Merge master.kernel.org:/home/rmk/linux-2.6-serial
[linux-2.6] / kernel / sched.c
1 /*
2  *  kernel/sched.c
3  *
4  *  Kernel scheduler and related syscalls
5  *
6  *  Copyright (C) 1991-2002  Linus Torvalds
7  *
8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
9  *              make semaphores SMP safe
10  *  1998-11-19  Implemented schedule_timeout() and related stuff
11  *              by Andrea Arcangeli
12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
13  *              hybrid priority-list and round-robin design with
14  *              an array-switch method of distributing timeslices
15  *              and per-CPU runqueues.  Cleanups and useful suggestions
16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
17  *  2003-09-03  Interactivity tuning by Con Kolivas.
18  *  2004-04-02  Scheduler domains code by Nick Piggin
19  */
20
21 #include <linux/mm.h>
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/completion.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/security.h>
33 #include <linux/notifier.h>
34 #include <linux/profile.h>
35 #include <linux/suspend.h>
36 #include <linux/blkdev.h>
37 #include <linux/delay.h>
38 #include <linux/smp.h>
39 #include <linux/threads.h>
40 #include <linux/timer.h>
41 #include <linux/rcupdate.h>
42 #include <linux/cpu.h>
43 #include <linux/cpuset.h>
44 #include <linux/percpu.h>
45 #include <linux/kthread.h>
46 #include <linux/seq_file.h>
47 #include <linux/syscalls.h>
48 #include <linux/times.h>
49 #include <linux/acct.h>
50 #include <asm/tlb.h>
51
52 #include <asm/unistd.h>
53
54 /*
55  * Convert user-nice values [ -20 ... 0 ... 19 ]
56  * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
57  * and back.
58  */
59 #define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
60 #define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
61 #define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
62
63 /*
64  * 'User priority' is the nice value converted to something we
65  * can work with better when scaling various scheduler parameters,
66  * it's a [ 0 ... 39 ] range.
67  */
68 #define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
69 #define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
70 #define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
71
72 /*
73  * Some helpers for converting nanosecond timing to jiffy resolution
74  */
75 #define NS_TO_JIFFIES(TIME)     ((TIME) / (1000000000 / HZ))
76 #define JIFFIES_TO_NS(TIME)     ((TIME) * (1000000000 / HZ))
77
78 /*
79  * These are the 'tuning knobs' of the scheduler:
80  *
81  * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
82  * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
83  * Timeslices get refilled after they expire.
84  */
85 #define MIN_TIMESLICE           max(5 * HZ / 1000, 1)
86 #define DEF_TIMESLICE           (100 * HZ / 1000)
87 #define ON_RUNQUEUE_WEIGHT       30
88 #define CHILD_PENALTY            95
89 #define PARENT_PENALTY          100
90 #define EXIT_WEIGHT               3
91 #define PRIO_BONUS_RATIO         25
92 #define MAX_BONUS               (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
93 #define INTERACTIVE_DELTA         2
94 #define MAX_SLEEP_AVG           (DEF_TIMESLICE * MAX_BONUS)
95 #define STARVATION_LIMIT        (MAX_SLEEP_AVG)
96 #define NS_MAX_SLEEP_AVG        (JIFFIES_TO_NS(MAX_SLEEP_AVG))
97
98 /*
99  * If a task is 'interactive' then we reinsert it in the active
100  * array after it has expired its current timeslice. (it will not
101  * continue to run immediately, it will still roundrobin with
102  * other interactive tasks.)
103  *
104  * This part scales the interactivity limit depending on niceness.
105  *
106  * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
107  * Here are a few examples of different nice levels:
108  *
109  *  TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
110  *  TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
111  *  TASK_INTERACTIVE(  0): [1,1,1,1,0,0,0,0,0,0,0]
112  *  TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
113  *  TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
114  *
115  * (the X axis represents the possible -5 ... 0 ... +5 dynamic
116  *  priority range a task can explore, a value of '1' means the
117  *  task is rated interactive.)
118  *
119  * Ie. nice +19 tasks can never get 'interactive' enough to be
120  * reinserted into the active array. And only heavily CPU-hog nice -20
121  * tasks will be expired. Default nice 0 tasks are somewhere between,
122  * it takes some effort for them to get interactive, but it's not
123  * too hard.
124  */
125
126 #define CURRENT_BONUS(p) \
127         (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
128                 MAX_SLEEP_AVG)
129
130 #define GRANULARITY     (10 * HZ / 1000 ? : 1)
131
132 #ifdef CONFIG_SMP
133 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
134                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
135                         num_online_cpus())
136 #else
137 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
138                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
139 #endif
140
141 #define SCALE(v1,v1_max,v2_max) \
142         (v1) * (v2_max) / (v1_max)
143
144 #define DELTA(p) \
145         (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
146
147 #define TASK_INTERACTIVE(p) \
148         ((p)->prio <= (p)->static_prio - DELTA(p))
149
150 #define INTERACTIVE_SLEEP(p) \
151         (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
152                 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
153
154 #define TASK_PREEMPTS_CURR(p, rq) \
155         ((p)->prio < (rq)->curr->prio)
156
157 /*
158  * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
159  * to time slice values: [800ms ... 100ms ... 5ms]
160  *
161  * The higher a thread's priority, the bigger timeslices
162  * it gets during one round of execution. But even the lowest
163  * priority thread gets MIN_TIMESLICE worth of execution time.
164  */
165
166 #define SCALE_PRIO(x, prio) \
167         max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
168
169 static unsigned int task_timeslice(task_t *p)
170 {
171         if (p->static_prio < NICE_TO_PRIO(0))
172                 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
173         else
174                 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
175 }
176 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran)       \
177                                 < (long long) (sd)->cache_hot_time)
178
179 /*
180  * These are the runqueue data structures:
181  */
182
183 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
184
185 typedef struct runqueue runqueue_t;
186
187 struct prio_array {
188         unsigned int nr_active;
189         unsigned long bitmap[BITMAP_SIZE];
190         struct list_head queue[MAX_PRIO];
191 };
192
193 /*
194  * This is the main, per-CPU runqueue data structure.
195  *
196  * Locking rule: those places that want to lock multiple runqueues
197  * (such as the load balancing or the thread migration code), lock
198  * acquire operations must be ordered by ascending &runqueue.
199  */
200 struct runqueue {
201         spinlock_t lock;
202
203         /*
204          * nr_running and cpu_load should be in the same cacheline because
205          * remote CPUs use both these fields when doing load calculation.
206          */
207         unsigned long nr_running;
208 #ifdef CONFIG_SMP
209         unsigned long cpu_load[3];
210 #endif
211         unsigned long long nr_switches;
212
213         /*
214          * This is part of a global counter where only the total sum
215          * over all CPUs matters. A task can increase this counter on
216          * one CPU and if it got migrated afterwards it may decrease
217          * it on another CPU. Always updated under the runqueue lock:
218          */
219         unsigned long nr_uninterruptible;
220
221         unsigned long expired_timestamp;
222         unsigned long long timestamp_last_tick;
223         task_t *curr, *idle;
224         struct mm_struct *prev_mm;
225         prio_array_t *active, *expired, arrays[2];
226         int best_expired_prio;
227         atomic_t nr_iowait;
228
229 #ifdef CONFIG_SMP
230         struct sched_domain *sd;
231
232         /* For active balancing */
233         int active_balance;
234         int push_cpu;
235
236         task_t *migration_thread;
237         struct list_head migration_queue;
238 #endif
239
240 #ifdef CONFIG_SCHEDSTATS
241         /* latency stats */
242         struct sched_info rq_sched_info;
243
244         /* sys_sched_yield() stats */
245         unsigned long yld_exp_empty;
246         unsigned long yld_act_empty;
247         unsigned long yld_both_empty;
248         unsigned long yld_cnt;
249
250         /* schedule() stats */
251         unsigned long sched_switch;
252         unsigned long sched_cnt;
253         unsigned long sched_goidle;
254
255         /* try_to_wake_up() stats */
256         unsigned long ttwu_cnt;
257         unsigned long ttwu_local;
258 #endif
259 };
260
261 static DEFINE_PER_CPU(struct runqueue, runqueues);
262
263 /*
264  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
265  * See detach_destroy_domains: synchronize_sched for details.
266  *
267  * The domain tree of any CPU may only be accessed from within
268  * preempt-disabled sections.
269  */
270 #define for_each_domain(cpu, domain) \
271 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
272
273 #define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
274 #define this_rq()               (&__get_cpu_var(runqueues))
275 #define task_rq(p)              cpu_rq(task_cpu(p))
276 #define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
277
278 #ifndef prepare_arch_switch
279 # define prepare_arch_switch(next)      do { } while (0)
280 #endif
281 #ifndef finish_arch_switch
282 # define finish_arch_switch(prev)       do { } while (0)
283 #endif
284
285 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
286 static inline int task_running(runqueue_t *rq, task_t *p)
287 {
288         return rq->curr == p;
289 }
290
291 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
292 {
293 }
294
295 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
296 {
297 #ifdef CONFIG_DEBUG_SPINLOCK
298         /* this is a valid case when another task releases the spinlock */
299         rq->lock.owner = current;
300 #endif
301         spin_unlock_irq(&rq->lock);
302 }
303
304 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
305 static inline int task_running(runqueue_t *rq, task_t *p)
306 {
307 #ifdef CONFIG_SMP
308         return p->oncpu;
309 #else
310         return rq->curr == p;
311 #endif
312 }
313
314 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
315 {
316 #ifdef CONFIG_SMP
317         /*
318          * We can optimise this out completely for !SMP, because the
319          * SMP rebalancing from interrupt is the only thing that cares
320          * here.
321          */
322         next->oncpu = 1;
323 #endif
324 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
325         spin_unlock_irq(&rq->lock);
326 #else
327         spin_unlock(&rq->lock);
328 #endif
329 }
330
331 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
332 {
333 #ifdef CONFIG_SMP
334         /*
335          * After ->oncpu is cleared, the task can be moved to a different CPU.
336          * We must ensure this doesn't happen until the switch is completely
337          * finished.
338          */
339         smp_wmb();
340         prev->oncpu = 0;
341 #endif
342 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
343         local_irq_enable();
344 #endif
345 }
346 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
347
348 /*
349  * task_rq_lock - lock the runqueue a given task resides on and disable
350  * interrupts.  Note the ordering: we can safely lookup the task_rq without
351  * explicitly disabling preemption.
352  */
353 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
354         __acquires(rq->lock)
355 {
356         struct runqueue *rq;
357
358 repeat_lock_task:
359         local_irq_save(*flags);
360         rq = task_rq(p);
361         spin_lock(&rq->lock);
362         if (unlikely(rq != task_rq(p))) {
363                 spin_unlock_irqrestore(&rq->lock, *flags);
364                 goto repeat_lock_task;
365         }
366         return rq;
367 }
368
369 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
370         __releases(rq->lock)
371 {
372         spin_unlock_irqrestore(&rq->lock, *flags);
373 }
374
375 #ifdef CONFIG_SCHEDSTATS
376 /*
377  * bump this up when changing the output format or the meaning of an existing
378  * format, so that tools can adapt (or abort)
379  */
380 #define SCHEDSTAT_VERSION 12
381
382 static int show_schedstat(struct seq_file *seq, void *v)
383 {
384         int cpu;
385
386         seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
387         seq_printf(seq, "timestamp %lu\n", jiffies);
388         for_each_online_cpu(cpu) {
389                 runqueue_t *rq = cpu_rq(cpu);
390 #ifdef CONFIG_SMP
391                 struct sched_domain *sd;
392                 int dcnt = 0;
393 #endif
394
395                 /* runqueue-specific stats */
396                 seq_printf(seq,
397                     "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
398                     cpu, rq->yld_both_empty,
399                     rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
400                     rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
401                     rq->ttwu_cnt, rq->ttwu_local,
402                     rq->rq_sched_info.cpu_time,
403                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
404
405                 seq_printf(seq, "\n");
406
407 #ifdef CONFIG_SMP
408                 /* domain-specific stats */
409                 preempt_disable();
410                 for_each_domain(cpu, sd) {
411                         enum idle_type itype;
412                         char mask_str[NR_CPUS];
413
414                         cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
415                         seq_printf(seq, "domain%d %s", dcnt++, mask_str);
416                         for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
417                                         itype++) {
418                                 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
419                                     sd->lb_cnt[itype],
420                                     sd->lb_balanced[itype],
421                                     sd->lb_failed[itype],
422                                     sd->lb_imbalance[itype],
423                                     sd->lb_gained[itype],
424                                     sd->lb_hot_gained[itype],
425                                     sd->lb_nobusyq[itype],
426                                     sd->lb_nobusyg[itype]);
427                         }
428                         seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
429                             sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
430                             sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
431                             sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
432                             sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
433                 }
434                 preempt_enable();
435 #endif
436         }
437         return 0;
438 }
439
440 static int schedstat_open(struct inode *inode, struct file *file)
441 {
442         unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
443         char *buf = kmalloc(size, GFP_KERNEL);
444         struct seq_file *m;
445         int res;
446
447         if (!buf)
448                 return -ENOMEM;
449         res = single_open(file, show_schedstat, NULL);
450         if (!res) {
451                 m = file->private_data;
452                 m->buf = buf;
453                 m->size = size;
454         } else
455                 kfree(buf);
456         return res;
457 }
458
459 struct file_operations proc_schedstat_operations = {
460         .open    = schedstat_open,
461         .read    = seq_read,
462         .llseek  = seq_lseek,
463         .release = single_release,
464 };
465
466 # define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
467 # define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
468 #else /* !CONFIG_SCHEDSTATS */
469 # define schedstat_inc(rq, field)       do { } while (0)
470 # define schedstat_add(rq, field, amt)  do { } while (0)
471 #endif
472
473 /*
474  * rq_lock - lock a given runqueue and disable interrupts.
475  */
476 static inline runqueue_t *this_rq_lock(void)
477         __acquires(rq->lock)
478 {
479         runqueue_t *rq;
480
481         local_irq_disable();
482         rq = this_rq();
483         spin_lock(&rq->lock);
484
485         return rq;
486 }
487
488 #ifdef CONFIG_SCHEDSTATS
489 /*
490  * Called when a process is dequeued from the active array and given
491  * the cpu.  We should note that with the exception of interactive
492  * tasks, the expired queue will become the active queue after the active
493  * queue is empty, without explicitly dequeuing and requeuing tasks in the
494  * expired queue.  (Interactive tasks may be requeued directly to the
495  * active queue, thus delaying tasks in the expired queue from running;
496  * see scheduler_tick()).
497  *
498  * This function is only called from sched_info_arrive(), rather than
499  * dequeue_task(). Even though a task may be queued and dequeued multiple
500  * times as it is shuffled about, we're really interested in knowing how
501  * long it was from the *first* time it was queued to the time that it
502  * finally hit a cpu.
503  */
504 static inline void sched_info_dequeued(task_t *t)
505 {
506         t->sched_info.last_queued = 0;
507 }
508
509 /*
510  * Called when a task finally hits the cpu.  We can now calculate how
511  * long it was waiting to run.  We also note when it began so that we
512  * can keep stats on how long its timeslice is.
513  */
514 static inline void sched_info_arrive(task_t *t)
515 {
516         unsigned long now = jiffies, diff = 0;
517         struct runqueue *rq = task_rq(t);
518
519         if (t->sched_info.last_queued)
520                 diff = now - t->sched_info.last_queued;
521         sched_info_dequeued(t);
522         t->sched_info.run_delay += diff;
523         t->sched_info.last_arrival = now;
524         t->sched_info.pcnt++;
525
526         if (!rq)
527                 return;
528
529         rq->rq_sched_info.run_delay += diff;
530         rq->rq_sched_info.pcnt++;
531 }
532
533 /*
534  * Called when a process is queued into either the active or expired
535  * array.  The time is noted and later used to determine how long we
536  * had to wait for us to reach the cpu.  Since the expired queue will
537  * become the active queue after active queue is empty, without dequeuing
538  * and requeuing any tasks, we are interested in queuing to either. It
539  * is unusual but not impossible for tasks to be dequeued and immediately
540  * requeued in the same or another array: this can happen in sched_yield(),
541  * set_user_nice(), and even load_balance() as it moves tasks from runqueue
542  * to runqueue.
543  *
544  * This function is only called from enqueue_task(), but also only updates
545  * the timestamp if it is already not set.  It's assumed that
546  * sched_info_dequeued() will clear that stamp when appropriate.
547  */
548 static inline void sched_info_queued(task_t *t)
549 {
550         if (!t->sched_info.last_queued)
551                 t->sched_info.last_queued = jiffies;
552 }
553
554 /*
555  * Called when a process ceases being the active-running process, either
556  * voluntarily or involuntarily.  Now we can calculate how long we ran.
557  */
558 static inline void sched_info_depart(task_t *t)
559 {
560         struct runqueue *rq = task_rq(t);
561         unsigned long diff = jiffies - t->sched_info.last_arrival;
562
563         t->sched_info.cpu_time += diff;
564
565         if (rq)
566                 rq->rq_sched_info.cpu_time += diff;
567 }
568
569 /*
570  * Called when tasks are switched involuntarily due, typically, to expiring
571  * their time slice.  (This may also be called when switching to or from
572  * the idle task.)  We are only called when prev != next.
573  */
574 static inline void sched_info_switch(task_t *prev, task_t *next)
575 {
576         struct runqueue *rq = task_rq(prev);
577
578         /*
579          * prev now departs the cpu.  It's not interesting to record
580          * stats about how efficient we were at scheduling the idle
581          * process, however.
582          */
583         if (prev != rq->idle)
584                 sched_info_depart(prev);
585
586         if (next != rq->idle)
587                 sched_info_arrive(next);
588 }
589 #else
590 #define sched_info_queued(t)            do { } while (0)
591 #define sched_info_switch(t, next)      do { } while (0)
592 #endif /* CONFIG_SCHEDSTATS */
593
594 /*
595  * Adding/removing a task to/from a priority array:
596  */
597 static void dequeue_task(struct task_struct *p, prio_array_t *array)
598 {
599         array->nr_active--;
600         list_del(&p->run_list);
601         if (list_empty(array->queue + p->prio))
602                 __clear_bit(p->prio, array->bitmap);
603 }
604
605 static void enqueue_task(struct task_struct *p, prio_array_t *array)
606 {
607         sched_info_queued(p);
608         list_add_tail(&p->run_list, array->queue + p->prio);
609         __set_bit(p->prio, array->bitmap);
610         array->nr_active++;
611         p->array = array;
612 }
613
614 /*
615  * Put task to the end of the run list without the overhead of dequeue
616  * followed by enqueue.
617  */
618 static void requeue_task(struct task_struct *p, prio_array_t *array)
619 {
620         list_move_tail(&p->run_list, array->queue + p->prio);
621 }
622
623 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
624 {
625         list_add(&p->run_list, array->queue + p->prio);
626         __set_bit(p->prio, array->bitmap);
627         array->nr_active++;
628         p->array = array;
629 }
630
631 /*
632  * effective_prio - return the priority that is based on the static
633  * priority but is modified by bonuses/penalties.
634  *
635  * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
636  * into the -5 ... 0 ... +5 bonus/penalty range.
637  *
638  * We use 25% of the full 0...39 priority range so that:
639  *
640  * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
641  * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
642  *
643  * Both properties are important to certain workloads.
644  */
645 static int effective_prio(task_t *p)
646 {
647         int bonus, prio;
648
649         if (rt_task(p))
650                 return p->prio;
651
652         bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
653
654         prio = p->static_prio - bonus;
655         if (prio < MAX_RT_PRIO)
656                 prio = MAX_RT_PRIO;
657         if (prio > MAX_PRIO-1)
658                 prio = MAX_PRIO-1;
659         return prio;
660 }
661
662 /*
663  * __activate_task - move a task to the runqueue.
664  */
665 static inline void __activate_task(task_t *p, runqueue_t *rq)
666 {
667         enqueue_task(p, rq->active);
668         rq->nr_running++;
669 }
670
671 /*
672  * __activate_idle_task - move idle task to the _front_ of runqueue.
673  */
674 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
675 {
676         enqueue_task_head(p, rq->active);
677         rq->nr_running++;
678 }
679
680 static int recalc_task_prio(task_t *p, unsigned long long now)
681 {
682         /* Caller must always ensure 'now >= p->timestamp' */
683         unsigned long long __sleep_time = now - p->timestamp;
684         unsigned long sleep_time;
685
686         if (__sleep_time > NS_MAX_SLEEP_AVG)
687                 sleep_time = NS_MAX_SLEEP_AVG;
688         else
689                 sleep_time = (unsigned long)__sleep_time;
690
691         if (likely(sleep_time > 0)) {
692                 /*
693                  * User tasks that sleep a long time are categorised as
694                  * idle and will get just interactive status to stay active &
695                  * prevent them suddenly becoming cpu hogs and starving
696                  * other processes.
697                  */
698                 if (p->mm && p->activated != -1 &&
699                         sleep_time > INTERACTIVE_SLEEP(p)) {
700                                 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
701                                                 DEF_TIMESLICE);
702                 } else {
703                         /*
704                          * The lower the sleep avg a task has the more
705                          * rapidly it will rise with sleep time.
706                          */
707                         sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
708
709                         /*
710                          * Tasks waking from uninterruptible sleep are
711                          * limited in their sleep_avg rise as they
712                          * are likely to be waiting on I/O
713                          */
714                         if (p->activated == -1 && p->mm) {
715                                 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
716                                         sleep_time = 0;
717                                 else if (p->sleep_avg + sleep_time >=
718                                                 INTERACTIVE_SLEEP(p)) {
719                                         p->sleep_avg = INTERACTIVE_SLEEP(p);
720                                         sleep_time = 0;
721                                 }
722                         }
723
724                         /*
725                          * This code gives a bonus to interactive tasks.
726                          *
727                          * The boost works by updating the 'average sleep time'
728                          * value here, based on ->timestamp. The more time a
729                          * task spends sleeping, the higher the average gets -
730                          * and the higher the priority boost gets as well.
731                          */
732                         p->sleep_avg += sleep_time;
733
734                         if (p->sleep_avg > NS_MAX_SLEEP_AVG)
735                                 p->sleep_avg = NS_MAX_SLEEP_AVG;
736                 }
737         }
738
739         return effective_prio(p);
740 }
741
742 /*
743  * activate_task - move a task to the runqueue and do priority recalculation
744  *
745  * Update all the scheduling statistics stuff. (sleep average
746  * calculation, priority modifiers, etc.)
