aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/sched.c
diff options
context:
space:
mode:
Diffstat (limited to 'kernel/sched.c')
-rw-r--r--kernel/sched.c1919
1 files changed, 79 insertions, 1840 deletions
diff --git a/kernel/sched.c b/kernel/sched.c
index 64298a52eaa6..13a2acf18b2d 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -1805,6 +1805,51 @@ static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1805 raw_spin_unlock(&busiest->lock); 1805 raw_spin_unlock(&busiest->lock);
1806 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); 1806 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1807} 1807}
1808
1809/*
1810 * double_rq_lock - safely lock two runqueues
1811 *
1812 * Note this does not disable interrupts like task_rq_lock,
1813 * you need to do so manually before calling.
1814 */
1815static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1816 __acquires(rq1->lock)
1817 __acquires(rq2->lock)
1818{
1819 BUG_ON(!irqs_disabled());
1820 if (rq1 == rq2) {
1821 raw_spin_lock(&rq1->lock);
1822 __acquire(rq2->lock); /* Fake it out ;) */
1823 } else {
1824 if (rq1 < rq2) {
1825 raw_spin_lock(&rq1->lock);
1826 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1827 } else {
1828 raw_spin_lock(&rq2->lock);
1829 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1830 }
1831 }
1832 update_rq_clock(rq1);
1833 update_rq_clock(rq2);
1834}
1835
1836/*
1837 * double_rq_unlock - safely unlock two runqueues
1838 *
1839 * Note this does not restore interrupts like task_rq_unlock,
1840 * you need to do so manually after calling.
1841 */
1842static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1843 __releases(rq1->lock)
1844 __releases(rq2->lock)
1845{
1846 raw_spin_unlock(&rq1->lock);
1847 if (rq1 != rq2)
1848 raw_spin_unlock(&rq2->lock);
1849 else
1850 __release(rq2->lock);
1851}
1852
1808#endif 1853#endif
1809 1854
1810#ifdef CONFIG_FAIR_GROUP_SCHED 1855#ifdef CONFIG_FAIR_GROUP_SCHED
@@ -1834,18 +1879,14 @@ static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1834#endif 1879#endif
1835} 1880}
1836 1881
1837#include "sched_stats.h" 1882static const struct sched_class rt_sched_class;
1838#include "sched_idletask.c"
1839#include "sched_fair.c"
1840#include "sched_rt.c"
1841#ifdef CONFIG_SCHED_DEBUG
1842# include "sched_debug.c"
1843#endif
1844 1883
1845#define sched_class_highest (&rt_sched_class) 1884#define sched_class_highest (&rt_sched_class)
1846#define for_each_class(class) \ 1885#define for_each_class(class) \
1847 for (class = sched_class_highest; class; class = class->next) 1886 for (class = sched_class_highest; class; class = class->next)
1848 1887
1888#include "sched_stats.h"
1889
1849static void inc_nr_running(struct rq *rq) 1890static void inc_nr_running(struct rq *rq)
1850{ 1891{
1851 rq->nr_running++; 1892 rq->nr_running++;
@@ -1912,6 +1953,37 @@ static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1912} 1953}
1913 1954
1914/* 1955/*
1956 * activate_task - move a task to the runqueue.
1957 */
1958static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1959{
1960 if (task_contributes_to_load(p))
1961 rq->nr_uninterruptible--;
1962
1963 enqueue_task(rq, p, wakeup);
1964 inc_nr_running(rq);
1965}
1966
1967/*
1968 * deactivate_task - remove a task from the runqueue.
1969 */
1970static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1971{
1972 if (task_contributes_to_load(p))
1973 rq->nr_uninterruptible++;
1974
1975 dequeue_task(rq, p, sleep);
1976 dec_nr_running(rq);
1977}
1978
1979#include "sched_idletask.c"
1980#include "sched_fair.c"
1981#include "sched_rt.c"
1982#ifdef CONFIG_SCHED_DEBUG
1983# include "sched_debug.c"
1984#endif
1985
1986/*
1915 * __normal_prio - return the priority that is based on the static prio 1987 * __normal_prio - return the priority that is based on the static prio
1916 */ 1988 */
1917static inline int __normal_prio(struct task_struct *p) 1989static inline int __normal_prio(struct task_struct *p)
@@ -1957,30 +2029,6 @@ static int effective_prio(struct task_struct *p)
1957 return p->prio; 2029 return p->prio;
1958} 2030}
1959 2031
1960/*
1961 * activate_task - move a task to the runqueue.
1962 */
1963static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1964{
1965 if (task_contributes_to_load(p))
1966 rq->nr_uninterruptible--;
1967
1968 enqueue_task(rq, p, wakeup);
1969 inc_nr_running(rq);
1970}
1971
1972/*
1973 * deactivate_task - remove a task from the runqueue.
1974 */
1975static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1976{
1977 if (task_contributes_to_load(p))
1978 rq->nr_uninterruptible++;
1979
1980 dequeue_task(rq, p, sleep);
1981 dec_nr_running(rq);
1982}
1983
1984/** 2032/**
1985 * task_curr - is this task currently executing on a CPU? 2033 * task_curr - is this task currently executing on a CPU?
1986 * @p: the task in question. 2034 * @p: the task in question.
@@ -3088,50 +3136,6 @@ static void update_cpu_load(struct rq *this_rq)
3088#ifdef CONFIG_SMP 3136#ifdef CONFIG_SMP
3089 3137
3090/* 3138/*
3091 * double_rq_lock - safely lock two runqueues
3092 *
3093 * Note this does not disable interrupts like task_rq_lock,
3094 * you need to do so manually before calling.
3095 */
3096static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3097 __acquires(rq1->lock)
3098 __acquires(rq2->lock)
3099{
3100 BUG_ON(!irqs_disabled());
3101 if (rq1 == rq2) {
3102 raw_spin_lock(&rq1->lock);
3103 __acquire(rq2->lock); /* Fake it out ;) */
3104 } else {
3105 if (rq1 < rq2) {
3106 raw_spin_lock(&rq1->lock);
3107 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3108 } else {
3109 raw_spin_lock(&rq2->lock);
3110 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3111 }
3112 }
3113 update_rq_clock(rq1);
3114 update_rq_clock(rq2);
3115}
3116
3117/*
3118 * double_rq_unlock - safely unlock two runqueues
3119 *
3120 * Note this does not restore interrupts like task_rq_unlock,
3121 * you need to do so manually after calling.
3122 */
3123static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3124 __releases(rq1->lock)
3125 __releases(rq2->lock)
3126{
3127 raw_spin_unlock(&rq1->lock);
3128 if (rq1 != rq2)
3129 raw_spin_unlock(&rq2->lock);
3130 else
3131 __release(rq2->lock);
3132}
3133
3134/*
3135 * sched_exec - execve() is a valuable balancing opportunity, because at 3139 * sched_exec - execve() is a valuable balancing opportunity, because at
3136 * this point the task has the smallest effective memory and cache footprint. 3140 * this point the task has the smallest effective memory and cache footprint.
3137 */ 3141 */
@@ -3179,1771 +3183,6 @@ again:
3179 task_rq_unlock(rq, &flags); 3183 task_rq_unlock(rq, &flags);
3180} 3184}
3181 3185
3182/*
3183 * pull_task - move a task from a remote runqueue to the local runqueue.
3184 * Both runqueues must be locked.
3185 */
3186static void pull_task(struct rq *src_rq, struct task_struct *p,
3187 struct rq *this_rq, int this_cpu)
3188{
3189 deactivate_task(src_rq, p, 0);
3190 set_task_cpu(p, this_cpu);
3191 activate_task(this_rq, p, 0);
3192 check_preempt_curr(this_rq, p, 0);
3193}
3194
3195/*
3196 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3197 */
3198static
3199int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3200 struct sched_domain *sd, enum cpu_idle_type idle,
3201 int *all_pinned)
3202{
3203 int tsk_cache_hot = 0;
3204 /*
3205 * We do not migrate tasks that are:
3206 * 1) running (obviously), or
3207 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3208 * 3) are cache-hot on their current CPU.
