aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/sched.c
diff options
context:
space:
mode:
Diffstat (limited to 'kernel/sched.c')
-rw-r--r--kernel/sched.c3170
1 files changed, 732 insertions, 2438 deletions
diff --git a/kernel/sched.c b/kernel/sched.c
index adb5e923cc61..5e3c509e0efe 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -71,6 +71,7 @@
71#include <linux/debugfs.h> 71#include <linux/debugfs.h>
72#include <linux/ctype.h> 72#include <linux/ctype.h>
73#include <linux/ftrace.h> 73#include <linux/ftrace.h>
74#include <linux/slab.h>
74 75
75#include <asm/tlb.h> 76#include <asm/tlb.h>
76#include <asm/irq_regs.h> 77#include <asm/irq_regs.h>
@@ -144,7 +145,7 @@ struct rt_prio_array {
144 145
145struct rt_bandwidth { 146struct rt_bandwidth {
146 /* nests inside the rq lock: */ 147 /* nests inside the rq lock: */
147 spinlock_t rt_runtime_lock; 148 raw_spinlock_t rt_runtime_lock;
148 ktime_t rt_period; 149 ktime_t rt_period;
149 u64 rt_runtime; 150 u64 rt_runtime;
150 struct hrtimer rt_period_timer; 151 struct hrtimer rt_period_timer;
@@ -181,7 +182,7 @@ void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
181 rt_b->rt_period = ns_to_ktime(period); 182 rt_b->rt_period = ns_to_ktime(period);
182 rt_b->rt_runtime = runtime; 183 rt_b->rt_runtime = runtime;
183 184
184 spin_lock_init(&rt_b->rt_runtime_lock); 185 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 186
186 hrtimer_init(&rt_b->rt_period_timer, 187 hrtimer_init(&rt_b->rt_period_timer,
187 CLOCK_MONOTONIC, HRTIMER_MODE_REL); 188 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
@@ -203,7 +204,7 @@ static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
203 if (hrtimer_active(&rt_b->rt_period_timer)) 204 if (hrtimer_active(&rt_b->rt_period_timer))
204 return; 205 return;
205 206
206 spin_lock(&rt_b->rt_runtime_lock); 207 raw_spin_lock(&rt_b->rt_runtime_lock);
207 for (;;) { 208 for (;;) {
208 unsigned long delta; 209 unsigned long delta;
209 ktime_t soft, hard; 210 ktime_t soft, hard;
@@ -220,7 +221,7 @@ static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
220 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta, 221 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
221 HRTIMER_MODE_ABS_PINNED, 0); 222 HRTIMER_MODE_ABS_PINNED, 0);
222 } 223 }
223 spin_unlock(&rt_b->rt_runtime_lock); 224 raw_spin_unlock(&rt_b->rt_runtime_lock);
224} 225}
225 226
226#ifdef CONFIG_RT_GROUP_SCHED 227#ifdef CONFIG_RT_GROUP_SCHED
@@ -236,7 +237,7 @@ static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
236 */ 237 */
237static DEFINE_MUTEX(sched_domains_mutex); 238static DEFINE_MUTEX(sched_domains_mutex);
238 239
239#ifdef CONFIG_GROUP_SCHED 240#ifdef CONFIG_CGROUP_SCHED
240 241
241#include <linux/cgroup.h> 242#include <linux/cgroup.h>
242 243
@@ -246,13 +247,7 @@ static LIST_HEAD(task_groups);
246 247
247/* task group related information */ 248/* task group related information */
248struct task_group { 249struct task_group {
249#ifdef CONFIG_CGROUP_SCHED
250 struct cgroup_subsys_state css; 250 struct cgroup_subsys_state css;
251#endif
252
253#ifdef CONFIG_USER_SCHED
254 uid_t uid;
255#endif
256 251
257#ifdef CONFIG_FAIR_GROUP_SCHED 252#ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */ 253 /* schedulable entities of this group on each cpu */
@@ -277,35 +272,7 @@ struct task_group {
277 struct list_head children; 272 struct list_head children;
278}; 273};
279 274
280#ifdef CONFIG_USER_SCHED
281
282/* Helper function to pass uid information to create_sched_user() */
283void set_tg_uid(struct user_struct *user)
284{
285 user->tg->uid = user->uid;
286}
287
288/*
289 * Root task group.
290 * Every UID task group (including init_task_group aka UID-0) will
291 * be a child to this group.
292 */
293struct task_group root_task_group;
294
295#ifdef CONFIG_FAIR_GROUP_SCHED
296/* Default task group's sched entity on each cpu */
297static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
298/* Default task group's cfs_rq on each cpu */
299static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
300#endif /* CONFIG_FAIR_GROUP_SCHED */
301
302#ifdef CONFIG_RT_GROUP_SCHED
303static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
304static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
305#endif /* CONFIG_RT_GROUP_SCHED */
306#else /* !CONFIG_USER_SCHED */
307#define root_task_group init_task_group 275#define root_task_group init_task_group
308#endif /* CONFIG_USER_SCHED */
309 276
310/* task_group_lock serializes add/remove of task groups and also changes to 277/* task_group_lock serializes add/remove of task groups and also changes to
311 * a task group's cpu shares. 278 * a task group's cpu shares.
@@ -321,11 +288,7 @@ static int root_task_group_empty(void)
321} 288}
322#endif 289#endif
323 290
324#ifdef CONFIG_USER_SCHED
325# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
326#else /* !CONFIG_USER_SCHED */
327# define INIT_TASK_GROUP_LOAD NICE_0_LOAD 291# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
328#endif /* CONFIG_USER_SCHED */
329 292
330/* 293/*
331 * A weight of 0 or 1 can cause arithmetics problems. 294 * A weight of 0 or 1 can cause arithmetics problems.
@@ -351,11 +314,7 @@ static inline struct task_group *task_group(struct task_struct *p)
351{ 314{
352 struct task_group *tg; 315 struct task_group *tg;
353 316
354#ifdef CONFIG_USER_SCHED 317#ifdef CONFIG_CGROUP_SCHED
355 rcu_read_lock();
356 tg = __task_cred(p)->user->tg;
357 rcu_read_unlock();
358#elif defined(CONFIG_CGROUP_SCHED)
359 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id), 318 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
360 struct task_group, css); 319 struct task_group, css);
361#else 320#else
@@ -367,6 +326,15 @@ static inline struct task_group *task_group(struct task_struct *p)
367/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ 326/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
368static inline void set_task_rq(struct task_struct *p, unsigned int cpu) 327static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
369{ 328{
329 /*
330 * Strictly speaking this rcu_read_lock() is not needed since the
331 * task_group is tied to the cgroup, which in turn can never go away
332 * as long as there are tasks attached to it.
333 *
334 * However since task_group() uses task_subsys_state() which is an
335 * rcu_dereference() user, this quiets CONFIG_PROVE_RCU.
336 */
337 rcu_read_lock();
370#ifdef CONFIG_FAIR_GROUP_SCHED 338#ifdef CONFIG_FAIR_GROUP_SCHED
371 p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; 339 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
372 p->se.parent = task_group(p)->se[cpu]; 340 p->se.parent = task_group(p)->se[cpu];
@@ -376,6 +344,7 @@ static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
376 p->rt.rt_rq = task_group(p)->rt_rq[cpu]; 344 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
377 p->rt.parent = task_group(p)->rt_se[cpu]; 345 p->rt.parent = task_group(p)->rt_se[cpu];
378#endif 346#endif
347 rcu_read_unlock();
379} 348}
380 349
381#else 350#else
@@ -386,7 +355,7 @@ static inline struct task_group *task_group(struct task_struct *p)
386 return NULL; 355 return NULL;
387} 356}
388 357
389#endif /* CONFIG_GROUP_SCHED */ 358#endif /* CONFIG_CGROUP_SCHED */
390 359
391/* CFS-related fields in a runqueue */ 360/* CFS-related fields in a runqueue */
392struct cfs_rq { 361struct cfs_rq {
@@ -473,7 +442,7 @@ struct rt_rq {
473 u64 rt_time; 442 u64 rt_time;
474 u64 rt_runtime; 443 u64 rt_runtime;
475 /* Nests inside the rq lock: */ 444 /* Nests inside the rq lock: */
476 spinlock_t rt_runtime_lock; 445 raw_spinlock_t rt_runtime_lock;
477 446
478#ifdef CONFIG_RT_GROUP_SCHED 447#ifdef CONFIG_RT_GROUP_SCHED
479 unsigned long rt_nr_boosted; 448 unsigned long rt_nr_boosted;
@@ -481,7 +450,6 @@ struct rt_rq {
481 struct rq *rq; 450 struct rq *rq;
482 struct list_head leaf_rt_rq_list; 451 struct list_head leaf_rt_rq_list;
483 struct task_group *tg; 452 struct task_group *tg;
484 struct sched_rt_entity *rt_se;
485#endif 453#endif
486}; 454};
487 455
@@ -534,7 +502,7 @@ static struct root_domain def_root_domain;
534 */ 502 */
535struct rq { 503struct rq {
536 /* runqueue lock: */ 504 /* runqueue lock: */
537 spinlock_t lock; 505 raw_spinlock_t lock;
538 506
539 /* 507 /*
540 * nr_running and cpu_load should be in the same cacheline because 508 * nr_running and cpu_load should be in the same cacheline because
@@ -544,14 +512,12 @@ struct rq {
544 #define CPU_LOAD_IDX_MAX 5 512 #define CPU_LOAD_IDX_MAX 5
545 unsigned long cpu_load[CPU_LOAD_IDX_MAX]; 513 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
546#ifdef CONFIG_NO_HZ 514#ifdef CONFIG_NO_HZ
547 unsigned long last_tick_seen;
548 unsigned char in_nohz_recently; 515 unsigned char in_nohz_recently;
549#endif 516#endif
550 /* capture load from *all* tasks on this cpu: */ 517 /* capture load from *all* tasks on this cpu: */
551 struct load_weight load; 518 struct load_weight load;
552 unsigned long nr_load_updates; 519 unsigned long nr_load_updates;
553 u64 nr_switches; 520 u64 nr_switches;
554 u64 nr_migrations_in;
555 521
556 struct cfs_rq cfs; 522 struct cfs_rq cfs;
557 struct rt_rq rt; 523 struct rt_rq rt;
@@ -601,6 +567,8 @@ struct rq {
601 567
602 u64 rt_avg; 568 u64 rt_avg;
603 u64 age_stamp; 569 u64 age_stamp;
570 u64 idle_stamp;
571 u64 avg_idle;
604#endif 572#endif
605 573
606 /* calc_load related fields */ 574 /* calc_load related fields */
@@ -655,6 +623,11 @@ static inline int cpu_of(struct rq *rq)
655#endif 623#endif
656} 624}
657 625
626#define rcu_dereference_check_sched_domain(p) \
627 rcu_dereference_check((p), \
628 rcu_read_lock_sched_held() || \
629 lockdep_is_held(&sched_domains_mutex))
630
658/* 631/*
659 * The domain tree (rq->sd) is protected by RCU's quiescent state transition. 632 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
660 * See detach_destroy_domains: synchronize_sched for details. 633 * See detach_destroy_domains: synchronize_sched for details.
@@ -663,7 +636,7 @@ static inline int cpu_of(struct rq *rq)
663 * preempt-disabled sections. 636 * preempt-disabled sections.
664 */ 637 */
665#define for_each_domain(cpu, __sd) \ 638#define for_each_domain(cpu, __sd) \
666 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) 639 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
667 640
668#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) 641#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
669#define this_rq() (&__get_cpu_var(runqueues)) 642#define this_rq() (&__get_cpu_var(runqueues))
@@ -695,7 +668,7 @@ inline void update_rq_clock(struct rq *rq)
695 */ 668 */
696int runqueue_is_locked(int cpu) 669int runqueue_is_locked(int cpu)
697{ 670{
698 return spin_is_locked(&cpu_rq(cpu)->lock); 671 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
699} 672}
700 673
701/* 674/*
@@ -782,7 +755,7 @@ sched_feat_write(struct file *filp, const char __user *ubuf,
782 if (!sched_feat_names[i]) 755 if (!sched_feat_names[i])
783 return -EINVAL; 756 return -EINVAL;
784 757
785 filp->f_pos += cnt; 758 *ppos += cnt;
786 759
787 return cnt; 760 return cnt;
788} 761}
@@ -824,6 +797,7 @@ const_debug unsigned int sysctl_sched_nr_migrate = 32;
824 * default: 0.25ms 797 * default: 0.25ms
825 */ 798 */
826unsigned int sysctl_sched_shares_ratelimit = 250000; 799unsigned int sysctl_sched_shares_ratelimit = 250000;
800unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
827 801
828/* 802/*
829 * Inject some fuzzyness into changing the per-cpu group shares 803 * Inject some fuzzyness into changing the per-cpu group shares
@@ -902,7 +876,7 @@ static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
902 */ 876 */
903 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); 877 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
904 878
905 spin_unlock_irq(&rq->lock); 879 raw_spin_unlock_irq(&rq->lock);
906} 880}
907 881
908#else /* __ARCH_WANT_UNLOCKED_CTXSW */ 882#else /* __ARCH_WANT_UNLOCKED_CTXSW */
@@ -926,9 +900,9 @@ static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
926 next->oncpu = 1; 900 next->oncpu = 1;
927#endif 901#endif
928#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 902#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
929 spin_unlock_irq(&rq->lock); 903 raw_spin_unlock_irq(&rq->lock);
930#else 904#else
931 spin_unlock(&rq->lock); 905 raw_spin_unlock(&rq->lock);
932#endif 906#endif
933} 907}
934 908
@@ -950,18 +924,35 @@ static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
950#endif /* __ARCH_WANT_UNLOCKED_CTXSW */ 924#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
951 925
952/* 926/*
927 * Check whether the task is waking, we use this to synchronize against
928 * ttwu() so that task_cpu() reports a stable number.
929 *
930 * We need to make an exception for PF_STARTING tasks because the fork
931 * path might require task_rq_lock() to work, eg. it can call
932 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
933 */
934static inline int task_is_waking(struct task_struct *p)
935{
936 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
937}
938
939/*
953 * __task_rq_lock - lock the runqueue a given task resides on. 940 * __task_rq_lock - lock the runqueue a given task resides on.
954 * Must be called interrupts disabled. 941 * Must be called interrupts disabled.
955 */ 942 */
956static inline struct rq *__task_rq_lock(struct task_struct *p) 943static inline struct rq *__task_rq_lock(struct task_struct *p)
957 __acquires(rq->lock) 944 __acquires(rq->lock)
958{ 945{
946 struct rq *rq;
947
959 for (;;) { 948 for (;;) {
960 struct rq *rq = task_rq(p); 949 while (task_is_waking(p))
961 spin_lock(&rq->lock); 950 cpu_relax();
962 if (likely(rq == task_rq(p))) 951 rq = task_rq(p);
952 raw_spin_lock(&rq->lock);
953 if (likely(rq == task_rq(p) && !task_is_waking(p)))
963 return rq; 954 return rq;
964 spin_unlock(&rq->lock); 955 raw_spin_unlock(&rq->lock);
965 } 956 }
966} 957}
967 958
@@ -976,12 +967,14 @@ static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
976 struct rq *rq; 967 struct rq *rq;
977 968
978 for (;;) { 969 for (;;) {
970 while (task_is_waking(p))
971 cpu_relax();
979 local_irq_save(*flags); 972 local_irq_save(*flags);
980 rq = task_rq(p); 973 rq = task_rq(p);
981 spin_lock(&rq->lock); 974 raw_spin_lock(&rq->lock);
982 if (likely(rq == task_rq(p))) 975 if (likely(rq == task_rq(p) && !task_is_waking(p)))
983 return rq; 976 return rq;
984 spin_unlock_irqrestore(&rq->lock, *flags); 977 raw_spin_unlock_irqrestore(&rq->lock, *flags);
985 } 978 }
986} 979}
987 980
@@ -990,19 +983,19 @@ void task_rq_unlock_wait(struct task_struct *p)
990 struct rq *rq = task_rq(p); 983 struct rq *rq = task_rq(p);
991 984
992 smp_mb(); /* spin-unlock-wait is not a full memory barrier */ 985 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
993 spin_unlock_wait(&rq->lock); 986 raw_spin_unlock_wait(&rq->lock);
994} 987}
995 988
996static void __task_rq_unlock(struct rq *rq) 989static void __task_rq_unlock(struct rq *rq)
997 __releases(rq->lock) 990 __releases(rq->lock)
998{ 991{
999 spin_unlock(&rq->lock); 992 raw_spin_unlock(&rq->lock);
1000} 993}
1001 994
1002static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) 995static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1003 __releases(rq->lock) 996 __releases(rq->lock)
1004{ 997{
1005 spin_unlock_irqrestore(&rq->lock, *flags); 998 raw_spin_unlock_irqrestore(&rq->lock, *flags);
1006} 999}
1007 1000
1008/* 1001/*
@@ -1015,7 +1008,7 @@ static struct rq *this_rq_lock(void)
1015 1008
1016 local_irq_disable(); 1009 local_irq_disable();
1017 rq = this_rq(); 1010 rq = this_rq();
1018 spin_lock(&rq->lock); 1011 raw_spin_lock(&rq->lock);
1019 1012
1020 return rq; 1013 return rq;
1021} 1014}
@@ -1062,10 +1055,10 @@ static enum hrtimer_restart hrtick(struct hrtimer *timer)
1062 1055
1063 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 1056 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1064 1057
1065 spin_lock(&rq->lock); 1058 raw_spin_lock(&rq->lock);
1066 update_rq_clock(rq); 1059 update_rq_clock(rq);
1067 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 1060 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1068 spin_unlock(&rq->lock); 1061 raw_spin_unlock(&rq->lock);
1069 1062
1070 return HRTIMER_NORESTART; 1063 return HRTIMER_NORESTART;
1071} 1064}
@@ -1078,10 +1071,10 @@ static void __hrtick_start(void *arg)
1078{ 1071{
1079 struct rq *rq = arg; 1072 struct rq *rq = arg;
1080 1073
1081 spin_lock(&rq->lock); 1074 raw_spin_lock(&rq->lock);
1082 hrtimer_restart(&rq->hrtick_timer); 1075 hrtimer_restart(&rq->hrtick_timer);
1083 rq->hrtick_csd_pending = 0; 1076 rq->hrtick_csd_pending = 0;
1084 spin_unlock(&rq->lock); 1077 raw_spin_unlock(&rq->lock);
1085} 1078}
1086 1079
1087/* 1080/*
@@ -1188,7 +1181,7 @@ static void resched_task(struct task_struct *p)
1188{ 1181{
1189 int cpu; 1182 int cpu;
1190 1183
1191 assert_spin_locked(&task_rq(p)->lock); 1184 assert_raw_spin_locked(&task_rq(p)->lock);
1192 1185
1193 if (test_tsk_need_resched(p)) 1186 if (test_tsk_need_resched(p))
1194 return; 1187 return;
@@ -1210,10 +1203,10 @@ static void resched_cpu(int cpu)
1210 struct rq *rq = cpu_rq(cpu); 1203 struct rq *rq = cpu_rq(cpu);
1211 unsigned long flags; 1204 unsigned long flags;
1212 1205
1213 if (!spin_trylock_irqsave(&rq->lock, flags)) 1206 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1214 return; 1207 return;
1215 resched_task(cpu_curr(cpu)); 1208 resched_task(cpu_curr(cpu));
1216 spin_unlock_irqrestore(&rq->lock, flags); 1209 raw_spin_unlock_irqrestore(&rq->lock, flags);
1217} 1210}
1218 1211
1219#ifdef CONFIG_NO_HZ 1212#ifdef CONFIG_NO_HZ
@@ -1282,7 +1275,7 @@ static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1282#else /* !CONFIG_SMP */ 1275#else /* !CONFIG_SMP */
1283static void resched_task(struct task_struct *p) 1276static void resched_task(struct task_struct *p)
1284{ 1277{
1285 assert_spin_locked(&task_rq(p)->lock); 1278 assert_raw_spin_locked(&task_rq(p)->lock);
1286 set_tsk_need_resched(p); 1279 set_tsk_need_resched(p);
1287} 1280}
1288 1281
@@ -1399,32 +1392,6 @@ static const u32 prio_to_wmult[40] = {
1399 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, 1392 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1400}; 1393};
1401 1394
1402static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1403
1404/*
1405 * runqueue iterator, to support SMP load-balancing between different
1406 * scheduling classes, without having to expose their internal data
1407 * structures to the load-balancing proper:
1408 */
1409struct rq_iterator {
1410 void *arg;
1411 struct task_struct *(*start)(void *);
1412 struct task_struct *(*next)(void *);
1413};
1414
1415#ifdef CONFIG_SMP
1416static unsigned long
1417balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1418 unsigned long max_load_move, struct sched_domain *sd,
1419 enum cpu_idle_type idle, int *all_pinned,
1420 int *this_best_prio, struct rq_iterator *iterator);
1421
1422static int
1423iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1424 struct sched_domain *sd, enum cpu_idle_type idle,
1425 struct rq_iterator *iterator);
1426#endif
1427
1428/* Time spent by the tasks of the cpu accounting group executing in ... */ 1395/* Time spent by the tasks of the cpu accounting group executing in ... */
1429enum cpuacct_stat_index { 1396enum cpuacct_stat_index {
1430 CPUACCT_STAT_USER, /* ... user mode */ 1397 CPUACCT_STAT_USER, /* ... user mode */
@@ -1540,7 +1507,7 @@ static unsigned long target_load(int cpu, int type)
1540 1507
1541static struct sched_group *group_of(int cpu) 1508static struct sched_group *group_of(int cpu)
1542{ 1509{
1543 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd); 1510 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1544 1511
1545 if (!sd) 1512 if (!sd)
1546 return NULL; 1513 return NULL;
@@ -1575,7 +1542,7 @@ static unsigned long cpu_avg_load_per_task(int cpu)
1575 1542
1576#ifdef CONFIG_FAIR_GROUP_SCHED 1543#ifdef CONFIG_FAIR_GROUP_SCHED
1577 1544
1578static __read_mostly unsigned long *update_shares_data; 1545static __read_mostly unsigned long __percpu *update_shares_data;
1579 1546
1580static void __set_se_shares(struct sched_entity *se, unsigned long shares); 1547static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1581 1548
@@ -1609,11 +1576,11 @@ static void update_group_shares_cpu(struct task_group *tg, int cpu,
1609 struct rq *rq = cpu_rq(cpu); 1576 struct rq *rq = cpu_rq(cpu);
1610 unsigned long flags; 1577 unsigned long flags;
1611 1578
1612 spin_lock_irqsave(&rq->lock, flags); 1579 raw_spin_lock_irqsave(&rq->lock, flags);
1613 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight; 1580 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1614 tg->cfs_rq[cpu]->shares = boost ? 0 : shares; 1581 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1615 __set_se_shares(tg->se[cpu], shares); 1582 __set_se_shares(tg->se[cpu], shares);
1616 spin_unlock_irqrestore(&rq->lock, flags); 1583 raw_spin_unlock_irqrestore(&rq->lock, flags);
1617 } 1584 }
1618} 1585}
1619 1586
@@ -1624,7 +1591,7 @@ static void update_group_shares_cpu(struct task_group *tg, int cpu,
1624 */ 1591 */
1625static int tg_shares_up(struct task_group *tg, void *data) 1592static int tg_shares_up(struct task_group *tg, void *data)
1626{ 1593{
1627 unsigned long weight, rq_weight = 0, shares = 0; 1594 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1628 unsigned long *usd_rq_weight; 1595 unsigned long *usd_rq_weight;
1629 struct sched_domain *sd = data; 1596 struct sched_domain *sd = data;
1630 unsigned long flags; 1597 unsigned long flags;
@@ -1640,6 +1607,7 @@ static int tg_shares_up(struct task_group *tg, void *data)
1640 weight = tg->cfs_rq[i]->load.weight; 1607 weight = tg->cfs_rq[i]->load.weight;
1641 usd_rq_weight[i] = weight; 1608 usd_rq_weight[i] = weight;
1642 1609
1610 rq_weight += weight;
1643 /* 1611 /*
1644 * If there are currently no tasks on the cpu pretend there 1612 * If there are currently no tasks on the cpu pretend there
1645 * is one of average load so that when a new task gets to 1613 * is one of average load so that when a new task gets to
@@ -1648,10 +1616,13 @@ static int tg_shares_up(struct task_group *tg, void *data)
1648 if (!weight) 1616 if (!weight)
1649 weight = NICE_0_LOAD; 1617 weight = NICE_0_LOAD;
1650 1618
1651 rq_weight += weight; 1619 sum_weight += weight;
1652 shares += tg->cfs_rq[i]->shares; 1620 shares += tg->cfs_rq[i]->shares;
1653 } 1621 }
1654 1622
1623 if (!rq_weight)
1624 rq_weight = sum_weight;
1625
1655 if ((!shares && rq_weight) || shares > tg->shares) 1626 if ((!shares && rq_weight) || shares > tg->shares)
1656 shares = tg->shares; 1627 shares = tg->shares;
1657 1628
@@ -1706,16 +1677,6 @@ static void update_shares(struct sched_domain *sd)
1706 } 1677 }
1707} 1678}
1708 1679
1709static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1710{
1711 if (root_task_group_empty())
1712 return;
1713
1714 spin_unlock(&rq->lock);
1715 update_shares(sd);
1716 spin_lock(&rq->lock);
1717}
1718
1719static void update_h_load(long cpu) 1680static void update_h_load(long cpu)
1720{ 1681{
1721 if (root_task_group_empty()) 1682 if (root_task_group_empty())
@@ -1730,10 +1691,6 @@ static inline void update_shares(struct sched_domain *sd)
1730{ 1691{
1731} 1692}
1732 1693
1733static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1734{
1735}
1736
1737#endif 1694#endif
1738 1695
1739#ifdef CONFIG_PREEMPT 1696#ifdef CONFIG_PREEMPT
@@ -1753,7 +1710,7 @@ static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1753 __acquires(busiest->lock) 1710 __acquires(busiest->lock)
1754 __acquires(this_rq->lock) 1711 __acquires(this_rq->lock)
1755{ 1712{
1756 spin_unlock(&this_rq->lock); 1713 raw_spin_unlock(&this_rq->lock);
1757 double_rq_lock(this_rq, busiest); 1714 double_rq_lock(this_rq, busiest);
1758 1715
1759 return 1; 1716 return 1;
@@ -1774,14 +1731,16 @@ static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1774{ 1731{
1775 int ret = 0; 1732 int ret = 0;
1776 1733
1777 if (unlikely(!spin_trylock(&busiest->lock))) { 1734 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1778 if (busiest < this_rq) { 1735 if (busiest < this_rq) {
1779 spin_unlock(&this_rq->lock); 1736 raw_spin_unlock(&this_rq->lock);
1780 spin_lock(&busiest->lock); 1737 raw_spin_lock(&busiest->lock);
1781 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING); 1738 raw_spin_lock_nested(&this_rq->lock,
1739 SINGLE_DEPTH_NESTING);
1782 ret = 1; 1740 ret = 1;
1783 } else 1741 } else
1784 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING); 1742 raw_spin_lock_nested(&busiest->lock,
1743 SINGLE_DEPTH_NESTING);
1785 } 1744 }
1786 return ret; 1745 return ret;
1787} 1746}
@@ -1795,7 +1754,7 @@ static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1795{ 1754{
1796 if (unlikely(!irqs_disabled())) { 1755 if (unlikely(!irqs_disabled())) {
1797 /* printk() doesn't work good under rq->lock */ 1756 /* printk() doesn't work good under rq->lock */
1798 spin_unlock(&this_rq->lock); 1757 raw_spin_unlock(&this_rq->lock);
1799 BUG_ON(1); 1758 BUG_ON(1);
1800 } 1759 }
1801 1760
@@ -1805,9 +1764,54 @@ static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1805static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) 1764static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1806 __releases(busiest->lock) 1765 __releases(busiest->lock)
1807{ 1766{
1808 spin_unlock(&busiest->lock); 1767 raw_spin_unlock(&busiest->lock);
1809 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); 1768 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1810} 1769}
1770
1771/*
1772 * double_rq_lock - safely lock two runqueues
1773 *
1774 * Note this does not disable interrupts like task_rq_lock,
1775 * you need to do so manually before calling.
