c0c60c926d5eafe1ebd57a0f298324933cb46507
[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/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/security.h>
34 #include <linux/notifier.h>
35 #include <linux/profile.h>
36 #include <linux/suspend.h>
37 #include <linux/blkdev.h>
38 #include <linux/delay.h>
39 #include <linux/smp.h>
40 #include <linux/threads.h>
41 #include <linux/timer.h>
42 #include <linux/rcupdate.h>
43 #include <linux/cpu.h>
44 #include <linux/cpuset.h>
45 #include <linux/percpu.h>
46 #include <linux/kthread.h>
47 #include <linux/seq_file.h>
48 #include <linux/syscalls.h>
49 #include <linux/times.h>
50 #include <linux/acct.h>
51 #include <asm/tlb.h>
52
53 #include <asm/unistd.h>
54
55 /*
56  * Convert user-nice values [ -20 ... 0 ... 19 ]
57  * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
58  * and back.
59  */
60 #define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
61 #define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
62 #define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
63
64 /*
65  * 'User priority' is the nice value converted to something we
66  * can work with better when scaling various scheduler parameters,
67  * it's a [ 0 ... 39 ] range.
68  */
69 #define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
70 #define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
71 #define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
72
73 /*
74  * Some helpers for converting nanosecond timing to jiffy resolution
75  */
76 #define NS_TO_JIFFIES(TIME)     ((TIME) / (1000000000 / HZ))
77 #define JIFFIES_TO_NS(TIME)     ((TIME) * (1000000000 / HZ))
78
79 /*
80  * These are the 'tuning knobs' of the scheduler:
81  *
82  * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
83  * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
84  * Timeslices get refilled after they expire.
85  */
86 #define MIN_TIMESLICE           max(5 * HZ / 1000, 1)
87 #define DEF_TIMESLICE           (100 * HZ / 1000)
88 #define ON_RUNQUEUE_WEIGHT       30
89 #define CHILD_PENALTY            95
90 #define PARENT_PENALTY          100
91 #define EXIT_WEIGHT               3
92 #define PRIO_BONUS_RATIO         25
93 #define MAX_BONUS               (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
94 #define INTERACTIVE_DELTA         2
95 #define MAX_SLEEP_AVG           (DEF_TIMESLICE * MAX_BONUS)
96 #define STARVATION_LIMIT        (MAX_SLEEP_AVG)
97 #define NS_MAX_SLEEP_AVG        (JIFFIES_TO_NS(MAX_SLEEP_AVG))
98
99 /*
100  * If a task is 'interactive' then we reinsert it in the active
101  * array after it has expired its current timeslice. (it will not
102  * continue to run immediately, it will still roundrobin with
103  * other interactive tasks.)
104  *
105  * This part scales the interactivity limit depending on niceness.
106  *
107  * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
108  * Here are a few examples of different nice levels:
109  *
110  *  TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
111  *  TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
112  *  TASK_INTERACTIVE(  0): [1,1,1,1,0,0,0,0,0,0,0]
113  *  TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
114  *  TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
115  *
116  * (the X axis represents the possible -5 ... 0 ... +5 dynamic
117  *  priority range a task can explore, a value of '1' means the
118  *  task is rated interactive.)
119  *
120  * Ie. nice +19 tasks can never get 'interactive' enough to be
121  * reinserted into the active array. And only heavily CPU-hog nice -20
122  * tasks will be expired. Default nice 0 tasks are somewhere between,
123  * it takes some effort for them to get interactive, but it's not
124  * too hard.
125  */
126
127 #define CURRENT_BONUS(p) \
128         (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
129                 MAX_SLEEP_AVG)
130
131 #define GRANULARITY     (10 * HZ / 1000 ? : 1)
132
133 #ifdef CONFIG_SMP
134 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
135                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
136                         num_online_cpus())
137 #else
138 #define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
139                 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
140 #endif
141
142 #define SCALE(v1,v1_max,v2_max) \
143         (v1) * (v2_max) / (v1_max)
144
145 #define DELTA(p) \
146         (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
147
148 #define TASK_INTERACTIVE(p) \
149         ((p)->prio <= (p)->static_prio - DELTA(p))
150
151 #define INTERACTIVE_SLEEP(p) \
152         (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
153                 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
154
155 #define TASK_PREEMPTS_CURR(p, rq) \
156         ((p)->prio < (rq)->curr->prio)
157
158 /*
159  * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
160  * to time slice values: [800ms ... 100ms ... 5ms]
161  *
162  * The higher a thread's priority, the bigger timeslices
163  * it gets during one round of execution. But even the lowest
164  * priority thread gets MIN_TIMESLICE worth of execution time.
165  */
166
167 #define SCALE_PRIO(x, prio) \
168         max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
169
170 static unsigned int task_timeslice(task_t *p)
171 {
172         if (p->static_prio < NICE_TO_PRIO(0))
173                 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
174         else
175                 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
176 }
177 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran)       \
178                                 < (long long) (sd)->cache_hot_time)
179
180 void __put_task_struct_cb(struct rcu_head *rhp)
181 {
182         __put_task_struct(container_of(rhp, struct task_struct, rcu));
183 }
184
185 EXPORT_SYMBOL_GPL(__put_task_struct_cb);
186
187 /*
188  * These are the runqueue data structures:
189  */
190
191 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
192
193 typedef struct runqueue runqueue_t;
194
195 struct prio_array {
196         unsigned int nr_active;
197         unsigned long bitmap[BITMAP_SIZE];
198         struct list_head queue[MAX_PRIO];
199 };
200
201 /*
202  * This is the main, per-CPU runqueue data structure.
203  *
204  * Locking rule: those places that want to lock multiple runqueues
205  * (such as the load balancing or the thread migration code), lock
206  * acquire operations must be ordered by ascending &runqueue.
207  */
208 struct runqueue {
209         spinlock_t lock;
210
211         /*
212          * nr_running and cpu_load should be in the same cacheline because
213          * remote CPUs use both these fields when doing load calculation.
214          */
215         unsigned long nr_running;
216 #ifdef CONFIG_SMP
217         unsigned long prio_bias;
218         unsigned long cpu_load[3];
219 #endif
220         unsigned long long nr_switches;
221
222         /*
223          * This is part of a global counter where only the total sum
224          * over all CPUs matters. A task can increase this counter on
225          * one CPU and if it got migrated afterwards it may decrease
226          * it on another CPU. Always updated under the runqueue lock:
227          */
228         unsigned long nr_uninterruptible;
229
230         unsigned long expired_timestamp;
231         unsigned long long timestamp_last_tick;
232         task_t *curr, *idle;
233         struct mm_struct *prev_mm;
234         prio_array_t *active, *expired, arrays[2];
235         int best_expired_prio;
236         atomic_t nr_iowait;
237
238 #ifdef CONFIG_SMP
239         struct sched_domain *sd;
240
241         /* For active balancing */
242         int active_balance;
243         int push_cpu;
244
245         task_t *migration_thread;
246         struct list_head migration_queue;
247 #endif
248
249 #ifdef CONFIG_SCHEDSTATS
250         /* latency stats */
251         struct sched_info rq_sched_info;
252
253         /* sys_sched_yield() stats */
254         unsigned long yld_exp_empty;
255         unsigned long yld_act_empty;
256         unsigned long yld_both_empty;
257         unsigned long yld_cnt;
258
259         /* schedule() stats */
260         unsigned long sched_switch;
261         unsigned long sched_cnt;
262         unsigned long sched_goidle;
263
264         /* try_to_wake_up() stats */
265         unsigned long ttwu_cnt;
266         unsigned long ttwu_local;
267 #endif
268 };
269
270 static DEFINE_PER_CPU(struct runqueue, runqueues);
271
272 /*
273  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
274  * See detach_destroy_domains: synchronize_sched for details.
275  *
276  * The domain tree of any CPU may only be accessed from within
277  * preempt-disabled sections.
278  */
279 #define for_each_domain(cpu, domain) \
280 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
281
282 #define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
283 #define this_rq()               (&__get_cpu_var(runqueues))
284 #define task_rq(p)              cpu_rq(task_cpu(p))
285 #define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
286
287 #ifndef prepare_arch_switch
288 # define prepare_arch_switch(next)      do { } while (0)
289 #endif
290 #ifndef finish_arch_switch
291 # define finish_arch_switch(prev)       do { } while (0)
292 #endif
293
294 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
295 static inline int task_running(runqueue_t *rq, task_t *p)
296 {
297         return rq->curr == p;
298 }
299
300 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
301 {
302 }
303
304 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
305 {
306 #ifdef CONFIG_DEBUG_SPINLOCK
307         /* this is a valid case when another task releases the spinlock */
308         rq->lock.owner = current;
309 #endif
310         spin_unlock_irq(&rq->lock);
311 }
312
313 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
314 static inline int task_running(runqueue_t *rq, task_t *p)
315 {
316 #ifdef CONFIG_SMP
317         return p->oncpu;
318 #else
319         return rq->curr == p;
320 #endif
321 }
322
323 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
324 {
325 #ifdef CONFIG_SMP
326         /*
327          * We can optimise this out completely for !SMP, because the
328          * SMP rebalancing from interrupt is the only thing that cares
329          * here.
330          */
331         next->oncpu = 1;
332 #endif
333 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
334         spin_unlock_irq(&rq->lock);
335 #else
336         spin_unlock(&rq->lock);
337 #endif
338 }
339
340 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
341 {
342 #ifdef CONFIG_SMP
343         /*
344          * After ->oncpu is cleared, the task can be moved to a different CPU.
345          * We must ensure this doesn't happen until the switch is completely
346          * finished.
347          */
348         smp_wmb();
349         prev->oncpu = 0;
350 #endif
351 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
352         local_irq_enable();
353 #endif
354 }
355 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
356
357 /*
358  * task_rq_lock - lock the runqueue a given task resides on and disable
359  * interrupts.  Note the ordering: we can safely lookup the task_rq without
360  * explicitly disabling preemption.
361  */
362 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
363         __acquires(rq->lock)
364 {
365         struct runqueue *rq;
366
367 repeat_lock_task:
368         local_irq_save(*flags);
369         rq = task_rq(p);
370         spin_lock(&rq->lock);
371         if (unlikely(rq != task_rq(p))) {
372                 spin_unlock_irqrestore(&rq->lock, *flags);
373                 goto repeat_lock_task;
374         }
375         return rq;
376 }
377
378 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
379         __releases(rq->lock)
380 {
381         spin_unlock_irqrestore(&rq->lock, *flags);
382 }
383
384 #ifdef CONFIG_SCHEDSTATS
385 /*
386  * bump this up when changing the output format or the meaning of an existing
387  * format, so that tools can adapt (or abort)
388  */
389 #define SCHEDSTAT_VERSION 12
390
391 static int show_schedstat(struct seq_file *seq, void *v)
392 {
393         int cpu;
394
395         seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
396         seq_printf(seq, "timestamp %lu\n", jiffies);
397         for_each_online_cpu(cpu) {
398                 runqueue_t *rq = cpu_rq(cpu);
399 #ifdef CONFIG_SMP
400                 struct sched_domain *sd;
401                 int dcnt = 0;
402 #endif
403
404                 /* runqueue-specific stats */
405                 seq_printf(seq,
406                     "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
407                     cpu, rq->yld_both_empty,
408                     rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
409                     rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
410                     rq->ttwu_cnt, rq->ttwu_local,
411                     rq->rq_sched_info.cpu_time,
412                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
413
414                 seq_printf(seq, "\n");
415
416 #ifdef CONFIG_SMP
417                 /* domain-specific stats */
418                 preempt_disable();
419                 for_each_domain(cpu, sd) {
420                         enum idle_type itype;
421                         char mask_str[NR_CPUS];
422
423                         cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
424                         seq_printf(seq, "domain%d %s", dcnt++, mask_str);
425                         for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
426                                         itype++) {
427                                 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
428                                     sd->lb_cnt[itype],
429                                     sd->lb_balanced[itype],
430                                     sd->lb_failed[itype],
431                                     sd->lb_imbalance[itype],
432                                     sd->lb_gained[itype],
433                                     sd->lb_hot_gained[itype],
434                                     sd->lb_nobusyq[itype],
435                                     sd->lb_nobusyg[itype]);
436                         }
437                         seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
438                             sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
439                             sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
440                             sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
441                             sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
442                 }
443                 preempt_enable();
444 #endif
445         }
446         return 0;
447 }
448
449 static int schedstat_open(struct inode *inode, struct file *file)
450 {
451         unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
452         char *buf = kmalloc(size, GFP_KERNEL);
453         struct seq_file *m;
454         int res;
455
456         if (!buf)
457                 return -ENOMEM;
458         res = single_open(file, show_schedstat, NULL);
459         if (!res) {
460                 m = file->private_data;
461                 m->buf = buf;
462                 m->size = size;
463         } else
464                 kfree(buf);
465         return res;
466 }
467
468 struct file_operations proc_schedstat_operations = {
469         .open    = schedstat_open,
470         .read    = seq_read,
471         .llseek  = seq_lseek,
472         .release = single_release,
473 };
474
475 # define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
476 # define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
477 #else /* !CONFIG_SCHEDSTATS */
478 # define schedstat_inc(rq, field)       do { } while (0)
479 # define schedstat_add(rq, field, amt)  do { } while (0)
480 #endif
481
482 /*
483  * rq_lock - lock a given runqueue and disable interrupts.