3209 */
3210 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3211 schedstat_inc(p, se.nr_failed_migrations_affine);
3212 return 0;
3213 }
3214 *all_pinned = 0;
3215
3216 if (task_running(rq, p)) {
3217 schedstat_inc(p, se.nr_failed_migrations_running);
3218 return 0;
3219 }
3220
3221 /*
3222 * Aggressive migration if:
3223 * 1) task is cache cold, or
3224 * 2) too many balance attempts have failed.
3225 */
3226
3227 tsk_cache_hot = task_hot(p, rq->clock, sd);
3228 if (!tsk_cache_hot ||
3229 sd->nr_balance_failed > sd->cache_nice_tries) {
3230#ifdef CONFIG_SCHEDSTATS
3231 if (tsk_cache_hot) {
3232 schedstat_inc(sd, lb_hot_gained[idle]);
3233 schedstat_inc(p, se.nr_forced_migrations);
3234 }
3235#endif
3236 return 1;
3237 }
3238
3239 if (tsk_cache_hot) {
3240 schedstat_inc(p, se.nr_failed_migrations_hot);
3241 return 0;
3242 }
3243 return 1;
3244}
3245
3246static unsigned long
3247balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3248 unsigned long max_load_move, struct sched_domain *sd,
3249 enum cpu_idle_type idle, int *all_pinned,
3250 int *this_best_prio, struct rq_iterator *iterator)
3251{
3252 int loops = 0, pulled = 0, pinned = 0;
3253 struct task_struct *p;
3254 long rem_load_move = max_load_move;
3255
3256 if (max_load_move == 0)
3257 goto out;
3258
3259 pinned = 1;
3260
3261 /*
3262 * Start the load-balancing iterator:
3263 */
3264 p = iterator->start(iterator->arg);
3265next:
3266 if (!p || loops++ > sysctl_sched_nr_migrate)
3267 goto out;
3268
3269 if ((p->se.load.weight >> 1) > rem_load_move ||
3270 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3271 p = iterator->next(iterator->arg);
3272 goto next;
3273 }
3274
3275 pull_task(busiest, p, this_rq, this_cpu);
3276 pulled++;
3277 rem_load_move -= p->se.load.weight;
3278
3279#ifdef CONFIG_PREEMPT
3280 /*
3281 * NEWIDLE balancing is a source of latency, so preemptible kernels
3282 * will stop after the first task is pulled to minimize the critical
3283 * section.
3284 */
3285 if (idle == CPU_NEWLY_IDLE)
3286 goto out;
3287#endif
3288
3289 /*
3290 * We only want to steal up to the prescribed amount of weighted load.
3291 */
3292 if (rem_load_move > 0) {
3293 if (p->prio < *this_best_prio)
3294 *this_best_prio = p->prio;
3295 p = iterator->next(iterator->arg);
3296 goto next;
3297 }
3298out:
3299 /*
3300 * Right now, this is one of only two places pull_task() is called,
3301 * so we can safely collect pull_task() stats here rather than
3302 * inside pull_task().
3303 */
3304 schedstat_add(sd, lb_gained[idle], pulled);
3305
3306 if (all_pinned)
3307 *all_pinned = pinned;
3308
3309 return max_load_move - rem_load_move;
3310}
3311
3312/*
3313 * move_tasks tries to move up to max_load_move weighted load from busiest to
3314 * this_rq, as part of a balancing operation within domain "sd".
3315 * Returns 1 if successful and 0 otherwise.
3316 *
3317 * Called with both runqueues locked.
3318 */
3319static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3320 unsigned long max_load_move,
3321 struct sched_domain *sd, enum cpu_idle_type idle,
3322 int *all_pinned)
3323{
3324 const struct sched_class *class = sched_class_highest;
3325 unsigned long total_load_moved = 0;
3326 int this_best_prio = this_rq->curr->prio;
3327
3328 do {
3329 total_load_moved +=
3330 class->load_balance(this_rq, this_cpu, busiest,
3331 max_load_move - total_load_moved,
3332 sd, idle, all_pinned, &this_best_prio);
3333 class = class->next;
3334
3335#ifdef CONFIG_PREEMPT
3336 /*
3337 * NEWIDLE balancing is a source of latency, so preemptible
3338 * kernels will stop after the first task is pulled to minimize
3339 * the critical section.
3340 */
3341 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3342 break;
3343#endif
3344 } while (class && max_load_move > total_load_moved);
3345
3346 return total_load_moved > 0;
3347}
3348
3349static int
3350iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3351 struct sched_domain *sd, enum cpu_idle_type idle,
3352 struct rq_iterator *iterator)
3353{
3354 struct task_struct *p = iterator->start(iterator->arg);
3355 int pinned = 0;
3356
3357 while (p) {
3358 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3359 pull_task(busiest, p, this_rq, this_cpu);
3360 /*
3361 * Right now, this is only the second place pull_task()
3362 * is called, so we can safely collect pull_task()
3363 * stats here rather than inside pull_task().
3364 */
3365 schedstat_inc(sd, lb_gained[idle]);
3366
3367 return 1;
3368 }
3369 p = iterator->next(iterator->arg);
3370 }
3371
3372 return 0;
3373}
3374
3375/*
3376 * move_one_task tries to move exactly one task from busiest to this_rq, as
3377 * part of active balancing operations within "domain".
3378 * Returns 1 if successful and 0 otherwise.
3379 *
3380 * Called with both runqueues locked.
3381 */
3382static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3383 struct sched_domain *sd, enum cpu_idle_type idle)
3384{
3385 const struct sched_class *class;
3386
3387 for_each_class(class) {
3388 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3389 return 1;
3390 }
3391
3392 return 0;
3393}
3394/********** Helpers for find_busiest_group ************************/
3395/*
3396 * sd_lb_stats - Structure to store the statistics of a sched_domain
3397 * during load balancing.
3398 */
3399struct sd_lb_stats {
3400 struct sched_group *busiest; /* Busiest group in this sd */
3401 struct sched_group *this; /* Local group in this sd */
3402 unsigned long total_load; /* Total load of all groups in sd */
3403 unsigned long total_pwr; /* Total power of all groups in sd */
3404 unsigned long avg_load; /* Average load across all groups in sd */
3405
3406 /** Statistics of this group */
3407 unsigned long this_load;
3408 unsigned long this_load_per_task;
3409 unsigned long this_nr_running;
3410
3411 /* Statistics of the busiest group */
3412 unsigned long max_load;
3413 unsigned long busiest_load_per_task;
3414 unsigned long busiest_nr_running;
3415
3416 int group_imb; /* Is there imbalance in this sd */
3417#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3418 int power_savings_balance; /* Is powersave balance needed for this sd */
3419 struct sched_group *group_min; /* Least loaded group in sd */
3420 struct sched_group *group_leader; /* Group which relieves group_min */
3421 unsigned long min_load_per_task; /* load_per_task in group_min */
3422 unsigned long leader_nr_running; /* Nr running of group_leader */
3423 unsigned long min_nr_running; /* Nr running of group_min */
3424#endif
3425};
3426
3427/*
3428 * sg_lb_stats - stats of a sched_group required for load_balancing
3429 */
3430struct sg_lb_stats {
3431 unsigned long avg_load; /*Avg load across the CPUs of the group */
3432 unsigned long group_load; /* Total load over the CPUs of the group */
3433 unsigned long sum_nr_running; /* Nr tasks running in the group */
3434 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3435 unsigned long group_capacity;
3436 int group_imb; /* Is there an imbalance in the group ? */
3437};
3438
3439/**
3440 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3441 * @group: The group whose first cpu is to be returned.
