Merge rsync://bughost.org/repos/ieee80211-delta/
[linux-2.6] / 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 #ifdef CONFIG_DEBUG_SPINLOCK
298         /* this is a valid case when another task releases the spinlock */
299         rq->lock.owner = current;
300 #endif
301         spin_unlock_irq(&rq->lock);
302 }
303
304 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
305 static inline int task_running(runqueue_t *rq, task_t *p)
306 {
307 #ifdef CONFIG_SMP
308         return p->oncpu;
309 #else
310         return rq->curr == p;
311 #endif
312 }
313
314 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
315 {
316 #ifdef CONFIG_SMP
317         /*
318          * We can optimise this out completely for !SMP, because the
319          * SMP rebalancing from interrupt is the only thing that cares
320          * here.
321          */
322         next->oncpu = 1;
323 #endif
324 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
325         spin_unlock_irq(&rq->lock);
326 #else
327         spin_unlock(&rq->lock);
328 #endif
329 }
330
331 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
332 {
333 #ifdef CONFIG_SMP
334         /*
335          * After ->oncpu is cleared, the task can be moved to a different CPU.
336          * We must ensure this doesn't happen until the switch is completely
337          * finished.
338          */
339         smp_wmb();
340         prev->oncpu = 0;
341 #endif
342 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
343         local_irq_enable();
344 #endif
345 }
346 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
347
348 /*
349  * task_rq_lock - lock the runqueue a given task resides on and disable
350  * interrupts.  Note the ordering: we can safely lookup the task_rq without
351  * explicitly disabling preemption.
352  */
353 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
354         __acquires(rq->lock)
355 {
356         struct runqueue *rq;
357
358 repeat_lock_task:
359         local_irq_save(*flags);
360         rq = task_rq(p);
361         spin_lock(&rq->lock);
362         if (unlikely(rq != task_rq(p))) {
363                 spin_unlock_irqrestore(&rq->lock, *flags);
364                 goto repeat_lock_task;
365         }
366         return rq;
367 }
368
369 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
370         __releases(rq->lock)
371 {
372         spin_unlock_irqrestore(&rq->lock, *flags);
373 }
374
375 #ifdef CONFIG_SCHEDSTATS
376 /*
377  * bump this up when changing the output format or the meaning of an existing
378  * format, so that tools can adapt (or abort)
379  */
380 #define SCHEDSTAT_VERSION 12
381
382 static int show_schedstat(struct seq_file *seq, void *v)
383 {
384         int cpu;
385
386         seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
387         seq_printf(seq, "timestamp %lu\n", jiffies);
388         for_each_online_cpu(cpu) {
389                 runqueue_t *rq = cpu_rq(cpu);
390 #ifdef CONFIG_SMP
391                 struct sched_domain *sd;
392                 int dcnt = 0;
393 #endif
394
395                 /* runqueue-specific stats */
396                 seq_printf(seq,
397                     "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
398                     cpu, rq->yld_both_empty,
399                     rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
400                     rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
401                     rq->ttwu_cnt, rq->ttwu_local,
402                     rq->rq_sched_info.cpu_time,
403                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
404
405                 seq_printf(seq, "\n");
406
407 #ifdef CONFIG_SMP
408                 /* domain-specific stats */
409                 preempt_disable();
410                 for_each_domain(cpu, sd) {
411                         enum idle_type itype;
412                         char mask_str[NR_CPUS];
413
414                         cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
415                         seq_printf(seq, "domain%d %s", dcnt++, mask_str);
416                         for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
417                                         itype++) {
418                                 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
419                                     sd->lb_cnt[itype],
420                                     sd->lb_balanced[itype],
421                                     sd->lb_failed[itype],
422                                     sd->lb_imbalance[itype],
423                                     sd->lb_gained[itype],
424                                     sd->lb_hot_gained[itype],
425                                     sd->lb_nobusyq[itype],
426                                     sd->lb_nobusyg[itype]);
427                         }
428                         seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
429                             sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
430                             sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
431                             sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
432                             sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
433                 }
434                 preempt_enable();
435 #endif
436         }
437         return 0;
438 }
439
440 static int schedstat_open(struct inode *inode, struct file *file)
441 {
442         unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
443         char *buf = kmalloc(size, GFP_KERNEL);
444         struct seq_file *m;
445         int res;
446
447         if (!buf)
448                 return -ENOMEM;
449         res = single_open(file, show_schedstat, NULL);
450         if (!res) {
451                 m = file->private_data;
452                 m->buf = buf;
453                 m->size = size;
454         } else
455                 kfree(buf);
456         return res;
457 }
458
459 struct file_operations proc_schedstat_operations = {
460         .open    = schedstat_open,
461         .read    = seq_read,
462         .llseek  = seq_lseek,
463         .release = single_release,
464 };
465
466 # define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
467 # define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
468 #else /* !CONFIG_SCHEDSTATS */
469 # define schedstat_inc(rq, field)       do { } while (0)
470 # define schedstat_add(rq, field, amt)  do { } while (0)
471 #endif
472
473 /*
474  * rq_lock - lock a given runqueue and disable interrupts.
475  */
476 static inline runqueue_t *this_rq_lock(void)
477         __acquires(rq->lock)
478 {
479         runqueue_t *rq;
480
481         local_irq_disable();
482         rq = this_rq();
483         spin_lock(&rq->lock);
484
485         return rq;
486 }
487
488 #ifdef CONFIG_SCHEDSTATS
489 /*
490  * Called when a process is dequeued from the active array and given
491  * the cpu.  We should note that with the exception of interactive
492  * tasks, the expired queue will become the active queue after the active
493  * queue is empty, without explicitly dequeuing and requeuing tasks in the
494  * expired queue.  (Interactive tasks may be requeued directly to the
495  * active queue, thus delaying tasks in the expired queue from running;
496  * see scheduler_tick()).
497  *
498  * This function is only called from sched_info_arrive(), rather than
499  * dequeue_task(). Even though a task may be queued and dequeued multiple
500  * times as it is shuffled about, we're really interested in knowing how
501  * long it was from the *first* time it was queued to the time that it
502  * finally hit a cpu.
503  */
504 static inline void sched_info_dequeued(task_t *t)
505 {
506         t->sched_info.last_queued = 0;
507 }
508
509 /*
510  * Called when a task finally hits the cpu.  We can now calculate how
511  * long it was waiting to run.  We also note when it began so that we
512  * can keep stats on how long its timeslice is.
513  */
514 static inline void sched_info_arrive(task_t *t)
515 {
516         unsigned long now = jiffies, diff = 0;
517         struct runqueue *rq = task_rq(t);
518
519         if (t->sched_info.last_queued)
520                 diff = now - t->sched_info.last_queued;
521         sched_info_dequeued(t);
522         t->sched_info.run_delay += diff;
523         t->sched_info.last_arrival = now;
524         t->sched_info.pcnt++;
525
526         if (!rq)
527                 return;
528
529         rq->rq_sched_info.run_delay += diff;
530         rq->rq_sched_info.pcnt++;
531 }
532
533 /*
534  * Called when a process is queued into either the active or expired
535  * array.  The time is noted and later used to determine how long we
536  * had to wait for us to reach the cpu.  Since the expired queue will
537  * become the active queue after active queue is empty, without dequeuing
538  * and requeuing any tasks, we are interested in queuing to either. It
539  * is unusual but not impossible for tasks to be dequeued and immediately
540  * requeued in the same or another array: this can happen in sched_yield(),
541  * set_user_nice(), and even load_balance() as it moves tasks from runqueue
542  * to runqueue.
543  *
544  * This function is only called from enqueue_task(), but also only updates
545  * the timestamp if it is already not set.  It's assumed that
546  * sched_info_dequeued() will clear that stamp when appropriate.
547  */
548 static inline void sched_info_queued(task_t *t)
549 {
550         if (!t->sched_info.last_queued)
551                 t->sched_info.last_queued = jiffies;
552 }
553
554 /*
555  * Called when a process ceases being the active-running process, either
556  * voluntarily or involuntarily.  Now we can calculate how long we ran.
557  */
558 static inline void sched_info_depart(task_t *t)
559 {
560         struct runqueue *rq = task_rq(t);
561         unsigned long diff = jiffies - t->sched_info.last_arrival;
562
563         t->sched_info.cpu_time += diff;
564
565         if (rq)
566                 rq->rq_sched_info.cpu_time += diff;
567 }
568
569 /*
570  * Called when tasks are switched involuntarily due, typically, to expiring
571  * their time slice.  (This may also be called when switching to or from
572  * the idle task.)  We are only called when prev != next.
573  */
574 static inline void sched_info_switch(task_t *prev, task_t *next)
575 {
576         struct runqueue *rq = task_rq(prev);
577
578         /*
579          * prev now departs the cpu.  It's not interesting to record
580          * stats about how efficient we were at scheduling the idle
581          * process, however.
582          */
583         if (prev != rq->idle)
584                 sched_info_depart(prev);
585
586         if (next != rq->idle)
587                 sched_info_arrive(next);
588 }
589 #else
590 #define sched_info_queued(t)            do { } while (0)
591 #define sched_info_switch(t, next)      do { } while (0)
592 #endif /* CONFIG_SCHEDSTATS */
593
594 /*
595  * Adding/removing a task to/from a priority array:
596  */
597 static void dequeue_task(struct task_struct *p, prio_array_t *array)
598 {
599         array->nr_active--;
600         list_del(&p->run_list);
601         if (list_empty(array->queue + p->prio))
602                 __clear_bit(p->prio, array->bitmap);
603 }
604
605 static void enqueue_task(struct task_struct *p, prio_array_t *array)
606 {
607         sched_info_queued(p);
608         list_add_tail(&p->run_list, array->queue + p->prio);
609         __set_bit(p->prio, array->bitmap);
610         array->nr_active++;
611         p->array = array;
612 }
613
614 /*
615  * Put task to the end of the run list without the overhead of dequeue
616  * followed by enqueue.
617  */
618 static void requeue_task(struct task_struct *p, prio_array_t *array)
619 {
620         list_move_tail(&p->run_list, array->queue + p->prio);
621 }
622
623 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
624 {
625         list_add(&p->run_list, array->queue + p->prio);
626         __set_bit(p->prio, array->bitmap);
627         array->nr_active++;
628         p->array = array;
629 }
630
631 /*
632  * effective_prio - return the priority that is based on the static
633  * priority but is modified by bonuses/penalties.
634  *
635  * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
636  * into the -5 ... 0 ... +5 bonus/penalty range.
637  *
638  * We use 25% of the full 0...39 priority range so that:
639  *
640  * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
641  * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
642  *
643  * Both properties are important to certain workloads.
644  */
645 static int effective_prio(task_t *p)
646 {
647         int bonus, prio;
648
649         if (rt_task(p))
650                 return p->prio;
651
652         bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
653
654         prio = p->static_prio - bonus;
655         if (prio < MAX_RT_PRIO)
656                 prio = MAX_RT_PRIO;
657         if (prio > MAX_PRIO-1)
658                 prio = MAX_PRIO-1;
659         return prio;
660 }
661
662 /*
663  * __activate_task - move a task to the runqueue.
664  */
665 static inline void __activate_task(task_t *p, runqueue_t *rq)
666 {
667         enqueue_task(p, rq->active);
668         rq->nr_running++;
669 }
670
671 /*
672  * __activate_idle_task - move idle task to the _front_ of runqueue.
673  */
674 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
675 {
676         enqueue_task_head(p, rq->active);
677         rq->nr_running++;
678 }
679
680 static int recalc_task_prio(task_t *p, unsigned long long now)
681 {
682         /* Caller must always ensure 'now >= p->timestamp' */
683         unsigned long long __sleep_time = now - p->timestamp;
684         unsigned long sleep_time;
685
686         if (__sleep_time > NS_MAX_SLEEP_AVG)
687                 sleep_time = NS_MAX_SLEEP_AVG;
688         else
689                 sleep_time = (unsigned long)__sleep_time;
690
691         if (likely(sleep_time > 0)) {
692                 /*
693                  * User tasks that sleep a long time are categorised as
694                  * idle and will get just interactive status to stay active &
695                  * prevent them suddenly becoming cpu hogs and starving
696                  * other processes.
697                  */
698                 if (p->mm && p->activated != -1 &&
699                         sleep_time > INTERACTIVE_SLEEP(p)) {
700                                 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
701                                                 DEF_TIMESLICE);
702                 } else {
703                         /*
704                          * The lower the sleep avg a task has the more
705                          * rapidly it will rise with sleep time.
706                          */
707                         sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
708
709                         /*
710                          * Tasks waking from uninterruptible sleep are
711                          * limited in their sleep_avg rise as they
712                          * are likely to be waiting on I/O
713                          */
714                         if (p->activated == -1 && p->mm) {
715                                 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
716                                         sleep_time = 0;
717                                 else if (p->sleep_avg + sleep_time >=
718                                                 INTERACTIVE_SLEEP(p)) {
719                                         p->sleep_avg = INTERACTIVE_SLEEP(p);
720                                         sleep_time = 0;
721                                 }
722                         }
723
724                         /*
725                          * This code gives a bonus to interactive tasks.
726                          *
727                          * The boost works by updating the 'average sleep time'
728                          * value here, based on ->timestamp. The more time a
729                          * task spends sleeping, the higher the average gets -
730                          * and the higher the priority boost gets as well.
731                          */
732                         p->sleep_avg += sleep_time;
733
734                         if (p->sleep_avg > NS_MAX_SLEEP_AVG)
735                                 p->sleep_avg = NS_MAX_SLEEP_AVG;
736                 }
737         }
738
739         return effective_prio(p);
740 }
741
742 /*
743  * activate_task - move a task to the runqueue and do priority recalculation
744  *
745  * Update all the scheduling statistics stuff. (sleep average
746  * calculation, priority modifiers, etc.)
747  */
748 static void activate_task(task_t *p, runqueue_t *rq, int local)
749 {
750         unsigned long long now;
751
752         now = sched_clock();
753 #ifdef CONFIG_SMP
754         if (!local) {
755                 /* Compensate for drifting sched_clock */
756                 runqueue_t *this_rq = this_rq();
757                 now = (now - this_rq->timestamp_last_tick)
758                         + rq->timestamp_last_tick;
759         }
760 #endif
761
762         p->prio = recalc_task_prio(p, now);
763
764         /*
765          * This checks to make sure it's not an uninterruptible task
766          * that is now waking up.
767          */
768         if (!p->activated) {
769                 /*
770                  * Tasks which were woken up by interrupts (ie. hw events)
771                  * are most likely of interactive nature. So we give them
772                  * the credit of extending their sleep time to the period
773                  * of time they spend on the runqueue, waiting for execution
774                  * on a CPU, first time around:
775                  */
776                 if (in_interrupt())
777                         p->activated = 2;
778                 else {
779                         /*
780                          * Normal first-time wakeups get a credit too for
781                          * on-runqueue time, but it will be weighted down:
782                          */
783                         p->activated = 1;
784                 }
785         }
786         p->timestamp = now;
787
788         __activate_task(p, rq);
789 }
790
791 /*
792  * deactivate_task - remove a task from the runqueue.
793  */
794 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
795 {
796         rq->nr_running--;
797         dequeue_task(p, p->array);
798         p->array = NULL;
799 }
800
801 /*
802  * resched_task - mark a task 'to be rescheduled now'.
803  *
804  * On UP this means the setting of the need_resched flag, on SMP it
805  * might also involve a cross-CPU call to trigger the scheduler on
806  * the target CPU.
807  */
808 #ifdef CONFIG_SMP
809 static void resched_task(task_t *p)
810 {
811         int need_resched, nrpolling;
812
813         assert_spin_locked(&task_rq(p)->lock);
814
815         /* minimise the chance of sending an interrupt to poll_idle() */
816         nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
817         need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
818         nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
819
820         if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
821                 smp_send_reschedule(task_cpu(p));
822 }
823 #else
824 static inline void resched_task(task_t *p)
825 {
826         set_tsk_need_resched(p);
827 }
828 #endif
829
830 /**
831  * task_curr - is this task currently executing on a CPU?
832  * @p: the task in question.
833  */
834 inline int task_curr(const task_t *p)
835 {
836         return cpu_curr(task_cpu(p)) == p;
837 }
838
839 #ifdef CONFIG_SMP
840 typedef struct {
841         struct list_head list;
842
843         task_t *task;
844         int dest_cpu;
845
846         struct completion done;
847 } migration_req_t;
848
849 /*
850  * The task's runqueue lock must be held.
851  * Returns true if you have to wait for migration thread.
852  */
853 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
854 {
855         runqueue_t *rq = task_rq(p);
856
857         /*
858          * If the task is not on a runqueue (and not running), then
859          * it is sufficient to simply update the task's cpu field.
860          */
861         if (!p->array && !task_running(rq, p)) {
862                 set_task_cpu(p, dest_cpu);
863                 return 0;
864         }
865
866         init_completion(&req->done);
867         req->task = p;
868         req->dest_cpu = dest_cpu;
869         list_add(&req->list, &rq->migration_queue);
870         return 1;
871 }
872
873 /*
874  * wait_task_inactive - wait for a thread to unschedule.
875  *
876  * The caller must ensure that the task *will* unschedule sometime soon,
877  * else this function might spin for a *long* time. This function can't
878  * be called with interrupts off, or it may introduce deadlock with
879  * smp_call_function() if an IPI is sent by the same process we are
880  * waiting to become inactive.
881  */
882 void wait_task_inactive(task_t *p)
883 {
884         unsigned long flags;
885         runqueue_t *rq;
886         int preempted;
887
888 repeat:
889         rq = task_rq_lock(p, &flags);
890         /* Must be off runqueue entirely, not preempted. */
891         if (unlikely(p->array || task_running(rq, p))) {
892                 /* If it's preempted, we yield.  It could be a while. */
893                 preempted = !task_running(rq, p);
894                 task_rq_unlock(rq, &flags);
895                 cpu_relax();
896                 if (preempted)
897                         yield();
898                 goto repeat;
899         }
900         task_rq_unlock(rq, &flags);
901 }
902
903 /***
904  * kick_process - kick a running thread to enter/exit the kernel
905  * @p: the to-be-kicked thread
906  *
907  * Cause a process which is running on another CPU to enter
908  * kernel-mode, without any delay. (to get signals handled.)
909  *
910  * NOTE: this function doesnt have to take the runqueue lock,
911  * because all it wants to ensure is that the remote task enters
912  * the kernel. If the IPI races and the task has been migrated
913  * to another CPU then no harm is done and the purpose has been
914  * achieved as well.
915  */
916 void kick_process(task_t *p)
917 {
918         int cpu;
919
920         preempt_disable();
921         cpu = task_cpu(p);
922         if ((cpu != smp_processor_id()) && task_curr(p))
923                 smp_send_reschedule(cpu);
924         preempt_enable();
925 }
926
927 /*
928  * Return a low guess at the load of a migration-source cpu.
929  *
930  * We want to under-estimate the load of migration sources, to
931  * balance conservatively.
932  */
933 static inline unsigned long source_load(int cpu, int type)
934 {
935         runqueue_t *rq = cpu_rq(cpu);
936         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
937         if (type == 0)
938                 return load_now;
939
940         return min(rq->cpu_load[type-1], load_now);
941 }
942
943 /*
944  * Return a high guess at the load of a migration-target cpu
945  */
946 static inline unsigned long target_load(int cpu, int type)
947 {
948         runqueue_t *rq = cpu_rq(cpu);
949         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
950         if (type == 0)
951                 return load_now;
952
953         return max(rq->cpu_load[type-1], load_now);
954 }
955
956 /*
957  * find_idlest_group finds and returns the least busy CPU group within the
958  * domain.
