[PATCH] sched: less newidle locking
[pandora-kernel.git] / kernel / sched.c
1 /*
2  *  kernel/sched.c
3  *
4  *  Kernel scheduler and related syscalls
5  *
6  *  Copyright (C) 1991-2002  Linus Torvalds
7  *
8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
9  *              make semaphores SMP safe
10  *  1998-11-19  Implemented schedule_timeout() and related stuff
11  *              by Andrea Arcangeli
12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
13  *              hybrid priority-list and round-robin design with
14  *              an array-switch method of distributing timeslices
15  *              and per-CPU runqueues.  Cleanups and useful suggestions
16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
17  *  2003-09-03  Interactivity tuning by Con Kolivas.
18  *  2004-04-02  Scheduler domains code by Nick Piggin
19  */
20
21 #include <linux/mm.h>
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/completion.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/security.h>
33 #include <linux/notifier.h>
34 #include <linux/profile.h>
35 #include <linux/suspend.h>
36 #include <linux/blkdev.h>
37 #include <linux/delay.h>
38 #include <linux/smp.h>
39 #include <linux/threads.h>
40 #include <linux/timer.h>
41 #include <linux/rcupdate.h>
42 #include <linux/cpu.h>
43 #include <linux/cpuset.h>
44 #include <linux/percpu.h>
45 #include <linux/kthread.h>
46 #include <linux/seq_file.h>
47 #include <linux/syscalls.h>
48 #include <linux/times.h>
49 #include <linux/acct.h>
50 #include <asm/tlb.h>
51
52 #include <asm/unistd.h>
53
54 /*
55  * Convert user-nice values [ -20 ... 0 ... 19 ]
56  * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
57  * and back.
58  */
59 #define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
60 #define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
61 #define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
62
63 /*
64  * 'User priority' is the nice value converted to something we
65  * can work with better when scaling various scheduler parameters,
66  * it's a [ 0 ... 39 ] range.
67  */
68 #define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
69 #define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
70 #define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
71
72 /*
73  * Some helpers for converting nanosecond timing to jiffy resolution
74  */
75 #define NS_TO_JIFFIES(TIME)     ((TIME) / (1000000000 / HZ))
76 #define JIFFIES_TO_NS(TIME)     ((TIME) * (1000000000 / HZ))
77
78 /*
79  * These are the 'tuning knobs' of the scheduler:
80  *
81  * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
82  * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
83  * Timeslices get refilled after they expire.
84  */
85 #define MIN_TIMESLICE           max(5 * HZ / 1000, 1)
86 #define DEF_TIMESLICE           (100 * HZ / 1000)
87 #define ON_RUNQUEUE_WEIGHT       30
88 #define CHILD_PENALTY            95
89 #define PARENT_PENALTY          100
90 #define EXIT_WEIGHT               3
91 #define PRIO_BONUS_RATIO         25
92 #define MAX_BONUS               (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
93 #define INTERACTIVE_DELTA         2
94 #define MAX_SLEEP_AVG           (DEF_TIMESLICE * MAX_BONUS)
95 #define STARVATION_LIMIT        (MAX_SLEEP_AVG)
96 #define NS_MAX_SLEEP_AVG        (JIFFIES_TO_NS(MAX_SLEEP_AVG))
97
98 /*
99  * If a task is 'interactive' then we reinsert it in the active
100  * array after it has expired its current timeslice. (it will not
101  * continue to run immediately, it will still roundrobin with
102  * other interactive tasks.)
103  *
104  * This part scales the interactivity limit depending on niceness.
105  *
106  * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
107  * Here are a few examples of different nice levels:
108  *
109  *  TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
110  *  TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
111  *  TASK_INTERACTIVE(  0): [1,1,1,1,0,0,0,0,0,0,0]
112  *  TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
113  *  TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
114  *
115  * (the X axis represents the possible -5 ... 0 ... +5 dynamic
116  *  priority range a task can explore, a value of '1' means the
117  *  task is rated interactive.)
118  *
119  * Ie. nice +19 tasks can never get 'interactive' enough to be
120  * reinserted into the active array. And only heavily CPU-hog nice -20
121  * tasks will be expired. Default nice 0 tasks are somewhere between,
122  * it takes some effort for them to get interactive, but it's not
123  * too hard.
124  */
125
126 #define CURRENT_BONUS(p) \
127         (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
128                 MAX_SLEEP_AVG)
129
130 #define GRANULARITY     (10 * HZ / 1000 ? : 1)
131
132 #ifdef CONFIG_SMP
133 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
134                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
135                         num_online_cpus())
136 #else
137 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
138                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
139 #endif
140
141 #define SCALE(v1,v1_max,v2_max) \
142         (v1) * (v2_max) / (v1_max)
143
144 #define DELTA(p) \
145         (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
146
147 #define TASK_INTERACTIVE(p) \
148         ((p)->prio <= (p)->static_prio - DELTA(p))
149
150 #define INTERACTIVE_SLEEP(p) \
151         (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
152                 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
153
154 #define TASK_PREEMPTS_CURR(p, rq) \
155         ((p)->prio < (rq)->curr->prio)
156
157 /*
158  * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
159  * to time slice values: [800ms ... 100ms ... 5ms]
160  *
161  * The higher a thread's priority, the bigger timeslices
162  * it gets during one round of execution. But even the lowest
163  * priority thread gets MIN_TIMESLICE worth of execution time.
164  */
165
166 #define SCALE_PRIO(x, prio) \
167         max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
168
169 static unsigned int task_timeslice(task_t *p)
170 {
171         if (p->static_prio < NICE_TO_PRIO(0))
172                 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
173         else
174                 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
175 }
176 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran)       \
177                                 < (long long) (sd)->cache_hot_time)
178
179 /*
180  * These are the runqueue data structures:
181  */
182
183 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
184
185 typedef struct runqueue runqueue_t;
186
187 struct prio_array {
188         unsigned int nr_active;
189         unsigned long bitmap[BITMAP_SIZE];
190         struct list_head queue[MAX_PRIO];
191 };
192
193 /*
194  * This is the main, per-CPU runqueue data structure.
195  *
196  * Locking rule: those places that want to lock multiple runqueues
197  * (such as the load balancing or the thread migration code), lock
198  * acquire operations must be ordered by ascending &runqueue.
199  */
200 struct runqueue {
201         spinlock_t lock;
202
203         /*
204          * nr_running and cpu_load should be in the same cacheline because
205          * remote CPUs use both these fields when doing load calculation.
206          */
207         unsigned long nr_running;
208 #ifdef CONFIG_SMP
209         unsigned long cpu_load[3];
210 #endif
211         unsigned long long nr_switches;
212
213         /*
214          * This is part of a global counter where only the total sum
215          * over all CPUs matters. A task can increase this counter on
216          * one CPU and if it got migrated afterwards it may decrease
217          * it on another CPU. Always updated under the runqueue lock:
218          */
219         unsigned long nr_uninterruptible;
220
221         unsigned long expired_timestamp;
222         unsigned long long timestamp_last_tick;
223         task_t *curr, *idle;
224         struct mm_struct *prev_mm;
225         prio_array_t *active, *expired, arrays[2];
226         int best_expired_prio;
227         atomic_t nr_iowait;
228
229 #ifdef CONFIG_SMP
230         struct sched_domain *sd;
231
232         /* For active balancing */
233         int active_balance;
234         int push_cpu;
235
236         task_t *migration_thread;
237         struct list_head migration_queue;
238 #endif
239
240 #ifdef CONFIG_SCHEDSTATS
241         /* latency stats */
242         struct sched_info rq_sched_info;
243
244         /* sys_sched_yield() stats */
245         unsigned long yld_exp_empty;
246         unsigned long yld_act_empty;
247         unsigned long yld_both_empty;
248         unsigned long yld_cnt;
249
250         /* schedule() stats */
251         unsigned long sched_switch;
252         unsigned long sched_cnt;
253         unsigned long sched_goidle;
254
255         /* try_to_wake_up() stats */
256         unsigned long ttwu_cnt;
257         unsigned long ttwu_local;
258 #endif
259 };
260
261 static DEFINE_PER_CPU(struct runqueue, runqueues);
262
263 /*
264  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
265  * See detach_destroy_domains: synchronize_sched for details.
266  *
267  * The domain tree of any CPU may only be accessed from within
268  * preempt-disabled sections.
269  */
270 #define for_each_domain(cpu, domain) \
271 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
272
273 #define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
274 #define this_rq()               (&__get_cpu_var(runqueues))
275 #define task_rq(p)              cpu_rq(task_cpu(p))
276 #define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
277
278 #ifndef prepare_arch_switch
279 # define prepare_arch_switch(next)      do { } while (0)
280 #endif
281 #ifndef finish_arch_switch
282 # define finish_arch_switch(prev)       do { } while (0)
283 #endif
284
285 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
286 static inline int task_running(runqueue_t *rq, task_t *p)
287 {
288         return rq->curr == p;
289 }
290
291 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
292 {
293 }
294
295 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
296 {
297         spin_unlock_irq(&rq->lock);
298 }
299
300 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
301 static inline int task_running(runqueue_t *rq, task_t *p)
302 {
303 #ifdef CONFIG_SMP
304         return p->oncpu;
305 #else
306         return rq->curr == p;
307 #endif
308 }
309
310 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
311 {
312 #ifdef CONFIG_SMP
313         /*
314          * We can optimise this out completely for !SMP, because the
315          * SMP rebalancing from interrupt is the only thing that cares
316          * here.
317          */
318         next->oncpu = 1;
319 #endif
320 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
321         spin_unlock_irq(&rq->lock);
322 #else
323         spin_unlock(&rq->lock);
324 #endif
325 }
326
327 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
328 {
329 #ifdef CONFIG_SMP
330         /*
331          * After ->oncpu is cleared, the task can be moved to a different CPU.
332          * We must ensure this doesn't happen until the switch is completely
333          * finished.
334          */
335         smp_wmb();
336         prev->oncpu = 0;
337 #endif
338 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
339         local_irq_enable();
340 #endif
341 }
342 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
343
344 /*
345  * task_rq_lock - lock the runqueue a given task resides on and disable
346  * interrupts.  Note the ordering: we can safely lookup the task_rq without
347  * explicitly disabling preemption.
348  */
349 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
350         __acquires(rq->lock)
351 {
352         struct runqueue *rq;
353
354 repeat_lock_task:
355         local_irq_save(*flags);
356         rq = task_rq(p);
357         spin_lock(&rq->lock);
358         if (unlikely(rq != task_rq(p))) {
359                 spin_unlock_irqrestore(&rq->lock, *flags);
360                 goto repeat_lock_task;
361         }
362         return rq;
363 }
364
365 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
366         __releases(rq->lock)
367 {
368         spin_unlock_irqrestore(&rq->lock, *flags);
369 }
370
371 #ifdef CONFIG_SCHEDSTATS
372 /*
373  * bump this up when changing the output format or the meaning of an existing
374  * format, so that tools can adapt (or abort)
375  */
376 #define SCHEDSTAT_VERSION 12
377
378 static int show_schedstat(struct seq_file *seq, void *v)
379 {
380         int cpu;
381
382         seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
383         seq_printf(seq, "timestamp %lu\n", jiffies);
384         for_each_online_cpu(cpu) {
385                 runqueue_t *rq = cpu_rq(cpu);
386 #ifdef CONFIG_SMP
387                 struct sched_domain *sd;
388                 int dcnt = 0;
389 #endif
390
391                 /* runqueue-specific stats */
392                 seq_printf(seq,
393                     "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
394                     cpu, rq->yld_both_empty,
395                     rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
396                     rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
397                     rq->ttwu_cnt, rq->ttwu_local,
398                     rq->rq_sched_info.cpu_time,
399                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
400
401                 seq_printf(seq, "\n");
402
403 #ifdef CONFIG_SMP
404                 /* domain-specific stats */
405                 preempt_disable();
406                 for_each_domain(cpu, sd) {
407                         enum idle_type itype;
408                         char mask_str[NR_CPUS];
409
410                         cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
411                         seq_printf(seq, "domain%d %s", dcnt++, mask_str);
412                         for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
413                                         itype++) {
414                                 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
415                                     sd->lb_cnt[itype],
416                                     sd->lb_balanced[itype],
417                                     sd->lb_failed[itype],
418                                     sd->lb_imbalance[itype],
419                                     sd->lb_gained[itype],
420                                     sd->lb_hot_gained[itype],
421                                     sd->lb_nobusyq[itype],
422                                     sd->lb_nobusyg[itype]);
423                         }
424                         seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
425                             sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
426                             sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
427                             sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
428                             sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
429                 }
430                 preempt_enable();
431 #endif
432         }
433         return 0;
434 }
435
436 static int schedstat_open(struct inode *inode, struct file *file)
437 {
438         unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
439         char *buf = kmalloc(size, GFP_KERNEL);
440         struct seq_file *m;
441         int res;
442
443         if (!buf)
444                 return -ENOMEM;
445         res = single_open(file, show_schedstat, NULL);
446         if (!res) {
447                 m = file->private_data;
448                 m->buf = buf;
449                 m->size = size;
450         } else
451                 kfree(buf);
452         return res;
453 }
454
455 struct file_operations proc_schedstat_operations = {
456         .open    = schedstat_open,
457         .read    = seq_read,
458         .llseek  = seq_lseek,
459         .release = single_release,
460 };
461
462 # define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
463 # define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
464 #else /* !CONFIG_SCHEDSTATS */
465 # define schedstat_inc(rq, field)       do { } while (0)
466 # define schedstat_add(rq, field, amt)  do { } while (0)
467 #endif
468
469 /*
470  * rq_lock - lock a given runqueue and disable interrupts.