3442 */
3443static inline unsigned int group_first_cpu(struct sched_group *group)
3444{
3445 return cpumask_first(sched_group_cpus(group));
3446}
3447
3448/**
3449 * get_sd_load_idx - Obtain the load index for a given sched domain.
3450 * @sd: The sched_domain whose load_idx is to be obtained.
3451 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3452 */
3453static inline int get_sd_load_idx(struct sched_domain *sd,
3454 enum cpu_idle_type idle)
3455{
3456 int load_idx;
3457
3458 switch (idle) {
3459 case CPU_NOT_IDLE:
3460 load_idx = sd->busy_idx;
3461 break;
3462
3463 case CPU_NEWLY_IDLE:
3464 load_idx = sd->newidle_idx;
3465 break;
3466 default:
3467 load_idx = sd->idle_idx;
3468 break;
3469 }
3470
3471 return load_idx;
3472}
3473
3474
3475#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3476/**
3477 * init_sd_power_savings_stats - Initialize power savings statistics for
3478 * the given sched_domain, during load balancing.
3479 *
3480 * @sd: Sched domain whose power-savings statistics are to be initialized.
3481 * @sds: Variable containing the statistics for sd.
3482 * @idle: Idle status of the CPU at which we're performing load-balancing.
3483 */
3484static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3485 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3486{
3487 /*
3488 * Busy processors will not participate in power savings
3489 * balance.
3490 */
3491 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3492 sds->power_savings_balance = 0;
3493 else {
3494 sds->power_savings_balance = 1;
3495 sds->min_nr_running = ULONG_MAX;
3496 sds->leader_nr_running = 0;
3497 }
3498}
3499
3500/**
3501 * update_sd_power_savings_stats - Update the power saving stats for a
3502 * sched_domain while performing load balancing.
3503 *
3504 * @group: sched_group belonging to the sched_domain under consideration.
3505 * @sds: Variable containing the statistics of the sched_domain
3506 * @local_group: Does group contain the CPU for which we're performing
3507 * load balancing ?
3508 * @sgs: Variable containing the statistics of the group.
3509 */
3510static inline void update_sd_power_savings_stats(struct sched_group *group,
3511 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3512{
3513
3514 if (!sds->power_savings_balance)
3515 return;
3516
3517 /*
3518 * If the local group is idle or completely loaded
3519 * no need to do power savings balance at this domain
3520 */
3521 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3522 !sds->this_nr_running))
3523 sds->power_savings_balance = 0;
3524
3525 /*
3526 * If a group is already running at full capacity or idle,
3527 * don't include that group in power savings calculations
3528 */
3529 if (!sds->power_savings_balance ||
3530 sgs->sum_nr_running >= sgs->group_capacity ||
3531 !sgs->sum_nr_running)
3532 return;
3533
3534 /*
3535 * Calculate the group which has the least non-idle load.
3536 * This is the group from where we need to pick up the load
3537 * for saving power
3538 */
3539 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3540 (sgs->sum_nr_running == sds->min_nr_running &&
3541 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3542 sds->group_min = group;
3543 sds->min_nr_running = sgs->sum_nr_running;
3544 sds->min_load_per_task = sgs->sum_weighted_load /
3545 sgs->sum_nr_running;
3546 }
3547
3548 /*
3549 * Calculate the group which is almost near its
3550 * capacity but still has some space to pick up some load
3551 * from other group and save more power
3552 */
3553 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3554 return;
3555
3556 if (sgs->sum_nr_running > sds->leader_nr_running ||
3557 (sgs->sum_nr_running == sds->leader_nr_running &&
3558 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3559 sds->group_leader = group;
3560 sds->leader_nr_running = sgs->sum_nr_running;
3561 }
3562}
3563
3564/**
3565 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3566 * @sds: Variable containing the statistics of the sched_domain
3567 * under consideration.
3568 * @this_cpu: Cpu at which we're currently performing load-balancing.
3569 * @imbalance: Variable to store the imbalance.
3570 *
3571 * Description:
3572 * Check if we have potential to perform some power-savings balance.
3573 * If yes, set the busiest group to be the least loaded group in the
3574 * sched_domain, so that it's CPUs can be put to idle.
3575 *
3576 * Returns 1 if there is potential to perform power-savings balance.
3577 * Else returns 0.
3578 */
3579static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3580 int this_cpu, unsigned long *imbalance)
3581{
3582 if (!sds->power_savings_balance)
3583 return 0;
3584
3585 if (sds->this != sds->group_leader ||
3586 sds->group_leader == sds->group_min)
3587 return 0;
3588
3589 *imbalance = sds->min_load_per_task;
3590 sds->busiest = sds->group_min;
3591
3592 return 1;
3593
3594}
3595#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3596static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3597 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3598{
3599 return;
3600}
3601
3602static inline void update_sd_power_savings_stats(struct sched_group *group,
3603 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3604{
3605 return;
3606}
3607
3608static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3609 int this_cpu, unsigned long *imbalance)
3610{
3611 return 0;
3612}
3613#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3614
3615
3616unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3617{
3618 return SCHED_LOAD_SCALE;
3619}
3620
3621unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3622{
3623 return default_scale_freq_power(sd, cpu);
3624}
3625
3626unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3627{
3628 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3629 unsigned long smt_gain = sd->smt_gain;
3630
3631 smt_gain /= weight;
3632
3633 return smt_gain;
3634}
3635
3636unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3637{
3638 return default_scale_smt_power(sd, cpu);
3639}
3640
3641unsigned long scale_rt_power(int cpu)
3642{
3643 struct rq *rq = cpu_rq(cpu);
3644 u64 total, available;
3645
3646 sched_avg_update(rq);
3647
3648 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3649 available = total - rq->rt_avg;
3650
3651 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3652 total = SCHED_LOAD_SCALE;
3653
3654 total >>= SCHED_LOAD_SHIFT;
3655
3656 return div_u64(available, total);
3657}
3658
3659static void update_cpu_power(struct sched_domain *sd, int cpu)
3660{
3661 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3662 unsigned long power = SCHED_LOAD_SCALE;
3663 struct sched_group *sdg = sd->groups;
3664
3665 if (sched_feat(ARCH_POWER))
3666 power *= arch_scale_freq_power(sd, cpu);
3667 else
3668 power *= default_scale_freq_power(sd, cpu);
3669
3670 power >>= SCHED_LOAD_SHIFT;
3671
3672 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3673 if (sched_feat(ARCH_POWER))
3674 power *= arch_scale_smt_power(sd, cpu);
3675 else
3676 power *= default_scale_smt_power(sd, cpu);
3677
3678 power >>= SCHED_LOAD_SHIFT;
3679 }
3680
3681 power *= scale_rt_power(cpu);
3682 power >>= SCHED_LOAD_SHIFT;
3683
3684 if (!power)
3685 power = 1;
3686
3687 sdg->cpu_power = power;
3688}
3689
3690static void update_group_power(struct sched_domain *sd, int cpu)
3691{
3692 struct sched_domain *child = sd->child;
3693 struct sched_group *group, *sdg = sd->groups;
3694 unsigned long power;
3695
3696 if (!child) {
3697 update_cpu_power(sd, cpu);
3698 return;
3699 }
3700
3701 power = 0;
3702
3703 group = child->groups;
3704 do {
3705 power += group->cpu_power;
3706 group = group->next;
3707 } while (group != child->groups);
3708
3709 sdg->cpu_power = power;
3710}
3711
3712/**
3713 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3714 * @sd: The sched_domain whose statistics are to be updated.
3715 * @group: sched_group whose statistics are to be updated.
3716 * @this_cpu: Cpu for which load balance is currently performed.
3717 * @idle: Idle status of this_cpu
3718 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3719 * @sd_idle: Idle status of the sched_domain containing group.
3720 * @local_group: Does group contain this_cpu.
3721 * @cpus: Set of cpus considered for load balancing.
