[PATCH] m68k: introduce setup_thread_stack() and end_of_stack()
[linux-2.6.git] / kernel / sched.c
1 /*
2  *  kernel/sched.c
3  *
4  *  Kernel scheduler and related syscalls
5  *
6  *  Copyright (C) 1991-2002  Linus Torvalds
7  *
8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
9  *              make semaphores SMP safe
10  *  1998-11-19  Implemented schedule_timeout() and related stuff
11  *              by Andrea Arcangeli
12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
13  *              hybrid priority-list and round-robin design with
14  *              an array-switch method of distributing timeslices
15  *              and per-CPU runqueues.  Cleanups and useful suggestions
16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
17  *  2003-09-03  Interactivity tuning by Con Kolivas.
18  *  2004-04-02  Scheduler domains code by Nick Piggin
19  */
20
21 #include <linux/mm.h>
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/completion.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/security.h>
33 #include <linux/notifier.h>
34 #include <linux/profile.h>
35 #include <linux/suspend.h>
36 #include <linux/blkdev.h>
37 #include <linux/delay.h>
38 #include <linux/smp.h>
39 #include <linux/threads.h>
40 #include <linux/timer.h>
41 #include <linux/rcupdate.h>
42 #include <linux/cpu.h>
43 #include <linux/cpuset.h>
44 #include <linux/percpu.h>
45 #include <linux/kthread.h>
46 #include <linux/seq_file.h>
47 #include <linux/syscalls.h>
48 #include <linux/times.h>
49 #include <linux/acct.h>
50 #include <asm/tlb.h>
51
52 #include <asm/unistd.h>
53
54 /*
55  * Convert user-nice values [ -20 ... 0 ... 19 ]
56  * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
57  * and back.
58  */
59 #define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
60 #define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
61 #define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
62
63 /*
64  * 'User priority' is the nice value converted to something we
65  * can work with better when scaling various scheduler parameters,
66  * it's a [ 0 ... 39 ] range.
67  */
68 #define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
69 #define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
70 #define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
71
72 /*
73  * Some helpers for converting nanosecond timing to jiffy resolution
74  */
75 #define NS_TO_JIFFIES(TIME)     ((TIME) / (1000000000 / HZ))
76 #define JIFFIES_TO_NS(TIME)     ((TIME) * (1000000000 / HZ))
77
78 /*
79  * These are the 'tuning knobs' of the scheduler:
80  *
81  * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
82  * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
83  * Timeslices get refilled after they expire.
84  */
85 #define MIN_TIMESLICE           max(5 * HZ / 1000, 1)
86 #define DEF_TIMESLICE           (100 * HZ / 1000)
87 #define ON_RUNQUEUE_WEIGHT       30
88 #define CHILD_PENALTY            95
89 #define PARENT_PENALTY          100
90 #define EXIT_WEIGHT               3
91 #define PRIO_BONUS_RATIO         25
92 #define MAX_BONUS               (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
93 #define INTERACTIVE_DELTA         2
94 #define MAX_SLEEP_AVG           (DEF_TIMESLICE * MAX_BONUS)
95 #define STARVATION_LIMIT        (MAX_SLEEP_AVG)
96 #define NS_MAX_SLEEP_AVG        (JIFFIES_TO_NS(MAX_SLEEP_AVG))
97
98 /*
99  * If a task is 'interactive' then we reinsert it in the active
100  * array after it has expired its current timeslice. (it will not
101  * continue to run immediately, it will still roundrobin with
102  * other interactive tasks.)
103  *
104  * This part scales the interactivity limit depending on niceness.
105  *
106  * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
107  * Here are a few examples of different nice levels:
108  *
109  *  TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
110  *  TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
111  *  TASK_INTERACTIVE(  0): [1,1,1,1,0,0,0,0,0,0,0]
112  *  TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
113  *  TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
114  *
115  * (the X axis represents the possible -5 ... 0 ... +5 dynamic
116  *  priority range a task can explore, a value of '1' means the
117  *  task is rated interactive.)
118  *
119  * Ie. nice +19 tasks can never get 'interactive' enough to be
120  * reinserted into the active array. And only heavily CPU-hog nice -20
121  * tasks will be expired. Default nice 0 tasks are somewhere between,
122  * it takes some effort for them to get interactive, but it's not
123  * too hard.
124  */
125
126 #define CURRENT_BONUS(p) \
127         (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
128                 MAX_SLEEP_AVG)
129
130 #define GRANULARITY     (10 * HZ / 1000 ? : 1)
131
132 #ifdef CONFIG_SMP
133 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
134                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
135                         num_online_cpus())
136 #else
137 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
138                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
139 #endif
140
141 #define SCALE(v1,v1_max,v2_max) \
142         (v1) * (v2_max) / (v1_max)
143
144 #define DELTA(p) \
145         (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
146
147 #define TASK_INTERACTIVE(p) \
148         ((p)->prio <= (p)->static_prio - DELTA(p))
149
150 #define INTERACTIVE_SLEEP(p) \
151         (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
152                 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
153
154 #define TASK_PREEMPTS_CURR(p, rq) \
155         ((p)->prio < (rq)->curr->prio)
156
157 /*
158  * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
159  * to time slice values: [800ms ... 100ms ... 5ms]
160  *
161  * The higher a thread's priority, the bigger timeslices
162  * it gets during one round of execution. But even the lowest
163  * priority thread gets MIN_TIMESLICE worth of execution time.
164  */
165
166 #define SCALE_PRIO(x, prio) \
167         max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
168
169 static unsigned int task_timeslice(task_t *p)
170 {
171         if (p->static_prio < NICE_TO_PRIO(0))
172                 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
173         else
174                 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
175 }
176 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran)       \
177                                 < (long long) (sd)->cache_hot_time)
178
179 /*
180  * These are the runqueue data structures:
181  */
182
183 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
184
185 typedef struct runqueue runqueue_t;
186
187 struct prio_array {
188         unsigned int nr_active;
189         unsigned long bitmap[BITMAP_SIZE];
190         struct list_head queue[MAX_PRIO];
191 };
192
193 /*
194  * This is the main, per-CPU runqueue data structure.
195  *
196  * Locking rule: those places that want to lock multiple runqueues
197  * (such as the load balancing or the thread migration code), lock
198  * acquire operations must be ordered by ascending &runqueue.
199  */
200 struct runqueue {
201         spinlock_t lock;
202
203         /*
204          * nr_running and cpu_load should be in the same cacheline because
205          * remote CPUs use both these fields when doing load calculation.
206          */
207         unsigned long nr_running;
208 #ifdef CONFIG_SMP
209         unsigned long prio_bias;
210         unsigned long cpu_load[3];
211 #endif
212         unsigned long long nr_switches;
213
214         /*
215          * This is part of a global counter where only the total sum
216          * over all CPUs matters. A task can increase this counter on
217          * one CPU and if it got migrated afterwards it may decrease
218          * it on another CPU. Always updated under the runqueue lock:
219          */
220         unsigned long nr_uninterruptible;
221
222         unsigned long expired_timestamp;
223         unsigned long long timestamp_last_tick;
224         task_t *curr, *idle;
225         struct mm_struct *prev_mm;
226         prio_array_t *active, *expired, arrays[2];
227         int best_expired_prio;
228         atomic_t nr_iowait;
229
230 #ifdef CONFIG_SMP
231         struct sched_domain *sd;
232
233         /* For active balancing */
234         int active_balance;
235         int push_cpu;
236
237         task_t *migration_thread;
238         struct list_head migration_queue;
239 #endif
240
241 #ifdef CONFIG_SCHEDSTATS
242         /* latency stats */
243         struct sched_info rq_sched_info;
244
245         /* sys_sched_yield() stats */
246         unsigned long yld_exp_empty;
247         unsigned long yld_act_empty;
248         unsigned long yld_both_empty;
249         unsigned long yld_cnt;
250
251         /* schedule() stats */
252         unsigned long sched_switch;
253         unsigned long sched_cnt;
254         unsigned long sched_goidle;
255
256         /* try_to_wake_up() stats */
257         unsigned long ttwu_cnt;
258         unsigned long ttwu_local;
259 #endif
260 };
261
262 static DEFINE_PER_CPU(struct runqueue, runqueues);
263
264 /*
265  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
266  * See detach_destroy_domains: synchronize_sched for details.
267  *
268  * The domain tree of any CPU may only be accessed from within
269  * preempt-disabled sections.
270  */
271 #define for_each_domain(cpu, domain) \
272 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
273
274 #define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
275 #define this_rq()               (&__get_cpu_var(runqueues))
276 #define task_rq(p)              cpu_rq(task_cpu(p))
277 #define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
278
279 #ifndef prepare_arch_switch
280 # define prepare_arch_switch(next)      do { } while (0)
281 #endif
282 #ifndef finish_arch_switch
283 # define finish_arch_switch(prev)       do { } while (0)
284 #endif
285
286 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
287 static inline int task_running(runqueue_t *rq, task_t *p)
288 {
289         return rq->curr == p;
290 }
291
292 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
293 {
294 }
295
296 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
297 {
298 #ifdef CONFIG_DEBUG_SPINLOCK
299         /* this is a valid case when another task releases the spinlock */
300         rq->lock.owner = current;
301 #endif
302         spin_unlock_irq(&rq->lock);
303 }
304
305 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
306 static inline int task_running(runqueue_t *rq, task_t *p)
307 {
308 #ifdef CONFIG_SMP
309         return p->oncpu;
310 #else
311         return rq->curr == p;
312 #endif
313 }
314
315 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
316 {
317 #ifdef CONFIG_SMP
318         /*
319          * We can optimise this out completely for !SMP, because the
320          * SMP rebalancing from interrupt is the only thing that cares
321          * here.
322          */
323         next->oncpu = 1;
324 #endif
325 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
326         spin_unlock_irq(&rq->lock);
327 #else
328         spin_unlock(&rq->lock);
329 #endif
330 }
331
332 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
333 {
334 #ifdef CONFIG_SMP
335         /*
336          * After ->oncpu is cleared, the task can be moved to a different CPU.
337          * We must ensure this doesn't happen until the switch is completely
338          * finished.
339          */
340         smp_wmb();
341         prev->oncpu = 0;
342 #endif
343 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
344         local_irq_enable();
345 #endif
346 }
347 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
348
349 /*
350  * task_rq_lock - lock the runqueue a given task resides on and disable
351  * interrupts.  Note the ordering: we can safely lookup the task_rq without
352  * explicitly disabling preemption.
353  */
354 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
355         __acquires(rq->lock)
356 {
357         struct runqueue *rq;
358
359 repeat_lock_task:
360         local_irq_save(*flags);
361         rq = task_rq(p);
362         spin_lock(&rq->lock);
363         if (unlikely(rq != task_rq(p))) {
364                 spin_unlock_irqrestore(&rq->lock, *flags);
365                 goto repeat_lock_task;
366         }
367         return rq;
368 }
369
370 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
371         __releases(rq->lock)
372 {
373         spin_unlock_irqrestore(&rq->lock, *flags);
374 }
375
376 #ifdef CONFIG_SCHEDSTATS
377 /*
378  * bump this up when changing the output format or the meaning of an existing
379  * format, so that tools can adapt (or abort)
380  */
381 #define SCHEDSTAT_VERSION 12
382
383 static int show_schedstat(struct seq_file *seq, void *v)
384 {
385         int cpu;
386
387         seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
388         seq_printf(seq, "timestamp %lu\n", jiffies);
389         for_each_online_cpu(cpu) {
390                 runqueue_t *rq = cpu_rq(cpu);
391 #ifdef CONFIG_SMP
392                 struct sched_domain *sd;
393                 int dcnt = 0;
394 #endif
395
396                 /* runqueue-specific stats */
397                 seq_printf(seq,
398                     "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
399                     cpu, rq->yld_both_empty,
400                     rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
401                     rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
402                     rq->ttwu_cnt, rq->ttwu_local,
403                     rq->rq_sched_info.cpu_time,
404                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
405
406                 seq_printf(seq, "\n");
407
408 #ifdef CONFIG_SMP
409                 /* domain-specific stats */
410                 preempt_disable();
411                 for_each_domain(cpu, sd) {
412                         enum idle_type itype;
413                         char mask_str[NR_CPUS];
414
415                         cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
416                         seq_printf(seq, "domain%d %s", dcnt++, mask_str);
417                         for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
418                                         itype++) {
419                                 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
420                                     sd->lb_cnt[itype],
421                                     sd->lb_balanced[itype],
422                                     sd->lb_failed[itype],
423                                     sd->lb_imbalance[itype],
424                                     sd->lb_gained[itype],
425                                     sd->lb_hot_gained[itype],
426                                     sd->lb_nobusyq[itype],
427                                     sd->lb_nobusyg[itype]);
428                         }
429                         seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
430                             sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
431                             sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
432                             sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
433                             sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
434                 }
435                 preempt_enable();
436 #endif
437         }
438         return 0;
439 }
440
441 static int schedstat_open(struct inode *inode, struct file *file)
442 {
443         unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
444         char *buf = kmalloc(size, GFP_KERNEL);
445         struct seq_file *m;
446         int res;
447
448         if (!buf)
449                 return -ENOMEM;
450         res = single_open(file, show_schedstat, NULL);
451         if (!res) {
452                 m = file->private_data;
453                 m->buf = buf;
454                 m->size = size;
455         } else
456                 kfree(buf);
457         return res;
458 }
459
460 struct file_operations proc_schedstat_operations = {
461         .open    = schedstat_open,
462         .read    = seq_read,
463         .llseek  = seq_lseek,
464         .release = single_release,
465 };
466
467 # define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
468 # define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
469 #else /* !CONFIG_SCHEDSTATS */
470 # define schedstat_inc(rq, field)       do { } while (0)
471 # define schedstat_add(rq, field, amt)  do { } while (0)
472 #endif
473
474 /*
475  * rq_lock - lock a given runqueue and disable interrupts.
476  */
477 static inline runqueue_t *this_rq_lock(void)
478         __acquires(rq->lock)
479 {
480         runqueue_t *rq;
481
482         local_irq_disable();
483         rq = this_rq();
484         spin_lock(&rq->lock);
485
486         return rq;
487 }
488
489 #ifdef CONFIG_SCHEDSTATS
490 /*
491  * Called when a process is dequeued from the active array and given
492  * the cpu.  We should note that with the exception of interactive
493  * tasks, the expired queue will become the active queue after the active
494  * queue is empty, without explicitly dequeuing and requeuing tasks in the
495  * expired queue.  (Interactive tasks may be requeued directly to the
496  * active queue, thus delaying tasks in the expired queue from running;
497  * see scheduler_tick()).
498  *
499  * This function is only called from sched_info_arrive(), rather than
500  * dequeue_task(). Even though a task may be queued and dequeued multiple
501  * times as it is shuffled about, we're really interested in knowing how
502  * long it was from the *first* time it was queued to the time that it
503  * finally hit a cpu.
504  */
505 static inline void sched_info_dequeued(task_t *t)
506 {
507         t->sched_info.last_queued = 0;
508 }
509
510 /*
511  * Called when a task finally hits the cpu.  We can now calculate how
512  * long it was waiting to run.  We also note when it began so that we
513  * can keep stats on how long its timeslice is.
514  */
515 static inline void sched_info_arrive(task_t *t)
516 {
517         unsigned long now = jiffies, diff = 0;
518         struct runqueue *rq = task_rq(t);
519
520         if (t->sched_info.last_queued)
521                 diff = now - t->sched_info.last_queued;
522         sched_info_dequeued(t);
523         t->sched_info.run_delay += diff;
524         t->sched_info.last_arrival = now;
525         t->sched_info.pcnt++;
526
527         if (!rq)
528                 return;
529
530         rq->rq_sched_info.run_delay += diff;
531         rq->rq_sched_info.pcnt++;
532 }
533
534 /*
535  * Called when a process is queued into either the active or expired
536  * array.  The time is noted and later used to determine how long we
537  * had to wait for us to reach the cpu.  Since the expired queue will
538  * become the active queue after active queue is empty, without dequeuing
539  * and requeuing any tasks, we are interested in queuing to either. It
540  * is unusual but not impossible for tasks to be dequeued and immediately
541  * requeued in the same or another array: this can happen in sched_yield(),
542  * set_user_nice(), and even load_balance() as it moves tasks from runqueue
543  * to runqueue.
544  *
545  * This function is only called from enqueue_task(), but also only updates
546  * the timestamp if it is already not set.  It's assumed that
547  * sched_info_dequeued() will clear that stamp when appropriate.
548  */
549 static inline void sched_info_queued(task_t *t)
550 {
551         if (!t->sched_info.last_queued)
552                 t->sched_info.last_queued = jiffies;
553 }
554
555 /*
556  * Called when a process ceases being the active-running process, either
557  * voluntarily or involuntarily.  Now we can calculate how long we ran.
558  */
559 static inline void sched_info_depart(task_t *t)
560 {
561         struct runqueue *rq = task_rq(t);
562         unsigned long diff = jiffies - t->sched_info.last_arrival;
563
564         t->sched_info.cpu_time += diff;
565
566         if (rq)
567                 rq->rq_sched_info.cpu_time += diff;
568 }
569
570 /*
571  * Called when tasks are switched involuntarily due, typically, to expiring
572  * their time slice.  (This may also be called when switching to or from
573  * the idle task.)  We are only called when prev != next.
574  */
575 static inline void sched_info_switch(task_t *prev, task_t *next)
576 {
577         struct runqueue *rq = task_rq(prev);
578
579         /*
580          * prev now departs the cpu.  It's not interesting to record
581          * stats about how efficient we were at scheduling the idle
582          * process, however.
583          */
584         if (prev != rq->idle)
585                 sched_info_depart(prev);
586
587         if (next != rq->idle)
588                 sched_info_arrive(next);
589 }
590 #else
591 #define sched_info_queued(t)            do { } while (0)
592 #define sched_info_switch(t, next)      do { } while (0)
593 #endif /* CONFIG_SCHEDSTATS */
594
595 /*
596  * Adding/removing a task to/from a priority array:
597  */
598 static void dequeue_task(struct task_struct *p, prio_array_t *array)
599 {
600         array->nr_active--;
601         list_del(&p->run_list);
602         if (list_empty(array->queue + p->prio))
603                 __clear_bit(p->prio, array->bitmap);
604 }
605
606 static void enqueue_task(struct task_struct *p, prio_array_t *array)
607 {
608         sched_info_queued(p);
609         list_add_tail(&p->run_list, array->queue + p->prio);
610         __set_bit(p->prio, array->bitmap);
611         array->nr_active++;
612         p->array = array;
613 }
614
615 /*
616  * Put task to the end of the run list without the overhead of dequeue
617  * followed by enqueue.
618  */
619 static void requeue_task(struct task_struct *p, prio_array_t *array)
620 {
621         list_move_tail(&p->run_list, array->queue + p->prio);
622 }
623
624 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
625 {
626         list_add(&p->run_list, array->queue + p->prio);
627         __set_bit(p->prio, array->bitmap);
628         array->nr_active++;
629         p->array = array;
630 }
631
632 /*
633  * effective_prio - return the priority that is based on the static
634  * priority but is modified by bonuses/penalties.
635  *
636  * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
637  * into the -5 ... 0 ... +5 bonus/penalty range.
638  *
639  * We use 25% of the full 0...39 priority range so that:
640  *
641  * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
642  * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
643  *
644  * Both properties are important to certain workloads.
645  */
646 static int effective_prio(task_t *p)
647 {
648         int bonus, prio;
649
650         if (rt_task(p))
651                 return p->prio;
652
653         bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
654
655         prio = p->static_prio - bonus;
656         if (prio < MAX_RT_PRIO)
657                 prio = MAX_RT_PRIO;
658         if (prio > MAX_PRIO-1)
659                 prio = MAX_PRIO-1;
660         return prio;
661 }
662
663 #ifdef CONFIG_SMP
664 static inline void inc_prio_bias(runqueue_t *rq, int prio)
665 {
666         rq->prio_bias += MAX_PRIO - prio;
667 }
668
669 static inline void dec_prio_bias(runqueue_t *rq, int prio)
670 {
671         rq->prio_bias -= MAX_PRIO - prio;
672 }
673
674 static inline void inc_nr_running(task_t *p, runqueue_t *rq)
675 {
676         rq->nr_running++;
677         if (rt_task(p)) {
678                 if (p != rq->migration_thread)
679                         /*
680                          * The migration thread does the actual balancing. Do
681                          * not bias by its priority as the ultra high priority
682                          * will skew balancing adversely.
683                          */
684                         inc_prio_bias(rq, p->prio);
685         } else
686                 inc_prio_bias(rq, p->static_prio);
687 }
688
689 static inline void dec_nr_running(task_t *p, runqueue_t *rq)
690 {
691         rq->nr_running--;
692         if (rt_task(p)) {
693                 if (p != rq->migration_thread)
694                         dec_prio_bias(rq, p->prio);
695         } else
696                 dec_prio_bias(rq, p->static_prio);
697 }
698 #else
699 static inline void inc_prio_bias(runqueue_t *rq, int prio)
700 {
701 }
702
703 static inline void dec_prio_bias(runqueue_t *rq, int prio)
704 {
705 }
706
707 static inline void inc_nr_running(task_t *p, runqueue_t *rq)
708 {
709         rq->nr_running++;
710 }
711
712 static inline void dec_nr_running(task_t *p, runqueue_t *rq)
713 {
714         rq->nr_running--;
715 }
716 #endif
717
718 /*
719  * __activate_task - move a task to the runqueue.
720  */
721 static inline void __activate_task(task_t *p, runqueue_t *rq)
722 {
723         enqueue_task(p, rq->active);
724         inc_nr_running(p, rq);
725 }
726
727 /*
728  * __activate_idle_task - move idle task to the _front_ of runqueue.
729  */
730 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
731 {
732         enqueue_task_head(p, rq->active);
733         inc_nr_running(p, rq);
734 }
735
736 static int recalc_task_prio(task_t *p, unsigned long long now)
737 {
738         /* Caller must always ensure 'now >= p->timestamp' */
739         unsigned long long __sleep_time = now - p->timestamp;
740         unsigned long sleep_time;
741
742         if (__sleep_time > NS_MAX_SLEEP_AVG)
743                 sleep_time = NS_MAX_SLEEP_AVG;
744         else
745                 sleep_time = (unsigned long)__sleep_time;
746
747         if (likely(sleep_time > 0)) {
748                 /*
749                  * User tasks that sleep a long time are categorised as
750                  * idle and will get just interactive status to stay active &
751                  * prevent them suddenly becoming cpu hogs and starving
752                  * other processes.