959  */
960 static struct sched_group *
961 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
962 {
963         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
964         unsigned long min_load = ULONG_MAX, this_load = 0;
965         int load_idx = sd->forkexec_idx;
966         int imbalance = 100 + (sd->imbalance_pct-100)/2;
967
968         do {
969                 unsigned long load, avg_load;
970                 int local_group;
971                 int i;
972
973                 /* Skip over this group if it has no CPUs allowed */
974                 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
975                         goto nextgroup;
976
977                 local_group = cpu_isset(this_cpu, group->cpumask);
978
979                 /* Tally up the load of all CPUs in the group */
980                 avg_load = 0;
981
982                 for_each_cpu_mask(i, group->cpumask) {
983                         /* Bias balancing toward cpus of our domain */
984                         if (local_group)
985                                 load = source_load(i, load_idx);
986                         else
987                                 load = target_load(i, load_idx);
988
989                         avg_load += load;
990                 }
991
992                 /* Adjust by relative CPU power of the group */
993                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
994
995                 if (local_group) {
996                         this_load = avg_load;
997                         this = group;
998                 } else if (avg_load < min_load) {
999                         min_load = avg_load;
1000                         idlest = group;
1001                 }
1002 nextgroup:
1003                 group = group->next;
1004         } while (group != sd->groups);
1005
1006         if (!idlest || 100*this_load < imbalance*min_load)
1007                 return NULL;
1008         return idlest;
1009 }
1010
1011 /*
1012  * find_idlest_queue - find the idlest runqueue among the cpus in group.
1013  */
1014 static int
1015 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1016 {
1017         cpumask_t tmp;
1018         unsigned long load, min_load = ULONG_MAX;
1019         int idlest = -1;
1020         int i;
1021
1022         /* Traverse only the allowed CPUs */
1023         cpus_and(tmp, group->cpumask, p->cpus_allowed);
1024
1025         for_each_cpu_mask(i, tmp) {
1026                 load = source_load(i, 0);
1027
1028                 if (load < min_load || (load == min_load && i == this_cpu)) {
1029                         min_load = load;
1030                         idlest = i;
1031                 }
1032         }
1033
1034         return idlest;
1035 }
1036
1037 /*
1038  * sched_balance_self: balance the current task (running on cpu) in domains
1039  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1040  * SD_BALANCE_EXEC.
1041  *
1042  * Balance, ie. select the least loaded group.
1043  *
1044  * Returns the target CPU number, or the same CPU if no balancing is needed.
1045  *
1046  * preempt must be disabled.
1047  */
1048 static int sched_balance_self(int cpu, int flag)
1049 {
1050         struct task_struct *t = current;
1051         struct sched_domain *tmp, *sd = NULL;
1052
1053         for_each_domain(cpu, tmp)
1054                 if (tmp->flags & flag)
1055                         sd = tmp;
1056
1057         while (sd) {
1058                 cpumask_t span;
1059                 struct sched_group *group;
1060                 int new_cpu;
1061                 int weight;
1062
1063                 span = sd->span;
1064                 group = find_idlest_group(sd, t, cpu);
1065                 if (!group)
1066                         goto nextlevel;
1067
1068                 new_cpu = find_idlest_cpu(group, t, cpu);
1069                 if (new_cpu == -1 || new_cpu == cpu)
1070                         goto nextlevel;
1071
1072                 /* Now try balancing at a lower domain level */
1073                 cpu = new_cpu;
1074 nextlevel:
1075                 sd = NULL;
1076                 weight = cpus_weight(span);
1077                 for_each_domain(cpu, tmp) {
1078                         if (weight <= cpus_weight(tmp->span))
1079                                 break;
1080                         if (tmp->flags & flag)
1081                                 sd = tmp;
1082                 }
1083                 /* while loop will break here if sd == NULL */
1084         }
1085
1086         return cpu;
1087 }
1088
1089 #endif /* CONFIG_SMP */
1090
1091 /*
1092  * wake_idle() will wake a task on an idle cpu if task->cpu is
1093  * not idle and an idle cpu is available.  The span of cpus to
1094  * search starts with cpus closest then further out as needed,
1095  * so we always favor a closer, idle cpu.
1096  *
1097  * Returns the CPU we should wake onto.
1098  */
1099 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1100 static int wake_idle(int cpu, task_t *p)
1101 {
1102         cpumask_t tmp;
1103         struct sched_domain *sd;
1104         int i;
1105
1106         if (idle_cpu(cpu))
1107                 return cpu;
1108
1109         for_each_domain(cpu, sd) {
1110                 if (sd->flags & SD_WAKE_IDLE) {
1111                         cpus_and(tmp, sd->span, p->cpus_allowed);
1112                         for_each_cpu_mask(i, tmp) {
1113                                 if (idle_cpu(i))
1114                                         return i;
1115                         }
1116                 }
1117                 else
1118                         break;
1119         }
1120         return cpu;
1121 }
1122 #else
1123 static inline int wake_idle(int cpu, task_t *p)
1124 {
1125         return cpu;
1126 }
1127 #endif
1128
1129 /***
1130  * try_to_wake_up - wake up a thread
1131  * @p: the to-be-woken-up thread
1132  * @state: the mask of task states that can be woken
1133  * @sync: do a synchronous wakeup?
1134  *
1135  * Put it on the run-queue if it's not already there. The "current"
1136  * thread is always on the run-queue (except when the actual
1137  * re-schedule is in progress), and as such you're allowed to do
1138  * the simpler "current->state = TASK_RUNNING" to mark yourself
1139  * runnable without the overhead of this.
1140  *
1141  * returns failure only if the task is already active.
1142  */
1143 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1144 {
1145         int cpu, this_cpu, success = 0;
1146         unsigned long flags;
1147         long old_state;
1148         runqueue_t *rq;
1149 #ifdef CONFIG_SMP
1150         unsigned long load, this_load;
1151         struct sched_domain *sd, *this_sd = NULL;
1152         int new_cpu;
1153 #endif
1154
1155         rq = task_rq_lock(p, &flags);
1156         old_state = p->state;
1157         if (!(old_state & state))
1158                 goto out;
1159
1160         if (p->array)
1161                 goto out_running;
1162
1163         cpu = task_cpu(p);
1164         this_cpu = smp_processor_id();
1165
1166 #ifdef CONFIG_SMP
1167         if (unlikely(task_running(rq, p)))
1168                 goto out_activate;
1169
1170         new_cpu = cpu;
1171
1172         schedstat_inc(rq, ttwu_cnt);
1173         if (cpu == this_cpu) {
1174                 schedstat_inc(rq, ttwu_local);
1175                 goto out_set_cpu;
1176         }
1177
1178         for_each_domain(this_cpu, sd) {
1179                 if (cpu_isset(cpu, sd->span)) {
1180                         schedstat_inc(sd, ttwu_wake_remote);
1181                         this_sd = sd;
1182                         break;
1183                 }
1184         }
1185
1186         if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1187                 goto out_set_cpu;
1188
1189         /*
1190          * Check for affine wakeup and passive balancing possibilities.
1191          */
1192         if (this_sd) {
1193                 int idx = this_sd->wake_idx;
1194                 unsigned int imbalance;
1195
1196                 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1197
1198                 load = source_load(cpu, idx);
1199                 this_load = target_load(this_cpu, idx);
1200
1201                 new_cpu = this_cpu; /* Wake to this CPU if we can */
1202
1203                 if (this_sd->flags & SD_WAKE_AFFINE) {
1204                         unsigned long tl = this_load;
1205                         /*
1206                          * If sync wakeup then subtract the (maximum possible)
1207                          * effect of the currently running task from the load
1208                          * of the current CPU:
1209                          */
1210                         if (sync)
1211                                 tl -= SCHED_LOAD_SCALE;
1212
1213                         if ((tl <= load &&
1214                                 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1215                                 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1216                                 /*
1217                                  * This domain has SD_WAKE_AFFINE and
1218                                  * p is cache cold in this domain, and
1219                                  * there is no bad imbalance.
1220                                  */
1221                                 schedstat_inc(this_sd, ttwu_move_affine);
1222                                 goto out_set_cpu;
1223                         }
1224                 }
1225
1226                 /*
1227                  * Start passive balancing when half the imbalance_pct
1228                  * limit is reached.
1229                  */
1230                 if (this_sd->flags & SD_WAKE_BALANCE) {
1231                         if (imbalance*this_load <= 100*load) {
1232                                 schedstat_inc(this_sd, ttwu_move_balance);
1233                                 goto out_set_cpu;
1234                         }
1235                 }
1236         }
1237
1238         new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1239 out_set_cpu:
1240         new_cpu = wake_idle(new_cpu, p);
1241         if (new_cpu != cpu) {
1242                 set_task_cpu(p, new_cpu);
1243                 task_rq_unlock(rq, &flags);
1244                 /* might preempt at this point */
1245                 rq = task_rq_lock(p, &flags);
1246                 old_state = p->state;
1247                 if (!(old_state & state))
1248                         goto out;
1249                 if (p->array)
1250                         goto out_running;
1251
1252                 this_cpu = smp_processor_id();
1253                 cpu = task_cpu(p);
1254         }
1255
1256 out_activate:
1257 #endif /* CONFIG_SMP */
1258         if (old_state == TASK_UNINTERRUPTIBLE) {
1259                 rq->nr_uninterruptible--;
1260                 /*
1261                  * Tasks on involuntary sleep don't earn
1262                  * sleep_avg beyond just interactive state.
1263                  */
1264                 p->activated = -1;
1265         }
1266
1267         /*
1268          * Tasks that have marked their sleep as noninteractive get
1269          * woken up without updating their sleep average. (i.e. their
1270          * sleep is handled in a priority-neutral manner, no priority
1271          * boost and no penalty.)
1272          */
1273         if (old_state & TASK_NONINTERACTIVE)
1274                 __activate_task(p, rq);
1275         else
1276                 activate_task(p, rq, cpu == this_cpu);
1277         /*
1278          * Sync wakeups (i.e. those types of wakeups where the waker
1279          * has indicated that it will leave the CPU in short order)
1280          * don't trigger a preemption, if the woken up task will run on
1281          * this cpu. (in this case the 'I will reschedule' promise of
1282          * the waker guarantees that the freshly woken up task is going
1283          * to be considered on this CPU.)
1284          */
1285         if (!sync || cpu != this_cpu) {
1286                 if (TASK_PREEMPTS_CURR(p, rq))
1287                         resched_task(rq->curr);
1288         }
1289         success = 1;
1290
1291 out_running:
1292         p->state = TASK_RUNNING;
1293 out:
1294         task_rq_unlock(rq, &flags);
1295
1296         return success;
1297 }
1298
1299 int fastcall wake_up_process(task_t *p)
1300 {
1301         return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1302                                  TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1303 }
1304
1305 EXPORT_SYMBOL(wake_up_process);
1306
1307 int fastcall wake_up_state(task_t *p, unsigned int state)
1308 {
1309         return try_to_wake_up(p, state, 0);
1310 }
1311
1312 /*
1313  * Perform scheduler related setup for a newly forked process p.
1314  * p is forked by current.
1315  */
1316 void fastcall sched_fork(task_t *p, int clone_flags)
1317 {
1318         int cpu = get_cpu();
1319
1320 #ifdef CONFIG_SMP
1321         cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1322 #endif
1323         set_task_cpu(p, cpu);
1324
1325         /*
1326          * We mark the process as running here, but have not actually
1327          * inserted it onto the runqueue yet. This guarantees that
1328          * nobody will actually run it, and a signal or other external
1329          * event cannot wake it up and insert it on the runqueue either.
1330          */
1331         p->state = TASK_RUNNING;
1332         INIT_LIST_HEAD(&p->run_list);
1333         p->array = NULL;
1334 #ifdef CONFIG_SCHEDSTATS
1335         memset(&p->sched_info, 0, sizeof(p->sched_info));
1336 #endif
1337 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1338         p->oncpu = 0;
1339 #endif
1340 #ifdef CONFIG_PREEMPT
1341         /* Want to start with kernel preemption disabled. */
1342         p->thread_info->preempt_count = 1;
1343 #endif
1344         /*
1345          * Share the timeslice between parent and child, thus the
1346          * total amount of pending timeslices in the system doesn't change,
1347          * resulting in more scheduling fairness.
1348          */
1349         local_irq_disable();
1350         p->time_slice = (current->time_slice + 1) >> 1;
1351         /*
1352          * The remainder of the first timeslice might be recovered by
1353          * the parent if the child exits early enough.
1354          */
1355         p->first_time_slice = 1;
1356         current->time_slice >>= 1;
1357         p->timestamp = sched_clock();
1358         if (unlikely(!current->time_slice)) {
1359                 /*
1360                  * This case is rare, it happens when the parent has only
1361                  * a single jiffy left from its timeslice. Taking the
1362                  * runqueue lock is not a problem.
1363                  */
1364                 current->time_slice = 1;
1365                 scheduler_tick();
1366         }
1367         local_irq_enable();
1368         put_cpu();
1369 }
1370
1371 /*
1372  * wake_up_new_task - wake up a newly created task for the first time.
1373  *
1374  * This function will do some initial scheduler statistics housekeeping
1375  * that must be done for every newly created context, then puts the task
1376  * on the runqueue and wakes it.
1377  */
1378 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1379 {
1380         unsigned long flags;
1381         int this_cpu, cpu;
1382         runqueue_t *rq, *this_rq;
1383
1384         rq = task_rq_lock(p, &flags);
1385         BUG_ON(p->state != TASK_RUNNING);
1386         this_cpu = smp_processor_id();
1387         cpu = task_cpu(p);
1388
1389         /*
1390          * We decrease the sleep average of forking parents
1391          * and children as well, to keep max-interactive tasks
1392          * from forking tasks that are max-interactive. The parent
1393          * (current) is done further down, under its lock.
1394          */
1395         p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1396                 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1397
1398         p->prio = effective_prio(p);
1399
1400         if (likely(cpu == this_cpu)) {
1401                 if (!(clone_flags & CLONE_VM)) {
1402                         /*
1403                          * The VM isn't cloned, so we're in a good position to
1404                          * do child-runs-first in anticipation of an exec. This
1405                          * usually avoids a lot of COW overhead.
1406                          */
1407                         if (unlikely(!current->array))
1408                                 __activate_task(p, rq);
1409                         else {
1410                                 p->prio = current->prio;
1411                                 list_add_tail(&p->run_list, &current->run_list);
1412                                 p->array = current->array;
1413                                 p->array->nr_active++;
1414                                 rq->nr_running++;
1415                         }
1416                         set_need_resched();
1417                 } else
1418                         /* Run child last */
1419                         __activate_task(p, rq);
1420                 /*
1421                  * We skip the following code due to cpu == this_cpu
1422                  *
1423                  *   task_rq_unlock(rq, &flags);
1424                  *   this_rq = task_rq_lock(current, &flags);
1425                  */
1426                 this_rq = rq;
1427         } else {
1428                 this_rq = cpu_rq(this_cpu);
1429
1430                 /*
1431                  * Not the local CPU - must adjust timestamp. This should
1432                  * get optimised away in the !CONFIG_SMP case.
1433                  */
1434                 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1435                                         + rq->timestamp_last_tick;
1436                 __activate_task(p, rq);
1437                 if (TASK_PREEMPTS_CURR(p, rq))
1438                         resched_task(rq->curr);
1439
1440                 /*
1441                  * Parent and child are on different CPUs, now get the
1442                  * parent runqueue to update the parent's ->sleep_avg:
1443                  */
1444                 task_rq_unlock(rq, &flags);
1445                 this_rq = task_rq_lock(current, &flags);
1446         }
1447         current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1448                 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1449         task_rq_unlock(this_rq, &flags);
1450 }
1451
1452 /*
1453  * Potentially available exiting-child timeslices are
1454  * retrieved here - this way the parent does not get
1455  * penalized for creating too many threads.
1456  *
1457  * (this cannot be used to 'generate' timeslices
1458  * artificially, because any timeslice recovered here
1459  * was given away by the parent in the first place.)
1460  */
1461 void fastcall sched_exit(task_t *p)
1462 {
1463         unsigned long flags;
1464         runqueue_t *rq;
1465
1466         /*
1467          * If the child was a (relative-) CPU hog then decrease
1468          * the sleep_avg of the parent as well.
1469          */
1470         rq = task_rq_lock(p->parent, &flags);
1471         if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1472                 p->parent->time_slice += p->time_slice;
1473                 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1474                         p->parent->time_slice = task_timeslice(p);
1475         }
1476         if (p->sleep_avg < p->parent->sleep_avg)
1477                 p->parent->sleep_avg = p->parent->sleep_avg /
1478                 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1479                 (EXIT_WEIGHT + 1);
1480         task_rq_unlock(rq, &flags);
1481 }
1482
1483 /**
1484  * prepare_task_switch - prepare to switch tasks
1485  * @rq: the runqueue preparing to switch
1486  * @next: the task we are going to switch to.
1487  *
1488  * This is called with the rq lock held and interrupts off. It must
1489  * be paired with a subsequent finish_task_switch after the context
1490  * switch.
1491  *
1492  * prepare_task_switch sets up locking and calls architecture specific
1493  * hooks.
1494  */
1495 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1496 {
1497         prepare_lock_switch(rq, next);
1498         prepare_arch_switch(next);
1499 }
1500
1501 /**
1502  * finish_task_switch - clean up after a task-switch
1503  * @rq: runqueue associated with task-switch
1504  * @prev: the thread we just switched away from.
1505  *
1506  * finish_task_switch must be called after the context switch, paired
1507  * with a prepare_task_switch call before the context switch.
1508  * finish_task_switch will reconcile locking set up by prepare_task_switch,
1509  * and do any other architecture-specific cleanup actions.
1510  *
1511  * Note that we may have delayed dropping an mm in context_switch(). If
1512  * so, we finish that here outside of the runqueue lock.  (Doing it
1513  * with the lock held can cause deadlocks; see schedule() for
1514  * details.)
1515  */
1516 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1517         __releases(rq->lock)
1518 {
1519         struct mm_struct *mm = rq->prev_mm;
1520         unsigned long prev_task_flags;
1521
1522         rq->prev_mm = NULL;
1523
1524         /*
1525          * A task struct has one reference for the use as "current".
1526          * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1527          * calls schedule one last time. The schedule call will never return,
1528          * and the scheduled task must drop that reference.
1529          * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1530          * still held, otherwise prev could be scheduled on another cpu, die
1531          * there before we look at prev->state, and then the reference would
1532          * be dropped twice.
1533          *              Manfred Spraul <manfred@colorfullife.com>
1534          */
1535         prev_task_flags = prev->flags;
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, int *sd_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         unsigned long max_pull;
1914         int load_idx;
1915
1916         max_load = this_load = total_load = total_pwr = 0;
1917         if (idle == NOT_IDLE)
1918                 load_idx = sd->busy_idx;
1919         else if (idle == NEWLY_IDLE)
1920                 load_idx = sd->newidle_idx;
1921         else
1922                 load_idx = sd->idle_idx;
1923
1924         do {
1925                 unsigned long load;
1926                 int local_group;
1927                 int i;
1928
1929                 local_group = cpu_isset(this_cpu, group->cpumask);
1930
1931                 /* Tally up the load of all CPUs in the group */
1932                 avg_load = 0;
1933
1934                 for_each_cpu_mask(i, group->cpumask) {
1935                         if (*sd_idle && !idle_cpu(i))
1936                                 *sd_idle = 0;
1937
1938                         /* Bias balancing toward cpus of our domain */
1939                         if (local_group)
1940                                 load = target_load(i, load_idx);
1941                         else
1942                                 load = source_load(i, load_idx);
1943
1944                         avg_load += load;
1945                 }
1946
1947                 total_load += avg_load;
1948                 total_pwr += group->cpu_power;
1949
1950                 /* Adjust by relative CPU power of the group */
1951                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1952
1953                 if (local_group) {
1954                         this_load = avg_load;
1955                         this = group;
1956                 } else if (avg_load > max_load) {
1957                         max_load = avg_load;
1958                         busiest = group;
1959                 }
1960                 group = group->next;
1961         } while (group != sd->groups);
1962
1963         if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
1964                 goto out_balanced;
1965
1966         avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1967
1968         if (this_load >= avg_load ||
1969                         100*max_load <= sd->imbalance_pct*this_load)
1970                 goto out_balanced;
1971
1972         /*
1973          * We're trying to get all the cpus to the average_load, so we don't
1974          * want to push ourselves above the average load, nor do we wish to
1975          * reduce the max loaded cpu below the average load, as either of these
1976          * actions would just result in more rebalancing later, and ping-pong
1977          * tasks around. Thus we look for the minimum possible imbalance.
