[PATCH] sched: HT optimisation
[linux-2.6.git] / kernel / sched.c
1 /*
2  *  kernel/sched.c
3  *
4  *  Kernel scheduler and related syscalls
5  *
6  *  Copyright (C) 1991-2002  Linus Torvalds
7  *
8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
9  *              make semaphores SMP safe
10  *  1998-11-19  Implemented schedule_timeout() and related stuff
11  *              by Andrea Arcangeli
12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
13  *              hybrid priority-list and round-robin design with
14  *              an array-switch method of distributing timeslices
15  *              and per-CPU runqueues.  Cleanups and useful suggestions
16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
17  *  2003-09-03  Interactivity tuning by Con Kolivas.
18  *  2004-04-02  Scheduler domains code by Nick Piggin
19  */
20
21 #include <linux/mm.h>
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/completion.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/security.h>
33 #include <linux/notifier.h>
34 #include <linux/profile.h>
35 #include <linux/suspend.h>
36 #include <linux/blkdev.h>
37 #include <linux/delay.h>
38 #include <linux/smp.h>
39 #include <linux/threads.h>
40 #include <linux/timer.h>
41 #include <linux/rcupdate.h>
42 #include <linux/cpu.h>
43 #include <linux/cpuset.h>
44 #include <linux/percpu.h>
45 #include <linux/kthread.h>
46 #include <linux/seq_file.h>
47 #include <linux/syscalls.h>
48 #include <linux/times.h>
49 #include <linux/acct.h>
50 #include <asm/tlb.h>
51
52 #include <asm/unistd.h>
53
54 /*
55  * Convert user-nice values [ -20 ... 0 ... 19 ]
56  * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
57  * and back.
58  */
59 #define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
60 #define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
61 #define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
62
63 /*
64  * 'User priority' is the nice value converted to something we
65  * can work with better when scaling various scheduler parameters,
66  * it's a [ 0 ... 39 ] range.
67  */
68 #define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
69 #define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
70 #define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
71
72 /*
73  * Some helpers for converting nanosecond timing to jiffy resolution
74  */
75 #define NS_TO_JIFFIES(TIME)     ((TIME) / (1000000000 / HZ))
76 #define JIFFIES_TO_NS(TIME)     ((TIME) * (1000000000 / HZ))
77
78 /*
79  * These are the 'tuning knobs' of the scheduler:
80  *
81  * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
82  * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
83  * Timeslices get refilled after they expire.
84  */
85 #define MIN_TIMESLICE           max(5 * HZ / 1000, 1)
86 #define DEF_TIMESLICE           (100 * HZ / 1000)
87 #define ON_RUNQUEUE_WEIGHT       30
88 #define CHILD_PENALTY            95
89 #define PARENT_PENALTY          100
90 #define EXIT_WEIGHT               3
91 #define PRIO_BONUS_RATIO         25
92 #define MAX_BONUS               (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
93 #define INTERACTIVE_DELTA         2
94 #define MAX_SLEEP_AVG           (DEF_TIMESLICE * MAX_BONUS)
95 #define STARVATION_LIMIT        (MAX_SLEEP_AVG)
96 #define NS_MAX_SLEEP_AVG        (JIFFIES_TO_NS(MAX_SLEEP_AVG))
97
98 /*
99  * If a task is 'interactive' then we reinsert it in the active
100  * array after it has expired its current timeslice. (it will not
101  * continue to run immediately, it will still roundrobin with
102  * other interactive tasks.)
103  *
104  * This part scales the interactivity limit depending on niceness.
105  *
106  * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
107  * Here are a few examples of different nice levels:
108  *
109  *  TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
110  *  TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
111  *  TASK_INTERACTIVE(  0): [1,1,1,1,0,0,0,0,0,0,0]
112  *  TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
113  *  TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
114  *
115  * (the X axis represents the possible -5 ... 0 ... +5 dynamic
116  *  priority range a task can explore, a value of '1' means the
117  *  task is rated interactive.)
118  *
119  * Ie. nice +19 tasks can never get 'interactive' enough to be
120  * reinserted into the active array. And only heavily CPU-hog nice -20
121  * tasks will be expired. Default nice 0 tasks are somewhere between,
122  * it takes some effort for them to get interactive, but it's not
123  * too hard.
124  */
125
126 #define CURRENT_BONUS(p) \
127         (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
128                 MAX_SLEEP_AVG)
129
130 #define GRANULARITY     (10 * HZ / 1000 ? : 1)
131
132 #ifdef CONFIG_SMP
133 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
134                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
135                         num_online_cpus())
136 #else
137 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
138                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
139 #endif
140
141 #define SCALE(v1,v1_max,v2_max) \
142         (v1) * (v2_max) / (v1_max)
143
144 #define DELTA(p) \
145         (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
146
147 #define TASK_INTERACTIVE(p) \
148         ((p)->prio <= (p)->static_prio - DELTA(p))
149
150 #define INTERACTIVE_SLEEP(p) \
151         (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
152                 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
153
154 #define TASK_PREEMPTS_CURR(p, rq) \
155         ((p)->prio < (rq)->curr->prio)
156
157 /*
158  * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
159  * to time slice values: [800ms ... 100ms ... 5ms]
160  *
161  * The higher a thread's priority, the bigger timeslices
162  * it gets during one round of execution. But even the lowest
163  * priority thread gets MIN_TIMESLICE worth of execution time.
164  */
165
166 #define SCALE_PRIO(x, prio) \
167         max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
168
169 static unsigned int task_timeslice(task_t *p)
170 {
171         if (p->static_prio < NICE_TO_PRIO(0))
172                 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
173         else
174                 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
175 }
176 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran)       \
177                                 < (long long) (sd)->cache_hot_time)
178
179 /*
180  * These are the runqueue data structures:
181  */
182
183 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
184
185 typedef struct runqueue runqueue_t;
186
187 struct prio_array {
188         unsigned int nr_active;
189         unsigned long bitmap[BITMAP_SIZE];
190         struct list_head queue[MAX_PRIO];
191 };
192
193 /*
194  * This is the main, per-CPU runqueue data structure.
195  *
196  * Locking rule: those places that want to lock multiple runqueues
197  * (such as the load balancing or the thread migration code), lock
198  * acquire operations must be ordered by ascending &runqueue.
199  */
200 struct runqueue {
201         spinlock_t lock;
202
203         /*
204          * nr_running and cpu_load should be in the same cacheline because
205          * remote CPUs use both these fields when doing load calculation.
206          */
207         unsigned long nr_running;
208 #ifdef CONFIG_SMP
209         unsigned long cpu_load[3];
210 #endif
211         unsigned long long nr_switches;
212
213         /*
214          * This is part of a global counter where only the total sum
215          * over all CPUs matters. A task can increase this counter on
216          * one CPU and if it got migrated afterwards it may decrease
217          * it on another CPU. Always updated under the runqueue lock:
218          */
219         unsigned long nr_uninterruptible;
220
221         unsigned long expired_timestamp;
222         unsigned long long timestamp_last_tick;
223         task_t *curr, *idle;
224         struct mm_struct *prev_mm;
225         prio_array_t *active, *expired, arrays[2];
226         int best_expired_prio;
227         atomic_t nr_iowait;
228
229 #ifdef CONFIG_SMP
230         struct sched_domain *sd;
231
232         /* For active balancing */
233         int active_balance;
234         int push_cpu;
235
236         task_t *migration_thread;
237         struct list_head migration_queue;
238 #endif
239
240 #ifdef CONFIG_SCHEDSTATS
241         /* latency stats */
242         struct sched_info rq_sched_info;
243
244         /* sys_sched_yield() stats */
245         unsigned long yld_exp_empty;
246         unsigned long yld_act_empty;
247         unsigned long yld_both_empty;
248         unsigned long yld_cnt;
249
250         /* schedule() stats */
251         unsigned long sched_switch;
252         unsigned long sched_cnt;
253         unsigned long sched_goidle;
254
255         /* try_to_wake_up() stats */
256         unsigned long ttwu_cnt;
257         unsigned long ttwu_local;
258 #endif
259 };
260
261 static DEFINE_PER_CPU(struct runqueue, runqueues);
262
263 /*
264  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
265  * See detach_destroy_domains: synchronize_sched for details.
266  *
267  * The domain tree of any CPU may only be accessed from within
268  * preempt-disabled sections.
269  */
270 #define for_each_domain(cpu, domain) \
271 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
272
273 #define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
274 #define this_rq()               (&__get_cpu_var(runqueues))
275 #define task_rq(p)              cpu_rq(task_cpu(p))
276 #define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
277
278 #ifndef prepare_arch_switch
279 # define prepare_arch_switch(next)      do { } while (0)
280 #endif
281 #ifndef finish_arch_switch
282 # define finish_arch_switch(prev)       do { } while (0)
283 #endif
284
285 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
286 static inline int task_running(runqueue_t *rq, task_t *p)
287 {
288         return rq->curr == p;
289 }
290
291 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
292 {
293 }
294
295 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
296 {
297         spin_unlock_irq(&rq->lock);
298 }
299
300 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
301 static inline int task_running(runqueue_t *rq, task_t *p)
302 {
303 #ifdef CONFIG_SMP
304         return p->oncpu;
305 #else
306         return rq->curr == p;
307 #endif
308 }
309
310 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
311 {
312 #ifdef CONFIG_SMP
313         /*
314          * We can optimise this out completely for !SMP, because the
315          * SMP rebalancing from interrupt is the only thing that cares
316          * here.
317          */
318         next->oncpu = 1;
319 #endif
320 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
321         spin_unlock_irq(&rq->lock);
322 #else
323         spin_unlock(&rq->lock);
324 #endif
325 }
326
327 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
328 {
329 #ifdef CONFIG_SMP
330         /*
331          * After ->oncpu is cleared, the task can be moved to a different CPU.
332          * We must ensure this doesn't happen until the switch is completely
333          * finished.
334          */
335         smp_wmb();
336         prev->oncpu = 0;
337 #endif
338 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
339         local_irq_enable();
340 #endif
341 }
342 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
343
344 /*
345  * task_rq_lock - lock the runqueue a given task resides on and disable
346  * interrupts.  Note the ordering: we can safely lookup the task_rq without
347  * explicitly disabling preemption.
348  */
349 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
350         __acquires(rq->lock)
351 {
352         struct runqueue *rq;
353
354 repeat_lock_task:
355         local_irq_save(*flags);
356         rq = task_rq(p);
357         spin_lock(&rq->lock);
358         if (unlikely(rq != task_rq(p))) {
359                 spin_unlock_irqrestore(&rq->lock, *flags);
360                 goto repeat_lock_task;
361         }
362         return rq;
363 }
364
365 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
366         __releases(rq->lock)
367 {
368         spin_unlock_irqrestore(&rq->lock, *flags);
369 }
370
371 #ifdef CONFIG_SCHEDSTATS
372 /*
373  * bump this up when changing the output format or the meaning of an existing
374  * format, so that tools can adapt (or abort)
375  */
376 #define SCHEDSTAT_VERSION 12
377
378 static int show_schedstat(struct seq_file *seq, void *v)
379 {
380         int cpu;
381
382         seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
383         seq_printf(seq, "timestamp %lu\n", jiffies);
384         for_each_online_cpu(cpu) {
385                 runqueue_t *rq = cpu_rq(cpu);
386 #ifdef CONFIG_SMP
387                 struct sched_domain *sd;
388                 int dcnt = 0;
389 #endif
390
391                 /* runqueue-specific stats */
392                 seq_printf(seq,
393                     "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
394                     cpu, rq->yld_both_empty,
395                     rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
396                     rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
397                     rq->ttwu_cnt, rq->ttwu_local,
398                     rq->rq_sched_info.cpu_time,
399                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
400
401                 seq_printf(seq, "\n");
402
403 #ifdef CONFIG_SMP
404                 /* domain-specific stats */
405                 preempt_disable();
406                 for_each_domain(cpu, sd) {
407                         enum idle_type itype;
408                         char mask_str[NR_CPUS];
409
410                         cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
411                         seq_printf(seq, "domain%d %s", dcnt++, mask_str);
412                         for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
413                                         itype++) {
414                                 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
415                                     sd->lb_cnt[itype],
416                                     sd->lb_balanced[itype],
417                                     sd->lb_failed[itype],
418                                     sd->lb_imbalance[itype],
419                                     sd->lb_gained[itype],
420                                     sd->lb_hot_gained[itype],
421                                     sd->lb_nobusyq[itype],
422                                     sd->lb_nobusyg[itype]);
423                         }
424                         seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
425                             sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
426                             sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
427                             sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
428                             sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
429                 }
430                 preempt_enable();
431 #endif
432         }
433         return 0;
434 }
435
436 static int schedstat_open(struct inode *inode, struct file *file)
437 {
438         unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
439         char *buf = kmalloc(size, GFP_KERNEL);
440         struct seq_file *m;
441         int res;
442
443         if (!buf)
444                 return -ENOMEM;
445         res = single_open(file, show_schedstat, NULL);
446         if (!res) {
447                 m = file->private_data;
448                 m->buf = buf;
449                 m->size = size;
450         } else
451                 kfree(buf);
452         return res;
453 }
454
455 struct file_operations proc_schedstat_operations = {
456         .open    = schedstat_open,
457         .read    = seq_read,
458         .llseek  = seq_lseek,
459         .release = single_release,
460 };
461
462 # define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
463 # define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
464 #else /* !CONFIG_SCHEDSTATS */
465 # define schedstat_inc(rq, field)       do { } while (0)
466 # define schedstat_add(rq, field, amt)  do { } while (0)
467 #endif
468
469 /*
470  * rq_lock - lock a given runqueue and disable interrupts.
471  */
472 static inline runqueue_t *this_rq_lock(void)
473         __acquires(rq->lock)
474 {
475         runqueue_t *rq;
476
477         local_irq_disable();
478         rq = this_rq();
479         spin_lock(&rq->lock);
480
481         return rq;
482 }
483
484 #ifdef CONFIG_SCHEDSTATS
485 /*
486  * Called when a process is dequeued from the active array and given
487  * the cpu.  We should note that with the exception of interactive
488  * tasks, the expired queue will become the active queue after the active
489  * queue is empty, without explicitly dequeuing and requeuing tasks in the
490  * expired queue.  (Interactive tasks may be requeued directly to the
491  * active queue, thus delaying tasks in the expired queue from running;
492  * see scheduler_tick()).
493  *
494  * This function is only called from sched_info_arrive(), rather than
495  * dequeue_task(). Even though a task may be queued and dequeued multiple
496  * times as it is shuffled about, we're really interested in knowing how
497  * long it was from the *first* time it was queued to the time that it
498  * finally hit a cpu.
499  */
500 static inline void sched_info_dequeued(task_t *t)
501 {
502         t->sched_info.last_queued = 0;
503 }
504
505 /*
506  * Called when a task finally hits the cpu.  We can now calculate how
507  * long it was waiting to run.  We also note when it began so that we
508  * can keep stats on how long its timeslice is.
509  */
510 static inline void sched_info_arrive(task_t *t)
511 {
512         unsigned long now = jiffies, diff = 0;
513         struct runqueue *rq = task_rq(t);
514
515         if (t->sched_info.last_queued)
516                 diff = now - t->sched_info.last_queued;
517         sched_info_dequeued(t);
518         t->sched_info.run_delay += diff;
519         t->sched_info.last_arrival = now;
520         t->sched_info.pcnt++;
521
522         if (!rq)
523                 return;
524
525         rq->rq_sched_info.run_delay += diff;
526         rq->rq_sched_info.pcnt++;
527 }
528
529 /*
530  * Called when a process is queued into either the active or expired
531  * array.  The time is noted and later used to determine how long we
532  * had to wait for us to reach the cpu.  Since the expired queue will
533  * become the active queue after active queue is empty, without dequeuing
534  * and requeuing any tasks, we are interested in queuing to either. It
535  * is unusual but not impossible for tasks to be dequeued and immediately
536  * requeued in the same or another array: this can happen in sched_yield(),
537  * set_user_nice(), and even load_balance() as it moves tasks from runqueue
538  * to runqueue.
539  *
540  * This function is only called from enqueue_task(), but also only updates
541  * the timestamp if it is already not set.  It's assumed that
542  * sched_info_dequeued() will clear that stamp when appropriate.
543  */
544 static inline void sched_info_queued(task_t *t)
545 {
546         if (!t->sched_info.last_queued)
547                 t->sched_info.last_queued = jiffies;
548 }
549
550 /*
551  * Called when a process ceases being the active-running process, either
552  * voluntarily or involuntarily.  Now we can calculate how long we ran.
553  */
554 static inline void sched_info_depart(task_t *t)
555 {
556         struct runqueue *rq = task_rq(t);
557         unsigned long diff = jiffies - t->sched_info.last_arrival;
558
559         t->sched_info.cpu_time += diff;
560
561         if (rq)
562                 rq->rq_sched_info.cpu_time += diff;
563 }
564
565 /*
566  * Called when tasks are switched involuntarily due, typically, to expiring
567  * their time slice.  (This may also be called when switching to or from
568  * the idle task.)  We are only called when prev != next.
569  */
570 static inline void sched_info_switch(task_t *prev, task_t *next)
571 {
572         struct runqueue *rq = task_rq(prev);
573
574         /*
575          * prev now departs the cpu.  It's not interesting to record
576          * stats about how efficient we were at scheduling the idle
577          * process, however.
578          */
579         if (prev != rq->idle)
580                 sched_info_depart(prev);
581
582         if (next != rq->idle)
583                 sched_info_arrive(next);
584 }
585 #else
586 #define sched_info_queued(t)            do { } while (0)
587 #define sched_info_switch(t, next)      do { } while (0)
588 #endif /* CONFIG_SCHEDSTATS */
589
590 /*
591  * Adding/removing a task to/from a priority array:
592  */
593 static void dequeue_task(struct task_struct *p, prio_array_t *array)
594 {
595         array->nr_active--;
596         list_del(&p->run_list);
597         if (list_empty(array->queue + p->prio))
598                 __clear_bit(p->prio, array->bitmap);
599 }
600
601 static void enqueue_task(struct task_struct *p, prio_array_t *array)
602 {
603         sched_info_queued(p);
604         list_add_tail(&p->run_list, array->queue + p->prio);
605         __set_bit(p->prio, array->bitmap);
606         array->nr_active++;
607         p->array = array;
608 }
609
610 /*
611  * Put task to the end of the run list without the overhead of dequeue
612  * followed by enqueue.
613  */
614 static void requeue_task(struct task_struct *p, prio_array_t *array)
615 {
616         list_move_tail(&p->run_list, array->queue + p->prio);
617 }
618
619 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
620 {
621         list_add(&p->run_list, array->queue + p->prio);
622         __set_bit(p->prio, array->bitmap);
623         array->nr_active++;
624         p->array = array;
625 }
626
627 /*
628  * effective_prio - return the priority that is based on the static
629  * priority but is modified by bonuses/penalties.
630  *
631  * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
632  * into the -5 ... 0 ... +5 bonus/penalty range.
633  *
634  * We use 25% of the full 0...39 priority range so that:
635  *
636  * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
637  * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
638  *
639  * Both properties are important to certain workloads.
640  */
641 static int effective_prio(task_t *p)
642 {
643         int bonus, prio;
644
645         if (rt_task(p))
646                 return p->prio;
647
648         bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
649
650         prio = p->static_prio - bonus;
651         if (prio < MAX_RT_PRIO)
652                 prio = MAX_RT_PRIO;
653         if (prio > MAX_PRIO-1)
654                 prio = MAX_PRIO-1;
655         return prio;
656 }
657
658 /*
659  * __activate_task - move a task to the runqueue.
660  */
661 static inline void __activate_task(task_t *p, runqueue_t *rq)
662 {
663         enqueue_task(p, rq->active);
664         rq->nr_running++;
665 }
666
667 /*
668  * __activate_idle_task - move idle task to the _front_ of runqueue.
669  */
670 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
671 {
672         enqueue_task_head(p, rq->active);
673         rq->nr_running++;
674 }
675
676 static int recalc_task_prio(task_t *p, unsigned long long now)
677 {
678         /* Caller must always ensure 'now >= p->timestamp' */
679         unsigned long long __sleep_time = now - p->timestamp;
680         unsigned long sleep_time;
681
682         if (__sleep_time > NS_MAX_SLEEP_AVG)
683                 sleep_time = NS_MAX_SLEEP_AVG;
684         else
685                 sleep_time = (unsigned long)__sleep_time;
686
687         if (likely(sleep_time > 0)) {
688                 /*
689                  * User tasks that sleep a long time are categorised as
690                  * idle and will get just interactive status to stay active &
691                  * prevent them suddenly becoming cpu hogs and starving
692                  * other processes.
693                  */
694                 if (p->mm && p->activated != -1 &&
695                         sleep_time > INTERACTIVE_SLEEP(p)) {
696                                 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
697                                                 DEF_TIMESLICE);
698                 } else {
699                         /*
700                          * The lower the sleep avg a task has the more
701                          * rapidly it will rise with sleep time.
702                          */
703                         sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
704
705                         /*
706                          * Tasks waking from uninterruptible sleep are
707                          * limited in their sleep_avg rise as they
708                          * are likely to be waiting on I/O
709                          */
710                         if (p->activated == -1 && p->mm) {
711                                 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
712                                         sleep_time = 0;
713                                 else if (p->sleep_avg + sleep_time >=
714                                                 INTERACTIVE_SLEEP(p)) {
715                                         p->sleep_avg = INTERACTIVE_SLEEP(p);
716                                         sleep_time = 0;
717                                 }
718                         }
719
720                         /*
721                          * This code gives a bonus to interactive tasks.
722                          *
723                          * The boost works by updating the 'average sleep time'
724                          * value here, based on ->timestamp. The more time a
725                          * task spends sleeping, the higher the average gets -
726                          * and the higher the priority boost gets as well.
