
| 1236 | * wheel for the next timer event. |
| 1237 | */ |
| 1238 | void wake_up_idle_cpu(int cpu) |
| 1239 | { |
| 1240 | struct rq *rq = cpu_rq(cpu); |
| 1241 | |
| 1242 | if (cpu == smp_processor_id()) |
| 1243 | return; |
| 1244 | |
| 1245 | /* |
| 1246 | * This is safe, as this function is called with the timer |
| 1247 | * wheel base lock of (cpu) held. When the CPU is on the way |
| 1248 | * to idle and has not yet set rq->curr to idle then it will |
| 1249 | * be serialized on the timer wheel base lock and take the new |
| 1250 | * timer into account automatically. |
| 1251 | */ |
| 1252 | if (rq->curr != rq->idle) |
| 1253 | return; |
| 1254 | |
| 1255 | /* |
| 1256 | * We can set TIF_RESCHED on the idle task of the other CPU |
| 1257 | * lockless. The worst case is that the other CPU runs the |
| 1258 | * idle task through an additional NOOP schedule() |
| 1259 | */ |
| 1260 | set_tsk_need_resched(rq->idle); |
| 1261 | |
| 1262 | /* NEED_RESCHED must be visible before we test polling */ |
| 1263 | smp_mb(); |
| 1264 | if (!tsk_is_polling(rq->idle)) |
| 1265 | smp_send_reschedule(cpu); |
| 1266 | } |
| 1267 | #endif /* CONFIG_NO_HZ */ |
| 1268 | |
| 1269 | static u64 sched_avg_period(void) |
| 1270 | { |
| 1271 | return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; |
| 1272 | } |
| 1273 | |
| 1274 | static void sched_avg_update(struct rq *rq) |
| 1275 | { |
| 1276 | s64 period = sched_avg_period(); |
| 1277 | |
| 1278 | while ((s64)(rq->clock - rq->age_stamp) > period) { |
| 1279 | rq->age_stamp += period; |
| 1280 | rq->rt_avg /= 2; |
| 1281 | } |
| 1282 | } |
| 1283 | |
| 1284 | static void sched_rt_avg_update(struct rq *rq, u64 rt_delta) |
| 1285 | { |
| 1286 | rq->rt_avg += rt_delta; |
| 1287 | sched_avg_update(rq); |
| 1288 | } |
| 1289 | |
| 1290 | #else /* !CONFIG_SMP */ |
| 1291 | static void resched_task(struct task_struct *p) |
| 1292 | { |
| 1293 | assert_spin_locked(&task_rq(p)->lock); |
| 1294 | set_tsk_need_resched(p); |
| 1295 | } |
| 1296 | |
| 1297 | static void sched_rt_avg_update(struct rq *rq, u64 rt_delta) |
| 1298 | { |
| 1299 | } |
| 1300 | #endif /* CONFIG_SMP */ |
| 1301 | |
| 1302 | #if BITS_PER_LONG == 32 |
| 1303 | # define WMULT_CONST (~0UL) |
| 1304 | #else |
| 1305 | # define WMULT_CONST (1UL << 32) |
| 1306 | #endif |
| 1307 | |
| 1308 | #define WMULT_SHIFT 32 |
| 1309 | |
| 1310 | /* |
| 1311 | * Shift right and round: |
| 1312 | */ |
| 1313 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) |
| 1314 | |
| 1315 | /* |
| 1316 | * delta *= weight / lw |
| 1317 | */ |
| 1318 | static unsigned long |
| 1319 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, |
| 1320 | struct load_weight *lw) |
| 1321 | { |
| 1322 | u64 tmp; |
| 1323 | |
| 1324 | if (!lw->inv_weight) { |
| 1325 | if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST)) |
| 1326 | lw->inv_weight = 1; |
| 1327 | else |
| 1328 | lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2) |
| 1329 | / (lw->weight+1); |
| 1330 | } |
| 1331 | |
| 1332 | tmp = (u64)delta_exec * weight; |
| 1333 | /* |
| 1334 | * Check whether we'd overflow the 64-bit multiplication: |
| 1335 | */ |
| 1336 | if (unlikely(tmp > WMULT_CONST)) |
| 1337 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, |
| 1338 | WMULT_SHIFT/2); |
| 1339 | else |
| 1340 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); |
| 1341 | |
| 1342 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); |
| 1343 | } |
| 1344 | |
| 1345 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
| 1346 | { |
| 1347 | lw->weight += inc; |
| 1348 | lw->inv_weight = 0; |
| 1349 | } |
| 1350 | |
| 1351 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) |
| 1352 | { |
| 1353 | lw->weight -= dec; |
| 1354 | lw->inv_weight = 0; |
| 1355 | } |
| 1356 | |
| 1357 | /* |
| 1358 | * To aid in avoiding the subversion of "niceness" due to uneven distribution |
| 1359 | * of tasks with abnormal "nice" values across CPUs the contribution that |
| 1360 | * each task makes to its run queue's load is weighted according to its |
| 1361 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a |
| 1362 | * scaled version of the new time slice allocation that they receive on time |
| 1363 | * slice expiry etc. |
| 1364 | */ |
| 1365 | |
| 1366 | #define WEIGHT_IDLEPRIO 3 |
| 1367 | #define WMULT_IDLEPRIO 1431655765 |
| 1368 | |
| 1369 | /* |
| 1370 | * Nice levels are multiplicative, with a gentle 10% change for every |
| 1371 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to |
| 1372 | * nice 1, it will get ~10% less CPU time than another CPU-bound task |
| 1373 | * that remained on nice 0. |
| 1374 | * |
| 1375 | * The "10% effect" is relative and cumulative: from _any_ nice level, |
| 1376 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level |
| 1377 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. |
| 1378 | * If a task goes up by ~10% and another task goes down by ~10% then |
| 1379 | * the relative distance between them is ~25%.) |
| 1380 | */ |
| 1381 | static const int prio_to_weight[40] = { |
| 1382 | /* -20 */ 88761, 71755, 56483, 46273, 36291, |
| 1383 | /* -15 */ 29154, 23254, 18705, 14949, 11916, |
| 1384 | /* -10 */ 9548, 7620, 6100, 4904, 3906, |
| 1385 | /* -5 */ 3121, 2501, 1991, 1586, 1277, |
| 1386 | /* 0 */ 1024, 820, 655, 526, 423, |
| 1387 | /* 5 */ 335, 272, 215, 172, 137, |
| 1388 | /* 10 */ 110, 87, 70, 56, 45, |
| 1389 | /* 15 */ 36, 29, 23, 18, 15, |
| 1390 | }; |
| 1391 | |
| 1392 | /* |
| 1393 | * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. |
| 1394 | * |
| 1395 | * In cases where the weight does not change often, we can use the |
| 1396 | * precalculated inverse to speed up arithmetics by turning divisions |
| 1397 | * into multiplications: |
| 1398 | */ |
| 1399 | static const u32 prio_to_wmult[40] = { |
| 1400 | /* -20 */ 48388, 59856, 76040, 92818, 118348, |
| 1401 | /* -15 */ 147320, 184698, 229616, 287308, 360437, |
| 1402 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, |
| 1403 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, |
| 1404 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, |
| 1405 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, |
| 1406 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, |
| 1407 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, |
| 1408 | }; |
| 1409 | |
| 1410 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup); |
| 1411 | |
| 1412 | /* |
| 1413 | * runqueue iterator, to support SMP load-balancing between different |
| 1414 | * scheduling classes, without having to expose their internal data |
| 1415 | * structures to the load-balancing proper: |
| 1416 | */ |
| 1417 | struct rq_iterator { |
| 1418 | void *arg; |
| 1419 | struct task_struct *(*start)(void *); |
| 1420 | struct task_struct *(*next)(void *); |
| 1421 | }; |
| 1422 | |
| 1423 | #ifdef CONFIG_SMP |
| 1424 | static unsigned long |
| 1425 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 1426 | unsigned long max_load_move, struct sched_domain *sd, |
| 1427 | enum cpu_idle_type idle, int *all_pinned, |
| 1428 | int *this_best_prio, struct rq_iterator *iterator); |
| 1429 | |
| 1430 | static int |
| 1431 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 1432 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 1433 | struct rq_iterator *iterator); |
| 1434 | #endif |
| 1435 | |
| 1436 | /* Time spent by the tasks of the cpu accounting group executing in ... */ |
| 1437 | enum cpuacct_stat_index { |
| 1438 | CPUACCT_STAT_USER, /* ... user mode */ |
| 1439 | CPUACCT_STAT_SYSTEM, /* ... kernel mode */ |
| 1440 | |
| 1441 | CPUACCT_STAT_NSTATS, |
| 1442 | }; |
| 1443 | |
| 1444 | #ifdef CONFIG_CGROUP_CPUACCT |
| 1445 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime); |
| 1446 | static void cpuacct_update_stats(struct task_struct *tsk, |
| 1447 | enum cpuacct_stat_index idx, cputime_t val); |
| 1448 | #else |
| 1449 | static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} |
| 1450 | static inline void cpuacct_update_stats(struct task_struct *tsk, |
| 1451 | enum cpuacct_stat_index idx, cputime_t val) {} |
| 1452 | #endif |
| 1453 | |
| 1454 | static inline void inc_cpu_load(struct rq *rq, unsigned long load) |
| 1455 | { |
| 1456 | update_load_add(&rq->load, load); |
| 1457 | } |
| 1458 | |
| 1459 | static inline void dec_cpu_load(struct rq *rq, unsigned long load) |
| 1460 | { |
| 1461 | update_load_sub(&rq->load, load); |
| 1462 | } |
| 1463 | |
| 1464 | #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED) |
| 1465 | typedef int (*tg_visitor)(struct task_group *, void *); |
| 1466 | |
| 1467 | /* |
| 1468 | * Iterate the full tree, calling @down when first entering a node and @up when |
| 1469 | * leaving it for the final time. |
| 1470 | */ |
| 1471 | static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) |
| 1472 | { |
| 1473 | struct task_group *parent, *child; |
| 1474 | int ret; |
| 1475 | |
| 1476 | rcu_read_lock(); |
| 1477 | parent = &root_task_group; |
| 1478 | down: |
| 1479 | ret = (*down)(parent, data); |
| 1480 | if (ret) |
| 1481 | goto out_unlock; |
| 1482 | list_for_each_entry_rcu(child, &parent->children, siblings) { |
| 1483 | parent = child; |
| 1484 | goto down; |
| 1485 | |
| 1486 | up: |
| 1487 | continue; |
| 1488 | } |
| 1489 | ret = (*up)(parent, data); |
| 1490 | if (ret) |
| 1491 | goto out_unlock; |
| 1492 | |
| 1493 | child = parent; |
| 1494 | parent = parent->parent; |
| 1495 | if (parent) |
| 1496 | goto up; |
| 1497 | out_unlock: |
| 1498 | rcu_read_unlock(); |
| 1499 | |
| 1500 | return ret; |
| 1501 | } |
| 1502 | |
| 1503 | static int tg_nop(struct task_group *tg, void *data) |
| 1504 | { |
| 1505 | return 0; |
| 1506 | } |
| 1507 | #endif |
| 1508 | |
| 1509 | #ifdef CONFIG_SMP |
| 1510 | /* Used instead of source_load when we know the type == 0 */ |
| 1511 | static unsigned long weighted_cpuload(const int cpu) |
| 1512 | { |
| 1513 | return cpu_rq(cpu)->load.weight; |
| 1514 | } |
| 1515 | |
| 1516 | /* |
| 1517 | * Return a low guess at the load of a migration-source cpu weighted |
| 1518 | * according to the scheduling class and "nice" value. |
| 1519 | * |
| 1520 | * We want to under-estimate the load of migration sources, to |
| 1521 | * balance conservatively. |
| 1522 | */ |
| 1523 | static unsigned long source_load(int cpu, int type) |
| 1524 | { |
| 1525 | struct rq *rq = cpu_rq(cpu); |
| 1526 | unsigned long total = weighted_cpuload(cpu); |
| 1527 | |
| 1528 | if (type == 0 || !sched_feat(LB_BIAS)) |
| 1529 | return total; |
| 1530 | |
| 1531 | return min(rq->cpu_load[type-1], total); |
| 1532 | } |
| 1533 | |
| 1534 | /* |
| 1535 | * Return a high guess at the load of a migration-target cpu weighted |
| 1536 | * according to the scheduling class and "nice" value. |
| 1537 | */ |
| 1538 | static unsigned long target_load(int cpu, int type) |
| 1539 | { |
| 1540 | struct rq *rq = cpu_rq(cpu); |
| 1541 | unsigned long total = weighted_cpuload(cpu); |
| 1542 | |
| 1543 | if (type == 0 || !sched_feat(LB_BIAS)) |
| 1544 | return total; |
| 1545 | |
| 1546 | return max(rq->cpu_load[type-1], total); |
| 1547 | } |
| 1548 | |
| 1549 | static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd); |
| 1550 | |
| 1551 | static unsigned long cpu_avg_load_per_task(int cpu) |
| 1552 | { |
| 1553 | struct rq *rq = cpu_rq(cpu); |
| 1554 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); |
| 1555 | |
| 1556 | if (nr_running) |
| 1557 | rq->avg_load_per_task = rq->load.weight / nr_running; |
| 1558 | else |
| 1559 | rq->avg_load_per_task = 0; |
| 1560 | |
| 1561 | return rq->avg_load_per_task; |
| 1562 | } |
| 1563 | |
| 1564 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1565 | |
| 1566 | struct update_shares_data { |
| 1567 | unsigned long rq_weight[NR_CPUS]; |
| 1568 | }; |
| 1569 | |
| 1570 | static DEFINE_PER_CPU(struct update_shares_data, update_shares_data); |
| 1571 | |
| 1572 | static void __set_se_shares(struct sched_entity *se, unsigned long shares); |
| 1573 | |
| 1574 | /* |
| 1575 | * Calculate and set the cpu's group shares. |
| 1576 | */ |
| 1577 | static void update_group_shares_cpu(struct task_group *tg, int cpu, |
| 1578 | unsigned long sd_shares, |
| 1579 | unsigned long sd_rq_weight, |
| 1580 | struct update_shares_data *usd) |
| 1581 | { |
| 1582 | unsigned long shares, rq_weight; |
| 1583 | int boost = 0; |
| 1584 | |
| 1585 | rq_weight = usd->rq_weight[cpu]; |
| 1586 | if (!rq_weight) { |
| 1587 | boost = 1; |
| 1588 | rq_weight = NICE_0_LOAD; |
| 1589 | } |
| 1590 | |
| 1591 | /* |
| 1592 | * \Sum_j shares_j * rq_weight_i |
| 1593 | * shares_i = ----------------------------- |
| 1594 | * \Sum_j rq_weight_j |
| 1595 | */ |
| 1596 | shares = (sd_shares * rq_weight) / sd_rq_weight; |
| 1597 | shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES); |
| 1598 | |
| 1599 | if (abs(shares - tg->se[cpu]->load.weight) > |
| 1600 | sysctl_sched_shares_thresh) { |
| 1601 | struct rq *rq = cpu_rq(cpu); |
| 1602 | unsigned long flags; |
| 1603 | |
| 1604 | spin_lock_irqsave(&rq->lock, flags); |
| 1605 | tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight; |
| 1606 | tg->cfs_rq[cpu]->shares = boost ? 0 : shares; |
| 1607 | __set_se_shares(tg->se[cpu], shares); |
| 1608 | spin_unlock_irqrestore(&rq->lock, flags); |
| 1609 | } |
| 1610 | } |
| 1611 | |
| 1612 | /* |
| 1613 | * Re-compute the task group their per cpu shares over the given domain. |
| 1614 | * This needs to be done in a bottom-up fashion because the rq weight of a |
| 1615 | * parent group depends on the shares of its child groups. |
| 1616 | */ |
| 1617 | static int tg_shares_up(struct task_group *tg, void *data) |
| 1618 | { |
| 1619 | unsigned long weight, rq_weight = 0, shares = 0; |
| 1620 | struct update_shares_data *usd; |
| 1621 | struct sched_domain *sd = data; |
| 1622 | unsigned long flags; |
| 1623 | int i; |
| 1624 | |
| 1625 | if (!tg->se[0]) |
| 1626 | return 0; |
| 1627 | |
| 1628 | local_irq_save(flags); |
| 1629 | usd = &__get_cpu_var(update_shares_data); |
| 1630 | |
| 1631 | for_each_cpu(i, sched_domain_span(sd)) { |
| 1632 | weight = tg->cfs_rq[i]->load.weight; |
| 1633 | usd->rq_weight[i] = weight; |
| 1634 | |
| 1635 | /* |
| 1636 | * If there are currently no tasks on the cpu pretend there |
| 1637 | * is one of average load so that when a new task gets to |
| 1638 | * run here it will not get delayed by group starvation. |
| 1639 | */ |
| 1640 | if (!weight) |
| 1641 | weight = NICE_0_LOAD; |
| 1642 | |
| 1643 | rq_weight += weight; |
| 1644 | shares += tg->cfs_rq[i]->shares; |
| 1645 | } |
| 1646 | |
| 1647 | if ((!shares && rq_weight) || shares > tg->shares) |
| 1648 | shares = tg->shares; |
| 1649 | |
| 1650 | if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE)) |
| 1651 | shares = tg->shares; |
| 1652 | |
| 1653 | for_each_cpu(i, sched_domain_span(sd)) |
| 1654 | update_group_shares_cpu(tg, i, shares, rq_weight, usd); |
| 1655 | |
| 1656 | local_irq_restore(flags); |
| 1657 | |
| 1658 | return 0; |
| 1659 | } |
| 1660 | |
| 1661 | /* |
| 1662 | * Compute the cpu's hierarchical load factor for each task group. |
| 1663 | * This needs to be done in a top-down fashion because the load of a child |
| 1664 | * group is a fraction of its parents load. |
| 1665 | */ |
| 1666 | static int tg_load_down(struct task_group *tg, void *data) |
| 1667 | { |
| 1668 | unsigned long load; |
| 1669 | long cpu = (long)data; |
| 1670 | |
| 1671 | if (!tg->parent) { |
| 1672 | load = cpu_rq(cpu)->load.weight; |
| 1673 | } else { |
| 1674 | load = tg->parent->cfs_rq[cpu]->h_load; |
| 1675 | load *= tg->cfs_rq[cpu]->shares; |
| 1676 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; |
| 1677 | } |
| 1678 | |
| 1679 | tg->cfs_rq[cpu]->h_load = load; |
| 1680 | |
| 1681 | return 0; |
| 1682 | } |
| 1683 | |
| 1684 | static void update_shares(struct sched_domain *sd) |
| 1685 | { |
| 1686 | s64 elapsed; |
| 1687 | u64 now; |
| 1688 | |
| 1689 | if (root_task_group_empty()) |
| 1690 | return; |
| 1691 | |
| 1692 | now = cpu_clock(raw_smp_processor_id()); |
| 1693 | elapsed = now - sd->last_update; |
| 1694 | |
| 1695 | if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) { |
| 1696 | sd->last_update = now; |
| 1697 | walk_tg_tree(tg_nop, tg_shares_up, sd); |
| 1698 | } |
| 1699 | } |
| 1700 | |
| 1701 | static void update_shares_locked(struct rq *rq, struct sched_domain *sd) |
| 1702 | { |
| 1703 | if (root_task_group_empty()) |
| 1704 | return; |
| 1705 | |
| 1706 | spin_unlock(&rq->lock); |
| 1707 | update_shares(sd); |
| 1708 | spin_lock(&rq->lock); |
| 1709 | } |
| 1710 | |
| 1711 | static void update_h_load(long cpu) |
| 1712 | { |
| 1713 | if (root_task_group_empty()) |
| 1714 | return; |
| 1715 | |
| 1716 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); |
| 1717 | } |
| 1718 | |
| 1719 | #else |
| 1720 | |
| 1721 | static inline void update_shares(struct sched_domain *sd) |
| 1722 | { |
| 1723 | } |
| 1724 | |
| 1725 | static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd) |
| 1726 | { |
| 1727 | } |
| 1728 | |
| 1729 | #endif |
| 1730 | |
| 1731 | #ifdef CONFIG_PREEMPT |
| 1732 | |
| 1733 | static void double_rq_lock(struct rq *rq1, struct rq *rq2); |
| 1734 | |
| 1735 | /* |
| 1736 | * fair double_lock_balance: Safely acquires both rq->locks in a fair |
| 1737 | * way at the expense of forcing extra atomic operations in all |
| 1738 | * invocations. This assures that the double_lock is acquired using the |
| 1739 | * same underlying policy as the spinlock_t on this architecture, which |
| 1740 | * reduces latency compared to the unfair variant below. However, it |
| 1741 | * also adds more overhead and therefore may reduce throughput. |
| 1742 | */ |
| 1743 | static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) |
| 1744 | __releases(this_rq->lock) |
| 1745 | __acquires(busiest->lock) |
| 1746 | __acquires(this_rq->lock) |
| 1747 | { |
| 1748 | spin_unlock(&this_rq->lock); |
| 1749 | double_rq_lock(this_rq, busiest); |
| 1750 | |
| 1751 | return 1; |
| 1752 | } |
| 1753 | |
| 1754 | #else |
| 1755 | /* |
| 1756 | * Unfair double_lock_balance: Optimizes throughput at the expense of |
| 1757 | * latency by eliminating extra atomic operations when the locks are |
| 1758 | * already in proper order on entry. This favors lower cpu-ids and will |
| 1759 | * grant the double lock to lower cpus over higher ids under contention, |
| 1760 | * regardless of entry order into the function. |
| 1761 | */ |
| 1762 | static int _double_lock_balance(struct rq *this_rq, struct rq *busiest) |
| 1763 | __releases(this_rq->lock) |
| 1764 | __acquires(busiest->lock) |
| 1765 | __acquires(this_rq->lock) |
| 1766 | { |
| 1767 | int ret = 0; |
| 1768 | |
| 1769 | if (unlikely(!spin_trylock(&busiest->lock))) { |
| 1770 | if (busiest < this_rq) { |
| 1771 | spin_unlock(&this_rq->lock); |
| 1772 | spin_lock(&busiest->lock); |
| 1773 | spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING); |
| 1774 | ret = 1; |
| 1775 | } else |
| 1776 | spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING); |
| 1777 | } |
| 1778 | return ret; |
| 1779 | } |
| 1780 | |
| 1781 | #endif /* CONFIG_PREEMPT */ |
| 1782 | |
| 1783 | /* |
| 1784 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. |
| 1785 | */ |
| 1786 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest) |
| 1787 | { |
| 1788 | if (unlikely(!irqs_disabled())) { |
| 1789 | /* printk() doesn't work good under rq->lock */ |
| 1790 | spin_unlock(&this_rq->lock); |
| 1791 | BUG_ON(1); |
| 1792 | } |
| 1793 | |
| 1794 | return _double_lock_balance(this_rq, busiest); |
| 1795 | } |
| 1796 | |
| 1797 | static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) |
| 1798 | __releases(busiest->lock) |
| 1799 | { |
| 1800 | spin_unlock(&busiest->lock); |
| 1801 | lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); |
| 1802 | } |
| 1803 | #endif |
| 1804 | |
| 1805 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1806 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) |
| 1807 | { |
| 1808 | #ifdef CONFIG_SMP |
| 1809 | cfs_rq->shares = shares; |
| 1810 | #endif |
| 1811 | } |
| 1812 | #endif |
| 1813 | |
| 1814 | static void calc_load_account_active(struct rq *this_rq); |
| 1815 | |
| 1816 | #include "sched_stats.h" |
| 1817 | #include "sched_idletask.c" |
| 1818 | #include "sched_fair.c" |
| 1819 | #include "sched_rt.c" |
| 1820 | #ifdef CONFIG_SCHED_DEBUG |
| 1821 | # include "sched_debug.c" |
| 1822 | #endif |
| 1823 | |
| 1824 | #define sched_class_highest (&rt_sched_class) |
| 1825 | #define for_each_class(class) \ |
| 1826 | for (class = sched_class_highest; class; class = class->next) |
| 1827 | |
| 1828 | static void inc_nr_running(struct rq *rq) |
| 1829 | { |
| 1830 | rq->nr_running++; |
| 1831 | } |
| 1832 | |
| 1833 | static void dec_nr_running(struct rq *rq) |
| 1834 | { |
| 1835 | rq->nr_running--; |
| 1836 | } |
| 1837 | |
| 1838 | static void set_load_weight(struct task_struct *p) |
| 1839 | { |
| 1840 | if (task_has_rt_policy(p)) { |
| 1841 | p->se.load.weight = prio_to_weight[0] * 2; |
| 1842 | p->se.load.inv_weight = prio_to_wmult[0] >> 1; |
| 1843 | return; |
| 1844 | } |
| 1845 | |
| 1846 | /* |
| 1847 | * SCHED_IDLE tasks get minimal weight: |
| 1848 | */ |
| 1849 | if (p->policy == SCHED_IDLE) { |
| 1850 | p->se.load.weight = WEIGHT_IDLEPRIO; |
| 1851 | p->se.load.inv_weight = WMULT_IDLEPRIO; |
| 1852 | return; |
| 1853 | } |
| 1854 | |
| 1855 | p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO]; |
| 1856 | p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO]; |
| 1857 | } |
| 1858 | |
| 1859 | static void update_avg(u64 *avg, u64 sample) |
| 1860 | { |
| 1861 | s64 diff = sample - *avg; |
| 1862 | *avg += diff >> 3; |
| 1863 | } |
| 1864 | |
| 1865 | static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup) |
| 1866 | { |
| 1867 | if (wakeup) |
| 1868 | p->se.start_runtime = p->se.sum_exec_runtime; |
| 1869 | |
| 1870 | sched_info_queued(p); |
| 1871 | p->sched_class->enqueue_task(rq, p, wakeup); |
| 1872 | p->se.on_rq = 1; |
| 1873 | } |
| 1874 | |
| 1875 | static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep) |
| 1876 | { |
| 1877 | if (sleep) { |
| 1878 | if (p->se.last_wakeup) { |
| 1879 | update_avg(&p->se.avg_overlap, |
| 1880 | p->se.sum_exec_runtime - p->se.last_wakeup); |
| 1881 | p->se.last_wakeup = 0; |
| 1882 | } else { |
| 1883 | update_avg(&p->se.avg_wakeup, |
| 1884 | sysctl_sched_wakeup_granularity); |
| 1885 | } |
| 1886 | } |
| 1887 | |
| 1888 | sched_info_dequeued(p); |
| 1889 | p->sched_class->dequeue_task(rq, p, sleep); |
| 1890 | p->se.on_rq = 0; |
| 1891 | } |
| 1892 | |
| 1893 | /* |
| 1894 | * __normal_prio - return the priority that is based on the static prio |
| 1895 | */ |
| 1896 | static inline int __normal_prio(struct task_struct *p) |
| 1897 | { |
| 1898 | return p->static_prio; |
| 1899 | } |
| 1900 | |
| 1901 | /* |
| 1902 | * Calculate the expected normal priority: i.e. priority |
| 1903 | * without taking RT-inheritance into account. Might be |
| 1904 | * boosted by interactivity modifiers. Changes upon fork, |
| 1905 | * setprio syscalls, and whenever the interactivity |
| 1906 | * estimator recalculates. |
| 1907 | */ |
| 1908 | static inline int normal_prio(struct task_struct *p) |
| 1909 | { |
| 1910 | int prio; |
| 1911 | |
| 1912 | if (task_has_rt_policy(p)) |
| 1913 | prio = MAX_RT_PRIO-1 - p->rt_priority; |
| 1914 | else |
| 1915 | prio = __normal_prio(p); |
| 1916 | return prio; |
| 1917 | } |
| 1918 | |
| 1919 | /* |
| 1920 | * Calculate the current priority, i.e. the priority |
| 1921 | * taken into account by the scheduler. This value might |
| 1922 | * be boosted by RT tasks, or might be boosted by |
| 1923 | * interactivity modifiers. Will be RT if the task got |
| 1924 | * RT-boosted. If not then it returns p->normal_prio. |
| 1925 | */ |
| 1926 | static int effective_prio(struct task_struct *p) |
| 1927 | { |
| 1928 | p->normal_prio = normal_prio(p); |
| 1929 | /* |
| 1930 | * If we are RT tasks or we were boosted to RT priority, |
| 1931 | * keep the priority unchanged. Otherwise, update priority |
| 1932 | * to the normal priority: |
| 1933 | */ |
| 1934 | if (!rt_prio(p->prio)) |
| 1935 | return p->normal_prio; |
| 1936 | return p->prio; |
| 1937 | } |
| 1938 | |
| 1939 | /* |
| 1940 | * activate_task - move a task to the runqueue. |
| 1941 | */ |
| 1942 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup) |
| 1943 | { |
| 1944 | if (task_contributes_to_load(p)) |
| 1945 | rq->nr_uninterruptible--; |
| 1946 | |
| 1947 | enqueue_task(rq, p, wakeup); |
| 1948 | inc_nr_running(rq); |
| 1949 | } |
| 1950 | |
| 1951 | /* |
| 1952 | * deactivate_task - remove a task from the runqueue. |
| 1953 | */ |
| 1954 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep) |
| 1955 | { |
| 1956 | if (task_contributes_to_load(p)) |
| 1957 | rq->nr_uninterruptible++; |
| 1958 | |
| 1959 | dequeue_task(rq, p, sleep); |
| 1960 | dec_nr_running(rq); |
| 1961 | } |
| 1962 | |
| 1963 | /** |
| 1964 | * task_curr - is this task currently executing on a CPU? |
| 1965 | * @p: the task in question. |
| 1966 | */ |
| 1967 | inline int task_curr(const struct task_struct *p) |
| 1968 | { |
| 1969 | return cpu_curr(task_cpu(p)) == p; |
| 1970 | } |
| 1971 | |
| 1972 | static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) |
| 1973 | { |
| 1974 | set_task_rq(p, cpu); |
| 1975 | #ifdef CONFIG_SMP |
| 1976 | /* |
| 1977 | * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be |
| 1978 | * successfuly executed on another CPU. We must ensure that updates of |
| 1979 | * per-task data have been completed by this moment. |
| 1980 | */ |
| 1981 | smp_wmb(); |
| 1982 | task_thread_info(p)->cpu = cpu; |
| 1983 | #endif |
| 1984 | } |
| 1985 | |
| 1986 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, |
| 1987 | const struct sched_class *prev_class, |
| 1988 | int oldprio, int running) |
| 1989 | { |
| 1990 | if (prev_class != p->sched_class) { |
| 1991 | if (prev_class->switched_from) |
| 1992 | prev_class->switched_from(rq, p, running); |
| 1993 | p->sched_class->switched_to(rq, p, running); |
| 1994 | } else |
| 1995 | p->sched_class->prio_changed(rq, p, oldprio, running); |
| 1996 | } |
| 1997 | |
| 1998 | #ifdef CONFIG_SMP |
| 1999 | /* |
| 2000 | * Is this task likely cache-hot: |
| 2001 | */ |
| 2002 | static int |
| 2003 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) |
| 2004 | { |
| 2005 | s64 delta; |
| 2006 | |
| 2007 | /* |
| 2008 | * Buddy candidates are cache hot: |
| 2009 | */ |
| 2010 | if (sched_feat(CACHE_HOT_BUDDY) && |
| 2011 | (&p->se == cfs_rq_of(&p->se)->next || |
| 2012 | &p->se == cfs_rq_of(&p->se)->last)) |
| 2013 | return 1; |
| 2014 | |
| 2015 | if (p->sched_class != &fair_sched_class) |
| 2016 | return 0; |
| 2017 | |
| 2018 | if (sysctl_sched_migration_cost == -1) |
| 2019 | return 1; |
| 2020 | if (sysctl_sched_migration_cost == 0) |
| 2021 | return 0; |
| 2022 | |
| 2023 | delta = now - p->se.exec_start; |
| 2024 | |
| 2025 | return delta < (s64)sysctl_sched_migration_cost; |
| 2026 | } |
| 2027 | |
| 2028 | |
| 2029 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
| 2030 | { |
| 2031 | int old_cpu = task_cpu(p); |
| 2032 | struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu); |
| 2033 | struct cfs_rq *old_cfsrq = task_cfs_rq(p), |
| 2034 | *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu); |
| 2035 | u64 clock_offset; |
| 2036 | |
| 2037 | clock_offset = old_rq->clock - new_rq->clock; |
| 2038 | |
| 2039 | trace_sched_migrate_task(p, new_cpu); |
| 2040 | |
| 2041 | #ifdef CONFIG_SCHEDSTATS |
| 2042 | if (p->se.wait_start) |
| 2043 | p->se.wait_start -= clock_offset; |
| 2044 | if (p->se.sleep_start) |
| 2045 | p->se.sleep_start -= clock_offset; |
| 2046 | if (p->se.block_start) |
| 2047 | p->se.block_start -= clock_offset; |
| 2048 | #endif |
| 2049 | if (old_cpu != new_cpu) { |
| 2050 | p->se.nr_migrations++; |
| 2051 | new_rq->nr_migrations_in++; |
| 2052 | #ifdef CONFIG_SCHEDSTATS |
| 2053 | if (task_hot(p, old_rq->clock, NULL)) |
| 2054 | schedstat_inc(p, se.nr_forced2_migrations); |
| 2055 | #endif |
| 2056 | perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS, |
| 2057 | 1, 1, NULL, 0); |
| 2058 | } |
| 2059 | p->se.vruntime -= old_cfsrq->min_vruntime - |
| 2060 | new_cfsrq->min_vruntime; |
| 2061 | |
| 2062 | __set_task_cpu(p, new_cpu); |
| 2063 | } |
| 2064 | |
| 2065 | struct migration_req { |
| 2066 | struct list_head list; |
| 2067 | |
| 2068 | struct task_struct *task; |
| 2069 | int dest_cpu; |
| 2070 | |
| 2071 | struct completion done; |
| 2072 | }; |
| 2073 | |
| 2074 | /* |
| 2075 | * The task's runqueue lock must be held. |
| 2076 | * Returns true if you have to wait for migration thread. |
| 2077 | */ |
| 2078 | static int |
| 2079 | migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req) |
| 2080 | { |
| 2081 | struct rq *rq = task_rq(p); |
| 2082 | |
| 2083 | /* |
| 2084 | * If the task is not on a runqueue (and not running), then |
| 2085 | * it is sufficient to simply update the task's cpu field. |
| 2086 | */ |
| 2087 | if (!p->se.on_rq && !task_running(rq, p)) { |
| 2088 | set_task_cpu(p, dest_cpu); |
| 2089 | return 0; |
| 2090 | } |
| 2091 | |
| 2092 | init_completion(&req->done); |
| 2093 | req->task = p; |
| 2094 | req->dest_cpu = dest_cpu; |
| 2095 | list_add(&req->list, &rq->migration_queue); |
| 2096 | |
| 2097 | return 1; |
| 2098 | } |
| 2099 | |
| 2100 | /* |
| 2101 | * wait_task_context_switch - wait for a thread to complete at least one |
| 2102 | * context switch. |
| 2103 | * |
| 2104 | * @p must not be current. |
| 2105 | */ |
| 2106 | void wait_task_context_switch(struct task_struct *p) |
| 2107 | { |
| 2108 | unsigned long nvcsw, nivcsw, flags; |
| 2109 | int running; |
| 2110 | struct rq *rq; |
| 2111 | |
| 2112 | nvcsw = p->nvcsw; |
| 2113 | nivcsw = p->nivcsw; |
| 2114 | for (;;) { |
| 2115 | /* |
| 2116 | * The runqueue is assigned before the actual context |
| 2117 | * switch. We need to take the runqueue lock. |
| 2118 | * |
| 2119 | * We could check initially without the lock but it is |
| 2120 | * very likely that we need to take the lock in every |
| 2121 | * iteration. |
| 2122 | */ |
| 2123 | rq = task_rq_lock(p, &flags); |
| 2124 | running = task_running(rq, p); |
| 2125 | task_rq_unlock(rq, &flags); |
| 2126 | |
| 2127 | if (likely(!running)) |
| 2128 | break; |
| 2129 | /* |
| 2130 | * The switch count is incremented before the actual |
| 2131 | * context switch. We thus wait for two switches to be |
| 2132 | * sure at least one completed. |
| 2133 | */ |
| 2134 | if ((p->nvcsw - nvcsw) > 1) |
| 2135 | break; |
| 2136 | if ((p->nivcsw - nivcsw) > 1) |
| 2137 | break; |
| 2138 | |
| 2139 | cpu_relax(); |
| 2140 | } |
| 2141 | } |
| 2142 | |
| 2143 | /* |
| 2144 | * wait_task_inactive - wait for a thread to unschedule. |
| 2145 | * |
| 2146 | * If @match_state is nonzero, it's the @p->state value just checked and |
| 2147 | * not expected to change. If it changes, i.e. @p might have woken up, |
| 2148 | * then return zero. When we succeed in waiting for @p to be off its CPU, |
| 2149 | * we return a positive number (its total switch count). If a second call |
| 2150 | * a short while later returns the same number, the caller can be sure that |
| 2151 | * @p has remained unscheduled the whole time. |
| 2152 | * |
| 2153 | * The caller must ensure that the task *will* unschedule sometime soon, |
| 2154 | * else this function might spin for a *long* time. This function can't |
| 2155 | * be called with interrupts off, or it may introduce deadlock with |
| 2156 | * smp_call_function() if an IPI is sent by the same process we are |
| 2157 | * waiting to become inactive. |
| 2158 | */ |
| 2159 | unsigned long wait_task_inactive(struct task_struct *p, long match_state) |
| 2160 | { |
| 2161 | unsigned long flags; |
| 2162 | int running, on_rq; |
| 2163 | unsigned long ncsw; |
| 2164 | struct rq *rq; |
| 2165 | |
| 2166 | for (;;) { |
| 2167 | /* |
| 2168 | * We do the initial early heuristics without holding |
| 2169 | * any task-queue locks at all. We'll only try to get |
| 2170 | * the runqueue lock when things look like they will |
| 2171 | * work out! |
| 2172 | */ |
| 2173 | rq = task_rq(p); |
| 2174 | |
| 2175 | /* |
| 2176 | * If the task is actively running on another CPU |
| 2177 | * still, just relax and busy-wait without holding |
| 2178 | * any locks. |
| 2179 | * |
| 2180 | * NOTE! Since we don't hold any locks, it's not |
| 2181 | * even sure that "rq" stays as the right runqueue! |
| 2182 | * But we don't care, since "task_running()" will |
| 2183 | * return false if the runqueue has changed and p |
| 2184 | * is actually now running somewhere else! |
| 2185 | */ |
| 2186 | while (task_running(rq, p)) { |
| 2187 | if (match_state && unlikely(p->state != match_state)) |
| 2188 | return 0; |
| 2189 | cpu_relax(); |
| 2190 | } |
| 2191 | |
| 2192 | /* |
| 2193 | * Ok, time to look more closely! We need the rq |
| 2194 | * lock now, to be *sure*. If we're wrong, we'll |
| 2195 | * just go back and repeat. |
| 2196 | */ |
| 2197 | rq = task_rq_lock(p, &flags); |
| 2198 | trace_sched_wait_task(rq, p); |
| 2199 | running = task_running(rq, p); |
| 2200 | on_rq = p->se.on_rq; |
| 2201 | ncsw = 0; |
| 2202 | if (!match_state || p->state == match_state) |
| 2203 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ |
| 2204 | task_rq_unlock(rq, &flags); |
| 2205 | |
| 2206 | /* |
| 2207 | * If it changed from the expected state, bail out now. |
| 2208 | */ |
| 2209 | if (unlikely(!ncsw)) |
| 2210 | break; |
| 2211 | |
| 2212 | /* |
| 2213 | * Was it really running after all now that we |
| 2214 | * checked with the proper locks actually held? |
| 2215 | * |
| 2216 | * Oops. Go back and try again.. |
| 2217 | */ |
| 2218 | if (unlikely(running)) { |
| 2219 | cpu_relax(); |
| 2220 | continue; |
| 2221 | } |
| 2222 | |
| 2223 | /* |
| 2224 | * It's not enough that it's not actively running, |
| 2225 | * it must be off the runqueue _entirely_, and not |
| 2226 | * preempted! |
| 2227 | * |
| 2228 | * So if it was still runnable (but just not actively |
| 2229 | * running right now), it's preempted, and we should |
| 2230 | * yield - it could be a while. |
| 2231 | */ |
| 2232 | if (unlikely(on_rq)) { |
| 2233 | schedule_timeout_uninterruptible(1); |
| 2234 | continue; |
| 2235 | } |
| 2236 | |
| 2237 | /* |
| 2238 | * Ahh, all good. It wasn't running, and it wasn't |
| 2239 | * runnable, which means that it will never become |
| 2240 | * running in the future either. We're all done! |
| 2241 | */ |
| 2242 | break; |
| 2243 | } |
| 2244 | |
| 2245 | return ncsw; |
| 2246 | } |
| 2247 | |
| 2248 | /*** |
| 2249 | * kick_process - kick a running thread to enter/exit the kernel |
| 2250 | * @p: the to-be-kicked thread |
| 2251 | * |
| 2252 | * Cause a process which is running on another CPU to enter |
| 2253 | * kernel-mode, without any delay. (to get signals handled.) |
| 2254 | * |
| 2255 | * NOTE: this function doesnt have to take the runqueue lock, |
| 2256 | * because all it wants to ensure is that the remote task enters |
| 2257 | * the kernel. If the IPI races and the task has been migrated |
| 2258 | * to another CPU then no harm is done and the purpose has been |
| 2259 | * achieved as well. |
| 2260 | */ |
| 2261 | void kick_process(struct task_struct *p) |
| 2262 | { |
| 2263 | int cpu; |
| 2264 | |
| 2265 | preempt_disable(); |
| 2266 | cpu = task_cpu(p); |
| 2267 | if ((cpu != smp_processor_id()) && task_curr(p)) |
| 2268 | smp_send_reschedule(cpu); |
| 2269 | preempt_enable(); |
| 2270 | } |
| 2271 | EXPORT_SYMBOL_GPL(kick_process); |
| 2272 | |
| 2273 | /* |
| 2274 | * find_idlest_group finds and returns the least busy CPU group within the |
| 2275 | * domain. |
| 2276 | */ |
| 2277 | static struct sched_group * |
| 2278 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
| 2279 | { |
| 2280 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; |
| 2281 | unsigned long min_load = ULONG_MAX, this_load = 0; |
| 2282 | int load_idx = sd->forkexec_idx; |
| 2283 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
| 2284 | |
| 2285 | do { |
| 2286 | unsigned long load, avg_load; |
| 2287 | int local_group; |
| 2288 | int i; |
| 2289 | |
| 2290 | /* Skip over this group if it has no CPUs allowed */ |
| 2291 | if (!cpumask_intersects(sched_group_cpus(group), |
| 2292 | &p->cpus_allowed)) |
| 2293 | continue; |
| 2294 | |
| 2295 | local_group = cpumask_test_cpu(this_cpu, |
| 2296 | sched_group_cpus(group)); |
| 2297 | |
| 2298 | /* Tally up the load of all CPUs in the group */ |
| 2299 | avg_load = 0; |
| 2300 | |
| 2301 | for_each_cpu(i, sched_group_cpus(group)) { |
| 2302 | /* Bias balancing toward cpus of our domain */ |
| 2303 | if (local_group) |
| 2304 | load = source_load(i, load_idx); |
| 2305 | else |
| 2306 | load = target_load(i, load_idx); |
| 2307 | |
| 2308 | avg_load += load; |
| 2309 | } |
| 2310 | |
| 2311 | /* Adjust by relative CPU power of the group */ |
| 2312 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; |
| 2313 | |
| 2314 | if (local_group) { |
| 2315 | this_load = avg_load; |
| 2316 | this = group; |
| 2317 | } else if (avg_load < min_load) { |
| 2318 | min_load = avg_load; |
| 2319 | idlest = group; |
| 2320 | } |
| 2321 | } while (group = group->next, group != sd->groups); |
| 2322 | |
| 2323 | if (!idlest || 100*this_load < imbalance*min_load) |
| 2324 | return NULL; |
| 2325 | return idlest; |
| 2326 | } |
| 2327 | |
| 2328 | /* |
| 2329 | * find_idlest_cpu - find the idlest cpu among the cpus in group. |
| 2330 | */ |
| 2331 | static int |
| 2332 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
| 2333 | { |
| 2334 | unsigned long load, min_load = ULONG_MAX; |
| 2335 | int idlest = -1; |
| 2336 | int i; |
| 2337 | |
| 2338 | /* Traverse only the allowed CPUs */ |
| 2339 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { |
| 2340 | load = weighted_cpuload(i); |
| 2341 | |
| 2342 | if (load < min_load || (load == min_load && i == this_cpu)) { |
| 2343 | min_load = load; |
| 2344 | idlest = i; |
| 2345 | } |
| 2346 | } |
| 2347 | |
| 2348 | return idlest; |
| 2349 | } |
| 2350 | |
| 2351 | /* |
| 2352 | * sched_balance_self: balance the current task (running on cpu) in domains |
| 2353 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and |
| 2354 | * SD_BALANCE_EXEC. |
| 2355 | * |
| 2356 | * Balance, ie. select the least loaded group. |
| 2357 | * |
| 2358 | * Returns the target CPU number, or the same CPU if no balancing is needed. |
| 2359 | * |
| 2360 | * preempt must be disabled. |
| 2361 | */ |
| 2362 | static int sched_balance_self(int cpu, int flag) |
| 2363 | { |
| 2364 | struct task_struct *t = current; |
| 2365 | struct sched_domain *tmp, *sd = NULL; |
| 2366 | |
| 2367 | for_each_domain(cpu, tmp) { |
| 2368 | /* |
| 2369 | * If power savings logic is enabled for a domain, stop there. |
| 2370 | */ |
| 2371 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
| 2372 | break; |
| 2373 | if (tmp->flags & flag) |
| 2374 | sd = tmp; |
| 2375 | } |
| 2376 | |
| 2377 | if (sd) |
| 2378 | update_shares(sd); |
| 2379 | |
| 2380 | while (sd) { |
| 2381 | struct sched_group *group; |
| 2382 | int new_cpu, weight; |
| 2383 | |
| 2384 | if (!(sd->flags & flag)) { |
| 2385 | sd = sd->child; |
| 2386 | continue; |
| 2387 | } |
| 2388 | |
| 2389 | group = find_idlest_group(sd, t, cpu); |
| 2390 | if (!group) { |
| 2391 | sd = sd->child; |
| 2392 | continue; |
| 2393 | } |
| 2394 | |
| 2395 | new_cpu = find_idlest_cpu(group, t, cpu); |
| 2396 | if (new_cpu == -1 || new_cpu == cpu) { |
| 2397 | /* Now try balancing at a lower domain level of cpu */ |
| 2398 | sd = sd->child; |
| 2399 | continue; |
| 2400 | } |
| 2401 | |
| 2402 | /* Now try balancing at a lower domain level of new_cpu */ |
| 2403 | cpu = new_cpu; |
| 2404 | weight = cpumask_weight(sched_domain_span(sd)); |
| 2405 | sd = NULL; |
| 2406 | for_each_domain(cpu, tmp) { |
| 2407 | if (weight <= cpumask_weight(sched_domain_span(tmp))) |
| 2408 | break; |
| 2409 | if (tmp->flags & flag) |
| 2410 | sd = tmp; |
| 2411 | } |
| 2412 | /* while loop will break here if sd == NULL */ |
| 2413 | } |
| 2414 | |
| 2415 | return cpu; |
| 2416 | } |
| 2417 | |
| 2418 | #endif /* CONFIG_SMP */ |
| 2419 | |
| 2420 | /** |
| 2421 | * task_oncpu_function_call - call a function on the cpu on which a task runs |
| 2422 | * @p: the task to evaluate |
| 2423 | * @func: the function to be called |
| 2424 | * @info: the function call argument |
| 2425 | * |
| 2426 | * Calls the function @func when the task is currently running. This might |
| 2427 | * be on the current CPU, which just calls the function directly |
| 2428 | */ |
| 2429 | void task_oncpu_function_call(struct task_struct *p, |
| 2430 | void (*func) (void *info), void *info) |
| 2431 | { |
| 2432 | int cpu; |
| 2433 | |
| 2434 | preempt_disable(); |
| 2435 | cpu = task_cpu(p); |
| 2436 | if (task_curr(p)) |
| 2437 | smp_call_function_single(cpu, func, info, 1); |
| 2438 | preempt_enable(); |
| 2439 | } |
| 2440 | |
| 2441 | /*** |
| 2442 | * try_to_wake_up - wake up a thread |
| 2443 | * @p: the to-be-woken-up thread |
| 2444 | * @state: the mask of task states that can be woken |
| 2445 | * @sync: do a synchronous wakeup? |
| 2446 | * |
| 2447 | * Put it on the run-queue if it's not already there. The "current" |
| 2448 | * thread is always on the run-queue (except when the actual |
| 2449 | * re-schedule is in progress), and as such you're allowed to do |
| 2450 | * the simpler "current->state = TASK_RUNNING" to mark yourself |
| 2451 | * runnable without the overhead of this. |
| 2452 | * |
| 2453 | * returns failure only if the task is already active. |
| 2454 | */ |
| 2455 | static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync) |
| 2456 | { |
| 2457 | int cpu, orig_cpu, this_cpu, success = 0; |
| 2458 | unsigned long flags; |
| 2459 | long old_state; |
| 2460 | struct rq *rq; |
| 2461 | |
| 2462 | if (!sched_feat(SYNC_WAKEUPS)) |
| 2463 | sync = 0; |
| 2464 | |
| 2465 | #ifdef CONFIG_SMP |
| 2466 | if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) { |
| 2467 | struct sched_domain *sd; |
| 2468 | |
| 2469 | this_cpu = raw_smp_processor_id(); |
| 2470 | cpu = task_cpu(p); |
| 2471 | |
| 2472 | for_each_domain(this_cpu, sd) { |