484  */
485 static inline runqueue_t *this_rq_lock(void)
486         __acquires(rq->lock)
487 {
488         runqueue_t *rq;
489
490         local_irq_disable();
491         rq = this_rq();
492         spin_lock(&rq->lock);
493
494         return rq;
495 }
496
497 #ifdef CONFIG_SCHEDSTATS
498 /*
499  * Called when a process is dequeued from the active array and given
500  * the cpu.  We should note that with the exception of interactive
501  * tasks, the expired queue will become the active queue after the active
502  * queue is empty, without explicitly dequeuing and requeuing tasks in the
503  * expired queue.  (Interactive tasks may be requeued directly to the
504  * active queue, thus delaying tasks in the expired queue from running;
505  * see scheduler_tick()).
506  *
507  * This function is only called from sched_info_arrive(), rather than
508  * dequeue_task(). Even though a task may be queued and dequeued multiple
509  * times as it is shuffled about, we're really interested in knowing how
510  * long it was from the *first* time it was queued to the time that it
511  * finally hit a cpu.
512  */
513 static inline void sched_info_dequeued(task_t *t)
514 {
515         t->sched_info.last_queued = 0;
516 }
517
518 /*
519  * Called when a task finally hits the cpu.  We can now calculate how
520  * long it was waiting to run.  We also note when it began so that we
521  * can keep stats on how long its timeslice is.
522  */
523 static inline void sched_info_arrive(task_t *t)
524 {
525         unsigned long now = jiffies, diff = 0;
526         struct runqueue *rq = task_rq(t);
527
528         if (t->sched_info.last_queued)
529                 diff = now - t->sched_info.last_queued;
530         sched_info_dequeued(t);
531         t->sched_info.run_delay += diff;
532         t->sched_info.last_arrival = now;
533         t->sched_info.pcnt++;
534
535         if (!rq)
536                 return;
537
538         rq->rq_sched_info.run_delay += diff;
539         rq->rq_sched_info.pcnt++;
540 }
541
542 /*
543  * Called when a process is queued into either the active or expired
544  * array.  The time is noted and later used to determine how long we
545  * had to wait for us to reach the cpu.  Since the expired queue will
546  * become the active queue after active queue is empty, without dequeuing
547  * and requeuing any tasks, we are interested in queuing to either. It
548  * is unusual but not impossible for tasks to be dequeued and immediately
549  * requeued in the same or another array: this can happen in sched_yield(),
550  * set_user_nice(), and even load_balance() as it moves tasks from runqueue
551  * to runqueue.
552  *
553  * This function is only called from enqueue_task(), but also only updates
554  * the timestamp if it is already not set.  It's assumed that
555  * sched_info_dequeued() will clear that stamp when appropriate.
556  */
557 static inline void sched_info_queued(task_t *t)
558 {
559         if (!t->sched_info.last_queued)
560                 t->sched_info.last_queued = jiffies;
561 }
562
563 /*
564  * Called when a process ceases being the active-running process, either
565  * voluntarily or involuntarily.  Now we can calculate how long we ran.
566  */
567 static inline void sched_info_depart(task_t *t)
568 {
569         struct runqueue *rq = task_rq(t);
570         unsigned long diff = jiffies - t->sched_info.last_arrival;
571
572         t->sched_info.cpu_time += diff;
573
574         if (rq)
575                 rq->rq_sched_info.cpu_time += diff;
576 }
577
578 /*
579  * Called when tasks are switched involuntarily due, typically, to expiring
580  * their time slice.  (This may also be called when switching to or from
581  * the idle task.)  We are only called when prev != next.
582  */
583 static inline void sched_info_switch(task_t *prev, task_t *next)
584 {
585         struct runqueue *rq = task_rq(prev);
586
587         /*
588          * prev now departs the cpu.  It's not interesting to record
589          * stats about how efficient we were at scheduling the idle
590          * process, however.
591          */
592         if (prev != rq->idle)
593                 sched_info_depart(prev);
594
595         if (next != rq->idle)
596                 sched_info_arrive(next);
597 }
598 #else
599 #define sched_info_queued(t)            do { } while (0)
600 #define sched_info_switch(t, next)      do { } while (0)
601 #endif /* CONFIG_SCHEDSTATS */
602
603 /*
604  * Adding/removing a task to/from a priority array:
605  */
606 static void dequeue_task(struct task_struct *p, prio_array_t *array)
607 {
608         array->nr_active--;
609         list_del(&p->run_list);
610         if (list_empty(array->queue + p->prio))
611                 __clear_bit(p->prio, array->bitmap);
612 }
613
614 static void enqueue_task(struct task_struct *p, prio_array_t *array)
615 {
616         sched_info_queued(p);
617         list_add_tail(&p->run_list, array->queue + p->prio);
618         __set_bit(p->prio, array->bitmap);
619         array->nr_active++;
620         p->array = array;
621 }
622
623 /*
624  * Put task to the end of the run list without the overhead of dequeue
625  * followed by enqueue.
626  */
627 static void requeue_task(struct task_struct *p, prio_array_t *array)
628 {
629         list_move_tail(&p->run_list, array->queue + p->prio);
630 }
631
632 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
633 {
634         list_add(&p->run_list, array->queue + p->prio);
635         __set_bit(p->prio, array->bitmap);
636         array->nr_active++;
637         p->array = array;
638 }
639
640 /*
641  * effective_prio - return the priority that is based on the static
642  * priority but is modified by bonuses/penalties.
643  *
644  * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
645  * into the -5 ... 0 ... +5 bonus/penalty range.
646  *
647  * We use 25% of the full 0...39 priority range so that:
648  *
649  * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
650  * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
651  *
652  * Both properties are important to certain workloads.
653  */
654 static int effective_prio(task_t *p)
655 {
656         int bonus, prio;
657
658         if (rt_task(p))
659                 return p->prio;
660
661         bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
662
663         prio = p->static_prio - bonus;
664         if (prio < MAX_RT_PRIO)
665                 prio = MAX_RT_PRIO;
666         if (prio > MAX_PRIO-1)
667                 prio = MAX_PRIO-1;
668         return prio;
669 }
670
671 #ifdef CONFIG_SMP
672 static inline void inc_prio_bias(runqueue_t *rq, int prio)
673 {
674         rq->prio_bias += MAX_PRIO - prio;
675 }
676
677 static inline void dec_prio_bias(runqueue_t *rq, int prio)
678 {
679         rq->prio_bias -= MAX_PRIO - prio;
680 }
681
682 static inline void inc_nr_running(task_t *p, runqueue_t *rq)
683 {
684         rq->nr_running++;
685         if (rt_task(p)) {
686                 if (p != rq->migration_thread)
687                         /*
688                          * The migration thread does the actual balancing. Do
689                          * not bias by its priority as the ultra high priority
690                          * will skew balancing adversely.
691                          */
692                         inc_prio_bias(rq, p->prio);
693         } else
694                 inc_prio_bias(rq, p->static_prio);
695 }
696
697 static inline void dec_nr_running(task_t *p, runqueue_t *rq)
698 {
699         rq->nr_running--;
700         if (rt_task(p)) {
701                 if (p != rq->migration_thread)
702                         dec_prio_bias(rq, p->prio);
703         } else
704                 dec_prio_bias(rq, p->static_prio);
705 }
706 #else
707 static inline void inc_prio_bias(runqueue_t *rq, int prio)
708 {
709 }
710
711 static inline void dec_prio_bias(runqueue_t *rq, int prio)
712 {
713 }
714
715 static inline void inc_nr_running(task_t *p, runqueue_t *rq)
716 {
717         rq->nr_running++;
718 }
719
720 static inline void dec_nr_running(task_t *p, runqueue_t *rq)
721 {
722         rq->nr_running--;
723 }
724 #endif
725
726 /*
727  * __activate_task - move a task to the runqueue.
728  */
729 static inline void __activate_task(task_t *p, runqueue_t *rq)
730 {
731         enqueue_task(p, rq->active);
732         inc_nr_running(p, rq);
733 }
734
735 /*
736  * __activate_idle_task - move idle task to the _front_ of runqueue.
737  */
738 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
739 {
740         enqueue_task_head(p, rq->active);
741         inc_nr_running(p, rq);
742 }
743
744 static int recalc_task_prio(task_t *p, unsigned long long now)
745 {
746         /* Caller must always ensure 'now >= p->timestamp' */
747         unsigned long long __sleep_time = now - p->timestamp;
748         unsigned long sleep_time;
749
750         if (__sleep_time > NS_MAX_SLEEP_AVG)
751                 sleep_time = NS_MAX_SLEEP_AVG;
752         else
753                 sleep_time = (unsigned long)__sleep_time;
754
755         if (likely(sleep_time > 0)) {
756                 /*
757                  * User tasks that sleep a long time are categorised as
758                  * idle and will get just interactive status to stay active &
759                  * prevent them suddenly becoming cpu hogs and starving
760                  * other processes.
761                  */
762                 if (p->mm && p->activated != -1 &&
763                         sleep_time > INTERACTIVE_SLEEP(p)) {
764                                 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
765                                                 DEF_TIMESLICE);
766                 } else {
767                         /*
768                          * The lower the sleep avg a task has the more
769                          * rapidly it will rise with sleep time.
770                          */
771                         sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
772
773                         /*
774                          * Tasks waking from uninterruptible sleep are
775                          * limited in their sleep_avg rise as they
776                          * are likely to be waiting on I/O
777                          */
778                         if (p->activated == -1 && p->mm) {
779                                 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
780                                         sleep_time = 0;
781                                 else if (p->sleep_avg + sleep_time >=
782                                                 INTERACTIVE_SLEEP(p)) {
783                                         p->sleep_avg = INTERACTIVE_SLEEP(p);
784                                         sleep_time = 0;
785                                 }
786                         }
787
788                         /*
789                          * This code gives a bonus to interactive tasks.
790                          *
791                          * The boost works by updating the 'average sleep time'
792                          * value here, based on ->timestamp. The more time a
793                          * task spends sleeping, the higher the average gets -
794                          * and the higher the priority boost gets as well.
795                          */
796                         p->sleep_avg += sleep_time;
797
798                         if (p->sleep_avg > NS_MAX_SLEEP_AVG)
799                                 p->sleep_avg = NS_MAX_SLEEP_AVG;
800                 }
801         }
802
803         return effective_prio(p);
804 }
805
806 /*
807  * activate_task - move a task to the runqueue and do priority recalculation
808  *
809  * Update all the scheduling statistics stuff. (sleep average
810  * calculation, priority modifiers, etc.)
811  */
812 static void activate_task(task_t *p, runqueue_t *rq, int local)
813 {
814         unsigned long long now;
815
816         now = sched_clock();
817 #ifdef CONFIG_SMP
818         if (!local) {
819                 /* Compensate for drifting sched_clock */
820                 runqueue_t *this_rq = this_rq();
821                 now = (now - this_rq->timestamp_last_tick)
822                         + rq->timestamp_last_tick;
823         }
824 #endif
825
826         if (!rt_task(p))
827                 p->prio = recalc_task_prio(p, now);
828
829         /*
830          * This checks to make sure it's not an uninterruptible task
831          * that is now waking up.
832          */
833         if (!p->activated) {
834                 /*
835                  * Tasks which were woken up by interrupts (ie. hw events)
836                  * are most likely of interactive nature. So we give them
837                  * the credit of extending their sleep time to the period
838                  * of time they spend on the runqueue, waiting for execution
839                  * on a CPU, first time around:
840                  */
841                 if (in_interrupt())
842                         p->activated = 2;
843                 else {
844                         /*
845                          * Normal first-time wakeups get a credit too for
846                          * on-runqueue time, but it will be weighted down:
847                          */
848                         p->activated = 1;
849                 }
850         }
851         p->timestamp = now;
852
853         __activate_task(p, rq);
854 }
855
856 /*
857  * deactivate_task - remove a task from the runqueue.
858  */
859 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
860 {
861         dec_nr_running(p, rq);
862         dequeue_task(p, p->array);
863         p->array = NULL;
864 }
865
866 /*
867  * resched_task - mark a task 'to be rescheduled now'.
868  *
869  * On UP this means the setting of the need_resched flag, on SMP it
870  * might also involve a cross-CPU call to trigger the scheduler on
871  * the target CPU.
872  */
873 #ifdef CONFIG_SMP
874 static void resched_task(task_t *p)
875 {
876         int cpu;
877
878         assert_spin_locked(&task_rq(p)->lock);
879
880         if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
881                 return;
882
883         set_tsk_thread_flag(p, TIF_NEED_RESCHED);
884
885         cpu = task_cpu(p);
886         if (cpu == smp_processor_id())
887                 return;
888
889         /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
890         smp_mb();
891         if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG))
892                 smp_send_reschedule(cpu);
893 }
894 #else
895 static inline void resched_task(task_t *p)
896 {
897         assert_spin_locked(&task_rq(p)->lock);
898         set_tsk_need_resched(p);
899 }
900 #endif
901
902 /**
903  * task_curr - is this task currently executing on a CPU?
904  * @p: the task in question.
905  */
906 inline int task_curr(const task_t *p)
907 {
908         return cpu_curr(task_cpu(p)) == p;
909 }
910
911 #ifdef CONFIG_SMP
912 typedef struct {
913         struct list_head list;
914
915         task_t *task;
916         int dest_cpu;
917
918         struct completion done;
919 } migration_req_t;
920
921 /*
922  * The task's runqueue lock must be held.
923  * Returns true if you have to wait for migration thread.
924  */
925 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
926 {
927         runqueue_t *rq = task_rq(p);
928
929         /*
930          * If the task is not on a runqueue (and not running), then
931          * it is sufficient to simply update the task's cpu field.