1776 */
1777static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1778 __acquires(rq1->lock)
1779 __acquires(rq2->lock)
1780{
1781 BUG_ON(!irqs_disabled());
1782 if (rq1 == rq2) {
1783 raw_spin_lock(&rq1->lock);
1784 __acquire(rq2->lock); /* Fake it out ;) */
1785 } else {
1786 if (rq1 < rq2) {
1787 raw_spin_lock(&rq1->lock);
1788 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1789 } else {
1790 raw_spin_lock(&rq2->lock);
1791 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1792 }
1793 }
1794 update_rq_clock(rq1);
1795 update_rq_clock(rq2);
1796}
1797
1798/*
1799 * double_rq_unlock - safely unlock two runqueues
1800 *
1801 * Note this does not restore interrupts like task_rq_unlock,
1802 * you need to do so manually after calling.
1803 */
1804static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1805 __releases(rq1->lock)
1806 __releases(rq2->lock)
1807{
1808 raw_spin_unlock(&rq1->lock);
1809 if (rq1 != rq2)
1810 raw_spin_unlock(&rq2->lock);
1811 else
1812 __release(rq2->lock);
1813}
1814
1811#endif 1815#endif
1812 1816
1813#ifdef CONFIG_FAIR_GROUP_SCHED 1817#ifdef CONFIG_FAIR_GROUP_SCHED
@@ -1820,20 +1824,31 @@ static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1820#endif 1824#endif
1821 1825
1822static void calc_load_account_active(struct rq *this_rq); 1826static void calc_load_account_active(struct rq *this_rq);
1827static void update_sysctl(void);
1828static int get_update_sysctl_factor(void);
1823 1829
1824#include "sched_stats.h" 1830static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1825#include "sched_idletask.c" 1831{
1826#include "sched_fair.c" 1832 set_task_rq(p, cpu);
1827#include "sched_rt.c" 1833#ifdef CONFIG_SMP
1828#include "../litmus/sched_litmus.c" 1834 /*
1829#ifdef CONFIG_SCHED_DEBUG 1835 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1830# include "sched_debug.c" 1836 * successfuly executed on another CPU. We must ensure that updates of
1837 * per-task data have been completed by this moment.
1838 */
1839 smp_wmb();
1840 task_thread_info(p)->cpu = cpu;
1831#endif 1841#endif
1842}
1843
1844static const struct sched_class rt_sched_class;
1832 1845
1833#define sched_class_highest (&litmus_sched_class) 1846#define sched_class_highest (&litmus_sched_class)
1834#define for_each_class(class) \ 1847#define for_each_class(class) \
1835 for (class = sched_class_highest; class; class = class->next) 1848 for (class = sched_class_highest; class; class = class->next)
1836 1849
1850#include "sched_stats.h"
1851
1837static void inc_nr_running(struct rq *rq) 1852static void inc_nr_running(struct rq *rq)
1838{ 1853{
1839 rq->nr_running++; 1854 rq->nr_running++;
@@ -1871,13 +1886,14 @@ static void update_avg(u64 *avg, u64 sample)
1871 *avg += diff >> 3; 1886 *avg += diff >> 3;
1872} 1887}
1873 1888
1874static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup) 1889static void
1890enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1875{ 1891{
1876 if (wakeup) 1892 if (wakeup)
1877 p->se.start_runtime = p->se.sum_exec_runtime; 1893 p->se.start_runtime = p->se.sum_exec_runtime;
1878 1894
1879 sched_info_queued(p); 1895 sched_info_queued(p);
1880 p->sched_class->enqueue_task(rq, p, wakeup); 1896 p->sched_class->enqueue_task(rq, p, wakeup, head);
1881 p->se.on_rq = 1; 1897 p->se.on_rq = 1;
1882} 1898}
1883 1899
@@ -1900,6 +1916,38 @@ static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1900} 1916}
1901 1917
1902/* 1918/*
1919 * activate_task - move a task to the runqueue.
1920 */
1921static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1922{
1923 if (task_contributes_to_load(p))
1924 rq->nr_uninterruptible--;
1925
1926 enqueue_task(rq, p, wakeup, false);
1927 inc_nr_running(rq);
1928}
1929
1930/*
1931 * deactivate_task - remove a task from the runqueue.
1932 */
1933static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1934{
1935 if (task_contributes_to_load(p))
1936 rq->nr_uninterruptible++;
1937
1938 dequeue_task(rq, p, sleep);
1939 dec_nr_running(rq);
1940}
1941
1942#include "sched_idletask.c"
1943#include "sched_fair.c"
1944#include "sched_rt.c"
1945#include "../litmus/sched_litmus.c"
1946#ifdef CONFIG_SCHED_DEBUG
1947# include "sched_debug.c"
1948#endif
1949
1950/*
1903 * __normal_prio - return the priority that is based on the static prio 1951 * __normal_prio - return the priority that is based on the static prio
1904 */ 1952 */
1905static inline int __normal_prio(struct task_struct *p) 1953static inline int __normal_prio(struct task_struct *p)
@@ -1945,30 +1993,6 @@ static int effective_prio(struct task_struct *p)
1945 return p->prio; 1993 return p->prio;
1946} 1994}
1947 1995
1948/*
1949 * activate_task - move a task to the runqueue.
1950 */
1951static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1952{
1953 if (task_contributes_to_load(p))
1954 rq->nr_uninterruptible--;
1955
1956 enqueue_task(rq, p, wakeup);
1957 inc_nr_running(rq);
1958}
1959
1960/*
1961 * deactivate_task - remove a task from the runqueue.
1962 */
1963static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1964{
1965 if (task_contributes_to_load(p))
1966 rq->nr_uninterruptible++;
1967
1968 dequeue_task(rq, p, sleep);
1969 dec_nr_running(rq);
1970}
1971
1972/** 1996/**
1973 * task_curr - is this task currently executing on a CPU? 1997 * task_curr - is this task currently executing on a CPU?
1974 * @p: the task in question. 1998 * @p: the task in question.
@@ -1978,20 +2002,6 @@ inline int task_curr(const struct task_struct *p)
1978 return cpu_curr(task_cpu(p)) == p; 2002 return cpu_curr(task_cpu(p)) == p;
1979} 2003}
1980 2004
1981static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1982{
1983 set_task_rq(p, cpu);
1984#ifdef CONFIG_SMP
1985 /*
1986 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1987 * successfuly executed on another CPU. We must ensure that updates of
1988 * per-task data have been completed by this moment.
1989 */
1990 smp_wmb();
1991 task_thread_info(p)->cpu = cpu;
1992#endif
1993}
1994
1995static inline void check_class_changed(struct rq *rq, struct task_struct *p, 2005static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1996 const struct sched_class *prev_class, 2006 const struct sched_class *prev_class,
1997 int oldprio, int running) 2007 int oldprio, int running)
@@ -2004,38 +2014,6 @@ static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2004 p->sched_class->prio_changed(rq, p, oldprio, running); 2014 p->sched_class->prio_changed(rq, p, oldprio, running);
2005} 2015}
2006 2016
2007/**
2008 * kthread_bind - bind a just-created kthread to a cpu.
2009 * @p: thread created by kthread_create().
2010 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2011 *
2012 * Description: This function is equivalent to set_cpus_allowed(),
2013 * except that @cpu doesn't need to be online, and the thread must be
2014 * stopped (i.e., just returned from kthread_create()).
2015 *
2016 * Function lives here instead of kthread.c because it messes with
2017 * scheduler internals which require locking.
2018 */
2019void kthread_bind(struct task_struct *p, unsigned int cpu)
2020{
2021 struct rq *rq = cpu_rq(cpu);
2022 unsigned long flags;
2023
2024 /* Must have done schedule() in kthread() before we set_task_cpu */
2025 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2026 WARN_ON(1);
2027 return;
2028 }
2029
2030 spin_lock_irqsave(&rq->lock, flags);
2031 set_task_cpu(p, cpu);
2032 p->cpus_allowed = cpumask_of_cpu(cpu);
2033 p->rt.nr_cpus_allowed = 1;
2034 p->flags |= PF_THREAD_BOUND;
2035 spin_unlock_irqrestore(&rq->lock, flags);
2036}
2037EXPORT_SYMBOL(kthread_bind);
2038
2039#ifdef CONFIG_SMP 2017#ifdef CONFIG_SMP
2040/* 2018/*
2041 * Is this task likely cache-hot: 2019 * Is this task likely cache-hot:
@@ -2045,6 +2023,9 @@ task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2045{ 2023{
2046 s64 delta; 2024 s64 delta;
2047 2025
2026 if (p->sched_class != &fair_sched_class)
2027 return 0;
2028
2048 /* 2029 /*
2049 * Buddy candidates are cache hot: 2030 * Buddy candidates are cache hot:
2050 */ 2031 */
@@ -2053,9 +2034,6 @@ task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2053 &p->se == cfs_rq_of(&p->se)->last)) 2034 &p->se == cfs_rq_of(&p->se)->last))
2054 return 1; 2035 return 1;
2055 2036
2056 if (p->sched_class != &fair_sched_class)
2057 return 0;
2058
2059 if (sysctl_sched_migration_cost == -1) 2037 if (sysctl_sched_migration_cost == -1)
2060 return 1; 2038 return 1;
2061 if (sysctl_sched_migration_cost == 0) 2039 if (sysctl_sched_migration_cost == 0)
@@ -2066,39 +2044,23 @@ task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2066 return delta < (s64)sysctl_sched_migration_cost; 2044 return delta < (s64)sysctl_sched_migration_cost;
2067} 2045}
2068 2046
2069
2070void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 2047void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2071{ 2048{
2072 int old_cpu = task_cpu(p); 2049#ifdef CONFIG_SCHED_DEBUG
2073 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu); 2050 /*
2074 struct cfs_rq *old_cfsrq = task_cfs_rq(p), 2051 * We should never call set_task_cpu() on a blocked task,
2075 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu); 2052 * ttwu() will sort out the placement.
2076 u64 clock_offset; 2053 */
2077 2054 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2078 clock_offset = old_rq->clock - new_rq->clock; 2055 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2056#endif
2079 2057
2080 trace_sched_migrate_task(p, new_cpu); 2058 trace_sched_migrate_task(p, new_cpu);
2081 2059
2082#ifdef CONFIG_SCHEDSTATS 2060 if (task_cpu(p) != new_cpu) {
2083 if (p->se.wait_start)
2084 p->se.wait_start -= clock_offset;
2085 if (p->se.sleep_start)
2086 p->se.sleep_start -= clock_offset;
2087 if (p->se.block_start)
2088 p->se.block_start -= clock_offset;
2089#endif
2090 if (old_cpu != new_cpu) {
2091 p->se.nr_migrations++; 2061 p->se.nr_migrations++;
2092 new_rq->nr_migrations_in++; 2062 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2093#ifdef CONFIG_SCHEDSTATS
2094 if (task_hot(p, old_rq->clock, NULL))
2095 schedstat_inc(p, se.nr_forced2_migrations);
2096#endif
2097 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2098 1, 1, NULL, 0);
2099 } 2063 }
2100 p->se.vruntime -= old_cfsrq->min_vruntime -
2101 new_cfsrq->min_vruntime;
2102 2064
2103 __set_task_cpu(p, new_cpu); 2065 __set_task_cpu(p, new_cpu);
2104} 2066}
@@ -2123,12 +2085,10 @@ migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2123 2085
2124 /* 2086 /*
2125 * If the task is not on a runqueue (and not running), then 2087 * If the task is not on a runqueue (and not running), then
2126 * it is sufficient to simply update the task's cpu field. 2088 * the next wake-up will properly place the task.
2127 */ 2089 */
2128 if (!p->se.on_rq && !task_running(rq, p)) { 2090 if (!p->se.on_rq && !task_running(rq, p))
2129 set_task_cpu(p, dest_cpu);
2130 return 0; 2091 return 0;
2131 }
2132 2092
2133 init_completion(&req->done); 2093 init_completion(&req->done);
2134 req->task = p; 2094 req->task = p;
@@ -2333,6 +2293,75 @@ void task_oncpu_function_call(struct task_struct *p,
2333 preempt_enable(); 2293 preempt_enable();
2334} 2294}
2335 2295
2296#ifdef CONFIG_SMP
2297static int select_fallback_rq(int cpu, struct task_struct *p)
2298{
2299 int dest_cpu;
2300 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2301
2302 /* Look for allowed, online CPU in same node. */
2303 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2304 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2305 return dest_cpu;
2306
2307 /* Any allowed, online CPU? */
2308 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2309 if (dest_cpu < nr_cpu_ids)
2310 return dest_cpu;
2311
2312 /* No more Mr. Nice Guy. */
2313 if (dest_cpu >= nr_cpu_ids) {
2314 rcu_read_lock();
2315 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2316 rcu_read_unlock();
2317 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2318
2319 /*
2320 * Don't tell them about moving exiting tasks or
2321 * kernel threads (both mm NULL), since they never
2322 * leave kernel.
2323 */
2324 if (p->mm && printk_ratelimit()) {
2325 printk(KERN_INFO "process %d (%s) no "
2326 "longer affine to cpu%d\n",
2327 task_pid_nr(p), p->comm, cpu);
2328 }
2329 }
2330
2331 return dest_cpu;
2332}
2333
2334/*
2335 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2336 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2337 * by:
2338 *
2339 * exec: is unstable, retry loop
2340 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2341 */
2342static inline
2343int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2344{
2345 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2346
2347 /*
2348 * In order not to call set_task_cpu() on a blocking task we need
2349 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2350 * cpu.
2351 *
2352 * Since this is common to all placement strategies, this lives here.
2353 *
2354 * [ this allows ->select_task() to simply return task_cpu(p) and
2355 * not worry about this generic constraint ]
2356 */
2357 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2358 !cpu_online(cpu)))
2359 cpu = select_fallback_rq(task_cpu(p), p);
2360
2361 return cpu;
2362}
2363#endif
2364
2336/*** 2365/***
2337 * try_to_wake_up - wake up a thread 2366 * try_to_wake_up - wake up a thread
2338 * @p: the to-be-woken-up thread 2367 * @p: the to-be-woken-up thread
@@ -2352,7 +2381,7 @@ static int try_to_wake_up(struct task_struct *p, unsigned int state,
2352{ 2381{
2353 int cpu, orig_cpu, this_cpu, success = 0; 2382 int cpu, orig_cpu, this_cpu, success = 0;
2354 unsigned long flags; 2383 unsigned long flags;
2355 struct rq *rq, *orig_rq; 2384 struct rq *rq;
2356 2385
2357 if (is_realtime(p)) 2386 if (is_realtime(p))
2358 TRACE_TASK(p, "try_to_wake_up() state:%d\n", p->state); 2387 TRACE_TASK(p, "try_to_wake_up() state:%d\n", p->state);
@@ -2363,7 +2392,7 @@ static int try_to_wake_up(struct task_struct *p, unsigned int state,
2363 this_cpu = get_cpu(); 2392 this_cpu = get_cpu();
2364 2393
2365 smp_wmb(); 2394 smp_wmb();
2366 rq = orig_rq = task_rq_lock(p, &flags); 2395 rq = task_rq_lock(p, &flags);
2367 update_rq_clock(rq); 2396 update_rq_clock(rq);
2368 if (!(p->state & state)) 2397 if (!(p->state & state))
2369 goto out; 2398 goto out;
@@ -2387,19 +2416,34 @@ static int try_to_wake_up(struct task_struct *p, unsigned int state,
2387 if (task_contributes_to_load(p)) 2416 if (task_contributes_to_load(p))
2388 rq->nr_uninterruptible--; 2417 rq->nr_uninterruptible--;
2389 p->state = TASK_WAKING; 2418 p->state = TASK_WAKING;
2390 task_rq_unlock(rq, &flags);
2391 2419
2392 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 2420 if (p->sched_class->task_waking)
2393 if (cpu != orig_cpu) 2421 p->sched_class->task_waking(rq, p);
2394 set_task_cpu(p, cpu);
2395 2422
2396 rq = task_rq_lock(p, &flags); 2423 __task_rq_unlock(rq);
2397 2424
2398 if (rq != orig_rq) 2425 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2399 update_rq_clock(rq); 2426 if (cpu != orig_cpu) {
2427 /*
2428 * Since we migrate the task without holding any rq->lock,
2429 * we need to be careful with task_rq_lock(), since that
2430 * might end up locking an invalid rq.
2431 */
2432 set_task_cpu(p, cpu);
2433 }
2400 2434
2435 rq = cpu_rq(cpu);
2436 raw_spin_lock(&rq->lock);
2437 update_rq_clock(rq);
2438
2439 /*
2440 * We migrated the task without holding either rq->lock, however
2441 * since the task is not on the task list itself, nobody else
2442 * will try and migrate the task, hence the rq should match the
2443 * cpu we just moved it to.
2444 */
2445 WARN_ON(task_cpu(p) != cpu);
2401 WARN_ON(p->state != TASK_WAKING); 2446 WARN_ON(p->state != TASK_WAKING);
2402 cpu = task_cpu(p);
2403 2447
2404#ifdef CONFIG_SCHEDSTATS 2448#ifdef CONFIG_SCHEDSTATS
2405 schedstat_inc(rq, ttwu_count); 2449 schedstat_inc(rq, ttwu_count);
@@ -2452,8 +2496,19 @@ out_running:
2452 2496
2453 p->state = TASK_RUNNING; 2497 p->state = TASK_RUNNING;
2454#ifdef CONFIG_SMP 2498#ifdef CONFIG_SMP
2455 if (p->sched_class->task_wake_up) 2499 if (p->sched_class->task_woken)
2456 p->sched_class->task_wake_up(rq, p); 2500 p->sched_class->task_woken(rq, p);
2501
2502 if (unlikely(rq->idle_stamp)) {
2503 u64 delta = rq->clock - rq->idle_stamp;
2504 u64 max = 2*sysctl_sched_migration_cost;
2505
2506 if (delta > max)
2507 rq->avg_idle = max;
2508 else
2509 update_avg(&rq->avg_idle, delta);
2510 rq->idle_stamp = 0;
2511 }
2457#endif 2512#endif
2458out: 2513out:
2459 if (is_realtime(p)) 2514 if (is_realtime(p))
@@ -2502,7 +2557,6 @@ static void __sched_fork(struct task_struct *p)
2502 p->se.avg_overlap = 0; 2557 p->se.avg_overlap = 0;
2503 p->se.start_runtime = 0; 2558 p->se.start_runtime = 0;
2504 p->se.avg_wakeup = sysctl_sched_wakeup_granularity; 2559 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2505 p->se.avg_running = 0;
2506 2560
2507#ifdef CONFIG_SCHEDSTATS 2561#ifdef CONFIG_SCHEDSTATS
2508 p->se.wait_start = 0; 2562 p->se.wait_start = 0;
@@ -2524,7 +2578,6 @@ static void __sched_fork(struct task_struct *p)
2524 p->se.nr_failed_migrations_running = 0; 2578 p->se.nr_failed_migrations_running = 0;
2525 p->se.nr_failed_migrations_hot = 0; 2579 p->se.nr_failed_migrations_hot = 0;
2526 p->se.nr_forced_migrations = 0; 2580 p->se.nr_forced_migrations = 0;
2527 p->se.nr_forced2_migrations = 0;
2528 2581
2529 p->se.nr_wakeups = 0; 2582 p->se.nr_wakeups = 0;
2530 p->se.nr_wakeups_sync = 0; 2583 p->se.nr_wakeups_sync = 0;
@@ -2545,14 +2598,6 @@ static void __sched_fork(struct task_struct *p)
2545#ifdef CONFIG_PREEMPT_NOTIFIERS 2598#ifdef CONFIG_PREEMPT_NOTIFIERS
2546 INIT_HLIST_HEAD(&p->preempt_notifiers); 2599 INIT_HLIST_HEAD(&p->preempt_notifiers);
2547#endif 2600#endif
2548
2549 /*
2550 * We mark the process as running here, but have not actually
2551 * inserted it onto the runqueue yet. This guarantees that
2552 * nobody will actually run it, and a signal or other external
2553 * event cannot wake it up and insert it on the runqueue either.
2554 */
2555 p->state = TASK_RUNNING;
2556} 2601}
2557 2602
2558/* 2603/*
@@ -2563,6 +2608,12 @@ void sched_fork(struct task_struct *p, int clone_flags)
2563 int cpu = get_cpu(); 2608 int cpu = get_cpu();
2564 2609
2565 __sched_fork(p); 2610 __sched_fork(p);
2611 /*
2612 * We mark the process as waking here. This guarantees that
2613 * nobody will actually run it, and a signal or other external
2614 * event cannot wake it up and insert it on the runqueue either.
2615 */
2616 p->state = TASK_WAKING;
2566 2617
2567 /* 2618 /*
2568 * Revert to default priority/policy on fork if requested. 2619 * Revert to default priority/policy on fork if requested.