471  */
472 static inline runqueue_t *this_rq_lock(void)
473         __acquires(rq->lock)
474 {
475         runqueue_t *rq;
476
477         local_irq_disable();
478         rq = this_rq();
479         spin_lock(&rq->lock);
480
481         return rq;
482 }
483
484 #ifdef CONFIG_SCHEDSTATS
485 /*
486  * Called when a process is dequeued from the active array and given
487  * the cpu.  We should note that with the exception of interactive
488  * tasks, the expired queue will become the active queue after the active
489  * queue is empty, without explicitly dequeuing and requeuing tasks in the
490  * expired queue.  (Interactive tasks may be requeued directly to the
491  * active queue, thus delaying tasks in the expired queue from running;
492  * see scheduler_tick()).
493  *
494  * This function is only called from sched_info_arrive(), rather than
495  * dequeue_task(). Even though a task may be queued and dequeued multiple
496  * times as it is shuffled about, we're really interested in knowing how
497  * long it was from the *first* time it was queued to the time that it
498  * finally hit a cpu.
499  */
500 static inline void sched_info_dequeued(task_t *t)
501 {
502         t->sched_info.last_queued = 0;
503 }
504
505 /*
506  * Called when a task finally hits the cpu.  We can now calculate how
507  * long it was waiting to run.  We also note when it began so that we
508  * can keep stats on how long its timeslice is.
509  */
510 static inline void sched_info_arrive(task_t *t)
511 {
512         unsigned long now = jiffies, diff = 0;
513         struct runqueue *rq = task_rq(t);
514
515         if (t->sched_info.last_queued)
516                 diff = now - t->sched_info.last_queued;
517         sched_info_dequeued(t);
518         t->sched_info.run_delay += diff;
519         t->sched_info.last_arrival = now;
520         t->sched_info.pcnt++;
521
522         if (!rq)
523                 return;
524
525         rq->rq_sched_info.run_delay += diff;
526         rq->rq_sched_info.pcnt++;
527 }
528
529 /*
530  * Called when a process is queued into either the active or expired
531  * array.  The time is noted and later used to determine how long we
532  * had to wait for us to reach the cpu.  Since the expired queue will
533  * become the active queue after active queue is empty, without dequeuing
534  * and requeuing any tasks, we are interested in queuing to either. It
535  * is unusual but not impossible for tasks to be dequeued and immediately
536  * requeued in the same or another array: this can happen in sched_yield(),
537  * set_user_nice(), and even load_balance() as it moves tasks from runqueue
538  * to runqueue.
539  *
540  * This function is only called from enqueue_task(), but also only updates
541  * the timestamp if it is already not set.  It's assumed that
542  * sched_info_dequeued() will clear that stamp when appropriate.
543  */
544 static inline void sched_info_queued(task_t *t)
545 {
546         if (!t->sched_info.last_queued)
547                 t->sched_info.last_queued = jiffies;
548 }
549
550 /*
551  * Called when a process ceases being the active-running process, either
552  * voluntarily or involuntarily.  Now we can calculate how long we ran.
553  */
554 static inline void sched_info_depart(task_t *t)
555 {
556         struct runqueue *rq = task_rq(t);
557         unsigned long diff = jiffies - t->sched_info.last_arrival;
558
559         t->sched_info.cpu_time += diff;
560
561         if (rq)
562                 rq->rq_sched_info.cpu_time += diff;
563 }
564
565 /*
566  * Called when tasks are switched involuntarily due, typically, to expiring
567  * their time slice.  (This may also be called when switching to or from
568  * the idle task.)  We are only called when prev != next.
569  */
570 static inline void sched_info_switch(task_t *prev, task_t *next)
571 {
572         struct runqueue *rq = task_rq(prev);
573
574         /*
575          * prev now departs the cpu.  It's not interesting to record
576          * stats about how efficient we were at scheduling the idle
577          * process, however.
578          */
579         if (prev != rq->idle)
580                 sched_info_depart(prev);
581
582         if (next != rq->idle)
583                 sched_info_arrive(next);
584 }
585 #else
586 #define sched_info_queued(t)            do { } while (0)
587 #define sched_info_switch(t, next)      do { } while (0)
588 #endif /* CONFIG_SCHEDSTATS */
589
590 /*
591  * Adding/removing a task to/from a priority array:
592  */
593 static void dequeue_task(struct task_struct *p, prio_array_t *array)
594 {
595         array->nr_active--;
596         list_del(&p->run_list);
597         if (list_empty(array->queue + p->prio))
598                 __clear_bit(p->prio, array->bitmap);
599 }
600
601 static void enqueue_task(struct task_struct *p, prio_array_t *array)
602 {
603         sched_info_queued(p);
604         list_add_tail(&p->run_list, array->queue + p->prio);
605         __set_bit(p->prio, array->bitmap);
606         array->nr_active++;
607         p->array = array;
608 }
609
610 /*
611  * Put task to the end of the run list without the overhead of dequeue
612  * followed by enqueue.
613  */
614 static void requeue_task(struct task_struct *p, prio_array_t *array)
615 {
616         list_move_tail(&p->run_list, array->queue + p->prio);
617 }
618
619 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
620 {
621         list_add(&p->run_list, array->queue + p->prio);
622         __set_bit(p->prio, array->bitmap);
623         array->nr_active++;
624         p->array = array;
625 }
626
627 /*
628  * effective_prio - return the priority that is based on the static
629  * priority but is modified by bonuses/penalties.
630  *
631  * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
632  * into the -5 ... 0 ... +5 bonus/penalty range.
633  *
634  * We use 25% of the full 0...39 priority range so that:
635  *
636  * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
637  * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
638  *
639  * Both properties are important to certain workloads.
640  */
641 static int effective_prio(task_t *p)
642 {
643         int bonus, prio;
644
645         if (rt_task(p))
646                 return p->prio;
647
648         bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
649
650         prio = p->static_prio - bonus;
651         if (prio < MAX_RT_PRIO)
652                 prio = MAX_RT_PRIO;
653         if (prio > MAX_PRIO-1)
654                 prio = MAX_PRIO-1;
655         return prio;
656 }
657
658 /*
659  * __activate_task - move a task to the runqueue.
660  */
661 static inline void __activate_task(task_t *p, runqueue_t *rq)
662 {
663         enqueue_task(p, rq->active);
664         rq->nr_running++;
665 }
666
667 /*
668  * __activate_idle_task - move idle task to the _front_ of runqueue.
669  */
670 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
671 {
672         enqueue_task_head(p, rq->active);
673         rq->nr_running++;
674 }
675
676 static int recalc_task_prio(task_t *p, unsigned long long now)
677 {
678         /* Caller must always ensure 'now >= p->timestamp' */
679         unsigned long long __sleep_time = now - p->timestamp;
680         unsigned long sleep_time;
681
682         if (__sleep_time > NS_MAX_SLEEP_AVG)
683                 sleep_time = NS_MAX_SLEEP_AVG;
684         else
685                 sleep_time = (unsigned long)__sleep_time;
686
687         if (likely(sleep_time > 0)) {
688                 /*
689                  * User tasks that sleep a long time are categorised as
690                  * idle and will get just interactive status to stay active &
691                  * prevent them suddenly becoming cpu hogs and starving
692                  * other processes.
693                  */
694                 if (p->mm && p->activated != -1 &&
695                         sleep_time > INTERACTIVE_SLEEP(p)) {
696                                 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
697                                                 DEF_TIMESLICE);
698                 } else {
699                         /*
700                          * The lower the sleep avg a task has the more
701                          * rapidly it will rise with sleep time.
702                          */
703                         sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
704
705                         /*
706                          * Tasks waking from uninterruptible sleep are
707                          * limited in their sleep_avg rise as they
708                          * are likely to be waiting on I/O
709                          */
710                         if (p->activated == -1 && p->mm) {
711                                 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
712                                         sleep_time = 0;
713                                 else if (p->sleep_avg + sleep_time >=
714                                                 INTERACTIVE_SLEEP(p)) {
715                                         p->sleep_avg = INTERACTIVE_SLEEP(p);
716                                         sleep_time = 0;
717                                 }
718                         }
719
720                         /*
721                          * This code gives a bonus to interactive tasks.
722                          *
723                          * The boost works by updating the 'average sleep time'
724                          * value here, based on ->timestamp. The more time a
725                          * task spends sleeping, the higher the average gets -
726                          * and the higher the priority boost gets as well.
727                          */
728                         p->sleep_avg += sleep_time;
729
730                         if (p->sleep_avg > NS_MAX_SLEEP_AVG)
731                                 p->sleep_avg = NS_MAX_SLEEP_AVG;
732                 }
733         }
734
735         return effective_prio(p);
736 }
737
738 /*
739  * activate_task - move a task to the runqueue and do priority recalculation
740  *
741  * Update all the scheduling statistics stuff. (sleep average
742  * calculation, priority modifiers, etc.)
743  */
744 static void activate_task(task_t *p, runqueue_t *rq, int local)
745 {
746         unsigned long long now;
747
748         now = sched_clock();
749 #ifdef CONFIG_SMP
750         if (!local) {
751                 /* Compensate for drifting sched_clock */
752                 runqueue_t *this_rq = this_rq();
753                 now = (now - this_rq->timestamp_last_tick)
754                         + rq->timestamp_last_tick;
755         }
756 #endif
757
758         p->prio = recalc_task_prio(p, now);
759
760         /*
761          * This checks to make sure it's not an uninterruptible task
762          * that is now waking up.
763          */
764         if (!p->activated) {
765                 /*
766                  * Tasks which were woken up by interrupts (ie. hw events)
767                  * are most likely of interactive nature. So we give them
768                  * the credit of extending their sleep time to the period
769                  * of time they spend on the runqueue, waiting for execution
770                  * on a CPU, first time around:
771                  */
772                 if (in_interrupt())
773                         p->activated = 2;
774                 else {
775                         /*
776                          * Normal first-time wakeups get a credit too for
777                          * on-runqueue time, but it will be weighted down:
778                          */
779                         p->activated = 1;
780                 }
781         }
782         p->timestamp = now;
783
784         __activate_task(p, rq);
785 }
786
787 /*
788  * deactivate_task - remove a task from the runqueue.
789  */
790 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
791 {
792         rq->nr_running--;
793         dequeue_task(p, p->array);
794         p->array = NULL;
795 }
796
797 /*
798  * resched_task - mark a task 'to be rescheduled now'.
799  *
800  * On UP this means the setting of the need_resched flag, on SMP it
801  * might also involve a cross-CPU call to trigger the scheduler on
802  * the target CPU.
803  */
804 #ifdef CONFIG_SMP
805 static void resched_task(task_t *p)
806 {
807         int need_resched, nrpolling;
808
809         assert_spin_locked(&task_rq(p)->lock);
810
811         /* minimise the chance of sending an interrupt to poll_idle() */
812         nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
813         need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
814         nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
815
816         if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
817                 smp_send_reschedule(task_cpu(p));
818 }
819 #else
820 static inline void resched_task(task_t *p)
821 {
822         set_tsk_need_resched(p);
823 }
824 #endif
825
826 /**
827  * task_curr - is this task currently executing on a CPU?
828  * @p: the task in question.
829  */
830 inline int task_curr(const task_t *p)
831 {
832         return cpu_curr(task_cpu(p)) == p;
833 }
834
835 #ifdef CONFIG_SMP
836 typedef struct {
837         struct list_head list;
838
839         task_t *task;
840         int dest_cpu;
841
842         struct completion done;
843 } migration_req_t;
844
845 /*
846  * The task's runqueue lock must be held.
847  * Returns true if you have to wait for migration thread.
848  */
849 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
850 {
851         runqueue_t *rq = task_rq(p);
852
853         /*
854          * If the task is not on a runqueue (and not running), then
855          * it is sufficient to simply update the task's cpu field.
856          */
857         if (!p->array && !task_running(rq, p)) {
858                 set_task_cpu(p, dest_cpu);
859                 return 0;
860         }
861
862         init_completion(&req->done);
863         req->task = p;
864         req->dest_cpu = dest_cpu;
865         list_add(&req->list, &rq->migration_queue);
866         return 1;
867 }
868
869 /*
870  * wait_task_inactive - wait for a thread to unschedule.
871  *
872  * The caller must ensure that the task *will* unschedule sometime soon,
873  * else this function might spin for a *long* time. This function can't
874  * be called with interrupts off, or it may introduce deadlock with
875  * smp_call_function() if an IPI is sent by the same process we are
876  * waiting to become inactive.