932          */
933         if (!p->array && !task_running(rq, p)) {
934                 set_task_cpu(p, dest_cpu);
935                 return 0;
936         }
937
938         init_completion(&req->done);
939         req->task = p;
940         req->dest_cpu = dest_cpu;
941         list_add(&req->list, &rq->migration_queue);
942         return 1;
943 }
944
945 /*
946  * wait_task_inactive - wait for a thread to unschedule.
947  *
948  * The caller must ensure that the task *will* unschedule sometime soon,
949  * else this function might spin for a *long* time. This function can't
950  * be called with interrupts off, or it may introduce deadlock with
951  * smp_call_function() if an IPI is sent by the same process we are
952  * waiting to become inactive.
953  */
954 void wait_task_inactive(task_t *p)
955 {
956         unsigned long flags;
957         runqueue_t *rq;
958         int preempted;
959
960 repeat:
961         rq = task_rq_lock(p, &flags);
962         /* Must be off runqueue entirely, not preempted. */
963         if (unlikely(p->array || task_running(rq, p))) {
964                 /* If it's preempted, we yield.  It could be a while. */
965                 preempted = !task_running(rq, p);
966                 task_rq_unlock(rq, &flags);
967                 cpu_relax();
968                 if (preempted)
969                         yield();
970                 goto repeat;
971         }
972         task_rq_unlock(rq, &flags);
973 }
974
975 /***
976  * kick_process - kick a running thread to enter/exit the kernel
977  * @p: the to-be-kicked thread
978  *
979  * Cause a process which is running on another CPU to enter
980  * kernel-mode, without any delay. (to get signals handled.)
981  *
982  * NOTE: this function doesnt have to take the runqueue lock,
983  * because all it wants to ensure is that the remote task enters
984  * the kernel. If the IPI races and the task has been migrated
985  * to another CPU then no harm is done and the purpose has been
986  * achieved as well.
987  */
988 void kick_process(task_t *p)
989 {
990         int cpu;
991
992         preempt_disable();
993         cpu = task_cpu(p);
994         if ((cpu != smp_processor_id()) && task_curr(p))
995                 smp_send_reschedule(cpu);
996         preempt_enable();
997 }
998
999 /*
1000  * Return a low guess at the load of a migration-source cpu.
1001  *
1002  * We want to under-estimate the load of migration sources, to
1003  * balance conservatively.
1004  */
1005 static inline unsigned long __source_load(int cpu, int type, enum idle_type idle)
1006 {
1007         runqueue_t *rq = cpu_rq(cpu);
1008         unsigned long running = rq->nr_running;
1009         unsigned long source_load, cpu_load = rq->cpu_load[type-1],
1010                 load_now = running * SCHED_LOAD_SCALE;
1011
1012         if (type == 0)
1013                 source_load = load_now;
1014         else
1015                 source_load = min(cpu_load, load_now);
1016
1017         if (running > 1 || (idle == NOT_IDLE && running))
1018                 /*
1019                  * If we are busy rebalancing the load is biased by
1020                  * priority to create 'nice' support across cpus. When
1021                  * idle rebalancing we should only bias the source_load if
1022                  * there is more than one task running on that queue to
1023                  * prevent idle rebalance from trying to pull tasks from a
1024                  * queue with only one running task.
1025                  */
1026                 source_load = source_load * rq->prio_bias / running;
1027
1028         return source_load;
1029 }
1030
1031 static inline unsigned long source_load(int cpu, int type)
1032 {
1033         return __source_load(cpu, type, NOT_IDLE);
1034 }
1035
1036 /*
1037  * Return a high guess at the load of a migration-target cpu
1038  */
1039 static inline unsigned long __target_load(int cpu, int type, enum idle_type idle)
1040 {
1041         runqueue_t *rq = cpu_rq(cpu);
1042         unsigned long running = rq->nr_running;
1043         unsigned long target_load, cpu_load = rq->cpu_load[type-1],
1044                 load_now = running * SCHED_LOAD_SCALE;
1045
1046         if (type == 0)
1047                 target_load = load_now;
1048         else
1049                 target_load = max(cpu_load, load_now);
1050
1051         if (running > 1 || (idle == NOT_IDLE && running))
1052                 target_load = target_load * rq->prio_bias / running;
1053
1054         return target_load;
1055 }
1056
1057 static inline unsigned long target_load(int cpu, int type)
1058 {
1059         return __target_load(cpu, type, NOT_IDLE);
1060 }
1061
1062 /*
1063  * find_idlest_group finds and returns the least busy CPU group within the
1064  * domain.
1065  */
1066 static struct sched_group *
1067 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1068 {
1069         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1070         unsigned long min_load = ULONG_MAX, this_load = 0;
1071         int load_idx = sd->forkexec_idx;
1072         int imbalance = 100 + (sd->imbalance_pct-100)/2;
1073
1074         do {
1075                 unsigned long load, avg_load;
1076                 int local_group;
1077                 int i;
1078
1079                 /* Skip over this group if it has no CPUs allowed */
1080                 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1081                         goto nextgroup;
1082
1083                 local_group = cpu_isset(this_cpu, group->cpumask);
1084
1085                 /* Tally up the load of all CPUs in the group */
1086                 avg_load = 0;
1087
1088                 for_each_cpu_mask(i, group->cpumask) {
1089                         /* Bias balancing toward cpus of our domain */
1090                         if (local_group)
1091                                 load = source_load(i, load_idx);
1092                         else
1093                                 load = target_load(i, load_idx);
1094
1095                         avg_load += load;
1096                 }
1097
1098                 /* Adjust by relative CPU power of the group */
1099                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1100
1101                 if (local_group) {
1102                         this_load = avg_load;
1103                         this = group;
1104                 } else if (avg_load < min_load) {
1105                         min_load = avg_load;
1106                         idlest = group;
1107                 }
1108 nextgroup:
1109                 group = group->next;
1110         } while (group != sd->groups);
1111
1112         if (!idlest || 100*this_load < imbalance*min_load)
1113                 return NULL;
1114         return idlest;
1115 }
1116
1117 /*
1118  * find_idlest_queue - find the idlest runqueue among the cpus in group.
1119  */
1120 static int
1121 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1122 {
1123         cpumask_t tmp;
1124         unsigned long load, min_load = ULONG_MAX;
1125         int idlest = -1;
1126         int i;
1127
1128         /* Traverse only the allowed CPUs */
1129         cpus_and(tmp, group->cpumask, p->cpus_allowed);
1130
1131         for_each_cpu_mask(i, tmp) {
1132                 load = source_load(i, 0);
1133
1134                 if (load < min_load || (load == min_load && i == this_cpu)) {
1135                         min_load = load;
1136                         idlest = i;
1137                 }
1138         }
1139
1140         return idlest;
1141 }
1142
1143 /*
1144  * sched_balance_self: balance the current task (running on cpu) in domains
1145  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1146  * SD_BALANCE_EXEC.
1147  *
1148  * Balance, ie. select the least loaded group.
1149  *
1150  * Returns the target CPU number, or the same CPU if no balancing is needed.
1151  *
1152  * preempt must be disabled.
1153  */
1154 static int sched_balance_self(int cpu, int flag)
1155 {
1156         struct task_struct *t = current;
1157         struct sched_domain *tmp, *sd = NULL;
1158
1159         for_each_domain(cpu, tmp)
1160                 if (tmp->flags & flag)
1161                         sd = tmp;
1162
1163         while (sd) {
1164                 cpumask_t span;
1165                 struct sched_group *group;
1166                 int new_cpu;
1167                 int weight;
1168
1169                 span = sd->span;
1170                 group = find_idlest_group(sd, t, cpu);
1171                 if (!group)
1172                         goto nextlevel;
1173
1174                 new_cpu = find_idlest_cpu(group, t, cpu);
1175                 if (new_cpu == -1 || new_cpu == cpu)
1176                         goto nextlevel;
1177
1178                 /* Now try balancing at a lower domain level */
1179                 cpu = new_cpu;
1180 nextlevel:
1181                 sd = NULL;
1182                 weight = cpus_weight(span);
1183                 for_each_domain(cpu, tmp) {
1184                         if (weight <= cpus_weight(tmp->span))
1185                                 break;
1186                         if (tmp->flags & flag)
1187                                 sd = tmp;
1188                 }
1189                 /* while loop will break here if sd == NULL */
1190         }
1191
1192         return cpu;
1193 }
1194
1195 #endif /* CONFIG_SMP */
1196
1197 /*
1198  * wake_idle() will wake a task on an idle cpu if task->cpu is
1199  * not idle and an idle cpu is available.  The span of cpus to
1200  * search starts with cpus closest then further out as needed,
1201  * so we always favor a closer, idle cpu.
1202  *
1203  * Returns the CPU we should wake onto.
1204  */
1205 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1206 static int wake_idle(int cpu, task_t *p)
1207 {
1208         cpumask_t tmp;
1209         struct sched_domain *sd;
1210         int i;
1211
1212         if (idle_cpu(cpu))
1213                 return cpu;
1214
1215         for_each_domain(cpu, sd) {
1216                 if (sd->flags & SD_WAKE_IDLE) {
1217                         cpus_and(tmp, sd->span, p->cpus_allowed);
1218                         for_each_cpu_mask(i, tmp) {
1219                                 if (idle_cpu(i))
1220                                         return i;
1221                         }
1222                 }
1223                 else
1224                         break;
1225         }
1226         return cpu;
1227 }
1228 #else
1229 static inline int wake_idle(int cpu, task_t *p)
1230 {
1231         return cpu;
1232 }
1233 #endif
1234
1235 /***
1236  * try_to_wake_up - wake up a thread
1237  * @p: the to-be-woken-up thread
1238  * @state: the mask of task states that can be woken
1239  * @sync: do a synchronous wakeup?
1240  *
1241  * Put it on the run-queue if it's not already there. The "current"
1242  * thread is always on the run-queue (except when the actual
1243  * re-schedule is in progress), and as such you're allowed to do
1244  * the simpler "current->state = TASK_RUNNING" to mark yourself
1245  * runnable without the overhead of this.
1246  *
1247  * returns failure only if the task is already active.
1248  */
1249 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1250 {
1251         int cpu, this_cpu, success = 0;
1252         unsigned long flags;
1253         long old_state;
1254         runqueue_t *rq;
1255 #ifdef CONFIG_SMP
1256         unsigned long load, this_load;
1257         struct sched_domain *sd, *this_sd = NULL;
1258         int new_cpu;
1259 #endif
1260
1261         rq = task_rq_lock(p, &flags);
1262         old_state = p->state;
1263         if (!(old_state & state))
1264                 goto out;
1265
1266         if (p->array)
1267                 goto out_running;
1268
1269         cpu = task_cpu(p);
1270         this_cpu = smp_processor_id();
1271
1272 #ifdef CONFIG_SMP
1273         if (unlikely(task_running(rq, p)))
1274                 goto out_activate;
1275
1276         new_cpu = cpu;
1277
1278         schedstat_inc(rq, ttwu_cnt);
1279         if (cpu == this_cpu) {
1280                 schedstat_inc(rq, ttwu_local);
1281                 goto out_set_cpu;
1282         }
1283
1284         for_each_domain(this_cpu, sd) {
1285                 if (cpu_isset(cpu, sd->span)) {
1286                         schedstat_inc(sd, ttwu_wake_remote);
1287                         this_sd = sd;
1288                         break;
1289                 }
1290         }
1291
1292         if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1293                 goto out_set_cpu;
1294
1295         /*
1296          * Check for affine wakeup and passive balancing possibilities.
1297          */
1298         if (this_sd) {
1299                 int idx = this_sd->wake_idx;
1300                 unsigned int imbalance;
1301
1302                 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1303
1304                 load = source_load(cpu, idx);
1305                 this_load = target_load(this_cpu, idx);
1306
1307                 new_cpu = this_cpu; /* Wake to this CPU if we can */
1308
1309                 if (this_sd->flags & SD_WAKE_AFFINE) {
1310                         unsigned long tl = this_load;
1311                         /*
1312                          * If sync wakeup then subtract the (maximum possible)
1313                          * effect of the currently running task from the load
1314                          * of the current CPU:
1315                          */
1316                         if (sync)
1317                                 tl -= SCHED_LOAD_SCALE;
1318
1319                         if ((tl <= load &&
1320                                 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1321                                 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1322                                 /*
1323                                  * This domain has SD_WAKE_AFFINE and
1324                                  * p is cache cold in this domain, and
1325                                  * there is no bad imbalance.
1326                                  */
1327                                 schedstat_inc(this_sd, ttwu_move_affine);
1328                                 goto out_set_cpu;
1329                         }
1330                 }
1331
1332                 /*
1333                  * Start passive balancing when half the imbalance_pct
1334                  * limit is reached.
1335                  */
1336                 if (this_sd->flags & SD_WAKE_BALANCE) {
1337                         if (imbalance*this_load <= 100*load) {
1338                                 schedstat_inc(this_sd, ttwu_move_balance);
1339                                 goto out_set_cpu;
1340                         }
1341                 }
1342         }
1343
1344         new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1345 out_set_cpu:
1346         new_cpu = wake_idle(new_cpu, p);
1347         if (new_cpu != cpu) {
1348                 set_task_cpu(p, new_cpu);
1349                 task_rq_unlock(rq, &flags);
1350                 /* might preempt at this point */
1351                 rq = task_rq_lock(p, &flags);
1352                 old_state = p->state;
1353                 if (!(old_state & state))
1354                         goto out;
1355                 if (p->array)
1356                         goto out_running;
1357
1358                 this_cpu = smp_processor_id();
1359                 cpu = task_cpu(p);
1360         }
1361
1362 out_activate:
1363 #endif /* CONFIG_SMP */
1364         if (old_state == TASK_UNINTERRUPTIBLE) {
1365                 rq->nr_uninterruptible--;
1366                 /*
1367                  * Tasks on involuntary sleep don't earn
1368                  * sleep_avg beyond just interactive state.