1978          * Negative imbalances (*we* are more loaded than anyone else) will
1979          * be counted as no imbalance for these purposes -- we can't fix that
1980          * by pulling tasks to us.  Be careful of negative numbers as they'll
1981          * appear as very large values with unsigned longs.
1982          */
1983
1984         /* Don't want to pull so many tasks that a group would go idle */
1985         max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
1986
1987         /* How much load to actually move to equalise the imbalance */
1988         *imbalance = min(max_pull * busiest->cpu_power,
1989                                 (avg_load - this_load) * this->cpu_power)
1990                         / SCHED_LOAD_SCALE;
1991
1992         if (*imbalance < SCHED_LOAD_SCALE) {
1993                 unsigned long pwr_now = 0, pwr_move = 0;
1994                 unsigned long tmp;
1995
1996                 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1997                         *imbalance = 1;
1998                         return busiest;
1999                 }
2000
2001                 /*
2002                  * OK, we don't have enough imbalance to justify moving tasks,
2003                  * however we may be able to increase total CPU power used by
2004                  * moving them.
2005                  */
2006
2007                 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2008                 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2009                 pwr_now /= SCHED_LOAD_SCALE;
2010
2011                 /* Amount of load we'd subtract */
2012                 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2013                 if (max_load > tmp)
2014                         pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2015                                                         max_load - tmp);
2016
2017                 /* Amount of load we'd add */
2018                 if (max_load*busiest->cpu_power <
2019                                 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2020                         tmp = max_load*busiest->cpu_power/this->cpu_power;
2021                 else
2022                         tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2023                 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2024                 pwr_move /= SCHED_LOAD_SCALE;
2025
2026                 /* Move if we gain throughput */
2027                 if (pwr_move <= pwr_now)
2028                         goto out_balanced;
2029
2030                 *imbalance = 1;
2031                 return busiest;
2032         }
2033
2034         /* Get rid of the scaling factor, rounding down as we divide */
2035         *imbalance = *imbalance / SCHED_LOAD_SCALE;
2036         return busiest;
2037
2038 out_balanced:
2039
2040         *imbalance = 0;
2041         return NULL;
2042 }
2043
2044 /*
2045  * find_busiest_queue - find the busiest runqueue among the cpus in group.
2046  */
2047 static runqueue_t *find_busiest_queue(struct sched_group *group)
2048 {
2049         unsigned long load, max_load = 0;
2050         runqueue_t *busiest = NULL;
2051         int i;
2052
2053         for_each_cpu_mask(i, group->cpumask) {
2054                 load = source_load(i, 0);
2055
2056                 if (load > max_load) {
2057                         max_load = load;
2058                         busiest = cpu_rq(i);
2059                 }
2060         }
2061
2062         return busiest;
2063 }
2064
2065 /*
2066  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2067  * so long as it is large enough.
2068  */
2069 #define MAX_PINNED_INTERVAL     512
2070
2071 /*
2072  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2073  * tasks if there is an imbalance.
2074  *
2075  * Called with this_rq unlocked.
2076  */
2077 static int load_balance(int this_cpu, runqueue_t *this_rq,
2078                         struct sched_domain *sd, enum idle_type idle)
2079 {
2080         struct sched_group *group;
2081         runqueue_t *busiest;
2082         unsigned long imbalance;
2083         int nr_moved, all_pinned = 0;
2084         int active_balance = 0;
2085         int sd_idle = 0;
2086
2087         if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2088                 sd_idle = 1;
2089
2090         schedstat_inc(sd, lb_cnt[idle]);
2091
2092         group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2093         if (!group) {
2094                 schedstat_inc(sd, lb_nobusyg[idle]);
2095                 goto out_balanced;
2096         }
2097
2098         busiest = find_busiest_queue(group);
2099         if (!busiest) {
2100                 schedstat_inc(sd, lb_nobusyq[idle]);
2101                 goto out_balanced;
2102         }
2103
2104         BUG_ON(busiest == this_rq);
2105
2106         schedstat_add(sd, lb_imbalance[idle], imbalance);
2107
2108         nr_moved = 0;
2109         if (busiest->nr_running > 1) {
2110                 /*
2111                  * Attempt to move tasks. If find_busiest_group has found
2112                  * an imbalance but busiest->nr_running <= 1, the group is
2113                  * still unbalanced. nr_moved simply stays zero, so it is
2114                  * correctly treated as an imbalance.
2115                  */
2116                 double_rq_lock(this_rq, busiest);
2117                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2118                                         imbalance, sd, idle, &all_pinned);
2119                 double_rq_unlock(this_rq, busiest);
2120
2121                 /* All tasks on this runqueue were pinned by CPU affinity */
2122                 if (unlikely(all_pinned))
2123                         goto out_balanced;
2124         }
2125
2126         if (!nr_moved) {
2127                 schedstat_inc(sd, lb_failed[idle]);
2128                 sd->nr_balance_failed++;
2129
2130                 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2131
2132                         spin_lock(&busiest->lock);
2133
2134                         /* don't kick the migration_thread, if the curr
2135                          * task on busiest cpu can't be moved to this_cpu
2136                          */
2137                         if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2138                                 spin_unlock(&busiest->lock);
2139                                 all_pinned = 1;
2140                                 goto out_one_pinned;
2141                         }
2142
2143                         if (!busiest->active_balance) {
2144                                 busiest->active_balance = 1;
2145                                 busiest->push_cpu = this_cpu;
2146                                 active_balance = 1;
2147                         }
2148                         spin_unlock(&busiest->lock);
2149                         if (active_balance)
2150                                 wake_up_process(busiest->migration_thread);
2151
2152                         /*
2153                          * We've kicked active balancing, reset the failure
2154                          * counter.
2155                          */
2156                         sd->nr_balance_failed = sd->cache_nice_tries+1;
2157                 }
2158         } else
2159                 sd->nr_balance_failed = 0;
2160
2161         if (likely(!active_balance)) {
2162                 /* We were unbalanced, so reset the balancing interval */
2163                 sd->balance_interval = sd->min_interval;
2164         } else {
2165                 /*
2166                  * If we've begun active balancing, start to back off. This
2167                  * case may not be covered by the all_pinned logic if there
2168                  * is only 1 task on the busy runqueue (because we don't call
2169                  * move_tasks).
2170                  */
2171                 if (sd->balance_interval < sd->max_interval)
2172                         sd->balance_interval *= 2;
2173         }
2174
2175         if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2176                 return -1;
2177         return nr_moved;
2178
2179 out_balanced:
2180         schedstat_inc(sd, lb_balanced[idle]);
2181
2182         sd->nr_balance_failed = 0;
2183
2184 out_one_pinned:
2185         /* tune up the balancing interval */
2186         if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2187                         (sd->balance_interval < sd->max_interval))
2188                 sd->balance_interval *= 2;
2189
2190         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2191                 return -1;
2192         return 0;
2193 }
2194
2195 /*
2196  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2197  * tasks if there is an imbalance.
2198  *
2199  * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2200  * this_rq is locked.
2201  */
2202 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2203                                 struct sched_domain *sd)
2204 {
2205         struct sched_group *group;
2206         runqueue_t *busiest = NULL;
2207         unsigned long imbalance;
2208         int nr_moved = 0;
2209         int sd_idle = 0;
2210
2211         if (sd->flags & SD_SHARE_CPUPOWER)
2212                 sd_idle = 1;
2213
2214         schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2215         group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2216         if (!group) {
2217                 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2218                 goto out_balanced;
2219         }
2220
2221         busiest = find_busiest_queue(group);
2222         if (!busiest) {
2223                 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2224                 goto out_balanced;
2225         }
2226
2227         BUG_ON(busiest == this_rq);
2228
2229         schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2230
2231         nr_moved = 0;
2232         if (busiest->nr_running > 1) {
2233                 /* Attempt to move tasks */
2234                 double_lock_balance(this_rq, busiest);
2235                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2236                                         imbalance, sd, NEWLY_IDLE, NULL);
2237                 spin_unlock(&busiest->lock);
2238         }
2239
2240         if (!nr_moved) {
2241                 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2242                 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2243                         return -1;
2244         } else
2245                 sd->nr_balance_failed = 0;
2246
2247         return nr_moved;
2248
2249 out_balanced:
2250         schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2251         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2252                 return -1;
2253         sd->nr_balance_failed = 0;
2254         return 0;
2255 }
2256
2257 /*
2258  * idle_balance is called by schedule() if this_cpu is about to become
2259  * idle. Attempts to pull tasks from other CPUs.
2260  */
2261 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2262 {
2263         struct sched_domain *sd;
2264
2265         for_each_domain(this_cpu, sd) {
2266                 if (sd->flags & SD_BALANCE_NEWIDLE) {
2267                         if (load_balance_newidle(this_cpu, this_rq, sd)) {
2268                                 /* We've pulled tasks over so stop searching */
2269                                 break;
2270                         }
2271                 }
2272         }
2273 }
2274
2275 /*
2276  * active_load_balance is run by migration threads. It pushes running tasks
2277  * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2278  * running on each physical CPU where possible, and avoids physical /
2279  * logical imbalances.
2280  *
2281  * Called with busiest_rq locked.
2282  */
2283 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2284 {
2285         struct sched_domain *sd;
2286         runqueue_t *target_rq;
2287         int target_cpu = busiest_rq->push_cpu;
2288
2289         if (busiest_rq->nr_running <= 1)
2290                 /* no task to move */
2291                 return;
2292
2293         target_rq = cpu_rq(target_cpu);
2294
2295         /*
2296          * This condition is "impossible", if it occurs
2297          * we need to fix it.  Originally reported by
2298          * Bjorn Helgaas on a 128-cpu setup.
2299          */
2300         BUG_ON(busiest_rq == target_rq);
2301
2302         /* move a task from busiest_rq to target_rq */
2303         double_lock_balance(busiest_rq, target_rq);
2304
2305         /* Search for an sd spanning us and the target CPU. */
2306         for_each_domain(target_cpu, sd)
2307                 if ((sd->flags & SD_LOAD_BALANCE) &&
2308                         cpu_isset(busiest_cpu, sd->span))
2309                                 break;
2310
2311         if (unlikely(sd == NULL))
2312                 goto out;
2313
2314         schedstat_inc(sd, alb_cnt);
2315
2316         if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2317                 schedstat_inc(sd, alb_pushed);
2318         else
2319                 schedstat_inc(sd, alb_failed);
2320 out:
2321         spin_unlock(&target_rq->lock);
2322 }
2323
2324 /*
2325  * rebalance_tick will get called every timer tick, on every CPU.
2326  *
2327  * It checks each scheduling domain to see if it is due to be balanced,
2328  * and initiates a balancing operation if so.
2329  *
2330  * Balancing parameters are set up in arch_init_sched_domains.
2331  */
2332
2333 /* Don't have all balancing operations going off at once */
2334 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2335
2336 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2337                            enum idle_type idle)
2338 {
2339         unsigned long old_load, this_load;
2340         unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2341         struct sched_domain *sd;
2342         int i;
2343
2344         this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2345         /* Update our load */
2346         for (i = 0; i < 3; i++) {
2347                 unsigned long new_load = this_load;
2348                 int scale = 1 << i;
2349                 old_load = this_rq->cpu_load[i];
2350                 /*
2351                  * Round up the averaging division if load is increasing. This
2352                  * prevents us from getting stuck on 9 if the load is 10, for
2353                  * example.
2354                  */
2355                 if (new_load > old_load)
2356                         new_load += scale-1;
2357                 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2358         }
2359
2360         for_each_domain(this_cpu, sd) {
2361                 unsigned long interval;
2362
2363                 if (!(sd->flags & SD_LOAD_BALANCE))
2364                         continue;
2365
2366                 interval = sd->balance_interval;
2367                 if (idle != SCHED_IDLE)
2368                         interval *= sd->busy_factor;
2369
2370                 /* scale ms to jiffies */
2371                 interval = msecs_to_jiffies(interval);
2372                 if (unlikely(!interval))
2373                         interval = 1;
2374
2375                 if (j - sd->last_balance >= interval) {
2376                         if (load_balance(this_cpu, this_rq, sd, idle)) {
2377                                 /*
2378                                  * We've pulled tasks over so either we're no
2379                                  * longer idle, or one of our SMT siblings is
2380                                  * not idle.
2381                                  */
2382                                 idle = NOT_IDLE;
2383                         }
2384                         sd->last_balance += interval;
2385                 }
2386         }
2387 }
2388 #else
2389 /*
2390  * on UP we do not need to balance between CPUs:
2391  */
2392 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2393 {
2394 }
2395 static inline void idle_balance(int cpu, runqueue_t *rq)
2396 {
2397 }
2398 #endif
2399
2400 static inline int wake_priority_sleeper(runqueue_t *rq)
2401 {
2402         int ret = 0;
2403 #ifdef CONFIG_SCHED_SMT
2404         spin_lock(&rq->lock);
2405         /*
2406          * If an SMT sibling task has been put to sleep for priority
2407          * reasons reschedule the idle task to see if it can now run.
2408          */
2409         if (rq->nr_running) {
2410                 resched_task(rq->idle);
2411                 ret = 1;
2412         }
2413         spin_unlock(&rq->lock);
2414 #endif
2415         return ret;
2416 }
2417
2418 DEFINE_PER_CPU(struct kernel_stat, kstat);
2419
2420 EXPORT_PER_CPU_SYMBOL(kstat);
2421
2422 /*
2423  * This is called on clock ticks and on context switches.
2424  * Bank in p->sched_time the ns elapsed since the last tick or switch.
2425  */
2426 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2427                                     unsigned long long now)
2428 {
2429         unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2430         p->sched_time += now - last;
2431 }
2432
2433 /*
2434  * Return current->sched_time plus any more ns on the sched_clock
2435  * that have not yet been banked.
2436  */
2437 unsigned long long current_sched_time(const task_t *tsk)
2438 {
2439         unsigned long long ns;
2440         unsigned long flags;
2441         local_irq_save(flags);
2442         ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2443         ns = tsk->sched_time + (sched_clock() - ns);
2444         local_irq_restore(flags);
2445         return ns;
2446 }
2447
2448 /*
2449  * We place interactive tasks back into the active array, if possible.
2450  *
2451  * To guarantee that this does not starve expired tasks we ignore the
2452  * interactivity of a task if the first expired task had to wait more
2453  * than a 'reasonable' amount of time. This deadline timeout is
2454  * load-dependent, as the frequency of array switched decreases with
2455  * increasing number of running tasks. We also ignore the interactivity
2456  * if a better static_prio task has expired:
2457  */
2458 #define EXPIRED_STARVING(rq) \
2459         ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2460                 (jiffies - (rq)->expired_timestamp >= \
2461                         STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2462                         ((rq)->curr->static_prio > (rq)->best_expired_prio))
2463
2464 /*
2465  * Account user cpu time to a process.
2466  * @p: the process that the cpu time gets accounted to
2467  * @hardirq_offset: the offset to subtract from hardirq_count()
2468  * @cputime: the cpu time spent in user space since the last update
2469  */
2470 void account_user_time(struct task_struct *p, cputime_t cputime)
2471 {
2472         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2473         cputime64_t tmp;
2474
2475         p->utime = cputime_add(p->utime, cputime);
2476
2477         /* Add user time to cpustat. */
2478         tmp = cputime_to_cputime64(cputime);
2479         if (TASK_NICE(p) > 0)
2480                 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2481         else
2482                 cpustat->user = cputime64_add(cpustat->user, tmp);
2483 }
2484
2485 /*
2486  * Account system cpu time to a process.
2487  * @p: the process that the cpu time gets accounted to
2488  * @hardirq_offset: the offset to subtract from hardirq_count()
2489  * @cputime: the cpu time spent in kernel space since the last update
2490  */
2491 void account_system_time(struct task_struct *p, int hardirq_offset,
2492                          cputime_t cputime)
2493 {
2494         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2495         runqueue_t *rq = this_rq();
2496         cputime64_t tmp;
2497
2498         p->stime = cputime_add(p->stime, cputime);
2499
2500         /* Add system time to cpustat. */
2501         tmp = cputime_to_cputime64(cputime);
2502         if (hardirq_count() - hardirq_offset)
2503                 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2504         else if (softirq_count())
2505                 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2506         else if (p != rq->idle)
2507                 cpustat->system = cputime64_add(cpustat->system, tmp);
2508         else if (atomic_read(&rq->nr_iowait) > 0)
2509                 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2510         else
2511                 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2512         /* Account for system time used */
2513         acct_update_integrals(p);
2514 }
2515
2516 /*
2517  * Account for involuntary wait time.
2518  * @p: the process from which the cpu time has been stolen
2519  * @steal: the cpu time spent in involuntary wait
2520  */
2521 void account_steal_time(struct task_struct *p, cputime_t steal)
2522 {
2523         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2524         cputime64_t tmp = cputime_to_cputime64(steal);
2525         runqueue_t *rq = this_rq();
2526
2527         if (p == rq->idle) {
2528                 p->stime = cputime_add(p->stime, steal);
2529                 if (atomic_read(&rq->nr_iowait) > 0)
2530                         cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2531                 else
2532                         cpustat->idle = cputime64_add(cpustat->idle, tmp);
2533         } else
2534                 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2535 }
2536
2537 /*
2538  * This function gets called by the timer code, with HZ frequency.
2539  * We call it with interrupts disabled.
2540  *
2541  * It also gets called by the fork code, when changing the parent's
2542  * timeslices.
2543  */
2544 void scheduler_tick(void)
2545 {
2546         int cpu = smp_processor_id();
2547         runqueue_t *rq = this_rq();
2548         task_t *p = current;
2549         unsigned long long now = sched_clock();
2550
2551         update_cpu_clock(p, rq, now);
2552
2553         rq->timestamp_last_tick = now;
2554
2555         if (p == rq->idle) {
2556                 if (wake_priority_sleeper(rq))
2557                         goto out;
2558                 rebalance_tick(cpu, rq, SCHED_IDLE);
2559                 return;
2560         }
2561
2562         /* Task might have expired already, but not scheduled off yet */
2563         if (p->array != rq->active) {
2564                 set_tsk_need_resched(p);
2565                 goto out;
2566         }
2567         spin_lock(&rq->lock);
2568         /*
2569          * The task was running during this tick - update the
2570          * time slice counter. Note: we do not update a thread's
2571          * priority until it either goes to sleep or uses up its
2572          * timeslice. This makes it possible for interactive tasks
2573          * to use up their timeslices at their highest priority levels.
2574          */
2575         if (rt_task(p)) {
2576                 /*
2577                  * RR tasks need a special form of timeslice management.
2578                  * FIFO tasks have no timeslices.