727                          */
728                         p->sleep_avg += sleep_time;
729
730                         if (p->sleep_avg > NS_MAX_SLEEP_AVG)
731                                 p->sleep_avg = NS_MAX_SLEEP_AVG;
732                 }
733         }
734
735         return effective_prio(p);
736 }
737
738 /*
739  * activate_task - move a task to the runqueue and do priority recalculation
740  *
741  * Update all the scheduling statistics stuff. (sleep average
742  * calculation, priority modifiers, etc.)
743  */
744 static void activate_task(task_t *p, runqueue_t *rq, int local)
745 {
746         unsigned long long now;
747
748         now = sched_clock();
749 #ifdef CONFIG_SMP
750         if (!local) {
751                 /* Compensate for drifting sched_clock */
752                 runqueue_t *this_rq = this_rq();
753                 now = (now - this_rq->timestamp_last_tick)
754                         + rq->timestamp_last_tick;
755         }
756 #endif
757
758         p->prio = recalc_task_prio(p, now);
759
760         /*
761          * This checks to make sure it's not an uninterruptible task
762          * that is now waking up.
763          */
764         if (!p->activated) {
765                 /*
766                  * Tasks which were woken up by interrupts (ie. hw events)
767                  * are most likely of interactive nature. So we give them
768                  * the credit of extending their sleep time to the period
769                  * of time they spend on the runqueue, waiting for execution
770                  * on a CPU, first time around:
771                  */
772                 if (in_interrupt())
773                         p->activated = 2;
774                 else {
775                         /*
776                          * Normal first-time wakeups get a credit too for
777                          * on-runqueue time, but it will be weighted down:
778                          */
779                         p->activated = 1;
780                 }
781         }
782         p->timestamp = now;
783
784         __activate_task(p, rq);
785 }
786
787 /*
788  * deactivate_task - remove a task from the runqueue.
789  */
790 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
791 {
792         rq->nr_running--;
793         dequeue_task(p, p->array);
794         p->array = NULL;
795 }
796
797 /*
798  * resched_task - mark a task 'to be rescheduled now'.
799  *
800  * On UP this means the setting of the need_resched flag, on SMP it
801  * might also involve a cross-CPU call to trigger the scheduler on
802  * the target CPU.
803  */
804 #ifdef CONFIG_SMP
805 static void resched_task(task_t *p)
806 {
807         int need_resched, nrpolling;
808
809         assert_spin_locked(&task_rq(p)->lock);
810
811         /* minimise the chance of sending an interrupt to poll_idle() */
812         nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
813         need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
814         nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
815
816         if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
817                 smp_send_reschedule(task_cpu(p));
818 }
819 #else
820 static inline void resched_task(task_t *p)
821 {
822         set_tsk_need_resched(p);
823 }
824 #endif
825
826 /**
827  * task_curr - is this task currently executing on a CPU?
828  * @p: the task in question.
829  */
830 inline int task_curr(const task_t *p)
831 {
832         return cpu_curr(task_cpu(p)) == p;
833 }
834
835 #ifdef CONFIG_SMP
836 typedef struct {
837         struct list_head list;
838
839         task_t *task;
840         int dest_cpu;
841
842         struct completion done;
843 } migration_req_t;
844
845 /*
846  * The task's runqueue lock must be held.
847  * Returns true if you have to wait for migration thread.
848  */
849 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
850 {
851         runqueue_t *rq = task_rq(p);
852
853         /*
854          * If the task is not on a runqueue (and not running), then
855          * it is sufficient to simply update the task's cpu field.
856          */
857         if (!p->array && !task_running(rq, p)) {
858                 set_task_cpu(p, dest_cpu);
859                 return 0;
860         }
861
862         init_completion(&req->done);
863         req->task = p;
864         req->dest_cpu = dest_cpu;
865         list_add(&req->list, &rq->migration_queue);
866         return 1;
867 }
868
869 /*
870  * wait_task_inactive - wait for a thread to unschedule.
871  *
872  * The caller must ensure that the task *will* unschedule sometime soon,
873  * else this function might spin for a *long* time. This function can't
874  * be called with interrupts off, or it may introduce deadlock with
875  * smp_call_function() if an IPI is sent by the same process we are
876  * waiting to become inactive.
877  */
878 void wait_task_inactive(task_t *p)
879 {
880         unsigned long flags;
881         runqueue_t *rq;
882         int preempted;
883
884 repeat:
885         rq = task_rq_lock(p, &flags);
886         /* Must be off runqueue entirely, not preempted. */
887         if (unlikely(p->array || task_running(rq, p))) {
888                 /* If it's preempted, we yield.  It could be a while. */
889                 preempted = !task_running(rq, p);
890                 task_rq_unlock(rq, &flags);
891                 cpu_relax();
892                 if (preempted)
893                         yield();
894                 goto repeat;
895         }
896         task_rq_unlock(rq, &flags);
897 }
898
899 /***
900  * kick_process - kick a running thread to enter/exit the kernel
901  * @p: the to-be-kicked thread
902  *
903  * Cause a process which is running on another CPU to enter
904  * kernel-mode, without any delay. (to get signals handled.)
905  *
906  * NOTE: this function doesnt have to take the runqueue lock,
907  * because all it wants to ensure is that the remote task enters
908  * the kernel. If the IPI races and the task has been migrated
909  * to another CPU then no harm is done and the purpose has been
910  * achieved as well.
911  */
912 void kick_process(task_t *p)
913 {
914         int cpu;
915
916         preempt_disable();
917         cpu = task_cpu(p);
918         if ((cpu != smp_processor_id()) && task_curr(p))
919                 smp_send_reschedule(cpu);
920         preempt_enable();
921 }
922
923 /*
924  * Return a low guess at the load of a migration-source cpu.
925  *
926  * We want to under-estimate the load of migration sources, to
927  * balance conservatively.
928  */
929 static inline unsigned long source_load(int cpu, int type)
930 {
931         runqueue_t *rq = cpu_rq(cpu);
932         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
933         if (type == 0)
934                 return load_now;
935
936         return min(rq->cpu_load[type-1], load_now);
937 }
938
939 /*
940  * Return a high guess at the load of a migration-target cpu
941  */
942 static inline unsigned long target_load(int cpu, int type)
943 {
944         runqueue_t *rq = cpu_rq(cpu);
945         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
946         if (type == 0)
947                 return load_now;
948
949         return max(rq->cpu_load[type-1], load_now);
950 }
951
952 /*
953  * find_idlest_group finds and returns the least busy CPU group within the
954  * domain.
955  */
956 static struct sched_group *
957 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
958 {
959         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
960         unsigned long min_load = ULONG_MAX, this_load = 0;
961         int load_idx = sd->forkexec_idx;
962         int imbalance = 100 + (sd->imbalance_pct-100)/2;
963
964         do {
965                 unsigned long load, avg_load;
966                 int local_group;
967                 int i;
968
969                 /* Skip over this group if it has no CPUs allowed */
970                 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
971                         goto nextgroup;
972
973                 local_group = cpu_isset(this_cpu, group->cpumask);
974
975                 /* Tally up the load of all CPUs in the group */
976                 avg_load = 0;
977
978                 for_each_cpu_mask(i, group->cpumask) {
979                         /* Bias balancing toward cpus of our domain */
980                         if (local_group)
981                                 load = source_load(i, load_idx);
982                         else
983                                 load = target_load(i, load_idx);
984
985                         avg_load += load;
986                 }
987
988                 /* Adjust by relative CPU power of the group */
989                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
990
991                 if (local_group) {
992                         this_load = avg_load;
993                         this = group;
994                 } else if (avg_load < min_load) {
995                         min_load = avg_load;
996                         idlest = group;
997                 }
998 nextgroup:
999                 group = group->next;
1000         } while (group != sd->groups);
1001
1002         if (!idlest || 100*this_load < imbalance*min_load)
1003                 return NULL;
1004         return idlest;
1005 }
1006
1007 /*
1008  * find_idlest_queue - find the idlest runqueue among the cpus in group.
1009  */
1010 static int
1011 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1012 {
1013         cpumask_t tmp;
1014         unsigned long load, min_load = ULONG_MAX;
1015         int idlest = -1;
1016         int i;
1017
1018         /* Traverse only the allowed CPUs */
1019         cpus_and(tmp, group->cpumask, p->cpus_allowed);
1020
1021         for_each_cpu_mask(i, tmp) {
1022                 load = source_load(i, 0);
1023
1024                 if (load < min_load || (load == min_load && i == this_cpu)) {
1025                         min_load = load;
1026                         idlest = i;
1027                 }
1028         }
1029
1030         return idlest;
1031 }
1032
1033 /*
1034  * sched_balance_self: balance the current task (running on cpu) in domains
1035  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1036  * SD_BALANCE_EXEC.
1037  *
1038  * Balance, ie. select the least loaded group.
1039  *
1040  * Returns the target CPU number, or the same CPU if no balancing is needed.
1041  *
1042  * preempt must be disabled.
1043  */
1044 static int sched_balance_self(int cpu, int flag)
1045 {
1046         struct task_struct *t = current;
1047         struct sched_domain *tmp, *sd = NULL;
1048
1049         for_each_domain(cpu, tmp)
1050                 if (tmp->flags & flag)
1051                         sd = tmp;
1052
1053         while (sd) {
1054                 cpumask_t span;
1055                 struct sched_group *group;
1056                 int new_cpu;
1057                 int weight;
1058
1059                 span = sd->span;
1060                 group = find_idlest_group(sd, t, cpu);
1061                 if (!group)
1062                         goto nextlevel;
1063
1064                 new_cpu = find_idlest_cpu(group, t, cpu);
1065                 if (new_cpu == -1 || new_cpu == cpu)
1066                         goto nextlevel;
1067
1068                 /* Now try balancing at a lower domain level */
1069                 cpu = new_cpu;
1070 nextlevel:
1071                 sd = NULL;
1072                 weight = cpus_weight(span);
1073                 for_each_domain(cpu, tmp) {
1074                         if (weight <= cpus_weight(tmp->span))
1075                                 break;
1076                         if (tmp->flags & flag)
1077                                 sd = tmp;
1078                 }
1079                 /* while loop will break here if sd == NULL */
1080         }
1081
1082         return cpu;
1083 }
1084
1085 #endif /* CONFIG_SMP */
1086
1087 /*
1088  * wake_idle() will wake a task on an idle cpu if task->cpu is
1089  * not idle and an idle cpu is available.  The span of cpus to
1090  * search starts with cpus closest then further out as needed,
1091  * so we always favor a closer, idle cpu.
1092  *
1093  * Returns the CPU we should wake onto.
1094  */
1095 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1096 static int wake_idle(int cpu, task_t *p)
1097 {
1098         cpumask_t tmp;
1099         struct sched_domain *sd;
1100         int i;
1101
1102         if (idle_cpu(cpu))
1103                 return cpu;
1104
1105         for_each_domain(cpu, sd) {
1106                 if (sd->flags & SD_WAKE_IDLE) {
1107                         cpus_and(tmp, sd->span, p->cpus_allowed);
1108                         for_each_cpu_mask(i, tmp) {
1109                                 if (idle_cpu(i))
1110                                         return i;
1111                         }
1112                 }
1113                 else
1114                         break;
1115         }
1116         return cpu;
1117 }
1118 #else
1119 static inline int wake_idle(int cpu, task_t *p)
1120 {
1121         return cpu;
1122 }
1123 #endif
1124
1125 /***
1126  * try_to_wake_up - wake up a thread
1127  * @p: the to-be-woken-up thread
1128  * @state: the mask of task states that can be woken
1129  * @sync: do a synchronous wakeup?
1130  *
1131  * Put it on the run-queue if it's not already there. The "current"
1132  * thread is always on the run-queue (except when the actual
1133  * re-schedule is in progress), and as such you're allowed to do
1134  * the simpler "current->state = TASK_RUNNING" to mark yourself
1135  * runnable without the overhead of this.
1136  *
1137  * returns failure only if the task is already active.
1138  */
1139 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1140 {
1141         int cpu, this_cpu, success = 0;
1142         unsigned long flags;
1143         long old_state;
1144         runqueue_t *rq;
1145 #ifdef CONFIG_SMP
1146         unsigned long load, this_load;
1147         struct sched_domain *sd, *this_sd = NULL;
1148         int new_cpu;
1149 #endif
1150
1151         rq = task_rq_lock(p, &flags);
1152         old_state = p->state;
1153         if (!(old_state & state))
1154                 goto out;
1155
1156         if (p->array)
1157                 goto out_running;
1158
1159         cpu = task_cpu(p);
1160         this_cpu = smp_processor_id();
1161
1162 #ifdef CONFIG_SMP
1163         if (unlikely(task_running(rq, p)))
1164                 goto out_activate;
1165
1166         new_cpu = cpu;
1167
1168         schedstat_inc(rq, ttwu_cnt);
1169         if (cpu == this_cpu) {
1170                 schedstat_inc(rq, ttwu_local);
1171                 goto out_set_cpu;
1172         }
1173
1174         for_each_domain(this_cpu, sd) {
1175                 if (cpu_isset(cpu, sd->span)) {
1176                         schedstat_inc(sd, ttwu_wake_remote);
1177                         this_sd = sd;
1178                         break;
1179                 }
1180         }
1181
1182         if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1183                 goto out_set_cpu;
1184
1185         /*
1186          * Check for affine wakeup and passive balancing possibilities.
1187          */
1188         if (this_sd) {
1189                 int idx = this_sd->wake_idx;
1190                 unsigned int imbalance;
1191
1192                 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1193
1194                 load = source_load(cpu, idx);
1195                 this_load = target_load(this_cpu, idx);
1196
1197                 new_cpu = this_cpu; /* Wake to this CPU if we can */
1198
1199                 if (this_sd->flags & SD_WAKE_AFFINE) {
1200                         unsigned long tl = this_load;
1201                         /*
1202                          * If sync wakeup then subtract the (maximum possible)
1203                          * effect of the currently running task from the load
1204                          * of the current CPU:
1205                          */
1206                         if (sync)
1207                                 tl -= SCHED_LOAD_SCALE;
1208
1209                         if ((tl <= load &&
1210                                 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1211                                 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1212                                 /*
1213                                  * This domain has SD_WAKE_AFFINE and
1214                                  * p is cache cold in this domain, and
1215                                  * there is no bad imbalance.
1216                                  */
1217                                 schedstat_inc(this_sd, ttwu_move_affine);
1218                                 goto out_set_cpu;
1219                         }
1220                 }
1221
1222                 /*
1223                  * Start passive balancing when half the imbalance_pct
1224                  * limit is reached.
1225                  */
1226                 if (this_sd->flags & SD_WAKE_BALANCE) {
1227                         if (imbalance*this_load <= 100*load) {
1228                                 schedstat_inc(this_sd, ttwu_move_balance);
1229                                 goto out_set_cpu;
1230                         }
1231                 }
1232         }
1233
1234         new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1235 out_set_cpu:
1236         new_cpu = wake_idle(new_cpu, p);
1237         if (new_cpu != cpu) {
1238                 set_task_cpu(p, new_cpu);
1239                 task_rq_unlock(rq, &flags);
1240                 /* might preempt at this point */
1241                 rq = task_rq_lock(p, &flags);
1242                 old_state = p->state;
1243                 if (!(old_state & state))
1244                         goto out;
1245                 if (p->array)
1246                         goto out_running;
1247
1248                 this_cpu = smp_processor_id();
1249                 cpu = task_cpu(p);
1250         }
1251
1252 out_activate:
1253 #endif /* CONFIG_SMP */
1254         if (old_state == TASK_UNINTERRUPTIBLE) {
1255                 rq->nr_uninterruptible--;
1256                 /*
1257                  * Tasks on involuntary sleep don't earn
1258                  * sleep_avg beyond just interactive state.
1259                  */
1260                 p->activated = -1;
1261         }
1262
1263         /*
1264          * Tasks that have marked their sleep as noninteractive get
1265          * woken up without updating their sleep average. (i.e. their
1266          * sleep is handled in a priority-neutral manner, no priority
1267          * boost and no penalty.)
1268          */
1269         if (old_state & TASK_NONINTERACTIVE)
1270                 __activate_task(p, rq);
1271         else
1272                 activate_task(p, rq, cpu == this_cpu);
1273         /*
1274          * Sync wakeups (i.e. those types of wakeups where the waker
1275          * has indicated that it will leave the CPU in short order)
1276          * don't trigger a preemption, if the woken up task will run on
1277          * this cpu. (in this case the 'I will reschedule' promise of
1278          * the waker guarantees that the freshly woken up task is going
1279          * to be considered on this CPU.)
1280          */
1281         if (!sync || cpu != this_cpu) {
1282                 if (TASK_PREEMPTS_CURR(p, rq))
1283                         resched_task(rq->curr);
1284         }
1285         success = 1;
1286
1287 out_running:
1288         p->state = TASK_RUNNING;
1289 out:
1290         task_rq_unlock(rq, &flags);
1291
1292         return success;
1293 }
1294
1295 int fastcall wake_up_process(task_t *p)
1296 {
1297         return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1298                                  TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1299 }
1300
1301 EXPORT_SYMBOL(wake_up_process);
1302
1303 int fastcall wake_up_state(task_t *p, unsigned int state)
1304 {
1305         return try_to_wake_up(p, state, 0);
1306 }
1307
1308 /*
1309  * Perform scheduler related setup for a newly forked process p.
1310  * p is forked by current.
1311  */
1312 void fastcall sched_fork(task_t *p, int clone_flags)
1313 {
1314         int cpu = get_cpu();
1315
1316 #ifdef CONFIG_SMP
1317         cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1318 #endif
1319         set_task_cpu(p, cpu);
1320
1321         /*
1322          * We mark the process as running here, but have not actually
1323          * inserted it onto the runqueue yet. This guarantees that
1324          * nobody will actually run it, and a signal or other external
1325          * event cannot wake it up and insert it on the runqueue either.
1326          */
1327         p->state = TASK_RUNNING;
1328         INIT_LIST_HEAD(&p->run_list);
1329         p->array = NULL;
1330 #ifdef CONFIG_SCHEDSTATS
1331         memset(&p->sched_info, 0, sizeof(p->sched_info));
1332 #endif
1333 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1334         p->oncpu = 0;
1335 #endif
1336 #ifdef CONFIG_PREEMPT
1337         /* Want to start with kernel preemption disabled. */
1338         p->thread_info->preempt_count = 1;
1339 #endif
1340         /*
1341          * Share the timeslice between parent and child, thus the
1342          * total amount of pending timeslices in the system doesn't change,
1343          * resulting in more scheduling fairness.
1344          */
1345         local_irq_disable();
1346         p->time_slice = (current->time_slice + 1) >> 1;
1347         /*
1348          * The remainder of the first timeslice might be recovered by
1349          * the parent if the child exits early enough.
1350          */
1351         p->first_time_slice = 1;
1352         current->time_slice >>= 1;
1353         p->timestamp = sched_clock();
1354         if (unlikely(!current->time_slice)) {
1355                 /*
1356                  * This case is rare, it happens when the parent has only
1357                  * a single jiffy left from its timeslice. Taking the
1358                  * runqueue lock is not a problem.
1359                  */
1360                 current->time_slice = 1;
1361                 scheduler_tick();
1362         }
1363         local_irq_enable();
1364         put_cpu();
1365 }
1366
1367 /*
1368  * wake_up_new_task - wake up a newly created task for the first time.
1369  *
1370  * This function will do some initial scheduler statistics housekeeping
1371  * that must be done for every newly created context, then puts the task
1372  * on the runqueue and wakes it.
1373  */
1374 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1375 {
1376         unsigned long flags;
1377         int this_cpu, cpu;
1378         runqueue_t *rq, *this_rq;
1379
1380         rq = task_rq_lock(p, &flags);
1381         BUG_ON(p->state != TASK_RUNNING);
1382         this_cpu = smp_processor_id();
1383         cpu = task_cpu(p);
1384
1385         /*
1386          * We decrease the sleep average of forking parents
1387          * and children as well, to keep max-interactive tasks
1388          * from forking tasks that are max-interactive. The parent
1389          * (current) is done further down, under its lock.
1390          */
1391         p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1392                 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1393
1394         p->prio = effective_prio(p);
1395
1396         if (likely(cpu == this_cpu)) {
1397                 if (!(clone_flags & CLONE_VM)) {
1398                         /*
1399                          * The VM isn't cloned, so we're in a good position to
1400                          * do child-runs-first in anticipation of an exec. This
1401                          * usually avoids a lot of COW overhead.
1402                          */
1403                         if (unlikely(!current->array))
1404                                 __activate_task(p, rq);
1405                         else {
1406                                 p->prio = current->prio;
1407                                 list_add_tail(&p->run_list, &current->run_list);
1408                                 p->array = current->array;
1409                                 p->array->nr_active++;
1410                                 rq->nr_running++;
1411                         }
1412                         set_need_resched();
1413                 } else
1414                         /* Run child last */
1415                         __activate_task(p, rq);
1416                 /*
1417                  * We skip the following code due to cpu == this_cpu
1418                  *
1419                  *   task_rq_unlock(rq, &flags);
1420                  *   this_rq = task_rq_lock(current, &flags);
1421                  */
1422                 this_rq = rq;
1423         } else {
1424                 this_rq = cpu_rq(this_cpu);
1425
1426                 /*
1427                  * Not the local CPU - must adjust timestamp. This should
1428                  * get optimised away in the !CONFIG_SMP case.
1429                  */
1430                 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1431                                         + rq->timestamp_last_tick;
1432                 __activate_task(p, rq);
1433                 if (TASK_PREEMPTS_CURR(p, rq))
1434                         resched_task(rq->curr);
1435
1436                 /*
1437                  * Parent and child are on different CPUs, now get the
1438                  * parent runqueue to update the parent's ->sleep_avg:
1439                  */
1440                 task_rq_unlock(rq, &flags);
1441                 this_rq = task_rq_lock(current, &flags);
1442         }
1443         current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1444                 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1445         task_rq_unlock(this_rq, &flags);
1446 }
1447
1448 /*
1449  * Potentially available exiting-child timeslices are
1450  * retrieved here - this way the parent does not get
1451  * penalized for creating too many threads.
1452  *
1453  * (this cannot be used to 'generate' timeslices
1454  * artificially, because any timeslice recovered here
1455  * was given away by the parent in the first place.)