747  */
748 static void activate_task(task_t *p, runqueue_t *rq, int local)
749 {
750         unsigned long long now;
751
752         now = sched_clock();
753 #ifdef CONFIG_SMP
754         if (!local) {
755                 /* Compensate for drifting sched_clock */
756                 runqueue_t *this_rq = this_rq();
757                 now = (now - this_rq->timestamp_last_tick)
758                         + rq->timestamp_last_tick;
759         }
760 #endif
761
762         p->prio = recalc_task_prio(p, now);
763
764         /*
765          * This checks to make sure it's not an uninterruptible task
766          * that is now waking up.
767          */
768         if (!p->activated) {
769                 /*
770                  * Tasks which were woken up by interrupts (ie. hw events)
771                  * are most likely of interactive nature. So we give them
772                  * the credit of extending their sleep time to the period
773                  * of time they spend on the runqueue, waiting for execution
774                  * on a CPU, first time around:
775                  */
776                 if (in_interrupt())
777                         p->activated = 2;
778                 else {
779                         /*
780                          * Normal first-time wakeups get a credit too for
781                          * on-runqueue time, but it will be weighted down:
782                          */
783                         p->activated = 1;
784                 }
785         }
786         p->timestamp = now;
787
788         __activate_task(p, rq);
789 }
790
791 /*
792  * deactivate_task - remove a task from the runqueue.
793  */
794 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
795 {
796         rq->nr_running--;
797         dequeue_task(p, p->array);
798         p->array = NULL;
799 }
800
801 /*
802  * resched_task - mark a task 'to be rescheduled now'.
803  *
804  * On UP this means the setting of the need_resched flag, on SMP it
805  * might also involve a cross-CPU call to trigger the scheduler on
806  * the target CPU.
807  */
808 #ifdef CONFIG_SMP
809 static void resched_task(task_t *p)
810 {
811         int need_resched, nrpolling;
812
813         assert_spin_locked(&task_rq(p)->lock);
814
815         /* minimise the chance of sending an interrupt to poll_idle() */
816         nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
817         need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
818         nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
819
820         if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
821                 smp_send_reschedule(task_cpu(p));
822 }
823 #else
824 static inline void resched_task(task_t *p)
825 {
826         set_tsk_need_resched(p);
827 }
828 #endif
829
830 /**
831  * task_curr - is this task currently executing on a CPU?
832  * @p: the task in question.
833  */
834 inline int task_curr(const task_t *p)
835 {
836         return cpu_curr(task_cpu(p)) == p;
837 }
838
839 #ifdef CONFIG_SMP
840 typedef struct {
841         struct list_head list;
842
843         task_t *task;
844         int dest_cpu;
845
846         struct completion done;
847 } migration_req_t;
848
849 /*
850  * The task's runqueue lock must be held.
851  * Returns true if you have to wait for migration thread.
852  */
853 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
854 {
855         runqueue_t *rq = task_rq(p);
856
857         /*
858          * If the task is not on a runqueue (and not running), then
859          * it is sufficient to simply update the task's cpu field.
860          */
861         if (!p->array && !task_running(rq, p)) {
862                 set_task_cpu(p, dest_cpu);
863                 return 0;
864         }
865
866         init_completion(&req->done);
867         req->task = p;
868         req->dest_cpu = dest_cpu;
869         list_add(&req->list, &rq->migration_queue);
870         return 1;
871 }
872
873 /*
874  * wait_task_inactive - wait for a thread to unschedule.
875  *
876  * The caller must ensure that the task *will* unschedule sometime soon,
877  * else this function might spin for a *long* time. This function can't
878  * be called with interrupts off, or it may introduce deadlock with
879  * smp_call_function() if an IPI is sent by the same process we are
880  * waiting to become inactive.
881  */
882 void wait_task_inactive(task_t *p)
883 {
884         unsigned long flags;
885         runqueue_t *rq;
886         int preempted;
887
888 repeat:
889         rq = task_rq_lock(p, &flags);
890         /* Must be off runqueue entirely, not preempted. */
891         if (unlikely(p->array || task_running(rq, p))) {
892                 /* If it's preempted, we yield.  It could be a while. */
893                 preempted = !task_running(rq, p);
894                 task_rq_unlock(rq, &flags);
895                 cpu_relax();
896                 if (preempted)
897                         yield();
898                 goto repeat;
899         }
900         task_rq_unlock(rq, &flags);
901 }
902
903 /***
904  * kick_process - kick a running thread to enter/exit the kernel
905  * @p: the to-be-kicked thread
906  *
907  * Cause a process which is running on another CPU to enter
908  * kernel-mode, without any delay. (to get signals handled.)
909  *
910  * NOTE: this function doesnt have to take the runqueue lock,
911  * because all it wants to ensure is that the remote task enters
912  * the kernel. If the IPI races and the task has been migrated
913  * to another CPU then no harm is done and the purpose has been
914  * achieved as well.
915  */
916 void kick_process(task_t *p)
917 {
918         int cpu;
919
920         preempt_disable();
921         cpu = task_cpu(p);
922         if ((cpu != smp_processor_id()) && task_curr(p))
923                 smp_send_reschedule(cpu);
924         preempt_enable();
925 }
926
927 /*
928  * Return a low guess at the load of a migration-source cpu.
929  *
930  * We want to under-estimate the load of migration sources, to
931  * balance conservatively.
932  */
933 static inline unsigned long source_load(int cpu, int type)
934 {
935         runqueue_t *rq = cpu_rq(cpu);
936         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
937         if (type == 0)
938                 return load_now;
939
940         return min(rq->cpu_load[type-1], load_now);
941 }
942
943 /*
944  * Return a high guess at the load of a migration-target cpu
945  */
946 static inline unsigned long target_load(int cpu, int type)
947 {
948         runqueue_t *rq = cpu_rq(cpu);
949         unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
950         if (type == 0)
951                 return load_now;
952
953         return max(rq->cpu_load[type-1], load_now);
954 }
955
956 /*
957  * find_idlest_group finds and returns the least busy CPU group within the
958  * domain.
959  */
960 static struct sched_group *
961 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
962 {
963         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
964         unsigned long min_load = ULONG_MAX, this_load = 0;
965         int load_idx = sd->forkexec_idx;
966         int imbalance = 100 + (sd->imbalance_pct-100)/2;
967
968         do {
969                 unsigned long load, avg_load;
970                 int local_group;
971                 int i;
972
973                 /* Skip over this group if it has no CPUs allowed */
974                 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
975                         goto nextgroup;
976
977                 local_group = cpu_isset(this_cpu, group->cpumask);
978
979                 /* Tally up the load of all CPUs in the group */
980                 avg_load = 0;
981
982                 for_each_cpu_mask(i, group->cpumask) {
983                         /* Bias balancing toward cpus of our domain */
984                         if (local_group)
985                                 load = source_load(i, load_idx);
986                         else
987                                 load = target_load(i, load_idx);
988
989                         avg_load += load;
990                 }
991
992                 /* Adjust by relative CPU power of the group */
993                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
994
995                 if (local_group) {
996                         this_load = avg_load;
997                         this = group;
998                 } else if (avg_load < min_load) {
999                         min_load = avg_load;
1000                         idlest = group;
1001                 }
1002 nextgroup:
1003                 group = group->next;
1004         } while (group != sd->groups);
1005
1006         if (!idlest || 100*this_load < imbalance*min_load)
1007                 return NULL;
1008         return idlest;
1009 }
1010
1011 /*
1012  * find_idlest_queue - find the idlest runqueue among the cpus in group.
1013  */
1014 static int
1015 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1016 {
1017         cpumask_t tmp;
1018         unsigned long load, min_load = ULONG_MAX;
1019         int idlest = -1;
1020         int i;
1021
1022         /* Traverse only the allowed CPUs */
1023         cpus_and(tmp, group->cpumask, p->cpus_allowed);
1024
1025         for_each_cpu_mask(i, tmp) {
1026                 load = source_load(i, 0);
1027
1028                 if (load < min_load || (load == min_load && i == this_cpu)) {
1029                         min_load = load;
1030                         idlest = i;
1031                 }
1032         }
1033
1034         return idlest;
1035 }
1036
1037 /*
1038  * sched_balance_self: balance the current task (running on cpu) in domains
1039  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1040  * SD_BALANCE_EXEC.
1041  *
1042  * Balance, ie. select the least loaded group.
1043  *
1044  * Returns the target CPU number, or the same CPU if no balancing is needed.
1045  *
1046  * preempt must be disabled.
1047  */
1048 static int sched_balance_self(int cpu, int flag)
1049 {
1050         struct task_struct *t = current;
1051         struct sched_domain *tmp, *sd = NULL;
1052
1053         for_each_domain(cpu, tmp)
1054                 if (tmp->flags & flag)
1055                         sd = tmp;
1056
1057         while (sd) {
1058                 cpumask_t span;
1059                 struct sched_group *group;
1060                 int new_cpu;
1061                 int weight;
1062
1063                 span = sd->span;
1064                 group = find_idlest_group(sd, t, cpu);
1065                 if (!group)
1066                         goto nextlevel;
1067
1068                 new_cpu = find_idlest_cpu(group, t, cpu);
1069                 if (new_cpu == -1 || new_cpu == cpu)
1070                         goto nextlevel;
1071
1072                 /* Now try balancing at a lower domain level */
1073                 cpu = new_cpu;
1074 nextlevel:
1075                 sd = NULL;
1076                 weight = cpus_weight(span);
1077                 for_each_domain(cpu, tmp) {
1078                         if (weight <= cpus_weight(tmp->span))
1079                                 break;
1080                         if (tmp->flags & flag)
1081                                 sd = tmp;
1082                 }
1083                 /* while loop will break here if sd == NULL */
1084         }
1085
1086         return cpu;
1087 }
1088
1089 #endif /* CONFIG_SMP */
1090
1091 /*
1092  * wake_idle() will wake a task on an idle cpu if task->cpu is
1093  * not idle and an idle cpu is available.  The span of cpus to
1094  * search starts with cpus closest then further out as needed,
1095  * so we always favor a closer, idle cpu.
1096  *
1097  * Returns the CPU we should wake onto.
1098  */
1099 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1100 static int wake_idle(int cpu, task_t *p)
1101 {
1102         cpumask_t tmp;
1103         struct sched_domain *sd;
1104         int i;
1105
1106         if (idle_cpu(cpu))
1107                 return cpu;
1108
1109         for_each_domain(cpu, sd) {
1110                 if (sd->flags & SD_WAKE_IDLE) {
1111                         cpus_and(tmp, sd->span, p->cpus_allowed);
1112                         for_each_cpu_mask(i, tmp) {
1113                                 if (idle_cpu(i))
1114                                         return i;
1115                         }
1116                 }
1117                 else
1118                         break;
1119         }
1120         return cpu;
1121 }
1122 #else
1123 static inline int wake_idle(int cpu, task_t *p)
1124 {
1125         return cpu;
1126 }
1127 #endif
1128
1129 /***
1130  * try_to_wake_up - wake up a thread
1131  * @p: the to-be-woken-up thread
1132  * @state: the mask of task states that can be woken
1133  * @sync: do a synchronous wakeup?
1134  *
1135  * Put it on the run-queue if it's not already there. The "current"
1136  * thread is always on the run-queue (except when the actual
1137  * re-schedule is in progress), and as such you're allowed to do
1138  * the simpler "current->state = TASK_RUNNING" to mark yourself
1139  * runnable without the overhead of this.
1140  *
1141  * returns failure only if the task is already active.
1142  */
1143 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1144 {
1145         int cpu, this_cpu, success = 0;
1146         unsigned long flags;
1147         long old_state;
1148         runqueue_t *rq;
1149 #ifdef CONFIG_SMP
1150         unsigned long load, this_load;
1151         struct sched_domain *sd, *this_sd = NULL;
1152         int new_cpu;
1153 #endif
1154
1155         rq = task_rq_lock(p, &flags);
1156         old_state = p->state;
1157         if (!(old_state & state))
1158                 goto out;
1159
1160         if (p->array)
1161                 goto out_running;
1162
1163         cpu = task_cpu(p);
1164         this_cpu = smp_processor_id();
1165
1166 #ifdef CONFIG_SMP
1167         if (unlikely(task_running(rq, p)))
1168                 goto out_activate;
1169
1170         new_cpu = cpu;
1171
1172         schedstat_inc(rq, ttwu_cnt);
1173         if (cpu == this_cpu) {
1174                 schedstat_inc(rq, ttwu_local);
1175                 goto out_set_cpu;
1176         }
1177
1178         for_each_domain(this_cpu, sd) {
1179                 if (cpu_isset(cpu, sd->span)) {
1180                         schedstat_inc(sd, ttwu_wake_remote);
1181                         this_sd = sd;
1182                         break;
1183                 }
1184         }
1185
1186         if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1187                 goto out_set_cpu;
1188
1189         /*
1190          * Check for affine wakeup and passive balancing possibilities.
1191          */
1192         if (this_sd) {
1193                 int idx = this_sd->wake_idx;
1194                 unsigned int imbalance;
1195
1196                 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1197
1198                 load = source_load(cpu, idx);
1199                 this_load = target_load(this_cpu, idx);
1200
1201                 new_cpu = this_cpu; /* Wake to this CPU if we can */
1202
1203                 if (this_sd->flags & SD_WAKE_AFFINE) {
1204                         unsigned long tl = this_load;
1205                         /*
1206                          * If sync wakeup then subtract the (maximum possible)
1207                          * effect of the currently running task from the load
1208                          * of the current CPU:
1209                          */
1210                         if (sync)
1211                                 tl -= SCHED_LOAD_SCALE;
1212
1213                         if ((tl <= load &&
1214                                 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1215                                 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1216                                 /*
1217                                  * This domain has SD_WAKE_AFFINE and
1218                                  * p is cache cold in this domain, and
1219                                  * there is no bad imbalance.
1220                                  */
1221                                 schedstat_inc(this_sd, ttwu_move_affine);
1222                                 goto out_set_cpu;
1223                         }
1224                 }
1225
1226                 /*
1227                  * Start passive balancing when half the imbalance_pct
1228                  * limit is reached.
1229                  */
1230                 if (this_sd->flags & SD_WAKE_BALANCE) {
1231                         if (imbalance*this_load <= 100*load) {
1232                                 schedstat_inc(this_sd, ttwu_move_balance);
1233                                 goto out_set_cpu;
1234                         }
1235                 }
1236         }
1237
1238         new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1239 out_set_cpu:
1240         new_cpu = wake_idle(new_cpu, p);
1241         if (new_cpu != cpu) {
1242                 set_task_cpu(p, new_cpu);
1243                 task_rq_unlock(rq, &flags);
1244                 /* might preempt at this point */
1245                 rq = task_rq_lock(p, &flags);
1246                 old_state = p->state;
1247                 if (!(old_state & state))
1248                         goto out;
1249                 if (p->array)
1250                         goto out_running;
1251
1252                 this_cpu = smp_processor_id();
1253                 cpu = task_cpu(p);
1254         }
1255
1256 out_activate:
1257 #endif /* CONFIG_SMP */
1258         if (old_state == TASK_UNINTERRUPTIBLE) {
1259                 rq->nr_uninterruptible--;
1260                 /*
1261                  * Tasks on involuntary sleep don't earn
1262                  * sleep_avg beyond just interactive state.
1263                  */
1264                 p->activated = -1;
1265         }
1266
1267         /*
1268          * Tasks that have marked their sleep as noninteractive get
1269          * woken up without updating their sleep average. (i.e. their
1270          * sleep is handled in a priority-neutral manner, no priority
1271          * boost and no penalty.)
1272          */
1273         if (old_state & TASK_NONINTERACTIVE)
1274                 __activate_task(p, rq);
1275         else
1276                 activate_task(p, rq, cpu == this_cpu);
1277         /*
1278          * Sync wakeups (i.e. those types of wakeups where the waker
1279          * has indicated that it will leave the CPU in short order)
1280          * don't trigger a preemption, if the woken up task will run on
1281          * this cpu. (in this case the 'I will reschedule' promise of
1282          * the waker guarantees that the freshly woken up task is going
1283          * to be considered on this CPU.)
1284          */
1285         if (!sync || cpu != this_cpu) {
1286                 if (TASK_PREEMPTS_CURR(p, rq))
1287                         resched_task(rq->curr);
1288         }
1289         success = 1;
1290
1291 out_running:
1292         p->state = TASK_RUNNING;
1293 out:
1294         task_rq_unlock(rq, &flags);
1295
1296         return success;
1297 }
1298
1299 int fastcall wake_up_process(task_t *p)
1300 {
1301         return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1302                                  TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1303 }
1304
1305 EXPORT_SYMBOL(wake_up_process);
1306
1307 int fastcall wake_up_state(task_t *p, unsigned int state)
1308 {
1309         return try_to_wake_up(p, state, 0);
1310 }
1311
1312 /*
1313  * Perform scheduler related setup for a newly forked process p.
1314  * p is forked by current.
1315  */
1316 void fastcall sched_fork(task_t *p, int clone_flags)
1317 {
1318         int cpu = get_cpu();
1319
1320 #ifdef CONFIG_SMP
1321         cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1322 #endif
1323         set_task_cpu(p, cpu);
1324
1325         /*
1326          * We mark the process as running here, but have not actually
1327          * inserted it onto the runqueue yet. This guarantees that
1328          * nobody will actually run it, and a signal or other external
1329          * event cannot wake it up and insert it on the runqueue either.
1330          */
1331         p->state = TASK_RUNNING;
1332         INIT_LIST_HEAD(&p->run_list);
1333         p->array = NULL;
1334 #ifdef CONFIG_SCHEDSTATS
1335         memset(&p->sched_info, 0, sizeof(p->sched_info));
1336 #endif
1337 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1338         p->oncpu = 0;
1339 #endif
1340 #ifdef CONFIG_PREEMPT
1341         /* Want to start with kernel preemption disabled. */
1342         p->thread_info->preempt_count = 1;
1343 #endif
1344         /*
1345          * Share the timeslice between parent and child, thus the
1346          * total amount of pending timeslices in the system doesn't change,
1347          * resulting in more scheduling fairness.
1348          */
1349         local_irq_disable();
1350         p->time_slice = (current->time_slice + 1) >> 1;
1351         /*
1352          * The remainder of the first timeslice might be recovered by
1353          * the parent if the child exits early enough.
1354          */
1355         p->first_time_slice = 1;
1356         current->time_slice >>= 1;
1357         p->timestamp = sched_clock();
1358         if (unlikely(!current->time_slice)) {
1359                 /*
1360                  * This case is rare, it happens when the parent has only
1361                  * a single jiffy left from its timeslice. Taking the
1362                  * runqueue lock is not a problem.
1363                  */
1364                 current->time_slice = 1;
1365                 scheduler_tick();
1366         }
1367         local_irq_enable();
1368         put_cpu();
1369 }
1370
1371 /*
1372  * wake_up_new_task - wake up a newly created task for the first time.
1373  *
1374  * This function will do some initial scheduler statistics housekeeping
1375  * that must be done for every newly created context, then puts the task
1376  * on the runqueue and wakes it.
1377  */
1378 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1379 {
1380         unsigned long flags;
1381         int this_cpu, cpu;
1382         runqueue_t *rq, *this_rq;
1383
1384         rq = task_rq_lock(p, &flags);
1385         BUG_ON(p->state != TASK_RUNNING);
1386         this_cpu = smp_processor_id();
1387         cpu = task_cpu(p);
1388
1389         /*
1390          * We decrease the sleep average of forking parents
1391          * and children as well, to keep max-interactive tasks
1392          * from forking tasks that are max-interactive. The parent
1393          * (current) is done further down, under its lock.
1394          */
1395         p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1396                 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1397
1398         p->prio = effective_prio(p);
1399
1400         if (likely(cpu == this_cpu)) {
1401                 if (!(clone_flags & CLONE_VM)) {
1402                         /*
1403                          * The VM isn't cloned, so we're in a good position to
1404                          * do child-runs-first in anticipation of an exec. This
1405                          * usually avoids a lot of COW overhead.
1406                          */
1407                         if (unlikely(!current->array))
1408                                 __activate_task(p, rq);
1409                         else {
1410                                 p->prio = current->prio;
1411                                 list_add_tail(&p->run_list, &current->run_list);
1412                                 p->array = current->array;
1413                                 p->array->nr_active++;
1414                                 rq->nr_running++;
1415                         }
1416                         set_need_resched();
1417                 } else
1418                         /* Run child last */
1419                         __activate_task(p, rq);
1420                 /*
1421                  * We skip the following code due to cpu == this_cpu
1422                  *
1423                  *   task_rq_unlock(rq, &flags);
1424                  *   this_rq = task_rq_lock(current, &flags);
1425                  */
1426                 this_rq = rq;
1427         } else {
1428                 this_rq = cpu_rq(this_cpu);
1429
1430                 /*
1431                  * Not the local CPU - must adjust timestamp. This should
1432                  * get optimised away in the !CONFIG_SMP case.
1433                  */
1434                 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1435                                         + rq->timestamp_last_tick;
1436                 __activate_task(p, rq);
1437                 if (TASK_PREEMPTS_CURR(p, rq))
1438                         resched_task(rq->curr);
1439
1440                 /*
1441                  * Parent and child are on different CPUs, now get the
1442                  * parent runqueue to update the parent's ->sleep_avg:
1443                  */
1444                 task_rq_unlock(rq, &flags);
1445                 this_rq = task_rq_lock(current, &flags);
1446         }
1447         current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1448                 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1449         task_rq_unlock(this_rq, &flags);
1450 }
1451
1452 /*
1453  * Potentially available exiting-child timeslices are
1454  * retrieved here - this way the parent does not get
1455  * penalized for creating too many threads.
1456  *
1457  * (this cannot be used to 'generate' timeslices
1458  * artificially, because any timeslice recovered here
1459  * was given away by the parent in the first place.)
1460  */
1461 void fastcall sched_exit(task_t *p)
1462 {
1463         unsigned long flags;
1464         runqueue_t *rq;
1465
1466         /*
1467          * If the child was a (relative-) CPU hog then decrease
1468          * the sleep_avg of the parent as well.