1369                  */
1370                 p->activated = -1;
1371         }
1372
1373         /*
1374          * Tasks that have marked their sleep as noninteractive get
1375          * woken up without updating their sleep average. (i.e. their
1376          * sleep is handled in a priority-neutral manner, no priority
1377          * boost and no penalty.)
1378          */
1379         if (old_state & TASK_NONINTERACTIVE)
1380                 __activate_task(p, rq);
1381         else
1382                 activate_task(p, rq, cpu == this_cpu);
1383         /*
1384          * Sync wakeups (i.e. those types of wakeups where the waker
1385          * has indicated that it will leave the CPU in short order)
1386          * don't trigger a preemption, if the woken up task will run on
1387          * this cpu. (in this case the 'I will reschedule' promise of
1388          * the waker guarantees that the freshly woken up task is going
1389          * to be considered on this CPU.)
1390          */
1391         if (!sync || cpu != this_cpu) {
1392                 if (TASK_PREEMPTS_CURR(p, rq))
1393                         resched_task(rq->curr);
1394         }
1395         success = 1;
1396
1397 out_running:
1398         p->state = TASK_RUNNING;
1399 out:
1400         task_rq_unlock(rq, &flags);
1401
1402         return success;
1403 }
1404
1405 int fastcall wake_up_process(task_t *p)
1406 {
1407         return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1408                                  TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1409 }
1410
1411 EXPORT_SYMBOL(wake_up_process);
1412
1413 int fastcall wake_up_state(task_t *p, unsigned int state)
1414 {
1415         return try_to_wake_up(p, state, 0);
1416 }
1417
1418 /*
1419  * Perform scheduler related setup for a newly forked process p.
1420  * p is forked by current.
1421  */
1422 void fastcall sched_fork(task_t *p, int clone_flags)
1423 {
1424         int cpu = get_cpu();
1425
1426 #ifdef CONFIG_SMP
1427         cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1428 #endif
1429         set_task_cpu(p, cpu);
1430
1431         /*
1432          * We mark the process as running here, but have not actually
1433          * inserted it onto the runqueue yet. This guarantees that
1434          * nobody will actually run it, and a signal or other external
1435          * event cannot wake it up and insert it on the runqueue either.
1436          */
1437         p->state = TASK_RUNNING;
1438         INIT_LIST_HEAD(&p->run_list);
1439         p->array = NULL;
1440 #ifdef CONFIG_SCHEDSTATS
1441         memset(&p->sched_info, 0, sizeof(p->sched_info));
1442 #endif
1443 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1444         p->oncpu = 0;
1445 #endif
1446 #ifdef CONFIG_PREEMPT
1447         /* Want to start with kernel preemption disabled. */
1448         task_thread_info(p)->preempt_count = 1;
1449 #endif
1450         /*
1451          * Share the timeslice between parent and child, thus the
1452          * total amount of pending timeslices in the system doesn't change,
1453          * resulting in more scheduling fairness.
1454          */
1455         local_irq_disable();
1456         p->time_slice = (current->time_slice + 1) >> 1;
1457         /*
1458          * The remainder of the first timeslice might be recovered by
1459          * the parent if the child exits early enough.
1460          */
1461         p->first_time_slice = 1;
1462         current->time_slice >>= 1;
1463         p->timestamp = sched_clock();
1464         if (unlikely(!current->time_slice)) {
1465                 /*
1466                  * This case is rare, it happens when the parent has only
1467                  * a single jiffy left from its timeslice. Taking the
1468                  * runqueue lock is not a problem.
1469                  */
1470                 current->time_slice = 1;
1471                 scheduler_tick();
1472         }
1473         local_irq_enable();
1474         put_cpu();
1475 }
1476
1477 /*
1478  * wake_up_new_task - wake up a newly created task for the first time.
1479  *
1480  * This function will do some initial scheduler statistics housekeeping
1481  * that must be done for every newly created context, then puts the task
1482  * on the runqueue and wakes it.
1483  */
1484 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1485 {
1486         unsigned long flags;
1487         int this_cpu, cpu;
1488         runqueue_t *rq, *this_rq;
1489
1490         rq = task_rq_lock(p, &flags);
1491         BUG_ON(p->state != TASK_RUNNING);
1492         this_cpu = smp_processor_id();
1493         cpu = task_cpu(p);
1494
1495         /*
1496          * We decrease the sleep average of forking parents
1497          * and children as well, to keep max-interactive tasks
1498          * from forking tasks that are max-interactive. The parent
1499          * (current) is done further down, under its lock.
1500          */
1501         p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1502                 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1503
1504         p->prio = effective_prio(p);
1505
1506         if (likely(cpu == this_cpu)) {
1507                 if (!(clone_flags & CLONE_VM)) {
1508                         /*
1509                          * The VM isn't cloned, so we're in a good position to
1510                          * do child-runs-first in anticipation of an exec. This
1511                          * usually avoids a lot of COW overhead.
1512                          */
1513                         if (unlikely(!current->array))
1514                                 __activate_task(p, rq);
1515                         else {
1516                                 p->prio = current->prio;
1517                                 list_add_tail(&p->run_list, &current->run_list);
1518                                 p->array = current->array;
1519                                 p->array->nr_active++;
1520                                 inc_nr_running(p, rq);
1521                         }
1522                         set_need_resched();
1523                 } else
1524                         /* Run child last */
1525                         __activate_task(p, rq);
1526                 /*
1527                  * We skip the following code due to cpu == this_cpu
1528                  *
1529                  *   task_rq_unlock(rq, &flags);
1530                  *   this_rq = task_rq_lock(current, &flags);
1531                  */
1532                 this_rq = rq;
1533         } else {
1534                 this_rq = cpu_rq(this_cpu);
1535
1536                 /*
1537                  * Not the local CPU - must adjust timestamp. This should
1538                  * get optimised away in the !CONFIG_SMP case.
1539                  */
1540                 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1541                                         + rq->timestamp_last_tick;
1542                 __activate_task(p, rq);
1543                 if (TASK_PREEMPTS_CURR(p, rq))
1544                         resched_task(rq->curr);
1545
1546                 /*
1547                  * Parent and child are on different CPUs, now get the
1548                  * parent runqueue to update the parent's ->sleep_avg:
1549                  */
1550                 task_rq_unlock(rq, &flags);
1551                 this_rq = task_rq_lock(current, &flags);
1552         }
1553         current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1554                 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1555         task_rq_unlock(this_rq, &flags);
1556 }
1557
1558 /*
1559  * Potentially available exiting-child timeslices are
1560  * retrieved here - this way the parent does not get
1561  * penalized for creating too many threads.
1562  *
1563  * (this cannot be used to 'generate' timeslices
1564  * artificially, because any timeslice recovered here
1565  * was given away by the parent in the first place.)
1566  */
1567 void fastcall sched_exit(task_t *p)
1568 {
1569         unsigned long flags;
1570         runqueue_t *rq;
1571
1572         /*
1573          * If the child was a (relative-) CPU hog then decrease
1574          * the sleep_avg of the parent as well.
1575          */
1576         rq = task_rq_lock(p->parent, &flags);
1577         if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1578                 p->parent->time_slice += p->time_slice;
1579                 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1580                         p->parent->time_slice = task_timeslice(p);
1581         }
1582         if (p->sleep_avg < p->parent->sleep_avg)
1583                 p->parent->sleep_avg = p->parent->sleep_avg /
1584                 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1585                 (EXIT_WEIGHT + 1);
1586         task_rq_unlock(rq, &flags);
1587 }
1588
1589 /**
1590  * prepare_task_switch - prepare to switch tasks
1591  * @rq: the runqueue preparing to switch
1592  * @next: the task we are going to switch to.
1593  *
1594  * This is called with the rq lock held and interrupts off. It must
1595  * be paired with a subsequent finish_task_switch after the context
1596  * switch.
1597  *
1598  * prepare_task_switch sets up locking and calls architecture specific
1599  * hooks.
1600  */
1601 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1602 {
1603         prepare_lock_switch(rq, next);
1604         prepare_arch_switch(next);
1605 }
1606
1607 /**
1608  * finish_task_switch - clean up after a task-switch
1609  * @rq: runqueue associated with task-switch
1610  * @prev: the thread we just switched away from.
1611  *
1612  * finish_task_switch must be called after the context switch, paired
1613  * with a prepare_task_switch call before the context switch.
1614  * finish_task_switch will reconcile locking set up by prepare_task_switch,
1615  * and do any other architecture-specific cleanup actions.
1616  *
1617  * Note that we may have delayed dropping an mm in context_switch(). If
1618  * so, we finish that here outside of the runqueue lock.  (Doing it
1619  * with the lock held can cause deadlocks; see schedule() for
1620  * details.)
1621  */
1622 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1623         __releases(rq->lock)
1624 {
1625         struct mm_struct *mm = rq->prev_mm;
1626         unsigned long prev_task_flags;
1627
1628         rq->prev_mm = NULL;
1629
1630         /*
1631          * A task struct has one reference for the use as "current".
1632          * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1633          * calls schedule one last time. The schedule call will never return,
1634          * and the scheduled task must drop that reference.
1635          * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1636          * still held, otherwise prev could be scheduled on another cpu, die
1637          * there before we look at prev->state, and then the reference would
1638          * be dropped twice.
1639          *              Manfred Spraul <manfred@colorfullife.com>
1640          */
1641         prev_task_flags = prev->flags;
1642         finish_arch_switch(prev);
1643         finish_lock_switch(rq, prev);
1644         if (mm)
1645                 mmdrop(mm);
1646         if (unlikely(prev_task_flags & PF_DEAD))
1647                 put_task_struct(prev);
1648 }
1649
1650 /**
1651  * schedule_tail - first thing a freshly forked thread must call.
1652  * @prev: the thread we just switched away from.
1653  */
1654 asmlinkage void schedule_tail(task_t *prev)
1655         __releases(rq->lock)
1656 {
1657         runqueue_t *rq = this_rq();
1658         finish_task_switch(rq, prev);
1659 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1660         /* In this case, finish_task_switch does not reenable preemption */
1661         preempt_enable();
1662 #endif
1663         if (current->set_child_tid)
1664                 put_user(current->pid, current->set_child_tid);
1665 }
1666
1667 /*
1668  * context_switch - switch to the new MM and the new
1669  * thread's register state.
1670  */
1671 static inline
1672 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1673 {
1674         struct mm_struct *mm = next->mm;
1675         struct mm_struct *oldmm = prev->active_mm;
1676
1677         if (unlikely(!mm)) {
1678                 next->active_mm = oldmm;
1679                 atomic_inc(&oldmm->mm_count);
1680                 enter_lazy_tlb(oldmm, next);
1681         } else
1682                 switch_mm(oldmm, mm, next);
1683
1684         if (unlikely(!prev->mm)) {
1685                 prev->active_mm = NULL;
1686                 WARN_ON(rq->prev_mm);
1687                 rq->prev_mm = oldmm;
1688         }
1689
1690         /* Here we just switch the register state and the stack. */
1691         switch_to(prev, next, prev);
1692
1693         return prev;
1694 }
1695
1696 /*
1697  * nr_running, nr_uninterruptible and nr_context_switches:
1698  *
1699  * externally visible scheduler statistics: current number of runnable
1700  * threads, current number of uninterruptible-sleeping threads, total
1701  * number of context switches performed since bootup.
1702  */
1703 unsigned long nr_running(void)
1704 {
1705         unsigned long i, sum = 0;
1706
1707         for_each_online_cpu(i)
1708                 sum += cpu_rq(i)->nr_running;
1709
1710         return sum;
1711 }
1712
1713 unsigned long nr_uninterruptible(void)
1714 {
1715         unsigned long i, sum = 0;
1716
1717         for_each_cpu(i)
1718                 sum += cpu_rq(i)->nr_uninterruptible;
1719
1720         /*
1721          * Since we read the counters lockless, it might be slightly
1722          * inaccurate. Do not allow it to go below zero though:
1723          */
1724         if (unlikely((long)sum < 0))
1725                 sum = 0;
1726
1727         return sum;
1728 }
1729
1730 unsigned long long nr_context_switches(void)
1731 {
1732         unsigned long long i, sum = 0;
1733
1734         for_each_cpu(i)
1735                 sum += cpu_rq(i)->nr_switches;
1736
1737         return sum;
1738 }
1739
1740 unsigned long nr_iowait(void)
1741 {
1742         unsigned long i, sum = 0;
1743
1744         for_each_cpu(i)
1745                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1746
1747         return sum;
1748 }
1749
1750 #ifdef CONFIG_SMP
1751
1752 /*
1753  * double_rq_lock - safely lock two runqueues
1754  *
1755  * Note this does not disable interrupts like task_rq_lock,
1756  * you need to do so manually before calling.
1757  */
1758 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1759         __acquires(rq1->lock)
1760         __acquires(rq2->lock)
1761 {
1762         if (rq1 == rq2) {
1763                 spin_lock(&rq1->lock);
1764                 __acquire(rq2->lock);   /* Fake it out ;) */
1765         } else {
1766                 if (rq1 < rq2) {
1767                         spin_lock(&rq1->lock);
1768                         spin_lock(&rq2->lock);
1769                 } else {
1770                         spin_lock(&rq2->lock);
1771                         spin_lock(&rq1->lock);
1772                 }
1773         }
1774 }
1775
1776 /*
1777  * double_rq_unlock - safely unlock two runqueues
1778  *
1779  * Note this does not restore interrupts like task_rq_unlock,
1780  * you need to do so manually after calling.