877  */
878 void wait_task_inactive(task_t *p)
879 {
880         unsigned long flags;
881         runqueue_t *rq;
882         int preempted;
883
884 repeat:
885         rq = task_rq_lock(p, &flags);
886         /* Must be off runqueue entirely, not preempted. */
887         if (unlikely(p->array || task_running(rq, p))) {
888                 /* If it's preempted, we yield.  It could be a while. */
889                 preempted = !task_running(rq, p);
890                 task_rq_unlock(rq, &flags);
891                 cpu_relax();
892                 if (preempted)
893                         yield();
894                 goto repeat;
895         }
896         task_rq_unlock(rq, &flags);
897 }
898
899 /***
900  * kick_process - kick a running thread to enter/exit the kernel
901  * @p: the to-be-kicked thread
902  *
903  * Cause a process which is running on another CPU to enter
904  * kernel-mode, without any delay. (to get signals handled.)
905  *
906  * NOTE: this function doesnt have to take the runqueue lock,
907  * because all it wants to ensure is that the remote task enters
908  * the kernel. If the IPI races and the task has been migrated
909  * to another CPU then no harm is done and the purpose has been
910  * achieved as well.
911  */
912 void kick_process(task_t *p)
913 {
914         int cpu;
915
916         preempt_disable();
917         cpu = task_cpu(p);
918         if ((cpu != smp_processor_id()) && task_curr(p))
919                 smp_send_reschedule(cpu);
920         preempt_enable();
921 }
922
923 /*
924  * Return a low guess at the load of a migration-source cpu.
925  *
926  * We want to under-estimate the load of migration sources, to
927  * balance conservatively.
928  */
929 static inline unsigned long source_load(int cpu, int type)
930 {
931         runqueue_t *rq = cpu_rq(cpu);
932         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
933         if (type == 0)
934                 return load_now;
935
936         return min(rq->cpu_load[type-1], load_now);
937 }
938
939 /*
940  * Return a high guess at the load of a migration-target cpu
941  */
942 static inline unsigned long target_load(int cpu, int type)
943 {
944         runqueue_t *rq = cpu_rq(cpu);
945         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
946         if (type == 0)
947                 return load_now;
948
949         return max(rq->cpu_load[type-1], load_now);
950 }
951
952 /*
953  * find_idlest_group finds and returns the least busy CPU group within the
954  * domain.
955  */
956 static struct sched_group *
957 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
958 {
959         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
960         unsigned long min_load = ULONG_MAX, this_load = 0;
961         int load_idx = sd->forkexec_idx;
962         int imbalance = 100 + (sd->imbalance_pct-100)/2;
963
964         do {
965                 unsigned long load, avg_load;
966                 int local_group;
967                 int i;
968
969                 /* Skip over this group if it has no CPUs allowed */
970                 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
971                         goto nextgroup;
972
973                 local_group = cpu_isset(this_cpu, group->cpumask);
974
975                 /* Tally up the load of all CPUs in the group */
976                 avg_load = 0;
977
978                 for_each_cpu_mask(i, group->cpumask) {
979                         /* Bias balancing toward cpus of our domain */
980                         if (local_group)
981                                 load = source_load(i, load_idx);
982                         else
983                                 load = target_load(i, load_idx);
984
985                         avg_load += load;
986                 }
987
988                 /* Adjust by relative CPU power of the group */
989                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
990
991                 if (local_group) {
992                         this_load = avg_load;
993                         this = group;
994                 } else if (avg_load < min_load) {
995                         min_load = avg_load;
996                         idlest = group;
997                 }
998 nextgroup:
999                 group = group->next;
1000         } while (group != sd->groups);
1001
1002         if (!idlest || 100*this_load < imbalance*min_load)
1003                 return NULL;
1004         return idlest;
1005 }
1006
1007 /*
1008  * find_idlest_queue - find the idlest runqueue among the cpus in group.
1009  */
1010 static int
1011 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1012 {
1013         cpumask_t tmp;
1014         unsigned long load, min_load = ULONG_MAX;
1015         int idlest = -1;
1016         int i;
1017
1018         /* Traverse only the allowed CPUs */
1019         cpus_and(tmp, group->cpumask, p->cpus_allowed);
1020
1021         for_each_cpu_mask(i, tmp) {
1022                 load = source_load(i, 0);
1023
1024                 if (load < min_load || (load == min_load && i == this_cpu)) {
1025                         min_load = load;
1026                         idlest = i;
1027                 }
1028         }
1029
1030         return idlest;
1031 }
1032
1033 /*
1034  * sched_balance_self: balance the current task (running on cpu) in domains
1035  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1036  * SD_BALANCE_EXEC.
1037  *
1038  * Balance, ie. select the least loaded group.
1039  *
1040  * Returns the target CPU number, or the same CPU if no balancing is needed.
1041  *
1042  * preempt must be disabled.
1043  */
1044 static int sched_balance_self(int cpu, int flag)
1045 {
1046         struct task_struct *t = current;
1047         struct sched_domain *tmp, *sd = NULL;
1048
1049         for_each_domain(cpu, tmp)
1050                 if (tmp->flags & flag)
1051                         sd = tmp;
1052
1053         while (sd) {
1054                 cpumask_t span;
1055                 struct sched_group *group;
1056                 int new_cpu;
1057                 int weight;
1058
1059                 span = sd->span;
1060                 group = find_idlest_group(sd, t, cpu);
1061                 if (!group)
1062                         goto nextlevel;
1063
1064                 new_cpu = find_idlest_cpu(group, t, cpu);
1065                 if (new_cpu == -1 || new_cpu == cpu)
1066                         goto nextlevel;
1067
1068                 /* Now try balancing at a lower domain level */
1069                 cpu = new_cpu;
1070 nextlevel:
1071                 sd = NULL;
1072                 weight = cpus_weight(span);
1073                 for_each_domain(cpu, tmp) {
1074                         if (weight <= cpus_weight(tmp->span))
1075                                 break;
1076                         if (tmp->flags & flag)
1077                                 sd = tmp;
1078                 }
1079                 /* while loop will break here if sd == NULL */
1080         }
1081
1082         return cpu;
1083 }
1084
1085 #endif /* CONFIG_SMP */
1086
1087 /*
1088  * wake_idle() will wake a task on an idle cpu if task->cpu is
1089  * not idle and an idle cpu is available.  The span of cpus to
1090  * search starts with cpus closest then further out as needed,
1091  * so we always favor a closer, idle cpu.
1092  *
1093  * Returns the CPU we should wake onto.
1094  */
1095 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1096 static int wake_idle(int cpu, task_t *p)
1097 {
1098         cpumask_t tmp;
1099         struct sched_domain *sd;
1100         int i;
1101
1102         if (idle_cpu(cpu))
1103                 return cpu;
1104
1105         for_each_domain(cpu, sd) {
1106                 if (sd->flags & SD_WAKE_IDLE) {
1107                         cpus_and(tmp, sd->span, p->cpus_allowed);
1108                         for_each_cpu_mask(i, tmp) {
1109                                 if (idle_cpu(i))
1110                                         return i;
1111                         }
1112                 }
1113                 else
1114                         break;
1115         }
1116         return cpu;
1117 }
1118 #else
1119 static inline int wake_idle(int cpu, task_t *p)
1120 {
1121         return cpu;
1122 }
1123 #endif
1124
1125 /***
1126  * try_to_wake_up - wake up a thread
1127  * @p: the to-be-woken-up thread
1128  * @state: the mask of task states that can be woken
1129  * @sync: do a synchronous wakeup?
1130  *
1131  * Put it on the run-queue if it's not already there. The "current"
1132  * thread is always on the run-queue (except when the actual
1133  * re-schedule is in progress), and as such you're allowed to do
1134  * the simpler "current->state = TASK_RUNNING" to mark yourself
1135  * runnable without the overhead of this.
1136  *
1137  * returns failure only if the task is already active.
1138  */
1139 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1140 {
1141         int cpu, this_cpu, success = 0;
1142         unsigned long flags;
1143         long old_state;
1144         runqueue_t *rq;
1145 #ifdef CONFIG_SMP
1146         unsigned long load, this_load;
1147         struct sched_domain *sd, *this_sd = NULL;
1148         int new_cpu;
1149 #endif
1150
1151         rq = task_rq_lock(p, &flags);
1152         old_state = p->state;
1153         if (!(old_state & state))
1154                 goto out;
1155
1156         if (p->array)
1157                 goto out_running;
1158
1159         cpu = task_cpu(p);
1160         this_cpu = smp_processor_id();
1161
1162 #ifdef CONFIG_SMP
1163         if (unlikely(task_running(rq, p)))
1164                 goto out_activate;
1165
1166         new_cpu = cpu;
1167
1168         schedstat_inc(rq, ttwu_cnt);
1169         if (cpu == this_cpu) {
1170                 schedstat_inc(rq, ttwu_local);
1171                 goto out_set_cpu;
1172         }
1173
1174         for_each_domain(this_cpu, sd) {
1175                 if (cpu_isset(cpu, sd->span)) {
1176                         schedstat_inc(sd, ttwu_wake_remote);
1177                         this_sd = sd;
1178                         break;
1179                 }
1180         }
1181
1182         if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1183                 goto out_set_cpu;
1184
1185         /*
1186          * Check for affine wakeup and passive balancing possibilities.
1187          */
1188         if (this_sd) {
1189                 int idx = this_sd->wake_idx;
1190                 unsigned int imbalance;
1191
1192                 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1193
1194                 load = source_load(cpu, idx);
1195                 this_load = target_load(this_cpu, idx);
1196
1197                 new_cpu = this_cpu; /* Wake to this CPU if we can */
1198
1199                 if (this_sd->flags & SD_WAKE_AFFINE) {
1200                         unsigned long tl = this_load;
1201                         /*
1202                          * If sync wakeup then subtract the (maximum possible)
1203                          * effect of the currently running task from the load
1204                          * of the current CPU:
1205                          */
1206                         if (sync)
1207                                 tl -= SCHED_LOAD_SCALE;
1208
1209                         if ((tl <= load &&
1210                                 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1211                                 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1212                                 /*
1213                                  * This domain has SD_WAKE_AFFINE and
1214                                  * p is cache cold in this domain, and
1215                                  * there is no bad imbalance.
1216                                  */
1217                                 schedstat_inc(this_sd, ttwu_move_affine);
1218                                 goto out_set_cpu;
1219                         }
1220                 }
1221
1222                 /*
1223                  * Start passive balancing when half the imbalance_pct
1224                  * limit is reached.
1225                  */
1226                 if (this_sd->flags & SD_WAKE_BALANCE) {
1227                         if (imbalance*this_load <= 100*load) {
1228                                 schedstat_inc(this_sd, ttwu_move_balance);
1229                                 goto out_set_cpu;
1230                         }
1231                 }
1232         }
1233
1234         new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1235 out_set_cpu:
1236         new_cpu = wake_idle(new_cpu, p);
1237         if (new_cpu != cpu) {
1238                 set_task_cpu(p, new_cpu);
1239                 task_rq_unlock(rq, &flags);
1240                 /* might preempt at this point */
1241                 rq = task_rq_lock(p, &flags);
1242                 old_state = p->state;
1243                 if (!(old_state & state))
1244                         goto out;
1245                 if (p->array)
1246                         goto out_running;
1247
1248                 this_cpu = smp_processor_id();
1249                 cpu = task_cpu(p);
1250         }
1251
1252 out_activate:
1253 #endif /* CONFIG_SMP */
1254         if (old_state == TASK_UNINTERRUPTIBLE) {
1255                 rq->nr_uninterruptible--;
1256                 /*
1257                  * Tasks on involuntary sleep don't earn
1258                  * sleep_avg beyond just interactive state.
1259                  */
1260                 p->activated = -1;
1261         }
1262
1263         /*
1264          * Tasks that have marked their sleep as noninteractive get
1265          * woken up without updating their sleep average. (i.e. their
1266          * sleep is handled in a priority-neutral manner, no priority
1267          * boost and no penalty.)
1268          */
1269         if (old_state & TASK_NONINTERACTIVE)
1270                 __activate_task(p, rq);
1271         else
1272                 activate_task(p, rq, cpu == this_cpu);
1273         /*
1274          * Sync wakeups (i.e. those types of wakeups where the waker
1275          * has indicated that it will leave the CPU in short order)
1276          * don't trigger a preemption, if the woken up task will run on
1277          * this cpu. (in this case the 'I will reschedule' promise of
1278          * the waker guarantees that the freshly woken up task is going
1279          * to be considered on this CPU.)
1280          */
1281         if (!sync || cpu != this_cpu) {
1282                 if (TASK_PREEMPTS_CURR(p, rq))
1283                         resched_task(rq->curr);
1284         }
1285         success = 1;
1286
1287 out_running:
1288         p->state = TASK_RUNNING;
1289 out:
1290         task_rq_unlock(rq, &flags);
1291
1292         return success;
1293 }
1294
1295 int fastcall wake_up_process(task_t *p)
1296 {
1297         return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1298                                  TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1299 }
1300
1301 EXPORT_SYMBOL(wake_up_process);
1302
1303 int fastcall wake_up_state(task_t *p, unsigned int state)
1304 {
1305         return try_to_wake_up(p, state, 0);
1306 }
1307
1308 /*
1309  * Perform scheduler related setup for a newly forked process p.
1310  * p is forked by current.