753                  */
754                 if (p->mm && p->activated != -1 &&
755                         sleep_time > INTERACTIVE_SLEEP(p)) {
756                                 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
757                                                 DEF_TIMESLICE);
758                 } else {
759                         /*
760                          * The lower the sleep avg a task has the more
761                          * rapidly it will rise with sleep time.
762                          */
763                         sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
764
765                         /*
766                          * Tasks waking from uninterruptible sleep are
767                          * limited in their sleep_avg rise as they
768                          * are likely to be waiting on I/O
769                          */
770                         if (p->activated == -1 && p->mm) {
771                                 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
772                                         sleep_time = 0;
773                                 else if (p->sleep_avg + sleep_time >=
774                                                 INTERACTIVE_SLEEP(p)) {
775                                         p->sleep_avg = INTERACTIVE_SLEEP(p);
776                                         sleep_time = 0;
777                                 }
778                         }
779
780                         /*
781                          * This code gives a bonus to interactive tasks.
782                          *
783                          * The boost works by updating the 'average sleep time'
784                          * value here, based on ->timestamp. The more time a
785                          * task spends sleeping, the higher the average gets -
786                          * and the higher the priority boost gets as well.
787                          */
788                         p->sleep_avg += sleep_time;
789
790                         if (p->sleep_avg > NS_MAX_SLEEP_AVG)
791                                 p->sleep_avg = NS_MAX_SLEEP_AVG;
792                 }
793         }
794
795         return effective_prio(p);
796 }
797
798 /*
799  * activate_task - move a task to the runqueue and do priority recalculation
800  *
801  * Update all the scheduling statistics stuff. (sleep average
802  * calculation, priority modifiers, etc.)
803  */
804 static void activate_task(task_t *p, runqueue_t *rq, int local)
805 {
806         unsigned long long now;
807
808         now = sched_clock();
809 #ifdef CONFIG_SMP
810         if (!local) {
811                 /* Compensate for drifting sched_clock */
812                 runqueue_t *this_rq = this_rq();
813                 now = (now - this_rq->timestamp_last_tick)
814                         + rq->timestamp_last_tick;
815         }
816 #endif
817
818         if (!rt_task(p))
819                 p->prio = recalc_task_prio(p, now);
820
821         /*
822          * This checks to make sure it's not an uninterruptible task
823          * that is now waking up.
824          */
825         if (!p->activated) {
826                 /*
827                  * Tasks which were woken up by interrupts (ie. hw events)
828                  * are most likely of interactive nature. So we give them
829                  * the credit of extending their sleep time to the period
830                  * of time they spend on the runqueue, waiting for execution
831                  * on a CPU, first time around:
832                  */
833                 if (in_interrupt())
834                         p->activated = 2;
835                 else {
836                         /*
837                          * Normal first-time wakeups get a credit too for
838                          * on-runqueue time, but it will be weighted down:
839                          */
840                         p->activated = 1;
841                 }
842         }
843         p->timestamp = now;
844
845         __activate_task(p, rq);
846 }
847
848 /*
849  * deactivate_task - remove a task from the runqueue.
850  */
851 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
852 {
853         dec_nr_running(p, rq);
854         dequeue_task(p, p->array);
855         p->array = NULL;
856 }
857
858 /*
859  * resched_task - mark a task 'to be rescheduled now'.
860  *
861  * On UP this means the setting of the need_resched flag, on SMP it
862  * might also involve a cross-CPU call to trigger the scheduler on
863  * the target CPU.
864  */
865 #ifdef CONFIG_SMP
866 static void resched_task(task_t *p)
867 {
868         int cpu;
869
870         assert_spin_locked(&task_rq(p)->lock);
871
872         if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
873                 return;
874
875         set_tsk_thread_flag(p, TIF_NEED_RESCHED);
876
877         cpu = task_cpu(p);
878         if (cpu == smp_processor_id())
879                 return;
880
881         /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
882         smp_mb();
883         if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG))
884                 smp_send_reschedule(cpu);
885 }
886 #else
887 static inline void resched_task(task_t *p)
888 {
889         assert_spin_locked(&task_rq(p)->lock);
890         set_tsk_need_resched(p);
891 }
892 #endif
893
894 /**
895  * task_curr - is this task currently executing on a CPU?
896  * @p: the task in question.
897  */
898 inline int task_curr(const task_t *p)
899 {
900         return cpu_curr(task_cpu(p)) == p;
901 }
902
903 #ifdef CONFIG_SMP
904 typedef struct {
905         struct list_head list;
906
907         task_t *task;
908         int dest_cpu;
909
910         struct completion done;
911 } migration_req_t;
912
913 /*
914  * The task's runqueue lock must be held.
915  * Returns true if you have to wait for migration thread.
916  */
917 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
918 {
919         runqueue_t *rq = task_rq(p);
920
921         /*
922          * If the task is not on a runqueue (and not running), then
923          * it is sufficient to simply update the task's cpu field.
924          */
925         if (!p->array && !task_running(rq, p)) {
926                 set_task_cpu(p, dest_cpu);
927                 return 0;
928         }
929
930         init_completion(&req->done);
931         req->task = p;
932         req->dest_cpu = dest_cpu;
933         list_add(&req->list, &rq->migration_queue);
934         return 1;
935 }
936
937 /*
938  * wait_task_inactive - wait for a thread to unschedule.
939  *
940  * The caller must ensure that the task *will* unschedule sometime soon,
941  * else this function might spin for a *long* time. This function can't
942  * be called with interrupts off, or it may introduce deadlock with
943  * smp_call_function() if an IPI is sent by the same process we are
944  * waiting to become inactive.
945  */
946 void wait_task_inactive(task_t *p)
947 {
948         unsigned long flags;
949         runqueue_t *rq;
950         int preempted;
951
952 repeat:
953         rq = task_rq_lock(p, &flags);
954         /* Must be off runqueue entirely, not preempted. */
955         if (unlikely(p->array || task_running(rq, p))) {
956                 /* If it's preempted, we yield.  It could be a while. */
957                 preempted = !task_running(rq, p);
958                 task_rq_unlock(rq, &flags);
959                 cpu_relax();
960                 if (preempted)
961                         yield();
962                 goto repeat;
963         }
964         task_rq_unlock(rq, &flags);
965 }
966
967 /***
968  * kick_process - kick a running thread to enter/exit the kernel
969  * @p: the to-be-kicked thread
970  *
971  * Cause a process which is running on another CPU to enter
972  * kernel-mode, without any delay. (to get signals handled.)
973  *
974  * NOTE: this function doesnt have to take the runqueue lock,
975  * because all it wants to ensure is that the remote task enters
976  * the kernel. If the IPI races and the task has been migrated
977  * to another CPU then no harm is done and the purpose has been
978  * achieved as well.
979  */
980 void kick_process(task_t *p)
981 {
982         int cpu;
983
984         preempt_disable();
985         cpu = task_cpu(p);
986         if ((cpu != smp_processor_id()) && task_curr(p))
987                 smp_send_reschedule(cpu);
988         preempt_enable();
989 }
990
991 /*
992  * Return a low guess at the load of a migration-source cpu.
993  *
994  * We want to under-estimate the load of migration sources, to
995  * balance conservatively.
996  */
997 static inline unsigned long __source_load(int cpu, int type, enum idle_type idle)
998 {
999         runqueue_t *rq = cpu_rq(cpu);
1000         unsigned long running = rq->nr_running;
1001         unsigned long source_load, cpu_load = rq->cpu_load[type-1],
1002                 load_now = running * SCHED_LOAD_SCALE;
1003
1004         if (type == 0)
1005                 source_load = load_now;
1006         else
1007                 source_load = min(cpu_load, load_now);
1008
1009         if (running > 1 || (idle == NOT_IDLE && running))
1010                 /*
1011                  * If we are busy rebalancing the load is biased by
1012                  * priority to create 'nice' support across cpus. When
1013                  * idle rebalancing we should only bias the source_load if
1014                  * there is more than one task running on that queue to
1015                  * prevent idle rebalance from trying to pull tasks from a
1016                  * queue with only one running task.
1017                  */
1018                 source_load = source_load * rq->prio_bias / running;
1019
1020         return source_load;
1021 }
1022
1023 static inline unsigned long source_load(int cpu, int type)
1024 {
1025         return __source_load(cpu, type, NOT_IDLE);
1026 }
1027
1028 /*
1029  * Return a high guess at the load of a migration-target cpu
1030  */
1031 static inline unsigned long __target_load(int cpu, int type, enum idle_type idle)
1032 {
1033         runqueue_t *rq = cpu_rq(cpu);
1034         unsigned long running = rq->nr_running;
1035         unsigned long target_load, cpu_load = rq->cpu_load[type-1],
1036                 load_now = running * SCHED_LOAD_SCALE;
1037
1038         if (type == 0)
1039                 target_load = load_now;
1040         else
1041                 target_load = max(cpu_load, load_now);
1042
1043         if (running > 1 || (idle == NOT_IDLE && running))
1044                 target_load = target_load * rq->prio_bias / running;
1045
1046         return target_load;
1047 }
1048
1049 static inline unsigned long target_load(int cpu, int type)
1050 {
1051         return __target_load(cpu, type, NOT_IDLE);
1052 }
1053
1054 /*
1055  * find_idlest_group finds and returns the least busy CPU group within the
1056  * domain.
1057  */
1058 static struct sched_group *
1059 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1060 {
1061         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1062         unsigned long min_load = ULONG_MAX, this_load = 0;
1063         int load_idx = sd->forkexec_idx;
1064         int imbalance = 100 + (sd->imbalance_pct-100)/2;
1065
1066         do {
1067                 unsigned long load, avg_load;
1068                 int local_group;
1069                 int i;
1070
1071                 /* Skip over this group if it has no CPUs allowed */
1072                 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1073                         goto nextgroup;
1074
1075                 local_group = cpu_isset(this_cpu, group->cpumask);
1076
1077                 /* Tally up the load of all CPUs in the group */
1078                 avg_load = 0;
1079
1080                 for_each_cpu_mask(i, group->cpumask) {
1081                         /* Bias balancing toward cpus of our domain */
1082                         if (local_group)
1083                                 load = source_load(i, load_idx);
1084                         else
1085                                 load = target_load(i, load_idx);
1086
1087                         avg_load += load;
1088                 }
1089
1090                 /* Adjust by relative CPU power of the group */
1091                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1092
1093                 if (local_group) {
1094                         this_load = avg_load;
1095                         this = group;
1096                 } else if (avg_load < min_load) {
1097                         min_load = avg_load;
1098                         idlest = group;
1099                 }
1100 nextgroup:
1101                 group = group->next;
1102         } while (group != sd->groups);
1103
1104         if (!idlest || 100*this_load < imbalance*min_load)
1105                 return NULL;
1106         return idlest;
1107 }
1108
1109 /*
1110  * find_idlest_queue - find the idlest runqueue among the cpus in group.
1111  */
1112 static int
1113 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1114 {
1115         cpumask_t tmp;
1116         unsigned long load, min_load = ULONG_MAX;
1117         int idlest = -1;
1118         int i;
1119
1120         /* Traverse only the allowed CPUs */
1121         cpus_and(tmp, group->cpumask, p->cpus_allowed);
1122
1123         for_each_cpu_mask(i, tmp) {
1124                 load = source_load(i, 0);
1125
1126                 if (load < min_load || (load == min_load && i == this_cpu)) {
1127                         min_load = load;
1128                         idlest = i;
1129                 }
1130         }
1131
1132         return idlest;
1133 }
1134
1135 /*
1136  * sched_balance_self: balance the current task (running on cpu) in domains
1137  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1138  * SD_BALANCE_EXEC.
1139  *
1140  * Balance, ie. select the least loaded group.
1141  *
1142  * Returns the target CPU number, or the same CPU if no balancing is needed.
1143  *
1144  * preempt must be disabled.
1145  */
1146 static int sched_balance_self(int cpu, int flag)
1147 {
1148         struct task_struct *t = current;
1149         struct sched_domain *tmp, *sd = NULL;
1150
1151         for_each_domain(cpu, tmp)
1152                 if (tmp->flags & flag)
1153                         sd = tmp;
1154
1155         while (sd) {
1156                 cpumask_t span;
1157                 struct sched_group *group;
1158                 int new_cpu;
1159                 int weight;
1160
1161                 span = sd->span;
1162                 group = find_idlest_group(sd, t, cpu);
1163                 if (!group)
1164                         goto nextlevel;
1165
1166                 new_cpu = find_idlest_cpu(group, t, cpu);
1167                 if (new_cpu == -1 || new_cpu == cpu)
1168                         goto nextlevel;
1169
1170                 /* Now try balancing at a lower domain level */
1171                 cpu = new_cpu;
1172 nextlevel:
1173                 sd = NULL;
1174                 weight = cpus_weight(span);
1175                 for_each_domain(cpu, tmp) {
1176                         if (weight <= cpus_weight(tmp->span))
1177                                 break;
1178                         if (tmp->flags & flag)
1179                                 sd = tmp;
1180                 }
1181                 /* while loop will break here if sd == NULL */
1182         }
1183
1184         return cpu;
1185 }
1186
1187 #endif /* CONFIG_SMP */
1188
1189 /*
1190  * wake_idle() will wake a task on an idle cpu if task->cpu is
1191  * not idle and an idle cpu is available.  The span of cpus to
1192  * search starts with cpus closest then further out as needed,
1193  * so we always favor a closer, idle cpu.
1194  *
1195  * Returns the CPU we should wake onto.
1196  */
1197 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1198 static int wake_idle(int cpu, task_t *p)
1199 {
1200         cpumask_t tmp;
1201         struct sched_domain *sd;
1202         int i;
1203
1204         if (idle_cpu(cpu))
1205                 return cpu;
1206
1207         for_each_domain(cpu, sd) {
1208                 if (sd->flags & SD_WAKE_IDLE) {
1209                         cpus_and(tmp, sd->span, p->cpus_allowed);
1210                         for_each_cpu_mask(i, tmp) {
1211                                 if (idle_cpu(i))
1212                                         return i;
1213                         }
1214                 }
1215                 else
1216                         break;
1217         }
1218         return cpu;
1219 }
1220 #else
1221 static inline int wake_idle(int cpu, task_t *p)
1222 {
1223         return cpu;
1224 }
1225 #endif
1226
1227 /***
1228  * try_to_wake_up - wake up a thread
1229  * @p: the to-be-woken-up thread
1230  * @state: the mask of task states that can be woken
1231  * @sync: do a synchronous wakeup?
1232  *
1233  * Put it on the run-queue if it's not already there. The "current"
1234  * thread is always on the run-queue (except when the actual
1235  * re-schedule is in progress), and as such you're allowed to do
1236  * the simpler "current->state = TASK_RUNNING" to mark yourself
1237  * runnable without the overhead of this.
1238  *
1239  * returns failure only if the task is already active.
1240  */
1241 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1242 {
1243         int cpu, this_cpu, success = 0;
1244         unsigned long flags;
1245         long old_state;
1246         runqueue_t *rq;
1247 #ifdef CONFIG_SMP
1248         unsigned long load, this_load;
1249         struct sched_domain *sd, *this_sd = NULL;
1250         int new_cpu;
1251 #endif
1252
1253         rq = task_rq_lock(p, &flags);
1254         old_state = p->state;
1255         if (!(old_state & state))
1256                 goto out;
1257
1258         if (p->array)
1259                 goto out_running;
1260
1261         cpu = task_cpu(p);
1262         this_cpu = smp_processor_id();
1263
1264 #ifdef CONFIG_SMP
1265         if (unlikely(task_running(rq, p)))
1266                 goto out_activate;
1267
1268         new_cpu = cpu;
1269
1270         schedstat_inc(rq, ttwu_cnt);
1271         if (cpu == this_cpu) {
1272                 schedstat_inc(rq, ttwu_local);
1273                 goto out_set_cpu;
1274         }
1275
1276         for_each_domain(this_cpu, sd) {
1277                 if (cpu_isset(cpu, sd->span)) {
1278                         schedstat_inc(sd, ttwu_wake_remote);
1279                         this_sd = sd;
1280                         break;
1281                 }
1282         }
1283
1284         if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1285                 goto out_set_cpu;
1286
1287         /*
1288          * Check for affine wakeup and passive balancing possibilities.
1289          */
1290         if (this_sd) {
1291                 int idx = this_sd->wake_idx;
1292                 unsigned int imbalance;
1293
1294                 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1295
1296                 load = source_load(cpu, idx);
1297                 this_load = target_load(this_cpu, idx);
1298
1299                 new_cpu = this_cpu; /* Wake to this CPU if we can */
1300
1301                 if (this_sd->flags & SD_WAKE_AFFINE) {
1302                         unsigned long tl = this_load;
1303                         /*
1304                          * If sync wakeup then subtract the (maximum possible)
1305                          * effect of the currently running task from the load
1306                          * of the current CPU:
1307                          */
1308                         if (sync)
1309                                 tl -= SCHED_LOAD_SCALE;
1310
1311                         if ((tl <= load &&
1312                                 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1313                                 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1314                                 /*
1315                                  * This domain has SD_WAKE_AFFINE and
1316                                  * p is cache cold in this domain, and
1317                                  * there is no bad imbalance.
1318                                  */
1319                                 schedstat_inc(this_sd, ttwu_move_affine);
1320                                 goto out_set_cpu;
1321                         }
1322                 }
1323
1324                 /*
1325                  * Start passive balancing when half the imbalance_pct
1326                  * limit is reached.
1327                  */
1328                 if (this_sd->flags & SD_WAKE_BALANCE) {
1329                         if (imbalance*this_load <= 100*load) {
1330                                 schedstat_inc(this_sd, ttwu_move_balance);
1331                                 goto out_set_cpu;
1332                         }
1333                 }
1334         }
1335
1336         new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1337 out_set_cpu:
1338         new_cpu = wake_idle(new_cpu, p);
1339         if (new_cpu != cpu) {
1340                 set_task_cpu(p, new_cpu);
1341                 task_rq_unlock(rq, &flags);
1342                 /* might preempt at this point */
1343                 rq = task_rq_lock(p, &flags);
1344                 old_state = p->state;
1345                 if (!(old_state & state))
1346                         goto out;
1347                 if (p->array)
1348                         goto out_running;
1349
1350                 this_cpu = smp_processor_id();
1351                 cpu = task_cpu(p);
1352         }
1353
1354 out_activate:
1355 #endif /* CONFIG_SMP */
1356         if (old_state == TASK_UNINTERRUPTIBLE) {
1357                 rq->nr_uninterruptible--;
1358                 /*
1359                  * Tasks on involuntary sleep don't earn
1360                  * sleep_avg beyond just interactive state.
1361                  */
1362                 p->activated = -1;
1363         }
1364
1365         /*
1366          * Tasks that have marked their sleep as noninteractive get
1367          * woken up without updating their sleep average. (i.e. their
1368          * sleep is handled in a priority-neutral manner, no priority
1369          * boost and no penalty.)
1370          */
1371         if (old_state & TASK_NONINTERACTIVE)
1372                 __activate_task(p, rq);
1373         else
1374                 activate_task(p, rq, cpu == this_cpu);
1375         /*
1376          * Sync wakeups (i.e. those types of wakeups where the waker
1377          * has indicated that it will leave the CPU in short order)
1378          * don't trigger a preemption, if the woken up task will run on
1379          * this cpu. (in this case the 'I will reschedule' promise of
1380          * the waker guarantees that the freshly woken up task is going
1381          * to be considered on this CPU.)
1382          */
1383         if (!sync || cpu != this_cpu) {
1384                 if (TASK_PREEMPTS_CURR(p, rq))
1385                         resched_task(rq->curr);
1386         }
1387         success = 1;
1388
1389 out_running:
1390         p->state = TASK_RUNNING;
1391 out:
1392         task_rq_unlock(rq, &flags);
1393
1394         return success;
1395 }
1396
1397 int fastcall wake_up_process(task_t *p)
1398 {
1399         return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1400                                  TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1401 }
1402
1403 EXPORT_SYMBOL(wake_up_process);
1404
1405 int fastcall wake_up_state(task_t *p, unsigned int state)
1406 {
1407         return try_to_wake_up(p, state, 0);
1408 }
1409
1410 /*
1411  * Perform scheduler related setup for a newly forked process p.
1412  * p is forked by current.
1413  */
1414 void fastcall sched_fork(task_t *p, int clone_flags)
1415 {
1416         int cpu = get_cpu();
1417
1418 #ifdef CONFIG_SMP
1419         cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1420 #endif
1421         set_task_cpu(p, cpu);
1422
1423         /*
1424          * We mark the process as running here, but have not actually
1425          * inserted it onto the runqueue yet. This guarantees that
1426          * nobody will actually run it, and a signal or other external
1427          * event cannot wake it up and insert it on the runqueue either.
1428          */
1429         p->state = TASK_RUNNING;
1430         INIT_LIST_HEAD(&p->run_list);
1431         p->array = NULL;
1432 #ifdef CONFIG_SCHEDSTATS
1433         memset(&p->sched_info, 0, sizeof(p->sched_info));
1434 #endif
1435 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1436         p->oncpu = 0;
1437 #endif
1438 #ifdef CONFIG_PREEMPT
1439         /* Want to start with kernel preemption disabled. */
1440         task_thread_info(p)->preempt_count = 1;
1441 #endif
1442         /*
1443          * Share the timeslice between parent and child, thus the
1444          * total amount of pending timeslices in the system doesn't change,
1445          * resulting in more scheduling fairness.
1446          */
1447         local_irq_disable();
1448         p->time_slice = (current->time_slice + 1) >> 1;
1449         /*
1450          * The remainder of the first timeslice might be recovered by
1451          * the parent if the child exits early enough.
1452          */
1453         p->first_time_slice = 1;
1454         current->time_slice >>= 1;
1455         p->timestamp = sched_clock();
1456         if (unlikely(!current->time_slice)) {
1457                 /*
1458                  * This case is rare, it happens when the parent has only
1459                  * a single jiffy left from its timeslice. Taking the
1460                  * runqueue lock is not a problem.
1461                  */
1462                 current->time_slice = 1;
1463                 scheduler_tick();
1464         }
1465         local_irq_enable();
1466         put_cpu();
1467 }
1468
1469 /*
1470  * wake_up_new_task - wake up a newly created task for the first time.
1471  *
1472  * This function will do some initial scheduler statistics housekeeping
1473  * that must be done for every newly created context, then puts the task
1474  * on the runqueue and wakes it.
1475  */
1476 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1477 {
1478         unsigned long flags;
1479         int this_cpu, cpu;
1480         runqueue_t *rq, *this_rq;
1481
1482         rq = task_rq_lock(p, &flags);
1483         BUG_ON(p->state != TASK_RUNNING);
1484         this_cpu = smp_processor_id();
1485         cpu = task_cpu(p);
1486
1487         /*
1488          * We decrease the sleep average of forking parents
1489          * and children as well, to keep max-interactive tasks
1490          * from forking tasks that are max-interactive. The parent
1491          * (current) is done further down, under its lock.