1781  */
1782 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1783         __releases(rq1->lock)
1784         __releases(rq2->lock)
1785 {
1786         spin_unlock(&rq1->lock);
1787         if (rq1 != rq2)
1788                 spin_unlock(&rq2->lock);
1789         else
1790                 __release(rq2->lock);
1791 }
1792
1793 /*
1794  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1795  */
1796 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1797         __releases(this_rq->lock)
1798         __acquires(busiest->lock)
1799         __acquires(this_rq->lock)
1800 {
1801         if (unlikely(!spin_trylock(&busiest->lock))) {
1802                 if (busiest < this_rq) {
1803                         spin_unlock(&this_rq->lock);
1804                         spin_lock(&busiest->lock);
1805                         spin_lock(&this_rq->lock);
1806                 } else
1807                         spin_lock(&busiest->lock);
1808         }
1809 }
1810
1811 /*
1812  * If dest_cpu is allowed for this process, migrate the task to it.
1813  * This is accomplished by forcing the cpu_allowed mask to only
1814  * allow dest_cpu, which will force the cpu onto dest_cpu.  Then
1815  * the cpu_allowed mask is restored.
1816  */
1817 static void sched_migrate_task(task_t *p, int dest_cpu)
1818 {
1819         migration_req_t req;
1820         runqueue_t *rq;
1821         unsigned long flags;
1822
1823         rq = task_rq_lock(p, &flags);
1824         if (!cpu_isset(dest_cpu, p->cpus_allowed)
1825             || unlikely(cpu_is_offline(dest_cpu)))
1826                 goto out;
1827
1828         /* force the process onto the specified CPU */
1829         if (migrate_task(p, dest_cpu, &req)) {
1830                 /* Need to wait for migration thread (might exit: take ref). */
1831                 struct task_struct *mt = rq->migration_thread;
1832                 get_task_struct(mt);
1833                 task_rq_unlock(rq, &flags);
1834                 wake_up_process(mt);
1835                 put_task_struct(mt);
1836                 wait_for_completion(&req.done);
1837                 return;
1838         }
1839 out:
1840         task_rq_unlock(rq, &flags);
1841 }
1842
1843 /*
1844  * sched_exec - execve() is a valuable balancing opportunity, because at
1845  * this point the task has the smallest effective memory and cache footprint.
1846  */
1847 void sched_exec(void)
1848 {
1849         int new_cpu, this_cpu = get_cpu();
1850         new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1851         put_cpu();
1852         if (new_cpu != this_cpu)
1853                 sched_migrate_task(current, new_cpu);
1854 }
1855
1856 /*
1857  * pull_task - move a task from a remote runqueue to the local runqueue.
1858  * Both runqueues must be locked.
1859  */
1860 static inline
1861 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1862                runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1863 {
1864         dequeue_task(p, src_array);
1865         dec_nr_running(p, src_rq);
1866         set_task_cpu(p, this_cpu);
1867         inc_nr_running(p, this_rq);
1868         enqueue_task(p, this_array);
1869         p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1870                                 + this_rq->timestamp_last_tick;
1871         /*
1872          * Note that idle threads have a prio of MAX_PRIO, for this test
1873          * to be always true for them.
1874          */
1875         if (TASK_PREEMPTS_CURR(p, this_rq))
1876                 resched_task(this_rq->curr);
1877 }
1878
1879 /*
1880  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1881  */
1882 static inline
1883 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1884                      struct sched_domain *sd, enum idle_type idle,
1885                      int *all_pinned)
1886 {
1887         /*
1888          * We do not migrate tasks that are:
1889          * 1) running (obviously), or
1890          * 2) cannot be migrated to this CPU due to cpus_allowed, or
1891          * 3) are cache-hot on their current CPU.
1892          */
1893         if (!cpu_isset(this_cpu, p->cpus_allowed))
1894                 return 0;
1895         *all_pinned = 0;
1896
1897         if (task_running(rq, p))
1898                 return 0;
1899
1900         /*
1901          * Aggressive migration if:
1902          * 1) task is cache cold, or
1903          * 2) too many balance attempts have failed.
1904          */
1905
1906         if (sd->nr_balance_failed > sd->cache_nice_tries)
1907                 return 1;
1908
1909         if (task_hot(p, rq->timestamp_last_tick, sd))
1910                 return 0;
1911         return 1;
1912 }
1913
1914 /*
1915  * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1916  * as part of a balancing operation within "domain". Returns the number of
1917  * tasks moved.
1918  *
1919  * Called with both runqueues locked.
1920  */
1921 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1922                       unsigned long max_nr_move, struct sched_domain *sd,
1923                       enum idle_type idle, int *all_pinned)
1924 {
1925         prio_array_t *array, *dst_array;
1926         struct list_head *head, *curr;
1927         int idx, pulled = 0, pinned = 0;
1928         task_t *tmp;
1929
1930         if (max_nr_move == 0)
1931                 goto out;
1932
1933         pinned = 1;
1934
1935         /*
1936          * We first consider expired tasks. Those will likely not be
1937          * executed in the near future, and they are most likely to
1938          * be cache-cold, thus switching CPUs has the least effect
1939          * on them.
1940          */
1941         if (busiest->expired->nr_active) {
1942                 array = busiest->expired;
1943                 dst_array = this_rq->expired;
1944         } else {
1945                 array = busiest->active;
1946                 dst_array = this_rq->active;
1947         }
1948
1949 new_array:
1950         /* Start searching at priority 0: */
1951         idx = 0;
1952 skip_bitmap:
1953         if (!idx)
1954                 idx = sched_find_first_bit(array->bitmap);
1955         else
1956                 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1957         if (idx >= MAX_PRIO) {
1958                 if (array == busiest->expired && busiest->active->nr_active) {
1959                         array = busiest->active;
1960                         dst_array = this_rq->active;
1961                         goto new_array;
1962                 }
1963                 goto out;
1964         }
1965
1966         head = array->queue + idx;
1967         curr = head->prev;
1968 skip_queue:
1969         tmp = list_entry(curr, task_t, run_list);
1970
1971         curr = curr->prev;
1972
1973         if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1974                 if (curr != head)
1975                         goto skip_queue;
1976                 idx++;
1977                 goto skip_bitmap;
1978         }
1979
1980 #ifdef CONFIG_SCHEDSTATS
1981         if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1982                 schedstat_inc(sd, lb_hot_gained[idle]);
1983 #endif
1984
1985         pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1986         pulled++;
1987
1988         /* We only want to steal up to the prescribed number of tasks. */
1989         if (pulled < max_nr_move) {
1990                 if (curr != head)
1991                         goto skip_queue;
1992                 idx++;
1993                 goto skip_bitmap;
1994         }
1995 out:
1996         /*
1997          * Right now, this is the only place pull_task() is called,
1998          * so we can safely collect pull_task() stats here rather than
1999          * inside pull_task().
2000          */
2001         schedstat_add(sd, lb_gained[idle], pulled);
2002
2003         if (all_pinned)
2004                 *all_pinned = pinned;
2005         return pulled;
2006 }
2007
2008 /*
2009  * find_busiest_group finds and returns the busiest CPU group within the
2010  * domain. It calculates and returns the number of tasks which should be
2011  * moved to restore balance via the imbalance parameter.
2012  */
2013 static struct sched_group *
2014 find_busiest_group(struct sched_domain *sd, int this_cpu,
2015                    unsigned long *imbalance, enum idle_type idle, int *sd_idle)
2016 {
2017         struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2018         unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2019         unsigned long max_pull;
2020         int load_idx;
2021
2022         max_load = this_load = total_load = total_pwr = 0;
2023         if (idle == NOT_IDLE)
2024                 load_idx = sd->busy_idx;
2025         else if (idle == NEWLY_IDLE)
2026                 load_idx = sd->newidle_idx;
2027         else
2028                 load_idx = sd->idle_idx;
2029
2030         do {
2031                 unsigned long load;
2032                 int local_group;
2033                 int i;
2034
2035                 local_group = cpu_isset(this_cpu, group->cpumask);
2036
2037                 /* Tally up the load of all CPUs in the group */
2038                 avg_load = 0;
2039
2040                 for_each_cpu_mask(i, group->cpumask) {
2041                         if (*sd_idle && !idle_cpu(i))
2042                                 *sd_idle = 0;
2043
2044                         /* Bias balancing toward cpus of our domain */
2045                         if (local_group)
2046                                 load = __target_load(i, load_idx, idle);
2047                         else
2048                                 load = __source_load(i, load_idx, idle);
2049
2050                         avg_load += load;
2051                 }
2052
2053                 total_load += avg_load;
2054                 total_pwr += group->cpu_power;
2055
2056                 /* Adjust by relative CPU power of the group */
2057                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2058
2059                 if (local_group) {
2060                         this_load = avg_load;
2061                         this = group;
2062                 } else if (avg_load > max_load) {
2063                         max_load = avg_load;
2064                         busiest = group;
2065                 }
2066                 group = group->next;
2067         } while (group != sd->groups);
2068
2069         if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
2070                 goto out_balanced;
2071
2072         avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2073
2074         if (this_load >= avg_load ||
2075                         100*max_load <= sd->imbalance_pct*this_load)
2076                 goto out_balanced;
2077
2078         /*
2079          * We're trying to get all the cpus to the average_load, so we don't
2080          * want to push ourselves above the average load, nor do we wish to
2081          * reduce the max loaded cpu below the average load, as either of these
2082          * actions would just result in more rebalancing later, and ping-pong
2083          * tasks around. Thus we look for the minimum possible imbalance.
2084          * Negative imbalances (*we* are more loaded than anyone else) will
2085          * be counted as no imbalance for these purposes -- we can't fix that
2086          * by pulling tasks to us.  Be careful of negative numbers as they'll
2087          * appear as very large values with unsigned longs.
2088          */
2089
2090         /* Don't want to pull so many tasks that a group would go idle */
2091         max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
2092
2093         /* How much load to actually move to equalise the imbalance */
2094         *imbalance = min(max_pull * busiest->cpu_power,
2095                                 (avg_load - this_load) * this->cpu_power)
2096                         / SCHED_LOAD_SCALE;
2097
2098         if (*imbalance < SCHED_LOAD_SCALE) {
2099                 unsigned long pwr_now = 0, pwr_move = 0;
2100                 unsigned long tmp;
2101
2102                 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2103                         *imbalance = 1;
2104                         return busiest;
2105                 }
2106
2107                 /*
2108                  * OK, we don't have enough imbalance to justify moving tasks,
2109                  * however we may be able to increase total CPU power used by
2110                  * moving them.
2111                  */
2112
2113                 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2114                 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2115                 pwr_now /= SCHED_LOAD_SCALE;
2116
2117                 /* Amount of load we'd subtract */
2118                 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2119                 if (max_load > tmp)
2120                         pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2121                                                         max_load - tmp);
2122
2123                 /* Amount of load we'd add */
2124                 if (max_load*busiest->cpu_power <
2125                                 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2126                         tmp = max_load*busiest->cpu_power/this->cpu_power;
2127                 else
2128                         tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2129                 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2130                 pwr_move /= SCHED_LOAD_SCALE;
2131
2132                 /* Move if we gain throughput */
2133                 if (pwr_move <= pwr_now)
2134                         goto out_balanced;
2135
2136                 *imbalance = 1;
2137                 return busiest;
2138         }
2139
2140         /* Get rid of the scaling factor, rounding down as we divide */
2141         *imbalance = *imbalance / SCHED_LOAD_SCALE;
2142         return busiest;
2143
2144 out_balanced:
2145
2146         *imbalance = 0;
2147         return NULL;
2148 }
2149
2150 /*
2151  * find_busiest_queue - find the busiest runqueue among the cpus in group.
2152  */
2153 static runqueue_t *find_busiest_queue(struct sched_group *group,
2154         enum idle_type idle)
2155 {
2156         unsigned long load, max_load = 0;
2157         runqueue_t *busiest = NULL;
2158         int i;
2159
2160         for_each_cpu_mask(i, group->cpumask) {
2161                 load = __source_load(i, 0, idle);
2162
2163                 if (load > max_load) {
2164                         max_load = load;
2165                         busiest = cpu_rq(i);
2166                 }
2167         }
2168
2169         return busiest;
2170 }
2171
2172 /*
2173  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2174  * so long as it is large enough.
2175  */
2176 #define MAX_PINNED_INTERVAL     512
2177
2178 /*
2179  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2180  * tasks if there is an imbalance.
2181  *
2182  * Called with this_rq unlocked.
2183  */
2184 static int load_balance(int this_cpu, runqueue_t *this_rq,
2185                         struct sched_domain *sd, enum idle_type idle)
2186 {
2187         struct sched_group *group;
2188         runqueue_t *busiest;
2189         unsigned long imbalance;
2190         int nr_moved, all_pinned = 0;
2191         int active_balance = 0;
2192         int sd_idle = 0;
2193
2194         if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2195                 sd_idle = 1;
2196
2197         schedstat_inc(sd, lb_cnt[idle]);
2198
2199         group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2200         if (!group) {
2201                 schedstat_inc(sd, lb_nobusyg[idle]);
2202                 goto out_balanced;
2203         }
2204
2205         busiest = find_busiest_queue(group, idle);
2206         if (!busiest) {
2207                 schedstat_inc(sd, lb_nobusyq[idle]);
2208                 goto out_balanced;
2209         }
2210
2211         BUG_ON(busiest == this_rq);
2212
2213         schedstat_add(sd, lb_imbalance[idle], imbalance);
2214
2215         nr_moved = 0;
2216         if (busiest->nr_running > 1) {
2217                 /*
2218                  * Attempt to move tasks. If find_busiest_group has found
2219                  * an imbalance but busiest->nr_running <= 1, the group is
2220                  * still unbalanced. nr_moved simply stays zero, so it is
2221                  * correctly treated as an imbalance.