@@ -2594,9 +2645,9 @@ void sched_fork(struct task_struct *p, int clone_flags)
2594 if (!rt_prio(p->prio)) 2645 if (!rt_prio(p->prio))
2595 p->sched_class = &fair_sched_class; 2646 p->sched_class = &fair_sched_class;
2596 2647
2597#ifdef CONFIG_SMP 2648 if (p->sched_class->task_fork)
2598 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0); 2649 p->sched_class->task_fork(p);
2599#endif 2650
2600 set_task_cpu(p, cpu); 2651 set_task_cpu(p, cpu);
2601 2652
2602#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 2653#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
@@ -2626,28 +2677,41 @@ void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2626{ 2677{
2627 unsigned long flags; 2678 unsigned long flags;
2628 struct rq *rq; 2679 struct rq *rq;
2680 int cpu __maybe_unused = get_cpu();
2629 2681
2630 rq = task_rq_lock(p, &flags); 2682#ifdef CONFIG_SMP
2631 BUG_ON(p->state != TASK_RUNNING); 2683 /*
2632 update_rq_clock(rq); 2684 * Fork balancing, do it here and not earlier because:
2685 * - cpus_allowed can change in the fork path
2686 * - any previously selected cpu might disappear through hotplug
2687 *
2688 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2689 * ->cpus_allowed is stable, we have preemption disabled, meaning
2690 * cpu_online_mask is stable.
2691 */
2692 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2693 set_task_cpu(p, cpu);
2694#endif
2633 2695
2634 if (!p->sched_class->task_new || !current->se.on_rq) { 2696 /*
2635 activate_task(rq, p, 0); 2697 * Since the task is not on the rq and we still have TASK_WAKING set
2636 } else { 2698 * nobody else will migrate this task.
2637 /* 2699 */
2638 * Let the scheduling class do new task startup 2700 rq = cpu_rq(cpu);
2639 * management (if any): 2701 raw_spin_lock_irqsave(&rq->lock, flags);
2640 */ 2702
2641 p->sched_class->task_new(rq, p); 2703 BUG_ON(p->state != TASK_WAKING);
2642 inc_nr_running(rq); 2704 p->state = TASK_RUNNING;
2643 } 2705 update_rq_clock(rq);
2706 activate_task(rq, p, 0);
2644 trace_sched_wakeup_new(rq, p, 1); 2707 trace_sched_wakeup_new(rq, p, 1);
2645 check_preempt_curr(rq, p, WF_FORK); 2708 check_preempt_curr(rq, p, WF_FORK);
2646#ifdef CONFIG_SMP 2709#ifdef CONFIG_SMP
2647 if (p->sched_class->task_wake_up) 2710 if (p->sched_class->task_woken)
2648 p->sched_class->task_wake_up(rq, p); 2711 p->sched_class->task_woken(rq, p);
2649#endif 2712#endif
2650 task_rq_unlock(rq, &flags); 2713 task_rq_unlock(rq, &flags);
2714 put_cpu();
2651} 2715}
2652 2716
2653#ifdef CONFIG_PREEMPT_NOTIFIERS 2717#ifdef CONFIG_PREEMPT_NOTIFIERS
@@ -2768,7 +2832,13 @@ static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2768 finish_arch_switch(prev); 2832 finish_arch_switch(prev);
2769 litmus->finish_switch(prev); 2833 litmus->finish_switch(prev);
2770 prev->rt_param.stack_in_use = NO_CPU; 2834 prev->rt_param.stack_in_use = NO_CPU;
2771 perf_event_task_sched_in(current, cpu_of(rq)); 2835#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2836 local_irq_disable();
2837#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2838 perf_event_task_sched_in(current);
2839#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2840 local_irq_enable();
2841#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2772 finish_lock_switch(rq, prev); 2842 finish_lock_switch(rq, prev);
2773 2843
2774 fire_sched_in_preempt_notifiers(current); 2844 fire_sched_in_preempt_notifiers(current);
@@ -2808,10 +2878,10 @@ static inline void post_schedule(struct rq *rq)
2808 if (rq->post_schedule) { 2878 if (rq->post_schedule) {
2809 unsigned long flags; 2879 unsigned long flags;
2810 2880
2811 spin_lock_irqsave(&rq->lock, flags); 2881 raw_spin_lock_irqsave(&rq->lock, flags);
2812 if (rq->curr->sched_class->post_schedule) 2882 if (rq->curr->sched_class->post_schedule)
2813 rq->curr->sched_class->post_schedule(rq); 2883 rq->curr->sched_class->post_schedule(rq);
2814 spin_unlock_irqrestore(&rq->lock, flags); 2884 raw_spin_unlock_irqrestore(&rq->lock, flags);
2815 2885
2816 rq->post_schedule = 0; 2886 rq->post_schedule = 0;
2817 } 2887 }
@@ -2875,14 +2945,14 @@ context_switch(struct rq *rq, struct task_struct *prev,
2875 */ 2945 */
2876 arch_start_context_switch(prev); 2946 arch_start_context_switch(prev);
2877 2947
2878 if (unlikely(!mm)) { 2948 if (likely(!mm)) {
2879 next->active_mm = oldmm; 2949 next->active_mm = oldmm;
2880 atomic_inc(&oldmm->mm_count); 2950 atomic_inc(&oldmm->mm_count);
2881 enter_lazy_tlb(oldmm, next); 2951 enter_lazy_tlb(oldmm, next);
2882 } else 2952 } else
2883 switch_mm(oldmm, mm, next); 2953 switch_mm(oldmm, mm, next);
2884 2954
2885 if (unlikely(!prev->mm)) { 2955 if (likely(!prev->mm)) {
2886 prev->active_mm = NULL; 2956 prev->active_mm = NULL;
2887 rq->prev_mm = oldmm; 2957 rq->prev_mm = oldmm;
2888 } 2958 }
@@ -3045,15 +3115,6 @@ static void calc_load_account_active(struct rq *this_rq)
3045} 3115}
3046 3116
3047/* 3117/*
3048 * Externally visible per-cpu scheduler statistics:
3049 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3050 */
3051u64 cpu_nr_migrations(int cpu)
3052{
3053 return cpu_rq(cpu)->nr_migrations_in;
3054}
3055
3056/*
3057 * Update rq->cpu_load[] statistics. This function is usually called every 3118 * Update rq->cpu_load[] statistics. This function is usually called every
3058 * scheduler tick (TICK_NSEC). 3119 * scheduler tick (TICK_NSEC).
3059 */ 3120 */
@@ -3091,65 +3152,36 @@ static void update_cpu_load(struct rq *this_rq)
3091#ifdef CONFIG_SMP 3152#ifdef CONFIG_SMP
3092 3153
3093/* 3154/*
3094 * double_rq_lock - safely lock two runqueues 3155 * sched_exec - execve() is a valuable balancing opportunity, because at
3095 * 3156 * this point the task has the smallest effective memory and cache footprint.
3096 * Note this does not disable interrupts like task_rq_lock,
3097 * you need to do so manually before calling.
3098 */
3099static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3100 __acquires(rq1->lock)
3101 __acquires(rq2->lock)
3102{
3103 BUG_ON(!irqs_disabled());
3104 if (rq1 == rq2) {
3105 spin_lock(&rq1->lock);
3106 __acquire(rq2->lock); /* Fake it out ;) */
3107 } else {
3108 if (rq1 < rq2) {
3109 spin_lock(&rq1->lock);
3110 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3111 } else {
3112 spin_lock(&rq2->lock);
3113 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3114 }
3115 }
3116 update_rq_clock(rq1);
3117 update_rq_clock(rq2);
3118}
3119
3120/*
3121 * double_rq_unlock - safely unlock two runqueues
3122 *
3123 * Note this does not restore interrupts like task_rq_unlock,
3124 * you need to do so manually after calling.
3125 */
3126static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3127 __releases(rq1->lock)
3128 __releases(rq2->lock)
3129{
3130 spin_unlock(&rq1->lock);
3131 if (rq1 != rq2)
3132 spin_unlock(&rq2->lock);
3133 else
3134 __release(rq2->lock);
3135}
3136
3137/*
3138 * If dest_cpu is allowed for this process, migrate the task to it.
3139 * This is accomplished by forcing the cpu_allowed mask to only
3140 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3141 * the cpu_allowed mask is restored.
3142 */ 3157 */
3143static void sched_migrate_task(struct task_struct *p, int dest_cpu) 3158void sched_exec(void)
3144{ 3159{
3160 struct task_struct *p = current;
3145 struct migration_req req; 3161 struct migration_req req;
3162 int dest_cpu, this_cpu;
3146 unsigned long flags; 3163 unsigned long flags;
3147 struct rq *rq; 3164 struct rq *rq;
3148 3165
3166again:
3167 this_cpu = get_cpu();
3168 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3169 if (dest_cpu == this_cpu) {
3170 put_cpu();
3171 return;
3172 }
3173
3149 rq = task_rq_lock(p, &flags); 3174 rq = task_rq_lock(p, &flags);
3175 put_cpu();
3176
3177 /*
3178 * select_task_rq() can race against ->cpus_allowed
3179 */
3150 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed) 3180 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3151 || unlikely(!cpu_active(dest_cpu))) 3181 || unlikely(!cpu_active(dest_cpu))) {
3152 goto out; 3182 task_rq_unlock(rq, &flags);
3183 goto again;
3184 }
3153 3185
3154 /* force the process onto the specified CPU */ 3186 /* force the process onto the specified CPU */
3155 if (migrate_task(p, dest_cpu, &req)) { 3187 if (migrate_task(p, dest_cpu, &req)) {
@@ -3164,1784 +3196,9 @@ static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3164 3196
3165 return; 3197 return;
3166 } 3198 }
3167out:
3168 task_rq_unlock(rq, &flags); 3199 task_rq_unlock(rq, &flags);
3169} 3200}
3170 3201
3171/*
3172 * sched_exec - execve() is a valuable balancing opportunity, because at
3173 * this point the task has the smallest effective memory and cache footprint.
3174 */
3175void sched_exec(void)
3176{
3177 int new_cpu, this_cpu = get_cpu();
3178 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
3179 put_cpu();
3180 if (new_cpu != this_cpu)
3181 sched_migrate_task(current, new_cpu);
3182}
3183
3184/*
3185 * pull_task - move a task from a remote runqueue to the local runqueue.
3186 * Both runqueues must be locked.
3187 */
3188static void pull_task(struct rq *src_rq, struct task_struct *p,
3189 struct rq *this_rq, int this_cpu)
3190{
3191 deactivate_task(src_rq, p, 0);
3192 set_task_cpu(p, this_cpu);
3193 activate_task(this_rq, p, 0);
3194 /*
3195 * Note that idle threads have a prio of MAX_PRIO, for this test
3196 * to be always true for them.
3197 */
3198 check_preempt_curr(this_rq, p, 0);
3199}
3200
3201/*
3202 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3203 */
3204static
3205int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3206 struct sched_domain *sd, enum cpu_idle_type idle,
3207 int *all_pinned)
3208{
3209 int tsk_cache_hot = 0;
3210 /*
3211 * We do not migrate tasks that are:
3212 * 1) running (obviously), or
3213 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3214 * 3) are cache-hot on their current CPU.
3215 */
3216 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3217 schedstat_inc(p, se.nr_failed_migrations_affine);
3218 return 0;
3219 }
3220 *all_pinned = 0;
3221
3222 if (task_running(rq, p)) {
3223 schedstat_inc(p, se.nr_failed_migrations_running);
3224 return 0;
3225 }
3226
3227 /*
3228 * Aggressive migration if:
3229 * 1) task is cache cold, or
3230 * 2) too many balance attempts have failed.
3231 */
3232
3233 tsk_cache_hot = task_hot(p, rq->clock, sd);
3234 if (!tsk_cache_hot ||
3235 sd->nr_balance_failed > sd->cache_nice_tries) {
3236#ifdef CONFIG_SCHEDSTATS
3237 if (tsk_cache_hot) {
3238 schedstat_inc(sd, lb_hot_gained[idle]);
3239 schedstat_inc(p, se.nr_forced_migrations);
3240 }
3241#endif
3242 return 1;
3243 }
3244
3245 if (tsk_cache_hot) {
3246 schedstat_inc(p, se.nr_failed_migrations_hot);
3247 return 0;
3248 }
3249 return 1;
3250}
3251
3252static unsigned long
3253balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3254 unsigned long max_load_move, struct sched_domain *sd,
3255 enum cpu_idle_type idle, int *all_pinned,
3256 int *this_best_prio, struct rq_iterator *iterator)
3257{
3258 int loops = 0, pulled = 0, pinned = 0;
3259 struct task_struct *p;
3260 long rem_load_move = max_load_move;
3261
3262 if (max_load_move == 0)
3263 goto out;
3264
3265 pinned = 1;
3266
3267 /*
3268 * Start the load-balancing iterator:
3269 */
3270 p = iterator->start(iterator->arg);
3271next:
3272 if (!p || loops++ > sysctl_sched_nr_migrate)
3273 goto out;
3274
3275 if ((p->se.load.weight >> 1) > rem_load_move ||
3276 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3277 p = iterator->next(iterator->arg);
3278 goto next;
3279 }
3280
3281 pull_task(busiest, p, this_rq, this_cpu);
3282 pulled++;
3283 rem_load_move -= p->se.load.weight;
3284
3285#ifdef CONFIG_PREEMPT
3286 /*
3287 * NEWIDLE balancing is a source of latency, so preemptible kernels
3288 * will stop after the first task is pulled to minimize the critical
3289 * section.
3290 */
3291 if (idle == CPU_NEWLY_IDLE)
3292 goto out;
3293#endif
3294
3295 /*
3296 * We only want to steal up to the prescribed amount of weighted load.
3297 */
3298 if (rem_load_move > 0) {
3299 if (p->prio < *this_best_prio)
3300 *this_best_prio = p->prio;
3301 p = iterator->next(iterator->arg);
3302 goto next;
3303 }
3304out:
3305 /*
3306 * Right now, this is one of only two places pull_task() is called,
3307 * so we can safely collect pull_task() stats here rather than
3308 * inside pull_task().
3309 */
3310 schedstat_add(sd, lb_gained[idle], pulled);
3311
3312 if (all_pinned)
3313 *all_pinned = pinned;
3314
3315 return max_load_move - rem_load_move;
3316}
3317
3318/*
3319 * move_tasks tries to move up to max_load_move weighted load from busiest to
3320 * this_rq, as part of a balancing operation within domain "sd".
3321 * Returns 1 if successful and 0 otherwise.
3322 *
3323 * Called with both runqueues locked.
3324 */
3325static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3326 unsigned long max_load_move,
3327 struct sched_domain *sd, enum cpu_idle_type idle,
3328 int *all_pinned)
3329{
3330 const struct sched_class *class = sched_class_highest;
3331 unsigned long total_load_moved = 0;
3332 int this_best_prio = this_rq->curr->prio;
3333
3334 do {
3335 total_load_moved +=
3336 class->load_balance(this_rq, this_cpu, busiest,
3337 max_load_move - total_load_moved,
3338 sd, idle, all_pinned, &this_best_prio);
3339 class = class->next;
3340
3341#ifdef CONFIG_PREEMPT
3342 /*
3343 * NEWIDLE balancing is a source of latency, so preemptible
3344 * kernels will stop after the first task is pulled to minimize
3345 * the critical section.
3346 */
3347 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3348 break;
3349#endif
3350 } while (class && max_load_move > total_load_moved);
3351
3352 return total_load_moved > 0;
3353}
3354
3355static int
3356iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3357 struct sched_domain *sd, enum cpu_idle_type idle,
3358 struct rq_iterator *iterator)
3359{
3360 struct task_struct *p = iterator->start(iterator->arg);
3361 int pinned = 0;
3362
3363 while (p) {
3364 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3365 pull_task(busiest, p, this_rq, this_cpu);
3366 /*
3367 * Right now, this is only the second place pull_task()
3368 * is called, so we can safely collect pull_task()
3369 * stats here rather than inside pull_task().
3370 */
3371 schedstat_inc(sd, lb_gained[idle]);
3372
3373 return 1;
3374 }
3375 p = iterator->next(iterator->arg);
3376 }
3377
3378 return 0;
3379}
3380
3381/*
3382 * move_one_task tries to move exactly one task from busiest to this_rq, as
3383 * part of active balancing operations within "domain".
3384 * Returns 1 if successful and 0 otherwise.
3385 *
3386 * Called with both runqueues locked.
3387 */
3388static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3389 struct sched_domain *sd, enum cpu_idle_type idle)
3390{
3391 const struct sched_class *class;
3392
3393 for_each_class(class) {
3394 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3395 return 1;
3396 }
3397
3398 return 0;
3399}
3400/********** Helpers for find_busiest_group ************************/
3401/*
3402 * sd_lb_stats - Structure to store the statistics of a sched_domain
3403 * during load balancing.
3404 */
3405struct sd_lb_stats {
3406 struct sched_group *busiest; /* Busiest group in this sd */
3407 struct sched_group *this; /* Local group in this sd */
3408 unsigned long total_load; /* Total load of all groups in sd */
3409 unsigned long total_pwr; /* Total power of all groups in sd */
3410 unsigned long avg_load; /* Average load across all groups in sd */
3411
3412 /** Statistics of this group */
3413 unsigned long this_load;
3414 unsigned long this_load_per_task;
3415 unsigned long this_nr_running;
3416
3417 /* Statistics of the busiest group */
3418 unsigned long max_load;
3419 unsigned long busiest_load_per_task;
3420 unsigned long busiest_nr_running;
3421
3422 int group_imb; /* Is there imbalance in this sd */
3423#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3424 int power_savings_balance; /* Is powersave balance needed for this sd */
3425 struct sched_group *group_min; /* Least loaded group in sd */
3426 struct sched_group *group_leader; /* Group which relieves group_min */
3427 unsigned long min_load_per_task; /* load_per_task in group_min */
3428 unsigned long leader_nr_running; /* Nr running of group_leader */
3429 unsigned long min_nr_running; /* Nr running of group_min */
3430#endif
3431};
3432
3433/*
3434 * sg_lb_stats - stats of a sched_group required for load_balancing
3435 */
3436struct sg_lb_stats {
3437 unsigned long avg_load; /*Avg load across the CPUs of the group */
3438 unsigned long group_load; /* Total load over the CPUs of the group */
3439 unsigned long sum_nr_running; /* Nr tasks running in the group */
3440 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3441 unsigned long group_capacity;
3442 int group_imb; /* Is there an imbalance in the group ? */
3443};
3444
3445/**
3446 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3447 * @group: The group whose first cpu is to be returned.
3448 */
3449static inline unsigned int group_first_cpu(struct sched_group *group)
3450{
3451 return cpumask_first(sched_group_cpus(group));
3452}
3453
3454/**
3455 * get_sd_load_idx - Obtain the load index for a given sched domain.
3456 * @sd: The sched_domain whose load_idx is to be obtained.
3457 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3458 */
3459static inline int get_sd_load_idx(struct sched_domain *sd,
3460 enum cpu_idle_type idle)
3461{
3462 int load_idx;
3463
3464 switch (idle) {
3465 case CPU_NOT_IDLE:
3466 load_idx = sd->busy_idx;
3467 break;
3468
3469 case CPU_NEWLY_IDLE:
3470 load_idx = sd->newidle_idx;
3471 break;
3472 default:
3473 load_idx = sd->idle_idx;
3474 break;
3475 }
3476
3477 return load_idx;
3478}
3479
3480
3481#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3482/**
3483 * init_sd_power_savings_stats - Initialize power savings statistics for
3484 * the given sched_domain, during load balancing.
3485 *
3486 * @sd: Sched domain whose power-savings statistics are to be initialized.
3487 * @sds: Variable containing the statistics for sd.
3488 * @idle: Idle status of the CPU at which we're performing load-balancing.
3489 */
3490static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3491 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3492{
3493 /*
3494 * Busy processors will not participate in power savings
3495 * balance.
3496 */
3497 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3498 sds->power_savings_balance = 0;
3499 else {
3500 sds->power_savings_balance = 1;
3501 sds->min_nr_running = ULONG_MAX;
3502 sds->leader_nr_running = 0;
3503 }
3504}
3505
3506/**
3507 * update_sd_power_savings_stats - Update the power saving stats for a
3508 * sched_domain while performing load balancing.
3509 *
3510 * @group: sched_group belonging to the sched_domain under consideration.
3511 * @sds: Variable containing the statistics of the sched_domain
3512 * @local_group: Does group contain the CPU for which we're performing
3513 * load balancing ?
3514 * @sgs: Variable containing the statistics of the group.
3515 */
3516static inline void update_sd_power_savings_stats(struct sched_group *group,
3517 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3518{
3519
3520 if (!sds->power_savings_balance)
3521 return;
3522
3523 /*
3524 * If the local group is idle or completely loaded
3525 * no need to do power savings balance at this domain
3526 */
3527 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3528 !sds->this_nr_running))
3529 sds->power_savings_balance = 0;
3530
3531 /*
3532 * If a group is already running at full capacity or idle,
3533 * don't include that group in power savings calculations
3534 */
3535 if (!sds->power_savings_balance ||
3536 sgs->sum_nr_running >= sgs->group_capacity ||
3537 !sgs->sum_nr_running)
3538 return;
3539
3540 /*
3541 * Calculate the group which has the least non-idle load.
3542 * This is the group from where we need to pick up the load
3543 * for saving power
3544 */
3545 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3546 (sgs->sum_nr_running == sds->min_nr_running &&
3547 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3548 sds->group_min = group;
3549 sds->min_nr_running = sgs->sum_nr_running;
3550 sds->min_load_per_task = sgs->sum_weighted_load /
3551 sgs->sum_nr_running;
3552 }
3553
3554 /*
3555 * Calculate the group which is almost near its
3556 * capacity but still has some space to pick up some load
3557 * from other group and save more power
3558 */
3559 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3560 return;
3561
3562 if (sgs->sum_nr_running > sds->leader_nr_running ||
3563 (sgs->sum_nr_running == sds->leader_nr_running &&
3564 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3565 sds->group_leader = group;
3566 sds->leader_nr_running = sgs->sum_nr_running;
3567 }
3568}
3569
3570/**
3571 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3572 * @sds: Variable containing the statistics of the sched_domain
3573 * under consideration.
3574 * @this_cpu: Cpu at which we're currently performing load-balancing.
3575 * @imbalance: Variable to store the imbalance.
3576 *
3577 * Description:
3578 * Check if we have potential to perform some power-savings balance.
3579 * If yes, set the busiest group to be the least loaded group in the
3580 * sched_domain, so that it's CPUs can be put to idle.
3581 *
3582 * Returns 1 if there is potential to perform power-savings balance.
3583 * Else returns 0.
3584 */
3585static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3586 int this_cpu, unsigned long *imbalance)
3587{
3588 if (!sds->power_savings_balance)
3589 return 0;
3590
3591 if (sds->this != sds->group_leader ||
3592 sds->group_leader == sds->group_min)
3593 return 0;
3594
3595 *imbalance = sds->min_load_per_task;
3596 sds->busiest = sds->group_min;
3597
3598 return 1;
3599
3600}
3601#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3602static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3603 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3604{
3605 return;
3606}
3607
3608static inline void update_sd_power_savings_stats(struct sched_group *group,
3609 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3610{
3611 return;
3612}
3613
3614static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3615 int this_cpu, unsigned long *imbalance)
3616{
3617 return 0;
3618}
3619#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3620
3621
3622unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3623{
3624 return SCHED_LOAD_SCALE;
3625}
3626
3627unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3628{
3629 return default_scale_freq_power(sd, cpu);
3630}
3631
3632unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3633{
3634 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3635 unsigned long smt_gain = sd->smt_gain;
3636
3637 smt_gain /= weight;
3638
3639 return smt_gain;
3640}
3641
3642unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3643{
3644 return default_scale_smt_power(sd, cpu);
3645}
3646
3647unsigned long scale_rt_power(int cpu)
3648{
3649 struct rq *rq = cpu_rq(cpu);
3650 u64 total, available;
3651
3652 sched_avg_update(rq);
3653
3654 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3655 available = total - rq->rt_avg;
3656
3657 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3658 total = SCHED_LOAD_SCALE;
3659
3660 total >>= SCHED_LOAD_SHIFT;
3661
3662 return div_u64(available, total);
3663}
3664
3665static void update_cpu_power(struct sched_domain *sd, int cpu)
3666{
3667 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3668 unsigned long power = SCHED_LOAD_SCALE;
3669 struct sched_group *sdg = sd->groups;
3670
3671 if (sched_feat(ARCH_POWER))
3672 power *= arch_scale_freq_power(sd, cpu);
3673 else
3674 power *= default_scale_freq_power(sd, cpu);
3675
3676 power >>= SCHED_LOAD_SHIFT;
3677
3678 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3679 if (sched_feat(ARCH_POWER))
3680 power *= arch_scale_smt_power(sd, cpu);
3681 else
3682 power *= default_scale_smt_power(sd, cpu);
3683
3684 power >>= SCHED_LOAD_SHIFT;
3685 }
3686
3687 power *= scale_rt_power(cpu);
3688 power >>= SCHED_LOAD_SHIFT;
3689
3690 if (!power)
3691 power = 1;
3692
3693 sdg->cpu_power = power;
3694}
3695
3696static void update_group_power(struct sched_domain *sd, int cpu)
3697{
3698 struct sched_domain *child = sd->child;
3699 struct sched_group *group, *sdg = sd->groups;
3700 unsigned long power;
3701
3702 if (!child) {
3703 update_cpu_power(sd, cpu);
3704 return;
3705 }
3706
3707 power = 0;
3708
3709 group = child->groups;
3710 do {
3711 power += group->cpu_power;
3712 group = group->next;
3713 } while (group != child->groups);
3714
3715 sdg->cpu_power = power;
3716}
3717
3718/**
3719 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3720 * @sd: The sched_domain whose statistics are to be updated.