3722 * @balance: Should we balance.
3723 * @sgs: variable to hold the statistics for this group.
3724 */
3725static inline void update_sg_lb_stats(struct sched_domain *sd,
3726 struct sched_group *group, int this_cpu,
3727 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3728 int local_group, const struct cpumask *cpus,
3729 int *balance, struct sg_lb_stats *sgs)
3730{
3731 unsigned long load, max_cpu_load, min_cpu_load;
3732 int i;
3733 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3734 unsigned long sum_avg_load_per_task;
3735 unsigned long avg_load_per_task;
3736
3737 if (local_group) {
3738 balance_cpu = group_first_cpu(group);
3739 if (balance_cpu == this_cpu)
3740 update_group_power(sd, this_cpu);
3741 }
3742
3743 /* Tally up the load of all CPUs in the group */
3744 sum_avg_load_per_task = avg_load_per_task = 0;
3745 max_cpu_load = 0;
3746 min_cpu_load = ~0UL;
3747
3748 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3749 struct rq *rq = cpu_rq(i);
3750
3751 if (*sd_idle && rq->nr_running)
3752 *sd_idle = 0;
3753
3754 /* Bias balancing toward cpus of our domain */
3755 if (local_group) {
3756 if (idle_cpu(i) && !first_idle_cpu) {
3757 first_idle_cpu = 1;
3758 balance_cpu = i;
3759 }
3760
3761 load = target_load(i, load_idx);
3762 } else {
3763 load = source_load(i, load_idx);
3764 if (load > max_cpu_load)
3765 max_cpu_load = load;
3766 if (min_cpu_load > load)
3767 min_cpu_load = load;
3768 }
3769
3770 sgs->group_load += load;
3771 sgs->sum_nr_running += rq->nr_running;
3772 sgs->sum_weighted_load += weighted_cpuload(i);
3773
3774 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3775 }
3776
3777 /*
3778 * First idle cpu or the first cpu(busiest) in this sched group
3779 * is eligible for doing load balancing at this and above
3780 * domains. In the newly idle case, we will allow all the cpu's
3781 * to do the newly idle load balance.
3782 */
3783 if (idle != CPU_NEWLY_IDLE && local_group &&
3784 balance_cpu != this_cpu && balance) {
3785 *balance = 0;
3786 return;
3787 }
3788
3789 /* Adjust by relative CPU power of the group */
3790 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3791
3792
3793 /*
3794 * Consider the group unbalanced when the imbalance is larger
3795 * than the average weight of two tasks.
3796 *
3797 * APZ: with cgroup the avg task weight can vary wildly and
3798 * might not be a suitable number - should we keep a
3799 * normalized nr_running number somewhere that negates
3800 * the hierarchy?
3801 */
3802 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3803 group->cpu_power;
3804
3805 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3806 sgs->group_imb = 1;
3807
3808 sgs->group_capacity =
3809 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3810}
3811
3812/**
3813 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3814 * @sd: sched_domain whose statistics are to be updated.
3815 * @this_cpu: Cpu for which load balance is currently performed.
3816 * @idle: Idle status of this_cpu
3817 * @sd_idle: Idle status of the sched_domain containing group.
3818 * @cpus: Set of cpus considered for load balancing.
3819 * @balance: Should we balance.
3820 * @sds: variable to hold the statistics for this sched_domain.
3821 */
3822static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3823 enum cpu_idle_type idle, int *sd_idle,
3824 const struct cpumask *cpus, int *balance,
3825 struct sd_lb_stats *sds)
3826{
3827 struct sched_domain *child = sd->child;
3828 struct sched_group *group = sd->groups;
3829 struct sg_lb_stats sgs;
3830 int load_idx, prefer_sibling = 0;
3831
3832 if (child && child->flags & SD_PREFER_SIBLING)
3833 prefer_sibling = 1;
3834
3835 init_sd_power_savings_stats(sd, sds, idle);
3836 load_idx = get_sd_load_idx(sd, idle);
3837
3838 do {
3839 int local_group;
3840
3841 local_group = cpumask_test_cpu(this_cpu,
3842 sched_group_cpus(group));
3843 memset(&sgs, 0, sizeof(sgs));
3844 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3845 local_group, cpus, balance, &sgs);
3846
3847 if (local_group && balance && !(*balance))
3848 return;
3849
3850 sds->total_load += sgs.group_load;
3851 sds->total_pwr += group->cpu_power;
3852
3853 /*
3854 * In case the child domain prefers tasks go to siblings
3855 * first, lower the group capacity to one so that we'll try
3856 * and move all the excess tasks away.
3857 */
3858 if (prefer_sibling)
3859 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3860
3861 if (local_group) {
3862 sds->this_load = sgs.avg_load;
3863 sds->this = group;
3864 sds->this_nr_running = sgs.sum_nr_running;
3865 sds->this_load_per_task = sgs.sum_weighted_load;
3866 } else if (sgs.avg_load > sds->max_load &&
3867 (sgs.sum_nr_running > sgs.group_capacity ||
3868 sgs.group_imb)) {
3869 sds->max_load = sgs.avg_load;
3870 sds->busiest = group;
3871 sds->busiest_nr_running = sgs.sum_nr_running;
3872 sds->busiest_load_per_task = sgs.sum_weighted_load;
3873 sds->group_imb = sgs.group_imb;
3874 }
3875
3876 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3877 group = group->next;
3878 } while (group != sd->groups);
3879}
3880
3881/**
3882 * fix_small_imbalance - Calculate the minor imbalance that exists
3883 * amongst the groups of a sched_domain, during
3884 * load balancing.
3885 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3886 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3887 * @imbalance: Variable to store the imbalance.
3888 */
3889static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3890 int this_cpu, unsigned long *imbalance)
3891{
3892 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3893 unsigned int imbn = 2;
3894
3895 if (sds->this_nr_running) {
3896 sds->this_load_per_task /= sds->this_nr_running;
3897 if (sds->busiest_load_per_task >
3898 sds->this_load_per_task)
3899 imbn = 1;
3900 } else
3901 sds->this_load_per_task =
3902 cpu_avg_load_per_task(this_cpu);
3903
3904 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3905 sds->busiest_load_per_task * imbn) {
3906 *imbalance = sds->busiest_load_per_task;
3907 return;
3908 }
3909
3910 /*
3911 * OK, we don't have enough imbalance to justify moving tasks,
3912 * however we may be able to increase total CPU power used by
3913 * moving them.
3914 */
3915
3916 pwr_now += sds->busiest->cpu_power *
3917 min(sds->busiest_load_per_task, sds->max_load);
3918 pwr_now += sds->this->cpu_power *
3919 min(sds->this_load_per_task, sds->this_load);
3920 pwr_now /= SCHED_LOAD_SCALE;
3921
3922 /* Amount of load we'd subtract */
3923 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3924 sds->busiest->cpu_power;
3925 if (sds->max_load > tmp)
3926 pwr_move += sds->busiest->cpu_power *
3927 min(sds->busiest_load_per_task, sds->max_load - tmp);
3928
3929 /* Amount of load we'd add */
3930 if (sds->max_load * sds->busiest->cpu_power <
3931 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3932 tmp = (sds->max_load * sds->busiest->cpu_power) /
3933 sds->this->cpu_power;
3934 else
3935 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3936 sds->this->cpu_power;
3937 pwr_move += sds->this->cpu_power *
3938 min(sds->this_load_per_task, sds->this_load + tmp);
3939 pwr_move /= SCHED_LOAD_SCALE;
3940
3941 /* Move if we gain throughput */
3942 if (pwr_move > pwr_now)
3943 *imbalance = sds->busiest_load_per_task;
3944}
3945
3946/**
3947 * calculate_imbalance - Calculate the amount of imbalance present within the
3948 * groups of a given sched_domain during load balance.