2222                  */
2223                 double_rq_lock(this_rq, busiest);
2224                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2225                                         imbalance, sd, idle, &all_pinned);
2226                 double_rq_unlock(this_rq, busiest);
2227
2228                 /* All tasks on this runqueue were pinned by CPU affinity */
2229                 if (unlikely(all_pinned))
2230                         goto out_balanced;
2231         }
2232
2233         if (!nr_moved) {
2234                 schedstat_inc(sd, lb_failed[idle]);
2235                 sd->nr_balance_failed++;
2236
2237                 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2238
2239                         spin_lock(&busiest->lock);
2240
2241                         /* don't kick the migration_thread, if the curr
2242                          * task on busiest cpu can't be moved to this_cpu
2243                          */
2244                         if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2245                                 spin_unlock(&busiest->lock);
2246                                 all_pinned = 1;
2247                                 goto out_one_pinned;
2248                         }
2249
2250                         if (!busiest->active_balance) {
2251                                 busiest->active_balance = 1;
2252                                 busiest->push_cpu = this_cpu;
2253                                 active_balance = 1;
2254                         }
2255                         spin_unlock(&busiest->lock);
2256                         if (active_balance)
2257                                 wake_up_process(busiest->migration_thread);
2258
2259                         /*
2260                          * We've kicked active balancing, reset the failure
2261                          * counter.
2262                          */
2263                         sd->nr_balance_failed = sd->cache_nice_tries+1;
2264                 }
2265         } else
2266                 sd->nr_balance_failed = 0;
2267
2268         if (likely(!active_balance)) {
2269                 /* We were unbalanced, so reset the balancing interval */
2270                 sd->balance_interval = sd->min_interval;
2271         } else {
2272                 /*
2273                  * If we've begun active balancing, start to back off. This
2274                  * case may not be covered by the all_pinned logic if there
2275                  * is only 1 task on the busy runqueue (because we don't call
2276                  * move_tasks).
2277                  */
2278                 if (sd->balance_interval < sd->max_interval)
2279                         sd->balance_interval *= 2;
2280         }
2281
2282         if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2283                 return -1;
2284         return nr_moved;
2285
2286 out_balanced:
2287         schedstat_inc(sd, lb_balanced[idle]);
2288
2289         sd->nr_balance_failed = 0;
2290
2291 out_one_pinned:
2292         /* tune up the balancing interval */
2293         if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2294                         (sd->balance_interval < sd->max_interval))
2295                 sd->balance_interval *= 2;
2296
2297         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2298                 return -1;
2299         return 0;
2300 }
2301
2302 /*
2303  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2304  * tasks if there is an imbalance.
2305  *
2306  * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2307  * this_rq is locked.
2308  */
2309 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2310                                 struct sched_domain *sd)
2311 {
2312         struct sched_group *group;
2313         runqueue_t *busiest = NULL;
2314         unsigned long imbalance;
2315         int nr_moved = 0;
2316         int sd_idle = 0;
2317
2318         if (sd->flags & SD_SHARE_CPUPOWER)
2319                 sd_idle = 1;
2320
2321         schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2322         group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2323         if (!group) {
2324                 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2325                 goto out_balanced;
2326         }
2327
2328         busiest = find_busiest_queue(group, NEWLY_IDLE);
2329         if (!busiest) {
2330                 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2331                 goto out_balanced;
2332         }
2333
2334         BUG_ON(busiest == this_rq);
2335
2336         schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2337
2338         nr_moved = 0;
2339         if (busiest->nr_running > 1) {
2340                 /* Attempt to move tasks */
2341                 double_lock_balance(this_rq, busiest);
2342                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2343                                         imbalance, sd, NEWLY_IDLE, NULL);
2344                 spin_unlock(&busiest->lock);
2345         }
2346
2347         if (!nr_moved) {
2348                 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2349                 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2350                         return -1;
2351         } else
2352                 sd->nr_balance_failed = 0;
2353
2354         return nr_moved;
2355
2356 out_balanced:
2357         schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2358         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2359                 return -1;
2360         sd->nr_balance_failed = 0;
2361         return 0;
2362 }
2363
2364 /*
2365  * idle_balance is called by schedule() if this_cpu is about to become
2366  * idle. Attempts to pull tasks from other CPUs.
2367  */
2368 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2369 {
2370         struct sched_domain *sd;
2371
2372         for_each_domain(this_cpu, sd) {
2373                 if (sd->flags & SD_BALANCE_NEWIDLE) {
2374                         if (load_balance_newidle(this_cpu, this_rq, sd)) {
2375                                 /* We've pulled tasks over so stop searching */
2376                                 break;
2377                         }
2378                 }
2379         }
2380 }
2381
2382 /*
2383  * active_load_balance is run by migration threads. It pushes running tasks
2384  * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2385  * running on each physical CPU where possible, and avoids physical /
2386  * logical imbalances.
2387  *
2388  * Called with busiest_rq locked.
2389  */
2390 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2391 {
2392         struct sched_domain *sd;
2393         runqueue_t *target_rq;
2394         int target_cpu = busiest_rq->push_cpu;
2395
2396         if (busiest_rq->nr_running <= 1)
2397                 /* no task to move */
2398                 return;
2399
2400         target_rq = cpu_rq(target_cpu);
2401
2402         /*
2403          * This condition is "impossible", if it occurs
2404          * we need to fix it.  Originally reported by
2405          * Bjorn Helgaas on a 128-cpu setup.
2406          */
2407         BUG_ON(busiest_rq == target_rq);
2408
2409         /* move a task from busiest_rq to target_rq */
2410         double_lock_balance(busiest_rq, target_rq);
2411
2412         /* Search for an sd spanning us and the target CPU. */
2413         for_each_domain(target_cpu, sd)
2414                 if ((sd->flags & SD_LOAD_BALANCE) &&
2415                         cpu_isset(busiest_cpu, sd->span))
2416                                 break;
2417
2418         if (unlikely(sd == NULL))
2419                 goto out;
2420
2421         schedstat_inc(sd, alb_cnt);
2422
2423         if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2424                 schedstat_inc(sd, alb_pushed);
2425         else
2426                 schedstat_inc(sd, alb_failed);
2427 out:
2428         spin_unlock(&target_rq->lock);
2429 }
2430
2431 /*
2432  * rebalance_tick will get called every timer tick, on every CPU.
2433  *
2434  * It checks each scheduling domain to see if it is due to be balanced,
2435  * and initiates a balancing operation if so.
2436  *
2437  * Balancing parameters are set up in arch_init_sched_domains.
2438  */
2439
2440 /* Don't have all balancing operations going off at once */
2441 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2442
2443 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2444                            enum idle_type idle)
2445 {
2446         unsigned long old_load, this_load;
2447         unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2448         struct sched_domain *sd;
2449         int i;
2450
2451         this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2452         /* Update our load */
2453         for (i = 0; i < 3; i++) {
2454                 unsigned long new_load = this_load;
2455                 int scale = 1 << i;
2456                 old_load = this_rq->cpu_load[i];
2457                 /*
2458                  * Round up the averaging division if load is increasing. This
2459                  * prevents us from getting stuck on 9 if the load is 10, for
2460                  * example.
2461                  */
2462                 if (new_load > old_load)
2463                         new_load += scale-1;
2464                 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2465         }
2466
2467         for_each_domain(this_cpu, sd) {
2468                 unsigned long interval;
2469
2470                 if (!(sd->flags & SD_LOAD_BALANCE))
2471                         continue;
2472
2473                 interval = sd->balance_interval;
2474                 if (idle != SCHED_IDLE)
2475                         interval *= sd->busy_factor;
2476
2477                 /* scale ms to jiffies */
2478                 interval = msecs_to_jiffies(interval);
2479                 if (unlikely(!interval))
2480                         interval = 1;
2481
2482                 if (j - sd->last_balance >= interval) {
2483                         if (load_balance(this_cpu, this_rq, sd, idle)) {
2484                                 /*
2485                                  * We've pulled tasks over so either we're no
2486                                  * longer idle, or one of our SMT siblings is
2487                                  * not idle.
2488                                  */
2489                                 idle = NOT_IDLE;
2490                         }
2491                         sd->last_balance += interval;
2492                 }
2493         }
2494 }
2495 #else
2496 /*
2497  * on UP we do not need to balance between CPUs:
2498  */
2499 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2500 {
2501 }
2502 static inline void idle_balance(int cpu, runqueue_t *rq)
2503 {
2504 }
2505 #endif
2506
2507 static inline int wake_priority_sleeper(runqueue_t *rq)
2508 {
2509         int ret = 0;
2510 #ifdef CONFIG_SCHED_SMT
2511         spin_lock(&rq->lock);
2512         /*
2513          * If an SMT sibling task has been put to sleep for priority
2514          * reasons reschedule the idle task to see if it can now run.
2515          */
2516         if (rq->nr_running) {
2517                 resched_task(rq->idle);
2518                 ret = 1;
2519         }
2520         spin_unlock(&rq->lock);
2521 #endif
2522         return ret;
2523 }
2524
2525 DEFINE_PER_CPU(struct kernel_stat, kstat);
2526
2527 EXPORT_PER_CPU_SYMBOL(kstat);
2528
2529 /*
2530  * This is called on clock ticks and on context switches.
2531  * Bank in p->sched_time the ns elapsed since the last tick or switch.
2532  */
2533 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2534                                     unsigned long long now)
2535 {
2536         unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2537         p->sched_time += now - last;
2538 }
2539
2540 /*
2541  * Return current->sched_time plus any more ns on the sched_clock
2542  * that have not yet been banked.
2543  */
2544 unsigned long long current_sched_time(const task_t *tsk)
2545 {
2546         unsigned long long ns;
2547         unsigned long flags;
2548         local_irq_save(flags);
2549         ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2550         ns = tsk->sched_time + (sched_clock() - ns);
2551         local_irq_restore(flags);
2552         return ns;
2553 }
2554
2555 /*
2556  * We place interactive tasks back into the active array, if possible.
2557  *
2558  * To guarantee that this does not starve expired tasks we ignore the
2559  * interactivity of a task if the first expired task had to wait more
2560  * than a 'reasonable' amount of time. This deadline timeout is
2561  * load-dependent, as the frequency of array switched decreases with
2562  * increasing number of running tasks. We also ignore the interactivity
2563  * if a better static_prio task has expired:
2564  */
2565 #define EXPIRED_STARVING(rq) \
2566         ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2567                 (jiffies - (rq)->expired_timestamp >= \
2568                         STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2569                         ((rq)->curr->static_prio > (rq)->best_expired_prio))
2570
2571 /*
2572  * Account user cpu time to a process.
2573  * @p: the process that the cpu time gets accounted to
2574  * @hardirq_offset: the offset to subtract from hardirq_count()
2575  * @cputime: the cpu time spent in user space since the last update
2576  */
2577 void account_user_time(struct task_struct *p, cputime_t cputime)
2578 {
2579         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2580         cputime64_t tmp;
2581
2582         p->utime = cputime_add(p->utime, cputime);
2583
2584         /* Add user time to cpustat. */
2585         tmp = cputime_to_cputime64(cputime);
2586         if (TASK_NICE(p) > 0)
2587                 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2588         else
2589                 cpustat->user = cputime64_add(cpustat->user, tmp);
2590 }
2591
2592 /*
2593  * Account system cpu time to a process.
2594  * @p: the process that the cpu time gets accounted to
2595  * @hardirq_offset: the offset to subtract from hardirq_count()
2596  * @cputime: the cpu time spent in kernel space since the last update
2597  */
2598 void account_system_time(struct task_struct *p, int hardirq_offset,
2599                          cputime_t cputime)
2600 {
2601         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2602         runqueue_t *rq = this_rq();
2603         cputime64_t tmp;
2604
2605         p->stime = cputime_add(p->stime, cputime);
2606
2607         /* Add system time to cpustat. */
2608         tmp = cputime_to_cputime64(cputime);
2609         if (hardirq_count() - hardirq_offset)
2610                 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2611         else if (softirq_count())
2612                 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2613         else if (p != rq->idle)
2614                 cpustat->system = cputime64_add(cpustat->system, tmp);
2615         else if (atomic_read(&rq->nr_iowait) > 0)
2616                 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2617         else
2618                 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2619         /* Account for system time used */
2620         acct_update_integrals(p);
2621 }
2622
2623 /*
2624  * Account for involuntary wait time.
2625  * @p: the process from which the cpu time has been stolen
2626  * @steal: the cpu time spent in involuntary wait
2627  */
2628 void account_steal_time(struct task_struct *p, cputime_t steal)
2629 {
2630         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2631         cputime64_t tmp = cputime_to_cputime64(steal);
2632         runqueue_t *rq = this_rq();
2633
2634         if (p == rq->idle) {
2635                 p->stime = cputime_add(p->stime, steal);
2636                 if (atomic_read(&rq->nr_iowait) > 0)
2637                         cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2638                 else
2639                         cpustat->idle = cputime64_add(cpustat->idle, tmp);
2640         } else
2641                 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2642 }
2643
2644 /*
2645  * This function gets called by the timer code, with HZ frequency.
2646  * We call it with interrupts disabled.