3721 * @group: sched_group whose statistics are to be updated.
3722 * @this_cpu: Cpu for which load balance is currently performed.
3723 * @idle: Idle status of this_cpu
3724 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3725 * @sd_idle: Idle status of the sched_domain containing group.
3726 * @local_group: Does group contain this_cpu.
3727 * @cpus: Set of cpus considered for load balancing.
3728 * @balance: Should we balance.
3729 * @sgs: variable to hold the statistics for this group.
3730 */
3731static inline void update_sg_lb_stats(struct sched_domain *sd,
3732 struct sched_group *group, int this_cpu,
3733 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3734 int local_group, const struct cpumask *cpus,
3735 int *balance, struct sg_lb_stats *sgs)
3736{
3737 unsigned long load, max_cpu_load, min_cpu_load;
3738 int i;
3739 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3740 unsigned long sum_avg_load_per_task;
3741 unsigned long avg_load_per_task;
3742
3743 if (local_group) {
3744 balance_cpu = group_first_cpu(group);
3745 if (balance_cpu == this_cpu)
3746 update_group_power(sd, this_cpu);
3747 }
3748
3749 /* Tally up the load of all CPUs in the group */
3750 sum_avg_load_per_task = avg_load_per_task = 0;
3751 max_cpu_load = 0;
3752 min_cpu_load = ~0UL;
3753
3754 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3755 struct rq *rq = cpu_rq(i);
3756
3757 if (*sd_idle && rq->nr_running)
3758 *sd_idle = 0;
3759
3760 /* Bias balancing toward cpus of our domain */
3761 if (local_group) {
3762 if (idle_cpu(i) && !first_idle_cpu) {
3763 first_idle_cpu = 1;
3764 balance_cpu = i;
3765 }
3766
3767 load = target_load(i, load_idx);
3768 } else {
3769 load = source_load(i, load_idx);
3770 if (load > max_cpu_load)
3771 max_cpu_load = load;
3772 if (min_cpu_load > load)
3773 min_cpu_load = load;
3774 }
3775
3776 sgs->group_load += load;
3777 sgs->sum_nr_running += rq->nr_running;
3778 sgs->sum_weighted_load += weighted_cpuload(i);
3779
3780 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3781 }
3782
3783 /*
3784 * First idle cpu or the first cpu(busiest) in this sched group
3785 * is eligible for doing load balancing at this and above
3786 * domains. In the newly idle case, we will allow all the cpu's
3787 * to do the newly idle load balance.
3788 */
3789 if (idle != CPU_NEWLY_IDLE && local_group &&
3790 balance_cpu != this_cpu && balance) {
3791 *balance = 0;
3792 return;
3793 }
3794
3795 /* Adjust by relative CPU power of the group */
3796 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3797
3798
3799 /*
3800 * Consider the group unbalanced when the imbalance is larger
3801 * than the average weight of two tasks.
3802 *
3803 * APZ: with cgroup the avg task weight can vary wildly and
3804 * might not be a suitable number - should we keep a
3805 * normalized nr_running number somewhere that negates
3806 * the hierarchy?
3807 */
3808 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3809 group->cpu_power;
3810
3811 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3812 sgs->group_imb = 1;
3813
3814 sgs->group_capacity =
3815 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3816}
3817
3818/**
3819 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3820 * @sd: sched_domain whose statistics are to be updated.
3821 * @this_cpu: Cpu for which load balance is currently performed.
3822 * @idle: Idle status of this_cpu
3823 * @sd_idle: Idle status of the sched_domain containing group.
3824 * @cpus: Set of cpus considered for load balancing.
3825 * @balance: Should we balance.
3826 * @sds: variable to hold the statistics for this sched_domain.
3827 */
3828static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3829 enum cpu_idle_type idle, int *sd_idle,
3830 const struct cpumask *cpus, int *balance,
3831 struct sd_lb_stats *sds)
3832{
3833 struct sched_domain *child = sd->child;
3834 struct sched_group *group = sd->groups;
3835 struct sg_lb_stats sgs;
3836 int load_idx, prefer_sibling = 0;
3837
3838 if (child && child->flags & SD_PREFER_SIBLING)
3839 prefer_sibling = 1;
3840
3841 init_sd_power_savings_stats(sd, sds, idle);
3842 load_idx = get_sd_load_idx(sd, idle);
3843
3844 do {
3845 int local_group;
3846
3847 local_group = cpumask_test_cpu(this_cpu,
3848 sched_group_cpus(group));
3849 memset(&sgs, 0, sizeof(sgs));
3850 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3851 local_group, cpus, balance, &sgs);
3852
3853 if (local_group && balance && !(*balance))
3854 return;
3855
3856 sds->total_load += sgs.group_load;
3857 sds->total_pwr += group->cpu_power;
3858
3859 /*
3860 * In case the child domain prefers tasks go to siblings
3861 * first, lower the group capacity to one so that we'll try
3862 * and move all the excess tasks away.
3863 */
3864 if (prefer_sibling)
3865 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3866
3867 if (local_group) {
3868 sds->this_load = sgs.avg_load;
3869 sds->this = group;
3870 sds->this_nr_running = sgs.sum_nr_running;
3871 sds->this_load_per_task = sgs.sum_weighted_load;
3872 } else if (sgs.avg_load > sds->max_load &&
3873 (sgs.sum_nr_running > sgs.group_capacity ||
3874 sgs.group_imb)) {
3875 sds->max_load = sgs.avg_load;
3876 sds->busiest = group;
3877 sds->busiest_nr_running = sgs.sum_nr_running;
3878 sds->busiest_load_per_task = sgs.sum_weighted_load;
3879 sds->group_imb = sgs.group_imb;
3880 }
3881
3882 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3883 group = group->next;
3884 } while (group != sd->groups);
3885}
3886
3887/**
3888 * fix_small_imbalance - Calculate the minor imbalance that exists
3889 * amongst the groups of a sched_domain, during
3890 * load balancing.
3891 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3892 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3893 * @imbalance: Variable to store the imbalance.
3894 */
3895static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3896 int this_cpu, unsigned long *imbalance)
3897{
3898 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3899 unsigned int imbn = 2;
3900
3901 if (sds->this_nr_running) {
3902 sds->this_load_per_task /= sds->this_nr_running;
3903 if (sds->busiest_load_per_task >
3904 sds->this_load_per_task)
3905 imbn = 1;
3906 } else
3907 sds->this_load_per_task =
3908 cpu_avg_load_per_task(this_cpu);
3909
3910 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3911 sds->busiest_load_per_task * imbn) {
3912 *imbalance = sds->busiest_load_per_task;
3913 return;
3914 }
3915
3916 /*
3917 * OK, we don't have enough imbalance to justify moving tasks,
3918 * however we may be able to increase total CPU power used by
3919 * moving them.
3920 */
3921
3922 pwr_now += sds->busiest->cpu_power *
3923 min(sds->busiest_load_per_task, sds->max_load);
3924 pwr_now += sds->this->cpu_power *
3925 min(sds->this_load_per_task, sds->this_load);
3926 pwr_now /= SCHED_LOAD_SCALE;
3927
3928 /* Amount of load we'd subtract */
3929 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3930 sds->busiest->cpu_power;
3931 if (sds->max_load > tmp)
3932 pwr_move += sds->busiest->cpu_power *
3933 min(sds->busiest_load_per_task, sds->max_load - tmp);
3934
3935 /* Amount of load we'd add */
3936 if (sds->max_load * sds->busiest->cpu_power <
3937 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3938 tmp = (sds->max_load * sds->busiest->cpu_power) /
3939 sds->this->cpu_power;
3940 else
3941 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3942 sds->this->cpu_power;
3943 pwr_move += sds->this->cpu_power *
3944 min(sds->this_load_per_task, sds->this_load + tmp);
3945 pwr_move /= SCHED_LOAD_SCALE;
3946
3947 /* Move if we gain throughput */
3948 if (pwr_move > pwr_now)
3949 *imbalance = sds->busiest_load_per_task;
3950}
3951
3952/**
3953 * calculate_imbalance - Calculate the amount of imbalance present within the
3954 * groups of a given sched_domain during load balance.
3955 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3956 * @this_cpu: Cpu for which currently load balance is being performed.
3957 * @imbalance: The variable to store the imbalance.
3958 */
3959static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3960 unsigned long *imbalance)
3961{
3962 unsigned long max_pull;
3963 /*
3964 * In the presence of smp nice balancing, certain scenarios can have
3965 * max load less than avg load(as we skip the groups at or below
3966 * its cpu_power, while calculating max_load..)
3967 */
3968 if (sds->max_load < sds->avg_load) {
3969 *imbalance = 0;
3970 return fix_small_imbalance(sds, this_cpu, imbalance);
3971 }
3972
3973 /* Don't want to pull so many tasks that a group would go idle */
3974 max_pull = min(sds->max_load - sds->avg_load,
3975 sds->max_load - sds->busiest_load_per_task);
3976
3977 /* How much load to actually move to equalise the imbalance */
3978 *imbalance = min(max_pull * sds->busiest->cpu_power,
3979 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3980 / SCHED_LOAD_SCALE;
3981
3982 /*
3983 * if *imbalance is less than the average load per runnable task
3984 * there is no gaurantee that any tasks will be moved so we'll have
3985 * a think about bumping its value to force at least one task to be
3986 * moved
3987 */
3988 if (*imbalance < sds->busiest_load_per_task)
3989 return fix_small_imbalance(sds, this_cpu, imbalance);
3990
3991}
3992/******* find_busiest_group() helpers end here *********************/
3993
3994/**
3995 * find_busiest_group - Returns the busiest group within the sched_domain
3996 * if there is an imbalance. If there isn't an imbalance, and
3997 * the user has opted for power-savings, it returns a group whose
3998 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3999 * such a group exists.
4000 *
4001 * Also calculates the amount of weighted load which should be moved
4002 * to restore balance.
4003 *
4004 * @sd: The sched_domain whose busiest group is to be returned.
4005 * @this_cpu: The cpu for which load balancing is currently being performed.
4006 * @imbalance: Variable which stores amount of weighted load which should
4007 * be moved to restore balance/put a group to idle.
4008 * @idle: The idle status of this_cpu.
4009 * @sd_idle: The idleness of sd
4010 * @cpus: The set of CPUs under consideration for load-balancing.
4011 * @balance: Pointer to a variable indicating if this_cpu
4012 * is the appropriate cpu to perform load balancing at this_level.
4013 *
4014 * Returns: - the busiest group if imbalance exists.
4015 * - If no imbalance and user has opted for power-savings balance,
4016 * return the least loaded group whose CPUs can be
4017 * put to idle by rebalancing its tasks onto our group.
4018 */
4019static struct sched_group *
4020find_busiest_group(struct sched_domain *sd, int this_cpu,
4021 unsigned long *imbalance, enum cpu_idle_type idle,
4022 int *sd_idle, const struct cpumask *cpus, int *balance)
4023{
4024 struct sd_lb_stats sds;
4025
4026 memset(&sds, 0, sizeof(sds));
4027
4028 /*
4029 * Compute the various statistics relavent for load balancing at
4030 * this level.
4031 */
4032 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4033 balance, &sds);
4034
4035 /* Cases where imbalance does not exist from POV of this_cpu */
4036 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4037 * at this level.
4038 * 2) There is no busy sibling group to pull from.
4039 * 3) This group is the busiest group.
4040 * 4) This group is more busy than the avg busieness at this
4041 * sched_domain.
4042 * 5) The imbalance is within the specified limit.
4043 * 6) Any rebalance would lead to ping-pong
4044 */
4045 if (balance && !(*balance))
4046 goto ret;
4047
4048 if (!sds.busiest || sds.busiest_nr_running == 0)
4049 goto out_balanced;
4050
4051 if (sds.this_load >= sds.max_load)
4052 goto out_balanced;
4053
4054 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4055
4056 if (sds.this_load >= sds.avg_load)
4057 goto out_balanced;
4058
4059 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4060 goto out_balanced;
4061
4062 sds.busiest_load_per_task /= sds.busiest_nr_running;
4063 if (sds.group_imb)
4064 sds.busiest_load_per_task =
4065 min(sds.busiest_load_per_task, sds.avg_load);
4066
4067 /*
4068 * We're trying to get all the cpus to the average_load, so we don't
4069 * want to push ourselves above the average load, nor do we wish to
4070 * reduce the max loaded cpu below the average load, as either of these
4071 * actions would just result in more rebalancing later, and ping-pong
4072 * tasks around. Thus we look for the minimum possible imbalance.
4073 * Negative imbalances (*we* are more loaded than anyone else) will
4074 * be counted as no imbalance for these purposes -- we can't fix that
4075 * by pulling tasks to us. Be careful of negative numbers as they'll
4076 * appear as very large values with unsigned longs.
4077 */
4078 if (sds.max_load <= sds.busiest_load_per_task)
4079 goto out_balanced;
4080
4081 /* Looks like there is an imbalance. Compute it */
4082 calculate_imbalance(&sds, this_cpu, imbalance);
4083 return sds.busiest;
4084
4085out_balanced:
4086 /*
4087 * There is no obvious imbalance. But check if we can do some balancing
4088 * to save power.
4089 */
4090 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4091 return sds.busiest;
4092ret:
4093 *imbalance = 0;
4094 return NULL;
4095}
4096
4097/*
4098 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4099 */
4100static struct rq *
4101find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4102 unsigned long imbalance, const struct cpumask *cpus)
4103{
4104 struct rq *busiest = NULL, *rq;
4105 unsigned long max_load = 0;
4106 int i;
4107
4108 for_each_cpu(i, sched_group_cpus(group)) {
4109 unsigned long power = power_of(i);
4110 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4111 unsigned long wl;
4112
4113 if (!cpumask_test_cpu(i, cpus))
4114 continue;
4115
4116 rq = cpu_rq(i);
4117 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4118 wl /= power;
4119
4120 if (capacity && rq->nr_running == 1 && wl > imbalance)
4121 continue;
4122
4123 if (wl > max_load) {
4124 max_load = wl;
4125 busiest = rq;
4126 }
4127 }
4128
4129 return busiest;
4130}
4131
4132/*
4133 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4134 * so long as it is large enough.
4135 */
4136#define MAX_PINNED_INTERVAL 512
4137
4138/* Working cpumask for load_balance and load_balance_newidle. */
4139static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4140
4141/*
4142 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4143 * tasks if there is an imbalance.
4144 */
4145static int load_balance(int this_cpu, struct rq *this_rq,
4146 struct sched_domain *sd, enum cpu_idle_type idle,
4147 int *balance)
4148{
4149 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4150 struct sched_group *group;
4151 unsigned long imbalance;
4152 struct rq *busiest;
4153 unsigned long flags;
4154 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4155
4156 cpumask_setall(cpus);
4157
4158 /*
4159 * When power savings policy is enabled for the parent domain, idle
4160 * sibling can pick up load irrespective of busy siblings. In this case,
4161 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4162 * portraying it as CPU_NOT_IDLE.
4163 */
4164 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4165 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4166 sd_idle = 1;
4167
4168 schedstat_inc(sd, lb_count[idle]);
4169
4170redo:
4171 update_shares(sd);
4172 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4173 cpus, balance);
4174
4175 if (*balance == 0)
4176 goto out_balanced;
4177
4178 if (!group) {
4179 schedstat_inc(sd, lb_nobusyg[idle]);
4180 goto out_balanced;
4181 }
4182
4183 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4184 if (!busiest) {
4185 schedstat_inc(sd, lb_nobusyq[idle]);
4186 goto out_balanced;
4187 }
4188
4189 BUG_ON(busiest == this_rq);
4190
4191 schedstat_add(sd, lb_imbalance[idle], imbalance);
4192
4193 ld_moved = 0;
4194 if (busiest->nr_running > 1) {
4195 /*
4196 * Attempt to move tasks. If find_busiest_group has found
4197 * an imbalance but busiest->nr_running <= 1, the group is
4198 * still unbalanced. ld_moved simply stays zero, so it is
4199 * correctly treated as an imbalance.
4200 */
4201 local_irq_save(flags);
4202 double_rq_lock(this_rq, busiest);
4203 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4204 imbalance, sd, idle, &all_pinned);
4205 double_rq_unlock(this_rq, busiest);
4206 local_irq_restore(flags);
4207
4208 /*
4209 * some other cpu did the load balance for us.
4210 */
4211 if (ld_moved && this_cpu != smp_processor_id())
4212 resched_cpu(this_cpu);
4213
4214 /* All tasks on this runqueue were pinned by CPU affinity */
4215 if (unlikely(all_pinned)) {
4216 cpumask_clear_cpu(cpu_of(busiest), cpus);
4217 if (!cpumask_empty(cpus))
4218 goto redo;
4219 goto out_balanced;
4220 }
4221 }
4222
4223 if (!ld_moved) {
4224 schedstat_inc(sd, lb_failed[idle]);
4225 sd->nr_balance_failed++;
4226
4227 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4228
4229 spin_lock_irqsave(&busiest->lock, flags);
4230
4231 /* don't kick the migration_thread, if the curr
4232 * task on busiest cpu can't be moved to this_cpu
4233 */
4234 if (!cpumask_test_cpu(this_cpu,
4235 &busiest->curr->cpus_allowed)) {
4236 spin_unlock_irqrestore(&busiest->lock, flags);
4237 all_pinned = 1;
4238 goto out_one_pinned;
4239 }
4240
4241 if (!busiest->active_balance) {
4242 busiest->active_balance = 1;
4243 busiest->push_cpu = this_cpu;
4244 active_balance = 1;
4245 }
4246 spin_unlock_irqrestore(&busiest->lock, flags);
4247 if (active_balance)
4248 wake_up_process(busiest->migration_thread);
4249
4250 /*
4251 * We've kicked active balancing, reset the failure
4252 * counter.
4253 */
4254 sd->nr_balance_failed = sd->cache_nice_tries+1;
4255 }
4256 } else
4257 sd->nr_balance_failed = 0;
4258
4259 if (likely(!active_balance)) {
4260 /* We were unbalanced, so reset the balancing interval */
4261 sd->balance_interval = sd->min_interval;
4262 } else {
4263 /*
4264 * If we've begun active balancing, start to back off. This
4265 * case may not be covered by the all_pinned logic if there
4266 * is only 1 task on the busy runqueue (because we don't call
4267 * move_tasks).
4268 */
4269 if (sd->balance_interval < sd->max_interval)
4270 sd->balance_interval *= 2;
4271 }
4272
4273 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4274 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4275 ld_moved = -1;
4276
4277 goto out;
4278
4279out_balanced:
4280 schedstat_inc(sd, lb_balanced[idle]);
4281
4282 sd->nr_balance_failed = 0;
4283
4284out_one_pinned:
4285 /* tune up the balancing interval */
4286 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4287 (sd->balance_interval < sd->max_interval))
4288 sd->balance_interval *= 2;
4289
4290 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4291 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4292 ld_moved = -1;
4293 else
4294 ld_moved = 0;
4295out:
4296 if (ld_moved)
4297 update_shares(sd);
4298 return ld_moved;
4299}
4300
4301/*
4302 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4303 * tasks if there is an imbalance.
4304 *
4305 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4306 * this_rq is locked.
4307 */
4308static int
4309load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4310{
4311 struct sched_group *group;
4312 struct rq *busiest = NULL;
4313 unsigned long imbalance;
4314 int ld_moved = 0;
4315 int sd_idle = 0;
4316 int all_pinned = 0;
4317 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4318
4319 cpumask_setall(cpus);
4320
4321 /*
4322 * When power savings policy is enabled for the parent domain, idle
4323 * sibling can pick up load irrespective of busy siblings. In this case,
4324 * let the state of idle sibling percolate up as IDLE, instead of
4325 * portraying it as CPU_NOT_IDLE.
4326 */
4327 if (sd->flags & SD_SHARE_CPUPOWER &&
4328 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4329 sd_idle = 1;
4330
4331 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4332redo:
4333 update_shares_locked(this_rq, sd);
4334 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4335 &sd_idle, cpus, NULL);
4336 if (!group) {
4337 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4338 goto out_balanced;
4339 }
4340
4341 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4342 if (!busiest) {
4343 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4344 goto out_balanced;
4345 }
4346
4347 BUG_ON(busiest == this_rq);
4348
4349 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4350
4351 ld_moved = 0;
4352 if (busiest->nr_running > 1) {
4353 /* Attempt to move tasks */
4354 double_lock_balance(this_rq, busiest);
4355 /* this_rq->clock is already updated */
4356 update_rq_clock(busiest);
4357 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4358 imbalance, sd, CPU_NEWLY_IDLE,
4359 &all_pinned);
4360 double_unlock_balance(this_rq, busiest);
4361
4362 if (unlikely(all_pinned)) {
4363 cpumask_clear_cpu(cpu_of(busiest), cpus);
4364 if (!cpumask_empty(cpus))
4365 goto redo;
4366 }
4367 }
4368
4369 if (!ld_moved) {
4370 int active_balance = 0;
4371
4372 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4373 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4374 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4375 return -1;
4376
4377 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4378 return -1;
4379
4380 if (sd->nr_balance_failed++ < 2)
4381 return -1;
4382
4383 /*
4384 * The only task running in a non-idle cpu can be moved to this
4385 * cpu in an attempt to completely freeup the other CPU
4386 * package. The same method used to move task in load_balance()
4387 * have been extended for load_balance_newidle() to speedup
4388 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4389 *
4390 * The package power saving logic comes from
4391 * find_busiest_group(). If there are no imbalance, then
4392 * f_b_g() will return NULL. However when sched_mc={1,2} then
4393 * f_b_g() will select a group from which a running task may be
4394 * pulled to this cpu in order to make the other package idle.
4395 * If there is no opportunity to make a package idle and if
4396 * there are no imbalance, then f_b_g() will return NULL and no
4397 * action will be taken in load_balance_newidle().
4398 *
4399 * Under normal task pull operation due to imbalance, there
4400 * will be more than one task in the source run queue and
4401 * move_tasks() will succeed. ld_moved will be true and this
4402 * active balance code will not be triggered.
4403 */
4404
4405 /* Lock busiest in correct order while this_rq is held */
4406 double_lock_balance(this_rq, busiest);
4407
4408 /*
4409 * don't kick the migration_thread, if the curr
4410 * task on busiest cpu can't be moved to this_cpu
4411 */
4412 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4413 double_unlock_balance(this_rq, busiest);
4414 all_pinned = 1;
4415 return ld_moved;
4416 }
4417
4418 if (!busiest->active_balance) {
4419 busiest->active_balance = 1;
4420 busiest->push_cpu = this_cpu;
4421 active_balance = 1;
4422 }
4423
4424 double_unlock_balance(this_rq, busiest);
4425 /*
4426 * Should not call ttwu while holding a rq->lock
4427 */
4428 spin_unlock(&this_rq->lock);
4429 if (active_balance)
4430 wake_up_process(busiest->migration_thread);
4431 spin_lock(&this_rq->lock);
4432
4433 } else
4434 sd->nr_balance_failed = 0;
4435
4436 update_shares_locked(this_rq, sd);
4437 return ld_moved;
4438
4439out_balanced:
4440 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4441 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4442 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4443 return -1;
4444 sd->nr_balance_failed = 0;
4445
4446 return 0;
4447}
4448
4449/*
4450 * idle_balance is called by schedule() if this_cpu is about to become
4451 * idle. Attempts to pull tasks from other CPUs.
4452 */
4453static void idle_balance(int this_cpu, struct rq *this_rq)
4454{
4455 struct sched_domain *sd;
4456 int pulled_task = 0;
4457 unsigned long next_balance = jiffies + HZ;
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 break;
4475 }
4476 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4477 /*
4478 * We are going idle. next_balance may be set based on
4479 * a busy processor. So reset next_balance.
4480 */
4481 this_rq->next_balance = next_balance;
4482 }
4483}
4484
4485/*
4486 * active_load_balance is run by migration threads. It pushes running tasks
4487 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4488 * running on each physical CPU where possible, and avoids physical /
4489 * logical imbalances.
4490 *
4491 * Called with busiest_rq locked.