1456  */
1457 void fastcall sched_exit(task_t *p)
1458 {
1459         unsigned long flags;
1460         runqueue_t *rq;
1461
1462         /*
1463          * If the child was a (relative-) CPU hog then decrease
1464          * the sleep_avg of the parent as well.
1465          */
1466         rq = task_rq_lock(p->parent, &flags);
1467         if (p->first_time_slice) {
1468                 p->parent->time_slice += p->time_slice;
1469                 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1470                         p->parent->time_slice = task_timeslice(p);
1471         }
1472         if (p->sleep_avg < p->parent->sleep_avg)
1473                 p->parent->sleep_avg = p->parent->sleep_avg /
1474                 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1475                 (EXIT_WEIGHT + 1);
1476         task_rq_unlock(rq, &flags);
1477 }
1478
1479 /**
1480  * prepare_task_switch - prepare to switch tasks
1481  * @rq: the runqueue preparing to switch
1482  * @next: the task we are going to switch to.
1483  *
1484  * This is called with the rq lock held and interrupts off. It must
1485  * be paired with a subsequent finish_task_switch after the context
1486  * switch.
1487  *
1488  * prepare_task_switch sets up locking and calls architecture specific
1489  * hooks.
1490  */
1491 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1492 {
1493         prepare_lock_switch(rq, next);
1494         prepare_arch_switch(next);
1495 }
1496
1497 /**
1498  * finish_task_switch - clean up after a task-switch
1499  * @rq: runqueue associated with task-switch
1500  * @prev: the thread we just switched away from.
1501  *
1502  * finish_task_switch must be called after the context switch, paired
1503  * with a prepare_task_switch call before the context switch.
1504  * finish_task_switch will reconcile locking set up by prepare_task_switch,
1505  * and do any other architecture-specific cleanup actions.
1506  *
1507  * Note that we may have delayed dropping an mm in context_switch(). If
1508  * so, we finish that here outside of the runqueue lock.  (Doing it
1509  * with the lock held can cause deadlocks; see schedule() for
1510  * details.)
1511  */
1512 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1513         __releases(rq->lock)
1514 {
1515         struct mm_struct *mm = rq->prev_mm;
1516         unsigned long prev_task_flags;
1517
1518         rq->prev_mm = NULL;
1519
1520         /*
1521          * A task struct has one reference for the use as "current".
1522          * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1523          * calls schedule one last time. The schedule call will never return,
1524          * and the scheduled task must drop that reference.
1525          * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1526          * still held, otherwise prev could be scheduled on another cpu, die
1527          * there before we look at prev->state, and then the reference would
1528          * be dropped twice.
1529          *              Manfred Spraul <manfred@colorfullife.com>
1530          */
1531         prev_task_flags = prev->flags;
1532 #ifdef CONFIG_DEBUG_SPINLOCK
1533         /* this is a valid case when another task releases the spinlock */
1534         rq->lock.owner = current;
1535 #endif
1536         finish_arch_switch(prev);
1537         finish_lock_switch(rq, prev);
1538         if (mm)
1539                 mmdrop(mm);
1540         if (unlikely(prev_task_flags & PF_DEAD))
1541                 put_task_struct(prev);
1542 }
1543
1544 /**
1545  * schedule_tail - first thing a freshly forked thread must call.
1546  * @prev: the thread we just switched away from.
1547  */
1548 asmlinkage void schedule_tail(task_t *prev)
1549         __releases(rq->lock)
1550 {
1551         runqueue_t *rq = this_rq();
1552         finish_task_switch(rq, prev);
1553 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1554         /* In this case, finish_task_switch does not reenable preemption */
1555         preempt_enable();
1556 #endif
1557         if (current->set_child_tid)
1558                 put_user(current->pid, current->set_child_tid);
1559 }
1560
1561 /*
1562  * context_switch - switch to the new MM and the new
1563  * thread's register state.
1564  */
1565 static inline
1566 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1567 {
1568         struct mm_struct *mm = next->mm;
1569         struct mm_struct *oldmm = prev->active_mm;
1570
1571         if (unlikely(!mm)) {
1572                 next->active_mm = oldmm;
1573                 atomic_inc(&oldmm->mm_count);
1574                 enter_lazy_tlb(oldmm, next);
1575         } else
1576                 switch_mm(oldmm, mm, next);
1577
1578         if (unlikely(!prev->mm)) {
1579                 prev->active_mm = NULL;
1580                 WARN_ON(rq->prev_mm);
1581                 rq->prev_mm = oldmm;
1582         }
1583
1584         /* Here we just switch the register state and the stack. */
1585         switch_to(prev, next, prev);
1586
1587         return prev;
1588 }
1589
1590 /*
1591  * nr_running, nr_uninterruptible and nr_context_switches:
1592  *
1593  * externally visible scheduler statistics: current number of runnable
1594  * threads, current number of uninterruptible-sleeping threads, total
1595  * number of context switches performed since bootup.
1596  */
1597 unsigned long nr_running(void)
1598 {
1599         unsigned long i, sum = 0;
1600
1601         for_each_online_cpu(i)
1602                 sum += cpu_rq(i)->nr_running;
1603
1604         return sum;
1605 }
1606
1607 unsigned long nr_uninterruptible(void)
1608 {
1609         unsigned long i, sum = 0;
1610
1611         for_each_cpu(i)
1612                 sum += cpu_rq(i)->nr_uninterruptible;
1613
1614         /*
1615          * Since we read the counters lockless, it might be slightly
1616          * inaccurate. Do not allow it to go below zero though:
1617          */
1618         if (unlikely((long)sum < 0))
1619                 sum = 0;
1620
1621         return sum;
1622 }
1623
1624 unsigned long long nr_context_switches(void)
1625 {
1626         unsigned long long i, sum = 0;
1627
1628         for_each_cpu(i)
1629                 sum += cpu_rq(i)->nr_switches;
1630
1631         return sum;
1632 }
1633
1634 unsigned long nr_iowait(void)
1635 {
1636         unsigned long i, sum = 0;
1637
1638         for_each_cpu(i)
1639                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1640
1641         return sum;
1642 }
1643
1644 #ifdef CONFIG_SMP
1645
1646 /*
1647  * double_rq_lock - safely lock two runqueues
1648  *
1649  * Note this does not disable interrupts like task_rq_lock,
1650  * you need to do so manually before calling.
1651  */
1652 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1653         __acquires(rq1->lock)
1654         __acquires(rq2->lock)
1655 {
1656         if (rq1 == rq2) {
1657                 spin_lock(&rq1->lock);
1658                 __acquire(rq2->lock);   /* Fake it out ;) */
1659         } else {
1660                 if (rq1 < rq2) {
1661                         spin_lock(&rq1->lock);
1662                         spin_lock(&rq2->lock);
1663                 } else {
1664                         spin_lock(&rq2->lock);
1665                         spin_lock(&rq1->lock);
1666                 }
1667         }
1668 }
1669
1670 /*
1671  * double_rq_unlock - safely unlock two runqueues
1672  *
1673  * Note this does not restore interrupts like task_rq_unlock,
1674  * you need to do so manually after calling.
1675  */
1676 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1677         __releases(rq1->lock)
1678         __releases(rq2->lock)
1679 {
1680         spin_unlock(&rq1->lock);
1681         if (rq1 != rq2)
1682                 spin_unlock(&rq2->lock);
1683         else
1684                 __release(rq2->lock);
1685 }
1686
1687 /*
1688  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1689  */
1690 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1691         __releases(this_rq->lock)
1692         __acquires(busiest->lock)
1693         __acquires(this_rq->lock)
1694 {
1695         if (unlikely(!spin_trylock(&busiest->lock))) {
1696                 if (busiest < this_rq) {
1697                         spin_unlock(&this_rq->lock);
1698                         spin_lock(&busiest->lock);
1699                         spin_lock(&this_rq->lock);
1700                 } else
1701                         spin_lock(&busiest->lock);
1702         }
1703 }
1704
1705 /*
1706  * If dest_cpu is allowed for this process, migrate the task to it.
1707  * This is accomplished by forcing the cpu_allowed mask to only
1708  * allow dest_cpu, which will force the cpu onto dest_cpu.  Then
1709  * the cpu_allowed mask is restored.
1710  */
1711 static void sched_migrate_task(task_t *p, int dest_cpu)
1712 {
1713         migration_req_t req;
1714         runqueue_t *rq;
1715         unsigned long flags;
1716
1717         rq = task_rq_lock(p, &flags);
1718         if (!cpu_isset(dest_cpu, p->cpus_allowed)
1719             || unlikely(cpu_is_offline(dest_cpu)))
1720                 goto out;
1721
1722         /* force the process onto the specified CPU */
1723         if (migrate_task(p, dest_cpu, &req)) {
1724                 /* Need to wait for migration thread (might exit: take ref). */
1725                 struct task_struct *mt = rq->migration_thread;
1726                 get_task_struct(mt);
1727                 task_rq_unlock(rq, &flags);
1728                 wake_up_process(mt);
1729                 put_task_struct(mt);
1730                 wait_for_completion(&req.done);
1731                 return;
1732         }
1733 out:
1734         task_rq_unlock(rq, &flags);
1735 }
1736
1737 /*
1738  * sched_exec - execve() is a valuable balancing opportunity, because at
1739  * this point the task has the smallest effective memory and cache footprint.
1740  */
1741 void sched_exec(void)
1742 {
1743         int new_cpu, this_cpu = get_cpu();
1744         new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1745         put_cpu();
1746         if (new_cpu != this_cpu)
1747                 sched_migrate_task(current, new_cpu);
1748 }
1749
1750 /*
1751  * pull_task - move a task from a remote runqueue to the local runqueue.
1752  * Both runqueues must be locked.
1753  */
1754 static inline
1755 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1756                runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1757 {
1758         dequeue_task(p, src_array);
1759         src_rq->nr_running--;
1760         set_task_cpu(p, this_cpu);
1761         this_rq->nr_running++;
1762         enqueue_task(p, this_array);
1763         p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1764                                 + this_rq->timestamp_last_tick;
1765         /*
1766          * Note that idle threads have a prio of MAX_PRIO, for this test
1767          * to be always true for them.
1768          */
1769         if (TASK_PREEMPTS_CURR(p, this_rq))
1770                 resched_task(this_rq->curr);
1771 }
1772
1773 /*
1774  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1775  */
1776 static inline
1777 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1778                      struct sched_domain *sd, enum idle_type idle,
1779                      int *all_pinned)
1780 {
1781         /*
1782          * We do not migrate tasks that are:
1783          * 1) running (obviously), or
1784          * 2) cannot be migrated to this CPU due to cpus_allowed, or
1785          * 3) are cache-hot on their current CPU.
1786          */
1787         if (!cpu_isset(this_cpu, p->cpus_allowed))
1788                 return 0;
1789         *all_pinned = 0;
1790
1791         if (task_running(rq, p))
1792                 return 0;
1793
1794         /*
1795          * Aggressive migration if:
1796          * 1) task is cache cold, or
1797          * 2) too many balance attempts have failed.
1798          */
1799
1800         if (sd->nr_balance_failed > sd->cache_nice_tries)
1801                 return 1;
1802
1803         if (task_hot(p, rq->timestamp_last_tick, sd))
1804                 return 0;
1805         return 1;
1806 }
1807
1808 /*
1809  * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1810  * as part of a balancing operation within "domain". Returns the number of
1811  * tasks moved.
1812  *
1813  * Called with both runqueues locked.
1814  */
1815 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1816                       unsigned long max_nr_move, struct sched_domain *sd,
1817                       enum idle_type idle, int *all_pinned)
1818 {
1819         prio_array_t *array, *dst_array;
1820         struct list_head *head, *curr;
1821         int idx, pulled = 0, pinned = 0;
1822         task_t *tmp;
1823
1824         if (max_nr_move == 0)
1825                 goto out;
1826
1827         pinned = 1;
1828
1829         /*
1830          * We first consider expired tasks. Those will likely not be
1831          * executed in the near future, and they are most likely to
1832          * be cache-cold, thus switching CPUs has the least effect
1833          * on them.
1834          */
1835         if (busiest->expired->nr_active) {
1836                 array = busiest->expired;
1837                 dst_array = this_rq->expired;
1838         } else {
1839                 array = busiest->active;
1840                 dst_array = this_rq->active;
1841         }
1842
1843 new_array:
1844         /* Start searching at priority 0: */
1845         idx = 0;
1846 skip_bitmap:
1847         if (!idx)
1848                 idx = sched_find_first_bit(array->bitmap);
1849         else
1850                 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1851         if (idx >= MAX_PRIO) {
1852                 if (array == busiest->expired && busiest->active->nr_active) {
1853                         array = busiest->active;
1854                         dst_array = this_rq->active;
1855                         goto new_array;
1856                 }
1857                 goto out;
1858         }
1859
1860         head = array->queue + idx;
1861         curr = head->prev;
1862 skip_queue:
1863         tmp = list_entry(curr, task_t, run_list);
1864
1865         curr = curr->prev;
1866
1867         if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1868                 if (curr != head)
1869                         goto skip_queue;
1870                 idx++;
1871                 goto skip_bitmap;
1872         }
1873
1874 #ifdef CONFIG_SCHEDSTATS
1875         if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1876                 schedstat_inc(sd, lb_hot_gained[idle]);
1877 #endif
1878
1879         pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1880         pulled++;
1881
1882         /* We only want to steal up to the prescribed number of tasks. */
1883         if (pulled < max_nr_move) {
1884                 if (curr != head)
1885                         goto skip_queue;
1886                 idx++;
1887                 goto skip_bitmap;
1888         }
1889 out:
1890         /*
1891          * Right now, this is the only place pull_task() is called,
1892          * so we can safely collect pull_task() stats here rather than
1893          * inside pull_task().
1894          */
1895         schedstat_add(sd, lb_gained[idle], pulled);
1896
1897         if (all_pinned)
1898                 *all_pinned = pinned;
1899         return pulled;
1900 }
1901
1902 /*
1903  * find_busiest_group finds and returns the busiest CPU group within the
1904  * domain. It calculates and returns the number of tasks which should be
1905  * moved to restore balance via the imbalance parameter.
1906  */
1907 static struct sched_group *
1908 find_busiest_group(struct sched_domain *sd, int this_cpu,
1909                    unsigned long *imbalance, enum idle_type idle, int *sd_idle)
1910 {
1911         struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1912         unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1913         int load_idx;
1914
1915         max_load = this_load = total_load = total_pwr = 0;
1916         if (idle == NOT_IDLE)
1917                 load_idx = sd->busy_idx;
1918         else if (idle == NEWLY_IDLE)
1919                 load_idx = sd->newidle_idx;
1920         else
1921                 load_idx = sd->idle_idx;
1922
1923         do {
1924                 unsigned long load;
1925                 int local_group;
1926                 int i;
1927
1928                 local_group = cpu_isset(this_cpu, group->cpumask);
1929
1930                 /* Tally up the load of all CPUs in the group */
1931                 avg_load = 0;
1932
1933                 for_each_cpu_mask(i, group->cpumask) {
1934                         if (*sd_idle && !idle_cpu(i))
1935                                 *sd_idle = 0;
1936
1937                         /* Bias balancing toward cpus of our domain */
1938                         if (local_group)
1939                                 load = target_load(i, load_idx);
1940                         else
1941                                 load = source_load(i, load_idx);
1942
1943                         avg_load += load;
1944                 }
1945
1946                 total_load += avg_load;
1947                 total_pwr += group->cpu_power;
1948
1949                 /* Adjust by relative CPU power of the group */
1950                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1951
1952                 if (local_group) {
1953                         this_load = avg_load;
1954                         this = group;
1955                 } else if (avg_load > max_load) {
1956                         max_load = avg_load;
1957                         busiest = group;
1958                 }
1959                 group = group->next;
1960         } while (group != sd->groups);
1961
1962         if (!busiest || this_load >= max_load)
1963                 goto out_balanced;
1964
1965         avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1966
1967         if (this_load >= avg_load ||
1968                         100*max_load <= sd->imbalance_pct*this_load)
1969                 goto out_balanced;
1970
1971         /*
1972          * We're trying to get all the cpus to the average_load, so we don't
1973          * want to push ourselves above the average load, nor do we wish to
1974          * reduce the max loaded cpu below the average load, as either of these
1975          * actions would just result in more rebalancing later, and ping-pong
1976          * tasks around. Thus we look for the minimum possible imbalance.
1977          * Negative imbalances (*we* are more loaded than anyone else) will
1978          * be counted as no imbalance for these purposes -- we can't fix that
1979          * by pulling tasks to us.  Be careful of negative numbers as they'll
1980          * appear as very large values with unsigned longs.
1981          */
1982         /* How much load to actually move to equalise the imbalance */
1983         *imbalance = min((max_load - avg_load) * busiest->cpu_power,
1984                                 (avg_load - this_load) * this->cpu_power)
1985                         / SCHED_LOAD_SCALE;
1986
1987         if (*imbalance < SCHED_LOAD_SCALE) {
1988                 unsigned long pwr_now = 0, pwr_move = 0;
1989                 unsigned long tmp;
1990
1991                 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1992                         *imbalance = 1;
1993                         return busiest;
1994                 }
1995
1996                 /*
1997                  * OK, we don't have enough imbalance to justify moving tasks,
1998                  * however we may be able to increase total CPU power used by
1999                  * moving them.
2000                  */
2001
2002                 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2003                 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2004                 pwr_now /= SCHED_LOAD_SCALE;
2005
2006                 /* Amount of load we'd subtract */
2007                 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2008                 if (max_load > tmp)
2009                         pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2010                                                         max_load - tmp);
2011
2012                 /* Amount of load we'd add */
2013                 if (max_load*busiest->cpu_power <
2014                                 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2015                         tmp = max_load*busiest->cpu_power/this->cpu_power;
2016                 else
2017                         tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2018                 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2019                 pwr_move /= SCHED_LOAD_SCALE;
2020
2021                 /* Move if we gain throughput */
2022                 if (pwr_move <= pwr_now)
2023                         goto out_balanced;
2024
2025                 *imbalance = 1;
2026                 return busiest;
2027         }
2028
2029         /* Get rid of the scaling factor, rounding down as we divide */
2030         *imbalance = *imbalance / SCHED_LOAD_SCALE;
2031         return busiest;
2032
2033 out_balanced:
2034
2035         *imbalance = 0;
2036         return NULL;
2037 }
2038
2039 /*
2040  * find_busiest_queue - find the busiest runqueue among the cpus in group.
2041  */
2042 static runqueue_t *find_busiest_queue(struct sched_group *group)
2043 {
2044         unsigned long load, max_load = 0;
2045         runqueue_t *busiest = NULL;
2046         int i;
2047
2048         for_each_cpu_mask(i, group->cpumask) {
2049                 load = source_load(i, 0);
2050
2051                 if (load > max_load) {
2052                         max_load = load;
2053                         busiest = cpu_rq(i);
2054                 }
2055         }
2056
2057         return busiest;
2058 }
2059
2060 /*
2061  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2062  * so long as it is large enough.
2063  */
2064 #define MAX_PINNED_INTERVAL     512
2065
2066 /*
2067  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2068  * tasks if there is an imbalance.
2069  *
2070  * Called with this_rq unlocked.
2071  */
2072 static int load_balance(int this_cpu, runqueue_t *this_rq,
2073                         struct sched_domain *sd, enum idle_type idle)
2074 {
2075         struct sched_group *group;
2076         runqueue_t *busiest;
2077         unsigned long imbalance;
2078         int nr_moved, all_pinned = 0;
2079         int active_balance = 0;
2080         int sd_idle = 0;
2081
2082         if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2083                 sd_idle = 1;
2084
2085         schedstat_inc(sd, lb_cnt[idle]);
2086
2087         group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2088         if (!group) {
2089                 schedstat_inc(sd, lb_nobusyg[idle]);
2090                 goto out_balanced;
2091         }
2092
2093         busiest = find_busiest_queue(group);
2094         if (!busiest) {
2095                 schedstat_inc(sd, lb_nobusyq[idle]);
2096                 goto out_balanced;
2097         }
2098
2099         BUG_ON(busiest == this_rq);
2100
2101         schedstat_add(sd, lb_imbalance[idle], imbalance);
2102
2103         nr_moved = 0;
2104         if (busiest->nr_running > 1) {
2105                 /*
2106                  * Attempt to move tasks. If find_busiest_group has found
2107                  * an imbalance but busiest->nr_running <= 1, the group is
2108                  * still unbalanced. nr_moved simply stays zero, so it is
2109                  * correctly treated as an imbalance.
2110                  */
2111                 double_rq_lock(this_rq, busiest);
2112                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2113                                         imbalance, sd, idle, &all_pinned);
2114                 double_rq_unlock(this_rq, busiest);
2115
2116                 /* All tasks on this runqueue were pinned by CPU affinity */
2117                 if (unlikely(all_pinned))
2118                         goto out_balanced;
2119         }
2120
2121         if (!nr_moved) {
2122                 schedstat_inc(sd, lb_failed[idle]);
2123                 sd->nr_balance_failed++;
2124
2125                 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2126
2127                         spin_lock(&busiest->lock);
2128                         if (!busiest->active_balance) {
2129                                 busiest->active_balance = 1;
2130                                 busiest->push_cpu = this_cpu;
2131                                 active_balance = 1;
2132                         }
2133                         spin_unlock(&busiest->lock);
2134                         if (active_balance)
2135                                 wake_up_process(busiest->migration_thread);
2136
2137                         /*
2138                          * We've kicked active balancing, reset the failure
2139                          * counter.
2140                          */
2141                         sd->nr_balance_failed = sd->cache_nice_tries+1;
2142                 }
2143         } else
2144                 sd->nr_balance_failed = 0;
2145
2146         if (likely(!active_balance)) {
2147                 /* We were unbalanced, so reset the balancing interval */
2148                 sd->balance_interval = sd->min_interval;
2149         } else {
2150                 /*
2151                  * If we've begun active balancing, start to back off. This
2152                  * case may not be covered by the all_pinned logic if there
2153                  * is only 1 task on the busy runqueue (because we don't call
2154                  * move_tasks).
