Merge with ../linux-2.6-smp
[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 inline 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;
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 #define for_each_domain(cpu, domain) \
264         for (domain = cpu_rq(cpu)->sd; domain; domain = domain->parent)
265
266 #define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
267 #define this_rq()               (&__get_cpu_var(runqueues))
268 #define task_rq(p)              cpu_rq(task_cpu(p))
269 #define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
270
271 /*
272  * Default context-switch locking:
273  */
274 #ifndef prepare_arch_switch
275 # define prepare_arch_switch(rq, next)  do { } while (0)
276 # define finish_arch_switch(rq, next)   spin_unlock_irq(&(rq)->lock)
277 # define task_running(rq, p)            ((rq)->curr == (p))
278 #endif
279
280 /*
281  * task_rq_lock - lock the runqueue a given task resides on and disable
282  * interrupts.  Note the ordering: we can safely lookup the task_rq without
283  * explicitly disabling preemption.
284  */
285 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
286         __acquires(rq->lock)
287 {
288         struct runqueue *rq;
289
290 repeat_lock_task:
291         local_irq_save(*flags);
292         rq = task_rq(p);
293         spin_lock(&rq->lock);
294         if (unlikely(rq != task_rq(p))) {
295                 spin_unlock_irqrestore(&rq->lock, *flags);
296                 goto repeat_lock_task;
297         }
298         return rq;
299 }
300
301 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
302         __releases(rq->lock)
303 {
304         spin_unlock_irqrestore(&rq->lock, *flags);
305 }
306
307 #ifdef CONFIG_SCHEDSTATS
308 /*
309  * bump this up when changing the output format or the meaning of an existing
310  * format, so that tools can adapt (or abort)
311  */
312 #define SCHEDSTAT_VERSION 11
313
314 static int show_schedstat(struct seq_file *seq, void *v)
315 {
316         int cpu;
317
318         seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
319         seq_printf(seq, "timestamp %lu\n", jiffies);
320         for_each_online_cpu(cpu) {
321                 runqueue_t *rq = cpu_rq(cpu);
322 #ifdef CONFIG_SMP
323                 struct sched_domain *sd;
324                 int dcnt = 0;
325 #endif
326
327                 /* runqueue-specific stats */
328                 seq_printf(seq,
329                     "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
330                     cpu, rq->yld_both_empty,
331                     rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
332                     rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
333                     rq->ttwu_cnt, rq->ttwu_local,
334                     rq->rq_sched_info.cpu_time,
335                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
336
337                 seq_printf(seq, "\n");
338
339 #ifdef CONFIG_SMP
340                 /* domain-specific stats */
341                 for_each_domain(cpu, sd) {
342                         enum idle_type itype;
343                         char mask_str[NR_CPUS];
344
345                         cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
346                         seq_printf(seq, "domain%d %s", dcnt++, mask_str);
347                         for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
348                                         itype++) {
349                                 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
350                                     sd->lb_cnt[itype],
351                                     sd->lb_balanced[itype],
352                                     sd->lb_failed[itype],
353                                     sd->lb_imbalance[itype],
354                                     sd->lb_gained[itype],
355                                     sd->lb_hot_gained[itype],
356                                     sd->lb_nobusyq[itype],
357                                     sd->lb_nobusyg[itype]);
358                         }
359                         seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu\n",
360                             sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
361                             sd->sbe_pushed, sd->sbe_attempts,
362                             sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
363                 }
364 #endif
365         }
366         return 0;
367 }
368
369 static int schedstat_open(struct inode *inode, struct file *file)
370 {
371         unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
372         char *buf = kmalloc(size, GFP_KERNEL);
373         struct seq_file *m;
374         int res;
375
376         if (!buf)
377                 return -ENOMEM;
378         res = single_open(file, show_schedstat, NULL);
379         if (!res) {
380                 m = file->private_data;
381                 m->buf = buf;
382                 m->size = size;
383         } else
384                 kfree(buf);
385         return res;
386 }
387
388 struct file_operations proc_schedstat_operations = {
389         .open    = schedstat_open,
390         .read    = seq_read,
391         .llseek  = seq_lseek,
392         .release = single_release,
393 };
394
395 # define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
396 # define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
397 #else /* !CONFIG_SCHEDSTATS */
398 # define schedstat_inc(rq, field)       do { } while (0)
399 # define schedstat_add(rq, field, amt)  do { } while (0)
400 #endif
401
402 /*
403  * rq_lock - lock a given runqueue and disable interrupts.
404  */
405 static inline runqueue_t *this_rq_lock(void)
406         __acquires(rq->lock)
407 {
408         runqueue_t *rq;
409
410         local_irq_disable();
411         rq = this_rq();
412         spin_lock(&rq->lock);
413
414         return rq;
415 }
416
417 #ifdef CONFIG_SCHED_SMT
418 static int cpu_and_siblings_are_idle(int cpu)
419 {
420         int sib;
421         for_each_cpu_mask(sib, cpu_sibling_map[cpu]) {
422                 if (idle_cpu(sib))
423                         continue;
424                 return 0;
425         }
426
427         return 1;
428 }
429 #else
430 #define cpu_and_siblings_are_idle(A) idle_cpu(A)
431 #endif
432
433 #ifdef CONFIG_SCHEDSTATS
434 /*
435  * Called when a process is dequeued from the active array and given
436  * the cpu.  We should note that with the exception of interactive
437  * tasks, the expired queue will become the active queue after the active
438  * queue is empty, without explicitly dequeuing and requeuing tasks in the
439  * expired queue.  (Interactive tasks may be requeued directly to the
440  * active queue, thus delaying tasks in the expired queue from running;
441  * see scheduler_tick()).
442  *
443  * This function is only called from sched_info_arrive(), rather than
444  * dequeue_task(). Even though a task may be queued and dequeued multiple
445  * times as it is shuffled about, we're really interested in knowing how
446  * long it was from the *first* time it was queued to the time that it
447  * finally hit a cpu.
448  */
449 static inline void sched_info_dequeued(task_t *t)
450 {
451         t->sched_info.last_queued = 0;
452 }
453
454 /*
455  * Called when a task finally hits the cpu.  We can now calculate how
456  * long it was waiting to run.  We also note when it began so that we
457  * can keep stats on how long its timeslice is.
458  */
459 static inline void sched_info_arrive(task_t *t)
460 {
461         unsigned long now = jiffies, diff = 0;
462         struct runqueue *rq = task_rq(t);
463
464         if (t->sched_info.last_queued)
465                 diff = now - t->sched_info.last_queued;
466         sched_info_dequeued(t);
467         t->sched_info.run_delay += diff;
468         t->sched_info.last_arrival = now;
469         t->sched_info.pcnt++;
470
471         if (!rq)
472                 return;
473
474         rq->rq_sched_info.run_delay += diff;
475         rq->rq_sched_info.pcnt++;
476 }
477
478 /*
479  * Called when a process is queued into either the active or expired
480  * array.  The time is noted and later used to determine how long we
481  * had to wait for us to reach the cpu.  Since the expired queue will
482  * become the active queue after active queue is empty, without dequeuing
483  * and requeuing any tasks, we are interested in queuing to either. It
484  * is unusual but not impossible for tasks to be dequeued and immediately
485  * requeued in the same or another array: this can happen in sched_yield(),
486  * set_user_nice(), and even load_balance() as it moves tasks from runqueue
487  * to runqueue.
488  *
489  * This function is only called from enqueue_task(), but also only updates
490  * the timestamp if it is already not set.  It's assumed that
491  * sched_info_dequeued() will clear that stamp when appropriate.
492  */
493 static inline void sched_info_queued(task_t *t)
494 {
495         if (!t->sched_info.last_queued)
496                 t->sched_info.last_queued = jiffies;
497 }
498
499 /*
500  * Called when a process ceases being the active-running process, either
501  * voluntarily or involuntarily.  Now we can calculate how long we ran.
502  */
503 static inline void sched_info_depart(task_t *t)
504 {
505         struct runqueue *rq = task_rq(t);
506         unsigned long diff = jiffies - t->sched_info.last_arrival;
507
508         t->sched_info.cpu_time += diff;
509
510         if (rq)
511                 rq->rq_sched_info.cpu_time += diff;
512 }
513
514 /*
515  * Called when tasks are switched involuntarily due, typically, to expiring
516  * their time slice.  (This may also be called when switching to or from
517  * the idle task.)  We are only called when prev != next.
518  */
519 static inline void sched_info_switch(task_t *prev, task_t *next)
520 {
521         struct runqueue *rq = task_rq(prev);
522
523         /*
524          * prev now departs the cpu.  It's not interesting to record
525          * stats about how efficient we were at scheduling the idle
526          * process, however.
527          */
528         if (prev != rq->idle)
529                 sched_info_depart(prev);
530
531         if (next != rq->idle)
532                 sched_info_arrive(next);
533 }
534 #else
535 #define sched_info_queued(t)            do { } while (0)
536 #define sched_info_switch(t, next)      do { } while (0)
537 #endif /* CONFIG_SCHEDSTATS */
538
539 /*
540  * Adding/removing a task to/from a priority array:
541  */
542 static void dequeue_task(struct task_struct *p, prio_array_t *array)
543 {
544         array->nr_active--;
545         list_del(&p->run_list);
546         if (list_empty(array->queue + p->prio))
547                 __clear_bit(p->prio, array->bitmap);
548 }
549
550 static void enqueue_task(struct task_struct *p, prio_array_t *array)
551 {
552         sched_info_queued(p);
553         list_add_tail(&p->run_list, array->queue + p->prio);
554         __set_bit(p->prio, array->bitmap);
555         array->nr_active++;
556         p->array = array;
557 }
558
559 /*
560  * Put task to the end of the run list without the overhead of dequeue
561  * followed by enqueue.
562  */
563 static void requeue_task(struct task_struct *p, prio_array_t *array)
564 {
565         list_move_tail(&p->run_list, array->queue + p->prio);
566 }
567
568 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
569 {
570         list_add(&p->run_list, array->queue + p->prio);
571         __set_bit(p->prio, array->bitmap);
572         array->nr_active++;
573         p->array = array;
574 }
575
576 /*
577  * effective_prio - return the priority that is based on the static
578  * priority but is modified by bonuses/penalties.
579  *
580  * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
581  * into the -5 ... 0 ... +5 bonus/penalty range.
582  *
583  * We use 25% of the full 0...39 priority range so that:
584  *
585  * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
586  * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
587  *
588  * Both properties are important to certain workloads.
589  */
590 static int effective_prio(task_t *p)
591 {
592         int bonus, prio;
593
594         if (rt_task(p))
595                 return p->prio;
596
597         bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
598
599         prio = p->static_prio - bonus;
600         if (prio < MAX_RT_PRIO)
601                 prio = MAX_RT_PRIO;
602         if (prio > MAX_PRIO-1)
603                 prio = MAX_PRIO-1;
604         return prio;
605 }
606
607 /*
608  * __activate_task - move a task to the runqueue.
609  */
610 static inline void __activate_task(task_t *p, runqueue_t *rq)
611 {
612         enqueue_task(p, rq->active);
613         rq->nr_running++;
614 }
615
616 /*
617  * __activate_idle_task - move idle task to the _front_ of runqueue.
618  */
619 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
620 {
621         enqueue_task_head(p, rq->active);
622         rq->nr_running++;
623 }
624
625 static void recalc_task_prio(task_t *p, unsigned long long now)
626 {
627         /* Caller must always ensure 'now >= p->timestamp' */
628         unsigned long long __sleep_time = now - p->timestamp;
629         unsigned long sleep_time;
630
631         if (__sleep_time > NS_MAX_SLEEP_AVG)
632                 sleep_time = NS_MAX_SLEEP_AVG;
633         else
634                 sleep_time = (unsigned long)__sleep_time;
635
636         if (likely(sleep_time > 0)) {
637                 /*
638                  * User tasks that sleep a long time are categorised as
639                  * idle and will get just interactive status to stay active &
640                  * prevent them suddenly becoming cpu hogs and starving
641                  * other processes.
642                  */
643                 if (p->mm && p->activated != -1 &&
644                         sleep_time > INTERACTIVE_SLEEP(p)) {
645                                 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
646                                                 DEF_TIMESLICE);
647                 } else {
648                         /*
649                          * The lower the sleep avg a task has the more
650                          * rapidly it will rise with sleep time.
651                          */
652                         sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
653
654                         /*
655                          * Tasks waking from uninterruptible sleep are
656                          * limited in their sleep_avg rise as they
657                          * are likely to be waiting on I/O
658                          */
659                         if (p->activated == -1 && p->mm) {
660                                 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
661                                         sleep_time = 0;
662                                 else if (p->sleep_avg + sleep_time >=
663                                                 INTERACTIVE_SLEEP(p)) {
664                                         p->sleep_avg = INTERACTIVE_SLEEP(p);
665                                         sleep_time = 0;
666                                 }
667                         }
668
669                         /*
670                          * This code gives a bonus to interactive tasks.
671                          *
672                          * The boost works by updating the 'average sleep time'
673                          * value here, based on ->timestamp. The more time a
674                          * task spends sleeping, the higher the average gets -
675                          * and the higher the priority boost gets as well.
676                          */
677                         p->sleep_avg += sleep_time;
678
679                         if (p->sleep_avg > NS_MAX_SLEEP_AVG)
680                                 p->sleep_avg = NS_MAX_SLEEP_AVG;
681                 }
682         }
683
684         p->prio = effective_prio(p);
685 }
686
687 /*
688  * activate_task - move a task to the runqueue and do priority recalculation
689  *
690  * Update all the scheduling statistics stuff. (sleep average
691  * calculation, priority modifiers, etc.)
692  */
693 static void activate_task(task_t *p, runqueue_t *rq, int local)
694 {
695         unsigned long long now;
696
697         now = sched_clock();
698 #ifdef CONFIG_SMP
699         if (!local) {
700                 /* Compensate for drifting sched_clock */
701                 runqueue_t *this_rq = this_rq();
702                 now = (now - this_rq->timestamp_last_tick)
703                         + rq->timestamp_last_tick;
704         }
705 #endif
706
707         recalc_task_prio(p, now);
708
709         /*
710          * This checks to make sure it's not an uninterruptible task
711          * that is now waking up.
712          */
713         if (!p->activated) {
714                 /*
715                  * Tasks which were woken up by interrupts (ie. hw events)
716                  * are most likely of interactive nature. So we give them
717                  * the credit of extending their sleep time to the period
718                  * of time they spend on the runqueue, waiting for execution
719                  * on a CPU, first time around:
720                  */
721                 if (in_interrupt())
722                         p->activated = 2;
723                 else {
724                         /*
725                          * Normal first-time wakeups get a credit too for
726                          * on-runqueue time, but it will be weighted down:
727                          */
728                         p->activated = 1;
729                 }
730         }
731         p->timestamp = now;
732
733         __activate_task(p, rq);
734 }
735
736 /*
737  * deactivate_task - remove a task from the runqueue.
738  */
739 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
740 {
741         rq->nr_running--;
742         dequeue_task(p, p->array);
743         p->array = NULL;
744 }
745
746 /*
747  * resched_task - mark a task 'to be rescheduled now'.
748  *
749  * On UP this means the setting of the need_resched flag, on SMP it
750  * might also involve a cross-CPU call to trigger the scheduler on
751  * the target CPU.
752  */
753 #ifdef CONFIG_SMP
754 static void resched_task(task_t *p)
755 {
756         int need_resched, nrpolling;
757
758         assert_spin_locked(&task_rq(p)->lock);
759
760         /* minimise the chance of sending an interrupt to poll_idle() */
761         nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
762         need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
763         nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
764
765         if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
766                 smp_send_reschedule(task_cpu(p));
767 }
768 #else
769 static inline void resched_task(task_t *p)
770 {
771         set_tsk_need_resched(p);
772 }
773 #endif
774
775 /**
776  * task_curr - is this task currently executing on a CPU?
777  * @p: the task in question.
778  */
779 inline int task_curr(const task_t *p)
780 {
781         return cpu_curr(task_cpu(p)) == p;
782 }
783
784 #ifdef CONFIG_SMP
785 enum request_type {
786         REQ_MOVE_TASK,
787         REQ_SET_DOMAIN,
788 };
789
790 typedef struct {
791         struct list_head list;
792         enum request_type type;
793
794         /* For REQ_MOVE_TASK */
795         task_t *task;
796         int dest_cpu;
797
798         /* For REQ_SET_DOMAIN */
799         struct sched_domain *sd;
800
801         struct completion done;
802 } migration_req_t;
803
804 /*
805  * The task's runqueue lock must be held.
806  * Returns true if you have to wait for migration thread.
807  */
808 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
809 {
810         runqueue_t *rq = task_rq(p);
811
812         /*
813          * If the task is not on a runqueue (and not running), then
814          * it is sufficient to simply update the task's cpu field.
815          */
816         if (!p->array && !task_running(rq, p)) {
817                 set_task_cpu(p, dest_cpu);
818                 return 0;
819         }
820
821         init_completion(&req->done);
822         req->type = REQ_MOVE_TASK;
823         req->task = p;
824         req->dest_cpu = dest_cpu;
825         list_add(&req->list, &rq->migration_queue);
826         return 1;
827 }
828
829 /*
830  * wait_task_inactive - wait for a thread to unschedule.
831  *
832  * The caller must ensure that the task *will* unschedule sometime soon,
833  * else this function might spin for a *long* time. This function can't
834  * be called with interrupts off, or it may introduce deadlock with
835  * smp_call_function() if an IPI is sent by the same process we are
836  * waiting to become inactive.
837  */
838 void wait_task_inactive(task_t * p)
839 {
840         unsigned long flags;
841         runqueue_t *rq;
842         int preempted;
843
844 repeat:
845         rq = task_rq_lock(p, &flags);
846         /* Must be off runqueue entirely, not preempted. */
847         if (unlikely(p->array || task_running(rq, p))) {
848                 /* If it's preempted, we yield.  It could be a while. */
849                 preempted = !task_running(rq, p);
850                 task_rq_unlock(rq, &flags);
851                 cpu_relax();
852                 if (preempted)
853                         yield();
854                 goto repeat;
855         }
856         task_rq_unlock(rq, &flags);
857 }
858
859 /***
860  * kick_process - kick a running thread to enter/exit the kernel
861  * @p: the to-be-kicked thread
862  *
863  * Cause a process which is running on another CPU to enter
864  * kernel-mode, without any delay. (to get signals handled.)
865  *
866  * NOTE: this function doesnt have to take the runqueue lock,
867  * because all it wants to ensure is that the remote task enters
868  * the kernel. If the IPI races and the task has been migrated
869  * to another CPU then no harm is done and the purpose has been
870  * achieved as well.
871  */
872 void kick_process(task_t *p)
873 {
874         int cpu;
875
876         preempt_disable();
877         cpu = task_cpu(p);
878         if ((cpu != smp_processor_id()) && task_curr(p))
879                 smp_send_reschedule(cpu);
880         preempt_enable();
881 }
882
883 /*
884  * Return a low guess at the load of a migration-source cpu.
885  *
886  * We want to under-estimate the load of migration sources, to
887  * balance conservatively.