1492          */
1493         p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1494                 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1495
1496         p->prio = effective_prio(p);
1497
1498         if (likely(cpu == this_cpu)) {
1499                 if (!(clone_flags & CLONE_VM)) {
1500                         /*
1501                          * The VM isn't cloned, so we're in a good position to
1502                          * do child-runs-first in anticipation of an exec. This
1503                          * usually avoids a lot of COW overhead.
1504                          */
1505                         if (unlikely(!current->array))
1506                                 __activate_task(p, rq);
1507                         else {
1508                                 p->prio = current->prio;
1509                                 list_add_tail(&p->run_list, &current->run_list);
1510                                 p->array = current->array;
1511                                 p->array->nr_active++;
1512                                 inc_nr_running(p, rq);
1513                         }
1514                         set_need_resched();
1515                 } else
1516                         /* Run child last */
1517                         __activate_task(p, rq);
1518                 /*
1519                  * We skip the following code due to cpu == this_cpu
1520                  *
1521                  *   task_rq_unlock(rq, &flags);
1522                  *   this_rq = task_rq_lock(current, &flags);
1523                  */
1524                 this_rq = rq;
1525         } else {
1526                 this_rq = cpu_rq(this_cpu);
1527
1528                 /*
1529                  * Not the local CPU - must adjust timestamp. This should
1530                  * get optimised away in the !CONFIG_SMP case.
1531                  */
1532                 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1533                                         + rq->timestamp_last_tick;
1534                 __activate_task(p, rq);
1535                 if (TASK_PREEMPTS_CURR(p, rq))
1536                         resched_task(rq->curr);
1537
1538                 /*
1539                  * Parent and child are on different CPUs, now get the
1540                  * parent runqueue to update the parent's ->sleep_avg:
1541                  */
1542                 task_rq_unlock(rq, &flags);
1543                 this_rq = task_rq_lock(current, &flags);
1544         }
1545         current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1546                 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1547         task_rq_unlock(this_rq, &flags);
1548 }
1549
1550 /*
1551  * Potentially available exiting-child timeslices are
1552  * retrieved here - this way the parent does not get
1553  * penalized for creating too many threads.
1554  *
1555  * (this cannot be used to 'generate' timeslices
1556  * artificially, because any timeslice recovered here
1557  * was given away by the parent in the first place.)
1558  */
1559 void fastcall sched_exit(task_t *p)
1560 {
1561         unsigned long flags;
1562         runqueue_t *rq;
1563
1564         /*
1565          * If the child was a (relative-) CPU hog then decrease
1566          * the sleep_avg of the parent as well.
1567          */
1568         rq = task_rq_lock(p->parent, &flags);
1569         if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1570                 p->parent->time_slice += p->time_slice;
1571                 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1572                         p->parent->time_slice = task_timeslice(p);
1573         }
1574         if (p->sleep_avg < p->parent->sleep_avg)
1575                 p->parent->sleep_avg = p->parent->sleep_avg /
1576                 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1577                 (EXIT_WEIGHT + 1);
1578         task_rq_unlock(rq, &flags);
1579 }
1580
1581 /**
1582  * prepare_task_switch - prepare to switch tasks
1583  * @rq: the runqueue preparing to switch
1584  * @next: the task we are going to switch to.
1585  *
1586  * This is called with the rq lock held and interrupts off. It must
1587  * be paired with a subsequent finish_task_switch after the context
1588  * switch.
1589  *
1590  * prepare_task_switch sets up locking and calls architecture specific
1591  * hooks.
1592  */
1593 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1594 {
1595         prepare_lock_switch(rq, next);
1596         prepare_arch_switch(next);
1597 }
1598
1599 /**
1600  * finish_task_switch - clean up after a task-switch
1601  * @rq: runqueue associated with task-switch
1602  * @prev: the thread we just switched away from.
1603  *
1604  * finish_task_switch must be called after the context switch, paired
1605  * with a prepare_task_switch call before the context switch.
1606  * finish_task_switch will reconcile locking set up by prepare_task_switch,
1607  * and do any other architecture-specific cleanup actions.
1608  *
1609  * Note that we may have delayed dropping an mm in context_switch(). If
1610  * so, we finish that here outside of the runqueue lock.  (Doing it
1611  * with the lock held can cause deadlocks; see schedule() for
1612  * details.)
1613  */
1614 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1615         __releases(rq->lock)
1616 {
1617         struct mm_struct *mm = rq->prev_mm;
1618         unsigned long prev_task_flags;
1619
1620         rq->prev_mm = NULL;
1621
1622         /*
1623          * A task struct has one reference for the use as "current".
1624          * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1625          * calls schedule one last time. The schedule call will never return,
1626          * and the scheduled task must drop that reference.
1627          * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1628          * still held, otherwise prev could be scheduled on another cpu, die
1629          * there before we look at prev->state, and then the reference would
1630          * be dropped twice.
1631          *              Manfred Spraul <manfred@colorfullife.com>
1632          */
1633         prev_task_flags = prev->flags;
1634         finish_arch_switch(prev);
1635         finish_lock_switch(rq, prev);
1636         if (mm)
1637                 mmdrop(mm);
1638         if (unlikely(prev_task_flags & PF_DEAD))
1639                 put_task_struct(prev);
1640 }
1641
1642 /**
1643  * schedule_tail - first thing a freshly forked thread must call.
1644  * @prev: the thread we just switched away from.
1645  */
1646 asmlinkage void schedule_tail(task_t *prev)
1647         __releases(rq->lock)
1648 {
1649         runqueue_t *rq = this_rq();
1650         finish_task_switch(rq, prev);
1651 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1652         /* In this case, finish_task_switch does not reenable preemption */
1653         preempt_enable();
1654 #endif
1655         if (current->set_child_tid)
1656                 put_user(current->pid, current->set_child_tid);
1657 }
1658
1659 /*
1660  * context_switch - switch to the new MM and the new
1661  * thread's register state.
1662  */
1663 static inline
1664 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1665 {
1666         struct mm_struct *mm = next->mm;
1667         struct mm_struct *oldmm = prev->active_mm;
1668
1669         if (unlikely(!mm)) {
1670                 next->active_mm = oldmm;
1671                 atomic_inc(&oldmm->mm_count);
1672                 enter_lazy_tlb(oldmm, next);
1673         } else
1674                 switch_mm(oldmm, mm, next);
1675
1676         if (unlikely(!prev->mm)) {
1677                 prev->active_mm = NULL;
1678                 WARN_ON(rq->prev_mm);
1679                 rq->prev_mm = oldmm;
1680         }
1681
1682         /* Here we just switch the register state and the stack. */
1683         switch_to(prev, next, prev);
1684
1685         return prev;
1686 }
1687
1688 /*
1689  * nr_running, nr_uninterruptible and nr_context_switches:
1690  *
1691  * externally visible scheduler statistics: current number of runnable
1692  * threads, current number of uninterruptible-sleeping threads, total
1693  * number of context switches performed since bootup.
1694  */
1695 unsigned long nr_running(void)
1696 {
1697         unsigned long i, sum = 0;
1698
1699         for_each_online_cpu(i)
1700                 sum += cpu_rq(i)->nr_running;
1701
1702         return sum;
1703 }
1704
1705 unsigned long nr_uninterruptible(void)
1706 {
1707         unsigned long i, sum = 0;
1708
1709         for_each_cpu(i)
1710                 sum += cpu_rq(i)->nr_uninterruptible;
1711
1712         /*
1713          * Since we read the counters lockless, it might be slightly
1714          * inaccurate. Do not allow it to go below zero though:
1715          */
1716         if (unlikely((long)sum < 0))
1717                 sum = 0;
1718
1719         return sum;
1720 }
1721
1722 unsigned long long nr_context_switches(void)
1723 {
1724         unsigned long long i, sum = 0;
1725
1726         for_each_cpu(i)
1727                 sum += cpu_rq(i)->nr_switches;
1728
1729         return sum;
1730 }
1731
1732 unsigned long nr_iowait(void)
1733 {
1734         unsigned long i, sum = 0;
1735
1736         for_each_cpu(i)
1737                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1738
1739         return sum;
1740 }
1741
1742 #ifdef CONFIG_SMP
1743
1744 /*
1745  * double_rq_lock - safely lock two runqueues
1746  *
1747  * Note this does not disable interrupts like task_rq_lock,
1748  * you need to do so manually before calling.
1749  */
1750 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1751         __acquires(rq1->lock)
1752         __acquires(rq2->lock)
1753 {
1754         if (rq1 == rq2) {
1755                 spin_lock(&rq1->lock);
1756                 __acquire(rq2->lock);   /* Fake it out ;) */
1757         } else {
1758                 if (rq1 < rq2) {
1759                         spin_lock(&rq1->lock);
1760                         spin_lock(&rq2->lock);
1761                 } else {
1762                         spin_lock(&rq2->lock);
1763                         spin_lock(&rq1->lock);
1764                 }
1765         }
1766 }
1767
1768 /*
1769  * double_rq_unlock - safely unlock two runqueues
1770  *
1771  * Note this does not restore interrupts like task_rq_unlock,
1772  * you need to do so manually after calling.
1773  */
1774 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1775         __releases(rq1->lock)
1776         __releases(rq2->lock)
1777 {
1778         spin_unlock(&rq1->lock);
1779         if (rq1 != rq2)
1780                 spin_unlock(&rq2->lock);
1781         else
1782                 __release(rq2->lock);
1783 }
1784
1785 /*
1786  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1787  */
1788 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1789         __releases(this_rq->lock)
1790         __acquires(busiest->lock)
1791         __acquires(this_rq->lock)
1792 {
1793         if (unlikely(!spin_trylock(&busiest->lock))) {
1794                 if (busiest < this_rq) {
1795                         spin_unlock(&this_rq->lock);
1796                         spin_lock(&busiest->lock);
1797                         spin_lock(&this_rq->lock);
1798                 } else
1799                         spin_lock(&busiest->lock);
1800         }
1801 }
1802
1803 /*
1804  * If dest_cpu is allowed for this process, migrate the task to it.
1805  * This is accomplished by forcing the cpu_allowed mask to only
1806  * allow dest_cpu, which will force the cpu onto dest_cpu.  Then
1807  * the cpu_allowed mask is restored.
1808  */
1809 static void sched_migrate_task(task_t *p, int dest_cpu)
1810 {
1811         migration_req_t req;
1812         runqueue_t *rq;
1813         unsigned long flags;
1814
1815         rq = task_rq_lock(p, &flags);
1816         if (!cpu_isset(dest_cpu, p->cpus_allowed)
1817             || unlikely(cpu_is_offline(dest_cpu)))
1818                 goto out;
1819
1820         /* force the process onto the specified CPU */
1821         if (migrate_task(p, dest_cpu, &req)) {
1822                 /* Need to wait for migration thread (might exit: take ref). */
1823                 struct task_struct *mt = rq->migration_thread;
1824                 get_task_struct(mt);
1825                 task_rq_unlock(rq, &flags);
1826                 wake_up_process(mt);
1827                 put_task_struct(mt);
1828                 wait_for_completion(&req.done);
1829                 return;
1830         }
1831 out:
1832         task_rq_unlock(rq, &flags);
1833 }
1834
1835 /*
1836  * sched_exec - execve() is a valuable balancing opportunity, because at
1837  * this point the task has the smallest effective memory and cache footprint.
1838  */
1839 void sched_exec(void)
1840 {
1841         int new_cpu, this_cpu = get_cpu();
1842         new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1843         put_cpu();
1844         if (new_cpu != this_cpu)
1845                 sched_migrate_task(current, new_cpu);
1846 }
1847
1848 /*
1849  * pull_task - move a task from a remote runqueue to the local runqueue.
1850  * Both runqueues must be locked.
1851  */
1852 static inline
1853 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1854                runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1855 {
1856         dequeue_task(p, src_array);
1857         dec_nr_running(p, src_rq);
1858         set_task_cpu(p, this_cpu);
1859         inc_nr_running(p, this_rq);
1860         enqueue_task(p, this_array);
1861         p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1862                                 + this_rq->timestamp_last_tick;
1863         /*
1864          * Note that idle threads have a prio of MAX_PRIO, for this test
1865          * to be always true for them.
1866          */
1867         if (TASK_PREEMPTS_CURR(p, this_rq))
1868                 resched_task(this_rq->curr);
1869 }
1870
1871 /*
1872  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1873  */
1874 static inline
1875 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1876                      struct sched_domain *sd, enum idle_type idle,
1877                      int *all_pinned)
1878 {
1879         /*
1880          * We do not migrate tasks that are:
1881          * 1) running (obviously), or
1882          * 2) cannot be migrated to this CPU due to cpus_allowed, or
1883          * 3) are cache-hot on their current CPU.
1884          */
1885         if (!cpu_isset(this_cpu, p->cpus_allowed))
1886                 return 0;
1887         *all_pinned = 0;
1888
1889         if (task_running(rq, p))
1890                 return 0;
1891
1892         /*
1893          * Aggressive migration if:
1894          * 1) task is cache cold, or
1895          * 2) too many balance attempts have failed.
1896          */
1897
1898         if (sd->nr_balance_failed > sd->cache_nice_tries)
1899                 return 1;
1900
1901         if (task_hot(p, rq->timestamp_last_tick, sd))
1902                 return 0;
1903         return 1;
1904 }
1905
1906 /*
1907  * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1908  * as part of a balancing operation within "domain". Returns the number of
1909  * tasks moved.
1910  *
1911  * Called with both runqueues locked.
1912  */
1913 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1914                       unsigned long max_nr_move, struct sched_domain *sd,
1915                       enum idle_type idle, int *all_pinned)
1916 {
1917         prio_array_t *array, *dst_array;
1918         struct list_head *head, *curr;
1919         int idx, pulled = 0, pinned = 0;
1920         task_t *tmp;
1921
1922         if (max_nr_move == 0)
1923                 goto out;
1924
1925         pinned = 1;
1926
1927         /*
1928          * We first consider expired tasks. Those will likely not be
1929          * executed in the near future, and they are most likely to
1930          * be cache-cold, thus switching CPUs has the least effect
1931          * on them.
1932          */
1933         if (busiest->expired->nr_active) {
1934                 array = busiest->expired;
1935                 dst_array = this_rq->expired;
1936         } else {
1937                 array = busiest->active;
1938                 dst_array = this_rq->active;
1939         }
1940
1941 new_array:
1942         /* Start searching at priority 0: */
1943         idx = 0;
1944 skip_bitmap:
1945         if (!idx)
1946                 idx = sched_find_first_bit(array->bitmap);
1947         else
1948                 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1949         if (idx >= MAX_PRIO) {
1950                 if (array == busiest->expired && busiest->active->nr_active) {
1951                         array = busiest->active;
1952                         dst_array = this_rq->active;
1953                         goto new_array;
1954                 }
1955                 goto out;
1956         }
1957
1958         head = array->queue + idx;
1959         curr = head->prev;
1960 skip_queue:
1961         tmp = list_entry(curr, task_t, run_list);
1962
1963         curr = curr->prev;
1964
1965         if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1966                 if (curr != head)
1967                         goto skip_queue;
1968                 idx++;
1969                 goto skip_bitmap;
1970         }
1971
1972 #ifdef CONFIG_SCHEDSTATS
1973         if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1974                 schedstat_inc(sd, lb_hot_gained[idle]);
1975 #endif
1976
1977         pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1978         pulled++;
1979
1980         /* We only want to steal up to the prescribed number of tasks. */
1981         if (pulled < max_nr_move) {
1982                 if (curr != head)
1983                         goto skip_queue;
1984                 idx++;
1985                 goto skip_bitmap;
1986         }
1987 out:
1988         /*
1989          * Right now, this is the only place pull_task() is called,
1990          * so we can safely collect pull_task() stats here rather than
1991          * inside pull_task().
1992          */
1993         schedstat_add(sd, lb_gained[idle], pulled);
1994
1995         if (all_pinned)
1996                 *all_pinned = pinned;
1997         return pulled;
1998 }
1999
2000 /*
2001  * find_busiest_group finds and returns the busiest CPU group within the
2002  * domain. It calculates and returns the number of tasks which should be
2003  * moved to restore balance via the imbalance parameter.
2004  */
2005 static struct sched_group *
2006 find_busiest_group(struct sched_domain *sd, int this_cpu,
2007                    unsigned long *imbalance, enum idle_type idle, int *sd_idle)
2008 {
2009         struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2010         unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2011         unsigned long max_pull;
2012         int load_idx;
2013
2014         max_load = this_load = total_load = total_pwr = 0;
2015         if (idle == NOT_IDLE)
2016                 load_idx = sd->busy_idx;
2017         else if (idle == NEWLY_IDLE)
2018                 load_idx = sd->newidle_idx;
2019         else
2020                 load_idx = sd->idle_idx;
2021
2022         do {
2023                 unsigned long load;
2024                 int local_group;
2025                 int i;
2026
2027                 local_group = cpu_isset(this_cpu, group->cpumask);
2028
2029                 /* Tally up the load of all CPUs in the group */
2030                 avg_load = 0;
2031
2032                 for_each_cpu_mask(i, group->cpumask) {
2033                         if (*sd_idle && !idle_cpu(i))
2034                                 *sd_idle = 0;
2035
2036                         /* Bias balancing toward cpus of our domain */
2037                         if (local_group)
2038                                 load = __target_load(i, load_idx, idle);
2039                         else
2040                                 load = __source_load(i, load_idx, idle);
2041
2042                         avg_load += load;
2043                 }
2044
2045                 total_load += avg_load;
2046                 total_pwr += group->cpu_power;
2047
2048                 /* Adjust by relative CPU power of the group */
2049                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2050
2051                 if (local_group) {
2052                         this_load = avg_load;
2053                         this = group;
2054                 } else if (avg_load > max_load) {
2055                         max_load = avg_load;
2056                         busiest = group;
2057                 }
2058                 group = group->next;
2059         } while (group != sd->groups);
2060
2061         if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
2062                 goto out_balanced;
2063
2064         avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2065
2066         if (this_load >= avg_load ||
2067                         100*max_load <= sd->imbalance_pct*this_load)
2068                 goto out_balanced;
2069
2070         /*
2071          * We're trying to get all the cpus to the average_load, so we don't
2072          * want to push ourselves above the average load, nor do we wish to
2073          * reduce the max loaded cpu below the average load, as either of these
2074          * actions would just result in more rebalancing later, and ping-pong
2075          * tasks around. Thus we look for the minimum possible imbalance.
2076          * Negative imbalances (*we* are more loaded than anyone else) will
2077          * be counted as no imbalance for these purposes -- we can't fix that
2078          * by pulling tasks to us.  Be careful of negative numbers as they'll
2079          * appear as very large values with unsigned longs.
2080          */
2081
2082         /* Don't want to pull so many tasks that a group would go idle */
2083         max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
2084
2085         /* How much load to actually move to equalise the imbalance */
2086         *imbalance = min(max_pull * busiest->cpu_power,
2087                                 (avg_load - this_load) * this->cpu_power)
2088                         / SCHED_LOAD_SCALE;
2089
2090         if (*imbalance < SCHED_LOAD_SCALE) {
2091                 unsigned long pwr_now = 0, pwr_move = 0;
2092                 unsigned long tmp;
2093
2094                 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2095                         *imbalance = 1;
2096                         return busiest;
2097                 }
2098
2099                 /*
2100                  * OK, we don't have enough imbalance to justify moving tasks,
2101                  * however we may be able to increase total CPU power used by
2102                  * moving them.
2103                  */
2104
2105                 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2106                 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2107                 pwr_now /= SCHED_LOAD_SCALE;
2108
2109                 /* Amount of load we'd subtract */
2110                 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2111                 if (max_load > tmp)
2112                         pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2113                                                         max_load - tmp);
2114
2115                 /* Amount of load we'd add */
2116                 if (max_load*busiest->cpu_power <
2117                                 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2118                         tmp = max_load*busiest->cpu_power/this->cpu_power;
2119                 else
2120                         tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2121                 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2122                 pwr_move /= SCHED_LOAD_SCALE;
2123
2124                 /* Move if we gain throughput */
2125                 if (pwr_move <= pwr_now)
2126                         goto out_balanced;
2127
2128                 *imbalance = 1;
2129                 return busiest;
2130         }
2131
2132         /* Get rid of the scaling factor, rounding down as we divide */
2133         *imbalance = *imbalance / SCHED_LOAD_SCALE;
2134         return busiest;
2135
2136 out_balanced:
2137
2138         *imbalance = 0;
2139         return NULL;
2140 }
2141
2142 /*
2143  * find_busiest_queue - find the busiest runqueue among the cpus in group.
2144  */
2145 static runqueue_t *find_busiest_queue(struct sched_group *group,
2146         enum idle_type idle)
2147 {
2148         unsigned long load, max_load = 0;
2149         runqueue_t *busiest = NULL;
2150         int i;
2151
2152         for_each_cpu_mask(i, group->cpumask) {
2153                 load = __source_load(i, 0, idle);
2154
2155                 if (load > max_load) {
2156                         max_load = load;
2157                         busiest = cpu_rq(i);
2158                 }
2159         }
2160
2161         return busiest;
2162 }
2163
2164 /*
2165  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2166  * so long as it is large enough.
2167  */
2168 #define MAX_PINNED_INTERVAL     512
2169
2170 /*
2171  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2172  * tasks if there is an imbalance.
2173  *
2174  * Called with this_rq unlocked.
2175  */
2176 static int load_balance(int this_cpu, runqueue_t *this_rq,
2177                         struct sched_domain *sd, enum idle_type idle)
2178 {
2179         struct sched_group *group;
2180         runqueue_t *busiest;
2181         unsigned long imbalance;
2182         int nr_moved, all_pinned = 0;
2183         int active_balance = 0;
2184         int sd_idle = 0;
2185
2186         if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2187                 sd_idle = 1;
2188
2189         schedstat_inc(sd, lb_cnt[idle]);
2190
2191         group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2192         if (!group) {
2193                 schedstat_inc(sd, lb_nobusyg[idle]);
2194                 goto out_balanced;
2195         }
2196
2197         busiest = find_busiest_queue(group, idle);
2198         if (!busiest) {
2199                 schedstat_inc(sd, lb_nobusyq[idle]);
2200                 goto out_balanced;
2201         }
2202
2203         BUG_ON(busiest == this_rq);
2204
2205         schedstat_add(sd, lb_imbalance[idle], imbalance);
2206
2207         nr_moved = 0;
2208         if (busiest->nr_running > 1) {
2209                 /*
2210                  * Attempt to move tasks. If find_busiest_group has found
2211                  * an imbalance but busiest->nr_running <= 1, the group is
2212                  * still unbalanced. nr_moved simply stays zero, so it is
2213                  * correctly treated as an imbalance.