1311  */
1312 void fastcall sched_fork(task_t *p, int clone_flags)
1313 {
1314         int cpu = get_cpu();
1315
1316 #ifdef CONFIG_SMP
1317         cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1318 #endif
1319         set_task_cpu(p, cpu);
1320
1321         /*
1322          * We mark the process as running here, but have not actually
1323          * inserted it onto the runqueue yet. This guarantees that
1324          * nobody will actually run it, and a signal or other external
1325          * event cannot wake it up and insert it on the runqueue either.
1326          */
1327         p->state = TASK_RUNNING;
1328         INIT_LIST_HEAD(&p->run_list);
1329         p->array = NULL;
1330 #ifdef CONFIG_SCHEDSTATS
1331         memset(&p->sched_info, 0, sizeof(p->sched_info));
1332 #endif
1333 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1334         p->oncpu = 0;
1335 #endif
1336 #ifdef CONFIG_PREEMPT
1337         /* Want to start with kernel preemption disabled. */
1338         p->thread_info->preempt_count = 1;
1339 #endif
1340         /*
1341          * Share the timeslice between parent and child, thus the
1342          * total amount of pending timeslices in the system doesn't change,
1343          * resulting in more scheduling fairness.
1344          */
1345         local_irq_disable();
1346         p->time_slice = (current->time_slice + 1) >> 1;
1347         /*
1348          * The remainder of the first timeslice might be recovered by
1349          * the parent if the child exits early enough.
1350          */
1351         p->first_time_slice = 1;
1352         current->time_slice >>= 1;
1353         p->timestamp = sched_clock();
1354         if (unlikely(!current->time_slice)) {
1355                 /*
1356                  * This case is rare, it happens when the parent has only
1357                  * a single jiffy left from its timeslice. Taking the
1358                  * runqueue lock is not a problem.
1359                  */
1360                 current->time_slice = 1;
1361                 scheduler_tick();
1362         }
1363         local_irq_enable();
1364         put_cpu();
1365 }
1366
1367 /*
1368  * wake_up_new_task - wake up a newly created task for the first time.
1369  *
1370  * This function will do some initial scheduler statistics housekeeping
1371  * that must be done for every newly created context, then puts the task
1372  * on the runqueue and wakes it.
1373  */
1374 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1375 {
1376         unsigned long flags;
1377         int this_cpu, cpu;
1378         runqueue_t *rq, *this_rq;
1379
1380         rq = task_rq_lock(p, &flags);
1381         BUG_ON(p->state != TASK_RUNNING);
1382         this_cpu = smp_processor_id();
1383         cpu = task_cpu(p);
1384
1385         /*
1386          * We decrease the sleep average of forking parents
1387          * and children as well, to keep max-interactive tasks
1388          * from forking tasks that are max-interactive. The parent
1389          * (current) is done further down, under its lock.
1390          */
1391         p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1392                 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1393
1394         p->prio = effective_prio(p);
1395
1396         if (likely(cpu == this_cpu)) {
1397                 if (!(clone_flags & CLONE_VM)) {
1398                         /*
1399                          * The VM isn't cloned, so we're in a good position to
1400                          * do child-runs-first in anticipation of an exec. This
1401                          * usually avoids a lot of COW overhead.
1402                          */
1403                         if (unlikely(!current->array))
1404                                 __activate_task(p, rq);
1405                         else {
1406                                 p->prio = current->prio;
1407                                 list_add_tail(&p->run_list, &current->run_list);
1408                                 p->array = current->array;
1409                                 p->array->nr_active++;
1410                                 rq->nr_running++;
1411                         }
1412                         set_need_resched();
1413                 } else
1414                         /* Run child last */
1415                         __activate_task(p, rq);
1416                 /*
1417                  * We skip the following code due to cpu == this_cpu
1418                  *
1419                  *   task_rq_unlock(rq, &flags);
1420                  *   this_rq = task_rq_lock(current, &flags);
1421                  */
1422                 this_rq = rq;
1423         } else {
1424                 this_rq = cpu_rq(this_cpu);
1425
1426                 /*
1427                  * Not the local CPU - must adjust timestamp. This should
1428                  * get optimised away in the !CONFIG_SMP case.
1429                  */
1430                 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1431                                         + rq->timestamp_last_tick;
1432                 __activate_task(p, rq);
1433                 if (TASK_PREEMPTS_CURR(p, rq))
1434                         resched_task(rq->curr);
1435
1436                 /*
1437                  * Parent and child are on different CPUs, now get the
1438                  * parent runqueue to update the parent's ->sleep_avg:
1439                  */
1440                 task_rq_unlock(rq, &flags);
1441                 this_rq = task_rq_lock(current, &flags);
1442         }
1443         current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1444                 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1445         task_rq_unlock(this_rq, &flags);
1446 }
1447
1448 /*
1449  * Potentially available exiting-child timeslices are
1450  * retrieved here - this way the parent does not get
1451  * penalized for creating too many threads.
1452  *
1453  * (this cannot be used to 'generate' timeslices
1454  * artificially, because any timeslice recovered here
1455  * was given away by the parent in the first place.)
1456  */
1457 void fastcall sched_exit(task_t *p)
1458 {
1459         unsigned long flags;
1460         runqueue_t *rq;
1461
1462         /*
1463          * If the child was a (relative-) CPU hog then decrease
1464          * the sleep_avg of the parent as well.
1465          */
1466         rq = task_rq_lock(p->parent, &flags);
1467         if (p->first_time_slice) {
1468                 p->parent->time_slice += p->time_slice;
1469                 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1470                         p->parent->time_slice = task_timeslice(p);
1471         }
1472         if (p->sleep_avg < p->parent->sleep_avg)
1473                 p->parent->sleep_avg = p->parent->sleep_avg /
1474                 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1475                 (EXIT_WEIGHT + 1);
1476         task_rq_unlock(rq, &flags);
1477 }
1478
1479 /**
1480  * prepare_task_switch - prepare to switch tasks
1481  * @rq: the runqueue preparing to switch
1482  * @next: the task we are going to switch to.
1483  *
1484  * This is called with the rq lock held and interrupts off. It must
1485  * be paired with a subsequent finish_task_switch after the context
1486  * switch.
1487  *
1488  * prepare_task_switch sets up locking and calls architecture specific
1489  * hooks.
1490  */
1491 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1492 {
1493         prepare_lock_switch(rq, next);
1494         prepare_arch_switch(next);
1495 }
1496
1497 /**
1498  * finish_task_switch - clean up after a task-switch
1499  * @rq: runqueue associated with task-switch
1500  * @prev: the thread we just switched away from.
1501  *
1502  * finish_task_switch must be called after the context switch, paired
1503  * with a prepare_task_switch call before the context switch.
1504  * finish_task_switch will reconcile locking set up by prepare_task_switch,
1505  * and do any other architecture-specific cleanup actions.
1506  *
1507  * Note that we may have delayed dropping an mm in context_switch(). If
1508  * so, we finish that here outside of the runqueue lock.  (Doing it
1509  * with the lock held can cause deadlocks; see schedule() for
1510  * details.)
1511  */
1512 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1513         __releases(rq->lock)
1514 {
1515         struct mm_struct *mm = rq->prev_mm;
1516         unsigned long prev_task_flags;
1517
1518         rq->prev_mm = NULL;
1519
1520         /*
1521          * A task struct has one reference for the use as "current".
1522          * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1523          * calls schedule one last time. The schedule call will never return,
1524          * and the scheduled task must drop that reference.
1525          * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1526          * still held, otherwise prev could be scheduled on another cpu, die
1527          * there before we look at prev->state, and then the reference would
1528          * be dropped twice.
1529          *              Manfred Spraul <manfred@colorfullife.com>
1530          */
1531         prev_task_flags = prev->flags;
1532 #ifdef CONFIG_DEBUG_SPINLOCK
1533         /* this is a valid case when another task releases the spinlock */
1534         rq->lock.owner = current;
1535 #endif
1536         finish_arch_switch(prev);
1537         finish_lock_switch(rq, prev);
1538         if (mm)
1539                 mmdrop(mm);
1540         if (unlikely(prev_task_flags & PF_DEAD))
1541                 put_task_struct(prev);
1542 }
1543
1544 /**
1545  * schedule_tail - first thing a freshly forked thread must call.
1546  * @prev: the thread we just switched away from.
1547  */
1548 asmlinkage void schedule_tail(task_t *prev)
1549         __releases(rq->lock)
1550 {
1551         runqueue_t *rq = this_rq();
1552         finish_task_switch(rq, prev);
1553 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1554         /* In this case, finish_task_switch does not reenable preemption */
1555         preempt_enable();
1556 #endif
1557         if (current->set_child_tid)
1558                 put_user(current->pid, current->set_child_tid);
1559 }
1560
1561 /*
1562  * context_switch - switch to the new MM and the new
1563  * thread's register state.
1564  */
1565 static inline
1566 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1567 {
1568         struct mm_struct *mm = next->mm;
1569         struct mm_struct *oldmm = prev->active_mm;
1570
1571         if (unlikely(!mm)) {
1572                 next->active_mm = oldmm;
1573                 atomic_inc(&oldmm->mm_count);
1574                 enter_lazy_tlb(oldmm, next);
1575         } else
1576                 switch_mm(oldmm, mm, next);
1577
1578         if (unlikely(!prev->mm)) {
1579                 prev->active_mm = NULL;
1580                 WARN_ON(rq->prev_mm);
1581                 rq->prev_mm = oldmm;
1582         }
1583
1584         /* Here we just switch the register state and the stack. */
1585         switch_to(prev, next, prev);
1586
1587         return prev;
1588 }
1589
1590 /*
1591  * nr_running, nr_uninterruptible and nr_context_switches:
1592  *
1593  * externally visible scheduler statistics: current number of runnable
1594  * threads, current number of uninterruptible-sleeping threads, total
1595  * number of context switches performed since bootup.
1596  */
1597 unsigned long nr_running(void)
1598 {
1599         unsigned long i, sum = 0;
1600
1601         for_each_online_cpu(i)
1602                 sum += cpu_rq(i)->nr_running;
1603
1604         return sum;
1605 }
1606
1607 unsigned long nr_uninterruptible(void)
1608 {
1609         unsigned long i, sum = 0;
1610
1611         for_each_cpu(i)
1612                 sum += cpu_rq(i)->nr_uninterruptible;
1613
1614         /*
1615          * Since we read the counters lockless, it might be slightly
1616          * inaccurate. Do not allow it to go below zero though:
1617          */
1618         if (unlikely((long)sum < 0))
1619                 sum = 0;
1620
1621         return sum;
1622 }
1623
1624 unsigned long long nr_context_switches(void)
1625 {
1626         unsigned long long i, sum = 0;
1627
1628         for_each_cpu(i)
1629                 sum += cpu_rq(i)->nr_switches;
1630
1631         return sum;
1632 }
1633
1634 unsigned long nr_iowait(void)
1635 {
1636         unsigned long i, sum = 0;
1637
1638         for_each_cpu(i)
1639                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1640
1641         return sum;
1642 }
1643
1644 #ifdef CONFIG_SMP
1645
1646 /*
1647  * double_rq_lock - safely lock two runqueues
1648  *
1649  * Note this does not disable interrupts like task_rq_lock,
1650  * you need to do so manually before calling.
1651  */
1652 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1653         __acquires(rq1->lock)
1654         __acquires(rq2->lock)
1655 {
1656         if (rq1 == rq2) {
1657                 spin_lock(&rq1->lock);
1658                 __acquire(rq2->lock);   /* Fake it out ;) */
1659         } else {
1660                 if (rq1 < rq2) {
1661                         spin_lock(&rq1->lock);
1662                         spin_lock(&rq2->lock);
1663                 } else {
1664                         spin_lock(&rq2->lock);
1665                         spin_lock(&rq1->lock);
1666                 }
1667         }
1668 }
1669
1670 /*
1671  * double_rq_unlock - safely unlock two runqueues
1672  *
1673  * Note this does not restore interrupts like task_rq_unlock,
1674  * you need to do so manually after calling.
1675  */
1676 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1677         __releases(rq1->lock)
1678         __releases(rq2->lock)
1679 {
1680         spin_unlock(&rq1->lock);
1681         if (rq1 != rq2)
1682                 spin_unlock(&rq2->lock);
1683         else
1684                 __release(rq2->lock);
1685 }
1686
1687 /*
1688  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1689  */
1690 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1691         __releases(this_rq->lock)
1692         __acquires(busiest->lock)
1693         __acquires(this_rq->lock)
1694 {
1695         if (unlikely(!spin_trylock(&busiest->lock))) {
1696                 if (busiest < this_rq) {
1697                         spin_unlock(&this_rq->lock);
1698                         spin_lock(&busiest->lock);
1699                         spin_lock(&this_rq->lock);
1700                 } else
1701                         spin_lock(&busiest->lock);
1702         }
1703 }
1704
1705 /*
1706  * If dest_cpu is allowed for this process, migrate the task to it.
1707  * This is accomplished by forcing the cpu_allowed mask to only
1708  * allow dest_cpu, which will force the cpu onto dest_cpu.  Then
1709  * the cpu_allowed mask is restored.