2155                  */
2156                 if (sd->balance_interval < sd->max_interval)
2157                         sd->balance_interval *= 2;
2158         }
2159
2160         if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2161                 return -1;
2162         return nr_moved;
2163
2164 out_balanced:
2165         schedstat_inc(sd, lb_balanced[idle]);
2166
2167         sd->nr_balance_failed = 0;
2168         /* tune up the balancing interval */
2169         if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2170                         (sd->balance_interval < sd->max_interval))
2171                 sd->balance_interval *= 2;
2172
2173         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2174                 return -1;
2175         return 0;
2176 }
2177
2178 /*
2179  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2180  * tasks if there is an imbalance.
2181  *
2182  * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2183  * this_rq is locked.
2184  */
2185 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2186                                 struct sched_domain *sd)
2187 {
2188         struct sched_group *group;
2189         runqueue_t *busiest = NULL;
2190         unsigned long imbalance;
2191         int nr_moved = 0;
2192         int sd_idle = 0;
2193
2194         if (sd->flags & SD_SHARE_CPUPOWER)
2195                 sd_idle = 1;
2196
2197         schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2198         group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2199         if (!group) {
2200                 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2201                 goto out_balanced;
2202         }
2203
2204         busiest = find_busiest_queue(group);
2205         if (!busiest) {
2206                 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2207                 goto out_balanced;
2208         }
2209
2210         BUG_ON(busiest == this_rq);
2211
2212         schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2213
2214         nr_moved = 0;
2215         if (busiest->nr_running > 1) {
2216                 /* Attempt to move tasks */
2217                 double_lock_balance(this_rq, busiest);
2218                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2219                                         imbalance, sd, NEWLY_IDLE, NULL);
2220                 spin_unlock(&busiest->lock);
2221         }
2222
2223         if (!nr_moved) {
2224                 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2225                 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2226                         return -1;
2227         } else
2228                 sd->nr_balance_failed = 0;
2229
2230         return nr_moved;
2231
2232 out_balanced:
2233         schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2234         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2235                 return -1;
2236         sd->nr_balance_failed = 0;
2237         return 0;
2238 }
2239
2240 /*
2241  * idle_balance is called by schedule() if this_cpu is about to become
2242  * idle. Attempts to pull tasks from other CPUs.
2243  */
2244 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2245 {
2246         struct sched_domain *sd;
2247
2248         for_each_domain(this_cpu, sd) {
2249                 if (sd->flags & SD_BALANCE_NEWIDLE) {
2250                         if (load_balance_newidle(this_cpu, this_rq, sd)) {
2251                                 /* We've pulled tasks over so stop searching */
2252                                 break;
2253                         }
2254                 }
2255         }
2256 }
2257
2258 /*
2259  * active_load_balance is run by migration threads. It pushes running tasks
2260  * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2261  * running on each physical CPU where possible, and avoids physical /
2262  * logical imbalances.
2263  *
2264  * Called with busiest_rq locked.
2265  */
2266 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2267 {
2268         struct sched_domain *sd;
2269         runqueue_t *target_rq;
2270         int target_cpu = busiest_rq->push_cpu;
2271
2272         if (busiest_rq->nr_running <= 1)
2273                 /* no task to move */
2274                 return;
2275
2276         target_rq = cpu_rq(target_cpu);
2277
2278         /*
2279          * This condition is "impossible", if it occurs
2280          * we need to fix it.  Originally reported by
2281          * Bjorn Helgaas on a 128-cpu setup.
2282          */
2283         BUG_ON(busiest_rq == target_rq);
2284
2285         /* move a task from busiest_rq to target_rq */
2286         double_lock_balance(busiest_rq, target_rq);
2287
2288         /* Search for an sd spanning us and the target CPU. */
2289         for_each_domain(target_cpu, sd)
2290                 if ((sd->flags & SD_LOAD_BALANCE) &&
2291                         cpu_isset(busiest_cpu, sd->span))
2292                                 break;
2293
2294         if (unlikely(sd == NULL))
2295                 goto out;
2296
2297         schedstat_inc(sd, alb_cnt);
2298
2299         if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2300                 schedstat_inc(sd, alb_pushed);
2301         else
2302                 schedstat_inc(sd, alb_failed);
2303 out:
2304         spin_unlock(&target_rq->lock);
2305 }
2306
2307 /*
2308  * rebalance_tick will get called every timer tick, on every CPU.
2309  *
2310  * It checks each scheduling domain to see if it is due to be balanced,
2311  * and initiates a balancing operation if so.
2312  *
2313  * Balancing parameters are set up in arch_init_sched_domains.
2314  */
2315
2316 /* Don't have all balancing operations going off at once */
2317 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2318
2319 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2320                            enum idle_type idle)
2321 {
2322         unsigned long old_load, this_load;
2323         unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2324         struct sched_domain *sd;
2325         int i;
2326
2327         this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2328         /* Update our load */
2329         for (i = 0; i < 3; i++) {
2330                 unsigned long new_load = this_load;
2331                 int scale = 1 << i;
2332                 old_load = this_rq->cpu_load[i];
2333                 /*
2334                  * Round up the averaging division if load is increasing. This
2335                  * prevents us from getting stuck on 9 if the load is 10, for
2336                  * example.
2337                  */
2338                 if (new_load > old_load)
2339                         new_load += scale-1;
2340                 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2341         }
2342
2343         for_each_domain(this_cpu, sd) {
2344                 unsigned long interval;
2345
2346                 if (!(sd->flags & SD_LOAD_BALANCE))
2347                         continue;
2348
2349                 interval = sd->balance_interval;
2350                 if (idle != SCHED_IDLE)
2351                         interval *= sd->busy_factor;
2352
2353                 /* scale ms to jiffies */
2354                 interval = msecs_to_jiffies(interval);
2355                 if (unlikely(!interval))
2356                         interval = 1;
2357
2358                 if (j - sd->last_balance >= interval) {
2359                         if (load_balance(this_cpu, this_rq, sd, idle)) {
2360                                 /* We've pulled tasks over so either we're no
2361                                  * longer idle, or one of our SMT siblings is
2362                                  * not idle.
2363                                  */
2364                                 idle = NOT_IDLE;
2365                         }
2366                         sd->last_balance += interval;
2367                 }
2368         }
2369 }
2370 #else
2371 /*
2372  * on UP we do not need to balance between CPUs:
2373  */
2374 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2375 {
2376 }
2377 static inline void idle_balance(int cpu, runqueue_t *rq)
2378 {
2379 }
2380 #endif
2381
2382 static inline int wake_priority_sleeper(runqueue_t *rq)
2383 {
2384         int ret = 0;
2385 #ifdef CONFIG_SCHED_SMT
2386         spin_lock(&rq->lock);
2387         /*
2388          * If an SMT sibling task has been put to sleep for priority
2389          * reasons reschedule the idle task to see if it can now run.
2390          */
2391         if (rq->nr_running) {
2392                 resched_task(rq->idle);
2393                 ret = 1;
2394         }
2395         spin_unlock(&rq->lock);
2396 #endif
2397         return ret;
2398 }
2399
2400 DEFINE_PER_CPU(struct kernel_stat, kstat);
2401
2402 EXPORT_PER_CPU_SYMBOL(kstat);
2403
2404 /*
2405  * This is called on clock ticks and on context switches.
2406  * Bank in p->sched_time the ns elapsed since the last tick or switch.
2407  */
2408 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2409                                     unsigned long long now)
2410 {
2411         unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2412         p->sched_time += now - last;
2413 }
2414
2415 /*
2416  * Return current->sched_time plus any more ns on the sched_clock
2417  * that have not yet been banked.
2418  */
2419 unsigned long long current_sched_time(const task_t *tsk)
2420 {
2421         unsigned long long ns;
2422         unsigned long flags;
2423         local_irq_save(flags);
2424         ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2425         ns = tsk->sched_time + (sched_clock() - ns);
2426         local_irq_restore(flags);
2427         return ns;
2428 }
2429
2430 /*
2431  * We place interactive tasks back into the active array, if possible.
2432  *
2433  * To guarantee that this does not starve expired tasks we ignore the
2434  * interactivity of a task if the first expired task had to wait more
2435  * than a 'reasonable' amount of time. This deadline timeout is
2436  * load-dependent, as the frequency of array switched decreases with
2437  * increasing number of running tasks. We also ignore the interactivity
2438  * if a better static_prio task has expired:
2439  */
2440 #define EXPIRED_STARVING(rq) \
2441         ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2442                 (jiffies - (rq)->expired_timestamp >= \
2443                         STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2444                         ((rq)->curr->static_prio > (rq)->best_expired_prio))
2445
2446 /*
2447  * Account user cpu time to a process.
2448  * @p: the process that the cpu time gets accounted to
2449  * @hardirq_offset: the offset to subtract from hardirq_count()
2450  * @cputime: the cpu time spent in user space since the last update
2451  */
2452 void account_user_time(struct task_struct *p, cputime_t cputime)
2453 {
2454         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2455         cputime64_t tmp;
2456
2457         p->utime = cputime_add(p->utime, cputime);
2458
2459         /* Add user time to cpustat. */
2460         tmp = cputime_to_cputime64(cputime);
2461         if (TASK_NICE(p) > 0)
2462                 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2463         else
2464                 cpustat->user = cputime64_add(cpustat->user, tmp);
2465 }
2466
2467 /*
2468  * Account system cpu time to a process.
2469  * @p: the process that the cpu time gets accounted to
2470  * @hardirq_offset: the offset to subtract from hardirq_count()
2471  * @cputime: the cpu time spent in kernel space since the last update
2472  */
2473 void account_system_time(struct task_struct *p, int hardirq_offset,
2474                          cputime_t cputime)
2475 {
2476         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2477         runqueue_t *rq = this_rq();
2478         cputime64_t tmp;
2479
2480         p->stime = cputime_add(p->stime, cputime);
2481
2482         /* Add system time to cpustat. */
2483         tmp = cputime_to_cputime64(cputime);
2484         if (hardirq_count() - hardirq_offset)
2485                 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2486         else if (softirq_count())
2487                 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2488         else if (p != rq->idle)
2489                 cpustat->system = cputime64_add(cpustat->system, tmp);
2490         else if (atomic_read(&rq->nr_iowait) > 0)
2491                 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2492         else
2493                 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2494         /* Account for system time used */
2495         acct_update_integrals(p);
2496         /* Update rss highwater mark */
2497         update_mem_hiwater(p);
2498 }
2499
2500 /*
2501  * Account for involuntary wait time.
2502  * @p: the process from which the cpu time has been stolen
2503  * @steal: the cpu time spent in involuntary wait
2504  */
2505 void account_steal_time(struct task_struct *p, cputime_t steal)
2506 {
2507         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2508         cputime64_t tmp = cputime_to_cputime64(steal);
2509         runqueue_t *rq = this_rq();
2510
2511         if (p == rq->idle) {
2512                 p->stime = cputime_add(p->stime, steal);
2513                 if (atomic_read(&rq->nr_iowait) > 0)
2514                         cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2515                 else
2516                         cpustat->idle = cputime64_add(cpustat->idle, tmp);
2517         } else
2518                 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2519 }
2520
2521 /*
2522  * This function gets called by the timer code, with HZ frequency.
2523  * We call it with interrupts disabled.
2524  *
2525  * It also gets called by the fork code, when changing the parent's
2526  * timeslices.
2527  */
2528 void scheduler_tick(void)
2529 {
2530         int cpu = smp_processor_id();
2531         runqueue_t *rq = this_rq();
2532         task_t *p = current;
2533         unsigned long long now = sched_clock();
2534
2535         update_cpu_clock(p, rq, now);
2536
2537         rq->timestamp_last_tick = now;
2538
2539         if (p == rq->idle) {
2540                 if (wake_priority_sleeper(rq))
2541                         goto out;
2542                 rebalance_tick(cpu, rq, SCHED_IDLE);
2543                 return;
2544         }
2545
2546         /* Task might have expired already, but not scheduled off yet */
2547         if (p->array != rq->active) {
2548                 set_tsk_need_resched(p);
2549                 goto out;
2550         }
2551         spin_lock(&rq->lock);
2552         /*
2553          * The task was running during this tick - update the
2554          * time slice counter. Note: we do not update a thread's
2555          * priority until it either goes to sleep or uses up its
2556          * timeslice. This makes it possible for interactive tasks
2557          * to use up their timeslices at their highest priority levels.
2558          */
2559         if (rt_task(p)) {
2560                 /*
2561                  * RR tasks need a special form of timeslice management.
2562                  * FIFO tasks have no timeslices.
2563                  */
2564                 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2565                         p->time_slice = task_timeslice(p);
2566                         p->first_time_slice = 0;
2567                         set_tsk_need_resched(p);
2568
2569                         /* put it at the end of the queue: */
2570                         requeue_task(p, rq->active);
2571                 }
2572                 goto out_unlock;
2573         }
2574         if (!--p->time_slice) {
2575                 dequeue_task(p, rq->active);
2576                 set_tsk_need_resched(p);
2577                 p->prio = effective_prio(p);
2578                 p->time_slice = task_timeslice(p);
2579                 p->first_time_slice = 0;
2580
2581                 if (!rq->expired_timestamp)
2582                         rq->expired_timestamp = jiffies;
2583                 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2584                         enqueue_task(p, rq->expired);
2585                         if (p->static_prio < rq->best_expired_prio)
2586                                 rq->best_expired_prio = p->static_prio;
2587                 } else
2588                         enqueue_task(p, rq->active);
2589         } else {
2590                 /*
2591                  * Prevent a too long timeslice allowing a task to monopolize
2592                  * the CPU. We do this by splitting up the timeslice into
2593                  * smaller pieces.
2594                  *
2595                  * Note: this does not mean the task's timeslices expire or
2596                  * get lost in any way, they just might be preempted by
2597                  * another task of equal priority. (one with higher
2598                  * priority would have preempted this task already.) We
2599                  * requeue this task to the end of the list on this priority
2600                  * level, which is in essence a round-robin of tasks with
2601                  * equal priority.
2602                  *
2603                  * This only applies to tasks in the interactive
2604                  * delta range with at least TIMESLICE_GRANULARITY to requeue.
2605                  */
2606                 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2607                         p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2608                         (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2609                         (p->array == rq->active)) {
2610
2611                         requeue_task(p, rq->active);
2612                         set_tsk_need_resched(p);
2613                 }
2614         }
2615 out_unlock:
2616         spin_unlock(&rq->lock);
2617 out:
2618         rebalance_tick(cpu, rq, NOT_IDLE);
2619 }
2620
2621 #ifdef CONFIG_SCHED_SMT
2622 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2623 {
2624         /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2625         if (rq->curr == rq->idle && rq->nr_running)
2626                 resched_task(rq->idle);
2627 }
2628
2629 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2630 {
2631         struct sched_domain *tmp, *sd = NULL;
2632         cpumask_t sibling_map;
2633         int i;
2634
2635         for_each_domain(this_cpu, tmp)
2636                 if (tmp->flags & SD_SHARE_CPUPOWER)
2637                         sd = tmp;
2638
2639         if (!sd)
2640                 return;
2641
2642         /*
2643          * Unlock the current runqueue because we have to lock in
2644          * CPU order to avoid deadlocks. Caller knows that we might
2645          * unlock. We keep IRQs disabled.
2646          */
2647         spin_unlock(&this_rq->lock);
2648
2649         sibling_map = sd->span;
2650
2651         for_each_cpu_mask(i, sibling_map)
2652                 spin_lock(&cpu_rq(i)->lock);
2653         /*
2654          * We clear this CPU from the mask. This both simplifies the
2655          * inner loop and keps this_rq locked when we exit:
2656          */
2657         cpu_clear(this_cpu, sibling_map);
2658
2659         for_each_cpu_mask(i, sibling_map) {
2660                 runqueue_t *smt_rq = cpu_rq(i);
2661
2662                 wakeup_busy_runqueue(smt_rq);
2663         }
2664
2665         for_each_cpu_mask(i, sibling_map)
2666                 spin_unlock(&cpu_rq(i)->lock);
2667         /*
2668          * We exit with this_cpu's rq still held and IRQs
2669          * still disabled:
2670          */
2671 }
2672
2673 /*
2674  * number of 'lost' timeslices this task wont be able to fully
2675  * utilize, if another task runs on a sibling. This models the
2676  * slowdown effect of other tasks running on siblings:
2677  */
2678 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2679 {
2680         return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2681 }
2682
2683 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2684 {
2685         struct sched_domain *tmp, *sd = NULL;
2686         cpumask_t sibling_map;
2687         prio_array_t *array;
2688         int ret = 0, i;
2689         task_t *p;
2690
2691         for_each_domain(this_cpu, tmp)
2692                 if (tmp->flags & SD_SHARE_CPUPOWER)
2693                         sd = tmp;
2694
2695         if (!sd)
2696                 return 0;
2697
2698         /*
2699          * The same locking rules and details apply as for
2700          * wake_sleeping_dependent():
2701          */
2702         spin_unlock(&this_rq->lock);
2703         sibling_map = sd->span;
2704         for_each_cpu_mask(i, sibling_map)
2705                 spin_lock(&cpu_rq(i)->lock);
2706         cpu_clear(this_cpu, sibling_map);
2707
2708         /*
2709          * Establish next task to be run - it might have gone away because
2710          * we released the runqueue lock above:
2711          */
2712         if (!this_rq->nr_running)
2713                 goto out_unlock;
2714         array = this_rq->active;
2715         if (!array->nr_active)
2716                 array = this_rq->expired;
2717         BUG_ON(!array->nr_active);
2718
2719         p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2720                 task_t, run_list);
2721
2722         for_each_cpu_mask(i, sibling_map) {
2723                 runqueue_t *smt_rq = cpu_rq(i);
2724                 task_t *smt_curr = smt_rq->curr;
2725
2726                 /* Kernel threads do not participate in dependent sleeping */
2727                 if (!p->mm || !smt_curr->mm || rt_task(p))
2728                         goto check_smt_task;
2729
2730                 /*
2731                  * If a user task with lower static priority than the
2732                  * running task on the SMT sibling is trying to schedule,
2733                  * delay it till there is proportionately less timeslice
2734                  * left of the sibling task to prevent a lower priority
2735                  * task from using an unfair proportion of the
2736                  * physical cpu's resources. -ck
2737                  */
2738                 if (rt_task(smt_curr)) {
2739                         /*
2740                          * With real time tasks we run non-rt tasks only
2741                          * per_cpu_gain% of the time.
2742                          */
2743                         if ((jiffies % DEF_TIMESLICE) >
2744                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2745                                         ret = 1;
2746                 } else
2747                         if (smt_curr->static_prio < p->static_prio &&
2748                                 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2749                                 smt_slice(smt_curr, sd) > task_timeslice(p))
2750                                         ret = 1;
2751
2752 check_smt_task:
2753                 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2754                         rt_task(smt_curr))
2755                                 continue;
2756                 if (!p->mm) {
2757                         wakeup_busy_runqueue(smt_rq);
2758                         continue;
2759                 }
2760
2761                 /*
2762                  * Reschedule a lower priority task on the SMT sibling for
2763                  * it to be put to sleep, or wake it up if it has been put to
2764                  * sleep for priority reasons to see if it should run now.
2765                  */
2766                 if (rt_task(p)) {
2767                         if ((jiffies % DEF_TIMESLICE) >
2768                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2769                                         resched_task(smt_curr);
2770                 } else {
2771                         if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2772                                 smt_slice(p, sd) > task_timeslice(smt_curr))
2773                                         resched_task(smt_curr);
2774                         else
2775                                 wakeup_busy_runqueue(smt_rq);
2776                 }
2777         }
2778 out_unlock:
2779         for_each_cpu_mask(i, sibling_map)
2780                 spin_unlock(&cpu_rq(i)->lock);
2781         return ret;
2782 }
2783 #else
2784 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2785 {
2786 }
2787
2788 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2789 {
2790         return 0;
2791 }
2792 #endif
2793
2794 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2795
2796 void fastcall add_preempt_count(int val)
2797 {
2798         /*
2799          * Underflow?
2800          */
2801         BUG_ON((preempt_count() < 0));
2802         preempt_count() += val;
2803         /*
2804          * Spinlock count overflowing soon?
2805          */
2806         BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2807 }
2808 EXPORT_SYMBOL(add_preempt_count);
2809
2810 void fastcall sub_preempt_count(int val)
2811 {
2812         /*
2813          * Underflow?
2814          */
2815         BUG_ON(val > preempt_count());
2816         /*
2817          * Is the spinlock portion underflowing?
2818          */
2819         BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2820         preempt_count() -= val;
2821 }
2822 EXPORT_SYMBOL(sub_preempt_count);
2823
2824 #endif
2825
2826 /*
2827  * schedule() is the main scheduler function.
2828  */
2829 asmlinkage void __sched schedule(void)
2830 {
2831         long *switch_count;
2832         task_t *prev, *next;
2833         runqueue_t *rq;
2834         prio_array_t *array;
2835         struct list_head *queue;
2836         unsigned long long now;
2837         unsigned long run_time;
2838         int cpu, idx, new_prio;
2839
2840         /*
2841          * Test if we are atomic.  Since do_exit() needs to call into
2842          * schedule() atomically, we ignore that path for now.
2843          * Otherwise, whine if we are scheduling when we should not be.
2844          */
2845         if (likely(!current->exit_state)) {
2846                 if (unlikely(in_atomic())) {
2847                         printk(KERN_ERR "scheduling while atomic: "
2848                                 "%s/0x%08x/%d\n",
2849                                 current->comm, preempt_count(), current->pid);
2850                         dump_stack();
2851                 }
2852         }
2853         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2854
2855 need_resched:
2856         preempt_disable();
2857         prev = current;
2858         release_kernel_lock(prev);
2859 need_resched_nonpreemptible:
2860         rq = this_rq();
2861
2862         /*
2863          * The idle thread is not allowed to schedule!
2864          * Remove this check after it has been exercised a bit.