4492 */
4493static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4494{
4495 int target_cpu = busiest_rq->push_cpu;
4496 struct sched_domain *sd;
4497 struct rq *target_rq;
4498
4499 /* Is there any task to move? */
4500 if (busiest_rq->nr_running <= 1)
4501 return;
4502
4503 target_rq = cpu_rq(target_cpu);
4504
4505 /*
4506 * This condition is "impossible", if it occurs
4507 * we need to fix it. Originally reported by
4508 * Bjorn Helgaas on a 128-cpu setup.
4509 */
4510 BUG_ON(busiest_rq == target_rq);
4511
4512 /* move a task from busiest_rq to target_rq */
4513 double_lock_balance(busiest_rq, target_rq);
4514 update_rq_clock(busiest_rq);
4515 update_rq_clock(target_rq);
4516
4517 /* Search for an sd spanning us and the target CPU. */
4518 for_each_domain(target_cpu, sd) {
4519 if ((sd->flags & SD_LOAD_BALANCE) &&
4520 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4521 break;
4522 }
4523
4524 if (likely(sd)) {
4525 schedstat_inc(sd, alb_count);
4526
4527 if (move_one_task(target_rq, target_cpu, busiest_rq,
4528 sd, CPU_IDLE))
4529 schedstat_inc(sd, alb_pushed);
4530 else
4531 schedstat_inc(sd, alb_failed);
4532 }
4533 double_unlock_balance(busiest_rq, target_rq);
4534}
4535
4536#ifdef CONFIG_NO_HZ
4537static struct {
4538 atomic_t load_balancer;
4539 cpumask_var_t cpu_mask;
4540 cpumask_var_t ilb_grp_nohz_mask;
4541} nohz ____cacheline_aligned = {
4542 .load_balancer = ATOMIC_INIT(-1),
4543};
4544
4545int get_nohz_load_balancer(void)
4546{
4547 return atomic_read(&nohz.load_balancer);
4548}
4549
4550#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4551/**
4552 * lowest_flag_domain - Return lowest sched_domain containing flag.
4553 * @cpu: The cpu whose lowest level of sched domain is to
4554 * be returned.
4555 * @flag: The flag to check for the lowest sched_domain
4556 * for the given cpu.
4557 *
4558 * Returns the lowest sched_domain of a cpu which contains the given flag.
4559 */
4560static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4561{
4562 struct sched_domain *sd;
4563
4564 for_each_domain(cpu, sd)
4565 if (sd && (sd->flags & flag))
4566 break;
4567
4568 return sd;
4569}
4570
4571/**
4572 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4573 * @cpu: The cpu whose domains we're iterating over.
4574 * @sd: variable holding the value of the power_savings_sd
4575 * for cpu.
4576 * @flag: The flag to filter the sched_domains to be iterated.
4577 *
4578 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4579 * set, starting from the lowest sched_domain to the highest.
4580 */
4581#define for_each_flag_domain(cpu, sd, flag) \
4582 for (sd = lowest_flag_domain(cpu, flag); \
4583 (sd && (sd->flags & flag)); sd = sd->parent)
4584
4585/**
4586 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4587 * @ilb_group: group to be checked for semi-idleness
4588 *
4589 * Returns: 1 if the group is semi-idle. 0 otherwise.
4590 *
4591 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4592 * and atleast one non-idle CPU. This helper function checks if the given
4593 * sched_group is semi-idle or not.
4594 */
4595static inline int is_semi_idle_group(struct sched_group *ilb_group)
4596{
4597 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4598 sched_group_cpus(ilb_group));
4599
4600 /*
4601 * A sched_group is semi-idle when it has atleast one busy cpu
4602 * and atleast one idle cpu.
4603 */
4604 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4605 return 0;
4606
4607 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4608 return 0;
4609
4610 return 1;
4611}
4612/**
4613 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4614 * @cpu: The cpu which is nominating a new idle_load_balancer.
4615 *
4616 * Returns: Returns the id of the idle load balancer if it exists,
4617 * Else, returns >= nr_cpu_ids.
4618 *
4619 * This algorithm picks the idle load balancer such that it belongs to a
4620 * semi-idle powersavings sched_domain. The idea is to try and avoid
4621 * completely idle packages/cores just for the purpose of idle load balancing
4622 * when there are other idle cpu's which are better suited for that job.
4623 */
4624static int find_new_ilb(int cpu)
4625{
4626 struct sched_domain *sd;
4627 struct sched_group *ilb_group;
4628
4629 /*
4630 * Have idle load balancer selection from semi-idle packages only
4631 * when power-aware load balancing is enabled
4632 */
4633 if (!(sched_smt_power_savings || sched_mc_power_savings))
4634 goto out_done;
4635
4636 /*
4637 * Optimize for the case when we have no idle CPUs or only one
4638 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4639 */
4640 if (cpumask_weight(nohz.cpu_mask) < 2)
4641 goto out_done;
4642
4643 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4644 ilb_group = sd->groups;
4645
4646 do {
4647 if (is_semi_idle_group(ilb_group))
4648 return cpumask_first(nohz.ilb_grp_nohz_mask);
4649
4650 ilb_group = ilb_group->next;
4651
4652 } while (ilb_group != sd->groups);
4653 }
4654
4655out_done:
4656 return cpumask_first(nohz.cpu_mask);
4657}
4658#else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4659static inline int find_new_ilb(int call_cpu)
4660{
4661 return cpumask_first(nohz.cpu_mask);
4662}
4663#endif
4664
4665/*
4666 * This routine will try to nominate the ilb (idle load balancing)
4667 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4668 * load balancing on behalf of all those cpus. If all the cpus in the system
4669 * go into this tickless mode, then there will be no ilb owner (as there is
4670 * no need for one) and all the cpus will sleep till the next wakeup event
4671 * arrives...
4672 *
4673 * For the ilb owner, tick is not stopped. And this tick will be used
4674 * for idle load balancing. ilb owner will still be part of
4675 * nohz.cpu_mask..
4676 *
4677 * While stopping the tick, this cpu will become the ilb owner if there
4678 * is no other owner. And will be the owner till that cpu becomes busy
4679 * or if all cpus in the system stop their ticks at which point
4680 * there is no need for ilb owner.
4681 *
4682 * When the ilb owner becomes busy, it nominates another owner, during the
4683 * next busy scheduler_tick()
4684 */
4685int select_nohz_load_balancer(int stop_tick)
4686{
4687 int cpu = smp_processor_id();
4688
4689 if (stop_tick) {
4690 cpu_rq(cpu)->in_nohz_recently = 1;
4691
4692 if (!cpu_active(cpu)) {
4693 if (atomic_read(&nohz.load_balancer) != cpu)
4694 return 0;
4695
4696 /*
4697 * If we are going offline and still the leader,
4698 * give up!
4699 */
4700 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4701 BUG();
4702
4703 return 0;
4704 }
4705
4706 cpumask_set_cpu(cpu, nohz.cpu_mask);
4707
4708 /* time for ilb owner also to sleep */
4709 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4710 if (atomic_read(&nohz.load_balancer) == cpu)
4711 atomic_set(&nohz.load_balancer, -1);
4712 return 0;
4713 }
4714
4715 if (atomic_read(&nohz.load_balancer) == -1) {
4716 /* make me the ilb owner */
4717 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4718 return 1;
4719 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4720 int new_ilb;
4721
4722 if (!(sched_smt_power_savings ||
4723 sched_mc_power_savings))
4724 return 1;
4725 /*
4726 * Check to see if there is a more power-efficient
4727 * ilb.
4728 */
4729 new_ilb = find_new_ilb(cpu);
4730 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4731 atomic_set(&nohz.load_balancer, -1);
4732 resched_cpu(new_ilb);
4733 return 0;
4734 }
4735 return 1;
4736 }
4737 } else {
4738 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4739 return 0;
4740
4741 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4742
4743 if (atomic_read(&nohz.load_balancer) == cpu)
4744 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4745 BUG();
4746 }
4747 return 0;
4748}
4749#endif
4750
4751static DEFINE_SPINLOCK(balancing);
4752
4753/*
4754 * It checks each scheduling domain to see if it is due to be balanced,
4755 * and initiates a balancing operation if so.
4756 *
4757 * Balancing parameters are set up in arch_init_sched_domains.
4758 */
4759static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4760{
4761 int balance = 1;
4762 struct rq *rq = cpu_rq(cpu);
4763 unsigned long interval;
4764 struct sched_domain *sd;
4765 /* Earliest time when we have to do rebalance again */
4766 unsigned long next_balance = jiffies + 60*HZ;
4767 int update_next_balance = 0;
4768 int need_serialize;
4769
4770 for_each_domain(cpu, sd) {
4771 if (!(sd->flags & SD_LOAD_BALANCE))
4772 continue;
4773
4774 interval = sd->balance_interval;
4775 if (idle != CPU_IDLE)
4776 interval *= sd->busy_factor;
4777
4778 /* scale ms to jiffies */
4779 interval = msecs_to_jiffies(interval);
4780 if (unlikely(!interval))
4781 interval = 1;
4782 if (interval > HZ*NR_CPUS/10)
4783 interval = HZ*NR_CPUS/10;
4784
4785 need_serialize = sd->flags & SD_SERIALIZE;
4786
4787 if (need_serialize) {
4788 if (!spin_trylock(&balancing))
4789 goto out;
4790 }
4791
4792 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4793 if (load_balance(cpu, rq, sd, idle, &balance)) {
4794 /*
4795 * We've pulled tasks over so either we're no
4796 * longer idle, or one of our SMT siblings is
4797 * not idle.
4798 */
4799 idle = CPU_NOT_IDLE;
4800 }
4801 sd->last_balance = jiffies;
4802 }
4803 if (need_serialize)
4804 spin_unlock(&balancing);
4805out:
4806 if (time_after(next_balance, sd->last_balance + interval)) {
4807 next_balance = sd->last_balance + interval;
4808 update_next_balance = 1;
4809 }
4810
4811 /*
4812 * Stop the load balance at this level. There is another
4813 * CPU in our sched group which is doing load balancing more
4814 * actively.
4815 */
4816 if (!balance)
4817 break;
4818 }
4819
4820 /*
4821 * next_balance will be updated only when there is a need.
4822 * When the cpu is attached to null domain for ex, it will not be
4823 * updated.
4824 */
4825 if (likely(update_next_balance))
4826 rq->next_balance = next_balance;
4827}
4828
4829/*
4830 * run_rebalance_domains is triggered when needed from the scheduler tick.
4831 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4832 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4833 */
4834static void run_rebalance_domains(struct softirq_action *h)
4835{
4836 int this_cpu = smp_processor_id();
4837 struct rq *this_rq = cpu_rq(this_cpu);
4838 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4839 CPU_IDLE : CPU_NOT_IDLE;
4840
4841 rebalance_domains(this_cpu, idle);
4842
4843#ifdef CONFIG_NO_HZ
4844 /*
4845 * If this cpu is the owner for idle load balancing, then do the
4846 * balancing on behalf of the other idle cpus whose ticks are
4847 * stopped.
4848 */
4849 if (this_rq->idle_at_tick &&
4850 atomic_read(&nohz.load_balancer) == this_cpu) {
4851 struct rq *rq;
4852 int balance_cpu;
4853
4854 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4855 if (balance_cpu == this_cpu)
4856 continue;
4857
4858 /*
4859 * If this cpu gets work to do, stop the load balancing
4860 * work being done for other cpus. Next load
4861 * balancing owner will pick it up.
4862 */
4863 if (need_resched())
4864 break;
4865
4866 rebalance_domains(balance_cpu, CPU_IDLE);
4867
4868 rq = cpu_rq(balance_cpu);
4869 if (time_after(this_rq->next_balance, rq->next_balance))
4870 this_rq->next_balance = rq->next_balance;
4871 }
4872 }
4873#endif
4874}
4875
4876static inline int on_null_domain(int cpu)
4877{
4878 return !rcu_dereference(cpu_rq(cpu)->sd);
4879}
4880
4881/*
4882 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4883 *
4884 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4885 * idle load balancing owner or decide to stop the periodic load balancing,
4886 * if the whole system is idle.
4887 */
4888static inline void trigger_load_balance(struct rq *rq, int cpu)
4889{
4890#ifdef CONFIG_NO_HZ
4891 /*
4892 * If we were in the nohz mode recently and busy at the current
4893 * scheduler tick, then check if we need to nominate new idle
4894 * load balancer.
4895 */
4896 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4897 rq->in_nohz_recently = 0;
4898
4899 if (atomic_read(&nohz.load_balancer) == cpu) {
4900 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4901 atomic_set(&nohz.load_balancer, -1);
4902 }
4903
4904 if (atomic_read(&nohz.load_balancer) == -1) {
4905 int ilb = find_new_ilb(cpu);
4906
4907 if (ilb < nr_cpu_ids)
4908 resched_cpu(ilb);
4909 }
4910 }
4911
4912 /*
4913 * If this cpu is idle and doing idle load balancing for all the
4914 * cpus with ticks stopped, is it time for that to stop?
4915 */
4916 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4917 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4918 resched_cpu(cpu);
4919 return;
4920 }
4921
4922 /*
4923 * If this cpu is idle and the idle load balancing is done by
4924 * someone else, then no need raise the SCHED_SOFTIRQ
4925 */
4926 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4927 cpumask_test_cpu(cpu, nohz.cpu_mask))
4928 return;
4929#endif
4930 /* Don't need to rebalance while attached to NULL domain */
4931 if (time_after_eq(jiffies, rq->next_balance) &&
4932 likely(!on_null_domain(cpu)))
4933 raise_softirq(SCHED_SOFTIRQ);
4934}
4935
4936#else /* CONFIG_SMP */
4937
4938/*
4939 * on UP we do not need to balance between CPUs:
4940 */
4941static inline void idle_balance(int cpu, struct rq *rq)
4942{
4943}
4944
4945#endif 3202#endif
4946 3203
4947DEFINE_PER_CPU(struct kernel_stat, kstat); 3204DEFINE_PER_CPU(struct kernel_stat, kstat);
@@ -5073,8 +3330,13 @@ static void account_guest_time(struct task_struct *p, cputime_t cputime,
5073 p->gtime = cputime_add(p->gtime, cputime); 3330 p->gtime = cputime_add(p->gtime, cputime);
5074 3331
5075 /* Add guest time to cpustat. */ 3332 /* Add guest time to cpustat. */
5076 cpustat->user = cputime64_add(cpustat->user, tmp); 3333 if (TASK_NICE(p) > 0) {
5077 cpustat->guest = cputime64_add(cpustat->guest, tmp); 3334 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3335 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3336 } else {
3337 cpustat->user = cputime64_add(cpustat->user, tmp);
3338 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3339 }
5078} 3340}
5079 3341
5080/* 3342/*
@@ -5189,60 +3451,86 @@ void account_idle_ticks(unsigned long ticks)
5189 * Use precise platform statistics if available: 3451 * Use precise platform statistics if available:
5190 */ 3452 */
5191#ifdef CONFIG_VIRT_CPU_ACCOUNTING 3453#ifdef CONFIG_VIRT_CPU_ACCOUNTING
5192cputime_t task_utime(struct task_struct *p) 3454void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5193{ 3455{
5194 return p->utime; 3456 *ut = p->utime;
3457 *st = p->stime;
5195} 3458}
5196 3459
5197cputime_t task_stime(struct task_struct *p) 3460void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5198{ 3461{
5199 return p->stime; 3462 struct task_cputime cputime;
3463
3464 thread_group_cputime(p, &cputime);
3465
3466 *ut = cputime.utime;
3467 *st = cputime.stime;
5200} 3468}
5201#else 3469#else
5202cputime_t task_utime(struct task_struct *p) 3470
3471#ifndef nsecs_to_cputime
3472# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3473#endif
3474
3475void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5203{ 3476{
5204 clock_t utime = cputime_to_clock_t(p->utime), 3477 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5205 total = utime + cputime_to_clock_t(p->stime);
5206 u64 temp;
5207 3478
5208 /* 3479 /*
5209 * Use CFS's precise accounting: 3480 * Use CFS's precise accounting:
5210 */ 3481 */
5211 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime); 3482 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5212 3483
5213 if (total) { 3484 if (total) {
5214 temp *= utime; 3485 u64 temp;
3486
3487 temp = (u64)(rtime * utime);
5215 do_div(temp, total); 3488 do_div(temp, total);
5216 } 3489 utime = (cputime_t)temp;
5217 utime = (clock_t)temp; 3490 } else
3491 utime = rtime;
3492
3493 /*
3494 * Compare with previous values, to keep monotonicity:
3495 */
3496 p->prev_utime = max(p->prev_utime, utime);
3497 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5218 3498
5219 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime)); 3499 *ut = p->prev_utime;
5220 return p->prev_utime; 3500 *st = p->prev_stime;
5221} 3501}
5222 3502
5223cputime_t task_stime(struct task_struct *p) 3503/*
3504 * Must be called with siglock held.
3505 */
3506void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5224{ 3507{
5225 clock_t stime; 3508 struct signal_struct *sig = p->signal;
3509 struct task_cputime cputime;
3510 cputime_t rtime, utime, total;
5226 3511
5227 /* 3512 thread_group_cputime(p, &cputime);
5228 * Use CFS's precise accounting. (we subtract utime from
5229 * the total, to make sure the total observed by userspace
5230 * grows monotonically - apps rely on that):
5231 */
5232 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5233 cputime_to_clock_t(task_utime(p));
5234 3513
5235 if (stime >= 0) 3514 total = cputime_add(cputime.utime, cputime.stime);
5236 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime)); 3515 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5237 3516
5238 return p->prev_stime; 3517 if (total) {
5239} 3518 u64 temp;
5240#endif
5241 3519
5242inline cputime_t task_gtime(struct task_struct *p) 3520 temp = (u64)(rtime * cputime.utime);
5243{ 3521 do_div(temp, total);
5244 return p->gtime; 3522 utime = (cputime_t)temp;
3523 } else
3524 utime = rtime;
3525
3526 sig->prev_utime = max(sig->prev_utime, utime);
3527 sig->prev_stime = max(sig->prev_stime,
3528 cputime_sub(rtime, sig->prev_utime));
3529
3530 *ut = sig->prev_utime;
3531 *st = sig->prev_stime;
5245} 3532}
3533#endif
5246 3534
5247/* 3535/*
5248 * This function gets called by the timer code, with HZ frequency. 3536 * This function gets called by the timer code, with HZ frequency.
@@ -5261,7 +3549,7 @@ void scheduler_tick(void)
5261 3549
5262 TS_TICK_START(current); 3550 TS_TICK_START(current);
5263 3551
5264 spin_lock(&rq->lock); 3552 raw_spin_lock(&rq->lock);
5265 update_rq_clock(rq); 3553 update_rq_clock(rq);
5266 update_cpu_load(rq); 3554 update_cpu_load(rq);
5267 curr->sched_class->task_tick(rq, curr, 0); 3555 curr->sched_class->task_tick(rq, curr, 0);
@@ -5269,9 +3557,9 @@ void scheduler_tick(void)
5269 /* litmus_tick may force current to resched */ 3557 /* litmus_tick may force current to resched */
5270 litmus_tick(rq, curr); 3558 litmus_tick(rq, curr);
5271 3559
5272 spin_unlock(&rq->lock); 3560 raw_spin_unlock(&rq->lock);
5273 3561
5274 perf_event_task_tick(curr, cpu); 3562 perf_event_task_tick(curr);
5275 3563
5276#ifdef CONFIG_SMP 3564#ifdef CONFIG_SMP
5277 rq->idle_at_tick = idle_cpu(cpu); 3565 rq->idle_at_tick = idle_cpu(cpu);
@@ -5385,13 +3673,14 @@ static inline void schedule_debug(struct task_struct *prev)
5385#endif 3673#endif
5386} 3674}
5387 3675
5388static void put_prev_task(struct rq *rq, struct task_struct *p) 3676static void put_prev_task(struct rq *rq, struct task_struct *prev)
5389{ 3677{
5390 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime; 3678 if (prev->state == TASK_RUNNING) {
3679 u64 runtime = prev->se.sum_exec_runtime;
5391 3680
5392 update_avg(&p->se.avg_running, runtime); 3681 runtime -= prev->se.prev_sum_exec_runtime;
3682 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5393 3683
5394 if (p->state == TASK_RUNNING) {
5395 /* 3684 /*
5396 * In order to avoid avg_overlap growing stale when we are 3685 * In order to avoid avg_overlap growing stale when we are
5397 * indeed overlapping and hence not getting put to sleep, grow 3686 * indeed overlapping and hence not getting put to sleep, grow
@@ -5401,12 +3690,9 @@ static void put_prev_task(struct rq *rq, struct task_struct *p)
5401 * correlates to the amount of cache footprint a task can 3690 * correlates to the amount of cache footprint a task can
5402 * build up. 3691 * build up.
5403 */ 3692 */
5404 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost); 3693 update_avg(&prev->se.avg_overlap, runtime);
5405 update_avg(&p->se.avg_overlap, runtime);
5406 } else {
5407 update_avg(&p->se.avg_running, 0);
5408 } 3694 }
5409 p->sched_class->put_prev_task(rq, p); 3695 prev->sched_class->put_prev_task(rq, prev);
5410} 3696}
5411 3697
5412/* 3698/*
@@ -5477,7 +3763,7 @@ need_resched_nonpreemptible:
5477 if (sched_feat(HRTICK)) 3763 if (sched_feat(HRTICK))
5478 hrtick_clear(rq); 3764 hrtick_clear(rq);
5479 3765
5480 spin_lock_irq(&rq->lock); 3766 raw_spin_lock_irq(&rq->lock);
5481 update_rq_clock(rq); 3767 update_rq_clock(rq);
5482 clear_tsk_need_resched(prev); 3768 clear_tsk_need_resched(prev);
5483 3769
@@ -5499,7 +3785,7 @@ need_resched_nonpreemptible:
5499 3785
5500 if (likely(prev != next)) { 3786 if (likely(prev != next)) {
5501 sched_info_switch(prev, next); 3787 sched_info_switch(prev, next);
5502 perf_event_task_sched_out(prev, next, cpu); 3788 perf_event_task_sched_out(prev, next);
5503 3789
5504 rq->nr_switches++; 3790 rq->nr_switches++;
5505 rq->curr = next; 3791 rq->curr = next;
@@ -5517,7 +3803,7 @@ need_resched_nonpreemptible:
5517 rq = cpu_rq(cpu); 3803 rq = cpu_rq(cpu);
5518 } else { 3804 } else {
5519 TS_SCHED_END(prev); 3805 TS_SCHED_END(prev);
5520 spin_unlock_irq(&rq->lock); 3806 raw_spin_unlock_irq(&rq->lock);
5521 } 3807 }
5522 3808
5523 sched_trace_task_switch_to(current); 3809 sched_trace_task_switch_to(current);
@@ -5525,11 +3811,12 @@ need_resched_nonpreemptible:
5525 post_schedule(rq); 3811 post_schedule(rq);
5526 3812
5527 if (unlikely(reacquire_kernel_lock(current) < 0)) { 3813 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3814 prev = rq->curr;
3815 switch_count = &prev->nivcsw;
5528 goto need_resched_nonpreemptible; 3816 goto need_resched_nonpreemptible;
5529 } 3817 }
5530 3818
5531 preempt_enable_no_resched(); 3819 preempt_enable_no_resched();
5532
5533 if (need_resched()) 3820 if (need_resched())
5534 goto need_resched; 3821 goto need_resched;
5535 3822
@@ -5538,7 +3825,7 @@ need_resched_nonpreemptible:
5538} 3825}
5539EXPORT_SYMBOL(schedule); 3826EXPORT_SYMBOL(schedule);
5540 3827
5541#ifdef CONFIG_SMP 3828#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5542/* 3829/*
5543 * Look out! "owner" is an entirely speculative pointer 3830 * Look out! "owner" is an entirely speculative pointer
5544 * access and not reliable. 3831 * access and not reliable.
@@ -5558,7 +3845,7 @@ int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5558 * the mutex owner just released it and exited. 3845 * the mutex owner just released it and exited.
5559 */ 3846 */
5560 if (probe_kernel_address(&owner->cpu, cpu)) 3847 if (probe_kernel_address(&owner->cpu, cpu))
5561 goto out; 3848 return 0;
5562#else 3849#else
5563 cpu = owner->cpu; 3850 cpu = owner->cpu;
5564#endif 3851#endif
@@ -5568,14 +3855,14 @@ int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5568 * the cpu field may no longer be valid. 3855 * the cpu field may no longer be valid.
5569 */ 3856 */
5570 if (cpu >= nr_cpumask_bits) 3857 if (cpu >= nr_cpumask_bits)
5571 goto out; 3858 return 0;
5572 3859
5573 /* 3860 /*
5574 * We need to validate that we can do a 3861 * We need to validate that we can do a
5575 * get_cpu() and that we have the percpu area. 3862 * get_cpu() and that we have the percpu area.