2214                  */
2215                 double_rq_lock(this_rq, busiest);
2216                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2217                                         imbalance, sd, idle, &all_pinned);
2218                 double_rq_unlock(this_rq, busiest);
2219
2220                 /* All tasks on this runqueue were pinned by CPU affinity */
2221                 if (unlikely(all_pinned))
2222                         goto out_balanced;
2223         }
2224
2225         if (!nr_moved) {
2226                 schedstat_inc(sd, lb_failed[idle]);
2227                 sd->nr_balance_failed++;
2228
2229                 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2230
2231                         spin_lock(&busiest->lock);
2232
2233                         /* don't kick the migration_thread, if the curr
2234                          * task on busiest cpu can't be moved to this_cpu
2235                          */
2236                         if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2237                                 spin_unlock(&busiest->lock);
2238                                 all_pinned = 1;
2239                                 goto out_one_pinned;
2240                         }
2241
2242                         if (!busiest->active_balance) {
2243                                 busiest->active_balance = 1;
2244                                 busiest->push_cpu = this_cpu;
2245                                 active_balance = 1;
2246                         }
2247                         spin_unlock(&busiest->lock);
2248                         if (active_balance)
2249                                 wake_up_process(busiest->migration_thread);
2250
2251                         /*
2252                          * We've kicked active balancing, reset the failure
2253                          * counter.
2254                          */
2255                         sd->nr_balance_failed = sd->cache_nice_tries+1;
2256                 }
2257         } else
2258                 sd->nr_balance_failed = 0;
2259
2260         if (likely(!active_balance)) {
2261                 /* We were unbalanced, so reset the balancing interval */
2262                 sd->balance_interval = sd->min_interval;
2263         } else {
2264                 /*
2265                  * If we've begun active balancing, start to back off. This
2266                  * case may not be covered by the all_pinned logic if there
2267                  * is only 1 task on the busy runqueue (because we don't call
2268                  * move_tasks).
2269                  */
2270                 if (sd->balance_interval < sd->max_interval)
2271                         sd->balance_interval *= 2;
2272         }
2273
2274         if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2275                 return -1;
2276         return nr_moved;
2277
2278 out_balanced:
2279         schedstat_inc(sd, lb_balanced[idle]);
2280
2281         sd->nr_balance_failed = 0;
2282
2283 out_one_pinned:
2284         /* tune up the balancing interval */
2285         if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2286                         (sd->balance_interval < sd->max_interval))
2287                 sd->balance_interval *= 2;
2288
2289         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2290                 return -1;
2291         return 0;
2292 }
2293
2294 /*
2295  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2296  * tasks if there is an imbalance.
2297  *
2298  * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2299  * this_rq is locked.
2300  */
2301 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2302                                 struct sched_domain *sd)
2303 {
2304         struct sched_group *group;
2305         runqueue_t *busiest = NULL;
2306         unsigned long imbalance;
2307         int nr_moved = 0;
2308         int sd_idle = 0;
2309
2310         if (sd->flags & SD_SHARE_CPUPOWER)
2311                 sd_idle = 1;
2312
2313         schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2314         group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2315         if (!group) {
2316                 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2317                 goto out_balanced;
2318         }
2319
2320         busiest = find_busiest_queue(group, NEWLY_IDLE);
2321         if (!busiest) {
2322                 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2323                 goto out_balanced;
2324         }
2325
2326         BUG_ON(busiest == this_rq);
2327
2328         schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2329
2330         nr_moved = 0;
2331         if (busiest->nr_running > 1) {
2332                 /* Attempt to move tasks */
2333                 double_lock_balance(this_rq, busiest);
2334                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2335                                         imbalance, sd, NEWLY_IDLE, NULL);
2336                 spin_unlock(&busiest->lock);
2337         }
2338
2339         if (!nr_moved) {
2340                 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2341                 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2342                         return -1;
2343         } else
2344                 sd->nr_balance_failed = 0;
2345
2346         return nr_moved;
2347
2348 out_balanced:
2349         schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2350         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2351                 return -1;
2352         sd->nr_balance_failed = 0;
2353         return 0;
2354 }
2355
2356 /*
2357  * idle_balance is called by schedule() if this_cpu is about to become
2358  * idle. Attempts to pull tasks from other CPUs.
2359  */
2360 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2361 {
2362         struct sched_domain *sd;
2363
2364         for_each_domain(this_cpu, sd) {
2365                 if (sd->flags & SD_BALANCE_NEWIDLE) {
2366                         if (load_balance_newidle(this_cpu, this_rq, sd)) {
2367                                 /* We've pulled tasks over so stop searching */
2368                                 break;
2369                         }
2370                 }
2371         }
2372 }
2373
2374 /*
2375  * active_load_balance is run by migration threads. It pushes running tasks
2376  * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2377  * running on each physical CPU where possible, and avoids physical /
2378  * logical imbalances.
2379  *
2380  * Called with busiest_rq locked.
2381  */
2382 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2383 {
2384         struct sched_domain *sd;
2385         runqueue_t *target_rq;
2386         int target_cpu = busiest_rq->push_cpu;
2387
2388         if (busiest_rq->nr_running <= 1)
2389                 /* no task to move */
2390                 return;
2391
2392         target_rq = cpu_rq(target_cpu);
2393
2394         /*
2395          * This condition is "impossible", if it occurs
2396          * we need to fix it.  Originally reported by
2397          * Bjorn Helgaas on a 128-cpu setup.
2398          */
2399         BUG_ON(busiest_rq == target_rq);
2400
2401         /* move a task from busiest_rq to target_rq */
2402         double_lock_balance(busiest_rq, target_rq);
2403
2404         /* Search for an sd spanning us and the target CPU. */
2405         for_each_domain(target_cpu, sd)
2406                 if ((sd->flags & SD_LOAD_BALANCE) &&
2407                         cpu_isset(busiest_cpu, sd->span))
2408                                 break;
2409
2410         if (unlikely(sd == NULL))
2411                 goto out;
2412
2413         schedstat_inc(sd, alb_cnt);
2414
2415         if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2416                 schedstat_inc(sd, alb_pushed);
2417         else
2418                 schedstat_inc(sd, alb_failed);
2419 out:
2420         spin_unlock(&target_rq->lock);
2421 }
2422
2423 /*
2424  * rebalance_tick will get called every timer tick, on every CPU.
2425  *
2426  * It checks each scheduling domain to see if it is due to be balanced,
2427  * and initiates a balancing operation if so.
2428  *
2429  * Balancing parameters are set up in arch_init_sched_domains.
2430  */
2431
2432 /* Don't have all balancing operations going off at once */
2433 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2434
2435 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2436                            enum idle_type idle)
2437 {
2438         unsigned long old_load, this_load;
2439         unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2440         struct sched_domain *sd;
2441         int i;
2442
2443         this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2444         /* Update our load */
2445         for (i = 0; i < 3; i++) {
2446                 unsigned long new_load = this_load;
2447                 int scale = 1 << i;
2448                 old_load = this_rq->cpu_load[i];
2449                 /*
2450                  * Round up the averaging division if load is increasing. This
2451                  * prevents us from getting stuck on 9 if the load is 10, for
2452                  * example.
2453                  */
2454                 if (new_load > old_load)
2455                         new_load += scale-1;
2456                 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2457         }
2458
2459         for_each_domain(this_cpu, sd) {
2460                 unsigned long interval;
2461
2462                 if (!(sd->flags & SD_LOAD_BALANCE))
2463                         continue;
2464
2465                 interval = sd->balance_interval;
2466                 if (idle != SCHED_IDLE)
2467                         interval *= sd->busy_factor;
2468
2469                 /* scale ms to jiffies */
2470                 interval = msecs_to_jiffies(interval);
2471                 if (unlikely(!interval))
2472                         interval = 1;
2473
2474                 if (j - sd->last_balance >= interval) {
2475                         if (load_balance(this_cpu, this_rq, sd, idle)) {
2476                                 /*
2477                                  * We've pulled tasks over so either we're no
2478                                  * longer idle, or one of our SMT siblings is
2479                                  * not idle.
2480                                  */
2481                                 idle = NOT_IDLE;
2482                         }
2483                         sd->last_balance += interval;
2484                 }
2485         }
2486 }
2487 #else
2488 /*
2489  * on UP we do not need to balance between CPUs:
2490  */
2491 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2492 {
2493 }
2494 static inline void idle_balance(int cpu, runqueue_t *rq)
2495 {
2496 }
2497 #endif
2498
2499 static inline int wake_priority_sleeper(runqueue_t *rq)
2500 {
2501         int ret = 0;
2502 #ifdef CONFIG_SCHED_SMT
2503         spin_lock(&rq->lock);
2504         /*
2505          * If an SMT sibling task has been put to sleep for priority
2506          * reasons reschedule the idle task to see if it can now run.
2507          */
2508         if (rq->nr_running) {
2509                 resched_task(rq->idle);
2510                 ret = 1;
2511         }
2512         spin_unlock(&rq->lock);
2513 #endif
2514         return ret;
2515 }
2516
2517 DEFINE_PER_CPU(struct kernel_stat, kstat);
2518
2519 EXPORT_PER_CPU_SYMBOL(kstat);
2520
2521 /*
2522  * This is called on clock ticks and on context switches.
2523  * Bank in p->sched_time the ns elapsed since the last tick or switch.
2524  */
2525 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2526                                     unsigned long long now)
2527 {
2528         unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2529         p->sched_time += now - last;
2530 }
2531
2532 /*
2533  * Return current->sched_time plus any more ns on the sched_clock
2534  * that have not yet been banked.
2535  */
2536 unsigned long long current_sched_time(const task_t *tsk)
2537 {
2538         unsigned long long ns;
2539         unsigned long flags;
2540         local_irq_save(flags);
2541         ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2542         ns = tsk->sched_time + (sched_clock() - ns);
2543         local_irq_restore(flags);
2544         return ns;
2545 }
2546
2547 /*
2548  * We place interactive tasks back into the active array, if possible.
2549  *
2550  * To guarantee that this does not starve expired tasks we ignore the
2551  * interactivity of a task if the first expired task had to wait more
2552  * than a 'reasonable' amount of time. This deadline timeout is
2553  * load-dependent, as the frequency of array switched decreases with
2554  * increasing number of running tasks. We also ignore the interactivity
2555  * if a better static_prio task has expired:
2556  */
2557 #define EXPIRED_STARVING(rq) \
2558         ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2559                 (jiffies - (rq)->expired_timestamp >= \
2560                         STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2561                         ((rq)->curr->static_prio > (rq)->best_expired_prio))
2562
2563 /*
2564  * Account user cpu time to a process.
2565  * @p: the process that the cpu time gets accounted to
2566  * @hardirq_offset: the offset to subtract from hardirq_count()
2567  * @cputime: the cpu time spent in user space since the last update
2568  */
2569 void account_user_time(struct task_struct *p, cputime_t cputime)
2570 {
2571         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2572         cputime64_t tmp;
2573
2574         p->utime = cputime_add(p->utime, cputime);
2575
2576         /* Add user time to cpustat. */
2577         tmp = cputime_to_cputime64(cputime);
2578         if (TASK_NICE(p) > 0)
2579                 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2580         else
2581                 cpustat->user = cputime64_add(cpustat->user, tmp);
2582 }
2583
2584 /*
2585  * Account system cpu time to a process.
2586  * @p: the process that the cpu time gets accounted to
2587  * @hardirq_offset: the offset to subtract from hardirq_count()
2588  * @cputime: the cpu time spent in kernel space since the last update
2589  */
2590 void account_system_time(struct task_struct *p, int hardirq_offset,
2591                          cputime_t cputime)
2592 {
2593         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2594         runqueue_t *rq = this_rq();
2595         cputime64_t tmp;
2596
2597         p->stime = cputime_add(p->stime, cputime);
2598
2599         /* Add system time to cpustat. */
2600         tmp = cputime_to_cputime64(cputime);
2601         if (hardirq_count() - hardirq_offset)
2602                 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2603         else if (softirq_count())
2604                 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2605         else if (p != rq->idle)
2606                 cpustat->system = cputime64_add(cpustat->system, tmp);
2607         else if (atomic_read(&rq->nr_iowait) > 0)
2608                 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2609         else
2610                 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2611         /* Account for system time used */
2612         acct_update_integrals(p);
2613 }
2614
2615 /*
2616  * Account for involuntary wait time.
2617  * @p: the process from which the cpu time has been stolen
2618  * @steal: the cpu time spent in involuntary wait
2619  */
2620 void account_steal_time(struct task_struct *p, cputime_t steal)
2621 {
2622         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2623         cputime64_t tmp = cputime_to_cputime64(steal);
2624         runqueue_t *rq = this_rq();
2625
2626         if (p == rq->idle) {
2627                 p->stime = cputime_add(p->stime, steal);
2628                 if (atomic_read(&rq->nr_iowait) > 0)
2629                         cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2630                 else
2631                         cpustat->idle = cputime64_add(cpustat->idle, tmp);
2632         } else
2633                 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2634 }
2635
2636 /*
2637  * This function gets called by the timer code, with HZ frequency.
2638  * We call it with interrupts disabled.
2639  *
2640  * It also gets called by the fork code, when changing the parent's
2641  * timeslices.
2642  */
2643 void scheduler_tick(void)
2644 {
2645         int cpu = smp_processor_id();
2646         runqueue_t *rq = this_rq();
2647         task_t *p = current;
2648         unsigned long long now = sched_clock();
2649
2650         update_cpu_clock(p, rq, now);
2651
2652         rq->timestamp_last_tick = now;
2653
2654         if (p == rq->idle) {
2655                 if (wake_priority_sleeper(rq))
2656                         goto out;
2657                 rebalance_tick(cpu, rq, SCHED_IDLE);
2658                 return;
2659         }
2660
2661         /* Task might have expired already, but not scheduled off yet */
2662         if (p->array != rq->active) {
2663                 set_tsk_need_resched(p);
2664                 goto out;
2665         }
2666         spin_lock(&rq->lock);
2667         /*
2668          * The task was running during this tick - update the
2669          * time slice counter. Note: we do not update a thread's
2670          * priority until it either goes to sleep or uses up its
2671          * timeslice. This makes it possible for interactive tasks
2672          * to use up their timeslices at their highest priority levels.
2673          */
2674         if (rt_task(p)) {
2675                 /*
2676                  * RR tasks need a special form of timeslice management.
2677                  * FIFO tasks have no timeslices.
2678                  */
2679                 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2680                         p->time_slice = task_timeslice(p);
2681                         p->first_time_slice = 0;
2682                         set_tsk_need_resched(p);
2683
2684                         /* put it at the end of the queue: */
2685                         requeue_task(p, rq->active);
2686                 }
2687                 goto out_unlock;
2688         }
2689         if (!--p->time_slice) {
2690                 dequeue_task(p, rq->active);
2691                 set_tsk_need_resched(p);
2692                 p->prio = effective_prio(p);
2693                 p->time_slice = task_timeslice(p);
2694                 p->first_time_slice = 0;
2695
2696                 if (!rq->expired_timestamp)
2697                         rq->expired_timestamp = jiffies;
2698                 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2699                         enqueue_task(p, rq->expired);
2700                         if (p->static_prio < rq->best_expired_prio)
2701                                 rq->best_expired_prio = p->static_prio;
2702                 } else
2703                         enqueue_task(p, rq->active);
2704         } else {
2705                 /*
2706                  * Prevent a too long timeslice allowing a task to monopolize
2707                  * the CPU. We do this by splitting up the timeslice into
2708                  * smaller pieces.
2709                  *
2710                  * Note: this does not mean the task's timeslices expire or
2711                  * get lost in any way, they just might be preempted by
2712                  * another task of equal priority. (one with higher
2713                  * priority would have preempted this task already.) We
2714                  * requeue this task to the end of the list on this priority
2715                  * level, which is in essence a round-robin of tasks with
2716                  * equal priority.
2717                  *
2718                  * This only applies to tasks in the interactive
2719                  * delta range with at least TIMESLICE_GRANULARITY to requeue.
2720                  */
2721                 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2722                         p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2723                         (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2724                         (p->array == rq->active)) {
2725
2726                         requeue_task(p, rq->active);
2727                         set_tsk_need_resched(p);
2728                 }
2729         }
2730 out_unlock:
2731         spin_unlock(&rq->lock);
2732 out:
2733         rebalance_tick(cpu, rq, NOT_IDLE);
2734 }
2735
2736 #ifdef CONFIG_SCHED_SMT
2737 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2738 {
2739         /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2740         if (rq->curr == rq->idle && rq->nr_running)
2741                 resched_task(rq->idle);
2742 }
2743
2744 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2745 {
2746         struct sched_domain *tmp, *sd = NULL;
2747         cpumask_t sibling_map;
2748         int i;
2749
2750         for_each_domain(this_cpu, tmp)
2751                 if (tmp->flags & SD_SHARE_CPUPOWER)
2752                         sd = tmp;
2753
2754         if (!sd)
2755                 return;
2756
2757         /*
2758          * Unlock the current runqueue because we have to lock in
2759          * CPU order to avoid deadlocks. Caller knows that we might
2760          * unlock. We keep IRQs disabled.
2761          */
2762         spin_unlock(&this_rq->lock);
2763
2764         sibling_map = sd->span;
2765
2766         for_each_cpu_mask(i, sibling_map)
2767                 spin_lock(&cpu_rq(i)->lock);
2768         /*
2769          * We clear this CPU from the mask. This both simplifies the
2770          * inner loop and keps this_rq locked when we exit:
2771          */
2772         cpu_clear(this_cpu, sibling_map);
2773
2774         for_each_cpu_mask(i, sibling_map) {
2775                 runqueue_t *smt_rq = cpu_rq(i);
2776
2777                 wakeup_busy_runqueue(smt_rq);
2778         }
2779
2780         for_each_cpu_mask(i, sibling_map)
2781                 spin_unlock(&cpu_rq(i)->lock);
2782         /*
2783          * We exit with this_cpu's rq still held and IRQs
2784          * still disabled:
2785          */
2786 }
2787
2788 /*
2789  * number of 'lost' timeslices this task wont be able to fully
2790  * utilize, if another task runs on a sibling. This models the
2791  * slowdown effect of other tasks running on siblings:
2792  */
2793 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2794 {
2795         return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2796 }
2797
2798 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2799 {
2800         struct sched_domain *tmp, *sd = NULL;
2801         cpumask_t sibling_map;
2802         prio_array_t *array;
2803         int ret = 0, i;
2804         task_t *p;
2805
2806         for_each_domain(this_cpu, tmp)
2807                 if (tmp->flags & SD_SHARE_CPUPOWER)
2808                         sd = tmp;
2809
2810         if (!sd)
2811                 return 0;
2812
2813         /*
2814          * The same locking rules and details apply as for
2815          * wake_sleeping_dependent():
2816          */
2817         spin_unlock(&this_rq->lock);
2818         sibling_map = sd->span;
2819         for_each_cpu_mask(i, sibling_map)
2820                 spin_lock(&cpu_rq(i)->lock);
2821         cpu_clear(this_cpu, sibling_map);
2822
2823         /*
2824          * Establish next task to be run - it might have gone away because
2825          * we released the runqueue lock above:
2826          */
2827         if (!this_rq->nr_running)
2828                 goto out_unlock;
2829         array = this_rq->active;
2830         if (!array->nr_active)
2831                 array = this_rq->expired;
2832         BUG_ON(!array->nr_active);
2833
2834         p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2835                 task_t, run_list);
2836
2837         for_each_cpu_mask(i, sibling_map) {
2838                 runqueue_t *smt_rq = cpu_rq(i);
2839                 task_t *smt_curr = smt_rq->curr;
2840
2841                 /* Kernel threads do not participate in dependent sleeping */
2842                 if (!p->mm || !smt_curr->mm || rt_task(p))
2843                         goto check_smt_task;
2844
2845                 /*
2846                  * If a user task with lower static priority than the
2847                  * running task on the SMT sibling is trying to schedule,
2848                  * delay it till there is proportionately less timeslice
2849                  * left of the sibling task to prevent a lower priority
2850                  * task from using an unfair proportion of the
2851                  * physical cpu's resources. -ck
2852                  */
2853                 if (rt_task(smt_curr)) {
2854                         /*
2855                          * With real time tasks we run non-rt tasks only
2856                          * per_cpu_gain% of the time.
2857                          */
2858                         if ((jiffies % DEF_TIMESLICE) >
2859                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2860                                         ret = 1;
2861                 } else
2862                         if (smt_curr->static_prio < p->static_prio &&
2863                                 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2864                                 smt_slice(smt_curr, sd) > task_timeslice(p))
2865                                         ret = 1;
2866
2867 check_smt_task:
2868                 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2869                         rt_task(smt_curr))
2870                                 continue;
2871                 if (!p->mm) {
2872                         wakeup_busy_runqueue(smt_rq);
2873                         continue;
2874                 }
2875
2876                 /*
2877                  * Reschedule a lower priority task on the SMT sibling for
2878                  * it to be put to sleep, or wake it up if it has been put to
2879                  * sleep for priority reasons to see if it should run now.
2880                  */
2881                 if (rt_task(p)) {
2882                         if ((jiffies % DEF_TIMESLICE) >
2883                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2884                                         resched_task(smt_curr);
2885                 } else {
2886                         if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2887                                 smt_slice(p, sd) > task_timeslice(smt_curr))
2888                                         resched_task(smt_curr);
2889                         else
2890                                 wakeup_busy_runqueue(smt_rq);
2891                 }
2892         }
2893 out_unlock:
2894         for_each_cpu_mask(i, sibling_map)
2895                 spin_unlock(&cpu_rq(i)->lock);
2896         return ret;
2897 }
2898 #else
2899 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2900 {
2901 }
2902
2903 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2904 {
2905         return 0;
2906 }
2907 #endif
2908
2909 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2910
2911 void fastcall add_preempt_count(int val)
2912 {
2913         /*
2914          * Underflow?
2915          */
2916         BUG_ON((preempt_count() < 0));
2917         preempt_count() += val;
2918         /*
2919          * Spinlock count overflowing soon?
2920          */
2921         BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2922 }
2923 EXPORT_SYMBOL(add_preempt_count);
2924
2925 void fastcall sub_preempt_count(int val)
2926 {
2927         /*
2928          * Underflow?
2929          */
2930         BUG_ON(val > preempt_count());
2931         /*
2932          * Is the spinlock portion underflowing?
2933          */
2934         BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2935         preempt_count() -= val;
2936 }
2937 EXPORT_SYMBOL(sub_preempt_count);
2938
2939 #endif
2940
2941 /*
2942  * schedule() is the main scheduler function.