| 2473 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
| 2474 | update_shares(sd); |
| 2475 | break; |
| 2476 | } |
| 2477 | } |
| 2478 | } |
| 2479 | #endif |
| 2480 | |
| 2481 | smp_wmb(); |
| 2482 | rq = task_rq_lock(p, &flags); |
| 2483 | update_rq_clock(rq); |
| 2484 | old_state = p->state; |
| 2485 | if (!(old_state & state)) |
| 2486 | goto out; |
| 2487 | |
| 2488 | if (p->se.on_rq) |
| 2489 | goto out_running; |
| 2490 | |
| 2491 | cpu = task_cpu(p); |
| 2492 | orig_cpu = cpu; |
| 2493 | this_cpu = smp_processor_id(); |
| 2494 | |
| 2495 | #ifdef CONFIG_SMP |
| 2496 | if (unlikely(task_running(rq, p))) |
| 2497 | goto out_activate; |
| 2498 | |
| 2499 | cpu = p->sched_class->select_task_rq(p, sync); |
| 2500 | if (cpu != orig_cpu) { |
| 2501 | set_task_cpu(p, cpu); |
| 2502 | task_rq_unlock(rq, &flags); |
| 2503 | /* might preempt at this point */ |
| 2504 | rq = task_rq_lock(p, &flags); |
| 2505 | old_state = p->state; |
| 2506 | if (!(old_state & state)) |
| 2507 | goto out; |
| 2508 | if (p->se.on_rq) |
| 2509 | goto out_running; |
| 2510 | |
| 2511 | this_cpu = smp_processor_id(); |
| 2512 | cpu = task_cpu(p); |
| 2513 | } |
| 2514 | |
| 2515 | #ifdef CONFIG_SCHEDSTATS |
| 2516 | schedstat_inc(rq, ttwu_count); |
| 2517 | if (cpu == this_cpu) |
| 2518 | schedstat_inc(rq, ttwu_local); |
| 2519 | else { |
| 2520 | struct sched_domain *sd; |
| 2521 | for_each_domain(this_cpu, sd) { |
| 2522 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
| 2523 | schedstat_inc(sd, ttwu_wake_remote); |
| 2524 | break; |
| 2525 | } |
| 2526 | } |
| 2527 | } |
| 2528 | #endif /* CONFIG_SCHEDSTATS */ |
| 2529 | |
| 2530 | out_activate: |
| 2531 | #endif /* CONFIG_SMP */ |
| 2532 | schedstat_inc(p, se.nr_wakeups); |
| 2533 | if (sync) |
| 2534 | schedstat_inc(p, se.nr_wakeups_sync); |
| 2535 | if (orig_cpu != cpu) |
| 2536 | schedstat_inc(p, se.nr_wakeups_migrate); |
| 2537 | if (cpu == this_cpu) |
| 2538 | schedstat_inc(p, se.nr_wakeups_local); |
| 2539 | else |
| 2540 | schedstat_inc(p, se.nr_wakeups_remote); |
| 2541 | activate_task(rq, p, 1); |
| 2542 | success = 1; |
| 2543 | |
| 2544 | /* |
| 2545 | * Only attribute actual wakeups done by this task. |
| 2546 | */ |
| 2547 | if (!in_interrupt()) { |
| 2548 | struct sched_entity *se = ¤t->se; |
| 2549 | u64 sample = se->sum_exec_runtime; |
| 2550 | |
| 2551 | if (se->last_wakeup) |
| 2552 | sample -= se->last_wakeup; |
| 2553 | else |
| 2554 | sample -= se->start_runtime; |
| 2555 | update_avg(&se->avg_wakeup, sample); |
| 2556 | |
| 2557 | se->last_wakeup = se->sum_exec_runtime; |
| 2558 | } |
| 2559 | |
| 2560 | out_running: |
| 2561 | trace_sched_wakeup(rq, p, success); |
| 2562 | check_preempt_curr(rq, p, sync); |
| 2563 | |
| 2564 | p->state = TASK_RUNNING; |
| 2565 | #ifdef CONFIG_SMP |
| 2566 | if (p->sched_class->task_wake_up) |
| 2567 | p->sched_class->task_wake_up(rq, p); |
| 2568 | #endif |
| 2569 | out: |
| 2570 | task_rq_unlock(rq, &flags); |
| 2571 | |
| 2572 | return success; |
| 2573 | } |
| 2574 | |
| 2575 | /** |
| 2576 | * wake_up_process - Wake up a specific process |
| 2577 | * @p: The process to be woken up. |
| 2578 | * |
| 2579 | * Attempt to wake up the nominated process and move it to the set of runnable |
| 2580 | * processes. Returns 1 if the process was woken up, 0 if it was already |
| 2581 | * running. |
| 2582 | * |
| 2583 | * It may be assumed that this function implies a write memory barrier before |
| 2584 | * changing the task state if and only if any tasks are woken up. |
| 2585 | */ |
| 2586 | int wake_up_process(struct task_struct *p) |
| 2587 | { |
| 2588 | return try_to_wake_up(p, TASK_ALL, 0); |
| 2589 | } |
| 2590 | EXPORT_SYMBOL(wake_up_process); |
| 2591 | |
| 2592 | int wake_up_state(struct task_struct *p, unsigned int state) |
| 2593 | { |
| 2594 | return try_to_wake_up(p, state, 0); |
| 2595 | } |
| 2596 | |
| 2597 | /* |
| 2598 | * Perform scheduler related setup for a newly forked process p. |
| 2599 | * p is forked by current. |
| 2600 | * |
| 2601 | * __sched_fork() is basic setup used by init_idle() too: |
| 2602 | */ |
| 2603 | static void __sched_fork(struct task_struct *p) |
| 2604 | { |
| 2605 | p->se.exec_start = 0; |
| 2606 | p->se.sum_exec_runtime = 0; |
| 2607 | p->se.prev_sum_exec_runtime = 0; |
| 2608 | p->se.nr_migrations = 0; |
| 2609 | p->se.last_wakeup = 0; |
| 2610 | p->se.avg_overlap = 0; |
| 2611 | p->se.start_runtime = 0; |
| 2612 | p->se.avg_wakeup = sysctl_sched_wakeup_granularity; |
| 2613 | |
| 2614 | #ifdef CONFIG_SCHEDSTATS |
| 2615 | p->se.wait_start = 0; |
| 2616 | p->se.wait_max = 0; |
| 2617 | p->se.wait_count = 0; |
| 2618 | p->se.wait_sum = 0; |
| 2619 | |
| 2620 | p->se.sleep_start = 0; |
| 2621 | p->se.sleep_max = 0; |
| 2622 | p->se.sum_sleep_runtime = 0; |
| 2623 | |
| 2624 | p->se.block_start = 0; |
| 2625 | p->se.block_max = 0; |
| 2626 | p->se.exec_max = 0; |
| 2627 | p->se.slice_max = 0; |
| 2628 | |
| 2629 | p->se.nr_migrations_cold = 0; |
| 2630 | p->se.nr_failed_migrations_affine = 0; |
| 2631 | p->se.nr_failed_migrations_running = 0; |
| 2632 | p->se.nr_failed_migrations_hot = 0; |
| 2633 | p->se.nr_forced_migrations = 0; |
| 2634 | p->se.nr_forced2_migrations = 0; |
| 2635 | |
| 2636 | p->se.nr_wakeups = 0; |
| 2637 | p->se.nr_wakeups_sync = 0; |
| 2638 | p->se.nr_wakeups_migrate = 0; |
| 2639 | p->se.nr_wakeups_local = 0; |
| 2640 | p->se.nr_wakeups_remote = 0; |
| 2641 | p->se.nr_wakeups_affine = 0; |
| 2642 | p->se.nr_wakeups_affine_attempts = 0; |
| 2643 | p->se.nr_wakeups_passive = 0; |
| 2644 | p->se.nr_wakeups_idle = 0; |
| 2645 | |
| 2646 | #endif |
| 2647 | |
| 2648 | INIT_LIST_HEAD(&p->rt.run_list); |
| 2649 | p->se.on_rq = 0; |
| 2650 | INIT_LIST_HEAD(&p->se.group_node); |
| 2651 | |
| 2652 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 2653 | INIT_HLIST_HEAD(&p->preempt_notifiers); |
| 2654 | #endif |
| 2655 | |
| 2656 | /* |
| 2657 | * We mark the process as running here, but have not actually |
| 2658 | * inserted it onto the runqueue yet. This guarantees that |
| 2659 | * nobody will actually run it, and a signal or other external |
| 2660 | * event cannot wake it up and insert it on the runqueue either. |
| 2661 | */ |
| 2662 | p->state = TASK_RUNNING; |
| 2663 | } |
| 2664 | |
| 2665 | /* |
| 2666 | * fork()/clone()-time setup: |
| 2667 | */ |
| 2668 | void sched_fork(struct task_struct *p, int clone_flags) |
| 2669 | { |
| 2670 | int cpu = get_cpu(); |
| 2671 | |
| 2672 | __sched_fork(p); |
| 2673 | |
| 2674 | #ifdef CONFIG_SMP |
| 2675 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); |
| 2676 | #endif |
| 2677 | set_task_cpu(p, cpu); |
| 2678 | |
| 2679 | /* |
| 2680 | * Make sure we do not leak PI boosting priority to the child. |
| 2681 | */ |
| 2682 | p->prio = current->normal_prio; |
| 2683 | |
| 2684 | /* |
| 2685 | * Revert to default priority/policy on fork if requested. |
| 2686 | */ |
| 2687 | if (unlikely(p->sched_reset_on_fork)) { |
| 2688 | if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) |
| 2689 | p->policy = SCHED_NORMAL; |
| 2690 | |
| 2691 | if (p->normal_prio < DEFAULT_PRIO) |
| 2692 | p->prio = DEFAULT_PRIO; |
| 2693 | |
| 2694 | if (PRIO_TO_NICE(p->static_prio) < 0) { |
| 2695 | p->static_prio = NICE_TO_PRIO(0); |
| 2696 | set_load_weight(p); |
| 2697 | } |
| 2698 | |
| 2699 | /* |
| 2700 | * We don't need the reset flag anymore after the fork. It has |
| 2701 | * fulfilled its duty: |
| 2702 | */ |
| 2703 | p->sched_reset_on_fork = 0; |
| 2704 | } |
| 2705 | |
| 2706 | if (!rt_prio(p->prio)) |
| 2707 | p->sched_class = &fair_sched_class; |
| 2708 | |
| 2709 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
| 2710 | if (likely(sched_info_on())) |
| 2711 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
| 2712 | #endif |
| 2713 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
| 2714 | p->oncpu = 0; |
| 2715 | #endif |
| 2716 | #ifdef CONFIG_PREEMPT |
| 2717 | /* Want to start with kernel preemption disabled. */ |
| 2718 | task_thread_info(p)->preempt_count = 1; |
| 2719 | #endif |
| 2720 | plist_node_init(&p->pushable_tasks, MAX_PRIO); |
| 2721 | |
| 2722 | put_cpu(); |
| 2723 | } |
| 2724 | |
| 2725 | /* |
| 2726 | * wake_up_new_task - wake up a newly created task for the first time. |
| 2727 | * |
| 2728 | * This function will do some initial scheduler statistics housekeeping |
| 2729 | * that must be done for every newly created context, then puts the task |
| 2730 | * on the runqueue and wakes it. |
| 2731 | */ |
| 2732 | void wake_up_new_task(struct task_struct *p, unsigned long clone_flags) |
| 2733 | { |
| 2734 | unsigned long flags; |
| 2735 | struct rq *rq; |
| 2736 | |
| 2737 | rq = task_rq_lock(p, &flags); |
| 2738 | BUG_ON(p->state != TASK_RUNNING); |
| 2739 | update_rq_clock(rq); |
| 2740 | |
| 2741 | p->prio = effective_prio(p); |
| 2742 | |
| 2743 | if (!p->sched_class->task_new || !current->se.on_rq) { |
| 2744 | activate_task(rq, p, 0); |
| 2745 | } else { |
| 2746 | /* |
| 2747 | * Let the scheduling class do new task startup |
| 2748 | * management (if any): |
| 2749 | */ |
| 2750 | p->sched_class->task_new(rq, p); |
| 2751 | inc_nr_running(rq); |
| 2752 | } |
| 2753 | trace_sched_wakeup_new(rq, p, 1); |
| 2754 | check_preempt_curr(rq, p, 0); |
| 2755 | #ifdef CONFIG_SMP |
| 2756 | if (p->sched_class->task_wake_up) |
| 2757 | p->sched_class->task_wake_up(rq, p); |
| 2758 | #endif |
| 2759 | task_rq_unlock(rq, &flags); |
| 2760 | } |
| 2761 | |
| 2762 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 2763 | |
| 2764 | /** |
| 2765 | * preempt_notifier_register - tell me when current is being preempted & rescheduled |
| 2766 | * @notifier: notifier struct to register |
| 2767 | */ |
| 2768 | void preempt_notifier_register(struct preempt_notifier *notifier) |
| 2769 | { |
| 2770 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); |
| 2771 | } |
| 2772 | EXPORT_SYMBOL_GPL(preempt_notifier_register); |
| 2773 | |
| 2774 | /** |
| 2775 | * preempt_notifier_unregister - no longer interested in preemption notifications |
| 2776 | * @notifier: notifier struct to unregister |
| 2777 | * |
| 2778 | * This is safe to call from within a preemption notifier. |
| 2779 | */ |
| 2780 | void preempt_notifier_unregister(struct preempt_notifier *notifier) |
| 2781 | { |
| 2782 | hlist_del(¬ifier->link); |
| 2783 | } |
| 2784 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); |
| 2785 | |
| 2786 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
| 2787 | { |
| 2788 | struct preempt_notifier *notifier; |
| 2789 | struct hlist_node *node; |
| 2790 | |
| 2791 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) |
| 2792 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); |
| 2793 | } |
| 2794 | |
| 2795 | static void |
| 2796 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
| 2797 | struct task_struct *next) |
| 2798 | { |
| 2799 | struct preempt_notifier *notifier; |
| 2800 | struct hlist_node *node; |
| 2801 | |
| 2802 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) |
| 2803 | notifier->ops->sched_out(notifier, next); |
| 2804 | } |
| 2805 | |
| 2806 | #else /* !CONFIG_PREEMPT_NOTIFIERS */ |
| 2807 | |
| 2808 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
| 2809 | { |
| 2810 | } |
| 2811 | |
| 2812 | static void |
| 2813 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
| 2814 | struct task_struct *next) |
| 2815 | { |
| 2816 | } |
| 2817 | |
| 2818 | #endif /* CONFIG_PREEMPT_NOTIFIERS */ |
| 2819 | |
| 2820 | /** |
| 2821 | * prepare_task_switch - prepare to switch tasks |
| 2822 | * @rq: the runqueue preparing to switch |
| 2823 | * @prev: the current task that is being switched out |
| 2824 | * @next: the task we are going to switch to. |
| 2825 | * |
| 2826 | * This is called with the rq lock held and interrupts off. It must |
| 2827 | * be paired with a subsequent finish_task_switch after the context |
| 2828 | * switch. |
| 2829 | * |
| 2830 | * prepare_task_switch sets up locking and calls architecture specific |
| 2831 | * hooks. |
| 2832 | */ |
| 2833 | static inline void |
| 2834 | prepare_task_switch(struct rq *rq, struct task_struct *prev, |
| 2835 | struct task_struct *next) |
| 2836 | { |
| 2837 | fire_sched_out_preempt_notifiers(prev, next); |
| 2838 | prepare_lock_switch(rq, next); |
| 2839 | prepare_arch_switch(next); |
| 2840 | } |
| 2841 | |
| 2842 | /** |
| 2843 | * finish_task_switch - clean up after a task-switch |
| 2844 | * @rq: runqueue associated with task-switch |
| 2845 | * @prev: the thread we just switched away from. |
| 2846 | * |
| 2847 | * finish_task_switch must be called after the context switch, paired |
| 2848 | * with a prepare_task_switch call before the context switch. |
| 2849 | * finish_task_switch will reconcile locking set up by prepare_task_switch, |
| 2850 | * and do any other architecture-specific cleanup actions. |
| 2851 | * |
| 2852 | * Note that we may have delayed dropping an mm in context_switch(). If |
| 2853 | * so, we finish that here outside of the runqueue lock. (Doing it |
| 2854 | * with the lock held can cause deadlocks; see schedule() for |
| 2855 | * details.) |
| 2856 | */ |
| 2857 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) |
| 2858 | __releases(rq->lock) |
| 2859 | { |
| 2860 | struct mm_struct *mm = rq->prev_mm; |
| 2861 | long prev_state; |
| 2862 | |
| 2863 | rq->prev_mm = NULL; |
| 2864 | |
| 2865 | /* |
| 2866 | * A task struct has one reference for the use as "current". |
| 2867 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
| 2868 | * schedule one last time. The schedule call will never return, and |
| 2869 | * the scheduled task must drop that reference. |
| 2870 | * The test for TASK_DEAD must occur while the runqueue locks are |
| 2871 | * still held, otherwise prev could be scheduled on another cpu, die |
| 2872 | * there before we look at prev->state, and then the reference would |
| 2873 | * be dropped twice. |
| 2874 | * Manfred Spraul <manfred@colorfullife.com> |
| 2875 | */ |
| 2876 | prev_state = prev->state; |
| 2877 | finish_arch_switch(prev); |
| 2878 | perf_counter_task_sched_in(current, cpu_of(rq)); |
| 2879 | finish_lock_switch(rq, prev); |
| 2880 | |
| 2881 | fire_sched_in_preempt_notifiers(current); |
| 2882 | if (mm) |
| 2883 | mmdrop(mm); |
| 2884 | if (unlikely(prev_state == TASK_DEAD)) { |
| 2885 | /* |
| 2886 | * Remove function-return probe instances associated with this |
| 2887 | * task and put them back on the free list. |
| 2888 | */ |
| 2889 | kprobe_flush_task(prev); |
| 2890 | put_task_struct(prev); |
| 2891 | } |
| 2892 | } |
| 2893 | |
| 2894 | #ifdef CONFIG_SMP |
| 2895 | |
| 2896 | /* assumes rq->lock is held */ |
| 2897 | static inline void pre_schedule(struct rq *rq, struct task_struct *prev) |
| 2898 | { |
| 2899 | if (prev->sched_class->pre_schedule) |
| 2900 | prev->sched_class->pre_schedule(rq, prev); |
| 2901 | } |
| 2902 | |
| 2903 | /* rq->lock is NOT held, but preemption is disabled */ |
| 2904 | static inline void post_schedule(struct rq *rq) |
| 2905 | { |
| 2906 | if (rq->post_schedule) { |
| 2907 | unsigned long flags; |
| 2908 | |
| 2909 | spin_lock_irqsave(&rq->lock, flags); |
| 2910 | if (rq->curr->sched_class->post_schedule) |
| 2911 | rq->curr->sched_class->post_schedule(rq); |
| 2912 | spin_unlock_irqrestore(&rq->lock, flags); |
| 2913 | |
| 2914 | rq->post_schedule = 0; |
| 2915 | } |
| 2916 | } |
| 2917 | |
| 2918 | #else |
| 2919 | |
| 2920 | static inline void pre_schedule(struct rq *rq, struct task_struct *p) |
| 2921 | { |
| 2922 | } |
| 2923 | |
| 2924 | static inline void post_schedule(struct rq *rq) |
| 2925 | { |
| 2926 | } |
| 2927 | |
| 2928 | #endif |
| 2929 | |
| 2930 | /** |
| 2931 | * schedule_tail - first thing a freshly forked thread must call. |
| 2932 | * @prev: the thread we just switched away from. |
| 2933 | */ |
| 2934 | asmlinkage void schedule_tail(struct task_struct *prev) |
| 2935 | __releases(rq->lock) |
| 2936 | { |
| 2937 | struct rq *rq = this_rq(); |
| 2938 | |
| 2939 | finish_task_switch(rq, prev); |
| 2940 | |
| 2941 | /* |
| 2942 | * FIXME: do we need to worry about rq being invalidated by the |
| 2943 | * task_switch? |
| 2944 | */ |
| 2945 | post_schedule(rq); |
| 2946 | |
| 2947 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW |
| 2948 | /* In this case, finish_task_switch does not reenable preemption */ |
| 2949 | preempt_enable(); |
| 2950 | #endif |
| 2951 | if (current->set_child_tid) |
| 2952 | put_user(task_pid_vnr(current), current->set_child_tid); |
| 2953 | } |
| 2954 | |
| 2955 | /* |
| 2956 | * context_switch - switch to the new MM and the new |
| 2957 | * thread's register state. |
| 2958 | */ |
| 2959 | static inline void |
| 2960 | context_switch(struct rq *rq, struct task_struct *prev, |
| 2961 | struct task_struct *next) |
| 2962 | { |
| 2963 | struct mm_struct *mm, *oldmm; |
| 2964 | |
| 2965 | prepare_task_switch(rq, prev, next); |
| 2966 | trace_sched_switch(rq, prev, next); |
| 2967 | mm = next->mm; |
| 2968 | oldmm = prev->active_mm; |
| 2969 | /* |
| 2970 | * For paravirt, this is coupled with an exit in switch_to to |
| 2971 | * combine the page table reload and the switch backend into |
| 2972 | * one hypercall. |
| 2973 | */ |
| 2974 | arch_start_context_switch(prev); |
| 2975 | |
| 2976 | if (unlikely(!mm)) { |
| 2977 | next->active_mm = oldmm; |
| 2978 | atomic_inc(&oldmm->mm_count); |
| 2979 | enter_lazy_tlb(oldmm, next); |
| 2980 | } else |
| 2981 | switch_mm(oldmm, mm, next); |
| 2982 | |
| 2983 | if (unlikely(!prev->mm)) { |
| 2984 | prev->active_mm = NULL; |
| 2985 | rq->prev_mm = oldmm; |
| 2986 | } |
| 2987 | /* |
| 2988 | * Since the runqueue lock will be released by the next |
| 2989 | * task (which is an invalid locking op but in the case |
| 2990 | * of the scheduler it's an obvious special-case), so we |
| 2991 | * do an early lockdep release here: |
| 2992 | */ |
| 2993 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
| 2994 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
| 2995 | #endif |
| 2996 | |
| 2997 | /* Here we just switch the register state and the stack. */ |
| 2998 | switch_to(prev, next, prev); |
| 2999 | |
| 3000 | barrier(); |
| 3001 | /* |
| 3002 | * this_rq must be evaluated again because prev may have moved |
| 3003 | * CPUs since it called schedule(), thus the 'rq' on its stack |
| 3004 | * frame will be invalid. |
| 3005 | */ |
| 3006 | finish_task_switch(this_rq(), prev); |
| 3007 | } |
| 3008 | |
| 3009 | /* |
| 3010 | * nr_running, nr_uninterruptible and nr_context_switches: |
| 3011 | * |
| 3012 | * externally visible scheduler statistics: current number of runnable |
| 3013 | * threads, current number of uninterruptible-sleeping threads, total |
| 3014 | * number of context switches performed since bootup. |
| 3015 | */ |
| 3016 | unsigned long nr_running(void) |
| 3017 | { |
| 3018 | unsigned long i, sum = 0; |
| 3019 | |
| 3020 | for_each_online_cpu(i) |
| 3021 | sum += cpu_rq(i)->nr_running; |
| 3022 | |
| 3023 | return sum; |
| 3024 | } |
| 3025 | |
| 3026 | unsigned long nr_uninterruptible(void) |
| 3027 | { |
| 3028 | unsigned long i, sum = 0; |
| 3029 | |
| 3030 | for_each_possible_cpu(i) |
| 3031 | sum += cpu_rq(i)->nr_uninterruptible; |
| 3032 | |
| 3033 | /* |
| 3034 | * Since we read the counters lockless, it might be slightly |
| 3035 | * inaccurate. Do not allow it to go below zero though: |
| 3036 | */ |
| 3037 | if (unlikely((long)sum < 0)) |
| 3038 | sum = 0; |
| 3039 | |
| 3040 | return sum; |
| 3041 | } |
| 3042 | |
| 3043 | unsigned long long nr_context_switches(void) |
| 3044 | { |
| 3045 | int i; |
| 3046 | unsigned long long sum = 0; |
| 3047 | |
| 3048 | for_each_possible_cpu(i) |
| 3049 | sum += cpu_rq(i)->nr_switches; |
| 3050 | |
| 3051 | return sum; |
| 3052 | } |
| 3053 | |
| 3054 | unsigned long nr_iowait(void) |
| 3055 | { |
| 3056 | unsigned long i, sum = 0; |
| 3057 | |
| 3058 | for_each_possible_cpu(i) |
| 3059 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
| 3060 | |
| 3061 | return sum; |
| 3062 | } |
| 3063 | |
| 3064 | /* Variables and functions for calc_load */ |
| 3065 | static atomic_long_t calc_load_tasks; |
| 3066 | static unsigned long calc_load_update; |
| 3067 | unsigned long avenrun[3]; |
| 3068 | EXPORT_SYMBOL(avenrun); |
| 3069 | |
| 3070 | /** |
| 3071 | * get_avenrun - get the load average array |
| 3072 | * @loads: pointer to dest load array |
| 3073 | * @offset: offset to add |
| 3074 | * @shift: shift count to shift the result left |
| 3075 | * |
| 3076 | * These values are estimates at best, so no need for locking. |
| 3077 | */ |
| 3078 | void get_avenrun(unsigned long *loads, unsigned long offset, int shift) |
| 3079 | { |
| 3080 | loads[0] = (avenrun[0] + offset) << shift; |
| 3081 | loads[1] = (avenrun[1] + offset) << shift; |
| 3082 | loads[2] = (avenrun[2] + offset) << shift; |
| 3083 | } |
| 3084 | |
| 3085 | static unsigned long |
| 3086 | calc_load(unsigned long load, unsigned long exp, unsigned long active) |
| 3087 | { |
| 3088 | load *= exp; |
| 3089 | load += active * (FIXED_1 - exp); |
| 3090 | return load >> FSHIFT; |
| 3091 | } |
| 3092 | |
| 3093 | /* |
| 3094 | * calc_load - update the avenrun load estimates 10 ticks after the |
| 3095 | * CPUs have updated calc_load_tasks. |
| 3096 | */ |
| 3097 | void calc_global_load(void) |
| 3098 | { |
| 3099 | unsigned long upd = calc_load_update + 10; |
| 3100 | long active; |
| 3101 | |
| 3102 | if (time_before(jiffies, upd)) |
| 3103 | return; |
| 3104 | |
| 3105 | active = atomic_long_read(&calc_load_tasks); |
| 3106 | active = active > 0 ? active * FIXED_1 : 0; |
| 3107 | |
| 3108 | avenrun[0] = calc_load(avenrun[0], EXP_1, active); |
| 3109 | avenrun[1] = calc_load(avenrun[1], EXP_5, active); |
| 3110 | avenrun[2] = calc_load(avenrun[2], EXP_15, active); |
| 3111 | |
| 3112 | calc_load_update += LOAD_FREQ; |
| 3113 | } |
| 3114 | |
| 3115 | /* |
| 3116 | * Either called from update_cpu_load() or from a cpu going idle |
| 3117 | */ |
| 3118 | static void calc_load_account_active(struct rq *this_rq) |
| 3119 | { |
| 3120 | long nr_active, delta; |
| 3121 | |
| 3122 | nr_active = this_rq->nr_running; |
| 3123 | nr_active += (long) this_rq->nr_uninterruptible; |
| 3124 | |
| 3125 | if (nr_active != this_rq->calc_load_active) { |
| 3126 | delta = nr_active - this_rq->calc_load_active; |
| 3127 | this_rq->calc_load_active = nr_active; |
| 3128 | atomic_long_add(delta, &calc_load_tasks); |
| 3129 | } |
| 3130 | } |
| 3131 | |
| 3132 | /* |
| 3133 | * Externally visible per-cpu scheduler statistics: |
| 3134 | * cpu_nr_migrations(cpu) - number of migrations into that cpu |
| 3135 | */ |
| 3136 | u64 cpu_nr_migrations(int cpu) |
| 3137 | { |
| 3138 | return cpu_rq(cpu)->nr_migrations_in; |
| 3139 | } |
| 3140 | |
| 3141 | /* |
| 3142 | * Update rq->cpu_load[] statistics. This function is usually called every |
| 3143 | * scheduler tick (TICK_NSEC). |
| 3144 | */ |
| 3145 | static void update_cpu_load(struct rq *this_rq) |
| 3146 | { |
| 3147 | unsigned long this_load = this_rq->load.weight; |
| 3148 | int i, scale; |
| 3149 | |
| 3150 | this_rq->nr_load_updates++; |
| 3151 | |
| 3152 | /* Update our load: */ |
| 3153 | for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { |
| 3154 | unsigned long old_load, new_load; |
| 3155 | |
| 3156 | /* scale is effectively 1 << i now, and >> i divides by scale */ |
| 3157 | |
| 3158 | old_load = this_rq->cpu_load[i]; |
| 3159 | new_load = this_load; |
| 3160 | /* |
| 3161 | * Round up the averaging division if load is increasing. This |
| 3162 | * prevents us from getting stuck on 9 if the load is 10, for |
| 3163 | * example. |
| 3164 | */ |
| 3165 | if (new_load > old_load) |
| 3166 | new_load += scale-1; |
| 3167 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i; |
| 3168 | } |
| 3169 | |
| 3170 | if (time_after_eq(jiffies, this_rq->calc_load_update)) { |
| 3171 | this_rq->calc_load_update += LOAD_FREQ; |
| 3172 | calc_load_account_active(this_rq); |
| 3173 | } |
| 3174 | } |
| 3175 | |
| 3176 | #ifdef CONFIG_SMP |
| 3177 | |
| 3178 | /* |
| 3179 | * double_rq_lock - safely lock two runqueues |
| 3180 | * |
| 3181 | * Note this does not disable interrupts like task_rq_lock, |
| 3182 | * you need to do so manually before calling. |
| 3183 | */ |
| 3184 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) |
| 3185 | __acquires(rq1->lock) |
| 3186 | __acquires(rq2->lock) |
| 3187 | { |
| 3188 | BUG_ON(!irqs_disabled()); |
| 3189 | if (rq1 == rq2) { |
| 3190 | spin_lock(&rq1->lock); |
| 3191 | __acquire(rq2->lock); /* Fake it out ;) */ |
| 3192 | } else { |
| 3193 | if (rq1 < rq2) { |
| 3194 | spin_lock(&rq1->lock); |
| 3195 | spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); |
| 3196 | } else { |
| 3197 | spin_lock(&rq2->lock); |
| 3198 | spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); |
| 3199 | } |
| 3200 | } |
| 3201 | update_rq_clock(rq1); |
| 3202 | update_rq_clock(rq2); |
| 3203 | } |
| 3204 | |
| 3205 | /* |
| 3206 | * double_rq_unlock - safely unlock two runqueues |
| 3207 | * |
| 3208 | * Note this does not restore interrupts like task_rq_unlock, |
| 3209 | * you need to do so manually after calling. |
| 3210 | */ |
| 3211 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) |
| 3212 | __releases(rq1->lock) |
| 3213 | __releases(rq2->lock) |
| 3214 | { |
| 3215 | spin_unlock(&rq1->lock); |
| 3216 | if (rq1 != rq2) |
| 3217 | spin_unlock(&rq2->lock); |
| 3218 | else |
| 3219 | __release(rq2->lock); |
| 3220 | } |
| 3221 | |
| 3222 | /* |
| 3223 | * If dest_cpu is allowed for this process, migrate the task to it. |
| 3224 | * This is accomplished by forcing the cpu_allowed mask to only |
| 3225 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then |
| 3226 | * the cpu_allowed mask is restored. |
| 3227 | */ |
| 3228 | static void sched_migrate_task(struct task_struct *p, int dest_cpu) |
| 3229 | { |
| 3230 | struct migration_req req; |
| 3231 | unsigned long flags; |
| 3232 | struct rq *rq; |
| 3233 | |
| 3234 | rq = task_rq_lock(p, &flags); |
| 3235 | if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed) |
| 3236 | || unlikely(!cpu_active(dest_cpu))) |
| 3237 | goto out; |
| 3238 | |
| 3239 | /* force the process onto the specified CPU */ |
| 3240 | if (migrate_task(p, dest_cpu, &req)) { |
| 3241 | /* Need to wait for migration thread (might exit: take ref). */ |
| 3242 | struct task_struct *mt = rq->migration_thread; |
| 3243 | |
| 3244 | get_task_struct(mt); |
| 3245 | task_rq_unlock(rq, &flags); |
| 3246 | wake_up_process(mt); |
| 3247 | put_task_struct(mt); |
| 3248 | wait_for_completion(&req.done); |
| 3249 | |
| 3250 | return; |
| 3251 | } |
| 3252 | out: |
| 3253 | task_rq_unlock(rq, &flags); |
| 3254 | } |
| 3255 | |
| 3256 | /* |
| 3257 | * sched_exec - execve() is a valuable balancing opportunity, because at |
| 3258 | * this point the task has the smallest effective memory and cache footprint. |
| 3259 | */ |
| 3260 | void sched_exec(void) |
| 3261 | { |
| 3262 | int new_cpu, this_cpu = get_cpu(); |
| 3263 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
| 3264 | put_cpu(); |
| 3265 | if (new_cpu != this_cpu) |
| 3266 | sched_migrate_task(current, new_cpu); |
| 3267 | } |
| 3268 | |
| 3269 | /* |
| 3270 | * pull_task - move a task from a remote runqueue to the local runqueue. |
| 3271 | * Both runqueues must be locked. |
| 3272 | */ |
| 3273 | static void pull_task(struct rq *src_rq, struct task_struct *p, |
| 3274 | struct rq *this_rq, int this_cpu) |
| 3275 | { |
| 3276 | deactivate_task(src_rq, p, 0); |
| 3277 | set_task_cpu(p, this_cpu); |
| 3278 | activate_task(this_rq, p, 0); |
| 3279 | /* |
| 3280 | * Note that idle threads have a prio of MAX_PRIO, for this test |
| 3281 | * to be always true for them. |
| 3282 | */ |
| 3283 | check_preempt_curr(this_rq, p, 0); |
| 3284 | } |
| 3285 | |
| 3286 | /* |
| 3287 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? |
| 3288 | */ |
| 3289 | static |
| 3290 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, |
| 3291 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 3292 | int *all_pinned) |
| 3293 | { |
| 3294 | int tsk_cache_hot = 0; |
| 3295 | /* |
| 3296 | * We do not migrate tasks that are: |
| 3297 | * 1) running (obviously), or |
| 3298 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
| 3299 | * 3) are cache-hot on their current CPU. |
| 3300 | */ |
| 3301 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { |
| 3302 | schedstat_inc(p, se.nr_failed_migrations_affine); |
| 3303 | return 0; |
| 3304 | } |
| 3305 | *all_pinned = 0; |
| 3306 | |
| 3307 | if (task_running(rq, p)) { |
| 3308 | schedstat_inc(p, se.nr_failed_migrations_running); |
| 3309 | return 0; |
| 3310 | } |
| 3311 | |
| 3312 | /* |
| 3313 | * Aggressive migration if: |
| 3314 | * 1) task is cache cold, or |
| 3315 | * 2) too many balance attempts have failed. |
| 3316 | */ |
| 3317 | |
| 3318 | tsk_cache_hot = task_hot(p, rq->clock, sd); |
| 3319 | if (!tsk_cache_hot || |
| 3320 | sd->nr_balance_failed > sd->cache_nice_tries) { |
| 3321 | #ifdef CONFIG_SCHEDSTATS |
| 3322 | if (tsk_cache_hot) { |
| 3323 | schedstat_inc(sd, lb_hot_gained[idle]); |
| 3324 | schedstat_inc(p, se.nr_forced_migrations); |
| 3325 | } |
| 3326 | #endif |
| 3327 | return 1; |
| 3328 | } |
| 3329 | |
| 3330 | if (tsk_cache_hot) { |
| 3331 | schedstat_inc(p, se.nr_failed_migrations_hot); |
| 3332 | return 0; |
| 3333 | } |
| 3334 | return 1; |
| 3335 | } |
| 3336 | |
| 3337 | static unsigned long |
| 3338 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 3339 | unsigned long max_load_move, struct sched_domain *sd, |
| 3340 | enum cpu_idle_type idle, int *all_pinned, |
| 3341 | int *this_best_prio, struct rq_iterator *iterator) |
| 3342 | { |
| 3343 | int loops = 0, pulled = 0, pinned = 0; |
| 3344 | struct task_struct *p; |
| 3345 | long rem_load_move = max_load_move; |
| 3346 | |
| 3347 | if (max_load_move == 0) |
| 3348 | goto out; |
| 3349 | |
| 3350 | pinned = 1; |
| 3351 | |
| 3352 | /* |
| 3353 | * Start the load-balancing iterator: |
| 3354 | */ |
| 3355 | p = iterator->start(iterator->arg); |
| 3356 | next: |
| 3357 | if (!p || loops++ > sysctl_sched_nr_migrate) |
| 3358 | goto out; |
| 3359 | |
| 3360 | if ((p->se.load.weight >> 1) > rem_load_move || |
| 3361 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { |
| 3362 | p = iterator->next(iterator->arg); |
| 3363 | goto next; |
| 3364 | } |
| 3365 | |
| 3366 | pull_task(busiest, p, this_rq, this_cpu); |
| 3367 | pulled++; |
| 3368 | rem_load_move -= p->se.load.weight; |
| 3369 | |
| 3370 | #ifdef CONFIG_PREEMPT |
| 3371 | /* |
| 3372 | * NEWIDLE balancing is a source of latency, so preemptible kernels |
| 3373 | * will stop after the first task is pulled to minimize the critical |
| 3374 | * section. |
| 3375 | */ |
| 3376 | if (idle == CPU_NEWLY_IDLE) |
| 3377 | goto out; |
| 3378 | #endif |
| 3379 | |
| 3380 | /* |
| 3381 | * We only want to steal up to the prescribed amount of weighted load. |
| 3382 | */ |
| 3383 | if (rem_load_move > 0) { |
| 3384 | if (p->prio < *this_best_prio) |
| 3385 | *this_best_prio = p->prio; |
| 3386 | p = iterator->next(iterator->arg); |
| 3387 | goto next; |
| 3388 | } |
| 3389 | out: |
| 3390 | /* |
| 3391 | * Right now, this is one of only two places pull_task() is called, |
| 3392 | * so we can safely collect pull_task() stats here rather than |
| 3393 | * inside pull_task(). |
| 3394 | */ |
| 3395 | schedstat_add(sd, lb_gained[idle], pulled); |
| 3396 | |
| 3397 | if (all_pinned) |
| 3398 | *all_pinned = pinned; |
| 3399 | |
| 3400 | return max_load_move - rem_load_move; |
| 3401 | } |
| 3402 | |
| 3403 | /* |
| 3404 | * move_tasks tries to move up to max_load_move weighted load from busiest to |
| 3405 | * this_rq, as part of a balancing operation within domain "sd". |
| 3406 | * Returns 1 if successful and 0 otherwise. |
| 3407 | * |
| 3408 | * Called with both runqueues locked. |
| 3409 | */ |
| 3410 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 3411 | unsigned long max_load_move, |
| 3412 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 3413 | int *all_pinned) |
| 3414 | { |
| 3415 | const struct sched_class *class = sched_class_highest; |
| 3416 | unsigned long total_load_moved = 0; |
| 3417 | int this_best_prio = this_rq->curr->prio; |
| 3418 | |
| 3419 | do { |
| 3420 | total_load_moved += |
| 3421 | class->load_balance(this_rq, this_cpu, busiest, |
| 3422 | max_load_move - total_load_moved, |
| 3423 | sd, idle, all_pinned, &this_best_prio); |
| 3424 | class = class->next; |
| 3425 | |
| 3426 | #ifdef CONFIG_PREEMPT |
| 3427 | /* |
| 3428 | * NEWIDLE balancing is a source of latency, so preemptible |
| 3429 | * kernels will stop after the first task is pulled to minimize |
| 3430 | * the critical section. |
| 3431 | */ |
| 3432 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) |
| 3433 | break; |
| 3434 | #endif |
| 3435 | } while (class && max_load_move > total_load_moved); |
| 3436 | |
| 3437 | return total_load_moved > 0; |
| 3438 | } |
| 3439 | |
| 3440 | static int |
| 3441 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 3442 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 3443 | struct rq_iterator *iterator) |
| 3444 | { |
| 3445 | struct task_struct *p = iterator->start(iterator->arg); |
| 3446 | int pinned = 0; |
| 3447 | |
| 3448 | while (p) { |
| 3449 | if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { |
| 3450 | pull_task(busiest, p, this_rq, this_cpu); |
| 3451 | /* |
| 3452 | * Right now, this is only the second place pull_task() |
| 3453 | * is called, so we can safely collect pull_task() |
| 3454 | * stats here rather than inside pull_task(). |
| 3455 | */ |
| 3456 | schedstat_inc(sd, lb_gained[idle]); |
| 3457 | |
| 3458 | return 1; |
| 3459 | } |
| 3460 | p = iterator->next(iterator->arg); |
| 3461 | } |
| 3462 | |
| 3463 | return 0; |
| 3464 | } |
| 3465 | |
| 3466 | /* |
| 3467 | * move_one_task tries to move exactly one task from busiest to this_rq, as |
| 3468 | * part of active balancing operations within "domain". |
| 3469 | * Returns 1 if successful and 0 otherwise. |
| 3470 | * |
| 3471 | * Called with both runqueues locked. |
| 3472 | */ |
| 3473 | static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 3474 | struct sched_domain *sd, enum cpu_idle_type idle) |
| 3475 | { |
| 3476 | const struct sched_class *class; |
| 3477 | |
| 3478 | for_each_class(class) { |
| 3479 | if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle)) |
| 3480 | return 1; |
| 3481 | } |
| 3482 | |
| 3483 | return 0; |
| 3484 | } |
| 3485 | /********** Helpers for find_busiest_group ************************/ |
| 3486 | /* |
| 3487 | * sd_lb_stats - Structure to store the statistics of a sched_domain |
| 3488 | * during load balancing. |
| 3489 | */ |
| 3490 | struct sd_lb_stats { |
| 3491 | struct sched_group *busiest; /* Busiest group in this sd */ |
| 3492 | struct sched_group *this; /* Local group in this sd */ |
| 3493 | unsigned long total_load; /* Total load of all groups in sd */ |
| 3494 | unsigned long total_pwr; /* Total power of all groups in sd */ |
| 3495 | unsigned long avg_load; /* Average load across all groups in sd */ |
| 3496 | |
| 3497 | /** Statistics of this group */ |
| 3498 | unsigned long this_load; |
| 3499 | unsigned long this_load_per_task; |
| 3500 | unsigned long this_nr_running; |
| 3501 | |
| 3502 | /* Statistics of the busiest group */ |
| 3503 | unsigned long max_load; |
| 3504 | unsigned long busiest_load_per_task; |
| 3505 | unsigned long busiest_nr_running; |
| 3506 | |
| 3507 | int group_imb; /* Is there imbalance in this sd */ |
| 3508 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
| 3509 | int power_savings_balance; /* Is powersave balance needed for this sd */ |
| 3510 | struct sched_group *group_min; /* Least loaded group in sd */ |
| 3511 | struct sched_group *group_leader; /* Group which relieves group_min */ |
| 3512 | unsigned long min_load_per_task; /* load_per_task in group_min */ |
| 3513 | unsigned long leader_nr_running; /* Nr running of group_leader */ |
| 3514 | unsigned long min_nr_running; /* Nr running of group_min */ |
| 3515 | #endif |
| 3516 | }; |
| 3517 | |
| 3518 | /* |
| 3519 | * sg_lb_stats - stats of a sched_group required for load_balancing |
| 3520 | */ |
| 3521 | struct sg_lb_stats { |
| 3522 | unsigned long avg_load; /*Avg load across the CPUs of the group */ |
| 3523 | unsigned long group_load; /* Total load over the CPUs of the group */ |
| 3524 | unsigned long sum_nr_running; /* Nr tasks running in the group */ |
| 3525 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
| 3526 | unsigned long group_capacity; |
| 3527 | int group_imb; /* Is there an imbalance in the group ? */ |
| 3528 | }; |
| 3529 | |
| 3530 | /** |
| 3531 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. |
| 3532 | * @group: The group whose first cpu is to be returned. |
| 3533 | */ |
| 3534 | static inline unsigned int group_first_cpu(struct sched_group *group) |
| 3535 | { |
| 3536 | return cpumask_first(sched_group_cpus(group)); |
| 3537 | } |
| 3538 | |
| 3539 | /** |
| 3540 | * get_sd_load_idx - Obtain the load index for a given sched domain. |
| 3541 | * @sd: The sched_domain whose load_idx is to be obtained. |
| 3542 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. |
| 3543 | */ |
| 3544 | static inline int get_sd_load_idx(struct sched_domain *sd, |
| 3545 | enum cpu_idle_type idle) |
| 3546 | { |
| 3547 | int load_idx; |
| 3548 | |
| 3549 | switch (idle) { |
| 3550 | case CPU_NOT_IDLE: |
| 3551 | load_idx = sd->busy_idx; |
| 3552 | break; |
| 3553 | |
| 3554 | case CPU_NEWLY_IDLE: |
| 3555 | load_idx = sd->newidle_idx; |
| 3556 | break; |
| 3557 | default: |
| 3558 | load_idx = sd->idle_idx; |
| 3559 | break; |
| 3560 | } |
| 3561 | |
| 3562 | return load_idx; |
| 3563 | } |
| 3564 | |
| 3565 | |
| 3566 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
| 3567 | /** |
| 3568 | * init_sd_power_savings_stats - Initialize power savings statistics for |
| 3569 | * the given sched_domain, during load balancing. |
| 3570 | * |
| 3571 | * @sd: Sched domain whose power-savings statistics are to be initialized. |
| 3572 | * @sds: Variable containing the statistics for sd. |
| 3573 | * @idle: Idle status of the CPU at which we're performing load-balancing. |
| 3574 | */ |
| 3575 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, |
| 3576 | struct sd_lb_stats *sds, enum cpu_idle_type idle) |
| 3577 | { |
| 3578 | /* |
| 3579 | * Busy processors will not participate in power savings |
| 3580 | * balance. |
| 3581 | */ |
| 3582 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) |
| 3583 | sds->power_savings_balance = 0; |
| 3584 | else { |
| 3585 | sds->power_savings_balance = 1; |
| 3586 | sds->min_nr_running = ULONG_MAX; |
| 3587 | sds->leader_nr_running = 0; |
| 3588 | } |
| 3589 | } |
| 3590 | |
| 3591 | /** |
| 3592 | * update_sd_power_savings_stats - Update the power saving stats for a |
| 3593 | * sched_domain while performing load balancing. |
| 3594 | * |
| 3595 | * @group: sched_group belonging to the sched_domain under consideration. |
| 3596 | * @sds: Variable containing the statistics of the sched_domain |
| 3597 | * @local_group: Does group contain the CPU for which we're performing |
| 3598 | * load balancing ? |
| 3599 | * @sgs: Variable containing the statistics of the group. |
| 3600 | */ |
| 3601 | static inline void update_sd_power_savings_stats(struct sched_group *group, |
| 3602 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) |
| 3603 | { |
| 3604 | |
| 3605 | if (!sds->power_savings_balance) |
| 3606 | return; |
| 3607 | |
| 3608 | /* |
| 3609 | * If the local group is idle or completely loaded |
| 3610 | * no need to do power savings balance at this domain |
| 3611 | */ |
| 3612 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || |
| 3613 | !sds->this_nr_running)) |
| 3614 | sds->power_savings_balance = 0; |
| 3615 | |
| 3616 | /* |
| 3617 | * If a group is already running at full capacity or idle, |
| 3618 | * don't include that group in power savings calculations |
| 3619 | */ |
| 3620 | if (!sds->power_savings_balance || |
| 3621 | sgs->sum_nr_running >= sgs->group_capacity || |
| 3622 | !sgs->sum_nr_running) |
| 3623 | return; |
| 3624 | |
| 3625 | /* |
| 3626 | * Calculate the group which has the least non-idle load. |
| 3627 | * This is the group from where we need to pick up the load |
| 3628 | * for saving power |
| 3629 | */ |
| 3630 | if ((sgs->sum_nr_running < sds->min_nr_running) || |
| 3631 | (sgs->sum_nr_running == sds->min_nr_running && |
| 3632 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { |
| 3633 | sds->group_min = group; |
| 3634 | sds->min_nr_running = sgs->sum_nr_running; |
| 3635 | sds->min_load_per_task = sgs->sum_weighted_load / |
| 3636 | sgs->sum_nr_running; |
| 3637 | } |
| 3638 | |
| 3639 | /* |
| 3640 | * Calculate the group which is almost near its |
| 3641 | * capacity but still has some space to pick up some load |
| 3642 | * from other group and save more power |
| 3643 | */ |
| 3644 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) |
| 3645 | return; |
| 3646 | |
| 3647 | if (sgs->sum_nr_running > sds->leader_nr_running || |
| 3648 | (sgs->sum_nr_running == sds->leader_nr_running && |
| 3649 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { |
| 3650 | sds->group_leader = group; |
| 3651 | sds->leader_nr_running = sgs->sum_nr_running; |
| 3652 | } |
| 3653 | } |
| 3654 | |
| 3655 | /** |
| 3656 | * check_power_save_busiest_group - see if there is potential for some power-savings balance |
| 3657 | * @sds: Variable containing the statistics of the sched_domain |
| 3658 | * under consideration. |
| 3659 | * @this_cpu: Cpu at which we're currently performing load-balancing. |
| 3660 | * @imbalance: Variable to store the imbalance. |
| 3661 | * |
| 3662 | * Description: |
| 3663 | * Check if we have potential to perform some power-savings balance. |
| 3664 | * If yes, set the busiest group to be the least loaded group in the |
| 3665 | * sched_domain, so that it's CPUs can be put to idle. |
| 3666 | * |
| 3667 | * Returns 1 if there is potential to perform power-savings balance. |
| 3668 | * Else returns 0. |
| 3669 | */ |
| 3670 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, |
| 3671 | int this_cpu, unsigned long *imbalance) |