5576 */ 3863 */
5577 if (!cpu_online(cpu)) 3864 if (!cpu_online(cpu))
5578 goto out; 3865 return 0;
5579 3866
5580 rq = cpu_rq(cpu); 3867 rq = cpu_rq(cpu);
5581 3868
@@ -5594,7 +3881,7 @@ int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5594 3881
5595 cpu_relax(); 3882 cpu_relax();
5596 } 3883 }
5597out: 3884
5598 return 1; 3885 return 1;
5599} 3886}
5600#endif 3887#endif
@@ -5953,14 +4240,15 @@ EXPORT_SYMBOL(wait_for_completion_killable);
5953 */ 4240 */
5954bool try_wait_for_completion(struct completion *x) 4241bool try_wait_for_completion(struct completion *x)
5955{ 4242{
4243 unsigned long flags;
5956 int ret = 1; 4244 int ret = 1;
5957 4245
5958 spin_lock_irq(&x->wait.lock); 4246 spin_lock_irqsave(&x->wait.lock, flags);
5959 if (!x->done) 4247 if (!x->done)
5960 ret = 0; 4248 ret = 0;
5961 else 4249 else
5962 x->done--; 4250 x->done--;
5963 spin_unlock_irq(&x->wait.lock); 4251 spin_unlock_irqrestore(&x->wait.lock, flags);
5964 return ret; 4252 return ret;
5965} 4253}
5966EXPORT_SYMBOL(try_wait_for_completion); 4254EXPORT_SYMBOL(try_wait_for_completion);
@@ -5975,12 +4263,13 @@ EXPORT_SYMBOL(try_wait_for_completion);
5975 */ 4263 */
5976bool completion_done(struct completion *x) 4264bool completion_done(struct completion *x)
5977{ 4265{
4266 unsigned long flags;
5978 int ret = 1; 4267 int ret = 1;
5979 4268
5980 spin_lock_irq(&x->wait.lock); 4269 spin_lock_irqsave(&x->wait.lock, flags);
5981 if (!x->done) 4270 if (!x->done)
5982 ret = 0; 4271 ret = 0;
5983 spin_unlock_irq(&x->wait.lock); 4272 spin_unlock_irqrestore(&x->wait.lock, flags);
5984 return ret; 4273 return ret;
5985} 4274}
5986EXPORT_SYMBOL(completion_done); 4275EXPORT_SYMBOL(completion_done);
@@ -6048,7 +4337,7 @@ void rt_mutex_setprio(struct task_struct *p, int prio)
6048 unsigned long flags; 4337 unsigned long flags;
6049 int oldprio, on_rq, running; 4338 int oldprio, on_rq, running;
6050 struct rq *rq; 4339 struct rq *rq;
6051 const struct sched_class *prev_class = p->sched_class; 4340 const struct sched_class *prev_class;
6052 4341
6053 BUG_ON(prio < 0 || prio > MAX_PRIO); 4342 BUG_ON(prio < 0 || prio > MAX_PRIO);
6054 4343
@@ -6056,6 +4345,7 @@ void rt_mutex_setprio(struct task_struct *p, int prio)
6056 update_rq_clock(rq); 4345 update_rq_clock(rq);
6057 4346
6058 oldprio = p->prio; 4347 oldprio = p->prio;
4348 prev_class = p->sched_class;
6059 on_rq = p->se.on_rq; 4349 on_rq = p->se.on_rq;
6060 running = task_current(rq, p); 4350 running = task_current(rq, p);
6061 if (on_rq) 4351 if (on_rq)
@@ -6073,7 +4363,7 @@ void rt_mutex_setprio(struct task_struct *p, int prio)
6073 if (running) 4363 if (running)
6074 p->sched_class->set_curr_task(rq); 4364 p->sched_class->set_curr_task(rq);
6075 if (on_rq) { 4365 if (on_rq) {
6076 enqueue_task(rq, p, 0); 4366 enqueue_task(rq, p, 0, oldprio < prio);
6077 4367
6078 check_class_changed(rq, p, prev_class, oldprio, running); 4368 check_class_changed(rq, p, prev_class, oldprio, running);
6079 } 4369 }
@@ -6117,7 +4407,7 @@ void set_user_nice(struct task_struct *p, long nice)
6117 delta = p->prio - old_prio; 4407 delta = p->prio - old_prio;
6118 4408
6119 if (on_rq) { 4409 if (on_rq) {
6120 enqueue_task(rq, p, 0); 4410 enqueue_task(rq, p, 0, false);
6121 /* 4411 /*
6122 * If the task increased its priority or is running and 4412 * If the task increased its priority or is running and
6123 * lowered its priority, then reschedule its CPU: 4413 * lowered its priority, then reschedule its CPU:
@@ -6140,7 +4430,7 @@ int can_nice(const struct task_struct *p, const int nice)
6140 /* convert nice value [19,-20] to rlimit style value [1,40] */ 4430 /* convert nice value [19,-20] to rlimit style value [1,40] */
6141 int nice_rlim = 20 - nice; 4431 int nice_rlim = 20 - nice;
6142 4432
6143 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || 4433 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6144 capable(CAP_SYS_NICE)); 4434 capable(CAP_SYS_NICE));
6145} 4435}
6146 4436
@@ -6243,25 +4533,16 @@ __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6243 BUG_ON(p->se.on_rq); 4533 BUG_ON(p->se.on_rq);
6244 4534
6245 p->policy = policy; 4535 p->policy = policy;
6246 switch (p->policy) {
6247 case SCHED_NORMAL:
6248 case SCHED_BATCH:
6249 case SCHED_IDLE:
6250 p->sched_class = &fair_sched_class;
6251 break;
6252 case SCHED_FIFO:
6253 case SCHED_RR:
6254 p->sched_class = &rt_sched_class;
6255 break;
6256 case SCHED_LITMUS:
6257 p->sched_class = &litmus_sched_class;
6258 break;
6259 }
6260
6261 p->rt_priority = prio; 4536 p->rt_priority = prio;
6262 p->normal_prio = normal_prio(p); 4537 p->normal_prio = normal_prio(p);
6263 /* we are holding p->pi_lock already */ 4538 /* we are holding p->pi_lock already */
6264 p->prio = rt_mutex_getprio(p); 4539 p->prio = rt_mutex_getprio(p);
4540 if (p->policy == SCHED_LITMUS)
4541 p->sched_class = &litmus_sched_class;
4542 else if (rt_prio(p->prio))
4543 p->sched_class = &rt_sched_class;
4544 else
4545 p->sched_class = &fair_sched_class;
6265 set_load_weight(p); 4546 set_load_weight(p);
6266} 4547}
6267 4548
@@ -6286,7 +4567,7 @@ static int __sched_setscheduler(struct task_struct *p, int policy,
6286{ 4567{
6287 int retval, oldprio, oldpolicy = -1, on_rq, running; 4568 int retval, oldprio, oldpolicy = -1, on_rq, running;
6288 unsigned long flags; 4569 unsigned long flags;
6289 const struct sched_class *prev_class = p->sched_class; 4570 const struct sched_class *prev_class;
6290 struct rq *rq; 4571 struct rq *rq;
6291 int reset_on_fork; 4572 int reset_on_fork;
6292 4573
@@ -6330,7 +4611,7 @@ recheck:
6330 4611
6331 if (!lock_task_sighand(p, &flags)) 4612 if (!lock_task_sighand(p, &flags))
6332 return -ESRCH; 4613 return -ESRCH;
6333 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; 4614 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
6334 unlock_task_sighand(p, &flags); 4615 unlock_task_sighand(p, &flags);
6335 4616
6336 /* can't set/change the rt policy */ 4617 /* can't set/change the rt policy */
@@ -6384,7 +4665,7 @@ recheck:
6384 * make sure no PI-waiters arrive (or leave) while we are 4665 * make sure no PI-waiters arrive (or leave) while we are
6385 * changing the priority of the task: 4666 * changing the priority of the task:
6386 */ 4667 */
6387 spin_lock_irqsave(&p->pi_lock, flags); 4668 raw_spin_lock_irqsave(&p->pi_lock, flags);
6388 /* 4669 /*
6389 * To be able to change p->policy safely, the apropriate 4670 * To be able to change p->policy safely, the apropriate
6390 * runqueue lock must be held. 4671 * runqueue lock must be held.
@@ -6394,7 +4675,7 @@ recheck:
6394 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 4675 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6395 policy = oldpolicy = -1; 4676 policy = oldpolicy = -1;
6396 __task_rq_unlock(rq); 4677 __task_rq_unlock(rq);
6397 spin_unlock_irqrestore(&p->pi_lock, flags); 4678 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6398 goto recheck; 4679 goto recheck;
6399 } 4680 }
6400 update_rq_clock(rq); 4681 update_rq_clock(rq);
@@ -6411,6 +4692,7 @@ recheck:
6411 litmus_exit_task(p); 4692 litmus_exit_task(p);
6412 4693
6413 oldprio = p->prio; 4694 oldprio = p->prio;
4695 prev_class = p->sched_class;
6414 __setscheduler(rq, p, policy, param->sched_priority); 4696 __setscheduler(rq, p, policy, param->sched_priority);
6415 4697
6416 if (policy == SCHED_LITMUS) { 4698 if (policy == SCHED_LITMUS) {
@@ -6427,7 +4709,7 @@ recheck:
6427 check_class_changed(rq, p, prev_class, oldprio, running); 4709 check_class_changed(rq, p, prev_class, oldprio, running);
6428 } 4710 }
6429 __task_rq_unlock(rq); 4711 __task_rq_unlock(rq);
6430 spin_unlock_irqrestore(&p->pi_lock, flags); 4712 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6431 4713
6432 rt_mutex_adjust_pi(p); 4714 rt_mutex_adjust_pi(p);
6433 4715
@@ -6527,7 +4809,7 @@ SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6527 return -EINVAL; 4809 return -EINVAL;
6528 4810
6529 retval = -ESRCH; 4811 retval = -ESRCH;
6530 read_lock(&tasklist_lock); 4812 rcu_read_lock();
6531 p = find_process_by_pid(pid); 4813 p = find_process_by_pid(pid);
6532 if (p) { 4814 if (p) {
6533 retval = security_task_getscheduler(p); 4815 retval = security_task_getscheduler(p);
@@ -6535,7 +4817,7 @@ SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6535 retval = p->policy 4817 retval = p->policy
6536 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4818 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6537 } 4819 }
6538 read_unlock(&tasklist_lock); 4820 rcu_read_unlock();
6539 return retval; 4821 return retval;
6540} 4822}
6541 4823
@@ -6553,7 +4835,7 @@ SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6553 if (!param || pid < 0) 4835 if (!param || pid < 0)
6554 return -EINVAL; 4836 return -EINVAL;
6555 4837
6556 read_lock(&tasklist_lock); 4838 rcu_read_lock();
6557 p = find_process_by_pid(pid); 4839 p = find_process_by_pid(pid);
6558 retval = -ESRCH; 4840 retval = -ESRCH;
6559 if (!p) 4841 if (!p)
@@ -6564,7 +4846,7 @@ SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6564 goto out_unlock; 4846 goto out_unlock;
6565 4847
6566 lp.sched_priority = p->rt_priority; 4848 lp.sched_priority = p->rt_priority;
6567 read_unlock(&tasklist_lock); 4849 rcu_read_unlock();
6568 4850
6569 /* 4851 /*
6570 * This one might sleep, we cannot do it with a spinlock held ... 4852 * This one might sleep, we cannot do it with a spinlock held ...
@@ -6574,7 +4856,7 @@ SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6574 return retval; 4856 return retval;
6575 4857
6576out_unlock: 4858out_unlock:
6577 read_unlock(&tasklist_lock); 4859 rcu_read_unlock();
6578 return retval; 4860 return retval;
6579} 4861}
6580 4862
@@ -6585,23 +4867,19 @@ long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6585 int retval; 4867 int retval;
6586 4868
6587 get_online_cpus(); 4869 get_online_cpus();
6588 read_lock(&tasklist_lock); 4870 rcu_read_lock();
6589 4871
6590 p = find_process_by_pid(pid); 4872 p = find_process_by_pid(pid);
6591 /* Don't set affinity if task not found and for LITMUS tasks */ 4873 /* Don't set affinity if task not found and for LITMUS tasks */
6592 if (!p || is_realtime(p)) { 4874 if (!p || is_realtime(p)) {
6593 read_unlock(&tasklist_lock); 4875 rcu_read_unlock();
6594 put_online_cpus(); 4876 put_online_cpus();
6595 return p ? -EPERM : -ESRCH; 4877 return p ? -EPERM : -ESRCH;
6596 } 4878 }
6597 4879
6598 /* 4880 /* Prevent p going away */
6599 * It is not safe to call set_cpus_allowed with the
6600 * tasklist_lock held. We will bump the task_struct's
6601 * usage count and then drop tasklist_lock.
6602 */
6603 get_task_struct(p); 4881 get_task_struct(p);
6604 read_unlock(&tasklist_lock); 4882 rcu_read_unlock();
6605 4883
6606 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4884 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6607 retval = -ENOMEM; 4885 retval = -ENOMEM;
@@ -6682,10 +4960,12 @@ SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6682long sched_getaffinity(pid_t pid, struct cpumask *mask) 4960long sched_getaffinity(pid_t pid, struct cpumask *mask)
6683{ 4961{
6684 struct task_struct *p; 4962 struct task_struct *p;
4963 unsigned long flags;
4964 struct rq *rq;
6685 int retval; 4965 int retval;
6686 4966
6687 get_online_cpus(); 4967 get_online_cpus();
6688 read_lock(&tasklist_lock); 4968 rcu_read_lock();
6689 4969
6690 retval = -ESRCH; 4970 retval = -ESRCH;
6691 p = find_process_by_pid(pid); 4971 p = find_process_by_pid(pid);
@@ -6696,10 +4976,12 @@ long sched_getaffinity(pid_t pid, struct cpumask *mask)
6696 if (retval) 4976 if (retval)
6697 goto out_unlock; 4977 goto out_unlock;
6698 4978
4979 rq = task_rq_lock(p, &flags);
6699 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4980 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4981 task_rq_unlock(rq, &flags);
6700 4982
6701out_unlock: 4983out_unlock:
6702 read_unlock(&tasklist_lock); 4984 rcu_read_unlock();
6703 put_online_cpus(); 4985 put_online_cpus();
6704 4986
6705 return retval; 4987 return retval;
@@ -6717,7 +4999,9 @@ SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6717 int ret; 4999 int ret;
6718 cpumask_var_t mask; 5000 cpumask_var_t mask;
6719 5001
6720 if (len < cpumask_size()) 5002 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5003 return -EINVAL;
5004 if (len & (sizeof(unsigned long)-1))
6721 return -EINVAL; 5005 return -EINVAL;
6722 5006
6723 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 5007 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
@@ -6725,10 +5009,12 @@ SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6725 5009
6726 ret = sched_getaffinity(pid, mask); 5010 ret = sched_getaffinity(pid, mask);
6727 if (ret == 0) { 5011 if (ret == 0) {
6728 if (copy_to_user(user_mask_ptr, mask, cpumask_size())) 5012 size_t retlen = min_t(size_t, len, cpumask_size());
5013
5014 if (copy_to_user(user_mask_ptr, mask, retlen))
6729 ret = -EFAULT; 5015 ret = -EFAULT;
6730 else 5016 else
6731 ret = cpumask_size(); 5017 ret = retlen;
6732 } 5018 }
6733 free_cpumask_var(mask); 5019 free_cpumask_var(mask);
6734 5020
@@ -6754,7 +5040,7 @@ SYSCALL_DEFINE0(sched_yield)
6754 */ 5040 */
6755 __release(rq->lock); 5041 __release(rq->lock);
6756 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 5042 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6757 _raw_spin_unlock(&rq->lock); 5043 do_raw_spin_unlock(&rq->lock);
6758 preempt_enable_no_resched(); 5044 preempt_enable_no_resched();
6759 5045
6760 schedule(); 5046 schedule();
@@ -6934,6 +5220,8 @@ SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6934{ 5220{
6935 struct task_struct *p; 5221 struct task_struct *p;
6936 unsigned int time_slice; 5222 unsigned int time_slice;
5223 unsigned long flags;
5224 struct rq *rq;
6937 int retval; 5225 int retval;
6938 struct timespec t; 5226 struct timespec t;
6939 5227
@@ -6941,7 +5229,7 @@ SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6941 return -EINVAL; 5229 return -EINVAL;
6942 5230
6943 retval = -ESRCH; 5231 retval = -ESRCH;
6944 read_lock(&tasklist_lock); 5232 rcu_read_lock();
6945 p = find_process_by_pid(pid); 5233 p = find_process_by_pid(pid);
6946 if (!p) 5234 if (!p)
6947 goto out_unlock; 5235 goto out_unlock;
@@ -6950,15 +5238,17 @@ SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6950 if (retval) 5238 if (retval)
6951 goto out_unlock; 5239 goto out_unlock;
6952 5240
6953 time_slice = p->sched_class->get_rr_interval(p); 5241 rq = task_rq_lock(p, &flags);
5242 time_slice = p->sched_class->get_rr_interval(rq, p);
5243 task_rq_unlock(rq, &flags);
6954 5244
6955 read_unlock(&tasklist_lock); 5245 rcu_read_unlock();
6956 jiffies_to_timespec(time_slice, &t); 5246 jiffies_to_timespec(time_slice, &t);
6957 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 5247 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6958 return retval; 5248 return retval;
6959 5249
6960out_unlock: 5250out_unlock:
6961 read_unlock(&tasklist_lock); 5251 rcu_read_unlock();
6962 return retval; 5252 return retval;
6963} 5253}
6964 5254
@@ -7024,7 +5314,7 @@ void show_state_filter(unsigned long state_filter)
7024 /* 5314 /*
7025 * Only show locks if all tasks are dumped: 5315 * Only show locks if all tasks are dumped:
7026 */ 5316 */
7027 if (state_filter == -1) 5317 if (!state_filter)
7028 debug_show_all_locks(); 5318 debug_show_all_locks();
7029} 5319}
7030 5320
@@ -7046,12 +5336,12 @@ void __cpuinit init_idle(struct task_struct *idle, int cpu)
7046 struct rq *rq = cpu_rq(cpu); 5336 struct rq *rq = cpu_rq(cpu);
7047 unsigned long flags; 5337 unsigned long flags;
7048 5338
7049 spin_lock_irqsave(&rq->lock, flags); 5339 raw_spin_lock_irqsave(&rq->lock, flags);
7050 5340
7051 __sched_fork(idle); 5341 __sched_fork(idle);
5342 idle->state = TASK_RUNNING;
7052 idle->se.exec_start = sched_clock(); 5343 idle->se.exec_start = sched_clock();
7053 5344
7054 idle->prio = idle->normal_prio = MAX_PRIO;
7055 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu)); 5345 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7056 __set_task_cpu(idle, cpu); 5346 __set_task_cpu(idle, cpu);
7057 5347
@@ -7059,7 +5349,7 @@ void __cpuinit init_idle(struct task_struct *idle, int cpu)
7059#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) 5349#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7060 idle->oncpu = 1; 5350 idle->oncpu = 1;
7061#endif 5351#endif
7062 spin_unlock_irqrestore(&rq->lock, flags); 5352 raw_spin_unlock_irqrestore(&rq->lock, flags);
7063 5353
7064 /* Set the preempt count _outside_ the spinlocks! */ 5354 /* Set the preempt count _outside_ the spinlocks! */
7065#if defined(CONFIG_PREEMPT) 5355#if defined(CONFIG_PREEMPT)
@@ -7092,22 +5382,43 @@ cpumask_var_t nohz_cpu_mask;
7092 * 5382 *
7093 * This idea comes from the SD scheduler of Con Kolivas: 5383 * This idea comes from the SD scheduler of Con Kolivas:
7094 */ 5384 */
7095static inline void sched_init_granularity(void) 5385static int get_update_sysctl_factor(void)
7096{ 5386{
7097 unsigned int factor = 1 + ilog2(num_online_cpus()); 5387 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7098 const unsigned long limit = 200000000; 5388 unsigned int factor;
5389
5390 switch (sysctl_sched_tunable_scaling) {
5391 case SCHED_TUNABLESCALING_NONE:
5392 factor = 1;
5393 break;
5394 case SCHED_TUNABLESCALING_LINEAR:
5395 factor = cpus;
5396 break;
5397 case SCHED_TUNABLESCALING_LOG:
5398 default:
5399 factor = 1 + ilog2(cpus);
5400 break;
5401 }
7099 5402
7100 sysctl_sched_min_granularity *= factor; 5403 return factor;
7101 if (sysctl_sched_min_granularity > limit) 5404}
7102 sysctl_sched_min_granularity = limit;
7103 5405
7104 sysctl_sched_latency *= factor; 5406static void update_sysctl(void)
7105 if (sysctl_sched_latency > limit) 5407{
7106 sysctl_sched_latency = limit; 5408 unsigned int factor = get_update_sysctl_factor();
7107 5409
7108 sysctl_sched_wakeup_granularity *= factor; 5410#define SET_SYSCTL(name) \
5411 (sysctl_##name = (factor) * normalized_sysctl_##name)
5412 SET_SYSCTL(sched_min_granularity);
5413 SET_SYSCTL(sched_latency);
5414 SET_SYSCTL(sched_wakeup_granularity);
5415 SET_SYSCTL(sched_shares_ratelimit);
5416#undef SET_SYSCTL
5417}
7109 5418
7110 sysctl_sched_shares_ratelimit *= factor; 5419static inline void sched_init_granularity(void)
5420{
5421 update_sysctl();
7111} 5422}
7112 5423
7113#ifdef CONFIG_SMP 5424#ifdef CONFIG_SMP
@@ -7144,7 +5455,8 @@ int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7144 int ret = 0; 5455 int ret = 0;
7145 5456
7146 rq = task_rq_lock(p, &flags); 5457 rq = task_rq_lock(p, &flags);
7147 if (!cpumask_intersects(new_mask, cpu_online_mask)) { 5458
5459 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7148 ret = -EINVAL; 5460 ret = -EINVAL;
7149 goto out; 5461 goto out;
7150 } 5462 }
@@ -7166,13 +5478,13 @@ int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7166 if (cpumask_test_cpu(task_cpu(p), new_mask)) 5478 if (cpumask_test_cpu(task_cpu(p), new_mask))
7167 goto out; 5479 goto out;
7168 5480
7169 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) { 5481 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7170 /* Need help from migration thread: drop lock and wait. */ 5482 /* Need help from migration thread: drop lock and wait. */
7171 struct task_struct *mt = rq->migration_thread; 5483 struct task_struct *mt = rq->migration_thread;
7172 5484
7173 get_task_struct(mt); 5485 get_task_struct(mt);
7174 task_rq_unlock(rq, &flags); 5486 task_rq_unlock(rq, &flags);
7175 wake_up_process(rq->migration_thread); 5487 wake_up_process(mt);
7176 put_task_struct(mt); 5488 put_task_struct(mt);
7177 wait_for_completion(&req.done); 5489 wait_for_completion(&req.done);
7178 tlb_migrate_finish(p->mm); 5490 tlb_migrate_finish(p->mm);
@@ -7199,7 +5511,7 @@ EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7199static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 5511static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7200{ 5512{
7201 struct rq *rq_dest, *rq_src; 5513 struct rq *rq_dest, *rq_src;
7202 int ret = 0, on_rq; 5514 int ret = 0;
7203 5515
7204 if (unlikely(!cpu_active(dest_cpu))) 5516 if (unlikely(!cpu_active(dest_cpu)))
7205 return ret; 5517 return ret;
@@ -7215,12 +5527,13 @@ static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7215 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) 5527 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7216 goto fail; 5528 goto fail;
7217 5529
7218 on_rq = p->se.on_rq; 5530 /*
7219 if (on_rq) 5531 * If we're not on a rq, the next wake-up will ensure we're
5532 * placed properly.