2943  */
2944 asmlinkage void __sched schedule(void)
2945 {
2946         long *switch_count;
2947         task_t *prev, *next;
2948         runqueue_t *rq;
2949         prio_array_t *array;
2950         struct list_head *queue;
2951         unsigned long long now;
2952         unsigned long run_time;
2953         int cpu, idx, new_prio;
2954
2955         /*
2956          * Test if we are atomic.  Since do_exit() needs to call into
2957          * schedule() atomically, we ignore that path for now.
2958          * Otherwise, whine if we are scheduling when we should not be.
2959          */
2960         if (likely(!current->exit_state)) {
2961                 if (unlikely(in_atomic())) {
2962                         printk(KERN_ERR "scheduling while atomic: "
2963                                 "%s/0x%08x/%d\n",
2964                                 current->comm, preempt_count(), current->pid);
2965                         dump_stack();
2966                 }
2967         }
2968         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2969
2970 need_resched:
2971         preempt_disable();
2972         prev = current;
2973         release_kernel_lock(prev);
2974 need_resched_nonpreemptible:
2975         rq = this_rq();
2976
2977         /*
2978          * The idle thread is not allowed to schedule!
2979          * Remove this check after it has been exercised a bit.
2980          */
2981         if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2982                 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2983                 dump_stack();
2984         }
2985
2986         schedstat_inc(rq, sched_cnt);
2987         now = sched_clock();
2988         if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2989                 run_time = now - prev->timestamp;
2990                 if (unlikely((long long)(now - prev->timestamp) < 0))
2991                         run_time = 0;
2992         } else
2993                 run_time = NS_MAX_SLEEP_AVG;
2994
2995         /*
2996          * Tasks charged proportionately less run_time at high sleep_avg to
2997          * delay them losing their interactive status
2998          */
2999         run_time /= (CURRENT_BONUS(prev) ? : 1);
3000
3001         spin_lock_irq(&rq->lock);
3002
3003         if (unlikely(prev->flags & PF_DEAD))
3004                 prev->state = EXIT_DEAD;
3005
3006         switch_count = &prev->nivcsw;
3007         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3008                 switch_count = &prev->nvcsw;
3009                 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3010                                 unlikely(signal_pending(prev))))
3011                         prev->state = TASK_RUNNING;
3012                 else {
3013                         if (prev->state == TASK_UNINTERRUPTIBLE)
3014                                 rq->nr_uninterruptible++;
3015                         deactivate_task(prev, rq);
3016                 }
3017         }
3018
3019         cpu = smp_processor_id();
3020         if (unlikely(!rq->nr_running)) {
3021 go_idle:
3022                 idle_balance(cpu, rq);
3023                 if (!rq->nr_running) {
3024                         next = rq->idle;
3025                         rq->expired_timestamp = 0;
3026                         wake_sleeping_dependent(cpu, rq);
3027                         /*
3028                          * wake_sleeping_dependent() might have released
3029                          * the runqueue, so break out if we got new
3030                          * tasks meanwhile:
3031                          */
3032                         if (!rq->nr_running)
3033                                 goto switch_tasks;
3034                 }
3035         } else {
3036                 if (dependent_sleeper(cpu, rq)) {
3037                         next = rq->idle;
3038                         goto switch_tasks;
3039                 }
3040                 /*
3041                  * dependent_sleeper() releases and reacquires the runqueue
3042                  * lock, hence go into the idle loop if the rq went
3043                  * empty meanwhile:
3044                  */
3045                 if (unlikely(!rq->nr_running))
3046                         goto go_idle;
3047         }
3048
3049         array = rq->active;
3050         if (unlikely(!array->nr_active)) {
3051                 /*
3052                  * Switch the active and expired arrays.
3053                  */
3054                 schedstat_inc(rq, sched_switch);
3055                 rq->active = rq->expired;
3056                 rq->expired = array;
3057                 array = rq->active;
3058                 rq->expired_timestamp = 0;
3059                 rq->best_expired_prio = MAX_PRIO;
3060         }
3061
3062         idx = sched_find_first_bit(array->bitmap);
3063         queue = array->queue + idx;
3064         next = list_entry(queue->next, task_t, run_list);
3065
3066         if (!rt_task(next) && next->activated > 0) {
3067                 unsigned long long delta = now - next->timestamp;
3068                 if (unlikely((long long)(now - next->timestamp) < 0))
3069                         delta = 0;
3070
3071                 if (next->activated == 1)
3072                         delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3073
3074                 array = next->array;
3075                 new_prio = recalc_task_prio(next, next->timestamp + delta);
3076
3077                 if (unlikely(next->prio != new_prio)) {
3078                         dequeue_task(next, array);
3079                         next->prio = new_prio;
3080                         enqueue_task(next, array);
3081                 } else
3082                         requeue_task(next, array);
3083         }
3084         next->activated = 0;
3085 switch_tasks:
3086         if (next == rq->idle)
3087                 schedstat_inc(rq, sched_goidle);
3088         prefetch(next);
3089         prefetch_stack(next);
3090         clear_tsk_need_resched(prev);
3091         rcu_qsctr_inc(task_cpu(prev));
3092
3093         update_cpu_clock(prev, rq, now);
3094
3095         prev->sleep_avg -= run_time;
3096         if ((long)prev->sleep_avg <= 0)
3097                 prev->sleep_avg = 0;
3098         prev->timestamp = prev->last_ran = now;
3099
3100         sched_info_switch(prev, next);
3101         if (likely(prev != next)) {
3102                 next->timestamp = now;
3103                 rq->nr_switches++;
3104                 rq->curr = next;
3105                 ++*switch_count;
3106
3107                 prepare_task_switch(rq, next);
3108                 prev = context_switch(rq, prev, next);
3109                 barrier();
3110                 /*
3111                  * this_rq must be evaluated again because prev may have moved
3112                  * CPUs since it called schedule(), thus the 'rq' on its stack
3113                  * frame will be invalid.
3114                  */
3115                 finish_task_switch(this_rq(), prev);
3116         } else
3117                 spin_unlock_irq(&rq->lock);
3118
3119         prev = current;
3120         if (unlikely(reacquire_kernel_lock(prev) < 0))
3121                 goto need_resched_nonpreemptible;
3122         preempt_enable_no_resched();
3123         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3124                 goto need_resched;
3125 }
3126
3127 EXPORT_SYMBOL(schedule);
3128
3129 #ifdef CONFIG_PREEMPT
3130 /*
3131  * this is is the entry point to schedule() from in-kernel preemption
3132  * off of preempt_enable.  Kernel preemptions off return from interrupt
3133  * occur there and call schedule directly.
3134  */
3135 asmlinkage void __sched preempt_schedule(void)
3136 {
3137         struct thread_info *ti = current_thread_info();
3138 #ifdef CONFIG_PREEMPT_BKL
3139         struct task_struct *task = current;
3140         int saved_lock_depth;
3141 #endif
3142         /*
3143          * If there is a non-zero preempt_count or interrupts are disabled,
3144          * we do not want to preempt the current task.  Just return..
3145          */
3146         if (unlikely(ti->preempt_count || irqs_disabled()))
3147                 return;
3148
3149 need_resched:
3150         add_preempt_count(PREEMPT_ACTIVE);
3151         /*
3152          * We keep the big kernel semaphore locked, but we
3153          * clear ->lock_depth so that schedule() doesnt
3154          * auto-release the semaphore:
3155          */
3156 #ifdef CONFIG_PREEMPT_BKL
3157         saved_lock_depth = task->lock_depth;
3158         task->lock_depth = -1;
3159 #endif
3160         schedule();
3161 #ifdef CONFIG_PREEMPT_BKL
3162         task->lock_depth = saved_lock_depth;
3163 #endif
3164         sub_preempt_count(PREEMPT_ACTIVE);
3165
3166         /* we could miss a preemption opportunity between schedule and now */
3167         barrier();
3168         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3169                 goto need_resched;
3170 }
3171
3172 EXPORT_SYMBOL(preempt_schedule);
3173
3174 /*
3175  * this is is the entry point to schedule() from kernel preemption
3176  * off of irq context.
3177  * Note, that this is called and return with irqs disabled. This will
3178  * protect us against recursive calling from irq.
3179  */
3180 asmlinkage void __sched preempt_schedule_irq(void)
3181 {
3182         struct thread_info *ti = current_thread_info();
3183 #ifdef CONFIG_PREEMPT_BKL
3184         struct task_struct *task = current;
3185         int saved_lock_depth;
3186 #endif
3187         /* Catch callers which need to be fixed*/
3188         BUG_ON(ti->preempt_count || !irqs_disabled());
3189
3190 need_resched:
3191         add_preempt_count(PREEMPT_ACTIVE);
3192         /*
3193          * We keep the big kernel semaphore locked, but we
3194          * clear ->lock_depth so that schedule() doesnt
3195          * auto-release the semaphore:
3196          */
3197 #ifdef CONFIG_PREEMPT_BKL
3198         saved_lock_depth = task->lock_depth;
3199         task->lock_depth = -1;
3200 #endif
3201         local_irq_enable();
3202         schedule();
3203         local_irq_disable();
3204 #ifdef CONFIG_PREEMPT_BKL
3205         task->lock_depth = saved_lock_depth;
3206 #endif
3207         sub_preempt_count(PREEMPT_ACTIVE);
3208
3209         /* we could miss a preemption opportunity between schedule and now */
3210         barrier();
3211         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3212                 goto need_resched;
3213 }
3214
3215 #endif /* CONFIG_PREEMPT */
3216
3217 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3218                           void *key)
3219 {
3220         task_t *p = curr->private;
3221         return try_to_wake_up(p, mode, sync);
3222 }
3223
3224 EXPORT_SYMBOL(default_wake_function);
3225
3226 /*
3227  * The core wakeup function.  Non-exclusive wakeups (nr_exclusive == 0) just
3228  * wake everything up.  If it's an exclusive wakeup (nr_exclusive == small +ve
3229  * number) then we wake all the non-exclusive tasks and one exclusive task.
3230  *
3231  * There are circumstances in which we can try to wake a task which has already
3232  * started to run but is not in state TASK_RUNNING.  try_to_wake_up() returns
3233  * zero in this (rare) case, and we handle it by continuing to scan the queue.
3234  */
3235 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3236                              int nr_exclusive, int sync, void *key)
3237 {
3238         struct list_head *tmp, *next;
3239
3240         list_for_each_safe(tmp, next, &q->task_list) {
3241                 wait_queue_t *curr;
3242                 unsigned flags;
3243                 curr = list_entry(tmp, wait_queue_t, task_list);
3244                 flags = curr->flags;
3245                 if (curr->func(curr, mode, sync, key) &&
3246                     (flags & WQ_FLAG_EXCLUSIVE) &&
3247                     !--nr_exclusive)
3248                         break;
3249         }
3250 }
3251
3252 /**
3253  * __wake_up - wake up threads blocked on a waitqueue.
3254  * @q: the waitqueue
3255  * @mode: which threads
3256  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3257  * @key: is directly passed to the wakeup function
3258  */
3259 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3260                         int nr_exclusive, void *key)
3261 {
3262         unsigned long flags;
3263
3264         spin_lock_irqsave(&q->lock, flags);
3265         __wake_up_common(q, mode, nr_exclusive, 0, key);
3266         spin_unlock_irqrestore(&q->lock, flags);
3267 }
3268
3269 EXPORT_SYMBOL(__wake_up);
3270
3271 /*
3272  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3273  */
3274 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3275 {
3276         __wake_up_common(q, mode, 1, 0, NULL);
3277 }
3278
3279 /**
3280  * __wake_up_sync - wake up threads blocked on a waitqueue.
3281  * @q: the waitqueue
3282  * @mode: which threads
3283  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3284  *
3285  * The sync wakeup differs that the waker knows that it will schedule
3286  * away soon, so while the target thread will be woken up, it will not
3287  * be migrated to another CPU - ie. the two threads are 'synchronized'
3288  * with each other. This can prevent needless bouncing between CPUs.
3289  *
3290  * On UP it can prevent extra preemption.
3291  */
3292 void fastcall
3293 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3294 {
3295         unsigned long flags;
3296         int sync = 1;
3297
3298         if (unlikely(!q))
3299                 return;
3300
3301         if (unlikely(!nr_exclusive))
3302                 sync = 0;
3303
3304         spin_lock_irqsave(&q->lock, flags);
3305         __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3306         spin_unlock_irqrestore(&q->lock, flags);
3307 }
3308 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3309
3310 void fastcall complete(struct completion *x)
3311 {
3312         unsigned long flags;
3313
3314         spin_lock_irqsave(&x->wait.lock, flags);
3315         x->done++;
3316         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3317                          1, 0, NULL);
3318         spin_unlock_irqrestore(&x->wait.lock, flags);
3319 }
3320 EXPORT_SYMBOL(complete);
3321
3322 void fastcall complete_all(struct completion *x)
3323 {
3324         unsigned long flags;
3325
3326         spin_lock_irqsave(&x->wait.lock, flags);
3327         x->done += UINT_MAX/2;
3328         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3329                          0, 0, NULL);
3330         spin_unlock_irqrestore(&x->wait.lock, flags);
3331 }
3332 EXPORT_SYMBOL(complete_all);
3333
3334 void fastcall __sched wait_for_completion(struct completion *x)
3335 {
3336         might_sleep();
3337         spin_lock_irq(&x->wait.lock);
3338         if (!x->done) {
3339                 DECLARE_WAITQUEUE(wait, current);
3340
3341                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3342                 __add_wait_queue_tail(&x->wait, &wait);
3343                 do {
3344                         __set_current_state(TASK_UNINTERRUPTIBLE);
3345                         spin_unlock_irq(&x->wait.lock);
3346                         schedule();
3347                         spin_lock_irq(&x->wait.lock);
3348                 } while (!x->done);
3349                 __remove_wait_queue(&x->wait, &wait);
3350         }
3351         x->done--;
3352         spin_unlock_irq(&x->wait.lock);
3353 }
3354 EXPORT_SYMBOL(wait_for_completion);
3355
3356 unsigned long fastcall __sched
3357 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3358 {
3359         might_sleep();
3360
3361         spin_lock_irq(&x->wait.lock);
3362         if (!x->done) {
3363                 DECLARE_WAITQUEUE(wait, current);
3364
3365                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3366                 __add_wait_queue_tail(&x->wait, &wait);
3367                 do {
3368                         __set_current_state(TASK_UNINTERRUPTIBLE);
3369                         spin_unlock_irq(&x->wait.lock);
3370                         timeout = schedule_timeout(timeout);
3371                         spin_lock_irq(&x->wait.lock);
3372                         if (!timeout) {
3373                                 __remove_wait_queue(&x->wait, &wait);
3374                                 goto out;
3375                         }
3376                 } while (!x->done);
3377                 __remove_wait_queue(&x->wait, &wait);
3378         }
3379         x->done--;
3380 out:
3381         spin_unlock_irq(&x->wait.lock);
3382         return timeout;
3383 }
3384 EXPORT_SYMBOL(wait_for_completion_timeout);
3385
3386 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3387 {
3388         int ret = 0;
3389
3390         might_sleep();
3391
3392         spin_lock_irq(&x->wait.lock);
3393         if (!x->done) {
3394                 DECLARE_WAITQUEUE(wait, current);
3395
3396                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3397                 __add_wait_queue_tail(&x->wait, &wait);
3398                 do {
3399                         if (signal_pending(current)) {
3400                                 ret = -ERESTARTSYS;
3401                                 __remove_wait_queue(&x->wait, &wait);
3402                                 goto out;
3403                         }
3404                         __set_current_state(TASK_INTERRUPTIBLE);
3405                         spin_unlock_irq(&x->wait.lock);
3406                         schedule();
3407                         spin_lock_irq(&x->wait.lock);
3408                 } while (!x->done);
3409                 __remove_wait_queue(&x->wait, &wait);
3410         }
3411         x->done--;
3412 out:
3413         spin_unlock_irq(&x->wait.lock);
3414
3415         return ret;
3416 }
3417 EXPORT_SYMBOL(wait_for_completion_interruptible);
3418
3419 unsigned long fastcall __sched
3420 wait_for_completion_interruptible_timeout(struct completion *x,
3421                                           unsigned long timeout)
3422 {
3423         might_sleep();
3424
3425         spin_lock_irq(&x->wait.lock);
3426         if (!x->done) {
3427                 DECLARE_WAITQUEUE(wait, current);
3428
3429                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3430                 __add_wait_queue_tail(&x->wait, &wait);
3431                 do {
3432                         if (signal_pending(current)) {
3433                                 timeout = -ERESTARTSYS;
3434                                 __remove_wait_queue(&x->wait, &wait);
3435                                 goto out;
3436                         }
3437                         __set_current_state(TASK_INTERRUPTIBLE);
3438                         spin_unlock_irq(&x->wait.lock);
3439                         timeout = schedule_timeout(timeout);
3440                         spin_lock_irq(&x->wait.lock);
3441                         if (!timeout) {
3442                                 __remove_wait_queue(&x->wait, &wait);
3443                                 goto out;
3444                         }
3445                 } while (!x->done);
3446                 __remove_wait_queue(&x->wait, &wait);
3447         }
3448         x->done--;
3449 out:
3450         spin_unlock_irq(&x->wait.lock);
3451         return timeout;
3452 }
3453 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3454
3455
3456 #define SLEEP_ON_VAR                                    \
3457         unsigned long flags;                            \
3458         wait_queue_t wait;                              \
3459         init_waitqueue_entry(&wait, current);
3460
3461 #define SLEEP_ON_HEAD                                   \
3462         spin_lock_irqsave(&q->lock,flags);              \
3463         __add_wait_queue(q, &wait);                     \
3464         spin_unlock(&q->lock);
3465
3466 #define SLEEP_ON_TAIL                                   \
3467         spin_lock_irq(&q->lock);                        \
3468         __remove_wait_queue(q, &wait);                  \
3469         spin_unlock_irqrestore(&q->lock, flags);
3470
3471 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3472 {
3473         SLEEP_ON_VAR
3474
3475         current->state = TASK_INTERRUPTIBLE;
3476
3477         SLEEP_ON_HEAD
3478         schedule();
3479         SLEEP_ON_TAIL
3480 }
3481
3482 EXPORT_SYMBOL(interruptible_sleep_on);
3483
3484 long fastcall __sched
3485 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3486 {
3487         SLEEP_ON_VAR
3488
3489         current->state = TASK_INTERRUPTIBLE;
3490
3491         SLEEP_ON_HEAD
3492         timeout = schedule_timeout(timeout);
3493         SLEEP_ON_TAIL
3494
3495         return timeout;
3496 }
3497
3498 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3499
3500 void fastcall __sched sleep_on(wait_queue_head_t *q)
3501 {
3502         SLEEP_ON_VAR
3503
3504         current->state = TASK_UNINTERRUPTIBLE;
3505
3506         SLEEP_ON_HEAD
3507         schedule();
3508         SLEEP_ON_TAIL
3509 }
3510
3511 EXPORT_SYMBOL(sleep_on);
3512
3513 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3514 {
3515         SLEEP_ON_VAR
3516
3517         current->state = TASK_UNINTERRUPTIBLE;
3518
3519         SLEEP_ON_HEAD
3520         timeout = schedule_timeout(timeout);
3521         SLEEP_ON_TAIL
3522
3523         return timeout;
3524 }
3525
3526 EXPORT_SYMBOL(sleep_on_timeout);
3527
3528 void set_user_nice(task_t *p, long nice)
3529 {
3530         unsigned long flags;
3531         prio_array_t *array;
3532         runqueue_t *rq;
3533         int old_prio, new_prio, delta;
3534
3535         if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3536                 return;
3537         /*
3538          * We have to be careful, if called from sys_setpriority(),
3539          * the task might be in the middle of scheduling on another CPU.
3540          */
3541         rq = task_rq_lock(p, &flags);
3542         /*
3543          * The RT priorities are set via sched_setscheduler(), but we still
3544          * allow the 'normal' nice value to be set - but as expected
3545          * it wont have any effect on scheduling until the task is
3546          * not SCHED_NORMAL:
3547          */
3548         if (rt_task(p)) {
3549                 p->static_prio = NICE_TO_PRIO(nice);
3550                 goto out_unlock;
3551         }
3552         array = p->array;
3553         if (array) {
3554                 dequeue_task(p, array);
3555                 dec_prio_bias(rq, p->static_prio);
3556         }
3557
3558         old_prio = p->prio;
3559         new_prio = NICE_TO_PRIO(nice);
3560         delta = new_prio - old_prio;
3561         p->static_prio = NICE_TO_PRIO(nice);
3562         p->prio += delta;
3563
3564         if (array) {
3565                 enqueue_task(p, array);
3566                 inc_prio_bias(rq, p->static_prio);
3567                 /*
3568                  * If the task increased its priority or is running and
3569                  * lowered its priority, then reschedule its CPU:
3570                  */
3571                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3572                         resched_task(rq->curr);
3573         }
3574 out_unlock:
3575         task_rq_unlock(rq, &flags);
3576 }
3577
3578 EXPORT_SYMBOL(set_user_nice);
3579
3580 /*
3581  * can_nice - check if a task can reduce its nice value
3582  * @p: task
3583  * @nice: nice value
3584  */
3585 int can_nice(const task_t *p, const int nice)
3586 {
3587         /* convert nice value [19,-20] to rlimit style value [1,40] */
3588         int nice_rlim = 20 - nice;
3589         return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3590                 capable(CAP_SYS_NICE));
3591 }
3592
3593 #ifdef __ARCH_WANT_SYS_NICE
3594
3595 /*
3596  * sys_nice - change the priority of the current process.
3597  * @increment: priority increment
3598  *
3599  * sys_setpriority is a more generic, but much slower function that
3600  * does similar things.
3601  */
3602 asmlinkage long sys_nice(int increment)
3603 {
3604         int retval;
3605         long nice;
3606
3607         /*
3608          * Setpriority might change our priority at the same moment.
3609          * We don't have to worry. Conceptually one call occurs first
3610          * and we have a single winner.
3611          */
3612         if (increment < -40)
3613                 increment = -40;
3614         if (increment > 40)
3615                 increment = 40;
3616
3617         nice = PRIO_TO_NICE(current->static_prio) + increment;
3618         if (nice < -20)
3619                 nice = -20;
3620         if (nice > 19)
3621                 nice = 19;
3622
3623         if (increment < 0 && !can_nice(current, nice))
3624                 return -EPERM;
3625
3626         retval = security_task_setnice(current, nice);
3627         if (retval)
3628                 return retval;
3629
3630         set_user_nice(current, nice);
3631         return 0;
3632 }
3633
3634 #endif
3635
3636 /**
3637  * task_prio - return the priority value of a given task.