3949 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3950 * @this_cpu: Cpu for which currently load balance is being performed.
3951 * @imbalance: The variable to store the imbalance.
3952 */
3953static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3954 unsigned long *imbalance)
3955{
3956 unsigned long max_pull;
3957 /*
3958 * In the presence of smp nice balancing, certain scenarios can have
3959 * max load less than avg load(as we skip the groups at or below
3960 * its cpu_power, while calculating max_load..)
3961 */
3962 if (sds->max_load < sds->avg_load) {
3963 *imbalance = 0;
3964 return fix_small_imbalance(sds, this_cpu, imbalance);
3965 }
3966
3967 /* Don't want to pull so many tasks that a group would go idle */
3968 max_pull = min(sds->max_load - sds->avg_load,
3969 sds->max_load - sds->busiest_load_per_task);
3970
3971 /* How much load to actually move to equalise the imbalance */
3972 *imbalance = min(max_pull * sds->busiest->cpu_power,
3973 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3974 / SCHED_LOAD_SCALE;
3975
3976 /*
3977 * if *imbalance is less than the average load per runnable task
3978 * there is no gaurantee that any tasks will be moved so we'll have
3979 * a think about bumping its value to force at least one task to be
3980 * moved
3981 */
3982 if (*imbalance < sds->busiest_load_per_task)
3983 return fix_small_imbalance(sds, this_cpu, imbalance);
3984
3985}
3986/******* find_busiest_group() helpers end here *********************/
3987
3988/**
3989 * find_busiest_group - Returns the busiest group within the sched_domain
3990 * if there is an imbalance. If there isn't an imbalance, and
3991 * the user has opted for power-savings, it returns a group whose
3992 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3993 * such a group exists.
3994 *
3995 * Also calculates the amount of weighted load which should be moved
3996 * to restore balance.
3997 *
3998 * @sd: The sched_domain whose busiest group is to be returned.
3999 * @this_cpu: The cpu for which load balancing is currently being performed.
4000 * @imbalance: Variable which stores amount of weighted load which should
4001 * be moved to restore balance/put a group to idle.
4002 * @idle: The idle status of this_cpu.
4003 * @sd_idle: The idleness of sd
4004 * @cpus: The set of CPUs under consideration for load-balancing.
4005 * @balance: Pointer to a variable indicating if this_cpu
4006 * is the appropriate cpu to perform load balancing at this_level.
4007 *
4008 * Returns: - the busiest group if imbalance exists.
4009 * - If no imbalance and user has opted for power-savings balance,
4010 * return the least loaded group whose CPUs can be
4011 * put to idle by rebalancing its tasks onto our group.
4012 */
4013static struct sched_group *
4014find_busiest_group(struct sched_domain *sd, int this_cpu,
4015 unsigned long *imbalance, enum cpu_idle_type idle,
4016 int *sd_idle, const struct cpumask *cpus, int *balance)
4017{
4018 struct sd_lb_stats sds;
4019
4020 memset(&sds, 0, sizeof(sds));
4021
4022 /*
4023 * Compute the various statistics relavent for load balancing at
4024 * this level.
4025 */
4026 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4027 balance, &sds);
4028
4029 /* Cases where imbalance does not exist from POV of this_cpu */
4030 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4031 * at this level.
4032 * 2) There is no busy sibling group to pull from.
4033 * 3) This group is the busiest group.
4034 * 4) This group is more busy than the avg busieness at this
4035 * sched_domain.
4036 * 5) The imbalance is within the specified limit.
4037 * 6) Any rebalance would lead to ping-pong
4038 */
4039 if (balance && !(*balance))
4040 goto ret;
4041
4042 if (!sds.busiest || sds.busiest_nr_running == 0)
4043 goto out_balanced;
4044
4045 if (sds.this_load >= sds.max_load)
4046 goto out_balanced;
4047
4048 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4049
4050 if (sds.this_load >= sds.avg_load)
4051 goto out_balanced;
4052
4053 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4054 goto out_balanced;
4055
4056 sds.busiest_load_per_task /= sds.busiest_nr_running;
4057 if (sds.group_imb)
4058 sds.busiest_load_per_task =
4059 min(sds.busiest_load_per_task, sds.avg_load);
4060
4061 /*
4062 * We're trying to get all the cpus to the average_load, so we don't
4063 * want to push ourselves above the average load, nor do we wish to
4064 * reduce the max loaded cpu below the average load, as either of these
4065 * actions would just result in more rebalancing later, and ping-pong
4066 * tasks around. Thus we look for the minimum possible imbalance.
4067 * Negative imbalances (*we* are more loaded than anyone else) will
4068 * be counted as no imbalance for these purposes -- we can't fix that
4069 * by pulling tasks to us. Be careful of negative numbers as they'll
4070 * appear as very large values with unsigned longs.
4071 */
4072 if (sds.max_load <= sds.busiest_load_per_task)
4073 goto out_balanced;
4074
4075 /* Looks like there is an imbalance. Compute it */
4076 calculate_imbalance(&sds, this_cpu, imbalance);
4077 return sds.busiest;
4078
4079out_balanced:
4080 /*
4081 * There is no obvious imbalance. But check if we can do some balancing
4082 * to save power.
4083 */
4084 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4085 return sds.busiest;
4086ret:
4087 *imbalance = 0;
4088 return NULL;
4089}
4090
4091/*
4092 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4093 */
4094static struct rq *
4095find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4096 unsigned long imbalance, const struct cpumask *cpus)
4097{
4098 struct rq *busiest = NULL, *rq;
4099 unsigned long max_load = 0;
4100 int i;
4101
4102 for_each_cpu(i, sched_group_cpus(group)) {
4103 unsigned long power = power_of(i);
4104 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4105 unsigned long wl;
4106
4107 if (!cpumask_test_cpu(i, cpus))
4108 continue;
4109
4110 rq = cpu_rq(i);
4111 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4112 wl /= power;
4113
4114 if (capacity && rq->nr_running == 1 && wl > imbalance)
4115 continue;
4116
4117 if (wl > max_load) {
4118 max_load = wl;
4119 busiest = rq;
4120 }
4121 }
4122
4123 return busiest;
4124}
4125
4126/*
4127 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4128 * so long as it is large enough.
4129 */
4130#define MAX_PINNED_INTERVAL 512
4131
4132/* Working cpumask for load_balance and load_balance_newidle. */
4133static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4134
4135/*
4136 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4137 * tasks if there is an imbalance.
4138 */
4139static int load_balance(int this_cpu, struct rq *this_rq,
4140 struct sched_domain *sd, enum cpu_idle_type idle,
4141 int *balance)
4142{
4143 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4144 struct sched_group *group;
4145 unsigned long imbalance;
4146 struct rq *busiest;
4147 unsigned long flags;
4148 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4149
4150 cpumask_copy(cpus, cpu_active_mask);
4151
4152 /*
4153 * When power savings policy is enabled for the parent domain, idle
4154 * sibling can pick up load irrespective of busy siblings. In this case,
4155 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4156 * portraying it as CPU_NOT_IDLE.
4157 */
4158 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4159 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4160 sd_idle = 1;
4161
4162 schedstat_inc(sd, lb_count[idle]);
4163
4164redo:
4165 update_shares(sd);
4166 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4167 cpus, balance);
4168
4169 if (*balance == 0)
4170 goto out_balanced;
4171
4172 if (!group) {
4173 schedstat_inc(sd, lb_nobusyg[idle]);
4174 goto out_balanced;
4175 }
4176
4177 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4178 if (!busiest) {
4179 schedstat_inc(sd, lb_nobusyq[idle]);
4180 goto out_balanced;
4181 }
4182
4183 BUG_ON(busiest == this_rq);
4184
4185 schedstat_add(sd, lb_imbalance[idle], imbalance);
4186
4187 ld_moved = 0;
4188 if (busiest->nr_running > 1) {
4189 /*
4190 * Attempt to move tasks. If find_busiest_group has found
4191 * an imbalance but busiest->nr_running <= 1, the group is
4192 * still unbalanced. ld_moved simply stays zero, so it is
4193 * correctly treated as an imbalance.