| 3672 | { |
| 3673 | if (!sds->power_savings_balance) |
| 3674 | return 0; |
| 3675 | |
| 3676 | if (sds->this != sds->group_leader || |
| 3677 | sds->group_leader == sds->group_min) |
| 3678 | return 0; |
| 3679 | |
| 3680 | *imbalance = sds->min_load_per_task; |
| 3681 | sds->busiest = sds->group_min; |
| 3682 | |
| 3683 | if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) { |
| 3684 | cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu = |
| 3685 | group_first_cpu(sds->group_leader); |
| 3686 | } |
| 3687 | |
| 3688 | return 1; |
| 3689 | |
| 3690 | } |
| 3691 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ |
| 3692 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, |
| 3693 | struct sd_lb_stats *sds, enum cpu_idle_type idle) |
| 3694 | { |
| 3695 | return; |
| 3696 | } |
| 3697 | |
| 3698 | static inline void update_sd_power_savings_stats(struct sched_group *group, |
| 3699 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) |
| 3700 | { |
| 3701 | return; |
| 3702 | } |
| 3703 | |
| 3704 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, |
| 3705 | int this_cpu, unsigned long *imbalance) |
| 3706 | { |
| 3707 | return 0; |
| 3708 | } |
| 3709 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ |
| 3710 | |
| 3711 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) |
| 3712 | { |
| 3713 | unsigned long weight = cpumask_weight(sched_domain_span(sd)); |
| 3714 | unsigned long smt_gain = sd->smt_gain; |
| 3715 | |
| 3716 | smt_gain /= weight; |
| 3717 | |
| 3718 | return smt_gain; |
| 3719 | } |
| 3720 | |
| 3721 | unsigned long scale_rt_power(int cpu) |
| 3722 | { |
| 3723 | struct rq *rq = cpu_rq(cpu); |
| 3724 | u64 total, available; |
| 3725 | |
| 3726 | sched_avg_update(rq); |
| 3727 | |
| 3728 | total = sched_avg_period() + (rq->clock - rq->age_stamp); |
| 3729 | available = total - rq->rt_avg; |
| 3730 | |
| 3731 | if (unlikely((s64)total < SCHED_LOAD_SCALE)) |
| 3732 | total = SCHED_LOAD_SCALE; |
| 3733 | |
| 3734 | total >>= SCHED_LOAD_SHIFT; |
| 3735 | |
| 3736 | return div_u64(available, total); |
| 3737 | } |
| 3738 | |
| 3739 | static void update_cpu_power(struct sched_domain *sd, int cpu) |
| 3740 | { |
| 3741 | unsigned long weight = cpumask_weight(sched_domain_span(sd)); |
| 3742 | unsigned long power = SCHED_LOAD_SCALE; |
| 3743 | struct sched_group *sdg = sd->groups; |
| 3744 | |
| 3745 | /* here we could scale based on cpufreq */ |
| 3746 | |
| 3747 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
| 3748 | power *= arch_scale_smt_power(sd, cpu); |
| 3749 | power >>= SCHED_LOAD_SHIFT; |
| 3750 | } |
| 3751 | |
| 3752 | power *= scale_rt_power(cpu); |
| 3753 | power >>= SCHED_LOAD_SHIFT; |
| 3754 | |
| 3755 | if (!power) |
| 3756 | power = 1; |
| 3757 | |
| 3758 | sdg->cpu_power = power; |
| 3759 | } |
| 3760 | |
| 3761 | static void update_group_power(struct sched_domain *sd, int cpu) |
| 3762 | { |
| 3763 | struct sched_domain *child = sd->child; |
| 3764 | struct sched_group *group, *sdg = sd->groups; |
| 3765 | unsigned long power; |
| 3766 | |
| 3767 | if (!child) { |
| 3768 | update_cpu_power(sd, cpu); |
| 3769 | return; |
| 3770 | } |
| 3771 | |
| 3772 | power = 0; |
| 3773 | |
| 3774 | group = child->groups; |
| 3775 | do { |
| 3776 | power += group->cpu_power; |
| 3777 | group = group->next; |
| 3778 | } while (group != child->groups); |
| 3779 | |
| 3780 | sdg->cpu_power = power; |
| 3781 | } |
| 3782 | |
| 3783 | /** |
| 3784 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. |
| 3785 | * @group: sched_group whose statistics are to be updated. |
| 3786 | * @this_cpu: Cpu for which load balance is currently performed. |
| 3787 | * @idle: Idle status of this_cpu |
| 3788 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
| 3789 | * @sd_idle: Idle status of the sched_domain containing group. |
| 3790 | * @local_group: Does group contain this_cpu. |
| 3791 | * @cpus: Set of cpus considered for load balancing. |
| 3792 | * @balance: Should we balance. |
| 3793 | * @sgs: variable to hold the statistics for this group. |
| 3794 | */ |
| 3795 | static inline void update_sg_lb_stats(struct sched_domain *sd, |
| 3796 | struct sched_group *group, int this_cpu, |
| 3797 | enum cpu_idle_type idle, int load_idx, int *sd_idle, |
| 3798 | int local_group, const struct cpumask *cpus, |
| 3799 | int *balance, struct sg_lb_stats *sgs) |
| 3800 | { |
| 3801 | unsigned long load, max_cpu_load, min_cpu_load; |
| 3802 | int i; |
| 3803 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
| 3804 | unsigned long sum_avg_load_per_task; |
| 3805 | unsigned long avg_load_per_task; |
| 3806 | |
| 3807 | if (local_group) { |
| 3808 | balance_cpu = group_first_cpu(group); |
| 3809 | if (balance_cpu == this_cpu) |
| 3810 | update_group_power(sd, this_cpu); |
| 3811 | } |
| 3812 | |
| 3813 | /* Tally up the load of all CPUs in the group */ |
| 3814 | sum_avg_load_per_task = avg_load_per_task = 0; |
| 3815 | max_cpu_load = 0; |
| 3816 | min_cpu_load = ~0UL; |
| 3817 | |
| 3818 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { |
| 3819 | struct rq *rq = cpu_rq(i); |
| 3820 | |
| 3821 | if (*sd_idle && rq->nr_running) |
| 3822 | *sd_idle = 0; |
| 3823 | |
| 3824 | /* Bias balancing toward cpus of our domain */ |
| 3825 | if (local_group) { |
| 3826 | if (idle_cpu(i) && !first_idle_cpu) { |
| 3827 | first_idle_cpu = 1; |
| 3828 | balance_cpu = i; |
| 3829 | } |
| 3830 | |
| 3831 | load = target_load(i, load_idx); |
| 3832 | } else { |
| 3833 | load = source_load(i, load_idx); |
| 3834 | if (load > max_cpu_load) |
| 3835 | max_cpu_load = load; |
| 3836 | if (min_cpu_load > load) |
| 3837 | min_cpu_load = load; |
| 3838 | } |
| 3839 | |
| 3840 | sgs->group_load += load; |
| 3841 | sgs->sum_nr_running += rq->nr_running; |
| 3842 | sgs->sum_weighted_load += weighted_cpuload(i); |
| 3843 | |
| 3844 | sum_avg_load_per_task += cpu_avg_load_per_task(i); |
| 3845 | } |
| 3846 | |
| 3847 | /* |
| 3848 | * First idle cpu or the first cpu(busiest) in this sched group |
| 3849 | * is eligible for doing load balancing at this and above |
| 3850 | * domains. In the newly idle case, we will allow all the cpu's |
| 3851 | * to do the newly idle load balance. |
| 3852 | */ |
| 3853 | if (idle != CPU_NEWLY_IDLE && local_group && |
| 3854 | balance_cpu != this_cpu && balance) { |
| 3855 | *balance = 0; |
| 3856 | return; |
| 3857 | } |
| 3858 | |
| 3859 | /* Adjust by relative CPU power of the group */ |
| 3860 | sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power; |
| 3861 | |
| 3862 | |
| 3863 | /* |
| 3864 | * Consider the group unbalanced when the imbalance is larger |
| 3865 | * than the average weight of two tasks. |
| 3866 | * |
| 3867 | * APZ: with cgroup the avg task weight can vary wildly and |
| 3868 | * might not be a suitable number - should we keep a |
| 3869 | * normalized nr_running number somewhere that negates |
| 3870 | * the hierarchy? |
| 3871 | */ |
| 3872 | avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) / |
| 3873 | group->cpu_power; |
| 3874 | |
| 3875 | if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task) |
| 3876 | sgs->group_imb = 1; |
| 3877 | |
| 3878 | sgs->group_capacity = |
| 3879 | DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE); |
| 3880 | } |
| 3881 | |
| 3882 | /** |
| 3883 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. |
| 3884 | * @sd: sched_domain whose statistics are to be updated. |
| 3885 | * @this_cpu: Cpu for which load balance is currently performed. |
| 3886 | * @idle: Idle status of this_cpu |
| 3887 | * @sd_idle: Idle status of the sched_domain containing group. |
| 3888 | * @cpus: Set of cpus considered for load balancing. |
| 3889 | * @balance: Should we balance. |
| 3890 | * @sds: variable to hold the statistics for this sched_domain. |
| 3891 | */ |
| 3892 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, |
| 3893 | enum cpu_idle_type idle, int *sd_idle, |
| 3894 | const struct cpumask *cpus, int *balance, |
| 3895 | struct sd_lb_stats *sds) |
| 3896 | { |
| 3897 | struct sched_domain *child = sd->child; |
| 3898 | struct sched_group *group = sd->groups; |
| 3899 | struct sg_lb_stats sgs; |
| 3900 | int load_idx, prefer_sibling = 0; |
| 3901 | |
| 3902 | if (child && child->flags & SD_PREFER_SIBLING) |
| 3903 | prefer_sibling = 1; |
| 3904 | |
| 3905 | init_sd_power_savings_stats(sd, sds, idle); |
| 3906 | load_idx = get_sd_load_idx(sd, idle); |
| 3907 | |
| 3908 | do { |
| 3909 | int local_group; |
| 3910 | |
| 3911 | local_group = cpumask_test_cpu(this_cpu, |
| 3912 | sched_group_cpus(group)); |
| 3913 | memset(&sgs, 0, sizeof(sgs)); |
| 3914 | update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle, |
| 3915 | local_group, cpus, balance, &sgs); |
| 3916 | |
| 3917 | if (local_group && balance && !(*balance)) |
| 3918 | return; |
| 3919 | |
| 3920 | sds->total_load += sgs.group_load; |
| 3921 | sds->total_pwr += group->cpu_power; |
| 3922 | |
| 3923 | /* |
| 3924 | * In case the child domain prefers tasks go to siblings |
| 3925 | * first, lower the group capacity to one so that we'll try |
| 3926 | * and move all the excess tasks away. |
| 3927 | */ |
| 3928 | if (prefer_sibling) |
| 3929 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
| 3930 | |
| 3931 | if (local_group) { |
| 3932 | sds->this_load = sgs.avg_load; |
| 3933 | sds->this = group; |
| 3934 | sds->this_nr_running = sgs.sum_nr_running; |
| 3935 | sds->this_load_per_task = sgs.sum_weighted_load; |
| 3936 | } else if (sgs.avg_load > sds->max_load && |
| 3937 | (sgs.sum_nr_running > sgs.group_capacity || |
| 3938 | sgs.group_imb)) { |
| 3939 | sds->max_load = sgs.avg_load; |
| 3940 | sds->busiest = group; |
| 3941 | sds->busiest_nr_running = sgs.sum_nr_running; |
| 3942 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
| 3943 | sds->group_imb = sgs.group_imb; |
| 3944 | } |
| 3945 | |
| 3946 | update_sd_power_savings_stats(group, sds, local_group, &sgs); |
| 3947 | group = group->next; |
| 3948 | } while (group != sd->groups); |
| 3949 | } |
| 3950 | |
| 3951 | /** |
| 3952 | * fix_small_imbalance - Calculate the minor imbalance that exists |
| 3953 | * amongst the groups of a sched_domain, during |
| 3954 | * load balancing. |
| 3955 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
| 3956 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. |
| 3957 | * @imbalance: Variable to store the imbalance. |
| 3958 | */ |
| 3959 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, |
| 3960 | int this_cpu, unsigned long *imbalance) |
| 3961 | { |
| 3962 | unsigned long tmp, pwr_now = 0, pwr_move = 0; |
| 3963 | unsigned int imbn = 2; |
| 3964 | |
| 3965 | if (sds->this_nr_running) { |
| 3966 | sds->this_load_per_task /= sds->this_nr_running; |
| 3967 | if (sds->busiest_load_per_task > |
| 3968 | sds->this_load_per_task) |
| 3969 | imbn = 1; |
| 3970 | } else |
| 3971 | sds->this_load_per_task = |
| 3972 | cpu_avg_load_per_task(this_cpu); |
| 3973 | |
| 3974 | if (sds->max_load - sds->this_load + sds->busiest_load_per_task >= |
| 3975 | sds->busiest_load_per_task * imbn) { |
| 3976 | *imbalance = sds->busiest_load_per_task; |
| 3977 | return; |
| 3978 | } |
| 3979 | |
| 3980 | /* |
| 3981 | * OK, we don't have enough imbalance to justify moving tasks, |
| 3982 | * however we may be able to increase total CPU power used by |
| 3983 | * moving them. |
| 3984 | */ |
| 3985 | |
| 3986 | pwr_now += sds->busiest->cpu_power * |
| 3987 | min(sds->busiest_load_per_task, sds->max_load); |
| 3988 | pwr_now += sds->this->cpu_power * |
| 3989 | min(sds->this_load_per_task, sds->this_load); |
| 3990 | pwr_now /= SCHED_LOAD_SCALE; |
| 3991 | |
| 3992 | /* Amount of load we'd subtract */ |
| 3993 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / |
| 3994 | sds->busiest->cpu_power; |
| 3995 | if (sds->max_load > tmp) |
| 3996 | pwr_move += sds->busiest->cpu_power * |
| 3997 | min(sds->busiest_load_per_task, sds->max_load - tmp); |
| 3998 | |
| 3999 | /* Amount of load we'd add */ |
| 4000 | if (sds->max_load * sds->busiest->cpu_power < |
| 4001 | sds->busiest_load_per_task * SCHED_LOAD_SCALE) |
| 4002 | tmp = (sds->max_load * sds->busiest->cpu_power) / |
| 4003 | sds->this->cpu_power; |
| 4004 | else |
| 4005 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / |
| 4006 | sds->this->cpu_power; |
| 4007 | pwr_move += sds->this->cpu_power * |
| 4008 | min(sds->this_load_per_task, sds->this_load + tmp); |
| 4009 | pwr_move /= SCHED_LOAD_SCALE; |
| 4010 | |
| 4011 | /* Move if we gain throughput */ |
| 4012 | if (pwr_move > pwr_now) |
| 4013 | *imbalance = sds->busiest_load_per_task; |
| 4014 | } |
| 4015 | |
| 4016 | /** |
| 4017 | * calculate_imbalance - Calculate the amount of imbalance present within the |
| 4018 | * groups of a given sched_domain during load balance. |
| 4019 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
| 4020 | * @this_cpu: Cpu for which currently load balance is being performed. |
| 4021 | * @imbalance: The variable to store the imbalance. |
| 4022 | */ |
| 4023 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, |
| 4024 | unsigned long *imbalance) |
| 4025 | { |
| 4026 | unsigned long max_pull; |
| 4027 | /* |
| 4028 | * In the presence of smp nice balancing, certain scenarios can have |
| 4029 | * max load less than avg load(as we skip the groups at or below |
| 4030 | * its cpu_power, while calculating max_load..) |
| 4031 | */ |
| 4032 | if (sds->max_load < sds->avg_load) { |
| 4033 | *imbalance = 0; |
| 4034 | return fix_small_imbalance(sds, this_cpu, imbalance); |
| 4035 | } |
| 4036 | |
| 4037 | /* Don't want to pull so many tasks that a group would go idle */ |
| 4038 | max_pull = min(sds->max_load - sds->avg_load, |
| 4039 | sds->max_load - sds->busiest_load_per_task); |
| 4040 | |
| 4041 | /* How much load to actually move to equalise the imbalance */ |
| 4042 | *imbalance = min(max_pull * sds->busiest->cpu_power, |
| 4043 | (sds->avg_load - sds->this_load) * sds->this->cpu_power) |
| 4044 | / SCHED_LOAD_SCALE; |
| 4045 | |
| 4046 | /* |
| 4047 | * if *imbalance is less than the average load per runnable task |
| 4048 | * there is no gaurantee that any tasks will be moved so we'll have |
| 4049 | * a think about bumping its value to force at least one task to be |
| 4050 | * moved |
| 4051 | */ |
| 4052 | if (*imbalance < sds->busiest_load_per_task) |
| 4053 | return fix_small_imbalance(sds, this_cpu, imbalance); |
| 4054 | |
| 4055 | } |
| 4056 | /******* find_busiest_group() helpers end here *********************/ |
| 4057 | |
| 4058 | /** |
| 4059 | * find_busiest_group - Returns the busiest group within the sched_domain |
| 4060 | * if there is an imbalance. If there isn't an imbalance, and |
| 4061 | * the user has opted for power-savings, it returns a group whose |
| 4062 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if |
| 4063 | * such a group exists. |
| 4064 | * |
| 4065 | * Also calculates the amount of weighted load which should be moved |
| 4066 | * to restore balance. |
| 4067 | * |
| 4068 | * @sd: The sched_domain whose busiest group is to be returned. |
| 4069 | * @this_cpu: The cpu for which load balancing is currently being performed. |
| 4070 | * @imbalance: Variable which stores amount of weighted load which should |
| 4071 | * be moved to restore balance/put a group to idle. |
| 4072 | * @idle: The idle status of this_cpu. |
| 4073 | * @sd_idle: The idleness of sd |
| 4074 | * @cpus: The set of CPUs under consideration for load-balancing. |
| 4075 | * @balance: Pointer to a variable indicating if this_cpu |
| 4076 | * is the appropriate cpu to perform load balancing at this_level. |
| 4077 | * |
| 4078 | * Returns: - the busiest group if imbalance exists. |
| 4079 | * - If no imbalance and user has opted for power-savings balance, |
| 4080 | * return the least loaded group whose CPUs can be |
| 4081 | * put to idle by rebalancing its tasks onto our group. |
| 4082 | */ |
| 4083 | static struct sched_group * |
| 4084 | find_busiest_group(struct sched_domain *sd, int this_cpu, |
| 4085 | unsigned long *imbalance, enum cpu_idle_type idle, |
| 4086 | int *sd_idle, const struct cpumask *cpus, int *balance) |
| 4087 | { |
| 4088 | struct sd_lb_stats sds; |
| 4089 | |
| 4090 | memset(&sds, 0, sizeof(sds)); |
| 4091 | |
| 4092 | /* |
| 4093 | * Compute the various statistics relavent for load balancing at |
| 4094 | * this level. |
| 4095 | */ |
| 4096 | update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus, |
| 4097 | balance, &sds); |
| 4098 | |
| 4099 | /* Cases where imbalance does not exist from POV of this_cpu */ |
| 4100 | /* 1) this_cpu is not the appropriate cpu to perform load balancing |
| 4101 | * at this level. |
| 4102 | * 2) There is no busy sibling group to pull from. |
| 4103 | * 3) This group is the busiest group. |
| 4104 | * 4) This group is more busy than the avg busieness at this |
| 4105 | * sched_domain. |
| 4106 | * 5) The imbalance is within the specified limit. |
| 4107 | * 6) Any rebalance would lead to ping-pong |
| 4108 | */ |
| 4109 | if (balance && !(*balance)) |
| 4110 | goto ret; |
| 4111 | |
| 4112 | if (!sds.busiest || sds.busiest_nr_running == 0) |
| 4113 | goto out_balanced; |
| 4114 | |
| 4115 | if (sds.this_load >= sds.max_load) |
| 4116 | goto out_balanced; |
| 4117 | |
| 4118 | sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr; |
| 4119 | |
| 4120 | if (sds.this_load >= sds.avg_load) |
| 4121 | goto out_balanced; |
| 4122 | |
| 4123 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) |
| 4124 | goto out_balanced; |
| 4125 | |
| 4126 | sds.busiest_load_per_task /= sds.busiest_nr_running; |
| 4127 | if (sds.group_imb) |
| 4128 | sds.busiest_load_per_task = |
| 4129 | min(sds.busiest_load_per_task, sds.avg_load); |
| 4130 | |
| 4131 | /* |
| 4132 | * We're trying to get all the cpus to the average_load, so we don't |
| 4133 | * want to push ourselves above the average load, nor do we wish to |
| 4134 | * reduce the max loaded cpu below the average load, as either of these |
| 4135 | * actions would just result in more rebalancing later, and ping-pong |
| 4136 | * tasks around. Thus we look for the minimum possible imbalance. |
| 4137 | * Negative imbalances (*we* are more loaded than anyone else) will |
| 4138 | * be counted as no imbalance for these purposes -- we can't fix that |
| 4139 | * by pulling tasks to us. Be careful of negative numbers as they'll |
| 4140 | * appear as very large values with unsigned longs. |
| 4141 | */ |
| 4142 | if (sds.max_load <= sds.busiest_load_per_task) |
| 4143 | goto out_balanced; |
| 4144 | |
| 4145 | /* Looks like there is an imbalance. Compute it */ |
| 4146 | calculate_imbalance(&sds, this_cpu, imbalance); |
| 4147 | return sds.busiest; |
| 4148 | |
| 4149 | out_balanced: |
| 4150 | /* |
| 4151 | * There is no obvious imbalance. But check if we can do some balancing |
| 4152 | * to save power. |
| 4153 | */ |
| 4154 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) |
| 4155 | return sds.busiest; |
| 4156 | ret: |
| 4157 | *imbalance = 0; |
| 4158 | return NULL; |
| 4159 | } |
| 4160 | |
| 4161 | static struct sched_group *group_of(int cpu) |
| 4162 | { |
| 4163 | struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd); |
| 4164 | |
| 4165 | if (!sd) |
| 4166 | return NULL; |
| 4167 | |
| 4168 | return sd->groups; |
| 4169 | } |
| 4170 | |
| 4171 | static unsigned long power_of(int cpu) |
| 4172 | { |
| 4173 | struct sched_group *group = group_of(cpu); |
| 4174 | |
| 4175 | if (!group) |
| 4176 | return SCHED_LOAD_SCALE; |
| 4177 | |
| 4178 | return group->cpu_power; |
| 4179 | } |
| 4180 | |
| 4181 | /* |
| 4182 | * find_busiest_queue - find the busiest runqueue among the cpus in group. |
| 4183 | */ |
| 4184 | static struct rq * |
| 4185 | find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle, |
| 4186 | unsigned long imbalance, const struct cpumask *cpus) |
| 4187 | { |
| 4188 | struct rq *busiest = NULL, *rq; |
| 4189 | unsigned long max_load = 0; |
| 4190 | int i; |
| 4191 | |
| 4192 | for_each_cpu(i, sched_group_cpus(group)) { |
| 4193 | unsigned long power = power_of(i); |
| 4194 | unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); |
| 4195 | unsigned long wl; |
| 4196 | |
| 4197 | if (!cpumask_test_cpu(i, cpus)) |
| 4198 | continue; |
| 4199 | |
| 4200 | rq = cpu_rq(i); |
| 4201 | wl = weighted_cpuload(i) * SCHED_LOAD_SCALE; |
| 4202 | wl /= power; |
| 4203 | |
| 4204 | if (capacity && rq->nr_running == 1 && wl > imbalance) |
| 4205 | continue; |
| 4206 | |
| 4207 | if (wl > max_load) { |
| 4208 | max_load = wl; |
| 4209 | busiest = rq; |
| 4210 | } |
| 4211 | } |
| 4212 | |
| 4213 | return busiest; |
| 4214 | } |
| 4215 | |
| 4216 | /* |
| 4217 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but |
| 4218 | * so long as it is large enough. |
| 4219 | */ |
| 4220 | #define MAX_PINNED_INTERVAL 512 |
| 4221 | |
| 4222 | /* Working cpumask for load_balance and load_balance_newidle. */ |
| 4223 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); |
| 4224 | |
| 4225 | /* |
| 4226 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
| 4227 | * tasks if there is an imbalance. |
| 4228 | */ |
| 4229 | static int load_balance(int this_cpu, struct rq *this_rq, |
| 4230 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 4231 | int *balance) |
| 4232 | { |
| 4233 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; |
| 4234 | struct sched_group *group; |
| 4235 | unsigned long imbalance; |
| 4236 | struct rq *busiest; |
| 4237 | unsigned long flags; |
| 4238 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); |
| 4239 | |
| 4240 | cpumask_setall(cpus); |
| 4241 | |
| 4242 | /* |
| 4243 | * When power savings policy is enabled for the parent domain, idle |
| 4244 | * sibling can pick up load irrespective of busy siblings. In this case, |
| 4245 | * let the state of idle sibling percolate up as CPU_IDLE, instead of |
| 4246 | * portraying it as CPU_NOT_IDLE. |
| 4247 | */ |
| 4248 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && |
| 4249 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 4250 | sd_idle = 1; |
| 4251 | |
| 4252 | schedstat_inc(sd, lb_count[idle]); |
| 4253 | |
| 4254 | redo: |
| 4255 | update_shares(sd); |
| 4256 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, |
| 4257 | cpus, balance); |
| 4258 | |
| 4259 | if (*balance == 0) |
| 4260 | goto out_balanced; |
| 4261 | |
| 4262 | if (!group) { |
| 4263 | schedstat_inc(sd, lb_nobusyg[idle]); |
| 4264 | goto out_balanced; |
| 4265 | } |
| 4266 | |
| 4267 | busiest = find_busiest_queue(group, idle, imbalance, cpus); |
| 4268 | if (!busiest) { |
| 4269 | schedstat_inc(sd, lb_nobusyq[idle]); |
| 4270 | goto out_balanced; |
| 4271 | } |
| 4272 | |
| 4273 | BUG_ON(busiest == this_rq); |
| 4274 | |
| 4275 | schedstat_add(sd, lb_imbalance[idle], imbalance); |
| 4276 | |
| 4277 | ld_moved = 0; |
| 4278 | if (busiest->nr_running > 1) { |
| 4279 | /* |
| 4280 | * Attempt to move tasks. If find_busiest_group has found |
| 4281 | * an imbalance but busiest->nr_running <= 1, the group is |
| 4282 | * still unbalanced. ld_moved simply stays zero, so it is |
| 4283 | * correctly treated as an imbalance. |
| 4284 | */ |
| 4285 | local_irq_save(flags); |
| 4286 | double_rq_lock(this_rq, busiest); |
| 4287 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
| 4288 | imbalance, sd, idle, &all_pinned); |
| 4289 | double_rq_unlock(this_rq, busiest); |
| 4290 | local_irq_restore(flags); |
| 4291 | |
| 4292 | /* |
| 4293 | * some other cpu did the load balance for us. |
| 4294 | */ |
| 4295 | if (ld_moved && this_cpu != smp_processor_id()) |
| 4296 | resched_cpu(this_cpu); |
| 4297 | |
| 4298 | /* All tasks on this runqueue were pinned by CPU affinity */ |
| 4299 | if (unlikely(all_pinned)) { |
| 4300 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
| 4301 | if (!cpumask_empty(cpus)) |
| 4302 | goto redo; |
| 4303 | goto out_balanced; |
| 4304 | } |
| 4305 | } |
| 4306 | |
| 4307 | if (!ld_moved) { |
| 4308 | schedstat_inc(sd, lb_failed[idle]); |
| 4309 | sd->nr_balance_failed++; |
| 4310 | |
| 4311 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { |
| 4312 | |
| 4313 | spin_lock_irqsave(&busiest->lock, flags); |
| 4314 | |
| 4315 | /* don't kick the migration_thread, if the curr |
| 4316 | * task on busiest cpu can't be moved to this_cpu |
| 4317 | */ |
| 4318 | if (!cpumask_test_cpu(this_cpu, |
| 4319 | &busiest->curr->cpus_allowed)) { |
| 4320 | spin_unlock_irqrestore(&busiest->lock, flags); |
| 4321 | all_pinned = 1; |
| 4322 | goto out_one_pinned; |
| 4323 | } |
| 4324 | |
| 4325 | if (!busiest->active_balance) { |
| 4326 | busiest->active_balance = 1; |
| 4327 | busiest->push_cpu = this_cpu; |
| 4328 | active_balance = 1; |
| 4329 | } |
| 4330 | spin_unlock_irqrestore(&busiest->lock, flags); |
| 4331 | if (active_balance) |
| 4332 | wake_up_process(busiest->migration_thread); |
| 4333 | |
| 4334 | /* |
| 4335 | * We've kicked active balancing, reset the failure |
| 4336 | * counter. |
| 4337 | */ |
| 4338 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
| 4339 | } |
| 4340 | } else |
| 4341 | sd->nr_balance_failed = 0; |
| 4342 | |
| 4343 | if (likely(!active_balance)) { |
| 4344 | /* We were unbalanced, so reset the balancing interval */ |
| 4345 | sd->balance_interval = sd->min_interval; |
| 4346 | } else { |
| 4347 | /* |
| 4348 | * If we've begun active balancing, start to back off. This |
| 4349 | * case may not be covered by the all_pinned logic if there |
| 4350 | * is only 1 task on the busy runqueue (because we don't call |
| 4351 | * move_tasks). |
| 4352 | */ |
| 4353 | if (sd->balance_interval < sd->max_interval) |
| 4354 | sd->balance_interval *= 2; |
| 4355 | } |
| 4356 | |
| 4357 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
| 4358 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 4359 | ld_moved = -1; |
| 4360 | |
| 4361 | goto out; |
| 4362 | |
| 4363 | out_balanced: |
| 4364 | schedstat_inc(sd, lb_balanced[idle]); |
| 4365 | |
| 4366 | sd->nr_balance_failed = 0; |
| 4367 | |
| 4368 | out_one_pinned: |
| 4369 | /* tune up the balancing interval */ |
| 4370 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
| 4371 | (sd->balance_interval < sd->max_interval)) |
| 4372 | sd->balance_interval *= 2; |
| 4373 | |
| 4374 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
| 4375 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 4376 | ld_moved = -1; |
| 4377 | else |
| 4378 | ld_moved = 0; |
| 4379 | out: |
| 4380 | if (ld_moved) |
| 4381 | update_shares(sd); |
| 4382 | return ld_moved; |
| 4383 | } |
| 4384 | |
| 4385 | /* |
| 4386 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
| 4387 | * tasks if there is an imbalance. |
| 4388 | * |
| 4389 | * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE). |
| 4390 | * this_rq is locked. |
| 4391 | */ |
| 4392 | static int |
| 4393 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd) |
| 4394 | { |
| 4395 | struct sched_group *group; |
| 4396 | struct rq *busiest = NULL; |
| 4397 | unsigned long imbalance; |
| 4398 | int ld_moved = 0; |
| 4399 | int sd_idle = 0; |
| 4400 | int all_pinned = 0; |
| 4401 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); |
| 4402 | |
| 4403 | cpumask_setall(cpus); |
| 4404 | |
| 4405 | /* |
| 4406 | * When power savings policy is enabled for the parent domain, idle |
| 4407 | * sibling can pick up load irrespective of busy siblings. In this case, |
| 4408 | * let the state of idle sibling percolate up as IDLE, instead of |
| 4409 | * portraying it as CPU_NOT_IDLE. |
| 4410 | */ |
| 4411 | if (sd->flags & SD_SHARE_CPUPOWER && |
| 4412 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 4413 | sd_idle = 1; |
| 4414 | |
| 4415 | schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]); |
| 4416 | redo: |
| 4417 | update_shares_locked(this_rq, sd); |
| 4418 | group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE, |
| 4419 | &sd_idle, cpus, NULL); |
| 4420 | if (!group) { |
| 4421 | schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]); |
| 4422 | goto out_balanced; |
| 4423 | } |
| 4424 | |
| 4425 | busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus); |
| 4426 | if (!busiest) { |
| 4427 | schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]); |
| 4428 | goto out_balanced; |
| 4429 | } |
| 4430 | |
| 4431 | BUG_ON(busiest == this_rq); |
| 4432 | |
| 4433 | schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance); |
| 4434 | |
| 4435 | ld_moved = 0; |
| 4436 | if (busiest->nr_running > 1) { |
| 4437 | /* Attempt to move tasks */ |
| 4438 | double_lock_balance(this_rq, busiest); |
| 4439 | /* this_rq->clock is already updated */ |
| 4440 | update_rq_clock(busiest); |
| 4441 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
| 4442 | imbalance, sd, CPU_NEWLY_IDLE, |
| 4443 | &all_pinned); |
| 4444 | double_unlock_balance(this_rq, busiest); |
| 4445 | |
| 4446 | if (unlikely(all_pinned)) { |
| 4447 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
| 4448 | if (!cpumask_empty(cpus)) |
| 4449 | goto redo; |
| 4450 | } |
| 4451 | } |
| 4452 | |
| 4453 | if (!ld_moved) { |
| 4454 | int active_balance = 0; |
| 4455 | |
| 4456 | schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]); |
| 4457 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
| 4458 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 4459 | return -1; |
| 4460 | |
| 4461 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) |
| 4462 | return -1; |
| 4463 | |
| 4464 | if (sd->nr_balance_failed++ < 2) |
| 4465 | return -1; |
| 4466 | |
| 4467 | /* |
| 4468 | * The only task running in a non-idle cpu can be moved to this |
| 4469 | * cpu in an attempt to completely freeup the other CPU |
| 4470 | * package. The same method used to move task in load_balance() |
| 4471 | * have been extended for load_balance_newidle() to speedup |
| 4472 | * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2) |
| 4473 | * |
| 4474 | * The package power saving logic comes from |
| 4475 | * find_busiest_group(). If there are no imbalance, then |
| 4476 | * f_b_g() will return NULL. However when sched_mc={1,2} then |
| 4477 | * f_b_g() will select a group from which a running task may be |
| 4478 | * pulled to this cpu in order to make the other package idle. |
| 4479 | * If there is no opportunity to make a package idle and if |
| 4480 | * there are no imbalance, then f_b_g() will return NULL and no |
| 4481 | * action will be taken in load_balance_newidle(). |
| 4482 | * |
| 4483 | * Under normal task pull operation due to imbalance, there |
| 4484 | * will be more than one task in the source run queue and |
| 4485 | * move_tasks() will succeed. ld_moved will be true and this |
| 4486 | * active balance code will not be triggered. |
| 4487 | */ |
| 4488 | |
| 4489 | /* Lock busiest in correct order while this_rq is held */ |
| 4490 | double_lock_balance(this_rq, busiest); |
| 4491 | |
| 4492 | /* |
| 4493 | * don't kick the migration_thread, if the curr |
| 4494 | * task on busiest cpu can't be moved to this_cpu |
| 4495 | */ |
| 4496 | if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) { |
| 4497 | double_unlock_balance(this_rq, busiest); |
| 4498 | all_pinned = 1; |
| 4499 | return ld_moved; |
| 4500 | } |
| 4501 | |
| 4502 | if (!busiest->active_balance) { |
| 4503 | busiest->active_balance = 1; |
| 4504 | busiest->push_cpu = this_cpu; |
| 4505 | active_balance = 1; |
| 4506 | } |
| 4507 | |
| 4508 | double_unlock_balance(this_rq, busiest); |
| 4509 | /* |
| 4510 | * Should not call ttwu while holding a rq->lock |
| 4511 | */ |
| 4512 | spin_unlock(&this_rq->lock); |
| 4513 | if (active_balance) |
| 4514 | wake_up_process(busiest->migration_thread); |
| 4515 | spin_lock(&this_rq->lock); |
| 4516 | |
| 4517 | } else |
| 4518 | sd->nr_balance_failed = 0; |
| 4519 | |
| 4520 | update_shares_locked(this_rq, sd); |
| 4521 | return ld_moved; |
| 4522 | |
| 4523 | out_balanced: |
| 4524 | schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]); |
| 4525 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
| 4526 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 4527 | return -1; |
| 4528 | sd->nr_balance_failed = 0; |
| 4529 | |
| 4530 | return 0; |
| 4531 | } |
| 4532 | |
| 4533 | /* |
| 4534 | * idle_balance is called by schedule() if this_cpu is about to become |
| 4535 | * idle. Attempts to pull tasks from other CPUs. |
| 4536 | */ |
| 4537 | static void idle_balance(int this_cpu, struct rq *this_rq) |
| 4538 | { |
| 4539 | struct sched_domain *sd; |
| 4540 | int pulled_task = 0; |
| 4541 | unsigned long next_balance = jiffies + HZ; |
| 4542 | |
| 4543 | for_each_domain(this_cpu, sd) { |
| 4544 | unsigned long interval; |
| 4545 | |
| 4546 | if (!(sd->flags & SD_LOAD_BALANCE)) |
| 4547 | continue; |
| 4548 | |
| 4549 | if (sd->flags & SD_BALANCE_NEWIDLE) |
| 4550 | /* If we've pulled tasks over stop searching: */ |
| 4551 | pulled_task = load_balance_newidle(this_cpu, this_rq, |
| 4552 | sd); |
| 4553 | |
| 4554 | interval = msecs_to_jiffies(sd->balance_interval); |
| 4555 | if (time_after(next_balance, sd->last_balance + interval)) |
| 4556 | next_balance = sd->last_balance + interval; |
| 4557 | if (pulled_task) |
| 4558 | break; |
| 4559 | } |
| 4560 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
| 4561 | /* |
| 4562 | * We are going idle. next_balance may be set based on |
| 4563 | * a busy processor. So reset next_balance. |
| 4564 | */ |
| 4565 | this_rq->next_balance = next_balance; |
| 4566 | } |
| 4567 | } |
| 4568 | |
| 4569 | /* |
| 4570 | * active_load_balance is run by migration threads. It pushes running tasks |
| 4571 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be |
| 4572 | * running on each physical CPU where possible, and avoids physical / |
| 4573 | * logical imbalances. |
| 4574 | * |
| 4575 | * Called with busiest_rq locked. |
| 4576 | */ |
| 4577 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) |
| 4578 | { |
| 4579 | int target_cpu = busiest_rq->push_cpu; |
| 4580 | struct sched_domain *sd; |
| 4581 | struct rq *target_rq; |
| 4582 | |
| 4583 | /* Is there any task to move? */ |
| 4584 | if (busiest_rq->nr_running <= 1) |
| 4585 | return; |
| 4586 | |
| 4587 | target_rq = cpu_rq(target_cpu); |
| 4588 | |
| 4589 | /* |
| 4590 | * This condition is "impossible", if it occurs |
| 4591 | * we need to fix it. Originally reported by |
| 4592 | * Bjorn Helgaas on a 128-cpu setup. |
| 4593 | */ |
| 4594 | BUG_ON(busiest_rq == target_rq); |
| 4595 | |
| 4596 | /* move a task from busiest_rq to target_rq */ |
| 4597 | double_lock_balance(busiest_rq, target_rq); |
| 4598 | update_rq_clock(busiest_rq); |
| 4599 | update_rq_clock(target_rq); |
| 4600 | |
| 4601 | /* Search for an sd spanning us and the target CPU. */ |
| 4602 | for_each_domain(target_cpu, sd) { |
| 4603 | if ((sd->flags & SD_LOAD_BALANCE) && |
| 4604 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) |
| 4605 | break; |
| 4606 | } |
| 4607 | |
| 4608 | if (likely(sd)) { |
| 4609 | schedstat_inc(sd, alb_count); |
| 4610 | |
| 4611 | if (move_one_task(target_rq, target_cpu, busiest_rq, |
| 4612 | sd, CPU_IDLE)) |
| 4613 | schedstat_inc(sd, alb_pushed); |
| 4614 | else |
| 4615 | schedstat_inc(sd, alb_failed); |
| 4616 | } |
| 4617 | double_unlock_balance(busiest_rq, target_rq); |
| 4618 | } |
| 4619 | |
| 4620 | #ifdef CONFIG_NO_HZ |
| 4621 | static struct { |
| 4622 | atomic_t load_balancer; |
| 4623 | cpumask_var_t cpu_mask; |
| 4624 | cpumask_var_t ilb_grp_nohz_mask; |
| 4625 | } nohz ____cacheline_aligned = { |
| 4626 | .load_balancer = ATOMIC_INIT(-1), |
| 4627 | }; |
| 4628 | |
| 4629 | int get_nohz_load_balancer(void) |
| 4630 | { |
| 4631 | return atomic_read(&nohz.load_balancer); |
| 4632 | } |
| 4633 | |
| 4634 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
| 4635 | /** |
| 4636 | * lowest_flag_domain - Return lowest sched_domain containing flag. |
| 4637 | * @cpu: The cpu whose lowest level of sched domain is to |
| 4638 | * be returned. |
| 4639 | * @flag: The flag to check for the lowest sched_domain |
| 4640 | * for the given cpu. |
| 4641 | * |
| 4642 | * Returns the lowest sched_domain of a cpu which contains the given flag. |
| 4643 | */ |
| 4644 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) |
| 4645 | { |
| 4646 | struct sched_domain *sd; |
| 4647 | |
| 4648 | for_each_domain(cpu, sd) |
| 4649 | if (sd && (sd->flags & flag)) |
| 4650 | break; |
| 4651 | |
| 4652 | return sd; |
| 4653 | } |
| 4654 | |
| 4655 | /** |
| 4656 | * for_each_flag_domain - Iterates over sched_domains containing the flag. |
| 4657 | * @cpu: The cpu whose domains we're iterating over. |
| 4658 | * @sd: variable holding the value of the power_savings_sd |
| 4659 | * for cpu. |
| 4660 | * @flag: The flag to filter the sched_domains to be iterated. |
| 4661 | * |
| 4662 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' |
| 4663 | * set, starting from the lowest sched_domain to the highest. |
| 4664 | */ |
| 4665 | #define for_each_flag_domain(cpu, sd, flag) \ |
| 4666 | for (sd = lowest_flag_domain(cpu, flag); \ |
| 4667 | (sd && (sd->flags & flag)); sd = sd->parent) |
| 4668 | |
| 4669 | /** |
| 4670 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. |
| 4671 | * @ilb_group: group to be checked for semi-idleness |
| 4672 | * |
| 4673 | * Returns: 1 if the group is semi-idle. 0 otherwise. |
| 4674 | * |
| 4675 | * We define a sched_group to be semi idle if it has atleast one idle-CPU |
| 4676 | * and atleast one non-idle CPU. This helper function checks if the given |
| 4677 | * sched_group is semi-idle or not. |
| 4678 | */ |
| 4679 | static inline int is_semi_idle_group(struct sched_group *ilb_group) |
| 4680 | { |
| 4681 | cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask, |
| 4682 | sched_group_cpus(ilb_group)); |
| 4683 | |
| 4684 | /* |
| 4685 | * A sched_group is semi-idle when it has atleast one busy cpu |
| 4686 | * and atleast one idle cpu. |
| 4687 | */ |
| 4688 | if (cpumask_empty(nohz.ilb_grp_nohz_mask)) |
| 4689 | return 0; |
| 4690 | |
| 4691 | if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group))) |
| 4692 | return 0; |
| 4693 | |
| 4694 | return 1; |
| 4695 | } |
| 4696 | /** |
| 4697 | * find_new_ilb - Finds the optimum idle load balancer for nomination. |
| 4698 | * @cpu: The cpu which is nominating a new idle_load_balancer. |
| 4699 | * |
| 4700 | * Returns: Returns the id of the idle load balancer if it exists, |
| 4701 | * Else, returns >= nr_cpu_ids. |
| 4702 | * |
| 4703 | * This algorithm picks the idle load balancer such that it belongs to a |
| 4704 | * semi-idle powersavings sched_domain. The idea is to try and avoid |
| 4705 | * completely idle packages/cores just for the purpose of idle load balancing |
| 4706 | * when there are other idle cpu's which are better suited for that job. |
| 4707 | */ |
| 4708 | static int find_new_ilb(int cpu) |
| 4709 | { |
| 4710 | struct sched_domain *sd; |
| 4711 | struct sched_group *ilb_group; |
| 4712 | |
| 4713 | /* |
| 4714 | * Have idle load balancer selection from semi-idle packages only |
| 4715 | * when power-aware load balancing is enabled |
| 4716 | */ |
| 4717 | if (!(sched_smt_power_savings || sched_mc_power_savings)) |
| 4718 | goto out_done; |
| 4719 | |
| 4720 | /* |
| 4721 | * Optimize for the case when we have no idle CPUs or only one |
| 4722 | * idle CPU. Don't walk the sched_domain hierarchy in such cases |
| 4723 | */ |
| 4724 | if (cpumask_weight(nohz.cpu_mask) < 2) |
| 4725 | goto out_done; |
| 4726 | |
| 4727 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { |
| 4728 | ilb_group = sd->groups; |
| 4729 | |
| 4730 | do { |
| 4731 | if (is_semi_idle_group(ilb_group)) |
| 4732 | return cpumask_first(nohz.ilb_grp_nohz_mask); |
| 4733 | |
| 4734 | ilb_group = ilb_group->next; |
| 4735 | |
| 4736 | } while (ilb_group != sd->groups); |
| 4737 | } |
| 4738 | |
| 4739 | out_done: |
| 4740 | return cpumask_first(nohz.cpu_mask); |
| 4741 | } |
| 4742 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ |
| 4743 | static inline int find_new_ilb(int call_cpu) |
| 4744 | { |
| 4745 | return cpumask_first(nohz.cpu_mask); |
| 4746 | } |
| 4747 | #endif |
| 4748 | |
| 4749 | /* |
| 4750 | * This routine will try to nominate the ilb (idle load balancing) |
| 4751 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle |
| 4752 | * load balancing on behalf of all those cpus. If all the cpus in the system |
| 4753 | * go into this tickless mode, then there will be no ilb owner (as there is |
| 4754 | * no need for one) and all the cpus will sleep till the next wakeup event |
| 4755 | * arrives... |
| 4756 | * |
| 4757 | * For the ilb owner, tick is not stopped. And this tick will be used |
| 4758 | * for idle load balancing. ilb owner will still be part of |
| 4759 | * nohz.cpu_mask.. |
| 4760 | * |
| 4761 | * While stopping the tick, this cpu will become the ilb owner if there |
| 4762 | * is no other owner. And will be the owner till that cpu becomes busy |
| 4763 | * or if all cpus in the system stop their ticks at which point |
| 4764 | * there is no need for ilb owner. |
| 4765 | * |
| 4766 | * When the ilb owner becomes busy, it nominates another owner, during the |
| 4767 | * next busy scheduler_tick() |
| 4768 | */ |
| 4769 | int select_nohz_load_balancer(int stop_tick) |
| 4770 | { |
| 4771 | int cpu = smp_processor_id(); |
| 4772 | |
| 4773 | if (stop_tick) { |
| 4774 | cpu_rq(cpu)->in_nohz_recently = 1; |
| 4775 | |
| 4776 | if (!cpu_active(cpu)) { |
| 4777 | if (atomic_read(&nohz.load_balancer) != cpu) |
| 4778 | return 0; |
| 4779 | |
| 4780 | /* |
| 4781 | * If we are going offline and still the leader, |
| 4782 | * give up! |
| 4783 | */ |
| 4784 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) |
| 4785 | BUG(); |
| 4786 | |
| 4787 | return 0; |
| 4788 | } |
| 4789 | |
| 4790 | cpumask_set_cpu(cpu, nohz.cpu_mask); |
| 4791 | |
| 4792 | /* time for ilb owner also to sleep */ |
| 4793 | if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) { |
| 4794 | if (atomic_read(&nohz.load_balancer) == cpu) |
| 4795 | atomic_set(&nohz.load_balancer, -1); |
| 4796 | return 0; |
| 4797 | } |
| 4798 | |
| 4799 | if (atomic_read(&nohz.load_balancer) == -1) { |
| 4800 | /* make me the ilb owner */ |
| 4801 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) |
| 4802 | return 1; |