2865          */
2866         if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2867                 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2868                 dump_stack();
2869         }
2870
2871         schedstat_inc(rq, sched_cnt);
2872         now = sched_clock();
2873         if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2874                 run_time = now - prev->timestamp;
2875                 if (unlikely((long long)(now - prev->timestamp) < 0))
2876                         run_time = 0;
2877         } else
2878                 run_time = NS_MAX_SLEEP_AVG;
2879
2880         /*
2881          * Tasks charged proportionately less run_time at high sleep_avg to
2882          * delay them losing their interactive status
2883          */
2884         run_time /= (CURRENT_BONUS(prev) ? : 1);
2885
2886         spin_lock_irq(&rq->lock);
2887
2888         if (unlikely(prev->flags & PF_DEAD))
2889                 prev->state = EXIT_DEAD;
2890
2891         switch_count = &prev->nivcsw;
2892         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2893                 switch_count = &prev->nvcsw;
2894                 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2895                                 unlikely(signal_pending(prev))))
2896                         prev->state = TASK_RUNNING;
2897                 else {
2898                         if (prev->state == TASK_UNINTERRUPTIBLE)
2899                                 rq->nr_uninterruptible++;
2900                         deactivate_task(prev, rq);
2901                 }
2902         }
2903
2904         cpu = smp_processor_id();
2905         if (unlikely(!rq->nr_running)) {
2906 go_idle:
2907                 idle_balance(cpu, rq);
2908                 if (!rq->nr_running) {
2909                         next = rq->idle;
2910                         rq->expired_timestamp = 0;
2911                         wake_sleeping_dependent(cpu, rq);
2912                         /*
2913                          * wake_sleeping_dependent() might have released
2914                          * the runqueue, so break out if we got new
2915                          * tasks meanwhile:
2916                          */
2917                         if (!rq->nr_running)
2918                                 goto switch_tasks;
2919                 }
2920         } else {
2921                 if (dependent_sleeper(cpu, rq)) {
2922                         next = rq->idle;
2923                         goto switch_tasks;
2924                 }
2925                 /*
2926                  * dependent_sleeper() releases and reacquires the runqueue
2927                  * lock, hence go into the idle loop if the rq went
2928                  * empty meanwhile:
2929                  */
2930                 if (unlikely(!rq->nr_running))
2931                         goto go_idle;
2932         }
2933
2934         array = rq->active;
2935         if (unlikely(!array->nr_active)) {
2936                 /*
2937                  * Switch the active and expired arrays.
2938                  */
2939                 schedstat_inc(rq, sched_switch);
2940                 rq->active = rq->expired;
2941                 rq->expired = array;
2942                 array = rq->active;
2943                 rq->expired_timestamp = 0;
2944                 rq->best_expired_prio = MAX_PRIO;
2945         }
2946
2947         idx = sched_find_first_bit(array->bitmap);
2948         queue = array->queue + idx;
2949         next = list_entry(queue->next, task_t, run_list);
2950
2951         if (!rt_task(next) && next->activated > 0) {
2952                 unsigned long long delta = now - next->timestamp;
2953                 if (unlikely((long long)(now - next->timestamp) < 0))
2954                         delta = 0;
2955
2956                 if (next->activated == 1)
2957                         delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2958
2959                 array = next->array;
2960                 new_prio = recalc_task_prio(next, next->timestamp + delta);
2961
2962                 if (unlikely(next->prio != new_prio)) {
2963                         dequeue_task(next, array);
2964                         next->prio = new_prio;
2965                         enqueue_task(next, array);
2966                 } else
2967                         requeue_task(next, array);
2968         }
2969         next->activated = 0;
2970 switch_tasks:
2971         if (next == rq->idle)
2972                 schedstat_inc(rq, sched_goidle);
2973         prefetch(next);
2974         prefetch_stack(next);
2975         clear_tsk_need_resched(prev);
2976         rcu_qsctr_inc(task_cpu(prev));
2977
2978         update_cpu_clock(prev, rq, now);
2979
2980         prev->sleep_avg -= run_time;
2981         if ((long)prev->sleep_avg <= 0)
2982                 prev->sleep_avg = 0;
2983         prev->timestamp = prev->last_ran = now;
2984
2985         sched_info_switch(prev, next);
2986         if (likely(prev != next)) {
2987                 next->timestamp = now;
2988                 rq->nr_switches++;
2989                 rq->curr = next;
2990                 ++*switch_count;
2991
2992                 prepare_task_switch(rq, next);
2993                 prev = context_switch(rq, prev, next);
2994                 barrier();
2995                 /*
2996                  * this_rq must be evaluated again because prev may have moved
2997                  * CPUs since it called schedule(), thus the 'rq' on its stack
2998                  * frame will be invalid.
2999                  */
3000                 finish_task_switch(this_rq(), prev);
3001         } else
3002                 spin_unlock_irq(&rq->lock);
3003
3004         prev = current;
3005         if (unlikely(reacquire_kernel_lock(prev) < 0))
3006                 goto need_resched_nonpreemptible;
3007         preempt_enable_no_resched();
3008         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3009                 goto need_resched;
3010 }
3011
3012 EXPORT_SYMBOL(schedule);
3013
3014 #ifdef CONFIG_PREEMPT
3015 /*
3016  * this is is the entry point to schedule() from in-kernel preemption
3017  * off of preempt_enable.  Kernel preemptions off return from interrupt
3018  * occur there and call schedule directly.
3019  */
3020 asmlinkage void __sched preempt_schedule(void)
3021 {
3022         struct thread_info *ti = current_thread_info();
3023 #ifdef CONFIG_PREEMPT_BKL
3024         struct task_struct *task = current;
3025         int saved_lock_depth;
3026 #endif
3027         /*
3028          * If there is a non-zero preempt_count or interrupts are disabled,
3029          * we do not want to preempt the current task.  Just return..
3030          */
3031         if (unlikely(ti->preempt_count || irqs_disabled()))
3032                 return;
3033
3034 need_resched:
3035         add_preempt_count(PREEMPT_ACTIVE);
3036         /*
3037          * We keep the big kernel semaphore locked, but we
3038          * clear ->lock_depth so that schedule() doesnt
3039          * auto-release the semaphore:
3040          */
3041 #ifdef CONFIG_PREEMPT_BKL
3042         saved_lock_depth = task->lock_depth;
3043         task->lock_depth = -1;
3044 #endif
3045         schedule();
3046 #ifdef CONFIG_PREEMPT_BKL
3047         task->lock_depth = saved_lock_depth;
3048 #endif
3049         sub_preempt_count(PREEMPT_ACTIVE);
3050
3051         /* we could miss a preemption opportunity between schedule and now */
3052         barrier();
3053         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3054                 goto need_resched;
3055 }
3056
3057 EXPORT_SYMBOL(preempt_schedule);
3058
3059 /*
3060  * this is is the entry point to schedule() from kernel preemption
3061  * off of irq context.
3062  * Note, that this is called and return with irqs disabled. This will
3063  * protect us against recursive calling from irq.
3064  */
3065 asmlinkage void __sched preempt_schedule_irq(void)
3066 {
3067         struct thread_info *ti = current_thread_info();
3068 #ifdef CONFIG_PREEMPT_BKL
3069         struct task_struct *task = current;
3070         int saved_lock_depth;
3071 #endif
3072         /* Catch callers which need to be fixed*/
3073         BUG_ON(ti->preempt_count || !irqs_disabled());
3074
3075 need_resched:
3076         add_preempt_count(PREEMPT_ACTIVE);
3077         /*
3078          * We keep the big kernel semaphore locked, but we
3079          * clear ->lock_depth so that schedule() doesnt
3080          * auto-release the semaphore:
3081          */
3082 #ifdef CONFIG_PREEMPT_BKL
3083         saved_lock_depth = task->lock_depth;
3084         task->lock_depth = -1;
3085 #endif
3086         local_irq_enable();
3087         schedule();
3088         local_irq_disable();
3089 #ifdef CONFIG_PREEMPT_BKL
3090         task->lock_depth = saved_lock_depth;
3091 #endif
3092         sub_preempt_count(PREEMPT_ACTIVE);
3093
3094         /* we could miss a preemption opportunity between schedule and now */
3095         barrier();
3096         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3097                 goto need_resched;
3098 }
3099
3100 #endif /* CONFIG_PREEMPT */
3101
3102 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3103                           void *key)
3104 {
3105         task_t *p = curr->private;
3106         return try_to_wake_up(p, mode, sync);
3107 }
3108
3109 EXPORT_SYMBOL(default_wake_function);
3110
3111 /*
3112  * The core wakeup function.  Non-exclusive wakeups (nr_exclusive == 0) just
3113  * wake everything up.  If it's an exclusive wakeup (nr_exclusive == small +ve
3114  * number) then we wake all the non-exclusive tasks and one exclusive task.
3115  *
3116  * There are circumstances in which we can try to wake a task which has already
3117  * started to run but is not in state TASK_RUNNING.  try_to_wake_up() returns
3118  * zero in this (rare) case, and we handle it by continuing to scan the queue.
3119  */
3120 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3121                              int nr_exclusive, int sync, void *key)
3122 {
3123         struct list_head *tmp, *next;
3124
3125         list_for_each_safe(tmp, next, &q->task_list) {
3126                 wait_queue_t *curr;
3127                 unsigned flags;
3128                 curr = list_entry(tmp, wait_queue_t, task_list);
3129                 flags = curr->flags;
3130                 if (curr->func(curr, mode, sync, key) &&
3131                     (flags & WQ_FLAG_EXCLUSIVE) &&
3132                     !--nr_exclusive)
3133                         break;
3134         }
3135 }
3136
3137 /**
3138  * __wake_up - wake up threads blocked on a waitqueue.
3139  * @q: the waitqueue
3140  * @mode: which threads
3141  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3142  * @key: is directly passed to the wakeup function
3143  */
3144 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3145                         int nr_exclusive, void *key)
3146 {
3147         unsigned long flags;
3148
3149         spin_lock_irqsave(&q->lock, flags);
3150         __wake_up_common(q, mode, nr_exclusive, 0, key);
3151         spin_unlock_irqrestore(&q->lock, flags);
3152 }
3153
3154 EXPORT_SYMBOL(__wake_up);
3155
3156 /*
3157  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3158  */
3159 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3160 {
3161         __wake_up_common(q, mode, 1, 0, NULL);
3162 }
3163
3164 /**
3165  * __wake_up_sync - wake up threads blocked on a waitqueue.
3166  * @q: the waitqueue
3167  * @mode: which threads
3168  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3169  *
3170  * The sync wakeup differs that the waker knows that it will schedule
3171  * away soon, so while the target thread will be woken up, it will not
3172  * be migrated to another CPU - ie. the two threads are 'synchronized'
3173  * with each other. This can prevent needless bouncing between CPUs.
3174  *
3175  * On UP it can prevent extra preemption.
3176  */
3177 void fastcall
3178 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3179 {
3180         unsigned long flags;
3181         int sync = 1;
3182
3183         if (unlikely(!q))
3184                 return;
3185
3186         if (unlikely(!nr_exclusive))
3187                 sync = 0;
3188
3189         spin_lock_irqsave(&q->lock, flags);
3190         __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3191         spin_unlock_irqrestore(&q->lock, flags);
3192 }
3193 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3194
3195 void fastcall complete(struct completion *x)
3196 {
3197         unsigned long flags;
3198
3199         spin_lock_irqsave(&x->wait.lock, flags);
3200         x->done++;
3201         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3202                          1, 0, NULL);
3203         spin_unlock_irqrestore(&x->wait.lock, flags);
3204 }
3205 EXPORT_SYMBOL(complete);
3206
3207 void fastcall complete_all(struct completion *x)
3208 {
3209         unsigned long flags;
3210
3211         spin_lock_irqsave(&x->wait.lock, flags);
3212         x->done += UINT_MAX/2;
3213         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3214                          0, 0, NULL);
3215         spin_unlock_irqrestore(&x->wait.lock, flags);
3216 }
3217 EXPORT_SYMBOL(complete_all);
3218
3219 void fastcall __sched wait_for_completion(struct completion *x)
3220 {
3221         might_sleep();
3222         spin_lock_irq(&x->wait.lock);
3223         if (!x->done) {
3224                 DECLARE_WAITQUEUE(wait, current);
3225
3226                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3227                 __add_wait_queue_tail(&x->wait, &wait);
3228                 do {
3229                         __set_current_state(TASK_UNINTERRUPTIBLE);
3230                         spin_unlock_irq(&x->wait.lock);
3231                         schedule();
3232                         spin_lock_irq(&x->wait.lock);
3233                 } while (!x->done);
3234                 __remove_wait_queue(&x->wait, &wait);
3235         }
3236         x->done--;
3237         spin_unlock_irq(&x->wait.lock);
3238 }
3239 EXPORT_SYMBOL(wait_for_completion);
3240
3241 unsigned long fastcall __sched
3242 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3243 {
3244         might_sleep();
3245
3246         spin_lock_irq(&x->wait.lock);
3247         if (!x->done) {
3248                 DECLARE_WAITQUEUE(wait, current);
3249
3250                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3251                 __add_wait_queue_tail(&x->wait, &wait);
3252                 do {
3253                         __set_current_state(TASK_UNINTERRUPTIBLE);
3254                         spin_unlock_irq(&x->wait.lock);
3255                         timeout = schedule_timeout(timeout);
3256                         spin_lock_irq(&x->wait.lock);
3257                         if (!timeout) {
3258                                 __remove_wait_queue(&x->wait, &wait);
3259                                 goto out;
3260                         }
3261                 } while (!x->done);
3262                 __remove_wait_queue(&x->wait, &wait);
3263         }
3264         x->done--;
3265 out:
3266         spin_unlock_irq(&x->wait.lock);
3267         return timeout;
3268 }
3269 EXPORT_SYMBOL(wait_for_completion_timeout);
3270
3271 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3272 {
3273         int ret = 0;
3274
3275         might_sleep();
3276
3277         spin_lock_irq(&x->wait.lock);
3278         if (!x->done) {
3279                 DECLARE_WAITQUEUE(wait, current);
3280
3281                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3282                 __add_wait_queue_tail(&x->wait, &wait);
3283                 do {
3284                         if (signal_pending(current)) {
3285                                 ret = -ERESTARTSYS;
3286                                 __remove_wait_queue(&x->wait, &wait);
3287                                 goto out;
3288                         }
3289                         __set_current_state(TASK_INTERRUPTIBLE);
3290                         spin_unlock_irq(&x->wait.lock);
3291                         schedule();
3292                         spin_lock_irq(&x->wait.lock);
3293                 } while (!x->done);
3294                 __remove_wait_queue(&x->wait, &wait);
3295         }
3296         x->done--;
3297 out:
3298         spin_unlock_irq(&x->wait.lock);
3299
3300         return ret;
3301 }
3302 EXPORT_SYMBOL(wait_for_completion_interruptible);
3303
3304 unsigned long fastcall __sched
3305 wait_for_completion_interruptible_timeout(struct completion *x,
3306                                           unsigned long timeout)
3307 {
3308         might_sleep();
3309
3310         spin_lock_irq(&x->wait.lock);
3311         if (!x->done) {
3312                 DECLARE_WAITQUEUE(wait, current);
3313
3314                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3315                 __add_wait_queue_tail(&x->wait, &wait);
3316                 do {
3317                         if (signal_pending(current)) {
3318                                 timeout = -ERESTARTSYS;
3319                                 __remove_wait_queue(&x->wait, &wait);
3320                                 goto out;
3321                         }
3322                         __set_current_state(TASK_INTERRUPTIBLE);
3323                         spin_unlock_irq(&x->wait.lock);
3324                         timeout = schedule_timeout(timeout);
3325                         spin_lock_irq(&x->wait.lock);
3326                         if (!timeout) {
3327                                 __remove_wait_queue(&x->wait, &wait);
3328                                 goto out;
3329                         }
3330                 } while (!x->done);
3331                 __remove_wait_queue(&x->wait, &wait);
3332         }
3333         x->done--;
3334 out:
3335         spin_unlock_irq(&x->wait.lock);
3336         return timeout;
3337 }
3338 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3339
3340
3341 #define SLEEP_ON_VAR                                    \
3342         unsigned long flags;                            \
3343         wait_queue_t wait;                              \
3344         init_waitqueue_entry(&wait, current);
3345
3346 #define SLEEP_ON_HEAD                                   \
3347         spin_lock_irqsave(&q->lock,flags);              \
3348         __add_wait_queue(q, &wait);                     \
3349         spin_unlock(&q->lock);
3350
3351 #define SLEEP_ON_TAIL                                   \
3352         spin_lock_irq(&q->lock);                        \
3353         __remove_wait_queue(q, &wait);                  \
3354         spin_unlock_irqrestore(&q->lock, flags);
3355
3356 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3357 {
3358         SLEEP_ON_VAR
3359
3360         current->state = TASK_INTERRUPTIBLE;
3361
3362         SLEEP_ON_HEAD
3363         schedule();
3364         SLEEP_ON_TAIL
3365 }
3366
3367 EXPORT_SYMBOL(interruptible_sleep_on);
3368
3369 long fastcall __sched
3370 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3371 {
3372         SLEEP_ON_VAR
3373
3374         current->state = TASK_INTERRUPTIBLE;
3375
3376         SLEEP_ON_HEAD
3377         timeout = schedule_timeout(timeout);
3378         SLEEP_ON_TAIL
3379
3380         return timeout;
3381 }
3382
3383 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3384
3385 void fastcall __sched sleep_on(wait_queue_head_t *q)
3386 {
3387         SLEEP_ON_VAR
3388
3389         current->state = TASK_UNINTERRUPTIBLE;
3390
3391         SLEEP_ON_HEAD
3392         schedule();
3393         SLEEP_ON_TAIL
3394 }
3395
3396 EXPORT_SYMBOL(sleep_on);
3397
3398 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3399 {
3400         SLEEP_ON_VAR
3401
3402         current->state = TASK_UNINTERRUPTIBLE;
3403
3404         SLEEP_ON_HEAD
3405         timeout = schedule_timeout(timeout);
3406         SLEEP_ON_TAIL
3407
3408         return timeout;
3409 }
3410
3411 EXPORT_SYMBOL(sleep_on_timeout);
3412
3413 void set_user_nice(task_t *p, long nice)
3414 {
3415         unsigned long flags;
3416         prio_array_t *array;
3417         runqueue_t *rq;
3418         int old_prio, new_prio, delta;
3419
3420         if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3421                 return;
3422         /*
3423          * We have to be careful, if called from sys_setpriority(),
3424          * the task might be in the middle of scheduling on another CPU.
3425          */
3426         rq = task_rq_lock(p, &flags);
3427         /*
3428          * The RT priorities are set via sched_setscheduler(), but we still
3429          * allow the 'normal' nice value to be set - but as expected
3430          * it wont have any effect on scheduling until the task is
3431          * not SCHED_NORMAL:
3432          */
3433         if (rt_task(p)) {
3434                 p->static_prio = NICE_TO_PRIO(nice);
3435                 goto out_unlock;
3436         }
3437         array = p->array;
3438         if (array)
3439                 dequeue_task(p, array);
3440
3441         old_prio = p->prio;
3442         new_prio = NICE_TO_PRIO(nice);
3443         delta = new_prio - old_prio;
3444         p->static_prio = NICE_TO_PRIO(nice);
3445         p->prio += delta;
3446
3447         if (array) {
3448                 enqueue_task(p, array);
3449                 /*
3450                  * If the task increased its priority or is running and
3451                  * lowered its priority, then reschedule its CPU:
3452                  */
3453                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3454                         resched_task(rq->curr);
3455         }
3456 out_unlock:
3457         task_rq_unlock(rq, &flags);
3458 }
3459
3460 EXPORT_SYMBOL(set_user_nice);
3461
3462 /*
3463  * can_nice - check if a task can reduce its nice value
3464  * @p: task
3465  * @nice: nice value
3466  */
3467 int can_nice(const task_t *p, const int nice)
3468 {
3469         /* convert nice value [19,-20] to rlimit style value [1,40] */
3470         int nice_rlim = 20 - nice;
3471         return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3472                 capable(CAP_SYS_NICE));
3473 }
3474
3475 #ifdef __ARCH_WANT_SYS_NICE
3476
3477 /*
3478  * sys_nice - change the priority of the current process.
3479  * @increment: priority increment
3480  *
3481  * sys_setpriority is a more generic, but much slower function that
3482  * does similar things.
3483  */
3484 asmlinkage long sys_nice(int increment)
3485 {
3486         int retval;
3487         long nice;
3488
3489         /*
3490          * Setpriority might change our priority at the same moment.
3491          * We don't have to worry. Conceptually one call occurs first
3492          * and we have a single winner.
3493          */
3494         if (increment < -40)
3495                 increment = -40;
3496         if (increment > 40)
3497                 increment = 40;
3498
3499         nice = PRIO_TO_NICE(current->static_prio) + increment;
3500         if (nice < -20)
3501                 nice = -20;
3502         if (nice > 19)
3503                 nice = 19;
3504
3505         if (increment < 0 && !can_nice(current, nice))
3506                 return -EPERM;
3507
3508         retval = security_task_setnice(current, nice);
3509         if (retval)
3510                 return retval;
3511
3512         set_user_nice(current, nice);
3513         return 0;
3514 }
3515
3516 #endif
3517
3518 /**
3519  * task_prio - return the priority value of a given task.
3520  * @p: the task in question.
3521  *
3522  * This is the priority value as seen by users in /proc.
3523  * RT tasks are offset by -200. Normal tasks are centered
3524  * around 0, value goes from -16 to +15.
3525  */
3526 int task_prio(const task_t *p)
3527 {
3528         return p->prio - MAX_RT_PRIO;
3529 }
3530
3531 /**
3532  * task_nice - return the nice value of a given task.
3533  * @p: the task in question.
3534  */
3535 int task_nice(const task_t *p)
3536 {
3537         return TASK_NICE(p);
3538 }
3539 EXPORT_SYMBOL_GPL(task_nice);
3540
3541 /**
3542  * idle_cpu - is a given cpu idle currently?
3543  * @cpu: the processor in question.