1469          */
1470         rq = task_rq_lock(p->parent, &flags);
1471         if (p->first_time_slice) {
1472                 p->parent->time_slice += p->time_slice;
1473                 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1474                         p->parent->time_slice = task_timeslice(p);
1475         }
1476         if (p->sleep_avg < p->parent->sleep_avg)
1477                 p->parent->sleep_avg = p->parent->sleep_avg /
1478                 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1479                 (EXIT_WEIGHT + 1);
1480         task_rq_unlock(rq, &flags);
1481 }
1482
1483 /**
1484  * prepare_task_switch - prepare to switch tasks
1485  * @rq: the runqueue preparing to switch
1486  * @next: the task we are going to switch to.
1487  *
1488  * This is called with the rq lock held and interrupts off. It must
1489  * be paired with a subsequent finish_task_switch after the context
1490  * switch.
1491  *
1492  * prepare_task_switch sets up locking and calls architecture specific
1493  * hooks.
1494  */
1495 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1496 {
1497         prepare_lock_switch(rq, next);
1498         prepare_arch_switch(next);
1499 }
1500
1501 /**
1502  * finish_task_switch - clean up after a task-switch
1503  * @rq: runqueue associated with task-switch
1504  * @prev: the thread we just switched away from.
1505  *
1506  * finish_task_switch must be called after the context switch, paired
1507  * with a prepare_task_switch call before the context switch.
1508  * finish_task_switch will reconcile locking set up by prepare_task_switch,
1509  * and do any other architecture-specific cleanup actions.
1510  *
1511  * Note that we may have delayed dropping an mm in context_switch(). If
1512  * so, we finish that here outside of the runqueue lock.  (Doing it
1513  * with the lock held can cause deadlocks; see schedule() for
1514  * details.)
1515  */
1516 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1517         __releases(rq->lock)
1518 {
1519         struct mm_struct *mm = rq->prev_mm;
1520         unsigned long prev_task_flags;
1521
1522         rq->prev_mm = NULL;
1523
1524         /*
1525          * A task struct has one reference for the use as "current".
1526          * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1527          * calls schedule one last time. The schedule call will never return,
1528          * and the scheduled task must drop that reference.
1529          * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1530          * still held, otherwise prev could be scheduled on another cpu, die
1531          * there before we look at prev->state, and then the reference would
1532          * be dropped twice.
1533          *              Manfred Spraul <manfred@colorfullife.com>
1534          */
1535         prev_task_flags = prev->flags;
1536         finish_arch_switch(prev);
1537         finish_lock_switch(rq, prev);
1538         if (mm)
1539                 mmdrop(mm);
1540         if (unlikely(prev_task_flags & PF_DEAD))
1541                 put_task_struct(prev);
1542 }
1543
1544 /**
1545  * schedule_tail - first thing a freshly forked thread must call.
1546  * @prev: the thread we just switched away from.
1547  */
1548 asmlinkage void schedule_tail(task_t *prev)
1549         __releases(rq->lock)
1550 {
1551         runqueue_t *rq = this_rq();
1552         finish_task_switch(rq, prev);
1553 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1554         /* In this case, finish_task_switch does not reenable preemption */
1555         preempt_enable();
1556 #endif
1557         if (current->set_child_tid)
1558                 put_user(current->pid, current->set_child_tid);
1559 }
1560
1561 /*
1562  * context_switch - switch to the new MM and the new
1563  * thread's register state.
1564  */
1565 static inline
1566 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1567 {
1568         struct mm_struct *mm = next->mm;
1569         struct mm_struct *oldmm = prev->active_mm;
1570
1571         if (unlikely(!mm)) {
1572                 next->active_mm = oldmm;
1573                 atomic_inc(&oldmm->mm_count);
1574                 enter_lazy_tlb(oldmm, next);
1575         } else
1576                 switch_mm(oldmm, mm, next);
1577
1578         if (unlikely(!prev->mm)) {
1579                 prev->active_mm = NULL;
1580                 WARN_ON(rq->prev_mm);
1581                 rq->prev_mm = oldmm;
1582         }
1583
1584         /* Here we just switch the register state and the stack. */
1585         switch_to(prev, next, prev);
1586
1587         return prev;
1588 }
1589
1590 /*
1591  * nr_running, nr_uninterruptible and nr_context_switches:
1592  *
1593  * externally visible scheduler statistics: current number of runnable
1594  * threads, current number of uninterruptible-sleeping threads, total
1595  * number of context switches performed since bootup.
1596  */
1597 unsigned long nr_running(void)
1598 {
1599         unsigned long i, sum = 0;
1600
1601         for_each_online_cpu(i)
1602                 sum += cpu_rq(i)->nr_running;
1603
1604         return sum;
1605 }
1606
1607 unsigned long nr_uninterruptible(void)
1608 {
1609         unsigned long i, sum = 0;
1610
1611         for_each_cpu(i)
1612                 sum += cpu_rq(i)->nr_uninterruptible;
1613
1614         /*
1615          * Since we read the counters lockless, it might be slightly
1616          * inaccurate. Do not allow it to go below zero though:
1617          */
1618         if (unlikely((long)sum < 0))
1619                 sum = 0;
1620
1621         return sum;
1622 }
1623
1624 unsigned long long nr_context_switches(void)
1625 {
1626         unsigned long long i, sum = 0;
1627
1628         for_each_cpu(i)
1629                 sum += cpu_rq(i)->nr_switches;
1630
1631         return sum;
1632 }
1633
1634 unsigned long nr_iowait(void)
1635 {
1636         unsigned long i, sum = 0;
1637
1638         for_each_cpu(i)
1639                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1640
1641         return sum;
1642 }
1643
1644 #ifdef CONFIG_SMP
1645
1646 /*
1647  * double_rq_lock - safely lock two runqueues
1648  *
1649  * Note this does not disable interrupts like task_rq_lock,
1650  * you need to do so manually before calling.
1651  */
1652 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1653         __acquires(rq1->lock)
1654         __acquires(rq2->lock)
1655 {
1656         if (rq1 == rq2) {
1657                 spin_lock(&rq1->lock);
1658                 __acquire(rq2->lock);   /* Fake it out ;) */
1659         } else {
1660                 if (rq1 < rq2) {
1661                         spin_lock(&rq1->lock);
1662                         spin_lock(&rq2->lock);
1663                 } else {
1664                         spin_lock(&rq2->lock);
1665                         spin_lock(&rq1->lock);
1666                 }
1667         }
1668 }
1669
1670 /*
1671  * double_rq_unlock - safely unlock two runqueues
1672  *
1673  * Note this does not restore interrupts like task_rq_unlock,
1674  * you need to do so manually after calling.
1675  */
1676 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1677         __releases(rq1->lock)
1678         __releases(rq2->lock)
1679 {
1680         spin_unlock(&rq1->lock);
1681         if (rq1 != rq2)
1682                 spin_unlock(&rq2->lock);
1683         else
1684                 __release(rq2->lock);
1685 }
1686
1687 /*
1688  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1689  */
1690 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1691         __releases(this_rq->lock)
1692         __acquires(busiest->lock)
1693         __acquires(this_rq->lock)
1694 {
1695         if (unlikely(!spin_trylock(&busiest->lock))) {
1696                 if (busiest < this_rq) {
1697                         spin_unlock(&this_rq->lock);
1698                         spin_lock(&busiest->lock);
1699                         spin_lock(&this_rq->lock);
1700                 } else
1701                         spin_lock(&busiest->lock);
1702         }
1703 }
1704
1705 /*
1706  * If dest_cpu is allowed for this process, migrate the task to it.
1707  * This is accomplished by forcing the cpu_allowed mask to only
1708  * allow dest_cpu, which will force the cpu onto dest_cpu.  Then
1709  * the cpu_allowed mask is restored.
1710  */
1711 static void sched_migrate_task(task_t *p, int dest_cpu)
1712 {
1713         migration_req_t req;
1714         runqueue_t *rq;
1715         unsigned long flags;
1716
1717         rq = task_rq_lock(p, &flags);
1718         if (!cpu_isset(dest_cpu, p->cpus_allowed)
1719             || unlikely(cpu_is_offline(dest_cpu)))
1720                 goto out;
1721
1722         /* force the process onto the specified CPU */
1723         if (migrate_task(p, dest_cpu, &req)) {
1724                 /* Need to wait for migration thread (might exit: take ref). */
1725                 struct task_struct *mt = rq->migration_thread;
1726                 get_task_struct(mt);
1727                 task_rq_unlock(rq, &flags);
1728                 wake_up_process(mt);
1729                 put_task_struct(mt);
1730                 wait_for_completion(&req.done);
1731                 return;
1732         }
1733 out:
1734         task_rq_unlock(rq, &flags);
1735 }
1736
1737 /*
1738  * sched_exec - execve() is a valuable balancing opportunity, because at
1739  * this point the task has the smallest effective memory and cache footprint.
1740  */
1741 void sched_exec(void)
1742 {
1743         int new_cpu, this_cpu = get_cpu();
1744         new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1745         put_cpu();
1746         if (new_cpu != this_cpu)
1747                 sched_migrate_task(current, new_cpu);
1748 }
1749
1750 /*
1751  * pull_task - move a task from a remote runqueue to the local runqueue.
1752  * Both runqueues must be locked.
1753  */
1754 static inline
1755 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1756                runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1757 {
1758         dequeue_task(p, src_array);
1759         src_rq->nr_running--;
1760         set_task_cpu(p, this_cpu);
1761         this_rq->nr_running++;
1762         enqueue_task(p, this_array);
1763         p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1764                                 + this_rq->timestamp_last_tick;
1765         /*
1766          * Note that idle threads have a prio of MAX_PRIO, for this test
1767          * to be always true for them.
1768          */
1769         if (TASK_PREEMPTS_CURR(p, this_rq))
1770                 resched_task(this_rq->curr);
1771 }
1772
1773 /*
1774  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1775  */
1776 static inline
1777 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1778                      struct sched_domain *sd, enum idle_type idle,
1779                      int *all_pinned)
1780 {
1781         /*
1782          * We do not migrate tasks that are:
1783          * 1) running (obviously), or
1784          * 2) cannot be migrated to this CPU due to cpus_allowed, or
1785          * 3) are cache-hot on their current CPU.
1786          */
1787         if (!cpu_isset(this_cpu, p->cpus_allowed))
1788                 return 0;
1789         *all_pinned = 0;
1790
1791         if (task_running(rq, p))
1792                 return 0;
1793
1794         /*
1795          * Aggressive migration if:
1796          * 1) task is cache cold, or
1797          * 2) too many balance attempts have failed.
1798          */
1799
1800         if (sd->nr_balance_failed > sd->cache_nice_tries)
1801                 return 1;
1802
1803         if (task_hot(p, rq->timestamp_last_tick, sd))
1804                 return 0;
1805         return 1;
1806 }
1807
1808 /*
1809  * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1810  * as part of a balancing operation within "domain". Returns the number of
1811  * tasks moved.
1812  *
1813  * Called with both runqueues locked.
1814  */
1815 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1816                       unsigned long max_nr_move, struct sched_domain *sd,
1817                       enum idle_type idle, int *all_pinned)
1818 {
1819         prio_array_t *array, *dst_array;
1820         struct list_head *head, *curr;
1821         int idx, pulled = 0, pinned = 0;
1822         task_t *tmp;
1823
1824         if (max_nr_move == 0)
1825                 goto out;
1826
1827         pinned = 1;
1828
1829         /*
1830          * We first consider expired tasks. Those will likely not be
1831          * executed in the near future, and they are most likely to
1832          * be cache-cold, thus switching CPUs has the least effect
1833          * on them.
1834          */
1835         if (busiest->expired->nr_active) {
1836                 array = busiest->expired;
1837                 dst_array = this_rq->expired;
1838         } else {
1839                 array = busiest->active;
1840                 dst_array = this_rq->active;
1841         }
1842
1843 new_array:
1844         /* Start searching at priority 0: */
1845         idx = 0;
1846 skip_bitmap:
1847         if (!idx)
1848                 idx = sched_find_first_bit(array->bitmap);
1849         else
1850                 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1851         if (idx >= MAX_PRIO) {
1852                 if (array == busiest->expired && busiest->active->nr_active) {
1853                         array = busiest->active;
1854                         dst_array = this_rq->active;
1855                         goto new_array;
1856                 }
1857                 goto out;
1858         }
1859
1860         head = array->queue + idx;
1861         curr = head->prev;
1862 skip_queue:
1863         tmp = list_entry(curr, task_t, run_list);
1864
1865         curr = curr->prev;
1866
1867         if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1868                 if (curr != head)
1869                         goto skip_queue;
1870                 idx++;
1871                 goto skip_bitmap;
1872         }
1873
1874 #ifdef CONFIG_SCHEDSTATS
1875         if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1876                 schedstat_inc(sd, lb_hot_gained[idle]);
1877 #endif
1878
1879         pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1880         pulled++;
1881
1882         /* We only want to steal up to the prescribed number of tasks. */
1883         if (pulled < max_nr_move) {
1884                 if (curr != head)
1885                         goto skip_queue;
1886                 idx++;
1887                 goto skip_bitmap;
1888         }
1889 out:
1890         /*
1891          * Right now, this is the only place pull_task() is called,
1892          * so we can safely collect pull_task() stats here rather than
1893          * inside pull_task().
1894          */
1895         schedstat_add(sd, lb_gained[idle], pulled);
1896
1897         if (all_pinned)
1898                 *all_pinned = pinned;
1899         return pulled;
1900 }
1901
1902 /*
1903  * find_busiest_group finds and returns the busiest CPU group within the
1904  * domain. It calculates and returns the number of tasks which should be
1905  * moved to restore balance via the imbalance parameter.
1906  */
1907 static struct sched_group *
1908 find_busiest_group(struct sched_domain *sd, int this_cpu,
1909                    unsigned long *imbalance, enum idle_type idle, int *sd_idle)
1910 {
1911         struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1912         unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1913         unsigned long max_pull;
1914         int load_idx;
1915
1916         max_load = this_load = total_load = total_pwr = 0;
1917         if (idle == NOT_IDLE)
1918                 load_idx = sd->busy_idx;
1919         else if (idle == NEWLY_IDLE)
1920                 load_idx = sd->newidle_idx;
1921         else
1922                 load_idx = sd->idle_idx;
1923
1924         do {
1925                 unsigned long load;
1926                 int local_group;
1927                 int i;
1928
1929                 local_group = cpu_isset(this_cpu, group->cpumask);
1930
1931                 /* Tally up the load of all CPUs in the group */
1932                 avg_load = 0;
1933
1934                 for_each_cpu_mask(i, group->cpumask) {
1935                         if (*sd_idle && !idle_cpu(i))
1936                                 *sd_idle = 0;
1937
1938                         /* Bias balancing toward cpus of our domain */
1939                         if (local_group)
1940                                 load = target_load(i, load_idx);
1941                         else
1942                                 load = source_load(i, load_idx);
1943
1944                         avg_load += load;
1945                 }
1946
1947                 total_load += avg_load;
1948                 total_pwr += group->cpu_power;
1949
1950                 /* Adjust by relative CPU power of the group */
1951                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1952
1953                 if (local_group) {
1954                         this_load = avg_load;
1955                         this = group;
1956                 } else if (avg_load > max_load) {
1957                         max_load = avg_load;
1958                         busiest = group;
1959                 }
1960                 group = group->next;
1961         } while (group != sd->groups);
1962
1963         if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
1964                 goto out_balanced;
1965
1966         avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1967
1968         if (this_load >= avg_load ||
1969                         100*max_load <= sd->imbalance_pct*this_load)
1970                 goto out_balanced;
1971
1972         /*
1973          * We're trying to get all the cpus to the average_load, so we don't
1974          * want to push ourselves above the average load, nor do we wish to
1975          * reduce the max loaded cpu below the average load, as either of these
1976          * actions would just result in more rebalancing later, and ping-pong
1977          * tasks around. Thus we look for the minimum possible imbalance.
1978          * Negative imbalances (*we* are more loaded than anyone else) will
1979          * be counted as no imbalance for these purposes -- we can't fix that
1980          * by pulling tasks to us.  Be careful of negative numbers as they'll
1981          * appear as very large values with unsigned longs.
1982          */
1983
1984         /* Don't want to pull so many tasks that a group would go idle */
1985         max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
1986
1987         /* How much load to actually move to equalise the imbalance */
1988         *imbalance = min(max_pull * busiest->cpu_power,
1989                                 (avg_load - this_load) * this->cpu_power)
1990                         / SCHED_LOAD_SCALE;
1991
1992         if (*imbalance < SCHED_LOAD_SCALE) {
1993                 unsigned long pwr_now = 0, pwr_move = 0;
1994                 unsigned long tmp;
1995
1996                 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1997                         *imbalance = 1;
1998                         return busiest;
1999                 }
2000
2001                 /*
2002                  * OK, we don't have enough imbalance to justify moving tasks,
2003                  * however we may be able to increase total CPU power used by
2004                  * moving them.
2005                  */
2006
2007                 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2008                 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2009                 pwr_now /= SCHED_LOAD_SCALE;
2010
2011                 /* Amount of load we'd subtract */
2012                 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2013                 if (max_load > tmp)
2014                         pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2015                                                         max_load - tmp);
2016
2017                 /* Amount of load we'd add */
2018                 if (max_load*busiest->cpu_power <
2019                                 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2020                         tmp = max_load*busiest->cpu_power/this->cpu_power;
2021                 else
2022                         tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2023                 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2024                 pwr_move /= SCHED_LOAD_SCALE;
2025
2026                 /* Move if we gain throughput */
2027                 if (pwr_move <= pwr_now)
2028                         goto out_balanced;
2029
2030                 *imbalance = 1;
2031                 return busiest;
2032         }
2033
2034         /* Get rid of the scaling factor, rounding down as we divide */
2035         *imbalance = *imbalance / SCHED_LOAD_SCALE;
2036         return busiest;
2037
2038 out_balanced:
2039
2040         *imbalance = 0;
2041         return NULL;
2042 }
2043
2044 /*
2045  * find_busiest_queue - find the busiest runqueue among the cpus in group.
2046  */
2047 static runqueue_t *find_busiest_queue(struct sched_group *group)
2048 {
2049         unsigned long load, max_load = 0;
2050         runqueue_t *busiest = NULL;
2051         int i;
2052
2053         for_each_cpu_mask(i, group->cpumask) {
2054                 load = source_load(i, 0);
2055
2056                 if (load > max_load) {
2057                         max_load = load;
2058                         busiest = cpu_rq(i);
2059                 }
2060         }
2061
2062         return busiest;
2063 }
2064
2065 /*
2066  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2067  * so long as it is large enough.
2068  */
2069 #define MAX_PINNED_INTERVAL     512
2070
2071 /*
2072  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2073  * tasks if there is an imbalance.
2074  *
2075  * Called with this_rq unlocked.
2076  */
2077 static int load_balance(int this_cpu, runqueue_t *this_rq,
2078                         struct sched_domain *sd, enum idle_type idle)
2079 {
2080         struct sched_group *group;
2081         runqueue_t *busiest;
2082         unsigned long imbalance;
2083         int nr_moved, all_pinned = 0;
2084         int active_balance = 0;
2085         int sd_idle = 0;
2086
2087         if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2088                 sd_idle = 1;
2089
2090         schedstat_inc(sd, lb_cnt[idle]);
2091
2092         group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2093         if (!group) {
2094                 schedstat_inc(sd, lb_nobusyg[idle]);
2095                 goto out_balanced;
2096         }
2097
2098         busiest = find_busiest_queue(group);
2099         if (!busiest) {
2100                 schedstat_inc(sd, lb_nobusyq[idle]);
2101                 goto out_balanced;
2102         }
2103
2104         BUG_ON(busiest == this_rq);
2105
2106         schedstat_add(sd, lb_imbalance[idle], imbalance);
2107
2108         nr_moved = 0;
2109         if (busiest->nr_running > 1) {
2110                 /*
2111                  * Attempt to move tasks. If find_busiest_group has found
2112                  * an imbalance but busiest->nr_running <= 1, the group is
2113                  * still unbalanced. nr_moved simply stays zero, so it is
2114                  * correctly treated as an imbalance.
2115                  */
2116                 double_rq_lock(this_rq, busiest);
2117                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2118                                         imbalance, sd, idle, &all_pinned);
2119                 double_rq_unlock(this_rq, busiest);
2120
2121                 /* All tasks on this runqueue were pinned by CPU affinity */
2122                 if (unlikely(all_pinned))
2123                         goto out_balanced;
2124         }
2125
2126         if (!nr_moved) {
2127                 schedstat_inc(sd, lb_failed[idle]);
2128                 sd->nr_balance_failed++;
2129
2130                 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2131
2132                         spin_lock(&busiest->lock);
2133
2134                         /* don't kick the migration_thread, if the curr
2135                          * task on busiest cpu can't be moved to this_cpu
2136                          */
2137                         if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2138                                 spin_unlock(&busiest->lock);
2139                                 all_pinned = 1;
2140                                 goto out_one_pinned;
2141                         }
2142
2143                         if (!busiest->active_balance) {
2144                                 busiest->active_balance = 1;
2145                                 busiest->push_cpu = this_cpu;
2146                                 active_balance = 1;
2147                         }
2148                         spin_unlock(&busiest->lock);
2149                         if (active_balance)
2150                                 wake_up_process(busiest->migration_thread);
2151
2152                         /*
2153                          * We've kicked active balancing, reset the failure
2154                          * counter.
2155                          */
2156                         sd->nr_balance_failed = sd->cache_nice_tries+1;
2157                 }
2158         } else
2159                 sd->nr_balance_failed = 0;
2160
2161         if (likely(!active_balance)) {
2162                 /* We were unbalanced, so reset the balancing interval */
2163                 sd->balance_interval = sd->min_interval;
2164         } else {
2165                 /*
2166                  * If we've begun active balancing, start to back off. This
2167                  * case may not be covered by the all_pinned logic if there
2168                  * is only 1 task on the busy runqueue (because we don't call
2169                  * move_tasks).