1710  */
1711 static void sched_migrate_task(task_t *p, int dest_cpu)
1712 {
1713         migration_req_t req;
1714         runqueue_t *rq;
1715         unsigned long flags;
1716
1717         rq = task_rq_lock(p, &flags);
1718         if (!cpu_isset(dest_cpu, p->cpus_allowed)
1719             || unlikely(cpu_is_offline(dest_cpu)))
1720                 goto out;
1721
1722         /* force the process onto the specified CPU */
1723         if (migrate_task(p, dest_cpu, &req)) {
1724                 /* Need to wait for migration thread (might exit: take ref). */
1725                 struct task_struct *mt = rq->migration_thread;
1726                 get_task_struct(mt);
1727                 task_rq_unlock(rq, &flags);
1728                 wake_up_process(mt);
1729                 put_task_struct(mt);
1730                 wait_for_completion(&req.done);
1731                 return;
1732         }
1733 out:
1734         task_rq_unlock(rq, &flags);
1735 }
1736
1737 /*
1738  * sched_exec - execve() is a valuable balancing opportunity, because at
1739  * this point the task has the smallest effective memory and cache footprint.
1740  */
1741 void sched_exec(void)
1742 {
1743         int new_cpu, this_cpu = get_cpu();
1744         new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1745         put_cpu();
1746         if (new_cpu != this_cpu)
1747                 sched_migrate_task(current, new_cpu);
1748 }
1749
1750 /*
1751  * pull_task - move a task from a remote runqueue to the local runqueue.
1752  * Both runqueues must be locked.
1753  */
1754 static inline
1755 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1756                runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1757 {
1758         dequeue_task(p, src_array);
1759         src_rq->nr_running--;
1760         set_task_cpu(p, this_cpu);
1761         this_rq->nr_running++;
1762         enqueue_task(p, this_array);
1763         p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1764                                 + this_rq->timestamp_last_tick;
1765         /*
1766          * Note that idle threads have a prio of MAX_PRIO, for this test
1767          * to be always true for them.
1768          */
1769         if (TASK_PREEMPTS_CURR(p, this_rq))
1770                 resched_task(this_rq->curr);
1771 }
1772
1773 /*
1774  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1775  */
1776 static inline
1777 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1778                      struct sched_domain *sd, enum idle_type idle,
1779                      int *all_pinned)
1780 {
1781         /*
1782          * We do not migrate tasks that are:
1783          * 1) running (obviously), or
1784          * 2) cannot be migrated to this CPU due to cpus_allowed, or
1785          * 3) are cache-hot on their current CPU.
1786          */
1787         if (!cpu_isset(this_cpu, p->cpus_allowed))
1788                 return 0;
1789         *all_pinned = 0;
1790
1791         if (task_running(rq, p))
1792                 return 0;
1793
1794         /*
1795          * Aggressive migration if:
1796          * 1) task is cache cold, or
1797          * 2) too many balance attempts have failed.
1798          */
1799
1800         if (sd->nr_balance_failed > sd->cache_nice_tries)
1801                 return 1;
1802
1803         if (task_hot(p, rq->timestamp_last_tick, sd))
1804                 return 0;
1805         return 1;
1806 }
1807
1808 /*
1809  * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1810  * as part of a balancing operation within "domain". Returns the number of
1811  * tasks moved.
1812  *
1813  * Called with both runqueues locked.
1814  */
1815 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1816                       unsigned long max_nr_move, struct sched_domain *sd,
1817                       enum idle_type idle, int *all_pinned)
1818 {
1819         prio_array_t *array, *dst_array;
1820         struct list_head *head, *curr;
1821         int idx, pulled = 0, pinned = 0;
1822         task_t *tmp;
1823
1824         if (max_nr_move == 0)
1825                 goto out;
1826
1827         pinned = 1;
1828
1829         /*
1830          * We first consider expired tasks. Those will likely not be
1831          * executed in the near future, and they are most likely to
1832          * be cache-cold, thus switching CPUs has the least effect
1833          * on them.
1834          */
1835         if (busiest->expired->nr_active) {
1836                 array = busiest->expired;
1837                 dst_array = this_rq->expired;
1838         } else {
1839                 array = busiest->active;
1840                 dst_array = this_rq->active;
1841         }
1842
1843 new_array:
1844         /* Start searching at priority 0: */
1845         idx = 0;
1846 skip_bitmap:
1847         if (!idx)
1848                 idx = sched_find_first_bit(array->bitmap);
1849         else
1850                 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1851         if (idx >= MAX_PRIO) {
1852                 if (array == busiest->expired && busiest->active->nr_active) {
1853                         array = busiest->active;
1854                         dst_array = this_rq->active;
1855                         goto new_array;
1856                 }
1857                 goto out;
1858         }
1859
1860         head = array->queue + idx;
1861         curr = head->prev;
1862 skip_queue:
1863         tmp = list_entry(curr, task_t, run_list);
1864
1865         curr = curr->prev;
1866
1867         if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1868                 if (curr != head)
1869                         goto skip_queue;
1870                 idx++;
1871                 goto skip_bitmap;
1872         }
1873
1874 #ifdef CONFIG_SCHEDSTATS
1875         if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1876                 schedstat_inc(sd, lb_hot_gained[idle]);
1877 #endif
1878
1879         pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1880         pulled++;
1881
1882         /* We only want to steal up to the prescribed number of tasks. */
1883         if (pulled < max_nr_move) {
1884                 if (curr != head)
1885                         goto skip_queue;
1886                 idx++;
1887                 goto skip_bitmap;
1888         }
1889 out:
1890         /*
1891          * Right now, this is the only place pull_task() is called,
1892          * so we can safely collect pull_task() stats here rather than
1893          * inside pull_task().
1894          */
1895         schedstat_add(sd, lb_gained[idle], pulled);
1896
1897         if (all_pinned)
1898                 *all_pinned = pinned;
1899         return pulled;
1900 }
1901
1902 /*
1903  * find_busiest_group finds and returns the busiest CPU group within the
1904  * domain. It calculates and returns the number of tasks which should be
1905  * moved to restore balance via the imbalance parameter.
1906  */
1907 static struct sched_group *
1908 find_busiest_group(struct sched_domain *sd, int this_cpu,
1909                    unsigned long *imbalance, enum idle_type idle)
1910 {
1911         struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1912         unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1913         int load_idx;
1914
1915         max_load = this_load = total_load = total_pwr = 0;
1916         if (idle == NOT_IDLE)
1917                 load_idx = sd->busy_idx;
1918         else if (idle == NEWLY_IDLE)
1919                 load_idx = sd->newidle_idx;
1920         else
1921                 load_idx = sd->idle_idx;
1922
1923         do {
1924                 unsigned long load;
1925                 int local_group;
1926                 int i;
1927
1928                 local_group = cpu_isset(this_cpu, group->cpumask);
1929
1930                 /* Tally up the load of all CPUs in the group */
1931                 avg_load = 0;
1932
1933                 for_each_cpu_mask(i, group->cpumask) {
1934                         /* Bias balancing toward cpus of our domain */
1935                         if (local_group)
1936                                 load = target_load(i, load_idx);
1937                         else
1938                                 load = source_load(i, load_idx);
1939
1940                         avg_load += load;
1941                 }
1942
1943                 total_load += avg_load;
1944                 total_pwr += group->cpu_power;
1945
1946                 /* Adjust by relative CPU power of the group */
1947                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1948
1949                 if (local_group) {
1950                         this_load = avg_load;
1951                         this = group;
1952                 } else if (avg_load > max_load) {
1953                         max_load = avg_load;
1954                         busiest = group;
1955                 }
1956                 group = group->next;
1957         } while (group != sd->groups);
1958
1959         if (!busiest || this_load >= max_load)
1960                 goto out_balanced;
1961
1962         avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1963
1964         if (this_load >= avg_load ||
1965                         100*max_load <= sd->imbalance_pct*this_load)
1966                 goto out_balanced;
1967
1968         /*
1969          * We're trying to get all the cpus to the average_load, so we don't
1970          * want to push ourselves above the average load, nor do we wish to
1971          * reduce the max loaded cpu below the average load, as either of these
1972          * actions would just result in more rebalancing later, and ping-pong
1973          * tasks around. Thus we look for the minimum possible imbalance.
1974          * Negative imbalances (*we* are more loaded than anyone else) will
1975          * be counted as no imbalance for these purposes -- we can't fix that
1976          * by pulling tasks to us.  Be careful of negative numbers as they'll
1977          * appear as very large values with unsigned longs.
1978          */
1979         /* How much load to actually move to equalise the imbalance */
1980         *imbalance = min((max_load - avg_load) * busiest->cpu_power,
1981                                 (avg_load - this_load) * this->cpu_power)
1982                         / SCHED_LOAD_SCALE;
1983
1984         if (*imbalance < SCHED_LOAD_SCALE) {
1985                 unsigned long pwr_now = 0, pwr_move = 0;
1986                 unsigned long tmp;
1987
1988                 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1989                         *imbalance = 1;
1990                         return busiest;
1991                 }
1992
1993                 /*
1994                  * OK, we don't have enough imbalance to justify moving tasks,
1995                  * however we may be able to increase total CPU power used by
1996                  * moving them.
1997                  */
1998
1999                 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2000                 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2001                 pwr_now /= SCHED_LOAD_SCALE;
2002
2003                 /* Amount of load we'd subtract */
2004                 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2005                 if (max_load > tmp)
2006                         pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2007                                                         max_load - tmp);
2008
2009                 /* Amount of load we'd add */
2010                 if (max_load*busiest->cpu_power <
2011                                 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2012                         tmp = max_load*busiest->cpu_power/this->cpu_power;
2013                 else
2014                         tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2015                 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2016                 pwr_move /= SCHED_LOAD_SCALE;
2017
2018                 /* Move if we gain throughput */
2019                 if (pwr_move <= pwr_now)
2020                         goto out_balanced;
2021
2022                 *imbalance = 1;
2023                 return busiest;
2024         }
2025
2026         /* Get rid of the scaling factor, rounding down as we divide */
2027         *imbalance = *imbalance / SCHED_LOAD_SCALE;
2028         return busiest;
2029
2030 out_balanced:
2031
2032         *imbalance = 0;
2033         return NULL;
2034 }
2035
2036 /*
2037  * find_busiest_queue - find the busiest runqueue among the cpus in group.
2038  */
2039 static runqueue_t *find_busiest_queue(struct sched_group *group)
2040 {
2041         unsigned long load, max_load = 0;
2042         runqueue_t *busiest = NULL;
2043         int i;
2044
2045         for_each_cpu_mask(i, group->cpumask) {
2046                 load = source_load(i, 0);
2047
2048                 if (load > max_load) {
2049                         max_load = load;
2050                         busiest = cpu_rq(i);
2051                 }
2052         }
2053
2054         return busiest;
2055 }
2056
2057 /*
2058  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2059  * so long as it is large enough.
2060  */
2061 #define MAX_PINNED_INTERVAL     512
2062
2063 /*
2064  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2065  * tasks if there is an imbalance.
2066  *
2067  * Called with this_rq unlocked.
2068  */
2069 static int load_balance(int this_cpu, runqueue_t *this_rq,
2070                         struct sched_domain *sd, enum idle_type idle)
2071 {
2072         struct sched_group *group;
2073         runqueue_t *busiest;
2074         unsigned long imbalance;
2075         int nr_moved, all_pinned = 0;
2076         int active_balance = 0;
2077
2078         spin_lock(&this_rq->lock);
2079         schedstat_inc(sd, lb_cnt[idle]);
2080
2081         group = find_busiest_group(sd, this_cpu, &imbalance, idle);
2082         if (!group) {
2083                 schedstat_inc(sd, lb_nobusyg[idle]);
2084                 goto out_balanced;
2085         }
2086
2087         busiest = find_busiest_queue(group);
2088         if (!busiest) {
2089                 schedstat_inc(sd, lb_nobusyq[idle]);
2090                 goto out_balanced;
2091         }
2092
2093         BUG_ON(busiest == this_rq);
2094
2095         schedstat_add(sd, lb_imbalance[idle], imbalance);
2096
2097         nr_moved = 0;
2098         if (busiest->nr_running > 1) {
2099                 /*
2100                  * Attempt to move tasks. If find_busiest_group has found
2101                  * an imbalance but busiest->nr_running <= 1, the group is
2102                  * still unbalanced. nr_moved simply stays zero, so it is
2103                  * correctly treated as an imbalance.
2104                  */
2105                 double_lock_balance(this_rq, busiest);
2106                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2107                                         imbalance, sd, idle, &all_pinned);
2108                 spin_unlock(&busiest->lock);
2109
2110                 /* All tasks on this runqueue were pinned by CPU affinity */
2111                 if (unlikely(all_pinned))
2112                         goto out_balanced;
2113         }
2114
2115         spin_unlock(&this_rq->lock);
2116
2117         if (!nr_moved) {
2118                 schedstat_inc(sd, lb_failed[idle]);
2119                 sd->nr_balance_failed++;
2120
2121                 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2122
2123                         spin_lock(&busiest->lock);
2124                         if (!busiest->active_balance) {
2125                                 busiest->active_balance = 1;
2126                                 busiest->push_cpu = this_cpu;
2127                                 active_balance = 1;
2128                         }
2129                         spin_unlock(&busiest->lock);
2130                         if (active_balance)
2131                                 wake_up_process(busiest->migration_thread);
2132
2133                         /*
2134                          * We've kicked active balancing, reset the failure
2135                          * counter.