3544  */
3545 int idle_cpu(int cpu)
3546 {
3547         return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3548 }
3549
3550 EXPORT_SYMBOL_GPL(idle_cpu);
3551
3552 /**
3553  * idle_task - return the idle task for a given cpu.
3554  * @cpu: the processor in question.
3555  */
3556 task_t *idle_task(int cpu)
3557 {
3558         return cpu_rq(cpu)->idle;
3559 }
3560
3561 /**
3562  * find_process_by_pid - find a process with a matching PID value.
3563  * @pid: the pid in question.
3564  */
3565 static inline task_t *find_process_by_pid(pid_t pid)
3566 {
3567         return pid ? find_task_by_pid(pid) : current;
3568 }
3569
3570 /* Actually do priority change: must hold rq lock. */
3571 static void __setscheduler(struct task_struct *p, int policy, int prio)
3572 {
3573         BUG_ON(p->array);
3574         p->policy = policy;
3575         p->rt_priority = prio;
3576         if (policy != SCHED_NORMAL)
3577                 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3578         else
3579                 p->prio = p->static_prio;
3580 }
3581
3582 /**
3583  * sched_setscheduler - change the scheduling policy and/or RT priority of
3584  * a thread.
3585  * @p: the task in question.
3586  * @policy: new policy.
3587  * @param: structure containing the new RT priority.
3588  */
3589 int sched_setscheduler(struct task_struct *p, int policy,
3590                        struct sched_param *param)
3591 {
3592         int retval;
3593         int oldprio, oldpolicy = -1;
3594         prio_array_t *array;
3595         unsigned long flags;
3596         runqueue_t *rq;
3597
3598 recheck:
3599         /* double check policy once rq lock held */
3600         if (policy < 0)
3601                 policy = oldpolicy = p->policy;
3602         else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3603                                 policy != SCHED_NORMAL)
3604                         return -EINVAL;
3605         /*
3606          * Valid priorities for SCHED_FIFO and SCHED_RR are
3607          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3608          */
3609         if (param->sched_priority < 0 ||
3610             (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3611             (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3612                 return -EINVAL;
3613         if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3614                 return -EINVAL;
3615
3616         /*
3617          * Allow unprivileged RT tasks to decrease priority:
3618          */
3619         if (!capable(CAP_SYS_NICE)) {
3620                 /* can't change policy */
3621                 if (policy != p->policy &&
3622                         !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3623                         return -EPERM;
3624                 /* can't increase priority */
3625                 if (policy != SCHED_NORMAL &&
3626                     param->sched_priority > p->rt_priority &&
3627                     param->sched_priority >
3628                                 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3629                         return -EPERM;
3630                 /* can't change other user's priorities */
3631                 if ((current->euid != p->euid) &&
3632                     (current->euid != p->uid))
3633                         return -EPERM;
3634         }
3635
3636         retval = security_task_setscheduler(p, policy, param);
3637         if (retval)
3638                 return retval;
3639         /*
3640          * To be able to change p->policy safely, the apropriate
3641          * runqueue lock must be held.
3642          */
3643         rq = task_rq_lock(p, &flags);
3644         /* recheck policy now with rq lock held */
3645         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3646                 policy = oldpolicy = -1;
3647                 task_rq_unlock(rq, &flags);
3648                 goto recheck;
3649         }
3650         array = p->array;
3651         if (array)
3652                 deactivate_task(p, rq);
3653         oldprio = p->prio;
3654         __setscheduler(p, policy, param->sched_priority);
3655         if (array) {
3656                 __activate_task(p, rq);
3657                 /*
3658                  * Reschedule if we are currently running on this runqueue and
3659                  * our priority decreased, or if we are not currently running on
3660                  * this runqueue and our priority is higher than the current's
3661                  */
3662                 if (task_running(rq, p)) {
3663                         if (p->prio > oldprio)
3664                                 resched_task(rq->curr);
3665                 } else if (TASK_PREEMPTS_CURR(p, rq))
3666                         resched_task(rq->curr);
3667         }
3668         task_rq_unlock(rq, &flags);
3669         return 0;
3670 }
3671 EXPORT_SYMBOL_GPL(sched_setscheduler);
3672
3673 static int
3674 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3675 {
3676         int retval;
3677         struct sched_param lparam;
3678         struct task_struct *p;
3679
3680         if (!param || pid < 0)
3681                 return -EINVAL;
3682         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3683                 return -EFAULT;
3684         read_lock_irq(&tasklist_lock);
3685         p = find_process_by_pid(pid);
3686         if (!p) {
3687                 read_unlock_irq(&tasklist_lock);
3688                 return -ESRCH;
3689         }
3690         retval = sched_setscheduler(p, policy, &lparam);
3691         read_unlock_irq(&tasklist_lock);
3692         return retval;
3693 }
3694
3695 /**
3696  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3697  * @pid: the pid in question.
3698  * @policy: new policy.
3699  * @param: structure containing the new RT priority.
3700  */
3701 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3702                                        struct sched_param __user *param)
3703 {
3704         return do_sched_setscheduler(pid, policy, param);
3705 }
3706
3707 /**
3708  * sys_sched_setparam - set/change the RT priority of a thread
3709  * @pid: the pid in question.
3710  * @param: structure containing the new RT priority.
3711  */
3712 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3713 {
3714         return do_sched_setscheduler(pid, -1, param);
3715 }
3716
3717 /**
3718  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3719  * @pid: the pid in question.
3720  */
3721 asmlinkage long sys_sched_getscheduler(pid_t pid)
3722 {
3723         int retval = -EINVAL;
3724         task_t *p;
3725
3726         if (pid < 0)
3727                 goto out_nounlock;
3728
3729         retval = -ESRCH;
3730         read_lock(&tasklist_lock);
3731         p = find_process_by_pid(pid);
3732         if (p) {
3733                 retval = security_task_getscheduler(p);
3734                 if (!retval)
3735                         retval = p->policy;
3736         }
3737         read_unlock(&tasklist_lock);
3738
3739 out_nounlock:
3740         return retval;
3741 }
3742
3743 /**
3744  * sys_sched_getscheduler - get the RT priority of a thread
3745  * @pid: the pid in question.
3746  * @param: structure containing the RT priority.
3747  */
3748 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3749 {
3750         struct sched_param lp;
3751         int retval = -EINVAL;
3752         task_t *p;
3753
3754         if (!param || pid < 0)
3755                 goto out_nounlock;
3756
3757         read_lock(&tasklist_lock);
3758         p = find_process_by_pid(pid);
3759         retval = -ESRCH;
3760         if (!p)
3761                 goto out_unlock;
3762
3763         retval = security_task_getscheduler(p);
3764         if (retval)
3765                 goto out_unlock;
3766
3767         lp.sched_priority = p->rt_priority;
3768         read_unlock(&tasklist_lock);
3769
3770         /*
3771          * This one might sleep, we cannot do it with a spinlock held ...
3772          */
3773         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3774
3775 out_nounlock:
3776         return retval;
3777
3778 out_unlock:
3779         read_unlock(&tasklist_lock);
3780         return retval;
3781 }
3782
3783 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3784 {
3785         task_t *p;
3786         int retval;
3787         cpumask_t cpus_allowed;
3788
3789         lock_cpu_hotplug();
3790         read_lock(&tasklist_lock);
3791
3792         p = find_process_by_pid(pid);
3793         if (!p) {
3794                 read_unlock(&tasklist_lock);
3795                 unlock_cpu_hotplug();
3796                 return -ESRCH;
3797         }
3798
3799         /*
3800          * It is not safe to call set_cpus_allowed with the
3801          * tasklist_lock held.  We will bump the task_struct's
3802          * usage count and then drop tasklist_lock.
3803          */
3804         get_task_struct(p);
3805         read_unlock(&tasklist_lock);
3806
3807         retval = -EPERM;
3808         if ((current->euid != p->euid) && (current->euid != p->uid) &&
3809                         !capable(CAP_SYS_NICE))
3810                 goto out_unlock;
3811
3812         cpus_allowed = cpuset_cpus_allowed(p);
3813         cpus_and(new_mask, new_mask, cpus_allowed);
3814         retval = set_cpus_allowed(p, new_mask);
3815
3816 out_unlock:
3817         put_task_struct(p);
3818         unlock_cpu_hotplug();
3819         return retval;
3820 }
3821
3822 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3823                              cpumask_t *new_mask)
3824 {
3825         if (len < sizeof(cpumask_t)) {
3826                 memset(new_mask, 0, sizeof(cpumask_t));
3827         } else if (len > sizeof(cpumask_t)) {
3828                 len = sizeof(cpumask_t);
3829         }
3830         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3831 }
3832
3833 /**
3834  * sys_sched_setaffinity - set the cpu affinity of a process
3835  * @pid: pid of the process
3836  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3837  * @user_mask_ptr: user-space pointer to the new cpu mask
3838  */
3839 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3840                                       unsigned long __user *user_mask_ptr)
3841 {
3842         cpumask_t new_mask;
3843         int retval;
3844
3845         retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3846         if (retval)
3847                 return retval;
3848
3849         return sched_setaffinity(pid, new_mask);
3850 }
3851
3852 /*
3853  * Represents all cpu's present in the system
3854  * In systems capable of hotplug, this map could dynamically grow
3855  * as new cpu's are detected in the system via any platform specific
3856  * method, such as ACPI for e.g.
3857  */
3858
3859 cpumask_t cpu_present_map;
3860 EXPORT_SYMBOL(cpu_present_map);
3861
3862 #ifndef CONFIG_SMP
3863 cpumask_t cpu_online_map = CPU_MASK_ALL;
3864 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3865 #endif
3866
3867 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3868 {
3869         int retval;
3870         task_t *p;
3871
3872         lock_cpu_hotplug();
3873         read_lock(&tasklist_lock);
3874
3875         retval = -ESRCH;
3876         p = find_process_by_pid(pid);
3877         if (!p)
3878                 goto out_unlock;
3879
3880         retval = 0;
3881         cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
3882
3883 out_unlock:
3884         read_unlock(&tasklist_lock);
3885         unlock_cpu_hotplug();
3886         if (retval)
3887                 return retval;
3888
3889         return 0;
3890 }
3891
3892 /**
3893  * sys_sched_getaffinity - get the cpu affinity of a process
3894  * @pid: pid of the process
3895  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3896  * @user_mask_ptr: user-space pointer to hold the current cpu mask
3897  */
3898 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3899                                       unsigned long __user *user_mask_ptr)
3900 {
3901         int ret;
3902         cpumask_t mask;
3903
3904         if (len < sizeof(cpumask_t))
3905                 return -EINVAL;
3906
3907         ret = sched_getaffinity(pid, &mask);
3908         if (ret < 0)
3909                 return ret;
3910
3911         if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3912                 return -EFAULT;
3913
3914         return sizeof(cpumask_t);
3915 }
3916
3917 /**
3918  * sys_sched_yield - yield the current processor to other threads.
3919  *
3920  * this function yields the current CPU by moving the calling thread
3921  * to the expired array. If there are no other threads running on this
3922  * CPU then this function will return.
3923  */
3924 asmlinkage long sys_sched_yield(void)
3925 {
3926         runqueue_t *rq = this_rq_lock();
3927         prio_array_t *array = current->array;
3928         prio_array_t *target = rq->expired;
3929
3930         schedstat_inc(rq, yld_cnt);
3931         /*
3932          * We implement yielding by moving the task into the expired
3933          * queue.
3934          *
3935          * (special rule: RT tasks will just roundrobin in the active
3936          *  array.)
3937          */
3938         if (rt_task(current))
3939                 target = rq->active;
3940
3941         if (current->array->nr_active == 1) {
3942                 schedstat_inc(rq, yld_act_empty);
3943                 if (!rq->expired->nr_active)
3944                         schedstat_inc(rq, yld_both_empty);
3945         } else if (!rq->expired->nr_active)
3946                 schedstat_inc(rq, yld_exp_empty);
3947
3948         if (array != target) {
3949                 dequeue_task(current, array);
3950                 enqueue_task(current, target);
3951         } else
3952                 /*
3953                  * requeue_task is cheaper so perform that if possible.
3954                  */
3955                 requeue_task(current, array);
3956
3957         /*
3958          * Since we are going to call schedule() anyway, there's
3959          * no need to preempt or enable interrupts:
3960          */
3961         __release(rq->lock);
3962         _raw_spin_unlock(&rq->lock);
3963         preempt_enable_no_resched();
3964
3965         schedule();
3966
3967         return 0;
3968 }
3969
3970 static inline void __cond_resched(void)
3971 {
3972         /*
3973          * The BKS might be reacquired before we have dropped
3974          * PREEMPT_ACTIVE, which could trigger a second
3975          * cond_resched() call.
3976          */
3977         if (unlikely(preempt_count()))
3978                 return;
3979         do {
3980                 add_preempt_count(PREEMPT_ACTIVE);
3981                 schedule();
3982                 sub_preempt_count(PREEMPT_ACTIVE);
3983         } while (need_resched());
3984 }
3985
3986 int __sched cond_resched(void)
3987 {
3988         if (need_resched()) {
3989                 __cond_resched();
3990                 return 1;
3991         }
3992         return 0;
3993 }
3994
3995 EXPORT_SYMBOL(cond_resched);
3996
3997 /*
3998  * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3999  * call schedule, and on return reacquire the lock.
4000  *
4001  * This works OK both with and without CONFIG_PREEMPT.  We do strange low-level
4002  * operations here to prevent schedule() from being called twice (once via
4003  * spin_unlock(), once by hand).
4004  */
4005 int cond_resched_lock(spinlock_t *lock)
4006 {
4007         int ret = 0;
4008
4009         if (need_lockbreak(lock)) {
4010                 spin_unlock(lock);
4011                 cpu_relax();
4012                 ret = 1;
4013                 spin_lock(lock);
4014         }
4015         if (need_resched()) {
4016                 _raw_spin_unlock(lock);
4017                 preempt_enable_no_resched();
4018                 __cond_resched();
4019                 ret = 1;
4020                 spin_lock(lock);
4021         }
4022         return ret;
4023 }
4024
4025 EXPORT_SYMBOL(cond_resched_lock);
4026
4027 int __sched cond_resched_softirq(void)
4028 {
4029         BUG_ON(!in_softirq());
4030
4031         if (need_resched()) {
4032                 __local_bh_enable();
4033                 __cond_resched();
4034                 local_bh_disable();
4035                 return 1;
4036         }
4037         return 0;
4038 }
4039
4040 EXPORT_SYMBOL(cond_resched_softirq);
4041
4042
4043 /**
4044  * yield - yield the current processor to other threads.
4045  *
4046  * this is a shortcut for kernel-space yielding - it marks the
4047  * thread runnable and calls sys_sched_yield().
4048  */
4049 void __sched yield(void)
4050 {
4051         set_current_state(TASK_RUNNING);
4052         sys_sched_yield();
4053 }
4054
4055 EXPORT_SYMBOL(yield);
4056
4057 /*
4058  * This task is about to go to sleep on IO.  Increment rq->nr_iowait so
4059  * that process accounting knows that this is a task in IO wait state.
4060  *
4061  * But don't do that if it is a deliberate, throttling IO wait (this task
4062  * has set its backing_dev_info: the queue against which it should throttle)
4063  */
4064 void __sched io_schedule(void)
4065 {
4066         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4067
4068         atomic_inc(&rq->nr_iowait);
4069         schedule();
4070         atomic_dec(&rq->nr_iowait);
4071 }
4072
4073 EXPORT_SYMBOL(io_schedule);
4074
4075 long __sched io_schedule_timeout(long timeout)
4076 {
4077         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4078         long ret;
4079
4080         atomic_inc(&rq->nr_iowait);
4081         ret = schedule_timeout(timeout);
4082         atomic_dec(&rq->nr_iowait);
4083         return ret;
4084 }
4085
4086 /**
4087  * sys_sched_get_priority_max - return maximum RT priority.
4088  * @policy: scheduling class.
4089  *
4090  * this syscall returns the maximum rt_priority that can be used
4091  * by a given scheduling class.
4092  */
4093 asmlinkage long sys_sched_get_priority_max(int policy)
4094 {
4095         int ret = -EINVAL;
4096
4097         switch (policy) {
4098         case SCHED_FIFO:
4099         case SCHED_RR:
4100                 ret = MAX_USER_RT_PRIO-1;
4101                 break;
4102         case SCHED_NORMAL:
4103                 ret = 0;
4104                 break;
4105         }
4106         return ret;
4107 }
4108
4109 /**
4110  * sys_sched_get_priority_min - return minimum RT priority.
4111  * @policy: scheduling class.
4112  *
4113  * this syscall returns the minimum rt_priority that can be used
4114  * by a given scheduling class.
4115  */
4116 asmlinkage long sys_sched_get_priority_min(int policy)
4117 {
4118         int ret = -EINVAL;
4119
4120         switch (policy) {
4121         case SCHED_FIFO:
4122         case SCHED_RR:
4123                 ret = 1;
4124                 break;
4125         case SCHED_NORMAL:
4126                 ret = 0;
4127         }
4128         return ret;
4129 }
4130
4131 /**
4132  * sys_sched_rr_get_interval - return the default timeslice of a process.
4133  * @pid: pid of the process.
4134  * @interval: userspace pointer to the timeslice value.
4135  *
4136  * this syscall writes the default timeslice value of a given process
4137  * into the user-space timespec buffer. A value of '0' means infinity.
4138  */
4139 asmlinkage
4140 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4141 {
4142         int retval = -EINVAL;
4143         struct timespec t;
4144         task_t *p;
4145
4146         if (pid < 0)
4147                 goto out_nounlock;
4148
4149         retval = -ESRCH;
4150         read_lock(&tasklist_lock);
4151         p = find_process_by_pid(pid);
4152         if (!p)
4153                 goto out_unlock;
4154
4155         retval = security_task_getscheduler(p);
4156         if (retval)
4157                 goto out_unlock;
4158
4159         jiffies_to_timespec(p->policy & SCHED_FIFO ?
4160                                 0 : task_timeslice(p), &t);
4161         read_unlock(&tasklist_lock);
4162         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4163 out_nounlock:
4164         return retval;
4165 out_unlock:
4166         read_unlock(&tasklist_lock);
4167         return retval;
4168 }
4169
4170 static inline struct task_struct *eldest_child(struct task_struct *p)
4171 {
4172         if (list_empty(&p->children)) return NULL;
4173         return list_entry(p->children.next,struct task_struct,sibling);
4174 }
4175
4176 static inline struct task_struct *older_sibling(struct task_struct *p)
4177 {
4178         if (p->sibling.prev==&p->parent->children) return NULL;
4179         return list_entry(p->sibling.prev,struct task_struct,sibling);
4180 }
4181
4182 static inline struct task_struct *younger_sibling(struct task_struct *p)
4183 {
4184         if (p->sibling.next==&p->parent->children) return NULL;
4185         return list_entry(p->sibling.next,struct task_struct,sibling);
4186 }
4187
4188 static void show_task(task_t *p)
4189 {
4190         task_t *relative;
4191         unsigned state;
4192         unsigned long free = 0;
4193         static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4194
4195         printk("%-13.13s ", p->comm);
4196         state = p->state ? __ffs(p->state) + 1 : 0;
4197         if (state < ARRAY_SIZE(stat_nam))
4198                 printk(stat_nam[state]);
4199         else
4200                 printk("?");
4201 #if (BITS_PER_LONG == 32)
4202         if (state == TASK_RUNNING)
4203                 printk(" running ");
4204         else
4205                 printk(" %08lX ", thread_saved_pc(p));
4206 #else
4207         if (state == TASK_RUNNING)
4208                 printk("  running task   ");
4209         else
4210                 printk(" %016lx ", thread_saved_pc(p));
4211 #endif
4212 #ifdef CONFIG_DEBUG_STACK_USAGE
4213         {
4214                 unsigned long *n = (unsigned long *) (p->thread_info+1);
4215                 while (!*n)
4216                         n++;
4217                 free = (unsigned long) n - (unsigned long)(p->thread_info+1);
4218         }
4219 #endif
4220         printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4221         if ((relative = eldest_child(p)))
4222                 printk("%5d ", relative->pid);
4223         else
4224                 printk("      ");
4225         if ((relative = younger_sibling(p)))
4226                 printk("%7d", relative->pid);
4227         else
4228                 printk("       ");
4229         if ((relative = older_sibling(p)))
4230                 printk(" %5d", relative->pid);
4231         else
4232                 printk("      ");
4233         if (!p->mm)
4234                 printk(" (L-TLB)\n");
4235         else
4236                 printk(" (NOTLB)\n");
4237
4238         if (state != TASK_RUNNING)
4239                 show_stack(p, NULL);
4240 }
4241
4242 void show_state(void)
4243 {
4244         task_t *g, *p;
4245
4246 #if (BITS_PER_LONG == 32)
4247         printk("\n"
4248                "                                               sibling\n");
4249         printk("  task             PC      pid father child younger older\n");
4250 #else
4251         printk("\n"
4252                "                                                       sibling\n");
4253         printk("  task                 PC          pid father child younger older\n");
4254 #endif
4255         read_lock(&tasklist_lock);
4256         do_each_thread(g, p) {
4257                 /*
4258                  * reset the NMI-timeout, listing all files on a slow
4259                  * console might take alot of time:
4260                  */
4261                 touch_nmi_watchdog();
4262                 show_task(p);
4263         } while_each_thread(g, p);
4264
4265         read_unlock(&tasklist_lock);
4266 }
4267
4268 /**
4269  * init_idle - set up an idle thread for a given CPU
4270  * @idle: task in question
4271  * @cpu: cpu the idle task belongs to
4272  *
4273  * NOTE: this function does not set the idle thread's NEED_RESCHED
4274  * flag, to make booting more robust.