5533 */
5534 if (p->se.on_rq) {
7220 deactivate_task(rq_src, p, 0); 5535 deactivate_task(rq_src, p, 0);
7221 5536 set_task_cpu(p, dest_cpu);
7222 set_task_cpu(p, dest_cpu);
7223 if (on_rq) {
7224 activate_task(rq_dest, p, 0); 5537 activate_task(rq_dest, p, 0);
7225 check_preempt_curr(rq_dest, p, 0); 5538 check_preempt_curr(rq_dest, p, 0);
7226 } 5539 }
@@ -7255,10 +5568,10 @@ static int migration_thread(void *data)
7255 struct migration_req *req; 5568 struct migration_req *req;
7256 struct list_head *head; 5569 struct list_head *head;
7257 5570
7258 spin_lock_irq(&rq->lock); 5571 raw_spin_lock_irq(&rq->lock);
7259 5572
7260 if (cpu_is_offline(cpu)) { 5573 if (cpu_is_offline(cpu)) {
7261 spin_unlock_irq(&rq->lock); 5574 raw_spin_unlock_irq(&rq->lock);
7262 break; 5575 break;
7263 } 5576 }
7264 5577
@@ -7270,7 +5583,7 @@ static int migration_thread(void *data)
7270 head = &rq->migration_queue; 5583 head = &rq->migration_queue;
7271 5584
7272 if (list_empty(head)) { 5585 if (list_empty(head)) {
7273 spin_unlock_irq(&rq->lock); 5586 raw_spin_unlock_irq(&rq->lock);
7274 schedule(); 5587 schedule();
7275 set_current_state(TASK_INTERRUPTIBLE); 5588 set_current_state(TASK_INTERRUPTIBLE);
7276 continue; 5589 continue;
@@ -7279,14 +5592,14 @@ static int migration_thread(void *data)
7279 list_del_init(head->next); 5592 list_del_init(head->next);
7280 5593
7281 if (req->task != NULL) { 5594 if (req->task != NULL) {
7282 spin_unlock(&rq->lock); 5595 raw_spin_unlock(&rq->lock);
7283 __migrate_task(req->task, cpu, req->dest_cpu); 5596 __migrate_task(req->task, cpu, req->dest_cpu);
7284 } else if (likely(cpu == (badcpu = smp_processor_id()))) { 5597 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7285 req->dest_cpu = RCU_MIGRATION_GOT_QS; 5598 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7286 spin_unlock(&rq->lock); 5599 raw_spin_unlock(&rq->lock);
7287 } else { 5600 } else {
7288 req->dest_cpu = RCU_MIGRATION_MUST_SYNC; 5601 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7289 spin_unlock(&rq->lock); 5602 raw_spin_unlock(&rq->lock);
7290 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu); 5603 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7291 } 5604 }
7292 local_irq_enable(); 5605 local_irq_enable();
@@ -7316,37 +5629,10 @@ static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7316static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) 5629static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7317{ 5630{
7318 int dest_cpu; 5631 int dest_cpu;
7319 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7320 5632
7321again: 5633again:
7322 /* Look for allowed, online CPU in same node. */ 5634 dest_cpu = select_fallback_rq(dead_cpu, p);
7323 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7324 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7325 goto move;
7326
7327 /* Any allowed, online CPU? */
7328 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7329 if (dest_cpu < nr_cpu_ids)
7330 goto move;
7331
7332 /* No more Mr. Nice Guy. */
7333 if (dest_cpu >= nr_cpu_ids) {
7334 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7335 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7336
7337 /*
7338 * Don't tell them about moving exiting tasks or
7339 * kernel threads (both mm NULL), since they never
7340 * leave kernel.
7341 */
7342 if (p->mm && printk_ratelimit()) {
7343 printk(KERN_INFO "process %d (%s) no "
7344 "longer affine to cpu%d\n",
7345 task_pid_nr(p), p->comm, dead_cpu);
7346 }
7347 }
7348 5635
7349move:
7350 /* It can have affinity changed while we were choosing. */ 5636 /* It can have affinity changed while we were choosing. */
7351 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu))) 5637 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7352 goto again; 5638 goto again;
@@ -7361,7 +5647,7 @@ move:
7361 */ 5647 */
7362static void migrate_nr_uninterruptible(struct rq *rq_src) 5648static void migrate_nr_uninterruptible(struct rq *rq_src)
7363{ 5649{
7364 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask)); 5650 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7365 unsigned long flags; 5651 unsigned long flags;
7366 5652
7367 local_irq_save(flags); 5653 local_irq_save(flags);
@@ -7409,14 +5695,14 @@ void sched_idle_next(void)
7409 * Strictly not necessary since rest of the CPUs are stopped by now 5695 * Strictly not necessary since rest of the CPUs are stopped by now
7410 * and interrupts disabled on the current cpu. 5696 * and interrupts disabled on the current cpu.
7411 */ 5697 */
7412 spin_lock_irqsave(&rq->lock, flags); 5698 raw_spin_lock_irqsave(&rq->lock, flags);
7413 5699
7414 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); 5700 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7415 5701
7416 update_rq_clock(rq); 5702 update_rq_clock(rq);
7417 activate_task(rq, p, 0); 5703 activate_task(rq, p, 0);
7418 5704
7419 spin_unlock_irqrestore(&rq->lock, flags); 5705 raw_spin_unlock_irqrestore(&rq->lock, flags);
7420} 5706}
7421 5707
7422/* 5708/*
@@ -7452,9 +5738,9 @@ static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7452 * that's OK. No task can be added to this CPU, so iteration is 5738 * that's OK. No task can be added to this CPU, so iteration is
7453 * fine. 5739 * fine.
7454 */ 5740 */
7455 spin_unlock_irq(&rq->lock); 5741 raw_spin_unlock_irq(&rq->lock);
7456 move_task_off_dead_cpu(dead_cpu, p); 5742 move_task_off_dead_cpu(dead_cpu, p);
7457 spin_lock_irq(&rq->lock); 5743 raw_spin_lock_irq(&rq->lock);
7458 5744
7459 put_task_struct(p); 5745 put_task_struct(p);
7460} 5746}
@@ -7495,17 +5781,16 @@ static struct ctl_table sd_ctl_dir[] = {
7495 .procname = "sched_domain", 5781 .procname = "sched_domain",
7496 .mode = 0555, 5782 .mode = 0555,
7497 }, 5783 },
7498 {0, }, 5784 {}
7499}; 5785};
7500 5786
7501static struct ctl_table sd_ctl_root[] = { 5787static struct ctl_table sd_ctl_root[] = {
7502 { 5788 {
7503 .ctl_name = CTL_KERN,
7504 .procname = "kernel", 5789 .procname = "kernel",
7505 .mode = 0555, 5790 .mode = 0555,
7506 .child = sd_ctl_dir, 5791 .child = sd_ctl_dir,
7507 }, 5792 },
7508 {0, }, 5793 {}
7509}; 5794};
7510 5795
7511static struct ctl_table *sd_alloc_ctl_entry(int n) 5796static struct ctl_table *sd_alloc_ctl_entry(int n)
@@ -7615,7 +5900,7 @@ static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7615static struct ctl_table_header *sd_sysctl_header; 5900static struct ctl_table_header *sd_sysctl_header;
7616static void register_sched_domain_sysctl(void) 5901static void register_sched_domain_sysctl(void)
7617{ 5902{
7618 int i, cpu_num = num_online_cpus(); 5903 int i, cpu_num = num_possible_cpus();
7619 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5904 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7620 char buf[32]; 5905 char buf[32];
7621 5906
@@ -7625,7 +5910,7 @@ static void register_sched_domain_sysctl(void)
7625 if (entry == NULL) 5910 if (entry == NULL)
7626 return; 5911 return;
7627 5912
7628 for_each_online_cpu(i) { 5913 for_each_possible_cpu(i) {
7629 snprintf(buf, 32, "cpu%d", i); 5914 snprintf(buf, 32, "cpu%d", i);
7630 entry->procname = kstrdup(buf, GFP_KERNEL); 5915 entry->procname = kstrdup(buf, GFP_KERNEL);
7631 entry->mode = 0555; 5916 entry->mode = 0555;
@@ -7721,13 +6006,13 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7721 6006
7722 /* Update our root-domain */ 6007 /* Update our root-domain */
7723 rq = cpu_rq(cpu); 6008 rq = cpu_rq(cpu);
7724 spin_lock_irqsave(&rq->lock, flags); 6009 raw_spin_lock_irqsave(&rq->lock, flags);
7725 if (rq->rd) { 6010 if (rq->rd) {
7726 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 6011 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7727 6012
7728 set_rq_online(rq); 6013 set_rq_online(rq);
7729 } 6014 }
7730 spin_unlock_irqrestore(&rq->lock, flags); 6015 raw_spin_unlock_irqrestore(&rq->lock, flags);
7731 break; 6016 break;
7732 6017
7733#ifdef CONFIG_HOTPLUG_CPU 6018#ifdef CONFIG_HOTPLUG_CPU
@@ -7752,14 +6037,13 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7752 put_task_struct(rq->migration_thread); 6037 put_task_struct(rq->migration_thread);
7753 rq->migration_thread = NULL; 6038 rq->migration_thread = NULL;
7754 /* Idle task back to normal (off runqueue, low prio) */ 6039 /* Idle task back to normal (off runqueue, low prio) */
7755 spin_lock_irq(&rq->lock); 6040 raw_spin_lock_irq(&rq->lock);
7756 update_rq_clock(rq); 6041 update_rq_clock(rq);
7757 deactivate_task(rq, rq->idle, 0); 6042 deactivate_task(rq, rq->idle, 0);
7758 rq->idle->static_prio = MAX_PRIO;
7759 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0); 6043 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7760 rq->idle->sched_class = &idle_sched_class; 6044 rq->idle->sched_class = &idle_sched_class;
7761 migrate_dead_tasks(cpu); 6045 migrate_dead_tasks(cpu);
7762 spin_unlock_irq(&rq->lock); 6046 raw_spin_unlock_irq(&rq->lock);
7763 cpuset_unlock(); 6047 cpuset_unlock();
7764 migrate_nr_uninterruptible(rq); 6048 migrate_nr_uninterruptible(rq);
7765 BUG_ON(rq->nr_running != 0); 6049 BUG_ON(rq->nr_running != 0);
@@ -7769,30 +6053,30 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7769 * they didn't take sched_hotcpu_mutex. Just wake up 6053 * they didn't take sched_hotcpu_mutex. Just wake up
7770 * the requestors. 6054 * the requestors.
7771 */ 6055 */
7772 spin_lock_irq(&rq->lock); 6056 raw_spin_lock_irq(&rq->lock);
7773 while (!list_empty(&rq->migration_queue)) { 6057 while (!list_empty(&rq->migration_queue)) {
7774 struct migration_req *req; 6058 struct migration_req *req;
7775 6059
7776 req = list_entry(rq->migration_queue.next, 6060 req = list_entry(rq->migration_queue.next,
7777 struct migration_req, list); 6061 struct migration_req, list);
7778 list_del_init(&req->list); 6062 list_del_init(&req->list);
7779 spin_unlock_irq(&rq->lock); 6063 raw_spin_unlock_irq(&rq->lock);
7780 complete(&req->done); 6064 complete(&req->done);
7781 spin_lock_irq(&rq->lock); 6065 raw_spin_lock_irq(&rq->lock);
7782 } 6066 }
7783 spin_unlock_irq(&rq->lock); 6067 raw_spin_unlock_irq(&rq->lock);
7784 break; 6068 break;
7785 6069
7786 case CPU_DYING: 6070 case CPU_DYING:
7787 case CPU_DYING_FROZEN: 6071 case CPU_DYING_FROZEN:
7788 /* Update our root-domain */ 6072 /* Update our root-domain */
7789 rq = cpu_rq(cpu); 6073 rq = cpu_rq(cpu);
7790 spin_lock_irqsave(&rq->lock, flags); 6074 raw_spin_lock_irqsave(&rq->lock, flags);
7791 if (rq->rd) { 6075 if (rq->rd) {
7792 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 6076 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7793 set_rq_offline(rq); 6077 set_rq_offline(rq);
7794 } 6078 }
7795 spin_unlock_irqrestore(&rq->lock, flags); 6079 raw_spin_unlock_irqrestore(&rq->lock, flags);
7796 break; 6080 break;
7797#endif 6081#endif
7798 } 6082 }
@@ -7829,6 +6113,16 @@ early_initcall(migration_init);
7829 6113
7830#ifdef CONFIG_SCHED_DEBUG 6114#ifdef CONFIG_SCHED_DEBUG
7831 6115
6116static __read_mostly int sched_domain_debug_enabled;
6117
6118static int __init sched_domain_debug_setup(char *str)
6119{
6120 sched_domain_debug_enabled = 1;
6121
6122 return 0;
6123}
6124early_param("sched_debug", sched_domain_debug_setup);
6125
7832static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 6126static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7833 struct cpumask *groupmask) 6127 struct cpumask *groupmask)
7834{ 6128{
@@ -7915,6 +6209,9 @@ static void sched_domain_debug(struct sched_domain *sd, int cpu)
7915 cpumask_var_t groupmask; 6209 cpumask_var_t groupmask;
7916 int level = 0; 6210 int level = 0;
7917 6211
6212 if (!sched_domain_debug_enabled)
6213 return;
6214
7918 if (!sd) { 6215 if (!sd) {
7919 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 6216 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7920 return; 6217 return;
@@ -7994,6 +6291,8 @@ sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7994 6291
7995static void free_rootdomain(struct root_domain *rd) 6292static void free_rootdomain(struct root_domain *rd)
7996{ 6293{
6294 synchronize_sched();
6295
7997 cpupri_cleanup(&rd->cpupri); 6296 cpupri_cleanup(&rd->cpupri);
7998 6297
7999 free_cpumask_var(rd->rto_mask); 6298 free_cpumask_var(rd->rto_mask);
@@ -8007,7 +6306,7 @@ static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8007 struct root_domain *old_rd = NULL; 6306 struct root_domain *old_rd = NULL;
8008 unsigned long flags; 6307 unsigned long flags;
8009 6308
8010 spin_lock_irqsave(&rq->lock, flags); 6309 raw_spin_lock_irqsave(&rq->lock, flags);
8011 6310
8012 if (rq->rd) { 6311 if (rq->rd) {
8013 old_rd = rq->rd; 6312 old_rd = rq->rd;
@@ -8033,7 +6332,7 @@ static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8033 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 6332 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8034 set_rq_online(rq); 6333 set_rq_online(rq);
8035 6334
8036 spin_unlock_irqrestore(&rq->lock, flags); 6335 raw_spin_unlock_irqrestore(&rq->lock, flags);
8037 6336
8038 if (old_rd) 6337 if (old_rd)
8039 free_rootdomain(old_rd); 6338 free_rootdomain(old_rd);
@@ -8134,6 +6433,7 @@ static cpumask_var_t cpu_isolated_map;
8134/* Setup the mask of cpus configured for isolated domains */ 6433/* Setup the mask of cpus configured for isolated domains */
8135static int __init isolated_cpu_setup(char *str) 6434static int __init isolated_cpu_setup(char *str)
8136{ 6435{
6436 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8137 cpulist_parse(str, cpu_isolated_map); 6437 cpulist_parse(str, cpu_isolated_map);
8138 return 1; 6438 return 1;
8139} 6439}
@@ -8318,14 +6618,14 @@ enum s_alloc {
8318 */ 6618 */
8319#ifdef CONFIG_SCHED_SMT 6619#ifdef CONFIG_SCHED_SMT
8320static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains); 6620static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8321static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus); 6621static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
8322 6622
8323static int 6623static int
8324cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map, 6624cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8325 struct sched_group **sg, struct cpumask *unused) 6625 struct sched_group **sg, struct cpumask *unused)
8326{ 6626{
8327 if (sg) 6627 if (sg)
8328 *sg = &per_cpu(sched_group_cpus, cpu).sg; 6628 *sg = &per_cpu(sched_groups, cpu).sg;
8329 return cpu; 6629 return cpu;
8330} 6630}
8331#endif /* CONFIG_SCHED_SMT */ 6631#endif /* CONFIG_SCHED_SMT */
@@ -8970,7 +7270,7 @@ static int build_sched_domains(const struct cpumask *cpu_map)
8970 return __build_sched_domains(cpu_map, NULL); 7270 return __build_sched_domains(cpu_map, NULL);
8971} 7271}
8972 7272
8973static struct cpumask *doms_cur; /* current sched domains */ 7273static cpumask_var_t *doms_cur; /* current sched domains */
8974static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 7274static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8975static struct sched_domain_attr *dattr_cur; 7275static struct sched_domain_attr *dattr_cur;
8976 /* attribues of custom domains in 'doms_cur' */ 7276 /* attribues of custom domains in 'doms_cur' */
@@ -8992,6 +7292,31 @@ int __attribute__((weak)) arch_update_cpu_topology(void)
8992 return 0; 7292 return 0;
8993} 7293}
8994 7294
7295cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7296{
7297 int i;
7298 cpumask_var_t *doms;
7299
7300 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7301 if (!doms)
7302 return NULL;
7303 for (i = 0; i < ndoms; i++) {
7304 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7305 free_sched_domains(doms, i);
7306 return NULL;
7307 }
7308 }
7309 return doms;
7310}
7311
7312void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7313{
7314 unsigned int i;
7315 for (i = 0; i < ndoms; i++)
7316 free_cpumask_var(doms[i]);
7317 kfree(doms);
7318}
7319
8995/* 7320/*
8996 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 7321 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8997 * For now this just excludes isolated cpus, but could be used to 7322 * For now this just excludes isolated cpus, but could be used to
@@ -9003,12 +7328,12 @@ static int arch_init_sched_domains(const struct cpumask *cpu_map)
9003 7328
9004 arch_update_cpu_topology(); 7329 arch_update_cpu_topology();
9005 ndoms_cur = 1; 7330 ndoms_cur = 1;
9006 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL); 7331 doms_cur = alloc_sched_domains(ndoms_cur);
9007 if (!doms_cur) 7332 if (!doms_cur)
9008 doms_cur = fallback_doms; 7333 doms_cur = &fallback_doms;
9009 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map); 7334 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
9010 dattr_cur = NULL; 7335 dattr_cur = NULL;
9011 err = build_sched_domains(doms_cur); 7336 err = build_sched_domains(doms_cur[0]);
9012 register_sched_domain_sysctl(); 7337 register_sched_domain_sysctl();
9013 7338
9014 return err; 7339 return err;
@@ -9058,19 +7383,19 @@ static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9058 * doms_new[] to the current sched domain partitioning, doms_cur[]. 7383 * doms_new[] to the current sched domain partitioning, doms_cur[].
9059 * It destroys each deleted domain and builds each new domain. 7384 * It destroys each deleted domain and builds each new domain.
9060 * 7385 *
9061 * 'doms_new' is an array of cpumask's of length 'ndoms_new'. 7386 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9062 * The masks don't intersect (don't overlap.) We should setup one 7387 * The masks don't intersect (don't overlap.) We should setup one
9063 * sched domain for each mask. CPUs not in any of the cpumasks will 7388 * sched domain for each mask. CPUs not in any of the cpumasks will
9064 * not be load balanced. If the same cpumask appears both in the 7389 * not be load balanced. If the same cpumask appears both in the
9065 * current 'doms_cur' domains and in the new 'doms_new', we can leave 7390 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9066 * it as it is. 7391 * it as it is.
9067 * 7392 *
9068 * The passed in 'doms_new' should be kmalloc'd. This routine takes 7393 * The passed in 'doms_new' should be allocated using
9069 * ownership of it and will kfree it when done with it. If the caller 7394 * alloc_sched_domains. This routine takes ownership of it and will
9070 * failed the kmalloc call, then it can pass in doms_new == NULL && 7395 * free_sched_domains it when done with it. If the caller failed the
9071 * ndoms_new == 1, and partition_sched_domains() will fallback to 7396 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9072 * the single partition 'fallback_doms', it also forces the domains 7397 * and partition_sched_domains() will fallback to the single partition
9073 * to be rebuilt. 7398 * 'fallback_doms', it also forces the domains to be rebuilt.
9074 * 7399 *
9075 * If doms_new == NULL it will be replaced with cpu_online_mask. 7400 * If doms_new == NULL it will be replaced with cpu_online_mask.
9076 * ndoms_new == 0 is a special case for destroying existing domains, 7401 * ndoms_new == 0 is a special case for destroying existing domains,
@@ -9078,8 +7403,7 @@ static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9078 * 7403 *
9079 * Call with hotplug lock held 7404 * Call with hotplug lock held
9080 */ 7405 */
9081/* FIXME: Change to struct cpumask *doms_new[] */ 7406void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9082void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9083 struct sched_domain_attr *dattr_new) 7407 struct sched_domain_attr *dattr_new)
9084{ 7408{
9085 int i, j, n; 7409 int i, j, n;
@@ -9098,40 +7422,40 @@ void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9098 /* Destroy deleted domains */ 7422 /* Destroy deleted domains */
9099 for (i = 0; i < ndoms_cur; i++) { 7423 for (i = 0; i < ndoms_cur; i++) {
9100 for (j = 0; j < n && !new_topology; j++) { 7424 for (j = 0; j < n && !new_topology; j++) {
9101 if (cpumask_equal(&doms_cur[i], &doms_new[j]) 7425 if (cpumask_equal(doms_cur[i], doms_new[j])
9102 && dattrs_equal(dattr_cur, i, dattr_new, j)) 7426 && dattrs_equal(dattr_cur, i, dattr_new, j))
9103 goto match1; 7427 goto match1;
9104 } 7428 }
9105 /* no match - a current sched domain not in new doms_new[] */ 7429 /* no match - a current sched domain not in new doms_new[] */
9106 detach_destroy_domains(doms_cur + i); 7430 detach_destroy_domains(doms_cur[i]);
9107match1: 7431match1:
9108 ; 7432 ;
9109 } 7433 }
9110 7434
9111 if (doms_new == NULL) { 7435 if (doms_new == NULL) {
9112 ndoms_cur = 0; 7436 ndoms_cur = 0;
9113 doms_new = fallback_doms; 7437 doms_new = &fallback_doms;
9114 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map); 7438 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9115 WARN_ON_ONCE(dattr_new); 7439 WARN_ON_ONCE(dattr_new);
9116 } 7440 }
9117 7441
9118 /* Build new domains */ 7442 /* Build new domains */
9119 for (i = 0; i < ndoms_new; i++) { 7443 for (i = 0; i < ndoms_new; i++) {
9120 for (j = 0; j < ndoms_cur && !new_topology; j++) { 7444 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9121 if (cpumask_equal(&doms_new[i], &doms_cur[j]) 7445 if (cpumask_equal(doms_new[i], doms_cur[j])
9122 && dattrs_equal(dattr_new, i, dattr_cur, j)) 7446 && dattrs_equal(dattr_new, i, dattr_cur, j))
9123 goto match2; 7447 goto match2;
9124 } 7448 }
9125 /* no match - add a new doms_new */ 7449 /* no match - add a new doms_new */
9126 __build_sched_domains(doms_new + i, 7450 __build_sched_domains(doms_new[i],
9127 dattr_new ? dattr_new + i : NULL); 7451 dattr_new ? dattr_new + i : NULL);
9128match2: 7452match2:
9129 ; 7453 ;
9130 } 7454 }
9131 7455
9132 /* Remember the new sched domains */ 7456 /* Remember the new sched domains */
9133 if (doms_cur != fallback_doms) 7457 if (doms_cur != &fallback_doms)
9134 kfree(doms_cur); 7458 free_sched_domains(doms_cur, ndoms_cur);
9135 kfree(dattr_cur); /* kfree(NULL) is safe */ 7459 kfree(dattr_cur); /* kfree(NULL) is safe */
9136 doms_cur = doms_new; 7460 doms_cur = doms_new;
9137 dattr_cur = dattr_new; 7461 dattr_cur = dattr_new;
@@ -9183,11 +7507,13 @@ static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9183 7507
9184#ifdef CONFIG_SCHED_MC 7508#ifdef CONFIG_SCHED_MC
9185static ssize_t sched_mc_power_savings_show(struct sysdev_class *class, 7509static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7510 struct sysdev_class_attribute *attr,
9186 char *page) 7511 char *page)
9187{ 7512{
9188 return sprintf(page, "%u\n", sched_mc_power_savings); 7513 return sprintf(page, "%u\n", sched_mc_power_savings);
9189} 7514}
9190static ssize_t sched_mc_power_savings_store(struct sysdev_class *class, 7515static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7516 struct sysdev_class_attribute *attr,
9191 const char *buf, size_t count) 7517 const char *buf, size_t count)
9192{ 7518{
9193 return sched_power_savings_store(buf, count, 0); 7519 return sched_power_savings_store(buf, count, 0);
@@ -9199,11 +7525,13 @@ static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9199 7525
9200#ifdef CONFIG_SCHED_SMT 7526#ifdef CONFIG_SCHED_SMT
9201static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev, 7527static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7528 struct sysdev_class_attribute *attr,
9202 char *page) 7529 char *page)
9203{ 7530{
9204 return sprintf(page, "%u\n", sched_smt_power_savings); 7531 return sprintf(page, "%u\n", sched_smt_power_savings);
9205} 7532}
9206static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev, 7533static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7534 struct sysdev_class_attribute *attr,
9207 const char *buf, size_t count) 7535 const char *buf, size_t count)
9208{ 7536{
9209 return sched_power_savings_store(buf, count, 1); 7537 return sched_power_savings_store(buf, count, 1);
@@ -9242,8 +7570,10 @@ static int update_sched_domains(struct notifier_block *nfb,
9242 switch (action) { 7570 switch (action) {
9243 case CPU_ONLINE: 7571 case CPU_ONLINE:
9244 case CPU_ONLINE_FROZEN: 7572 case CPU_ONLINE_FROZEN:
9245 case CPU_DEAD: 7573 case CPU_DOWN_PREPARE:
9246 case CPU_DEAD_FROZEN: 7574 case CPU_DOWN_PREPARE_FROZEN:
7575 case CPU_DOWN_FAILED:
7576 case CPU_DOWN_FAILED_FROZEN:
9247 partition_sched_domains(1, NULL, NULL); 7577 partition_sched_domains(1, NULL, NULL);
9248 return NOTIFY_OK; 7578 return NOTIFY_OK;
9249 7579
@@ -9290,7 +7620,7 @@ void __init sched_init_smp(void)
9290#endif 7620#endif
9291 get_online_cpus(); 7621 get_online_cpus();
9292 mutex_lock(&sched_domains_mutex); 7622 mutex_lock(&sched_domains_mutex);
9293 arch_init_sched_domains(cpu_online_mask); 7623 arch_init_sched_domains(cpu_active_mask);
9294 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 7624 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9295 if (cpumask_empty(non_isolated_cpus)) 7625 if (cpumask_empty(non_isolated_cpus))
9296 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 7626 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
@@ -9363,13 +7693,13 @@ static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9363#ifdef CONFIG_SMP 7693#ifdef CONFIG_SMP
9364 rt_rq->rt_nr_migratory = 0; 7694 rt_rq->rt_nr_migratory = 0;
9365 rt_rq->overloaded = 0; 7695 rt_rq->overloaded = 0;
9366 plist_head_init(&rt_rq->pushable_tasks, &rq->lock); 7696 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
9367#endif 7697#endif
9368 7698
9369 rt_rq->rt_time = 0; 7699 rt_rq->rt_time = 0;
9370 rt_rq->rt_throttled = 0; 7700 rt_rq->rt_throttled = 0;
9371 rt_rq->rt_runtime = 0; 7701 rt_rq->rt_runtime = 0;
9372 spin_lock_init(&rt_rq->rt_runtime_lock); 7702 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
9373 7703
9374#ifdef CONFIG_RT_GROUP_SCHED 7704#ifdef CONFIG_RT_GROUP_SCHED
9375 rt_rq->rt_nr_boosted = 0; 7705 rt_rq->rt_nr_boosted = 0;
@@ -9416,7 +7746,6 @@ static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9416 tg->rt_rq[cpu] = rt_rq; 7746 tg->rt_rq[cpu] = rt_rq;
9417 init_rt_rq(rt_rq, rq); 7747 init_rt_rq(rt_rq, rq);
9418 rt_rq->tg = tg; 7748 rt_rq->tg = tg;
9419 rt_rq->rt_se = rt_se;
9420 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; 7749 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9421 if (add) 7750 if (add)
9422 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list); 7751 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
@@ -9447,16 +7776,9 @@ void __init sched_init(void)
9447#ifdef CONFIG_RT_GROUP_SCHED 7776#ifdef CONFIG_RT_GROUP_SCHED
9448 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 7777 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9449#endif 7778#endif
9450#ifdef CONFIG_USER_SCHED
9451 alloc_size *= 2;
9452#endif
9453#ifdef CONFIG_CPUMASK_OFFSTACK 7779#ifdef CONFIG_CPUMASK_OFFSTACK
9454 alloc_size += num_possible_cpus() * cpumask_size(); 7780 alloc_size += num_possible_cpus() * cpumask_size();
9455#endif 7781#endif
9456 /*
9457 * As sched_init() is called before page_alloc is setup,
9458 * we use alloc_bootmem().