2170                  */
2171                 if (sd->balance_interval < sd->max_interval)
2172                         sd->balance_interval *= 2;
2173         }
2174
2175         if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2176                 return -1;
2177         return nr_moved;
2178
2179 out_balanced:
2180         schedstat_inc(sd, lb_balanced[idle]);
2181
2182         sd->nr_balance_failed = 0;
2183
2184 out_one_pinned:
2185         /* tune up the balancing interval */
2186         if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2187                         (sd->balance_interval < sd->max_interval))
2188                 sd->balance_interval *= 2;
2189
2190         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2191                 return -1;
2192         return 0;
2193 }
2194
2195 /*
2196  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2197  * tasks if there is an imbalance.
2198  *
2199  * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2200  * this_rq is locked.
2201  */
2202 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2203                                 struct sched_domain *sd)
2204 {
2205         struct sched_group *group;
2206         runqueue_t *busiest = NULL;
2207         unsigned long imbalance;
2208         int nr_moved = 0;
2209         int sd_idle = 0;
2210
2211         if (sd->flags & SD_SHARE_CPUPOWER)
2212                 sd_idle = 1;
2213
2214         schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2215         group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2216         if (!group) {
2217                 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2218                 goto out_balanced;
2219         }
2220
2221         busiest = find_busiest_queue(group);
2222         if (!busiest) {
2223                 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2224                 goto out_balanced;
2225         }
2226
2227         BUG_ON(busiest == this_rq);
2228
2229         schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2230
2231         nr_moved = 0;
2232         if (busiest->nr_running > 1) {
2233                 /* Attempt to move tasks */
2234                 double_lock_balance(this_rq, busiest);
2235                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2236                                         imbalance, sd, NEWLY_IDLE, NULL);
2237                 spin_unlock(&busiest->lock);
2238         }
2239
2240         if (!nr_moved) {
2241                 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2242                 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2243                         return -1;
2244         } else
2245                 sd->nr_balance_failed = 0;
2246
2247         return nr_moved;
2248
2249 out_balanced:
2250         schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2251         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2252                 return -1;
2253         sd->nr_balance_failed = 0;
2254         return 0;
2255 }
2256
2257 /*
2258  * idle_balance is called by schedule() if this_cpu is about to become
2259  * idle. Attempts to pull tasks from other CPUs.
2260  */
2261 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2262 {
2263         struct sched_domain *sd;
2264
2265         for_each_domain(this_cpu, sd) {
2266                 if (sd->flags & SD_BALANCE_NEWIDLE) {
2267                         if (load_balance_newidle(this_cpu, this_rq, sd)) {
2268                                 /* We've pulled tasks over so stop searching */
2269                                 break;
2270                         }
2271                 }
2272         }
2273 }
2274
2275 /*
2276  * active_load_balance is run by migration threads. It pushes running tasks
2277  * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2278  * running on each physical CPU where possible, and avoids physical /
2279  * logical imbalances.
2280  *
2281  * Called with busiest_rq locked.
2282  */
2283 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2284 {
2285         struct sched_domain *sd;
2286         runqueue_t *target_rq;
2287         int target_cpu = busiest_rq->push_cpu;
2288
2289         if (busiest_rq->nr_running <= 1)
2290                 /* no task to move */
2291                 return;
2292
2293         target_rq = cpu_rq(target_cpu);
2294
2295         /*
2296          * This condition is "impossible", if it occurs
2297          * we need to fix it.  Originally reported by
2298          * Bjorn Helgaas on a 128-cpu setup.
2299          */
2300         BUG_ON(busiest_rq == target_rq);
2301
2302         /* move a task from busiest_rq to target_rq */
2303         double_lock_balance(busiest_rq, target_rq);
2304
2305         /* Search for an sd spanning us and the target CPU. */
2306         for_each_domain(target_cpu, sd)
2307                 if ((sd->flags & SD_LOAD_BALANCE) &&
2308                         cpu_isset(busiest_cpu, sd->span))
2309                                 break;
2310
2311         if (unlikely(sd == NULL))
2312                 goto out;
2313
2314         schedstat_inc(sd, alb_cnt);
2315
2316         if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2317                 schedstat_inc(sd, alb_pushed);
2318         else
2319                 schedstat_inc(sd, alb_failed);
2320 out:
2321         spin_unlock(&target_rq->lock);
2322 }
2323
2324 /*
2325  * rebalance_tick will get called every timer tick, on every CPU.
2326  *
2327  * It checks each scheduling domain to see if it is due to be balanced,
2328  * and initiates a balancing operation if so.
2329  *
2330  * Balancing parameters are set up in arch_init_sched_domains.
2331  */
2332
2333 /* Don't have all balancing operations going off at once */
2334 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2335
2336 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2337                            enum idle_type idle)
2338 {
2339         unsigned long old_load, this_load;
2340         unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2341         struct sched_domain *sd;
2342         int i;
2343
2344         this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2345         /* Update our load */
2346         for (i = 0; i < 3; i++) {
2347                 unsigned long new_load = this_load;
2348                 int scale = 1 << i;
2349                 old_load = this_rq->cpu_load[i];
2350                 /*
2351                  * Round up the averaging division if load is increasing. This
2352                  * prevents us from getting stuck on 9 if the load is 10, for
2353                  * example.
2354                  */
2355                 if (new_load > old_load)
2356                         new_load += scale-1;
2357                 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2358         }
2359
2360         for_each_domain(this_cpu, sd) {
2361                 unsigned long interval;
2362
2363                 if (!(sd->flags & SD_LOAD_BALANCE))
2364                         continue;
2365
2366                 interval = sd->balance_interval;
2367                 if (idle != SCHED_IDLE)
2368                         interval *= sd->busy_factor;
2369
2370                 /* scale ms to jiffies */
2371                 interval = msecs_to_jiffies(interval);
2372                 if (unlikely(!interval))
2373                         interval = 1;
2374
2375                 if (j - sd->last_balance >= interval) {
2376                         if (load_balance(this_cpu, this_rq, sd, idle)) {
2377                                 /*
2378                                  * We've pulled tasks over so either we're no
2379                                  * longer idle, or one of our SMT siblings is
2380                                  * not idle.
2381                                  */
2382                                 idle = NOT_IDLE;
2383                         }
2384                         sd->last_balance += interval;
2385                 }
2386         }
2387 }
2388 #else
2389 /*
2390  * on UP we do not need to balance between CPUs:
2391  */
2392 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2393 {
2394 }
2395 static inline void idle_balance(int cpu, runqueue_t *rq)
2396 {
2397 }
2398 #endif
2399
2400 static inline int wake_priority_sleeper(runqueue_t *rq)
2401 {
2402         int ret = 0;
2403 #ifdef CONFIG_SCHED_SMT
2404         spin_lock(&rq->lock);
2405         /*
2406          * If an SMT sibling task has been put to sleep for priority
2407          * reasons reschedule the idle task to see if it can now run.
2408          */
2409         if (rq->nr_running) {
2410                 resched_task(rq->idle);
2411                 ret = 1;
2412         }
2413         spin_unlock(&rq->lock);
2414 #endif
2415         return ret;
2416 }
2417
2418 DEFINE_PER_CPU(struct kernel_stat, kstat);
2419
2420 EXPORT_PER_CPU_SYMBOL(kstat);
2421
2422 /*
2423  * This is called on clock ticks and on context switches.
2424  * Bank in p->sched_time the ns elapsed since the last tick or switch.
2425  */
2426 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2427                                     unsigned long long now)
2428 {
2429         unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2430         p->sched_time += now - last;
2431 }
2432
2433 /*
2434  * Return current->sched_time plus any more ns on the sched_clock
2435  * that have not yet been banked.
2436  */
2437 unsigned long long current_sched_time(const task_t *tsk)
2438 {
2439         unsigned long long ns;
2440         unsigned long flags;
2441         local_irq_save(flags);
2442         ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2443         ns = tsk->sched_time + (sched_clock() - ns);
2444         local_irq_restore(flags);
2445         return ns;
2446 }
2447
2448 /*
2449  * We place interactive tasks back into the active array, if possible.
2450  *
2451  * To guarantee that this does not starve expired tasks we ignore the
2452  * interactivity of a task if the first expired task had to wait more
2453  * than a 'reasonable' amount of time. This deadline timeout is
2454  * load-dependent, as the frequency of array switched decreases with
2455  * increasing number of running tasks. We also ignore the interactivity
2456  * if a better static_prio task has expired:
2457  */
2458 #define EXPIRED_STARVING(rq) \
2459         ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2460                 (jiffies - (rq)->expired_timestamp >= \
2461                         STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2462                         ((rq)->curr->static_prio > (rq)->best_expired_prio))
2463
2464 /*
2465  * Account user cpu time to a process.
2466  * @p: the process that the cpu time gets accounted to
2467  * @hardirq_offset: the offset to subtract from hardirq_count()
2468  * @cputime: the cpu time spent in user space since the last update
2469  */
2470 void account_user_time(struct task_struct *p, cputime_t cputime)
2471 {
2472         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2473         cputime64_t tmp;
2474
2475         p->utime = cputime_add(p->utime, cputime);
2476
2477         /* Add user time to cpustat. */
2478         tmp = cputime_to_cputime64(cputime);
2479         if (TASK_NICE(p) > 0)
2480                 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2481         else
2482                 cpustat->user = cputime64_add(cpustat->user, tmp);
2483 }
2484
2485 /*
2486  * Account system cpu time to a process.
2487  * @p: the process that the cpu time gets accounted to
2488  * @hardirq_offset: the offset to subtract from hardirq_count()
2489  * @cputime: the cpu time spent in kernel space since the last update
2490  */
2491 void account_system_time(struct task_struct *p, int hardirq_offset,
2492                          cputime_t cputime)
2493 {
2494         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2495         runqueue_t *rq = this_rq();
2496         cputime64_t tmp;
2497
2498         p->stime = cputime_add(p->stime, cputime);
2499
2500         /* Add system time to cpustat. */
2501         tmp = cputime_to_cputime64(cputime);
2502         if (hardirq_count() - hardirq_offset)
2503                 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2504         else if (softirq_count())
2505                 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2506         else if (p != rq->idle)
2507                 cpustat->system = cputime64_add(cpustat->system, tmp);
2508         else if (atomic_read(&rq->nr_iowait) > 0)
2509                 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2510         else
2511                 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2512         /* Account for system time used */
2513         acct_update_integrals(p);
2514         /* Update rss highwater mark */
2515         update_mem_hiwater(p);
2516 }
2517
2518 /*
2519  * Account for involuntary wait time.
2520  * @p: the process from which the cpu time has been stolen
2521  * @steal: the cpu time spent in involuntary wait
2522  */
2523 void account_steal_time(struct task_struct *p, cputime_t steal)
2524 {
2525         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2526         cputime64_t tmp = cputime_to_cputime64(steal);
2527         runqueue_t *rq = this_rq();
2528
2529         if (p == rq->idle) {
2530                 p->stime = cputime_add(p->stime, steal);
2531                 if (atomic_read(&rq->nr_iowait) > 0)
2532                         cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2533                 else
2534                         cpustat->idle = cputime64_add(cpustat->idle, tmp);
2535         } else
2536                 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2537 }
2538
2539 /*
2540  * This function gets called by the timer code, with HZ frequency.
2541  * We call it with interrupts disabled.
2542  *
2543  * It also gets called by the fork code, when changing the parent's
2544  * timeslices.
2545  */
2546 void scheduler_tick(void)
2547 {
2548         int cpu = smp_processor_id();
2549         runqueue_t *rq = this_rq();
2550         task_t *p = current;
2551         unsigned long long now = sched_clock();
2552
2553         update_cpu_clock(p, rq, now);
2554
2555         rq->timestamp_last_tick = now;
2556
2557         if (p == rq->idle) {
2558                 if (wake_priority_sleeper(rq))
2559                         goto out;
2560                 rebalance_tick(cpu, rq, SCHED_IDLE);
2561                 return;
2562         }
2563
2564         /* Task might have expired already, but not scheduled off yet */
2565         if (p->array != rq->active) {
2566                 set_tsk_need_resched(p);
2567                 goto out;
2568         }
2569         spin_lock(&rq->lock);
2570         /*
2571          * The task was running during this tick - update the
2572          * time slice counter. Note: we do not update a thread's
2573          * priority until it either goes to sleep or uses up its
2574          * timeslice. This makes it possible for interactive tasks
2575          * to use up their timeslices at their highest priority levels.
2576          */
2577         if (rt_task(p)) {
2578                 /*
2579                  * RR tasks need a special form of timeslice management.
2580                  * FIFO tasks have no timeslices.
2581                  */
2582                 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2583                         p->time_slice = task_timeslice(p);
2584                         p->first_time_slice = 0;
2585                         set_tsk_need_resched(p);
2586
2587                         /* put it at the end of the queue: */
2588                         requeue_task(p, rq->active);
2589                 }
2590                 goto out_unlock;
2591         }
2592         if (!--p->time_slice) {
2593                 dequeue_task(p, rq->active);
2594                 set_tsk_need_resched(p);
2595                 p->prio = effective_prio(p);
2596                 p->time_slice = task_timeslice(p);
2597                 p->first_time_slice = 0;
2598
2599                 if (!rq->expired_timestamp)
2600                         rq->expired_timestamp = jiffies;
2601                 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2602                         enqueue_task(p, rq->expired);
2603                         if (p->static_prio < rq->best_expired_prio)
2604                                 rq->best_expired_prio = p->static_prio;
2605                 } else
2606                         enqueue_task(p, rq->active);
2607         } else {
2608                 /*
2609                  * Prevent a too long timeslice allowing a task to monopolize
2610                  * the CPU. We do this by splitting up the timeslice into
2611                  * smaller pieces.
2612                  *
2613                  * Note: this does not mean the task's timeslices expire or
2614                  * get lost in any way, they just might be preempted by
2615                  * another task of equal priority. (one with higher
2616                  * priority would have preempted this task already.) We
2617                  * requeue this task to the end of the list on this priority
2618                  * level, which is in essence a round-robin of tasks with
2619                  * equal priority.
2620                  *
2621                  * This only applies to tasks in the interactive
2622                  * delta range with at least TIMESLICE_GRANULARITY to requeue.
2623                  */
2624                 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2625                         p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2626                         (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2627                         (p->array == rq->active)) {
2628
2629                         requeue_task(p, rq->active);
2630                         set_tsk_need_resched(p);
2631                 }
2632         }
2633 out_unlock:
2634         spin_unlock(&rq->lock);
2635 out:
2636         rebalance_tick(cpu, rq, NOT_IDLE);
2637 }
2638
2639 #ifdef CONFIG_SCHED_SMT
2640 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2641 {
2642         /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2643         if (rq->curr == rq->idle && rq->nr_running)
2644                 resched_task(rq->idle);
2645 }
2646
2647 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2648 {
2649         struct sched_domain *tmp, *sd = NULL;
2650         cpumask_t sibling_map;
2651         int i;
2652
2653         for_each_domain(this_cpu, tmp)
2654                 if (tmp->flags & SD_SHARE_CPUPOWER)
2655                         sd = tmp;
2656
2657         if (!sd)
2658                 return;
2659
2660         /*
2661          * Unlock the current runqueue because we have to lock in
2662          * CPU order to avoid deadlocks. Caller knows that we might
2663          * unlock. We keep IRQs disabled.
2664          */
2665         spin_unlock(&this_rq->lock);
2666
2667         sibling_map = sd->span;
2668
2669         for_each_cpu_mask(i, sibling_map)
2670                 spin_lock(&cpu_rq(i)->lock);
2671         /*
2672          * We clear this CPU from the mask. This both simplifies the
2673          * inner loop and keps this_rq locked when we exit:
2674          */
2675         cpu_clear(this_cpu, sibling_map);
2676
2677         for_each_cpu_mask(i, sibling_map) {
2678                 runqueue_t *smt_rq = cpu_rq(i);
2679
2680                 wakeup_busy_runqueue(smt_rq);
2681         }
2682
2683         for_each_cpu_mask(i, sibling_map)
2684                 spin_unlock(&cpu_rq(i)->lock);
2685         /*
2686          * We exit with this_cpu's rq still held and IRQs
2687          * still disabled:
2688          */
2689 }
2690
2691 /*
2692  * number of 'lost' timeslices this task wont be able to fully
2693  * utilize, if another task runs on a sibling. This models the
2694  * slowdown effect of other tasks running on siblings:
2695  */
2696 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2697 {
2698         return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2699 }
2700
2701 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2702 {
2703         struct sched_domain *tmp, *sd = NULL;
2704         cpumask_t sibling_map;
2705         prio_array_t *array;
2706         int ret = 0, i;
2707         task_t *p;
2708
2709         for_each_domain(this_cpu, tmp)
2710                 if (tmp->flags & SD_SHARE_CPUPOWER)
2711                         sd = tmp;
2712
2713         if (!sd)
2714                 return 0;
2715
2716         /*
2717          * The same locking rules and details apply as for
2718          * wake_sleeping_dependent():
2719          */
2720         spin_unlock(&this_rq->lock);
2721         sibling_map = sd->span;
2722         for_each_cpu_mask(i, sibling_map)
2723                 spin_lock(&cpu_rq(i)->lock);
2724         cpu_clear(this_cpu, sibling_map);
2725
2726         /*
2727          * Establish next task to be run - it might have gone away because
2728          * we released the runqueue lock above:
2729          */
2730         if (!this_rq->nr_running)
2731                 goto out_unlock;
2732         array = this_rq->active;
2733         if (!array->nr_active)
2734                 array = this_rq->expired;
2735         BUG_ON(!array->nr_active);
2736
2737         p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2738                 task_t, run_list);
2739
2740         for_each_cpu_mask(i, sibling_map) {
2741                 runqueue_t *smt_rq = cpu_rq(i);
2742                 task_t *smt_curr = smt_rq->curr;
2743
2744                 /* Kernel threads do not participate in dependent sleeping */
2745                 if (!p->mm || !smt_curr->mm || rt_task(p))
2746                         goto check_smt_task;
2747
2748                 /*
2749                  * If a user task with lower static priority than the
2750                  * running task on the SMT sibling is trying to schedule,
2751                  * delay it till there is proportionately less timeslice
2752                  * left of the sibling task to prevent a lower priority
2753                  * task from using an unfair proportion of the
2754                  * physical cpu's resources. -ck
2755                  */
2756                 if (rt_task(smt_curr)) {
2757                         /*
2758                          * With real time tasks we run non-rt tasks only
2759                          * per_cpu_gain% of the time.
2760                          */
2761                         if ((jiffies % DEF_TIMESLICE) >
2762                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2763                                         ret = 1;
2764                 } else
2765                         if (smt_curr->static_prio < p->static_prio &&
2766                                 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2767                                 smt_slice(smt_curr, sd) > task_timeslice(p))
2768                                         ret = 1;
2769
2770 check_smt_task:
2771                 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2772                         rt_task(smt_curr))
2773                                 continue;
2774                 if (!p->mm) {
2775                         wakeup_busy_runqueue(smt_rq);
2776                         continue;
2777                 }
2778
2779                 /*
2780                  * Reschedule a lower priority task on the SMT sibling for
2781                  * it to be put to sleep, or wake it up if it has been put to
2782                  * sleep for priority reasons to see if it should run now.
2783                  */
2784                 if (rt_task(p)) {
2785                         if ((jiffies % DEF_TIMESLICE) >
2786                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2787                                         resched_task(smt_curr);
2788                 } else {
2789                         if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2790                                 smt_slice(p, sd) > task_timeslice(smt_curr))
2791                                         resched_task(smt_curr);
2792                         else
2793                                 wakeup_busy_runqueue(smt_rq);
2794                 }
2795         }
2796 out_unlock:
2797         for_each_cpu_mask(i, sibling_map)
2798                 spin_unlock(&cpu_rq(i)->lock);
2799         return ret;
2800 }
2801 #else
2802 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2803 {
2804 }
2805
2806 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2807 {
2808         return 0;
2809 }
2810 #endif
2811
2812 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2813
2814 void fastcall add_preempt_count(int val)
2815 {
2816         /*
2817          * Underflow?
2818          */
2819         BUG_ON((preempt_count() < 0));
2820         preempt_count() += val;
2821         /*
2822          * Spinlock count overflowing soon?
2823          */
2824         BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2825 }
2826 EXPORT_SYMBOL(add_preempt_count);
2827
2828 void fastcall sub_preempt_count(int val)
2829 {
2830         /*
2831          * Underflow?
2832          */
2833         BUG_ON(val > preempt_count());
2834         /*
2835          * Is the spinlock portion underflowing?
2836          */
2837         BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2838         preempt_count() -= val;
2839 }
2840 EXPORT_SYMBOL(sub_preempt_count);
2841
2842 #endif
2843
2844 /*
2845  * schedule() is the main scheduler function.
2846  */
2847 asmlinkage void __sched schedule(void)
2848 {
2849         long *switch_count;
2850         task_t *prev, *next;
2851         runqueue_t *rq;
2852         prio_array_t *array;
2853         struct list_head *queue;
2854         unsigned long long now;
2855         unsigned long run_time;
2856         int cpu, idx, new_prio;
2857
2858         /*
2859          * Test if we are atomic.  Since do_exit() needs to call into
2860          * schedule() atomically, we ignore that path for now.
2861          * Otherwise, whine if we are scheduling when we should not be.
2862          */
2863         if (likely(!current->exit_state)) {
2864                 if (unlikely(in_atomic())) {
2865                         printk(KERN_ERR "scheduling while atomic: "
2866                                 "%s/0x%08x/%d\n",
2867                                 current->comm, preempt_count(), current->pid);
2868                         dump_stack();
2869                 }
2870         }
2871         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2872
2873 need_resched:
2874         preempt_disable();
2875         prev = current;
2876         release_kernel_lock(prev);
2877 need_resched_nonpreemptible:
2878         rq = this_rq();
2879
2880         /*
2881          * The idle thread is not allowed to schedule!
2882          * Remove this check after it has been exercised a bit.