| 4803 | } else if (atomic_read(&nohz.load_balancer) == cpu) { |
| 4804 | int new_ilb; |
| 4805 | |
| 4806 | if (!(sched_smt_power_savings || |
| 4807 | sched_mc_power_savings)) |
| 4808 | return 1; |
| 4809 | /* |
| 4810 | * Check to see if there is a more power-efficient |
| 4811 | * ilb. |
| 4812 | */ |
| 4813 | new_ilb = find_new_ilb(cpu); |
| 4814 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { |
| 4815 | atomic_set(&nohz.load_balancer, -1); |
| 4816 | resched_cpu(new_ilb); |
| 4817 | return 0; |
| 4818 | } |
| 4819 | return 1; |
| 4820 | } |
| 4821 | } else { |
| 4822 | if (!cpumask_test_cpu(cpu, nohz.cpu_mask)) |
| 4823 | return 0; |
| 4824 | |
| 4825 | cpumask_clear_cpu(cpu, nohz.cpu_mask); |
| 4826 | |
| 4827 | if (atomic_read(&nohz.load_balancer) == cpu) |
| 4828 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) |
| 4829 | BUG(); |
| 4830 | } |
| 4831 | return 0; |
| 4832 | } |
| 4833 | #endif |
| 4834 | |
| 4835 | static DEFINE_SPINLOCK(balancing); |
| 4836 | |
| 4837 | /* |
| 4838 | * It checks each scheduling domain to see if it is due to be balanced, |
| 4839 | * and initiates a balancing operation if so. |
| 4840 | * |
| 4841 | * Balancing parameters are set up in arch_init_sched_domains. |
| 4842 | */ |
| 4843 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) |
| 4844 | { |
| 4845 | int balance = 1; |
| 4846 | struct rq *rq = cpu_rq(cpu); |
| 4847 | unsigned long interval; |
| 4848 | struct sched_domain *sd; |
| 4849 | /* Earliest time when we have to do rebalance again */ |
| 4850 | unsigned long next_balance = jiffies + 60*HZ; |
| 4851 | int update_next_balance = 0; |
| 4852 | int need_serialize; |
| 4853 | |
| 4854 | for_each_domain(cpu, sd) { |
| 4855 | if (!(sd->flags & SD_LOAD_BALANCE)) |
| 4856 | continue; |
| 4857 | |
| 4858 | interval = sd->balance_interval; |
| 4859 | if (idle != CPU_IDLE) |
| 4860 | interval *= sd->busy_factor; |
| 4861 | |
| 4862 | /* scale ms to jiffies */ |
| 4863 | interval = msecs_to_jiffies(interval); |
| 4864 | if (unlikely(!interval)) |
| 4865 | interval = 1; |
| 4866 | if (interval > HZ*NR_CPUS/10) |
| 4867 | interval = HZ*NR_CPUS/10; |
| 4868 | |
| 4869 | need_serialize = sd->flags & SD_SERIALIZE; |
| 4870 | |
| 4871 | if (need_serialize) { |
| 4872 | if (!spin_trylock(&balancing)) |
| 4873 | goto out; |
| 4874 | } |
| 4875 | |
| 4876 | if (time_after_eq(jiffies, sd->last_balance + interval)) { |
| 4877 | if (load_balance(cpu, rq, sd, idle, &balance)) { |
| 4878 | /* |
| 4879 | * We've pulled tasks over so either we're no |
| 4880 | * longer idle, or one of our SMT siblings is |
| 4881 | * not idle. |
| 4882 | */ |
| 4883 | idle = CPU_NOT_IDLE; |
| 4884 | } |
| 4885 | sd->last_balance = jiffies; |
| 4886 | } |
| 4887 | if (need_serialize) |
| 4888 | spin_unlock(&balancing); |
| 4889 | out: |
| 4890 | if (time_after(next_balance, sd->last_balance + interval)) { |
| 4891 | next_balance = sd->last_balance + interval; |
| 4892 | update_next_balance = 1; |
| 4893 | } |
| 4894 | |
| 4895 | /* |
| 4896 | * Stop the load balance at this level. There is another |
| 4897 | * CPU in our sched group which is doing load balancing more |
| 4898 | * actively. |
| 4899 | */ |
| 4900 | if (!balance) |
| 4901 | break; |
| 4902 | } |
| 4903 | |
| 4904 | /* |
| 4905 | * next_balance will be updated only when there is a need. |
| 4906 | * When the cpu is attached to null domain for ex, it will not be |
| 4907 | * updated. |
| 4908 | */ |
| 4909 | if (likely(update_next_balance)) |
| 4910 | rq->next_balance = next_balance; |
| 4911 | } |
| 4912 | |
| 4913 | /* |
| 4914 | * run_rebalance_domains is triggered when needed from the scheduler tick. |
| 4915 | * In CONFIG_NO_HZ case, the idle load balance owner will do the |
| 4916 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
| 4917 | */ |
| 4918 | static void run_rebalance_domains(struct softirq_action *h) |
| 4919 | { |
| 4920 | int this_cpu = smp_processor_id(); |
| 4921 | struct rq *this_rq = cpu_rq(this_cpu); |
| 4922 | enum cpu_idle_type idle = this_rq->idle_at_tick ? |
| 4923 | CPU_IDLE : CPU_NOT_IDLE; |
| 4924 | |
| 4925 | rebalance_domains(this_cpu, idle); |
| 4926 | |
| 4927 | #ifdef CONFIG_NO_HZ |
| 4928 | /* |
| 4929 | * If this cpu is the owner for idle load balancing, then do the |
| 4930 | * balancing on behalf of the other idle cpus whose ticks are |
| 4931 | * stopped. |
| 4932 | */ |
| 4933 | if (this_rq->idle_at_tick && |
| 4934 | atomic_read(&nohz.load_balancer) == this_cpu) { |
| 4935 | struct rq *rq; |
| 4936 | int balance_cpu; |
| 4937 | |
| 4938 | for_each_cpu(balance_cpu, nohz.cpu_mask) { |
| 4939 | if (balance_cpu == this_cpu) |
| 4940 | continue; |
| 4941 | |
| 4942 | /* |
| 4943 | * If this cpu gets work to do, stop the load balancing |
| 4944 | * work being done for other cpus. Next load |
| 4945 | * balancing owner will pick it up. |
| 4946 | */ |
| 4947 | if (need_resched()) |
| 4948 | break; |
| 4949 | |
| 4950 | rebalance_domains(balance_cpu, CPU_IDLE); |
| 4951 | |
| 4952 | rq = cpu_rq(balance_cpu); |
| 4953 | if (time_after(this_rq->next_balance, rq->next_balance)) |
| 4954 | this_rq->next_balance = rq->next_balance; |
| 4955 | } |
| 4956 | } |
| 4957 | #endif |
| 4958 | } |
| 4959 | |
| 4960 | static inline int on_null_domain(int cpu) |
| 4961 | { |
| 4962 | return !rcu_dereference(cpu_rq(cpu)->sd); |
| 4963 | } |
| 4964 | |
| 4965 | /* |
| 4966 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. |
| 4967 | * |
| 4968 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new |
| 4969 | * idle load balancing owner or decide to stop the periodic load balancing, |
| 4970 | * if the whole system is idle. |
| 4971 | */ |
| 4972 | static inline void trigger_load_balance(struct rq *rq, int cpu) |
| 4973 | { |
| 4974 | #ifdef CONFIG_NO_HZ |
| 4975 | /* |
| 4976 | * If we were in the nohz mode recently and busy at the current |
| 4977 | * scheduler tick, then check if we need to nominate new idle |
| 4978 | * load balancer. |
| 4979 | */ |
| 4980 | if (rq->in_nohz_recently && !rq->idle_at_tick) { |
| 4981 | rq->in_nohz_recently = 0; |
| 4982 | |
| 4983 | if (atomic_read(&nohz.load_balancer) == cpu) { |
| 4984 | cpumask_clear_cpu(cpu, nohz.cpu_mask); |
| 4985 | atomic_set(&nohz.load_balancer, -1); |
| 4986 | } |
| 4987 | |
| 4988 | if (atomic_read(&nohz.load_balancer) == -1) { |
| 4989 | int ilb = find_new_ilb(cpu); |
| 4990 | |
| 4991 | if (ilb < nr_cpu_ids) |
| 4992 | resched_cpu(ilb); |
| 4993 | } |
| 4994 | } |
| 4995 | |
| 4996 | /* |
| 4997 | * If this cpu is idle and doing idle load balancing for all the |
| 4998 | * cpus with ticks stopped, is it time for that to stop? |
| 4999 | */ |
| 5000 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && |
| 5001 | cpumask_weight(nohz.cpu_mask) == num_online_cpus()) { |
| 5002 | resched_cpu(cpu); |
| 5003 | return; |
| 5004 | } |
| 5005 | |
| 5006 | /* |
| 5007 | * If this cpu is idle and the idle load balancing is done by |
| 5008 | * someone else, then no need raise the SCHED_SOFTIRQ |
| 5009 | */ |
| 5010 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && |
| 5011 | cpumask_test_cpu(cpu, nohz.cpu_mask)) |
| 5012 | return; |
| 5013 | #endif |
| 5014 | /* Don't need to rebalance while attached to NULL domain */ |
| 5015 | if (time_after_eq(jiffies, rq->next_balance) && |
| 5016 | likely(!on_null_domain(cpu))) |
| 5017 | raise_softirq(SCHED_SOFTIRQ); |
| 5018 | } |
| 5019 | |
| 5020 | #else /* CONFIG_SMP */ |
| 5021 | |
| 5022 | /* |
| 5023 | * on UP we do not need to balance between CPUs: |
| 5024 | */ |
| 5025 | static inline void idle_balance(int cpu, struct rq *rq) |
| 5026 | { |
| 5027 | } |
| 5028 | |
| 5029 | #endif |
| 5030 | |
| 5031 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
| 5032 | |
| 5033 | EXPORT_PER_CPU_SYMBOL(kstat); |
| 5034 | |
| 5035 | /* |
| 5036 | * Return any ns on the sched_clock that have not yet been accounted in |
| 5037 | * @p in case that task is currently running. |
| 5038 | * |
| 5039 | * Called with task_rq_lock() held on @rq. |
| 5040 | */ |
| 5041 | static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) |
| 5042 | { |
| 5043 | u64 ns = 0; |
| 5044 | |
| 5045 | if (task_current(rq, p)) { |
| 5046 | update_rq_clock(rq); |
| 5047 | ns = rq->clock - p->se.exec_start; |
| 5048 | if ((s64)ns < 0) |
| 5049 | ns = 0; |
| 5050 | } |
| 5051 | |
| 5052 | return ns; |
| 5053 | } |
| 5054 | |
| 5055 | unsigned long long task_delta_exec(struct task_struct *p) |
| 5056 | { |
| 5057 | unsigned long flags; |
| 5058 | struct rq *rq; |
| 5059 | u64 ns = 0; |
| 5060 | |
| 5061 | rq = task_rq_lock(p, &flags); |
| 5062 | ns = do_task_delta_exec(p, rq); |
| 5063 | task_rq_unlock(rq, &flags); |
| 5064 | |
| 5065 | return ns; |
| 5066 | } |
| 5067 | |
| 5068 | /* |
| 5069 | * Return accounted runtime for the task. |
| 5070 | * In case the task is currently running, return the runtime plus current's |
| 5071 | * pending runtime that have not been accounted yet. |
| 5072 | */ |
| 5073 | unsigned long long task_sched_runtime(struct task_struct *p) |
| 5074 | { |
| 5075 | unsigned long flags; |
| 5076 | struct rq *rq; |
| 5077 | u64 ns = 0; |
| 5078 | |
| 5079 | rq = task_rq_lock(p, &flags); |
| 5080 | ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); |
| 5081 | task_rq_unlock(rq, &flags); |
| 5082 | |
| 5083 | return ns; |
| 5084 | } |
| 5085 | |
| 5086 | /* |
| 5087 | * Return sum_exec_runtime for the thread group. |
| 5088 | * In case the task is currently running, return the sum plus current's |
| 5089 | * pending runtime that have not been accounted yet. |
| 5090 | * |
| 5091 | * Note that the thread group might have other running tasks as well, |
| 5092 | * so the return value not includes other pending runtime that other |
| 5093 | * running tasks might have. |
| 5094 | */ |
| 5095 | unsigned long long thread_group_sched_runtime(struct task_struct *p) |
| 5096 | { |
| 5097 | struct task_cputime totals; |
| 5098 | unsigned long flags; |
| 5099 | struct rq *rq; |
| 5100 | u64 ns; |
| 5101 | |
| 5102 | rq = task_rq_lock(p, &flags); |
| 5103 | thread_group_cputime(p, &totals); |
| 5104 | ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq); |
| 5105 | task_rq_unlock(rq, &flags); |
| 5106 | |
| 5107 | return ns; |
| 5108 | } |
| 5109 | |
| 5110 | /* |
| 5111 | * Account user cpu time to a process. |
| 5112 | * @p: the process that the cpu time gets accounted to |
| 5113 | * @cputime: the cpu time spent in user space since the last update |
| 5114 | * @cputime_scaled: cputime scaled by cpu frequency |
| 5115 | */ |
| 5116 | void account_user_time(struct task_struct *p, cputime_t cputime, |
| 5117 | cputime_t cputime_scaled) |
| 5118 | { |
| 5119 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 5120 | cputime64_t tmp; |
| 5121 | |
| 5122 | /* Add user time to process. */ |
| 5123 | p->utime = cputime_add(p->utime, cputime); |
| 5124 | p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); |
| 5125 | account_group_user_time(p, cputime); |
| 5126 | |
| 5127 | /* Add user time to cpustat. */ |
| 5128 | tmp = cputime_to_cputime64(cputime); |
| 5129 | if (TASK_NICE(p) > 0) |
| 5130 | cpustat->nice = cputime64_add(cpustat->nice, tmp); |
| 5131 | else |
| 5132 | cpustat->user = cputime64_add(cpustat->user, tmp); |
| 5133 | |
| 5134 | cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime); |
| 5135 | /* Account for user time used */ |
| 5136 | acct_update_integrals(p); |
| 5137 | } |
| 5138 | |
| 5139 | /* |
| 5140 | * Account guest cpu time to a process. |
| 5141 | * @p: the process that the cpu time gets accounted to |
| 5142 | * @cputime: the cpu time spent in virtual machine since the last update |
| 5143 | * @cputime_scaled: cputime scaled by cpu frequency |
| 5144 | */ |
| 5145 | static void account_guest_time(struct task_struct *p, cputime_t cputime, |
| 5146 | cputime_t cputime_scaled) |
| 5147 | { |
| 5148 | cputime64_t tmp; |
| 5149 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 5150 | |
| 5151 | tmp = cputime_to_cputime64(cputime); |
| 5152 | |
| 5153 | /* Add guest time to process. */ |
| 5154 | p->utime = cputime_add(p->utime, cputime); |
| 5155 | p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); |
| 5156 | account_group_user_time(p, cputime); |
| 5157 | p->gtime = cputime_add(p->gtime, cputime); |
| 5158 | |
| 5159 | /* Add guest time to cpustat. */ |
| 5160 | cpustat->user = cputime64_add(cpustat->user, tmp); |
| 5161 | cpustat->guest = cputime64_add(cpustat->guest, tmp); |
| 5162 | } |
| 5163 | |
| 5164 | /* |
| 5165 | * Account system cpu time to a process. |
| 5166 | * @p: the process that the cpu time gets accounted to |
| 5167 | * @hardirq_offset: the offset to subtract from hardirq_count() |
| 5168 | * @cputime: the cpu time spent in kernel space since the last update |
| 5169 | * @cputime_scaled: cputime scaled by cpu frequency |
| 5170 | */ |
| 5171 | void account_system_time(struct task_struct *p, int hardirq_offset, |
| 5172 | cputime_t cputime, cputime_t cputime_scaled) |
| 5173 | { |
| 5174 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 5175 | cputime64_t tmp; |
| 5176 | |
| 5177 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { |
| 5178 | account_guest_time(p, cputime, cputime_scaled); |
| 5179 | return; |
| 5180 | } |
| 5181 | |
| 5182 | /* Add system time to process. */ |
| 5183 | p->stime = cputime_add(p->stime, cputime); |
| 5184 | p->stimescaled = cputime_add(p->stimescaled, cputime_scaled); |
| 5185 | account_group_system_time(p, cputime); |
| 5186 | |
| 5187 | /* Add system time to cpustat. */ |
| 5188 | tmp = cputime_to_cputime64(cputime); |
| 5189 | if (hardirq_count() - hardirq_offset) |
| 5190 | cpustat->irq = cputime64_add(cpustat->irq, tmp); |
| 5191 | else if (softirq_count()) |
| 5192 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); |
| 5193 | else |
| 5194 | cpustat->system = cputime64_add(cpustat->system, tmp); |
| 5195 | |
| 5196 | cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime); |
| 5197 | |
| 5198 | /* Account for system time used */ |
| 5199 | acct_update_integrals(p); |
| 5200 | } |
| 5201 | |
| 5202 | /* |
| 5203 | * Account for involuntary wait time. |
| 5204 | * @steal: the cpu time spent in involuntary wait |
| 5205 | */ |
| 5206 | void account_steal_time(cputime_t cputime) |
| 5207 | { |
| 5208 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 5209 | cputime64_t cputime64 = cputime_to_cputime64(cputime); |
| 5210 | |
| 5211 | cpustat->steal = cputime64_add(cpustat->steal, cputime64); |
| 5212 | } |
| 5213 | |
| 5214 | /* |
| 5215 | * Account for idle time. |
| 5216 | * @cputime: the cpu time spent in idle wait |
| 5217 | */ |
| 5218 | void account_idle_time(cputime_t cputime) |
| 5219 | { |
| 5220 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 5221 | cputime64_t cputime64 = cputime_to_cputime64(cputime); |
| 5222 | struct rq *rq = this_rq(); |
| 5223 | |
| 5224 | if (atomic_read(&rq->nr_iowait) > 0) |
| 5225 | cpustat->iowait = cputime64_add(cpustat->iowait, cputime64); |
| 5226 | else |
| 5227 | cpustat->idle = cputime64_add(cpustat->idle, cputime64); |
| 5228 | } |
| 5229 | |
| 5230 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
| 5231 | |
| 5232 | /* |
| 5233 | * Account a single tick of cpu time. |
| 5234 | * @p: the process that the cpu time gets accounted to |
| 5235 | * @user_tick: indicates if the tick is a user or a system tick |
| 5236 | */ |
| 5237 | void account_process_tick(struct task_struct *p, int user_tick) |
| 5238 | { |
| 5239 | cputime_t one_jiffy = jiffies_to_cputime(1); |
| 5240 | cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy); |
| 5241 | struct rq *rq = this_rq(); |
| 5242 | |
| 5243 | if (user_tick) |
| 5244 | account_user_time(p, one_jiffy, one_jiffy_scaled); |
| 5245 | else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) |
| 5246 | account_system_time(p, HARDIRQ_OFFSET, one_jiffy, |
| 5247 | one_jiffy_scaled); |
| 5248 | else |
| 5249 | account_idle_time(one_jiffy); |
| 5250 | } |
| 5251 | |
| 5252 | /* |
| 5253 | * Account multiple ticks of steal time. |
| 5254 | * @p: the process from which the cpu time has been stolen |
| 5255 | * @ticks: number of stolen ticks |
| 5256 | */ |
| 5257 | void account_steal_ticks(unsigned long ticks) |
| 5258 | { |
| 5259 | account_steal_time(jiffies_to_cputime(ticks)); |
| 5260 | } |
| 5261 | |
| 5262 | /* |
| 5263 | * Account multiple ticks of idle time. |
| 5264 | * @ticks: number of stolen ticks |
| 5265 | */ |
| 5266 | void account_idle_ticks(unsigned long ticks) |
| 5267 | { |
| 5268 | account_idle_time(jiffies_to_cputime(ticks)); |
| 5269 | } |
| 5270 | |
| 5271 | #endif |
| 5272 | |
| 5273 | /* |
| 5274 | * Use precise platform statistics if available: |
| 5275 | */ |
| 5276 | #ifdef CONFIG_VIRT_CPU_ACCOUNTING |
| 5277 | cputime_t task_utime(struct task_struct *p) |
| 5278 | { |
| 5279 | return p->utime; |
| 5280 | } |
| 5281 | |
| 5282 | cputime_t task_stime(struct task_struct *p) |
| 5283 | { |
| 5284 | return p->stime; |
| 5285 | } |
| 5286 | #else |
| 5287 | cputime_t task_utime(struct task_struct *p) |
| 5288 | { |
| 5289 | clock_t utime = cputime_to_clock_t(p->utime), |
| 5290 | total = utime + cputime_to_clock_t(p->stime); |
| 5291 | u64 temp; |
| 5292 | |
| 5293 | /* |
| 5294 | * Use CFS's precise accounting: |
| 5295 | */ |
| 5296 | temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime); |
| 5297 | |
| 5298 | if (total) { |
| 5299 | temp *= utime; |
| 5300 | do_div(temp, total); |
| 5301 | } |
| 5302 | utime = (clock_t)temp; |
| 5303 | |
| 5304 | p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime)); |
| 5305 | return p->prev_utime; |
| 5306 | } |
| 5307 | |
| 5308 | cputime_t task_stime(struct task_struct *p) |
| 5309 | { |
| 5310 | clock_t stime; |
| 5311 | |
| 5312 | /* |
| 5313 | * Use CFS's precise accounting. (we subtract utime from |
| 5314 | * the total, to make sure the total observed by userspace |
| 5315 | * grows monotonically - apps rely on that): |
| 5316 | */ |
| 5317 | stime = nsec_to_clock_t(p->se.sum_exec_runtime) - |
| 5318 | cputime_to_clock_t(task_utime(p)); |
| 5319 | |
| 5320 | if (stime >= 0) |
| 5321 | p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime)); |
| 5322 | |
| 5323 | return p->prev_stime; |
| 5324 | } |
| 5325 | #endif |
| 5326 | |
| 5327 | inline cputime_t task_gtime(struct task_struct *p) |
| 5328 | { |
| 5329 | return p->gtime; |
| 5330 | } |
| 5331 | |
| 5332 | /* |
| 5333 | * This function gets called by the timer code, with HZ frequency. |
| 5334 | * We call it with interrupts disabled. |
| 5335 | * |
| 5336 | * It also gets called by the fork code, when changing the parent's |
| 5337 | * timeslices. |
| 5338 | */ |
| 5339 | void scheduler_tick(void) |
| 5340 | { |
| 5341 | int cpu = smp_processor_id(); |
| 5342 | struct rq *rq = cpu_rq(cpu); |
| 5343 | struct task_struct *curr = rq->curr; |
| 5344 | |
| 5345 | sched_clock_tick(); |
| 5346 | |
| 5347 | spin_lock(&rq->lock); |
| 5348 | update_rq_clock(rq); |
| 5349 | update_cpu_load(rq); |
| 5350 | curr->sched_class->task_tick(rq, curr, 0); |
| 5351 | spin_unlock(&rq->lock); |
| 5352 | |
| 5353 | perf_counter_task_tick(curr, cpu); |
| 5354 | |
| 5355 | #ifdef CONFIG_SMP |
| 5356 | rq->idle_at_tick = idle_cpu(cpu); |
| 5357 | trigger_load_balance(rq, cpu); |
| 5358 | #endif |
| 5359 | } |
| 5360 | |
| 5361 | notrace unsigned long get_parent_ip(unsigned long addr) |
| 5362 | { |
| 5363 | if (in_lock_functions(addr)) { |
| 5364 | addr = CALLER_ADDR2; |
| 5365 | if (in_lock_functions(addr)) |
| 5366 | addr = CALLER_ADDR3; |
| 5367 | } |
| 5368 | return addr; |
| 5369 | } |
| 5370 | |
| 5371 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ |
| 5372 | defined(CONFIG_PREEMPT_TRACER)) |
| 5373 | |
| 5374 | void __kprobes add_preempt_count(int val) |
| 5375 | { |
| 5376 | #ifdef CONFIG_DEBUG_PREEMPT |
| 5377 | /* |
| 5378 | * Underflow? |
| 5379 | */ |
| 5380 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
| 5381 | return; |
| 5382 | #endif |
| 5383 | preempt_count() += val; |
| 5384 | #ifdef CONFIG_DEBUG_PREEMPT |
| 5385 | /* |
| 5386 | * Spinlock count overflowing soon? |
| 5387 | */ |
| 5388 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
| 5389 | PREEMPT_MASK - 10); |
| 5390 | #endif |
| 5391 | if (preempt_count() == val) |
| 5392 | trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); |
| 5393 | } |
| 5394 | EXPORT_SYMBOL(add_preempt_count); |
| 5395 | |
| 5396 | void __kprobes sub_preempt_count(int val) |
| 5397 | { |
| 5398 | #ifdef CONFIG_DEBUG_PREEMPT |
| 5399 | /* |
| 5400 | * Underflow? |
| 5401 | */ |
| 5402 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
| 5403 | return; |
| 5404 | /* |
| 5405 | * Is the spinlock portion underflowing? |
| 5406 | */ |
| 5407 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
| 5408 | !(preempt_count() & PREEMPT_MASK))) |
| 5409 | return; |
| 5410 | #endif |
| 5411 | |
| 5412 | if (preempt_count() == val) |
| 5413 | trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); |
| 5414 | preempt_count() -= val; |
| 5415 | } |
| 5416 | EXPORT_SYMBOL(sub_preempt_count); |
| 5417 | |
| 5418 | #endif |
| 5419 | |
| 5420 | /* |
| 5421 | * Print scheduling while atomic bug: |
| 5422 | */ |
| 5423 | static noinline void __schedule_bug(struct task_struct *prev) |
| 5424 | { |
| 5425 | struct pt_regs *regs = get_irq_regs(); |
| 5426 | |
| 5427 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", |
| 5428 | prev->comm, prev->pid, preempt_count()); |
| 5429 | |
| 5430 | debug_show_held_locks(prev); |
| 5431 | print_modules(); |
| 5432 | if (irqs_disabled()) |
| 5433 | print_irqtrace_events(prev); |
| 5434 | |
| 5435 | if (regs) |
| 5436 | show_regs(regs); |
| 5437 | else |
| 5438 | dump_stack(); |
| 5439 | } |
| 5440 | |
| 5441 | /* |
| 5442 | * Various schedule()-time debugging checks and statistics: |
| 5443 | */ |
| 5444 | static inline void schedule_debug(struct task_struct *prev) |
| 5445 | { |
| 5446 | /* |
| 5447 | * Test if we are atomic. Since do_exit() needs to call into |
| 5448 | * schedule() atomically, we ignore that path for now. |
| 5449 | * Otherwise, whine if we are scheduling when we should not be. |
| 5450 | */ |
| 5451 | if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) |
| 5452 | __schedule_bug(prev); |
| 5453 | |
| 5454 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
| 5455 | |
| 5456 | schedstat_inc(this_rq(), sched_count); |
| 5457 | #ifdef CONFIG_SCHEDSTATS |
| 5458 | if (unlikely(prev->lock_depth >= 0)) { |
| 5459 | schedstat_inc(this_rq(), bkl_count); |
| 5460 | schedstat_inc(prev, sched_info.bkl_count); |
| 5461 | } |
| 5462 | #endif |
| 5463 | } |
| 5464 | |
| 5465 | static void put_prev_task(struct rq *rq, struct task_struct *prev) |
| 5466 | { |
| 5467 | if (prev->state == TASK_RUNNING) { |
| 5468 | u64 runtime = prev->se.sum_exec_runtime; |
| 5469 | |
| 5470 | runtime -= prev->se.prev_sum_exec_runtime; |
| 5471 | runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost); |
| 5472 | |
| 5473 | /* |
| 5474 | * In order to avoid avg_overlap growing stale when we are |
| 5475 | * indeed overlapping and hence not getting put to sleep, grow |
| 5476 | * the avg_overlap on preemption. |
| 5477 | * |
| 5478 | * We use the average preemption runtime because that |
| 5479 | * correlates to the amount of cache footprint a task can |
| 5480 | * build up. |
| 5481 | */ |
| 5482 | update_avg(&prev->se.avg_overlap, runtime); |
| 5483 | } |
| 5484 | prev->sched_class->put_prev_task(rq, prev); |
| 5485 | } |
| 5486 | |
| 5487 | /* |
| 5488 | * Pick up the highest-prio task: |
| 5489 | */ |
| 5490 | static inline struct task_struct * |
| 5491 | pick_next_task(struct rq *rq) |
| 5492 | { |
| 5493 | const struct sched_class *class; |
| 5494 | struct task_struct *p; |
| 5495 | |
| 5496 | /* |
| 5497 | * Optimization: we know that if all tasks are in |
| 5498 | * the fair class we can call that function directly: |
| 5499 | */ |
| 5500 | if (likely(rq->nr_running == rq->cfs.nr_running)) { |
| 5501 | p = fair_sched_class.pick_next_task(rq); |
| 5502 | if (likely(p)) |
| 5503 | return p; |
| 5504 | } |
| 5505 | |
| 5506 | class = sched_class_highest; |
| 5507 | for ( ; ; ) { |
| 5508 | p = class->pick_next_task(rq); |
| 5509 | if (p) |
| 5510 | return p; |
| 5511 | /* |
| 5512 | * Will never be NULL as the idle class always |
| 5513 | * returns a non-NULL p: |
| 5514 | */ |
| 5515 | class = class->next; |
| 5516 | } |
| 5517 | } |
| 5518 | |
| 5519 | /* |
| 5520 | * schedule() is the main scheduler function. |
| 5521 | */ |
| 5522 | asmlinkage void __sched schedule(void) |
| 5523 | { |
| 5524 | struct task_struct *prev, *next; |
| 5525 | unsigned long *switch_count; |
| 5526 | struct rq *rq; |
| 5527 | int cpu; |
| 5528 | |
| 5529 | need_resched: |
| 5530 | preempt_disable(); |
| 5531 | cpu = smp_processor_id(); |
| 5532 | rq = cpu_rq(cpu); |
| 5533 | rcu_sched_qs(cpu); |
| 5534 | prev = rq->curr; |
| 5535 | switch_count = &prev->nivcsw; |
| 5536 | |
| 5537 | release_kernel_lock(prev); |
| 5538 | need_resched_nonpreemptible: |
| 5539 | |
| 5540 | schedule_debug(prev); |
| 5541 | |
| 5542 | if (sched_feat(HRTICK)) |
| 5543 | hrtick_clear(rq); |
| 5544 | |
| 5545 | spin_lock_irq(&rq->lock); |
| 5546 | update_rq_clock(rq); |
| 5547 | clear_tsk_need_resched(prev); |
| 5548 | |
| 5549 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
| 5550 | if (unlikely(signal_pending_state(prev->state, prev))) |
| 5551 | prev->state = TASK_RUNNING; |
| 5552 | else |
| 5553 | deactivate_task(rq, prev, 1); |
| 5554 | switch_count = &prev->nvcsw; |
| 5555 | } |
| 5556 | |
| 5557 | pre_schedule(rq, prev); |
| 5558 | |
| 5559 | if (unlikely(!rq->nr_running)) |
| 5560 | idle_balance(cpu, rq); |
| 5561 | |
| 5562 | put_prev_task(rq, prev); |
| 5563 | next = pick_next_task(rq); |
| 5564 | |
| 5565 | if (likely(prev != next)) { |
| 5566 | sched_info_switch(prev, next); |
| 5567 | perf_counter_task_sched_out(prev, next, cpu); |
| 5568 | |
| 5569 | rq->nr_switches++; |
| 5570 | rq->curr = next; |
| 5571 | ++*switch_count; |
| 5572 | |
| 5573 | context_switch(rq, prev, next); /* unlocks the rq */ |
| 5574 | /* |
| 5575 | * the context switch might have flipped the stack from under |
| 5576 | * us, hence refresh the local variables. |
| 5577 | */ |
| 5578 | cpu = smp_processor_id(); |
| 5579 | rq = cpu_rq(cpu); |
| 5580 | } else |
| 5581 | spin_unlock_irq(&rq->lock); |
| 5582 | |
| 5583 | post_schedule(rq); |
| 5584 | |
| 5585 | if (unlikely(reacquire_kernel_lock(current) < 0)) |
| 5586 | goto need_resched_nonpreemptible; |
| 5587 | |
| 5588 | preempt_enable_no_resched(); |
| 5589 | if (need_resched()) |
| 5590 | goto need_resched; |
| 5591 | } |
| 5592 | EXPORT_SYMBOL(schedule); |
| 5593 | |
| 5594 | #ifdef CONFIG_SMP |
| 5595 | /* |
| 5596 | * Look out! "owner" is an entirely speculative pointer |
| 5597 | * access and not reliable. |
| 5598 | */ |
| 5599 | int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner) |
| 5600 | { |
| 5601 | unsigned int cpu; |
| 5602 | struct rq *rq; |
| 5603 | |
| 5604 | if (!sched_feat(OWNER_SPIN)) |
| 5605 | return 0; |
| 5606 | |
| 5607 | #ifdef CONFIG_DEBUG_PAGEALLOC |
| 5608 | /* |
| 5609 | * Need to access the cpu field knowing that |
| 5610 | * DEBUG_PAGEALLOC could have unmapped it if |
| 5611 | * the mutex owner just released it and exited. |
| 5612 | */ |
| 5613 | if (probe_kernel_address(&owner->cpu, cpu)) |
| 5614 | goto out; |
| 5615 | #else |
| 5616 | cpu = owner->cpu; |
| 5617 | #endif |
| 5618 | |
| 5619 | /* |
| 5620 | * Even if the access succeeded (likely case), |
| 5621 | * the cpu field may no longer be valid. |
| 5622 | */ |
| 5623 | if (cpu >= nr_cpumask_bits) |
| 5624 | goto out; |
| 5625 | |
| 5626 | /* |
| 5627 | * We need to validate that we can do a |
| 5628 | * get_cpu() and that we have the percpu area. |
| 5629 | */ |
| 5630 | if (!cpu_online(cpu)) |
| 5631 | goto out; |
| 5632 | |
| 5633 | rq = cpu_rq(cpu); |
| 5634 | |
| 5635 | for (;;) { |
| 5636 | /* |
| 5637 | * Owner changed, break to re-assess state. |
| 5638 | */ |
| 5639 | if (lock->owner != owner) |
| 5640 | break; |
| 5641 | |
| 5642 | /* |
| 5643 | * Is that owner really running on that cpu? |
| 5644 | */ |
| 5645 | if (task_thread_info(rq->curr) != owner || need_resched()) |
| 5646 | return 0; |
| 5647 | |
| 5648 | cpu_relax(); |
| 5649 | } |
| 5650 | out: |
| 5651 | return 1; |
| 5652 | } |
| 5653 | #endif |
| 5654 | |
| 5655 | #ifdef CONFIG_PREEMPT |
| 5656 | /* |
| 5657 | * this is the entry point to schedule() from in-kernel preemption |
| 5658 | * off of preempt_enable. Kernel preemptions off return from interrupt |
| 5659 | * occur there and call schedule directly. |
| 5660 | */ |
| 5661 | asmlinkage void __sched preempt_schedule(void) |
| 5662 | { |
| 5663 | struct thread_info *ti = current_thread_info(); |
| 5664 | |
| 5665 | /* |
| 5666 | * If there is a non-zero preempt_count or interrupts are disabled, |
| 5667 | * we do not want to preempt the current task. Just return.. |
| 5668 | */ |
| 5669 | if (likely(ti->preempt_count || irqs_disabled())) |
| 5670 | return; |
| 5671 | |
| 5672 | do { |
| 5673 | add_preempt_count(PREEMPT_ACTIVE); |
| 5674 | schedule(); |
| 5675 | sub_preempt_count(PREEMPT_ACTIVE); |
| 5676 | |
| 5677 | /* |
| 5678 | * Check again in case we missed a preemption opportunity |
| 5679 | * between schedule and now. |
| 5680 | */ |
| 5681 | barrier(); |
| 5682 | } while (need_resched()); |
| 5683 | } |
| 5684 | EXPORT_SYMBOL(preempt_schedule); |
| 5685 | |
| 5686 | /* |
| 5687 | * this is the entry point to schedule() from kernel preemption |
| 5688 | * off of irq context. |
| 5689 | * Note, that this is called and return with irqs disabled. This will |
| 5690 | * protect us against recursive calling from irq. |
| 5691 | */ |
| 5692 | asmlinkage void __sched preempt_schedule_irq(void) |
| 5693 | { |
| 5694 | struct thread_info *ti = current_thread_info(); |
| 5695 | |
| 5696 | /* Catch callers which need to be fixed */ |
| 5697 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
| 5698 | |
| 5699 | do { |
| 5700 | add_preempt_count(PREEMPT_ACTIVE); |
| 5701 | local_irq_enable(); |
| 5702 | schedule(); |
| 5703 | local_irq_disable(); |
| 5704 | sub_preempt_count(PREEMPT_ACTIVE); |
| 5705 | |
| 5706 | /* |
| 5707 | * Check again in case we missed a preemption opportunity |
| 5708 | * between schedule and now. |
| 5709 | */ |
| 5710 | barrier(); |
| 5711 | } while (need_resched()); |
| 5712 | } |
| 5713 | |
| 5714 | #endif /* CONFIG_PREEMPT */ |
| 5715 | |
| 5716 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
| 5717 | void *key) |
| 5718 | { |
| 5719 | return try_to_wake_up(curr->private, mode, sync); |
| 5720 | } |
| 5721 | EXPORT_SYMBOL(default_wake_function); |
| 5722 | |
| 5723 | /* |
| 5724 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
| 5725 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve |
| 5726 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
| 5727 | * |
| 5728 | * There are circumstances in which we can try to wake a task which has already |
| 5729 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
| 5730 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
| 5731 | */ |
| 5732 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, |
| 5733 | int nr_exclusive, int sync, void *key) |
| 5734 | { |
| 5735 | wait_queue_t *curr, *next; |
| 5736 | |
| 5737 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { |
| 5738 | unsigned flags = curr->flags; |
| 5739 | |
| 5740 | if (curr->func(curr, mode, sync, key) && |
| 5741 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) |
| 5742 | break; |
| 5743 | } |
| 5744 | } |
| 5745 | |
| 5746 | /** |
| 5747 | * __wake_up - wake up threads blocked on a waitqueue. |
| 5748 | * @q: the waitqueue |
| 5749 | * @mode: which threads |
| 5750 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
| 5751 | * @key: is directly passed to the wakeup function |
| 5752 | * |
| 5753 | * It may be assumed that this function implies a write memory barrier before |
| 5754 | * changing the task state if and only if any tasks are woken up. |
| 5755 | */ |
| 5756 | void __wake_up(wait_queue_head_t *q, unsigned int mode, |
| 5757 | int nr_exclusive, void *key) |
| 5758 | { |
| 5759 | unsigned long flags; |
| 5760 | |
| 5761 | spin_lock_irqsave(&q->lock, flags); |
| 5762 | __wake_up_common(q, mode, nr_exclusive, 0, key); |
| 5763 | spin_unlock_irqrestore(&q->lock, flags); |
| 5764 | } |
| 5765 | EXPORT_SYMBOL(__wake_up); |
| 5766 | |
| 5767 | /* |
| 5768 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. |
| 5769 | */ |
| 5770 | void __wake_up_locked(wait_queue_head_t *q, unsigned int mode) |
| 5771 | { |
| 5772 | __wake_up_common(q, mode, 1, 0, NULL); |
| 5773 | } |
| 5774 | |
| 5775 | void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) |
| 5776 | { |
| 5777 | __wake_up_common(q, mode, 1, 0, key); |
| 5778 | } |
| 5779 | |
| 5780 | /** |
| 5781 | * __wake_up_sync_key - wake up threads blocked on a waitqueue. |
| 5782 | * @q: the waitqueue |
| 5783 | * @mode: which threads |
| 5784 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
| 5785 | * @key: opaque value to be passed to wakeup targets |
| 5786 | * |
| 5787 | * The sync wakeup differs that the waker knows that it will schedule |
| 5788 | * away soon, so while the target thread will be woken up, it will not |
| 5789 | * be migrated to another CPU - ie. the two threads are 'synchronized' |
| 5790 | * with each other. This can prevent needless bouncing between CPUs. |
| 5791 | * |
| 5792 | * On UP it can prevent extra preemption. |
| 5793 | * |
| 5794 | * It may be assumed that this function implies a write memory barrier before |
| 5795 | * changing the task state if and only if any tasks are woken up. |
| 5796 | */ |
| 5797 | void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, |
| 5798 | int nr_exclusive, void *key) |
| 5799 | { |
| 5800 | unsigned long flags; |
| 5801 | int sync = 1; |
| 5802 | |
| 5803 | if (unlikely(!q)) |
| 5804 | return; |
| 5805 | |
| 5806 | if (unlikely(!nr_exclusive)) |
| 5807 | sync = 0; |
| 5808 | |
| 5809 | spin_lock_irqsave(&q->lock, flags); |
| 5810 | __wake_up_common(q, mode, nr_exclusive, sync, key); |
| 5811 | spin_unlock_irqrestore(&q->lock, flags); |
| 5812 | } |
| 5813 | EXPORT_SYMBOL_GPL(__wake_up_sync_key); |
| 5814 | |
| 5815 | /* |
| 5816 | * __wake_up_sync - see __wake_up_sync_key() |
| 5817 | */ |
| 5818 | void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
| 5819 | { |
| 5820 | __wake_up_sync_key(q, mode, nr_exclusive, NULL); |
| 5821 | } |
| 5822 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ |
| 5823 | |
| 5824 | /** |
| 5825 | * complete: - signals a single thread waiting on this completion |
| 5826 | * @x: holds the state of this particular completion |
| 5827 | * |
| 5828 | * This will wake up a single thread waiting on this completion. Threads will be |
| 5829 | * awakened in the same order in which they were queued. |
| 5830 | * |
| 5831 | * See also complete_all(), wait_for_completion() and related routines. |
| 5832 | * |
| 5833 | * It may be assumed that this function implies a write memory barrier before |
| 5834 | * changing the task state if and only if any tasks are woken up. |
| 5835 | */ |
| 5836 | void complete(struct completion *x) |
| 5837 | { |
| 5838 | unsigned long flags; |
| 5839 | |
| 5840 | spin_lock_irqsave(&x->wait.lock, flags); |
| 5841 | x->done++; |
| 5842 | __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); |
| 5843 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 5844 | } |
| 5845 | EXPORT_SYMBOL(complete); |
| 5846 | |
| 5847 | /** |
| 5848 | * complete_all: - signals all threads waiting on this completion |
| 5849 | * @x: holds the state of this particular completion |
| 5850 | * |
| 5851 | * This will wake up all threads waiting on this particular completion event. |
| 5852 | * |
| 5853 | * It may be assumed that this function implies a write memory barrier before |
| 5854 | * changing the task state if and only if any tasks are woken up. |
| 5855 | */ |
| 5856 | void complete_all(struct completion *x) |
| 5857 | { |
| 5858 | unsigned long flags; |
| 5859 | |
| 5860 | spin_lock_irqsave(&x->wait.lock, flags); |
| 5861 | x->done += UINT_MAX/2; |
| 5862 | __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); |
| 5863 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 5864 | } |
| 5865 | EXPORT_SYMBOL(complete_all); |
| 5866 | |
| 5867 | static inline long __sched |
| 5868 | do_wait_for_common(struct completion *x, long timeout, int state) |
| 5869 | { |
| 5870 | if (!x->done) { |
| 5871 | DECLARE_WAITQUEUE(wait, current); |
| 5872 | |
| 5873 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 5874 | __add_wait_queue_tail(&x->wait, &wait); |
| 5875 | do { |
| 5876 | if (signal_pending_state(state, current)) { |
| 5877 | timeout = -ERESTARTSYS; |
| 5878 | break; |
| 5879 | } |
| 5880 | __set_current_state(state); |
| 5881 | spin_unlock_irq(&x->wait.lock); |
| 5882 | timeout = schedule_timeout(timeout); |
| 5883 | spin_lock_irq(&x->wait.lock); |
| 5884 | } while (!x->done && timeout); |
| 5885 | __remove_wait_queue(&x->wait, &wait); |
| 5886 | if (!x->done) |
| 5887 | return timeout; |
| 5888 | } |
| 5889 | x->done--; |
| 5890 | return timeout ?: 1; |
| 5891 | } |
| 5892 | |
| 5893 | static long __sched |
| 5894 | wait_for_common(struct completion *x, long timeout, int state) |
| 5895 | { |
| 5896 | might_sleep(); |
| 5897 | |
| 5898 | spin_lock_irq(&x->wait.lock); |
| 5899 | timeout = do_wait_for_common(x, timeout, state); |
| 5900 | spin_unlock_irq(&x->wait.lock); |
| 5901 | return timeout; |
| 5902 | } |
| 5903 | |
| 5904 | /** |
| 5905 | * wait_for_completion: - waits for completion of a task |
| 5906 | * @x: holds the state of this particular completion |
| 5907 | * |
| 5908 | * This waits to be signaled for completion of a specific task. It is NOT |
| 5909 | * interruptible and there is no timeout. |
| 5910 | * |
| 5911 | * See also similar routines (i.e. wait_for_completion_timeout()) with timeout |
| 5912 | * and interrupt capability. Also see complete(). |
| 5913 | */ |
| 5914 | void __sched wait_for_completion(struct completion *x) |
| 5915 | { |
| 5916 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); |
| 5917 | } |
| 5918 | EXPORT_SYMBOL(wait_for_completion); |
| 5919 | |
| 5920 | /** |
| 5921 | * wait_for_completion_timeout: - waits for completion of a task (w/timeout) |
| 5922 | * @x: holds the state of this particular completion |
| 5923 | * @timeout: timeout value in jiffies |
| 5924 | * |
| 5925 | * This waits for either a completion of a specific task to be signaled or for a |
| 5926 | * specified timeout to expire. The timeout is in jiffies. It is not |
| 5927 | * interruptible. |
| 5928 | */ |
| 5929 | unsigned long __sched |
| 5930 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
| 5931 | { |
| 5932 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); |
| 5933 | } |
| 5934 | EXPORT_SYMBOL(wait_for_completion_timeout); |
| 5935 | |
| 5936 | /** |
| 5937 | * wait_for_completion_interruptible: - waits for completion of a task (w/intr) |
| 5938 | * @x: holds the state of this particular completion |
| 5939 | * |
| 5940 | * This waits for completion of a specific task to be signaled. It is |
| 5941 | * interruptible. |
| 5942 | */ |
| 5943 | int __sched wait_for_completion_interruptible(struct completion *x) |
| 5944 | { |
| 5945 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); |
| 5946 | if (t == -ERESTARTSYS) |
| 5947 | return t; |
| 5948 | return 0; |
| 5949 | } |
| 5950 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
| 5951 | |
| 5952 | /** |
| 5953 | * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) |
| 5954 | * @x: holds the state of this particular completion |
| 5955 | * @timeout: timeout value in jiffies |
| 5956 | * |
| 5957 | * This waits for either a completion of a specific task to be signaled or for a |
| 5958 | * specified timeout to expire. It is interruptible. The timeout is in jiffies. |
| 5959 | */ |
| 5960 | unsigned long __sched |
| 5961 | wait_for_completion_interruptible_timeout(struct completion *x, |
| 5962 | unsigned long timeout) |
| 5963 | { |
| 5964 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); |
| 5965 | } |
| 5966 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
| 5967 | |
| 5968 | /** |
| 5969 | * wait_for_completion_killable: - waits for completion of a task (killable) |
| 5970 | * @x: holds the state of this particular completion |
| 5971 | * |
| 5972 | * This waits to be signaled for completion of a specific task. It can be |
| 5973 | * interrupted by a kill signal. |
| 5974 | */ |
| 5975 | int __sched wait_for_completion_killable(struct completion *x) |
| 5976 | { |
| 5977 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); |
| 5978 | if (t == -ERESTARTSYS) |
| 5979 | return t; |
| 5980 | return 0; |
| 5981 | } |
| 5982 | EXPORT_SYMBOL(wait_for_completion_killable); |
| 5983 | |
| 5984 | /** |
| 5985 | * try_wait_for_completion - try to decrement a completion without blocking |
| 5986 | * @x: completion structure |
| 5987 | * |
| 5988 | * Returns: 0 if a decrement cannot be done without blocking |
| 5989 | * 1 if a decrement succeeded. |
| 5990 | * |
| 5991 | * If a completion is being used as a counting completion, |
| 5992 | * attempt to decrement the counter without blocking. This |
| 5993 | * enables us to avoid waiting if the resource the completion |
| 5994 | * is protecting is not available. |
| 5995 | */ |
| 5996 | bool try_wait_for_completion(struct completion *x) |
| 5997 | { |
| 5998 | int ret = 1; |
| 5999 | |
| 6000 | spin_lock_irq(&x->wait.lock); |
| 6001 | if (!x->done) |
| 6002 | ret = 0; |
| 6003 | else |
| 6004 | x->done--; |
| 6005 | spin_unlock_irq(&x->wait.lock); |
| 6006 | return ret; |
| 6007 | } |
| 6008 | EXPORT_SYMBOL(try_wait_for_completion); |
| 6009 | |
| 6010 | /** |
| 6011 | * completion_done - Test to see if a completion has any waiters |
| 6012 | * @x: completion structure |
| 6013 | * |
| 6014 | * Returns: 0 if there are waiters (wait_for_completion() in progress) |
| 6015 | * 1 if there are no waiters. |
| 6016 | * |
| 6017 | */ |
| 6018 | bool completion_done(struct completion *x) |
| 6019 | { |
| 6020 | int ret = 1; |
| 6021 | |
| 6022 | spin_lock_irq(&x->wait.lock); |
| 6023 | if (!x->done) |
| 6024 | ret = 0; |
| 6025 | spin_unlock_irq(&x->wait.lock); |
| 6026 | return ret; |