3638  * @p: the task in question.
3639  *
3640  * This is the priority value as seen by users in /proc.
3641  * RT tasks are offset by -200. Normal tasks are centered
3642  * around 0, value goes from -16 to +15.
3643  */
3644 int task_prio(const task_t *p)
3645 {
3646         return p->prio - MAX_RT_PRIO;
3647 }
3648
3649 /**
3650  * task_nice - return the nice value of a given task.
3651  * @p: the task in question.
3652  */
3653 int task_nice(const task_t *p)
3654 {
3655         return TASK_NICE(p);
3656 }
3657 EXPORT_SYMBOL_GPL(task_nice);
3658
3659 /**
3660  * idle_cpu - is a given cpu idle currently?
3661  * @cpu: the processor in question.
3662  */
3663 int idle_cpu(int cpu)
3664 {
3665         return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3666 }
3667
3668 /**
3669  * idle_task - return the idle task for a given cpu.
3670  * @cpu: the processor in question.
3671  */
3672 task_t *idle_task(int cpu)
3673 {
3674         return cpu_rq(cpu)->idle;
3675 }
3676
3677 /**
3678  * find_process_by_pid - find a process with a matching PID value.
3679  * @pid: the pid in question.
3680  */
3681 static inline task_t *find_process_by_pid(pid_t pid)
3682 {
3683         return pid ? find_task_by_pid(pid) : current;
3684 }
3685
3686 /* Actually do priority change: must hold rq lock. */
3687 static void __setscheduler(struct task_struct *p, int policy, int prio)
3688 {
3689         BUG_ON(p->array);
3690         p->policy = policy;
3691         p->rt_priority = prio;
3692         if (policy != SCHED_NORMAL)
3693                 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3694         else
3695                 p->prio = p->static_prio;
3696 }
3697
3698 /**
3699  * sched_setscheduler - change the scheduling policy and/or RT priority of
3700  * a thread.
3701  * @p: the task in question.
3702  * @policy: new policy.
3703  * @param: structure containing the new RT priority.
3704  */
3705 int sched_setscheduler(struct task_struct *p, int policy,
3706                        struct sched_param *param)
3707 {
3708         int retval;
3709         int oldprio, oldpolicy = -1;
3710         prio_array_t *array;
3711         unsigned long flags;
3712         runqueue_t *rq;
3713
3714 recheck:
3715         /* double check policy once rq lock held */
3716         if (policy < 0)
3717                 policy = oldpolicy = p->policy;
3718         else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3719                                 policy != SCHED_NORMAL)
3720                         return -EINVAL;
3721         /*
3722          * Valid priorities for SCHED_FIFO and SCHED_RR are
3723          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3724          */
3725         if (param->sched_priority < 0 ||
3726             (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3727             (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3728                 return -EINVAL;
3729         if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3730                 return -EINVAL;
3731
3732         /*
3733          * Allow unprivileged RT tasks to decrease priority:
3734          */
3735         if (!capable(CAP_SYS_NICE)) {
3736                 /* can't change policy */
3737                 if (policy != p->policy &&
3738                         !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3739                         return -EPERM;
3740                 /* can't increase priority */
3741                 if (policy != SCHED_NORMAL &&
3742                     param->sched_priority > p->rt_priority &&
3743                     param->sched_priority >
3744                                 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3745                         return -EPERM;
3746                 /* can't change other user's priorities */
3747                 if ((current->euid != p->euid) &&
3748                     (current->euid != p->uid))
3749                         return -EPERM;
3750         }
3751
3752         retval = security_task_setscheduler(p, policy, param);
3753         if (retval)
3754                 return retval;
3755         /*
3756          * To be able to change p->policy safely, the apropriate
3757          * runqueue lock must be held.
3758          */
3759         rq = task_rq_lock(p, &flags);
3760         /* recheck policy now with rq lock held */
3761         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3762                 policy = oldpolicy = -1;
3763                 task_rq_unlock(rq, &flags);
3764                 goto recheck;
3765         }
3766         array = p->array;
3767         if (array)
3768                 deactivate_task(p, rq);
3769         oldprio = p->prio;
3770         __setscheduler(p, policy, param->sched_priority);
3771         if (array) {
3772                 __activate_task(p, rq);
3773                 /*
3774                  * Reschedule if we are currently running on this runqueue and
3775                  * our priority decreased, or if we are not currently running on
3776                  * this runqueue and our priority is higher than the current's
3777                  */
3778                 if (task_running(rq, p)) {
3779                         if (p->prio > oldprio)
3780                                 resched_task(rq->curr);
3781                 } else if (TASK_PREEMPTS_CURR(p, rq))
3782                         resched_task(rq->curr);
3783         }
3784         task_rq_unlock(rq, &flags);
3785         return 0;
3786 }
3787 EXPORT_SYMBOL_GPL(sched_setscheduler);
3788
3789 static int
3790 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3791 {
3792         int retval;
3793         struct sched_param lparam;
3794         struct task_struct *p;
3795
3796         if (!param || pid < 0)
3797                 return -EINVAL;
3798         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3799                 return -EFAULT;
3800         read_lock_irq(&tasklist_lock);
3801         p = find_process_by_pid(pid);
3802         if (!p) {
3803                 read_unlock_irq(&tasklist_lock);
3804                 return -ESRCH;
3805         }
3806         retval = sched_setscheduler(p, policy, &lparam);
3807         read_unlock_irq(&tasklist_lock);
3808         return retval;
3809 }
3810
3811 /**
3812  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3813  * @pid: the pid in question.
3814  * @policy: new policy.
3815  * @param: structure containing the new RT priority.
3816  */
3817 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3818                                        struct sched_param __user *param)
3819 {
3820         return do_sched_setscheduler(pid, policy, param);
3821 }
3822
3823 /**
3824  * sys_sched_setparam - set/change the RT priority of a thread
3825  * @pid: the pid in question.
3826  * @param: structure containing the new RT priority.
3827  */
3828 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3829 {
3830         return do_sched_setscheduler(pid, -1, param);
3831 }
3832
3833 /**
3834  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3835  * @pid: the pid in question.
3836  */
3837 asmlinkage long sys_sched_getscheduler(pid_t pid)
3838 {
3839         int retval = -EINVAL;
3840         task_t *p;
3841
3842         if (pid < 0)
3843                 goto out_nounlock;
3844
3845         retval = -ESRCH;
3846         read_lock(&tasklist_lock);
3847         p = find_process_by_pid(pid);
3848         if (p) {
3849                 retval = security_task_getscheduler(p);
3850                 if (!retval)
3851                         retval = p->policy;
3852         }
3853         read_unlock(&tasklist_lock);
3854
3855 out_nounlock:
3856         return retval;
3857 }
3858
3859 /**
3860  * sys_sched_getscheduler - get the RT priority of a thread
3861  * @pid: the pid in question.
3862  * @param: structure containing the RT priority.
3863  */
3864 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3865 {
3866         struct sched_param lp;
3867         int retval = -EINVAL;
3868         task_t *p;
3869
3870         if (!param || pid < 0)
3871                 goto out_nounlock;
3872
3873         read_lock(&tasklist_lock);
3874         p = find_process_by_pid(pid);
3875         retval = -ESRCH;
3876         if (!p)
3877                 goto out_unlock;
3878
3879         retval = security_task_getscheduler(p);
3880         if (retval)
3881                 goto out_unlock;
3882
3883         lp.sched_priority = p->rt_priority;
3884         read_unlock(&tasklist_lock);
3885
3886         /*
3887          * This one might sleep, we cannot do it with a spinlock held ...
3888          */
3889         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3890
3891 out_nounlock:
3892         return retval;
3893
3894 out_unlock:
3895         read_unlock(&tasklist_lock);
3896         return retval;
3897 }
3898
3899 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3900 {
3901         task_t *p;
3902         int retval;
3903         cpumask_t cpus_allowed;
3904
3905         lock_cpu_hotplug();
3906         read_lock(&tasklist_lock);
3907
3908         p = find_process_by_pid(pid);
3909         if (!p) {
3910                 read_unlock(&tasklist_lock);
3911                 unlock_cpu_hotplug();
3912                 return -ESRCH;
3913         }
3914
3915         /*
3916          * It is not safe to call set_cpus_allowed with the
3917          * tasklist_lock held.  We will bump the task_struct's
3918          * usage count and then drop tasklist_lock.
3919          */
3920         get_task_struct(p);
3921         read_unlock(&tasklist_lock);
3922
3923         retval = -EPERM;
3924         if ((current->euid != p->euid) && (current->euid != p->uid) &&
3925                         !capable(CAP_SYS_NICE))
3926                 goto out_unlock;
3927
3928         cpus_allowed = cpuset_cpus_allowed(p);
3929         cpus_and(new_mask, new_mask, cpus_allowed);
3930         retval = set_cpus_allowed(p, new_mask);
3931
3932 out_unlock:
3933         put_task_struct(p);
3934         unlock_cpu_hotplug();
3935         return retval;
3936 }
3937
3938 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3939                              cpumask_t *new_mask)
3940 {
3941         if (len < sizeof(cpumask_t)) {
3942                 memset(new_mask, 0, sizeof(cpumask_t));
3943         } else if (len > sizeof(cpumask_t)) {
3944                 len = sizeof(cpumask_t);
3945         }
3946         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3947 }
3948
3949 /**
3950  * sys_sched_setaffinity - set the cpu affinity of a process
3951  * @pid: pid of the process
3952  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3953  * @user_mask_ptr: user-space pointer to the new cpu mask
3954  */
3955 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3956                                       unsigned long __user *user_mask_ptr)
3957 {
3958         cpumask_t new_mask;
3959         int retval;
3960
3961         retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3962         if (retval)
3963                 return retval;
3964
3965         return sched_setaffinity(pid, new_mask);
3966 }
3967
3968 /*
3969  * Represents all cpu's present in the system
3970  * In systems capable of hotplug, this map could dynamically grow
3971  * as new cpu's are detected in the system via any platform specific
3972  * method, such as ACPI for e.g.
3973  */
3974
3975 cpumask_t cpu_present_map;
3976 EXPORT_SYMBOL(cpu_present_map);
3977
3978 #ifndef CONFIG_SMP
3979 cpumask_t cpu_online_map = CPU_MASK_ALL;
3980 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3981 #endif
3982
3983 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3984 {
3985         int retval;
3986         task_t *p;
3987
3988         lock_cpu_hotplug();
3989         read_lock(&tasklist_lock);
3990
3991         retval = -ESRCH;
3992         p = find_process_by_pid(pid);
3993         if (!p)
3994                 goto out_unlock;
3995
3996         retval = 0;
3997         cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
3998
3999 out_unlock:
4000         read_unlock(&tasklist_lock);
4001         unlock_cpu_hotplug();
4002         if (retval)
4003                 return retval;
4004
4005         return 0;
4006 }
4007
4008 /**
4009  * sys_sched_getaffinity - get the cpu affinity of a process
4010  * @pid: pid of the process
4011  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4012  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4013  */
4014 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4015                                       unsigned long __user *user_mask_ptr)
4016 {
4017         int ret;
4018         cpumask_t mask;
4019
4020         if (len < sizeof(cpumask_t))
4021                 return -EINVAL;
4022
4023         ret = sched_getaffinity(pid, &mask);
4024         if (ret < 0)
4025                 return ret;
4026
4027         if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4028                 return -EFAULT;
4029
4030         return sizeof(cpumask_t);
4031 }
4032
4033 /**
4034  * sys_sched_yield - yield the current processor to other threads.
4035  *
4036  * this function yields the current CPU by moving the calling thread
4037  * to the expired array. If there are no other threads running on this
4038  * CPU then this function will return.
4039  */
4040 asmlinkage long sys_sched_yield(void)
4041 {
4042         runqueue_t *rq = this_rq_lock();
4043         prio_array_t *array = current->array;
4044         prio_array_t *target = rq->expired;
4045
4046         schedstat_inc(rq, yld_cnt);
4047         /*
4048          * We implement yielding by moving the task into the expired
4049          * queue.
4050          *
4051          * (special rule: RT tasks will just roundrobin in the active
4052          *  array.)
4053          */
4054         if (rt_task(current))
4055                 target = rq->active;
4056
4057         if (array->nr_active == 1) {
4058                 schedstat_inc(rq, yld_act_empty);
4059                 if (!rq->expired->nr_active)
4060                         schedstat_inc(rq, yld_both_empty);
4061         } else if (!rq->expired->nr_active)
4062                 schedstat_inc(rq, yld_exp_empty);
4063
4064         if (array != target) {
4065                 dequeue_task(current, array);
4066                 enqueue_task(current, target);
4067         } else
4068                 /*
4069                  * requeue_task is cheaper so perform that if possible.
4070                  */
4071                 requeue_task(current, array);
4072
4073         /*
4074          * Since we are going to call schedule() anyway, there's
4075          * no need to preempt or enable interrupts:
4076          */
4077         __release(rq->lock);
4078         _raw_spin_unlock(&rq->lock);
4079         preempt_enable_no_resched();
4080
4081         schedule();
4082
4083         return 0;
4084 }
4085
4086 static inline void __cond_resched(void)
4087 {
4088         /*
4089          * The BKS might be reacquired before we have dropped
4090          * PREEMPT_ACTIVE, which could trigger a second
4091          * cond_resched() call.
4092          */
4093         if (unlikely(preempt_count()))
4094                 return;
4095         do {
4096                 add_preempt_count(PREEMPT_ACTIVE);
4097                 schedule();
4098                 sub_preempt_count(PREEMPT_ACTIVE);
4099         } while (need_resched());
4100 }
4101
4102 int __sched cond_resched(void)
4103 {
4104         if (need_resched()) {
4105                 __cond_resched();
4106                 return 1;
4107         }
4108         return 0;
4109 }
4110
4111 EXPORT_SYMBOL(cond_resched);
4112
4113 /*
4114  * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4115  * call schedule, and on return reacquire the lock.
4116  *
4117  * This works OK both with and without CONFIG_PREEMPT.  We do strange low-level
4118  * operations here to prevent schedule() from being called twice (once via
4119  * spin_unlock(), once by hand).
4120  */
4121 int cond_resched_lock(spinlock_t *lock)
4122 {
4123         int ret = 0;
4124
4125         if (need_lockbreak(lock)) {
4126                 spin_unlock(lock);
4127                 cpu_relax();
4128                 ret = 1;
4129                 spin_lock(lock);
4130         }
4131         if (need_resched()) {
4132                 _raw_spin_unlock(lock);
4133                 preempt_enable_no_resched();
4134                 __cond_resched();
4135                 ret = 1;
4136                 spin_lock(lock);
4137         }
4138         return ret;
4139 }
4140
4141 EXPORT_SYMBOL(cond_resched_lock);
4142
4143 int __sched cond_resched_softirq(void)
4144 {
4145         BUG_ON(!in_softirq());
4146
4147         if (need_resched()) {
4148                 __local_bh_enable();
4149                 __cond_resched();
4150                 local_bh_disable();
4151                 return 1;
4152         }
4153         return 0;
4154 }
4155
4156 EXPORT_SYMBOL(cond_resched_softirq);
4157
4158
4159 /**
4160  * yield - yield the current processor to other threads.
4161  *
4162  * this is a shortcut for kernel-space yielding - it marks the
4163  * thread runnable and calls sys_sched_yield().
4164  */
4165 void __sched yield(void)
4166 {
4167         set_current_state(TASK_RUNNING);
4168         sys_sched_yield();
4169 }
4170
4171 EXPORT_SYMBOL(yield);
4172
4173 /*
4174  * This task is about to go to sleep on IO.  Increment rq->nr_iowait so
4175  * that process accounting knows that this is a task in IO wait state.
4176  *
4177  * But don't do that if it is a deliberate, throttling IO wait (this task
4178  * has set its backing_dev_info: the queue against which it should throttle)
4179  */
4180 void __sched io_schedule(void)
4181 {
4182         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4183
4184         atomic_inc(&rq->nr_iowait);
4185         schedule();
4186         atomic_dec(&rq->nr_iowait);
4187 }
4188
4189 EXPORT_SYMBOL(io_schedule);
4190
4191 long __sched io_schedule_timeout(long timeout)
4192 {
4193         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4194         long ret;
4195
4196         atomic_inc(&rq->nr_iowait);
4197         ret = schedule_timeout(timeout);
4198         atomic_dec(&rq->nr_iowait);
4199         return ret;
4200 }
4201
4202 /**
4203  * sys_sched_get_priority_max - return maximum RT priority.
4204  * @policy: scheduling class.
4205  *
4206  * this syscall returns the maximum rt_priority that can be used
4207  * by a given scheduling class.
4208  */
4209 asmlinkage long sys_sched_get_priority_max(int policy)
4210 {
4211         int ret = -EINVAL;
4212
4213         switch (policy) {
4214         case SCHED_FIFO:
4215         case SCHED_RR:
4216                 ret = MAX_USER_RT_PRIO-1;
4217                 break;
4218         case SCHED_NORMAL:
4219                 ret = 0;
4220                 break;
4221         }
4222         return ret;
4223 }
4224
4225 /**
4226  * sys_sched_get_priority_min - return minimum RT priority.
4227  * @policy: scheduling class.
4228  *
4229  * this syscall returns the minimum rt_priority that can be used
4230  * by a given scheduling class.
4231  */
4232 asmlinkage long sys_sched_get_priority_min(int policy)
4233 {
4234         int ret = -EINVAL;
4235
4236         switch (policy) {
4237         case SCHED_FIFO:
4238         case SCHED_RR:
4239                 ret = 1;
4240                 break;
4241         case SCHED_NORMAL:
4242                 ret = 0;
4243         }
4244         return ret;
4245 }
4246
4247 /**
4248  * sys_sched_rr_get_interval - return the default timeslice of a process.
4249  * @pid: pid of the process.
4250  * @interval: userspace pointer to the timeslice value.
4251  *
4252  * this syscall writes the default timeslice value of a given process
4253  * into the user-space timespec buffer. A value of '0' means infinity.
4254  */
4255 asmlinkage
4256 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4257 {
4258         int retval = -EINVAL;
4259         struct timespec t;
4260         task_t *p;
4261
4262         if (pid < 0)
4263                 goto out_nounlock;
4264
4265         retval = -ESRCH;
4266         read_lock(&tasklist_lock);
4267         p = find_process_by_pid(pid);
4268         if (!p)
4269                 goto out_unlock;
4270
4271         retval = security_task_getscheduler(p);
4272         if (retval)
4273                 goto out_unlock;
4274
4275         jiffies_to_timespec(p->policy & SCHED_FIFO ?
4276                                 0 : task_timeslice(p), &t);
4277         read_unlock(&tasklist_lock);
4278         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4279 out_nounlock:
4280         return retval;
4281 out_unlock:
4282         read_unlock(&tasklist_lock);
4283         return retval;
4284 }
4285
4286 static inline struct task_struct *eldest_child(struct task_struct *p)
4287 {
4288         if (list_empty(&p->children)) return NULL;
4289         return list_entry(p->children.next,struct task_struct,sibling);
4290 }
4291
4292 static inline struct task_struct *older_sibling(struct task_struct *p)
4293 {
4294         if (p->sibling.prev==&p->parent->children) return NULL;
4295         return list_entry(p->sibling.prev,struct task_struct,sibling);
4296 }
4297
4298 static inline struct task_struct *younger_sibling(struct task_struct *p)
4299 {
4300         if (p->sibling.next==&p->parent->children) return NULL;
4301         return list_entry(p->sibling.next,struct task_struct,sibling);
4302 }
4303
4304 static void show_task(task_t *p)
4305 {
4306         task_t *relative;
4307         unsigned state;
4308         unsigned long free = 0;
4309         static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4310
4311         printk("%-13.13s ", p->comm);
4312         state = p->state ? __ffs(p->state) + 1 : 0;
4313         if (state < ARRAY_SIZE(stat_nam))
4314                 printk(stat_nam[state]);
4315         else
4316                 printk("?");
4317 #if (BITS_PER_LONG == 32)
4318         if (state == TASK_RUNNING)
4319                 printk(" running ");
4320         else
4321                 printk(" %08lX ", thread_saved_pc(p));
4322 #else
4323         if (state == TASK_RUNNING)
4324                 printk("  running task   ");
4325         else
4326                 printk(" %016lx ", thread_saved_pc(p));
4327 #endif
4328 #ifdef CONFIG_DEBUG_STACK_USAGE
4329         {
4330                 unsigned long *n = end_of_stack(p);
4331                 while (!*n)
4332                         n++;
4333                 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4334         }
4335 #endif
4336         printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4337         if ((relative = eldest_child(p)))
4338                 printk("%5d ", relative->pid);
4339         else
4340                 printk("      ");
4341         if ((relative = younger_sibling(p)))
4342                 printk("%7d", relative->pid);
4343         else
4344                 printk("       ");
4345         if ((relative = older_sibling(p)))
4346                 printk(" %5d", relative->pid);
4347         else
4348                 printk("      ");
4349         if (!p->mm)
4350                 printk(" (L-TLB)\n");
4351         else
4352                 printk(" (NOTLB)\n");
4353
4354         if (state != TASK_RUNNING)
4355                 show_stack(p, NULL);
4356 }
4357
4358 void show_state(void)
4359 {
4360         task_t *g, *p;
4361
4362 #if (BITS_PER_LONG == 32)
4363         printk("\n"
4364                "                                               sibling\n");
4365         printk("  task             PC      pid father child younger older\n");
4366 #else
4367         printk("\n"
4368                "                                                       sibling\n");
4369         printk("  task                 PC          pid father child younger older\n");
4370 #endif
4371         read_lock(&tasklist_lock);
4372         do_each_thread(g, p) {
4373                 /*
4374                  * reset the NMI-timeout, listing all files on a slow
4375                  * console might take alot of time:
4376                  */
4377                 touch_nmi_watchdog();
4378                 show_task(p);
4379         } while_each_thread(g, p);
4380
4381         read_unlock(&tasklist_lock);
4382 }
4383
4384 /**
4385  * init_idle - set up an idle thread for a given CPU
4386  * @idle: task in question
4387  * @cpu: cpu the idle task belongs to
4388  *
4389  * NOTE: this function does not set the idle thread's NEED_RESCHED
4390  * flag, to make booting more robust.