2883          */
2884         if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2885                 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2886                 dump_stack();
2887         }
2888
2889         schedstat_inc(rq, sched_cnt);
2890         now = sched_clock();
2891         if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2892                 run_time = now - prev->timestamp;
2893                 if (unlikely((long long)(now - prev->timestamp) < 0))
2894                         run_time = 0;
2895         } else
2896                 run_time = NS_MAX_SLEEP_AVG;
2897
2898         /*
2899          * Tasks charged proportionately less run_time at high sleep_avg to
2900          * delay them losing their interactive status
2901          */
2902         run_time /= (CURRENT_BONUS(prev) ? : 1);
2903
2904         spin_lock_irq(&rq->lock);
2905
2906         if (unlikely(prev->flags & PF_DEAD))
2907                 prev->state = EXIT_DEAD;
2908
2909         switch_count = &prev->nivcsw;
2910         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2911                 switch_count = &prev->nvcsw;
2912                 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2913                                 unlikely(signal_pending(prev))))
2914                         prev->state = TASK_RUNNING;
2915                 else {
2916                         if (prev->state == TASK_UNINTERRUPTIBLE)
2917                                 rq->nr_uninterruptible++;
2918                         deactivate_task(prev, rq);
2919                 }
2920         }
2921
2922         cpu = smp_processor_id();
2923         if (unlikely(!rq->nr_running)) {
2924 go_idle:
2925                 idle_balance(cpu, rq);
2926                 if (!rq->nr_running) {
2927                         next = rq->idle;
2928                         rq->expired_timestamp = 0;
2929                         wake_sleeping_dependent(cpu, rq);
2930                         /*
2931                          * wake_sleeping_dependent() might have released
2932                          * the runqueue, so break out if we got new
2933                          * tasks meanwhile:
2934                          */
2935                         if (!rq->nr_running)
2936                                 goto switch_tasks;
2937                 }
2938         } else {
2939                 if (dependent_sleeper(cpu, rq)) {
2940                         next = rq->idle;
2941                         goto switch_tasks;
2942                 }
2943                 /*
2944                  * dependent_sleeper() releases and reacquires the runqueue
2945                  * lock, hence go into the idle loop if the rq went
2946                  * empty meanwhile:
2947                  */
2948                 if (unlikely(!rq->nr_running))
2949                         goto go_idle;
2950         }
2951
2952         array = rq->active;
2953         if (unlikely(!array->nr_active)) {
2954                 /*
2955                  * Switch the active and expired arrays.
2956                  */
2957                 schedstat_inc(rq, sched_switch);
2958                 rq->active = rq->expired;
2959                 rq->expired = array;
2960                 array = rq->active;
2961                 rq->expired_timestamp = 0;
2962                 rq->best_expired_prio = MAX_PRIO;
2963         }
2964
2965         idx = sched_find_first_bit(array->bitmap);
2966         queue = array->queue + idx;
2967         next = list_entry(queue->next, task_t, run_list);
2968
2969         if (!rt_task(next) && next->activated > 0) {
2970                 unsigned long long delta = now - next->timestamp;
2971                 if (unlikely((long long)(now - next->timestamp) < 0))
2972                         delta = 0;
2973
2974                 if (next->activated == 1)
2975                         delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2976
2977                 array = next->array;
2978                 new_prio = recalc_task_prio(next, next->timestamp + delta);
2979
2980                 if (unlikely(next->prio != new_prio)) {
2981                         dequeue_task(next, array);
2982                         next->prio = new_prio;
2983                         enqueue_task(next, array);
2984                 } else
2985                         requeue_task(next, array);
2986         }
2987         next->activated = 0;
2988 switch_tasks:
2989         if (next == rq->idle)
2990                 schedstat_inc(rq, sched_goidle);
2991         prefetch(next);
2992         prefetch_stack(next);
2993         clear_tsk_need_resched(prev);
2994         rcu_qsctr_inc(task_cpu(prev));
2995
2996         update_cpu_clock(prev, rq, now);
2997
2998         prev->sleep_avg -= run_time;
2999         if ((long)prev->sleep_avg <= 0)
3000                 prev->sleep_avg = 0;
3001         prev->timestamp = prev->last_ran = now;
3002
3003         sched_info_switch(prev, next);
3004         if (likely(prev != next)) {
3005                 next->timestamp = now;
3006                 rq->nr_switches++;
3007                 rq->curr = next;
3008                 ++*switch_count;
3009
3010                 prepare_task_switch(rq, next);
3011                 prev = context_switch(rq, prev, next);
3012                 barrier();
3013                 /*
3014                  * this_rq must be evaluated again because prev may have moved
3015                  * CPUs since it called schedule(), thus the 'rq' on its stack
3016                  * frame will be invalid.
3017                  */
3018                 finish_task_switch(this_rq(), prev);
3019         } else
3020                 spin_unlock_irq(&rq->lock);
3021
3022         prev = current;
3023         if (unlikely(reacquire_kernel_lock(prev) < 0))
3024                 goto need_resched_nonpreemptible;
3025         preempt_enable_no_resched();
3026         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3027                 goto need_resched;
3028 }
3029
3030 EXPORT_SYMBOL(schedule);
3031
3032 #ifdef CONFIG_PREEMPT
3033 /*
3034  * this is is the entry point to schedule() from in-kernel preemption
3035  * off of preempt_enable.  Kernel preemptions off return from interrupt
3036  * occur there and call schedule directly.
3037  */
3038 asmlinkage void __sched preempt_schedule(void)
3039 {
3040         struct thread_info *ti = current_thread_info();
3041 #ifdef CONFIG_PREEMPT_BKL
3042         struct task_struct *task = current;
3043         int saved_lock_depth;
3044 #endif
3045         /*
3046          * If there is a non-zero preempt_count or interrupts are disabled,
3047          * we do not want to preempt the current task.  Just return..
3048          */
3049         if (unlikely(ti->preempt_count || irqs_disabled()))
3050                 return;
3051
3052 need_resched:
3053         add_preempt_count(PREEMPT_ACTIVE);
3054         /*
3055          * We keep the big kernel semaphore locked, but we
3056          * clear ->lock_depth so that schedule() doesnt
3057          * auto-release the semaphore:
3058          */
3059 #ifdef CONFIG_PREEMPT_BKL
3060         saved_lock_depth = task->lock_depth;
3061         task->lock_depth = -1;
3062 #endif
3063         schedule();
3064 #ifdef CONFIG_PREEMPT_BKL
3065         task->lock_depth = saved_lock_depth;
3066 #endif
3067         sub_preempt_count(PREEMPT_ACTIVE);
3068
3069         /* we could miss a preemption opportunity between schedule and now */
3070         barrier();
3071         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3072                 goto need_resched;
3073 }
3074
3075 EXPORT_SYMBOL(preempt_schedule);
3076
3077 /*
3078  * this is is the entry point to schedule() from kernel preemption
3079  * off of irq context.
3080  * Note, that this is called and return with irqs disabled. This will
3081  * protect us against recursive calling from irq.
3082  */
3083 asmlinkage void __sched preempt_schedule_irq(void)
3084 {
3085         struct thread_info *ti = current_thread_info();
3086 #ifdef CONFIG_PREEMPT_BKL
3087         struct task_struct *task = current;
3088         int saved_lock_depth;
3089 #endif
3090         /* Catch callers which need to be fixed*/
3091         BUG_ON(ti->preempt_count || !irqs_disabled());
3092
3093 need_resched:
3094         add_preempt_count(PREEMPT_ACTIVE);
3095         /*
3096          * We keep the big kernel semaphore locked, but we
3097          * clear ->lock_depth so that schedule() doesnt
3098          * auto-release the semaphore:
3099          */
3100 #ifdef CONFIG_PREEMPT_BKL
3101         saved_lock_depth = task->lock_depth;
3102         task->lock_depth = -1;
3103 #endif
3104         local_irq_enable();
3105         schedule();
3106         local_irq_disable();
3107 #ifdef CONFIG_PREEMPT_BKL
3108         task->lock_depth = saved_lock_depth;
3109 #endif
3110         sub_preempt_count(PREEMPT_ACTIVE);
3111
3112         /* we could miss a preemption opportunity between schedule and now */
3113         barrier();
3114         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3115                 goto need_resched;
3116 }
3117
3118 #endif /* CONFIG_PREEMPT */
3119
3120 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3121                           void *key)
3122 {
3123         task_t *p = curr->private;
3124         return try_to_wake_up(p, mode, sync);
3125 }
3126
3127 EXPORT_SYMBOL(default_wake_function);
3128
3129 /*
3130  * The core wakeup function.  Non-exclusive wakeups (nr_exclusive == 0) just
3131  * wake everything up.  If it's an exclusive wakeup (nr_exclusive == small +ve
3132  * number) then we wake all the non-exclusive tasks and one exclusive task.
3133  *
3134  * There are circumstances in which we can try to wake a task which has already
3135  * started to run but is not in state TASK_RUNNING.  try_to_wake_up() returns
3136  * zero in this (rare) case, and we handle it by continuing to scan the queue.
3137  */
3138 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3139                              int nr_exclusive, int sync, void *key)
3140 {
3141         struct list_head *tmp, *next;
3142
3143         list_for_each_safe(tmp, next, &q->task_list) {
3144                 wait_queue_t *curr;
3145                 unsigned flags;
3146                 curr = list_entry(tmp, wait_queue_t, task_list);
3147                 flags = curr->flags;
3148                 if (curr->func(curr, mode, sync, key) &&
3149                     (flags & WQ_FLAG_EXCLUSIVE) &&
3150                     !--nr_exclusive)
3151                         break;
3152         }
3153 }
3154
3155 /**
3156  * __wake_up - wake up threads blocked on a waitqueue.
3157  * @q: the waitqueue
3158  * @mode: which threads
3159  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3160  * @key: is directly passed to the wakeup function
3161  */
3162 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3163                         int nr_exclusive, void *key)
3164 {
3165         unsigned long flags;
3166
3167         spin_lock_irqsave(&q->lock, flags);
3168         __wake_up_common(q, mode, nr_exclusive, 0, key);
3169         spin_unlock_irqrestore(&q->lock, flags);
3170 }
3171
3172 EXPORT_SYMBOL(__wake_up);
3173
3174 /*
3175  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3176  */
3177 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3178 {
3179         __wake_up_common(q, mode, 1, 0, NULL);
3180 }
3181
3182 /**
3183  * __wake_up_sync - wake up threads blocked on a waitqueue.
3184  * @q: the waitqueue
3185  * @mode: which threads
3186  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3187  *
3188  * The sync wakeup differs that the waker knows that it will schedule
3189  * away soon, so while the target thread will be woken up, it will not
3190  * be migrated to another CPU - ie. the two threads are 'synchronized'
3191  * with each other. This can prevent needless bouncing between CPUs.
3192  *
3193  * On UP it can prevent extra preemption.
3194  */
3195 void fastcall
3196 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3197 {
3198         unsigned long flags;
3199         int sync = 1;
3200
3201         if (unlikely(!q))
3202                 return;
3203
3204         if (unlikely(!nr_exclusive))
3205                 sync = 0;
3206
3207         spin_lock_irqsave(&q->lock, flags);
3208         __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3209         spin_unlock_irqrestore(&q->lock, flags);
3210 }
3211 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3212
3213 void fastcall complete(struct completion *x)
3214 {
3215         unsigned long flags;
3216
3217         spin_lock_irqsave(&x->wait.lock, flags);
3218         x->done++;
3219         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3220                          1, 0, NULL);
3221         spin_unlock_irqrestore(&x->wait.lock, flags);
3222 }
3223 EXPORT_SYMBOL(complete);
3224
3225 void fastcall complete_all(struct completion *x)
3226 {
3227         unsigned long flags;
3228
3229         spin_lock_irqsave(&x->wait.lock, flags);
3230         x->done += UINT_MAX/2;
3231         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3232                          0, 0, NULL);
3233         spin_unlock_irqrestore(&x->wait.lock, flags);
3234 }
3235 EXPORT_SYMBOL(complete_all);
3236
3237 void fastcall __sched wait_for_completion(struct completion *x)
3238 {
3239         might_sleep();
3240         spin_lock_irq(&x->wait.lock);
3241         if (!x->done) {
3242                 DECLARE_WAITQUEUE(wait, current);
3243
3244                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3245                 __add_wait_queue_tail(&x->wait, &wait);
3246                 do {
3247                         __set_current_state(TASK_UNINTERRUPTIBLE);
3248                         spin_unlock_irq(&x->wait.lock);
3249                         schedule();
3250                         spin_lock_irq(&x->wait.lock);
3251                 } while (!x->done);
3252                 __remove_wait_queue(&x->wait, &wait);
3253         }
3254         x->done--;
3255         spin_unlock_irq(&x->wait.lock);
3256 }
3257 EXPORT_SYMBOL(wait_for_completion);
3258
3259 unsigned long fastcall __sched
3260 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3261 {
3262         might_sleep();
3263
3264         spin_lock_irq(&x->wait.lock);
3265         if (!x->done) {
3266                 DECLARE_WAITQUEUE(wait, current);
3267
3268                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3269                 __add_wait_queue_tail(&x->wait, &wait);
3270                 do {
3271                         __set_current_state(TASK_UNINTERRUPTIBLE);
3272                         spin_unlock_irq(&x->wait.lock);
3273                         timeout = schedule_timeout(timeout);
3274                         spin_lock_irq(&x->wait.lock);
3275                         if (!timeout) {
3276                                 __remove_wait_queue(&x->wait, &wait);
3277                                 goto out;
3278                         }
3279                 } while (!x->done);
3280                 __remove_wait_queue(&x->wait, &wait);
3281         }
3282         x->done--;
3283 out:
3284         spin_unlock_irq(&x->wait.lock);
3285         return timeout;
3286 }
3287 EXPORT_SYMBOL(wait_for_completion_timeout);
3288
3289 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3290 {
3291         int ret = 0;
3292
3293         might_sleep();
3294
3295         spin_lock_irq(&x->wait.lock);
3296         if (!x->done) {
3297                 DECLARE_WAITQUEUE(wait, current);
3298
3299                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3300                 __add_wait_queue_tail(&x->wait, &wait);
3301                 do {
3302                         if (signal_pending(current)) {
3303                                 ret = -ERESTARTSYS;
3304                                 __remove_wait_queue(&x->wait, &wait);
3305                                 goto out;
3306                         }
3307                         __set_current_state(TASK_INTERRUPTIBLE);
3308                         spin_unlock_irq(&x->wait.lock);
3309                         schedule();
3310                         spin_lock_irq(&x->wait.lock);
3311                 } while (!x->done);
3312                 __remove_wait_queue(&x->wait, &wait);
3313         }
3314         x->done--;
3315 out:
3316         spin_unlock_irq(&x->wait.lock);
3317
3318         return ret;
3319 }
3320 EXPORT_SYMBOL(wait_for_completion_interruptible);
3321
3322 unsigned long fastcall __sched
3323 wait_for_completion_interruptible_timeout(struct completion *x,
3324                                           unsigned long timeout)
3325 {
3326         might_sleep();
3327
3328         spin_lock_irq(&x->wait.lock);
3329         if (!x->done) {
3330                 DECLARE_WAITQUEUE(wait, current);
3331
3332                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3333                 __add_wait_queue_tail(&x->wait, &wait);
3334                 do {
3335                         if (signal_pending(current)) {
3336                                 timeout = -ERESTARTSYS;
3337                                 __remove_wait_queue(&x->wait, &wait);
3338                                 goto out;
3339                         }
3340                         __set_current_state(TASK_INTERRUPTIBLE);
3341                         spin_unlock_irq(&x->wait.lock);
3342                         timeout = schedule_timeout(timeout);
3343                         spin_lock_irq(&x->wait.lock);
3344                         if (!timeout) {
3345                                 __remove_wait_queue(&x->wait, &wait);
3346                                 goto out;
3347                         }
3348                 } while (!x->done);
3349                 __remove_wait_queue(&x->wait, &wait);
3350         }
3351         x->done--;
3352 out:
3353         spin_unlock_irq(&x->wait.lock);
3354         return timeout;
3355 }
3356 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3357
3358
3359 #define SLEEP_ON_VAR                                    \
3360         unsigned long flags;                            \
3361         wait_queue_t wait;                              \
3362         init_waitqueue_entry(&wait, current);
3363
3364 #define SLEEP_ON_HEAD                                   \
3365         spin_lock_irqsave(&q->lock,flags);              \
3366         __add_wait_queue(q, &wait);                     \
3367         spin_unlock(&q->lock);
3368
3369 #define SLEEP_ON_TAIL                                   \
3370         spin_lock_irq(&q->lock);                        \
3371         __remove_wait_queue(q, &wait);                  \
3372         spin_unlock_irqrestore(&q->lock, flags);
3373
3374 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3375 {
3376         SLEEP_ON_VAR
3377
3378         current->state = TASK_INTERRUPTIBLE;
3379
3380         SLEEP_ON_HEAD
3381         schedule();
3382         SLEEP_ON_TAIL
3383 }
3384
3385 EXPORT_SYMBOL(interruptible_sleep_on);
3386
3387 long fastcall __sched
3388 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3389 {
3390         SLEEP_ON_VAR
3391
3392         current->state = TASK_INTERRUPTIBLE;
3393
3394         SLEEP_ON_HEAD
3395         timeout = schedule_timeout(timeout);
3396         SLEEP_ON_TAIL
3397
3398         return timeout;
3399 }
3400
3401 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3402
3403 void fastcall __sched sleep_on(wait_queue_head_t *q)
3404 {
3405         SLEEP_ON_VAR
3406
3407         current->state = TASK_UNINTERRUPTIBLE;
3408
3409         SLEEP_ON_HEAD
3410         schedule();
3411         SLEEP_ON_TAIL
3412 }
3413
3414 EXPORT_SYMBOL(sleep_on);
3415
3416 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3417 {
3418         SLEEP_ON_VAR
3419
3420         current->state = TASK_UNINTERRUPTIBLE;
3421
3422         SLEEP_ON_HEAD
3423         timeout = schedule_timeout(timeout);
3424         SLEEP_ON_TAIL
3425
3426         return timeout;
3427 }
3428
3429 EXPORT_SYMBOL(sleep_on_timeout);
3430
3431 void set_user_nice(task_t *p, long nice)
3432 {
3433         unsigned long flags;
3434         prio_array_t *array;
3435         runqueue_t *rq;
3436         int old_prio, new_prio, delta;
3437
3438         if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3439                 return;
3440         /*
3441          * We have to be careful, if called from sys_setpriority(),
3442          * the task might be in the middle of scheduling on another CPU.
3443          */
3444         rq = task_rq_lock(p, &flags);
3445         /*
3446          * The RT priorities are set via sched_setscheduler(), but we still
3447          * allow the 'normal' nice value to be set - but as expected
3448          * it wont have any effect on scheduling until the task is
3449          * not SCHED_NORMAL:
3450          */
3451         if (rt_task(p)) {
3452                 p->static_prio = NICE_TO_PRIO(nice);
3453                 goto out_unlock;
3454         }
3455         array = p->array;
3456         if (array)
3457                 dequeue_task(p, array);
3458
3459         old_prio = p->prio;
3460         new_prio = NICE_TO_PRIO(nice);
3461         delta = new_prio - old_prio;
3462         p->static_prio = NICE_TO_PRIO(nice);
3463         p->prio += delta;
3464
3465         if (array) {
3466                 enqueue_task(p, array);
3467                 /*
3468                  * If the task increased its priority or is running and
3469                  * lowered its priority, then reschedule its CPU:
3470                  */
3471                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3472                         resched_task(rq->curr);
3473         }
3474 out_unlock:
3475         task_rq_unlock(rq, &flags);
3476 }
3477
3478 EXPORT_SYMBOL(set_user_nice);
3479
3480 /*
3481  * can_nice - check if a task can reduce its nice value
3482  * @p: task
3483  * @nice: nice value
3484  */
3485 int can_nice(const task_t *p, const int nice)
3486 {
3487         /* convert nice value [19,-20] to rlimit style value [1,40] */
3488         int nice_rlim = 20 - nice;
3489         return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3490                 capable(CAP_SYS_NICE));
3491 }
3492
3493 #ifdef __ARCH_WANT_SYS_NICE
3494
3495 /*
3496  * sys_nice - change the priority of the current process.
3497  * @increment: priority increment
3498  *
3499  * sys_setpriority is a more generic, but much slower function that
3500  * does similar things.
3501  */
3502 asmlinkage long sys_nice(int increment)
3503 {
3504         int retval;
3505         long nice;
3506
3507         /*
3508          * Setpriority might change our priority at the same moment.
3509          * We don't have to worry. Conceptually one call occurs first
3510          * and we have a single winner.
3511          */
3512         if (increment < -40)
3513                 increment = -40;
3514         if (increment > 40)
3515                 increment = 40;
3516
3517         nice = PRIO_TO_NICE(current->static_prio) + increment;
3518         if (nice < -20)
3519                 nice = -20;
3520         if (nice > 19)
3521                 nice = 19;
3522
3523         if (increment < 0 && !can_nice(current, nice))
3524                 return -EPERM;
3525
3526         retval = security_task_setnice(current, nice);
3527         if (retval)
3528                 return retval;
3529
3530         set_user_nice(current, nice);
3531         return 0;
3532 }
3533
3534 #endif
3535
3536 /**
3537  * task_prio - return the priority value of a given task.
3538  * @p: the task in question.
3539  *
3540  * This is the priority value as seen by users in /proc.
3541  * RT tasks are offset by -200. Normal tasks are centered
3542  * around 0, value goes from -16 to +15.
3543  */
3544 int task_prio(const task_t *p)
3545 {
3546         return p->prio - MAX_RT_PRIO;
3547 }
3548
3549 /**
3550  * task_nice - return the nice value of a given task.
3551  * @p: the task in question.
3552  */
3553 int task_nice(const task_t *p)
3554 {
3555         return TASK_NICE(p);
3556 }
3557 EXPORT_SYMBOL_GPL(task_nice);
3558
3559 /**
3560  * idle_cpu - is a given cpu idle currently?
3561  * @cpu: the processor in question.
3562  */
3563 int idle_cpu(int cpu)
3564 {
3565         return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3566 }
3567
3568 EXPORT_SYMBOL_GPL(idle_cpu);
3569
3570 /**
3571  * idle_task - return the idle task for a given cpu.
3572  * @cpu: the processor in question.
3573  */
3574 task_t *idle_task(int cpu)
3575 {
3576         return cpu_rq(cpu)->idle;
3577 }
3578
3579 /**
3580  * find_process_by_pid - find a process with a matching PID value.
3581  * @pid: the pid in question.