| 6027 | } |
| 6028 | EXPORT_SYMBOL(completion_done); |
| 6029 | |
| 6030 | static long __sched |
| 6031 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) |
| 6032 | { |
| 6033 | unsigned long flags; |
| 6034 | wait_queue_t wait; |
| 6035 | |
| 6036 | init_waitqueue_entry(&wait, current); |
| 6037 | |
| 6038 | __set_current_state(state); |
| 6039 | |
| 6040 | spin_lock_irqsave(&q->lock, flags); |
| 6041 | __add_wait_queue(q, &wait); |
| 6042 | spin_unlock(&q->lock); |
| 6043 | timeout = schedule_timeout(timeout); |
| 6044 | spin_lock_irq(&q->lock); |
| 6045 | __remove_wait_queue(q, &wait); |
| 6046 | spin_unlock_irqrestore(&q->lock, flags); |
| 6047 | |
| 6048 | return timeout; |
| 6049 | } |
| 6050 | |
| 6051 | void __sched interruptible_sleep_on(wait_queue_head_t *q) |
| 6052 | { |
| 6053 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
| 6054 | } |
| 6055 | EXPORT_SYMBOL(interruptible_sleep_on); |
| 6056 | |
| 6057 | long __sched |
| 6058 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
| 6059 | { |
| 6060 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); |
| 6061 | } |
| 6062 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
| 6063 | |
| 6064 | void __sched sleep_on(wait_queue_head_t *q) |
| 6065 | { |
| 6066 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
| 6067 | } |
| 6068 | EXPORT_SYMBOL(sleep_on); |
| 6069 | |
| 6070 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
| 6071 | { |
| 6072 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); |
| 6073 | } |
| 6074 | EXPORT_SYMBOL(sleep_on_timeout); |
| 6075 | |
| 6076 | #ifdef CONFIG_RT_MUTEXES |
| 6077 | |
| 6078 | /* |
| 6079 | * rt_mutex_setprio - set the current priority of a task |
| 6080 | * @p: task |
| 6081 | * @prio: prio value (kernel-internal form) |
| 6082 | * |
| 6083 | * This function changes the 'effective' priority of a task. It does |
| 6084 | * not touch ->normal_prio like __setscheduler(). |
| 6085 | * |
| 6086 | * Used by the rt_mutex code to implement priority inheritance logic. |
| 6087 | */ |
| 6088 | void rt_mutex_setprio(struct task_struct *p, int prio) |
| 6089 | { |
| 6090 | unsigned long flags; |
| 6091 | int oldprio, on_rq, running; |
| 6092 | struct rq *rq; |
| 6093 | const struct sched_class *prev_class = p->sched_class; |
| 6094 | |
| 6095 | BUG_ON(prio < 0 || prio > MAX_PRIO); |
| 6096 | |
| 6097 | rq = task_rq_lock(p, &flags); |
| 6098 | update_rq_clock(rq); |
| 6099 | |
| 6100 | oldprio = p->prio; |
| 6101 | on_rq = p->se.on_rq; |
| 6102 | running = task_current(rq, p); |
| 6103 | if (on_rq) |
| 6104 | dequeue_task(rq, p, 0); |
| 6105 | if (running) |
| 6106 | p->sched_class->put_prev_task(rq, p); |
| 6107 | |
| 6108 | if (rt_prio(prio)) |
| 6109 | p->sched_class = &rt_sched_class; |
| 6110 | else |
| 6111 | p->sched_class = &fair_sched_class; |
| 6112 | |
| 6113 | p->prio = prio; |
| 6114 | |
| 6115 | if (running) |
| 6116 | p->sched_class->set_curr_task(rq); |
| 6117 | if (on_rq) { |
| 6118 | enqueue_task(rq, p, 0); |
| 6119 | |
| 6120 | check_class_changed(rq, p, prev_class, oldprio, running); |
| 6121 | } |
| 6122 | task_rq_unlock(rq, &flags); |
| 6123 | } |
| 6124 | |
| 6125 | #endif |
| 6126 | |
| 6127 | void set_user_nice(struct task_struct *p, long nice) |
| 6128 | { |
| 6129 | int old_prio, delta, on_rq; |
| 6130 | unsigned long flags; |
| 6131 | struct rq *rq; |
| 6132 | |
| 6133 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) |
| 6134 | return; |
| 6135 | /* |
| 6136 | * We have to be careful, if called from sys_setpriority(), |
| 6137 | * the task might be in the middle of scheduling on another CPU. |
| 6138 | */ |
| 6139 | rq = task_rq_lock(p, &flags); |
| 6140 | update_rq_clock(rq); |
| 6141 | /* |
| 6142 | * The RT priorities are set via sched_setscheduler(), but we still |
| 6143 | * allow the 'normal' nice value to be set - but as expected |
| 6144 | * it wont have any effect on scheduling until the task is |
| 6145 | * SCHED_FIFO/SCHED_RR: |
| 6146 | */ |
| 6147 | if (task_has_rt_policy(p)) { |
| 6148 | p->static_prio = NICE_TO_PRIO(nice); |
| 6149 | goto out_unlock; |
| 6150 | } |
| 6151 | on_rq = p->se.on_rq; |
| 6152 | if (on_rq) |
| 6153 | dequeue_task(rq, p, 0); |
| 6154 | |
| 6155 | p->static_prio = NICE_TO_PRIO(nice); |
| 6156 | set_load_weight(p); |
| 6157 | old_prio = p->prio; |
| 6158 | p->prio = effective_prio(p); |
| 6159 | delta = p->prio - old_prio; |
| 6160 | |
| 6161 | if (on_rq) { |
| 6162 | enqueue_task(rq, p, 0); |
| 6163 | /* |
| 6164 | * If the task increased its priority or is running and |
| 6165 | * lowered its priority, then reschedule its CPU: |
| 6166 | */ |
| 6167 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
| 6168 | resched_task(rq->curr); |
| 6169 | } |
| 6170 | out_unlock: |
| 6171 | task_rq_unlock(rq, &flags); |
| 6172 | } |
| 6173 | EXPORT_SYMBOL(set_user_nice); |
| 6174 | |
| 6175 | /* |
| 6176 | * can_nice - check if a task can reduce its nice value |
| 6177 | * @p: task |
| 6178 | * @nice: nice value |
| 6179 | */ |
| 6180 | int can_nice(const struct task_struct *p, const int nice) |
| 6181 | { |
| 6182 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
| 6183 | int nice_rlim = 20 - nice; |
| 6184 | |
| 6185 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
| 6186 | capable(CAP_SYS_NICE)); |
| 6187 | } |
| 6188 | |
| 6189 | #ifdef __ARCH_WANT_SYS_NICE |
| 6190 | |
| 6191 | /* |
| 6192 | * sys_nice - change the priority of the current process. |
| 6193 | * @increment: priority increment |
| 6194 | * |
| 6195 | * sys_setpriority is a more generic, but much slower function that |
| 6196 | * does similar things. |
| 6197 | */ |
| 6198 | SYSCALL_DEFINE1(nice, int, increment) |
| 6199 | { |
| 6200 | long nice, retval; |
| 6201 | |
| 6202 | /* |
| 6203 | * Setpriority might change our priority at the same moment. |
| 6204 | * We don't have to worry. Conceptually one call occurs first |
| 6205 | * and we have a single winner. |
| 6206 | */ |
| 6207 | if (increment < -40) |
| 6208 | increment = -40; |
| 6209 | if (increment > 40) |
| 6210 | increment = 40; |
| 6211 | |
| 6212 | nice = TASK_NICE(current) + increment; |
| 6213 | if (nice < -20) |
| 6214 | nice = -20; |
| 6215 | if (nice > 19) |
| 6216 | nice = 19; |
| 6217 | |
| 6218 | if (increment < 0 && !can_nice(current, nice)) |
| 6219 | return -EPERM; |
| 6220 | |
| 6221 | retval = security_task_setnice(current, nice); |
| 6222 | if (retval) |
| 6223 | return retval; |
| 6224 | |
| 6225 | set_user_nice(current, nice); |
| 6226 | return 0; |
| 6227 | } |
| 6228 | |
| 6229 | #endif |
| 6230 | |
| 6231 | /** |
| 6232 | * task_prio - return the priority value of a given task. |
| 6233 | * @p: the task in question. |
| 6234 | * |
| 6235 | * This is the priority value as seen by users in /proc. |
| 6236 | * RT tasks are offset by -200. Normal tasks are centered |
| 6237 | * around 0, value goes from -16 to +15. |
| 6238 | */ |
| 6239 | int task_prio(const struct task_struct *p) |
| 6240 | { |
| 6241 | return p->prio - MAX_RT_PRIO; |
| 6242 | } |
| 6243 | |
| 6244 | /** |
| 6245 | * task_nice - return the nice value of a given task. |
| 6246 | * @p: the task in question. |
| 6247 | */ |
| 6248 | int task_nice(const struct task_struct *p) |
| 6249 | { |
| 6250 | return TASK_NICE(p); |
| 6251 | } |
| 6252 | EXPORT_SYMBOL(task_nice); |
| 6253 | |
| 6254 | /** |
| 6255 | * idle_cpu - is a given cpu idle currently? |
| 6256 | * @cpu: the processor in question. |
| 6257 | */ |
| 6258 | int idle_cpu(int cpu) |
| 6259 | { |
| 6260 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; |
| 6261 | } |
| 6262 | |
| 6263 | /** |
| 6264 | * idle_task - return the idle task for a given cpu. |
| 6265 | * @cpu: the processor in question. |
| 6266 | */ |
| 6267 | struct task_struct *idle_task(int cpu) |
| 6268 | { |
| 6269 | return cpu_rq(cpu)->idle; |
| 6270 | } |
| 6271 | |
| 6272 | /** |
| 6273 | * find_process_by_pid - find a process with a matching PID value. |
| 6274 | * @pid: the pid in question. |
| 6275 | */ |
| 6276 | static struct task_struct *find_process_by_pid(pid_t pid) |
| 6277 | { |
| 6278 | return pid ? find_task_by_vpid(pid) : current; |
| 6279 | } |
| 6280 | |
| 6281 | /* Actually do priority change: must hold rq lock. */ |
| 6282 | static void |
| 6283 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) |
| 6284 | { |
| 6285 | BUG_ON(p->se.on_rq); |
| 6286 | |
| 6287 | p->policy = policy; |
| 6288 | switch (p->policy) { |
| 6289 | case SCHED_NORMAL: |
| 6290 | case SCHED_BATCH: |
| 6291 | case SCHED_IDLE: |
| 6292 | p->sched_class = &fair_sched_class; |
| 6293 | break; |
| 6294 | case SCHED_FIFO: |
| 6295 | case SCHED_RR: |
| 6296 | p->sched_class = &rt_sched_class; |
| 6297 | break; |
| 6298 | } |
| 6299 | |
| 6300 | p->rt_priority = prio; |
| 6301 | p->normal_prio = normal_prio(p); |
| 6302 | /* we are holding p->pi_lock already */ |
| 6303 | p->prio = rt_mutex_getprio(p); |
| 6304 | set_load_weight(p); |
| 6305 | } |
| 6306 | |
| 6307 | /* |
| 6308 | * check the target process has a UID that matches the current process's |
| 6309 | */ |
| 6310 | static bool check_same_owner(struct task_struct *p) |
| 6311 | { |
| 6312 | const struct cred *cred = current_cred(), *pcred; |
| 6313 | bool match; |
| 6314 | |
| 6315 | rcu_read_lock(); |
| 6316 | pcred = __task_cred(p); |
| 6317 | match = (cred->euid == pcred->euid || |
| 6318 | cred->euid == pcred->uid); |
| 6319 | rcu_read_unlock(); |
| 6320 | return match; |
| 6321 | } |
| 6322 | |
| 6323 | static int __sched_setscheduler(struct task_struct *p, int policy, |
| 6324 | struct sched_param *param, bool user) |
| 6325 | { |
| 6326 | int retval, oldprio, oldpolicy = -1, on_rq, running; |
| 6327 | unsigned long flags; |
| 6328 | const struct sched_class *prev_class = p->sched_class; |
| 6329 | struct rq *rq; |
| 6330 | int reset_on_fork; |
| 6331 | |
| 6332 | /* may grab non-irq protected spin_locks */ |
| 6333 | BUG_ON(in_interrupt()); |
| 6334 | recheck: |
| 6335 | /* double check policy once rq lock held */ |
| 6336 | if (policy < 0) { |
| 6337 | reset_on_fork = p->sched_reset_on_fork; |
| 6338 | policy = oldpolicy = p->policy; |
| 6339 | } else { |
| 6340 | reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); |
| 6341 | policy &= ~SCHED_RESET_ON_FORK; |
| 6342 | |
| 6343 | if (policy != SCHED_FIFO && policy != SCHED_RR && |
| 6344 | policy != SCHED_NORMAL && policy != SCHED_BATCH && |
| 6345 | policy != SCHED_IDLE) |
| 6346 | return -EINVAL; |
| 6347 | } |
| 6348 | |
| 6349 | /* |
| 6350 | * Valid priorities for SCHED_FIFO and SCHED_RR are |
| 6351 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, |
| 6352 | * SCHED_BATCH and SCHED_IDLE is 0. |
| 6353 | */ |
| 6354 | if (param->sched_priority < 0 || |
| 6355 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
| 6356 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
| 6357 | return -EINVAL; |
| 6358 | if (rt_policy(policy) != (param->sched_priority != 0)) |
| 6359 | return -EINVAL; |
| 6360 | |
| 6361 | /* |
| 6362 | * Allow unprivileged RT tasks to decrease priority: |
| 6363 | */ |
| 6364 | if (user && !capable(CAP_SYS_NICE)) { |
| 6365 | if (rt_policy(policy)) { |
| 6366 | unsigned long rlim_rtprio; |
| 6367 | |
| 6368 | if (!lock_task_sighand(p, &flags)) |
| 6369 | return -ESRCH; |
| 6370 | rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; |
| 6371 | unlock_task_sighand(p, &flags); |
| 6372 | |
| 6373 | /* can't set/change the rt policy */ |
| 6374 | if (policy != p->policy && !rlim_rtprio) |
| 6375 | return -EPERM; |
| 6376 | |
| 6377 | /* can't increase priority */ |
| 6378 | if (param->sched_priority > p->rt_priority && |
| 6379 | param->sched_priority > rlim_rtprio) |
| 6380 | return -EPERM; |
| 6381 | } |
| 6382 | /* |
| 6383 | * Like positive nice levels, dont allow tasks to |
| 6384 | * move out of SCHED_IDLE either: |
| 6385 | */ |
| 6386 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) |
| 6387 | return -EPERM; |
| 6388 | |
| 6389 | /* can't change other user's priorities */ |
| 6390 | if (!check_same_owner(p)) |
| 6391 | return -EPERM; |
| 6392 | |
| 6393 | /* Normal users shall not reset the sched_reset_on_fork flag */ |
| 6394 | if (p->sched_reset_on_fork && !reset_on_fork) |
| 6395 | return -EPERM; |
| 6396 | } |
| 6397 | |
| 6398 | if (user) { |
| 6399 | #ifdef CONFIG_RT_GROUP_SCHED |
| 6400 | /* |
| 6401 | * Do not allow realtime tasks into groups that have no runtime |
| 6402 | * assigned. |
| 6403 | */ |
| 6404 | if (rt_bandwidth_enabled() && rt_policy(policy) && |
| 6405 | task_group(p)->rt_bandwidth.rt_runtime == 0) |
| 6406 | return -EPERM; |
| 6407 | #endif |
| 6408 | |
| 6409 | retval = security_task_setscheduler(p, policy, param); |
| 6410 | if (retval) |
| 6411 | return retval; |
| 6412 | } |
| 6413 | |
| 6414 | /* |
| 6415 | * make sure no PI-waiters arrive (or leave) while we are |
| 6416 | * changing the priority of the task: |
| 6417 | */ |
| 6418 | spin_lock_irqsave(&p->pi_lock, flags); |
| 6419 | /* |
| 6420 | * To be able to change p->policy safely, the apropriate |
| 6421 | * runqueue lock must be held. |
| 6422 | */ |
| 6423 | rq = __task_rq_lock(p); |
| 6424 | /* recheck policy now with rq lock held */ |
| 6425 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { |
| 6426 | policy = oldpolicy = -1; |
| 6427 | __task_rq_unlock(rq); |
| 6428 | spin_unlock_irqrestore(&p->pi_lock, flags); |
| 6429 | goto recheck; |
| 6430 | } |
| 6431 | update_rq_clock(rq); |
| 6432 | on_rq = p->se.on_rq; |
| 6433 | running = task_current(rq, p); |
| 6434 | if (on_rq) |
| 6435 | deactivate_task(rq, p, 0); |
| 6436 | if (running) |
| 6437 | p->sched_class->put_prev_task(rq, p); |
| 6438 | |
| 6439 | p->sched_reset_on_fork = reset_on_fork; |
| 6440 | |
| 6441 | oldprio = p->prio; |
| 6442 | __setscheduler(rq, p, policy, param->sched_priority); |
| 6443 | |
| 6444 | if (running) |
| 6445 | p->sched_class->set_curr_task(rq); |
| 6446 | if (on_rq) { |
| 6447 | activate_task(rq, p, 0); |
| 6448 | |
| 6449 | check_class_changed(rq, p, prev_class, oldprio, running); |
| 6450 | } |
| 6451 | __task_rq_unlock(rq); |
| 6452 | spin_unlock_irqrestore(&p->pi_lock, flags); |
| 6453 | |
| 6454 | rt_mutex_adjust_pi(p); |
| 6455 | |
| 6456 | return 0; |
| 6457 | } |
| 6458 | |
| 6459 | /** |
| 6460 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. |
| 6461 | * @p: the task in question. |
| 6462 | * @policy: new policy. |
| 6463 | * @param: structure containing the new RT priority. |
| 6464 | * |
| 6465 | * NOTE that the task may be already dead. |
| 6466 | */ |
| 6467 | int sched_setscheduler(struct task_struct *p, int policy, |
| 6468 | struct sched_param *param) |
| 6469 | { |
| 6470 | return __sched_setscheduler(p, policy, param, true); |
| 6471 | } |
| 6472 | EXPORT_SYMBOL_GPL(sched_setscheduler); |
| 6473 | |
| 6474 | /** |
| 6475 | * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. |
| 6476 | * @p: the task in question. |
| 6477 | * @policy: new policy. |
| 6478 | * @param: structure containing the new RT priority. |
| 6479 | * |
| 6480 | * Just like sched_setscheduler, only don't bother checking if the |
| 6481 | * current context has permission. For example, this is needed in |
| 6482 | * stop_machine(): we create temporary high priority worker threads, |
| 6483 | * but our caller might not have that capability. |
| 6484 | */ |
| 6485 | int sched_setscheduler_nocheck(struct task_struct *p, int policy, |
| 6486 | struct sched_param *param) |
| 6487 | { |
| 6488 | return __sched_setscheduler(p, policy, param, false); |
| 6489 | } |
| 6490 | |
| 6491 | static int |
| 6492 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) |
| 6493 | { |
| 6494 | struct sched_param lparam; |
| 6495 | struct task_struct *p; |
| 6496 | int retval; |
| 6497 | |
| 6498 | if (!param || pid < 0) |
| 6499 | return -EINVAL; |
| 6500 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) |
| 6501 | return -EFAULT; |
| 6502 | |
| 6503 | rcu_read_lock(); |
| 6504 | retval = -ESRCH; |
| 6505 | p = find_process_by_pid(pid); |
| 6506 | if (p != NULL) |
| 6507 | retval = sched_setscheduler(p, policy, &lparam); |
| 6508 | rcu_read_unlock(); |
| 6509 | |
| 6510 | return retval; |
| 6511 | } |
| 6512 | |
| 6513 | /** |
| 6514 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority |
| 6515 | * @pid: the pid in question. |
| 6516 | * @policy: new policy. |
| 6517 | * @param: structure containing the new RT priority. |
| 6518 | */ |
| 6519 | SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, |
| 6520 | struct sched_param __user *, param) |
| 6521 | { |
| 6522 | /* negative values for policy are not valid */ |
| 6523 | if (policy < 0) |
| 6524 | return -EINVAL; |
| 6525 | |
| 6526 | return do_sched_setscheduler(pid, policy, param); |
| 6527 | } |
| 6528 | |
| 6529 | /** |
| 6530 | * sys_sched_setparam - set/change the RT priority of a thread |
| 6531 | * @pid: the pid in question. |
| 6532 | * @param: structure containing the new RT priority. |
| 6533 | */ |
| 6534 | SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) |
| 6535 | { |
| 6536 | return do_sched_setscheduler(pid, -1, param); |
| 6537 | } |
| 6538 | |
| 6539 | /** |
| 6540 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread |
| 6541 | * @pid: the pid in question. |
| 6542 | */ |
| 6543 | SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) |
| 6544 | { |
| 6545 | struct task_struct *p; |
| 6546 | int retval; |
| 6547 | |
| 6548 | if (pid < 0) |
| 6549 | return -EINVAL; |
| 6550 | |
| 6551 | retval = -ESRCH; |
| 6552 | read_lock(&tasklist_lock); |
| 6553 | p = find_process_by_pid(pid); |
| 6554 | if (p) { |
| 6555 | retval = security_task_getscheduler(p); |
| 6556 | if (!retval) |
| 6557 | retval = p->policy |
| 6558 | | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); |
| 6559 | } |
| 6560 | read_unlock(&tasklist_lock); |
| 6561 | return retval; |
| 6562 | } |
| 6563 | |
| 6564 | /** |
| 6565 | * sys_sched_getparam - get the RT priority of a thread |
| 6566 | * @pid: the pid in question. |
| 6567 | * @param: structure containing the RT priority. |
| 6568 | */ |
| 6569 | SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) |
| 6570 | { |
| 6571 | struct sched_param lp; |
| 6572 | struct task_struct *p; |
| 6573 | int retval; |
| 6574 | |
| 6575 | if (!param || pid < 0) |
| 6576 | return -EINVAL; |
| 6577 | |
| 6578 | read_lock(&tasklist_lock); |
| 6579 | p = find_process_by_pid(pid); |
| 6580 | retval = -ESRCH; |
| 6581 | if (!p) |
| 6582 | goto out_unlock; |
| 6583 | |
| 6584 | retval = security_task_getscheduler(p); |
| 6585 | if (retval) |
| 6586 | goto out_unlock; |
| 6587 | |
| 6588 | lp.sched_priority = p->rt_priority; |
| 6589 | read_unlock(&tasklist_lock); |
| 6590 | |
| 6591 | /* |
| 6592 | * This one might sleep, we cannot do it with a spinlock held ... |
| 6593 | */ |
| 6594 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; |
| 6595 | |
| 6596 | return retval; |
| 6597 | |
| 6598 | out_unlock: |
| 6599 | read_unlock(&tasklist_lock); |
| 6600 | return retval; |
| 6601 | } |
| 6602 | |
| 6603 | long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) |
| 6604 | { |
| 6605 | cpumask_var_t cpus_allowed, new_mask; |
| 6606 | struct task_struct *p; |
| 6607 | int retval; |
| 6608 | |
| 6609 | get_online_cpus(); |
| 6610 | read_lock(&tasklist_lock); |
| 6611 | |
| 6612 | p = find_process_by_pid(pid); |
| 6613 | if (!p) { |
| 6614 | read_unlock(&tasklist_lock); |
| 6615 | put_online_cpus(); |
| 6616 | return -ESRCH; |
| 6617 | } |
| 6618 | |
| 6619 | /* |
| 6620 | * It is not safe to call set_cpus_allowed with the |
| 6621 | * tasklist_lock held. We will bump the task_struct's |
| 6622 | * usage count and then drop tasklist_lock. |
| 6623 | */ |
| 6624 | get_task_struct(p); |
| 6625 | read_unlock(&tasklist_lock); |
| 6626 | |
| 6627 | if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { |
| 6628 | retval = -ENOMEM; |
| 6629 | goto out_put_task; |
| 6630 | } |
| 6631 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { |
| 6632 | retval = -ENOMEM; |
| 6633 | goto out_free_cpus_allowed; |
| 6634 | } |
| 6635 | retval = -EPERM; |
| 6636 | if (!check_same_owner(p) && !capable(CAP_SYS_NICE)) |
| 6637 | goto out_unlock; |
| 6638 | |
| 6639 | retval = security_task_setscheduler(p, 0, NULL); |
| 6640 | if (retval) |
| 6641 | goto out_unlock; |
| 6642 | |
| 6643 | cpuset_cpus_allowed(p, cpus_allowed); |
| 6644 | cpumask_and(new_mask, in_mask, cpus_allowed); |
| 6645 | again: |
| 6646 | retval = set_cpus_allowed_ptr(p, new_mask); |
| 6647 | |
| 6648 | if (!retval) { |
| 6649 | cpuset_cpus_allowed(p, cpus_allowed); |
| 6650 | if (!cpumask_subset(new_mask, cpus_allowed)) { |
| 6651 | /* |
| 6652 | * We must have raced with a concurrent cpuset |
| 6653 | * update. Just reset the cpus_allowed to the |
| 6654 | * cpuset's cpus_allowed |
| 6655 | */ |
| 6656 | cpumask_copy(new_mask, cpus_allowed); |
| 6657 | goto again; |
| 6658 | } |
| 6659 | } |
| 6660 | out_unlock: |
| 6661 | free_cpumask_var(new_mask); |
| 6662 | out_free_cpus_allowed: |
| 6663 | free_cpumask_var(cpus_allowed); |
| 6664 | out_put_task: |
| 6665 | put_task_struct(p); |
| 6666 | put_online_cpus(); |
| 6667 | return retval; |
| 6668 | } |
| 6669 | |
| 6670 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, |
| 6671 | struct cpumask *new_mask) |
| 6672 | { |
| 6673 | if (len < cpumask_size()) |
| 6674 | cpumask_clear(new_mask); |
| 6675 | else if (len > cpumask_size()) |
| 6676 | len = cpumask_size(); |
| 6677 | |
| 6678 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; |
| 6679 | } |
| 6680 | |
| 6681 | /** |
| 6682 | * sys_sched_setaffinity - set the cpu affinity of a process |
| 6683 | * @pid: pid of the process |
| 6684 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 6685 | * @user_mask_ptr: user-space pointer to the new cpu mask |
| 6686 | */ |
| 6687 | SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, |
| 6688 | unsigned long __user *, user_mask_ptr) |
| 6689 | { |
| 6690 | cpumask_var_t new_mask; |
| 6691 | int retval; |
| 6692 | |
| 6693 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) |
| 6694 | return -ENOMEM; |
| 6695 | |
| 6696 | retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); |
| 6697 | if (retval == 0) |
| 6698 | retval = sched_setaffinity(pid, new_mask); |
| 6699 | free_cpumask_var(new_mask); |
| 6700 | return retval; |
| 6701 | } |
| 6702 | |
| 6703 | long sched_getaffinity(pid_t pid, struct cpumask *mask) |
| 6704 | { |
| 6705 | struct task_struct *p; |
| 6706 | int retval; |
| 6707 | |
| 6708 | get_online_cpus(); |
| 6709 | read_lock(&tasklist_lock); |
| 6710 | |
| 6711 | retval = -ESRCH; |
| 6712 | p = find_process_by_pid(pid); |
| 6713 | if (!p) |
| 6714 | goto out_unlock; |
| 6715 | |
| 6716 | retval = security_task_getscheduler(p); |
| 6717 | if (retval) |
| 6718 | goto out_unlock; |
| 6719 | |
| 6720 | cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); |
| 6721 | |
| 6722 | out_unlock: |
| 6723 | read_unlock(&tasklist_lock); |
| 6724 | put_online_cpus(); |
| 6725 | |
| 6726 | return retval; |
| 6727 | } |
| 6728 | |
| 6729 | /** |
| 6730 | * sys_sched_getaffinity - get the cpu affinity of a process |
| 6731 | * @pid: pid of the process |
| 6732 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 6733 | * @user_mask_ptr: user-space pointer to hold the current cpu mask |
| 6734 | */ |
| 6735 | SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, |
| 6736 | unsigned long __user *, user_mask_ptr) |
| 6737 | { |
| 6738 | int ret; |
| 6739 | cpumask_var_t mask; |
| 6740 | |
| 6741 | if (len < cpumask_size()) |
| 6742 | return -EINVAL; |
| 6743 | |
| 6744 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) |
| 6745 | return -ENOMEM; |
| 6746 | |
| 6747 | ret = sched_getaffinity(pid, mask); |
| 6748 | if (ret == 0) { |
| 6749 | if (copy_to_user(user_mask_ptr, mask, cpumask_size())) |
| 6750 | ret = -EFAULT; |
| 6751 | else |
| 6752 | ret = cpumask_size(); |
| 6753 | } |
| 6754 | free_cpumask_var(mask); |
| 6755 | |
| 6756 | return ret; |
| 6757 | } |
| 6758 | |
| 6759 | /** |
| 6760 | * sys_sched_yield - yield the current processor to other threads. |
| 6761 | * |
| 6762 | * This function yields the current CPU to other tasks. If there are no |
| 6763 | * other threads running on this CPU then this function will return. |
| 6764 | */ |
| 6765 | SYSCALL_DEFINE0(sched_yield) |
| 6766 | { |
| 6767 | struct rq *rq = this_rq_lock(); |
| 6768 | |
| 6769 | schedstat_inc(rq, yld_count); |
| 6770 | current->sched_class->yield_task(rq); |
| 6771 | |
| 6772 | /* |
| 6773 | * Since we are going to call schedule() anyway, there's |
| 6774 | * no need to preempt or enable interrupts: |
| 6775 | */ |
| 6776 | __release(rq->lock); |
| 6777 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
| 6778 | _raw_spin_unlock(&rq->lock); |
| 6779 | preempt_enable_no_resched(); |
| 6780 | |
| 6781 | schedule(); |
| 6782 | |
| 6783 | return 0; |
| 6784 | } |
| 6785 | |
| 6786 | static inline int should_resched(void) |
| 6787 | { |
| 6788 | return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); |
| 6789 | } |
| 6790 | |
| 6791 | static void __cond_resched(void) |
| 6792 | { |
| 6793 | add_preempt_count(PREEMPT_ACTIVE); |
| 6794 | schedule(); |
| 6795 | sub_preempt_count(PREEMPT_ACTIVE); |
| 6796 | } |
| 6797 | |
| 6798 | int __sched _cond_resched(void) |
| 6799 | { |
| 6800 | if (should_resched()) { |
| 6801 | __cond_resched(); |
| 6802 | return 1; |
| 6803 | } |
| 6804 | return 0; |
| 6805 | } |
| 6806 | EXPORT_SYMBOL(_cond_resched); |
| 6807 | |
| 6808 | /* |
| 6809 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, |
| 6810 | * call schedule, and on return reacquire the lock. |
| 6811 | * |
| 6812 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
| 6813 | * operations here to prevent schedule() from being called twice (once via |
| 6814 | * spin_unlock(), once by hand). |
| 6815 | */ |
| 6816 | int __cond_resched_lock(spinlock_t *lock) |
| 6817 | { |
| 6818 | int resched = should_resched(); |
| 6819 | int ret = 0; |
| 6820 | |
| 6821 | lockdep_assert_held(lock); |
| 6822 | |
| 6823 | if (spin_needbreak(lock) || resched) { |
| 6824 | spin_unlock(lock); |
| 6825 | if (resched) |
| 6826 | __cond_resched(); |
| 6827 | else |
| 6828 | cpu_relax(); |
| 6829 | ret = 1; |
| 6830 | spin_lock(lock); |
| 6831 | } |
| 6832 | return ret; |
| 6833 | } |
| 6834 | EXPORT_SYMBOL(__cond_resched_lock); |
| 6835 | |
| 6836 | int __sched __cond_resched_softirq(void) |
| 6837 | { |
| 6838 | BUG_ON(!in_softirq()); |
| 6839 | |
| 6840 | if (should_resched()) { |
| 6841 | local_bh_enable(); |
| 6842 | __cond_resched(); |
| 6843 | local_bh_disable(); |
| 6844 | return 1; |
| 6845 | } |
| 6846 | return 0; |
| 6847 | } |
| 6848 | EXPORT_SYMBOL(__cond_resched_softirq); |
| 6849 | |
| 6850 | /** |
| 6851 | * yield - yield the current processor to other threads. |
| 6852 | * |
| 6853 | * This is a shortcut for kernel-space yielding - it marks the |
| 6854 | * thread runnable and calls sys_sched_yield(). |
| 6855 | */ |
| 6856 | void __sched yield(void) |
| 6857 | { |
| 6858 | set_current_state(TASK_RUNNING); |
| 6859 | sys_sched_yield(); |
| 6860 | } |
| 6861 | EXPORT_SYMBOL(yield); |
| 6862 | |
| 6863 | /* |
| 6864 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
| 6865 | * that process accounting knows that this is a task in IO wait state. |
| 6866 | * |
| 6867 | * But don't do that if it is a deliberate, throttling IO wait (this task |
| 6868 | * has set its backing_dev_info: the queue against which it should throttle) |
| 6869 | */ |
| 6870 | void __sched io_schedule(void) |
| 6871 | { |
| 6872 | struct rq *rq = raw_rq(); |
| 6873 | |
| 6874 | delayacct_blkio_start(); |
| 6875 | atomic_inc(&rq->nr_iowait); |
| 6876 | current->in_iowait = 1; |
| 6877 | schedule(); |
| 6878 | current->in_iowait = 0; |
| 6879 | atomic_dec(&rq->nr_iowait); |
| 6880 | delayacct_blkio_end(); |
| 6881 | } |
| 6882 | EXPORT_SYMBOL(io_schedule); |
| 6883 | |
| 6884 | long __sched io_schedule_timeout(long timeout) |
| 6885 | { |
| 6886 | struct rq *rq = raw_rq(); |
| 6887 | long ret; |
| 6888 | |
| 6889 | delayacct_blkio_start(); |
| 6890 | atomic_inc(&rq->nr_iowait); |
| 6891 | current->in_iowait = 1; |
| 6892 | ret = schedule_timeout(timeout); |
| 6893 | current->in_iowait = 0; |
| 6894 | atomic_dec(&rq->nr_iowait); |
| 6895 | delayacct_blkio_end(); |
| 6896 | return ret; |
| 6897 | } |
| 6898 | |
| 6899 | /** |
| 6900 | * sys_sched_get_priority_max - return maximum RT priority. |
| 6901 | * @policy: scheduling class. |
| 6902 | * |
| 6903 | * this syscall returns the maximum rt_priority that can be used |
| 6904 | * by a given scheduling class. |
| 6905 | */ |
| 6906 | SYSCALL_DEFINE1(sched_get_priority_max, int, policy) |
| 6907 | { |
| 6908 | int ret = -EINVAL; |
| 6909 | |
| 6910 | switch (policy) { |
| 6911 | case SCHED_FIFO: |
| 6912 | case SCHED_RR: |
| 6913 | ret = MAX_USER_RT_PRIO-1; |
| 6914 | break; |
| 6915 | case SCHED_NORMAL: |
| 6916 | case SCHED_BATCH: |
| 6917 | case SCHED_IDLE: |
| 6918 | ret = 0; |
| 6919 | break; |
| 6920 | } |
| 6921 | return ret; |
| 6922 | } |
| 6923 | |
| 6924 | /** |
| 6925 | * sys_sched_get_priority_min - return minimum RT priority. |
| 6926 | * @policy: scheduling class. |
| 6927 | * |
| 6928 | * this syscall returns the minimum rt_priority that can be used |
| 6929 | * by a given scheduling class. |
| 6930 | */ |
| 6931 | SYSCALL_DEFINE1(sched_get_priority_min, int, policy) |
| 6932 | { |
| 6933 | int ret = -EINVAL; |
| 6934 | |
| 6935 | switch (policy) { |
| 6936 | case SCHED_FIFO: |
| 6937 | case SCHED_RR: |
| 6938 | ret = 1; |
| 6939 | break; |
| 6940 | case SCHED_NORMAL: |
| 6941 | case SCHED_BATCH: |
| 6942 | case SCHED_IDLE: |
| 6943 | ret = 0; |
| 6944 | } |
| 6945 | return ret; |
| 6946 | } |
| 6947 | |
| 6948 | /** |
| 6949 | * sys_sched_rr_get_interval - return the default timeslice of a process. |
| 6950 | * @pid: pid of the process. |
| 6951 | * @interval: userspace pointer to the timeslice value. |
| 6952 | * |
| 6953 | * this syscall writes the default timeslice value of a given process |
| 6954 | * into the user-space timespec buffer. A value of '0' means infinity. |
| 6955 | */ |
| 6956 | SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, |
| 6957 | struct timespec __user *, interval) |
| 6958 | { |
| 6959 | struct task_struct *p; |
| 6960 | unsigned int time_slice; |
| 6961 | int retval; |
| 6962 | struct timespec t; |
| 6963 | |
| 6964 | if (pid < 0) |
| 6965 | return -EINVAL; |
| 6966 | |
| 6967 | retval = -ESRCH; |
| 6968 | read_lock(&tasklist_lock); |
| 6969 | p = find_process_by_pid(pid); |
| 6970 | if (!p) |
| 6971 | goto out_unlock; |
| 6972 | |
| 6973 | retval = security_task_getscheduler(p); |
| 6974 | if (retval) |
| 6975 | goto out_unlock; |
| 6976 | |
| 6977 | /* |
| 6978 | * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER |
| 6979 | * tasks that are on an otherwise idle runqueue: |
| 6980 | */ |
| 6981 | time_slice = 0; |
| 6982 | if (p->policy == SCHED_RR) { |
| 6983 | time_slice = DEF_TIMESLICE; |
| 6984 | } else if (p->policy != SCHED_FIFO) { |
| 6985 | struct sched_entity *se = &p->se; |
| 6986 | unsigned long flags; |
| 6987 | struct rq *rq; |
| 6988 | |
| 6989 | rq = task_rq_lock(p, &flags); |
| 6990 | if (rq->cfs.load.weight) |
| 6991 | time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); |
| 6992 | task_rq_unlock(rq, &flags); |
| 6993 | } |
| 6994 | read_unlock(&tasklist_lock); |
| 6995 | jiffies_to_timespec(time_slice, &t); |
| 6996 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
| 6997 | return retval; |
| 6998 | |
| 6999 | out_unlock: |
| 7000 | read_unlock(&tasklist_lock); |
| 7001 | return retval; |
| 7002 | } |
| 7003 | |
| 7004 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; |
| 7005 | |
| 7006 | void sched_show_task(struct task_struct *p) |
| 7007 | { |
| 7008 | unsigned long free = 0; |
| 7009 | unsigned state; |
| 7010 | |
| 7011 | state = p->state ? __ffs(p->state) + 1 : 0; |
| 7012 | printk(KERN_INFO "%-13.13s %c", p->comm, |
| 7013 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); |
| 7014 | #if BITS_PER_LONG == 32 |
| 7015 | if (state == TASK_RUNNING) |
| 7016 | printk(KERN_CONT " running "); |
| 7017 | else |
| 7018 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); |
| 7019 | #else |
| 7020 | if (state == TASK_RUNNING) |
| 7021 | printk(KERN_CONT " running task "); |
| 7022 | else |
| 7023 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); |
| 7024 | #endif |
| 7025 | #ifdef CONFIG_DEBUG_STACK_USAGE |
| 7026 | free = stack_not_used(p); |
| 7027 | #endif |
| 7028 | printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, |
| 7029 | task_pid_nr(p), task_pid_nr(p->real_parent), |
| 7030 | (unsigned long)task_thread_info(p)->flags); |
| 7031 | |
| 7032 | show_stack(p, NULL); |
| 7033 | } |
| 7034 | |
| 7035 | void show_state_filter(unsigned long state_filter) |
| 7036 | { |
| 7037 | struct task_struct *g, *p; |
| 7038 | |
| 7039 | #if BITS_PER_LONG == 32 |
| 7040 | printk(KERN_INFO |
| 7041 | " task PC stack pid father\n"); |
| 7042 | #else |
| 7043 | printk(KERN_INFO |
| 7044 | " task PC stack pid father\n"); |
| 7045 | #endif |
| 7046 | read_lock(&tasklist_lock); |
| 7047 | do_each_thread(g, p) { |
| 7048 | /* |
| 7049 | * reset the NMI-timeout, listing all files on a slow |
| 7050 | * console might take alot of time: |
| 7051 | */ |
| 7052 | touch_nmi_watchdog(); |
| 7053 | if (!state_filter || (p->state & state_filter)) |
| 7054 | sched_show_task(p); |
| 7055 | } while_each_thread(g, p); |
| 7056 | |
| 7057 | touch_all_softlockup_watchdogs(); |
| 7058 | |
| 7059 | #ifdef CONFIG_SCHED_DEBUG |
| 7060 | sysrq_sched_debug_show(); |
| 7061 | #endif |
| 7062 | read_unlock(&tasklist_lock); |
| 7063 | /* |
| 7064 | * Only show locks if all tasks are dumped: |
| 7065 | */ |
| 7066 | if (state_filter == -1) |
| 7067 | debug_show_all_locks(); |
| 7068 | } |
| 7069 | |
| 7070 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) |
| 7071 | { |
| 7072 | idle->sched_class = &idle_sched_class; |
| 7073 | } |
| 7074 | |
| 7075 | /** |
| 7076 | * init_idle - set up an idle thread for a given CPU |
| 7077 | * @idle: task in question |
| 7078 | * @cpu: cpu the idle task belongs to |
| 7079 | * |
| 7080 | * NOTE: this function does not set the idle thread's NEED_RESCHED |
| 7081 | * flag, to make booting more robust. |
| 7082 | */ |
| 7083 | void __cpuinit init_idle(struct task_struct *idle, int cpu) |
| 7084 | { |
| 7085 | struct rq *rq = cpu_rq(cpu); |
| 7086 | unsigned long flags; |
| 7087 | |
| 7088 | spin_lock_irqsave(&rq->lock, flags); |
| 7089 | |
| 7090 | __sched_fork(idle); |
| 7091 | idle->se.exec_start = sched_clock(); |
| 7092 | |
| 7093 | idle->prio = idle->normal_prio = MAX_PRIO; |
| 7094 | cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu)); |
| 7095 | __set_task_cpu(idle, cpu); |
| 7096 | |
| 7097 | rq->curr = rq->idle = idle; |
| 7098 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
| 7099 | idle->oncpu = 1; |
| 7100 | #endif |
| 7101 | spin_unlock_irqrestore(&rq->lock, flags); |
| 7102 | |
| 7103 | /* Set the preempt count _outside_ the spinlocks! */ |
| 7104 | #if defined(CONFIG_PREEMPT) |
| 7105 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
| 7106 | #else |
| 7107 | task_thread_info(idle)->preempt_count = 0; |
| 7108 | #endif |
| 7109 | /* |
| 7110 | * The idle tasks have their own, simple scheduling class: |
| 7111 | */ |
| 7112 | idle->sched_class = &idle_sched_class; |
| 7113 | ftrace_graph_init_task(idle); |
| 7114 | } |
| 7115 | |
| 7116 | /* |
| 7117 | * In a system that switches off the HZ timer nohz_cpu_mask |
| 7118 | * indicates which cpus entered this state. This is used |
| 7119 | * in the rcu update to wait only for active cpus. For system |
| 7120 | * which do not switch off the HZ timer nohz_cpu_mask should |
| 7121 | * always be CPU_BITS_NONE. |
| 7122 | */ |
| 7123 | cpumask_var_t nohz_cpu_mask; |
| 7124 | |
| 7125 | /* |
| 7126 | * Increase the granularity value when there are more CPUs, |
| 7127 | * because with more CPUs the 'effective latency' as visible |
| 7128 | * to users decreases. But the relationship is not linear, |
| 7129 | * so pick a second-best guess by going with the log2 of the |
| 7130 | * number of CPUs. |
| 7131 | * |
| 7132 | * This idea comes from the SD scheduler of Con Kolivas: |
| 7133 | */ |
| 7134 | static inline void sched_init_granularity(void) |
| 7135 | { |
| 7136 | unsigned int factor = 1 + ilog2(num_online_cpus()); |
| 7137 | const unsigned long limit = 200000000; |
| 7138 | |
| 7139 | sysctl_sched_min_granularity *= factor; |
| 7140 | if (sysctl_sched_min_granularity > limit) |
| 7141 | sysctl_sched_min_granularity = limit; |
| 7142 | |
| 7143 | sysctl_sched_latency *= factor; |
| 7144 | if (sysctl_sched_latency > limit) |
| 7145 | sysctl_sched_latency = limit; |
| 7146 | |
| 7147 | sysctl_sched_wakeup_granularity *= factor; |
| 7148 | |
| 7149 | sysctl_sched_shares_ratelimit *= factor; |
| 7150 | } |
| 7151 | |
| 7152 | #ifdef CONFIG_SMP |
| 7153 | /* |
| 7154 | * This is how migration works: |
| 7155 | * |
| 7156 | * 1) we queue a struct migration_req structure in the source CPU's |
| 7157 | * runqueue and wake up that CPU's migration thread. |
| 7158 | * 2) we down() the locked semaphore => thread blocks. |
| 7159 | * 3) migration thread wakes up (implicitly it forces the migrated |
| 7160 | * thread off the CPU) |
| 7161 | * 4) it gets the migration request and checks whether the migrated |
| 7162 | * task is still in the wrong runqueue. |
| 7163 | * 5) if it's in the wrong runqueue then the migration thread removes |
| 7164 | * it and puts it into the right queue. |
| 7165 | * 6) migration thread up()s the semaphore. |
| 7166 | * 7) we wake up and the migration is done. |
| 7167 | */ |
| 7168 | |
| 7169 | /* |
| 7170 | * Change a given task's CPU affinity. Migrate the thread to a |
| 7171 | * proper CPU and schedule it away if the CPU it's executing on |
| 7172 | * is removed from the allowed bitmask. |
| 7173 | * |
| 7174 | * NOTE: the caller must have a valid reference to the task, the |
| 7175 | * task must not exit() & deallocate itself prematurely. The |
| 7176 | * call is not atomic; no spinlocks may be held. |
| 7177 | */ |
| 7178 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) |
| 7179 | { |
| 7180 | struct migration_req req; |
| 7181 | unsigned long flags; |
| 7182 | struct rq *rq; |
| 7183 | int ret = 0; |
| 7184 | |
| 7185 | rq = task_rq_lock(p, &flags); |
| 7186 | if (!cpumask_intersects(new_mask, cpu_online_mask)) { |
| 7187 | ret = -EINVAL; |
| 7188 | goto out; |
| 7189 | } |
| 7190 | |
| 7191 | if (unlikely((p->flags & PF_THREAD_BOUND) && p != current && |
| 7192 | !cpumask_equal(&p->cpus_allowed, new_mask))) { |
| 7193 | ret = -EINVAL; |
| 7194 | goto out; |
| 7195 | } |
| 7196 | |
| 7197 | if (p->sched_class->set_cpus_allowed) |
| 7198 | p->sched_class->set_cpus_allowed(p, new_mask); |
| 7199 | else { |
| 7200 | cpumask_copy(&p->cpus_allowed, new_mask); |
| 7201 | p->rt.nr_cpus_allowed = cpumask_weight(new_mask); |
| 7202 | } |
| 7203 | |
| 7204 | /* Can the task run on the task's current CPU? If so, we're done */ |
| 7205 | if (cpumask_test_cpu(task_cpu(p), new_mask)) |
| 7206 | goto out; |
| 7207 | |
| 7208 | if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) { |
| 7209 | /* Need help from migration thread: drop lock and wait. */ |
| 7210 | struct task_struct *mt = rq->migration_thread; |
| 7211 | |
| 7212 | get_task_struct(mt); |
| 7213 | task_rq_unlock(rq, &flags); |
| 7214 | wake_up_process(rq->migration_thread); |
| 7215 | put_task_struct(mt); |
| 7216 | wait_for_completion(&req.done); |
| 7217 | tlb_migrate_finish(p->mm); |
| 7218 | return 0; |
| 7219 | } |
| 7220 | out: |
| 7221 | task_rq_unlock(rq, &flags); |
| 7222 | |
| 7223 | return ret; |
| 7224 | } |
| 7225 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); |
| 7226 | |
| 7227 | /* |
| 7228 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
| 7229 | * this because either it can't run here any more (set_cpus_allowed() |
| 7230 | * away from this CPU, or CPU going down), or because we're |
| 7231 | * attempting to rebalance this task on exec (sched_exec). |
| 7232 | * |
| 7233 | * So we race with normal scheduler movements, but that's OK, as long |
| 7234 | * as the task is no longer on this CPU. |
| 7235 | * |
| 7236 | * Returns non-zero if task was successfully migrated. |
| 7237 | */ |
| 7238 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
| 7239 | { |
| 7240 | struct rq *rq_dest, *rq_src; |
| 7241 | int ret = 0, on_rq; |
| 7242 | |
| 7243 | if (unlikely(!cpu_active(dest_cpu))) |
| 7244 | return ret; |
| 7245 | |
| 7246 | rq_src = cpu_rq(src_cpu); |
| 7247 | rq_dest = cpu_rq(dest_cpu); |
| 7248 | |
| 7249 | double_rq_lock(rq_src, rq_dest); |
| 7250 | /* Already moved. */ |
| 7251 | if (task_cpu(p) != src_cpu) |
| 7252 | goto done; |
| 7253 | /* Affinity changed (again). */ |
| 7254 | if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) |
| 7255 | goto fail; |
| 7256 | |
| 7257 | on_rq = p->se.on_rq; |
| 7258 | if (on_rq) |
| 7259 | deactivate_task(rq_src, p, 0); |
| 7260 | |
| 7261 | set_task_cpu(p, dest_cpu); |
| 7262 | if (on_rq) { |
| 7263 | activate_task(rq_dest, p, 0); |
| 7264 | check_preempt_curr(rq_dest, p, 0); |
| 7265 | } |
| 7266 | done: |
| 7267 | ret = 1; |
| 7268 | fail: |
| 7269 | double_rq_unlock(rq_src, rq_dest); |
| 7270 | return ret; |
| 7271 | } |
| 7272 | |
| 7273 | #define RCU_MIGRATION_IDLE 0 |
| 7274 | #define RCU_MIGRATION_NEED_QS 1 |
| 7275 | #define RCU_MIGRATION_GOT_QS 2 |
| 7276 | #define RCU_MIGRATION_MUST_SYNC 3 |
| 7277 | |
| 7278 | /* |
| 7279 | * migration_thread - this is a highprio system thread that performs |
| 7280 | * thread migration by bumping thread off CPU then 'pushing' onto |
| 7281 | * another runqueue. |
| 7282 | */ |
| 7283 | static int migration_thread(void *data) |
| 7284 | { |
| 7285 | int badcpu; |
| 7286 | int cpu = (long)data; |
| 7287 | struct rq *rq; |
| 7288 | |
| 7289 | rq = cpu_rq(cpu); |
| 7290 | BUG_ON(rq->migration_thread != current); |
| 7291 | |
| 7292 | set_current_state(TASK_INTERRUPTIBLE); |
| 7293 | while (!kthread_should_stop()) { |
| 7294 | struct migration_req *req; |
| 7295 | struct list_head *head; |
| 7296 | |
| 7297 | spin_lock_irq(&rq->lock); |
| 7298 | |
| 7299 | if (cpu_is_offline(cpu)) { |
| 7300 | spin_unlock_irq(&rq->lock); |
| 7301 | break; |