4391  */
4392 void __devinit init_idle(task_t *idle, int cpu)
4393 {
4394         runqueue_t *rq = cpu_rq(cpu);
4395         unsigned long flags;
4396
4397         idle->sleep_avg = 0;
4398         idle->array = NULL;
4399         idle->prio = MAX_PRIO;
4400         idle->state = TASK_RUNNING;
4401         idle->cpus_allowed = cpumask_of_cpu(cpu);
4402         set_task_cpu(idle, cpu);
4403
4404         spin_lock_irqsave(&rq->lock, flags);
4405         rq->curr = rq->idle = idle;
4406 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4407         idle->oncpu = 1;
4408 #endif
4409         spin_unlock_irqrestore(&rq->lock, flags);
4410
4411         /* Set the preempt count _outside_ the spinlocks! */
4412 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4413         task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4414 #else
4415         task_thread_info(idle)->preempt_count = 0;
4416 #endif
4417 }
4418
4419 /*
4420  * In a system that switches off the HZ timer nohz_cpu_mask
4421  * indicates which cpus entered this state. This is used
4422  * in the rcu update to wait only for active cpus. For system
4423  * which do not switch off the HZ timer nohz_cpu_mask should
4424  * always be CPU_MASK_NONE.
4425  */
4426 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4427
4428 #ifdef CONFIG_SMP
4429 /*
4430  * This is how migration works:
4431  *
4432  * 1) we queue a migration_req_t structure in the source CPU's
4433  *    runqueue and wake up that CPU's migration thread.
4434  * 2) we down() the locked semaphore => thread blocks.
4435  * 3) migration thread wakes up (implicitly it forces the migrated
4436  *    thread off the CPU)
4437  * 4) it gets the migration request and checks whether the migrated
4438  *    task is still in the wrong runqueue.
4439  * 5) if it's in the wrong runqueue then the migration thread removes
4440  *    it and puts it into the right queue.
4441  * 6) migration thread up()s the semaphore.
4442  * 7) we wake up and the migration is done.
4443  */
4444
4445 /*
4446  * Change a given task's CPU affinity. Migrate the thread to a
4447  * proper CPU and schedule it away if the CPU it's executing on
4448  * is removed from the allowed bitmask.
4449  *
4450  * NOTE: the caller must have a valid reference to the task, the
4451  * task must not exit() & deallocate itself prematurely.  The
4452  * call is not atomic; no spinlocks may be held.
4453  */
4454 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4455 {
4456         unsigned long flags;
4457         int ret = 0;
4458         migration_req_t req;
4459         runqueue_t *rq;
4460
4461         rq = task_rq_lock(p, &flags);
4462         if (!cpus_intersects(new_mask, cpu_online_map)) {
4463                 ret = -EINVAL;
4464                 goto out;
4465         }
4466
4467         p->cpus_allowed = new_mask;
4468         /* Can the task run on the task's current CPU? If so, we're done */
4469         if (cpu_isset(task_cpu(p), new_mask))
4470                 goto out;
4471
4472         if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4473                 /* Need help from migration thread: drop lock and wait. */
4474                 task_rq_unlock(rq, &flags);
4475                 wake_up_process(rq->migration_thread);
4476                 wait_for_completion(&req.done);
4477                 tlb_migrate_finish(p->mm);
4478                 return 0;
4479         }
4480 out:
4481         task_rq_unlock(rq, &flags);
4482         return ret;
4483 }
4484
4485 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4486
4487 /*
4488  * Move (not current) task off this cpu, onto dest cpu.  We're doing
4489  * this because either it can't run here any more (set_cpus_allowed()
4490  * away from this CPU, or CPU going down), or because we're
4491  * attempting to rebalance this task on exec (sched_exec).
4492  *
4493  * So we race with normal scheduler movements, but that's OK, as long
4494  * as the task is no longer on this CPU.
4495  */
4496 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4497 {
4498         runqueue_t *rq_dest, *rq_src;
4499
4500         if (unlikely(cpu_is_offline(dest_cpu)))
4501                 return;
4502
4503         rq_src = cpu_rq(src_cpu);
4504         rq_dest = cpu_rq(dest_cpu);
4505
4506         double_rq_lock(rq_src, rq_dest);
4507         /* Already moved. */
4508         if (task_cpu(p) != src_cpu)
4509                 goto out;
4510         /* Affinity changed (again). */
4511         if (!cpu_isset(dest_cpu, p->cpus_allowed))
4512                 goto out;
4513
4514         set_task_cpu(p, dest_cpu);
4515         if (p->array) {
4516                 /*
4517                  * Sync timestamp with rq_dest's before activating.
4518                  * The same thing could be achieved by doing this step
4519                  * afterwards, and pretending it was a local activate.
4520                  * This way is cleaner and logically correct.
4521                  */
4522                 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4523                                 + rq_dest->timestamp_last_tick;
4524                 deactivate_task(p, rq_src);
4525                 activate_task(p, rq_dest, 0);
4526                 if (TASK_PREEMPTS_CURR(p, rq_dest))
4527                         resched_task(rq_dest->curr);
4528         }
4529
4530 out:
4531         double_rq_unlock(rq_src, rq_dest);
4532 }
4533
4534 /*
4535  * migration_thread - this is a highprio system thread that performs
4536  * thread migration by bumping thread off CPU then 'pushing' onto
4537  * another runqueue.
4538  */
4539 static int migration_thread(void *data)
4540 {
4541         runqueue_t *rq;
4542         int cpu = (long)data;
4543
4544         rq = cpu_rq(cpu);
4545         BUG_ON(rq->migration_thread != current);
4546
4547         set_current_state(TASK_INTERRUPTIBLE);
4548         while (!kthread_should_stop()) {
4549                 struct list_head *head;
4550                 migration_req_t *req;
4551
4552                 try_to_freeze();
4553
4554                 spin_lock_irq(&rq->lock);
4555
4556                 if (cpu_is_offline(cpu)) {
4557                         spin_unlock_irq(&rq->lock);
4558                         goto wait_to_die;
4559                 }
4560
4561                 if (rq->active_balance) {
4562                         active_load_balance(rq, cpu);
4563                         rq->active_balance = 0;
4564                 }
4565
4566                 head = &rq->migration_queue;
4567
4568                 if (list_empty(head)) {
4569                         spin_unlock_irq(&rq->lock);
4570                         schedule();
4571                         set_current_state(TASK_INTERRUPTIBLE);
4572                         continue;
4573                 }
4574                 req = list_entry(head->next, migration_req_t, list);
4575                 list_del_init(head->next);
4576
4577                 spin_unlock(&rq->lock);
4578                 __migrate_task(req->task, cpu, req->dest_cpu);
4579                 local_irq_enable();
4580
4581                 complete(&req->done);
4582         }
4583         __set_current_state(TASK_RUNNING);
4584         return 0;
4585
4586 wait_to_die:
4587         /* Wait for kthread_stop */
4588         set_current_state(TASK_INTERRUPTIBLE);
4589         while (!kthread_should_stop()) {
4590                 schedule();
4591                 set_current_state(TASK_INTERRUPTIBLE);
4592         }
4593         __set_current_state(TASK_RUNNING);
4594         return 0;
4595 }
4596
4597 #ifdef CONFIG_HOTPLUG_CPU
4598 /* Figure out where task on dead CPU should go, use force if neccessary. */
4599 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4600 {
4601         int dest_cpu;
4602         cpumask_t mask;
4603
4604         /* On same node? */
4605         mask = node_to_cpumask(cpu_to_node(dead_cpu));
4606         cpus_and(mask, mask, tsk->cpus_allowed);
4607         dest_cpu = any_online_cpu(mask);
4608
4609         /* On any allowed CPU? */
4610         if (dest_cpu == NR_CPUS)
4611                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4612
4613         /* No more Mr. Nice Guy. */
4614         if (dest_cpu == NR_CPUS) {
4615                 cpus_setall(tsk->cpus_allowed);
4616                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4617
4618                 /*
4619                  * Don't tell them about moving exiting tasks or
4620                  * kernel threads (both mm NULL), since they never
4621                  * leave kernel.
4622                  */
4623                 if (tsk->mm && printk_ratelimit())
4624                         printk(KERN_INFO "process %d (%s) no "
4625                                "longer affine to cpu%d\n",
4626                                tsk->pid, tsk->comm, dead_cpu);
4627         }
4628         __migrate_task(tsk, dead_cpu, dest_cpu);
4629 }
4630
4631 /*
4632  * While a dead CPU has no uninterruptible tasks queued at this point,
4633  * it might still have a nonzero ->nr_uninterruptible counter, because
4634  * for performance reasons the counter is not stricly tracking tasks to
4635  * their home CPUs. So we just add the counter to another CPU's counter,
4636  * to keep the global sum constant after CPU-down:
4637  */
4638 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4639 {
4640         runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4641         unsigned long flags;
4642
4643         local_irq_save(flags);
4644         double_rq_lock(rq_src, rq_dest);
4645         rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4646         rq_src->nr_uninterruptible = 0;
4647         double_rq_unlock(rq_src, rq_dest);
4648         local_irq_restore(flags);
4649 }
4650
4651 /* Run through task list and migrate tasks from the dead cpu. */
4652 static void migrate_live_tasks(int src_cpu)
4653 {
4654         struct task_struct *tsk, *t;
4655
4656         write_lock_irq(&tasklist_lock);
4657
4658         do_each_thread(t, tsk) {
4659                 if (tsk == current)
4660                         continue;
4661
4662                 if (task_cpu(tsk) == src_cpu)
4663                         move_task_off_dead_cpu(src_cpu, tsk);
4664         } while_each_thread(t, tsk);
4665
4666         write_unlock_irq(&tasklist_lock);
4667 }
4668
4669 /* Schedules idle task to be the next runnable task on current CPU.
4670  * It does so by boosting its priority to highest possible and adding it to
4671  * the _front_ of runqueue. Used by CPU offline code.
4672  */
4673 void sched_idle_next(void)
4674 {
4675         int cpu = smp_processor_id();
4676         runqueue_t *rq = this_rq();
4677         struct task_struct *p = rq->idle;
4678         unsigned long flags;
4679
4680         /* cpu has to be offline */
4681         BUG_ON(cpu_online(cpu));
4682
4683         /* Strictly not necessary since rest of the CPUs are stopped by now
4684          * and interrupts disabled on current cpu.
4685          */
4686         spin_lock_irqsave(&rq->lock, flags);
4687
4688         __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4689         /* Add idle task to _front_ of it's priority queue */
4690         __activate_idle_task(p, rq);
4691
4692         spin_unlock_irqrestore(&rq->lock, flags);
4693 }
4694
4695 /* Ensures that the idle task is using init_mm right before its cpu goes
4696  * offline.
4697  */
4698 void idle_task_exit(void)
4699 {
4700         struct mm_struct *mm = current->active_mm;
4701
4702         BUG_ON(cpu_online(smp_processor_id()));
4703
4704         if (mm != &init_mm)
4705                 switch_mm(mm, &init_mm, current);
4706         mmdrop(mm);
4707 }
4708
4709 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4710 {
4711         struct runqueue *rq = cpu_rq(dead_cpu);
4712
4713         /* Must be exiting, otherwise would be on tasklist. */
4714         BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4715
4716         /* Cannot have done final schedule yet: would have vanished. */
4717         BUG_ON(tsk->flags & PF_DEAD);
4718
4719         get_task_struct(tsk);
4720
4721         /*
4722          * Drop lock around migration; if someone else moves it,
4723          * that's OK.  No task can be added to this CPU, so iteration is
4724          * fine.
4725          */
4726         spin_unlock_irq(&rq->lock);
4727         move_task_off_dead_cpu(dead_cpu, tsk);
4728         spin_lock_irq(&rq->lock);
4729
4730         put_task_struct(tsk);
4731 }
4732
4733 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4734 static void migrate_dead_tasks(unsigned int dead_cpu)
4735 {
4736         unsigned arr, i;
4737         struct runqueue *rq = cpu_rq(dead_cpu);
4738
4739         for (arr = 0; arr < 2; arr++) {
4740                 for (i = 0; i < MAX_PRIO; i++) {
4741                         struct list_head *list = &rq->arrays[arr].queue[i];
4742                         while (!list_empty(list))
4743                                 migrate_dead(dead_cpu,
4744                                              list_entry(list->next, task_t,
4745                                                         run_list));
4746                 }
4747         }
4748 }
4749 #endif /* CONFIG_HOTPLUG_CPU */
4750
4751 /*
4752  * migration_call - callback that gets triggered when a CPU is added.
4753  * Here we can start up the necessary migration thread for the new CPU.
4754  */
4755 static int migration_call(struct notifier_block *nfb, unsigned long action,
4756                           void *hcpu)
4757 {
4758         int cpu = (long)hcpu;
4759         struct task_struct *p;
4760         struct runqueue *rq;
4761         unsigned long flags;
4762
4763         switch (action) {
4764         case CPU_UP_PREPARE:
4765                 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4766                 if (IS_ERR(p))
4767                         return NOTIFY_BAD;
4768                 p->flags |= PF_NOFREEZE;
4769                 kthread_bind(p, cpu);
4770                 /* Must be high prio: stop_machine expects to yield to it. */
4771                 rq = task_rq_lock(p, &flags);
4772                 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4773                 task_rq_unlock(rq, &flags);
4774                 cpu_rq(cpu)->migration_thread = p;
4775                 break;
4776         case CPU_ONLINE:
4777                 /* Strictly unneccessary, as first user will wake it. */
4778                 wake_up_process(cpu_rq(cpu)->migration_thread);
4779                 break;
4780 #ifdef CONFIG_HOTPLUG_CPU
4781         case CPU_UP_CANCELED:
4782                 /* Unbind it from offline cpu so it can run.  Fall thru. */
4783                 kthread_bind(cpu_rq(cpu)->migration_thread,
4784                              any_online_cpu(cpu_online_map));
4785                 kthread_stop(cpu_rq(cpu)->migration_thread);
4786                 cpu_rq(cpu)->migration_thread = NULL;
4787                 break;
4788         case CPU_DEAD:
4789                 migrate_live_tasks(cpu);
4790                 rq = cpu_rq(cpu);
4791                 kthread_stop(rq->migration_thread);
4792                 rq->migration_thread = NULL;
4793                 /* Idle task back to normal (off runqueue, low prio) */
4794                 rq = task_rq_lock(rq->idle, &flags);
4795                 deactivate_task(rq->idle, rq);
4796                 rq->idle->static_prio = MAX_PRIO;
4797                 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4798                 migrate_dead_tasks(cpu);
4799                 task_rq_unlock(rq, &flags);
4800                 migrate_nr_uninterruptible(rq);
4801                 BUG_ON(rq->nr_running != 0);
4802
4803                 /* No need to migrate the tasks: it was best-effort if
4804                  * they didn't do lock_cpu_hotplug().  Just wake up
4805                  * the requestors. */
4806                 spin_lock_irq(&rq->lock);
4807                 while (!list_empty(&rq->migration_queue)) {
4808                         migration_req_t *req;
4809                         req = list_entry(rq->migration_queue.next,
4810                                          migration_req_t, list);
4811                         list_del_init(&req->list);
4812                         complete(&req->done);
4813                 }
4814                 spin_unlock_irq(&rq->lock);
4815                 break;
4816 #endif
4817         }
4818         return NOTIFY_OK;
4819 }
4820
4821 /* Register at highest priority so that task migration (migrate_all_tasks)
4822  * happens before everything else.
4823  */
4824 static struct notifier_block __devinitdata migration_notifier = {
4825         .notifier_call = migration_call,
4826         .priority = 10
4827 };
4828
4829 int __init migration_init(void)
4830 {
4831         void *cpu = (void *)(long)smp_processor_id();
4832         /* Start one for boot CPU. */
4833         migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4834         migration_call(&migration_notifier, CPU_ONLINE, cpu);
4835         register_cpu_notifier(&migration_notifier);
4836         return 0;
4837 }
4838 #endif
4839
4840 #ifdef CONFIG_SMP
4841 #undef SCHED_DOMAIN_DEBUG
4842 #ifdef SCHED_DOMAIN_DEBUG
4843 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4844 {
4845         int level = 0;
4846
4847         if (!sd) {
4848                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4849                 return;
4850         }
4851
4852         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4853
4854         do {
4855                 int i;
4856                 char str[NR_CPUS];
4857                 struct sched_group *group = sd->groups;
4858                 cpumask_t groupmask;
4859
4860                 cpumask_scnprintf(str, NR_CPUS, sd->span);
4861                 cpus_clear(groupmask);
4862
4863                 printk(KERN_DEBUG);
4864                 for (i = 0; i < level + 1; i++)
4865                         printk(" ");
4866                 printk("domain %d: ", level);
4867
4868                 if (!(sd->flags & SD_LOAD_BALANCE)) {
4869                         printk("does not load-balance\n");
4870                         if (sd->parent)
4871                                 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4872                         break;
4873                 }
4874
4875                 printk("span %s\n", str);
4876
4877                 if (!cpu_isset(cpu, sd->span))
4878                         printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4879                 if (!cpu_isset(cpu, group->cpumask))
4880                         printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4881
4882                 printk(KERN_DEBUG);
4883                 for (i = 0; i < level + 2; i++)
4884                         printk(" ");
4885                 printk("groups:");
4886                 do {
4887                         if (!group) {
4888                                 printk("\n");
4889                                 printk(KERN_ERR "ERROR: group is NULL\n");
4890                                 break;
4891                         }
4892
4893                         if (!group->cpu_power) {
4894                                 printk("\n");
4895                                 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4896                         }
4897
4898                         if (!cpus_weight(group->cpumask)) {
4899                                 printk("\n");
4900                                 printk(KERN_ERR "ERROR: empty group\n");
4901                         }
4902
4903                         if (cpus_intersects(groupmask, group->cpumask)) {
4904                                 printk("\n");
4905                                 printk(KERN_ERR "ERROR: repeated CPUs\n");
4906                         }
4907
4908                         cpus_or(groupmask, groupmask, group->cpumask);
4909
4910                         cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4911                         printk(" %s", str);
4912
4913                         group = group->next;
4914                 } while (group != sd->groups);
4915                 printk("\n");
4916
4917                 if (!cpus_equal(sd->span, groupmask))
4918                         printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4919
4920                 level++;
4921                 sd = sd->parent;
4922
4923                 if (sd) {
4924                         if (!cpus_subset(groupmask, sd->span))
4925                                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4926                 }
4927
4928         } while (sd);
4929 }
4930 #else
4931 #define sched_domain_debug(sd, cpu) {}
4932 #endif
4933
4934 static int sd_degenerate(struct sched_domain *sd)
4935 {
4936         if (cpus_weight(sd->span) == 1)
4937                 return 1;
4938
4939         /* Following flags need at least 2 groups */
4940         if (sd->flags & (SD_LOAD_BALANCE |
4941                          SD_BALANCE_NEWIDLE |
4942                          SD_BALANCE_FORK |
4943                          SD_BALANCE_EXEC)) {
4944                 if (sd->groups != sd->groups->next)
4945                         return 0;
4946         }
4947
4948         /* Following flags don't use groups */
4949         if (sd->flags & (SD_WAKE_IDLE |
4950                          SD_WAKE_AFFINE |
4951                          SD_WAKE_BALANCE))
4952                 return 0;
4953
4954         return 1;
4955 }
4956
4957 static int sd_parent_degenerate(struct sched_domain *sd,
4958                                                 struct sched_domain *parent)
4959 {
4960         unsigned long cflags = sd->flags, pflags = parent->flags;
4961
4962         if (sd_degenerate(parent))
4963                 return 1;
4964
4965         if (!cpus_equal(sd->span, parent->span))
4966                 return 0;
4967
4968         /* Does parent contain flags not in child? */
4969         /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4970         if (cflags & SD_WAKE_AFFINE)
4971                 pflags &= ~SD_WAKE_BALANCE;
4972         /* Flags needing groups don't count if only 1 group in parent */
4973         if (parent->groups == parent->groups->next) {
4974                 pflags &= ~(SD_LOAD_BALANCE |
4975                                 SD_BALANCE_NEWIDLE |
4976                                 SD_BALANCE_FORK |
4977                                 SD_BALANCE_EXEC);
4978         }
4979         if (~cflags & pflags)
4980                 return 0;
4981
4982         return 1;
4983 }
4984
4985 /*
4986  * Attach the domain 'sd' to 'cpu' as its base domain.  Callers must
4987  * hold the hotplug lock.
4988  */
4989 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4990 {
4991         runqueue_t *rq = cpu_rq(cpu);
4992         struct sched_domain *tmp;
4993
4994         /* Remove the sched domains which do not contribute to scheduling. */
4995         for (tmp = sd; tmp; tmp = tmp->parent) {
4996                 struct sched_domain *parent = tmp->parent;
4997                 if (!parent)
4998                         break;
4999                 if (sd_parent_degenerate(tmp, parent))
5000                         tmp->parent = parent->parent;
5001         }
5002
5003         if (sd && sd_degenerate(sd))
5004                 sd = sd->parent;
5005
5006         sched_domain_debug(sd, cpu);
5007
5008         rcu_assign_pointer(rq->sd, sd);
5009 }
5010
5011 /* cpus with isolated domains */
5012 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
5013
5014 /* Setup the mask of cpus configured for isolated domains */
5015 static int __init isolated_cpu_setup(char *str)
5016 {
5017         int ints[NR_CPUS], i;
5018
5019         str = get_options(str, ARRAY_SIZE(ints), ints);
5020         cpus_clear(cpu_isolated_map);
5021         for (i = 1; i <= ints[0]; i++)
5022                 if (ints[i] < NR_CPUS)
5023                         cpu_set(ints[i], cpu_isolated_map);
5024         return 1;
5025 }
5026
5027 __setup ("isolcpus=", isolated_cpu_setup);
5028
5029 /*
5030  * init_sched_build_groups takes an array of groups, the cpumask we wish
5031  * to span, and a pointer to a function which identifies what group a CPU
5032  * belongs to. The return value of group_fn must be a valid index into the
5033  * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5034  * keep track of groups covered with a cpumask_t).
5035  *
5036  * init_sched_build_groups will build a circular linked list of the groups
5037  * covered by the given span, and will set each group's ->cpumask correctly,
5038  * and ->cpu_power to 0.
5039  */
5040 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5041                                     int (*group_fn)(int cpu))
5042 {
5043         struct sched_group *first = NULL, *last = NULL;
5044         cpumask_t covered = CPU_MASK_NONE;
5045         int i;
5046
5047         for_each_cpu_mask(i, span) {
5048                 int group = group_fn(i);
5049                 struct sched_group *sg = &groups[group];
5050                 int j;
5051
5052                 if (cpu_isset(i, covered))
5053                         continue;
5054
5055                 sg->cpumask = CPU_MASK_NONE;
5056                 sg->cpu_power = 0;
5057
5058                 for_each_cpu_mask(j, span) {
5059                         if (group_fn(j) != group)
5060                                 continue;
5061
5062                         cpu_set(j, covered);
5063                         cpu_set(j, sg->cpumask);
5064                 }
5065                 if (!first)
5066                         first = sg;
5067                 if (last)
5068                         last->next = sg;
5069                 last = sg;
5070         }
5071         last->next = first;
5072 }
5073
5074 #define SD_NODES_PER_DOMAIN 16
5075
5076 #ifdef CONFIG_NUMA
5077 /**
5078  * find_next_best_node - find the next node to include in a sched_domain
5079  * @node: node whose sched_domain we're building
5080  * @used_nodes: nodes already in the sched_domain
5081  *
5082  * Find the next node to include in a given scheduling domain.  Simply
5083  * finds the closest node not already in the @used_nodes map.