888  */
889 static inline unsigned long source_load(int cpu)
890 {
891         runqueue_t *rq = cpu_rq(cpu);
892         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
893
894         return min(rq->cpu_load, load_now);
895 }
896
897 /*
898  * Return a high guess at the load of a migration-target cpu
899  */
900 static inline unsigned long target_load(int cpu)
901 {
902         runqueue_t *rq = cpu_rq(cpu);
903         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
904
905         return max(rq->cpu_load, load_now);
906 }
907
908 #endif
909
910 /*
911  * wake_idle() will wake a task on an idle cpu if task->cpu is
912  * not idle and an idle cpu is available.  The span of cpus to
913  * search starts with cpus closest then further out as needed,
914  * so we always favor a closer, idle cpu.
915  *
916  * Returns the CPU we should wake onto.
917  */
918 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
919 static int wake_idle(int cpu, task_t *p)
920 {
921         cpumask_t tmp;
922         struct sched_domain *sd;
923         int i;
924
925         if (idle_cpu(cpu))
926                 return cpu;
927
928         for_each_domain(cpu, sd) {
929                 if (sd->flags & SD_WAKE_IDLE) {
930                         cpus_and(tmp, sd->span, cpu_online_map);
931                         cpus_and(tmp, tmp, p->cpus_allowed);
932                         for_each_cpu_mask(i, tmp) {
933                                 if (idle_cpu(i))
934                                         return i;
935                         }
936                 }
937                 else break;
938         }
939         return cpu;
940 }
941 #else
942 static inline int wake_idle(int cpu, task_t *p)
943 {
944         return cpu;
945 }
946 #endif
947
948 /***
949  * try_to_wake_up - wake up a thread
950  * @p: the to-be-woken-up thread
951  * @state: the mask of task states that can be woken
952  * @sync: do a synchronous wakeup?
953  *
954  * Put it on the run-queue if it's not already there. The "current"
955  * thread is always on the run-queue (except when the actual
956  * re-schedule is in progress), and as such you're allowed to do
957  * the simpler "current->state = TASK_RUNNING" to mark yourself
958  * runnable without the overhead of this.
959  *
960  * returns failure only if the task is already active.
961  */
962 static int try_to_wake_up(task_t * p, unsigned int state, int sync)
963 {
964         int cpu, this_cpu, success = 0;
965         unsigned long flags;
966         long old_state;
967         runqueue_t *rq;
968 #ifdef CONFIG_SMP
969         unsigned long load, this_load;
970         struct sched_domain *sd;
971         int new_cpu;
972 #endif
973
974         rq = task_rq_lock(p, &flags);
975         old_state = p->state;
976         if (!(old_state & state))
977                 goto out;
978
979         if (p->array)
980                 goto out_running;
981
982         cpu = task_cpu(p);
983         this_cpu = smp_processor_id();
984
985 #ifdef CONFIG_SMP
986         if (unlikely(task_running(rq, p)))
987                 goto out_activate;
988
989 #ifdef CONFIG_SCHEDSTATS
990         schedstat_inc(rq, ttwu_cnt);
991         if (cpu == this_cpu) {
992                 schedstat_inc(rq, ttwu_local);
993         } else {
994                 for_each_domain(this_cpu, sd) {
995                         if (cpu_isset(cpu, sd->span)) {
996                                 schedstat_inc(sd, ttwu_wake_remote);
997                                 break;
998                         }
999                 }
1000         }
1001 #endif
1002
1003         new_cpu = cpu;
1004         if (cpu == this_cpu || unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1005                 goto out_set_cpu;
1006
1007         load = source_load(cpu);
1008         this_load = target_load(this_cpu);
1009
1010         /*
1011          * If sync wakeup then subtract the (maximum possible) effect of
1012          * the currently running task from the load of the current CPU:
1013          */
1014         if (sync)
1015                 this_load -= SCHED_LOAD_SCALE;
1016
1017         /* Don't pull the task off an idle CPU to a busy one */
1018         if (load < SCHED_LOAD_SCALE/2 && this_load > SCHED_LOAD_SCALE/2)
1019                 goto out_set_cpu;
1020
1021         new_cpu = this_cpu; /* Wake to this CPU if we can */
1022
1023         /*
1024          * Scan domains for affine wakeup and passive balancing
1025          * possibilities.
1026          */
1027         for_each_domain(this_cpu, sd) {
1028                 unsigned int imbalance;
1029                 /*
1030                  * Start passive balancing when half the imbalance_pct
1031                  * limit is reached.
1032                  */
1033                 imbalance = sd->imbalance_pct + (sd->imbalance_pct - 100) / 2;
1034
1035                 if ((sd->flags & SD_WAKE_AFFINE) &&
1036                                 !task_hot(p, rq->timestamp_last_tick, sd)) {
1037                         /*
1038                          * This domain has SD_WAKE_AFFINE and p is cache cold
1039                          * in this domain.
1040                          */
1041                         if (cpu_isset(cpu, sd->span)) {
1042                                 schedstat_inc(sd, ttwu_move_affine);
1043                                 goto out_set_cpu;
1044                         }
1045                 } else if ((sd->flags & SD_WAKE_BALANCE) &&
1046                                 imbalance*this_load <= 100*load) {
1047                         /*
1048                          * This domain has SD_WAKE_BALANCE and there is
1049                          * an imbalance.
1050                          */
1051                         if (cpu_isset(cpu, sd->span)) {
1052                                 schedstat_inc(sd, ttwu_move_balance);
1053                                 goto out_set_cpu;
1054                         }
1055                 }
1056         }
1057
1058         new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1059 out_set_cpu:
1060         new_cpu = wake_idle(new_cpu, p);
1061         if (new_cpu != cpu) {
1062                 set_task_cpu(p, new_cpu);
1063                 task_rq_unlock(rq, &flags);
1064                 /* might preempt at this point */
1065                 rq = task_rq_lock(p, &flags);
1066                 old_state = p->state;
1067                 if (!(old_state & state))
1068                         goto out;
1069                 if (p->array)
1070                         goto out_running;
1071
1072                 this_cpu = smp_processor_id();
1073                 cpu = task_cpu(p);
1074         }
1075
1076 out_activate:
1077 #endif /* CONFIG_SMP */
1078         if (old_state == TASK_UNINTERRUPTIBLE) {
1079                 rq->nr_uninterruptible--;
1080                 /*
1081                  * Tasks on involuntary sleep don't earn
1082                  * sleep_avg beyond just interactive state.
1083                  */
1084                 p->activated = -1;
1085         }
1086
1087         /*
1088          * Sync wakeups (i.e. those types of wakeups where the waker
1089          * has indicated that it will leave the CPU in short order)
1090          * don't trigger a preemption, if the woken up task will run on
1091          * this cpu. (in this case the 'I will reschedule' promise of
1092          * the waker guarantees that the freshly woken up task is going
1093          * to be considered on this CPU.)
1094          */
1095         activate_task(p, rq, cpu == this_cpu);
1096         if (!sync || cpu != this_cpu) {
1097                 if (TASK_PREEMPTS_CURR(p, rq))
1098                         resched_task(rq->curr);
1099         }
1100         success = 1;
1101
1102 out_running:
1103         p->state = TASK_RUNNING;
1104 out:
1105         task_rq_unlock(rq, &flags);
1106
1107         return success;
1108 }
1109
1110 int fastcall wake_up_process(task_t * p)
1111 {
1112         return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1113                                  TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1114 }
1115
1116 EXPORT_SYMBOL(wake_up_process);
1117
1118 int fastcall wake_up_state(task_t *p, unsigned int state)
1119 {
1120         return try_to_wake_up(p, state, 0);
1121 }
1122
1123 #ifdef CONFIG_SMP
1124 static int find_idlest_cpu(struct task_struct *p, int this_cpu,
1125                            struct sched_domain *sd);
1126 #endif
1127
1128 /*
1129  * Perform scheduler related setup for a newly forked process p.
1130  * p is forked by current.
1131  */
1132 void fastcall sched_fork(task_t *p)
1133 {
1134         /*
1135          * We mark the process as running here, but have not actually
1136          * inserted it onto the runqueue yet. This guarantees that
1137          * nobody will actually run it, and a signal or other external
1138          * event cannot wake it up and insert it on the runqueue either.
1139          */
1140         p->state = TASK_RUNNING;
1141         INIT_LIST_HEAD(&p->run_list);
1142         p->array = NULL;
1143         spin_lock_init(&p->switch_lock);
1144 #ifdef CONFIG_SCHEDSTATS
1145         memset(&p->sched_info, 0, sizeof(p->sched_info));
1146 #endif
1147 #ifdef CONFIG_PREEMPT
1148         /*
1149          * During context-switch we hold precisely one spinlock, which
1150          * schedule_tail drops. (in the common case it's this_rq()->lock,
1151          * but it also can be p->switch_lock.) So we compensate with a count
1152          * of 1. Also, we want to start with kernel preemption disabled.
1153          */
1154         p->thread_info->preempt_count = 1;
1155 #endif
1156         /*
1157          * Share the timeslice between parent and child, thus the
1158          * total amount of pending timeslices in the system doesn't change,
1159          * resulting in more scheduling fairness.
1160          */
1161         local_irq_disable();
1162         p->time_slice = (current->time_slice + 1) >> 1;
1163         /*
1164          * The remainder of the first timeslice might be recovered by
1165          * the parent if the child exits early enough.
1166          */
1167         p->first_time_slice = 1;
1168         current->time_slice >>= 1;
1169         p->timestamp = sched_clock();
1170         if (unlikely(!current->time_slice)) {
1171                 /*
1172                  * This case is rare, it happens when the parent has only
1173                  * a single jiffy left from its timeslice. Taking the
1174                  * runqueue lock is not a problem.
1175                  */
1176                 current->time_slice = 1;
1177                 preempt_disable();
1178                 scheduler_tick();
1179                 local_irq_enable();
1180                 preempt_enable();
1181         } else
1182                 local_irq_enable();
1183 }
1184
1185 /*
1186  * wake_up_new_task - wake up a newly created task for the first time.
1187  *
1188  * This function will do some initial scheduler statistics housekeeping
1189  * that must be done for every newly created context, then puts the task
1190  * on the runqueue and wakes it.
1191  */
1192 void fastcall wake_up_new_task(task_t * p, unsigned long clone_flags)
1193 {
1194         unsigned long flags;
1195         int this_cpu, cpu;
1196         runqueue_t *rq, *this_rq;
1197
1198         rq = task_rq_lock(p, &flags);
1199         cpu = task_cpu(p);
1200         this_cpu = smp_processor_id();
1201
1202         BUG_ON(p->state != TASK_RUNNING);
1203
1204         /*
1205          * We decrease the sleep average of forking parents
1206          * and children as well, to keep max-interactive tasks
1207          * from forking tasks that are max-interactive. The parent
1208          * (current) is done further down, under its lock.
1209          */
1210         p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1211                 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1212
1213         p->prio = effective_prio(p);
1214
1215         if (likely(cpu == this_cpu)) {
1216                 if (!(clone_flags & CLONE_VM)) {
1217                         /*
1218                          * The VM isn't cloned, so we're in a good position to
1219                          * do child-runs-first in anticipation of an exec. This
1220                          * usually avoids a lot of COW overhead.
1221                          */
1222                         if (unlikely(!current->array))
1223                                 __activate_task(p, rq);
1224                         else {
1225                                 p->prio = current->prio;
1226                                 list_add_tail(&p->run_list, &current->run_list);
1227                                 p->array = current->array;
1228                                 p->array->nr_active++;
1229                                 rq->nr_running++;
1230                         }
1231                         set_need_resched();
1232                 } else
1233                         /* Run child last */
1234                         __activate_task(p, rq);
1235                 /*
1236                  * We skip the following code due to cpu == this_cpu
1237                  *
1238                  *   task_rq_unlock(rq, &flags);
1239                  *   this_rq = task_rq_lock(current, &flags);
1240                  */
1241                 this_rq = rq;
1242         } else {
1243                 this_rq = cpu_rq(this_cpu);
1244
1245                 /*
1246                  * Not the local CPU - must adjust timestamp. This should
1247                  * get optimised away in the !CONFIG_SMP case.
1248                  */
1249                 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1250                                         + rq->timestamp_last_tick;
1251                 __activate_task(p, rq);
1252                 if (TASK_PREEMPTS_CURR(p, rq))
1253                         resched_task(rq->curr);
1254
1255                 /*
1256                  * Parent and child are on different CPUs, now get the
1257                  * parent runqueue to update the parent's ->sleep_avg:
1258                  */
1259                 task_rq_unlock(rq, &flags);
1260                 this_rq = task_rq_lock(current, &flags);
1261         }
1262         current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1263                 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1264         task_rq_unlock(this_rq, &flags);
1265 }
1266
1267 /*
1268  * Potentially available exiting-child timeslices are
1269  * retrieved here - this way the parent does not get
1270  * penalized for creating too many threads.
1271  *
1272  * (this cannot be used to 'generate' timeslices
1273  * artificially, because any timeslice recovered here
1274  * was given away by the parent in the first place.)
1275  */
1276 void fastcall sched_exit(task_t * p)
1277 {
1278         unsigned long flags;
1279         runqueue_t *rq;
1280
1281         /*
1282          * If the child was a (relative-) CPU hog then decrease
1283          * the sleep_avg of the parent as well.
1284          */
1285         rq = task_rq_lock(p->parent, &flags);
1286         if (p->first_time_slice) {
1287                 p->parent->time_slice += p->time_slice;
1288                 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1289                         p->parent->time_slice = task_timeslice(p);
1290         }
1291         if (p->sleep_avg < p->parent->sleep_avg)
1292                 p->parent->sleep_avg = p->parent->sleep_avg /
1293                 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1294                 (EXIT_WEIGHT + 1);
1295         task_rq_unlock(rq, &flags);
1296 }
1297
1298 /**
1299  * finish_task_switch - clean up after a task-switch
1300  * @prev: the thread we just switched away from.
1301  *
1302  * We enter this with the runqueue still locked, and finish_arch_switch()
1303  * will unlock it along with doing any other architecture-specific cleanup
1304  * actions.
1305  *
1306  * Note that we may have delayed dropping an mm in context_switch(). If
1307  * so, we finish that here outside of the runqueue lock.  (Doing it
1308  * with the lock held can cause deadlocks; see schedule() for
1309  * details.)
1310  */
1311 static inline void finish_task_switch(task_t *prev)
1312         __releases(rq->lock)
1313 {
1314         runqueue_t *rq = this_rq();
1315         struct mm_struct *mm = rq->prev_mm;
1316         unsigned long prev_task_flags;
1317
1318         rq->prev_mm = NULL;
1319
1320         /*
1321          * A task struct has one reference for the use as "current".
1322          * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1323          * calls schedule one last time. The schedule call will never return,
1324          * and the scheduled task must drop that reference.
1325          * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1326          * still held, otherwise prev could be scheduled on another cpu, die
1327          * there before we look at prev->state, and then the reference would
1328          * be dropped twice.
1329          *              Manfred Spraul <manfred@colorfullife.com>
1330          */
1331         prev_task_flags = prev->flags;
1332         finish_arch_switch(rq, prev);
1333         if (mm)
1334                 mmdrop(mm);
1335         if (unlikely(prev_task_flags & PF_DEAD))
1336                 put_task_struct(prev);
1337 }
1338
1339 /**
1340  * schedule_tail - first thing a freshly forked thread must call.
1341  * @prev: the thread we just switched away from.
1342  */
1343 asmlinkage void schedule_tail(task_t *prev)
1344         __releases(rq->lock)
1345 {
1346         finish_task_switch(prev);
1347
1348         if (current->set_child_tid)
1349                 put_user(current->pid, current->set_child_tid);
1350 }
1351
1352 /*
1353  * context_switch - switch to the new MM and the new
1354  * thread's register state.
1355  */
1356 static inline
1357 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1358 {
1359         struct mm_struct *mm = next->mm;
1360         struct mm_struct *oldmm = prev->active_mm;
1361
1362         if (unlikely(!mm)) {
1363                 next->active_mm = oldmm;
1364                 atomic_inc(&oldmm->mm_count);
1365                 enter_lazy_tlb(oldmm, next);
1366         } else
1367                 switch_mm(oldmm, mm, next);
1368
1369         if (unlikely(!prev->mm)) {
1370                 prev->active_mm = NULL;
1371                 WARN_ON(rq->prev_mm);
1372                 rq->prev_mm = oldmm;
1373         }
1374
1375         /* Here we just switch the register state and the stack. */
1376         switch_to(prev, next, prev);
1377
1378         return prev;
1379 }
1380
1381 /*
1382  * nr_running, nr_uninterruptible and nr_context_switches:
1383  *
1384  * externally visible scheduler statistics: current number of runnable
1385  * threads, current number of uninterruptible-sleeping threads, total
1386  * number of context switches performed since bootup.
1387  */
1388 unsigned long nr_running(void)
1389 {
1390         unsigned long i, sum = 0;
1391
1392         for_each_online_cpu(i)
1393                 sum += cpu_rq(i)->nr_running;
1394
1395         return sum;
1396 }
1397
1398 unsigned long nr_uninterruptible(void)
1399 {
1400         unsigned long i, sum = 0;
1401
1402         for_each_cpu(i)
1403                 sum += cpu_rq(i)->nr_uninterruptible;
1404
1405         /*
1406          * Since we read the counters lockless, it might be slightly
1407          * inaccurate. Do not allow it to go below zero though:
1408          */
1409         if (unlikely((long)sum < 0))
1410                 sum = 0;
1411
1412         return sum;
1413 }
1414
1415 unsigned long long nr_context_switches(void)
1416 {
1417         unsigned long long i, sum = 0;
1418
1419         for_each_cpu(i)
1420                 sum += cpu_rq(i)->nr_switches;
1421
1422         return sum;
1423 }
1424
1425 unsigned long nr_iowait(void)
1426 {
1427         unsigned long i, sum = 0;
1428
1429         for_each_cpu(i)
1430                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1431
1432         return sum;
1433 }
1434
1435 #ifdef CONFIG_SMP
1436
1437 /*
1438  * double_rq_lock - safely lock two runqueues
1439  *
1440  * Note this does not disable interrupts like task_rq_lock,
1441  * you need to do so manually before calling.
1442  */
1443 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1444         __acquires(rq1->lock)
1445         __acquires(rq2->lock)
1446 {
1447         if (rq1 == rq2) {
1448                 spin_lock(&rq1->lock);
1449                 __acquire(rq2->lock);   /* Fake it out ;) */
1450         } else {
1451                 if (rq1 < rq2) {
1452                         spin_lock(&rq1->lock);
1453                         spin_lock(&rq2->lock);
1454                 } else {
1455                         spin_lock(&rq2->lock);
1456                         spin_lock(&rq1->lock);
1457                 }
1458         }
1459 }
1460
1461 /*
1462  * double_rq_unlock - safely unlock two runqueues
1463  *
1464  * Note this does not restore interrupts like task_rq_unlock,
1465  * you need to do so manually after calling.
1466  */
1467 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1468         __releases(rq1->lock)
1469         __releases(rq2->lock)
1470 {
1471         spin_unlock(&rq1->lock);
1472         if (rq1 != rq2)
1473                 spin_unlock(&rq2->lock);
1474         else
1475                 __release(rq2->lock);
1476 }
1477
1478 /*
1479  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1480  */
1481 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1482         __releases(this_rq->lock)
1483         __acquires(busiest->lock)
1484         __acquires(this_rq->lock)
1485 {
1486         if (unlikely(!spin_trylock(&busiest->lock))) {
1487                 if (busiest < this_rq) {
1488                         spin_unlock(&this_rq->lock);
1489                         spin_lock(&busiest->lock);
1490                         spin_lock(&this_rq->lock);
1491                 } else
1492                         spin_lock(&busiest->lock);
1493         }
1494 }
1495
1496 /*
1497  * find_idlest_cpu - find the least busy runqueue.