| 7302 | } |
| 7303 | |
| 7304 | if (rq->active_balance) { |
| 7305 | active_load_balance(rq, cpu); |
| 7306 | rq->active_balance = 0; |
| 7307 | } |
| 7308 | |
| 7309 | head = &rq->migration_queue; |
| 7310 | |
| 7311 | if (list_empty(head)) { |
| 7312 | spin_unlock_irq(&rq->lock); |
| 7313 | schedule(); |
| 7314 | set_current_state(TASK_INTERRUPTIBLE); |
| 7315 | continue; |
| 7316 | } |
| 7317 | req = list_entry(head->next, struct migration_req, list); |
| 7318 | list_del_init(head->next); |
| 7319 | |
| 7320 | if (req->task != NULL) { |
| 7321 | spin_unlock(&rq->lock); |
| 7322 | __migrate_task(req->task, cpu, req->dest_cpu); |
| 7323 | } else if (likely(cpu == (badcpu = smp_processor_id()))) { |
| 7324 | req->dest_cpu = RCU_MIGRATION_GOT_QS; |
| 7325 | spin_unlock(&rq->lock); |
| 7326 | } else { |
| 7327 | req->dest_cpu = RCU_MIGRATION_MUST_SYNC; |
| 7328 | spin_unlock(&rq->lock); |
| 7329 | WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu); |
| 7330 | } |
| 7331 | local_irq_enable(); |
| 7332 | |
| 7333 | complete(&req->done); |
| 7334 | } |
| 7335 | __set_current_state(TASK_RUNNING); |
| 7336 | |
| 7337 | return 0; |
| 7338 | } |
| 7339 | |
| 7340 | #ifdef CONFIG_HOTPLUG_CPU |
| 7341 | |
| 7342 | static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu) |
| 7343 | { |
| 7344 | int ret; |
| 7345 | |
| 7346 | local_irq_disable(); |
| 7347 | ret = __migrate_task(p, src_cpu, dest_cpu); |
| 7348 | local_irq_enable(); |
| 7349 | return ret; |
| 7350 | } |
| 7351 | |
| 7352 | /* |
| 7353 | * Figure out where task on dead CPU should go, use force if necessary. |
| 7354 | */ |
| 7355 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) |
| 7356 | { |
| 7357 | int dest_cpu; |
| 7358 | const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu)); |
| 7359 | |
| 7360 | again: |
| 7361 | /* Look for allowed, online CPU in same node. */ |
| 7362 | for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask) |
| 7363 | if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) |
| 7364 | goto move; |
| 7365 | |
| 7366 | /* Any allowed, online CPU? */ |
| 7367 | dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask); |
| 7368 | if (dest_cpu < nr_cpu_ids) |
| 7369 | goto move; |
| 7370 | |
| 7371 | /* No more Mr. Nice Guy. */ |
| 7372 | if (dest_cpu >= nr_cpu_ids) { |
| 7373 | cpuset_cpus_allowed_locked(p, &p->cpus_allowed); |
| 7374 | dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed); |
| 7375 | |
| 7376 | /* |
| 7377 | * Don't tell them about moving exiting tasks or |
| 7378 | * kernel threads (both mm NULL), since they never |
| 7379 | * leave kernel. |
| 7380 | */ |
| 7381 | if (p->mm && printk_ratelimit()) { |
| 7382 | printk(KERN_INFO "process %d (%s) no " |
| 7383 | "longer affine to cpu%d\n", |
| 7384 | task_pid_nr(p), p->comm, dead_cpu); |
| 7385 | } |
| 7386 | } |
| 7387 | |
| 7388 | move: |
| 7389 | /* It can have affinity changed while we were choosing. */ |
| 7390 | if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu))) |
| 7391 | goto again; |
| 7392 | } |
| 7393 | |
| 7394 | /* |
| 7395 | * While a dead CPU has no uninterruptible tasks queued at this point, |
| 7396 | * it might still have a nonzero ->nr_uninterruptible counter, because |
| 7397 | * for performance reasons the counter is not stricly tracking tasks to |
| 7398 | * their home CPUs. So we just add the counter to another CPU's counter, |
| 7399 | * to keep the global sum constant after CPU-down: |
| 7400 | */ |
| 7401 | static void migrate_nr_uninterruptible(struct rq *rq_src) |
| 7402 | { |
| 7403 | struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask)); |
| 7404 | unsigned long flags; |
| 7405 | |
| 7406 | local_irq_save(flags); |
| 7407 | double_rq_lock(rq_src, rq_dest); |
| 7408 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; |
| 7409 | rq_src->nr_uninterruptible = 0; |
| 7410 | double_rq_unlock(rq_src, rq_dest); |
| 7411 | local_irq_restore(flags); |
| 7412 | } |
| 7413 | |
| 7414 | /* Run through task list and migrate tasks from the dead cpu. */ |
| 7415 | static void migrate_live_tasks(int src_cpu) |
| 7416 | { |
| 7417 | struct task_struct *p, *t; |
| 7418 | |
| 7419 | read_lock(&tasklist_lock); |
| 7420 | |
| 7421 | do_each_thread(t, p) { |
| 7422 | if (p == current) |
| 7423 | continue; |
| 7424 | |
| 7425 | if (task_cpu(p) == src_cpu) |
| 7426 | move_task_off_dead_cpu(src_cpu, p); |
| 7427 | } while_each_thread(t, p); |
| 7428 | |
| 7429 | read_unlock(&tasklist_lock); |
| 7430 | } |
| 7431 | |
| 7432 | /* |
| 7433 | * Schedules idle task to be the next runnable task on current CPU. |
| 7434 | * It does so by boosting its priority to highest possible. |
| 7435 | * Used by CPU offline code. |
| 7436 | */ |
| 7437 | void sched_idle_next(void) |
| 7438 | { |
| 7439 | int this_cpu = smp_processor_id(); |
| 7440 | struct rq *rq = cpu_rq(this_cpu); |
| 7441 | struct task_struct *p = rq->idle; |
| 7442 | unsigned long flags; |
| 7443 | |
| 7444 | /* cpu has to be offline */ |
| 7445 | BUG_ON(cpu_online(this_cpu)); |
| 7446 | |
| 7447 | /* |
| 7448 | * Strictly not necessary since rest of the CPUs are stopped by now |
| 7449 | * and interrupts disabled on the current cpu. |
| 7450 | */ |
| 7451 | spin_lock_irqsave(&rq->lock, flags); |
| 7452 | |
| 7453 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
| 7454 | |
| 7455 | update_rq_clock(rq); |
| 7456 | activate_task(rq, p, 0); |
| 7457 | |
| 7458 | spin_unlock_irqrestore(&rq->lock, flags); |
| 7459 | } |
| 7460 | |
| 7461 | /* |
| 7462 | * Ensures that the idle task is using init_mm right before its cpu goes |
| 7463 | * offline. |
| 7464 | */ |
| 7465 | void idle_task_exit(void) |
| 7466 | { |
| 7467 | struct mm_struct *mm = current->active_mm; |
| 7468 | |
| 7469 | BUG_ON(cpu_online(smp_processor_id())); |
| 7470 | |
| 7471 | if (mm != &init_mm) |
| 7472 | switch_mm(mm, &init_mm, current); |
| 7473 | mmdrop(mm); |
| 7474 | } |
| 7475 | |
| 7476 | /* called under rq->lock with disabled interrupts */ |
| 7477 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) |
| 7478 | { |
| 7479 | struct rq *rq = cpu_rq(dead_cpu); |
| 7480 | |
| 7481 | /* Must be exiting, otherwise would be on tasklist. */ |
| 7482 | BUG_ON(!p->exit_state); |
| 7483 | |
| 7484 | /* Cannot have done final schedule yet: would have vanished. */ |
| 7485 | BUG_ON(p->state == TASK_DEAD); |
| 7486 | |
| 7487 | get_task_struct(p); |
| 7488 | |
| 7489 | /* |
| 7490 | * Drop lock around migration; if someone else moves it, |
| 7491 | * that's OK. No task can be added to this CPU, so iteration is |
| 7492 | * fine. |
| 7493 | */ |
| 7494 | spin_unlock_irq(&rq->lock); |
| 7495 | move_task_off_dead_cpu(dead_cpu, p); |
| 7496 | spin_lock_irq(&rq->lock); |
| 7497 | |
| 7498 | put_task_struct(p); |
| 7499 | } |
| 7500 | |
| 7501 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ |
| 7502 | static void migrate_dead_tasks(unsigned int dead_cpu) |
| 7503 | { |
| 7504 | struct rq *rq = cpu_rq(dead_cpu); |
| 7505 | struct task_struct *next; |
| 7506 | |
| 7507 | for ( ; ; ) { |
| 7508 | if (!rq->nr_running) |
| 7509 | break; |
| 7510 | update_rq_clock(rq); |
| 7511 | next = pick_next_task(rq); |
| 7512 | if (!next) |
| 7513 | break; |
| 7514 | next->sched_class->put_prev_task(rq, next); |
| 7515 | migrate_dead(dead_cpu, next); |
| 7516 | |
| 7517 | } |
| 7518 | } |
| 7519 | |
| 7520 | /* |
| 7521 | * remove the tasks which were accounted by rq from calc_load_tasks. |
| 7522 | */ |
| 7523 | static void calc_global_load_remove(struct rq *rq) |
| 7524 | { |
| 7525 | atomic_long_sub(rq->calc_load_active, &calc_load_tasks); |
| 7526 | rq->calc_load_active = 0; |
| 7527 | } |
| 7528 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 7529 | |
| 7530 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) |
| 7531 | |
| 7532 | static struct ctl_table sd_ctl_dir[] = { |
| 7533 | { |
| 7534 | .procname = "sched_domain", |
| 7535 | .mode = 0555, |
| 7536 | }, |
| 7537 | {0, }, |
| 7538 | }; |
| 7539 | |
| 7540 | static struct ctl_table sd_ctl_root[] = { |
| 7541 | { |
| 7542 | .ctl_name = CTL_KERN, |
| 7543 | .procname = "kernel", |
| 7544 | .mode = 0555, |
| 7545 | .child = sd_ctl_dir, |
| 7546 | }, |
| 7547 | {0, }, |
| 7548 | }; |
| 7549 | |
| 7550 | static struct ctl_table *sd_alloc_ctl_entry(int n) |
| 7551 | { |
| 7552 | struct ctl_table *entry = |
| 7553 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); |
| 7554 | |
| 7555 | return entry; |
| 7556 | } |
| 7557 | |
| 7558 | static void sd_free_ctl_entry(struct ctl_table **tablep) |
| 7559 | { |
| 7560 | struct ctl_table *entry; |
| 7561 | |
| 7562 | /* |
| 7563 | * In the intermediate directories, both the child directory and |
| 7564 | * procname are dynamically allocated and could fail but the mode |
| 7565 | * will always be set. In the lowest directory the names are |
| 7566 | * static strings and all have proc handlers. |
| 7567 | */ |
| 7568 | for (entry = *tablep; entry->mode; entry++) { |
| 7569 | if (entry->child) |
| 7570 | sd_free_ctl_entry(&entry->child); |
| 7571 | if (entry->proc_handler == NULL) |
| 7572 | kfree(entry->procname); |
| 7573 | } |
| 7574 | |
| 7575 | kfree(*tablep); |
| 7576 | *tablep = NULL; |
| 7577 | } |
| 7578 | |
| 7579 | static void |
| 7580 | set_table_entry(struct ctl_table *entry, |
| 7581 | const char *procname, void *data, int maxlen, |
| 7582 | mode_t mode, proc_handler *proc_handler) |
| 7583 | { |
| 7584 | entry->procname = procname; |
| 7585 | entry->data = data; |
| 7586 | entry->maxlen = maxlen; |
| 7587 | entry->mode = mode; |
| 7588 | entry->proc_handler = proc_handler; |
| 7589 | } |
| 7590 | |
| 7591 | static struct ctl_table * |
| 7592 | sd_alloc_ctl_domain_table(struct sched_domain *sd) |
| 7593 | { |
| 7594 | struct ctl_table *table = sd_alloc_ctl_entry(13); |
| 7595 | |
| 7596 | if (table == NULL) |
| 7597 | return NULL; |
| 7598 | |
| 7599 | set_table_entry(&table[0], "min_interval", &sd->min_interval, |
| 7600 | sizeof(long), 0644, proc_doulongvec_minmax); |
| 7601 | set_table_entry(&table[1], "max_interval", &sd->max_interval, |
| 7602 | sizeof(long), 0644, proc_doulongvec_minmax); |
| 7603 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, |
| 7604 | sizeof(int), 0644, proc_dointvec_minmax); |
| 7605 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, |
| 7606 | sizeof(int), 0644, proc_dointvec_minmax); |
| 7607 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, |
| 7608 | sizeof(int), 0644, proc_dointvec_minmax); |
| 7609 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, |
| 7610 | sizeof(int), 0644, proc_dointvec_minmax); |
| 7611 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, |
| 7612 | sizeof(int), 0644, proc_dointvec_minmax); |
| 7613 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, |
| 7614 | sizeof(int), 0644, proc_dointvec_minmax); |
| 7615 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, |
| 7616 | sizeof(int), 0644, proc_dointvec_minmax); |
| 7617 | set_table_entry(&table[9], "cache_nice_tries", |
| 7618 | &sd->cache_nice_tries, |
| 7619 | sizeof(int), 0644, proc_dointvec_minmax); |
| 7620 | set_table_entry(&table[10], "flags", &sd->flags, |
| 7621 | sizeof(int), 0644, proc_dointvec_minmax); |
| 7622 | set_table_entry(&table[11], "name", sd->name, |
| 7623 | CORENAME_MAX_SIZE, 0444, proc_dostring); |
| 7624 | /* &table[12] is terminator */ |
| 7625 | |
| 7626 | return table; |
| 7627 | } |
| 7628 | |
| 7629 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) |
| 7630 | { |
| 7631 | struct ctl_table *entry, *table; |
| 7632 | struct sched_domain *sd; |
| 7633 | int domain_num = 0, i; |
| 7634 | char buf[32]; |
| 7635 | |
| 7636 | for_each_domain(cpu, sd) |
| 7637 | domain_num++; |
| 7638 | entry = table = sd_alloc_ctl_entry(domain_num + 1); |
| 7639 | if (table == NULL) |
| 7640 | return NULL; |
| 7641 | |
| 7642 | i = 0; |
| 7643 | for_each_domain(cpu, sd) { |
| 7644 | snprintf(buf, 32, "domain%d", i); |
| 7645 | entry->procname = kstrdup(buf, GFP_KERNEL); |
| 7646 | entry->mode = 0555; |
| 7647 | entry->child = sd_alloc_ctl_domain_table(sd); |
| 7648 | entry++; |
| 7649 | i++; |
| 7650 | } |
| 7651 | return table; |
| 7652 | } |
| 7653 | |
| 7654 | static struct ctl_table_header *sd_sysctl_header; |
| 7655 | static void register_sched_domain_sysctl(void) |
| 7656 | { |
| 7657 | int i, cpu_num = num_online_cpus(); |
| 7658 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); |
| 7659 | char buf[32]; |
| 7660 | |
| 7661 | WARN_ON(sd_ctl_dir[0].child); |
| 7662 | sd_ctl_dir[0].child = entry; |
| 7663 | |
| 7664 | if (entry == NULL) |
| 7665 | return; |
| 7666 | |
| 7667 | for_each_online_cpu(i) { |
| 7668 | snprintf(buf, 32, "cpu%d", i); |
| 7669 | entry->procname = kstrdup(buf, GFP_KERNEL); |
| 7670 | entry->mode = 0555; |
| 7671 | entry->child = sd_alloc_ctl_cpu_table(i); |
| 7672 | entry++; |
| 7673 | } |
| 7674 | |
| 7675 | WARN_ON(sd_sysctl_header); |
| 7676 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); |
| 7677 | } |
| 7678 | |
| 7679 | /* may be called multiple times per register */ |
| 7680 | static void unregister_sched_domain_sysctl(void) |
| 7681 | { |
| 7682 | if (sd_sysctl_header) |
| 7683 | unregister_sysctl_table(sd_sysctl_header); |
| 7684 | sd_sysctl_header = NULL; |
| 7685 | if (sd_ctl_dir[0].child) |
| 7686 | sd_free_ctl_entry(&sd_ctl_dir[0].child); |
| 7687 | } |
| 7688 | #else |
| 7689 | static void register_sched_domain_sysctl(void) |
| 7690 | { |
| 7691 | } |
| 7692 | static void unregister_sched_domain_sysctl(void) |
| 7693 | { |
| 7694 | } |
| 7695 | #endif |
| 7696 | |
| 7697 | static void set_rq_online(struct rq *rq) |
| 7698 | { |
| 7699 | if (!rq->online) { |
| 7700 | const struct sched_class *class; |
| 7701 | |
| 7702 | cpumask_set_cpu(rq->cpu, rq->rd->online); |
| 7703 | rq->online = 1; |
| 7704 | |
| 7705 | for_each_class(class) { |
| 7706 | if (class->rq_online) |
| 7707 | class->rq_online(rq); |
| 7708 | } |
| 7709 | } |
| 7710 | } |
| 7711 | |
| 7712 | static void set_rq_offline(struct rq *rq) |
| 7713 | { |
| 7714 | if (rq->online) { |
| 7715 | const struct sched_class *class; |
| 7716 | |
| 7717 | for_each_class(class) { |
| 7718 | if (class->rq_offline) |
| 7719 | class->rq_offline(rq); |
| 7720 | } |
| 7721 | |
| 7722 | cpumask_clear_cpu(rq->cpu, rq->rd->online); |
| 7723 | rq->online = 0; |
| 7724 | } |
| 7725 | } |
| 7726 | |
| 7727 | /* |
| 7728 | * migration_call - callback that gets triggered when a CPU is added. |
| 7729 | * Here we can start up the necessary migration thread for the new CPU. |
| 7730 | */ |
| 7731 | static int __cpuinit |
| 7732 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) |
| 7733 | { |
| 7734 | struct task_struct *p; |
| 7735 | int cpu = (long)hcpu; |
| 7736 | unsigned long flags; |
| 7737 | struct rq *rq; |
| 7738 | |
| 7739 | switch (action) { |
| 7740 | |
| 7741 | case CPU_UP_PREPARE: |
| 7742 | case CPU_UP_PREPARE_FROZEN: |
| 7743 | p = kthread_create(migration_thread, hcpu, "migration/%d", cpu); |
| 7744 | if (IS_ERR(p)) |
| 7745 | return NOTIFY_BAD; |
| 7746 | kthread_bind(p, cpu); |
| 7747 | /* Must be high prio: stop_machine expects to yield to it. */ |
| 7748 | rq = task_rq_lock(p, &flags); |
| 7749 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
| 7750 | task_rq_unlock(rq, &flags); |
| 7751 | get_task_struct(p); |
| 7752 | cpu_rq(cpu)->migration_thread = p; |
| 7753 | rq->calc_load_update = calc_load_update; |
| 7754 | break; |
| 7755 | |
| 7756 | case CPU_ONLINE: |
| 7757 | case CPU_ONLINE_FROZEN: |
| 7758 | /* Strictly unnecessary, as first user will wake it. */ |
| 7759 | wake_up_process(cpu_rq(cpu)->migration_thread); |
| 7760 | |
| 7761 | /* Update our root-domain */ |
| 7762 | rq = cpu_rq(cpu); |
| 7763 | spin_lock_irqsave(&rq->lock, flags); |
| 7764 | if (rq->rd) { |
| 7765 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
| 7766 | |
| 7767 | set_rq_online(rq); |
| 7768 | } |
| 7769 | spin_unlock_irqrestore(&rq->lock, flags); |
| 7770 | break; |
| 7771 | |
| 7772 | #ifdef CONFIG_HOTPLUG_CPU |
| 7773 | case CPU_UP_CANCELED: |
| 7774 | case CPU_UP_CANCELED_FROZEN: |
| 7775 | if (!cpu_rq(cpu)->migration_thread) |
| 7776 | break; |
| 7777 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
| 7778 | kthread_bind(cpu_rq(cpu)->migration_thread, |
| 7779 | cpumask_any(cpu_online_mask)); |
| 7780 | kthread_stop(cpu_rq(cpu)->migration_thread); |
| 7781 | put_task_struct(cpu_rq(cpu)->migration_thread); |
| 7782 | cpu_rq(cpu)->migration_thread = NULL; |
| 7783 | break; |
| 7784 | |
| 7785 | case CPU_DEAD: |
| 7786 | case CPU_DEAD_FROZEN: |
| 7787 | cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */ |
| 7788 | migrate_live_tasks(cpu); |
| 7789 | rq = cpu_rq(cpu); |
| 7790 | kthread_stop(rq->migration_thread); |
| 7791 | put_task_struct(rq->migration_thread); |
| 7792 | rq->migration_thread = NULL; |
| 7793 | /* Idle task back to normal (off runqueue, low prio) */ |
| 7794 | spin_lock_irq(&rq->lock); |
| 7795 | update_rq_clock(rq); |
| 7796 | deactivate_task(rq, rq->idle, 0); |
| 7797 | rq->idle->static_prio = MAX_PRIO; |
| 7798 | __setscheduler(rq, rq->idle, SCHED_NORMAL, 0); |
| 7799 | rq->idle->sched_class = &idle_sched_class; |
| 7800 | migrate_dead_tasks(cpu); |
| 7801 | spin_unlock_irq(&rq->lock); |
| 7802 | cpuset_unlock(); |
| 7803 | migrate_nr_uninterruptible(rq); |
| 7804 | BUG_ON(rq->nr_running != 0); |
| 7805 | calc_global_load_remove(rq); |
| 7806 | /* |
| 7807 | * No need to migrate the tasks: it was best-effort if |
| 7808 | * they didn't take sched_hotcpu_mutex. Just wake up |
| 7809 | * the requestors. |
| 7810 | */ |
| 7811 | spin_lock_irq(&rq->lock); |
| 7812 | while (!list_empty(&rq->migration_queue)) { |
| 7813 | struct migration_req *req; |
| 7814 | |
| 7815 | req = list_entry(rq->migration_queue.next, |
| 7816 | struct migration_req, list); |
| 7817 | list_del_init(&req->list); |
| 7818 | spin_unlock_irq(&rq->lock); |
| 7819 | complete(&req->done); |
| 7820 | spin_lock_irq(&rq->lock); |
| 7821 | } |
| 7822 | spin_unlock_irq(&rq->lock); |
| 7823 | break; |
| 7824 | |
| 7825 | case CPU_DYING: |
| 7826 | case CPU_DYING_FROZEN: |
| 7827 | /* Update our root-domain */ |
| 7828 | rq = cpu_rq(cpu); |
| 7829 | spin_lock_irqsave(&rq->lock, flags); |
| 7830 | if (rq->rd) { |
| 7831 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
| 7832 | set_rq_offline(rq); |
| 7833 | } |
| 7834 | spin_unlock_irqrestore(&rq->lock, flags); |
| 7835 | break; |
| 7836 | #endif |
| 7837 | } |
| 7838 | return NOTIFY_OK; |
| 7839 | } |
| 7840 | |
| 7841 | /* |
| 7842 | * Register at high priority so that task migration (migrate_all_tasks) |
| 7843 | * happens before everything else. This has to be lower priority than |
| 7844 | * the notifier in the perf_counter subsystem, though. |
| 7845 | */ |
| 7846 | static struct notifier_block __cpuinitdata migration_notifier = { |
| 7847 | .notifier_call = migration_call, |
| 7848 | .priority = 10 |
| 7849 | }; |
| 7850 | |
| 7851 | static int __init migration_init(void) |
| 7852 | { |
| 7853 | void *cpu = (void *)(long)smp_processor_id(); |
| 7854 | int err; |
| 7855 | |
| 7856 | /* Start one for the boot CPU: */ |
| 7857 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
| 7858 | BUG_ON(err == NOTIFY_BAD); |
| 7859 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
| 7860 | register_cpu_notifier(&migration_notifier); |
| 7861 | |
| 7862 | return 0; |
| 7863 | } |
| 7864 | early_initcall(migration_init); |
| 7865 | #endif |
| 7866 | |
| 7867 | #ifdef CONFIG_SMP |
| 7868 | |
| 7869 | #ifdef CONFIG_SCHED_DEBUG |
| 7870 | |
| 7871 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, |
| 7872 | struct cpumask *groupmask) |
| 7873 | { |
| 7874 | struct sched_group *group = sd->groups; |
| 7875 | char str[256]; |
| 7876 | |
| 7877 | cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); |
| 7878 | cpumask_clear(groupmask); |
| 7879 | |
| 7880 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); |
| 7881 | |
| 7882 | if (!(sd->flags & SD_LOAD_BALANCE)) { |
| 7883 | printk("does not load-balance\n"); |
| 7884 | if (sd->parent) |
| 7885 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" |
| 7886 | " has parent"); |
| 7887 | return -1; |
| 7888 | } |
| 7889 | |
| 7890 | printk(KERN_CONT "span %s level %s\n", str, sd->name); |
| 7891 | |
| 7892 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
| 7893 | printk(KERN_ERR "ERROR: domain->span does not contain " |
| 7894 | "CPU%d\n", cpu); |
| 7895 | } |
| 7896 | if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { |
| 7897 | printk(KERN_ERR "ERROR: domain->groups does not contain" |
| 7898 | " CPU%d\n", cpu); |
| 7899 | } |
| 7900 | |
| 7901 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); |
| 7902 | do { |
| 7903 | if (!group) { |
| 7904 | printk("\n"); |
| 7905 | printk(KERN_ERR "ERROR: group is NULL\n"); |
| 7906 | break; |
| 7907 | } |
| 7908 | |
| 7909 | if (!group->cpu_power) { |
| 7910 | printk(KERN_CONT "\n"); |
| 7911 | printk(KERN_ERR "ERROR: domain->cpu_power not " |
| 7912 | "set\n"); |
| 7913 | break; |
| 7914 | } |
| 7915 | |
| 7916 | if (!cpumask_weight(sched_group_cpus(group))) { |
| 7917 | printk(KERN_CONT "\n"); |
| 7918 | printk(KERN_ERR "ERROR: empty group\n"); |
| 7919 | break; |
| 7920 | } |
| 7921 | |
| 7922 | if (cpumask_intersects(groupmask, sched_group_cpus(group))) { |
| 7923 | printk(KERN_CONT "\n"); |
| 7924 | printk(KERN_ERR "ERROR: repeated CPUs\n"); |
| 7925 | break; |
| 7926 | } |
| 7927 | |
| 7928 | cpumask_or(groupmask, groupmask, sched_group_cpus(group)); |
| 7929 | |
| 7930 | cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); |
| 7931 | |
| 7932 | printk(KERN_CONT " %s", str); |
| 7933 | if (group->cpu_power != SCHED_LOAD_SCALE) { |
| 7934 | printk(KERN_CONT " (cpu_power = %d)", |
| 7935 | group->cpu_power); |
| 7936 | } |
| 7937 | |
| 7938 | group = group->next; |
| 7939 | } while (group != sd->groups); |
| 7940 | printk(KERN_CONT "\n"); |
| 7941 | |
| 7942 | if (!cpumask_equal(sched_domain_span(sd), groupmask)) |
| 7943 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
| 7944 | |
| 7945 | if (sd->parent && |
| 7946 | !cpumask_subset(groupmask, sched_domain_span(sd->parent))) |
| 7947 | printk(KERN_ERR "ERROR: parent span is not a superset " |
| 7948 | "of domain->span\n"); |
| 7949 | return 0; |
| 7950 | } |
| 7951 | |
| 7952 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
| 7953 | { |
| 7954 | cpumask_var_t groupmask; |
| 7955 | int level = 0; |
| 7956 | |
| 7957 | if (!sd) { |
| 7958 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); |
| 7959 | return; |
| 7960 | } |
| 7961 | |
| 7962 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
| 7963 | |
| 7964 | if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) { |
| 7965 | printk(KERN_DEBUG "Cannot load-balance (out of memory)\n"); |
| 7966 | return; |
| 7967 | } |
| 7968 | |
| 7969 | for (;;) { |
| 7970 | if (sched_domain_debug_one(sd, cpu, level, groupmask)) |
| 7971 | break; |
| 7972 | level++; |
| 7973 | sd = sd->parent; |
| 7974 | if (!sd) |
| 7975 | break; |
| 7976 | } |
| 7977 | free_cpumask_var(groupmask); |
| 7978 | } |
| 7979 | #else /* !CONFIG_SCHED_DEBUG */ |
| 7980 | # define sched_domain_debug(sd, cpu) do { } while (0) |
| 7981 | #endif /* CONFIG_SCHED_DEBUG */ |
| 7982 | |
| 7983 | static int sd_degenerate(struct sched_domain *sd) |
| 7984 | { |
| 7985 | if (cpumask_weight(sched_domain_span(sd)) == 1) |
| 7986 | return 1; |
| 7987 | |
| 7988 | /* Following flags need at least 2 groups */ |
| 7989 | if (sd->flags & (SD_LOAD_BALANCE | |
| 7990 | SD_BALANCE_NEWIDLE | |
| 7991 | SD_BALANCE_FORK | |
| 7992 | SD_BALANCE_EXEC | |
| 7993 | SD_SHARE_CPUPOWER | |
| 7994 | SD_SHARE_PKG_RESOURCES)) { |
| 7995 | if (sd->groups != sd->groups->next) |
| 7996 | return 0; |
| 7997 | } |
| 7998 | |
| 7999 | /* Following flags don't use groups */ |
| 8000 | if (sd->flags & (SD_WAKE_IDLE | |
| 8001 | SD_WAKE_AFFINE | |
| 8002 | SD_WAKE_BALANCE)) |
| 8003 | return 0; |
| 8004 | |
| 8005 | return 1; |
| 8006 | } |
| 8007 | |
| 8008 | static int |
| 8009 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) |
| 8010 | { |
| 8011 | unsigned long cflags = sd->flags, pflags = parent->flags; |
| 8012 | |
| 8013 | if (sd_degenerate(parent)) |
| 8014 | return 1; |
| 8015 | |
| 8016 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) |
| 8017 | return 0; |
| 8018 | |
| 8019 | /* Does parent contain flags not in child? */ |
| 8020 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ |
| 8021 | if (cflags & SD_WAKE_AFFINE) |
| 8022 | pflags &= ~SD_WAKE_BALANCE; |
| 8023 | /* Flags needing groups don't count if only 1 group in parent */ |
| 8024 | if (parent->groups == parent->groups->next) { |
| 8025 | pflags &= ~(SD_LOAD_BALANCE | |
| 8026 | SD_BALANCE_NEWIDLE | |
| 8027 | SD_BALANCE_FORK | |
| 8028 | SD_BALANCE_EXEC | |
| 8029 | SD_SHARE_CPUPOWER | |
| 8030 | SD_SHARE_PKG_RESOURCES); |
| 8031 | if (nr_node_ids == 1) |
| 8032 | pflags &= ~SD_SERIALIZE; |
| 8033 | } |
| 8034 | if (~cflags & pflags) |
| 8035 | return 0; |
| 8036 | |
| 8037 | return 1; |
| 8038 | } |
| 8039 | |
| 8040 | static void free_rootdomain(struct root_domain *rd) |
| 8041 | { |
| 8042 | cpupri_cleanup(&rd->cpupri); |
| 8043 | |
| 8044 | free_cpumask_var(rd->rto_mask); |
| 8045 | free_cpumask_var(rd->online); |
| 8046 | free_cpumask_var(rd->span); |
| 8047 | kfree(rd); |
| 8048 | } |
| 8049 | |
| 8050 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) |
| 8051 | { |
| 8052 | struct root_domain *old_rd = NULL; |
| 8053 | unsigned long flags; |
| 8054 | |
| 8055 | spin_lock_irqsave(&rq->lock, flags); |
| 8056 | |
| 8057 | if (rq->rd) { |
| 8058 | old_rd = rq->rd; |
| 8059 | |
| 8060 | if (cpumask_test_cpu(rq->cpu, old_rd->online)) |
| 8061 | set_rq_offline(rq); |
| 8062 | |
| 8063 | cpumask_clear_cpu(rq->cpu, old_rd->span); |
| 8064 | |
| 8065 | /* |
| 8066 | * If we dont want to free the old_rt yet then |
| 8067 | * set old_rd to NULL to skip the freeing later |
| 8068 | * in this function: |
| 8069 | */ |
| 8070 | if (!atomic_dec_and_test(&old_rd->refcount)) |
| 8071 | old_rd = NULL; |
| 8072 | } |
| 8073 | |
| 8074 | atomic_inc(&rd->refcount); |
| 8075 | rq->rd = rd; |
| 8076 | |
| 8077 | cpumask_set_cpu(rq->cpu, rd->span); |
| 8078 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) |
| 8079 | set_rq_online(rq); |
| 8080 | |
| 8081 | spin_unlock_irqrestore(&rq->lock, flags); |
| 8082 | |
| 8083 | if (old_rd) |
| 8084 | free_rootdomain(old_rd); |
| 8085 | } |
| 8086 | |
| 8087 | static int init_rootdomain(struct root_domain *rd, bool bootmem) |
| 8088 | { |
| 8089 | gfp_t gfp = GFP_KERNEL; |
| 8090 | |
| 8091 | memset(rd, 0, sizeof(*rd)); |
| 8092 | |
| 8093 | if (bootmem) |
| 8094 | gfp = GFP_NOWAIT; |
| 8095 | |
| 8096 | if (!alloc_cpumask_var(&rd->span, gfp)) |
| 8097 | goto out; |
| 8098 | if (!alloc_cpumask_var(&rd->online, gfp)) |
| 8099 | goto free_span; |
| 8100 | if (!alloc_cpumask_var(&rd->rto_mask, gfp)) |
| 8101 | goto free_online; |
| 8102 | |
| 8103 | if (cpupri_init(&rd->cpupri, bootmem) != 0) |
| 8104 | goto free_rto_mask; |
| 8105 | return 0; |
| 8106 | |
| 8107 | free_rto_mask: |
| 8108 | free_cpumask_var(rd->rto_mask); |
| 8109 | free_online: |
| 8110 | free_cpumask_var(rd->online); |
| 8111 | free_span: |
| 8112 | free_cpumask_var(rd->span); |
| 8113 | out: |
| 8114 | return -ENOMEM; |
| 8115 | } |
| 8116 | |
| 8117 | static void init_defrootdomain(void) |
| 8118 | { |
| 8119 | init_rootdomain(&def_root_domain, true); |
| 8120 | |
| 8121 | atomic_set(&def_root_domain.refcount, 1); |
| 8122 | } |
| 8123 | |
| 8124 | static struct root_domain *alloc_rootdomain(void) |
| 8125 | { |
| 8126 | struct root_domain *rd; |
| 8127 | |
| 8128 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); |
| 8129 | if (!rd) |
| 8130 | return NULL; |
| 8131 | |
| 8132 | if (init_rootdomain(rd, false) != 0) { |
| 8133 | kfree(rd); |
| 8134 | return NULL; |
| 8135 | } |
| 8136 | |
| 8137 | return rd; |
| 8138 | } |
| 8139 | |
| 8140 | /* |
| 8141 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
| 8142 | * hold the hotplug lock. |
| 8143 | */ |
| 8144 | static void |
| 8145 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) |
| 8146 | { |
| 8147 | struct rq *rq = cpu_rq(cpu); |
| 8148 | struct sched_domain *tmp; |
| 8149 | |
| 8150 | /* Remove the sched domains which do not contribute to scheduling. */ |
| 8151 | for (tmp = sd; tmp; ) { |
| 8152 | struct sched_domain *parent = tmp->parent; |
| 8153 | if (!parent) |
| 8154 | break; |
| 8155 | |
| 8156 | if (sd_parent_degenerate(tmp, parent)) { |
| 8157 | tmp->parent = parent->parent; |
| 8158 | if (parent->parent) |
| 8159 | parent->parent->child = tmp; |
| 8160 | } else |
| 8161 | tmp = tmp->parent; |
| 8162 | } |
| 8163 | |
| 8164 | if (sd && sd_degenerate(sd)) { |
| 8165 | sd = sd->parent; |
| 8166 | if (sd) |
| 8167 | sd->child = NULL; |
| 8168 | } |
| 8169 | |
| 8170 | sched_domain_debug(sd, cpu); |
| 8171 | |
| 8172 | rq_attach_root(rq, rd); |
| 8173 | rcu_assign_pointer(rq->sd, sd); |
| 8174 | } |
| 8175 | |
| 8176 | /* cpus with isolated domains */ |
| 8177 | static cpumask_var_t cpu_isolated_map; |
| 8178 | |
| 8179 | /* Setup the mask of cpus configured for isolated domains */ |
| 8180 | static int __init isolated_cpu_setup(char *str) |
| 8181 | { |
| 8182 | cpulist_parse(str, cpu_isolated_map); |
| 8183 | return 1; |
| 8184 | } |
| 8185 | |
| 8186 | __setup("isolcpus=", isolated_cpu_setup); |
| 8187 | |
| 8188 | /* |
| 8189 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer |
| 8190 | * to a function which identifies what group(along with sched group) a CPU |
| 8191 | * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids |
| 8192 | * (due to the fact that we keep track of groups covered with a struct cpumask). |
| 8193 | * |
| 8194 | * init_sched_build_groups will build a circular linked list of the groups |
| 8195 | * covered by the given span, and will set each group's ->cpumask correctly, |
| 8196 | * and ->cpu_power to 0. |
| 8197 | */ |
| 8198 | static void |
| 8199 | init_sched_build_groups(const struct cpumask *span, |
| 8200 | const struct cpumask *cpu_map, |
| 8201 | int (*group_fn)(int cpu, const struct cpumask *cpu_map, |
| 8202 | struct sched_group **sg, |
| 8203 | struct cpumask *tmpmask), |
| 8204 | struct cpumask *covered, struct cpumask *tmpmask) |
| 8205 | { |
| 8206 | struct sched_group *first = NULL, *last = NULL; |
| 8207 | int i; |
| 8208 | |
| 8209 | cpumask_clear(covered); |
| 8210 | |
| 8211 | for_each_cpu(i, span) { |
| 8212 | struct sched_group *sg; |
| 8213 | int group = group_fn(i, cpu_map, &sg, tmpmask); |
| 8214 | int j; |
| 8215 | |
| 8216 | if (cpumask_test_cpu(i, covered)) |
| 8217 | continue; |
| 8218 | |
| 8219 | cpumask_clear(sched_group_cpus(sg)); |
| 8220 | sg->cpu_power = 0; |
| 8221 | |
| 8222 | for_each_cpu(j, span) { |
| 8223 | if (group_fn(j, cpu_map, NULL, tmpmask) != group) |
| 8224 | continue; |
| 8225 | |
| 8226 | cpumask_set_cpu(j, covered); |
| 8227 | cpumask_set_cpu(j, sched_group_cpus(sg)); |
| 8228 | } |
| 8229 | if (!first) |
| 8230 | first = sg; |
| 8231 | if (last) |
| 8232 | last->next = sg; |
| 8233 | last = sg; |
| 8234 | } |
| 8235 | last->next = first; |
| 8236 | } |
| 8237 | |
| 8238 | #define SD_NODES_PER_DOMAIN 16 |
| 8239 | |
| 8240 | #ifdef CONFIG_NUMA |
| 8241 | |
| 8242 | /** |
| 8243 | * find_next_best_node - find the next node to include in a sched_domain |
| 8244 | * @node: node whose sched_domain we're building |
| 8245 | * @used_nodes: nodes already in the sched_domain |
| 8246 | * |
| 8247 | * Find the next node to include in a given scheduling domain. Simply |
| 8248 | * finds the closest node not already in the @used_nodes map. |
| 8249 | * |
| 8250 | * Should use nodemask_t. |
| 8251 | */ |
| 8252 | static int find_next_best_node(int node, nodemask_t *used_nodes) |
| 8253 | { |
| 8254 | int i, n, val, min_val, best_node = 0; |
| 8255 | |
| 8256 | min_val = INT_MAX; |
| 8257 | |
| 8258 | for (i = 0; i < nr_node_ids; i++) { |
| 8259 | /* Start at @node */ |
| 8260 | n = (node + i) % nr_node_ids; |
| 8261 | |
| 8262 | if (!nr_cpus_node(n)) |
| 8263 | continue; |
| 8264 | |
| 8265 | /* Skip already used nodes */ |
| 8266 | if (node_isset(n, *used_nodes)) |
| 8267 | continue; |
| 8268 | |
| 8269 | /* Simple min distance search */ |
| 8270 | val = node_distance(node, n); |
| 8271 | |
| 8272 | if (val < min_val) { |
| 8273 | min_val = val; |
| 8274 | best_node = n; |
| 8275 | } |
| 8276 | } |
| 8277 | |
| 8278 | node_set(best_node, *used_nodes); |
| 8279 | return best_node; |
| 8280 | } |
| 8281 | |
| 8282 | /** |
| 8283 | * sched_domain_node_span - get a cpumask for a node's sched_domain |
| 8284 | * @node: node whose cpumask we're constructing |
| 8285 | * @span: resulting cpumask |
| 8286 | * |
| 8287 | * Given a node, construct a good cpumask for its sched_domain to span. It |
| 8288 | * should be one that prevents unnecessary balancing, but also spreads tasks |
| 8289 | * out optimally. |
| 8290 | */ |
| 8291 | static void sched_domain_node_span(int node, struct cpumask *span) |
| 8292 | { |
| 8293 | nodemask_t used_nodes; |
| 8294 | int i; |
| 8295 | |
| 8296 | cpumask_clear(span); |
| 8297 | nodes_clear(used_nodes); |
| 8298 | |
| 8299 | cpumask_or(span, span, cpumask_of_node(node)); |
| 8300 | node_set(node, used_nodes); |
| 8301 | |
| 8302 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { |
| 8303 | int next_node = find_next_best_node(node, &used_nodes); |
| 8304 | |
| 8305 | cpumask_or(span, span, cpumask_of_node(next_node)); |
| 8306 | } |
| 8307 | } |
| 8308 | #endif /* CONFIG_NUMA */ |
| 8309 | |
| 8310 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
| 8311 | |
| 8312 | /* |
| 8313 | * The cpus mask in sched_group and sched_domain hangs off the end. |
| 8314 | * |
| 8315 | * ( See the the comments in include/linux/sched.h:struct sched_group |
| 8316 | * and struct sched_domain. ) |
| 8317 | */ |
| 8318 | struct static_sched_group { |
| 8319 | struct sched_group sg; |
| 8320 | DECLARE_BITMAP(cpus, CONFIG_NR_CPUS); |
| 8321 | }; |
| 8322 | |
| 8323 | struct static_sched_domain { |
| 8324 | struct sched_domain sd; |
| 8325 | DECLARE_BITMAP(span, CONFIG_NR_CPUS); |
| 8326 | }; |
| 8327 | |
| 8328 | struct s_data { |
| 8329 | #ifdef CONFIG_NUMA |
| 8330 | int sd_allnodes; |
| 8331 | cpumask_var_t domainspan; |
| 8332 | cpumask_var_t covered; |
| 8333 | cpumask_var_t notcovered; |
| 8334 | #endif |
| 8335 | cpumask_var_t nodemask; |
| 8336 | cpumask_var_t this_sibling_map; |
| 8337 | cpumask_var_t this_core_map; |
| 8338 | cpumask_var_t send_covered; |
| 8339 | cpumask_var_t tmpmask; |
| 8340 | struct sched_group **sched_group_nodes; |
| 8341 | struct root_domain *rd; |
| 8342 | }; |
| 8343 | |
| 8344 | enum s_alloc { |
| 8345 | sa_sched_groups = 0, |
| 8346 | sa_rootdomain, |
| 8347 | sa_tmpmask, |
| 8348 | sa_send_covered, |
| 8349 | sa_this_core_map, |
| 8350 | sa_this_sibling_map, |
| 8351 | sa_nodemask, |
| 8352 | sa_sched_group_nodes, |
| 8353 | #ifdef CONFIG_NUMA |
| 8354 | sa_notcovered, |
| 8355 | sa_covered, |
| 8356 | sa_domainspan, |
| 8357 | #endif |
| 8358 | sa_none, |
| 8359 | }; |
| 8360 | |
| 8361 | /* |
| 8362 | * SMT sched-domains: |
| 8363 | */ |
| 8364 | #ifdef CONFIG_SCHED_SMT |
| 8365 | static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains); |
| 8366 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus); |
| 8367 | |
| 8368 | static int |
| 8369 | cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map, |
| 8370 | struct sched_group **sg, struct cpumask *unused) |
| 8371 | { |
| 8372 | if (sg) |
| 8373 | *sg = &per_cpu(sched_group_cpus, cpu).sg; |
| 8374 | return cpu; |
| 8375 | } |
| 8376 | #endif /* CONFIG_SCHED_SMT */ |
| 8377 | |
| 8378 | /* |
| 8379 | * multi-core sched-domains: |
| 8380 | */ |
| 8381 | #ifdef CONFIG_SCHED_MC |
| 8382 | static DEFINE_PER_CPU(struct static_sched_domain, core_domains); |
| 8383 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_core); |
| 8384 | #endif /* CONFIG_SCHED_MC */ |
| 8385 | |
| 8386 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) |
| 8387 | static int |
| 8388 | cpu_to_core_group(int cpu, const struct cpumask *cpu_map, |
| 8389 | struct sched_group **sg, struct cpumask *mask) |
| 8390 | { |
| 8391 | int group; |
| 8392 | |
| 8393 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); |
| 8394 | group = cpumask_first(mask); |
| 8395 | if (sg) |
| 8396 | *sg = &per_cpu(sched_group_core, group).sg; |
| 8397 | return group; |
| 8398 | } |
| 8399 | #elif defined(CONFIG_SCHED_MC) |
| 8400 | static int |
| 8401 | cpu_to_core_group(int cpu, const struct cpumask *cpu_map, |
| 8402 | struct sched_group **sg, struct cpumask *unused) |
| 8403 | { |
| 8404 | if (sg) |
| 8405 | *sg = &per_cpu(sched_group_core, cpu).sg; |
| 8406 | return cpu; |
| 8407 | } |
| 8408 | #endif |
| 8409 | |
| 8410 | static DEFINE_PER_CPU(struct static_sched_domain, phys_domains); |
| 8411 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys); |
| 8412 | |
| 8413 | static int |
| 8414 | cpu_to_phys_group(int cpu, const struct cpumask *cpu_map, |
| 8415 | struct sched_group **sg, struct cpumask *mask) |
| 8416 | { |
| 8417 | int group; |
| 8418 | #ifdef CONFIG_SCHED_MC |
| 8419 | cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map); |
| 8420 | group = cpumask_first(mask); |
| 8421 | #elif defined(CONFIG_SCHED_SMT) |
| 8422 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); |
| 8423 | group = cpumask_first(mask); |
| 8424 | #else |
| 8425 | group = cpu; |
| 8426 | #endif |
| 8427 | if (sg) |
| 8428 | *sg = &per_cpu(sched_group_phys, group).sg; |
| 8429 | return group; |
| 8430 | } |
| 8431 | |
| 8432 | #ifdef CONFIG_NUMA |
| 8433 | /* |
| 8434 | * The init_sched_build_groups can't handle what we want to do with node |
| 8435 | * groups, so roll our own. Now each node has its own list of groups which |
| 8436 | * gets dynamically allocated. |
| 8437 | */ |
| 8438 | static DEFINE_PER_CPU(struct static_sched_domain, node_domains); |
| 8439 | static struct sched_group ***sched_group_nodes_bycpu; |
| 8440 | |
| 8441 | static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains); |
| 8442 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes); |
| 8443 | |
| 8444 | static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map, |
| 8445 | struct sched_group **sg, |
| 8446 | struct cpumask *nodemask) |
| 8447 | { |
| 8448 | int group; |
| 8449 | |
| 8450 | cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map); |
| 8451 | group = cpumask_first(nodemask); |
| 8452 | |
| 8453 | if (sg) |
| 8454 | *sg = &per_cpu(sched_group_allnodes, group).sg; |
| 8455 | return group; |
| 8456 | } |
| 8457 | |
| 8458 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
| 8459 | { |
| 8460 | struct sched_group *sg = group_head; |
| 8461 | int j; |
| 8462 | |
| 8463 | if (!sg) |
| 8464 | return; |
| 8465 | do { |
| 8466 | for_each_cpu(j, sched_group_cpus(sg)) { |
| 8467 | struct sched_domain *sd; |
| 8468 | |
| 8469 | sd = &per_cpu(phys_domains, j).sd; |
| 8470 | if (j != group_first_cpu(sd->groups)) { |
| 8471 | /* |
| 8472 | * Only add "power" once for each |
| 8473 | * physical package. |
| 8474 | */ |
| 8475 | continue; |
| 8476 | } |
| 8477 | |
| 8478 | sg->cpu_power += sd->groups->cpu_power; |
| 8479 | } |
| 8480 | sg = sg->next; |
| 8481 | } while (sg != group_head); |
| 8482 | } |
| 8483 | |
| 8484 | static int build_numa_sched_groups(struct s_data *d, |
| 8485 | const struct cpumask *cpu_map, int num) |
| 8486 | { |
| 8487 | struct sched_domain *sd; |
| 8488 | struct sched_group *sg, *prev; |
| 8489 | int n, j; |
| 8490 | |
| 8491 | cpumask_clear(d->covered); |
| 8492 | cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map); |
| 8493 | if (cpumask_empty(d->nodemask)) { |
| 8494 | d->sched_group_nodes[num] = NULL; |
| 8495 | goto out; |
| 8496 | } |
| 8497 | |
| 8498 | sched_domain_node_span(num, d->domainspan); |
| 8499 | cpumask_and(d->domainspan, d->domainspan, cpu_map); |
| 8500 | |
| 8501 | sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(), |
| 8502 | GFP_KERNEL, num); |
| 8503 | if (!sg) { |
| 8504 | printk(KERN_WARNING "Can not alloc domain group for node %d\n", |
| 8505 | num); |
| 8506 | return -ENOMEM; |
| 8507 | } |
| 8508 | d->sched_group_nodes[num] = sg; |
| 8509 | |
| 8510 | for_each_cpu(j, d->nodemask) { |
| 8511 | sd = &per_cpu(node_domains, j).sd; |
| 8512 | sd->groups = sg; |
| 8513 | } |
| 8514 | |
| 8515 | sg->cpu_power = 0; |
| 8516 | cpumask_copy(sched_group_cpus(sg), d->nodemask); |
| 8517 | sg->next = sg; |
| 8518 | cpumask_or(d->covered, d->covered, d->nodemask); |
| 8519 | |
| 8520 | prev = sg; |
| 8521 | for (j = 0; j < nr_node_ids; j++) { |
| 8522 | n = (num + j) % nr_node_ids; |
| 8523 | cpumask_complement(d->notcovered, d->covered); |
| 8524 | cpumask_and(d->tmpmask, d->notcovered, cpu_map); |
| 8525 | cpumask_and(d->tmpmask, d->tmpmask, d->domainspan); |
| 8526 | if (cpumask_empty(d->tmpmask)) |
| 8527 | break; |
| 8528 | cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n)); |
| 8529 | if (cpumask_empty(d->tmpmask)) |
| 8530 | continue; |
| 8531 | sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(), |
| 8532 | GFP_KERNEL, num); |
| 8533 | if (!sg) { |
| 8534 | printk(KERN_WARNING |
| 8535 | "Can not alloc domain group for node %d\n", j); |
| 8536 | return -ENOMEM; |
| 8537 | } |
| 8538 | sg->cpu_power = 0; |
| 8539 | cpumask_copy(sched_group_cpus(sg), d->tmpmask); |
| 8540 | sg->next = prev->next; |
| 8541 | cpumask_or(d->covered, d->covered, d->tmpmask); |
| 8542 | prev->next = sg; |
| 8543 | prev = sg; |
| 8544 | } |
| 8545 | out: |
| 8546 | return 0; |
| 8547 | } |
| 8548 | #endif /* CONFIG_NUMA */ |
| 8549 | |
| 8550 | #ifdef CONFIG_NUMA |
| 8551 | /* Free memory allocated for various sched_group structures */ |
| 8552 | static void free_sched_groups(const struct cpumask *cpu_map, |
| 8553 | struct cpumask *nodemask) |
| 8554 | { |
| 8555 | int cpu, i; |
| 8556 | |
| 8557 | for_each_cpu(cpu, cpu_map) { |
| 8558 | struct sched_group **sched_group_nodes |
| 8559 | = sched_group_nodes_bycpu[cpu]; |
| 8560 | |
| 8561 | if (!sched_group_nodes) |
| 8562 | continue; |
| 8563 | |
| 8564 | for (i = 0; i < nr_node_ids; i++) { |
| 8565 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; |
| 8566 | |
| 8567 | cpumask_and(nodemask, cpumask_of_node(i), cpu_map); |
| 8568 | if (cpumask_empty(nodemask)) |
| 8569 | continue; |
| 8570 | |