3582  */
3583 static inline task_t *find_process_by_pid(pid_t pid)
3584 {
3585         return pid ? find_task_by_pid(pid) : current;
3586 }
3587
3588 /* Actually do priority change: must hold rq lock. */
3589 static void __setscheduler(struct task_struct *p, int policy, int prio)
3590 {
3591         BUG_ON(p->array);
3592         p->policy = policy;
3593         p->rt_priority = prio;
3594         if (policy != SCHED_NORMAL)
3595                 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3596         else
3597                 p->prio = p->static_prio;
3598 }
3599
3600 /**
3601  * sched_setscheduler - change the scheduling policy and/or RT priority of
3602  * a thread.
3603  * @p: the task in question.
3604  * @policy: new policy.
3605  * @param: structure containing the new RT priority.
3606  */
3607 int sched_setscheduler(struct task_struct *p, int policy,
3608                        struct sched_param *param)
3609 {
3610         int retval;
3611         int oldprio, oldpolicy = -1;
3612         prio_array_t *array;
3613         unsigned long flags;
3614         runqueue_t *rq;
3615
3616 recheck:
3617         /* double check policy once rq lock held */
3618         if (policy < 0)
3619                 policy = oldpolicy = p->policy;
3620         else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3621                                 policy != SCHED_NORMAL)
3622                         return -EINVAL;
3623         /*
3624          * Valid priorities for SCHED_FIFO and SCHED_RR are
3625          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3626          */
3627         if (param->sched_priority < 0 ||
3628             (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3629             (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3630                 return -EINVAL;
3631         if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3632                 return -EINVAL;
3633
3634         /*
3635          * Allow unprivileged RT tasks to decrease priority:
3636          */
3637         if (!capable(CAP_SYS_NICE)) {
3638                 /* can't change policy */
3639                 if (policy != p->policy &&
3640                         !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3641                         return -EPERM;
3642                 /* can't increase priority */
3643                 if (policy != SCHED_NORMAL &&
3644                     param->sched_priority > p->rt_priority &&
3645                     param->sched_priority >
3646                                 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3647                         return -EPERM;
3648                 /* can't change other user's priorities */
3649                 if ((current->euid != p->euid) &&
3650                     (current->euid != p->uid))
3651                         return -EPERM;
3652         }
3653
3654         retval = security_task_setscheduler(p, policy, param);
3655         if (retval)
3656                 return retval;
3657         /*
3658          * To be able to change p->policy safely, the apropriate
3659          * runqueue lock must be held.
3660          */
3661         rq = task_rq_lock(p, &flags);
3662         /* recheck policy now with rq lock held */
3663         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3664                 policy = oldpolicy = -1;
3665                 task_rq_unlock(rq, &flags);
3666                 goto recheck;
3667         }
3668         array = p->array;
3669         if (array)
3670                 deactivate_task(p, rq);
3671         oldprio = p->prio;
3672         __setscheduler(p, policy, param->sched_priority);
3673         if (array) {
3674                 __activate_task(p, rq);
3675                 /*
3676                  * Reschedule if we are currently running on this runqueue and
3677                  * our priority decreased, or if we are not currently running on
3678                  * this runqueue and our priority is higher than the current's
3679                  */
3680                 if (task_running(rq, p)) {
3681                         if (p->prio > oldprio)
3682                                 resched_task(rq->curr);
3683                 } else if (TASK_PREEMPTS_CURR(p, rq))
3684                         resched_task(rq->curr);
3685         }
3686         task_rq_unlock(rq, &flags);
3687         return 0;
3688 }
3689 EXPORT_SYMBOL_GPL(sched_setscheduler);
3690
3691 static int
3692 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3693 {
3694         int retval;
3695         struct sched_param lparam;
3696         struct task_struct *p;
3697
3698         if (!param || pid < 0)
3699                 return -EINVAL;
3700         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3701                 return -EFAULT;
3702         read_lock_irq(&tasklist_lock);
3703         p = find_process_by_pid(pid);
3704         if (!p) {
3705                 read_unlock_irq(&tasklist_lock);
3706                 return -ESRCH;
3707         }
3708         retval = sched_setscheduler(p, policy, &lparam);
3709         read_unlock_irq(&tasklist_lock);
3710         return retval;
3711 }
3712
3713 /**
3714  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3715  * @pid: the pid in question.
3716  * @policy: new policy.
3717  * @param: structure containing the new RT priority.
3718  */
3719 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3720                                        struct sched_param __user *param)
3721 {
3722         return do_sched_setscheduler(pid, policy, param);
3723 }
3724
3725 /**
3726  * sys_sched_setparam - set/change the RT priority of a thread
3727  * @pid: the pid in question.
3728  * @param: structure containing the new RT priority.
3729  */
3730 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3731 {
3732         return do_sched_setscheduler(pid, -1, param);
3733 }
3734
3735 /**
3736  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3737  * @pid: the pid in question.
3738  */
3739 asmlinkage long sys_sched_getscheduler(pid_t pid)
3740 {
3741         int retval = -EINVAL;
3742         task_t *p;
3743
3744         if (pid < 0)
3745                 goto out_nounlock;
3746
3747         retval = -ESRCH;
3748         read_lock(&tasklist_lock);
3749         p = find_process_by_pid(pid);
3750         if (p) {
3751                 retval = security_task_getscheduler(p);
3752                 if (!retval)
3753                         retval = p->policy;
3754         }
3755         read_unlock(&tasklist_lock);
3756
3757 out_nounlock:
3758         return retval;
3759 }
3760
3761 /**
3762  * sys_sched_getscheduler - get the RT priority of a thread
3763  * @pid: the pid in question.
3764  * @param: structure containing the RT priority.
3765  */
3766 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3767 {
3768         struct sched_param lp;
3769         int retval = -EINVAL;
3770         task_t *p;
3771
3772         if (!param || pid < 0)
3773                 goto out_nounlock;
3774
3775         read_lock(&tasklist_lock);
3776         p = find_process_by_pid(pid);
3777         retval = -ESRCH;
3778         if (!p)
3779                 goto out_unlock;
3780
3781         retval = security_task_getscheduler(p);
3782         if (retval)
3783                 goto out_unlock;
3784
3785         lp.sched_priority = p->rt_priority;
3786         read_unlock(&tasklist_lock);
3787
3788         /*
3789          * This one might sleep, we cannot do it with a spinlock held ...
3790          */
3791         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3792
3793 out_nounlock:
3794         return retval;
3795
3796 out_unlock:
3797         read_unlock(&tasklist_lock);
3798         return retval;
3799 }
3800
3801 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3802 {
3803         task_t *p;
3804         int retval;
3805         cpumask_t cpus_allowed;
3806
3807         lock_cpu_hotplug();
3808         read_lock(&tasklist_lock);
3809
3810         p = find_process_by_pid(pid);
3811         if (!p) {
3812                 read_unlock(&tasklist_lock);
3813                 unlock_cpu_hotplug();
3814                 return -ESRCH;
3815         }
3816
3817         /*
3818          * It is not safe to call set_cpus_allowed with the
3819          * tasklist_lock held.  We will bump the task_struct's
3820          * usage count and then drop tasklist_lock.
3821          */
3822         get_task_struct(p);
3823         read_unlock(&tasklist_lock);
3824
3825         retval = -EPERM;
3826         if ((current->euid != p->euid) && (current->euid != p->uid) &&
3827                         !capable(CAP_SYS_NICE))
3828                 goto out_unlock;
3829
3830         cpus_allowed = cpuset_cpus_allowed(p);
3831         cpus_and(new_mask, new_mask, cpus_allowed);
3832         retval = set_cpus_allowed(p, new_mask);
3833
3834 out_unlock:
3835         put_task_struct(p);
3836         unlock_cpu_hotplug();
3837         return retval;
3838 }
3839
3840 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3841                              cpumask_t *new_mask)
3842 {
3843         if (len < sizeof(cpumask_t)) {
3844                 memset(new_mask, 0, sizeof(cpumask_t));
3845         } else if (len > sizeof(cpumask_t)) {
3846                 len = sizeof(cpumask_t);
3847         }
3848         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3849 }
3850
3851 /**
3852  * sys_sched_setaffinity - set the cpu affinity of a process
3853  * @pid: pid of the process
3854  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3855  * @user_mask_ptr: user-space pointer to the new cpu mask
3856  */
3857 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3858                                       unsigned long __user *user_mask_ptr)
3859 {
3860         cpumask_t new_mask;
3861         int retval;
3862
3863         retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3864         if (retval)
3865                 return retval;
3866
3867         return sched_setaffinity(pid, new_mask);
3868 }
3869
3870 /*
3871  * Represents all cpu's present in the system
3872  * In systems capable of hotplug, this map could dynamically grow
3873  * as new cpu's are detected in the system via any platform specific
3874  * method, such as ACPI for e.g.
3875  */
3876
3877 cpumask_t cpu_present_map;
3878 EXPORT_SYMBOL(cpu_present_map);
3879
3880 #ifndef CONFIG_SMP
3881 cpumask_t cpu_online_map = CPU_MASK_ALL;
3882 EXPORT_SYMBOL_GPL(cpu_online_map);
3883 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3884 #endif
3885
3886 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3887 {
3888         int retval;
3889         task_t *p;
3890
3891         lock_cpu_hotplug();
3892         read_lock(&tasklist_lock);
3893
3894         retval = -ESRCH;
3895         p = find_process_by_pid(pid);
3896         if (!p)
3897                 goto out_unlock;
3898
3899         retval = 0;
3900         cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
3901
3902 out_unlock:
3903         read_unlock(&tasklist_lock);
3904         unlock_cpu_hotplug();
3905         if (retval)
3906                 return retval;
3907
3908         return 0;
3909 }
3910
3911 /**
3912  * sys_sched_getaffinity - get the cpu affinity of a process
3913  * @pid: pid of the process
3914  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3915  * @user_mask_ptr: user-space pointer to hold the current cpu mask
3916  */
3917 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3918                                       unsigned long __user *user_mask_ptr)
3919 {
3920         int ret;
3921         cpumask_t mask;
3922
3923         if (len < sizeof(cpumask_t))
3924                 return -EINVAL;
3925
3926         ret = sched_getaffinity(pid, &mask);
3927         if (ret < 0)
3928                 return ret;
3929
3930         if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3931                 return -EFAULT;
3932
3933         return sizeof(cpumask_t);
3934 }
3935
3936 /**
3937  * sys_sched_yield - yield the current processor to other threads.
3938  *
3939  * this function yields the current CPU by moving the calling thread
3940  * to the expired array. If there are no other threads running on this
3941  * CPU then this function will return.
3942  */
3943 asmlinkage long sys_sched_yield(void)
3944 {
3945         runqueue_t *rq = this_rq_lock();
3946         prio_array_t *array = current->array;
3947         prio_array_t *target = rq->expired;
3948
3949         schedstat_inc(rq, yld_cnt);
3950         /*
3951          * We implement yielding by moving the task into the expired
3952          * queue.
3953          *
3954          * (special rule: RT tasks will just roundrobin in the active
3955          *  array.)
3956          */
3957         if (rt_task(current))
3958                 target = rq->active;
3959
3960         if (array->nr_active == 1) {
3961                 schedstat_inc(rq, yld_act_empty);
3962                 if (!rq->expired->nr_active)
3963                         schedstat_inc(rq, yld_both_empty);
3964         } else if (!rq->expired->nr_active)
3965                 schedstat_inc(rq, yld_exp_empty);
3966
3967         if (array != target) {
3968                 dequeue_task(current, array);
3969                 enqueue_task(current, target);
3970         } else
3971                 /*
3972                  * requeue_task is cheaper so perform that if possible.
3973                  */
3974                 requeue_task(current, array);
3975
3976         /*
3977          * Since we are going to call schedule() anyway, there's
3978          * no need to preempt or enable interrupts:
3979          */
3980         __release(rq->lock);
3981         _raw_spin_unlock(&rq->lock);
3982         preempt_enable_no_resched();
3983
3984         schedule();
3985
3986         return 0;
3987 }
3988
3989 static inline void __cond_resched(void)
3990 {
3991         /*
3992          * The BKS might be reacquired before we have dropped
3993          * PREEMPT_ACTIVE, which could trigger a second
3994          * cond_resched() call.
3995          */
3996         if (unlikely(preempt_count()))
3997                 return;
3998         do {
3999                 add_preempt_count(PREEMPT_ACTIVE);
4000                 schedule();
4001                 sub_preempt_count(PREEMPT_ACTIVE);
4002         } while (need_resched());
4003 }
4004
4005 int __sched cond_resched(void)
4006 {
4007         if (need_resched()) {
4008                 __cond_resched();
4009                 return 1;
4010         }
4011         return 0;
4012 }
4013
4014 EXPORT_SYMBOL(cond_resched);
4015
4016 /*
4017  * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4018  * call schedule, and on return reacquire the lock.
4019  *
4020  * This works OK both with and without CONFIG_PREEMPT.  We do strange low-level
4021  * operations here to prevent schedule() from being called twice (once via
4022  * spin_unlock(), once by hand).
4023  */
4024 int cond_resched_lock(spinlock_t *lock)
4025 {
4026         int ret = 0;
4027
4028         if (need_lockbreak(lock)) {
4029                 spin_unlock(lock);
4030                 cpu_relax();
4031                 ret = 1;
4032                 spin_lock(lock);
4033         }
4034         if (need_resched()) {
4035                 _raw_spin_unlock(lock);
4036                 preempt_enable_no_resched();
4037                 __cond_resched();
4038                 ret = 1;
4039                 spin_lock(lock);
4040         }
4041         return ret;
4042 }
4043
4044 EXPORT_SYMBOL(cond_resched_lock);
4045
4046 int __sched cond_resched_softirq(void)
4047 {
4048         BUG_ON(!in_softirq());
4049
4050         if (need_resched()) {
4051                 __local_bh_enable();
4052                 __cond_resched();
4053                 local_bh_disable();
4054                 return 1;
4055         }
4056         return 0;
4057 }
4058
4059 EXPORT_SYMBOL(cond_resched_softirq);
4060
4061
4062 /**
4063  * yield - yield the current processor to other threads.
4064  *
4065  * this is a shortcut for kernel-space yielding - it marks the
4066  * thread runnable and calls sys_sched_yield().
4067  */
4068 void __sched yield(void)
4069 {
4070         set_current_state(TASK_RUNNING);
4071         sys_sched_yield();
4072 }
4073
4074 EXPORT_SYMBOL(yield);
4075
4076 /*
4077  * This task is about to go to sleep on IO.  Increment rq->nr_iowait so
4078  * that process accounting knows that this is a task in IO wait state.
4079  *
4080  * But don't do that if it is a deliberate, throttling IO wait (this task
4081  * has set its backing_dev_info: the queue against which it should throttle)
4082  */
4083 void __sched io_schedule(void)
4084 {
4085         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4086
4087         atomic_inc(&rq->nr_iowait);
4088         schedule();
4089         atomic_dec(&rq->nr_iowait);
4090 }
4091
4092 EXPORT_SYMBOL(io_schedule);
4093
4094 long __sched io_schedule_timeout(long timeout)
4095 {
4096         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4097         long ret;
4098
4099         atomic_inc(&rq->nr_iowait);
4100         ret = schedule_timeout(timeout);
4101         atomic_dec(&rq->nr_iowait);
4102         return ret;
4103 }
4104
4105 /**
4106  * sys_sched_get_priority_max - return maximum RT priority.
4107  * @policy: scheduling class.
4108  *
4109  * this syscall returns the maximum rt_priority that can be used
4110  * by a given scheduling class.
4111  */
4112 asmlinkage long sys_sched_get_priority_max(int policy)
4113 {
4114         int ret = -EINVAL;
4115
4116         switch (policy) {
4117         case SCHED_FIFO:
4118         case SCHED_RR:
4119                 ret = MAX_USER_RT_PRIO-1;
4120                 break;
4121         case SCHED_NORMAL:
4122                 ret = 0;
4123                 break;
4124         }
4125         return ret;
4126 }
4127
4128 /**
4129  * sys_sched_get_priority_min - return minimum RT priority.
4130  * @policy: scheduling class.
4131  *
4132  * this syscall returns the minimum rt_priority that can be used
4133  * by a given scheduling class.
4134  */
4135 asmlinkage long sys_sched_get_priority_min(int policy)
4136 {
4137         int ret = -EINVAL;
4138
4139         switch (policy) {
4140         case SCHED_FIFO:
4141         case SCHED_RR:
4142                 ret = 1;
4143                 break;
4144         case SCHED_NORMAL:
4145                 ret = 0;
4146         }
4147         return ret;
4148 }
4149
4150 /**
4151  * sys_sched_rr_get_interval - return the default timeslice of a process.
4152  * @pid: pid of the process.
4153  * @interval: userspace pointer to the timeslice value.
4154  *
4155  * this syscall writes the default timeslice value of a given process
4156  * into the user-space timespec buffer. A value of '0' means infinity.
4157  */
4158 asmlinkage
4159 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4160 {
4161         int retval = -EINVAL;
4162         struct timespec t;
4163         task_t *p;
4164
4165         if (pid < 0)
4166                 goto out_nounlock;
4167
4168         retval = -ESRCH;
4169         read_lock(&tasklist_lock);
4170         p = find_process_by_pid(pid);
4171         if (!p)
4172                 goto out_unlock;
4173
4174         retval = security_task_getscheduler(p);
4175         if (retval)
4176                 goto out_unlock;
4177
4178         jiffies_to_timespec(p->policy & SCHED_FIFO ?
4179                                 0 : task_timeslice(p), &t);
4180         read_unlock(&tasklist_lock);
4181         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4182 out_nounlock:
4183         return retval;
4184 out_unlock:
4185         read_unlock(&tasklist_lock);
4186         return retval;
4187 }
4188
4189 static inline struct task_struct *eldest_child(struct task_struct *p)
4190 {
4191         if (list_empty(&p->children)) return NULL;
4192         return list_entry(p->children.next,struct task_struct,sibling);
4193 }
4194
4195 static inline struct task_struct *older_sibling(struct task_struct *p)
4196 {
4197         if (p->sibling.prev==&p->parent->children) return NULL;
4198         return list_entry(p->sibling.prev,struct task_struct,sibling);
4199 }
4200
4201 static inline struct task_struct *younger_sibling(struct task_struct *p)
4202 {
4203         if (p->sibling.next==&p->parent->children) return NULL;
4204         return list_entry(p->sibling.next,struct task_struct,sibling);
4205 }
4206
4207 static void show_task(task_t *p)
4208 {
4209         task_t *relative;
4210         unsigned state;
4211         unsigned long free = 0;
4212         static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4213
4214         printk("%-13.13s ", p->comm);
4215         state = p->state ? __ffs(p->state) + 1 : 0;
4216         if (state < ARRAY_SIZE(stat_nam))
4217                 printk(stat_nam[state]);
4218         else
4219                 printk("?");
4220 #if (BITS_PER_LONG == 32)
4221         if (state == TASK_RUNNING)
4222                 printk(" running ");
4223         else
4224                 printk(" %08lX ", thread_saved_pc(p));
4225 #else
4226         if (state == TASK_RUNNING)
4227                 printk("  running task   ");
4228         else
4229                 printk(" %016lx ", thread_saved_pc(p));
4230 #endif
4231 #ifdef CONFIG_DEBUG_STACK_USAGE
4232         {
4233                 unsigned long *n = (unsigned long *) (p->thread_info+1);
4234                 while (!*n)
4235                         n++;
4236                 free = (unsigned long) n - (unsigned long)(p->thread_info+1);
4237         }
4238 #endif
4239         printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4240         if ((relative = eldest_child(p)))
4241                 printk("%5d ", relative->pid);
4242         else
4243                 printk("      ");
4244         if ((relative = younger_sibling(p)))
4245                 printk("%7d", relative->pid);
4246         else
4247                 printk("       ");
4248         if ((relative = older_sibling(p)))
4249                 printk(" %5d", relative->pid);
4250         else
4251                 printk("      ");
4252         if (!p->mm)
4253                 printk(" (L-TLB)\n");
4254         else
4255                 printk(" (NOTLB)\n");
4256
4257         if (state != TASK_RUNNING)
4258                 show_stack(p, NULL);
4259 }
4260
4261 void show_state(void)
4262 {
4263         task_t *g, *p;
4264
4265 #if (BITS_PER_LONG == 32)
4266         printk("\n"
4267                "                                               sibling\n");
4268         printk("  task             PC      pid father child younger older\n");
4269 #else
4270         printk("\n"
4271                "                                                       sibling\n");
4272         printk("  task                 PC          pid father child younger older\n");
4273 #endif
4274         read_lock(&tasklist_lock);
4275         do_each_thread(g, p) {
4276                 /*
4277                  * reset the NMI-timeout, listing all files on a slow
4278                  * console might take alot of time:
4279                  */
4280                 touch_nmi_watchdog();
4281                 show_task(p);
4282         } while_each_thread(g, p);
4283
4284         read_unlock(&tasklist_lock);
4285 }
4286
4287 /**
4288  * init_idle - set up an idle thread for a given CPU
4289  * @idle: task in question
4290  * @cpu: cpu the idle task belongs to
4291  *
4292  * NOTE: this function does not set the idle thread's NEED_RESCHED
4293  * flag, to make booting more robust.
4294  */
4295 void __devinit init_idle(task_t *idle, int cpu)
4296 {
4297         runqueue_t *rq = cpu_rq(cpu);
4298         unsigned long flags;
4299
4300         idle->sleep_avg = 0;
4301         idle->array = NULL;
4302         idle->prio = MAX_PRIO;
4303         idle->state = TASK_RUNNING;
4304         idle->cpus_allowed = cpumask_of_cpu(cpu);
4305         set_task_cpu(idle, cpu);
4306
4307         spin_lock_irqsave(&rq->lock, flags);
4308         rq->curr = rq->idle = idle;
4309 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4310         idle->oncpu = 1;
4311 #endif
4312         spin_unlock_irqrestore(&rq->lock, flags);
4313
4314         /* Set the preempt count _outside_ the spinlocks! */
4315 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4316         idle->thread_info->preempt_count = (idle->lock_depth >= 0);
4317 #else
4318         idle->thread_info->preempt_count = 0;
4319 #endif
4320 }
4321
4322 /*
4323  * In a system that switches off the HZ timer nohz_cpu_mask
4324  * indicates which cpus entered this state. This is used
4325  * in the rcu update to wait only for active cpus. For system
4326  * which do not switch off the HZ timer nohz_cpu_mask should
4327  * always be CPU_MASK_NONE.