4194 */
4195 local_irq_save(flags);
4196 double_rq_lock(this_rq, busiest);
4197 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4198 imbalance, sd, idle, &all_pinned);
4199 double_rq_unlock(this_rq, busiest);
4200 local_irq_restore(flags);
4201
4202 /*
4203 * some other cpu did the load balance for us.
4204 */
4205 if (ld_moved && this_cpu != smp_processor_id())
4206 resched_cpu(this_cpu);
4207
4208 /* All tasks on this runqueue were pinned by CPU affinity */
4209 if (unlikely(all_pinned)) {
4210 cpumask_clear_cpu(cpu_of(busiest), cpus);
4211 if (!cpumask_empty(cpus))
4212 goto redo;
4213 goto out_balanced;
4214 }
4215 }
4216
4217 if (!ld_moved) {
4218 schedstat_inc(sd, lb_failed[idle]);
4219 sd->nr_balance_failed++;
4220
4221 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4222
4223 raw_spin_lock_irqsave(&busiest->lock, flags);
4224
4225 /* don't kick the migration_thread, if the curr
4226 * task on busiest cpu can't be moved to this_cpu
4227 */
4228 if (!cpumask_test_cpu(this_cpu,
4229 &busiest->curr->cpus_allowed)) {
4230 raw_spin_unlock_irqrestore(&busiest->lock,
4231 flags);
4232 all_pinned = 1;
4233 goto out_one_pinned;
4234 }
4235
4236 if (!busiest->active_balance) {
4237 busiest->active_balance = 1;
4238 busiest->push_cpu = this_cpu;
4239 active_balance = 1;
4240 }
4241 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4242 if (active_balance)
4243 wake_up_process(busiest->migration_thread);
4244
4245 /*
4246 * We've kicked active balancing, reset the failure
4247 * counter.
4248 */
4249 sd->nr_balance_failed = sd->cache_nice_tries+1;
4250 }
4251 } else
4252 sd->nr_balance_failed = 0;
4253
4254 if (likely(!active_balance)) {
4255 /* We were unbalanced, so reset the balancing interval */
4256 sd->balance_interval = sd->min_interval;
4257 } else {
4258 /*
4259 * If we've begun active balancing, start to back off. This
4260 * case may not be covered by the all_pinned logic if there
4261 * is only 1 task on the busy runqueue (because we don't call
4262 * move_tasks).
4263 */
4264 if (sd->balance_interval < sd->max_interval)
4265 sd->balance_interval *= 2;
4266 }
4267
4268 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4269 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4270 ld_moved = -1;
4271
4272 goto out;
4273
4274out_balanced:
4275 schedstat_inc(sd, lb_balanced[idle]);
4276
4277 sd->nr_balance_failed = 0;
4278
4279out_one_pinned:
4280 /* tune up the balancing interval */
4281 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4282 (sd->balance_interval < sd->max_interval))
4283 sd->balance_interval *= 2;
4284
4285 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4286 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4287 ld_moved = -1;
4288 else
4289 ld_moved = 0;
4290out:
4291 if (ld_moved)
4292 update_shares(sd);
4293 return ld_moved;
4294}
4295
4296/*
4297 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4298 * tasks if there is an imbalance.
4299 *
4300 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4301 * this_rq is locked.
4302 */
4303static int
4304load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4305{
4306 struct sched_group *group;
4307 struct rq *busiest = NULL;
4308 unsigned long imbalance;
4309 int ld_moved = 0;
4310 int sd_idle = 0;
4311 int all_pinned = 0;
4312 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4313
4314 cpumask_copy(cpus, cpu_active_mask);
4315
4316 /*
4317 * When power savings policy is enabled for the parent domain, idle
4318 * sibling can pick up load irrespective of busy siblings. In this case,
4319 * let the state of idle sibling percolate up as IDLE, instead of
4320 * portraying it as CPU_NOT_IDLE.
4321 */
4322 if (sd->flags & SD_SHARE_CPUPOWER &&
4323 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4324 sd_idle = 1;
4325
4326 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4327redo:
4328 update_shares_locked(this_rq, sd);
4329 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4330 &sd_idle, cpus, NULL);
4331 if (!group) {
4332 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4333 goto out_balanced;
4334 }
4335
4336 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4337 if (!busiest) {
4338 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4339 goto out_balanced;
4340 }
4341
4342 BUG_ON(busiest == this_rq);
4343
4344 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4345
4346 ld_moved = 0;
4347 if (busiest->nr_running > 1) {
4348 /* Attempt to move tasks */
4349 double_lock_balance(this_rq, busiest);
4350 /* this_rq->clock is already updated */
4351 update_rq_clock(busiest);
4352 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4353 imbalance, sd, CPU_NEWLY_IDLE,
4354 &all_pinned);
4355 double_unlock_balance(this_rq, busiest);
4356
4357 if (unlikely(all_pinned)) {
4358 cpumask_clear_cpu(cpu_of(busiest), cpus);
4359 if (!cpumask_empty(cpus))
4360 goto redo;
4361 }
4362 }
4363
4364 if (!ld_moved) {
4365 int active_balance = 0;
4366
4367 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4368 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4369 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4370 return -1;
4371
4372 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4373 return -1;
4374
4375 if (sd->nr_balance_failed++ < 2)
4376 return -1;
4377
4378 /*
4379 * The only task running in a non-idle cpu can be moved to this
4380 * cpu in an attempt to completely freeup the other CPU
4381 * package. The same method used to move task in load_balance()
4382 * have been extended for load_balance_newidle() to speedup
4383 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4384 *
4385 * The package power saving logic comes from
4386 * find_busiest_group(). If there are no imbalance, then
4387 * f_b_g() will return NULL. However when sched_mc={1,2} then
4388 * f_b_g() will select a group from which a running task may be
4389 * pulled to this cpu in order to make the other package idle.
4390 * If there is no opportunity to make a package idle and if
4391 * there are no imbalance, then f_b_g() will return NULL and no
4392 * action will be taken in load_balance_newidle().
4393 *
4394 * Under normal task pull operation due to imbalance, there
4395 * will be more than one task in the source run queue and
4396 * move_tasks() will succeed. ld_moved will be true and this
4397 * active balance code will not be triggered.
4398 */
4399
4400 /* Lock busiest in correct order while this_rq is held */
4401 double_lock_balance(this_rq, busiest);
4402
4403 /*
4404 * don't kick the migration_thread, if the curr
4405 * task on busiest cpu can't be moved to this_cpu
4406 */
4407 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4408 double_unlock_balance(this_rq, busiest);
4409 all_pinned = 1;
4410 return ld_moved;
4411 }
4412
4413 if (!busiest->active_balance) {
4414 busiest->active_balance = 1;
4415 busiest->push_cpu = this_cpu;
4416 active_balance = 1;
4417 }
4418
4419 double_unlock_balance(this_rq, busiest);
4420 /*
4421 * Should not call ttwu while holding a rq->lock
4422 */
4423 raw_spin_unlock(&this_rq->lock);
4424 if (active_balance)
4425 wake_up_process(busiest->migration_thread);
4426 raw_spin_lock(&this_rq->lock);
4427
4428 } else
4429 sd->nr_balance_failed = 0;
4430
4431 update_shares_locked(this_rq, sd);
4432 return ld_moved;
4433
4434out_balanced:
4435 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4436 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4437 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4438 return -1;
4439 sd->nr_balance_failed = 0;
4440
4441 return 0;
4442}
4443
4444/*
4445 * idle_balance is called by schedule() if this_cpu is about to become
4446 * idle. Attempts to pull tasks from other CPUs.