| 8571 | if (sg == NULL) |
| 8572 | continue; |
| 8573 | sg = sg->next; |
| 8574 | next_sg: |
| 8575 | oldsg = sg; |
| 8576 | sg = sg->next; |
| 8577 | kfree(oldsg); |
| 8578 | if (oldsg != sched_group_nodes[i]) |
| 8579 | goto next_sg; |
| 8580 | } |
| 8581 | kfree(sched_group_nodes); |
| 8582 | sched_group_nodes_bycpu[cpu] = NULL; |
| 8583 | } |
| 8584 | } |
| 8585 | #else /* !CONFIG_NUMA */ |
| 8586 | static void free_sched_groups(const struct cpumask *cpu_map, |
| 8587 | struct cpumask *nodemask) |
| 8588 | { |
| 8589 | } |
| 8590 | #endif /* CONFIG_NUMA */ |
| 8591 | |
| 8592 | /* |
| 8593 | * Initialize sched groups cpu_power. |
| 8594 | * |
| 8595 | * cpu_power indicates the capacity of sched group, which is used while |
| 8596 | * distributing the load between different sched groups in a sched domain. |
| 8597 | * Typically cpu_power for all the groups in a sched domain will be same unless |
| 8598 | * there are asymmetries in the topology. If there are asymmetries, group |
| 8599 | * having more cpu_power will pickup more load compared to the group having |
| 8600 | * less cpu_power. |
| 8601 | */ |
| 8602 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) |
| 8603 | { |
| 8604 | struct sched_domain *child; |
| 8605 | struct sched_group *group; |
| 8606 | long power; |
| 8607 | int weight; |
| 8608 | |
| 8609 | WARN_ON(!sd || !sd->groups); |
| 8610 | |
| 8611 | if (cpu != group_first_cpu(sd->groups)) |
| 8612 | return; |
| 8613 | |
| 8614 | child = sd->child; |
| 8615 | |
| 8616 | sd->groups->cpu_power = 0; |
| 8617 | |
| 8618 | if (!child) { |
| 8619 | power = SCHED_LOAD_SCALE; |
| 8620 | weight = cpumask_weight(sched_domain_span(sd)); |
| 8621 | /* |
| 8622 | * SMT siblings share the power of a single core. |
| 8623 | * Usually multiple threads get a better yield out of |
| 8624 | * that one core than a single thread would have, |
| 8625 | * reflect that in sd->smt_gain. |
| 8626 | */ |
| 8627 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
| 8628 | power *= sd->smt_gain; |
| 8629 | power /= weight; |
| 8630 | power >>= SCHED_LOAD_SHIFT; |
| 8631 | } |
| 8632 | sd->groups->cpu_power += power; |
| 8633 | return; |
| 8634 | } |
| 8635 | |
| 8636 | /* |
| 8637 | * Add cpu_power of each child group to this groups cpu_power. |
| 8638 | */ |
| 8639 | group = child->groups; |
| 8640 | do { |
| 8641 | sd->groups->cpu_power += group->cpu_power; |
| 8642 | group = group->next; |
| 8643 | } while (group != child->groups); |
| 8644 | } |
| 8645 | |
| 8646 | /* |
| 8647 | * Initializers for schedule domains |
| 8648 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() |
| 8649 | */ |
| 8650 | |
| 8651 | #ifdef CONFIG_SCHED_DEBUG |
| 8652 | # define SD_INIT_NAME(sd, type) sd->name = #type |
| 8653 | #else |
| 8654 | # define SD_INIT_NAME(sd, type) do { } while (0) |
| 8655 | #endif |
| 8656 | |
| 8657 | #define SD_INIT(sd, type) sd_init_##type(sd) |
| 8658 | |
| 8659 | #define SD_INIT_FUNC(type) \ |
| 8660 | static noinline void sd_init_##type(struct sched_domain *sd) \ |
| 8661 | { \ |
| 8662 | memset(sd, 0, sizeof(*sd)); \ |
| 8663 | *sd = SD_##type##_INIT; \ |
| 8664 | sd->level = SD_LV_##type; \ |
| 8665 | SD_INIT_NAME(sd, type); \ |
| 8666 | } |
| 8667 | |
| 8668 | SD_INIT_FUNC(CPU) |
| 8669 | #ifdef CONFIG_NUMA |
| 8670 | SD_INIT_FUNC(ALLNODES) |
| 8671 | SD_INIT_FUNC(NODE) |
| 8672 | #endif |
| 8673 | #ifdef CONFIG_SCHED_SMT |
| 8674 | SD_INIT_FUNC(SIBLING) |
| 8675 | #endif |
| 8676 | #ifdef CONFIG_SCHED_MC |
| 8677 | SD_INIT_FUNC(MC) |
| 8678 | #endif |
| 8679 | |
| 8680 | static int default_relax_domain_level = -1; |
| 8681 | |
| 8682 | static int __init setup_relax_domain_level(char *str) |
| 8683 | { |
| 8684 | unsigned long val; |
| 8685 | |
| 8686 | val = simple_strtoul(str, NULL, 0); |
| 8687 | if (val < SD_LV_MAX) |
| 8688 | default_relax_domain_level = val; |
| 8689 | |
| 8690 | return 1; |
| 8691 | } |
| 8692 | __setup("relax_domain_level=", setup_relax_domain_level); |
| 8693 | |
| 8694 | static void set_domain_attribute(struct sched_domain *sd, |
| 8695 | struct sched_domain_attr *attr) |
| 8696 | { |
| 8697 | int request; |
| 8698 | |
| 8699 | if (!attr || attr->relax_domain_level < 0) { |
| 8700 | if (default_relax_domain_level < 0) |
| 8701 | return; |
| 8702 | else |
| 8703 | request = default_relax_domain_level; |
| 8704 | } else |
| 8705 | request = attr->relax_domain_level; |
| 8706 | if (request < sd->level) { |
| 8707 | /* turn off idle balance on this domain */ |
| 8708 | sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE); |
| 8709 | } else { |
| 8710 | /* turn on idle balance on this domain */ |
| 8711 | sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE); |
| 8712 | } |
| 8713 | } |
| 8714 | |
| 8715 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what, |
| 8716 | const struct cpumask *cpu_map) |
| 8717 | { |
| 8718 | switch (what) { |
| 8719 | case sa_sched_groups: |
| 8720 | free_sched_groups(cpu_map, d->tmpmask); /* fall through */ |
| 8721 | d->sched_group_nodes = NULL; |
| 8722 | case sa_rootdomain: |
| 8723 | free_rootdomain(d->rd); /* fall through */ |
| 8724 | case sa_tmpmask: |
| 8725 | free_cpumask_var(d->tmpmask); /* fall through */ |
| 8726 | case sa_send_covered: |
| 8727 | free_cpumask_var(d->send_covered); /* fall through */ |
| 8728 | case sa_this_core_map: |
| 8729 | free_cpumask_var(d->this_core_map); /* fall through */ |
| 8730 | case sa_this_sibling_map: |
| 8731 | free_cpumask_var(d->this_sibling_map); /* fall through */ |
| 8732 | case sa_nodemask: |
| 8733 | free_cpumask_var(d->nodemask); /* fall through */ |
| 8734 | case sa_sched_group_nodes: |
| 8735 | #ifdef CONFIG_NUMA |
| 8736 | kfree(d->sched_group_nodes); /* fall through */ |
| 8737 | case sa_notcovered: |
| 8738 | free_cpumask_var(d->notcovered); /* fall through */ |
| 8739 | case sa_covered: |
| 8740 | free_cpumask_var(d->covered); /* fall through */ |
| 8741 | case sa_domainspan: |
| 8742 | free_cpumask_var(d->domainspan); /* fall through */ |
| 8743 | #endif |
| 8744 | case sa_none: |
| 8745 | break; |
| 8746 | } |
| 8747 | } |
| 8748 | |
| 8749 | static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, |
| 8750 | const struct cpumask *cpu_map) |
| 8751 | { |
| 8752 | #ifdef CONFIG_NUMA |
| 8753 | if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL)) |
| 8754 | return sa_none; |
| 8755 | if (!alloc_cpumask_var(&d->covered, GFP_KERNEL)) |
| 8756 | return sa_domainspan; |
| 8757 | if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL)) |
| 8758 | return sa_covered; |
| 8759 | /* Allocate the per-node list of sched groups */ |
| 8760 | d->sched_group_nodes = kcalloc(nr_node_ids, |
| 8761 | sizeof(struct sched_group *), GFP_KERNEL); |
| 8762 | if (!d->sched_group_nodes) { |
| 8763 | printk(KERN_WARNING "Can not alloc sched group node list\n"); |
| 8764 | return sa_notcovered; |
| 8765 | } |
| 8766 | sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes; |
| 8767 | #endif |
| 8768 | if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL)) |
| 8769 | return sa_sched_group_nodes; |
| 8770 | if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL)) |
| 8771 | return sa_nodemask; |
| 8772 | if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL)) |
| 8773 | return sa_this_sibling_map; |
| 8774 | if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL)) |
| 8775 | return sa_this_core_map; |
| 8776 | if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL)) |
| 8777 | return sa_send_covered; |
| 8778 | d->rd = alloc_rootdomain(); |
| 8779 | if (!d->rd) { |
| 8780 | printk(KERN_WARNING "Cannot alloc root domain\n"); |
| 8781 | return sa_tmpmask; |
| 8782 | } |
| 8783 | return sa_rootdomain; |
| 8784 | } |
| 8785 | |
| 8786 | static struct sched_domain *__build_numa_sched_domains(struct s_data *d, |
| 8787 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i) |
| 8788 | { |
| 8789 | struct sched_domain *sd = NULL; |
| 8790 | #ifdef CONFIG_NUMA |
| 8791 | struct sched_domain *parent; |
| 8792 | |
| 8793 | d->sd_allnodes = 0; |
| 8794 | if (cpumask_weight(cpu_map) > |
| 8795 | SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) { |
| 8796 | sd = &per_cpu(allnodes_domains, i).sd; |
| 8797 | SD_INIT(sd, ALLNODES); |
| 8798 | set_domain_attribute(sd, attr); |
| 8799 | cpumask_copy(sched_domain_span(sd), cpu_map); |
| 8800 | cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask); |
| 8801 | d->sd_allnodes = 1; |
| 8802 | } |
| 8803 | parent = sd; |
| 8804 | |
| 8805 | sd = &per_cpu(node_domains, i).sd; |
| 8806 | SD_INIT(sd, NODE); |
| 8807 | set_domain_attribute(sd, attr); |
| 8808 | sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd)); |
| 8809 | sd->parent = parent; |
| 8810 | if (parent) |
| 8811 | parent->child = sd; |
| 8812 | cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map); |
| 8813 | #endif |
| 8814 | return sd; |
| 8815 | } |
| 8816 | |
| 8817 | static struct sched_domain *__build_cpu_sched_domain(struct s_data *d, |
| 8818 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
| 8819 | struct sched_domain *parent, int i) |
| 8820 | { |
| 8821 | struct sched_domain *sd; |
| 8822 | sd = &per_cpu(phys_domains, i).sd; |
| 8823 | SD_INIT(sd, CPU); |
| 8824 | set_domain_attribute(sd, attr); |
| 8825 | cpumask_copy(sched_domain_span(sd), d->nodemask); |
| 8826 | sd->parent = parent; |
| 8827 | if (parent) |
| 8828 | parent->child = sd; |
| 8829 | cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask); |
| 8830 | return sd; |
| 8831 | } |
| 8832 | |
| 8833 | static struct sched_domain *__build_mc_sched_domain(struct s_data *d, |
| 8834 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
| 8835 | struct sched_domain *parent, int i) |
| 8836 | { |
| 8837 | struct sched_domain *sd = parent; |
| 8838 | #ifdef CONFIG_SCHED_MC |
| 8839 | sd = &per_cpu(core_domains, i).sd; |
| 8840 | SD_INIT(sd, MC); |
| 8841 | set_domain_attribute(sd, attr); |
| 8842 | cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i)); |
| 8843 | sd->parent = parent; |
| 8844 | parent->child = sd; |
| 8845 | cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask); |
| 8846 | #endif |
| 8847 | return sd; |
| 8848 | } |
| 8849 | |
| 8850 | static struct sched_domain *__build_smt_sched_domain(struct s_data *d, |
| 8851 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
| 8852 | struct sched_domain *parent, int i) |
| 8853 | { |
| 8854 | struct sched_domain *sd = parent; |
| 8855 | #ifdef CONFIG_SCHED_SMT |
| 8856 | sd = &per_cpu(cpu_domains, i).sd; |
| 8857 | SD_INIT(sd, SIBLING); |
| 8858 | set_domain_attribute(sd, attr); |
| 8859 | cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i)); |
| 8860 | sd->parent = parent; |
| 8861 | parent->child = sd; |
| 8862 | cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask); |
| 8863 | #endif |
| 8864 | return sd; |
| 8865 | } |
| 8866 | |
| 8867 | static void build_sched_groups(struct s_data *d, enum sched_domain_level l, |
| 8868 | const struct cpumask *cpu_map, int cpu) |
| 8869 | { |
| 8870 | switch (l) { |
| 8871 | #ifdef CONFIG_SCHED_SMT |
| 8872 | case SD_LV_SIBLING: /* set up CPU (sibling) groups */ |
| 8873 | cpumask_and(d->this_sibling_map, cpu_map, |
| 8874 | topology_thread_cpumask(cpu)); |
| 8875 | if (cpu == cpumask_first(d->this_sibling_map)) |
| 8876 | init_sched_build_groups(d->this_sibling_map, cpu_map, |
| 8877 | &cpu_to_cpu_group, |
| 8878 | d->send_covered, d->tmpmask); |
| 8879 | break; |
| 8880 | #endif |
| 8881 | #ifdef CONFIG_SCHED_MC |
| 8882 | case SD_LV_MC: /* set up multi-core groups */ |
| 8883 | cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu)); |
| 8884 | if (cpu == cpumask_first(d->this_core_map)) |
| 8885 | init_sched_build_groups(d->this_core_map, cpu_map, |
| 8886 | &cpu_to_core_group, |
| 8887 | d->send_covered, d->tmpmask); |
| 8888 | break; |
| 8889 | #endif |
| 8890 | case SD_LV_CPU: /* set up physical groups */ |
| 8891 | cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map); |
| 8892 | if (!cpumask_empty(d->nodemask)) |
| 8893 | init_sched_build_groups(d->nodemask, cpu_map, |
| 8894 | &cpu_to_phys_group, |
| 8895 | d->send_covered, d->tmpmask); |
| 8896 | break; |
| 8897 | #ifdef CONFIG_NUMA |
| 8898 | case SD_LV_ALLNODES: |
| 8899 | init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group, |
| 8900 | d->send_covered, d->tmpmask); |
| 8901 | break; |
| 8902 | #endif |
| 8903 | default: |
| 8904 | break; |
| 8905 | } |
| 8906 | } |
| 8907 | |
| 8908 | /* |
| 8909 | * Build sched domains for a given set of cpus and attach the sched domains |
| 8910 | * to the individual cpus |
| 8911 | */ |
| 8912 | static int __build_sched_domains(const struct cpumask *cpu_map, |
| 8913 | struct sched_domain_attr *attr) |
| 8914 | { |
| 8915 | enum s_alloc alloc_state = sa_none; |
| 8916 | struct s_data d; |
| 8917 | struct sched_domain *sd; |
| 8918 | int i; |
| 8919 | #ifdef CONFIG_NUMA |
| 8920 | d.sd_allnodes = 0; |
| 8921 | #endif |
| 8922 | |
| 8923 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map); |
| 8924 | if (alloc_state != sa_rootdomain) |
| 8925 | goto error; |
| 8926 | alloc_state = sa_sched_groups; |
| 8927 | |
| 8928 | /* |
| 8929 | * Set up domains for cpus specified by the cpu_map. |
| 8930 | */ |
| 8931 | for_each_cpu(i, cpu_map) { |
| 8932 | cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)), |
| 8933 | cpu_map); |
| 8934 | |
| 8935 | sd = __build_numa_sched_domains(&d, cpu_map, attr, i); |
| 8936 | sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i); |
| 8937 | sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i); |
| 8938 | sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i); |
| 8939 | } |
| 8940 | |
| 8941 | for_each_cpu(i, cpu_map) { |
| 8942 | build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i); |
| 8943 | build_sched_groups(&d, SD_LV_MC, cpu_map, i); |
| 8944 | } |
| 8945 | |
| 8946 | /* Set up physical groups */ |
| 8947 | for (i = 0; i < nr_node_ids; i++) |
| 8948 | build_sched_groups(&d, SD_LV_CPU, cpu_map, i); |
| 8949 | |
| 8950 | #ifdef CONFIG_NUMA |
| 8951 | /* Set up node groups */ |
| 8952 | if (d.sd_allnodes) |
| 8953 | build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0); |
| 8954 | |
| 8955 | for (i = 0; i < nr_node_ids; i++) |
| 8956 | if (build_numa_sched_groups(&d, cpu_map, i)) |
| 8957 | goto error; |
| 8958 | #endif |
| 8959 | |
| 8960 | /* Calculate CPU power for physical packages and nodes */ |
| 8961 | #ifdef CONFIG_SCHED_SMT |
| 8962 | for_each_cpu(i, cpu_map) { |
| 8963 | sd = &per_cpu(cpu_domains, i).sd; |
| 8964 | init_sched_groups_power(i, sd); |
| 8965 | } |
| 8966 | #endif |
| 8967 | #ifdef CONFIG_SCHED_MC |
| 8968 | for_each_cpu(i, cpu_map) { |
| 8969 | sd = &per_cpu(core_domains, i).sd; |
| 8970 | init_sched_groups_power(i, sd); |
| 8971 | } |
| 8972 | #endif |
| 8973 | |
| 8974 | for_each_cpu(i, cpu_map) { |
| 8975 | sd = &per_cpu(phys_domains, i).sd; |
| 8976 | init_sched_groups_power(i, sd); |
| 8977 | } |
| 8978 | |
| 8979 | #ifdef CONFIG_NUMA |
| 8980 | for (i = 0; i < nr_node_ids; i++) |
| 8981 | init_numa_sched_groups_power(d.sched_group_nodes[i]); |
| 8982 | |
| 8983 | if (d.sd_allnodes) { |
| 8984 | struct sched_group *sg; |
| 8985 | |
| 8986 | cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg, |
| 8987 | d.tmpmask); |
| 8988 | init_numa_sched_groups_power(sg); |
| 8989 | } |
| 8990 | #endif |
| 8991 | |
| 8992 | /* Attach the domains */ |
| 8993 | for_each_cpu(i, cpu_map) { |
| 8994 | #ifdef CONFIG_SCHED_SMT |
| 8995 | sd = &per_cpu(cpu_domains, i).sd; |
| 8996 | #elif defined(CONFIG_SCHED_MC) |
| 8997 | sd = &per_cpu(core_domains, i).sd; |
| 8998 | #else |
| 8999 | sd = &per_cpu(phys_domains, i).sd; |
| 9000 | #endif |
| 9001 | cpu_attach_domain(sd, d.rd, i); |
| 9002 | } |
| 9003 | |
| 9004 | d.sched_group_nodes = NULL; /* don't free this we still need it */ |
| 9005 | __free_domain_allocs(&d, sa_tmpmask, cpu_map); |
| 9006 | return 0; |
| 9007 | |
| 9008 | error: |
| 9009 | __free_domain_allocs(&d, alloc_state, cpu_map); |
| 9010 | return -ENOMEM; |
| 9011 | } |
| 9012 | |
| 9013 | static int build_sched_domains(const struct cpumask *cpu_map) |
| 9014 | { |
| 9015 | return __build_sched_domains(cpu_map, NULL); |
| 9016 | } |
| 9017 | |
| 9018 | static struct cpumask *doms_cur; /* current sched domains */ |
| 9019 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ |
| 9020 | static struct sched_domain_attr *dattr_cur; |
| 9021 | /* attribues of custom domains in 'doms_cur' */ |
| 9022 | |
| 9023 | /* |
| 9024 | * Special case: If a kmalloc of a doms_cur partition (array of |
| 9025 | * cpumask) fails, then fallback to a single sched domain, |
| 9026 | * as determined by the single cpumask fallback_doms. |
| 9027 | */ |
| 9028 | static cpumask_var_t fallback_doms; |
| 9029 | |
| 9030 | /* |
| 9031 | * arch_update_cpu_topology lets virtualized architectures update the |
| 9032 | * cpu core maps. It is supposed to return 1 if the topology changed |
| 9033 | * or 0 if it stayed the same. |
| 9034 | */ |
| 9035 | int __attribute__((weak)) arch_update_cpu_topology(void) |
| 9036 | { |
| 9037 | return 0; |
| 9038 | } |
| 9039 | |
| 9040 | /* |
| 9041 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
| 9042 | * For now this just excludes isolated cpus, but could be used to |
| 9043 | * exclude other special cases in the future. |
| 9044 | */ |
| 9045 | static int arch_init_sched_domains(const struct cpumask *cpu_map) |
| 9046 | { |
| 9047 | int err; |
| 9048 | |
| 9049 | arch_update_cpu_topology(); |
| 9050 | ndoms_cur = 1; |
| 9051 | doms_cur = kmalloc(cpumask_size(), GFP_KERNEL); |
| 9052 | if (!doms_cur) |
| 9053 | doms_cur = fallback_doms; |
| 9054 | cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map); |
| 9055 | dattr_cur = NULL; |
| 9056 | err = build_sched_domains(doms_cur); |
| 9057 | register_sched_domain_sysctl(); |
| 9058 | |
| 9059 | return err; |
| 9060 | } |
| 9061 | |
| 9062 | static void arch_destroy_sched_domains(const struct cpumask *cpu_map, |
| 9063 | struct cpumask *tmpmask) |
| 9064 | { |
| 9065 | free_sched_groups(cpu_map, tmpmask); |
| 9066 | } |
| 9067 | |
| 9068 | /* |
| 9069 | * Detach sched domains from a group of cpus specified in cpu_map |
| 9070 | * These cpus will now be attached to the NULL domain |
| 9071 | */ |
| 9072 | static void detach_destroy_domains(const struct cpumask *cpu_map) |
| 9073 | { |
| 9074 | /* Save because hotplug lock held. */ |
| 9075 | static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS); |
| 9076 | int i; |
| 9077 | |
| 9078 | for_each_cpu(i, cpu_map) |
| 9079 | cpu_attach_domain(NULL, &def_root_domain, i); |
| 9080 | synchronize_sched(); |
| 9081 | arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask)); |
| 9082 | } |
| 9083 | |
| 9084 | /* handle null as "default" */ |
| 9085 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, |
| 9086 | struct sched_domain_attr *new, int idx_new) |
| 9087 | { |
| 9088 | struct sched_domain_attr tmp; |
| 9089 | |
| 9090 | /* fast path */ |
| 9091 | if (!new && !cur) |
| 9092 | return 1; |
| 9093 | |
| 9094 | tmp = SD_ATTR_INIT; |
| 9095 | return !memcmp(cur ? (cur + idx_cur) : &tmp, |
| 9096 | new ? (new + idx_new) : &tmp, |
| 9097 | sizeof(struct sched_domain_attr)); |
| 9098 | } |
| 9099 | |
| 9100 | /* |
| 9101 | * Partition sched domains as specified by the 'ndoms_new' |
| 9102 | * cpumasks in the array doms_new[] of cpumasks. This compares |
| 9103 | * doms_new[] to the current sched domain partitioning, doms_cur[]. |
| 9104 | * It destroys each deleted domain and builds each new domain. |
| 9105 | * |
| 9106 | * 'doms_new' is an array of cpumask's of length 'ndoms_new'. |
| 9107 | * The masks don't intersect (don't overlap.) We should setup one |
| 9108 | * sched domain for each mask. CPUs not in any of the cpumasks will |
| 9109 | * not be load balanced. If the same cpumask appears both in the |
| 9110 | * current 'doms_cur' domains and in the new 'doms_new', we can leave |
| 9111 | * it as it is. |
| 9112 | * |
| 9113 | * The passed in 'doms_new' should be kmalloc'd. This routine takes |
| 9114 | * ownership of it and will kfree it when done with it. If the caller |
| 9115 | * failed the kmalloc call, then it can pass in doms_new == NULL && |
| 9116 | * ndoms_new == 1, and partition_sched_domains() will fallback to |
| 9117 | * the single partition 'fallback_doms', it also forces the domains |
| 9118 | * to be rebuilt. |
| 9119 | * |
| 9120 | * If doms_new == NULL it will be replaced with cpu_online_mask. |
| 9121 | * ndoms_new == 0 is a special case for destroying existing domains, |
| 9122 | * and it will not create the default domain. |
| 9123 | * |
| 9124 | * Call with hotplug lock held |
| 9125 | */ |
| 9126 | /* FIXME: Change to struct cpumask *doms_new[] */ |
| 9127 | void partition_sched_domains(int ndoms_new, struct cpumask *doms_new, |
| 9128 | struct sched_domain_attr *dattr_new) |
| 9129 | { |
| 9130 | int i, j, n; |
| 9131 | int new_topology; |
| 9132 | |
| 9133 | mutex_lock(&sched_domains_mutex); |
| 9134 | |
| 9135 | /* always unregister in case we don't destroy any domains */ |
| 9136 | unregister_sched_domain_sysctl(); |
| 9137 | |
| 9138 | /* Let architecture update cpu core mappings. */ |
| 9139 | new_topology = arch_update_cpu_topology(); |
| 9140 | |
| 9141 | n = doms_new ? ndoms_new : 0; |
| 9142 | |
| 9143 | /* Destroy deleted domains */ |
| 9144 | for (i = 0; i < ndoms_cur; i++) { |
| 9145 | for (j = 0; j < n && !new_topology; j++) { |
| 9146 | if (cpumask_equal(&doms_cur[i], &doms_new[j]) |
| 9147 | && dattrs_equal(dattr_cur, i, dattr_new, j)) |
| 9148 | goto match1; |
| 9149 | } |
| 9150 | /* no match - a current sched domain not in new doms_new[] */ |
| 9151 | detach_destroy_domains(doms_cur + i); |
| 9152 | match1: |
| 9153 | ; |
| 9154 | } |
| 9155 | |
| 9156 | if (doms_new == NULL) { |
| 9157 | ndoms_cur = 0; |
| 9158 | doms_new = fallback_doms; |
| 9159 | cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map); |
| 9160 | WARN_ON_ONCE(dattr_new); |
| 9161 | } |
| 9162 | |
| 9163 | /* Build new domains */ |
| 9164 | for (i = 0; i < ndoms_new; i++) { |
| 9165 | for (j = 0; j < ndoms_cur && !new_topology; j++) { |
| 9166 | if (cpumask_equal(&doms_new[i], &doms_cur[j]) |
| 9167 | && dattrs_equal(dattr_new, i, dattr_cur, j)) |
| 9168 | goto match2; |
| 9169 | } |
| 9170 | /* no match - add a new doms_new */ |
| 9171 | __build_sched_domains(doms_new + i, |
| 9172 | dattr_new ? dattr_new + i : NULL); |
| 9173 | match2: |
| 9174 | ; |
| 9175 | } |
| 9176 | |
| 9177 | /* Remember the new sched domains */ |
| 9178 | if (doms_cur != fallback_doms) |
| 9179 | kfree(doms_cur); |
| 9180 | kfree(dattr_cur); /* kfree(NULL) is safe */ |
| 9181 | doms_cur = doms_new; |
| 9182 | dattr_cur = dattr_new; |
| 9183 | ndoms_cur = ndoms_new; |
| 9184 | |
| 9185 | register_sched_domain_sysctl(); |
| 9186 | |
| 9187 | mutex_unlock(&sched_domains_mutex); |
| 9188 | } |
| 9189 | |
| 9190 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
| 9191 | static void arch_reinit_sched_domains(void) |
| 9192 | { |
| 9193 | get_online_cpus(); |
| 9194 | |
| 9195 | /* Destroy domains first to force the rebuild */ |
| 9196 | partition_sched_domains(0, NULL, NULL); |
| 9197 | |
| 9198 | rebuild_sched_domains(); |
| 9199 | put_online_cpus(); |
| 9200 | } |
| 9201 | |
| 9202 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) |
| 9203 | { |
| 9204 | unsigned int level = 0; |
| 9205 | |
| 9206 | if (sscanf(buf, "%u", &level) != 1) |
| 9207 | return -EINVAL; |
| 9208 | |
| 9209 | /* |
| 9210 | * level is always be positive so don't check for |
| 9211 | * level < POWERSAVINGS_BALANCE_NONE which is 0 |
| 9212 | * What happens on 0 or 1 byte write, |
| 9213 | * need to check for count as well? |
| 9214 | */ |
| 9215 | |
| 9216 | if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS) |
| 9217 | return -EINVAL; |
| 9218 | |
| 9219 | if (smt) |
| 9220 | sched_smt_power_savings = level; |
| 9221 | else |
| 9222 | sched_mc_power_savings = level; |
| 9223 | |
| 9224 | arch_reinit_sched_domains(); |
| 9225 | |
| 9226 | return count; |
| 9227 | } |
| 9228 | |
| 9229 | #ifdef CONFIG_SCHED_MC |
| 9230 | static ssize_t sched_mc_power_savings_show(struct sysdev_class *class, |
| 9231 | char *page) |
| 9232 | { |
| 9233 | return sprintf(page, "%u\n", sched_mc_power_savings); |
| 9234 | } |
| 9235 | static ssize_t sched_mc_power_savings_store(struct sysdev_class *class, |
| 9236 | const char *buf, size_t count) |
| 9237 | { |
| 9238 | return sched_power_savings_store(buf, count, 0); |
| 9239 | } |
| 9240 | static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644, |
| 9241 | sched_mc_power_savings_show, |
| 9242 | sched_mc_power_savings_store); |
| 9243 | #endif |
| 9244 | |
| 9245 | #ifdef CONFIG_SCHED_SMT |
| 9246 | static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev, |
| 9247 | char *page) |
| 9248 | { |
| 9249 | return sprintf(page, "%u\n", sched_smt_power_savings); |
| 9250 | } |
| 9251 | static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev, |
| 9252 | const char *buf, size_t count) |
| 9253 | { |
| 9254 | return sched_power_savings_store(buf, count, 1); |
| 9255 | } |
| 9256 | static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644, |
| 9257 | sched_smt_power_savings_show, |
| 9258 | sched_smt_power_savings_store); |
| 9259 | #endif |
| 9260 | |
| 9261 | int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) |
| 9262 | { |
| 9263 | int err = 0; |
| 9264 | |
| 9265 | #ifdef CONFIG_SCHED_SMT |
| 9266 | if (smt_capable()) |
| 9267 | err = sysfs_create_file(&cls->kset.kobj, |
| 9268 | &attr_sched_smt_power_savings.attr); |
| 9269 | #endif |
| 9270 | #ifdef CONFIG_SCHED_MC |
| 9271 | if (!err && mc_capable()) |
| 9272 | err = sysfs_create_file(&cls->kset.kobj, |
| 9273 | &attr_sched_mc_power_savings.attr); |
| 9274 | #endif |
| 9275 | return err; |
| 9276 | } |
| 9277 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ |
| 9278 | |
| 9279 | #ifndef CONFIG_CPUSETS |
| 9280 | /* |
| 9281 | * Add online and remove offline CPUs from the scheduler domains. |
| 9282 | * When cpusets are enabled they take over this function. |
| 9283 | */ |
| 9284 | static int update_sched_domains(struct notifier_block *nfb, |
| 9285 | unsigned long action, void *hcpu) |
| 9286 | { |
| 9287 | switch (action) { |
| 9288 | case CPU_ONLINE: |
| 9289 | case CPU_ONLINE_FROZEN: |
| 9290 | case CPU_DEAD: |
| 9291 | case CPU_DEAD_FROZEN: |
| 9292 | partition_sched_domains(1, NULL, NULL); |
| 9293 | return NOTIFY_OK; |
| 9294 | |
| 9295 | default: |
| 9296 | return NOTIFY_DONE; |
| 9297 | } |
| 9298 | } |
| 9299 | #endif |
| 9300 | |
| 9301 | static int update_runtime(struct notifier_block *nfb, |
| 9302 | unsigned long action, void *hcpu) |
| 9303 | { |
| 9304 | int cpu = (int)(long)hcpu; |
| 9305 | |
| 9306 | switch (action) { |
| 9307 | case CPU_DOWN_PREPARE: |
| 9308 | case CPU_DOWN_PREPARE_FROZEN: |
| 9309 | disable_runtime(cpu_rq(cpu)); |
| 9310 | return NOTIFY_OK; |
| 9311 | |
| 9312 | case CPU_DOWN_FAILED: |
| 9313 | case CPU_DOWN_FAILED_FROZEN: |
| 9314 | case CPU_ONLINE: |
| 9315 | case CPU_ONLINE_FROZEN: |
| 9316 | enable_runtime(cpu_rq(cpu)); |
| 9317 | return NOTIFY_OK; |
| 9318 | |
| 9319 | default: |
| 9320 | return NOTIFY_DONE; |
| 9321 | } |
| 9322 | } |
| 9323 | |
| 9324 | void __init sched_init_smp(void) |
| 9325 | { |
| 9326 | cpumask_var_t non_isolated_cpus; |
| 9327 | |
| 9328 | alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); |
| 9329 | |
| 9330 | #if defined(CONFIG_NUMA) |
| 9331 | sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **), |
| 9332 | GFP_KERNEL); |
| 9333 | BUG_ON(sched_group_nodes_bycpu == NULL); |
| 9334 | #endif |
| 9335 | get_online_cpus(); |
| 9336 | mutex_lock(&sched_domains_mutex); |
| 9337 | arch_init_sched_domains(cpu_online_mask); |
| 9338 | cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); |
| 9339 | if (cpumask_empty(non_isolated_cpus)) |
| 9340 | cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); |
| 9341 | mutex_unlock(&sched_domains_mutex); |
| 9342 | put_online_cpus(); |
| 9343 | |
| 9344 | #ifndef CONFIG_CPUSETS |
| 9345 | /* XXX: Theoretical race here - CPU may be hotplugged now */ |
| 9346 | hotcpu_notifier(update_sched_domains, 0); |
| 9347 | #endif |
| 9348 | |
| 9349 | /* RT runtime code needs to handle some hotplug events */ |
| 9350 | hotcpu_notifier(update_runtime, 0); |
| 9351 | |
| 9352 | init_hrtick(); |
| 9353 | |
| 9354 | /* Move init over to a non-isolated CPU */ |
| 9355 | if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) |
| 9356 | BUG(); |
| 9357 | sched_init_granularity(); |
| 9358 | free_cpumask_var(non_isolated_cpus); |
| 9359 | |
| 9360 | alloc_cpumask_var(&fallback_doms, GFP_KERNEL); |
| 9361 | init_sched_rt_class(); |
| 9362 | } |
| 9363 | #else |
| 9364 | void __init sched_init_smp(void) |
| 9365 | { |
| 9366 | sched_init_granularity(); |
| 9367 | } |
| 9368 | #endif /* CONFIG_SMP */ |
| 9369 | |
| 9370 | const_debug unsigned int sysctl_timer_migration = 1; |
| 9371 | |
| 9372 | int in_sched_functions(unsigned long addr) |
| 9373 | { |
| 9374 | return in_lock_functions(addr) || |
| 9375 | (addr >= (unsigned long)__sched_text_start |
| 9376 | && addr < (unsigned long)__sched_text_end); |
| 9377 | } |
| 9378 | |
| 9379 | static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq) |
| 9380 | { |
| 9381 | cfs_rq->tasks_timeline = RB_ROOT; |
| 9382 | INIT_LIST_HEAD(&cfs_rq->tasks); |
| 9383 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 9384 | cfs_rq->rq = rq; |
| 9385 | #endif |
| 9386 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
| 9387 | } |
| 9388 | |
| 9389 | static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) |
| 9390 | { |
| 9391 | struct rt_prio_array *array; |
| 9392 | int i; |
| 9393 | |
| 9394 | array = &rt_rq->active; |
| 9395 | for (i = 0; i < MAX_RT_PRIO; i++) { |
| 9396 | INIT_LIST_HEAD(array->queue + i); |
| 9397 | __clear_bit(i, array->bitmap); |
| 9398 | } |
| 9399 | /* delimiter for bitsearch: */ |
| 9400 | __set_bit(MAX_RT_PRIO, array->bitmap); |
| 9401 | |
| 9402 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
| 9403 | rt_rq->highest_prio.curr = MAX_RT_PRIO; |
| 9404 | #ifdef CONFIG_SMP |
| 9405 | rt_rq->highest_prio.next = MAX_RT_PRIO; |
| 9406 | #endif |
| 9407 | #endif |
| 9408 | #ifdef CONFIG_SMP |
| 9409 | rt_rq->rt_nr_migratory = 0; |
| 9410 | rt_rq->overloaded = 0; |
| 9411 | plist_head_init(&rt_rq->pushable_tasks, &rq->lock); |
| 9412 | #endif |
| 9413 | |
| 9414 | rt_rq->rt_time = 0; |
| 9415 | rt_rq->rt_throttled = 0; |
| 9416 | rt_rq->rt_runtime = 0; |
| 9417 | spin_lock_init(&rt_rq->rt_runtime_lock); |
| 9418 | |
| 9419 | #ifdef CONFIG_RT_GROUP_SCHED |
| 9420 | rt_rq->rt_nr_boosted = 0; |
| 9421 | rt_rq->rq = rq; |
| 9422 | #endif |
| 9423 | } |
| 9424 | |
| 9425 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 9426 | static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, |
| 9427 | struct sched_entity *se, int cpu, int add, |
| 9428 | struct sched_entity *parent) |
| 9429 | { |
| 9430 | struct rq *rq = cpu_rq(cpu); |
| 9431 | tg->cfs_rq[cpu] = cfs_rq; |
| 9432 | init_cfs_rq(cfs_rq, rq); |
| 9433 | cfs_rq->tg = tg; |
| 9434 | if (add) |
| 9435 | list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list); |
| 9436 | |
| 9437 | tg->se[cpu] = se; |
| 9438 | /* se could be NULL for init_task_group */ |
| 9439 | if (!se) |
| 9440 | return; |
| 9441 | |
| 9442 | if (!parent) |
| 9443 | se->cfs_rq = &rq->cfs; |
| 9444 | else |
| 9445 | se->cfs_rq = parent->my_q; |
| 9446 | |
| 9447 | se->my_q = cfs_rq; |
| 9448 | se->load.weight = tg->shares; |
| 9449 | se->load.inv_weight = 0; |
| 9450 | se->parent = parent; |
| 9451 | } |
| 9452 | #endif |
| 9453 | |
| 9454 | #ifdef CONFIG_RT_GROUP_SCHED |
| 9455 | static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, |
| 9456 | struct sched_rt_entity *rt_se, int cpu, int add, |
| 9457 | struct sched_rt_entity *parent) |
| 9458 | { |
| 9459 | struct rq *rq = cpu_rq(cpu); |
| 9460 | |
| 9461 | tg->rt_rq[cpu] = rt_rq; |
| 9462 | init_rt_rq(rt_rq, rq); |
| 9463 | rt_rq->tg = tg; |
| 9464 | rt_rq->rt_se = rt_se; |
| 9465 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; |
| 9466 | if (add) |
| 9467 | list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list); |
| 9468 | |
| 9469 | tg->rt_se[cpu] = rt_se; |
| 9470 | if (!rt_se) |
| 9471 | return; |
| 9472 | |
| 9473 | if (!parent) |
| 9474 | rt_se->rt_rq = &rq->rt; |
| 9475 | else |
| 9476 | rt_se->rt_rq = parent->my_q; |
| 9477 | |
| 9478 | rt_se->my_q = rt_rq; |
| 9479 | rt_se->parent = parent; |
| 9480 | INIT_LIST_HEAD(&rt_se->run_list); |
| 9481 | } |
| 9482 | #endif |
| 9483 | |
| 9484 | void __init sched_init(void) |
| 9485 | { |
| 9486 | int i, j; |
| 9487 | unsigned long alloc_size = 0, ptr; |
| 9488 | |
| 9489 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 9490 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); |
| 9491 | #endif |
| 9492 | #ifdef CONFIG_RT_GROUP_SCHED |
| 9493 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); |
| 9494 | #endif |
| 9495 | #ifdef CONFIG_USER_SCHED |
| 9496 | alloc_size *= 2; |
| 9497 | #endif |
| 9498 | #ifdef CONFIG_CPUMASK_OFFSTACK |
| 9499 | alloc_size += num_possible_cpus() * cpumask_size(); |
| 9500 | #endif |
| 9501 | /* |
| 9502 | * As sched_init() is called before page_alloc is setup, |
| 9503 | * we use alloc_bootmem(). |
| 9504 | */ |
| 9505 | if (alloc_size) { |
| 9506 | ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); |
| 9507 | |
| 9508 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 9509 | init_task_group.se = (struct sched_entity **)ptr; |
| 9510 | ptr += nr_cpu_ids * sizeof(void **); |
| 9511 | |
| 9512 | init_task_group.cfs_rq = (struct cfs_rq **)ptr; |
| 9513 | ptr += nr_cpu_ids * sizeof(void **); |
| 9514 | |
| 9515 | #ifdef CONFIG_USER_SCHED |
| 9516 | root_task_group.se = (struct sched_entity **)ptr; |
| 9517 | ptr += nr_cpu_ids * sizeof(void **); |
| 9518 | |
| 9519 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; |
| 9520 | ptr += nr_cpu_ids * sizeof(void **); |
| 9521 | #endif /* CONFIG_USER_SCHED */ |
| 9522 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 9523 | #ifdef CONFIG_RT_GROUP_SCHED |
| 9524 | init_task_group.rt_se = (struct sched_rt_entity **)ptr; |
| 9525 | ptr += nr_cpu_ids * sizeof(void **); |
| 9526 | |
| 9527 | init_task_group.rt_rq = (struct rt_rq **)ptr; |
| 9528 | ptr += nr_cpu_ids * sizeof(void **); |
| 9529 | |
| 9530 | #ifdef CONFIG_USER_SCHED |
| 9531 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; |
| 9532 | ptr += nr_cpu_ids * sizeof(void **); |
| 9533 | |
| 9534 | root_task_group.rt_rq = (struct rt_rq **)ptr; |
| 9535 | ptr += nr_cpu_ids * sizeof(void **); |
| 9536 | #endif /* CONFIG_USER_SCHED */ |
| 9537 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 9538 | #ifdef CONFIG_CPUMASK_OFFSTACK |
| 9539 | for_each_possible_cpu(i) { |
| 9540 | per_cpu(load_balance_tmpmask, i) = (void *)ptr; |
| 9541 | ptr += cpumask_size(); |
| 9542 | } |
| 9543 | #endif /* CONFIG_CPUMASK_OFFSTACK */ |
| 9544 | } |
| 9545 | |
| 9546 | #ifdef CONFIG_SMP |
| 9547 | init_defrootdomain(); |
| 9548 | #endif |
| 9549 | |
| 9550 | init_rt_bandwidth(&def_rt_bandwidth, |
| 9551 | global_rt_period(), global_rt_runtime()); |
| 9552 | |
| 9553 | #ifdef CONFIG_RT_GROUP_SCHED |
| 9554 | init_rt_bandwidth(&init_task_group.rt_bandwidth, |
| 9555 | global_rt_period(), global_rt_runtime()); |
| 9556 | #ifdef CONFIG_USER_SCHED |
| 9557 | init_rt_bandwidth(&root_task_group.rt_bandwidth, |
| 9558 | global_rt_period(), RUNTIME_INF); |
| 9559 | #endif /* CONFIG_USER_SCHED */ |
| 9560 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 9561 | |
| 9562 | #ifdef CONFIG_GROUP_SCHED |
| 9563 | list_add(&init_task_group.list, &task_groups); |
| 9564 | INIT_LIST_HEAD(&init_task_group.children); |
| 9565 | |
| 9566 | #ifdef CONFIG_USER_SCHED |
| 9567 | INIT_LIST_HEAD(&root_task_group.children); |
| 9568 | init_task_group.parent = &root_task_group; |
| 9569 | list_add(&init_task_group.siblings, &root_task_group.children); |
| 9570 | #endif /* CONFIG_USER_SCHED */ |
| 9571 | #endif /* CONFIG_GROUP_SCHED */ |
| 9572 | |
| 9573 | for_each_possible_cpu(i) { |
| 9574 | struct rq *rq; |
| 9575 | |
| 9576 | rq = cpu_rq(i); |
| 9577 | spin_lock_init(&rq->lock); |
| 9578 | rq->nr_running = 0; |
| 9579 | rq->calc_load_active = 0; |
| 9580 | rq->calc_load_update = jiffies + LOAD_FREQ; |
| 9581 | init_cfs_rq(&rq->cfs, rq); |
| 9582 | init_rt_rq(&rq->rt, rq); |
| 9583 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 9584 | init_task_group.shares = init_task_group_load; |
| 9585 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
| 9586 | #ifdef CONFIG_CGROUP_SCHED |
| 9587 | /* |
| 9588 | * How much cpu bandwidth does init_task_group get? |
| 9589 | * |
| 9590 | * In case of task-groups formed thr' the cgroup filesystem, it |
| 9591 | * gets 100% of the cpu resources in the system. This overall |
| 9592 | * system cpu resource is divided among the tasks of |
| 9593 | * init_task_group and its child task-groups in a fair manner, |
| 9594 | * based on each entity's (task or task-group's) weight |
| 9595 | * (se->load.weight). |
| 9596 | * |
| 9597 | * In other words, if init_task_group has 10 tasks of weight |
| 9598 | * 1024) and two child groups A0 and A1 (of weight 1024 each), |
| 9599 | * then A0's share of the cpu resource is: |
| 9600 | * |
| 9601 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% |
| 9602 | * |
| 9603 | * We achieve this by letting init_task_group's tasks sit |
| 9604 | * directly in rq->cfs (i.e init_task_group->se[] = NULL). |
| 9605 | */ |
| 9606 | init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL); |
| 9607 | #elif defined CONFIG_USER_SCHED |
| 9608 | root_task_group.shares = NICE_0_LOAD; |
| 9609 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL); |
| 9610 | /* |
| 9611 | * In case of task-groups formed thr' the user id of tasks, |
| 9612 | * init_task_group represents tasks belonging to root user. |
| 9613 | * Hence it forms a sibling of all subsequent groups formed. |
| 9614 | * In this case, init_task_group gets only a fraction of overall |
| 9615 | * system cpu resource, based on the weight assigned to root |
| 9616 | * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished |
| 9617 | * by letting tasks of init_task_group sit in a separate cfs_rq |
| 9618 | * (init_tg_cfs_rq) and having one entity represent this group of |
| 9619 | * tasks in rq->cfs (i.e init_task_group->se[] != NULL). |
| 9620 | */ |
| 9621 | init_tg_cfs_entry(&init_task_group, |
| 9622 | &per_cpu(init_tg_cfs_rq, i), |
| 9623 | &per_cpu(init_sched_entity, i), i, 1, |
| 9624 | root_task_group.se[i]); |
| 9625 | |
| 9626 | #endif |
| 9627 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 9628 | |
| 9629 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; |
| 9630 | #ifdef CONFIG_RT_GROUP_SCHED |
| 9631 | INIT_LIST_HEAD(&rq->leaf_rt_rq_list); |
| 9632 | #ifdef CONFIG_CGROUP_SCHED |
| 9633 | init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL); |
| 9634 | #elif defined CONFIG_USER_SCHED |
| 9635 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL); |
| 9636 | init_tg_rt_entry(&init_task_group, |
| 9637 | &per_cpu(init_rt_rq, i), |
| 9638 | &per_cpu(init_sched_rt_entity, i), i, 1, |
| 9639 | root_task_group.rt_se[i]); |
| 9640 | #endif |
| 9641 | #endif |
| 9642 | |
| 9643 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) |
| 9644 | rq->cpu_load[j] = 0; |
| 9645 | #ifdef CONFIG_SMP |
| 9646 | rq->sd = NULL; |
| 9647 | rq->rd = NULL; |
| 9648 | rq->post_schedule = 0; |
| 9649 | rq->active_balance = 0; |
| 9650 | rq->next_balance = jiffies; |
| 9651 | rq->push_cpu = 0; |
| 9652 | rq->cpu = i; |
| 9653 | rq->online = 0; |
| 9654 | rq->migration_thread = NULL; |
| 9655 | INIT_LIST_HEAD(&rq->migration_queue); |
| 9656 | rq_attach_root(rq, &def_root_domain); |
| 9657 | #endif |
| 9658 | init_rq_hrtick(rq); |
| 9659 | atomic_set(&rq->nr_iowait, 0); |
| 9660 | } |
| 9661 | |
| 9662 | set_load_weight(&init_task); |
| 9663 | |
| 9664 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 9665 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); |
| 9666 | #endif |
| 9667 | |
| 9668 | #ifdef CONFIG_SMP |
| 9669 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); |
| 9670 | #endif |
| 9671 | |
| 9672 | #ifdef CONFIG_RT_MUTEXES |
| 9673 | plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); |
| 9674 | #endif |
| 9675 | |
| 9676 | /* |
| 9677 | * The boot idle thread does lazy MMU switching as well: |
| 9678 | */ |
| 9679 | atomic_inc(&init_mm.mm_count); |
| 9680 | enter_lazy_tlb(&init_mm, current); |
| 9681 | |
| 9682 | /* |
| 9683 | * Make us the idle thread. Technically, schedule() should not be |
| 9684 | * called from this thread, however somewhere below it might be, |
| 9685 | * but because we are the idle thread, we just pick up running again |
| 9686 | * when this runqueue becomes "idle". |
| 9687 | */ |
| 9688 | init_idle(current, smp_processor_id()); |
| 9689 | |
| 9690 | calc_load_update = jiffies + LOAD_FREQ; |
| 9691 | |
| 9692 | /* |
| 9693 | * During early bootup we pretend to be a normal task: |
| 9694 | */ |
| 9695 | current->sched_class = &fair_sched_class; |
| 9696 | |
| 9697 | /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */ |
| 9698 | alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT); |
| 9699 | #ifdef CONFIG_SMP |
| 9700 | #ifdef CONFIG_NO_HZ |
| 9701 | alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT); |
| 9702 | alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT); |
| 9703 | #endif |
| 9704 | alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); |
| 9705 | #endif /* SMP */ |
| 9706 | |
| 9707 | perf_counter_init(); |
| 9708 | |
| 9709 | scheduler_running = 1; |
| 9710 | } |
| 9711 | |
| 9712 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
| 9713 | static inline int preempt_count_equals(int preempt_offset) |
| 9714 | { |
| 9715 | int nested = preempt_count() & ~PREEMPT_ACTIVE; |
| 9716 | |
| 9717 | return (nested == PREEMPT_INATOMIC_BASE + preempt_offset); |
| 9718 | } |
| 9719 | |
| 9720 | void __might_sleep(char *file, int line, int preempt_offset) |
| 9721 | { |
| 9722 | #ifdef in_atomic |
| 9723 | static unsigned long prev_jiffy; /* ratelimiting */ |
| 9724 | |
| 9725 | if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || |
| 9726 | system_state != SYSTEM_RUNNING || oops_in_progress) |