4328  */
4329 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4330
4331 #ifdef CONFIG_SMP
4332 /*
4333  * This is how migration works:
4334  *
4335  * 1) we queue a migration_req_t structure in the source CPU's
4336  *    runqueue and wake up that CPU's migration thread.
4337  * 2) we down() the locked semaphore => thread blocks.
4338  * 3) migration thread wakes up (implicitly it forces the migrated
4339  *    thread off the CPU)
4340  * 4) it gets the migration request and checks whether the migrated
4341  *    task is still in the wrong runqueue.
4342  * 5) if it's in the wrong runqueue then the migration thread removes
4343  *    it and puts it into the right queue.
4344  * 6) migration thread up()s the semaphore.
4345  * 7) we wake up and the migration is done.
4346  */
4347
4348 /*
4349  * Change a given task's CPU affinity. Migrate the thread to a
4350  * proper CPU and schedule it away if the CPU it's executing on
4351  * is removed from the allowed bitmask.
4352  *
4353  * NOTE: the caller must have a valid reference to the task, the
4354  * task must not exit() & deallocate itself prematurely.  The
4355  * call is not atomic; no spinlocks may be held.
4356  */
4357 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4358 {
4359         unsigned long flags;
4360         int ret = 0;
4361         migration_req_t req;
4362         runqueue_t *rq;
4363
4364         rq = task_rq_lock(p, &flags);
4365         if (!cpus_intersects(new_mask, cpu_online_map)) {
4366                 ret = -EINVAL;
4367                 goto out;
4368         }
4369
4370         p->cpus_allowed = new_mask;
4371         /* Can the task run on the task's current CPU? If so, we're done */
4372         if (cpu_isset(task_cpu(p), new_mask))
4373                 goto out;
4374
4375         if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4376                 /* Need help from migration thread: drop lock and wait. */
4377                 task_rq_unlock(rq, &flags);
4378                 wake_up_process(rq->migration_thread);
4379                 wait_for_completion(&req.done);
4380                 tlb_migrate_finish(p->mm);
4381                 return 0;
4382         }
4383 out:
4384         task_rq_unlock(rq, &flags);
4385         return ret;
4386 }
4387
4388 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4389
4390 /*
4391  * Move (not current) task off this cpu, onto dest cpu.  We're doing
4392  * this because either it can't run here any more (set_cpus_allowed()
4393  * away from this CPU, or CPU going down), or because we're
4394  * attempting to rebalance this task on exec (sched_exec).
4395  *
4396  * So we race with normal scheduler movements, but that's OK, as long
4397  * as the task is no longer on this CPU.
4398  */
4399 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4400 {
4401         runqueue_t *rq_dest, *rq_src;
4402
4403         if (unlikely(cpu_is_offline(dest_cpu)))
4404                 return;
4405
4406         rq_src = cpu_rq(src_cpu);
4407         rq_dest = cpu_rq(dest_cpu);
4408
4409         double_rq_lock(rq_src, rq_dest);
4410         /* Already moved. */
4411         if (task_cpu(p) != src_cpu)
4412                 goto out;
4413         /* Affinity changed (again). */
4414         if (!cpu_isset(dest_cpu, p->cpus_allowed))
4415                 goto out;
4416
4417         set_task_cpu(p, dest_cpu);
4418         if (p->array) {
4419                 /*
4420                  * Sync timestamp with rq_dest's before activating.
4421                  * The same thing could be achieved by doing this step
4422                  * afterwards, and pretending it was a local activate.
4423                  * This way is cleaner and logically correct.
4424                  */
4425                 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4426                                 + rq_dest->timestamp_last_tick;
4427                 deactivate_task(p, rq_src);
4428                 activate_task(p, rq_dest, 0);
4429                 if (TASK_PREEMPTS_CURR(p, rq_dest))
4430                         resched_task(rq_dest->curr);
4431         }
4432
4433 out:
4434         double_rq_unlock(rq_src, rq_dest);
4435 }
4436
4437 /*
4438  * migration_thread - this is a highprio system thread that performs
4439  * thread migration by bumping thread off CPU then 'pushing' onto
4440  * another runqueue.
4441  */
4442 static int migration_thread(void *data)
4443 {
4444         runqueue_t *rq;
4445         int cpu = (long)data;
4446
4447         rq = cpu_rq(cpu);
4448         BUG_ON(rq->migration_thread != current);
4449
4450         set_current_state(TASK_INTERRUPTIBLE);
4451         while (!kthread_should_stop()) {
4452                 struct list_head *head;
4453                 migration_req_t *req;
4454
4455                 try_to_freeze();
4456
4457                 spin_lock_irq(&rq->lock);
4458
4459                 if (cpu_is_offline(cpu)) {
4460                         spin_unlock_irq(&rq->lock);
4461                         goto wait_to_die;
4462                 }
4463
4464                 if (rq->active_balance) {
4465                         active_load_balance(rq, cpu);
4466                         rq->active_balance = 0;
4467                 }
4468
4469                 head = &rq->migration_queue;
4470
4471                 if (list_empty(head)) {
4472                         spin_unlock_irq(&rq->lock);
4473                         schedule();
4474                         set_current_state(TASK_INTERRUPTIBLE);
4475                         continue;
4476                 }
4477                 req = list_entry(head->next, migration_req_t, list);
4478                 list_del_init(head->next);
4479
4480                 spin_unlock(&rq->lock);
4481                 __migrate_task(req->task, cpu, req->dest_cpu);
4482                 local_irq_enable();
4483
4484                 complete(&req->done);
4485         }
4486         __set_current_state(TASK_RUNNING);
4487         return 0;
4488
4489 wait_to_die:
4490         /* Wait for kthread_stop */
4491         set_current_state(TASK_INTERRUPTIBLE);
4492         while (!kthread_should_stop()) {
4493                 schedule();
4494                 set_current_state(TASK_INTERRUPTIBLE);
4495         }
4496         __set_current_state(TASK_RUNNING);
4497         return 0;
4498 }
4499
4500 #ifdef CONFIG_HOTPLUG_CPU
4501 /* Figure out where task on dead CPU should go, use force if neccessary. */
4502 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4503 {
4504         int dest_cpu;
4505         cpumask_t mask;
4506
4507         /* On same node? */
4508         mask = node_to_cpumask(cpu_to_node(dead_cpu));
4509         cpus_and(mask, mask, tsk->cpus_allowed);
4510         dest_cpu = any_online_cpu(mask);
4511
4512         /* On any allowed CPU? */
4513         if (dest_cpu == NR_CPUS)
4514                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4515
4516         /* No more Mr. Nice Guy. */
4517         if (dest_cpu == NR_CPUS) {
4518                 cpus_setall(tsk->cpus_allowed);
4519                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4520
4521                 /*
4522                  * Don't tell them about moving exiting tasks or
4523                  * kernel threads (both mm NULL), since they never
4524                  * leave kernel.
4525                  */
4526                 if (tsk->mm && printk_ratelimit())
4527                         printk(KERN_INFO "process %d (%s) no "
4528                                "longer affine to cpu%d\n",
4529                                tsk->pid, tsk->comm, dead_cpu);
4530         }
4531         __migrate_task(tsk, dead_cpu, dest_cpu);
4532 }
4533
4534 /*
4535  * While a dead CPU has no uninterruptible tasks queued at this point,
4536  * it might still have a nonzero ->nr_uninterruptible counter, because
4537  * for performance reasons the counter is not stricly tracking tasks to
4538  * their home CPUs. So we just add the counter to another CPU's counter,
4539  * to keep the global sum constant after CPU-down:
4540  */
4541 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4542 {
4543         runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4544         unsigned long flags;
4545
4546         local_irq_save(flags);
4547         double_rq_lock(rq_src, rq_dest);
4548         rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4549         rq_src->nr_uninterruptible = 0;
4550         double_rq_unlock(rq_src, rq_dest);
4551         local_irq_restore(flags);
4552 }
4553
4554 /* Run through task list and migrate tasks from the dead cpu. */
4555 static void migrate_live_tasks(int src_cpu)
4556 {
4557         struct task_struct *tsk, *t;
4558
4559         write_lock_irq(&tasklist_lock);
4560
4561         do_each_thread(t, tsk) {
4562                 if (tsk == current)
4563                         continue;
4564
4565                 if (task_cpu(tsk) == src_cpu)
4566                         move_task_off_dead_cpu(src_cpu, tsk);
4567         } while_each_thread(t, tsk);
4568
4569         write_unlock_irq(&tasklist_lock);
4570 }
4571
4572 /* Schedules idle task to be the next runnable task on current CPU.
4573  * It does so by boosting its priority to highest possible and adding it to
4574  * the _front_ of runqueue. Used by CPU offline code.
4575  */
4576 void sched_idle_next(void)
4577 {
4578         int cpu = smp_processor_id();
4579         runqueue_t *rq = this_rq();
4580         struct task_struct *p = rq->idle;
4581         unsigned long flags;
4582
4583         /* cpu has to be offline */
4584         BUG_ON(cpu_online(cpu));
4585
4586         /* Strictly not necessary since rest of the CPUs are stopped by now
4587          * and interrupts disabled on current cpu.
4588          */
4589         spin_lock_irqsave(&rq->lock, flags);
4590
4591         __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4592         /* Add idle task to _front_ of it's priority queue */
4593         __activate_idle_task(p, rq);
4594
4595         spin_unlock_irqrestore(&rq->lock, flags);
4596 }
4597
4598 /* Ensures that the idle task is using init_mm right before its cpu goes
4599  * offline.
4600  */
4601 void idle_task_exit(void)
4602 {
4603         struct mm_struct *mm = current->active_mm;
4604
4605         BUG_ON(cpu_online(smp_processor_id()));
4606
4607         if (mm != &init_mm)
4608                 switch_mm(mm, &init_mm, current);
4609         mmdrop(mm);
4610 }
4611
4612 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4613 {
4614         struct runqueue *rq = cpu_rq(dead_cpu);
4615
4616         /* Must be exiting, otherwise would be on tasklist. */
4617         BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4618
4619         /* Cannot have done final schedule yet: would have vanished. */
4620         BUG_ON(tsk->flags & PF_DEAD);
4621
4622         get_task_struct(tsk);
4623
4624         /*
4625          * Drop lock around migration; if someone else moves it,
4626          * that's OK.  No task can be added to this CPU, so iteration is
4627          * fine.
4628          */
4629         spin_unlock_irq(&rq->lock);
4630         move_task_off_dead_cpu(dead_cpu, tsk);
4631         spin_lock_irq(&rq->lock);
4632
4633         put_task_struct(tsk);
4634 }
4635
4636 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4637 static void migrate_dead_tasks(unsigned int dead_cpu)
4638 {
4639         unsigned arr, i;
4640         struct runqueue *rq = cpu_rq(dead_cpu);
4641
4642         for (arr = 0; arr < 2; arr++) {
4643                 for (i = 0; i < MAX_PRIO; i++) {
4644                         struct list_head *list = &rq->arrays[arr].queue[i];
4645                         while (!list_empty(list))
4646                                 migrate_dead(dead_cpu,
4647                                              list_entry(list->next, task_t,
4648                                                         run_list));
4649                 }
4650         }
4651 }
4652 #endif /* CONFIG_HOTPLUG_CPU */
4653
4654 /*
4655  * migration_call - callback that gets triggered when a CPU is added.
4656  * Here we can start up the necessary migration thread for the new CPU.
4657  */
4658 static int migration_call(struct notifier_block *nfb, unsigned long action,
4659                           void *hcpu)
4660 {
4661         int cpu = (long)hcpu;
4662         struct task_struct *p;
4663         struct runqueue *rq;
4664         unsigned long flags;
4665
4666         switch (action) {
4667         case CPU_UP_PREPARE:
4668                 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4669                 if (IS_ERR(p))
4670                         return NOTIFY_BAD;
4671                 p->flags |= PF_NOFREEZE;
4672                 kthread_bind(p, cpu);
4673                 /* Must be high prio: stop_machine expects to yield to it. */
4674                 rq = task_rq_lock(p, &flags);
4675                 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4676                 task_rq_unlock(rq, &flags);
4677                 cpu_rq(cpu)->migration_thread = p;
4678                 break;
4679         case CPU_ONLINE:
4680                 /* Strictly unneccessary, as first user will wake it. */
4681                 wake_up_process(cpu_rq(cpu)->migration_thread);
4682                 break;
4683 #ifdef CONFIG_HOTPLUG_CPU
4684         case CPU_UP_CANCELED:
4685                 /* Unbind it from offline cpu so it can run.  Fall thru. */
4686                 kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id());
4687                 kthread_stop(cpu_rq(cpu)->migration_thread);
4688                 cpu_rq(cpu)->migration_thread = NULL;
4689                 break;
4690         case CPU_DEAD:
4691                 migrate_live_tasks(cpu);
4692                 rq = cpu_rq(cpu);
4693                 kthread_stop(rq->migration_thread);
4694                 rq->migration_thread = NULL;
4695                 /* Idle task back to normal (off runqueue, low prio) */
4696                 rq = task_rq_lock(rq->idle, &flags);
4697                 deactivate_task(rq->idle, rq);
4698                 rq->idle->static_prio = MAX_PRIO;
4699                 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4700                 migrate_dead_tasks(cpu);
4701                 task_rq_unlock(rq, &flags);
4702                 migrate_nr_uninterruptible(rq);
4703                 BUG_ON(rq->nr_running != 0);
4704
4705                 /* No need to migrate the tasks: it was best-effort if
4706                  * they didn't do lock_cpu_hotplug().  Just wake up
4707                  * the requestors. */
4708                 spin_lock_irq(&rq->lock);
4709                 while (!list_empty(&rq->migration_queue)) {
4710                         migration_req_t *req;
4711                         req = list_entry(rq->migration_queue.next,
4712                                          migration_req_t, list);
4713                         list_del_init(&req->list);
4714                         complete(&req->done);
4715                 }
4716                 spin_unlock_irq(&rq->lock);
4717                 break;
4718 #endif
4719         }
4720         return NOTIFY_OK;
4721 }
4722
4723 /* Register at highest priority so that task migration (migrate_all_tasks)
4724  * happens before everything else.
4725  */
4726 static struct notifier_block __devinitdata migration_notifier = {
4727         .notifier_call = migration_call,
4728         .priority = 10
4729 };
4730
4731 int __init migration_init(void)
4732 {
4733         void *cpu = (void *)(long)smp_processor_id();
4734         /* Start one for boot CPU. */
4735         migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4736         migration_call(&migration_notifier, CPU_ONLINE, cpu);
4737         register_cpu_notifier(&migration_notifier);
4738         return 0;
4739 }
4740 #endif
4741
4742 #ifdef CONFIG_SMP
4743 #undef SCHED_DOMAIN_DEBUG
4744 #ifdef SCHED_DOMAIN_DEBUG
4745 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4746 {
4747         int level = 0;
4748
4749         if (!sd) {
4750                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4751                 return;
4752         }
4753
4754         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4755
4756         do {
4757                 int i;
4758                 char str[NR_CPUS];
4759                 struct sched_group *group = sd->groups;
4760                 cpumask_t groupmask;
4761
4762                 cpumask_scnprintf(str, NR_CPUS, sd->span);
4763                 cpus_clear(groupmask);
4764
4765                 printk(KERN_DEBUG);
4766                 for (i = 0; i < level + 1; i++)
4767                         printk(" ");
4768                 printk("domain %d: ", level);
4769
4770                 if (!(sd->flags & SD_LOAD_BALANCE)) {
4771                         printk("does not load-balance\n");
4772                         if (sd->parent)
4773                                 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4774                         break;
4775                 }
4776
4777                 printk("span %s\n", str);
4778
4779                 if (!cpu_isset(cpu, sd->span))
4780                         printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4781                 if (!cpu_isset(cpu, group->cpumask))
4782                         printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4783
4784                 printk(KERN_DEBUG);
4785                 for (i = 0; i < level + 2; i++)
4786                         printk(" ");
4787                 printk("groups:");
4788                 do {
4789                         if (!group) {
4790                                 printk("\n");
4791                                 printk(KERN_ERR "ERROR: group is NULL\n");
4792                                 break;
4793                         }
4794
4795                         if (!group->cpu_power) {
4796                                 printk("\n");
4797                                 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4798                         }
4799
4800                         if (!cpus_weight(group->cpumask)) {
4801                                 printk("\n");
4802                                 printk(KERN_ERR "ERROR: empty group\n");
4803                         }
4804
4805                         if (cpus_intersects(groupmask, group->cpumask)) {
4806                                 printk("\n");
4807                                 printk(KERN_ERR "ERROR: repeated CPUs\n");
4808                         }
4809
4810                         cpus_or(groupmask, groupmask, group->cpumask);
4811
4812                         cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4813                         printk(" %s", str);
4814
4815                         group = group->next;
4816                 } while (group != sd->groups);
4817                 printk("\n");
4818
4819                 if (!cpus_equal(sd->span, groupmask))
4820                         printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4821
4822                 level++;
4823                 sd = sd->parent;
4824
4825                 if (sd) {
4826                         if (!cpus_subset(groupmask, sd->span))
4827                                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4828                 }
4829
4830         } while (sd);
4831 }
4832 #else
4833 #define sched_domain_debug(sd, cpu) {}
4834 #endif
4835
4836 static int sd_degenerate(struct sched_domain *sd)
4837 {
4838         if (cpus_weight(sd->span) == 1)
4839                 return 1;
4840
4841         /* Following flags need at least 2 groups */
4842         if (sd->flags & (SD_LOAD_BALANCE |
4843                          SD_BALANCE_NEWIDLE |
4844                          SD_BALANCE_FORK |
4845                          SD_BALANCE_EXEC)) {
4846                 if (sd->groups != sd->groups->next)
4847                         return 0;
4848         }
4849
4850         /* Following flags don't use groups */
4851         if (sd->flags & (SD_WAKE_IDLE |
4852                          SD_WAKE_AFFINE |
4853                          SD_WAKE_BALANCE))
4854                 return 0;
4855
4856         return 1;
4857 }
4858
4859 static int sd_parent_degenerate(struct sched_domain *sd,
4860                                                 struct sched_domain *parent)
4861 {
4862         unsigned long cflags = sd->flags, pflags = parent->flags;
4863
4864         if (sd_degenerate(parent))
4865                 return 1;
4866
4867         if (!cpus_equal(sd->span, parent->span))
4868                 return 0;
4869
4870         /* Does parent contain flags not in child? */
4871         /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4872         if (cflags & SD_WAKE_AFFINE)
4873                 pflags &= ~SD_WAKE_BALANCE;
4874         /* Flags needing groups don't count if only 1 group in parent */
4875         if (parent->groups == parent->groups->next) {
4876                 pflags &= ~(SD_LOAD_BALANCE |
4877                                 SD_BALANCE_NEWIDLE |
4878                                 SD_BALANCE_FORK |
4879                                 SD_BALANCE_EXEC);
4880         }
4881         if (~cflags & pflags)
4882                 return 0;
4883
4884         return 1;
4885 }
4886
4887 /*
4888  * Attach the domain 'sd' to 'cpu' as its base domain.  Callers must
4889  * hold the hotplug lock.
4890  */
4891 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4892 {
4893         runqueue_t *rq = cpu_rq(cpu);
4894         struct sched_domain *tmp;
4895
4896         /* Remove the sched domains which do not contribute to scheduling. */
4897         for (tmp = sd; tmp; tmp = tmp->parent) {
4898                 struct sched_domain *parent = tmp->parent;
4899                 if (!parent)
4900                         break;
4901                 if (sd_parent_degenerate(tmp, parent))
4902                         tmp->parent = parent->parent;
4903         }
4904
4905         if (sd && sd_degenerate(sd))
4906                 sd = sd->parent;
4907
4908         sched_domain_debug(sd, cpu);
4909
4910         rcu_assign_pointer(rq->sd, sd);
4911 }
4912
4913 /* cpus with isolated domains */
4914 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4915
4916 /* Setup the mask of cpus configured for isolated domains */
4917 static int __init isolated_cpu_setup(char *str)
4918 {
4919         int ints[NR_CPUS], i;
4920
4921         str = get_options(str, ARRAY_SIZE(ints), ints);
4922         cpus_clear(cpu_isolated_map);
4923         for (i = 1; i <= ints[0]; i++)
4924                 if (ints[i] < NR_CPUS)
4925                         cpu_set(ints[i], cpu_isolated_map);
4926         return 1;
4927 }
4928
4929 __setup ("isolcpus=", isolated_cpu_setup);
4930
4931 /*
4932  * init_sched_build_groups takes an array of groups, the cpumask we wish
4933  * to span, and a pointer to a function which identifies what group a CPU
4934  * belongs to. The return value of group_fn must be a valid index into the
4935  * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4936  * keep track of groups covered with a cpumask_t).
4937  *
4938  * init_sched_build_groups will build a circular linked list of the groups
4939  * covered by the given span, and will set each group's ->cpumask correctly,
4940  * and ->cpu_power to 0.
4941  */
4942 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
4943                                     int (*group_fn)(int cpu))
4944 {
4945         struct sched_group *first = NULL, *last = NULL;
4946         cpumask_t covered = CPU_MASK_NONE;
4947         int i;
4948
4949         for_each_cpu_mask(i, span) {
4950                 int group = group_fn(i);
4951                 struct sched_group *sg = &groups[group];
4952                 int j;
4953
4954                 if (cpu_isset(i, covered))
4955                         continue;
4956
4957                 sg->cpumask = CPU_MASK_NONE;
4958                 sg->cpu_power = 0;
4959
4960                 for_each_cpu_mask(j, span) {
4961                         if (group_fn(j) != group)
4962                                 continue;
4963
4964                         cpu_set(j, covered);
4965                         cpu_set(j, sg->cpumask);
4966                 }
4967                 if (!first)
4968                         first = sg;
4969                 if (last)
4970                         last->next = sg;
4971                 last = sg;
4972         }
4973         last->next = first;
4974 }
4975
4976 #define SD_NODES_PER_DOMAIN 16
4977
4978 #ifdef CONFIG_NUMA
4979 /**
4980  * find_next_best_node - find the next node to include in a sched_domain
4981  * @node: node whose sched_domain we're building
4982  * @used_nodes: nodes already in the sched_domain
4983  *
4984  * Find the next node to include in a given scheduling domain.  Simply
4985  * finds the closest node not already in the @used_nodes map.