5084  *
5085  * Should use nodemask_t.
5086  */
5087 static int find_next_best_node(int node, unsigned long *used_nodes)
5088 {
5089         int i, n, val, min_val, best_node = 0;
5090
5091         min_val = INT_MAX;
5092
5093         for (i = 0; i < MAX_NUMNODES; i++) {
5094                 /* Start at @node */
5095                 n = (node + i) % MAX_NUMNODES;
5096
5097                 if (!nr_cpus_node(n))
5098                         continue;
5099
5100                 /* Skip already used nodes */
5101                 if (test_bit(n, used_nodes))
5102                         continue;
5103
5104                 /* Simple min distance search */
5105                 val = node_distance(node, n);
5106
5107                 if (val < min_val) {
5108                         min_val = val;
5109                         best_node = n;
5110                 }
5111         }
5112
5113         set_bit(best_node, used_nodes);
5114         return best_node;
5115 }
5116
5117 /**
5118  * sched_domain_node_span - get a cpumask for a node's sched_domain
5119  * @node: node whose cpumask we're constructing
5120  * @size: number of nodes to include in this span
5121  *
5122  * Given a node, construct a good cpumask for its sched_domain to span.  It
5123  * should be one that prevents unnecessary balancing, but also spreads tasks
5124  * out optimally.
5125  */
5126 static cpumask_t sched_domain_node_span(int node)
5127 {
5128         int i;
5129         cpumask_t span, nodemask;
5130         DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5131
5132         cpus_clear(span);
5133         bitmap_zero(used_nodes, MAX_NUMNODES);
5134
5135         nodemask = node_to_cpumask(node);
5136         cpus_or(span, span, nodemask);
5137         set_bit(node, used_nodes);
5138
5139         for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5140                 int next_node = find_next_best_node(node, used_nodes);
5141                 nodemask = node_to_cpumask(next_node);
5142                 cpus_or(span, span, nodemask);
5143         }
5144
5145         return span;
5146 }
5147 #endif
5148
5149 /*
5150  * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5151  * can switch it on easily if needed.
5152  */
5153 #ifdef CONFIG_SCHED_SMT
5154 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5155 static struct sched_group sched_group_cpus[NR_CPUS];
5156 static int cpu_to_cpu_group(int cpu)
5157 {
5158         return cpu;
5159 }
5160 #endif
5161
5162 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5163 static struct sched_group sched_group_phys[NR_CPUS];
5164 static int cpu_to_phys_group(int cpu)
5165 {
5166 #ifdef CONFIG_SCHED_SMT
5167         return first_cpu(cpu_sibling_map[cpu]);
5168 #else
5169         return cpu;
5170 #endif
5171 }
5172
5173 #ifdef CONFIG_NUMA
5174 /*
5175  * The init_sched_build_groups can't handle what we want to do with node
5176  * groups, so roll our own. Now each node has its own list of groups which
5177  * gets dynamically allocated.
5178  */
5179 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5180 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5181
5182 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5183 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5184
5185 static int cpu_to_allnodes_group(int cpu)
5186 {
5187         return cpu_to_node(cpu);
5188 }
5189 #endif
5190
5191 /*
5192  * Build sched domains for a given set of cpus and attach the sched domains
5193  * to the individual cpus
5194  */
5195 void build_sched_domains(const cpumask_t *cpu_map)
5196 {
5197         int i;
5198 #ifdef CONFIG_NUMA
5199         struct sched_group **sched_group_nodes = NULL;
5200         struct sched_group *sched_group_allnodes = NULL;
5201
5202         /*
5203          * Allocate the per-node list of sched groups
5204          */
5205         sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5206                                            GFP_ATOMIC);
5207         if (!sched_group_nodes) {
5208                 printk(KERN_WARNING "Can not alloc sched group node list\n");
5209                 return;
5210         }
5211         sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5212 #endif
5213
5214         /*
5215          * Set up domains for cpus specified by the cpu_map.
5216          */
5217         for_each_cpu_mask(i, *cpu_map) {
5218                 int group;
5219                 struct sched_domain *sd = NULL, *p;
5220                 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5221
5222                 cpus_and(nodemask, nodemask, *cpu_map);
5223
5224 #ifdef CONFIG_NUMA
5225                 if (cpus_weight(*cpu_map)
5226                                 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5227                         if (!sched_group_allnodes) {
5228                                 sched_group_allnodes
5229                                         = kmalloc(sizeof(struct sched_group)
5230                                                         * MAX_NUMNODES,
5231                                                   GFP_KERNEL);
5232                                 if (!sched_group_allnodes) {
5233                                         printk(KERN_WARNING
5234                                         "Can not alloc allnodes sched group\n");
5235                                         break;
5236                                 }
5237                                 sched_group_allnodes_bycpu[i]
5238                                                 = sched_group_allnodes;
5239                         }
5240                         sd = &per_cpu(allnodes_domains, i);
5241                         *sd = SD_ALLNODES_INIT;
5242                         sd->span = *cpu_map;
5243                         group = cpu_to_allnodes_group(i);
5244                         sd->groups = &sched_group_allnodes[group];
5245                         p = sd;
5246                 } else
5247                         p = NULL;
5248
5249                 sd = &per_cpu(node_domains, i);
5250                 *sd = SD_NODE_INIT;
5251                 sd->span = sched_domain_node_span(cpu_to_node(i));
5252                 sd->parent = p;
5253                 cpus_and(sd->span, sd->span, *cpu_map);
5254 #endif
5255
5256                 p = sd;
5257                 sd = &per_cpu(phys_domains, i);
5258                 group = cpu_to_phys_group(i);
5259                 *sd = SD_CPU_INIT;
5260                 sd->span = nodemask;
5261                 sd->parent = p;
5262                 sd->groups = &sched_group_phys[group];
5263
5264 #ifdef CONFIG_SCHED_SMT
5265                 p = sd;
5266                 sd = &per_cpu(cpu_domains, i);
5267                 group = cpu_to_cpu_group(i);
5268                 *sd = SD_SIBLING_INIT;
5269                 sd->span = cpu_sibling_map[i];
5270                 cpus_and(sd->span, sd->span, *cpu_map);
5271                 sd->parent = p;
5272                 sd->groups = &sched_group_cpus[group];
5273 #endif
5274         }
5275
5276 #ifdef CONFIG_SCHED_SMT
5277         /* Set up CPU (sibling) groups */
5278         for_each_cpu_mask(i, *cpu_map) {
5279                 cpumask_t this_sibling_map = cpu_sibling_map[i];
5280                 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5281                 if (i != first_cpu(this_sibling_map))
5282                         continue;
5283
5284                 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5285                                                 &cpu_to_cpu_group);
5286         }
5287 #endif
5288
5289         /* Set up physical groups */
5290         for (i = 0; i < MAX_NUMNODES; i++) {
5291                 cpumask_t nodemask = node_to_cpumask(i);
5292
5293                 cpus_and(nodemask, nodemask, *cpu_map);
5294                 if (cpus_empty(nodemask))
5295                         continue;
5296
5297                 init_sched_build_groups(sched_group_phys, nodemask,
5298                                                 &cpu_to_phys_group);
5299         }
5300
5301 #ifdef CONFIG_NUMA
5302         /* Set up node groups */
5303         if (sched_group_allnodes)
5304                 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5305                                         &cpu_to_allnodes_group);
5306
5307         for (i = 0; i < MAX_NUMNODES; i++) {
5308                 /* Set up node groups */
5309                 struct sched_group *sg, *prev;
5310                 cpumask_t nodemask = node_to_cpumask(i);
5311                 cpumask_t domainspan;
5312                 cpumask_t covered = CPU_MASK_NONE;
5313                 int j;
5314
5315                 cpus_and(nodemask, nodemask, *cpu_map);
5316                 if (cpus_empty(nodemask)) {
5317                         sched_group_nodes[i] = NULL;
5318                         continue;
5319                 }
5320
5321                 domainspan = sched_domain_node_span(i);
5322                 cpus_and(domainspan, domainspan, *cpu_map);
5323
5324                 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5325                 sched_group_nodes[i] = sg;
5326                 for_each_cpu_mask(j, nodemask) {
5327                         struct sched_domain *sd;
5328                         sd = &per_cpu(node_domains, j);
5329                         sd->groups = sg;
5330                         if (sd->groups == NULL) {
5331                                 /* Turn off balancing if we have no groups */
5332                                 sd->flags = 0;
5333                         }
5334                 }
5335                 if (!sg) {
5336                         printk(KERN_WARNING
5337                         "Can not alloc domain group for node %d\n", i);
5338                         continue;
5339                 }
5340                 sg->cpu_power = 0;
5341                 sg->cpumask = nodemask;
5342                 cpus_or(covered, covered, nodemask);
5343                 prev = sg;
5344
5345                 for (j = 0; j < MAX_NUMNODES; j++) {
5346                         cpumask_t tmp, notcovered;
5347                         int n = (i + j) % MAX_NUMNODES;
5348
5349                         cpus_complement(notcovered, covered);
5350                         cpus_and(tmp, notcovered, *cpu_map);
5351                         cpus_and(tmp, tmp, domainspan);
5352                         if (cpus_empty(tmp))
5353                                 break;
5354
5355                         nodemask = node_to_cpumask(n);
5356                         cpus_and(tmp, tmp, nodemask);
5357                         if (cpus_empty(tmp))
5358                                 continue;
5359
5360                         sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5361                         if (!sg) {
5362                                 printk(KERN_WARNING
5363                                 "Can not alloc domain group for node %d\n", j);
5364                                 break;
5365                         }
5366                         sg->cpu_power = 0;
5367                         sg->cpumask = tmp;
5368                         cpus_or(covered, covered, tmp);
5369                         prev->next = sg;
5370                         prev = sg;
5371                 }
5372                 prev->next = sched_group_nodes[i];
5373         }
5374 #endif
5375
5376         /* Calculate CPU power for physical packages and nodes */
5377         for_each_cpu_mask(i, *cpu_map) {
5378                 int power;
5379                 struct sched_domain *sd;
5380 #ifdef CONFIG_SCHED_SMT
5381                 sd = &per_cpu(cpu_domains, i);
5382                 power = SCHED_LOAD_SCALE;
5383                 sd->groups->cpu_power = power;
5384 #endif
5385
5386                 sd = &per_cpu(phys_domains, i);
5387                 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5388                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5389                 sd->groups->cpu_power = power;
5390
5391 #ifdef CONFIG_NUMA
5392                 sd = &per_cpu(allnodes_domains, i);
5393                 if (sd->groups) {
5394                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5395                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5396                         sd->groups->cpu_power = power;
5397                 }
5398 #endif
5399         }
5400
5401 #ifdef CONFIG_NUMA
5402         for (i = 0; i < MAX_NUMNODES; i++) {
5403                 struct sched_group *sg = sched_group_nodes[i];
5404                 int j;
5405
5406                 if (sg == NULL)
5407                         continue;
5408 next_sg:
5409                 for_each_cpu_mask(j, sg->cpumask) {
5410                         struct sched_domain *sd;
5411                         int power;
5412
5413                         sd = &per_cpu(phys_domains, j);
5414                         if (j != first_cpu(sd->groups->cpumask)) {
5415                                 /*
5416                                  * Only add "power" once for each
5417                                  * physical package.
5418                                  */
5419                                 continue;
5420                         }
5421                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5422                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5423
5424                         sg->cpu_power += power;
5425                 }
5426                 sg = sg->next;
5427                 if (sg != sched_group_nodes[i])
5428                         goto next_sg;
5429         }
5430 #endif
5431
5432         /* Attach the domains */
5433         for_each_cpu_mask(i, *cpu_map) {
5434                 struct sched_domain *sd;
5435 #ifdef CONFIG_SCHED_SMT
5436                 sd = &per_cpu(cpu_domains, i);
5437 #else
5438                 sd = &per_cpu(phys_domains, i);
5439 #endif
5440                 cpu_attach_domain(sd, i);
5441         }
5442 }
5443 /*
5444  * Set up scheduler domains and groups.  Callers must hold the hotplug lock.
5445  */
5446 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5447 {
5448         cpumask_t cpu_default_map;
5449
5450         /*
5451          * Setup mask for cpus without special case scheduling requirements.
5452          * For now this just excludes isolated cpus, but could be used to
5453          * exclude other special cases in the future.
5454          */
5455         cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5456
5457         build_sched_domains(&cpu_default_map);
5458 }
5459
5460 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5461 {
5462 #ifdef CONFIG_NUMA
5463         int i;
5464         int cpu;
5465
5466         for_each_cpu_mask(cpu, *cpu_map) {
5467                 struct sched_group *sched_group_allnodes
5468                         = sched_group_allnodes_bycpu[cpu];
5469                 struct sched_group **sched_group_nodes
5470                         = sched_group_nodes_bycpu[cpu];
5471
5472                 if (sched_group_allnodes) {
5473                         kfree(sched_group_allnodes);
5474                         sched_group_allnodes_bycpu[cpu] = NULL;
5475                 }
5476
5477                 if (!sched_group_nodes)
5478                         continue;
5479
5480                 for (i = 0; i < MAX_NUMNODES; i++) {
5481                         cpumask_t nodemask = node_to_cpumask(i);
5482                         struct sched_group *oldsg, *sg = sched_group_nodes[i];
5483
5484                         cpus_and(nodemask, nodemask, *cpu_map);
5485                         if (cpus_empty(nodemask))
5486                                 continue;
5487
5488                         if (sg == NULL)
5489                                 continue;
5490                         sg = sg->next;
5491 next_sg:
5492                         oldsg = sg;
5493                         sg = sg->next;
5494                         kfree(oldsg);
5495                         if (oldsg != sched_group_nodes[i])
5496                                 goto next_sg;
5497                 }
5498                 kfree(sched_group_nodes);
5499                 sched_group_nodes_bycpu[cpu] = NULL;
5500         }
5501 #endif
5502 }
5503
5504 /*
5505  * Detach sched domains from a group of cpus specified in cpu_map
5506  * These cpus will now be attached to the NULL domain
5507  */
5508 static inline void detach_destroy_domains(const cpumask_t *cpu_map)
5509 {
5510         int i;
5511
5512         for_each_cpu_mask(i, *cpu_map)
5513                 cpu_attach_domain(NULL, i);
5514         synchronize_sched();
5515         arch_destroy_sched_domains(cpu_map);
5516 }
5517
5518 /*
5519  * Partition sched domains as specified by the cpumasks below.
5520  * This attaches all cpus from the cpumasks to the NULL domain,
5521  * waits for a RCU quiescent period, recalculates sched
5522  * domain information and then attaches them back to the
5523  * correct sched domains
5524  * Call with hotplug lock held
5525  */
5526 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
5527 {
5528         cpumask_t change_map;
5529
5530         cpus_and(*partition1, *partition1, cpu_online_map);
5531         cpus_and(*partition2, *partition2, cpu_online_map);
5532         cpus_or(change_map, *partition1, *partition2);
5533
5534         /* Detach sched domains from all of the affected cpus */
5535         detach_destroy_domains(&change_map);
5536         if (!cpus_empty(*partition1))
5537                 build_sched_domains(partition1);
5538         if (!cpus_empty(*partition2))
5539                 build_sched_domains(partition2);
5540 }
5541
5542 #ifdef CONFIG_HOTPLUG_CPU
5543 /*
5544  * Force a reinitialization of the sched domains hierarchy.  The domains
5545  * and groups cannot be updated in place without racing with the balancing
5546  * code, so we temporarily attach all running cpus to the NULL domain
5547  * which will prevent rebalancing while the sched domains are recalculated.
5548  */
5549 static int update_sched_domains(struct notifier_block *nfb,
5550                                 unsigned long action, void *hcpu)
5551 {
5552         switch (action) {
5553         case CPU_UP_PREPARE:
5554         case CPU_DOWN_PREPARE:
5555                 detach_destroy_domains(&cpu_online_map);
5556                 return NOTIFY_OK;
5557
5558         case CPU_UP_CANCELED:
5559         case CPU_DOWN_FAILED:
5560         case CPU_ONLINE:
5561         case CPU_DEAD:
5562                 /*
5563                  * Fall through and re-initialise the domains.
5564                  */
5565                 break;
5566         default:
5567                 return NOTIFY_DONE;
5568         }
5569
5570         /* The hotplug lock is already held by cpu_up/cpu_down */
5571         arch_init_sched_domains(&cpu_online_map);
5572
5573         return NOTIFY_OK;
5574 }
5575 #endif
5576
5577 void __init sched_init_smp(void)
5578 {
5579         lock_cpu_hotplug();
5580         arch_init_sched_domains(&cpu_online_map);
5581         unlock_cpu_hotplug();
5582         /* XXX: Theoretical race here - CPU may be hotplugged now */
5583         hotcpu_notifier(update_sched_domains, 0);
5584 }
5585 #else
5586 void __init sched_init_smp(void)
5587 {
5588 }
5589 #endif /* CONFIG_SMP */
5590
5591 int in_sched_functions(unsigned long addr)
5592 {
5593         /* Linker adds these: start and end of __sched functions */
5594         extern char __sched_text_start[], __sched_text_end[];
5595         return in_lock_functions(addr) ||
5596                 (addr >= (unsigned long)__sched_text_start
5597                 && addr < (unsigned long)__sched_text_end);
5598 }
5599
5600 void __init sched_init(void)
5601 {
5602         runqueue_t *rq;
5603         int i, j, k;
5604
5605         for (i = 0; i < NR_CPUS; i++) {
5606                 prio_array_t *array;
5607
5608                 rq = cpu_rq(i);
5609                 spin_lock_init(&rq->lock);
5610                 rq->nr_running = 0;
5611                 rq->active = rq->arrays;
5612                 rq->expired = rq->arrays + 1;
5613                 rq->best_expired_prio = MAX_PRIO;
5614
5615 #ifdef CONFIG_SMP
5616                 rq->sd = NULL;
5617                 for (j = 1; j < 3; j++)
5618                         rq->cpu_load[j] = 0;
5619                 rq->active_balance = 0;
5620                 rq->push_cpu = 0;
5621                 rq->migration_thread = NULL;
5622                 INIT_LIST_HEAD(&rq->migration_queue);
5623 #endif
5624                 atomic_set(&rq->nr_iowait, 0);
5625
5626                 for (j = 0; j < 2; j++) {
5627                         array = rq->arrays + j;
5628                         for (k = 0; k < MAX_PRIO; k++) {
5629                                 INIT_LIST_HEAD(array->queue + k);
5630                                 __clear_bit(k, array->bitmap);
5631                         }
5632                         // delimiter for bitsearch
5633                         __set_bit(MAX_PRIO, array->bitmap);
5634                 }
5635         }
5636
5637         /*
5638          * The boot idle thread does lazy MMU switching as well:
5639          */
5640         atomic_inc(&init_mm.mm_count);
5641         enter_lazy_tlb(&init_mm, current);
5642
5643         /*
5644          * Make us the idle thread. Technically, schedule() should not be
5645          * called from this thread, however somewhere below it might be,
5646          * but because we are the idle thread, we just pick up running again
5647          * when this runqueue becomes "idle".
5648          */
5649         init_idle(current, smp_processor_id());
5650 }
5651
5652 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5653 void __might_sleep(char *file, int line)
5654 {
5655 #if defined(in_atomic)
5656         static unsigned long prev_jiffy;        /* ratelimiting */
5657
5658         if ((in_atomic() || irqs_disabled()) &&
5659             system_state == SYSTEM_RUNNING && !oops_in_progress) {
5660                 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5661                         return;
5662                 prev_jiffy = jiffies;
5663                 printk(KERN_ERR "Debug: sleeping function called from invalid"
5664                                 " context at %s:%d\n", file, line);
5665                 printk("in_atomic():%d, irqs_disabled():%d\n",
5666                         in_atomic(), irqs_disabled());
5667                 dump_stack();
5668         }
5669 #endif
5670 }
5671 EXPORT_SYMBOL(__might_sleep);
5672 #endif
5673
5674 #ifdef CONFIG_MAGIC_SYSRQ
5675 void normalize_rt_tasks(void)
5676 {
5677         struct task_struct *p;
5678         prio_array_t *array;
5679         unsigned long flags;
5680         runqueue_t *rq;
5681
5682         read_lock_irq(&tasklist_lock);
5683         for_each_process (p) {
5684                 if (!rt_task(p))
5685                         continue;
5686
5687                 rq = task_rq_lock(p, &flags);
5688
5689                 array = p->array;
5690                 if (array)
5691                         deactivate_task(p, task_rq(p));
5692                 __setscheduler(p, SCHED_NORMAL, 0);
5693                 if (array) {
5694                         __activate_task(p, task_rq(p));
5695                         resched_task(rq->curr);
5696                 }
5697
5698                 task_rq_unlock(rq, &flags);
5699         }
5700         read_unlock_irq(&tasklist_lock);
5701 }
5702
5703 #endif /* CONFIG_MAGIC_SYSRQ */
5704
5705 #ifdef CONFIG_IA64
5706 /*
5707  * These functions are only useful for the IA64 MCA handling.
5708  *
5709  * They can only be called when the whole system has been
5710  * stopped - every CPU needs to be quiescent, and no scheduling
5711  * activity can take place. Using them for anything else would
5712  * be a serious bug, and as a result, they aren't even visible
5713  * under any other configuration.
5714  */
5715
5716 /**
5717  * curr_task - return the current task for a given cpu.
5718  * @cpu: the processor in question.
5719  *
5720  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5721  */
5722 task_t *curr_task(int cpu)
5723 {
5724         return cpu_curr(cpu);
5725 }
5726
5727 /**
5728  * set_curr_task - set the current task for a given cpu.
5729  * @cpu: the processor in question.
5730  * @p: the task pointer to set.
5731  *
5732  * Description: This function must only be used when non-maskable interrupts
5733  * are serviced on a separate stack.  It allows the architecture to switch the
5734  * notion of the current task on a cpu in a non-blocking manner.  This function
5735  * must be called with all CPU's synchronized, and interrupts disabled, the
5736  * and caller must save the original value of the current task (see
5737  * curr_task() above) and restore that value before reenabling interrupts and
5738  * re-starting the system.
5739  *
5740  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5741  */
5742 void set_curr_task(int cpu, task_t *p)
5743 {
5744         cpu_curr(cpu) = p;
5745 }
5746
5747 #endif