2136                          */
2137                         sd->nr_balance_failed = sd->cache_nice_tries+1;
2138                 }
2139         } else
2140                 sd->nr_balance_failed = 0;
2141
2142         if (likely(!active_balance)) {
2143                 /* We were unbalanced, so reset the balancing interval */
2144                 sd->balance_interval = sd->min_interval;
2145         } else {
2146                 /*
2147                  * If we've begun active balancing, start to back off. This
2148                  * case may not be covered by the all_pinned logic if there
2149                  * is only 1 task on the busy runqueue (because we don't call
2150                  * move_tasks).
2151                  */
2152                 if (sd->balance_interval < sd->max_interval)
2153                         sd->balance_interval *= 2;
2154         }
2155
2156         return nr_moved;
2157
2158 out_balanced:
2159         spin_unlock(&this_rq->lock);
2160
2161         schedstat_inc(sd, lb_balanced[idle]);
2162
2163         sd->nr_balance_failed = 0;
2164         /* tune up the balancing interval */
2165         if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2166                         (sd->balance_interval < sd->max_interval))
2167                 sd->balance_interval *= 2;
2168
2169         return 0;
2170 }
2171
2172 /*
2173  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2174  * tasks if there is an imbalance.
2175  *
2176  * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2177  * this_rq is locked.
2178  */
2179 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2180                                 struct sched_domain *sd)
2181 {
2182         struct sched_group *group;
2183         runqueue_t *busiest = NULL;
2184         unsigned long imbalance;
2185         int nr_moved = 0;
2186
2187         schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2188         group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);
2189         if (!group) {
2190                 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2191                 goto out_balanced;
2192         }
2193
2194         busiest = find_busiest_queue(group);
2195         if (!busiest) {
2196                 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2197                 goto out_balanced;
2198         }
2199
2200         BUG_ON(busiest == this_rq);
2201
2202         schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2203
2204         nr_moved = 0;
2205         if (busiest->nr_running > 1) {
2206                 /* Attempt to move tasks */
2207                 double_lock_balance(this_rq, busiest);
2208                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2209                                         imbalance, sd, NEWLY_IDLE, NULL);
2210                 spin_unlock(&busiest->lock);
2211         }
2212
2213         if (!nr_moved)
2214                 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2215         else
2216                 sd->nr_balance_failed = 0;
2217
2218         return nr_moved;
2219
2220 out_balanced:
2221         schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2222         sd->nr_balance_failed = 0;
2223         return 0;
2224 }
2225
2226 /*
2227  * idle_balance is called by schedule() if this_cpu is about to become
2228  * idle. Attempts to pull tasks from other CPUs.
2229  */
2230 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2231 {
2232         struct sched_domain *sd;
2233
2234         for_each_domain(this_cpu, sd) {
2235                 if (sd->flags & SD_BALANCE_NEWIDLE) {
2236                         if (load_balance_newidle(this_cpu, this_rq, sd)) {
2237                                 /* We've pulled tasks over so stop searching */
2238                                 break;
2239                         }
2240                 }
2241         }
2242 }
2243
2244 /*
2245  * active_load_balance is run by migration threads. It pushes running tasks
2246  * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2247  * running on each physical CPU where possible, and avoids physical /
2248  * logical imbalances.
2249  *
2250  * Called with busiest_rq locked.
2251  */
2252 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2253 {
2254         struct sched_domain *sd;
2255         runqueue_t *target_rq;
2256         int target_cpu = busiest_rq->push_cpu;
2257
2258         if (busiest_rq->nr_running <= 1)
2259                 /* no task to move */
2260                 return;
2261
2262         target_rq = cpu_rq(target_cpu);
2263
2264         /*
2265          * This condition is "impossible", if it occurs
2266          * we need to fix it.  Originally reported by
2267          * Bjorn Helgaas on a 128-cpu setup.
2268          */
2269         BUG_ON(busiest_rq == target_rq);
2270
2271         /* move a task from busiest_rq to target_rq */
2272         double_lock_balance(busiest_rq, target_rq);
2273
2274         /* Search for an sd spanning us and the target CPU. */
2275         for_each_domain(target_cpu, sd)
2276                 if ((sd->flags & SD_LOAD_BALANCE) &&
2277                         cpu_isset(busiest_cpu, sd->span))
2278                                 break;
2279
2280         if (unlikely(sd == NULL))
2281                 goto out;
2282
2283         schedstat_inc(sd, alb_cnt);
2284
2285         if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2286                 schedstat_inc(sd, alb_pushed);
2287         else
2288                 schedstat_inc(sd, alb_failed);
2289 out:
2290         spin_unlock(&target_rq->lock);
2291 }
2292
2293 /*
2294  * rebalance_tick will get called every timer tick, on every CPU.
2295  *
2296  * It checks each scheduling domain to see if it is due to be balanced,
2297  * and initiates a balancing operation if so.
2298  *
2299  * Balancing parameters are set up in arch_init_sched_domains.
2300  */
2301
2302 /* Don't have all balancing operations going off at once */
2303 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2304
2305 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2306                            enum idle_type idle)
2307 {
2308         unsigned long old_load, this_load;
2309         unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2310         struct sched_domain *sd;
2311         int i;
2312
2313         this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2314         /* Update our load */
2315         for (i = 0; i < 3; i++) {
2316                 unsigned long new_load = this_load;
2317                 int scale = 1 << i;
2318                 old_load = this_rq->cpu_load[i];
2319                 /*
2320                  * Round up the averaging division if load is increasing. This
2321                  * prevents us from getting stuck on 9 if the load is 10, for
2322                  * example.
2323                  */
2324                 if (new_load > old_load)
2325                         new_load += scale-1;
2326                 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2327         }
2328
2329         for_each_domain(this_cpu, sd) {
2330                 unsigned long interval;
2331
2332                 if (!(sd->flags & SD_LOAD_BALANCE))
2333                         continue;
2334
2335                 interval = sd->balance_interval;
2336                 if (idle != SCHED_IDLE)
2337                         interval *= sd->busy_factor;
2338
2339                 /* scale ms to jiffies */
2340                 interval = msecs_to_jiffies(interval);
2341                 if (unlikely(!interval))
2342                         interval = 1;
2343
2344                 if (j - sd->last_balance >= interval) {
2345                         if (load_balance(this_cpu, this_rq, sd, idle)) {
2346                                 /* We've pulled tasks over so no longer idle */
2347                                 idle = NOT_IDLE;
2348                         }
2349                         sd->last_balance += interval;
2350                 }
2351         }
2352 }
2353 #else
2354 /*
2355  * on UP we do not need to balance between CPUs:
2356  */
2357 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2358 {
2359 }
2360 static inline void idle_balance(int cpu, runqueue_t *rq)
2361 {
2362 }
2363 #endif
2364
2365 static inline int wake_priority_sleeper(runqueue_t *rq)
2366 {
2367         int ret = 0;
2368 #ifdef CONFIG_SCHED_SMT
2369         spin_lock(&rq->lock);
2370         /*
2371          * If an SMT sibling task has been put to sleep for priority
2372          * reasons reschedule the idle task to see if it can now run.
2373          */
2374         if (rq->nr_running) {
2375                 resched_task(rq->idle);
2376                 ret = 1;
2377         }
2378         spin_unlock(&rq->lock);
2379 #endif
2380         return ret;
2381 }
2382
2383 DEFINE_PER_CPU(struct kernel_stat, kstat);
2384
2385 EXPORT_PER_CPU_SYMBOL(kstat);
2386
2387 /*
2388  * This is called on clock ticks and on context switches.
2389  * Bank in p->sched_time the ns elapsed since the last tick or switch.
2390  */
2391 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2392                                     unsigned long long now)
2393 {
2394         unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2395         p->sched_time += now - last;
2396 }
2397
2398 /*
2399  * Return current->sched_time plus any more ns on the sched_clock
2400  * that have not yet been banked.
2401  */
2402 unsigned long long current_sched_time(const task_t *tsk)
2403 {
2404         unsigned long long ns;
2405         unsigned long flags;
2406         local_irq_save(flags);
2407         ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2408         ns = tsk->sched_time + (sched_clock() - ns);
2409         local_irq_restore(flags);
2410         return ns;
2411 }
2412
2413 /*
2414  * We place interactive tasks back into the active array, if possible.
2415  *
2416  * To guarantee that this does not starve expired tasks we ignore the
2417  * interactivity of a task if the first expired task had to wait more
2418  * than a 'reasonable' amount of time. This deadline timeout is
2419  * load-dependent, as the frequency of array switched decreases with
2420  * increasing number of running tasks. We also ignore the interactivity
2421  * if a better static_prio task has expired:
2422  */
2423 #define EXPIRED_STARVING(rq) \
2424         ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2425                 (jiffies - (rq)->expired_timestamp >= \
2426                         STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2427                         ((rq)->curr->static_prio > (rq)->best_expired_prio))
2428
2429 /*
2430  * Account user cpu time to a process.
2431  * @p: the process that the cpu time gets accounted to
2432  * @hardirq_offset: the offset to subtract from hardirq_count()
2433  * @cputime: the cpu time spent in user space since the last update
2434  */
2435 void account_user_time(struct task_struct *p, cputime_t cputime)
2436 {
2437         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2438         cputime64_t tmp;
2439
2440         p->utime = cputime_add(p->utime, cputime);
2441
2442         /* Add user time to cpustat. */
2443         tmp = cputime_to_cputime64(cputime);
2444         if (TASK_NICE(p) > 0)
2445                 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2446         else
2447                 cpustat->user = cputime64_add(cpustat->user, tmp);
2448 }
2449
2450 /*
2451  * Account system cpu time to a process.
2452  * @p: the process that the cpu time gets accounted to
2453  * @hardirq_offset: the offset to subtract from hardirq_count()
2454  * @cputime: the cpu time spent in kernel space since the last update
2455  */
2456 void account_system_time(struct task_struct *p, int hardirq_offset,
2457                          cputime_t cputime)
2458 {
2459         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2460         runqueue_t *rq = this_rq();
2461         cputime64_t tmp;
2462
2463         p->stime = cputime_add(p->stime, cputime);
2464
2465         /* Add system time to cpustat. */
2466         tmp = cputime_to_cputime64(cputime);
2467         if (hardirq_count() - hardirq_offset)
2468                 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2469         else if (softirq_count())
2470                 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2471         else if (p != rq->idle)
2472                 cpustat->system = cputime64_add(cpustat->system, tmp);
2473         else if (atomic_read(&rq->nr_iowait) > 0)
2474                 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2475         else
2476                 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2477         /* Account for system time used */
2478         acct_update_integrals(p);
2479         /* Update rss highwater mark */
2480         update_mem_hiwater(p);
2481 }
2482
2483 /*
2484  * Account for involuntary wait time.
2485  * @p: the process from which the cpu time has been stolen
2486  * @steal: the cpu time spent in involuntary wait
2487  */
2488 void account_steal_time(struct task_struct *p, cputime_t steal)
2489 {
2490         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2491         cputime64_t tmp = cputime_to_cputime64(steal);
2492         runqueue_t *rq = this_rq();
2493
2494         if (p == rq->idle) {
2495                 p->stime = cputime_add(p->stime, steal);
2496                 if (atomic_read(&rq->nr_iowait) > 0)
2497                         cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2498                 else
2499                         cpustat->idle = cputime64_add(cpustat->idle, tmp);
2500         } else
2501                 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2502 }
2503
2504 /*
2505  * This function gets called by the timer code, with HZ frequency.
2506  * We call it with interrupts disabled.
2507  *
2508  * It also gets called by the fork code, when changing the parent's
2509  * timeslices.
2510  */
2511 void scheduler_tick(void)
2512 {
2513         int cpu = smp_processor_id();
2514         runqueue_t *rq = this_rq();
2515         task_t *p = current;
2516         unsigned long long now = sched_clock();
2517
2518         update_cpu_clock(p, rq, now);
2519
2520         rq->timestamp_last_tick = now;
2521
2522         if (p == rq->idle) {
2523                 if (wake_priority_sleeper(rq))
2524                         goto out;
2525                 rebalance_tick(cpu, rq, SCHED_IDLE);
2526                 return;
2527         }
2528
2529         /* Task might have expired already, but not scheduled off yet */
2530         if (p->array != rq->active) {
2531                 set_tsk_need_resched(p);
2532                 goto out;
2533         }
2534         spin_lock(&rq->lock);
2535         /*
2536          * The task was running during this tick - update the
2537          * time slice counter. Note: we do not update a thread's
2538          * priority until it either goes to sleep or uses up its
2539          * timeslice. This makes it possible for interactive tasks
2540          * to use up their timeslices at their highest priority levels.
2541          */
2542         if (rt_task(p)) {
2543                 /*
2544                  * RR tasks need a special form of timeslice management.
2545                  * FIFO tasks have no timeslices.