2579                  */
2580                 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2581                         p->time_slice = task_timeslice(p);
2582                         p->first_time_slice = 0;
2583                         set_tsk_need_resched(p);
2584
2585                         /* put it at the end of the queue: */
2586                         requeue_task(p, rq->active);
2587                 }
2588                 goto out_unlock;
2589         }
2590         if (!--p->time_slice) {
2591                 dequeue_task(p, rq->active);
2592                 set_tsk_need_resched(p);
2593                 p->prio = effective_prio(p);
2594                 p->time_slice = task_timeslice(p);
2595                 p->first_time_slice = 0;
2596
2597                 if (!rq->expired_timestamp)
2598                         rq->expired_timestamp = jiffies;
2599                 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2600                         enqueue_task(p, rq->expired);
2601                         if (p->static_prio < rq->best_expired_prio)
2602                                 rq->best_expired_prio = p->static_prio;
2603                 } else
2604                         enqueue_task(p, rq->active);
2605         } else {
2606                 /*
2607                  * Prevent a too long timeslice allowing a task to monopolize
2608                  * the CPU. We do this by splitting up the timeslice into
2609                  * smaller pieces.
2610                  *
2611                  * Note: this does not mean the task's timeslices expire or
2612                  * get lost in any way, they just might be preempted by
2613                  * another task of equal priority. (one with higher
2614                  * priority would have preempted this task already.) We
2615                  * requeue this task to the end of the list on this priority
2616                  * level, which is in essence a round-robin of tasks with
2617                  * equal priority.
2618                  *
2619                  * This only applies to tasks in the interactive
2620                  * delta range with at least TIMESLICE_GRANULARITY to requeue.
2621                  */
2622                 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2623                         p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2624                         (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2625                         (p->array == rq->active)) {
2626
2627                         requeue_task(p, rq->active);
2628                         set_tsk_need_resched(p);
2629                 }
2630         }
2631 out_unlock:
2632         spin_unlock(&rq->lock);
2633 out:
2634         rebalance_tick(cpu, rq, NOT_IDLE);
2635 }
2636
2637 #ifdef CONFIG_SCHED_SMT
2638 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2639 {
2640         /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2641         if (rq->curr == rq->idle && rq->nr_running)
2642                 resched_task(rq->idle);
2643 }
2644
2645 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2646 {
2647         struct sched_domain *tmp, *sd = NULL;
2648         cpumask_t sibling_map;
2649         int i;
2650
2651         for_each_domain(this_cpu, tmp)
2652                 if (tmp->flags & SD_SHARE_CPUPOWER)
2653                         sd = tmp;
2654
2655         if (!sd)
2656                 return;
2657
2658         /*
2659          * Unlock the current runqueue because we have to lock in
2660          * CPU order to avoid deadlocks. Caller knows that we might
2661          * unlock. We keep IRQs disabled.
2662          */
2663         spin_unlock(&this_rq->lock);
2664
2665         sibling_map = sd->span;
2666
2667         for_each_cpu_mask(i, sibling_map)
2668                 spin_lock(&cpu_rq(i)->lock);
2669         /*
2670          * We clear this CPU from the mask. This both simplifies the
2671          * inner loop and keps this_rq locked when we exit:
2672          */
2673         cpu_clear(this_cpu, sibling_map);
2674
2675         for_each_cpu_mask(i, sibling_map) {
2676                 runqueue_t *smt_rq = cpu_rq(i);
2677
2678                 wakeup_busy_runqueue(smt_rq);
2679         }
2680
2681         for_each_cpu_mask(i, sibling_map)
2682                 spin_unlock(&cpu_rq(i)->lock);
2683         /*
2684          * We exit with this_cpu's rq still held and IRQs
2685          * still disabled:
2686          */
2687 }
2688
2689 /*
2690  * number of 'lost' timeslices this task wont be able to fully
2691  * utilize, if another task runs on a sibling. This models the
2692  * slowdown effect of other tasks running on siblings:
2693  */
2694 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2695 {
2696         return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2697 }
2698
2699 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2700 {
2701         struct sched_domain *tmp, *sd = NULL;
2702         cpumask_t sibling_map;
2703         prio_array_t *array;
2704         int ret = 0, i;
2705         task_t *p;
2706
2707         for_each_domain(this_cpu, tmp)
2708                 if (tmp->flags & SD_SHARE_CPUPOWER)
2709                         sd = tmp;
2710
2711         if (!sd)
2712                 return 0;
2713
2714         /*
2715          * The same locking rules and details apply as for
2716          * wake_sleeping_dependent():
2717          */
2718         spin_unlock(&this_rq->lock);
2719         sibling_map = sd->span;
2720         for_each_cpu_mask(i, sibling_map)
2721                 spin_lock(&cpu_rq(i)->lock);
2722         cpu_clear(this_cpu, sibling_map);
2723
2724         /*
2725          * Establish next task to be run - it might have gone away because
2726          * we released the runqueue lock above:
2727          */
2728         if (!this_rq->nr_running)
2729                 goto out_unlock;
2730         array = this_rq->active;
2731         if (!array->nr_active)
2732                 array = this_rq->expired;
2733         BUG_ON(!array->nr_active);
2734
2735         p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2736                 task_t, run_list);
2737
2738         for_each_cpu_mask(i, sibling_map) {
2739                 runqueue_t *smt_rq = cpu_rq(i);
2740                 task_t *smt_curr = smt_rq->curr;
2741
2742                 /* Kernel threads do not participate in dependent sleeping */
2743                 if (!p->mm || !smt_curr->mm || rt_task(p))
2744                         goto check_smt_task;
2745
2746                 /*
2747                  * If a user task with lower static priority than the
2748                  * running task on the SMT sibling is trying to schedule,
2749                  * delay it till there is proportionately less timeslice
2750                  * left of the sibling task to prevent a lower priority
2751                  * task from using an unfair proportion of the
2752                  * physical cpu's resources. -ck
2753                  */
2754                 if (rt_task(smt_curr)) {
2755                         /*
2756                          * With real time tasks we run non-rt tasks only
2757                          * per_cpu_gain% of the time.
2758                          */
2759                         if ((jiffies % DEF_TIMESLICE) >
2760                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2761                                         ret = 1;
2762                 } else
2763                         if (smt_curr->static_prio < p->static_prio &&
2764                                 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2765                                 smt_slice(smt_curr, sd) > task_timeslice(p))
2766                                         ret = 1;
2767
2768 check_smt_task:
2769                 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2770                         rt_task(smt_curr))
2771                                 continue;
2772                 if (!p->mm) {
2773                         wakeup_busy_runqueue(smt_rq);
2774                         continue;
2775                 }
2776
2777                 /*
2778                  * Reschedule a lower priority task on the SMT sibling for
2779                  * it to be put to sleep, or wake it up if it has been put to
2780                  * sleep for priority reasons to see if it should run now.
2781                  */
2782                 if (rt_task(p)) {
2783                         if ((jiffies % DEF_TIMESLICE) >
2784                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2785                                         resched_task(smt_curr);
2786                 } else {
2787                         if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2788                                 smt_slice(p, sd) > task_timeslice(smt_curr))
2789                                         resched_task(smt_curr);
2790                         else
2791                                 wakeup_busy_runqueue(smt_rq);
2792                 }
2793         }
2794 out_unlock:
2795         for_each_cpu_mask(i, sibling_map)
2796                 spin_unlock(&cpu_rq(i)->lock);
2797         return ret;
2798 }
2799 #else
2800 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2801 {
2802 }
2803
2804 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2805 {
2806         return 0;
2807 }
2808 #endif
2809
2810 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2811
2812 void fastcall add_preempt_count(int val)
2813 {
2814         /*
2815          * Underflow?
2816          */
2817         BUG_ON((preempt_count() < 0));
2818         preempt_count() += val;
2819         /*
2820          * Spinlock count overflowing soon?
2821          */
2822         BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2823 }
2824 EXPORT_SYMBOL(add_preempt_count);
2825
2826 void fastcall sub_preempt_count(int val)
2827 {
2828         /*
2829          * Underflow?
2830          */
2831         BUG_ON(val > preempt_count());
2832         /*
2833          * Is the spinlock portion underflowing?
2834          */
2835         BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2836         preempt_count() -= val;
2837 }
2838 EXPORT_SYMBOL(sub_preempt_count);
2839
2840 #endif
2841
2842 /*
2843  * schedule() is the main scheduler function.
2844  */
2845 asmlinkage void __sched schedule(void)
2846 {
2847         long *switch_count;
2848         task_t *prev, *next;
2849         runqueue_t *rq;
2850         prio_array_t *array;
2851         struct list_head *queue;
2852         unsigned long long now;
2853         unsigned long run_time;
2854         int cpu, idx, new_prio;
2855
2856         /*
2857          * Test if we are atomic.  Since do_exit() needs to call into
2858          * schedule() atomically, we ignore that path for now.
2859          * Otherwise, whine if we are scheduling when we should not be.
2860          */
2861         if (likely(!current->exit_state)) {
2862                 if (unlikely(in_atomic())) {
2863                         printk(KERN_ERR "scheduling while atomic: "
2864                                 "%s/0x%08x/%d\n",
2865                                 current->comm, preempt_count(), current->pid);
2866                         dump_stack();
2867                 }
2868         }
2869         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2870
2871 need_resched:
2872         preempt_disable();
2873         prev = current;
2874         release_kernel_lock(prev);
2875 need_resched_nonpreemptible:
2876         rq = this_rq();
2877
2878         /*
2879          * The idle thread is not allowed to schedule!
2880          * Remove this check after it has been exercised a bit.
2881          */
2882         if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2883                 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2884                 dump_stack();
2885         }
2886
2887         schedstat_inc(rq, sched_cnt);
2888         now = sched_clock();
2889         if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2890                 run_time = now - prev->timestamp;
2891                 if (unlikely((long long)(now - prev->timestamp) < 0))
2892                         run_time = 0;
2893         } else
2894                 run_time = NS_MAX_SLEEP_AVG;
2895
2896         /*
2897          * Tasks charged proportionately less run_time at high sleep_avg to
2898          * delay them losing their interactive status
2899          */
2900         run_time /= (CURRENT_BONUS(prev) ? : 1);
2901
2902         spin_lock_irq(&rq->lock);
2903
2904         if (unlikely(prev->flags & PF_DEAD))
2905                 prev->state = EXIT_DEAD;
2906
2907         switch_count = &prev->nivcsw;
2908         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2909                 switch_count = &prev->nvcsw;
2910                 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2911                                 unlikely(signal_pending(prev))))
2912                         prev->state = TASK_RUNNING;
2913                 else {
2914                         if (prev->state == TASK_UNINTERRUPTIBLE)
2915                                 rq->nr_uninterruptible++;
2916                         deactivate_task(prev, rq);
2917                 }
2918         }
2919
2920         cpu = smp_processor_id();
2921         if (unlikely(!rq->nr_running)) {
2922 go_idle:
2923                 idle_balance(cpu, rq);
2924                 if (!rq->nr_running) {
2925                         next = rq->idle;
2926                         rq->expired_timestamp = 0;
2927                         wake_sleeping_dependent(cpu, rq);
2928                         /*
2929                          * wake_sleeping_dependent() might have released
2930                          * the runqueue, so break out if we got new
2931                          * tasks meanwhile:
2932                          */
2933                         if (!rq->nr_running)
2934                                 goto switch_tasks;
2935                 }
2936         } else {
2937                 if (dependent_sleeper(cpu, rq)) {
2938                         next = rq->idle;
2939                         goto switch_tasks;
2940                 }
2941                 /*
2942                  * dependent_sleeper() releases and reacquires the runqueue
2943                  * lock, hence go into the idle loop if the rq went
2944                  * empty meanwhile:
2945                  */
2946                 if (unlikely(!rq->nr_running))
2947                         goto go_idle;
2948         }
2949
2950         array = rq->active;
2951         if (unlikely(!array->nr_active)) {
2952                 /*
2953                  * Switch the active and expired arrays.
2954                  */
2955                 schedstat_inc(rq, sched_switch);
2956                 rq->active = rq->expired;
2957                 rq->expired = array;
2958                 array = rq->active;
2959                 rq->expired_timestamp = 0;
2960                 rq->best_expired_prio = MAX_PRIO;
2961         }
2962
2963         idx = sched_find_first_bit(array->bitmap);
2964         queue = array->queue + idx;
2965         next = list_entry(queue->next, task_t, run_list);
2966
2967         if (!rt_task(next) && next->activated > 0) {
2968                 unsigned long long delta = now - next->timestamp;
2969                 if (unlikely((long long)(now - next->timestamp) < 0))
2970                         delta = 0;
2971
2972                 if (next->activated == 1)
2973                         delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2974
2975                 array = next->array;
2976                 new_prio = recalc_task_prio(next, next->timestamp + delta);
2977
2978                 if (unlikely(next->prio != new_prio)) {
2979                         dequeue_task(next, array);
2980                         next->prio = new_prio;
2981                         enqueue_task(next, array);
2982                 } else
2983                         requeue_task(next, array);
2984         }
2985         next->activated = 0;
2986 switch_tasks:
2987         if (next == rq->idle)
2988                 schedstat_inc(rq, sched_goidle);
2989         prefetch(next);
2990         prefetch_stack(next);
2991         clear_tsk_need_resched(prev);
2992         rcu_qsctr_inc(task_cpu(prev));
2993
2994         update_cpu_clock(prev, rq, now);
2995
2996         prev->sleep_avg -= run_time;
2997         if ((long)prev->sleep_avg <= 0)
2998                 prev->sleep_avg = 0;
2999         prev->timestamp = prev->last_ran = now;
3000
3001         sched_info_switch(prev, next);
3002         if (likely(prev != next)) {
3003                 next->timestamp = now;
3004                 rq->nr_switches++;
3005                 rq->curr = next;
3006                 ++*switch_count;
3007
3008                 prepare_task_switch(rq, next);
3009                 prev = context_switch(rq, prev, next);
3010                 barrier();
3011                 /*
3012                  * this_rq must be evaluated again because prev may have moved
3013                  * CPUs since it called schedule(), thus the 'rq' on its stack
3014                  * frame will be invalid.
3015                  */
3016                 finish_task_switch(this_rq(), prev);
3017         } else
3018                 spin_unlock_irq(&rq->lock);
3019
3020         prev = current;
3021         if (unlikely(reacquire_kernel_lock(prev) < 0))
3022                 goto need_resched_nonpreemptible;
3023         preempt_enable_no_resched();
3024         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3025                 goto need_resched;
3026 }
3027
3028 EXPORT_SYMBOL(schedule);
3029
3030 #ifdef CONFIG_PREEMPT
3031 /*
3032  * this is is the entry point to schedule() from in-kernel preemption
3033  * off of preempt_enable.  Kernel preemptions off return from interrupt
3034  * occur there and call schedule directly.
3035  */
3036 asmlinkage void __sched preempt_schedule(void)
3037 {
3038         struct thread_info *ti = current_thread_info();
3039 #ifdef CONFIG_PREEMPT_BKL
3040         struct task_struct *task = current;
3041         int saved_lock_depth;
3042 #endif
3043         /*
3044          * If there is a non-zero preempt_count or interrupts are disabled,
3045          * we do not want to preempt the current task.  Just return..
3046          */
3047         if (unlikely(ti->preempt_count || irqs_disabled()))
3048                 return;
3049
3050 need_resched:
3051         add_preempt_count(PREEMPT_ACTIVE);
3052         /*
3053          * We keep the big kernel semaphore locked, but we
3054          * clear ->lock_depth so that schedule() doesnt
3055          * auto-release the semaphore:
3056          */
3057 #ifdef CONFIG_PREEMPT_BKL
3058         saved_lock_depth = task->lock_depth;
3059         task->lock_depth = -1;
3060 #endif
3061         schedule();
3062 #ifdef CONFIG_PREEMPT_BKL
3063         task->lock_depth = saved_lock_depth;
3064 #endif
3065         sub_preempt_count(PREEMPT_ACTIVE);
3066
3067         /* we could miss a preemption opportunity between schedule and now */
3068         barrier();
3069         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3070                 goto need_resched;
3071 }
3072
3073 EXPORT_SYMBOL(preempt_schedule);
3074
3075 /*
3076  * this is is the entry point to schedule() from kernel preemption
3077  * off of irq context.
3078  * Note, that this is called and return with irqs disabled. This will
3079  * protect us against recursive calling from irq.
3080  */
3081 asmlinkage void __sched preempt_schedule_irq(void)
3082 {
3083         struct thread_info *ti = current_thread_info();
3084 #ifdef CONFIG_PREEMPT_BKL
3085         struct task_struct *task = current;
3086         int saved_lock_depth;
3087 #endif
3088         /* Catch callers which need to be fixed*/
3089         BUG_ON(ti->preempt_count || !irqs_disabled());
3090
3091 need_resched:
3092         add_preempt_count(PREEMPT_ACTIVE);
3093         /*
3094          * We keep the big kernel semaphore locked, but we
3095          * clear ->lock_depth so that schedule() doesnt
3096          * auto-release the semaphore:
3097          */
3098 #ifdef CONFIG_PREEMPT_BKL
3099         saved_lock_depth = task->lock_depth;
3100         task->lock_depth = -1;
3101 #endif
3102         local_irq_enable();
3103         schedule();
3104         local_irq_disable();
3105 #ifdef CONFIG_PREEMPT_BKL
3106         task->lock_depth = saved_lock_depth;
3107 #endif
3108         sub_preempt_count(PREEMPT_ACTIVE);
3109
3110         /* we could miss a preemption opportunity between schedule and now */
3111         barrier();
3112         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3113                 goto need_resched;
3114 }
3115
3116 #endif /* CONFIG_PREEMPT */
3117
3118 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3119                           void *key)
3120 {
3121         task_t *p = curr->private;
3122         return try_to_wake_up(p, mode, sync);
3123 }
3124
3125 EXPORT_SYMBOL(default_wake_function);
3126
3127 /*
3128  * The core wakeup function.  Non-exclusive wakeups (nr_exclusive == 0) just
3129  * wake everything up.  If it's an exclusive wakeup (nr_exclusive == small +ve
3130  * number) then we wake all the non-exclusive tasks and one exclusive task.
3131  *
3132  * There are circumstances in which we can try to wake a task which has already
3133  * started to run but is not in state TASK_RUNNING.  try_to_wake_up() returns
3134  * zero in this (rare) case, and we handle it by continuing to scan the queue.
3135  */
3136 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3137                              int nr_exclusive, int sync, void *key)
3138 {
3139         struct list_head *tmp, *next;
3140
3141         list_for_each_safe(tmp, next, &q->task_list) {
3142                 wait_queue_t *curr;
3143                 unsigned flags;
3144                 curr = list_entry(tmp, wait_queue_t, task_list);
3145                 flags = curr->flags;
3146                 if (curr->func(curr, mode, sync, key) &&
3147                     (flags & WQ_FLAG_EXCLUSIVE) &&
3148                     !--nr_exclusive)
3149                         break;
3150         }
3151 }
3152
3153 /**
3154  * __wake_up - wake up threads blocked on a waitqueue.
3155  * @q: the waitqueue
3156  * @mode: which threads
3157  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3158  * @key: is directly passed to the wakeup function
3159  */
3160 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3161                         int nr_exclusive, void *key)
3162 {
3163         unsigned long flags;
3164
3165         spin_lock_irqsave(&q->lock, flags);
3166         __wake_up_common(q, mode, nr_exclusive, 0, key);
3167         spin_unlock_irqrestore(&q->lock, flags);
3168 }
3169
3170 EXPORT_SYMBOL(__wake_up);
3171
3172 /*
3173  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3174  */
3175 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3176 {
3177         __wake_up_common(q, mode, 1, 0, NULL);
3178 }
3179
3180 /**
3181  * __wake_up_sync - wake up threads blocked on a waitqueue.