| 9727 | return; |
| 9728 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
| 9729 | return; |
| 9730 | prev_jiffy = jiffies; |
| 9731 | |
| 9732 | printk(KERN_ERR |
| 9733 | "BUG: sleeping function called from invalid context at %s:%d\n", |
| 9734 | file, line); |
| 9735 | printk(KERN_ERR |
| 9736 | "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", |
| 9737 | in_atomic(), irqs_disabled(), |
| 9738 | current->pid, current->comm); |
| 9739 | |
| 9740 | debug_show_held_locks(current); |
| 9741 | if (irqs_disabled()) |
| 9742 | print_irqtrace_events(current); |
| 9743 | dump_stack(); |
| 9744 | #endif |
| 9745 | } |
| 9746 | EXPORT_SYMBOL(__might_sleep); |
| 9747 | #endif |
| 9748 | |
| 9749 | #ifdef CONFIG_MAGIC_SYSRQ |
| 9750 | static void normalize_task(struct rq *rq, struct task_struct *p) |
| 9751 | { |
| 9752 | int on_rq; |
| 9753 | |
| 9754 | update_rq_clock(rq); |
| 9755 | on_rq = p->se.on_rq; |
| 9756 | if (on_rq) |
| 9757 | deactivate_task(rq, p, 0); |
| 9758 | __setscheduler(rq, p, SCHED_NORMAL, 0); |
| 9759 | if (on_rq) { |
| 9760 | activate_task(rq, p, 0); |
| 9761 | resched_task(rq->curr); |
| 9762 | } |
| 9763 | } |
| 9764 | |
| 9765 | void normalize_rt_tasks(void) |
| 9766 | { |
| 9767 | struct task_struct *g, *p; |
| 9768 | unsigned long flags; |
| 9769 | struct rq *rq; |
| 9770 | |
| 9771 | read_lock_irqsave(&tasklist_lock, flags); |
| 9772 | do_each_thread(g, p) { |
| 9773 | /* |
| 9774 | * Only normalize user tasks: |
| 9775 | */ |
| 9776 | if (!p->mm) |
| 9777 | continue; |
| 9778 | |
| 9779 | p->se.exec_start = 0; |
| 9780 | #ifdef CONFIG_SCHEDSTATS |
| 9781 | p->se.wait_start = 0; |
| 9782 | p->se.sleep_start = 0; |
| 9783 | p->se.block_start = 0; |
| 9784 | #endif |
| 9785 | |
| 9786 | if (!rt_task(p)) { |
| 9787 | /* |
| 9788 | * Renice negative nice level userspace |
| 9789 | * tasks back to 0: |
| 9790 | */ |
| 9791 | if (TASK_NICE(p) < 0 && p->mm) |
| 9792 | set_user_nice(p, 0); |
| 9793 | continue; |
| 9794 | } |
| 9795 | |
| 9796 | spin_lock(&p->pi_lock); |
| 9797 | rq = __task_rq_lock(p); |
| 9798 | |
| 9799 | normalize_task(rq, p); |
| 9800 | |
| 9801 | __task_rq_unlock(rq); |
| 9802 | spin_unlock(&p->pi_lock); |
| 9803 | } while_each_thread(g, p); |
| 9804 | |
| 9805 | read_unlock_irqrestore(&tasklist_lock, flags); |
| 9806 | } |
| 9807 | |
| 9808 | #endif /* CONFIG_MAGIC_SYSRQ */ |
| 9809 | |
| 9810 | #ifdef CONFIG_IA64 |
| 9811 | /* |
| 9812 | * These functions are only useful for the IA64 MCA handling. |
| 9813 | * |
| 9814 | * They can only be called when the whole system has been |
| 9815 | * stopped - every CPU needs to be quiescent, and no scheduling |
| 9816 | * activity can take place. Using them for anything else would |
| 9817 | * be a serious bug, and as a result, they aren't even visible |
| 9818 | * under any other configuration. |
| 9819 | */ |
| 9820 | |
| 9821 | /** |
| 9822 | * curr_task - return the current task for a given cpu. |
| 9823 | * @cpu: the processor in question. |
| 9824 | * |
| 9825 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 9826 | */ |
| 9827 | struct task_struct *curr_task(int cpu) |
| 9828 | { |
| 9829 | return cpu_curr(cpu); |
| 9830 | } |
| 9831 | |
| 9832 | /** |
| 9833 | * set_curr_task - set the current task for a given cpu. |
| 9834 | * @cpu: the processor in question. |
| 9835 | * @p: the task pointer to set. |
| 9836 | * |
| 9837 | * Description: This function must only be used when non-maskable interrupts |
| 9838 | * are serviced on a separate stack. It allows the architecture to switch the |
| 9839 | * notion of the current task on a cpu in a non-blocking manner. This function |
| 9840 | * must be called with all CPU's synchronized, and interrupts disabled, the |
| 9841 | * and caller must save the original value of the current task (see |
| 9842 | * curr_task() above) and restore that value before reenabling interrupts and |
| 9843 | * re-starting the system. |
| 9844 | * |
| 9845 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 9846 | */ |
| 9847 | void set_curr_task(int cpu, struct task_struct *p) |
| 9848 | { |
| 9849 | cpu_curr(cpu) = p; |
| 9850 | } |
| 9851 | |
| 9852 | #endif |
| 9853 | |
| 9854 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 9855 | static void free_fair_sched_group(struct task_group *tg) |
| 9856 | { |
| 9857 | int i; |
| 9858 | |
| 9859 | for_each_possible_cpu(i) { |
| 9860 | if (tg->cfs_rq) |
| 9861 | kfree(tg->cfs_rq[i]); |
| 9862 | if (tg->se) |
| 9863 | kfree(tg->se[i]); |
| 9864 | } |
| 9865 | |
| 9866 | kfree(tg->cfs_rq); |
| 9867 | kfree(tg->se); |
| 9868 | } |
| 9869 | |
| 9870 | static |
| 9871 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) |
| 9872 | { |
| 9873 | struct cfs_rq *cfs_rq; |
| 9874 | struct sched_entity *se; |
| 9875 | struct rq *rq; |
| 9876 | int i; |
| 9877 | |
| 9878 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); |
| 9879 | if (!tg->cfs_rq) |
| 9880 | goto err; |
| 9881 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); |
| 9882 | if (!tg->se) |
| 9883 | goto err; |
| 9884 | |
| 9885 | tg->shares = NICE_0_LOAD; |
| 9886 | |
| 9887 | for_each_possible_cpu(i) { |
| 9888 | rq = cpu_rq(i); |
| 9889 | |
| 9890 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), |
| 9891 | GFP_KERNEL, cpu_to_node(i)); |
| 9892 | if (!cfs_rq) |
| 9893 | goto err; |
| 9894 | |
| 9895 | se = kzalloc_node(sizeof(struct sched_entity), |
| 9896 | GFP_KERNEL, cpu_to_node(i)); |
| 9897 | if (!se) |
| 9898 | goto err; |
| 9899 | |
| 9900 | init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]); |
| 9901 | } |
| 9902 | |
| 9903 | return 1; |
| 9904 | |
| 9905 | err: |
| 9906 | return 0; |
| 9907 | } |
| 9908 | |
| 9909 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) |
| 9910 | { |
| 9911 | list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list, |
| 9912 | &cpu_rq(cpu)->leaf_cfs_rq_list); |
| 9913 | } |
| 9914 | |
| 9915 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) |
| 9916 | { |
| 9917 | list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list); |
| 9918 | } |
| 9919 | #else /* !CONFG_FAIR_GROUP_SCHED */ |
| 9920 | static inline void free_fair_sched_group(struct task_group *tg) |
| 9921 | { |
| 9922 | } |
| 9923 | |
| 9924 | static inline |
| 9925 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) |
| 9926 | { |
| 9927 | return 1; |
| 9928 | } |
| 9929 | |
| 9930 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) |
| 9931 | { |
| 9932 | } |
| 9933 | |
| 9934 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) |
| 9935 | { |
| 9936 | } |
| 9937 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 9938 | |
| 9939 | #ifdef CONFIG_RT_GROUP_SCHED |
| 9940 | static void free_rt_sched_group(struct task_group *tg) |
| 9941 | { |
| 9942 | int i; |
| 9943 | |
| 9944 | destroy_rt_bandwidth(&tg->rt_bandwidth); |
| 9945 | |
| 9946 | for_each_possible_cpu(i) { |
| 9947 | if (tg->rt_rq) |
| 9948 | kfree(tg->rt_rq[i]); |
| 9949 | if (tg->rt_se) |
| 9950 | kfree(tg->rt_se[i]); |
| 9951 | } |
| 9952 | |
| 9953 | kfree(tg->rt_rq); |
| 9954 | kfree(tg->rt_se); |
| 9955 | } |
| 9956 | |
| 9957 | static |
| 9958 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) |
| 9959 | { |
| 9960 | struct rt_rq *rt_rq; |
| 9961 | struct sched_rt_entity *rt_se; |
| 9962 | struct rq *rq; |
| 9963 | int i; |
| 9964 | |
| 9965 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); |
| 9966 | if (!tg->rt_rq) |
| 9967 | goto err; |
| 9968 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); |
| 9969 | if (!tg->rt_se) |
| 9970 | goto err; |
| 9971 | |
| 9972 | init_rt_bandwidth(&tg->rt_bandwidth, |
| 9973 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); |
| 9974 | |
| 9975 | for_each_possible_cpu(i) { |
| 9976 | rq = cpu_rq(i); |
| 9977 | |
| 9978 | rt_rq = kzalloc_node(sizeof(struct rt_rq), |
| 9979 | GFP_KERNEL, cpu_to_node(i)); |
| 9980 | if (!rt_rq) |
| 9981 | goto err; |
| 9982 | |
| 9983 | rt_se = kzalloc_node(sizeof(struct sched_rt_entity), |
| 9984 | GFP_KERNEL, cpu_to_node(i)); |
| 9985 | if (!rt_se) |
| 9986 | goto err; |
| 9987 | |
| 9988 | init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]); |
| 9989 | } |
| 9990 | |
| 9991 | return 1; |
| 9992 | |
| 9993 | err: |
| 9994 | return 0; |
| 9995 | } |
| 9996 | |
| 9997 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) |
| 9998 | { |
| 9999 | list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list, |
| 10000 | &cpu_rq(cpu)->leaf_rt_rq_list); |
| 10001 | } |
| 10002 | |
| 10003 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) |
| 10004 | { |
| 10005 | list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list); |
| 10006 | } |
| 10007 | #else /* !CONFIG_RT_GROUP_SCHED */ |
| 10008 | static inline void free_rt_sched_group(struct task_group *tg) |
| 10009 | { |
| 10010 | } |
| 10011 | |
| 10012 | static inline |
| 10013 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) |
| 10014 | { |
| 10015 | return 1; |
| 10016 | } |
| 10017 | |
| 10018 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) |
| 10019 | { |
| 10020 | } |
| 10021 | |
| 10022 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) |
| 10023 | { |
| 10024 | } |
| 10025 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 10026 | |
| 10027 | #ifdef CONFIG_GROUP_SCHED |
| 10028 | static void free_sched_group(struct task_group *tg) |
| 10029 | { |
| 10030 | free_fair_sched_group(tg); |
| 10031 | free_rt_sched_group(tg); |
| 10032 | kfree(tg); |
| 10033 | } |
| 10034 | |
| 10035 | /* allocate runqueue etc for a new task group */ |
| 10036 | struct task_group *sched_create_group(struct task_group *parent) |
| 10037 | { |
| 10038 | struct task_group *tg; |
| 10039 | unsigned long flags; |
| 10040 | int i; |
| 10041 | |
| 10042 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); |
| 10043 | if (!tg) |
| 10044 | return ERR_PTR(-ENOMEM); |
| 10045 | |
| 10046 | if (!alloc_fair_sched_group(tg, parent)) |
| 10047 | goto err; |
| 10048 | |
| 10049 | if (!alloc_rt_sched_group(tg, parent)) |
| 10050 | goto err; |
| 10051 | |
| 10052 | spin_lock_irqsave(&task_group_lock, flags); |
| 10053 | for_each_possible_cpu(i) { |
| 10054 | register_fair_sched_group(tg, i); |
| 10055 | register_rt_sched_group(tg, i); |
| 10056 | } |
| 10057 | list_add_rcu(&tg->list, &task_groups); |
| 10058 | |
| 10059 | WARN_ON(!parent); /* root should already exist */ |
| 10060 | |
| 10061 | tg->parent = parent; |
| 10062 | INIT_LIST_HEAD(&tg->children); |
| 10063 | list_add_rcu(&tg->siblings, &parent->children); |
| 10064 | spin_unlock_irqrestore(&task_group_lock, flags); |
| 10065 | |
| 10066 | return tg; |
| 10067 | |
| 10068 | err: |
| 10069 | free_sched_group(tg); |
| 10070 | return ERR_PTR(-ENOMEM); |
| 10071 | } |
| 10072 | |
| 10073 | /* rcu callback to free various structures associated with a task group */ |
| 10074 | static void free_sched_group_rcu(struct rcu_head *rhp) |
| 10075 | { |
| 10076 | /* now it should be safe to free those cfs_rqs */ |
| 10077 | free_sched_group(container_of(rhp, struct task_group, rcu)); |
| 10078 | } |
| 10079 | |
| 10080 | /* Destroy runqueue etc associated with a task group */ |
| 10081 | void sched_destroy_group(struct task_group *tg) |
| 10082 | { |
| 10083 | unsigned long flags; |
| 10084 | int i; |
| 10085 | |
| 10086 | spin_lock_irqsave(&task_group_lock, flags); |
| 10087 | for_each_possible_cpu(i) { |
| 10088 | unregister_fair_sched_group(tg, i); |
| 10089 | unregister_rt_sched_group(tg, i); |
| 10090 | } |
| 10091 | list_del_rcu(&tg->list); |
| 10092 | list_del_rcu(&tg->siblings); |
| 10093 | spin_unlock_irqrestore(&task_group_lock, flags); |
| 10094 | |
| 10095 | /* wait for possible concurrent references to cfs_rqs complete */ |
| 10096 | call_rcu(&tg->rcu, free_sched_group_rcu); |
| 10097 | } |
| 10098 | |
| 10099 | /* change task's runqueue when it moves between groups. |
| 10100 | * The caller of this function should have put the task in its new group |
| 10101 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to |
| 10102 | * reflect its new group. |
| 10103 | */ |
| 10104 | void sched_move_task(struct task_struct *tsk) |
| 10105 | { |
| 10106 | int on_rq, running; |
| 10107 | unsigned long flags; |
| 10108 | struct rq *rq; |
| 10109 | |
| 10110 | rq = task_rq_lock(tsk, &flags); |
| 10111 | |
| 10112 | update_rq_clock(rq); |
| 10113 | |
| 10114 | running = task_current(rq, tsk); |
| 10115 | on_rq = tsk->se.on_rq; |
| 10116 | |
| 10117 | if (on_rq) |
| 10118 | dequeue_task(rq, tsk, 0); |
| 10119 | if (unlikely(running)) |
| 10120 | tsk->sched_class->put_prev_task(rq, tsk); |
| 10121 | |
| 10122 | set_task_rq(tsk, task_cpu(tsk)); |
| 10123 | |
| 10124 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 10125 | if (tsk->sched_class->moved_group) |
| 10126 | tsk->sched_class->moved_group(tsk); |
| 10127 | #endif |
| 10128 | |
| 10129 | if (unlikely(running)) |
| 10130 | tsk->sched_class->set_curr_task(rq); |
| 10131 | if (on_rq) |
| 10132 | enqueue_task(rq, tsk, 0); |
| 10133 | |
| 10134 | task_rq_unlock(rq, &flags); |
| 10135 | } |
| 10136 | #endif /* CONFIG_GROUP_SCHED */ |
| 10137 | |
| 10138 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 10139 | static void __set_se_shares(struct sched_entity *se, unsigned long shares) |
| 10140 | { |
| 10141 | struct cfs_rq *cfs_rq = se->cfs_rq; |
| 10142 | int on_rq; |
| 10143 | |
| 10144 | on_rq = se->on_rq; |
| 10145 | if (on_rq) |
| 10146 | dequeue_entity(cfs_rq, se, 0); |
| 10147 | |
| 10148 | se->load.weight = shares; |
| 10149 | se->load.inv_weight = 0; |
| 10150 | |
| 10151 | if (on_rq) |
| 10152 | enqueue_entity(cfs_rq, se, 0); |
| 10153 | } |
| 10154 | |
| 10155 | static void set_se_shares(struct sched_entity *se, unsigned long shares) |
| 10156 | { |
| 10157 | struct cfs_rq *cfs_rq = se->cfs_rq; |
| 10158 | struct rq *rq = cfs_rq->rq; |
| 10159 | unsigned long flags; |
| 10160 | |
| 10161 | spin_lock_irqsave(&rq->lock, flags); |
| 10162 | __set_se_shares(se, shares); |
| 10163 | spin_unlock_irqrestore(&rq->lock, flags); |
| 10164 | } |
| 10165 | |
| 10166 | static DEFINE_MUTEX(shares_mutex); |
| 10167 | |
| 10168 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) |
| 10169 | { |
| 10170 | int i; |
| 10171 | unsigned long flags; |
| 10172 | |
| 10173 | /* |
| 10174 | * We can't change the weight of the root cgroup. |
| 10175 | */ |
| 10176 | if (!tg->se[0]) |
| 10177 | return -EINVAL; |
| 10178 | |
| 10179 | if (shares < MIN_SHARES) |
| 10180 | shares = MIN_SHARES; |
| 10181 | else if (shares > MAX_SHARES) |
| 10182 | shares = MAX_SHARES; |
| 10183 | |
| 10184 | mutex_lock(&shares_mutex); |
| 10185 | if (tg->shares == shares) |
| 10186 | goto done; |
| 10187 | |
| 10188 | spin_lock_irqsave(&task_group_lock, flags); |
| 10189 | for_each_possible_cpu(i) |
| 10190 | unregister_fair_sched_group(tg, i); |
| 10191 | list_del_rcu(&tg->siblings); |
| 10192 | spin_unlock_irqrestore(&task_group_lock, flags); |
| 10193 | |
| 10194 | /* wait for any ongoing reference to this group to finish */ |
| 10195 | synchronize_sched(); |
| 10196 | |
| 10197 | /* |
| 10198 | * Now we are free to modify the group's share on each cpu |
| 10199 | * w/o tripping rebalance_share or load_balance_fair. |
| 10200 | */ |
| 10201 | tg->shares = shares; |
| 10202 | for_each_possible_cpu(i) { |
| 10203 | /* |
| 10204 | * force a rebalance |
| 10205 | */ |
| 10206 | cfs_rq_set_shares(tg->cfs_rq[i], 0); |
| 10207 | set_se_shares(tg->se[i], shares); |
| 10208 | } |
| 10209 | |
| 10210 | /* |
| 10211 | * Enable load balance activity on this group, by inserting it back on |
| 10212 | * each cpu's rq->leaf_cfs_rq_list. |
| 10213 | */ |
| 10214 | spin_lock_irqsave(&task_group_lock, flags); |
| 10215 | for_each_possible_cpu(i) |
| 10216 | register_fair_sched_group(tg, i); |
| 10217 | list_add_rcu(&tg->siblings, &tg->parent->children); |
| 10218 | spin_unlock_irqrestore(&task_group_lock, flags); |
| 10219 | done: |
| 10220 | mutex_unlock(&shares_mutex); |
| 10221 | return 0; |
| 10222 | } |
| 10223 | |
| 10224 | unsigned long sched_group_shares(struct task_group *tg) |
| 10225 | { |
| 10226 | return tg->shares; |
| 10227 | } |
| 10228 | #endif |
| 10229 | |
| 10230 | #ifdef CONFIG_RT_GROUP_SCHED |
| 10231 | /* |
| 10232 | * Ensure that the real time constraints are schedulable. |
| 10233 | */ |
| 10234 | static DEFINE_MUTEX(rt_constraints_mutex); |
| 10235 | |
| 10236 | static unsigned long to_ratio(u64 period, u64 runtime) |
| 10237 | { |
| 10238 | if (runtime == RUNTIME_INF) |
| 10239 | return 1ULL << 20; |
| 10240 | |
| 10241 | return div64_u64(runtime << 20, period); |
| 10242 | } |
| 10243 | |
| 10244 | /* Must be called with tasklist_lock held */ |
| 10245 | static inline int tg_has_rt_tasks(struct task_group *tg) |
| 10246 | { |
| 10247 | struct task_struct *g, *p; |
| 10248 | |
| 10249 | do_each_thread(g, p) { |
| 10250 | if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg) |
| 10251 | return 1; |
| 10252 | } while_each_thread(g, p); |
| 10253 | |
| 10254 | return 0; |
| 10255 | } |
| 10256 | |
| 10257 | struct rt_schedulable_data { |
| 10258 | struct task_group *tg; |
| 10259 | u64 rt_period; |
| 10260 | u64 rt_runtime; |
| 10261 | }; |
| 10262 | |
| 10263 | static int tg_schedulable(struct task_group *tg, void *data) |
| 10264 | { |
| 10265 | struct rt_schedulable_data *d = data; |
| 10266 | struct task_group *child; |
| 10267 | unsigned long total, sum = 0; |
| 10268 | u64 period, runtime; |
| 10269 | |
| 10270 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); |
| 10271 | runtime = tg->rt_bandwidth.rt_runtime; |
| 10272 | |
| 10273 | if (tg == d->tg) { |
| 10274 | period = d->rt_period; |
| 10275 | runtime = d->rt_runtime; |
| 10276 | } |
| 10277 | |
| 10278 | #ifdef CONFIG_USER_SCHED |
| 10279 | if (tg == &root_task_group) { |
| 10280 | period = global_rt_period(); |
| 10281 | runtime = global_rt_runtime(); |
| 10282 | } |
| 10283 | #endif |
| 10284 | |
| 10285 | /* |
| 10286 | * Cannot have more runtime than the period. |
| 10287 | */ |
| 10288 | if (runtime > period && runtime != RUNTIME_INF) |
| 10289 | return -EINVAL; |
| 10290 | |
| 10291 | /* |
| 10292 | * Ensure we don't starve existing RT tasks. |
| 10293 | */ |
| 10294 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) |
| 10295 | return -EBUSY; |
| 10296 | |
| 10297 | total = to_ratio(period, runtime); |
| 10298 | |
| 10299 | /* |
| 10300 | * Nobody can have more than the global setting allows. |
| 10301 | */ |
| 10302 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) |
| 10303 | return -EINVAL; |
| 10304 | |
| 10305 | /* |
| 10306 | * The sum of our children's runtime should not exceed our own. |
| 10307 | */ |
| 10308 | list_for_each_entry_rcu(child, &tg->children, siblings) { |
| 10309 | period = ktime_to_ns(child->rt_bandwidth.rt_period); |
| 10310 | runtime = child->rt_bandwidth.rt_runtime; |
| 10311 | |
| 10312 | if (child == d->tg) { |
| 10313 | period = d->rt_period; |
| 10314 | runtime = d->rt_runtime; |
| 10315 | } |
| 10316 | |
| 10317 | sum += to_ratio(period, runtime); |
| 10318 | } |
| 10319 | |
| 10320 | if (sum > total) |
| 10321 | return -EINVAL; |
| 10322 | |
| 10323 | return 0; |
| 10324 | } |
| 10325 | |
| 10326 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) |
| 10327 | { |
| 10328 | struct rt_schedulable_data data = { |
| 10329 | .tg = tg, |
| 10330 | .rt_period = period, |
| 10331 | .rt_runtime = runtime, |
| 10332 | }; |
| 10333 | |
| 10334 | return walk_tg_tree(tg_schedulable, tg_nop, &data); |
| 10335 | } |
| 10336 | |
| 10337 | static int tg_set_bandwidth(struct task_group *tg, |
| 10338 | u64 rt_period, u64 rt_runtime) |
| 10339 | { |
| 10340 | int i, err = 0; |
| 10341 | |
| 10342 | mutex_lock(&rt_constraints_mutex); |
| 10343 | read_lock(&tasklist_lock); |
| 10344 | err = __rt_schedulable(tg, rt_period, rt_runtime); |
| 10345 | if (err) |
| 10346 | goto unlock; |
| 10347 | |
| 10348 | spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
| 10349 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); |
| 10350 | tg->rt_bandwidth.rt_runtime = rt_runtime; |
| 10351 | |
| 10352 | for_each_possible_cpu(i) { |
| 10353 | struct rt_rq *rt_rq = tg->rt_rq[i]; |
| 10354 | |
| 10355 | spin_lock(&rt_rq->rt_runtime_lock); |
| 10356 | rt_rq->rt_runtime = rt_runtime; |
| 10357 | spin_unlock(&rt_rq->rt_runtime_lock); |
| 10358 | } |
| 10359 | spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
| 10360 | unlock: |
| 10361 | read_unlock(&tasklist_lock); |
| 10362 | mutex_unlock(&rt_constraints_mutex); |
| 10363 | |
| 10364 | return err; |
| 10365 | } |
| 10366 | |
| 10367 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) |
| 10368 | { |
| 10369 | u64 rt_runtime, rt_period; |
| 10370 | |
| 10371 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); |
| 10372 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; |
| 10373 | if (rt_runtime_us < 0) |
| 10374 | rt_runtime = RUNTIME_INF; |
| 10375 | |
| 10376 | return tg_set_bandwidth(tg, rt_period, rt_runtime); |
| 10377 | } |
| 10378 | |
| 10379 | long sched_group_rt_runtime(struct task_group *tg) |
| 10380 | { |
| 10381 | u64 rt_runtime_us; |
| 10382 | |
| 10383 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) |
| 10384 | return -1; |
| 10385 | |
| 10386 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; |
| 10387 | do_div(rt_runtime_us, NSEC_PER_USEC); |
| 10388 | return rt_runtime_us; |
| 10389 | } |
| 10390 | |
| 10391 | int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) |
| 10392 | { |
| 10393 | u64 rt_runtime, rt_period; |
| 10394 | |
| 10395 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; |
| 10396 | rt_runtime = tg->rt_bandwidth.rt_runtime; |
| 10397 | |
| 10398 | if (rt_period == 0) |
| 10399 | return -EINVAL; |
| 10400 | |
| 10401 | return tg_set_bandwidth(tg, rt_period, rt_runtime); |
| 10402 | } |
| 10403 | |
| 10404 | long sched_group_rt_period(struct task_group *tg) |
| 10405 | { |
| 10406 | u64 rt_period_us; |
| 10407 | |
| 10408 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); |
| 10409 | do_div(rt_period_us, NSEC_PER_USEC); |
| 10410 | return rt_period_us; |
| 10411 | } |
| 10412 | |
| 10413 | static int sched_rt_global_constraints(void) |
| 10414 | { |
| 10415 | u64 runtime, period; |
| 10416 | int ret = 0; |
| 10417 | |
| 10418 | if (sysctl_sched_rt_period <= 0) |
| 10419 | return -EINVAL; |
| 10420 | |
| 10421 | runtime = global_rt_runtime(); |
| 10422 | period = global_rt_period(); |
| 10423 | |
| 10424 | /* |
| 10425 | * Sanity check on the sysctl variables. |
| 10426 | */ |
| 10427 | if (runtime > period && runtime != RUNTIME_INF) |
| 10428 | return -EINVAL; |
| 10429 | |
| 10430 | mutex_lock(&rt_constraints_mutex); |
| 10431 | read_lock(&tasklist_lock); |
| 10432 | ret = __rt_schedulable(NULL, 0, 0); |
| 10433 | read_unlock(&tasklist_lock); |
| 10434 | mutex_unlock(&rt_constraints_mutex); |
| 10435 | |
| 10436 | return ret; |
| 10437 | } |
| 10438 | |
| 10439 | int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) |
| 10440 | { |
| 10441 | /* Don't accept realtime tasks when there is no way for them to run */ |
| 10442 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) |
| 10443 | return 0; |
| 10444 | |
| 10445 | return 1; |
| 10446 | } |
| 10447 | |
| 10448 | #else /* !CONFIG_RT_GROUP_SCHED */ |
| 10449 | static int sched_rt_global_constraints(void) |
| 10450 | { |
| 10451 | unsigned long flags; |
| 10452 | int i; |
| 10453 | |
| 10454 | if (sysctl_sched_rt_period <= 0) |
| 10455 | return -EINVAL; |
| 10456 | |
| 10457 | /* |
| 10458 | * There's always some RT tasks in the root group |
| 10459 | * -- migration, kstopmachine etc.. |
| 10460 | */ |
| 10461 | if (sysctl_sched_rt_runtime == 0) |
| 10462 | return -EBUSY; |
| 10463 | |
| 10464 | spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); |
| 10465 | for_each_possible_cpu(i) { |
| 10466 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; |
| 10467 | |
| 10468 | spin_lock(&rt_rq->rt_runtime_lock); |
| 10469 | rt_rq->rt_runtime = global_rt_runtime(); |
| 10470 | spin_unlock(&rt_rq->rt_runtime_lock); |
| 10471 | } |
| 10472 | spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); |
| 10473 | |
| 10474 | return 0; |
| 10475 | } |
| 10476 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 10477 | |
| 10478 | int sched_rt_handler(struct ctl_table *table, int write, |
| 10479 | struct file *filp, void __user *buffer, size_t *lenp, |
| 10480 | loff_t *ppos) |
| 10481 | { |
| 10482 | int ret; |
| 10483 | int old_period, old_runtime; |
| 10484 | static DEFINE_MUTEX(mutex); |
| 10485 | |
| 10486 | mutex_lock(&mutex); |
| 10487 | old_period = sysctl_sched_rt_period; |
| 10488 | old_runtime = sysctl_sched_rt_runtime; |
| 10489 | |
| 10490 | ret = proc_dointvec(table, write, filp, buffer, lenp, ppos); |
| 10491 | |
| 10492 | if (!ret && write) { |
| 10493 | ret = sched_rt_global_constraints(); |
| 10494 | if (ret) { |
| 10495 | sysctl_sched_rt_period = old_period; |
| 10496 | sysctl_sched_rt_runtime = old_runtime; |
| 10497 | } else { |
| 10498 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); |
| 10499 | def_rt_bandwidth.rt_period = |
| 10500 | ns_to_ktime(global_rt_period()); |
| 10501 | } |
| 10502 | } |
| 10503 | mutex_unlock(&mutex); |
| 10504 | |
| 10505 | return ret; |
| 10506 | } |
| 10507 | |
| 10508 | #ifdef CONFIG_CGROUP_SCHED |
| 10509 | |
| 10510 | /* return corresponding task_group object of a cgroup */ |
| 10511 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) |
| 10512 | { |
| 10513 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), |
| 10514 | struct task_group, css); |
| 10515 | } |
| 10516 | |
| 10517 | static struct cgroup_subsys_state * |
| 10518 | cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) |
| 10519 | { |
| 10520 | struct task_group *tg, *parent; |
| 10521 | |
| 10522 | if (!cgrp->parent) { |
| 10523 | /* This is early initialization for the top cgroup */ |
| 10524 | return &init_task_group.css; |
| 10525 | } |
| 10526 | |
| 10527 | parent = cgroup_tg(cgrp->parent); |
| 10528 | tg = sched_create_group(parent); |
| 10529 | if (IS_ERR(tg)) |
| 10530 | return ERR_PTR(-ENOMEM); |
| 10531 | |
| 10532 | return &tg->css; |
| 10533 | } |
| 10534 | |
| 10535 | static void |
| 10536 | cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) |
| 10537 | { |
| 10538 | struct task_group *tg = cgroup_tg(cgrp); |
| 10539 | |
| 10540 | sched_destroy_group(tg); |
| 10541 | } |
| 10542 | |
| 10543 | static int |
| 10544 | cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
| 10545 | struct task_struct *tsk) |
| 10546 | { |
| 10547 | #ifdef CONFIG_RT_GROUP_SCHED |
| 10548 | if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk)) |
| 10549 | return -EINVAL; |
| 10550 | #else |
| 10551 | /* We don't support RT-tasks being in separate groups */ |
| 10552 | if (tsk->sched_class != &fair_sched_class) |
| 10553 | return -EINVAL; |
| 10554 | #endif |
| 10555 | |
| 10556 | return 0; |
| 10557 | } |
| 10558 | |
| 10559 | static void |
| 10560 | cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
| 10561 | struct cgroup *old_cont, struct task_struct *tsk) |
| 10562 | { |
| 10563 | sched_move_task(tsk); |
| 10564 | } |
| 10565 | |
| 10566 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 10567 | static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, |
| 10568 | u64 shareval) |
| 10569 | { |
| 10570 | return sched_group_set_shares(cgroup_tg(cgrp), shareval); |
| 10571 | } |
| 10572 | |
| 10573 | static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) |
| 10574 | { |
| 10575 | struct task_group *tg = cgroup_tg(cgrp); |
| 10576 | |
| 10577 | return (u64) tg->shares; |
| 10578 | } |
| 10579 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 10580 | |
| 10581 | #ifdef CONFIG_RT_GROUP_SCHED |
| 10582 | static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, |
| 10583 | s64 val) |
| 10584 | { |
| 10585 | return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); |
| 10586 | } |
| 10587 | |
| 10588 | static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) |
| 10589 | { |
| 10590 | return sched_group_rt_runtime(cgroup_tg(cgrp)); |
| 10591 | } |
| 10592 | |
| 10593 | static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, |
| 10594 | u64 rt_period_us) |
| 10595 | { |
| 10596 | return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); |
| 10597 | } |
| 10598 | |
| 10599 | static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) |
| 10600 | { |
| 10601 | return sched_group_rt_period(cgroup_tg(cgrp)); |
| 10602 | } |
| 10603 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 10604 | |
| 10605 | static struct cftype cpu_files[] = { |
| 10606 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 10607 | { |
| 10608 | .name = "shares", |
| 10609 | .read_u64 = cpu_shares_read_u64, |
| 10610 | .write_u64 = cpu_shares_write_u64, |
| 10611 | }, |
| 10612 | #endif |
| 10613 | #ifdef CONFIG_RT_GROUP_SCHED |
| 10614 | { |
| 10615 | .name = "rt_runtime_us", |
| 10616 | .read_s64 = cpu_rt_runtime_read, |
| 10617 | .write_s64 = cpu_rt_runtime_write, |
| 10618 | }, |
| 10619 | { |
| 10620 | .name = "rt_period_us", |
| 10621 | .read_u64 = cpu_rt_period_read_uint, |
| 10622 | .write_u64 = cpu_rt_period_write_uint, |
| 10623 | }, |
| 10624 | #endif |
| 10625 | }; |
| 10626 | |
| 10627 | static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) |
| 10628 | { |
| 10629 | return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); |
| 10630 | } |
| 10631 | |
| 10632 | struct cgroup_subsys cpu_cgroup_subsys = { |
| 10633 | .name = "cpu", |
| 10634 | .create = cpu_cgroup_create, |
| 10635 | .destroy = cpu_cgroup_destroy, |
| 10636 | .can_attach = cpu_cgroup_can_attach, |
| 10637 | .attach = cpu_cgroup_attach, |
| 10638 | .populate = cpu_cgroup_populate, |
| 10639 | .subsys_id = cpu_cgroup_subsys_id, |
| 10640 | .early_init = 1, |
| 10641 | }; |
| 10642 | |
| 10643 | #endif /* CONFIG_CGROUP_SCHED */ |
| 10644 | |
| 10645 | #ifdef CONFIG_CGROUP_CPUACCT |
| 10646 | |
| 10647 | /* |
| 10648 | * CPU accounting code for task groups. |
| 10649 | * |
| 10650 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh |
| 10651 | * (balbir@in.ibm.com). |
| 10652 | */ |
| 10653 | |
| 10654 | /* track cpu usage of a group of tasks and its child groups */ |
| 10655 | struct cpuacct { |
| 10656 | struct cgroup_subsys_state css; |
| 10657 | /* cpuusage holds pointer to a u64-type object on every cpu */ |
| 10658 | u64 *cpuusage; |
| 10659 | struct percpu_counter cpustat[CPUACCT_STAT_NSTATS]; |
| 10660 | struct cpuacct *parent; |
| 10661 | }; |
| 10662 | |
| 10663 | struct cgroup_subsys cpuacct_subsys; |
| 10664 | |
| 10665 | /* return cpu accounting group corresponding to this container */ |
| 10666 | static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) |
| 10667 | { |
| 10668 | return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), |
| 10669 | struct cpuacct, css); |
| 10670 | } |
| 10671 | |
| 10672 | /* return cpu accounting group to which this task belongs */ |
| 10673 | static inline struct cpuacct *task_ca(struct task_struct *tsk) |
| 10674 | { |
| 10675 | return container_of(task_subsys_state(tsk, cpuacct_subsys_id), |
| 10676 | struct cpuacct, css); |
| 10677 | } |
| 10678 | |
| 10679 | /* create a new cpu accounting group */ |
| 10680 | static struct cgroup_subsys_state *cpuacct_create( |
| 10681 | struct cgroup_subsys *ss, struct cgroup *cgrp) |
| 10682 | { |
| 10683 | struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); |
| 10684 | int i; |
| 10685 | |
| 10686 | if (!ca) |
| 10687 | goto out; |
| 10688 | |
| 10689 | ca->cpuusage = alloc_percpu(u64); |
| 10690 | if (!ca->cpuusage) |
| 10691 | goto out_free_ca; |
| 10692 | |
| 10693 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) |
| 10694 | if (percpu_counter_init(&ca->cpustat[i], 0)) |
| 10695 | goto out_free_counters; |
| 10696 | |
| 10697 | if (cgrp->parent) |
| 10698 | ca->parent = cgroup_ca(cgrp->parent); |
| 10699 | |
| 10700 | return &ca->css; |
| 10701 | |
| 10702 | out_free_counters: |
| 10703 | while (--i >= 0) |
| 10704 | percpu_counter_destroy(&ca->cpustat[i]); |
| 10705 | free_percpu(ca->cpuusage); |
| 10706 | out_free_ca: |
| 10707 | kfree(ca); |
| 10708 | out: |
| 10709 | return ERR_PTR(-ENOMEM); |
| 10710 | } |
| 10711 | |
| 10712 | /* destroy an existing cpu accounting group */ |
| 10713 | static void |
| 10714 | cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) |
| 10715 | { |
| 10716 | struct cpuacct *ca = cgroup_ca(cgrp); |
| 10717 | int i; |
| 10718 | |
| 10719 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) |
| 10720 | percpu_counter_destroy(&ca->cpustat[i]); |
| 10721 | free_percpu(ca->cpuusage); |
| 10722 | kfree(ca); |
| 10723 | } |
| 10724 | |
| 10725 | static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) |
| 10726 | { |
| 10727 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); |
| 10728 | u64 data; |
| 10729 | |
| 10730 | #ifndef CONFIG_64BIT |
| 10731 | /* |
| 10732 | * Take rq->lock to make 64-bit read safe on 32-bit platforms. |
| 10733 | */ |
| 10734 | spin_lock_irq(&cpu_rq(cpu)->lock); |
| 10735 | data = *cpuusage; |
| 10736 | spin_unlock_irq(&cpu_rq(cpu)->lock); |
| 10737 | #else |
| 10738 | data = *cpuusage; |
| 10739 | #endif |
| 10740 | |
| 10741 | return data; |
| 10742 | } |
| 10743 | |
| 10744 | static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) |
| 10745 | { |
| 10746 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); |
| 10747 | |
| 10748 | #ifndef CONFIG_64BIT |
| 10749 | /* |
| 10750 | * Take rq->lock to make 64-bit write safe on 32-bit platforms. |
| 10751 | */ |
| 10752 | spin_lock_irq(&cpu_rq(cpu)->lock); |
| 10753 | *cpuusage = val; |
| 10754 | spin_unlock_irq(&cpu_rq(cpu)->lock); |
| 10755 | #else |
| 10756 | *cpuusage = val; |
| 10757 | #endif |
| 10758 | } |
| 10759 | |
| 10760 | /* return total cpu usage (in nanoseconds) of a group */ |
| 10761 | static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) |
| 10762 | { |
| 10763 | struct cpuacct *ca = cgroup_ca(cgrp); |
| 10764 | u64 totalcpuusage = 0; |
| 10765 | int i; |
| 10766 | |
| 10767 | for_each_present_cpu(i) |
| 10768 | totalcpuusage += cpuacct_cpuusage_read(ca, i); |
| 10769 | |
| 10770 | return totalcpuusage; |
| 10771 | } |
| 10772 | |
| 10773 | static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, |
| 10774 | u64 reset) |
| 10775 | { |
| 10776 | struct cpuacct *ca = cgroup_ca(cgrp); |
| 10777 | int err = 0; |
| 10778 | int i; |
| 10779 | |
| 10780 | if (reset) { |
| 10781 | err = -EINVAL; |
| 10782 | goto out; |
| 10783 | } |
| 10784 | |
| 10785 | for_each_present_cpu(i) |
| 10786 | cpuacct_cpuusage_write(ca, i, 0); |
| 10787 | |
| 10788 | out: |
| 10789 | return err; |
| 10790 | } |
| 10791 | |
| 10792 | static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, |
| 10793 | struct seq_file *m) |
| 10794 | { |
| 10795 | struct cpuacct *ca = cgroup_ca(cgroup); |
| 10796 | u64 percpu; |
| 10797 | int i; |
| 10798 | |
| 10799 | for_each_present_cpu(i) { |
| 10800 | percpu = cpuacct_cpuusage_read(ca, i); |
| 10801 | seq_printf(m, "%llu ", (unsigned long long) percpu); |
| 10802 | } |
| 10803 | seq_printf(m, "\n"); |
| 10804 | return 0; |
| 10805 | } |
| 10806 | |
| 10807 | static const char *cpuacct_stat_desc[] = { |
| 10808 | [CPUACCT_STAT_USER] = "user", |
| 10809 | [CPUACCT_STAT_SYSTEM] = "system", |
| 10810 | }; |
| 10811 | |
| 10812 | static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, |
| 10813 | struct cgroup_map_cb *cb) |
| 10814 | { |
| 10815 | struct cpuacct *ca = cgroup_ca(cgrp); |
| 10816 | int i; |
| 10817 | |
| 10818 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) { |
| 10819 | s64 val = percpu_counter_read(&ca->cpustat[i]); |
| 10820 | val = cputime64_to_clock_t(val); |
| 10821 | cb->fill(cb, cpuacct_stat_desc[i], val); |
| 10822 | } |
| 10823 | return 0; |
| 10824 | } |
| 10825 | |
| 10826 | static struct cftype files[] = { |
| 10827 | { |
| 10828 | .name = "usage", |
| 10829 | .read_u64 = cpuusage_read, |
| 10830 | .write_u64 = cpuusage_write, |
| 10831 | }, |
| 10832 | { |
| 10833 | .name = "usage_percpu", |
| 10834 | .read_seq_string = cpuacct_percpu_seq_read, |
| 10835 | }, |
| 10836 | { |
| 10837 | .name = "stat", |
| 10838 | .read_map = cpuacct_stats_show, |
| 10839 | }, |
| 10840 | }; |
| 10841 | |
| 10842 | static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) |
| 10843 | { |
| 10844 | return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); |
| 10845 | } |
| 10846 | |
| 10847 | /* |
| 10848 | * charge this task's execution time to its accounting group. |
| 10849 | * |
| 10850 | * called with rq->lock held. |
| 10851 | */ |
| 10852 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime) |
| 10853 | { |
| 10854 | struct cpuacct *ca; |
| 10855 | int cpu; |
| 10856 | |
| 10857 | if (unlikely(!cpuacct_subsys.active)) |
| 10858 | return; |
| 10859 | |
| 10860 | cpu = task_cpu(tsk); |
| 10861 | |
| 10862 | rcu_read_lock(); |
| 10863 | |
| 10864 | ca = task_ca(tsk); |
| 10865 | |
| 10866 | for (; ca; ca = ca->parent) { |
| 10867 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); |
| 10868 | *cpuusage += cputime; |
| 10869 | } |
| 10870 | |
| 10871 | rcu_read_unlock(); |
| 10872 | } |
| 10873 | |
| 10874 | /* |
| 10875 | * Charge the system/user time to the task's accounting group. |
| 10876 | */ |
| 10877 | static void cpuacct_update_stats(struct task_struct *tsk, |
| 10878 | enum cpuacct_stat_index idx, cputime_t val) |
| 10879 | { |
| 10880 | struct cpuacct *ca; |
| 10881 | |
| 10882 | if (unlikely(!cpuacct_subsys.active)) |
| 10883 | return; |
| 10884 | |
| 10885 | rcu_read_lock(); |
| 10886 | ca = task_ca(tsk); |
| 10887 | |
| 10888 | do { |
| 10889 | percpu_counter_add(&ca->cpustat[idx], val); |
| 10890 | ca = ca->parent; |
| 10891 | } while (ca); |
| 10892 | rcu_read_unlock(); |
| 10893 | } |
| 10894 | |
| 10895 | struct cgroup_subsys cpuacct_subsys = { |
| 10896 | .name = "cpuacct", |
| 10897 | .create = cpuacct_create, |
| 10898 | .destroy = cpuacct_destroy, |
| 10899 | .populate = cpuacct_populate, |
| 10900 | .subsys_id = cpuacct_subsys_id, |
| 10901 | }; |
| 10902 | #endif /* CONFIG_CGROUP_CPUACCT */ |
| 10903 | |
| 10904 | #ifndef CONFIG_SMP |
| 10905 | |
| 10906 | int rcu_expedited_torture_stats(char *page) |
| 10907 | { |
| 10908 | return 0; |
| 10909 | } |
| 10910 | EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats); |
| 10911 | |
| 10912 | void synchronize_sched_expedited(void) |
| 10913 | { |
| 10914 | } |
| 10915 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); |
| 10916 | |
| 10917 | #else /* #ifndef CONFIG_SMP */ |
| 10918 | |
| 10919 | static DEFINE_PER_CPU(struct migration_req, rcu_migration_req); |
| 10920 | static DEFINE_MUTEX(rcu_sched_expedited_mutex); |
| 10921 | |
| 10922 | #define RCU_EXPEDITED_STATE_POST -2 |
| 10923 | #define RCU_EXPEDITED_STATE_IDLE -1 |
| 10924 | |
| 10925 | static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE; |
| 10926 | |
| 10927 | int rcu_expedited_torture_stats(char *page) |
| 10928 | { |
| 10929 | int cnt = 0; |
| 10930 | int cpu; |
| 10931 | |
| 10932 | cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state); |
| 10933 | for_each_online_cpu(cpu) { |
| 10934 | cnt += sprintf(&page[cnt], " %d:%d", |
| 10935 | cpu, per_cpu(rcu_migration_req, cpu).dest_cpu); |
| 10936 | } |
| 10937 | cnt += sprintf(&page[cnt], "\n"); |
| 10938 | return cnt; |
| 10939 | } |
| 10940 | EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats); |
| 10941 | |
| 10942 | static long synchronize_sched_expedited_count; |
| 10943 | |
| 10944 | /* |
| 10945 | * Wait for an rcu-sched grace period to elapse, but use "big hammer" |
| 10946 | * approach to force grace period to end quickly. This consumes |
| 10947 | * significant time on all CPUs, and is thus not recommended for |
| 10948 | * any sort of common-case code. |
| 10949 | * |
| 10950 | * Note that it is illegal to call this function while holding any |
| 10951 | * lock that is acquired by a CPU-hotplug notifier. Failing to |
| 10952 | * observe this restriction will result in deadlock. |
| 10953 | */ |
| 10954 | void synchronize_sched_expedited(void) |
| 10955 | { |
| 10956 | int cpu; |
| 10957 | unsigned long flags; |
| 10958 | bool need_full_sync = 0; |
| 10959 | struct rq *rq; |
| 10960 | struct migration_req *req; |
| 10961 | long snap; |
| 10962 | int trycount = 0; |
| 10963 | |
| 10964 | smp_mb(); /* ensure prior mod happens before capturing snap. */ |
| 10965 | snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1; |
| 10966 | get_online_cpus(); |
| 10967 | while (!mutex_trylock(&rcu_sched_expedited_mutex)) { |
| 10968 | put_online_cpus(); |
| 10969 | if (trycount++ < 10) |
| 10970 | udelay(trycount * num_online_cpus()); |
| 10971 | else { |
| 10972 | synchronize_sched(); |
| 10973 | return; |
| 10974 | } |
| 10975 | if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) { |
| 10976 | smp_mb(); /* ensure test happens before caller kfree */ |
| 10977 | return; |
| 10978 | } |
| 10979 | get_online_cpus(); |
| 10980 | } |
| 10981 | rcu_expedited_state = RCU_EXPEDITED_STATE_POST; |
| 10982 | for_each_online_cpu(cpu) { |
| 10983 | rq = cpu_rq(cpu); |
| 10984 | req = &per_cpu(rcu_migration_req, cpu); |
| 10985 | init_completion(&req->done); |
| 10986 | req->task = NULL; |
| 10987 | req->dest_cpu = RCU_MIGRATION_NEED_QS; |
| 10988 | spin_lock_irqsave(&rq->lock, flags); |
| 10989 | list_add(&req->list, &rq->migration_queue); |
| 10990 | spin_unlock_irqrestore(&rq->lock, flags); |
| 10991 | wake_up_process(rq->migration_thread); |
| 10992 | } |
| 10993 | for_each_online_cpu(cpu) { |
| 10994 | rcu_expedited_state = cpu; |
| 10995 | req = &per_cpu(rcu_migration_req, cpu); |
| 10996 | rq = cpu_rq(cpu); |
| 10997 | wait_for_completion(&req->done); |
| 10998 | spin_lock_irqsave(&rq->lock, flags); |
| 10999 | if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC)) |
| 11000 | need_full_sync = 1; |
| 11001 | req->dest_cpu = RCU_MIGRATION_IDLE; |
| 11002 | spin_unlock_irqrestore(&rq->lock, flags); |
| 11003 | } |
| 11004 | rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE; |
| 11005 | mutex_unlock(&rcu_sched_expedited_mutex); |
| 11006 | put_online_cpus(); |
| 11007 | if (need_full_sync) |
| 11008 | synchronize_sched(); |
| 11009 | } |
| 11010 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); |
| 11011 | |
| 11012 | #endif /* #else #ifndef CONFIG_SMP */ |