4275  */
4276 void __devinit init_idle(task_t *idle, int cpu)
4277 {
4278         runqueue_t *rq = cpu_rq(cpu);
4279         unsigned long flags;
4280
4281         idle->sleep_avg = 0;
4282         idle->array = NULL;
4283         idle->prio = MAX_PRIO;
4284         idle->state = TASK_RUNNING;
4285         idle->cpus_allowed = cpumask_of_cpu(cpu);
4286         set_task_cpu(idle, cpu);
4287
4288         spin_lock_irqsave(&rq->lock, flags);
4289         rq->curr = rq->idle = idle;
4290 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4291         idle->oncpu = 1;
4292 #endif
4293         spin_unlock_irqrestore(&rq->lock, flags);
4294
4295         /* Set the preempt count _outside_ the spinlocks! */
4296 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4297         idle->thread_info->preempt_count = (idle->lock_depth >= 0);
4298 #else
4299         idle->thread_info->preempt_count = 0;
4300 #endif
4301 }
4302
4303 /*
4304  * In a system that switches off the HZ timer nohz_cpu_mask
4305  * indicates which cpus entered this state. This is used
4306  * in the rcu update to wait only for active cpus. For system
4307  * which do not switch off the HZ timer nohz_cpu_mask should
4308  * always be CPU_MASK_NONE.
4309  */
4310 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4311
4312 #ifdef CONFIG_SMP
4313 /*
4314  * This is how migration works:
4315  *
4316  * 1) we queue a migration_req_t structure in the source CPU's
4317  *    runqueue and wake up that CPU's migration thread.
4318  * 2) we down() the locked semaphore => thread blocks.
4319  * 3) migration thread wakes up (implicitly it forces the migrated
4320  *    thread off the CPU)
4321  * 4) it gets the migration request and checks whether the migrated
4322  *    task is still in the wrong runqueue.
4323  * 5) if it's in the wrong runqueue then the migration thread removes
4324  *    it and puts it into the right queue.
4325  * 6) migration thread up()s the semaphore.
4326  * 7) we wake up and the migration is done.
4327  */
4328
4329 /*
4330  * Change a given task's CPU affinity. Migrate the thread to a
4331  * proper CPU and schedule it away if the CPU it's executing on
4332  * is removed from the allowed bitmask.
4333  *
4334  * NOTE: the caller must have a valid reference to the task, the
4335  * task must not exit() & deallocate itself prematurely.  The
4336  * call is not atomic; no spinlocks may be held.
4337  */
4338 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4339 {
4340         unsigned long flags;
4341         int ret = 0;
4342         migration_req_t req;
4343         runqueue_t *rq;
4344
4345         rq = task_rq_lock(p, &flags);
4346         if (!cpus_intersects(new_mask, cpu_online_map)) {
4347                 ret = -EINVAL;
4348                 goto out;
4349         }
4350
4351         p->cpus_allowed = new_mask;
4352         /* Can the task run on the task's current CPU? If so, we're done */
4353         if (cpu_isset(task_cpu(p), new_mask))
4354                 goto out;
4355
4356         if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4357                 /* Need help from migration thread: drop lock and wait. */
4358                 task_rq_unlock(rq, &flags);
4359                 wake_up_process(rq->migration_thread);
4360                 wait_for_completion(&req.done);
4361                 tlb_migrate_finish(p->mm);
4362                 return 0;
4363         }
4364 out:
4365         task_rq_unlock(rq, &flags);
4366         return ret;
4367 }
4368
4369 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4370
4371 /*
4372  * Move (not current) task off this cpu, onto dest cpu.  We're doing
4373  * this because either it can't run here any more (set_cpus_allowed()
4374  * away from this CPU, or CPU going down), or because we're
4375  * attempting to rebalance this task on exec (sched_exec).
4376  *
4377  * So we race with normal scheduler movements, but that's OK, as long
4378  * as the task is no longer on this CPU.
4379  */
4380 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4381 {
4382         runqueue_t *rq_dest, *rq_src;
4383
4384         if (unlikely(cpu_is_offline(dest_cpu)))
4385                 return;
4386
4387         rq_src = cpu_rq(src_cpu);
4388         rq_dest = cpu_rq(dest_cpu);
4389
4390         double_rq_lock(rq_src, rq_dest);
4391         /* Already moved. */
4392         if (task_cpu(p) != src_cpu)
4393                 goto out;
4394         /* Affinity changed (again). */
4395         if (!cpu_isset(dest_cpu, p->cpus_allowed))
4396                 goto out;
4397
4398         set_task_cpu(p, dest_cpu);
4399         if (p->array) {
4400                 /*
4401                  * Sync timestamp with rq_dest's before activating.
4402                  * The same thing could be achieved by doing this step
4403                  * afterwards, and pretending it was a local activate.
4404                  * This way is cleaner and logically correct.
4405                  */
4406                 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4407                                 + rq_dest->timestamp_last_tick;
4408                 deactivate_task(p, rq_src);
4409                 activate_task(p, rq_dest, 0);
4410                 if (TASK_PREEMPTS_CURR(p, rq_dest))
4411                         resched_task(rq_dest->curr);
4412         }
4413
4414 out:
4415         double_rq_unlock(rq_src, rq_dest);
4416 }
4417
4418 /*
4419  * migration_thread - this is a highprio system thread that performs
4420  * thread migration by bumping thread off CPU then 'pushing' onto
4421  * another runqueue.
4422  */
4423 static int migration_thread(void *data)
4424 {
4425         runqueue_t *rq;
4426         int cpu = (long)data;
4427
4428         rq = cpu_rq(cpu);
4429         BUG_ON(rq->migration_thread != current);
4430
4431         set_current_state(TASK_INTERRUPTIBLE);
4432         while (!kthread_should_stop()) {
4433                 struct list_head *head;
4434                 migration_req_t *req;
4435
4436                 try_to_freeze();
4437
4438                 spin_lock_irq(&rq->lock);
4439
4440                 if (cpu_is_offline(cpu)) {
4441                         spin_unlock_irq(&rq->lock);
4442                         goto wait_to_die;
4443                 }
4444
4445                 if (rq->active_balance) {
4446                         active_load_balance(rq, cpu);
4447                         rq->active_balance = 0;
4448                 }
4449
4450                 head = &rq->migration_queue;
4451
4452                 if (list_empty(head)) {
4453                         spin_unlock_irq(&rq->lock);
4454                         schedule();
4455                         set_current_state(TASK_INTERRUPTIBLE);
4456                         continue;
4457                 }
4458                 req = list_entry(head->next, migration_req_t, list);
4459                 list_del_init(head->next);
4460
4461                 spin_unlock(&rq->lock);
4462                 __migrate_task(req->task, cpu, req->dest_cpu);
4463                 local_irq_enable();
4464
4465                 complete(&req->done);
4466         }
4467         __set_current_state(TASK_RUNNING);
4468         return 0;
4469
4470 wait_to_die:
4471         /* Wait for kthread_stop */
4472         set_current_state(TASK_INTERRUPTIBLE);
4473         while (!kthread_should_stop()) {
4474                 schedule();
4475                 set_current_state(TASK_INTERRUPTIBLE);
4476         }
4477         __set_current_state(TASK_RUNNING);
4478         return 0;
4479 }
4480
4481 #ifdef CONFIG_HOTPLUG_CPU
4482 /* Figure out where task on dead CPU should go, use force if neccessary. */
4483 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4484 {
4485         int dest_cpu;
4486         cpumask_t mask;
4487
4488         /* On same node? */
4489         mask = node_to_cpumask(cpu_to_node(dead_cpu));
4490         cpus_and(mask, mask, tsk->cpus_allowed);
4491         dest_cpu = any_online_cpu(mask);
4492
4493         /* On any allowed CPU? */
4494         if (dest_cpu == NR_CPUS)
4495                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4496
4497         /* No more Mr. Nice Guy. */
4498         if (dest_cpu == NR_CPUS) {
4499                 cpus_setall(tsk->cpus_allowed);
4500                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4501
4502                 /*
4503                  * Don't tell them about moving exiting tasks or
4504                  * kernel threads (both mm NULL), since they never
4505                  * leave kernel.
4506                  */
4507                 if (tsk->mm && printk_ratelimit())
4508                         printk(KERN_INFO "process %d (%s) no "
4509                                "longer affine to cpu%d\n",
4510                                tsk->pid, tsk->comm, dead_cpu);
4511         }
4512         __migrate_task(tsk, dead_cpu, dest_cpu);
4513 }
4514
4515 /*
4516  * While a dead CPU has no uninterruptible tasks queued at this point,
4517  * it might still have a nonzero ->nr_uninterruptible counter, because
4518  * for performance reasons the counter is not stricly tracking tasks to
4519  * their home CPUs. So we just add the counter to another CPU's counter,
4520  * to keep the global sum constant after CPU-down:
4521  */
4522 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4523 {
4524         runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4525         unsigned long flags;
4526
4527         local_irq_save(flags);
4528         double_rq_lock(rq_src, rq_dest);
4529         rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4530         rq_src->nr_uninterruptible = 0;
4531         double_rq_unlock(rq_src, rq_dest);
4532         local_irq_restore(flags);
4533 }
4534
4535 /* Run through task list and migrate tasks from the dead cpu. */
4536 static void migrate_live_tasks(int src_cpu)
4537 {
4538         struct task_struct *tsk, *t;
4539
4540         write_lock_irq(&tasklist_lock);
4541
4542         do_each_thread(t, tsk) {
4543                 if (tsk == current)
4544                         continue;
4545
4546                 if (task_cpu(tsk) == src_cpu)
4547                         move_task_off_dead_cpu(src_cpu, tsk);
4548         } while_each_thread(t, tsk);
4549
4550         write_unlock_irq(&tasklist_lock);
4551 }
4552
4553 /* Schedules idle task to be the next runnable task on current CPU.
4554  * It does so by boosting its priority to highest possible and adding it to
4555  * the _front_ of runqueue. Used by CPU offline code.
4556  */
4557 void sched_idle_next(void)
4558 {
4559         int cpu = smp_processor_id();
4560         runqueue_t *rq = this_rq();
4561         struct task_struct *p = rq->idle;
4562         unsigned long flags;
4563
4564         /* cpu has to be offline */
4565         BUG_ON(cpu_online(cpu));
4566
4567         /* Strictly not necessary since rest of the CPUs are stopped by now
4568          * and interrupts disabled on current cpu.
4569          */
4570         spin_lock_irqsave(&rq->lock, flags);
4571
4572         __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4573         /* Add idle task to _front_ of it's priority queue */
4574         __activate_idle_task(p, rq);
4575
4576         spin_unlock_irqrestore(&rq->lock, flags);
4577 }
4578
4579 /* Ensures that the idle task is using init_mm right before its cpu goes
4580  * offline.
4581  */
4582 void idle_task_exit(void)
4583 {
4584         struct mm_struct *mm = current->active_mm;
4585
4586         BUG_ON(cpu_online(smp_processor_id()));
4587
4588         if (mm != &init_mm)
4589                 switch_mm(mm, &init_mm, current);
4590         mmdrop(mm);
4591 }
4592
4593 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4594 {
4595         struct runqueue *rq = cpu_rq(dead_cpu);
4596
4597         /* Must be exiting, otherwise would be on tasklist. */
4598         BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4599
4600         /* Cannot have done final schedule yet: would have vanished. */
4601         BUG_ON(tsk->flags & PF_DEAD);
4602
4603         get_task_struct(tsk);
4604
4605         /*
4606          * Drop lock around migration; if someone else moves it,
4607          * that's OK.  No task can be added to this CPU, so iteration is
4608          * fine.
4609          */
4610         spin_unlock_irq(&rq->lock);
4611         move_task_off_dead_cpu(dead_cpu, tsk);
4612         spin_lock_irq(&rq->lock);
4613
4614         put_task_struct(tsk);
4615 }
4616
4617 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4618 static void migrate_dead_tasks(unsigned int dead_cpu)
4619 {
4620         unsigned arr, i;
4621         struct runqueue *rq = cpu_rq(dead_cpu);
4622
4623         for (arr = 0; arr < 2; arr++) {
4624                 for (i = 0; i < MAX_PRIO; i++) {
4625                         struct list_head *list = &rq->arrays[arr].queue[i];
4626                         while (!list_empty(list))
4627                                 migrate_dead(dead_cpu,
4628                                              list_entry(list->next, task_t,
4629                                                         run_list));
4630                 }
4631         }
4632 }
4633 #endif /* CONFIG_HOTPLUG_CPU */
4634
4635 /*
4636  * migration_call - callback that gets triggered when a CPU is added.
4637  * Here we can start up the necessary migration thread for the new CPU.
4638  */
4639 static int migration_call(struct notifier_block *nfb, unsigned long action,
4640                           void *hcpu)
4641 {
4642         int cpu = (long)hcpu;
4643         struct task_struct *p;
4644         struct runqueue *rq;
4645         unsigned long flags;
4646
4647         switch (action) {
4648         case CPU_UP_PREPARE:
4649                 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4650                 if (IS_ERR(p))
4651                         return NOTIFY_BAD;
4652                 p->flags |= PF_NOFREEZE;
4653                 kthread_bind(p, cpu);
4654                 /* Must be high prio: stop_machine expects to yield to it. */
4655                 rq = task_rq_lock(p, &flags);
4656                 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4657                 task_rq_unlock(rq, &flags);
4658                 cpu_rq(cpu)->migration_thread = p;
4659                 break;
4660         case CPU_ONLINE:
4661                 /* Strictly unneccessary, as first user will wake it. */
4662                 wake_up_process(cpu_rq(cpu)->migration_thread);
4663                 break;
4664 #ifdef CONFIG_HOTPLUG_CPU
4665         case CPU_UP_CANCELED:
4666                 /* Unbind it from offline cpu so it can run.  Fall thru. */
4667                 kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id());
4668                 kthread_stop(cpu_rq(cpu)->migration_thread);
4669                 cpu_rq(cpu)->migration_thread = NULL;
4670                 break;
4671         case CPU_DEAD:
4672                 migrate_live_tasks(cpu);
4673                 rq = cpu_rq(cpu);
4674                 kthread_stop(rq->migration_thread);
4675                 rq->migration_thread = NULL;
4676                 /* Idle task back to normal (off runqueue, low prio) */
4677                 rq = task_rq_lock(rq->idle, &flags);
4678                 deactivate_task(rq->idle, rq);
4679                 rq->idle->static_prio = MAX_PRIO;
4680                 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4681                 migrate_dead_tasks(cpu);
4682                 task_rq_unlock(rq, &flags);
4683                 migrate_nr_uninterruptible(rq);
4684                 BUG_ON(rq->nr_running != 0);
4685
4686                 /* No need to migrate the tasks: it was best-effort if
4687                  * they didn't do lock_cpu_hotplug().  Just wake up
4688                  * the requestors. */
4689                 spin_lock_irq(&rq->lock);
4690                 while (!list_empty(&rq->migration_queue)) {
4691                         migration_req_t *req;
4692                         req = list_entry(rq->migration_queue.next,
4693                                          migration_req_t, list);
4694                         list_del_init(&req->list);
4695                         complete(&req->done);
4696                 }
4697                 spin_unlock_irq(&rq->lock);
4698                 break;
4699 #endif
4700         }
4701         return NOTIFY_OK;
4702 }
4703
4704 /* Register at highest priority so that task migration (migrate_all_tasks)
4705  * happens before everything else.
4706  */
4707 static struct notifier_block __devinitdata migration_notifier = {
4708         .notifier_call = migration_call,
4709         .priority = 10
4710 };
4711
4712 int __init migration_init(void)
4713 {
4714         void *cpu = (void *)(long)smp_processor_id();
4715         /* Start one for boot CPU. */
4716         migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4717         migration_call(&migration_notifier, CPU_ONLINE, cpu);
4718         register_cpu_notifier(&migration_notifier);
4719         return 0;
4720 }
4721 #endif
4722
4723 #ifdef CONFIG_SMP
4724 #undef SCHED_DOMAIN_DEBUG
4725 #ifdef SCHED_DOMAIN_DEBUG
4726 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4727 {
4728         int level = 0;
4729
4730         if (!sd) {
4731                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4732                 return;
4733         }
4734
4735         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4736
4737         do {
4738                 int i;
4739                 char str[NR_CPUS];
4740                 struct sched_group *group = sd->groups;
4741                 cpumask_t groupmask;
4742
4743                 cpumask_scnprintf(str, NR_CPUS, sd->span);
4744                 cpus_clear(groupmask);
4745
4746                 printk(KERN_DEBUG);
4747                 for (i = 0; i < level + 1; i++)
4748                         printk(" ");
4749                 printk("domain %d: ", level);
4750
4751                 if (!(sd->flags & SD_LOAD_BALANCE)) {
4752                         printk("does not load-balance\n");
4753                         if (sd->parent)
4754                                 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4755                         break;
4756                 }
4757
4758                 printk("span %s\n", str);
4759
4760                 if (!cpu_isset(cpu, sd->span))
4761                         printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4762                 if (!cpu_isset(cpu, group->cpumask))
4763                         printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4764
4765                 printk(KERN_DEBUG);
4766                 for (i = 0; i < level + 2; i++)
4767                         printk(" ");
4768                 printk("groups:");
4769                 do {
4770                         if (!group) {
4771                                 printk("\n");
4772                                 printk(KERN_ERR "ERROR: group is NULL\n");
4773                                 break;
4774                         }
4775
4776                         if (!group->cpu_power) {
4777                                 printk("\n");
4778                                 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4779                         }
4780
4781                         if (!cpus_weight(group->cpumask)) {
4782                                 printk("\n");
4783                                 printk(KERN_ERR "ERROR: empty group\n");
4784                         }
4785
4786                         if (cpus_intersects(groupmask, group->cpumask)) {
4787                                 printk("\n");
4788                                 printk(KERN_ERR "ERROR: repeated CPUs\n");
4789                         }
4790
4791                         cpus_or(groupmask, groupmask, group->cpumask);
4792
4793                         cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4794                         printk(" %s", str);
4795
4796                         group = group->next;
4797                 } while (group != sd->groups);
4798                 printk("\n");
4799
4800                 if (!cpus_equal(sd->span, groupmask))
4801                         printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4802
4803                 level++;
4804                 sd = sd->parent;
4805
4806                 if (sd) {
4807                         if (!cpus_subset(groupmask, sd->span))
4808                                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4809                 }
4810
4811         } while (sd);
4812 }
4813 #else
4814 #define sched_domain_debug(sd, cpu) {}
4815 #endif
4816
4817 static int sd_degenerate(struct sched_domain *sd)
4818 {
4819         if (cpus_weight(sd->span) == 1)
4820                 return 1;
4821
4822         /* Following flags need at least 2 groups */
4823         if (sd->flags & (SD_LOAD_BALANCE |
4824                          SD_BALANCE_NEWIDLE |
4825                          SD_BALANCE_FORK |
4826                          SD_BALANCE_EXEC)) {
4827                 if (sd->groups != sd->groups->next)
4828                         return 0;
4829         }
4830
4831         /* Following flags don't use groups */
4832         if (sd->flags & (SD_WAKE_IDLE |
4833                          SD_WAKE_AFFINE |
4834                          SD_WAKE_BALANCE))
4835                 return 0;
4836
4837         return 1;
4838 }
4839
4840 static int sd_parent_degenerate(struct sched_domain *sd,
4841                                                 struct sched_domain *parent)
4842 {
4843         unsigned long cflags = sd->flags, pflags = parent->flags;
4844
4845         if (sd_degenerate(parent))
4846                 return 1;
4847
4848         if (!cpus_equal(sd->span, parent->span))
4849                 return 0;
4850
4851         /* Does parent contain flags not in child? */
4852         /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4853         if (cflags & SD_WAKE_AFFINE)
4854                 pflags &= ~SD_WAKE_BALANCE;
4855         /* Flags needing groups don't count if only 1 group in parent */
4856         if (parent->groups == parent->groups->next) {
4857                 pflags &= ~(SD_LOAD_BALANCE |
4858                                 SD_BALANCE_NEWIDLE |
4859                                 SD_BALANCE_FORK |
4860                                 SD_BALANCE_EXEC);
4861         }
4862         if (~cflags & pflags)
4863                 return 0;
4864
4865         return 1;
4866 }
4867
4868 /*
4869  * Attach the domain 'sd' to 'cpu' as its base domain.  Callers must
4870  * hold the hotplug lock.
4871  */
4872 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4873 {
4874         runqueue_t *rq = cpu_rq(cpu);
4875         struct sched_domain *tmp;
4876
4877         /* Remove the sched domains which do not contribute to scheduling. */
4878         for (tmp = sd; tmp; tmp = tmp->parent) {
4879                 struct sched_domain *parent = tmp->parent;
4880                 if (!parent)
4881                         break;
4882                 if (sd_parent_degenerate(tmp, parent))
4883                         tmp->parent = parent->parent;
4884         }
4885
4886         if (sd && sd_degenerate(sd))
4887                 sd = sd->parent;
4888
4889         sched_domain_debug(sd, cpu);
4890
4891         rcu_assign_pointer(rq->sd, sd);
4892 }
4893
4894 /* cpus with isolated domains */
4895 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4896
4897 /* Setup the mask of cpus configured for isolated domains */
4898 static int __init isolated_cpu_setup(char *str)
4899 {
4900         int ints[NR_CPUS], i;
4901
4902         str = get_options(str, ARRAY_SIZE(ints), ints);
4903         cpus_clear(cpu_isolated_map);
4904         for (i = 1; i <= ints[0]; i++)
4905                 if (ints[i] < NR_CPUS)
4906                         cpu_set(ints[i], cpu_isolated_map);
4907         return 1;
4908 }
4909
4910 __setup ("isolcpus=", isolated_cpu_setup);
4911
4912 /*
4913  * init_sched_build_groups takes an array of groups, the cpumask we wish
4914  * to span, and a pointer to a function which identifies what group a CPU
4915  * belongs to. The return value of group_fn must be a valid index into the
4916  * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4917  * keep track of groups covered with a cpumask_t).
4918  *
4919  * init_sched_build_groups will build a circular linked list of the groups
4920  * covered by the given span, and will set each group's ->cpumask correctly,
4921  * and ->cpu_power to 0.