3182  * @q: the waitqueue
3183  * @mode: which threads
3184  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3185  *
3186  * The sync wakeup differs that the waker knows that it will schedule
3187  * away soon, so while the target thread will be woken up, it will not
3188  * be migrated to another CPU - ie. the two threads are 'synchronized'
3189  * with each other. This can prevent needless bouncing between CPUs.
3190  *
3191  * On UP it can prevent extra preemption.
3192  */
3193 void fastcall
3194 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3195 {
3196         unsigned long flags;
3197         int sync = 1;
3198
3199         if (unlikely(!q))
3200                 return;
3201
3202         if (unlikely(!nr_exclusive))
3203                 sync = 0;
3204
3205         spin_lock_irqsave(&q->lock, flags);
3206         __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3207         spin_unlock_irqrestore(&q->lock, flags);
3208 }
3209 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3210
3211 void fastcall complete(struct completion *x)
3212 {
3213         unsigned long flags;
3214
3215         spin_lock_irqsave(&x->wait.lock, flags);
3216         x->done++;
3217         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3218                          1, 0, NULL);
3219         spin_unlock_irqrestore(&x->wait.lock, flags);
3220 }
3221 EXPORT_SYMBOL(complete);
3222
3223 void fastcall complete_all(struct completion *x)
3224 {
3225         unsigned long flags;
3226
3227         spin_lock_irqsave(&x->wait.lock, flags);
3228         x->done += UINT_MAX/2;
3229         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3230                          0, 0, NULL);
3231         spin_unlock_irqrestore(&x->wait.lock, flags);
3232 }
3233 EXPORT_SYMBOL(complete_all);
3234
3235 void fastcall __sched wait_for_completion(struct completion *x)
3236 {
3237         might_sleep();
3238         spin_lock_irq(&x->wait.lock);
3239         if (!x->done) {
3240                 DECLARE_WAITQUEUE(wait, current);
3241
3242                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3243                 __add_wait_queue_tail(&x->wait, &wait);
3244                 do {
3245                         __set_current_state(TASK_UNINTERRUPTIBLE);
3246                         spin_unlock_irq(&x->wait.lock);
3247                         schedule();
3248                         spin_lock_irq(&x->wait.lock);
3249                 } while (!x->done);
3250                 __remove_wait_queue(&x->wait, &wait);
3251         }
3252         x->done--;
3253         spin_unlock_irq(&x->wait.lock);
3254 }
3255 EXPORT_SYMBOL(wait_for_completion);
3256
3257 unsigned long fastcall __sched
3258 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3259 {
3260         might_sleep();
3261
3262         spin_lock_irq(&x->wait.lock);
3263         if (!x->done) {
3264                 DECLARE_WAITQUEUE(wait, current);
3265
3266                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3267                 __add_wait_queue_tail(&x->wait, &wait);
3268                 do {
3269                         __set_current_state(TASK_UNINTERRUPTIBLE);
3270                         spin_unlock_irq(&x->wait.lock);
3271                         timeout = schedule_timeout(timeout);
3272                         spin_lock_irq(&x->wait.lock);
3273                         if (!timeout) {
3274                                 __remove_wait_queue(&x->wait, &wait);
3275                                 goto out;
3276                         }
3277                 } while (!x->done);
3278                 __remove_wait_queue(&x->wait, &wait);
3279         }
3280         x->done--;
3281 out:
3282         spin_unlock_irq(&x->wait.lock);
3283         return timeout;
3284 }
3285 EXPORT_SYMBOL(wait_for_completion_timeout);
3286
3287 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3288 {
3289         int ret = 0;
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                                 ret = -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                         schedule();
3308                         spin_lock_irq(&x->wait.lock);
3309                 } while (!x->done);
3310                 __remove_wait_queue(&x->wait, &wait);
3311         }
3312         x->done--;
3313 out:
3314         spin_unlock_irq(&x->wait.lock);
3315
3316         return ret;
3317 }
3318 EXPORT_SYMBOL(wait_for_completion_interruptible);
3319
3320 unsigned long fastcall __sched
3321 wait_for_completion_interruptible_timeout(struct completion *x,
3322                                           unsigned long timeout)
3323 {
3324         might_sleep();
3325
3326         spin_lock_irq(&x->wait.lock);
3327         if (!x->done) {
3328                 DECLARE_WAITQUEUE(wait, current);
3329
3330                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3331                 __add_wait_queue_tail(&x->wait, &wait);
3332                 do {
3333                         if (signal_pending(current)) {
3334                                 timeout = -ERESTARTSYS;
3335                                 __remove_wait_queue(&x->wait, &wait);
3336                                 goto out;
3337                         }
3338                         __set_current_state(TASK_INTERRUPTIBLE);
3339                         spin_unlock_irq(&x->wait.lock);
3340                         timeout = schedule_timeout(timeout);
3341                         spin_lock_irq(&x->wait.lock);
3342                         if (!timeout) {
3343                                 __remove_wait_queue(&x->wait, &wait);
3344                                 goto out;
3345                         }
3346                 } while (!x->done);
3347                 __remove_wait_queue(&x->wait, &wait);
3348         }
3349         x->done--;
3350 out:
3351         spin_unlock_irq(&x->wait.lock);
3352         return timeout;
3353 }
3354 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3355
3356
3357 #define SLEEP_ON_VAR                                    \
3358         unsigned long flags;                            \
3359         wait_queue_t wait;                              \
3360         init_waitqueue_entry(&wait, current);
3361
3362 #define SLEEP_ON_HEAD                                   \
3363         spin_lock_irqsave(&q->lock,flags);              \
3364         __add_wait_queue(q, &wait);                     \
3365         spin_unlock(&q->lock);
3366
3367 #define SLEEP_ON_TAIL                                   \
3368         spin_lock_irq(&q->lock);                        \
3369         __remove_wait_queue(q, &wait);                  \
3370         spin_unlock_irqrestore(&q->lock, flags);
3371
3372 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3373 {
3374         SLEEP_ON_VAR
3375
3376         current->state = TASK_INTERRUPTIBLE;
3377
3378         SLEEP_ON_HEAD
3379         schedule();
3380         SLEEP_ON_TAIL
3381 }
3382
3383 EXPORT_SYMBOL(interruptible_sleep_on);
3384
3385 long fastcall __sched
3386 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3387 {
3388         SLEEP_ON_VAR
3389
3390         current->state = TASK_INTERRUPTIBLE;
3391
3392         SLEEP_ON_HEAD
3393         timeout = schedule_timeout(timeout);
3394         SLEEP_ON_TAIL
3395
3396         return timeout;
3397 }
3398
3399 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3400
3401 void fastcall __sched sleep_on(wait_queue_head_t *q)
3402 {
3403         SLEEP_ON_VAR
3404
3405         current->state = TASK_UNINTERRUPTIBLE;
3406
3407         SLEEP_ON_HEAD
3408         schedule();
3409         SLEEP_ON_TAIL
3410 }
3411
3412 EXPORT_SYMBOL(sleep_on);
3413
3414 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3415 {
3416         SLEEP_ON_VAR
3417
3418         current->state = TASK_UNINTERRUPTIBLE;
3419
3420         SLEEP_ON_HEAD
3421         timeout = schedule_timeout(timeout);
3422         SLEEP_ON_TAIL
3423
3424         return timeout;
3425 }
3426
3427 EXPORT_SYMBOL(sleep_on_timeout);
3428
3429 void set_user_nice(task_t *p, long nice)
3430 {
3431         unsigned long flags;
3432         prio_array_t *array;
3433         runqueue_t *rq;
3434         int old_prio, new_prio, delta;
3435
3436         if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3437                 return;
3438         /*
3439          * We have to be careful, if called from sys_setpriority(),
3440          * the task might be in the middle of scheduling on another CPU.
3441          */
3442         rq = task_rq_lock(p, &flags);
3443         /*
3444          * The RT priorities are set via sched_setscheduler(), but we still
3445          * allow the 'normal' nice value to be set - but as expected
3446          * it wont have any effect on scheduling until the task is
3447          * not SCHED_NORMAL:
3448          */
3449         if (rt_task(p)) {
3450                 p->static_prio = NICE_TO_PRIO(nice);
3451                 goto out_unlock;
3452         }
3453         array = p->array;
3454         if (array)
3455                 dequeue_task(p, array);
3456
3457         old_prio = p->prio;
3458         new_prio = NICE_TO_PRIO(nice);
3459         delta = new_prio - old_prio;
3460         p->static_prio = NICE_TO_PRIO(nice);
3461         p->prio += delta;
3462
3463         if (array) {
3464                 enqueue_task(p, array);
3465                 /*
3466                  * If the task increased its priority or is running and
3467                  * lowered its priority, then reschedule its CPU:
3468                  */
3469                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3470                         resched_task(rq->curr);
3471         }
3472 out_unlock:
3473         task_rq_unlock(rq, &flags);
3474 }
3475
3476 EXPORT_SYMBOL(set_user_nice);
3477
3478 /*
3479  * can_nice - check if a task can reduce its nice value
3480  * @p: task
3481  * @nice: nice value
3482  */
3483 int can_nice(const task_t *p, const int nice)
3484 {
3485         /* convert nice value [19,-20] to rlimit style value [1,40] */
3486         int nice_rlim = 20 - nice;
3487         return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3488                 capable(CAP_SYS_NICE));
3489 }
3490
3491 #ifdef __ARCH_WANT_SYS_NICE
3492
3493 /*
3494  * sys_nice - change the priority of the current process.
3495  * @increment: priority increment
3496  *
3497  * sys_setpriority is a more generic, but much slower function that
3498  * does similar things.
3499  */
3500 asmlinkage long sys_nice(int increment)
3501 {
3502         int retval;
3503         long nice;
3504
3505         /*
3506          * Setpriority might change our priority at the same moment.
3507          * We don't have to worry. Conceptually one call occurs first
3508          * and we have a single winner.
3509          */
3510         if (increment < -40)
3511                 increment = -40;
3512         if (increment > 40)
3513                 increment = 40;
3514
3515         nice = PRIO_TO_NICE(current->static_prio) + increment;
3516         if (nice < -20)
3517                 nice = -20;
3518         if (nice > 19)
3519                 nice = 19;
3520
3521         if (increment < 0 && !can_nice(current, nice))
3522                 return -EPERM;
3523
3524         retval = security_task_setnice(current, nice);
3525         if (retval)
3526                 return retval;
3527
3528         set_user_nice(current, nice);
3529         return 0;
3530 }
3531
3532 #endif
3533
3534 /**
3535  * task_prio - return the priority value of a given task.
3536  * @p: the task in question.
3537  *
3538  * This is the priority value as seen by users in /proc.
3539  * RT tasks are offset by -200. Normal tasks are centered
3540  * around 0, value goes from -16 to +15.
3541  */
3542 int task_prio(const task_t *p)
3543 {
3544         return p->prio - MAX_RT_PRIO;
3545 }
3546
3547 /**
3548  * task_nice - return the nice value of a given task.
3549  * @p: the task in question.
3550  */
3551 int task_nice(const task_t *p)
3552 {
3553         return TASK_NICE(p);
3554 }
3555 EXPORT_SYMBOL_GPL(task_nice);
3556
3557 /**
3558  * idle_cpu - is a given cpu idle currently?
3559  * @cpu: the processor in question.
3560  */
3561 int idle_cpu(int cpu)
3562 {
3563         return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3564 }
3565
3566 /**
3567  * idle_task - return the idle task for a given cpu.
3568  * @cpu: the processor in question.
3569  */
3570 task_t *idle_task(int cpu)
3571 {
3572         return cpu_rq(cpu)->idle;
3573 }
3574
3575 /**
3576  * find_process_by_pid - find a process with a matching PID value.
3577  * @pid: the pid in question.
3578  */
3579 static inline task_t *find_process_by_pid(pid_t pid)
3580 {
3581         return pid ? find_task_by_pid(pid) : current;
3582 }
3583
3584 /* Actually do priority change: must hold rq lock. */
3585 static void __setscheduler(struct task_struct *p, int policy, int prio)
3586 {
3587         BUG_ON(p->array);
3588         p->policy = policy;
3589         p->rt_priority = prio;
3590         if (policy != SCHED_NORMAL)
3591                 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3592         else
3593                 p->prio = p->static_prio;
3594 }
3595
3596 /**
3597  * sched_setscheduler - change the scheduling policy and/or RT priority of
3598  * a thread.
3599  * @p: the task in question.
3600  * @policy: new policy.
3601  * @param: structure containing the new RT priority.
3602  */
3603 int sched_setscheduler(struct task_struct *p, int policy,
3604                        struct sched_param *param)
3605 {
3606         int retval;
3607         int oldprio, oldpolicy = -1;
3608         prio_array_t *array;
3609         unsigned long flags;
3610         runqueue_t *rq;
3611
3612 recheck:
3613         /* double check policy once rq lock held */
3614         if (policy < 0)
3615                 policy = oldpolicy = p->policy;
3616         else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3617                                 policy != SCHED_NORMAL)
3618                         return -EINVAL;
3619         /*
3620          * Valid priorities for SCHED_FIFO and SCHED_RR are
3621          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3622          */
3623         if (param->sched_priority < 0 ||
3624             (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3625             (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3626                 return -EINVAL;
3627         if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3628                 return -EINVAL;
3629
3630         /*
3631          * Allow unprivileged RT tasks to decrease priority:
3632          */
3633         if (!capable(CAP_SYS_NICE)) {
3634                 /* can't change policy */
3635                 if (policy != p->policy &&
3636                         !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3637                         return -EPERM;
3638                 /* can't increase priority */
3639                 if (policy != SCHED_NORMAL &&
3640                     param->sched_priority > p->rt_priority &&
3641                     param->sched_priority >
3642                                 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3643                         return -EPERM;
3644                 /* can't change other user's priorities */
3645                 if ((current->euid != p->euid) &&
3646                     (current->euid != p->uid))
3647                         return -EPERM;
3648         }
3649
3650         retval = security_task_setscheduler(p, policy, param);
3651         if (retval)
3652                 return retval;
3653         /*
3654          * To be able to change p->policy safely, the apropriate
3655          * runqueue lock must be held.
3656          */
3657         rq = task_rq_lock(p, &flags);
3658         /* recheck policy now with rq lock held */
3659         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3660                 policy = oldpolicy = -1;
3661                 task_rq_unlock(rq, &flags);
3662                 goto recheck;
3663         }
3664         array = p->array;
3665         if (array)
3666                 deactivate_task(p, rq);
3667         oldprio = p->prio;
3668         __setscheduler(p, policy, param->sched_priority);
3669         if (array) {
3670                 __activate_task(p, rq);
3671                 /*
3672                  * Reschedule if we are currently running on this runqueue and
3673                  * our priority decreased, or if we are not currently running on
3674                  * this runqueue and our priority is higher than the current's
3675                  */
3676                 if (task_running(rq, p)) {
3677                         if (p->prio > oldprio)
3678                                 resched_task(rq->curr);
3679                 } else if (TASK_PREEMPTS_CURR(p, rq))
3680                         resched_task(rq->curr);
3681         }
3682         task_rq_unlock(rq, &flags);
3683         return 0;
3684 }
3685 EXPORT_SYMBOL_GPL(sched_setscheduler);
3686
3687 static int
3688 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3689 {
3690         int retval;
3691         struct sched_param lparam;
3692         struct task_struct *p;
3693
3694         if (!param || pid < 0)
3695                 return -EINVAL;
3696         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3697                 return -EFAULT;
3698         read_lock_irq(&tasklist_lock);
3699         p = find_process_by_pid(pid);
3700         if (!p) {
3701                 read_unlock_irq(&tasklist_lock);
3702                 return -ESRCH;
3703         }
3704         retval = sched_setscheduler(p, policy, &lparam);
3705         read_unlock_irq(&tasklist_lock);
3706         return retval;
3707 }
3708
3709 /**
3710  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3711  * @pid: the pid in question.
3712  * @policy: new policy.
3713  * @param: structure containing the new RT priority.
3714  */
3715 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3716                                        struct sched_param __user *param)
3717 {
3718         return do_sched_setscheduler(pid, policy, param);
3719 }
3720
3721 /**
3722  * sys_sched_setparam - set/change the RT priority of a thread
3723  * @pid: the pid in question.
3724  * @param: structure containing the new RT priority.
3725  */
3726 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3727 {
3728         return do_sched_setscheduler(pid, -1, param);
3729 }
3730
3731 /**
3732  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3733  * @pid: the pid in question.
3734  */
3735 asmlinkage long sys_sched_getscheduler(pid_t pid)
3736 {
3737         int retval = -EINVAL;
3738         task_t *p;
3739
3740         if (pid < 0)
3741                 goto out_nounlock;
3742
3743         retval = -ESRCH;
3744         read_lock(&tasklist_lock);
3745         p = find_process_by_pid(pid);
3746         if (p) {
3747                 retval = security_task_getscheduler(p);
3748                 if (!retval)
3749                         retval = p->policy;
3750         }
3751         read_unlock(&tasklist_lock);
3752
3753 out_nounlock:
3754         return retval;
3755 }
3756
3757 /**
3758  * sys_sched_getscheduler - get the RT priority of a thread
3759  * @pid: the pid in question.
3760  * @param: structure containing the RT priority.
3761  */
3762 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3763 {
3764         struct sched_param lp;
3765         int retval = -EINVAL;
3766         task_t *p;
3767
3768         if (!param || pid < 0)
3769                 goto out_nounlock;
3770
3771         read_lock(&tasklist_lock);
3772         p = find_process_by_pid(pid);
3773         retval = -ESRCH;
3774         if (!p)
3775                 goto out_unlock;
3776
3777         retval = security_task_getscheduler(p);
3778         if (retval)
3779                 goto out_unlock;
3780
3781         lp.sched_priority = p->rt_priority;
3782         read_unlock(&tasklist_lock);
3783
3784         /*
3785          * This one might sleep, we cannot do it with a spinlock held ...
3786          */
3787         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3788
3789 out_nounlock:
3790         return retval;
3791
3792 out_unlock:
3793         read_unlock(&tasklist_lock);
3794         return retval;
3795 }
3796
3797 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3798 {
3799         task_t *p;
3800         int retval;
3801         cpumask_t cpus_allowed;
3802
3803         lock_cpu_hotplug();
3804         read_lock(&tasklist_lock);
3805
3806         p = find_process_by_pid(pid);
3807         if (!p) {
3808                 read_unlock(&tasklist_lock);
3809                 unlock_cpu_hotplug();
3810                 return -ESRCH;
3811         }
3812
3813         /*
3814          * It is not safe to call set_cpus_allowed with the
3815          * tasklist_lock held.  We will bump the task_struct's
3816          * usage count and then drop tasklist_lock.
3817          */
3818         get_task_struct(p);
3819         read_unlock(&tasklist_lock);
3820
3821         retval = -EPERM;
3822         if ((current->euid != p->euid) && (current->euid != p->uid) &&
3823                         !capable(CAP_SYS_NICE))
3824                 goto out_unlock;
3825
3826         cpus_allowed = cpuset_cpus_allowed(p);
3827         cpus_and(new_mask, new_mask, cpus_allowed);
3828         retval = set_cpus_allowed(p, new_mask);
3829
3830 out_unlock:
3831         put_task_struct(p);
3832         unlock_cpu_hotplug();
3833         return retval;
3834 }
3835
3836 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3837                              cpumask_t *new_mask)
3838 {
3839         if (len < sizeof(cpumask_t)) {
3840                 memset(new_mask, 0, sizeof(cpumask_t));
3841         } else if (len > sizeof(cpumask_t)) {
3842                 len = sizeof(cpumask_t);
3843         }
3844         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3845 }
3846
3847 /**
3848  * sys_sched_setaffinity - set the cpu affinity of a process
3849  * @pid: pid of the process
3850  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3851  * @user_mask_ptr: user-space pointer to the new cpu mask
3852  */
3853 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3854                                       unsigned long __user *user_mask_ptr)
3855 {
3856         cpumask_t new_mask;
3857         int retval;
3858
3859         retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3860         if (retval)
3861                 return retval;
3862
3863         return sched_setaffinity(pid, new_mask);
3864 }
3865
3866 /*
3867  * Represents all cpu's present in the system
3868  * In systems capable of hotplug, this map could dynamically grow
3869  * as new cpu's are detected in the system via any platform specific
3870  * method, such as ACPI for e.g.