4447 */
4448static void idle_balance(int this_cpu, struct rq *this_rq)
4449{
4450 struct sched_domain *sd;
4451 int pulled_task = 0;
4452 unsigned long next_balance = jiffies + HZ;
4453
4454 this_rq->idle_stamp = this_rq->clock;
4455
4456 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4457 return;
4458
4459 for_each_domain(this_cpu, sd) {
4460 unsigned long interval;
4461
4462 if (!(sd->flags & SD_LOAD_BALANCE))
4463 continue;
4464
4465 if (sd->flags & SD_BALANCE_NEWIDLE)
4466 /* If we've pulled tasks over stop searching: */
4467 pulled_task = load_balance_newidle(this_cpu, this_rq,
4468 sd);
4469
4470 interval = msecs_to_jiffies(sd->balance_interval);
4471 if (time_after(next_balance, sd->last_balance + interval))
4472 next_balance = sd->last_balance + interval;
4473 if (pulled_task) {
4474 this_rq->idle_stamp = 0;
4475 break;
4476 }
4477 }
4478 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4479 /*
4480 * We are going idle. next_balance may be set based on
4481 * a busy processor. So reset next_balance.
4482 */
4483 this_rq->next_balance = next_balance;
4484 }
4485}
4486
4487/*
4488 * active_load_balance is run by migration threads. It pushes running tasks
4489 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4490 * running on each physical CPU where possible, and avoids physical /
4491 * logical imbalances.
4492 *
4493 * Called with busiest_rq locked.
4494 */
4495static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4496{
4497 int target_cpu = busiest_rq->push_cpu;
4498 struct sched_domain *sd;
4499 struct rq *target_rq;
4500
4501 /* Is there any task to move? */
4502 if (busiest_rq->nr_running <= 1)
4503 return;
4504
4505 target_rq = cpu_rq(target_cpu);
4506
4507 /*
4508 * This condition is "impossible", if it occurs
4509 * we need to fix it. Originally reported by
4510 * Bjorn Helgaas on a 128-cpu setup.
4511 */
4512 BUG_ON(busiest_rq == target_rq);
4513
4514 /* move a task from busiest_rq to target_rq */
4515 double_lock_balance(busiest_rq, target_rq);
4516 update_rq_clock(busiest_rq);
4517 update_rq_clock(target_rq);
4518
4519 /* Search for an sd spanning us and the target CPU. */
4520 for_each_domain(target_cpu, sd) {
4521 if ((sd->flags & SD_LOAD_BALANCE) &&
4522 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4523 break;
4524 }
4525
4526 if (likely(sd)) {
4527 schedstat_inc(sd, alb_count);
4528
4529 if (move_one_task(target_rq, target_cpu, busiest_rq,
4530 sd, CPU_IDLE))
4531 schedstat_inc(sd, alb_pushed);
4532 else
4533 schedstat_inc(sd, alb_failed);
4534 }
4535 double_unlock_balance(busiest_rq, target_rq);
4536}
4537
4538#ifdef CONFIG_NO_HZ
4539static struct {
4540 atomic_t load_balancer;
4541 cpumask_var_t cpu_mask;
4542 cpumask_var_t ilb_grp_nohz_mask;
4543} nohz ____cacheline_aligned = {
4544 .load_balancer = ATOMIC_INIT(-1),
4545};
4546
4547int get_nohz_load_balancer(void)
4548{
4549 return atomic_read(&nohz.load_balancer);
4550}
4551
4552#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4553/**
4554 * lowest_flag_domain - Return lowest sched_domain containing flag.
4555 * @cpu: The cpu whose lowest level of sched domain is to
4556 * be returned.
4557 * @flag: The flag to check for the lowest sched_domain
4558 * for the given cpu.
4559 *
4560 * Returns the lowest sched_domain of a cpu which contains the given flag.
4561 */
4562static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4563{
4564 struct sched_domain *sd;
4565
4566 for_each_domain(cpu, sd)
4567 if (sd && (sd->flags & flag))
4568 break;
4569
4570 return sd;
4571}
4572
4573/**
4574 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4575 * @cpu: The cpu whose domains we're iterating over.
4576 * @sd: variable holding the value of the power_savings_sd
4577 * for cpu.
4578 * @flag: The flag to filter the sched_domains to be iterated.
4579 *
4580 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4581 * set, starting from the lowest sched_domain to the highest.
4582 */
4583#define for_each_flag_domain(cpu, sd, flag) \
4584 for (sd = lowest_flag_domain(cpu, flag); \
4585 (sd && (sd->flags & flag)); sd = sd->parent)
4586
4587/**
4588 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4589 * @ilb_group: group to be checked for semi-idleness
4590 *
4591 * Returns: 1 if the group is semi-idle. 0 otherwise.
4592 *
4593 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4594 * and atleast one non-idle CPU. This helper function checks if the given
4595 * sched_group is semi-idle or not.
4596 */
4597static inline int is_semi_idle_group(struct sched_group *ilb_group)
4598{
4599 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4600 sched_group_cpus(ilb_group));
4601
4602 /*
4603 * A sched_group is semi-idle when it has atleast one busy cpu
4604 * and atleast one idle cpu.
4605 */
4606 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4607 return 0;
4608
4609 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4610 return 0;
4611
4612 return 1;
4613}
4614/**
4615 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4616 * @cpu: The cpu which is nominating a new idle_load_balancer.
4617 *
4618 * Returns: Returns the id of the idle load balancer if it exists,
4619 * Else, returns >= nr_cpu_ids.
4620 *
4621 * This algorithm picks the idle load balancer such that it belongs to a
4622 * semi-idle powersavings sched_domain. The idea is to try and avoid
4623 * completely idle packages/cores just for the purpose of idle load balancing
4624 * when there are other idle cpu's which are better suited for that job.
4625 */
4626static int find_new_ilb(int cpu)
4627{
4628 struct sched_domain *sd;
4629 struct sched_group *ilb_group;
4630
4631 /*
4632 * Have idle load balancer selection from semi-idle packages only
4633 * when power-aware load balancing is enabled
4634 */
4635 if (!(sched_smt_power_savings || sched_mc_power_savings))
4636 goto out_done;
4637
4638 /*
4639 * Optimize for the case when we have no idle CPUs or only one
4640 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4641 */
4642 if (cpumask_weight(nohz.cpu_mask) < 2)
4643 goto out_done;
4644
4645 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4646 ilb_group = sd->groups;
4647
4648 do {
4649 if (is_semi_idle_group(ilb_group))
4650 return cpumask_first(nohz.ilb_grp_nohz_mask);
4651
4652 ilb_group = ilb_group->next;
4653
4654 } while (ilb_group != sd->groups);
4655 }
4656
4657out_done:
4658 return cpumask_first(nohz.cpu_mask);
4659}
4660#else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4661static inline int find_new_ilb(int call_cpu)
4662{
4663 return cpumask_first(nohz.cpu_mask);
4664}
4665#endif
4666
4667/*
4668 * This routine will try to nominate the ilb (idle load balancing)
4669 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4670 * load balancing on behalf of all those cpus. If all the cpus in the system
4671 * go into this tickless mode, then there will be no ilb owner (as there is
4672 * no need for one) and all the cpus will sleep till the next wakeup event
4673 * arrives...
4674 *
4675 * For the ilb owner, tick is not stopped. And this tick will be used
4676 * for idle load balancing. ilb owner will still be part of
4677 * nohz.cpu_mask..
4678 *
4679 * While stopping the tick, this cpu will become the ilb owner if there
4680 * is no other owner. And will be the owner till that cpu becomes busy
4681 * or if all cpus in the system stop their ticks at which point
4682 * there is no need for ilb owner.