4922  */
4923 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
4924                                     int (*group_fn)(int cpu))
4925 {
4926         struct sched_group *first = NULL, *last = NULL;
4927         cpumask_t covered = CPU_MASK_NONE;
4928         int i;
4929
4930         for_each_cpu_mask(i, span) {
4931                 int group = group_fn(i);
4932                 struct sched_group *sg = &groups[group];
4933                 int j;
4934
4935                 if (cpu_isset(i, covered))
4936                         continue;
4937
4938                 sg->cpumask = CPU_MASK_NONE;
4939                 sg->cpu_power = 0;
4940
4941                 for_each_cpu_mask(j, span) {
4942                         if (group_fn(j) != group)
4943                                 continue;
4944
4945                         cpu_set(j, covered);
4946                         cpu_set(j, sg->cpumask);
4947                 }
4948                 if (!first)
4949                         first = sg;
4950                 if (last)
4951                         last->next = sg;
4952                 last = sg;
4953         }
4954         last->next = first;
4955 }
4956
4957 #define SD_NODES_PER_DOMAIN 16
4958
4959 #ifdef CONFIG_NUMA
4960 /**
4961  * find_next_best_node - find the next node to include in a sched_domain
4962  * @node: node whose sched_domain we're building
4963  * @used_nodes: nodes already in the sched_domain
4964  *
4965  * Find the next node to include in a given scheduling domain.  Simply
4966  * finds the closest node not already in the @used_nodes map.
4967  *
4968  * Should use nodemask_t.
4969  */
4970 static int find_next_best_node(int node, unsigned long *used_nodes)
4971 {
4972         int i, n, val, min_val, best_node = 0;
4973
4974         min_val = INT_MAX;
4975
4976         for (i = 0; i < MAX_NUMNODES; i++) {
4977                 /* Start at @node */
4978                 n = (node + i) % MAX_NUMNODES;
4979
4980                 if (!nr_cpus_node(n))
4981                         continue;
4982
4983                 /* Skip already used nodes */
4984                 if (test_bit(n, used_nodes))
4985                         continue;
4986
4987                 /* Simple min distance search */
4988                 val = node_distance(node, n);
4989
4990                 if (val < min_val) {
4991                         min_val = val;
4992                         best_node = n;
4993                 }
4994         }
4995
4996         set_bit(best_node, used_nodes);
4997         return best_node;
4998 }
4999
5000 /**
5001  * sched_domain_node_span - get a cpumask for a node's sched_domain
5002  * @node: node whose cpumask we're constructing
5003  * @size: number of nodes to include in this span
5004  *
5005  * Given a node, construct a good cpumask for its sched_domain to span.  It
5006  * should be one that prevents unnecessary balancing, but also spreads tasks
5007  * out optimally.
5008  */
5009 static cpumask_t sched_domain_node_span(int node)
5010 {
5011         int i;
5012         cpumask_t span, nodemask;
5013         DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5014
5015         cpus_clear(span);
5016         bitmap_zero(used_nodes, MAX_NUMNODES);
5017
5018         nodemask = node_to_cpumask(node);
5019         cpus_or(span, span, nodemask);
5020         set_bit(node, used_nodes);
5021
5022         for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5023                 int next_node = find_next_best_node(node, used_nodes);
5024                 nodemask = node_to_cpumask(next_node);
5025                 cpus_or(span, span, nodemask);
5026         }
5027
5028         return span;
5029 }
5030 #endif
5031
5032 /*
5033  * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5034  * can switch it on easily if needed.
5035  */
5036 #ifdef CONFIG_SCHED_SMT
5037 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5038 static struct sched_group sched_group_cpus[NR_CPUS];
5039 static int cpu_to_cpu_group(int cpu)
5040 {
5041         return cpu;
5042 }
5043 #endif
5044
5045 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5046 static struct sched_group sched_group_phys[NR_CPUS];
5047 static int cpu_to_phys_group(int cpu)
5048 {
5049 #ifdef CONFIG_SCHED_SMT
5050         return first_cpu(cpu_sibling_map[cpu]);
5051 #else
5052         return cpu;
5053 #endif
5054 }
5055
5056 #ifdef CONFIG_NUMA
5057 /*
5058  * The init_sched_build_groups can't handle what we want to do with node
5059  * groups, so roll our own. Now each node has its own list of groups which
5060  * gets dynamically allocated.
5061  */
5062 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5063 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5064
5065 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5066 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5067
5068 static int cpu_to_allnodes_group(int cpu)
5069 {
5070         return cpu_to_node(cpu);
5071 }
5072 #endif
5073
5074 /*
5075  * Build sched domains for a given set of cpus and attach the sched domains
5076  * to the individual cpus
5077  */
5078 void build_sched_domains(const cpumask_t *cpu_map)
5079 {
5080         int i;
5081 #ifdef CONFIG_NUMA
5082         struct sched_group **sched_group_nodes = NULL;
5083         struct sched_group *sched_group_allnodes = NULL;
5084
5085         /*
5086          * Allocate the per-node list of sched groups
5087          */
5088         sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5089                                            GFP_ATOMIC);
5090         if (!sched_group_nodes) {
5091                 printk(KERN_WARNING "Can not alloc sched group node list\n");
5092                 return;
5093         }
5094         sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5095 #endif
5096
5097         /*
5098          * Set up domains for cpus specified by the cpu_map.
5099          */
5100         for_each_cpu_mask(i, *cpu_map) {
5101                 int group;
5102                 struct sched_domain *sd = NULL, *p;
5103                 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5104
5105                 cpus_and(nodemask, nodemask, *cpu_map);
5106
5107 #ifdef CONFIG_NUMA
5108                 if (cpus_weight(*cpu_map)
5109                                 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5110                         if (!sched_group_allnodes) {
5111                                 sched_group_allnodes
5112                                         = kmalloc(sizeof(struct sched_group)
5113                                                         * MAX_NUMNODES,
5114                                                   GFP_KERNEL);
5115                                 if (!sched_group_allnodes) {
5116                                         printk(KERN_WARNING
5117                                         "Can not alloc allnodes sched group\n");
5118                                         break;
5119                                 }
5120                                 sched_group_allnodes_bycpu[i]
5121                                                 = sched_group_allnodes;
5122                         }
5123                         sd = &per_cpu(allnodes_domains, i);
5124                         *sd = SD_ALLNODES_INIT;
5125                         sd->span = *cpu_map;
5126                         group = cpu_to_allnodes_group(i);
5127                         sd->groups = &sched_group_allnodes[group];
5128                         p = sd;
5129                 } else
5130                         p = NULL;
5131
5132                 sd = &per_cpu(node_domains, i);
5133                 *sd = SD_NODE_INIT;
5134                 sd->span = sched_domain_node_span(cpu_to_node(i));
5135                 sd->parent = p;
5136                 cpus_and(sd->span, sd->span, *cpu_map);
5137 #endif
5138
5139                 p = sd;
5140                 sd = &per_cpu(phys_domains, i);
5141                 group = cpu_to_phys_group(i);
5142                 *sd = SD_CPU_INIT;
5143                 sd->span = nodemask;
5144                 sd->parent = p;
5145                 sd->groups = &sched_group_phys[group];
5146
5147 #ifdef CONFIG_SCHED_SMT
5148                 p = sd;
5149                 sd = &per_cpu(cpu_domains, i);
5150                 group = cpu_to_cpu_group(i);
5151                 *sd = SD_SIBLING_INIT;
5152                 sd->span = cpu_sibling_map[i];
5153                 cpus_and(sd->span, sd->span, *cpu_map);
5154                 sd->parent = p;
5155                 sd->groups = &sched_group_cpus[group];
5156 #endif
5157         }
5158
5159 #ifdef CONFIG_SCHED_SMT
5160         /* Set up CPU (sibling) groups */
5161         for_each_cpu_mask(i, *cpu_map) {
5162                 cpumask_t this_sibling_map = cpu_sibling_map[i];
5163                 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5164                 if (i != first_cpu(this_sibling_map))
5165                         continue;
5166
5167                 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5168                                                 &cpu_to_cpu_group);
5169         }
5170 #endif
5171
5172         /* Set up physical groups */
5173         for (i = 0; i < MAX_NUMNODES; i++) {
5174                 cpumask_t nodemask = node_to_cpumask(i);
5175
5176                 cpus_and(nodemask, nodemask, *cpu_map);
5177                 if (cpus_empty(nodemask))
5178                         continue;
5179
5180                 init_sched_build_groups(sched_group_phys, nodemask,
5181                                                 &cpu_to_phys_group);
5182         }
5183
5184 #ifdef CONFIG_NUMA
5185         /* Set up node groups */
5186         if (sched_group_allnodes)
5187                 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5188                                         &cpu_to_allnodes_group);
5189
5190         for (i = 0; i < MAX_NUMNODES; i++) {
5191                 /* Set up node groups */
5192                 struct sched_group *sg, *prev;
5193                 cpumask_t nodemask = node_to_cpumask(i);
5194                 cpumask_t domainspan;
5195                 cpumask_t covered = CPU_MASK_NONE;
5196                 int j;
5197
5198                 cpus_and(nodemask, nodemask, *cpu_map);
5199                 if (cpus_empty(nodemask)) {
5200                         sched_group_nodes[i] = NULL;
5201                         continue;
5202                 }
5203
5204                 domainspan = sched_domain_node_span(i);
5205                 cpus_and(domainspan, domainspan, *cpu_map);
5206
5207                 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5208                 sched_group_nodes[i] = sg;
5209                 for_each_cpu_mask(j, nodemask) {
5210                         struct sched_domain *sd;
5211                         sd = &per_cpu(node_domains, j);
5212                         sd->groups = sg;
5213                         if (sd->groups == NULL) {
5214                                 /* Turn off balancing if we have no groups */
5215                                 sd->flags = 0;
5216                         }
5217                 }
5218                 if (!sg) {
5219                         printk(KERN_WARNING
5220                         "Can not alloc domain group for node %d\n", i);
5221                         continue;
5222                 }
5223                 sg->cpu_power = 0;
5224                 sg->cpumask = nodemask;
5225                 cpus_or(covered, covered, nodemask);
5226                 prev = sg;
5227
5228                 for (j = 0; j < MAX_NUMNODES; j++) {
5229                         cpumask_t tmp, notcovered;
5230                         int n = (i + j) % MAX_NUMNODES;
5231
5232                         cpus_complement(notcovered, covered);
5233                         cpus_and(tmp, notcovered, *cpu_map);
5234                         cpus_and(tmp, tmp, domainspan);
5235                         if (cpus_empty(tmp))
5236                                 break;
5237
5238                         nodemask = node_to_cpumask(n);
5239                         cpus_and(tmp, tmp, nodemask);
5240                         if (cpus_empty(tmp))
5241                                 continue;
5242
5243                         sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5244                         if (!sg) {
5245                                 printk(KERN_WARNING
5246                                 "Can not alloc domain group for node %d\n", j);
5247                                 break;
5248                         }
5249                         sg->cpu_power = 0;
5250                         sg->cpumask = tmp;
5251                         cpus_or(covered, covered, tmp);
5252                         prev->next = sg;
5253                         prev = sg;
5254                 }
5255                 prev->next = sched_group_nodes[i];
5256         }
5257 #endif
5258
5259         /* Calculate CPU power for physical packages and nodes */
5260         for_each_cpu_mask(i, *cpu_map) {
5261                 int power;
5262                 struct sched_domain *sd;
5263 #ifdef CONFIG_SCHED_SMT
5264                 sd = &per_cpu(cpu_domains, i);
5265                 power = SCHED_LOAD_SCALE;
5266                 sd->groups->cpu_power = power;
5267 #endif
5268
5269                 sd = &per_cpu(phys_domains, i);
5270                 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5271                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5272                 sd->groups->cpu_power = power;
5273
5274 #ifdef CONFIG_NUMA
5275                 sd = &per_cpu(allnodes_domains, i);
5276                 if (sd->groups) {
5277                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5278                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5279                         sd->groups->cpu_power = power;
5280                 }
5281 #endif
5282         }
5283
5284 #ifdef CONFIG_NUMA
5285         for (i = 0; i < MAX_NUMNODES; i++) {
5286                 struct sched_group *sg = sched_group_nodes[i];
5287                 int j;
5288
5289                 if (sg == NULL)
5290                         continue;
5291 next_sg:
5292                 for_each_cpu_mask(j, sg->cpumask) {
5293                         struct sched_domain *sd;
5294                         int power;
5295
5296                         sd = &per_cpu(phys_domains, j);
5297                         if (j != first_cpu(sd->groups->cpumask)) {
5298                                 /*
5299                                  * Only add "power" once for each
5300                                  * physical package.
5301                                  */
5302                                 continue;
5303                         }
5304                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5305                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5306
5307                         sg->cpu_power += power;
5308                 }
5309                 sg = sg->next;
5310                 if (sg != sched_group_nodes[i])
5311                         goto next_sg;
5312         }
5313 #endif
5314
5315         /* Attach the domains */
5316         for_each_cpu_mask(i, *cpu_map) {
5317                 struct sched_domain *sd;
5318 #ifdef CONFIG_SCHED_SMT
5319                 sd = &per_cpu(cpu_domains, i);
5320 #else
5321                 sd = &per_cpu(phys_domains, i);
5322 #endif
5323                 cpu_attach_domain(sd, i);
5324         }
5325 }
5326 /*
5327  * Set up scheduler domains and groups.  Callers must hold the hotplug lock.
5328  */
5329 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5330 {
5331         cpumask_t cpu_default_map;
5332
5333         /*
5334          * Setup mask for cpus without special case scheduling requirements.
5335          * For now this just excludes isolated cpus, but could be used to
5336          * exclude other special cases in the future.
5337          */
5338         cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5339
5340         build_sched_domains(&cpu_default_map);
5341 }
5342
5343 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5344 {
5345 #ifdef CONFIG_NUMA
5346         int i;
5347         int cpu;
5348
5349         for_each_cpu_mask(cpu, *cpu_map) {
5350                 struct sched_group *sched_group_allnodes
5351                         = sched_group_allnodes_bycpu[cpu];
5352                 struct sched_group **sched_group_nodes
5353                         = sched_group_nodes_bycpu[cpu];
5354
5355                 if (sched_group_allnodes) {
5356                         kfree(sched_group_allnodes);
5357                         sched_group_allnodes_bycpu[cpu] = NULL;
5358                 }
5359
5360                 if (!sched_group_nodes)
5361                         continue;
5362
5363                 for (i = 0; i < MAX_NUMNODES; i++) {
5364                         cpumask_t nodemask = node_to_cpumask(i);
5365                         struct sched_group *oldsg, *sg = sched_group_nodes[i];
5366
5367                         cpus_and(nodemask, nodemask, *cpu_map);
5368                         if (cpus_empty(nodemask))
5369                                 continue;
5370
5371                         if (sg == NULL)
5372                                 continue;
5373                         sg = sg->next;
5374 next_sg:
5375                         oldsg = sg;
5376                         sg = sg->next;
5377                         kfree(oldsg);
5378                         if (oldsg != sched_group_nodes[i])
5379                                 goto next_sg;
5380                 }
5381                 kfree(sched_group_nodes);
5382                 sched_group_nodes_bycpu[cpu] = NULL;
5383         }
5384 #endif
5385 }
5386
5387 /*
5388  * Detach sched domains from a group of cpus specified in cpu_map
5389  * These cpus will now be attached to the NULL domain
5390  */
5391 static inline void detach_destroy_domains(const cpumask_t *cpu_map)
5392 {
5393         int i;
5394
5395         for_each_cpu_mask(i, *cpu_map)
5396                 cpu_attach_domain(NULL, i);
5397         synchronize_sched();
5398         arch_destroy_sched_domains(cpu_map);
5399 }
5400
5401 /*
5402  * Partition sched domains as specified by the cpumasks below.
5403  * This attaches all cpus from the cpumasks to the NULL domain,
5404  * waits for a RCU quiescent period, recalculates sched
5405  * domain information and then attaches them back to the
5406  * correct sched domains
5407  * Call with hotplug lock held
5408  */
5409 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
5410 {
5411         cpumask_t change_map;
5412
5413         cpus_and(*partition1, *partition1, cpu_online_map);
5414         cpus_and(*partition2, *partition2, cpu_online_map);
5415         cpus_or(change_map, *partition1, *partition2);
5416
5417         /* Detach sched domains from all of the affected cpus */
5418         detach_destroy_domains(&change_map);
5419         if (!cpus_empty(*partition1))
5420                 build_sched_domains(partition1);
5421         if (!cpus_empty(*partition2))
5422                 build_sched_domains(partition2);
5423 }
5424
5425 #ifdef CONFIG_HOTPLUG_CPU
5426 /*
5427  * Force a reinitialization of the sched domains hierarchy.  The domains
5428  * and groups cannot be updated in place without racing with the balancing
5429  * code, so we temporarily attach all running cpus to the NULL domain
5430  * which will prevent rebalancing while the sched domains are recalculated.
5431  */
5432 static int update_sched_domains(struct notifier_block *nfb,
5433                                 unsigned long action, void *hcpu)
5434 {
5435         switch (action) {
5436         case CPU_UP_PREPARE:
5437         case CPU_DOWN_PREPARE:
5438                 detach_destroy_domains(&cpu_online_map);
5439                 return NOTIFY_OK;
5440
5441         case CPU_UP_CANCELED:
5442         case CPU_DOWN_FAILED:
5443         case CPU_ONLINE:
5444         case CPU_DEAD:
5445                 /*
5446                  * Fall through and re-initialise the domains.
5447                  */
5448                 break;
5449         default:
5450                 return NOTIFY_DONE;
5451         }
5452
5453         /* The hotplug lock is already held by cpu_up/cpu_down */
5454         arch_init_sched_domains(&cpu_online_map);
5455
5456         return NOTIFY_OK;
5457 }
5458 #endif
5459
5460 void __init sched_init_smp(void)
5461 {
5462         lock_cpu_hotplug();
5463         arch_init_sched_domains(&cpu_online_map);
5464         unlock_cpu_hotplug();
5465         /* XXX: Theoretical race here - CPU may be hotplugged now */
5466         hotcpu_notifier(update_sched_domains, 0);
5467 }
5468 #else
5469 void __init sched_init_smp(void)
5470 {
5471 }
5472 #endif /* CONFIG_SMP */
5473
5474 int in_sched_functions(unsigned long addr)
5475 {
5476         /* Linker adds these: start and end of __sched functions */
5477         extern char __sched_text_start[], __sched_text_end[];
5478         return in_lock_functions(addr) ||
5479                 (addr >= (unsigned long)__sched_text_start
5480                 && addr < (unsigned long)__sched_text_end);
5481 }
5482
5483 void __init sched_init(void)
5484 {
5485         runqueue_t *rq;
5486         int i, j, k;
5487
5488         for (i = 0; i < NR_CPUS; i++) {
5489                 prio_array_t *array;
5490
5491                 rq = cpu_rq(i);
5492                 spin_lock_init(&rq->lock);
5493                 rq->nr_running = 0;
5494                 rq->active = rq->arrays;
5495                 rq->expired = rq->arrays + 1;
5496                 rq->best_expired_prio = MAX_PRIO;
5497
5498 #ifdef CONFIG_SMP
5499                 rq->sd = NULL;
5500                 for (j = 1; j < 3; j++)
5501                         rq->cpu_load[j] = 0;
5502                 rq->active_balance = 0;
5503                 rq->push_cpu = 0;
5504                 rq->migration_thread = NULL;
5505                 INIT_LIST_HEAD(&rq->migration_queue);
5506 #endif
5507                 atomic_set(&rq->nr_iowait, 0);
5508
5509                 for (j = 0; j < 2; j++) {
5510                         array = rq->arrays + j;
5511                         for (k = 0; k < MAX_PRIO; k++) {
5512                                 INIT_LIST_HEAD(array->queue + k);
5513                                 __clear_bit(k, array->bitmap);
5514                         }
5515                         // delimiter for bitsearch
5516                         __set_bit(MAX_PRIO, array->bitmap);
5517                 }
5518         }
5519
5520         /*
5521          * The boot idle thread does lazy MMU switching as well:
5522          */
5523         atomic_inc(&init_mm.mm_count);
5524         enter_lazy_tlb(&init_mm, current);
5525
5526         /*
5527          * Make us the idle thread. Technically, schedule() should not be
5528          * called from this thread, however somewhere below it might be,
5529          * but because we are the idle thread, we just pick up running again
5530          * when this runqueue becomes "idle".
5531          */
5532         init_idle(current, smp_processor_id());
5533 }
5534
5535 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5536 void __might_sleep(char *file, int line)
5537 {
5538 #if defined(in_atomic)
5539         static unsigned long prev_jiffy;        /* ratelimiting */
5540
5541         if ((in_atomic() || irqs_disabled()) &&
5542             system_state == SYSTEM_RUNNING && !oops_in_progress) {
5543                 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5544                         return;
5545                 prev_jiffy = jiffies;
5546                 printk(KERN_ERR "Debug: sleeping function called from invalid"
5547                                 " context at %s:%d\n", file, line);
5548                 printk("in_atomic():%d, irqs_disabled():%d\n",
5549                         in_atomic(), irqs_disabled());
5550                 dump_stack();
5551         }
5552 #endif
5553 }
5554 EXPORT_SYMBOL(__might_sleep);
5555 #endif
5556
5557 #ifdef CONFIG_MAGIC_SYSRQ
5558 void normalize_rt_tasks(void)
5559 {
5560         struct task_struct *p;
5561         prio_array_t *array;
5562         unsigned long flags;
5563         runqueue_t *rq;
5564
5565         read_lock_irq(&tasklist_lock);
5566         for_each_process (p) {
5567                 if (!rt_task(p))
5568                         continue;
5569
5570                 rq = task_rq_lock(p, &flags);
5571
5572                 array = p->array;
5573                 if (array)
5574                         deactivate_task(p, task_rq(p));
5575                 __setscheduler(p, SCHED_NORMAL, 0);
5576                 if (array) {
5577                         __activate_task(p, task_rq(p));
5578                         resched_task(rq->curr);
5579                 }
5580
5581                 task_rq_unlock(rq, &flags);
5582         }
5583         read_unlock_irq(&tasklist_lock);
5584 }
5585
5586 #endif /* CONFIG_MAGIC_SYSRQ */