3871  */
3872
3873 cpumask_t cpu_present_map;
3874 EXPORT_SYMBOL(cpu_present_map);
3875
3876 #ifndef CONFIG_SMP
3877 cpumask_t cpu_online_map = CPU_MASK_ALL;
3878 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3879 #endif
3880
3881 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3882 {
3883         int retval;
3884         task_t *p;
3885
3886         lock_cpu_hotplug();
3887         read_lock(&tasklist_lock);
3888
3889         retval = -ESRCH;
3890         p = find_process_by_pid(pid);
3891         if (!p)
3892                 goto out_unlock;
3893
3894         retval = 0;
3895         cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
3896
3897 out_unlock:
3898         read_unlock(&tasklist_lock);
3899         unlock_cpu_hotplug();
3900         if (retval)
3901                 return retval;
3902
3903         return 0;
3904 }
3905
3906 /**
3907  * sys_sched_getaffinity - get the cpu affinity of a process
3908  * @pid: pid of the process
3909  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3910  * @user_mask_ptr: user-space pointer to hold the current cpu mask
3911  */
3912 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3913                                       unsigned long __user *user_mask_ptr)
3914 {
3915         int ret;
3916         cpumask_t mask;
3917
3918         if (len < sizeof(cpumask_t))
3919                 return -EINVAL;
3920
3921         ret = sched_getaffinity(pid, &mask);
3922         if (ret < 0)
3923                 return ret;
3924
3925         if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3926                 return -EFAULT;
3927
3928         return sizeof(cpumask_t);
3929 }
3930
3931 /**
3932  * sys_sched_yield - yield the current processor to other threads.
3933  *
3934  * this function yields the current CPU by moving the calling thread
3935  * to the expired array. If there are no other threads running on this
3936  * CPU then this function will return.
3937  */
3938 asmlinkage long sys_sched_yield(void)
3939 {
3940         runqueue_t *rq = this_rq_lock();
3941         prio_array_t *array = current->array;
3942         prio_array_t *target = rq->expired;
3943
3944         schedstat_inc(rq, yld_cnt);
3945         /*
3946          * We implement yielding by moving the task into the expired
3947          * queue.
3948          *
3949          * (special rule: RT tasks will just roundrobin in the active
3950          *  array.)
3951          */
3952         if (rt_task(current))
3953                 target = rq->active;
3954
3955         if (array->nr_active == 1) {
3956                 schedstat_inc(rq, yld_act_empty);
3957                 if (!rq->expired->nr_active)
3958                         schedstat_inc(rq, yld_both_empty);
3959         } else if (!rq->expired->nr_active)
3960                 schedstat_inc(rq, yld_exp_empty);
3961
3962         if (array != target) {
3963                 dequeue_task(current, array);
3964                 enqueue_task(current, target);
3965         } else
3966                 /*
3967                  * requeue_task is cheaper so perform that if possible.
3968                  */
3969                 requeue_task(current, array);
3970
3971         /*
3972          * Since we are going to call schedule() anyway, there's
3973          * no need to preempt or enable interrupts:
3974          */
3975         __release(rq->lock);
3976         _raw_spin_unlock(&rq->lock);
3977         preempt_enable_no_resched();
3978
3979         schedule();
3980
3981         return 0;
3982 }
3983
3984 static inline void __cond_resched(void)
3985 {
3986         /*
3987          * The BKS might be reacquired before we have dropped
3988          * PREEMPT_ACTIVE, which could trigger a second
3989          * cond_resched() call.
3990          */
3991         if (unlikely(preempt_count()))
3992                 return;
3993         do {
3994                 add_preempt_count(PREEMPT_ACTIVE);
3995                 schedule();
3996                 sub_preempt_count(PREEMPT_ACTIVE);
3997         } while (need_resched());
3998 }
3999
4000 int __sched cond_resched(void)
4001 {
4002         if (need_resched()) {
4003                 __cond_resched();
4004                 return 1;
4005         }
4006         return 0;
4007 }
4008
4009 EXPORT_SYMBOL(cond_resched);
4010
4011 /*
4012  * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4013  * call schedule, and on return reacquire the lock.
4014  *
4015  * This works OK both with and without CONFIG_PREEMPT.  We do strange low-level
4016  * operations here to prevent schedule() from being called twice (once via
4017  * spin_unlock(), once by hand).
4018  */
4019 int cond_resched_lock(spinlock_t *lock)
4020 {
4021         int ret = 0;
4022
4023         if (need_lockbreak(lock)) {
4024                 spin_unlock(lock);
4025                 cpu_relax();
4026                 ret = 1;
4027                 spin_lock(lock);
4028         }
4029         if (need_resched()) {
4030                 _raw_spin_unlock(lock);
4031                 preempt_enable_no_resched();
4032                 __cond_resched();
4033                 ret = 1;
4034                 spin_lock(lock);
4035         }
4036         return ret;
4037 }
4038
4039 EXPORT_SYMBOL(cond_resched_lock);
4040
4041 int __sched cond_resched_softirq(void)
4042 {
4043         BUG_ON(!in_softirq());
4044
4045         if (need_resched()) {
4046                 __local_bh_enable();
4047                 __cond_resched();
4048                 local_bh_disable();
4049                 return 1;
4050         }
4051         return 0;
4052 }
4053
4054 EXPORT_SYMBOL(cond_resched_softirq);
4055
4056
4057 /**
4058  * yield - yield the current processor to other threads.
4059  *
4060  * this is a shortcut for kernel-space yielding - it marks the
4061  * thread runnable and calls sys_sched_yield().
4062  */
4063 void __sched yield(void)
4064 {
4065         set_current_state(TASK_RUNNING);
4066         sys_sched_yield();
4067 }
4068
4069 EXPORT_SYMBOL(yield);
4070
4071 /*
4072  * This task is about to go to sleep on IO.  Increment rq->nr_iowait so
4073  * that process accounting knows that this is a task in IO wait state.
4074  *
4075  * But don't do that if it is a deliberate, throttling IO wait (this task
4076  * has set its backing_dev_info: the queue against which it should throttle)
4077  */
4078 void __sched io_schedule(void)
4079 {
4080         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4081
4082         atomic_inc(&rq->nr_iowait);
4083         schedule();
4084         atomic_dec(&rq->nr_iowait);
4085 }
4086
4087 EXPORT_SYMBOL(io_schedule);
4088
4089 long __sched io_schedule_timeout(long timeout)
4090 {
4091         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4092         long ret;
4093
4094         atomic_inc(&rq->nr_iowait);
4095         ret = schedule_timeout(timeout);
4096         atomic_dec(&rq->nr_iowait);
4097         return ret;
4098 }
4099
4100 /**
4101  * sys_sched_get_priority_max - return maximum RT priority.
4102  * @policy: scheduling class.
4103  *
4104  * this syscall returns the maximum rt_priority that can be used
4105  * by a given scheduling class.
4106  */
4107 asmlinkage long sys_sched_get_priority_max(int policy)
4108 {
4109         int ret = -EINVAL;
4110
4111         switch (policy) {
4112         case SCHED_FIFO:
4113         case SCHED_RR:
4114                 ret = MAX_USER_RT_PRIO-1;
4115                 break;
4116         case SCHED_NORMAL:
4117                 ret = 0;
4118                 break;
4119         }
4120         return ret;
4121 }
4122
4123 /**
4124  * sys_sched_get_priority_min - return minimum RT priority.
4125  * @policy: scheduling class.
4126  *
4127  * this syscall returns the minimum rt_priority that can be used
4128  * by a given scheduling class.
4129  */
4130 asmlinkage long sys_sched_get_priority_min(int policy)
4131 {
4132         int ret = -EINVAL;
4133
4134         switch (policy) {
4135         case SCHED_FIFO:
4136         case SCHED_RR:
4137                 ret = 1;
4138                 break;
4139         case SCHED_NORMAL:
4140                 ret = 0;
4141         }
4142         return ret;
4143 }
4144
4145 /**
4146  * sys_sched_rr_get_interval - return the default timeslice of a process.
4147  * @pid: pid of the process.
4148  * @interval: userspace pointer to the timeslice value.
4149  *
4150  * this syscall writes the default timeslice value of a given process
4151  * into the user-space timespec buffer. A value of '0' means infinity.
4152  */
4153 asmlinkage
4154 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4155 {
4156         int retval = -EINVAL;
4157         struct timespec t;
4158         task_t *p;
4159
4160         if (pid < 0)
4161                 goto out_nounlock;
4162
4163         retval = -ESRCH;
4164         read_lock(&tasklist_lock);
4165         p = find_process_by_pid(pid);
4166         if (!p)
4167                 goto out_unlock;
4168
4169         retval = security_task_getscheduler(p);
4170         if (retval)
4171                 goto out_unlock;
4172
4173         jiffies_to_timespec(p->policy & SCHED_FIFO ?
4174                                 0 : task_timeslice(p), &t);
4175         read_unlock(&tasklist_lock);
4176         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4177 out_nounlock:
4178         return retval;
4179 out_unlock:
4180         read_unlock(&tasklist_lock);
4181         return retval;
4182 }
4183
4184 static inline struct task_struct *eldest_child(struct task_struct *p)
4185 {
4186         if (list_empty(&p->children)) return NULL;
4187         return list_entry(p->children.next,struct task_struct,sibling);
4188 }
4189
4190 static inline struct task_struct *older_sibling(struct task_struct *p)
4191 {
4192         if (p->sibling.prev==&p->parent->children) return NULL;
4193         return list_entry(p->sibling.prev,struct task_struct,sibling);
4194 }
4195
4196 static inline struct task_struct *younger_sibling(struct task_struct *p)
4197 {
4198         if (p->sibling.next==&p->parent->children) return NULL;
4199         return list_entry(p->sibling.next,struct task_struct,sibling);
4200 }
4201
4202 static void show_task(task_t *p)
4203 {
4204         task_t *relative;
4205         unsigned state;
4206         unsigned long free = 0;
4207         static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4208
4209         printk("%-13.13s ", p->comm);
4210         state = p->state ? __ffs(p->state) + 1 : 0;
4211         if (state < ARRAY_SIZE(stat_nam))
4212                 printk(stat_nam[state]);
4213         else
4214                 printk("?");
4215 #if (BITS_PER_LONG == 32)
4216         if (state == TASK_RUNNING)
4217                 printk(" running ");
4218         else
4219                 printk(" %08lX ", thread_saved_pc(p));
4220 #else
4221         if (state == TASK_RUNNING)
4222                 printk("  running task   ");
4223         else
4224                 printk(" %016lx ", thread_saved_pc(p));
4225 #endif
4226 #ifdef CONFIG_DEBUG_STACK_USAGE
4227         {
4228                 unsigned long *n = (unsigned long *) (p->thread_info+1);
4229                 while (!*n)
4230                         n++;
4231                 free = (unsigned long) n - (unsigned long)(p->thread_info+1);
4232         }
4233 #endif
4234         printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4235         if ((relative = eldest_child(p)))
4236                 printk("%5d ", relative->pid);
4237         else
4238                 printk("      ");
4239         if ((relative = younger_sibling(p)))
4240                 printk("%7d", relative->pid);
4241         else
4242                 printk("       ");
4243         if ((relative = older_sibling(p)))
4244                 printk(" %5d", relative->pid);
4245         else
4246                 printk("      ");
4247         if (!p->mm)
4248                 printk(" (L-TLB)\n");
4249         else
4250                 printk(" (NOTLB)\n");
4251
4252         if (state != TASK_RUNNING)
4253                 show_stack(p, NULL);
4254 }
4255
4256 void show_state(void)
4257 {
4258         task_t *g, *p;
4259
4260 #if (BITS_PER_LONG == 32)
4261         printk("\n"
4262                "                                               sibling\n");
4263         printk("  task             PC      pid father child younger older\n");
4264 #else
4265         printk("\n"
4266                "                                                       sibling\n");
4267         printk("  task                 PC          pid father child younger older\n");
4268 #endif
4269         read_lock(&tasklist_lock);
4270         do_each_thread(g, p) {
4271                 /*
4272                  * reset the NMI-timeout, listing all files on a slow
4273                  * console might take alot of time:
4274                  */
4275                 touch_nmi_watchdog();
4276                 show_task(p);
4277         } while_each_thread(g, p);
4278
4279         read_unlock(&tasklist_lock);
4280 }
4281
4282 /**
4283  * init_idle - set up an idle thread for a given CPU
4284  * @idle: task in question
4285  * @cpu: cpu the idle task belongs to
4286  *
4287  * NOTE: this function does not set the idle thread's NEED_RESCHED
4288  * flag, to make booting more robust.
4289  */
4290 void __devinit init_idle(task_t *idle, int cpu)
4291 {
4292         runqueue_t *rq = cpu_rq(cpu);
4293         unsigned long flags;
4294
4295         idle->sleep_avg = 0;
4296         idle->array = NULL;
4297         idle->prio = MAX_PRIO;
4298         idle->state = TASK_RUNNING;
4299         idle->cpus_allowed = cpumask_of_cpu(cpu);
4300         set_task_cpu(idle, cpu);
4301
4302         spin_lock_irqsave(&rq->lock, flags);
4303         rq->curr = rq->idle = idle;
4304 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4305         idle->oncpu = 1;
4306 #endif
4307         spin_unlock_irqrestore(&rq->lock, flags);
4308
4309         /* Set the preempt count _outside_ the spinlocks! */
4310 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4311         idle->thread_info->preempt_count = (idle->lock_depth >= 0);
4312 #else
4313         idle->thread_info->preempt_count = 0;
4314 #endif
4315 }
4316
4317 /*
4318  * In a system that switches off the HZ timer nohz_cpu_mask
4319  * indicates which cpus entered this state. This is used
4320  * in the rcu update to wait only for active cpus. For system
4321  * which do not switch off the HZ timer nohz_cpu_mask should
4322  * always be CPU_MASK_NONE.
4323  */
4324 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4325
4326 #ifdef CONFIG_SMP
4327 /*
4328  * This is how migration works:
4329  *
4330  * 1) we queue a migration_req_t structure in the source CPU's
4331  *    runqueue and wake up that CPU's migration thread.
4332  * 2) we down() the locked semaphore => thread blocks.
4333  * 3) migration thread wakes up (implicitly it forces the migrated
4334  *    thread off the CPU)
4335  * 4) it gets the migration request and checks whether the migrated
4336  *    task is still in the wrong runqueue.
4337  * 5) if it's in the wrong runqueue then the migration thread removes
4338  *    it and puts it into the right queue.
4339  * 6) migration thread up()s the semaphore.
4340  * 7) we wake up and the migration is done.
4341  */
4342
4343 /*
4344  * Change a given task's CPU affinity. Migrate the thread to a
4345  * proper CPU and schedule it away if the CPU it's executing on
4346  * is removed from the allowed bitmask.
4347  *
4348  * NOTE: the caller must have a valid reference to the task, the
4349  * task must not exit() & deallocate itself prematurely.  The
4350  * call is not atomic; no spinlocks may be held.
4351  */
4352 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4353 {
4354         unsigned long flags;
4355         int ret = 0;
4356         migration_req_t req;
4357         runqueue_t *rq;
4358
4359         rq = task_rq_lock(p, &flags);
4360         if (!cpus_intersects(new_mask, cpu_online_map)) {
4361                 ret = -EINVAL;
4362                 goto out;
4363         }
4364
4365         p->cpus_allowed = new_mask;
4366         /* Can the task run on the task's current CPU? If so, we're done */
4367         if (cpu_isset(task_cpu(p), new_mask))
4368                 goto out;
4369
4370         if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4371                 /* Need help from migration thread: drop lock and wait. */
4372                 task_rq_unlock(rq, &flags);
4373                 wake_up_process(rq->migration_thread);
4374                 wait_for_completion(&req.done);
4375                 tlb_migrate_finish(p->mm);
4376                 return 0;
4377         }
4378 out:
4379         task_rq_unlock(rq, &flags);
4380         return ret;
4381 }
4382
4383 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4384
4385 /*
4386  * Move (not current) task off this cpu, onto dest cpu.  We're doing
4387  * this because either it can't run here any more (set_cpus_allowed()
4388  * away from this CPU, or CPU going down), or because we're
4389  * attempting to rebalance this task on exec (sched_exec).
4390  *
4391  * So we race with normal scheduler movements, but that's OK, as long
4392  * as the task is no longer on this CPU.
4393  */
4394 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4395 {
4396         runqueue_t *rq_dest, *rq_src;
4397
4398         if (unlikely(cpu_is_offline(dest_cpu)))
4399                 return;
4400
4401         rq_src = cpu_rq(src_cpu);
4402         rq_dest = cpu_rq(dest_cpu);
4403
4404         double_rq_lock(rq_src, rq_dest);
4405         /* Already moved. */
4406         if (task_cpu(p) != src_cpu)
4407                 goto out;
4408         /* Affinity changed (again). */
4409         if (!cpu_isset(dest_cpu, p->cpus_allowed))
4410                 goto out;
4411
4412         set_task_cpu(p, dest_cpu);
4413         if (p->array) {
4414                 /*
4415                  * Sync timestamp with rq_dest's before activating.
4416                  * The same thing could be achieved by doing this step
4417                  * afterwards, and pretending it was a local activate.
4418                  * This way is cleaner and logically correct.
4419                  */
4420                 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4421                                 + rq_dest->timestamp_last_tick;
4422                 deactivate_task(p, rq_src);
4423                 activate_task(p, rq_dest, 0);
4424                 if (TASK_PREEMPTS_CURR(p, rq_dest))
4425                         resched_task(rq_dest->curr);
4426         }
4427
4428 out:
4429         double_rq_unlock(rq_src, rq_dest);
4430 }
4431
4432 /*
4433  * migration_thread - this is a highprio system thread that performs
4434  * thread migration by bumping thread off CPU then 'pushing' onto
4435  * another runqueue.