1498  */
1499 static int find_idlest_cpu(struct task_struct *p, int this_cpu,
1500                            struct sched_domain *sd)
1501 {
1502         unsigned long load, min_load, this_load;
1503         int i, min_cpu;
1504         cpumask_t mask;
1505
1506         min_cpu = UINT_MAX;
1507         min_load = ULONG_MAX;
1508
1509         cpus_and(mask, sd->span, p->cpus_allowed);
1510
1511         for_each_cpu_mask(i, mask) {
1512                 load = target_load(i);
1513
1514                 if (load < min_load) {
1515                         min_cpu = i;
1516                         min_load = load;
1517
1518                         /* break out early on an idle CPU: */
1519                         if (!min_load)
1520                                 break;
1521                 }
1522         }
1523
1524         /* add +1 to account for the new task */
1525         this_load = source_load(this_cpu) + SCHED_LOAD_SCALE;
1526
1527         /*
1528          * Would with the addition of the new task to the
1529          * current CPU there be an imbalance between this
1530          * CPU and the idlest CPU?
1531          *
1532          * Use half of the balancing threshold - new-context is
1533          * a good opportunity to balance.
1534          */
1535         if (min_load*(100 + (sd->imbalance_pct-100)/2) < this_load*100)
1536                 return min_cpu;
1537
1538         return this_cpu;
1539 }
1540
1541 /*
1542  * If dest_cpu is allowed for this process, migrate the task to it.
1543  * This is accomplished by forcing the cpu_allowed mask to only
1544  * allow dest_cpu, which will force the cpu onto dest_cpu.  Then
1545  * the cpu_allowed mask is restored.
1546  */
1547 static void sched_migrate_task(task_t *p, int dest_cpu)
1548 {
1549         migration_req_t req;
1550         runqueue_t *rq;
1551         unsigned long flags;
1552
1553         rq = task_rq_lock(p, &flags);
1554         if (!cpu_isset(dest_cpu, p->cpus_allowed)
1555             || unlikely(cpu_is_offline(dest_cpu)))
1556                 goto out;
1557
1558         /* force the process onto the specified CPU */
1559         if (migrate_task(p, dest_cpu, &req)) {
1560                 /* Need to wait for migration thread (might exit: take ref). */
1561                 struct task_struct *mt = rq->migration_thread;
1562                 get_task_struct(mt);
1563                 task_rq_unlock(rq, &flags);
1564                 wake_up_process(mt);
1565                 put_task_struct(mt);
1566                 wait_for_completion(&req.done);
1567                 return;
1568         }
1569 out:
1570         task_rq_unlock(rq, &flags);
1571 }
1572
1573 /*
1574  * sched_exec(): find the highest-level, exec-balance-capable
1575  * domain and try to migrate the task to the least loaded CPU.
1576  *
1577  * execve() is a valuable balancing opportunity, because at this point
1578  * the task has the smallest effective memory and cache footprint.
1579  */
1580 void sched_exec(void)
1581 {
1582         struct sched_domain *tmp, *sd = NULL;
1583         int new_cpu, this_cpu = get_cpu();
1584
1585         /* Prefer the current CPU if there's only this task running */
1586         if (this_rq()->nr_running <= 1)
1587                 goto out;
1588
1589         for_each_domain(this_cpu, tmp)
1590                 if (tmp->flags & SD_BALANCE_EXEC)
1591                         sd = tmp;
1592
1593         if (sd) {
1594                 schedstat_inc(sd, sbe_attempts);
1595                 new_cpu = find_idlest_cpu(current, this_cpu, sd);
1596                 if (new_cpu != this_cpu) {
1597                         schedstat_inc(sd, sbe_pushed);
1598                         put_cpu();
1599                         sched_migrate_task(current, new_cpu);
1600                         return;
1601                 }
1602         }
1603 out:
1604         put_cpu();
1605 }
1606
1607 /*
1608  * pull_task - move a task from a remote runqueue to the local runqueue.
1609  * Both runqueues must be locked.
1610  */
1611 static inline
1612 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1613                runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1614 {
1615         dequeue_task(p, src_array);
1616         src_rq->nr_running--;
1617         set_task_cpu(p, this_cpu);
1618         this_rq->nr_running++;
1619         enqueue_task(p, this_array);
1620         p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1621                                 + this_rq->timestamp_last_tick;
1622         /*
1623          * Note that idle threads have a prio of MAX_PRIO, for this test
1624          * to be always true for them.
1625          */
1626         if (TASK_PREEMPTS_CURR(p, this_rq))
1627                 resched_task(this_rq->curr);
1628 }
1629
1630 /*
1631  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1632  */
1633 static inline
1634 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1635                      struct sched_domain *sd, enum idle_type idle)
1636 {
1637         /*
1638          * We do not migrate tasks that are:
1639          * 1) running (obviously), or
1640          * 2) cannot be migrated to this CPU due to cpus_allowed, or
1641          * 3) are cache-hot on their current CPU.
1642          */
1643         if (task_running(rq, p))
1644                 return 0;
1645         if (!cpu_isset(this_cpu, p->cpus_allowed))
1646                 return 0;
1647
1648         /*
1649          * Aggressive migration if:
1650          * 1) the [whole] cpu is idle, or
1651          * 2) too many balance attempts have failed.
1652          */
1653
1654         if (cpu_and_siblings_are_idle(this_cpu) || \
1655                         sd->nr_balance_failed > sd->cache_nice_tries)
1656                 return 1;
1657
1658         if (task_hot(p, rq->timestamp_last_tick, sd))
1659                         return 0;
1660         return 1;
1661 }
1662
1663 /*
1664  * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1665  * as part of a balancing operation within "domain". Returns the number of
1666  * tasks moved.
1667  *
1668  * Called with both runqueues locked.
1669  */
1670 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1671                       unsigned long max_nr_move, struct sched_domain *sd,
1672                       enum idle_type idle)
1673 {
1674         prio_array_t *array, *dst_array;
1675         struct list_head *head, *curr;
1676         int idx, pulled = 0;
1677         task_t *tmp;
1678
1679         if (max_nr_move <= 0 || busiest->nr_running <= 1)
1680                 goto out;
1681
1682         /*
1683          * We first consider expired tasks. Those will likely not be
1684          * executed in the near future, and they are most likely to
1685          * be cache-cold, thus switching CPUs has the least effect
1686          * on them.
1687          */
1688         if (busiest->expired->nr_active) {
1689                 array = busiest->expired;
1690                 dst_array = this_rq->expired;
1691         } else {
1692                 array = busiest->active;
1693                 dst_array = this_rq->active;
1694         }
1695
1696 new_array:
1697         /* Start searching at priority 0: */
1698         idx = 0;
1699 skip_bitmap:
1700         if (!idx)
1701                 idx = sched_find_first_bit(array->bitmap);
1702         else
1703                 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1704         if (idx >= MAX_PRIO) {
1705                 if (array == busiest->expired && busiest->active->nr_active) {
1706                         array = busiest->active;
1707                         dst_array = this_rq->active;
1708                         goto new_array;
1709                 }
1710                 goto out;
1711         }
1712
1713         head = array->queue + idx;
1714         curr = head->prev;
1715 skip_queue:
1716         tmp = list_entry(curr, task_t, run_list);
1717
1718         curr = curr->prev;
1719
1720         if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle)) {
1721                 if (curr != head)
1722                         goto skip_queue;
1723                 idx++;
1724                 goto skip_bitmap;
1725         }
1726
1727 #ifdef CONFIG_SCHEDSTATS
1728         if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1729                 schedstat_inc(sd, lb_hot_gained[idle]);
1730 #endif
1731
1732         pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1733         pulled++;
1734
1735         /* We only want to steal up to the prescribed number of tasks. */
1736         if (pulled < max_nr_move) {
1737                 if (curr != head)
1738                         goto skip_queue;
1739                 idx++;
1740                 goto skip_bitmap;
1741         }
1742 out:
1743         /*
1744          * Right now, this is the only place pull_task() is called,
1745          * so we can safely collect pull_task() stats here rather than
1746          * inside pull_task().
1747          */
1748         schedstat_add(sd, lb_gained[idle], pulled);
1749         return pulled;
1750 }
1751
1752 /*
1753  * find_busiest_group finds and returns the busiest CPU group within the
1754  * domain. It calculates and returns the number of tasks which should be
1755  * moved to restore balance via the imbalance parameter.
1756  */
1757 static struct sched_group *
1758 find_busiest_group(struct sched_domain *sd, int this_cpu,
1759                    unsigned long *imbalance, enum idle_type idle)
1760 {
1761         struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1762         unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1763
1764         max_load = this_load = total_load = total_pwr = 0;
1765
1766         do {
1767                 unsigned long load;
1768                 int local_group;
1769                 int i;
1770
1771                 local_group = cpu_isset(this_cpu, group->cpumask);
1772
1773                 /* Tally up the load of all CPUs in the group */
1774                 avg_load = 0;
1775
1776                 for_each_cpu_mask(i, group->cpumask) {
1777                         /* Bias balancing toward cpus of our domain */
1778                         if (local_group)
1779                                 load = target_load(i);
1780                         else
1781                                 load = source_load(i);
1782
1783                         avg_load += load;
1784                 }
1785
1786                 total_load += avg_load;
1787                 total_pwr += group->cpu_power;
1788
1789                 /* Adjust by relative CPU power of the group */
1790                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1791
1792                 if (local_group) {
1793                         this_load = avg_load;
1794                         this = group;
1795                         goto nextgroup;
1796                 } else if (avg_load > max_load) {
1797                         max_load = avg_load;
1798                         busiest = group;
1799                 }
1800 nextgroup:
1801                 group = group->next;
1802         } while (group != sd->groups);
1803
1804         if (!busiest || this_load >= max_load)
1805                 goto out_balanced;
1806
1807         avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1808
1809         if (this_load >= avg_load ||
1810                         100*max_load <= sd->imbalance_pct*this_load)
1811                 goto out_balanced;
1812
1813         /*
1814          * We're trying to get all the cpus to the average_load, so we don't
1815          * want to push ourselves above the average load, nor do we wish to
1816          * reduce the max loaded cpu below the average load, as either of these
1817          * actions would just result in more rebalancing later, and ping-pong
1818          * tasks around. Thus we look for the minimum possible imbalance.
1819          * Negative imbalances (*we* are more loaded than anyone else) will
1820          * be counted as no imbalance for these purposes -- we can't fix that
1821          * by pulling tasks to us.  Be careful of negative numbers as they'll
1822          * appear as very large values with unsigned longs.
1823          */
1824         /* How much load to actually move to equalise the imbalance */
1825         *imbalance = min((max_load - avg_load) * busiest->cpu_power,
1826                                 (avg_load - this_load) * this->cpu_power)
1827                         / SCHED_LOAD_SCALE;
1828
1829         if (*imbalance < SCHED_LOAD_SCALE) {
1830                 unsigned long pwr_now = 0, pwr_move = 0;
1831                 unsigned long tmp;
1832
1833                 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1834                         *imbalance = 1;
1835                         return busiest;
1836                 }
1837
1838                 /*
1839                  * OK, we don't have enough imbalance to justify moving tasks,
1840                  * however we may be able to increase total CPU power used by
1841                  * moving them.
1842                  */
1843
1844                 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
1845                 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
1846                 pwr_now /= SCHED_LOAD_SCALE;
1847
1848                 /* Amount of load we'd subtract */
1849                 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
1850                 if (max_load > tmp)
1851                         pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
1852                                                         max_load - tmp);
1853
1854                 /* Amount of load we'd add */
1855                 if (max_load*busiest->cpu_power <
1856                                 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
1857                         tmp = max_load*busiest->cpu_power/this->cpu_power;
1858                 else
1859                         tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
1860                 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
1861                 pwr_move /= SCHED_LOAD_SCALE;
1862
1863                 /* Move if we gain throughput */
1864                 if (pwr_move <= pwr_now)
1865                         goto out_balanced;
1866
1867                 *imbalance = 1;
1868                 return busiest;
1869         }
1870
1871         /* Get rid of the scaling factor, rounding down as we divide */
1872         *imbalance = *imbalance / SCHED_LOAD_SCALE;
1873
1874         return busiest;
1875
1876 out_balanced:
1877         if (busiest && (idle == NEWLY_IDLE ||
1878                         (idle == SCHED_IDLE && max_load > SCHED_LOAD_SCALE)) ) {
1879                 *imbalance = 1;
1880                 return busiest;
1881         }
1882
1883         *imbalance = 0;
1884         return NULL;
1885 }
1886
1887 /*
1888  * find_busiest_queue - find the busiest runqueue among the cpus in group.
1889  */
1890 static runqueue_t *find_busiest_queue(struct sched_group *group)
1891 {
1892         unsigned long load, max_load = 0;
1893         runqueue_t *busiest = NULL;
1894         int i;
1895
1896         for_each_cpu_mask(i, group->cpumask) {
1897                 load = source_load(i);
1898
1899                 if (load > max_load) {
1900                         max_load = load;
1901                         busiest = cpu_rq(i);
1902                 }
1903         }
1904
1905         return busiest;
1906 }
1907
1908 /*
1909  * Check this_cpu to ensure it is balanced within domain. Attempt to move
1910  * tasks if there is an imbalance.
1911  *
1912  * Called with this_rq unlocked.
1913  */
1914 static int load_balance(int this_cpu, runqueue_t *this_rq,
1915                         struct sched_domain *sd, enum idle_type idle)
1916 {
1917         struct sched_group *group;
1918         runqueue_t *busiest;
1919         unsigned long imbalance;
1920         int nr_moved;
1921
1922         spin_lock(&this_rq->lock);
1923         schedstat_inc(sd, lb_cnt[idle]);
1924
1925         group = find_busiest_group(sd, this_cpu, &imbalance, idle);
1926         if (!group) {
1927                 schedstat_inc(sd, lb_nobusyg[idle]);
1928                 goto out_balanced;
1929         }
1930
1931         busiest = find_busiest_queue(group);
1932         if (!busiest) {
1933                 schedstat_inc(sd, lb_nobusyq[idle]);
1934                 goto out_balanced;
1935         }
1936
1937         /*
1938          * This should be "impossible", but since load
1939          * balancing is inherently racy and statistical,
1940          * it could happen in theory.
1941          */
1942         if (unlikely(busiest == this_rq)) {
1943                 WARN_ON(1);
1944                 goto out_balanced;
1945         }
1946
1947         schedstat_add(sd, lb_imbalance[idle], imbalance);
1948
1949         nr_moved = 0;
1950         if (busiest->nr_running > 1) {
1951                 /*
1952                  * Attempt to move tasks. If find_busiest_group has found
1953                  * an imbalance but busiest->nr_running <= 1, the group is
1954                  * still unbalanced. nr_moved simply stays zero, so it is
1955                  * correctly treated as an imbalance.
1956                  */
1957                 double_lock_balance(this_rq, busiest);
1958                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
1959                                                 imbalance, sd, idle);
1960                 spin_unlock(&busiest->lock);
1961         }
1962         spin_unlock(&this_rq->lock);
1963
1964         if (!nr_moved) {
1965                 schedstat_inc(sd, lb_failed[idle]);
1966                 sd->nr_balance_failed++;
1967
1968                 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
1969                         int wake = 0;
1970
1971                         spin_lock(&busiest->lock);
1972                         if (!busiest->active_balance) {
1973                                 busiest->active_balance = 1;
1974                                 busiest->push_cpu = this_cpu;
1975                                 wake = 1;
1976                         }
1977                         spin_unlock(&busiest->lock);
1978                         if (wake)
1979                                 wake_up_process(busiest->migration_thread);
1980
1981                         /*
1982                          * We've kicked active balancing, reset the failure
1983                          * counter.
1984                          */
1985                         sd->nr_balance_failed = sd->cache_nice_tries;
1986                 }
1987
1988                 /*
1989                  * We were unbalanced, but unsuccessful in move_tasks(),
1990                  * so bump the balance_interval to lessen the lock contention.
1991                  */
1992                 if (sd->balance_interval < sd->max_interval)
1993                         sd->balance_interval++;
1994         } else {
1995                 sd->nr_balance_failed = 0;
1996
1997                 /* We were unbalanced, so reset the balancing interval */
1998                 sd->balance_interval = sd->min_interval;
1999         }
2000
2001         return nr_moved;
2002
2003 out_balanced:
2004         spin_unlock(&this_rq->lock);
2005
2006         schedstat_inc(sd, lb_balanced[idle]);
2007
2008         /* tune up the balancing interval */
2009         if (sd->balance_interval < sd->max_interval)
2010                 sd->balance_interval *= 2;
2011
2012         return 0;
2013 }
2014
2015 /*
2016  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2017  * tasks if there is an imbalance.
2018  *
2019  * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2020  * this_rq is locked.
2021  */
2022 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2023                                 struct sched_domain *sd)
2024 {
2025         struct sched_group *group;
2026         runqueue_t *busiest = NULL;
2027         unsigned long imbalance;
2028         int nr_moved = 0;
2029
2030         schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2031         group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);
2032         if (!group) {
2033                 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2034                 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2035                 goto out;
2036         }
2037
2038         busiest = find_busiest_queue(group);
2039         if (!busiest || busiest == this_rq) {
2040                 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2041                 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2042                 goto out;
2043         }
2044
2045         /* Attempt to move tasks */
2046         double_lock_balance(this_rq, busiest);
2047
2048         schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2049         nr_moved = move_tasks(this_rq, this_cpu, busiest,
2050                                         imbalance, sd, NEWLY_IDLE);
2051         if (!nr_moved)
2052                 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2053
2054         spin_unlock(&busiest->lock);
2055
2056 out:
2057         return nr_moved;
2058 }
2059
2060 /*
2061  * idle_balance is called by schedule() if this_cpu is about to become
2062  * idle. Attempts to pull tasks from other CPUs.
2063  */
2064 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2065 {
2066         struct sched_domain *sd;
2067
2068         for_each_domain(this_cpu, sd) {
2069                 if (sd->flags & SD_BALANCE_NEWIDLE) {
2070                         if (load_balance_newidle(this_cpu, this_rq, sd)) {
2071                                 /* We've pulled tasks over so stop searching */
2072                                 break;
2073                         }
2074                 }
2075         }
2076 }
2077
2078 /*
2079  * active_load_balance is run by migration threads. It pushes running tasks
2080  * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2081  * running on each physical CPU where possible, and avoids physical /
2082  * logical imbalances.
2083  *
2084  * Called with busiest_rq locked.
2085  */
2086 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2087 {
2088         struct sched_domain *sd;
2089         struct sched_group *cpu_group;
2090         runqueue_t *target_rq;
2091         cpumask_t visited_cpus;
2092         int cpu;
2093
2094         /*
2095          * Search for suitable CPUs to push tasks to in successively higher
2096          * domains with SD_LOAD_BALANCE set.