2647  *
2648  * It also gets called by the fork code, when changing the parent's
2649  * timeslices.
2650  */
2651 void scheduler_tick(void)
2652 {
2653         int cpu = smp_processor_id();
2654         runqueue_t *rq = this_rq();
2655         task_t *p = current;
2656         unsigned long long now = sched_clock();
2657
2658         update_cpu_clock(p, rq, now);
2659
2660         rq->timestamp_last_tick = now;
2661
2662         if (p == rq->idle) {
2663                 if (wake_priority_sleeper(rq))
2664                         goto out;
2665                 rebalance_tick(cpu, rq, SCHED_IDLE);
2666                 return;
2667         }
2668
2669         /* Task might have expired already, but not scheduled off yet */
2670         if (p->array != rq->active) {
2671                 set_tsk_need_resched(p);
2672                 goto out;
2673         }
2674         spin_lock(&rq->lock);
2675         /*
2676          * The task was running during this tick - update the
2677          * time slice counter. Note: we do not update a thread's
2678          * priority until it either goes to sleep or uses up its
2679          * timeslice. This makes it possible for interactive tasks
2680          * to use up their timeslices at their highest priority levels.
2681          */
2682         if (rt_task(p)) {
2683                 /*
2684                  * RR tasks need a special form of timeslice management.
2685                  * FIFO tasks have no timeslices.
2686                  */
2687                 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2688                         p->time_slice = task_timeslice(p);
2689                         p->first_time_slice = 0;
2690                         set_tsk_need_resched(p);
2691
2692                         /* put it at the end of the queue: */
2693                         requeue_task(p, rq->active);
2694                 }
2695                 goto out_unlock;
2696         }
2697         if (!--p->time_slice) {
2698                 dequeue_task(p, rq->active);
2699                 set_tsk_need_resched(p);
2700                 p->prio = effective_prio(p);
2701                 p->time_slice = task_timeslice(p);
2702                 p->first_time_slice = 0;
2703
2704                 if (!rq->expired_timestamp)
2705                         rq->expired_timestamp = jiffies;
2706                 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2707                         enqueue_task(p, rq->expired);
2708                         if (p->static_prio < rq->best_expired_prio)
2709                                 rq->best_expired_prio = p->static_prio;
2710                 } else
2711                         enqueue_task(p, rq->active);
2712         } else {
2713                 /*
2714                  * Prevent a too long timeslice allowing a task to monopolize
2715                  * the CPU. We do this by splitting up the timeslice into
2716                  * smaller pieces.
2717                  *
2718                  * Note: this does not mean the task's timeslices expire or
2719                  * get lost in any way, they just might be preempted by
2720                  * another task of equal priority. (one with higher
2721                  * priority would have preempted this task already.) We
2722                  * requeue this task to the end of the list on this priority
2723                  * level, which is in essence a round-robin of tasks with
2724                  * equal priority.
2725                  *
2726                  * This only applies to tasks in the interactive
2727                  * delta range with at least TIMESLICE_GRANULARITY to requeue.
2728                  */
2729                 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2730                         p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2731                         (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2732                         (p->array == rq->active)) {
2733
2734                         requeue_task(p, rq->active);
2735                         set_tsk_need_resched(p);
2736                 }
2737         }
2738 out_unlock:
2739         spin_unlock(&rq->lock);
2740 out:
2741         rebalance_tick(cpu, rq, NOT_IDLE);
2742 }
2743
2744 #ifdef CONFIG_SCHED_SMT
2745 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2746 {
2747         /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2748         if (rq->curr == rq->idle && rq->nr_running)
2749                 resched_task(rq->idle);
2750 }
2751
2752 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2753 {
2754         struct sched_domain *tmp, *sd = NULL;
2755         cpumask_t sibling_map;
2756         int i;
2757
2758         for_each_domain(this_cpu, tmp)
2759                 if (tmp->flags & SD_SHARE_CPUPOWER)
2760                         sd = tmp;
2761
2762         if (!sd)
2763                 return;
2764
2765         /*
2766          * Unlock the current runqueue because we have to lock in
2767          * CPU order to avoid deadlocks. Caller knows that we might
2768          * unlock. We keep IRQs disabled.
2769          */
2770         spin_unlock(&this_rq->lock);
2771
2772         sibling_map = sd->span;
2773
2774         for_each_cpu_mask(i, sibling_map)
2775                 spin_lock(&cpu_rq(i)->lock);
2776         /*
2777          * We clear this CPU from the mask. This both simplifies the
2778          * inner loop and keps this_rq locked when we exit:
2779          */
2780         cpu_clear(this_cpu, sibling_map);
2781
2782         for_each_cpu_mask(i, sibling_map) {
2783                 runqueue_t *smt_rq = cpu_rq(i);
2784
2785                 wakeup_busy_runqueue(smt_rq);
2786         }
2787
2788         for_each_cpu_mask(i, sibling_map)
2789                 spin_unlock(&cpu_rq(i)->lock);
2790         /*
2791          * We exit with this_cpu's rq still held and IRQs
2792          * still disabled:
2793          */
2794 }
2795
2796 /*
2797  * number of 'lost' timeslices this task wont be able to fully
2798  * utilize, if another task runs on a sibling. This models the
2799  * slowdown effect of other tasks running on siblings:
2800  */
2801 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2802 {
2803         return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2804 }
2805
2806 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2807 {
2808         struct sched_domain *tmp, *sd = NULL;
2809         cpumask_t sibling_map;
2810         prio_array_t *array;
2811         int ret = 0, i;
2812         task_t *p;
2813
2814         for_each_domain(this_cpu, tmp)
2815                 if (tmp->flags & SD_SHARE_CPUPOWER)
2816                         sd = tmp;
2817
2818         if (!sd)
2819                 return 0;
2820
2821         /*
2822          * The same locking rules and details apply as for
2823          * wake_sleeping_dependent():
2824          */
2825         spin_unlock(&this_rq->lock);
2826         sibling_map = sd->span;
2827         for_each_cpu_mask(i, sibling_map)
2828                 spin_lock(&cpu_rq(i)->lock);
2829         cpu_clear(this_cpu, sibling_map);
2830
2831         /*
2832          * Establish next task to be run - it might have gone away because
2833          * we released the runqueue lock above:
2834          */
2835         if (!this_rq->nr_running)
2836                 goto out_unlock;
2837         array = this_rq->active;
2838         if (!array->nr_active)
2839                 array = this_rq->expired;
2840         BUG_ON(!array->nr_active);
2841
2842         p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2843                 task_t, run_list);
2844
2845         for_each_cpu_mask(i, sibling_map) {
2846                 runqueue_t *smt_rq = cpu_rq(i);
2847                 task_t *smt_curr = smt_rq->curr;
2848
2849                 /* Kernel threads do not participate in dependent sleeping */
2850                 if (!p->mm || !smt_curr->mm || rt_task(p))
2851                         goto check_smt_task;
2852
2853                 /*
2854                  * If a user task with lower static priority than the
2855                  * running task on the SMT sibling is trying to schedule,
2856                  * delay it till there is proportionately less timeslice
2857                  * left of the sibling task to prevent a lower priority
2858                  * task from using an unfair proportion of the
2859                  * physical cpu's resources. -ck
2860                  */
2861                 if (rt_task(smt_curr)) {
2862                         /*
2863                          * With real time tasks we run non-rt tasks only
2864                          * per_cpu_gain% of the time.
2865                          */
2866                         if ((jiffies % DEF_TIMESLICE) >
2867                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2868                                         ret = 1;
2869                 } else
2870                         if (smt_curr->static_prio < p->static_prio &&
2871                                 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2872                                 smt_slice(smt_curr, sd) > task_timeslice(p))
2873                                         ret = 1;
2874
2875 check_smt_task:
2876                 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2877                         rt_task(smt_curr))
2878                                 continue;
2879                 if (!p->mm) {
2880                         wakeup_busy_runqueue(smt_rq);
2881                         continue;
2882                 }
2883
2884                 /*
2885                  * Reschedule a lower priority task on the SMT sibling for
2886                  * it to be put to sleep, or wake it up if it has been put to
2887                  * sleep for priority reasons to see if it should run now.
2888                  */
2889                 if (rt_task(p)) {
2890                         if ((jiffies % DEF_TIMESLICE) >
2891                                 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2892                                         resched_task(smt_curr);
2893                 } else {
2894                         if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2895                                 smt_slice(p, sd) > task_timeslice(smt_curr))
2896                                         resched_task(smt_curr);
2897                         else
2898                                 wakeup_busy_runqueue(smt_rq);
2899                 }
2900         }
2901 out_unlock:
2902         for_each_cpu_mask(i, sibling_map)
2903                 spin_unlock(&cpu_rq(i)->lock);
2904         return ret;
2905 }
2906 #else
2907 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2908 {
2909 }
2910
2911 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2912 {
2913         return 0;
2914 }
2915 #endif
2916
2917 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2918
2919 void fastcall add_preempt_count(int val)
2920 {
2921         /*
2922          * Underflow?
2923          */
2924         BUG_ON((preempt_count() < 0));
2925         preempt_count() += val;
2926         /*
2927          * Spinlock count overflowing soon?
2928          */
2929         BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2930 }
2931 EXPORT_SYMBOL(add_preempt_count);
2932
2933 void fastcall sub_preempt_count(int val)
2934 {
2935         /*
2936          * Underflow?
2937          */
2938         BUG_ON(val > preempt_count());
2939         /*
2940          * Is the spinlock portion underflowing?
2941          */
2942         BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2943         preempt_count() -= val;
2944 }
2945 EXPORT_SYMBOL(sub_preempt_count);
2946
2947 #endif
2948
2949 /*
2950  * schedule() is the main scheduler function.
2951  */
2952 asmlinkage void __sched schedule(void)
2953 {
2954         long *switch_count;
2955         task_t *prev, *next;
2956         runqueue_t *rq;
2957         prio_array_t *array;
2958         struct list_head *queue;
2959         unsigned long long now;
2960         unsigned long run_time;
2961         int cpu, idx, new_prio;
2962
2963         /*
2964          * Test if we are atomic.  Since do_exit() needs to call into
2965          * schedule() atomically, we ignore that path for now.
2966          * Otherwise, whine if we are scheduling when we should not be.
2967          */
2968         if (likely(!current->exit_state)) {
2969                 if (unlikely(in_atomic())) {
2970                         printk(KERN_ERR "scheduling while atomic: "
2971                                 "%s/0x%08x/%d\n",
2972                                 current->comm, preempt_count(), current->pid);
2973                         dump_stack();
2974                 }
2975         }
2976         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2977
2978 need_resched:
2979         preempt_disable();
2980         prev = current;
2981         release_kernel_lock(prev);
2982 need_resched_nonpreemptible:
2983         rq = this_rq();
2984
2985         /*
2986          * The idle thread is not allowed to schedule!
2987          * Remove this check after it has been exercised a bit.
2988          */
2989         if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2990                 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2991                 dump_stack();
2992         }
2993
2994         schedstat_inc(rq, sched_cnt);
2995         now = sched_clock();
2996         if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2997                 run_time = now - prev->timestamp;
2998                 if (unlikely((long long)(now - prev->timestamp) < 0))
2999                         run_time = 0;
3000         } else
3001                 run_time = NS_MAX_SLEEP_AVG;
3002
3003         /*
3004          * Tasks charged proportionately less run_time at high sleep_avg to
3005          * delay them losing their interactive status
3006          */
3007         run_time /= (CURRENT_BONUS(prev) ? : 1);
3008
3009         spin_lock_irq(&rq->lock);
3010
3011         if (unlikely(prev->flags & PF_DEAD))
3012                 prev->state = EXIT_DEAD;
3013
3014         switch_count = &prev->nivcsw;
3015         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3016                 switch_count = &prev->nvcsw;
3017                 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3018                                 unlikely(signal_pending(prev))))
3019                         prev->state = TASK_RUNNING;
3020                 else {
3021                         if (prev->state == TASK_UNINTERRUPTIBLE)
3022                                 rq->nr_uninterruptible++;
3023                         deactivate_task(prev, rq);
3024                 }
3025         }
3026
3027         cpu = smp_processor_id();
3028         if (unlikely(!rq->nr_running)) {
3029 go_idle:
3030                 idle_balance(cpu, rq);
3031                 if (!rq->nr_running) {
3032                         next = rq->idle;
3033                         rq->expired_timestamp = 0;
3034                         wake_sleeping_dependent(cpu, rq);
3035                         /*
3036                          * wake_sleeping_dependent() might have released
3037                          * the runqueue, so break out if we got new
3038                          * tasks meanwhile:
3039                          */
3040                         if (!rq->nr_running)
3041                                 goto switch_tasks;
3042                 }
3043         } else {
3044                 if (dependent_sleeper(cpu, rq)) {
3045                         next = rq->idle;
3046                         goto switch_tasks;
3047                 }
3048                 /*
3049                  * dependent_sleeper() releases and reacquires the runqueue
3050                  * lock, hence go into the idle loop if the rq went
3051                  * empty meanwhile:
3052                  */
3053                 if (unlikely(!rq->nr_running))
3054                         goto go_idle;
3055         }
3056
3057         array = rq->active;
3058         if (unlikely(!array->nr_active)) {
3059                 /*
3060                  * Switch the active and expired arrays.