4986  *
4987  * Should use nodemask_t.
4988  */
4989 static int find_next_best_node(int node, unsigned long *used_nodes)
4990 {
4991         int i, n, val, min_val, best_node = 0;
4992
4993         min_val = INT_MAX;
4994
4995         for (i = 0; i < MAX_NUMNODES; i++) {
4996                 /* Start at @node */
4997                 n = (node + i) % MAX_NUMNODES;
4998
4999                 if (!nr_cpus_node(n))
5000                         continue;
5001
5002                 /* Skip already used nodes */
5003                 if (test_bit(n, used_nodes))
5004                         continue;
5005
5006                 /* Simple min distance search */
5007                 val = node_distance(node, n);
5008
5009                 if (val < min_val) {
5010                         min_val = val;
5011                         best_node = n;
5012                 }
5013         }
5014
5015         set_bit(best_node, used_nodes);
5016         return best_node;
5017 }
5018
5019 /**
5020  * sched_domain_node_span - get a cpumask for a node's sched_domain
5021  * @node: node whose cpumask we're constructing
5022  * @size: number of nodes to include in this span
5023  *
5024  * Given a node, construct a good cpumask for its sched_domain to span.  It
5025  * should be one that prevents unnecessary balancing, but also spreads tasks
5026  * out optimally.
5027  */
5028 static cpumask_t sched_domain_node_span(int node)
5029 {
5030         int i;
5031         cpumask_t span, nodemask;
5032         DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5033
5034         cpus_clear(span);
5035         bitmap_zero(used_nodes, MAX_NUMNODES);
5036
5037         nodemask = node_to_cpumask(node);
5038         cpus_or(span, span, nodemask);
5039         set_bit(node, used_nodes);
5040
5041         for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5042                 int next_node = find_next_best_node(node, used_nodes);
5043                 nodemask = node_to_cpumask(next_node);
5044                 cpus_or(span, span, nodemask);
5045         }
5046
5047         return span;
5048 }
5049 #endif
5050
5051 /*
5052  * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5053  * can switch it on easily if needed.
5054  */
5055 #ifdef CONFIG_SCHED_SMT
5056 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5057 static struct sched_group sched_group_cpus[NR_CPUS];
5058 static int cpu_to_cpu_group(int cpu)
5059 {
5060         return cpu;
5061 }
5062 #endif
5063
5064 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5065 static struct sched_group sched_group_phys[NR_CPUS];
5066 static int cpu_to_phys_group(int cpu)
5067 {
5068 #ifdef CONFIG_SCHED_SMT
5069         return first_cpu(cpu_sibling_map[cpu]);
5070 #else
5071         return cpu;
5072 #endif
5073 }
5074
5075 #ifdef CONFIG_NUMA
5076 /*
5077  * The init_sched_build_groups can't handle what we want to do with node
5078  * groups, so roll our own. Now each node has its own list of groups which
5079  * gets dynamically allocated.
5080  */
5081 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5082 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5083
5084 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5085 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5086
5087 static int cpu_to_allnodes_group(int cpu)
5088 {
5089         return cpu_to_node(cpu);
5090 }
5091 #endif
5092
5093 /*
5094  * Build sched domains for a given set of cpus and attach the sched domains
5095  * to the individual cpus
5096  */
5097 void build_sched_domains(const cpumask_t *cpu_map)
5098 {
5099         int i;
5100 #ifdef CONFIG_NUMA
5101         struct sched_group **sched_group_nodes = NULL;
5102         struct sched_group *sched_group_allnodes = NULL;
5103
5104         /*
5105          * Allocate the per-node list of sched groups
5106          */
5107         sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5108                                            GFP_ATOMIC);
5109         if (!sched_group_nodes) {
5110                 printk(KERN_WARNING "Can not alloc sched group node list\n");
5111                 return;
5112         }
5113         sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5114 #endif
5115
5116         /*
5117          * Set up domains for cpus specified by the cpu_map.
5118          */
5119         for_each_cpu_mask(i, *cpu_map) {
5120                 int group;
5121                 struct sched_domain *sd = NULL, *p;
5122                 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5123
5124                 cpus_and(nodemask, nodemask, *cpu_map);
5125
5126 #ifdef CONFIG_NUMA
5127                 if (cpus_weight(*cpu_map)
5128                                 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5129                         if (!sched_group_allnodes) {
5130                                 sched_group_allnodes
5131                                         = kmalloc(sizeof(struct sched_group)
5132                                                         * MAX_NUMNODES,
5133                                                   GFP_KERNEL);
5134                                 if (!sched_group_allnodes) {
5135                                         printk(KERN_WARNING
5136                                         "Can not alloc allnodes sched group\n");
5137                                         break;
5138                                 }
5139                                 sched_group_allnodes_bycpu[i]
5140                                                 = sched_group_allnodes;
5141                         }
5142                         sd = &per_cpu(allnodes_domains, i);
5143                         *sd = SD_ALLNODES_INIT;
5144                         sd->span = *cpu_map;
5145                         group = cpu_to_allnodes_group(i);
5146                         sd->groups = &sched_group_allnodes[group];
5147                         p = sd;
5148                 } else
5149                         p = NULL;
5150
5151                 sd = &per_cpu(node_domains, i);
5152                 *sd = SD_NODE_INIT;
5153                 sd->span = sched_domain_node_span(cpu_to_node(i));
5154                 sd->parent = p;
5155                 cpus_and(sd->span, sd->span, *cpu_map);
5156 #endif
5157
5158                 p = sd;
5159                 sd = &per_cpu(phys_domains, i);
5160                 group = cpu_to_phys_group(i);
5161                 *sd = SD_CPU_INIT;
5162                 sd->span = nodemask;
5163                 sd->parent = p;
5164                 sd->groups = &sched_group_phys[group];
5165
5166 #ifdef CONFIG_SCHED_SMT
5167                 p = sd;
5168                 sd = &per_cpu(cpu_domains, i);
5169                 group = cpu_to_cpu_group(i);
5170                 *sd = SD_SIBLING_INIT;
5171                 sd->span = cpu_sibling_map[i];
5172                 cpus_and(sd->span, sd->span, *cpu_map);
5173                 sd->parent = p;
5174                 sd->groups = &sched_group_cpus[group];
5175 #endif
5176         }
5177
5178 #ifdef CONFIG_SCHED_SMT
5179         /* Set up CPU (sibling) groups */
5180         for_each_cpu_mask(i, *cpu_map) {
5181                 cpumask_t this_sibling_map = cpu_sibling_map[i];
5182                 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5183                 if (i != first_cpu(this_sibling_map))
5184                         continue;
5185
5186                 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5187                                                 &cpu_to_cpu_group);
5188         }
5189 #endif
5190
5191         /* Set up physical groups */
5192         for (i = 0; i < MAX_NUMNODES; i++) {
5193                 cpumask_t nodemask = node_to_cpumask(i);
5194
5195                 cpus_and(nodemask, nodemask, *cpu_map);
5196                 if (cpus_empty(nodemask))
5197                         continue;
5198
5199                 init_sched_build_groups(sched_group_phys, nodemask,
5200                                                 &cpu_to_phys_group);
5201         }
5202
5203 #ifdef CONFIG_NUMA
5204         /* Set up node groups */
5205         if (sched_group_allnodes)
5206                 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5207                                         &cpu_to_allnodes_group);
5208
5209         for (i = 0; i < MAX_NUMNODES; i++) {
5210                 /* Set up node groups */
5211                 struct sched_group *sg, *prev;
5212                 cpumask_t nodemask = node_to_cpumask(i);
5213                 cpumask_t domainspan;
5214                 cpumask_t covered = CPU_MASK_NONE;
5215                 int j;
5216
5217                 cpus_and(nodemask, nodemask, *cpu_map);
5218                 if (cpus_empty(nodemask)) {
5219                         sched_group_nodes[i] = NULL;
5220                         continue;
5221                 }
5222
5223                 domainspan = sched_domain_node_span(i);
5224                 cpus_and(domainspan, domainspan, *cpu_map);
5225
5226                 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5227                 sched_group_nodes[i] = sg;
5228                 for_each_cpu_mask(j, nodemask) {
5229                         struct sched_domain *sd;
5230                         sd = &per_cpu(node_domains, j);
5231                         sd->groups = sg;
5232                         if (sd->groups == NULL) {
5233                                 /* Turn off balancing if we have no groups */
5234                                 sd->flags = 0;
5235                         }
5236                 }
5237                 if (!sg) {
5238                         printk(KERN_WARNING
5239                         "Can not alloc domain group for node %d\n", i);
5240                         continue;
5241                 }
5242                 sg->cpu_power = 0;
5243                 sg->cpumask = nodemask;
5244                 cpus_or(covered, covered, nodemask);
5245                 prev = sg;
5246
5247                 for (j = 0; j < MAX_NUMNODES; j++) {
5248                         cpumask_t tmp, notcovered;
5249                         int n = (i + j) % MAX_NUMNODES;
5250
5251                         cpus_complement(notcovered, covered);
5252                         cpus_and(tmp, notcovered, *cpu_map);
5253                         cpus_and(tmp, tmp, domainspan);
5254                         if (cpus_empty(tmp))
5255                                 break;
5256
5257                         nodemask = node_to_cpumask(n);
5258                         cpus_and(tmp, tmp, nodemask);
5259                         if (cpus_empty(tmp))
5260                                 continue;
5261
5262                         sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5263                         if (!sg) {
5264                                 printk(KERN_WARNING
5265                                 "Can not alloc domain group for node %d\n", j);
5266                                 break;
5267                         }
5268                         sg->cpu_power = 0;
5269                         sg->cpumask = tmp;
5270                         cpus_or(covered, covered, tmp);
5271                         prev->next = sg;
5272                         prev = sg;
5273                 }
5274                 prev->next = sched_group_nodes[i];
5275         }
5276 #endif
5277
5278         /* Calculate CPU power for physical packages and nodes */
5279         for_each_cpu_mask(i, *cpu_map) {
5280                 int power;
5281                 struct sched_domain *sd;
5282 #ifdef CONFIG_SCHED_SMT
5283                 sd = &per_cpu(cpu_domains, i);
5284                 power = SCHED_LOAD_SCALE;
5285                 sd->groups->cpu_power = power;
5286 #endif
5287
5288                 sd = &per_cpu(phys_domains, i);
5289                 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5290                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5291                 sd->groups->cpu_power = power;
5292
5293 #ifdef CONFIG_NUMA
5294                 sd = &per_cpu(allnodes_domains, i);
5295                 if (sd->groups) {
5296                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5297                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5298                         sd->groups->cpu_power = power;
5299                 }
5300 #endif
5301         }
5302
5303 #ifdef CONFIG_NUMA
5304         for (i = 0; i < MAX_NUMNODES; i++) {
5305                 struct sched_group *sg = sched_group_nodes[i];
5306                 int j;
5307
5308                 if (sg == NULL)
5309                         continue;
5310 next_sg:
5311                 for_each_cpu_mask(j, sg->cpumask) {
5312                         struct sched_domain *sd;
5313                         int power;
5314
5315                         sd = &per_cpu(phys_domains, j);
5316                         if (j != first_cpu(sd->groups->cpumask)) {
5317                                 /*
5318                                  * Only add "power" once for each
5319                                  * physical package.
5320                                  */
5321                                 continue;
5322                         }
5323                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5324                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5325
5326                         sg->cpu_power += power;
5327                 }
5328                 sg = sg->next;
5329                 if (sg != sched_group_nodes[i])
5330                         goto next_sg;
5331         }
5332 #endif
5333
5334         /* Attach the domains */
5335         for_each_cpu_mask(i, *cpu_map) {
5336                 struct sched_domain *sd;
5337 #ifdef CONFIG_SCHED_SMT
5338                 sd = &per_cpu(cpu_domains, i);
5339 #else
5340                 sd = &per_cpu(phys_domains, i);
5341 #endif
5342                 cpu_attach_domain(sd, i);
5343         }
5344 }
5345 /*
5346  * Set up scheduler domains and groups.  Callers must hold the hotplug lock.
5347  */
5348 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5349 {
5350         cpumask_t cpu_default_map;
5351
5352         /*
5353          * Setup mask for cpus without special case scheduling requirements.
5354          * For now this just excludes isolated cpus, but could be used to
5355          * exclude other special cases in the future.
5356          */
5357         cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5358
5359         build_sched_domains(&cpu_default_map);
5360 }
5361
5362 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5363 {
5364 #ifdef CONFIG_NUMA
5365         int i;
5366         int cpu;
5367
5368         for_each_cpu_mask(cpu, *cpu_map) {
5369                 struct sched_group *sched_group_allnodes
5370                         = sched_group_allnodes_bycpu[cpu];
5371                 struct sched_group **sched_group_nodes
5372                         = sched_group_nodes_bycpu[cpu];
5373
5374                 if (sched_group_allnodes) {
5375                         kfree(sched_group_allnodes);
5376                         sched_group_allnodes_bycpu[cpu] = NULL;
5377                 }
5378
5379                 if (!sched_group_nodes)
5380                         continue;
5381
5382                 for (i = 0; i < MAX_NUMNODES; i++) {
5383                         cpumask_t nodemask = node_to_cpumask(i);
5384                         struct sched_group *oldsg, *sg = sched_group_nodes[i];
5385
5386                         cpus_and(nodemask, nodemask, *cpu_map);
5387                         if (cpus_empty(nodemask))
5388                                 continue;
5389
5390                         if (sg == NULL)
5391                                 continue;
5392                         sg = sg->next;
5393 next_sg:
5394                         oldsg = sg;
5395                         sg = sg->next;
5396                         kfree(oldsg);
5397                         if (oldsg != sched_group_nodes[i])
5398                                 goto next_sg;
5399                 }
5400                 kfree(sched_group_nodes);
5401                 sched_group_nodes_bycpu[cpu] = NULL;
5402         }
5403 #endif
5404 }
5405
5406 /*
5407  * Detach sched domains from a group of cpus specified in cpu_map
5408  * These cpus will now be attached to the NULL domain
5409  */
5410 static inline void detach_destroy_domains(const cpumask_t *cpu_map)
5411 {
5412         int i;
5413
5414         for_each_cpu_mask(i, *cpu_map)
5415                 cpu_attach_domain(NULL, i);
5416         synchronize_sched();
5417         arch_destroy_sched_domains(cpu_map);
5418 }
5419
5420 /*
5421  * Partition sched domains as specified by the cpumasks below.
5422  * This attaches all cpus from the cpumasks to the NULL domain,
5423  * waits for a RCU quiescent period, recalculates sched
5424  * domain information and then attaches them back to the
5425  * correct sched domains
5426  * Call with hotplug lock held
5427  */
5428 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
5429 {
5430         cpumask_t change_map;
5431
5432         cpus_and(*partition1, *partition1, cpu_online_map);
5433         cpus_and(*partition2, *partition2, cpu_online_map);
5434         cpus_or(change_map, *partition1, *partition2);
5435
5436         /* Detach sched domains from all of the affected cpus */
5437         detach_destroy_domains(&change_map);
5438         if (!cpus_empty(*partition1))
5439                 build_sched_domains(partition1);
5440         if (!cpus_empty(*partition2))
5441                 build_sched_domains(partition2);
5442 }
5443
5444 #ifdef CONFIG_HOTPLUG_CPU
5445 /*
5446  * Force a reinitialization of the sched domains hierarchy.  The domains
5447  * and groups cannot be updated in place without racing with the balancing
5448  * code, so we temporarily attach all running cpus to the NULL domain
5449  * which will prevent rebalancing while the sched domains are recalculated.
5450  */
5451 static int update_sched_domains(struct notifier_block *nfb,
5452                                 unsigned long action, void *hcpu)
5453 {
5454         switch (action) {
5455         case CPU_UP_PREPARE:
5456         case CPU_DOWN_PREPARE:
5457                 detach_destroy_domains(&cpu_online_map);
5458                 return NOTIFY_OK;
5459
5460         case CPU_UP_CANCELED:
5461         case CPU_DOWN_FAILED:
5462         case CPU_ONLINE:
5463         case CPU_DEAD:
5464                 /*
5465                  * Fall through and re-initialise the domains.
5466                  */
5467                 break;
5468         default:
5469                 return NOTIFY_DONE;
5470         }
5471
5472         /* The hotplug lock is already held by cpu_up/cpu_down */
5473         arch_init_sched_domains(&cpu_online_map);
5474
5475         return NOTIFY_OK;
5476 }
5477 #endif
5478
5479 void __init sched_init_smp(void)
5480 {
5481         lock_cpu_hotplug();
5482         arch_init_sched_domains(&cpu_online_map);
5483         unlock_cpu_hotplug();
5484         /* XXX: Theoretical race here - CPU may be hotplugged now */
5485         hotcpu_notifier(update_sched_domains, 0);
5486 }
5487 #else
5488 void __init sched_init_smp(void)
5489 {
5490 }
5491 #endif /* CONFIG_SMP */
5492
5493 int in_sched_functions(unsigned long addr)
5494 {
5495         /* Linker adds these: start and end of __sched functions */
5496         extern char __sched_text_start[], __sched_text_end[];
5497         return in_lock_functions(addr) ||
5498                 (addr >= (unsigned long)__sched_text_start
5499                 && addr < (unsigned long)__sched_text_end);
5500 }
5501
5502 void __init sched_init(void)
5503 {
5504         runqueue_t *rq;
5505         int i, j, k;
5506
5507         for (i = 0; i < NR_CPUS; i++) {
5508                 prio_array_t *array;
5509
5510                 rq = cpu_rq(i);
5511                 spin_lock_init(&rq->lock);
5512                 rq->nr_running = 0;
5513                 rq->active = rq->arrays;
5514                 rq->expired = rq->arrays + 1;
5515                 rq->best_expired_prio = MAX_PRIO;
5516
5517 #ifdef CONFIG_SMP
5518                 rq->sd = NULL;
5519                 for (j = 1; j < 3; j++)
5520                         rq->cpu_load[j] = 0;
5521                 rq->active_balance = 0;
5522                 rq->push_cpu = 0;
5523                 rq->migration_thread = NULL;
5524                 INIT_LIST_HEAD(&rq->migration_queue);
5525 #endif
5526                 atomic_set(&rq->nr_iowait, 0);
5527
5528                 for (j = 0; j < 2; j++) {
5529                         array = rq->arrays + j;
5530                         for (k = 0; k < MAX_PRIO; k++) {
5531                                 INIT_LIST_HEAD(array->queue + k);
5532                                 __clear_bit(k, array->bitmap);
5533                         }
5534                         // delimiter for bitsearch
5535                         __set_bit(MAX_PRIO, array->bitmap);
5536                 }
5537         }
5538
5539         /*
5540          * The boot idle thread does lazy MMU switching as well:
5541          */
5542         atomic_inc(&init_mm.mm_count);
5543         enter_lazy_tlb(&init_mm, current);
5544
5545         /*
5546          * Make us the idle thread. Technically, schedule() should not be
5547          * called from this thread, however somewhere below it might be,
5548          * but because we are the idle thread, we just pick up running again
5549          * when this runqueue becomes "idle".
5550          */
5551         init_idle(current, smp_processor_id());
5552 }
5553
5554 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5555 void __might_sleep(char *file, int line)
5556 {
5557 #if defined(in_atomic)
5558         static unsigned long prev_jiffy;        /* ratelimiting */
5559
5560         if ((in_atomic() || irqs_disabled()) &&
5561             system_state == SYSTEM_RUNNING && !oops_in_progress) {
5562                 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5563                         return;
5564                 prev_jiffy = jiffies;
5565                 printk(KERN_ERR "Debug: sleeping function called from invalid"
5566                                 " context at %s:%d\n", file, line);
5567                 printk("in_atomic():%d, irqs_disabled():%d\n",
5568                         in_atomic(), irqs_disabled());
5569                 dump_stack();
5570         }
5571 #endif
5572 }
5573 EXPORT_SYMBOL(__might_sleep);
5574 #endif
5575
5576 #ifdef CONFIG_MAGIC_SYSRQ
5577 void normalize_rt_tasks(void)
5578 {
5579         struct task_struct *p;
5580         prio_array_t *array;
5581         unsigned long flags;
5582         runqueue_t *rq;
5583
5584         read_lock_irq(&tasklist_lock);
5585         for_each_process (p) {
5586                 if (!rt_task(p))
5587                         continue;
5588
5589                 rq = task_rq_lock(p, &flags);
5590
5591                 array = p->array;
5592                 if (array)
5593                         deactivate_task(p, task_rq(p));
5594                 __setscheduler(p, SCHED_NORMAL, 0);
5595                 if (array) {
5596                         __activate_task(p, task_rq(p));
5597                         resched_task(rq->curr);
5598                 }
5599
5600                 task_rq_unlock(rq, &flags);
5601         }
5602         read_unlock_irq(&tasklist_lock);
5603 }
5604
5605 #endif /* CONFIG_MAGIC_SYSRQ */
5606
5607 #ifdef CONFIG_IA64
5608 /*
5609  * These functions are only useful for the IA64 MCA handling.
5610  *
5611  * They can only be called when the whole system has been
5612  * stopped - every CPU needs to be quiescent, and no scheduling
5613  * activity can take place. Using them for anything else would
5614  * be a serious bug, and as a result, they aren't even visible
5615  * under any other configuration.
5616  */
5617
5618 /**
5619  * curr_task - return the current task for a given cpu.
5620  * @cpu: the processor in question.
5621  *
5622  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5623  */
5624 task_t *curr_task(int cpu)
5625 {
5626         return cpu_curr(cpu);
5627 }
5628
5629 /**
5630  * set_curr_task - set the current task for a given cpu.
5631  * @cpu: the processor in question.
5632  * @p: the task pointer to set.
5633  *
5634  * Description: This function must only be used when non-maskable interrupts
5635  * are serviced on a separate stack.  It allows the architecture to switch the
5636  * notion of the current task on a cpu in a non-blocking manner.  This function
5637  * must be called with all CPU's synchronized, and interrupts disabled, the
5638  * and caller must save the original value of the current task (see
5639  * curr_task() above) and restore that value before reenabling interrupts and
5640  * re-starting the system.
5641  *
5642  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5643  */
5644 void set_curr_task(int cpu, task_t *p)
5645 {
5646         cpu_curr(cpu) = p;
5647 }
5648
5649 #endif