2097          */
2098         visited_cpus = CPU_MASK_NONE;
2099         for_each_domain(busiest_cpu, sd) {
2100                 if (!(sd->flags & SD_LOAD_BALANCE))
2101                         /* no more domains to search */
2102                         break;
2103
2104                 schedstat_inc(sd, alb_cnt);
2105
2106                 cpu_group = sd->groups;
2107                 do {
2108                         for_each_cpu_mask(cpu, cpu_group->cpumask) {
2109                                 if (busiest_rq->nr_running <= 1)
2110                                         /* no more tasks left to move */
2111                                         return;
2112                                 if (cpu_isset(cpu, visited_cpus))
2113                                         continue;
2114                                 cpu_set(cpu, visited_cpus);
2115                                 if (!cpu_and_siblings_are_idle(cpu) || cpu == busiest_cpu)
2116                                         continue;
2117
2118                                 target_rq = cpu_rq(cpu);
2119                                 /*
2120                                  * This condition is "impossible", if it occurs
2121                                  * we need to fix it.  Originally reported by
2122                                  * Bjorn Helgaas on a 128-cpu setup.
2123                                  */
2124                                 BUG_ON(busiest_rq == target_rq);
2125
2126                                 /* move a task from busiest_rq to target_rq */
2127                                 double_lock_balance(busiest_rq, target_rq);
2128                                 if (move_tasks(target_rq, cpu, busiest_rq,
2129                                                 1, sd, SCHED_IDLE)) {
2130                                         schedstat_inc(sd, alb_pushed);
2131                                 } else {
2132                                         schedstat_inc(sd, alb_failed);
2133                                 }
2134                                 spin_unlock(&target_rq->lock);
2135                         }
2136                         cpu_group = cpu_group->next;
2137                 } while (cpu_group != sd->groups);
2138         }
2139 }
2140
2141 /*
2142  * rebalance_tick will get called every timer tick, on every CPU.
2143  *
2144  * It checks each scheduling domain to see if it is due to be balanced,
2145  * and initiates a balancing operation if so.
2146  *
2147  * Balancing parameters are set up in arch_init_sched_domains.
2148  */
2149
2150 /* Don't have all balancing operations going off at once */
2151 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2152
2153 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2154                            enum idle_type idle)
2155 {
2156         unsigned long old_load, this_load;
2157         unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2158         struct sched_domain *sd;
2159
2160         /* Update our load */
2161         old_load = this_rq->cpu_load;
2162         this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2163         /*
2164          * Round up the averaging division if load is increasing. This
2165          * prevents us from getting stuck on 9 if the load is 10, for
2166          * example.
2167          */
2168         if (this_load > old_load)
2169                 old_load++;
2170         this_rq->cpu_load = (old_load + this_load) / 2;
2171
2172         for_each_domain(this_cpu, sd) {
2173                 unsigned long interval;
2174
2175                 if (!(sd->flags & SD_LOAD_BALANCE))
2176                         continue;
2177
2178                 interval = sd->balance_interval;
2179                 if (idle != SCHED_IDLE)
2180                         interval *= sd->busy_factor;
2181
2182                 /* scale ms to jiffies */
2183                 interval = msecs_to_jiffies(interval);
2184                 if (unlikely(!interval))
2185                         interval = 1;
2186
2187                 if (j - sd->last_balance >= interval) {
2188                         if (load_balance(this_cpu, this_rq, sd, idle)) {
2189                                 /* We've pulled tasks over so no longer idle */
2190                                 idle = NOT_IDLE;
2191                         }
2192                         sd->last_balance += interval;
2193                 }
2194         }
2195 }
2196 #else
2197 /*
2198  * on UP we do not need to balance between CPUs:
2199  */
2200 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2201 {
2202 }
2203 static inline void idle_balance(int cpu, runqueue_t *rq)
2204 {
2205 }
2206 #endif
2207
2208 static inline int wake_priority_sleeper(runqueue_t *rq)
2209 {
2210         int ret = 0;
2211 #ifdef CONFIG_SCHED_SMT
2212         spin_lock(&rq->lock);
2213         /*
2214          * If an SMT sibling task has been put to sleep for priority
2215          * reasons reschedule the idle task to see if it can now run.
2216          */
2217         if (rq->nr_running) {
2218                 resched_task(rq->idle);
2219                 ret = 1;
2220         }
2221         spin_unlock(&rq->lock);
2222 #endif
2223         return ret;
2224 }
2225
2226 DEFINE_PER_CPU(struct kernel_stat, kstat);
2227
2228 EXPORT_PER_CPU_SYMBOL(kstat);
2229
2230 /*
2231  * This is called on clock ticks and on context switches.
2232  * Bank in p->sched_time the ns elapsed since the last tick or switch.
2233  */
2234 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2235                                     unsigned long long now)
2236 {
2237         unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2238         p->sched_time += now - last;
2239 }
2240
2241 /*
2242  * Return current->sched_time plus any more ns on the sched_clock
2243  * that have not yet been banked.
2244  */
2245 unsigned long long current_sched_time(const task_t *tsk)
2246 {
2247         unsigned long long ns;
2248         unsigned long flags;
2249         local_irq_save(flags);
2250         ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2251         ns = tsk->sched_time + (sched_clock() - ns);
2252         local_irq_restore(flags);
2253         return ns;
2254 }
2255
2256 /*
2257  * We place interactive tasks back into the active array, if possible.
2258  *
2259  * To guarantee that this does not starve expired tasks we ignore the
2260  * interactivity of a task if the first expired task had to wait more
2261  * than a 'reasonable' amount of time. This deadline timeout is
2262  * load-dependent, as the frequency of array switched decreases with
2263  * increasing number of running tasks. We also ignore the interactivity
2264  * if a better static_prio task has expired:
2265  */
2266 #define EXPIRED_STARVING(rq) \
2267         ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2268                 (jiffies - (rq)->expired_timestamp >= \
2269                         STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2270                         ((rq)->curr->static_prio > (rq)->best_expired_prio))
2271
2272 /*
2273  * Account user cpu time to a process.
2274  * @p: the process that the cpu time gets accounted to
2275  * @hardirq_offset: the offset to subtract from hardirq_count()
2276  * @cputime: the cpu time spent in user space since the last update
2277  */
2278 void account_user_time(struct task_struct *p, cputime_t cputime)
2279 {
2280         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2281         cputime64_t tmp;
2282
2283         p->utime = cputime_add(p->utime, cputime);
2284
2285         /* Add user time to cpustat. */
2286         tmp = cputime_to_cputime64(cputime);
2287         if (TASK_NICE(p) > 0)
2288                 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2289         else
2290                 cpustat->user = cputime64_add(cpustat->user, tmp);
2291 }
2292
2293 /*
2294  * Account system cpu time to a process.
2295  * @p: the process that the cpu time gets accounted to
2296  * @hardirq_offset: the offset to subtract from hardirq_count()
2297  * @cputime: the cpu time spent in kernel space since the last update
2298  */
2299 void account_system_time(struct task_struct *p, int hardirq_offset,
2300                          cputime_t cputime)
2301 {
2302         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2303         runqueue_t *rq = this_rq();
2304         cputime64_t tmp;
2305
2306         p->stime = cputime_add(p->stime, cputime);
2307
2308         /* Add system time to cpustat. */
2309         tmp = cputime_to_cputime64(cputime);
2310         if (hardirq_count() - hardirq_offset)
2311                 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2312         else if (softirq_count())
2313                 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2314         else if (p != rq->idle)
2315                 cpustat->system = cputime64_add(cpustat->system, tmp);
2316         else if (atomic_read(&rq->nr_iowait) > 0)
2317                 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2318         else
2319                 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2320         /* Account for system time used */
2321         acct_update_integrals(p);
2322         /* Update rss highwater mark */
2323         update_mem_hiwater(p);
2324 }
2325
2326 /*
2327  * Account for involuntary wait time.
2328  * @p: the process from which the cpu time has been stolen
2329  * @steal: the cpu time spent in involuntary wait
2330  */
2331 void account_steal_time(struct task_struct *p, cputime_t steal)
2332 {
2333         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2334         cputime64_t tmp = cputime_to_cputime64(steal);
2335         runqueue_t *rq = this_rq();
2336
2337         if (p == rq->idle) {
2338                 p->stime = cputime_add(p->stime, steal);
2339                 if (atomic_read(&rq->nr_iowait) > 0)
2340                         cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2341                 else
2342                         cpustat->idle = cputime64_add(cpustat->idle, tmp);
2343         } else
2344                 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2345 }
2346
2347 /*
2348  * This function gets called by the timer code, with HZ frequency.
2349  * We call it with interrupts disabled.
2350  *
2351  * It also gets called by the fork code, when changing the parent's
2352  * timeslices.
2353  */
2354 void scheduler_tick(void)
2355 {
2356         int cpu = smp_processor_id();
2357         runqueue_t *rq = this_rq();
2358         task_t *p = current;
2359         unsigned long long now = sched_clock();
2360
2361         update_cpu_clock(p, rq, now);
2362
2363         rq->timestamp_last_tick = now;
2364
2365         if (p == rq->idle) {
2366                 if (wake_priority_sleeper(rq))
2367                         goto out;
2368                 rebalance_tick(cpu, rq, SCHED_IDLE);
2369                 return;
2370         }
2371
2372         /* Task might have expired already, but not scheduled off yet */
2373         if (p->array != rq->active) {
2374                 set_tsk_need_resched(p);
2375                 goto out;
2376         }
2377         spin_lock(&rq->lock);
2378         /*
2379          * The task was running during this tick - update the
2380          * time slice counter. Note: we do not update a thread's
2381          * priority until it either goes to sleep or uses up its
2382          * timeslice. This makes it possible for interactive tasks
2383          * to use up their timeslices at their highest priority levels.
2384          */
2385         if (rt_task(p)) {
2386                 /*
2387                  * RR tasks need a special form of timeslice management.
2388                  * FIFO tasks have no timeslices.
2389                  */
2390                 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2391                         p->time_slice = task_timeslice(p);
2392                         p->first_time_slice = 0;
2393                         set_tsk_need_resched(p);
2394
2395                         /* put it at the end of the queue: */
2396                         requeue_task(p, rq->active);
2397                 }
2398                 goto out_unlock;
2399         }
2400         if (!--p->time_slice) {
2401                 dequeue_task(p, rq->active);
2402                 set_tsk_need_resched(p);
2403                 p->prio = effective_prio(p);
2404                 p->time_slice = task_timeslice(p);
2405                 p->first_time_slice = 0;
2406
2407                 if (!rq->expired_timestamp)
2408                         rq->expired_timestamp = jiffies;
2409                 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2410                         enqueue_task(p, rq->expired);
2411                         if (p->static_prio < rq->best_expired_prio)
2412                                 rq->best_expired_prio = p->static_prio;
2413                 } else
2414                         enqueue_task(p, rq->active);
2415         } else {
2416                 /*
2417                  * Prevent a too long timeslice allowing a task to monopolize
2418                  * the CPU. We do this by splitting up the timeslice into
2419                  * smaller pieces.
2420                  *
2421                  * Note: this does not mean the task's timeslices expire or
2422                  * get lost in any way, they just might be preempted by
2423                  * another task of equal priority. (one with higher
2424                  * priority would have preempted this task already.) We
2425                  * requeue this task to the end of the list on this priority
2426                  * level, which is in essence a round-robin of tasks with
2427                  * equal priority.
2428                  *
2429                  * This only applies to tasks in the interactive
2430                  * delta range with at least TIMESLICE_GRANULARITY to requeue.
2431                  */
2432                 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2433                         p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2434                         (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2435                         (p->array == rq->active)) {
2436
2437                         requeue_task(p, rq->active);
2438                         set_tsk_need_resched(p);
2439                 }
2440         }
2441 out_unlock:
2442         spin_unlock(&rq->lock);
2443 out:
2444         rebalance_tick(cpu, rq, NOT_IDLE);
2445 }
2446
2447 #ifdef CONFIG_SCHED_SMT
2448 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2449 {
2450         struct sched_domain *sd = this_rq->sd;
2451         cpumask_t sibling_map;
2452         int i;
2453
2454         if (!(sd->flags & SD_SHARE_CPUPOWER))
2455                 return;
2456
2457         /*
2458          * Unlock the current runqueue because we have to lock in
2459          * CPU order to avoid deadlocks. Caller knows that we might
2460          * unlock. We keep IRQs disabled.
2461          */
2462         spin_unlock(&this_rq->lock);
2463
2464         sibling_map = sd->span;
2465
2466         for_each_cpu_mask(i, sibling_map)
2467                 spin_lock(&cpu_rq(i)->lock);
2468         /*
2469          * We clear this CPU from the mask. This both simplifies the
2470          * inner loop and keps this_rq locked when we exit:
2471          */
2472         cpu_clear(this_cpu, sibling_map);
2473
2474         for_each_cpu_mask(i, sibling_map) {
2475                 runqueue_t *smt_rq = cpu_rq(i);
2476
2477                 /*
2478                  * If an SMT sibling task is sleeping due to priority
2479                  * reasons wake it up now.
2480                  */
2481                 if (smt_rq->curr == smt_rq->idle && smt_rq->nr_running)
2482                         resched_task(smt_rq->idle);
2483         }
2484
2485         for_each_cpu_mask(i, sibling_map)
2486                 spin_unlock(&cpu_rq(i)->lock);
2487         /*
2488          * We exit with this_cpu's rq still held and IRQs
2489          * still disabled:
2490          */
2491 }
2492
2493 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2494 {
2495         struct sched_domain *sd = this_rq->sd;
2496         cpumask_t sibling_map;
2497         prio_array_t *array;
2498         int ret = 0, i;
2499         task_t *p;
2500
2501         if (!(sd->flags & SD_SHARE_CPUPOWER))
2502                 return 0;
2503
2504         /*
2505          * The same locking rules and details apply as for
2506          * wake_sleeping_dependent():
2507          */
2508         spin_unlock(&this_rq->lock);
2509         sibling_map = sd->span;
2510         for_each_cpu_mask(i, sibling_map)
2511                 spin_lock(&cpu_rq(i)->lock);
2512         cpu_clear(this_cpu, sibling_map);
2513
2514         /*
2515          * Establish next task to be run - it might have gone away because
2516          * we released the runqueue lock above:
2517          */
2518         if (!this_rq->nr_running)
2519                 goto out_unlock;
2520         array = this_rq->active;
2521         if (!array->nr_active)
2522                 array = this_rq->expired;
2523         BUG_ON(!array->nr_active);
2524
2525         p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2526                 task_t, run_list);
2527
2528         for_each_cpu_mask(i, sibling_map) {
2529                 runqueue_t *smt_rq = cpu_rq(i);
2530                 task_t *smt_curr = smt_rq->curr;
2531
2532                 /*
2533                  * If a user task with lower static priority than the
2534                  * running task on the SMT sibling is trying to schedule,
2535                  * delay it till there is proportionately less timeslice
2536                  * left of the sibling task to prevent a lower priority
2537                  * task from using an unfair proportion of the
2538                  * physical cpu's resources. -ck
2539                  */
2540                 if (((smt_curr->time_slice * (100 - sd->per_cpu_gain) / 100) >
2541                         task_timeslice(p) || rt_task(smt_curr)) &&
2542                         p->mm && smt_curr->mm && !rt_task(p))
2543                                 ret = 1;
2544
2545                 /*
2546                  * Reschedule a lower priority task on the SMT sibling,
2547                  * or wake it up if it has been put to sleep for priority
2548                  * reasons.
2549                  */
2550                 if ((((p->time_slice * (100 - sd->per_cpu_gain) / 100) >
2551                         task_timeslice(smt_curr) || rt_task(p)) &&
2552                         smt_curr->mm && p->mm && !rt_task(smt_curr)) ||
2553                         (smt_curr == smt_rq->idle && smt_rq->nr_running))
2554                                 resched_task(smt_curr);
2555         }
2556 out_unlock:
2557         for_each_cpu_mask(i, sibling_map)
2558                 spin_unlock(&cpu_rq(i)->lock);
2559         return ret;
2560 }
2561 #else
2562 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2563 {
2564 }
2565
2566 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2567 {
2568         return 0;
2569 }
2570 #endif
2571
2572 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2573
2574 void fastcall add_preempt_count(int val)
2575 {
2576         /*
2577          * Underflow?
2578          */
2579         BUG_ON(((int)preempt_count() < 0));
2580         preempt_count() += val;
2581         /*
2582          * Spinlock count overflowing soon?
2583          */
2584         BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2585 }
2586 EXPORT_SYMBOL(add_preempt_count);
2587
2588 void fastcall sub_preempt_count(int val)
2589 {
2590         /*
2591          * Underflow?
2592          */
2593         BUG_ON(val > preempt_count());
2594         /*
2595          * Is the spinlock portion underflowing?
2596          */
2597         BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2598         preempt_count() -= val;
2599 }
2600 EXPORT_SYMBOL(sub_preempt_count);
2601
2602 #endif
2603
2604 /*
2605  * schedule() is the main scheduler function.
2606  */
2607 asmlinkage void __sched schedule(void)
2608 {
2609         long *switch_count;
2610         task_t *prev, *next;
2611         runqueue_t *rq;
2612         prio_array_t *array;
2613         struct list_head *queue;
2614         unsigned long long now;
2615         unsigned long run_time;
2616         int cpu, idx;
2617
2618         /*
2619          * Test if we are atomic.  Since do_exit() needs to call into
2620          * schedule() atomically, we ignore that path for now.
2621          * Otherwise, whine if we are scheduling when we should not be.
2622          */
2623         if (likely(!current->exit_state)) {
2624                 if (unlikely(in_atomic())) {
2625                         printk(KERN_ERR "scheduling while atomic: "
2626                                 "%s/0x%08x/%d\n",
2627                                 current->comm, preempt_count(), current->pid);
2628                         dump_stack();
2629                 }
2630         }
2631         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2632
2633 need_resched:
2634         preempt_disable();
2635         prev = current;
2636         release_kernel_lock(prev);
2637 need_resched_nonpreemptible:
2638         rq = this_rq();
2639
2640         /*
2641          * The idle thread is not allowed to schedule!
2642          * Remove this check after it has been exercised a bit.
2643          */
2644         if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2645                 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2646                 dump_stack();
2647         }
2648
2649         schedstat_inc(rq, sched_cnt);
2650         now = sched_clock();
2651         if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2652                 run_time = now - prev->timestamp;
2653                 if (unlikely((long long)(now - prev->timestamp) < 0))
2654                         run_time = 0;
2655         } else
2656                 run_time = NS_MAX_SLEEP_AVG;
2657
2658         /*
2659          * Tasks charged proportionately less run_time at high sleep_avg to
2660          * delay them losing their interactive status
2661          */
2662         run_time /= (CURRENT_BONUS(prev) ? : 1);
2663
2664         spin_lock_irq(&rq->lock);
2665
2666         if (unlikely(prev->flags & PF_DEAD))
2667                 prev->state = EXIT_DEAD;
2668
2669         switch_count = &prev->nivcsw;
2670         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2671                 switch_count = &prev->nvcsw;
2672                 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2673                                 unlikely(signal_pending(prev))))
2674                         prev->state = TASK_RUNNING;
2675                 else {
2676                         if (prev->state == TASK_UNINTERRUPTIBLE)
2677                                 rq->nr_uninterruptible++;
2678                         deactivate_task(prev, rq);
2679                 }
2680         }
2681
2682         cpu = smp_processor_id();
2683         if (unlikely(!rq->nr_running)) {
2684 go_idle:
2685                 idle_balance(cpu, rq);
2686                 if (!rq->nr_running) {
2687                         next = rq->idle;
2688                         rq->expired_timestamp = 0;
2689                         wake_sleeping_dependent(cpu, rq);
2690                         /*
2691                          * wake_sleeping_dependent() might have released
2692                          * the runqueue, so break out if we got new
2693                          * tasks meanwhile:
2694                          */
2695                         if (!rq->nr_running)
2696                                 goto switch_tasks;
2697                 }
2698         } else {
2699                 if (dependent_sleeper(cpu, rq)) {
2700                         next = rq->idle;
2701                         goto switch_tasks;
2702                 }
2703                 /*
2704                  * dependent_sleeper() releases and reacquires the runqueue
2705                  * lock, hence go into the idle loop if the rq went
2706                  * empty meanwhile:
2707                  */
2708                 if (unlikely(!rq->nr_running))
2709                         goto go_idle;
2710         }
2711
2712         array = rq->active;
2713         if (unlikely(!array->nr_active)) {
2714                 /*
2715                  * Switch the active and expired arrays.