4683 *
4684 * When the ilb owner becomes busy, it nominates another owner, during the
4685 * next busy scheduler_tick()
4686 */
4687int select_nohz_load_balancer(int stop_tick)
4688{
4689 int cpu = smp_processor_id();
4690
4691 if (stop_tick) {
4692 cpu_rq(cpu)->in_nohz_recently = 1;
4693
4694 if (!cpu_active(cpu)) {
4695 if (atomic_read(&nohz.load_balancer) != cpu)
4696 return 0;
4697
4698 /*
4699 * If we are going offline and still the leader,
4700 * give up!
4701 */
4702 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4703 BUG();
4704
4705 return 0;
4706 }
4707
4708 cpumask_set_cpu(cpu, nohz.cpu_mask);
4709
4710 /* time for ilb owner also to sleep */
4711 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4712 if (atomic_read(&nohz.load_balancer) == cpu)
4713 atomic_set(&nohz.load_balancer, -1);
4714 return 0;
4715 }
4716
4717 if (atomic_read(&nohz.load_balancer) == -1) {
4718 /* make me the ilb owner */
4719 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4720 return 1;
4721 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4722 int new_ilb;
4723
4724 if (!(sched_smt_power_savings ||
4725 sched_mc_power_savings))
4726 return 1;
4727 /*
4728 * Check to see if there is a more power-efficient
4729 * ilb.
4730 */
4731 new_ilb = find_new_ilb(cpu);
4732 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4733 atomic_set(&nohz.load_balancer, -1);
4734 resched_cpu(new_ilb);
4735 return 0;
4736 }
4737 return 1;
4738 }
4739 } else {
4740 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4741 return 0;
4742
4743 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4744
4745 if (atomic_read(&nohz.load_balancer) == cpu)
4746 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4747 BUG();
4748 }
4749 return 0;
4750}
4751#endif
4752
4753static DEFINE_SPINLOCK(balancing);
4754
4755/*
4756 * It checks each scheduling domain to see if it is due to be balanced,
4757 * and initiates a balancing operation if so.
4758 *
4759 * Balancing parameters are set up in arch_init_sched_domains.
4760 */
4761static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4762{
4763 int balance = 1;
4764 struct rq *rq = cpu_rq(cpu);
4765 unsigned long interval;
4766 struct sched_domain *sd;
4767 /* Earliest time when we have to do rebalance again */
4768 unsigned long next_balance = jiffies + 60*HZ;
4769 int update_next_balance = 0;
4770 int need_serialize;
4771
4772 for_each_domain(cpu, sd) {
4773 if (!(sd->flags & SD_LOAD_BALANCE))
4774 continue;
4775
4776 interval = sd->balance_interval;
4777 if (idle != CPU_IDLE)
4778 interval *= sd->busy_factor;
4779
4780 /* scale ms to jiffies */
4781 interval = msecs_to_jiffies(interval);
4782 if (unlikely(!interval))
4783 interval = 1;
4784 if (interval > HZ*NR_CPUS/10)
4785 interval = HZ*NR_CPUS/10;
4786
4787 need_serialize = sd->flags & SD_SERIALIZE;
4788
4789 if (need_serialize) {
4790 if (!spin_trylock(&balancing))
4791 goto out;
4792 }
4793
4794 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4795 if (load_balance(cpu, rq, sd, idle, &balance)) {
4796 /*
4797 * We've pulled tasks over so either we're no
4798 * longer idle, or one of our SMT siblings is
4799 * not idle.
4800 */
4801 idle = CPU_NOT_IDLE;
4802 }
4803 sd->last_balance = jiffies;
4804 }
4805 if (need_serialize)
4806 spin_unlock(&balancing);
4807out:
4808 if (time_after(next_balance, sd->last_balance + interval)) {
4809 next_balance = sd->last_balance + interval;
4810 update_next_balance = 1;
4811 }
4812
4813 /*
4814 * Stop the load balance at this level. There is another
4815 * CPU in our sched group which is doing load balancing more
4816 * actively.
4817 */
4818 if (!balance)
4819 break;
4820 }
4821
4822 /*
4823 * next_balance will be updated only when there is a need.
4824 * When the cpu is attached to null domain for ex, it will not be
4825 * updated.
4826 */
4827 if (likely(update_next_balance))
4828 rq->next_balance = next_balance;
4829}
4830
4831/*
4832 * run_rebalance_domains is triggered when needed from the scheduler tick.
4833 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4834 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4835 */
4836static void run_rebalance_domains(struct softirq_action *h)
4837{
4838 int this_cpu = smp_processor_id();
4839 struct rq *this_rq = cpu_rq(this_cpu);
4840 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4841 CPU_IDLE : CPU_NOT_IDLE;
4842
4843 rebalance_domains(this_cpu, idle);
4844
4845#ifdef CONFIG_NO_HZ
4846 /*
4847 * If this cpu is the owner for idle load balancing, then do the
4848 * balancing on behalf of the other idle cpus whose ticks are
4849 * stopped.
4850 */
4851 if (this_rq->idle_at_tick &&
4852 atomic_read(&nohz.load_balancer) == this_cpu) {
4853 struct rq *rq;
4854 int balance_cpu;
4855
4856 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4857 if (balance_cpu == this_cpu)
4858 continue;
4859
4860 /*
4861 * If this cpu gets work to do, stop the load balancing
4862 * work being done for other cpus. Next load
4863 * balancing owner will pick it up.
4864 */
4865 if (need_resched())
4866 break;
4867
4868 rebalance_domains(balance_cpu, CPU_IDLE);
4869
4870 rq = cpu_rq(balance_cpu);
4871 if (time_after(this_rq->next_balance, rq->next_balance))
4872 this_rq->next_balance = rq->next_balance;
4873 }
4874 }
4875#endif
4876}
4877
4878static inline int on_null_domain(int cpu)
4879{
4880 return !rcu_dereference(cpu_rq(cpu)->sd);
4881}
4882
4883/*
4884 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4885 *
4886 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4887 * idle load balancing owner or decide to stop the periodic load balancing,
4888 * if the whole system is idle.
4889 */
4890static inline void trigger_load_balance(struct rq *rq, int cpu)
4891{
4892#ifdef CONFIG_NO_HZ
4893 /*
4894 * If we were in the nohz mode recently and busy at the current
4895 * scheduler tick, then check if we need to nominate new idle
4896 * load balancer.
4897 */
4898 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4899 rq->in_nohz_recently = 0;
4900
4901 if (atomic_read(&nohz.load_balancer) == cpu) {
4902 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4903 atomic_set(&nohz.load_balancer, -1);
4904 }
4905
4906 if (atomic_read(&nohz.load_balancer) == -1) {
4907 int ilb = find_new_ilb(cpu);
4908
4909 if (ilb < nr_cpu_ids)
4910 resched_cpu(ilb);
4911 }
4912 }
4913
4914 /*
4915 * If this cpu is idle and doing idle load balancing for all the
4916 * cpus with ticks stopped, is it time for that to stop?
4917 */
4918 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4919 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4920 resched_cpu(cpu);
4921 return;
4922 }
4923
4924 /*
4925 * If this cpu is idle and the idle load balancing is done by
4926 * someone else, then no need raise the SCHED_SOFTIRQ
4927 */
4928 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4929 cpumask_test_cpu(cpu, nohz.cpu_mask))
4930 return;
4931#endif
4932 /* Don't need to rebalance while attached to NULL domain */
4933 if (time_after_eq(jiffies, rq->next_balance) &&
4934 likely(!on_null_domain(cpu)))
4935 raise_softirq(SCHED_SOFTIRQ);
4936}
4937
4938#else /* CONFIG_SMP */
4939
4940/*
4941 * on UP we do not need to balance between CPUs:
4942 */
4943static inline void idle_balance(int cpu, struct rq *rq)
4944{
4945}
4946
4947#endif 3186#endif
4948 3187
4949DEFINE_PER_CPU(struct kernel_stat, kstat); 3188DEFINE_PER_CPU(struct kernel_stat, kstat);
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-30 12:35:07 -0500