4436  */
4437 static int migration_thread(void *data)
4438 {
4439         runqueue_t *rq;
4440         int cpu = (long)data;
4441
4442         rq = cpu_rq(cpu);
4443         BUG_ON(rq->migration_thread != current);
4444
4445         set_current_state(TASK_INTERRUPTIBLE);
4446         while (!kthread_should_stop()) {
4447                 struct list_head *head;
4448                 migration_req_t *req;
4449
4450                 try_to_freeze();
4451
4452                 spin_lock_irq(&rq->lock);
4453
4454                 if (cpu_is_offline(cpu)) {
4455                         spin_unlock_irq(&rq->lock);
4456                         goto wait_to_die;
4457                 }
4458
4459                 if (rq->active_balance) {
4460                         active_load_balance(rq, cpu);
4461                         rq->active_balance = 0;
4462                 }
4463
4464                 head = &rq->migration_queue;
4465
4466                 if (list_empty(head)) {
4467                         spin_unlock_irq(&rq->lock);
4468                         schedule();
4469                         set_current_state(TASK_INTERRUPTIBLE);
4470                         continue;
4471                 }
4472                 req = list_entry(head->next, migration_req_t, list);
4473                 list_del_init(head->next);
4474
4475                 spin_unlock(&rq->lock);
4476                 __migrate_task(req->task, cpu, req->dest_cpu);
4477                 local_irq_enable();
4478
4479                 complete(&req->done);
4480         }
4481         __set_current_state(TASK_RUNNING);
4482         return 0;
4483
4484 wait_to_die:
4485         /* Wait for kthread_stop */
4486         set_current_state(TASK_INTERRUPTIBLE);
4487         while (!kthread_should_stop()) {
4488                 schedule();
4489                 set_current_state(TASK_INTERRUPTIBLE);
4490         }
4491         __set_current_state(TASK_RUNNING);
4492         return 0;
4493 }
4494
4495 #ifdef CONFIG_HOTPLUG_CPU
4496 /* Figure out where task on dead CPU should go, use force if neccessary. */
4497 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4498 {
4499         int dest_cpu;
4500         cpumask_t mask;
4501
4502         /* On same node? */
4503         mask = node_to_cpumask(cpu_to_node(dead_cpu));
4504         cpus_and(mask, mask, tsk->cpus_allowed);
4505         dest_cpu = any_online_cpu(mask);
4506
4507         /* On any allowed CPU? */
4508         if (dest_cpu == NR_CPUS)
4509                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4510
4511         /* No more Mr. Nice Guy. */
4512         if (dest_cpu == NR_CPUS) {
4513                 cpus_setall(tsk->cpus_allowed);
4514                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4515
4516                 /*
4517                  * Don't tell them about moving exiting tasks or
4518                  * kernel threads (both mm NULL), since they never
4519                  * leave kernel.
4520                  */
4521                 if (tsk->mm && printk_ratelimit())
4522                         printk(KERN_INFO "process %d (%s) no "
4523                                "longer affine to cpu%d\n",
4524                                tsk->pid, tsk->comm, dead_cpu);
4525         }
4526         __migrate_task(tsk, dead_cpu, dest_cpu);
4527 }
4528
4529 /*
4530  * While a dead CPU has no uninterruptible tasks queued at this point,
4531  * it might still have a nonzero ->nr_uninterruptible counter, because
4532  * for performance reasons the counter is not stricly tracking tasks to
4533  * their home CPUs. So we just add the counter to another CPU's counter,
4534  * to keep the global sum constant after CPU-down:
4535  */
4536 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4537 {
4538         runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4539         unsigned long flags;
4540
4541         local_irq_save(flags);
4542         double_rq_lock(rq_src, rq_dest);
4543         rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4544         rq_src->nr_uninterruptible = 0;
4545         double_rq_unlock(rq_src, rq_dest);
4546         local_irq_restore(flags);
4547 }
4548
4549 /* Run through task list and migrate tasks from the dead cpu. */
4550 static void migrate_live_tasks(int src_cpu)
4551 {
4552         struct task_struct *tsk, *t;
4553
4554         write_lock_irq(&tasklist_lock);
4555
4556         do_each_thread(t, tsk) {
4557                 if (tsk == current)
4558                         continue;
4559
4560                 if (task_cpu(tsk) == src_cpu)
4561                         move_task_off_dead_cpu(src_cpu, tsk);
4562         } while_each_thread(t, tsk);
4563
4564         write_unlock_irq(&tasklist_lock);
4565 }
4566
4567 /* Schedules idle task to be the next runnable task on current CPU.
4568  * It does so by boosting its priority to highest possible and adding it to
4569  * the _front_ of runqueue. Used by CPU offline code.
4570  */
4571 void sched_idle_next(void)
4572 {
4573         int cpu = smp_processor_id();
4574         runqueue_t *rq = this_rq();
4575         struct task_struct *p = rq->idle;
4576         unsigned long flags;
4577
4578         /* cpu has to be offline */
4579         BUG_ON(cpu_online(cpu));
4580
4581         /* Strictly not necessary since rest of the CPUs are stopped by now
4582          * and interrupts disabled on current cpu.
4583          */
4584         spin_lock_irqsave(&rq->lock, flags);
4585
4586         __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4587         /* Add idle task to _front_ of it's priority queue */
4588         __activate_idle_task(p, rq);
4589
4590         spin_unlock_irqrestore(&rq->lock, flags);
4591 }
4592
4593 /* Ensures that the idle task is using init_mm right before its cpu goes
4594  * offline.
4595  */
4596 void idle_task_exit(void)
4597 {
4598         struct mm_struct *mm = current->active_mm;
4599
4600         BUG_ON(cpu_online(smp_processor_id()));
4601
4602         if (mm != &init_mm)
4603                 switch_mm(mm, &init_mm, current);
4604         mmdrop(mm);
4605 }
4606
4607 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4608 {
4609         struct runqueue *rq = cpu_rq(dead_cpu);
4610
4611         /* Must be exiting, otherwise would be on tasklist. */
4612         BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4613
4614         /* Cannot have done final schedule yet: would have vanished. */
4615         BUG_ON(tsk->flags & PF_DEAD);
4616
4617         get_task_struct(tsk);
4618
4619         /*
4620          * Drop lock around migration; if someone else moves it,
4621          * that's OK.  No task can be added to this CPU, so iteration is
4622          * fine.
4623          */
4624         spin_unlock_irq(&rq->lock);
4625         move_task_off_dead_cpu(dead_cpu, tsk);
4626         spin_lock_irq(&rq->lock);
4627
4628         put_task_struct(tsk);
4629 }
4630
4631 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4632 static void migrate_dead_tasks(unsigned int dead_cpu)
4633 {
4634         unsigned arr, i;
4635         struct runqueue *rq = cpu_rq(dead_cpu);
4636
4637         for (arr = 0; arr < 2; arr++) {
4638                 for (i = 0; i < MAX_PRIO; i++) {
4639                         struct list_head *list = &rq->arrays[arr].queue[i];
4640                         while (!list_empty(list))
4641                                 migrate_dead(dead_cpu,
4642                                              list_entry(list->next, task_t,
4643                                                         run_list));
4644                 }
4645         }
4646 }
4647 #endif /* CONFIG_HOTPLUG_CPU */
4648
4649 /*
4650  * migration_call - callback that gets triggered when a CPU is added.
4651  * Here we can start up the necessary migration thread for the new CPU.
4652  */
4653 static int migration_call(struct notifier_block *nfb, unsigned long action,
4654                           void *hcpu)
4655 {
4656         int cpu = (long)hcpu;
4657         struct task_struct *p;
4658         struct runqueue *rq;
4659         unsigned long flags;
4660
4661         switch (action) {
4662         case CPU_UP_PREPARE:
4663                 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4664                 if (IS_ERR(p))
4665                         return NOTIFY_BAD;
4666                 p->flags |= PF_NOFREEZE;
4667                 kthread_bind(p, cpu);
4668                 /* Must be high prio: stop_machine expects to yield to it. */
4669                 rq = task_rq_lock(p, &flags);
4670                 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4671                 task_rq_unlock(rq, &flags);
4672                 cpu_rq(cpu)->migration_thread = p;
4673                 break;
4674         case CPU_ONLINE:
4675                 /* Strictly unneccessary, as first user will wake it. */
4676                 wake_up_process(cpu_rq(cpu)->migration_thread);
4677                 break;
4678 #ifdef CONFIG_HOTPLUG_CPU
4679         case CPU_UP_CANCELED:
4680                 /* Unbind it from offline cpu so it can run.  Fall thru. */
4681                 kthread_bind(cpu_rq(cpu)->migration_thread,
4682                              any_online_cpu(cpu_online_map));
4683                 kthread_stop(cpu_rq(cpu)->migration_thread);
4684                 cpu_rq(cpu)->migration_thread = NULL;
4685                 break;
4686         case CPU_DEAD:
4687                 migrate_live_tasks(cpu);
4688                 rq = cpu_rq(cpu);
4689                 kthread_stop(rq->migration_thread);
4690                 rq->migration_thread = NULL;
4691                 /* Idle task back to normal (off runqueue, low prio) */
4692                 rq = task_rq_lock(rq->idle, &flags);
4693                 deactivate_task(rq->idle, rq);
4694                 rq->idle->static_prio = MAX_PRIO;
4695                 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4696                 migrate_dead_tasks(cpu);
4697                 task_rq_unlock(rq, &flags);
4698                 migrate_nr_uninterruptible(rq);
4699                 BUG_ON(rq->nr_running != 0);
4700
4701                 /* No need to migrate the tasks: it was best-effort if
4702                  * they didn't do lock_cpu_hotplug().  Just wake up
4703                  * the requestors. */
4704                 spin_lock_irq(&rq->lock);
4705                 while (!list_empty(&rq->migration_queue)) {
4706                         migration_req_t *req;
4707                         req = list_entry(rq->migration_queue.next,
4708                                          migration_req_t, list);
4709                         list_del_init(&req->list);
4710                         complete(&req->done);
4711                 }
4712                 spin_unlock_irq(&rq->lock);
4713                 break;
4714 #endif
4715         }
4716         return NOTIFY_OK;
4717 }
4718
4719 /* Register at highest priority so that task migration (migrate_all_tasks)
4720  * happens before everything else.
4721  */
4722 static struct notifier_block __devinitdata migration_notifier = {
4723         .notifier_call = migration_call,
4724         .priority = 10
4725 };
4726
4727 int __init migration_init(void)
4728 {
4729         void *cpu = (void *)(long)smp_processor_id();
4730         /* Start one for boot CPU. */
4731         migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4732         migration_call(&migration_notifier, CPU_ONLINE, cpu);
4733         register_cpu_notifier(&migration_notifier);
4734         return 0;
4735 }
4736 #endif
4737
4738 #ifdef CONFIG_SMP
4739 #undef SCHED_DOMAIN_DEBUG
4740 #ifdef SCHED_DOMAIN_DEBUG
4741 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4742 {
4743         int level = 0;
4744
4745         if (!sd) {
4746                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4747                 return;
4748         }
4749
4750         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4751
4752         do {
4753                 int i;
4754                 char str[NR_CPUS];
4755                 struct sched_group *group = sd->groups;
4756                 cpumask_t groupmask;
4757
4758                 cpumask_scnprintf(str, NR_CPUS, sd->span);
4759                 cpus_clear(groupmask);
4760
4761                 printk(KERN_DEBUG);
4762                 for (i = 0; i < level + 1; i++)
4763                         printk(" ");
4764                 printk("domain %d: ", level);
4765
4766                 if (!(sd->flags & SD_LOAD_BALANCE)) {
4767                         printk("does not load-balance\n");
4768                         if (sd->parent)
4769                                 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4770                         break;
4771                 }
4772
4773                 printk("span %s\n", str);
4774
4775                 if (!cpu_isset(cpu, sd->span))
4776                         printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4777                 if (!cpu_isset(cpu, group->cpumask))
4778                         printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4779
4780                 printk(KERN_DEBUG);
4781                 for (i = 0; i < level + 2; i++)
4782                         printk(" ");
4783                 printk("groups:");
4784                 do {
4785                         if (!group) {
4786                                 printk("\n");
4787                                 printk(KERN_ERR "ERROR: group is NULL\n");
4788                                 break;
4789                         }
4790
4791                         if (!group->cpu_power) {
4792                                 printk("\n");
4793                                 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4794                         }
4795
4796                         if (!cpus_weight(group->cpumask)) {
4797                                 printk("\n");
4798                                 printk(KERN_ERR "ERROR: empty group\n");
4799                         }
4800
4801                         if (cpus_intersects(groupmask, group->cpumask)) {
4802                                 printk("\n");
4803                                 printk(KERN_ERR "ERROR: repeated CPUs\n");
4804                         }
4805
4806                         cpus_or(groupmask, groupmask, group->cpumask);
4807
4808                         cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4809                         printk(" %s", str);
4810
4811                         group = group->next;
4812                 } while (group != sd->groups);
4813                 printk("\n");
4814
4815                 if (!cpus_equal(sd->span, groupmask))
4816                         printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4817
4818                 level++;
4819                 sd = sd->parent;
4820
4821                 if (sd) {
4822                         if (!cpus_subset(groupmask, sd->span))
4823                                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4824                 }
4825
4826         } while (sd);
4827 }
4828 #else
4829 #define sched_domain_debug(sd, cpu) {}
4830 #endif
4831
4832 static int sd_degenerate(struct sched_domain *sd)
4833 {
4834         if (cpus_weight(sd->span) == 1)
4835                 return 1;
4836
4837         /* Following flags need at least 2 groups */
4838         if (sd->flags & (SD_LOAD_BALANCE |
4839                          SD_BALANCE_NEWIDLE |
4840                          SD_BALANCE_FORK |
4841                          SD_BALANCE_EXEC)) {
4842                 if (sd->groups != sd->groups->next)
4843                         return 0;
4844         }
4845
4846         /* Following flags don't use groups */
4847         if (sd->flags & (SD_WAKE_IDLE |
4848                          SD_WAKE_AFFINE |
4849                          SD_WAKE_BALANCE))
4850                 return 0;
4851
4852         return 1;
4853 }
4854
4855 static int sd_parent_degenerate(struct sched_domain *sd,
4856                                                 struct sched_domain *parent)
4857 {
4858         unsigned long cflags = sd->flags, pflags = parent->flags;
4859
4860         if (sd_degenerate(parent))
4861                 return 1;
4862
4863         if (!cpus_equal(sd->span, parent->span))
4864                 return 0;
4865
4866         /* Does parent contain flags not in child? */
4867         /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4868         if (cflags & SD_WAKE_AFFINE)
4869                 pflags &= ~SD_WAKE_BALANCE;
4870         /* Flags needing groups don't count if only 1 group in parent */
4871         if (parent->groups == parent->groups->next) {
4872                 pflags &= ~(SD_LOAD_BALANCE |
4873                                 SD_BALANCE_NEWIDLE |
4874                                 SD_BALANCE_FORK |
4875                                 SD_BALANCE_EXEC);
4876         }
4877         if (~cflags & pflags)
4878                 return 0;
4879
4880         return 1;
4881 }
4882
4883 /*
4884  * Attach the domain 'sd' to 'cpu' as its base domain.  Callers must
4885  * hold the hotplug lock.
4886  */
4887 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4888 {
4889         runqueue_t *rq = cpu_rq(cpu);
4890         struct sched_domain *tmp;
4891
4892         /* Remove the sched domains which do not contribute to scheduling. */
4893         for (tmp = sd; tmp; tmp = tmp->parent) {
4894                 struct sched_domain *parent = tmp->parent;
4895                 if (!parent)
4896                         break;
4897                 if (sd_parent_degenerate(tmp, parent))
4898                         tmp->parent = parent->parent;
4899         }
4900
4901         if (sd && sd_degenerate(sd))
4902                 sd = sd->parent;
4903
4904         sched_domain_debug(sd, cpu);
4905
4906         rcu_assign_pointer(rq->sd, sd);
4907 }
4908
4909 /* cpus with isolated domains */
4910 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4911
4912 /* Setup the mask of cpus configured for isolated domains */
4913 static int __init isolated_cpu_setup(char *str)
4914 {
4915         int ints[NR_CPUS], i;
4916
4917         str = get_options(str, ARRAY_SIZE(ints), ints);
4918         cpus_clear(cpu_isolated_map);
4919         for (i = 1; i <= ints[0]; i++)
4920                 if (ints[i] < NR_CPUS)
4921                         cpu_set(ints[i], cpu_isolated_map);
4922         return 1;
4923 }
4924
4925 __setup ("isolcpus=", isolated_cpu_setup);
4926
4927 /*
4928  * init_sched_build_groups takes an array of groups, the cpumask we wish
4929  * to span, and a pointer to a function which identifies what group a CPU
4930  * belongs to. The return value of group_fn must be a valid index into the
4931  * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4932  * keep track of groups covered with a cpumask_t).
4933  *
4934  * init_sched_build_groups will build a circular linked list of the groups
4935  * covered by the given span, and will set each group's ->cpumask correctly,
4936  * and ->cpu_power to 0.
4937  */
4938 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
4939                                     int (*group_fn)(int cpu))
4940 {
4941         struct sched_group *first = NULL, *last = NULL;
4942         cpumask_t covered = CPU_MASK_NONE;
4943         int i;
4944
4945         for_each_cpu_mask(i, span) {
4946                 int group = group_fn(i);
4947                 struct sched_group *sg = &groups[group];
4948                 int j;
4949
4950                 if (cpu_isset(i, covered))
4951                         continue;
4952
4953                 sg->cpumask = CPU_MASK_NONE;
4954                 sg->cpu_power = 0;
4955
4956                 for_each_cpu_mask(j, span) {
4957                         if (group_fn(j) != group)
4958                                 continue;
4959
4960                         cpu_set(j, covered);
4961                         cpu_set(j, sg->cpumask);
4962                 }
4963                 if (!first)
4964                         first = sg;
4965                 if (last)
4966                         last->next = sg;
4967                 last = sg;
4968         }
4969         last->next = first;
4970 }
4971
4972 #define SD_NODES_PER_DOMAIN 16
4973
4974 #ifdef CONFIG_NUMA
4975 /**
4976  * find_next_best_node - find the next node to include in a sched_domain
4977  * @node: node whose sched_domain we're building
4978  * @used_nodes: nodes already in the sched_domain
4979  *
4980  * Find the next node to include in a given scheduling domain.  Simply
4981  * finds the closest node not already in the @used_nodes map.
4982  *
4983  * Should use nodemask_t.
4984  */
4985 static int find_next_best_node(int node, unsigned long *used_nodes)
4986 {
4987         int i, n, val, min_val, best_node = 0;
4988
4989         min_val = INT_MAX;
4990
4991         for (i = 0; i < MAX_NUMNODES; i++) {
4992                 /* Start at @node */
4993                 n = (node + i) % MAX_NUMNODES;
4994
4995                 if (!nr_cpus_node(n))
4996                         continue;
4997
4998                 /* Skip already used nodes */
4999                 if (test_bit(n, used_nodes))
5000                         continue;
5001
5002                 /* Simple min distance search */
5003                 val = node_distance(node, n);
5004
5005                 if (val < min_val) {
5006                         min_val = val;
5007                         best_node = n;
5008                 }
5009         }
5010
5011         set_bit(best_node, used_nodes);
5012         return best_node;
5013 }
5014
5015 /**
5016  * sched_domain_node_span - get a cpumask for a node's sched_domain
5017  * @node: node whose cpumask we're constructing
5018  * @size: number of nodes to include in this span
5019  *
5020  * Given a node, construct a good cpumask for its sched_domain to span.  It
5021  * should be one that prevents unnecessary balancing, but also spreads tasks
5022  * out optimally.
5023  */
5024 static cpumask_t sched_domain_node_span(int node)
5025 {
5026         int i;
5027         cpumask_t span, nodemask;
5028         DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5029
5030         cpus_clear(span);
5031         bitmap_zero(used_nodes, MAX_NUMNODES);
5032
5033         nodemask = node_to_cpumask(node);
5034         cpus_or(span, span, nodemask);
5035         set_bit(node, used_nodes);
5036
5037         for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5038                 int next_node = find_next_best_node(node, used_nodes);
5039                 nodemask = node_to_cpumask(next_node);
5040                 cpus_or(span, span, nodemask);
5041         }
5042
5043         return span;
5044 }
5045 #endif
5046
5047 /*
5048  * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5049  * can switch it on easily if needed.
5050  */
5051 #ifdef CONFIG_SCHED_SMT
5052 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5053 static struct sched_group sched_group_cpus[NR_CPUS];
5054 static int cpu_to_cpu_group(int cpu)
5055 {
5056         return cpu;
5057 }
5058 #endif
5059
5060 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5061 static struct sched_group sched_group_phys[NR_CPUS];
5062 static int cpu_to_phys_group(int cpu)
5063 {
5064 #ifdef CONFIG_SCHED_SMT
5065         return first_cpu(cpu_sibling_map[cpu]);
5066 #else
5067         return cpu;
5068 #endif
5069 }
5070
5071 #ifdef CONFIG_NUMA
5072 /*
5073  * The init_sched_build_groups can't handle what we want to do with node
5074  * groups, so roll our own. Now each node has its own list of groups which
5075  * gets dynamically allocated.