2716                  */
2717                 schedstat_inc(rq, sched_switch);
2718                 rq->active = rq->expired;
2719                 rq->expired = array;
2720                 array = rq->active;
2721                 rq->expired_timestamp = 0;
2722                 rq->best_expired_prio = MAX_PRIO;
2723         }
2724
2725         idx = sched_find_first_bit(array->bitmap);
2726         queue = array->queue + idx;
2727         next = list_entry(queue->next, task_t, run_list);
2728
2729         if (!rt_task(next) && next->activated > 0) {
2730                 unsigned long long delta = now - next->timestamp;
2731                 if (unlikely((long long)(now - next->timestamp) < 0))
2732                         delta = 0;
2733
2734                 if (next->activated == 1)
2735                         delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2736
2737                 array = next->array;
2738                 dequeue_task(next, array);
2739                 recalc_task_prio(next, next->timestamp + delta);
2740                 enqueue_task(next, array);
2741         }
2742         next->activated = 0;
2743 switch_tasks:
2744         if (next == rq->idle)
2745                 schedstat_inc(rq, sched_goidle);
2746         prefetch(next);
2747         clear_tsk_need_resched(prev);
2748         rcu_qsctr_inc(task_cpu(prev));
2749
2750         update_cpu_clock(prev, rq, now);
2751
2752         prev->sleep_avg -= run_time;
2753         if ((long)prev->sleep_avg <= 0)
2754                 prev->sleep_avg = 0;
2755         prev->timestamp = prev->last_ran = now;
2756
2757         sched_info_switch(prev, next);
2758         if (likely(prev != next)) {
2759                 next->timestamp = now;
2760                 rq->nr_switches++;
2761                 rq->curr = next;
2762                 ++*switch_count;
2763
2764                 prepare_arch_switch(rq, next);
2765                 prev = context_switch(rq, prev, next);
2766                 barrier();
2767
2768                 finish_task_switch(prev);
2769         } else
2770                 spin_unlock_irq(&rq->lock);
2771
2772         prev = current;
2773         if (unlikely(reacquire_kernel_lock(prev) < 0))
2774                 goto need_resched_nonpreemptible;
2775         preempt_enable_no_resched();
2776         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2777                 goto need_resched;
2778 }
2779
2780 EXPORT_SYMBOL(schedule);
2781
2782 #ifdef CONFIG_PREEMPT
2783 /*
2784  * this is is the entry point to schedule() from in-kernel preemption
2785  * off of preempt_enable.  Kernel preemptions off return from interrupt
2786  * occur there and call schedule directly.
2787  */
2788 asmlinkage void __sched preempt_schedule(void)
2789 {
2790         struct thread_info *ti = current_thread_info();
2791 #ifdef CONFIG_PREEMPT_BKL
2792         struct task_struct *task = current;
2793         int saved_lock_depth;
2794 #endif
2795         /*
2796          * If there is a non-zero preempt_count or interrupts are disabled,
2797          * we do not want to preempt the current task.  Just return..
2798          */
2799         if (unlikely(ti->preempt_count || irqs_disabled()))
2800                 return;
2801
2802 need_resched:
2803         add_preempt_count(PREEMPT_ACTIVE);
2804         /*
2805          * We keep the big kernel semaphore locked, but we
2806          * clear ->lock_depth so that schedule() doesnt
2807          * auto-release the semaphore:
2808          */
2809 #ifdef CONFIG_PREEMPT_BKL
2810         saved_lock_depth = task->lock_depth;
2811         task->lock_depth = -1;
2812 #endif
2813         schedule();
2814 #ifdef CONFIG_PREEMPT_BKL
2815         task->lock_depth = saved_lock_depth;
2816 #endif
2817         sub_preempt_count(PREEMPT_ACTIVE);
2818
2819         /* we could miss a preemption opportunity between schedule and now */
2820         barrier();
2821         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2822                 goto need_resched;
2823 }
2824
2825 EXPORT_SYMBOL(preempt_schedule);
2826
2827 /*
2828  * this is is the entry point to schedule() from kernel preemption
2829  * off of irq context.
2830  * Note, that this is called and return with irqs disabled. This will
2831  * protect us against recursive calling from irq.
2832  */
2833 asmlinkage void __sched preempt_schedule_irq(void)
2834 {
2835         struct thread_info *ti = current_thread_info();
2836 #ifdef CONFIG_PREEMPT_BKL
2837         struct task_struct *task = current;
2838         int saved_lock_depth;
2839 #endif
2840         /* Catch callers which need to be fixed*/
2841         BUG_ON(ti->preempt_count || !irqs_disabled());
2842
2843 need_resched:
2844         add_preempt_count(PREEMPT_ACTIVE);
2845         /*
2846          * We keep the big kernel semaphore locked, but we
2847          * clear ->lock_depth so that schedule() doesnt
2848          * auto-release the semaphore:
2849          */
2850 #ifdef CONFIG_PREEMPT_BKL
2851         saved_lock_depth = task->lock_depth;
2852         task->lock_depth = -1;
2853 #endif
2854         local_irq_enable();
2855         schedule();
2856         local_irq_disable();
2857 #ifdef CONFIG_PREEMPT_BKL
2858         task->lock_depth = saved_lock_depth;
2859 #endif
2860         sub_preempt_count(PREEMPT_ACTIVE);
2861
2862         /* we could miss a preemption opportunity between schedule and now */
2863         barrier();
2864         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2865                 goto need_resched;
2866 }
2867
2868 #endif /* CONFIG_PREEMPT */
2869
2870 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, void *key)
2871 {
2872         task_t *p = curr->task;
2873         return try_to_wake_up(p, mode, sync);
2874 }
2875
2876 EXPORT_SYMBOL(default_wake_function);
2877
2878 /*
2879  * The core wakeup function.  Non-exclusive wakeups (nr_exclusive == 0) just
2880  * wake everything up.  If it's an exclusive wakeup (nr_exclusive == small +ve
2881  * number) then we wake all the non-exclusive tasks and one exclusive task.
2882  *
2883  * There are circumstances in which we can try to wake a task which has already
2884  * started to run but is not in state TASK_RUNNING.  try_to_wake_up() returns
2885  * zero in this (rare) case, and we handle it by continuing to scan the queue.
2886  */
2887 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2888                              int nr_exclusive, int sync, void *key)
2889 {
2890         struct list_head *tmp, *next;
2891
2892         list_for_each_safe(tmp, next, &q->task_list) {
2893                 wait_queue_t *curr;
2894                 unsigned flags;
2895                 curr = list_entry(tmp, wait_queue_t, task_list);
2896                 flags = curr->flags;
2897                 if (curr->func(curr, mode, sync, key) &&
2898                     (flags & WQ_FLAG_EXCLUSIVE) &&
2899                     !--nr_exclusive)
2900                         break;
2901         }
2902 }
2903
2904 /**
2905  * __wake_up - wake up threads blocked on a waitqueue.
2906  * @q: the waitqueue
2907  * @mode: which threads
2908  * @nr_exclusive: how many wake-one or wake-many threads to wake up
2909  * @key: is directly passed to the wakeup function
2910  */
2911 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
2912                                 int nr_exclusive, void *key)
2913 {
2914         unsigned long flags;
2915
2916         spin_lock_irqsave(&q->lock, flags);
2917         __wake_up_common(q, mode, nr_exclusive, 0, key);
2918         spin_unlock_irqrestore(&q->lock, flags);
2919 }
2920
2921 EXPORT_SYMBOL(__wake_up);
2922
2923 /*
2924  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2925  */
2926 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
2927 {
2928         __wake_up_common(q, mode, 1, 0, NULL);
2929 }
2930
2931 /**
2932  * __wake_up_sync - wake up threads blocked on a waitqueue.
2933  * @q: the waitqueue
2934  * @mode: which threads
2935  * @nr_exclusive: how many wake-one or wake-many threads to wake up
2936  *
2937  * The sync wakeup differs that the waker knows that it will schedule
2938  * away soon, so while the target thread will be woken up, it will not
2939  * be migrated to another CPU - ie. the two threads are 'synchronized'
2940  * with each other. This can prevent needless bouncing between CPUs.
2941  *
2942  * On UP it can prevent extra preemption.
2943  */
2944 void fastcall __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2945 {
2946         unsigned long flags;
2947         int sync = 1;
2948
2949         if (unlikely(!q))
2950                 return;
2951
2952         if (unlikely(!nr_exclusive))
2953                 sync = 0;
2954
2955         spin_lock_irqsave(&q->lock, flags);
2956         __wake_up_common(q, mode, nr_exclusive, sync, NULL);
2957         spin_unlock_irqrestore(&q->lock, flags);
2958 }
2959 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
2960
2961 void fastcall complete(struct completion *x)
2962 {
2963         unsigned long flags;
2964
2965         spin_lock_irqsave(&x->wait.lock, flags);
2966         x->done++;
2967         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
2968                          1, 0, NULL);
2969         spin_unlock_irqrestore(&x->wait.lock, flags);
2970 }
2971 EXPORT_SYMBOL(complete);
2972
2973 void fastcall complete_all(struct completion *x)
2974 {
2975         unsigned long flags;
2976
2977         spin_lock_irqsave(&x->wait.lock, flags);
2978         x->done += UINT_MAX/2;
2979         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
2980                          0, 0, NULL);
2981         spin_unlock_irqrestore(&x->wait.lock, flags);
2982 }
2983 EXPORT_SYMBOL(complete_all);
2984
2985 void fastcall __sched wait_for_completion(struct completion *x)
2986 {
2987         might_sleep();
2988         spin_lock_irq(&x->wait.lock);
2989         if (!x->done) {
2990                 DECLARE_WAITQUEUE(wait, current);
2991
2992                 wait.flags |= WQ_FLAG_EXCLUSIVE;
2993                 __add_wait_queue_tail(&x->wait, &wait);
2994                 do {
2995                         __set_current_state(TASK_UNINTERRUPTIBLE);
2996                         spin_unlock_irq(&x->wait.lock);
2997                         schedule();
2998                         spin_lock_irq(&x->wait.lock);
2999                 } while (!x->done);
3000                 __remove_wait_queue(&x->wait, &wait);
3001         }
3002         x->done--;
3003         spin_unlock_irq(&x->wait.lock);
3004 }
3005 EXPORT_SYMBOL(wait_for_completion);
3006
3007 unsigned long fastcall __sched
3008 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3009 {
3010         might_sleep();
3011
3012         spin_lock_irq(&x->wait.lock);
3013         if (!x->done) {
3014                 DECLARE_WAITQUEUE(wait, current);
3015
3016                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3017                 __add_wait_queue_tail(&x->wait, &wait);
3018                 do {
3019                         __set_current_state(TASK_UNINTERRUPTIBLE);
3020                         spin_unlock_irq(&x->wait.lock);
3021                         timeout = schedule_timeout(timeout);
3022                         spin_lock_irq(&x->wait.lock);
3023                         if (!timeout) {
3024                                 __remove_wait_queue(&x->wait, &wait);
3025                                 goto out;
3026                         }
3027                 } while (!x->done);
3028                 __remove_wait_queue(&x->wait, &wait);
3029         }
3030         x->done--;
3031 out:
3032         spin_unlock_irq(&x->wait.lock);
3033         return timeout;
3034 }
3035 EXPORT_SYMBOL(wait_for_completion_timeout);
3036
3037 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3038 {
3039         int ret = 0;
3040
3041         might_sleep();
3042
3043         spin_lock_irq(&x->wait.lock);
3044         if (!x->done) {
3045                 DECLARE_WAITQUEUE(wait, current);
3046
3047                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3048                 __add_wait_queue_tail(&x->wait, &wait);
3049                 do {
3050                         if (signal_pending(current)) {
3051                                 ret = -ERESTARTSYS;
3052                                 __remove_wait_queue(&x->wait, &wait);
3053                                 goto out;
3054                         }
3055                         __set_current_state(TASK_INTERRUPTIBLE);
3056                         spin_unlock_irq(&x->wait.lock);
3057                         schedule();
3058                         spin_lock_irq(&x->wait.lock);
3059                 } while (!x->done);
3060                 __remove_wait_queue(&x->wait, &wait);
3061         }
3062         x->done--;
3063 out:
3064         spin_unlock_irq(&x->wait.lock);
3065
3066         return ret;
3067 }
3068 EXPORT_SYMBOL(wait_for_completion_interruptible);
3069
3070 unsigned long fastcall __sched
3071 wait_for_completion_interruptible_timeout(struct completion *x,
3072                                           unsigned long timeout)
3073 {
3074         might_sleep();
3075
3076         spin_lock_irq(&x->wait.lock);
3077         if (!x->done) {
3078                 DECLARE_WAITQUEUE(wait, current);
3079
3080                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3081                 __add_wait_queue_tail(&x->wait, &wait);
3082                 do {
3083                         if (signal_pending(current)) {
3084                                 timeout = -ERESTARTSYS;
3085                                 __remove_wait_queue(&x->wait, &wait);
3086                                 goto out;
3087                         }
3088                         __set_current_state(TASK_INTERRUPTIBLE);
3089                         spin_unlock_irq(&x->wait.lock);
3090                         timeout = schedule_timeout(timeout);
3091                         spin_lock_irq(&x->wait.lock);
3092                         if (!timeout) {
3093                                 __remove_wait_queue(&x->wait, &wait);
3094                                 goto out;
3095                         }
3096                 } while (!x->done);
3097                 __remove_wait_queue(&x->wait, &wait);
3098         }
3099         x->done--;
3100 out:
3101         spin_unlock_irq(&x->wait.lock);
3102         return timeout;
3103 }
3104 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3105
3106
3107 #define SLEEP_ON_VAR                                    \
3108         unsigned long flags;                            \
3109         wait_queue_t wait;                              \
3110         init_waitqueue_entry(&wait, current);
3111
3112 #define SLEEP_ON_HEAD                                   \
3113         spin_lock_irqsave(&q->lock,flags);              \
3114         __add_wait_queue(q, &wait);                     \
3115         spin_unlock(&q->lock);
3116
3117 #define SLEEP_ON_TAIL                                   \
3118         spin_lock_irq(&q->lock);                        \
3119         __remove_wait_queue(q, &wait);                  \
3120         spin_unlock_irqrestore(&q->lock, flags);
3121
3122 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3123 {
3124         SLEEP_ON_VAR
3125
3126         current->state = TASK_INTERRUPTIBLE;
3127
3128         SLEEP_ON_HEAD
3129         schedule();
3130         SLEEP_ON_TAIL
3131 }
3132
3133 EXPORT_SYMBOL(interruptible_sleep_on);
3134
3135 long fastcall __sched interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3136 {
3137         SLEEP_ON_VAR
3138
3139         current->state = TASK_INTERRUPTIBLE;
3140
3141         SLEEP_ON_HEAD
3142         timeout = schedule_timeout(timeout);
3143         SLEEP_ON_TAIL
3144
3145         return timeout;
3146 }
3147
3148 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3149
3150 void fastcall __sched sleep_on(wait_queue_head_t *q)
3151 {
3152         SLEEP_ON_VAR
3153
3154         current->state = TASK_UNINTERRUPTIBLE;
3155
3156         SLEEP_ON_HEAD
3157         schedule();
3158         SLEEP_ON_TAIL
3159 }
3160
3161 EXPORT_SYMBOL(sleep_on);
3162
3163 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3164 {
3165         SLEEP_ON_VAR
3166
3167         current->state = TASK_UNINTERRUPTIBLE;
3168
3169         SLEEP_ON_HEAD
3170         timeout = schedule_timeout(timeout);
3171         SLEEP_ON_TAIL
3172
3173         return timeout;
3174 }
3175
3176 EXPORT_SYMBOL(sleep_on_timeout);
3177
3178 void set_user_nice(task_t *p, long nice)
3179 {
3180         unsigned long flags;
3181         prio_array_t *array;
3182         runqueue_t *rq;
3183         int old_prio, new_prio, delta;
3184
3185         if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3186                 return;
3187         /*
3188          * We have to be careful, if called from sys_setpriority(),
3189          * the task might be in the middle of scheduling on another CPU.
3190          */
3191         rq = task_rq_lock(p, &flags);
3192         /*
3193          * The RT priorities are set via sched_setscheduler(), but we still
3194          * allow the 'normal' nice value to be set - but as expected
3195          * it wont have any effect on scheduling until the task is
3196          * not SCHED_NORMAL:
3197          */
3198         if (rt_task(p)) {
3199                 p->static_prio = NICE_TO_PRIO(nice);
3200                 goto out_unlock;
3201         }
3202         array = p->array;
3203         if (array)
3204                 dequeue_task(p, array);
3205
3206         old_prio = p->prio;
3207         new_prio = NICE_TO_PRIO(nice);
3208         delta = new_prio - old_prio;
3209         p->static_prio = NICE_TO_PRIO(nice);
3210         p->prio += delta;
3211
3212         if (array) {
3213                 enqueue_task(p, array);
3214                 /*
3215                  * If the task increased its priority or is running and
3216                  * lowered its priority, then reschedule its CPU:
3217                  */
3218                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3219                         resched_task(rq->curr);
3220         }
3221 out_unlock:
3222         task_rq_unlock(rq, &flags);
3223 }
3224
3225 EXPORT_SYMBOL(set_user_nice);
3226
3227 /*
3228  * can_nice - check if a task can reduce its nice value
3229  * @p: task
3230  * @nice: nice value
3231  */
3232 int can_nice(const task_t *p, const int nice)
3233 {
3234         /* convert nice value [19,-20] to rlimit style value [0,39] */
3235         int nice_rlim = 19 - nice;
3236         return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3237                 capable(CAP_SYS_NICE));
3238 }
3239
3240 #ifdef __ARCH_WANT_SYS_NICE
3241
3242 /*
3243  * sys_nice - change the priority of the current process.
3244  * @increment: priority increment
3245  *
3246  * sys_setpriority is a more generic, but much slower function that
3247  * does similar things.
3248  */
3249 asmlinkage long sys_nice(int increment)
3250 {
3251         int retval;
3252         long nice;
3253
3254         /*
3255          * Setpriority might change our priority at the same moment.
3256          * We don't have to worry. Conceptually one call occurs first
3257          * and we have a single winner.