2546                  */
2547                 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2548                         p->time_slice = task_timeslice(p);
2549                         p->first_time_slice = 0;
2550                         set_tsk_need_resched(p);
2551
2552                         /* put it at the end of the queue: */
2553                         requeue_task(p, rq->active);
2554                 }
2555                 goto out_unlock;
2556         }
2557         if (!--p->time_slice) {
2558                 dequeue_task(p, rq->active);
2559                 set_tsk_need_resched(p);
2560                 p->prio = effective_prio(p);
2561                 p->time_slice = task_timeslice(p);
2562                 p->first_time_slice = 0;
2563
2564                 if (!rq->expired_timestamp)
2565                         rq->expired_timestamp = jiffies;
2566                 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2567                         enqueue_task(p, rq->expired);
2568                         if (p->static_prio < rq->best_expired_prio)
2569                                 rq->best_expired_prio = p->static_prio;
2570                 } else
2571                         enqueue_task(p, rq->active);
2572         } else {
2573                 /*
2574                  * Prevent a too long timeslice allowing a task to monopolize
2575                  * the CPU. We do this by splitting up the timeslice into
2576                  * smaller pieces.
2577                  *
2578                  * Note: this does not mean the task's timeslices expire or
2579                  * get lost in any way, they just might be preempted by
2580                  * another task of equal priority. (one with higher
2581                  * priority would have preempted this task already.) We
2582                  * requeue this task to the end of the list on this priority
2583                  * level, which is in essence a round-robin of tasks with
2584                  * equal priority.
2585                  *
2586                  * This only applies to tasks in the interactive
2587                  * delta range with at least TIMESLICE_GRANULARITY to requeue.
2588                  */
2589                 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2590                         p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2591                         (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2592                         (p->array == rq->active)) {
2593
2594                         requeue_task(p, rq->active);
2595                         set_tsk_need_resched(p);
2596                 }
2597         }
2598 out_unlock:
2599         spin_unlock(&rq->lock);
2600 out:
2601         rebalance_tick(cpu, rq, NOT_IDLE);
2602 }
2603
2604 #ifdef CONFIG_SCHED_SMT
2605 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2606 {
2607         /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2608         if (rq->curr == rq->idle && rq->nr_running)
2609                 resched_task(rq->idle);
2610 }
2611
2612 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2613 {
2614         struct sched_domain *tmp, *sd = NULL;
2615         cpumask_t sibling_map;
2616         int i;
2617
2618         for_each_domain(this_cpu, tmp)
2619                 if (tmp->flags & SD_SHARE_CPUPOWER)
2620                         sd = tmp;
2621
2622         if (!sd)
2623                 return;
2624
2625         /*
2626          * Unlock the current runqueue because we have to lock in
2627          * CPU order to avoid deadlocks. Caller knows that we might
2628          * unlock. We keep IRQs disabled.
2629          */
2630         spin_unlock(&this_rq->lock);
2631
2632         sibling_map = sd->span;
2633
2634         for_each_cpu_mask(i, sibling_map)
2635                 spin_lock(&cpu_rq(i)->lock);
2636         /*
2637          * We clear this CPU from the mask. This both simplifies the
2638          * inner loop and keps this_rq locked when we exit:
2639          */
2640         cpu_clear(this_cpu, sibling_map);
2641
2642         for_each_cpu_mask(i, sibling_map) {
2643                 runqueue_t *smt_rq = cpu_rq(i);
2644
2645                 wakeup_busy_runqueue(smt_rq);
2646         }
2647
2648         for_each_cpu_mask(i, sibling_map)
2649                 spin_unlock(&cpu_rq(i)->lock);
2650         /*
2651          * We exit with this_cpu's rq still held and IRQs
2652          * still disabled:
2653          */
2654 }
2655
2656 /*
2657  * number of 'lost' timeslices this task wont be able to fully
2658  * utilize, if another task runs on a sibling. This models the
2659  * slowdown effect of other tasks running on siblings:
2660  */
2661 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2662 {
2663         return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2664 }
2665
2666 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2667 {
2668         struct sched_domain *tmp, *sd = NULL;
2669         cpumask_t sibling_map;
2670         prio_array_t *array;
2671         int ret = 0, i;
2672         task_t *p;
2673
2674         for_each_domain(this_cpu, tmp)
2675                 if (tmp->flags & SD_SHARE_CPUPOWER)
2676                         sd = tmp;
2677
2678         if (!sd)
2679                 return 0;
2680
2681         /*
2682          * The same locking rules and details apply as for
2683          * wake_sleeping_dependent():
2684          */
2685         spin_unlock(&this_rq->lock);
2686         sibling_map = sd->span;
2687         for_each_cpu_mask(i, sibling_map)
2688                 spin_lock(&cpu_rq(i)->lock);
2689         cpu_clear(this_cpu, sibling_map);
2690
2691         /*
2692          * Establish next task to be run - it might have gone away because
2693          * we released the runqueue lock above:
2694          */
2695         if (!this_rq->nr_running)
2696                 goto out_unlock;
2697         array = this_rq->active;
2698         if (!array->nr_active)
2699                 array = this_rq->expired;
2700         BUG_ON(!array->nr_active);
2701
2702         p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2703                 task_t, run_list);
2704
2705         for_each_cpu_mask(i, sibling_map) {
2706                 runqueue_t *smt_rq = cpu_rq(i);
2707                 task_t *smt_curr = smt_rq->curr;
2708
2709                 /* Kernel threads do not participate in dependent sleeping */
2710                 if (!p->mm || !smt_curr->mm || rt_task(p))
2711                         goto check_smt_task;
2712
2713                 /*
2714                  * If a user task with lower static priority than the
2715                  * running task on the SMT sibling is trying to schedule,
2716                  * delay it till there is proportionately less timeslice
2717                  * left of the sibling task to prevent a lower priority
2718                  * task from using an unfair proportion of the
2719                  * physical cpu's resources. -ck
2720                  */
2721                 if (rt_task(smt_curr)) {
2722                         /*
2723                          * With real time tasks we run non-rt tasks only
2724                          * per_cpu_gain% of the time.
2725                          */
2726                         if ((jiffies % DEF_TIMESLICE) >
2727                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2728                                         ret = 1;
2729                 } else
2730                         if (smt_curr->static_prio < p->static_prio &&
2731                                 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2732                                 smt_slice(smt_curr, sd) > task_timeslice(p))
2733                                         ret = 1;
2734
2735 check_smt_task:
2736                 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2737                         rt_task(smt_curr))
2738                                 continue;
2739                 if (!p->mm) {
2740                         wakeup_busy_runqueue(smt_rq);
2741                         continue;
2742                 }
2743
2744                 /*
2745                  * Reschedule a lower priority task on the SMT sibling for
2746                  * it to be put to sleep, or wake it up if it has been put to
2747                  * sleep for priority reasons to see if it should run now.
2748                  */
2749                 if (rt_task(p)) {
2750                         if ((jiffies % DEF_TIMESLICE) >
2751                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2752                                         resched_task(smt_curr);
2753                 } else {
2754                         if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2755                                 smt_slice(p, sd) > task_timeslice(smt_curr))
2756                                         resched_task(smt_curr);
2757                         else
2758                                 wakeup_busy_runqueue(smt_rq);
2759                 }
2760         }
2761 out_unlock:
2762         for_each_cpu_mask(i, sibling_map)
2763                 spin_unlock(&cpu_rq(i)->lock);
2764         return ret;
2765 }
2766 #else
2767 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2768 {
2769 }
2770
2771 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2772 {
2773         return 0;
2774 }
2775 #endif
2776
2777 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2778
2779 void fastcall add_preempt_count(int val)
2780 {
2781         /*
2782          * Underflow?
2783          */
2784         BUG_ON((preempt_count() < 0));
2785         preempt_count() += val;
2786         /*
2787          * Spinlock count overflowing soon?
2788          */
2789         BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2790 }
2791 EXPORT_SYMBOL(add_preempt_count);
2792
2793 void fastcall sub_preempt_count(int val)
2794 {
2795         /*
2796          * Underflow?
2797          */
2798         BUG_ON(val > preempt_count());
2799         /*
2800          * Is the spinlock portion underflowing?
2801          */
2802         BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2803         preempt_count() -= val;
2804 }
2805 EXPORT_SYMBOL(sub_preempt_count);
2806
2807 #endif
2808
2809 /*
2810  * schedule() is the main scheduler function.
2811  */
2812 asmlinkage void __sched schedule(void)
2813 {
2814         long *switch_count;
2815         task_t *prev, *next;
2816         runqueue_t *rq;
2817         prio_array_t *array;
2818         struct list_head *queue;
2819         unsigned long long now;
2820         unsigned long run_time;
2821         int cpu, idx, new_prio;
2822
2823         /*
2824          * Test if we are atomic.  Since do_exit() needs to call into
2825          * schedule() atomically, we ignore that path for now.
2826          * Otherwise, whine if we are scheduling when we should not be.
2827          */
2828         if (likely(!current->exit_state)) {
2829                 if (unlikely(in_atomic())) {
2830                         printk(KERN_ERR "scheduling while atomic: "
2831                                 "%s/0x%08x/%d\n",
2832                                 current->comm, preempt_count(), current->pid);
2833                         dump_stack();
2834                 }
2835         }
2836         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2837
2838 need_resched:
2839         preempt_disable();
2840         prev = current;
2841         release_kernel_lock(prev);
2842 need_resched_nonpreemptible:
2843         rq = this_rq();
2844
2845         /*
2846          * The idle thread is not allowed to schedule!
2847          * Remove this check after it has been exercised a bit.
2848          */
2849         if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2850                 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2851                 dump_stack();
2852         }
2853
2854         schedstat_inc(rq, sched_cnt);
2855         now = sched_clock();
2856         if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2857                 run_time = now - prev->timestamp;
2858                 if (unlikely((long long)(now - prev->timestamp) < 0))
2859                         run_time = 0;
2860         } else
2861                 run_time = NS_MAX_SLEEP_AVG;
2862
2863         /*
2864          * Tasks charged proportionately less run_time at high sleep_avg to
2865          * delay them losing their interactive status
2866          */
2867         run_time /= (CURRENT_BONUS(prev) ? : 1);
2868
2869         spin_lock_irq(&rq->lock);
2870
2871         if (unlikely(prev->flags & PF_DEAD))
2872                 prev->state = EXIT_DEAD;
2873
2874         switch_count = &prev->nivcsw;
2875         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2876                 switch_count = &prev->nvcsw;
2877                 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2878                                 unlikely(signal_pending(prev))))
2879                         prev->state = TASK_RUNNING;
2880                 else {
2881                         if (prev->state == TASK_UNINTERRUPTIBLE)
2882                                 rq->nr_uninterruptible++;
2883                         deactivate_task(prev, rq);
2884                 }
2885         }
2886
2887         cpu = smp_processor_id();
2888         if (unlikely(!rq->nr_running)) {
2889 go_idle:
2890                 idle_balance(cpu, rq);
2891                 if (!rq->nr_running) {
2892                         next = rq->idle;
2893                         rq->expired_timestamp = 0;
2894                         wake_sleeping_dependent(cpu, rq);
2895                         /*
2896                          * wake_sleeping_dependent() might have released
2897                          * the runqueue, so break out if we got new
2898                          * tasks meanwhile:
2899                          */
2900                         if (!rq->nr_running)
2901                                 goto switch_tasks;
2902                 }
2903         } else {
2904                 if (dependent_sleeper(cpu, rq)) {
2905                         next = rq->idle;
2906                         goto switch_tasks;
2907                 }
2908                 /*
2909                  * dependent_sleeper() releases and reacquires the runqueue
2910                  * lock, hence go into the idle loop if the rq went
2911                  * empty meanwhile:
2912                  */
2913                 if (unlikely(!rq->nr_running))
2914                         goto go_idle;
2915         }
2916
2917         array = rq->active;
2918         if (unlikely(!array->nr_active)) {
2919                 /*
2920                  * Switch the active and expired arrays.
2921                  */
2922                 schedstat_inc(rq, sched_switch);
2923                 rq->active = rq->expired;
2924                 rq->expired = array;
2925                 array = rq->active;
2926                 rq->expired_timestamp = 0;
2927                 rq->best_expired_prio = MAX_PRIO;
2928         }
2929
2930         idx = sched_find_first_bit(array->bitmap);
2931         queue = array->queue + idx;
2932         next = list_entry(queue->next, task_t, run_list);
2933
2934         if (!rt_task(next) && next->activated > 0) {
2935                 unsigned long long delta = now - next->timestamp;
2936                 if (unlikely((long long)(now - next->timestamp) < 0))
2937                         delta = 0;
2938
2939                 if (next->activated == 1)
2940                         delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2941
2942                 array = next->array;
2943                 new_prio = recalc_task_prio(next, next->timestamp + delta);
2944
2945                 if (unlikely(next->prio != new_prio)) {
2946                         dequeue_task(next, array);
2947                         next->prio = new_prio;
2948                         enqueue_task(next, array);
2949                 } else
2950                         requeue_task(next, array);
2951         }
2952         next->activated = 0;
2953 switch_tasks:
2954         if (next == rq->idle)
2955                 schedstat_inc(rq, sched_goidle);
2956         prefetch(next);
2957         prefetch_stack(next);
2958         clear_tsk_need_resched(prev);
2959         rcu_qsctr_inc(task_cpu(prev));
2960
2961         update_cpu_clock(prev, rq, now);
2962
2963         prev->sleep_avg -= run_time;
2964         if ((long)prev->sleep_avg <= 0)
2965                 prev->sleep_avg = 0;
2966         prev->timestamp = prev->last_ran = now;
2967
2968         sched_info_switch(prev, next);
2969         if (likely(prev != next)) {
2970                 next->timestamp = now;
2971                 rq->nr_switches++;
2972                 rq->curr = next;
2973                 ++*switch_count;
2974
2975                 prepare_task_switch(rq, next);
2976                 prev = context_switch(rq, prev, next);
2977                 barrier();
2978                 /*
2979                  * this_rq must be evaluated again because prev may have moved
2980                  * CPUs since it called schedule(), thus the 'rq' on its stack
2981                  * frame will be invalid.