5076  */
5077 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5078 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5079
5080 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5081 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5082
5083 static int cpu_to_allnodes_group(int cpu)
5084 {
5085         return cpu_to_node(cpu);
5086 }
5087 #endif
5088
5089 /*
5090  * Build sched domains for a given set of cpus and attach the sched domains
5091  * to the individual cpus
5092  */
5093 void build_sched_domains(const cpumask_t *cpu_map)
5094 {
5095         int i;
5096 #ifdef CONFIG_NUMA
5097         struct sched_group **sched_group_nodes = NULL;
5098         struct sched_group *sched_group_allnodes = NULL;
5099
5100         /*
5101          * Allocate the per-node list of sched groups
5102          */
5103         sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5104                                            GFP_ATOMIC);
5105         if (!sched_group_nodes) {
5106                 printk(KERN_WARNING "Can not alloc sched group node list\n");
5107                 return;
5108         }
5109         sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5110 #endif
5111
5112         /*
5113          * Set up domains for cpus specified by the cpu_map.
5114          */
5115         for_each_cpu_mask(i, *cpu_map) {
5116                 int group;
5117                 struct sched_domain *sd = NULL, *p;
5118                 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5119
5120                 cpus_and(nodemask, nodemask, *cpu_map);
5121
5122 #ifdef CONFIG_NUMA
5123                 if (cpus_weight(*cpu_map)
5124                                 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5125                         if (!sched_group_allnodes) {
5126                                 sched_group_allnodes
5127                                         = kmalloc(sizeof(struct sched_group)
5128                                                         * MAX_NUMNODES,
5129                                                   GFP_KERNEL);
5130                                 if (!sched_group_allnodes) {
5131                                         printk(KERN_WARNING
5132                                         "Can not alloc allnodes sched group\n");
5133                                         break;
5134                                 }
5135                                 sched_group_allnodes_bycpu[i]
5136                                                 = sched_group_allnodes;
5137                         }
5138                         sd = &per_cpu(allnodes_domains, i);
5139                         *sd = SD_ALLNODES_INIT;
5140                         sd->span = *cpu_map;
5141                         group = cpu_to_allnodes_group(i);
5142                         sd->groups = &sched_group_allnodes[group];
5143                         p = sd;
5144                 } else
5145                         p = NULL;
5146
5147                 sd = &per_cpu(node_domains, i);
5148                 *sd = SD_NODE_INIT;
5149                 sd->span = sched_domain_node_span(cpu_to_node(i));
5150                 sd->parent = p;
5151                 cpus_and(sd->span, sd->span, *cpu_map);
5152 #endif
5153
5154                 p = sd;
5155                 sd = &per_cpu(phys_domains, i);
5156                 group = cpu_to_phys_group(i);
5157                 *sd = SD_CPU_INIT;
5158                 sd->span = nodemask;
5159                 sd->parent = p;
5160                 sd->groups = &sched_group_phys[group];
5161
5162 #ifdef CONFIG_SCHED_SMT
5163                 p = sd;
5164                 sd = &per_cpu(cpu_domains, i);
5165                 group = cpu_to_cpu_group(i);
5166                 *sd = SD_SIBLING_INIT;
5167                 sd->span = cpu_sibling_map[i];
5168                 cpus_and(sd->span, sd->span, *cpu_map);
5169                 sd->parent = p;
5170                 sd->groups = &sched_group_cpus[group];
5171 #endif
5172         }
5173
5174 #ifdef CONFIG_SCHED_SMT
5175         /* Set up CPU (sibling) groups */
5176         for_each_cpu_mask(i, *cpu_map) {
5177                 cpumask_t this_sibling_map = cpu_sibling_map[i];
5178                 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5179                 if (i != first_cpu(this_sibling_map))
5180                         continue;
5181
5182                 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5183                                                 &cpu_to_cpu_group);
5184         }
5185 #endif
5186
5187         /* Set up physical groups */
5188         for (i = 0; i < MAX_NUMNODES; i++) {
5189                 cpumask_t nodemask = node_to_cpumask(i);
5190
5191                 cpus_and(nodemask, nodemask, *cpu_map);
5192                 if (cpus_empty(nodemask))
5193                         continue;
5194
5195                 init_sched_build_groups(sched_group_phys, nodemask,
5196                                                 &cpu_to_phys_group);
5197         }
5198
5199 #ifdef CONFIG_NUMA
5200         /* Set up node groups */
5201         if (sched_group_allnodes)
5202                 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5203                                         &cpu_to_allnodes_group);
5204
5205         for (i = 0; i < MAX_NUMNODES; i++) {
5206                 /* Set up node groups */
5207                 struct sched_group *sg, *prev;
5208                 cpumask_t nodemask = node_to_cpumask(i);
5209                 cpumask_t domainspan;
5210                 cpumask_t covered = CPU_MASK_NONE;
5211                 int j;
5212
5213                 cpus_and(nodemask, nodemask, *cpu_map);
5214                 if (cpus_empty(nodemask)) {
5215                         sched_group_nodes[i] = NULL;
5216                         continue;
5217                 }
5218
5219                 domainspan = sched_domain_node_span(i);
5220                 cpus_and(domainspan, domainspan, *cpu_map);
5221
5222                 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5223                 sched_group_nodes[i] = sg;
5224                 for_each_cpu_mask(j, nodemask) {
5225                         struct sched_domain *sd;
5226                         sd = &per_cpu(node_domains, j);
5227                         sd->groups = sg;
5228                         if (sd->groups == NULL) {
5229                                 /* Turn off balancing if we have no groups */
5230                                 sd->flags = 0;
5231                         }
5232                 }
5233                 if (!sg) {
5234                         printk(KERN_WARNING
5235                         "Can not alloc domain group for node %d\n", i);
5236                         continue;
5237                 }
5238                 sg->cpu_power = 0;
5239                 sg->cpumask = nodemask;
5240                 cpus_or(covered, covered, nodemask);
5241                 prev = sg;
5242
5243                 for (j = 0; j < MAX_NUMNODES; j++) {
5244                         cpumask_t tmp, notcovered;
5245                         int n = (i + j) % MAX_NUMNODES;
5246
5247                         cpus_complement(notcovered, covered);
5248                         cpus_and(tmp, notcovered, *cpu_map);
5249                         cpus_and(tmp, tmp, domainspan);
5250                         if (cpus_empty(tmp))
5251                                 break;
5252
5253                         nodemask = node_to_cpumask(n);
5254                         cpus_and(tmp, tmp, nodemask);
5255                         if (cpus_empty(tmp))
5256                                 continue;
5257
5258                         sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5259                         if (!sg) {
5260                                 printk(KERN_WARNING
5261                                 "Can not alloc domain group for node %d\n", j);
5262                                 break;
5263                         }
5264                         sg->cpu_power = 0;
5265                         sg->cpumask = tmp;
5266                         cpus_or(covered, covered, tmp);
5267                         prev->next = sg;
5268                         prev = sg;
5269                 }
5270                 prev->next = sched_group_nodes[i];
5271         }
5272 #endif
5273
5274         /* Calculate CPU power for physical packages and nodes */
5275         for_each_cpu_mask(i, *cpu_map) {
5276                 int power;
5277                 struct sched_domain *sd;
5278 #ifdef CONFIG_SCHED_SMT
5279                 sd = &per_cpu(cpu_domains, i);
5280                 power = SCHED_LOAD_SCALE;
5281                 sd->groups->cpu_power = power;
5282 #endif
5283
5284                 sd = &per_cpu(phys_domains, i);
5285                 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5286                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5287                 sd->groups->cpu_power = power;
5288
5289 #ifdef CONFIG_NUMA
5290                 sd = &per_cpu(allnodes_domains, i);
5291                 if (sd->groups) {
5292                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5293                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5294                         sd->groups->cpu_power = power;
5295                 }
5296 #endif
5297         }
5298
5299 #ifdef CONFIG_NUMA
5300         for (i = 0; i < MAX_NUMNODES; i++) {
5301                 struct sched_group *sg = sched_group_nodes[i];
5302                 int j;
5303
5304                 if (sg == NULL)
5305                         continue;
5306 next_sg:
5307                 for_each_cpu_mask(j, sg->cpumask) {
5308                         struct sched_domain *sd;
5309                         int power;
5310
5311                         sd = &per_cpu(phys_domains, j);
5312                         if (j != first_cpu(sd->groups->cpumask)) {
5313                                 /*
5314                                  * Only add "power" once for each
5315                                  * physical package.
5316                                  */
5317                                 continue;
5318                         }
5319                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5320                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5321
5322                         sg->cpu_power += power;
5323                 }
5324                 sg = sg->next;
5325                 if (sg != sched_group_nodes[i])
5326                         goto next_sg;
5327         }
5328 #endif
5329
5330         /* Attach the domains */
5331         for_each_cpu_mask(i, *cpu_map) {
5332                 struct sched_domain *sd;
5333 #ifdef CONFIG_SCHED_SMT
5334                 sd = &per_cpu(cpu_domains, i);
5335 #else
5336                 sd = &per_cpu(phys_domains, i);
5337 #endif
5338                 cpu_attach_domain(sd, i);
5339         }
5340 }
5341 /*
5342  * Set up scheduler domains and groups.  Callers must hold the hotplug lock.
5343  */
5344 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5345 {
5346         cpumask_t cpu_default_map;
5347
5348         /*
5349          * Setup mask for cpus without special case scheduling requirements.
5350          * For now this just excludes isolated cpus, but could be used to
5351          * exclude other special cases in the future.
5352          */
5353         cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5354
5355         build_sched_domains(&cpu_default_map);
5356 }
5357
5358 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5359 {
5360 #ifdef CONFIG_NUMA
5361         int i;
5362         int cpu;
5363
5364         for_each_cpu_mask(cpu, *cpu_map) {
5365                 struct sched_group *sched_group_allnodes
5366                         = sched_group_allnodes_bycpu[cpu];
5367                 struct sched_group **sched_group_nodes
5368                         = sched_group_nodes_bycpu[cpu];
5369
5370                 if (sched_group_allnodes) {
5371                         kfree(sched_group_allnodes);
5372                         sched_group_allnodes_bycpu[cpu] = NULL;
5373                 }
5374
5375                 if (!sched_group_nodes)
5376                         continue;
5377
5378                 for (i = 0; i < MAX_NUMNODES; i++) {
5379                         cpumask_t nodemask = node_to_cpumask(i);
5380                         struct sched_group *oldsg, *sg = sched_group_nodes[i];
5381
5382                         cpus_and(nodemask, nodemask, *cpu_map);
5383                         if (cpus_empty(nodemask))
5384                                 continue;
5385
5386                         if (sg == NULL)
5387                                 continue;
5388                         sg = sg->next;
5389 next_sg:
5390                         oldsg = sg;
5391                         sg = sg->next;
5392                         kfree(oldsg);
5393                         if (oldsg != sched_group_nodes[i])
5394                                 goto next_sg;
5395                 }
5396                 kfree(sched_group_nodes);
5397                 sched_group_nodes_bycpu[cpu] = NULL;
5398         }
5399 #endif
5400 }
5401
5402 /*
5403  * Detach sched domains from a group of cpus specified in cpu_map
5404  * These cpus will now be attached to the NULL domain
5405  */
5406 static inline void detach_destroy_domains(const cpumask_t *cpu_map)
5407 {
5408         int i;
5409
5410         for_each_cpu_mask(i, *cpu_map)
5411                 cpu_attach_domain(NULL, i);
5412         synchronize_sched();
5413         arch_destroy_sched_domains(cpu_map);
5414 }
5415
5416 /*
5417  * Partition sched domains as specified by the cpumasks below.
5418  * This attaches all cpus from the cpumasks to the NULL domain,
5419  * waits for a RCU quiescent period, recalculates sched
5420  * domain information and then attaches them back to the
5421  * correct sched domains
5422  * Call with hotplug lock held
5423  */
5424 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
5425 {
5426         cpumask_t change_map;
5427
5428         cpus_and(*partition1, *partition1, cpu_online_map);
5429         cpus_and(*partition2, *partition2, cpu_online_map);
5430         cpus_or(change_map, *partition1, *partition2);
5431
5432         /* Detach sched domains from all of the affected cpus */
5433         detach_destroy_domains(&change_map);
5434         if (!cpus_empty(*partition1))
5435                 build_sched_domains(partition1);
5436         if (!cpus_empty(*partition2))
5437                 build_sched_domains(partition2);
5438 }
5439
5440 #ifdef CONFIG_HOTPLUG_CPU
5441 /*
5442  * Force a reinitialization of the sched domains hierarchy.  The domains
5443  * and groups cannot be updated in place without racing with the balancing
5444  * code, so we temporarily attach all running cpus to the NULL domain
5445  * which will prevent rebalancing while the sched domains are recalculated.
5446  */
5447 static int update_sched_domains(struct notifier_block *nfb,
5448                                 unsigned long action, void *hcpu)
5449 {
5450         switch (action) {
5451         case CPU_UP_PREPARE:
5452         case CPU_DOWN_PREPARE:
5453                 detach_destroy_domains(&cpu_online_map);
5454                 return NOTIFY_OK;
5455
5456         case CPU_UP_CANCELED:
5457         case CPU_DOWN_FAILED:
5458         case CPU_ONLINE:
5459         case CPU_DEAD:
5460                 /*
5461                  * Fall through and re-initialise the domains.
5462                  */
5463                 break;
5464         default:
5465                 return NOTIFY_DONE;
5466         }
5467
5468         /* The hotplug lock is already held by cpu_up/cpu_down */
5469         arch_init_sched_domains(&cpu_online_map);
5470
5471         return NOTIFY_OK;
5472 }
5473 #endif
5474
5475 void __init sched_init_smp(void)
5476 {
5477         lock_cpu_hotplug();
5478         arch_init_sched_domains(&cpu_online_map);
5479         unlock_cpu_hotplug();
5480         /* XXX: Theoretical race here - CPU may be hotplugged now */
5481         hotcpu_notifier(update_sched_domains, 0);
5482 }
5483 #else
5484 void __init sched_init_smp(void)
5485 {
5486 }
5487 #endif /* CONFIG_SMP */
5488
5489 int in_sched_functions(unsigned long addr)
5490 {
5491         /* Linker adds these: start and end of __sched functions */
5492         extern char __sched_text_start[], __sched_text_end[];
5493         return in_lock_functions(addr) ||
5494                 (addr >= (unsigned long)__sched_text_start
5495                 && addr < (unsigned long)__sched_text_end);
5496 }
5497
5498 void __init sched_init(void)
5499 {
5500         runqueue_t *rq;
5501         int i, j, k;
5502
5503         for (i = 0; i < NR_CPUS; i++) {
5504                 prio_array_t *array;
5505
5506                 rq = cpu_rq(i);
5507                 spin_lock_init(&rq->lock);
5508                 rq->nr_running = 0;
5509                 rq->active = rq->arrays;
5510                 rq->expired = rq->arrays + 1;
5511                 rq->best_expired_prio = MAX_PRIO;
5512
5513 #ifdef CONFIG_SMP
5514                 rq->sd = NULL;
5515                 for (j = 1; j < 3; j++)
5516                         rq->cpu_load[j] = 0;
5517                 rq->active_balance = 0;
5518                 rq->push_cpu = 0;
5519                 rq->migration_thread = NULL;
5520                 INIT_LIST_HEAD(&rq->migration_queue);
5521 #endif
5522                 atomic_set(&rq->nr_iowait, 0);
5523
5524                 for (j = 0; j < 2; j++) {
5525                         array = rq->arrays + j;
5526                         for (k = 0; k < MAX_PRIO; k++) {
5527                                 INIT_LIST_HEAD(array->queue + k);
5528                                 __clear_bit(k, array->bitmap);
5529                         }
5530                         // delimiter for bitsearch
5531                         __set_bit(MAX_PRIO, array->bitmap);
5532                 }
5533         }
5534
5535         /*
5536          * The boot idle thread does lazy MMU switching as well:
5537          */
5538         atomic_inc(&init_mm.mm_count);
5539         enter_lazy_tlb(&init_mm, current);
5540
5541         /*
5542          * Make us the idle thread. Technically, schedule() should not be
5543          * called from this thread, however somewhere below it might be,
5544          * but because we are the idle thread, we just pick up running again
5545          * when this runqueue becomes "idle".
5546          */
5547         init_idle(current, smp_processor_id());
5548 }
5549
5550 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5551 void __might_sleep(char *file, int line)
5552 {
5553 #if defined(in_atomic)
5554         static unsigned long prev_jiffy;        /* ratelimiting */
5555
5556         if ((in_atomic() || irqs_disabled()) &&
5557             system_state == SYSTEM_RUNNING && !oops_in_progress) {
5558                 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5559                         return;
5560                 prev_jiffy = jiffies;
5561                 printk(KERN_ERR "Debug: sleeping function called from invalid"
5562                                 " context at %s:%d\n", file, line);
5563                 printk("in_atomic():%d, irqs_disabled():%d\n",
5564                         in_atomic(), irqs_disabled());
5565                 dump_stack();
5566         }
5567 #endif
5568 }
5569 EXPORT_SYMBOL(__might_sleep);
5570 #endif
5571
5572 #ifdef CONFIG_MAGIC_SYSRQ
5573 void normalize_rt_tasks(void)
5574 {
5575         struct task_struct *p;
5576         prio_array_t *array;
5577         unsigned long flags;
5578         runqueue_t *rq;
5579
5580         read_lock_irq(&tasklist_lock);
5581         for_each_process (p) {
5582                 if (!rt_task(p))
5583                         continue;
5584
5585                 rq = task_rq_lock(p, &flags);
5586
5587                 array = p->array;
5588                 if (array)
5589                         deactivate_task(p, task_rq(p));
5590                 __setscheduler(p, SCHED_NORMAL, 0);
5591                 if (array) {
5592                         __activate_task(p, task_rq(p));
5593                         resched_task(rq->curr);
5594                 }
5595
5596                 task_rq_unlock(rq, &flags);
5597         }
5598         read_unlock_irq(&tasklist_lock);
5599 }
5600
5601 #endif /* CONFIG_MAGIC_SYSRQ */
5602
5603 #ifdef CONFIG_IA64
5604 /*
5605  * These functions are only useful for the IA64 MCA handling.
5606  *
5607  * They can only be called when the whole system has been
5608  * stopped - every CPU needs to be quiescent, and no scheduling
5609  * activity can take place. Using them for anything else would
5610  * be a serious bug, and as a result, they aren't even visible
5611  * under any other configuration.
5612  */
5613
5614 /**
5615  * curr_task - return the current task for a given cpu.
5616  * @cpu: the processor in question.
5617  *
5618  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5619  */
5620 task_t *curr_task(int cpu)
5621 {
5622         return cpu_curr(cpu);
5623 }
5624
5625 /**
5626  * set_curr_task - set the current task for a given cpu.
5627  * @cpu: the processor in question.
5628  * @p: the task pointer to set.
5629  *
5630  * Description: This function must only be used when non-maskable interrupts
5631  * are serviced on a separate stack.  It allows the architecture to switch the
5632  * notion of the current task on a cpu in a non-blocking manner.  This function
5633  * must be called with all CPU's synchronized, and interrupts disabled, the
5634  * and caller must save the original value of the current task (see
5635  * curr_task() above) and restore that value before reenabling interrupts and
5636  * re-starting the system.
5637  *
5638  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5639  */
5640 void set_curr_task(int cpu, task_t *p)
5641 {
5642         cpu_curr(cpu) = p;
5643 }
5644
5645 #endif