3258          */
3259         if (increment < -40)
3260                 increment = -40;
3261         if (increment > 40)
3262                 increment = 40;
3263
3264         nice = PRIO_TO_NICE(current->static_prio) + increment;
3265         if (nice < -20)
3266                 nice = -20;
3267         if (nice > 19)
3268                 nice = 19;
3269
3270         if (increment < 0 && !can_nice(current, nice))
3271                 return -EPERM;
3272
3273         retval = security_task_setnice(current, nice);
3274         if (retval)
3275                 return retval;
3276
3277         set_user_nice(current, nice);
3278         return 0;
3279 }
3280
3281 #endif
3282
3283 /**
3284  * task_prio - return the priority value of a given task.
3285  * @p: the task in question.
3286  *
3287  * This is the priority value as seen by users in /proc.
3288  * RT tasks are offset by -200. Normal tasks are centered
3289  * around 0, value goes from -16 to +15.
3290  */
3291 int task_prio(const task_t *p)
3292 {
3293         return p->prio - MAX_RT_PRIO;
3294 }
3295
3296 /**
3297  * task_nice - return the nice value of a given task.
3298  * @p: the task in question.
3299  */
3300 int task_nice(const task_t *p)
3301 {
3302         return TASK_NICE(p);
3303 }
3304
3305 /*
3306  * The only users of task_nice are binfmt_elf and binfmt_elf32.
3307  * binfmt_elf is no longer modular, but binfmt_elf32 still is.
3308  * Therefore, task_nice is needed if there is a compat_mode.
3309  */
3310 #ifdef CONFIG_COMPAT
3311 EXPORT_SYMBOL_GPL(task_nice);
3312 #endif
3313
3314 /**
3315  * idle_cpu - is a given cpu idle currently?
3316  * @cpu: the processor in question.
3317  */
3318 int idle_cpu(int cpu)
3319 {
3320         return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3321 }
3322
3323 EXPORT_SYMBOL_GPL(idle_cpu);
3324
3325 /**
3326  * idle_task - return the idle task for a given cpu.
3327  * @cpu: the processor in question.
3328  */
3329 task_t *idle_task(int cpu)
3330 {
3331         return cpu_rq(cpu)->idle;
3332 }
3333
3334 /**
3335  * find_process_by_pid - find a process with a matching PID value.
3336  * @pid: the pid in question.
3337  */
3338 static inline task_t *find_process_by_pid(pid_t pid)
3339 {
3340         return pid ? find_task_by_pid(pid) : current;
3341 }
3342
3343 /* Actually do priority change: must hold rq lock. */
3344 static void __setscheduler(struct task_struct *p, int policy, int prio)
3345 {
3346         BUG_ON(p->array);
3347         p->policy = policy;
3348         p->rt_priority = prio;
3349         if (policy != SCHED_NORMAL)
3350                 p->prio = MAX_USER_RT_PRIO-1 - p->rt_priority;
3351         else
3352                 p->prio = p->static_prio;
3353 }
3354
3355 /**
3356  * sched_setscheduler - change the scheduling policy and/or RT priority of
3357  * a thread.
3358  * @p: the task in question.
3359  * @policy: new policy.
3360  * @param: structure containing the new RT priority.
3361  */
3362 int sched_setscheduler(struct task_struct *p, int policy, struct sched_param *param)
3363 {
3364         int retval;
3365         int oldprio, oldpolicy = -1;
3366         prio_array_t *array;
3367         unsigned long flags;
3368         runqueue_t *rq;
3369
3370 recheck:
3371         /* double check policy once rq lock held */
3372         if (policy < 0)
3373                 policy = oldpolicy = p->policy;
3374         else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3375                                 policy != SCHED_NORMAL)
3376                         return -EINVAL;
3377         /*
3378          * Valid priorities for SCHED_FIFO and SCHED_RR are
3379          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3380          */
3381         if (param->sched_priority < 0 ||
3382             param->sched_priority > MAX_USER_RT_PRIO-1)
3383                 return -EINVAL;
3384         if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3385                 return -EINVAL;
3386
3387         if ((policy == SCHED_FIFO || policy == SCHED_RR) &&
3388             param->sched_priority > p->signal->rlim[RLIMIT_RTPRIO].rlim_cur &&
3389             !capable(CAP_SYS_NICE))
3390                 return -EPERM;
3391         if ((current->euid != p->euid) && (current->euid != p->uid) &&
3392             !capable(CAP_SYS_NICE))
3393                 return -EPERM;
3394
3395         retval = security_task_setscheduler(p, policy, param);
3396         if (retval)
3397                 return retval;
3398         /*
3399          * To be able to change p->policy safely, the apropriate
3400          * runqueue lock must be held.
3401          */
3402         rq = task_rq_lock(p, &flags);
3403         /* recheck policy now with rq lock held */
3404         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3405                 policy = oldpolicy = -1;
3406                 task_rq_unlock(rq, &flags);
3407                 goto recheck;
3408         }
3409         array = p->array;
3410         if (array)
3411                 deactivate_task(p, rq);
3412         oldprio = p->prio;
3413         __setscheduler(p, policy, param->sched_priority);
3414         if (array) {
3415                 __activate_task(p, rq);
3416                 /*
3417                  * Reschedule if we are currently running on this runqueue and
3418                  * our priority decreased, or if we are not currently running on
3419                  * this runqueue and our priority is higher than the current's
3420                  */
3421                 if (task_running(rq, p)) {
3422                         if (p->prio > oldprio)
3423                                 resched_task(rq->curr);
3424                 } else if (TASK_PREEMPTS_CURR(p, rq))
3425                         resched_task(rq->curr);
3426         }
3427         task_rq_unlock(rq, &flags);
3428         return 0;
3429 }
3430 EXPORT_SYMBOL_GPL(sched_setscheduler);
3431
3432 static int do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3433 {
3434         int retval;
3435         struct sched_param lparam;
3436         struct task_struct *p;
3437
3438         if (!param || pid < 0)
3439                 return -EINVAL;
3440         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3441                 return -EFAULT;
3442         read_lock_irq(&tasklist_lock);
3443         p = find_process_by_pid(pid);
3444         if (!p) {
3445                 read_unlock_irq(&tasklist_lock);
3446                 return -ESRCH;
3447         }
3448         retval = sched_setscheduler(p, policy, &lparam);
3449         read_unlock_irq(&tasklist_lock);
3450         return retval;
3451 }
3452
3453 /**
3454  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3455  * @pid: the pid in question.
3456  * @policy: new policy.
3457  * @param: structure containing the new RT priority.
3458  */
3459 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3460                                        struct sched_param __user *param)
3461 {
3462         return do_sched_setscheduler(pid, policy, param);
3463 }
3464
3465 /**
3466  * sys_sched_setparam - set/change the RT priority of a thread
3467  * @pid: the pid in question.
3468  * @param: structure containing the new RT priority.
3469  */
3470 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3471 {
3472         return do_sched_setscheduler(pid, -1, param);
3473 }
3474
3475 /**
3476  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3477  * @pid: the pid in question.
3478  */
3479 asmlinkage long sys_sched_getscheduler(pid_t pid)
3480 {
3481         int retval = -EINVAL;
3482         task_t *p;
3483
3484         if (pid < 0)
3485                 goto out_nounlock;
3486
3487         retval = -ESRCH;
3488         read_lock(&tasklist_lock);
3489         p = find_process_by_pid(pid);
3490         if (p) {
3491                 retval = security_task_getscheduler(p);
3492                 if (!retval)
3493                         retval = p->policy;
3494         }
3495         read_unlock(&tasklist_lock);
3496
3497 out_nounlock:
3498         return retval;
3499 }
3500
3501 /**
3502  * sys_sched_getscheduler - get the RT priority of a thread
3503  * @pid: the pid in question.
3504  * @param: structure containing the RT priority.
3505  */
3506 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3507 {
3508         struct sched_param lp;
3509         int retval = -EINVAL;
3510         task_t *p;
3511
3512         if (!param || pid < 0)
3513                 goto out_nounlock;
3514
3515         read_lock(&tasklist_lock);
3516         p = find_process_by_pid(pid);
3517         retval = -ESRCH;
3518         if (!p)
3519                 goto out_unlock;
3520
3521         retval = security_task_getscheduler(p);
3522         if (retval)
3523                 goto out_unlock;
3524
3525         lp.sched_priority = p->rt_priority;
3526         read_unlock(&tasklist_lock);
3527
3528         /*
3529          * This one might sleep, we cannot do it with a spinlock held ...
3530          */
3531         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3532
3533 out_nounlock:
3534         return retval;
3535
3536 out_unlock:
3537         read_unlock(&tasklist_lock);
3538         return retval;
3539 }
3540
3541 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3542 {
3543         task_t *p;
3544         int retval;
3545         cpumask_t cpus_allowed;
3546
3547         lock_cpu_hotplug();
3548         read_lock(&tasklist_lock);
3549
3550         p = find_process_by_pid(pid);
3551         if (!p) {
3552                 read_unlock(&tasklist_lock);
3553                 unlock_cpu_hotplug();
3554                 return -ESRCH;
3555         }
3556
3557         /*
3558          * It is not safe to call set_cpus_allowed with the
3559          * tasklist_lock held.  We will bump the task_struct's
3560          * usage count and then drop tasklist_lock.
3561          */
3562         get_task_struct(p);
3563         read_unlock(&tasklist_lock);
3564
3565         retval = -EPERM;
3566         if ((current->euid != p->euid) && (current->euid != p->uid) &&
3567                         !capable(CAP_SYS_NICE))
3568                 goto out_unlock;
3569
3570         cpus_allowed = cpuset_cpus_allowed(p);
3571         cpus_and(new_mask, new_mask, cpus_allowed);
3572         retval = set_cpus_allowed(p, new_mask);
3573
3574 out_unlock:
3575         put_task_struct(p);
3576         unlock_cpu_hotplug();
3577         return retval;
3578 }
3579
3580 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3581                              cpumask_t *new_mask)
3582 {
3583         if (len < sizeof(cpumask_t)) {
3584                 memset(new_mask, 0, sizeof(cpumask_t));
3585         } else if (len > sizeof(cpumask_t)) {
3586                 len = sizeof(cpumask_t);
3587         }
3588         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3589 }
3590
3591 /**
3592  * sys_sched_setaffinity - set the cpu affinity of a process
3593  * @pid: pid of the process
3594  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3595  * @user_mask_ptr: user-space pointer to the new cpu mask
3596  */
3597 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3598                                       unsigned long __user *user_mask_ptr)
3599 {
3600         cpumask_t new_mask;
3601         int retval;
3602
3603         retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3604         if (retval)
3605                 return retval;
3606
3607         return sched_setaffinity(pid, new_mask);
3608 }
3609
3610 /*
3611  * Represents all cpu's present in the system
3612  * In systems capable of hotplug, this map could dynamically grow
3613  * as new cpu's are detected in the system via any platform specific
3614  * method, such as ACPI for e.g.
3615  */
3616
3617 cpumask_t cpu_present_map;
3618 EXPORT_SYMBOL(cpu_present_map);
3619
3620 #ifndef CONFIG_SMP
3621 cpumask_t cpu_online_map = CPU_MASK_ALL;
3622 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3623 #endif
3624
3625 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3626 {
3627         int retval;
3628         task_t *p;
3629
3630         lock_cpu_hotplug();
3631         read_lock(&tasklist_lock);
3632
3633         retval = -ESRCH;
3634         p = find_process_by_pid(pid);
3635         if (!p)
3636                 goto out_unlock;
3637
3638         retval = 0;
3639         cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
3640
3641 out_unlock:
3642         read_unlock(&tasklist_lock);
3643         unlock_cpu_hotplug();
3644         if (retval)
3645                 return retval;
3646
3647         return 0;
3648 }
3649
3650 /**
3651  * sys_sched_getaffinity - get the cpu affinity of a process
3652  * @pid: pid of the process
3653  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3654  * @user_mask_ptr: user-space pointer to hold the current cpu mask
3655  */
3656 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3657                                       unsigned long __user *user_mask_ptr)
3658 {
3659         int ret;
3660         cpumask_t mask;
3661
3662         if (len < sizeof(cpumask_t))
3663                 return -EINVAL;
3664
3665         ret = sched_getaffinity(pid, &mask);
3666         if (ret < 0)
3667                 return ret;
3668
3669         if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3670                 return -EFAULT;
3671
3672         return sizeof(cpumask_t);
3673 }
3674
3675 /**
3676  * sys_sched_yield - yield the current processor to other threads.
3677  *
3678  * this function yields the current CPU by moving the calling thread
3679  * to the expired array. If there are no other threads running on this
3680  * CPU then this function will return.
3681  */
3682 asmlinkage long sys_sched_yield(void)
3683 {
3684         runqueue_t *rq = this_rq_lock();
3685         prio_array_t *array = current->array;
3686         prio_array_t *target = rq->expired;
3687
3688         schedstat_inc(rq, yld_cnt);
3689         /*
3690          * We implement yielding by moving the task into the expired
3691          * queue.
3692          *
3693          * (special rule: RT tasks will just roundrobin in the active
3694          *  array.)
3695          */
3696         if (rt_task(current))
3697                 target = rq->active;
3698
3699         if (current->array->nr_active == 1) {
3700                 schedstat_inc(rq, yld_act_empty);
3701                 if (!rq->expired->nr_active)
3702                         schedstat_inc(rq, yld_both_empty);
3703         } else if (!rq->expired->nr_active)
3704                 schedstat_inc(rq, yld_exp_empty);
3705
3706         if (array != target) {
3707                 dequeue_task(current, array);
3708                 enqueue_task(current, target);
3709         } else
3710                 /*
3711                  * requeue_task is cheaper so perform that if possible.
3712                  */
3713                 requeue_task(current, array);
3714
3715         /*
3716          * Since we are going to call schedule() anyway, there's
3717          * no need to preempt or enable interrupts:
3718          */
3719         __release(rq->lock);
3720         _raw_spin_unlock(&rq->lock);
3721         preempt_enable_no_resched();
3722
3723         schedule();
3724
3725         return 0;
3726 }
3727
3728 static inline void __cond_resched(void)
3729 {
3730         do {
3731                 add_preempt_count(PREEMPT_ACTIVE);
3732                 schedule();
3733                 sub_preempt_count(PREEMPT_ACTIVE);
3734         } while (need_resched());
3735 }
3736
3737 int __sched cond_resched(void)
3738 {
3739         if (need_resched()) {
3740                 __cond_resched();
3741                 return 1;
3742         }
3743         return 0;
3744 }
3745
3746 EXPORT_SYMBOL(cond_resched);
3747
3748 /*
3749  * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3750  * call schedule, and on return reacquire the lock.
3751  *
3752  * This works OK both with and without CONFIG_PREEMPT.  We do strange low-level
3753  * operations here to prevent schedule() from being called twice (once via
3754  * spin_unlock(), once by hand).
3755  */
3756 int cond_resched_lock(spinlock_t * lock)
3757 {
3758         int ret = 0;
3759
3760         if (need_lockbreak(lock)) {
3761                 spin_unlock(lock);
3762                 cpu_relax();
3763                 ret = 1;
3764                 spin_lock(lock);
3765         }
3766         if (need_resched()) {
3767                 _raw_spin_unlock(lock);
3768                 preempt_enable_no_resched();
3769                 __cond_resched();
3770                 ret = 1;
3771                 spin_lock(lock);
3772         }
3773         return ret;
3774 }
3775
3776 EXPORT_SYMBOL(cond_resched_lock);
3777
3778 int __sched cond_resched_softirq(void)
3779 {
3780         BUG_ON(!in_softirq());
3781
3782         if (need_resched()) {
3783                 __local_bh_enable();
3784                 __cond_resched();
3785                 local_bh_disable();
3786                 return 1;
3787         }
3788         return 0;
3789 }
3790
3791 EXPORT_SYMBOL(cond_resched_softirq);
3792
3793
3794 /**
3795  * yield - yield the current processor to other threads.
3796  *
3797  * this is a shortcut for kernel-space yielding - it marks the
3798  * thread runnable and calls sys_sched_yield().
3799  */
3800 void __sched yield(void)
3801 {
3802         set_current_state(TASK_RUNNING);
3803         sys_sched_yield();
3804 }
3805
3806 EXPORT_SYMBOL(yield);
3807
3808 /*
3809  * This task is about to go to sleep on IO.  Increment rq->nr_iowait so
3810  * that process accounting knows that this is a task in IO wait state.
3811  *
3812  * But don't do that if it is a deliberate, throttling IO wait (this task
3813  * has set its backing_dev_info: the queue against which it should throttle)
3814  */
3815 void __sched io_schedule(void)
3816 {
3817         struct runqueue *rq = &per_cpu(runqueues, _smp_processor_id());
3818
3819         atomic_inc(&rq->nr_iowait);
3820         schedule();
3821         atomic_dec(&rq->nr_iowait);
3822 }
3823
3824 EXPORT_SYMBOL(io_schedule);
3825
3826 long __sched io_schedule_timeout(long timeout)
3827 {
3828         struct runqueue *rq = &per_cpu(runqueues, _smp_processor_id());
3829         long ret;
3830
3831         atomic_inc(&rq->nr_iowait);
3832         ret = schedule_timeout(timeout);
3833         atomic_dec(&rq->nr_iowait);
3834         return ret;
3835 }
3836
3837 /**
3838  * sys_sched_get_priority_max - return maximum RT priority.
3839  * @policy: scheduling class.
3840  *
3841  * this syscall returns the maximum rt_priority that can be used
3842  * by a given scheduling class.
3843  */
3844 asmlinkage long sys_sched_get_priority_max(int policy)
3845 {
3846         int ret = -EINVAL;
3847
3848         switch (policy) {
3849         case SCHED_FIFO:
3850         case SCHED_RR:
3851                 ret = MAX_USER_RT_PRIO-1;
3852                 break;
3853         case SCHED_NORMAL:
3854                 ret = 0;
3855                 break;
3856         }
3857         return ret;
3858 }
3859
3860 /**
3861  * sys_sched_get_priority_min - return minimum RT priority.
3862  * @policy: scheduling class.
3863  *
3864  * this syscall returns the minimum rt_priority that can be used
3865  * by a given scheduling class.
3866  */
3867 asmlinkage long sys_sched_get_priority_min(int policy)
3868 {
3869         int ret = -EINVAL;
3870
3871         switch (policy) {
3872         case SCHED_FIFO:
3873         case SCHED_RR:
3874                 ret = 1;
3875                 break;
3876         case SCHED_NORMAL:
3877                 ret = 0;
3878         }
3879         return ret;
3880 }
3881
3882 /**
3883  * sys_sched_rr_get_interval - return the default timeslice of a process.
3884  * @pid: pid of the process.
3885  * @interval: userspace pointer to the timeslice value.
3886  *
3887  * this syscall writes the default timeslice value of a given process
3888  * into the user-space timespec buffer. A value of '0' means infinity.
3889  */
3890 asmlinkage
3891 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
3892 {
3893         int retval = -EINVAL;
3894         struct timespec t;
3895         task_t *p;
3896
3897         if (pid < 0)
3898                 goto out_nounlock;
3899
3900         retval = -ESRCH;
3901         read_lock(&tasklist_lock);
3902         p = find_process_by_pid(pid);
3903         if (!p)
3904                 goto out_unlock;
3905
3906         retval = security_task_getscheduler(p);
3907         if (retval)
3908                 goto out_unlock;
3909
3910         jiffies_to_timespec(p->policy & SCHED_FIFO ?