2982                  */
2983                 finish_task_switch(this_rq(), prev);
2984         } else
2985                 spin_unlock_irq(&rq->lock);
2986
2987         prev = current;
2988         if (unlikely(reacquire_kernel_lock(prev) < 0))
2989                 goto need_resched_nonpreemptible;
2990         preempt_enable_no_resched();
2991         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2992                 goto need_resched;
2993 }
2994
2995 EXPORT_SYMBOL(schedule);
2996
2997 #ifdef CONFIG_PREEMPT
2998 /*
2999  * this is is the entry point to schedule() from in-kernel preemption
3000  * off of preempt_enable.  Kernel preemptions off return from interrupt
3001  * occur there and call schedule directly.
3002  */
3003 asmlinkage void __sched preempt_schedule(void)
3004 {
3005         struct thread_info *ti = current_thread_info();
3006 #ifdef CONFIG_PREEMPT_BKL
3007         struct task_struct *task = current;
3008         int saved_lock_depth;
3009 #endif
3010         /*
3011          * If there is a non-zero preempt_count or interrupts are disabled,
3012          * we do not want to preempt the current task.  Just return..
3013          */
3014         if (unlikely(ti->preempt_count || irqs_disabled()))
3015                 return;
3016
3017 need_resched:
3018         add_preempt_count(PREEMPT_ACTIVE);
3019         /*
3020          * We keep the big kernel semaphore locked, but we
3021          * clear ->lock_depth so that schedule() doesnt
3022          * auto-release the semaphore:
3023          */
3024 #ifdef CONFIG_PREEMPT_BKL
3025         saved_lock_depth = task->lock_depth;
3026         task->lock_depth = -1;
3027 #endif
3028         schedule();
3029 #ifdef CONFIG_PREEMPT_BKL
3030         task->lock_depth = saved_lock_depth;
3031 #endif
3032         sub_preempt_count(PREEMPT_ACTIVE);
3033
3034         /* we could miss a preemption opportunity between schedule and now */
3035         barrier();
3036         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3037                 goto need_resched;
3038 }
3039
3040 EXPORT_SYMBOL(preempt_schedule);
3041
3042 /*
3043  * this is is the entry point to schedule() from kernel preemption
3044  * off of irq context.
3045  * Note, that this is called and return with irqs disabled. This will
3046  * protect us against recursive calling from irq.
3047  */
3048 asmlinkage void __sched preempt_schedule_irq(void)
3049 {
3050         struct thread_info *ti = current_thread_info();
3051 #ifdef CONFIG_PREEMPT_BKL
3052         struct task_struct *task = current;
3053         int saved_lock_depth;
3054 #endif
3055         /* Catch callers which need to be fixed*/
3056         BUG_ON(ti->preempt_count || !irqs_disabled());
3057
3058 need_resched:
3059         add_preempt_count(PREEMPT_ACTIVE);
3060         /*
3061          * We keep the big kernel semaphore locked, but we
3062          * clear ->lock_depth so that schedule() doesnt
3063          * auto-release the semaphore:
3064          */
3065 #ifdef CONFIG_PREEMPT_BKL
3066         saved_lock_depth = task->lock_depth;
3067         task->lock_depth = -1;
3068 #endif
3069         local_irq_enable();
3070         schedule();
3071         local_irq_disable();
3072 #ifdef CONFIG_PREEMPT_BKL
3073         task->lock_depth = saved_lock_depth;
3074 #endif
3075         sub_preempt_count(PREEMPT_ACTIVE);
3076
3077         /* we could miss a preemption opportunity between schedule and now */
3078         barrier();
3079         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3080                 goto need_resched;
3081 }
3082
3083 #endif /* CONFIG_PREEMPT */
3084
3085 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3086                           void *key)
3087 {
3088         task_t *p = curr->private;
3089         return try_to_wake_up(p, mode, sync);
3090 }
3091
3092 EXPORT_SYMBOL(default_wake_function);
3093
3094 /*
3095  * The core wakeup function.  Non-exclusive wakeups (nr_exclusive == 0) just
3096  * wake everything up.  If it's an exclusive wakeup (nr_exclusive == small +ve
3097  * number) then we wake all the non-exclusive tasks and one exclusive task.
3098  *
3099  * There are circumstances in which we can try to wake a task which has already
3100  * started to run but is not in state TASK_RUNNING.  try_to_wake_up() returns
3101  * zero in this (rare) case, and we handle it by continuing to scan the queue.
3102  */
3103 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3104                              int nr_exclusive, int sync, void *key)
3105 {
3106         struct list_head *tmp, *next;
3107
3108         list_for_each_safe(tmp, next, &q->task_list) {
3109                 wait_queue_t *curr;
3110                 unsigned flags;
3111                 curr = list_entry(tmp, wait_queue_t, task_list);
3112                 flags = curr->flags;
3113                 if (curr->func(curr, mode, sync, key) &&
3114                     (flags & WQ_FLAG_EXCLUSIVE) &&
3115                     !--nr_exclusive)
3116                         break;
3117         }
3118 }
3119
3120 /**
3121  * __wake_up - wake up threads blocked on a waitqueue.
3122  * @q: the waitqueue
3123  * @mode: which threads
3124  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3125  * @key: is directly passed to the wakeup function
3126  */
3127 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3128                         int nr_exclusive, void *key)
3129 {
3130         unsigned long flags;
3131
3132         spin_lock_irqsave(&q->lock, flags);
3133         __wake_up_common(q, mode, nr_exclusive, 0, key);
3134         spin_unlock_irqrestore(&q->lock, flags);
3135 }
3136
3137 EXPORT_SYMBOL(__wake_up);
3138
3139 /*
3140  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3141  */
3142 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3143 {
3144         __wake_up_common(q, mode, 1, 0, NULL);
3145 }
3146
3147 /**
3148  * __wake_up_sync - wake up threads blocked on a waitqueue.
3149  * @q: the waitqueue
3150  * @mode: which threads
3151  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3152  *
3153  * The sync wakeup differs that the waker knows that it will schedule
3154  * away soon, so while the target thread will be woken up, it will not
3155  * be migrated to another CPU - ie. the two threads are 'synchronized'
3156  * with each other. This can prevent needless bouncing between CPUs.
3157  *
3158  * On UP it can prevent extra preemption.
3159  */
3160 void fastcall
3161 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3162 {
3163         unsigned long flags;
3164         int sync = 1;
3165
3166         if (unlikely(!q))
3167                 return;
3168
3169         if (unlikely(!nr_exclusive))
3170                 sync = 0;
3171
3172         spin_lock_irqsave(&q->lock, flags);
3173         __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3174         spin_unlock_irqrestore(&q->lock, flags);
3175 }
3176 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3177
3178 void fastcall complete(struct completion *x)
3179 {
3180         unsigned long flags;
3181
3182         spin_lock_irqsave(&x->wait.lock, flags);
3183         x->done++;
3184         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3185                          1, 0, NULL);
3186         spin_unlock_irqrestore(&x->wait.lock, flags);
3187 }
3188 EXPORT_SYMBOL(complete);
3189
3190 void fastcall complete_all(struct completion *x)
3191 {
3192         unsigned long flags;
3193
3194         spin_lock_irqsave(&x->wait.lock, flags);
3195         x->done += UINT_MAX/2;
3196         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3197                          0, 0, NULL);
3198         spin_unlock_irqrestore(&x->wait.lock, flags);
3199 }
3200 EXPORT_SYMBOL(complete_all);
3201
3202 void fastcall __sched wait_for_completion(struct completion *x)
3203 {
3204         might_sleep();
3205         spin_lock_irq(&x->wait.lock);
3206         if (!x->done) {
3207                 DECLARE_WAITQUEUE(wait, current);
3208
3209                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3210                 __add_wait_queue_tail(&x->wait, &wait);
3211                 do {
3212                         __set_current_state(TASK_UNINTERRUPTIBLE);
3213                         spin_unlock_irq(&x->wait.lock);
3214                         schedule();
3215                         spin_lock_irq(&x->wait.lock);
3216                 } while (!x->done);
3217                 __remove_wait_queue(&x->wait, &wait);
3218         }
3219         x->done--;
3220         spin_unlock_irq(&x->wait.lock);
3221 }
3222 EXPORT_SYMBOL(wait_for_completion);
3223
3224 unsigned long fastcall __sched
3225 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3226 {
3227         might_sleep();
3228
3229         spin_lock_irq(&x->wait.lock);
3230         if (!x->done) {
3231                 DECLARE_WAITQUEUE(wait, current);
3232
3233                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3234                 __add_wait_queue_tail(&x->wait, &wait);
3235                 do {
3236                         __set_current_state(TASK_UNINTERRUPTIBLE);
3237                         spin_unlock_irq(&x->wait.lock);
3238                         timeout = schedule_timeout(timeout);
3239                         spin_lock_irq(&x->wait.lock);
3240                         if (!timeout) {
3241                                 __remove_wait_queue(&x->wait, &wait);
3242                                 goto out;
3243                         }
3244                 } while (!x->done);
3245                 __remove_wait_queue(&x->wait, &wait);
3246         }
3247         x->done--;
3248 out:
3249         spin_unlock_irq(&x->wait.lock);
3250         return timeout;
3251 }
3252 EXPORT_SYMBOL(wait_for_completion_timeout);
3253
3254 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3255 {
3256         int ret = 0;
3257
3258         might_sleep();
3259
3260         spin_lock_irq(&x->wait.lock);
3261         if (!x->done) {
3262                 DECLARE_WAITQUEUE(wait, current);
3263
3264                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3265                 __add_wait_queue_tail(&x->wait, &wait);
3266                 do {
3267                         if (signal_pending(current)) {
3268                                 ret = -ERESTARTSYS;
3269                                 __remove_wait_queue(&x->wait, &wait);
3270                                 goto out;
3271                         }
3272                         __set_current_state(TASK_INTERRUPTIBLE);
3273                         spin_unlock_irq(&x->wait.lock);
3274                         schedule();
3275                         spin_lock_irq(&x->wait.lock);
3276                 } while (!x->done);
3277                 __remove_wait_queue(&x->wait, &wait);
3278         }
3279         x->done--;
3280 out:
3281         spin_unlock_irq(&x->wait.lock);
3282
3283         return ret;
3284 }
3285 EXPORT_SYMBOL(wait_for_completion_interruptible);
3286
3287 unsigned long fastcall __sched
3288 wait_for_completion_interruptible_timeout(struct completion *x,
3289                                           unsigned long timeout)
3290 {
3291         might_sleep();
3292
3293         spin_lock_irq(&x->wait.lock);
3294         if (!x->done) {
3295                 DECLARE_WAITQUEUE(wait, current);
3296
3297                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3298                 __add_wait_queue_tail(&x->wait, &wait);
3299                 do {
3300                         if (signal_pending(current)) {
3301                                 timeout = -ERESTARTSYS;
3302                                 __remove_wait_queue(&x->wait, &wait);
3303                                 goto out;
3304                         }
3305                         __set_current_state(TASK_INTERRUPTIBLE);
3306                         spin_unlock_irq(&x->wait.lock);
3307                         timeout = schedule_timeout(timeout);
3308                         spin_lock_irq(&x->wait.lock);
3309                         if (!timeout) {
3310                                 __remove_wait_queue(&x->wait, &wait);
3311                                 goto out;
3312                         }
3313                 } while (!x->done);
3314                 __remove_wait_queue(&x->wait, &wait);
3315         }
3316         x->done--;
3317 out:
3318         spin_unlock_irq(&x->wait.lock);
3319         return timeout;
3320 }
3321 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3322
3323
3324 #define SLEEP_ON_VAR                                    \
3325         unsigned long flags;                            \
3326         wait_queue_t wait;                              \
3327         init_waitqueue_entry(&wait, current);
3328
3329 #define SLEEP_ON_HEAD                                   \
3330         spin_lock_irqsave(&q->lock,flags);              \
3331         __add_wait_queue(q, &wait);                     \
3332         spin_unlock(&q->lock);
3333
3334 #define SLEEP_ON_TAIL                                   \
3335         spin_lock_irq(&q->lock);                        \
3336         __remove_wait_queue(q, &wait);                  \
3337         spin_unlock_irqrestore(&q->lock, flags);