9459 */
9460 if (alloc_size) { 7782 if (alloc_size) {
9461 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 7783 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9462 7784
@@ -9467,13 +7789,6 @@ void __init sched_init(void)
9467 init_task_group.cfs_rq = (struct cfs_rq **)ptr; 7789 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9468 ptr += nr_cpu_ids * sizeof(void **); 7790 ptr += nr_cpu_ids * sizeof(void **);
9469 7791
9470#ifdef CONFIG_USER_SCHED
9471 root_task_group.se = (struct sched_entity **)ptr;
9472 ptr += nr_cpu_ids * sizeof(void **);
9473
9474 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9475 ptr += nr_cpu_ids * sizeof(void **);
9476#endif /* CONFIG_USER_SCHED */
9477#endif /* CONFIG_FAIR_GROUP_SCHED */ 7792#endif /* CONFIG_FAIR_GROUP_SCHED */
9478#ifdef CONFIG_RT_GROUP_SCHED 7793#ifdef CONFIG_RT_GROUP_SCHED
9479 init_task_group.rt_se = (struct sched_rt_entity **)ptr; 7794 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
@@ -9482,13 +7797,6 @@ void __init sched_init(void)
9482 init_task_group.rt_rq = (struct rt_rq **)ptr; 7797 init_task_group.rt_rq = (struct rt_rq **)ptr;
9483 ptr += nr_cpu_ids * sizeof(void **); 7798 ptr += nr_cpu_ids * sizeof(void **);
9484 7799
9485#ifdef CONFIG_USER_SCHED
9486 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9487 ptr += nr_cpu_ids * sizeof(void **);
9488
9489 root_task_group.rt_rq = (struct rt_rq **)ptr;
9490 ptr += nr_cpu_ids * sizeof(void **);
9491#endif /* CONFIG_USER_SCHED */
9492#endif /* CONFIG_RT_GROUP_SCHED */ 7800#endif /* CONFIG_RT_GROUP_SCHED */
9493#ifdef CONFIG_CPUMASK_OFFSTACK 7801#ifdef CONFIG_CPUMASK_OFFSTACK
9494 for_each_possible_cpu(i) { 7802 for_each_possible_cpu(i) {
@@ -9508,22 +7816,13 @@ void __init sched_init(void)
9508#ifdef CONFIG_RT_GROUP_SCHED 7816#ifdef CONFIG_RT_GROUP_SCHED
9509 init_rt_bandwidth(&init_task_group.rt_bandwidth, 7817 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9510 global_rt_period(), global_rt_runtime()); 7818 global_rt_period(), global_rt_runtime());
9511#ifdef CONFIG_USER_SCHED
9512 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9513 global_rt_period(), RUNTIME_INF);
9514#endif /* CONFIG_USER_SCHED */
9515#endif /* CONFIG_RT_GROUP_SCHED */ 7819#endif /* CONFIG_RT_GROUP_SCHED */
9516 7820
9517#ifdef CONFIG_GROUP_SCHED 7821#ifdef CONFIG_CGROUP_SCHED
9518 list_add(&init_task_group.list, &task_groups); 7822 list_add(&init_task_group.list, &task_groups);
9519 INIT_LIST_HEAD(&init_task_group.children); 7823 INIT_LIST_HEAD(&init_task_group.children);
9520 7824
9521#ifdef CONFIG_USER_SCHED 7825#endif /* CONFIG_CGROUP_SCHED */
9522 INIT_LIST_HEAD(&root_task_group.children);
9523 init_task_group.parent = &root_task_group;
9524 list_add(&init_task_group.siblings, &root_task_group.children);
9525#endif /* CONFIG_USER_SCHED */
9526#endif /* CONFIG_GROUP_SCHED */
9527 7826
9528#if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP 7827#if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9529 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long), 7828 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
@@ -9533,7 +7832,7 @@ void __init sched_init(void)
9533 struct rq *rq; 7832 struct rq *rq;
9534 7833
9535 rq = cpu_rq(i); 7834 rq = cpu_rq(i);
9536 spin_lock_init(&rq->lock); 7835 raw_spin_lock_init(&rq->lock);
9537 rq->nr_running = 0; 7836 rq->nr_running = 0;
9538 rq->calc_load_active = 0; 7837 rq->calc_load_active = 0;
9539 rq->calc_load_update = jiffies + LOAD_FREQ; 7838 rq->calc_load_update = jiffies + LOAD_FREQ;
@@ -9563,25 +7862,6 @@ void __init sched_init(void)
9563 * directly in rq->cfs (i.e init_task_group->se[] = NULL). 7862 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9564 */ 7863 */
9565 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL); 7864 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9566#elif defined CONFIG_USER_SCHED
9567 root_task_group.shares = NICE_0_LOAD;
9568 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9569 /*
9570 * In case of task-groups formed thr' the user id of tasks,
9571 * init_task_group represents tasks belonging to root user.
9572 * Hence it forms a sibling of all subsequent groups formed.
9573 * In this case, init_task_group gets only a fraction of overall
9574 * system cpu resource, based on the weight assigned to root
9575 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9576 * by letting tasks of init_task_group sit in a separate cfs_rq
9577 * (init_tg_cfs_rq) and having one entity represent this group of
9578 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9579 */
9580 init_tg_cfs_entry(&init_task_group,
9581 &per_cpu(init_tg_cfs_rq, i),
9582 &per_cpu(init_sched_entity, i), i, 1,
9583 root_task_group.se[i]);
9584
9585#endif 7865#endif
9586#endif /* CONFIG_FAIR_GROUP_SCHED */ 7866#endif /* CONFIG_FAIR_GROUP_SCHED */
9587 7867
@@ -9590,12 +7870,6 @@ void __init sched_init(void)
9590 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 7870 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9591#ifdef CONFIG_CGROUP_SCHED 7871#ifdef CONFIG_CGROUP_SCHED
9592 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL); 7872 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9593#elif defined CONFIG_USER_SCHED
9594 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9595 init_tg_rt_entry(&init_task_group,
9596 &per_cpu(init_rt_rq, i),
9597 &per_cpu(init_sched_rt_entity, i), i, 1,
9598 root_task_group.rt_se[i]);
9599#endif 7873#endif
9600#endif 7874#endif
9601 7875
@@ -9611,6 +7885,8 @@ void __init sched_init(void)
9611 rq->cpu = i; 7885 rq->cpu = i;
9612 rq->online = 0; 7886 rq->online = 0;
9613 rq->migration_thread = NULL; 7887 rq->migration_thread = NULL;
7888 rq->idle_stamp = 0;
7889 rq->avg_idle = 2*sysctl_sched_migration_cost;
9614 INIT_LIST_HEAD(&rq->migration_queue); 7890 INIT_LIST_HEAD(&rq->migration_queue);
9615 rq_attach_root(rq, &def_root_domain); 7891 rq_attach_root(rq, &def_root_domain);
9616#endif 7892#endif
@@ -9629,7 +7905,7 @@ void __init sched_init(void)
9629#endif 7905#endif
9630 7906
9631#ifdef CONFIG_RT_MUTEXES 7907#ifdef CONFIG_RT_MUTEXES
9632 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); 7908 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
9633#endif 7909#endif
9634 7910
9635 /* 7911 /*
@@ -9660,7 +7936,9 @@ void __init sched_init(void)
9660 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT); 7936 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9661 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT); 7937 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9662#endif 7938#endif
9663 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7939 /* May be allocated at isolcpus cmdline parse time */
7940 if (cpu_isolated_map == NULL)
7941 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9664#endif /* SMP */ 7942#endif /* SMP */
9665 7943
9666 perf_event_init(); 7944 perf_event_init();
@@ -9671,12 +7949,12 @@ void __init sched_init(void)
9671#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP 7949#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9672static inline int preempt_count_equals(int preempt_offset) 7950static inline int preempt_count_equals(int preempt_offset)
9673{ 7951{
9674 int nested = preempt_count() & ~PREEMPT_ACTIVE; 7952 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
9675 7953
9676 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset); 7954 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9677} 7955}
9678 7956
9679void __might_sleep(char *file, int line, int preempt_offset) 7957void __might_sleep(const char *file, int line, int preempt_offset)
9680{ 7958{
9681#ifdef in_atomic 7959#ifdef in_atomic
9682 static unsigned long prev_jiffy; /* ratelimiting */ 7960 static unsigned long prev_jiffy; /* ratelimiting */
@@ -9752,13 +8030,13 @@ void normalize_rt_tasks(void)
9752 continue; 8030 continue;
9753 } 8031 }
9754 8032
9755 spin_lock(&p->pi_lock); 8033 raw_spin_lock(&p->pi_lock);
9756 rq = __task_rq_lock(p); 8034 rq = __task_rq_lock(p);
9757 8035
9758 normalize_task(rq, p); 8036 normalize_task(rq, p);
9759 8037
9760 __task_rq_unlock(rq); 8038 __task_rq_unlock(rq);
9761 spin_unlock(&p->pi_lock); 8039 raw_spin_unlock(&p->pi_lock);
9762 } while_each_thread(g, p); 8040 } while_each_thread(g, p);
9763 8041
9764 read_unlock_irqrestore(&tasklist_lock, flags); 8042 read_unlock_irqrestore(&tasklist_lock, flags);
@@ -9854,13 +8132,15 @@ int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9854 se = kzalloc_node(sizeof(struct sched_entity), 8132 se = kzalloc_node(sizeof(struct sched_entity),
9855 GFP_KERNEL, cpu_to_node(i)); 8133 GFP_KERNEL, cpu_to_node(i));
9856 if (!se) 8134 if (!se)
9857 goto err; 8135 goto err_free_rq;
9858 8136
9859 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]); 8137 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9860 } 8138 }
9861 8139
9862 return 1; 8140 return 1;
9863 8141
8142 err_free_rq:
8143 kfree(cfs_rq);
9864 err: 8144 err:
9865 return 0; 8145 return 0;
9866} 8146}
@@ -9942,13 +8222,15 @@ int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9942 rt_se = kzalloc_node(sizeof(struct sched_rt_entity), 8222 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9943 GFP_KERNEL, cpu_to_node(i)); 8223 GFP_KERNEL, cpu_to_node(i));
9944 if (!rt_se) 8224 if (!rt_se)
9945 goto err; 8225 goto err_free_rq;
9946 8226
9947 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]); 8227 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9948 } 8228 }
9949 8229
9950 return 1; 8230 return 1;
9951 8231
8232 err_free_rq:
8233 kfree(rt_rq);
9952 err: 8234 err:
9953 return 0; 8235 return 0;
9954} 8236}
@@ -9983,7 +8265,7 @@ static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9983} 8265}
9984#endif /* CONFIG_RT_GROUP_SCHED */ 8266#endif /* CONFIG_RT_GROUP_SCHED */
9985 8267
9986#ifdef CONFIG_GROUP_SCHED 8268#ifdef CONFIG_CGROUP_SCHED
9987static void free_sched_group(struct task_group *tg) 8269static void free_sched_group(struct task_group *tg)
9988{ 8270{
9989 free_fair_sched_group(tg); 8271 free_fair_sched_group(tg);
@@ -10082,17 +8364,17 @@ void sched_move_task(struct task_struct *tsk)
10082 8364
10083#ifdef CONFIG_FAIR_GROUP_SCHED 8365#ifdef CONFIG_FAIR_GROUP_SCHED
10084 if (tsk->sched_class->moved_group) 8366 if (tsk->sched_class->moved_group)
10085 tsk->sched_class->moved_group(tsk); 8367 tsk->sched_class->moved_group(tsk, on_rq);
10086#endif 8368#endif
10087 8369
10088 if (unlikely(running)) 8370 if (unlikely(running))
10089 tsk->sched_class->set_curr_task(rq); 8371 tsk->sched_class->set_curr_task(rq);
10090 if (on_rq) 8372 if (on_rq)
10091 enqueue_task(rq, tsk, 0); 8373 enqueue_task(rq, tsk, 0, false);
10092 8374
10093 task_rq_unlock(rq, &flags); 8375 task_rq_unlock(rq, &flags);
10094} 8376}
10095#endif /* CONFIG_GROUP_SCHED */ 8377#endif /* CONFIG_CGROUP_SCHED */
10096 8378
10097#ifdef CONFIG_FAIR_GROUP_SCHED 8379#ifdef CONFIG_FAIR_GROUP_SCHED
10098static void __set_se_shares(struct sched_entity *se, unsigned long shares) 8380static void __set_se_shares(struct sched_entity *se, unsigned long shares)
@@ -10117,9 +8399,9 @@ static void set_se_shares(struct sched_entity *se, unsigned long shares)
10117 struct rq *rq = cfs_rq->rq; 8399 struct rq *rq = cfs_rq->rq;
10118 unsigned long flags; 8400 unsigned long flags;
10119 8401
10120 spin_lock_irqsave(&rq->lock, flags); 8402 raw_spin_lock_irqsave(&rq->lock, flags);
10121 __set_se_shares(se, shares); 8403 __set_se_shares(se, shares);
10122 spin_unlock_irqrestore(&rq->lock, flags); 8404 raw_spin_unlock_irqrestore(&rq->lock, flags);
10123} 8405}
10124 8406
10125static DEFINE_MUTEX(shares_mutex); 8407static DEFINE_MUTEX(shares_mutex);
@@ -10234,13 +8516,6 @@ static int tg_schedulable(struct task_group *tg, void *data)
10234 runtime = d->rt_runtime; 8516 runtime = d->rt_runtime;
10235 } 8517 }
10236 8518
10237#ifdef CONFIG_USER_SCHED
10238 if (tg == &root_task_group) {
10239 period = global_rt_period();
10240 runtime = global_rt_runtime();
10241 }
10242#endif
10243
10244 /* 8519 /*
10245 * Cannot have more runtime than the period. 8520 * Cannot have more runtime than the period.
10246 */ 8521 */
@@ -10304,18 +8579,18 @@ static int tg_set_bandwidth(struct task_group *tg,
10304 if (err) 8579 if (err)
10305 goto unlock; 8580 goto unlock;
10306 8581
10307 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 8582 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10308 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 8583 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10309 tg->rt_bandwidth.rt_runtime = rt_runtime; 8584 tg->rt_bandwidth.rt_runtime = rt_runtime;
10310 8585
10311 for_each_possible_cpu(i) { 8586 for_each_possible_cpu(i) {
10312 struct rt_rq *rt_rq = tg->rt_rq[i]; 8587 struct rt_rq *rt_rq = tg->rt_rq[i];
10313 8588
10314 spin_lock(&rt_rq->rt_runtime_lock); 8589 raw_spin_lock(&rt_rq->rt_runtime_lock);
10315 rt_rq->rt_runtime = rt_runtime; 8590 rt_rq->rt_runtime = rt_runtime;
10316 spin_unlock(&rt_rq->rt_runtime_lock); 8591 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10317 } 8592 }
10318 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 8593 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10319 unlock: 8594 unlock:
10320 read_unlock(&tasklist_lock); 8595 read_unlock(&tasklist_lock);
10321 mutex_unlock(&rt_constraints_mutex); 8596 mutex_unlock(&rt_constraints_mutex);
@@ -10420,15 +8695,15 @@ static int sched_rt_global_constraints(void)
10420 if (sysctl_sched_rt_runtime == 0) 8695 if (sysctl_sched_rt_runtime == 0)
10421 return -EBUSY; 8696 return -EBUSY;
10422 8697
10423 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 8698 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10424 for_each_possible_cpu(i) { 8699 for_each_possible_cpu(i) {
10425 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 8700 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10426 8701
10427 spin_lock(&rt_rq->rt_runtime_lock); 8702 raw_spin_lock(&rt_rq->rt_runtime_lock);
10428 rt_rq->rt_runtime = global_rt_runtime(); 8703 rt_rq->rt_runtime = global_rt_runtime();
10429 spin_unlock(&rt_rq->rt_runtime_lock); 8704 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10430 } 8705 }
10431 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 8706 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10432 8707
10433 return 0; 8708 return 0;
10434} 8709}
@@ -10643,7 +8918,7 @@ struct cgroup_subsys cpu_cgroup_subsys = {
10643struct cpuacct { 8918struct cpuacct {
10644 struct cgroup_subsys_state css; 8919 struct cgroup_subsys_state css;
10645 /* cpuusage holds pointer to a u64-type object on every cpu */ 8920 /* cpuusage holds pointer to a u64-type object on every cpu */
10646 u64 *cpuusage; 8921 u64 __percpu *cpuusage;
10647 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS]; 8922 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10648 struct cpuacct *parent; 8923 struct cpuacct *parent;
10649}; 8924};
@@ -10719,9 +8994,9 @@ static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10719 /* 8994 /*
10720 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 8995 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10721 */ 8996 */
10722 spin_lock_irq(&cpu_rq(cpu)->lock); 8997 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10723 data = *cpuusage; 8998 data = *cpuusage;
10724 spin_unlock_irq(&cpu_rq(cpu)->lock); 8999 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10725#else 9000#else
10726 data = *cpuusage; 9001 data = *cpuusage;
10727#endif 9002#endif
@@ -10737,9 +9012,9 @@ static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10737 /* 9012 /*
10738 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 9013 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10739 */ 9014 */
10740 spin_lock_irq(&cpu_rq(cpu)->lock); 9015 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10741 *cpuusage = val; 9016 *cpuusage = val;
10742 spin_unlock_irq(&cpu_rq(cpu)->lock); 9017 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10743#else 9018#else
10744 *cpuusage = val; 9019 *cpuusage = val;
10745#endif 9020#endif
@@ -10860,12 +9135,30 @@ static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10860} 9135}
10861 9136
10862/* 9137/*
9138 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9139 * in cputime_t units. As a result, cpuacct_update_stats calls
9140 * percpu_counter_add with values large enough to always overflow the
9141 * per cpu batch limit causing bad SMP scalability.
9142 *
9143 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9144 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9145 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9146 */
9147#ifdef CONFIG_SMP
9148#define CPUACCT_BATCH \
9149 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9150#else
9151#define CPUACCT_BATCH 0
9152#endif
9153
9154/*
10863 * Charge the system/user time to the task's accounting group. 9155 * Charge the system/user time to the task's accounting group.
10864 */ 9156 */
10865static void cpuacct_update_stats(struct task_struct *tsk, 9157static void cpuacct_update_stats(struct task_struct *tsk,
10866 enum cpuacct_stat_index idx, cputime_t val) 9158 enum cpuacct_stat_index idx, cputime_t val)
10867{ 9159{
10868 struct cpuacct *ca; 9160 struct cpuacct *ca;
9161 int batch = CPUACCT_BATCH;
10869 9162
10870 if (unlikely(!cpuacct_subsys.active)) 9163 if (unlikely(!cpuacct_subsys.active))
10871 return; 9164 return;
@@ -10874,7 +9167,7 @@ static void cpuacct_update_stats(struct task_struct *tsk,
10874 ca = task_ca(tsk); 9167 ca = task_ca(tsk);
10875 9168
10876 do { 9169 do {
10877 percpu_counter_add(&ca->cpustat[idx], val); 9170 __percpu_counter_add(&ca->cpustat[idx], val, batch);
10878 ca = ca->parent; 9171 ca = ca->parent;
10879 } while (ca); 9172 } while (ca);
10880 rcu_read_unlock(); 9173 rcu_read_unlock();
@@ -10973,9 +9266,9 @@ void synchronize_sched_expedited(void)
10973 init_completion(&req->done); 9266 init_completion(&req->done);
10974 req->task = NULL; 9267 req->task = NULL;
10975 req->dest_cpu = RCU_MIGRATION_NEED_QS; 9268 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10976 spin_lock_irqsave(&rq->lock, flags); 9269 raw_spin_lock_irqsave(&rq->lock, flags);
10977 list_add(&req->list, &rq->migration_queue); 9270 list_add(&req->list, &rq->migration_queue);
10978 spin_unlock_irqrestore(&rq->lock, flags); 9271 raw_spin_unlock_irqrestore(&rq->lock, flags);
10979 wake_up_process(rq->migration_thread); 9272 wake_up_process(rq->migration_thread);
10980 } 9273 }
10981 for_each_online_cpu(cpu) { 9274 for_each_online_cpu(cpu) {
@@ -10983,13 +9276,14 @@ void synchronize_sched_expedited(void)
10983 req = &per_cpu(rcu_migration_req, cpu); 9276 req = &per_cpu(rcu_migration_req, cpu);
10984 rq = cpu_rq(cpu); 9277 rq = cpu_rq(cpu);
10985 wait_for_completion(&req->done); 9278 wait_for_completion(&req->done);
10986 spin_lock_irqsave(&rq->lock, flags); 9279 raw_spin_lock_irqsave(&rq->lock, flags);
10987 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC)) 9280 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10988 need_full_sync = 1; 9281 need_full_sync = 1;
10989 req->dest_cpu = RCU_MIGRATION_IDLE; 9282 req->dest_cpu = RCU_MIGRATION_IDLE;
10990 spin_unlock_irqrestore(&rq->lock, flags); 9283 raw_spin_unlock_irqrestore(&rq->lock, flags);
10991 } 9284 }
10992 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE; 9285 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9286 synchronize_sched_expedited_count++;
10993 mutex_unlock(&rcu_sched_expedited_mutex); 9287 mutex_unlock(&rcu_sched_expedited_mutex);
10994 put_online_cpus(); 9288 put_online_cpus();
10995 if (need_full_sync) 9289 if (need_full_sync)
generated by cgit v1.2.2 (git 2.25.0) at 2025-09-12 10:05:59 -0400