3911                                 0 : task_timeslice(p), &t);
3912         read_unlock(&tasklist_lock);
3913         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
3914 out_nounlock:
3915         return retval;
3916 out_unlock:
3917         read_unlock(&tasklist_lock);
3918         return retval;
3919 }
3920
3921 static inline struct task_struct *eldest_child(struct task_struct *p)
3922 {
3923         if (list_empty(&p->children)) return NULL;
3924         return list_entry(p->children.next,struct task_struct,sibling);
3925 }
3926
3927 static inline struct task_struct *older_sibling(struct task_struct *p)
3928 {
3929         if (p->sibling.prev==&p->parent->children) return NULL;
3930         return list_entry(p->sibling.prev,struct task_struct,sibling);
3931 }
3932
3933 static inline struct task_struct *younger_sibling(struct task_struct *p)
3934 {
3935         if (p->sibling.next==&p->parent->children) return NULL;
3936         return list_entry(p->sibling.next,struct task_struct,sibling);
3937 }
3938
3939 static void show_task(task_t * p)
3940 {
3941         task_t *relative;
3942         unsigned state;
3943         unsigned long free = 0;
3944         static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
3945
3946         printk("%-13.13s ", p->comm);
3947         state = p->state ? __ffs(p->state) + 1 : 0;
3948         if (state < ARRAY_SIZE(stat_nam))
3949                 printk(stat_nam[state]);
3950         else
3951                 printk("?");
3952 #if (BITS_PER_LONG == 32)
3953         if (state == TASK_RUNNING)
3954                 printk(" running ");
3955         else
3956                 printk(" %08lX ", thread_saved_pc(p));
3957 #else
3958         if (state == TASK_RUNNING)
3959                 printk("  running task   ");
3960         else
3961                 printk(" %016lx ", thread_saved_pc(p));
3962 #endif
3963 #ifdef CONFIG_DEBUG_STACK_USAGE
3964         {
3965                 unsigned long * n = (unsigned long *) (p->thread_info+1);
3966                 while (!*n)
3967                         n++;
3968                 free = (unsigned long) n - (unsigned long)(p->thread_info+1);
3969         }
3970 #endif
3971         printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
3972         if ((relative = eldest_child(p)))
3973                 printk("%5d ", relative->pid);
3974         else
3975                 printk("      ");
3976         if ((relative = younger_sibling(p)))
3977                 printk("%7d", relative->pid);
3978         else
3979                 printk("       ");
3980         if ((relative = older_sibling(p)))
3981                 printk(" %5d", relative->pid);
3982         else
3983                 printk("      ");
3984         if (!p->mm)
3985                 printk(" (L-TLB)\n");
3986         else
3987                 printk(" (NOTLB)\n");
3988
3989         if (state != TASK_RUNNING)
3990                 show_stack(p, NULL);
3991 }
3992
3993 void show_state(void)
3994 {
3995         task_t *g, *p;
3996
3997 #if (BITS_PER_LONG == 32)
3998         printk("\n"
3999                "                                               sibling\n");
4000         printk("  task             PC      pid father child younger older\n");
4001 #else
4002         printk("\n"
4003                "                                                       sibling\n");
4004         printk("  task                 PC          pid father child younger older\n");
4005 #endif
4006         read_lock(&tasklist_lock);
4007         do_each_thread(g, p) {
4008                 /*
4009                  * reset the NMI-timeout, listing all files on a slow
4010                  * console might take alot of time:
4011                  */
4012                 touch_nmi_watchdog();
4013                 show_task(p);
4014         } while_each_thread(g, p);
4015
4016         read_unlock(&tasklist_lock);
4017 }
4018
4019 void __devinit init_idle(task_t *idle, int cpu)
4020 {
4021         runqueue_t *rq = cpu_rq(cpu);
4022         unsigned long flags;
4023
4024         idle->sleep_avg = 0;
4025         idle->array = NULL;
4026         idle->prio = MAX_PRIO;
4027         idle->state = TASK_RUNNING;
4028         idle->cpus_allowed = cpumask_of_cpu(cpu);
4029         set_task_cpu(idle, cpu);
4030
4031         spin_lock_irqsave(&rq->lock, flags);
4032         rq->curr = rq->idle = idle;
4033         set_tsk_need_resched(idle);
4034         spin_unlock_irqrestore(&rq->lock, flags);
4035
4036         /* Set the preempt count _outside_ the spinlocks! */
4037 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4038         idle->thread_info->preempt_count = (idle->lock_depth >= 0);
4039 #else
4040         idle->thread_info->preempt_count = 0;
4041 #endif
4042 }
4043
4044 /*
4045  * In a system that switches off the HZ timer nohz_cpu_mask
4046  * indicates which cpus entered this state. This is used
4047  * in the rcu update to wait only for active cpus. For system
4048  * which do not switch off the HZ timer nohz_cpu_mask should
4049  * always be CPU_MASK_NONE.
4050  */
4051 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4052
4053 #ifdef CONFIG_SMP
4054 /*
4055  * This is how migration works:
4056  *
4057  * 1) we queue a migration_req_t structure in the source CPU's
4058  *    runqueue and wake up that CPU's migration thread.
4059  * 2) we down() the locked semaphore => thread blocks.
4060  * 3) migration thread wakes up (implicitly it forces the migrated
4061  *    thread off the CPU)
4062  * 4) it gets the migration request and checks whether the migrated
4063  *    task is still in the wrong runqueue.
4064  * 5) if it's in the wrong runqueue then the migration thread removes
4065  *    it and puts it into the right queue.
4066  * 6) migration thread up()s the semaphore.
4067  * 7) we wake up and the migration is done.
4068  */
4069
4070 /*
4071  * Change a given task's CPU affinity. Migrate the thread to a
4072  * proper CPU and schedule it away if the CPU it's executing on
4073  * is removed from the allowed bitmask.
4074  *
4075  * NOTE: the caller must have a valid reference to the task, the
4076  * task must not exit() & deallocate itself prematurely.  The
4077  * call is not atomic; no spinlocks may be held.
4078  */
4079 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4080 {
4081         unsigned long flags;
4082         int ret = 0;
4083         migration_req_t req;
4084         runqueue_t *rq;
4085
4086         rq = task_rq_lock(p, &flags);
4087         if (!cpus_intersects(new_mask, cpu_online_map)) {
4088                 ret = -EINVAL;
4089                 goto out;
4090         }
4091
4092         p->cpus_allowed = new_mask;
4093         /* Can the task run on the task's current CPU? If so, we're done */
4094         if (cpu_isset(task_cpu(p), new_mask))
4095                 goto out;
4096
4097         if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4098                 /* Need help from migration thread: drop lock and wait. */
4099                 task_rq_unlock(rq, &flags);
4100                 wake_up_process(rq->migration_thread);
4101                 wait_for_completion(&req.done);
4102                 tlb_migrate_finish(p->mm);
4103                 return 0;
4104         }
4105 out:
4106         task_rq_unlock(rq, &flags);
4107         return ret;
4108 }
4109
4110 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4111
4112 /*
4113  * Move (not current) task off this cpu, onto dest cpu.  We're doing
4114  * this because either it can't run here any more (set_cpus_allowed()
4115  * away from this CPU, or CPU going down), or because we're
4116  * attempting to rebalance this task on exec (sched_exec).
4117  *
4118  * So we race with normal scheduler movements, but that's OK, as long
4119  * as the task is no longer on this CPU.
4120  */
4121 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4122 {
4123         runqueue_t *rq_dest, *rq_src;
4124
4125         if (unlikely(cpu_is_offline(dest_cpu)))
4126                 return;
4127
4128         rq_src = cpu_rq(src_cpu);
4129         rq_dest = cpu_rq(dest_cpu);
4130
4131         double_rq_lock(rq_src, rq_dest);
4132         /* Already moved. */
4133         if (task_cpu(p) != src_cpu)
4134                 goto out;
4135         /* Affinity changed (again). */
4136         if (!cpu_isset(dest_cpu, p->cpus_allowed))
4137                 goto out;
4138
4139         set_task_cpu(p, dest_cpu);
4140         if (p->array) {
4141                 /*
4142                  * Sync timestamp with rq_dest's before activating.
4143                  * The same thing could be achieved by doing this step
4144                  * afterwards, and pretending it was a local activate.
4145                  * This way is cleaner and logically correct.
4146                  */
4147                 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4148                                 + rq_dest->timestamp_last_tick;
4149                 deactivate_task(p, rq_src);
4150                 activate_task(p, rq_dest, 0);
4151                 if (TASK_PREEMPTS_CURR(p, rq_dest))
4152                         resched_task(rq_dest->curr);
4153         }
4154
4155 out:
4156         double_rq_unlock(rq_src, rq_dest);
4157 }
4158
4159 /*
4160  * migration_thread - this is a highprio system thread that performs
4161  * thread migration by bumping thread off CPU then 'pushing' onto
4162  * another runqueue.
4163  */
4164 static int migration_thread(void * data)
4165 {
4166         runqueue_t *rq;
4167         int cpu = (long)data;
4168
4169         rq = cpu_rq(cpu);
4170         BUG_ON(rq->migration_thread != current);
4171
4172         set_current_state(TASK_INTERRUPTIBLE);
4173         while (!kthread_should_stop()) {
4174                 struct list_head *head;
4175                 migration_req_t *req;
4176
4177                 if (current->flags & PF_FREEZE)
4178                         refrigerator(PF_FREEZE);
4179
4180                 spin_lock_irq(&rq->lock);
4181
4182                 if (cpu_is_offline(cpu)) {
4183                         spin_unlock_irq(&rq->lock);
4184                         goto wait_to_die;
4185                 }
4186
4187                 if (rq->active_balance) {
4188                         active_load_balance(rq, cpu);
4189                         rq->active_balance = 0;
4190                 }
4191
4192                 head = &rq->migration_queue;
4193
4194                 if (list_empty(head)) {
4195                         spin_unlock_irq(&rq->lock);
4196                         schedule();
4197                         set_current_state(TASK_INTERRUPTIBLE);
4198                         continue;
4199                 }
4200                 req = list_entry(head->next, migration_req_t, list);
4201                 list_del_init(head->next);
4202
4203                 if (req->type == REQ_MOVE_TASK) {
4204                         spin_unlock(&rq->lock);
4205                         __migrate_task(req->task, cpu, req->dest_cpu);
4206                         local_irq_enable();
4207                 } else if (req->type == REQ_SET_DOMAIN) {
4208                         rq->sd = req->sd;
4209                         spin_unlock_irq(&rq->lock);
4210                 } else {
4211                         spin_unlock_irq(&rq->lock);
4212                         WARN_ON(1);
4213                 }
4214
4215                 complete(&req->done);
4216         }
4217         __set_current_state(TASK_RUNNING);
4218         return 0;
4219
4220 wait_to_die:
4221         /* Wait for kthread_stop */
4222         set_current_state(TASK_INTERRUPTIBLE);
4223         while (!kthread_should_stop()) {
4224                 schedule();
4225                 set_current_state(TASK_INTERRUPTIBLE);
4226         }
4227         __set_current_state(TASK_RUNNING);
4228         return 0;
4229 }
4230
4231 #ifdef CONFIG_HOTPLUG_CPU
4232 /* Figure out where task on dead CPU should go, use force if neccessary. */
4233 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4234 {
4235         int dest_cpu;
4236         cpumask_t mask;
4237
4238         /* On same node? */
4239         mask = node_to_cpumask(cpu_to_node(dead_cpu));
4240         cpus_and(mask, mask, tsk->cpus_allowed);
4241         dest_cpu = any_online_cpu(mask);
4242
4243         /* On any allowed CPU? */
4244         if (dest_cpu == NR_CPUS)
4245                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4246
4247         /* No more Mr. Nice Guy. */
4248         if (dest_cpu == NR_CPUS) {
4249                 cpus_setall(tsk->cpus_allowed);
4250                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4251
4252                 /*
4253                  * Don't tell them about moving exiting tasks or
4254                  * kernel threads (both mm NULL), since they never
4255                  * leave kernel.
4256                  */
4257                 if (tsk->mm && printk_ratelimit())
4258                         printk(KERN_INFO "process %d (%s) no "
4259                                "longer affine to cpu%d\n",
4260                                tsk->pid, tsk->comm, dead_cpu);
4261         }
4262         __migrate_task(tsk, dead_cpu, dest_cpu);
4263 }
4264
4265 /*
4266  * While a dead CPU has no uninterruptible tasks queued at this point,
4267  * it might still have a nonzero ->nr_uninterruptible counter, because
4268  * for performance reasons the counter is not stricly tracking tasks to
4269  * their home CPUs. So we just add the counter to another CPU's counter,
4270  * to keep the global sum constant after CPU-down:
4271  */
4272 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4273 {
4274         runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4275         unsigned long flags;
4276
4277         local_irq_save(flags);
4278         double_rq_lock(rq_src, rq_dest);
4279         rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4280         rq_src->nr_uninterruptible = 0;
4281         double_rq_unlock(rq_src, rq_dest);
4282         local_irq_restore(flags);
4283 }
4284
4285 /* Run through task list and migrate tasks from the dead cpu. */
4286 static void migrate_live_tasks(int src_cpu)
4287 {
4288         struct task_struct *tsk, *t;
4289
4290         write_lock_irq(&tasklist_lock);
4291
4292         do_each_thread(t, tsk) {
4293                 if (tsk == current)
4294                         continue;
4295
4296                 if (task_cpu(tsk) == src_cpu)
4297                         move_task_off_dead_cpu(src_cpu, tsk);
4298         } while_each_thread(t, tsk);
4299
4300         write_unlock_irq(&tasklist_lock);
4301 }
4302
4303 /* Schedules idle task to be the next runnable task on current CPU.
4304  * It does so by boosting its priority to highest possible and adding it to
4305  * the _front_ of runqueue. Used by CPU offline code.
4306  */
4307 void sched_idle_next(void)
4308 {
4309         int cpu = smp_processor_id();
4310         runqueue_t *rq = this_rq();
4311         struct task_struct *p = rq->idle;
4312         unsigned long flags;
4313
4314         /* cpu has to be offline */
4315         BUG_ON(cpu_online(cpu));
4316
4317         /* Strictly not necessary since rest of the CPUs are stopped by now
4318          * and interrupts disabled on current cpu.
4319          */
4320         spin_lock_irqsave(&rq->lock, flags);
4321
4322         __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4323         /* Add idle task to _front_ of it's priority queue */
4324         __activate_idle_task(p, rq);
4325
4326         spin_unlock_irqrestore(&rq->lock, flags);
4327 }
4328
4329 /* Ensures that the idle task is using init_mm right before its cpu goes
4330  * offline.
4331  */
4332 void idle_task_exit(void)
4333 {
4334         struct mm_struct *mm = current->active_mm;
4335
4336         BUG_ON(cpu_online(smp_processor_id()));
4337
4338         if (mm != &init_mm)
4339                 switch_mm(mm, &init_mm, current);
4340         mmdrop(mm);
4341 }
4342
4343 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4344 {
4345         struct runqueue *rq = cpu_rq(dead_cpu);
4346
4347         /* Must be exiting, otherwise would be on tasklist. */
4348         BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4349
4350         /* Cannot have done final schedule yet: would have vanished. */
4351         BUG_ON(tsk->flags & PF_DEAD);
4352
4353         get_task_struct(tsk);
4354
4355         /*
4356          * Drop lock around migration; if someone else moves it,
4357          * that's OK.  No task can be added to this CPU, so iteration is
4358          * fine.
4359          */
4360         spin_unlock_irq(&rq->lock);
4361         move_task_off_dead_cpu(dead_cpu, tsk);
4362         spin_lock_irq(&rq->lock);
4363
4364         put_task_struct(tsk);
4365 }
4366
4367 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4368 static void migrate_dead_tasks(unsigned int dead_cpu)
4369 {
4370         unsigned arr, i;
4371         struct runqueue *rq = cpu_rq(dead_cpu);
4372
4373         for (arr = 0; arr < 2; arr++) {
4374                 for (i = 0; i < MAX_PRIO; i++) {
4375                         struct list_head *list = &rq->arrays[arr].queue[i];
4376                         while (!list_empty(list))
4377                                 migrate_dead(dead_cpu,
4378                                              list_entry(list->next, task_t,
4379                                                         run_list));
4380                 }
4381         }
4382 }
4383 #endif /* CONFIG_HOTPLUG_CPU */
4384
4385 /*
4386  * migration_call - callback that gets triggered when a CPU is added.
4387  * Here we can start up the necessary migration thread for the new CPU.
4388  */
4389 static int migration_call(struct notifier_block *nfb, unsigned long action,
4390                           void *hcpu)
4391 {
4392         int cpu = (long)hcpu;
4393         struct task_struct *p;
4394         struct runqueue *rq;
4395         unsigned long flags;
4396
4397         switch (action) {
4398         case CPU_UP_PREPARE:
4399                 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4400                 if (IS_ERR(p))
4401                         return NOTIFY_BAD;
4402                 p->flags |= PF_NOFREEZE;
4403                 kthread_bind(p, cpu);
4404                 /* Must be high prio: stop_machine expects to yield to it. */
4405                 rq = task_rq_lock(p, &flags);
4406                 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4407                 task_rq_unlock(rq, &flags);
4408                 cpu_rq(cpu)->migration_thread = p;
4409                 break;
4410         case CPU_ONLINE:
4411                 /* Strictly unneccessary, as first user will wake it. */
4412                 wake_up_process(cpu_rq(cpu)->migration_thread);
4413                 break;
4414 #ifdef CONFIG_HOTPLUG_CPU
4415         case CPU_UP_CANCELED:
4416                 /* Unbind it from offline cpu so it can run.  Fall thru. */
4417                 kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id());
4418                 kthread_stop(cpu_rq(cpu)->migration_thread);
4419                 cpu_rq(cpu)->migration_thread = NULL;
4420                 break;
4421         case CPU_DEAD:
4422                 migrate_live_tasks(cpu);
4423                 rq = cpu_rq(cpu);
4424                 kthread_stop(rq->migration_thread);
4425                 rq->migration_thread = NULL;
4426                 /* Idle task back to normal (off runqueue, low prio) */
4427                 rq = task_rq_lock(rq->idle, &flags);
4428                 deactivate_task(rq->idle, rq);
4429                 rq->idle->static_prio = MAX_PRIO;
4430                 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4431                 migrate_dead_tasks(cpu);
4432                 task_rq_unlock(rq, &flags);
4433                 migrate_nr_uninterruptible(rq);
4434                 BUG_ON(rq->nr_running != 0);
4435
4436                 /* No need to migrate the tasks: it was best-effort if
4437                  * they didn't do lock_cpu_hotplug().  Just wake up
4438                  * the requestors. */
4439                 spin_lock_irq(&rq->lock);
4440                 while (!list_empty(&rq->migration_queue)) {
4441                         migration_req_t *req;
4442                         req = list_entry(rq->migration_queue.next,
4443                                          migration_req_t, list);
4444                         BUG_ON(req->type != REQ_MOVE_TASK);
4445                         list_del_init(&req->list);
4446                         complete(&req->done);
4447                 }
4448                 spin_unlock_irq(&rq->lock);
4449                 break;
4450 #endif
4451         }
4452         return NOTIFY_OK;
4453 }
4454
4455 /* Register at highest priority so that task migration (migrate_all_tasks)
4456  * happens before everything else.
4457  */
4458 static struct notifier_block __devinitdata migration_notifier = {
4459         .notifier_call = migration_call,
4460         .priority = 10
4461 };
4462
4463 int __init migration_init(void)
4464 {
4465         void *cpu = (void *)(long)smp_processor_id();
4466         /* Start one for boot CPU. */
4467         migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4468         migration_call(&migration_notifier, CPU_ONLINE, cpu);
4469         register_cpu_notifier(&migration_notifier);
4470         return 0;
4471 }
4472 #endif
4473
4474 #ifdef CONFIG_SMP
4475 #define SCHED_DOMAIN_DEBUG
4476 #ifdef SCHED_DOMAIN_DEBUG
4477 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4478 {
4479         int level = 0;
4480
4481         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4482
4483         do {
4484                 int i;
4485                 char str[NR_CPUS];
4486                 struct sched_group *group = sd->groups;
4487                 cpumask_t groupmask;
4488
4489                 cpumask_scnprintf(str, NR_CPUS, sd->span);
4490                 cpus_clear(groupmask);
4491
4492                 printk(KERN_DEBUG);
4493                 for (i = 0; i < level + 1; i++)
4494                         printk(" ");
4495                 printk("domain %d: ", level);
4496
4497                 if (!(sd->flags & SD_LOAD_BALANCE)) {
4498                         printk("does not load-balance\n");
4499                         if (sd->parent)
4500                                 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4501                         break;
4502                 }
4503
4504                 printk("span %s\n", str);
4505
4506                 if (!cpu_isset(cpu, sd->span))
4507                         printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4508                 if (!cpu_isset(cpu, group->cpumask))
4509                         printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4510
4511                 printk(KERN_DEBUG);
4512                 for (i = 0; i < level + 2; i++)
4513                         printk(" ");
4514                 printk("groups:");
4515                 do {
4516                         if (!group) {
4517                                 printk("\n");
4518                                 printk(KERN_ERR "ERROR: group is NULL\n");
4519                                 break;
4520                         }
4521
4522                         if (!group->cpu_power) {
4523                                 printk("\n");
4524                                 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4525                         }
4526
4527                         if (!cpus_weight(group->cpumask)) {
4528                                 printk("\n");
4529                                 printk(KERN_ERR "ERROR: empty group\n");
4530                         }
4531
4532                         if (cpus_intersects(groupmask, group->cpumask)) {
4533                                 printk("\n");
4534                                 printk(KERN_ERR "ERROR: repeated CPUs\n");
4535                         }
4536
4537                         cpus_or(groupmask, groupmask, group->cpumask);
4538
4539                         cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4540                         printk(" %s", str);
4541
4542                         group = group->next;
4543                 } while (group != sd->groups);
4544                 printk("\n");
4545
4546                 if (!cpus_equal(sd->span, groupmask))
4547                         printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4548
4549                 level++;
4550                 sd = sd->parent;
4551
4552                 if (sd) {
4553                         if (!cpus_subset(groupmask, sd->span))
4554                                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4555                 }
4556
4557         } while (sd);
4558 }
4559 #else
4560 #define sched_domain_debug(sd, cpu) {}
4561 #endif
4562
4563 /*
4564  * Attach the domain 'sd' to 'cpu' as its base domain.  Callers must
4565  * hold the hotplug lock.
4566  */
4567 void __devinit cpu_attach_domain(struct sched_domain *sd, int cpu)
4568 {
4569         migration_req_t req;
4570         unsigned long flags;
4571         runqueue_t *rq = cpu_rq(cpu);
4572         int local = 1;
4573
4574         sched_domain_debug(sd, cpu);
4575
4576         spin_lock_irqsave(&rq->lock, flags);
4577
4578         if (cpu == smp_processor_id() || !cpu_online(cpu)) {
4579                 rq->sd = sd;
4580         } else {
4581                 init_completion(&req.done);
4582                 req.type = REQ_SET_DOMAIN;
4583                 req.sd = sd;
4584                 list_add(&req.list, &rq->migration_queue);
4585                 local = 0;
4586         }
4587
4588         spin_unlock_irqrestore(&rq->lock, flags);
4589
4590         if (!local) {
4591                 wake_up_process(rq->migration_thread);
4592                 wait_for_completion(&req.done);
4593         }
4594 }
4595
4596 /* cpus with isolated domains */
4597 cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4598
4599 /* Setup the mask of cpus configured for isolated domains */
4600 static int __init isolated_cpu_setup(char *str)
4601 {
4602         int ints[NR_CPUS], i;
4603
4604         str = get_options(str, ARRAY_SIZE(ints), ints);
4605         cpus_clear(cpu_isolated_map);
4606         for (i = 1; i <= ints[0]; i++)
4607                 if (ints[i] < NR_CPUS)
4608                         cpu_set(ints[i], cpu_isolated_map);
4609         return 1;
4610 }
4611
4612 __setup ("isolcpus=", isolated_cpu_setup);
4613
4614 /*
4615  * init_sched_build_groups takes an array of groups, the cpumask we wish
4616  * to span, and a pointer to a function which identifies what group a CPU
4617  * belongs to. The return value of group_fn must be a valid index into the
4618  * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4619  * keep track of groups covered with a cpumask_t).
4620  *
4621  * init_sched_build_groups will build a circular linked list of the groups
4622  * covered by the given span, and will set each group's ->cpumask correctly,
4623  * and ->cpu_power to 0.
4624  */
4625 void __devinit init_sched_build_groups(struct sched_group groups[],
4626                         cpumask_t span, int (*group_fn)(int cpu))
4627 {
4628         struct sched_group *first = NULL, *last = NULL;
4629         cpumask_t covered = CPU_MASK_NONE;
4630         int i;
4631
4632         for_each_cpu_mask(i, span) {
4633                 int group = group_fn(i);
4634                 struct sched_group *sg = &groups[group];
4635                 int j;
4636
4637                 if (cpu_isset(i, covered))
4638                         continue;
4639
4640                 sg->cpumask = CPU_MASK_NONE;
4641                 sg->cpu_power = 0;
4642
4643                 for_each_cpu_mask(j, span) {
4644                         if (group_fn(j) != group)
4645                                 continue;
4646
4647                         cpu_set(j, covered);
4648                         cpu_set(j, sg->cpumask);
4649                 }
4650                 if (!first)
4651                         first = sg;
4652                 if (last)
4653                         last->next = sg;
4654                 last = sg;
4655         }
4656         last->next = first;
4657 }
4658
4659
4660 #ifdef ARCH_HAS_SCHED_DOMAIN
4661 extern void __devinit arch_init_sched_domains(void);
4662 extern void __devinit arch_destroy_sched_domains(void);
4663 #else
4664 #ifdef CONFIG_SCHED_SMT
4665 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
4666 static struct sched_group sched_group_cpus[NR_CPUS];
4667 static int __devinit cpu_to_cpu_group(int cpu)
4668 {
4669         return cpu;
4670 }
4671 #endif
4672
4673 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
4674 static struct sched_group sched_group_phys[NR_CPUS];
4675 static int __devinit cpu_to_phys_group(int cpu)
4676 {
4677 #ifdef CONFIG_SCHED_SMT
4678         return first_cpu(cpu_sibling_map[cpu]);
4679 #else
4680         return cpu;
4681 #endif
4682 }
4683
4684 #ifdef CONFIG_NUMA
4685
4686 static DEFINE_PER_CPU(struct sched_domain, node_domains);
4687 static struct sched_group sched_group_nodes[MAX_NUMNODES];
4688 static int __devinit cpu_to_node_group(int cpu)
4689 {
4690         return cpu_to_node(cpu);
4691 }
4692 #endif
4693
4694 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4695 /*
4696  * The domains setup code relies on siblings not spanning
4697  * multiple nodes. Make sure the architecture has a proper
4698  * siblings map:
4699  */
4700 static void check_sibling_maps(void)
4701 {
4702         int i, j;
4703
4704         for_each_online_cpu(i) {
4705                 for_each_cpu_mask(j, cpu_sibling_map[i]) {
4706                         if (cpu_to_node(i) != cpu_to_node(j)) {
4707                                 printk(KERN_INFO "warning: CPU %d siblings map "
4708                                         "to different node - isolating "
4709                                         "them.\n", i);
4710                                 cpu_sibling_map[i] = cpumask_of_cpu(i);
4711                                 break;
4712                         }
4713                 }
4714         }
4715 }
4716 #endif
4717
4718 /*
4719  * Set up scheduler domains and groups.  Callers must hold the hotplug lock.
4720  */
4721 static void __devinit arch_init_sched_domains(void)
4722 {
4723         int i;
4724         cpumask_t cpu_default_map;
4725
4726 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4727         check_sibling_maps();
4728 #endif
4729         /*
4730          * Setup mask for cpus without special case scheduling requirements.
4731          * For now this just excludes isolated cpus, but could be used to
4732          * exclude other special cases in the future.
4733          */
4734         cpus_complement(cpu_default_map, cpu_isolated_map);
4735         cpus_and(cpu_default_map, cpu_default_map, cpu_online_map);
4736
4737         /*
4738          * Set up domains. Isolated domains just stay on the dummy domain.
4739          */
4740         for_each_cpu_mask(i, cpu_default_map) {
4741                 int group;
4742                 struct sched_domain *sd = NULL, *p;
4743                 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
4744
4745                 cpus_and(nodemask, nodemask, cpu_default_map);
4746
4747 #ifdef CONFIG_NUMA
4748                 sd = &per_cpu(node_domains, i);
4749                 group = cpu_to_node_group(i);
4750                 *sd = SD_NODE_INIT;
4751                 sd->span = cpu_default_map;
4752                 sd->groups = &sched_group_nodes[group];
4753 #endif
4754
4755                 p = sd;
4756                 sd = &per_cpu(phys_domains, i);
4757                 group = cpu_to_phys_group(i);
4758                 *sd = SD_CPU_INIT;
4759                 sd->span = nodemask;
4760                 sd->parent = p;
4761                 sd->groups = &sched_group_phys[group];
4762
4763 #ifdef CONFIG_SCHED_SMT
4764                 p = sd;
4765                 sd = &per_cpu(cpu_domains, i);
4766                 group = cpu_to_cpu_group(i);
4767                 *sd = SD_SIBLING_INIT;
4768                 sd->span = cpu_sibling_map[i];
4769                 cpus_and(sd->span, sd->span, cpu_default_map);
4770                 sd->parent = p;
4771                 sd->groups = &sched_group_cpus[group];
4772 #endif
4773         }
4774
4775 #ifdef CONFIG_SCHED_SMT
4776         /* Set up CPU (sibling) groups */
4777         for_each_online_cpu(i) {
4778                 cpumask_t this_sibling_map = cpu_sibling_map[i];
4779                 cpus_and(this_sibling_map, this_sibling_map, cpu_default_map);
4780                 if (i != first_cpu(this_sibling_map))
4781                         continue;
4782
4783                 init_sched_build_groups(sched_group_cpus, this_sibling_map,
4784                                                 &cpu_to_cpu_group);
4785         }
4786 #endif
4787
4788         /* Set up physical groups */
4789         for (i = 0; i < MAX_NUMNODES; i++) {
4790                 cpumask_t nodemask = node_to_cpumask(i);
4791
4792                 cpus_and(nodemask, nodemask, cpu_default_map);
4793                 if (cpus_empty(nodemask))
4794                         continue;
4795
4796                 init_sched_build_groups(sched_group_phys, nodemask,
4797                                                 &cpu_to_phys_group);
4798         }
4799
4800 #ifdef CONFIG_NUMA
4801         /* Set up node groups */
4802         init_sched_build_groups(sched_group_nodes, cpu_default_map,
4803                                         &cpu_to_node_group);
4804 #endif
4805
4806         /* Calculate CPU power for physical packages and nodes */
4807         for_each_cpu_mask(i, cpu_default_map) {
4808                 int power;
4809                 struct sched_domain *sd;
4810 #ifdef CONFIG_SCHED_SMT
4811                 sd = &per_cpu(cpu_domains, i);
4812                 power = SCHED_LOAD_SCALE;
4813                 sd->groups->cpu_power = power;
4814 #endif
4815
4816                 sd = &per_cpu(phys_domains, i);
4817                 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
4818                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
4819                 sd->groups->cpu_power = power;
4820
4821 #ifdef CONFIG_NUMA
4822                 if (i == first_cpu(sd->groups->cpumask)) {
4823                         /* Only add "power" once for each physical package. */
4824                         sd = &per_cpu(node_domains, i);
4825                         sd->groups->cpu_power += power;
4826                 }
4827 #endif
4828         }
4829
4830         /* Attach the domains */
4831         for_each_online_cpu(i) {
4832                 struct sched_domain *sd;
4833 #ifdef CONFIG_SCHED_SMT
4834                 sd = &per_cpu(cpu_domains, i);
4835 #else
4836                 sd = &per_cpu(phys_domains, i);
4837 #endif
4838                 cpu_attach_domain(sd, i);
4839         }
4840 }
4841
4842 #ifdef CONFIG_HOTPLUG_CPU
4843 static void __devinit arch_destroy_sched_domains(void)
4844 {
4845         /* Do nothing: everything is statically allocated. */
4846 }
4847 #endif
4848
4849 #endif /* ARCH_HAS_SCHED_DOMAIN */
4850
4851 /*
4852  * Initial dummy domain for early boot and for hotplug cpu. Being static,
4853  * it is initialized to zero, so all balancing flags are cleared which is
4854  * what we want.
4855  */
4856 static struct sched_domain sched_domain_dummy;
4857
4858 #ifdef CONFIG_HOTPLUG_CPU
4859 /*
4860  * Force a reinitialization of the sched domains hierarchy.  The domains
4861  * and groups cannot be updated in place without racing with the balancing
4862  * code, so we temporarily attach all running cpus to a "dummy" domain
4863  * which will prevent rebalancing while the sched domains are recalculated.
4864  */
4865 static int update_sched_domains(struct notifier_block *nfb,
4866                                 unsigned long action, void *hcpu)
4867 {
4868         int i;
4869
4870         switch (action) {
4871         case CPU_UP_PREPARE:
4872         case CPU_DOWN_PREPARE:
4873                 for_each_online_cpu(i)
4874                         cpu_attach_domain(&sched_domain_dummy, i);
4875                 arch_destroy_sched_domains();
4876                 return NOTIFY_OK;
4877
4878         case CPU_UP_CANCELED:
4879         case CPU_DOWN_FAILED:
4880         case CPU_ONLINE:
4881         case CPU_DEAD:
4882                 /*
4883                  * Fall through and re-initialise the domains.
4884                  */
4885                 break;
4886         default:
4887                 return NOTIFY_DONE;
4888         }
4889
4890         /* The hotplug lock is already held by cpu_up/cpu_down */
4891         arch_init_sched_domains();
4892
4893         return NOTIFY_OK;
4894 }
4895 #endif
4896
4897 void __init sched_init_smp(void)
4898 {
4899         lock_cpu_hotplug();
4900         arch_init_sched_domains();
4901         unlock_cpu_hotplug();
4902         /* XXX: Theoretical race here - CPU may be hotplugged now */
4903         hotcpu_notifier(update_sched_domains, 0);
4904 }
4905 #else
4906 void __init sched_init_smp(void)
4907 {
4908 }
4909 #endif /* CONFIG_SMP */
4910
4911 int in_sched_functions(unsigned long addr)
4912 {
4913         /* Linker adds these: start and end of __sched functions */
4914         extern char __sched_text_start[], __sched_text_end[];
4915         return in_lock_functions(addr) ||
4916                 (addr >= (unsigned long)__sched_text_start
4917                 && addr < (unsigned long)__sched_text_end);
4918 }
4919
4920 void __init sched_init(void)
4921 {
4922         runqueue_t *rq;
4923         int i, j, k;
4924
4925         for (i = 0; i < NR_CPUS; i++) {
4926                 prio_array_t *array;
4927
4928                 rq = cpu_rq(i);
4929                 spin_lock_init(&rq->lock);
4930                 rq->active = rq->arrays;
4931                 rq->expired = rq->arrays + 1;
4932                 rq->best_expired_prio = MAX_PRIO;
4933
4934 #ifdef CONFIG_SMP
4935                 rq->sd = &sched_domain_dummy;
4936                 rq->cpu_load = 0;
4937                 rq->active_balance = 0;
4938                 rq->push_cpu = 0;
4939                 rq->migration_thread = NULL;
4940                 INIT_LIST_HEAD(&rq->migration_queue);
4941 #endif
4942                 atomic_set(&rq->nr_iowait, 0);
4943
4944                 for (j = 0; j < 2; j++) {
4945                         array = rq->arrays + j;
4946                         for (k = 0; k < MAX_PRIO; k++) {
4947                                 INIT_LIST_HEAD(array->queue + k);
4948                                 __clear_bit(k, array->bitmap);
4949                         }
4950                         // delimiter for bitsearch
4951                         __set_bit(MAX_PRIO, array->bitmap);
4952                 }
4953         }
4954
4955         /*
4956          * The boot idle thread does lazy MMU switching as well:
4957          */
4958         atomic_inc(&init_mm.mm_count);
4959         enter_lazy_tlb(&init_mm, current);
4960
4961         /*
4962          * Make us the idle thread. Technically, schedule() should not be
4963          * called from this thread, however somewhere below it might be,
4964          * but because we are the idle thread, we just pick up running again
4965          * when this runqueue becomes "idle".
4966          */
4967         init_idle(current, smp_processor_id());
4968 }
4969
4970 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4971 void __might_sleep(char *file, int line)
4972 {
4973 #if defined(in_atomic)
4974         static unsigned long prev_jiffy;        /* ratelimiting */
4975
4976         if ((in_atomic() || irqs_disabled()) &&
4977             system_state == SYSTEM_RUNNING && !oops_in_progress) {
4978                 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
4979                         return;
4980                 prev_jiffy = jiffies;
4981                 printk(KERN_ERR "Debug: sleeping function called from invalid"
4982                                 " context at %s:%d\n", file, line);
4983                 printk("in_atomic():%d, irqs_disabled():%d\n",
4984                         in_atomic(), irqs_disabled());
4985                 dump_stack();
4986         }
4987 #endif
4988 }
4989 EXPORT_SYMBOL(__might_sleep);
4990 #endif
4991
4992 #ifdef CONFIG_MAGIC_SYSRQ
4993 void normalize_rt_tasks(void)
4994 {
4995         struct task_struct *p;
4996         prio_array_t *array;
4997         unsigned long flags;
4998         runqueue_t *rq;
4999
5000         read_lock_irq(&tasklist_lock);
5001         for_each_process (p) {
5002                 if (!rt_task(p))
5003                         continue;
5004
5005                 rq = task_rq_lock(p, &flags);
5006
5007                 array = p->array;
5008                 if (array)
5009                         deactivate_task(p, task_rq(p));
5010                 __setscheduler(p, SCHED_NORMAL, 0);
5011                 if (array) {
5012                         __activate_task(p, task_rq(p));
5013                         resched_task(rq->curr);
5014                 }
5015
5016                 task_rq_unlock(rq, &flags);
5017         }
5018         read_unlock_irq(&tasklist_lock);
5019 }
5020
5021 #endif /* CONFIG_MAGIC_SYSRQ */