3061                  */
3062                 schedstat_inc(rq, sched_switch);
3063                 rq->active = rq->expired;
3064                 rq->expired = array;
3065                 array = rq->active;
3066                 rq->expired_timestamp = 0;
3067                 rq->best_expired_prio = MAX_PRIO;
3068         }
3069
3070         idx = sched_find_first_bit(array->bitmap);
3071         queue = array->queue + idx;
3072         next = list_entry(queue->next, task_t, run_list);
3073
3074         if (!rt_task(next) && next->activated > 0) {
3075                 unsigned long long delta = now - next->timestamp;
3076                 if (unlikely((long long)(now - next->timestamp) < 0))
3077                         delta = 0;
3078
3079                 if (next->activated == 1)
3080                         delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3081
3082                 array = next->array;
3083                 new_prio = recalc_task_prio(next, next->timestamp + delta);
3084
3085                 if (unlikely(next->prio != new_prio)) {
3086                         dequeue_task(next, array);
3087                         next->prio = new_prio;
3088                         enqueue_task(next, array);
3089                 } else
3090                         requeue_task(next, array);
3091         }
3092         next->activated = 0;
3093 switch_tasks:
3094         if (next == rq->idle)
3095                 schedstat_inc(rq, sched_goidle);
3096         prefetch(next);
3097         prefetch_stack(next);
3098         clear_tsk_need_resched(prev);
3099         rcu_qsctr_inc(task_cpu(prev));
3100
3101         update_cpu_clock(prev, rq, now);
3102
3103         prev->sleep_avg -= run_time;
3104         if ((long)prev->sleep_avg <= 0)
3105                 prev->sleep_avg = 0;
3106         prev->timestamp = prev->last_ran = now;
3107
3108         sched_info_switch(prev, next);
3109         if (likely(prev != next)) {
3110                 next->timestamp = now;
3111                 rq->nr_switches++;
3112                 rq->curr = next;
3113                 ++*switch_count;
3114
3115                 prepare_task_switch(rq, next);
3116                 prev = context_switch(rq, prev, next);
3117                 barrier();
3118                 /*
3119                  * this_rq must be evaluated again because prev may have moved
3120                  * CPUs since it called schedule(), thus the 'rq' on its stack
3121                  * frame will be invalid.
3122                  */
3123                 finish_task_switch(this_rq(), prev);
3124         } else
3125                 spin_unlock_irq(&rq->lock);
3126
3127         prev = current;
3128         if (unlikely(reacquire_kernel_lock(prev) < 0))
3129                 goto need_resched_nonpreemptible;
3130         preempt_enable_no_resched();
3131         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3132                 goto need_resched;
3133 }
3134
3135 EXPORT_SYMBOL(schedule);
3136
3137 #ifdef CONFIG_PREEMPT
3138 /*
3139  * this is is the entry point to schedule() from in-kernel preemption
3140  * off of preempt_enable.  Kernel preemptions off return from interrupt
3141  * occur there and call schedule directly.
3142  */
3143 asmlinkage void __sched preempt_schedule(void)
3144 {
3145         struct thread_info *ti = current_thread_info();
3146 #ifdef CONFIG_PREEMPT_BKL
3147         struct task_struct *task = current;
3148         int saved_lock_depth;
3149 #endif
3150         /*
3151          * If there is a non-zero preempt_count or interrupts are disabled,
3152          * we do not want to preempt the current task.  Just return..
3153          */
3154         if (unlikely(ti->preempt_count || irqs_disabled()))
3155                 return;
3156
3157 need_resched:
3158         add_preempt_count(PREEMPT_ACTIVE);
3159         /*
3160          * We keep the big kernel semaphore locked, but we
3161          * clear ->lock_depth so that schedule() doesnt
3162          * auto-release the semaphore:
3163          */
3164 #ifdef CONFIG_PREEMPT_BKL
3165         saved_lock_depth = task->lock_depth;
3166         task->lock_depth = -1;
3167 #endif
3168         schedule();
3169 #ifdef CONFIG_PREEMPT_BKL
3170         task->lock_depth = saved_lock_depth;
3171 #endif
3172         sub_preempt_count(PREEMPT_ACTIVE);
3173
3174         /* we could miss a preemption opportunity between schedule and now */
3175         barrier();
3176         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3177                 goto need_resched;
3178 }
3179
3180 EXPORT_SYMBOL(preempt_schedule);
3181
3182 /*
3183  * this is is the entry point to schedule() from kernel preemption
3184  * off of irq context.
3185  * Note, that this is called and return with irqs disabled. This will
3186  * protect us against recursive calling from irq.
3187  */
3188 asmlinkage void __sched preempt_schedule_irq(void)
3189 {
3190         struct thread_info *ti = current_thread_info();
3191 #ifdef CONFIG_PREEMPT_BKL
3192         struct task_struct *task = current;
3193         int saved_lock_depth;
3194 #endif
3195         /* Catch callers which need to be fixed*/
3196         BUG_ON(ti->preempt_count || !irqs_disabled());
3197
3198 need_resched:
3199         add_preempt_count(PREEMPT_ACTIVE);
3200         /*
3201          * We keep the big kernel semaphore locked, but we
3202          * clear ->lock_depth so that schedule() doesnt
3203          * auto-release the semaphore:
3204          */
3205 #ifdef CONFIG_PREEMPT_BKL
3206         saved_lock_depth = task->lock_depth;
3207         task->lock_depth = -1;
3208 #endif
3209         local_irq_enable();
3210         schedule();
3211         local_irq_disable();
3212 #ifdef CONFIG_PREEMPT_BKL
3213         task->lock_depth = saved_lock_depth;
3214 #endif
3215         sub_preempt_count(PREEMPT_ACTIVE);
3216
3217         /* we could miss a preemption opportunity between schedule and now */
3218         barrier();
3219         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3220                 goto need_resched;
3221 }
3222
3223 #endif /* CONFIG_PREEMPT */
3224
3225 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3226                           void *key)
3227 {
3228         task_t *p = curr->private;
3229         return try_to_wake_up(p, mode, sync);
3230 }
3231
3232 EXPORT_SYMBOL(default_wake_function);
3233
3234 /*
3235  * The core wakeup function.  Non-exclusive wakeups (nr_exclusive == 0) just
3236  * wake everything up.  If it's an exclusive wakeup (nr_exclusive == small +ve
3237  * number) then we wake all the non-exclusive tasks and one exclusive task.
3238  *
3239  * There are circumstances in which we can try to wake a task which has already
3240  * started to run but is not in state TASK_RUNNING.  try_to_wake_up() returns
3241  * zero in this (rare) case, and we handle it by continuing to scan the queue.
3242  */
3243 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3244                              int nr_exclusive, int sync, void *key)
3245 {
3246         struct list_head *tmp, *next;
3247
3248         list_for_each_safe(tmp, next, &q->task_list) {
3249                 wait_queue_t *curr;
3250                 unsigned flags;
3251                 curr = list_entry(tmp, wait_queue_t, task_list);
3252                 flags = curr->flags;
3253                 if (curr->func(curr, mode, sync, key) &&
3254                     (flags & WQ_FLAG_EXCLUSIVE) &&
3255                     !--nr_exclusive)
3256                         break;
3257         }
3258 }
3259
3260 /**
3261  * __wake_up - wake up threads blocked on a waitqueue.
3262  * @q: the waitqueue
3263  * @mode: which threads
3264  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3265  * @key: is directly passed to the wakeup function
3266  */
3267 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3268                         int nr_exclusive, void *key)
3269 {
3270         unsigned long flags;
3271
3272         spin_lock_irqsave(&q->lock, flags);
3273         __wake_up_common(q, mode, nr_exclusive, 0, key);
3274         spin_unlock_irqrestore(&q->lock, flags);
3275 }
3276
3277 EXPORT_SYMBOL(__wake_up);
3278
3279 /*
3280  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3281  */
3282 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3283 {
3284         __wake_up_common(q, mode, 1, 0, NULL);
3285 }
3286
3287 /**
3288  * __wake_up_sync - wake up threads blocked on a waitqueue.
3289  * @q: the waitqueue
3290  * @mode: which threads
3291  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3292  *
3293  * The sync wakeup differs that the waker knows that it will schedule
3294  * away soon, so while the target thread will be woken up, it will not
3295  * be migrated to another CPU - ie. the two threads are 'synchronized'
3296  * with each other. This can prevent needless bouncing between CPUs.
3297  *
3298  * On UP it can prevent extra preemption.
3299  */
3300 void fastcall
3301 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3302 {
3303         unsigned long flags;
3304         int sync = 1;
3305
3306         if (unlikely(!q))
3307                 return;
3308
3309         if (unlikely(!nr_exclusive))
3310                 sync = 0;
3311
3312         spin_lock_irqsave(&q->lock, flags);
3313         __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3314         spin_unlock_irqrestore(&q->lock, flags);
3315 }
3316 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3317
3318 void fastcall complete(struct completion *x)
3319 {
3320         unsigned long flags;
3321
3322         spin_lock_irqsave(&x->wait.lock, flags);
3323         x->done++;
3324         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3325                          1, 0, NULL);
3326         spin_unlock_irqrestore(&x->wait.lock, flags);
3327 }
3328 EXPORT_SYMBOL(complete);
3329
3330 void fastcall complete_all(struct completion *x)
3331 {
3332         unsigned long flags;
3333
3334         spin_lock_irqsave(&x->wait.lock, flags);
3335         x->done += UINT_MAX/2;
3336         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3337                          0, 0, NULL);
3338         spin_unlock_irqrestore(&x->wait.lock, flags);
3339 }
3340 EXPORT_SYMBOL(complete_all);
3341
3342 void fastcall __sched wait_for_completion(struct completion *x)
3343 {
3344         might_sleep();
3345         spin_lock_irq(&x->wait.lock);
3346         if (!x->done) {
3347                 DECLARE_WAITQUEUE(wait, current);
3348
3349                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3350                 __add_wait_queue_tail(&x->wait, &wait);
3351                 do {
3352                         __set_current_state(TASK_UNINTERRUPTIBLE);
3353                         spin_unlock_irq(&x->wait.lock);
3354                         schedule();
3355                         spin_lock_irq(&x->wait.lock);
3356                 } while (!x->done);
3357                 __remove_wait_queue(&x->wait, &wait);
3358         }
3359         x->done--;
3360         spin_unlock_irq(&x->wait.lock);
3361 }
3362 EXPORT_SYMBOL(wait_for_completion);
3363
3364 unsigned long fastcall __sched
3365 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3366 {
3367         might_sleep();
3368
3369         spin_lock_irq(&x->wait.lock);
3370         if (!x->done) {
3371                 DECLARE_WAITQUEUE(wait, current);
3372
3373                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3374                 __add_wait_queue_tail(&x->wait, &wait);
3375                 do {
3376                         __set_current_state(TASK_UNINTERRUPTIBLE);
3377                         spin_unlock_irq(&x->wait.lock);
3378                         timeout = schedule_timeout(timeout);
3379                         spin_lock_irq(&x->wait.lock);
3380                         if (!timeout) {
3381                                 __remove_wait_queue(&x->wait, &wait);
3382                                 goto out;
3383                         }
3384                 } while (!x->done);
3385                 __remove_wait_queue(&x->wait, &wait);
3386         }
3387         x->done--;
3388 out:
3389         spin_unlock_irq(&x->wait.lock);
3390         return timeout;
3391 }
3392 EXPORT_SYMBOL(wait_for_completion_timeout);
3393
3394 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3395 {
3396         int ret = 0;
3397
3398         might_sleep();
3399
3400         spin_lock_irq(&x->wait.lock);
3401         if (!x->done) {
3402                 DECLARE_WAITQUEUE(wait, current);
3403
3404                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3405                 __add_wait_queue_tail(&x->wait, &wait);
3406                 do {
3407                         if (signal_pending(current)) {
3408                                 ret = -ERESTARTSYS;
3409                                 __remove_wait_queue(&x->wait, &wait);
3410                                 goto out;
3411                         }
3412                         __set_current_state(TASK_INTERRUPTIBLE);
3413                         spin_unlock_irq(&x->wait.lock);
3414                         schedule();
3415                         spin_lock_irq(&x->wait.lock);
3416                 } while (!x->done);
3417                 __remove_wait_queue(&x->wait, &wait);
3418         }
3419         x->done--;
3420 out:
3421         spin_unlock_irq(&x->wait.lock);
3422
3423         return ret;
3424 }
3425 EXPORT_SYMBOL(wait_for_completion_interruptible);
3426
3427 unsigned long fastcall __sched
3428 wait_for_completion_interruptible_timeout(struct completion *x,
3429                                           unsigned long timeout)
3430 {
3431         might_sleep();
3432
3433         spin_lock_irq(&x->wait.lock);
3434         if (!x->done) {
3435                 DECLARE_WAITQUEUE(wait, current);
3436
3437                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3438                 __add_wait_queue_tail(&x->wait, &wait);
3439                 do {
3440                         if (signal_pending(current)) {
3441                                 timeout = -ERESTARTSYS;
3442                                 __remove_wait_queue(&x->wait, &wait);
3443                                 goto out;
3444                         }
3445                         __set_current_state(TASK_INTERRUPTIBLE);
3446                         spin_unlock_irq(&x->wait.lock);
3447                         timeout = schedule_timeout(timeout);
3448                         spin_lock_irq(&x->wait.lock);
3449                         if (!timeout) {
3450                                 __remove_wait_queue(&x->wait, &wait);
3451                                 goto out;
3452                         }
3453                 } while (!x->done);
3454                 __remove_wait_queue(&x->wait, &wait);
3455         }
3456         x->done--;
3457 out:
3458         spin_unlock_irq(&x->wait.lock);
3459         return timeout;
3460 }
3461 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3462
3463
3464 #define SLEEP_ON_VAR                                    \
3465         unsigned long flags;                            \
3466         wait_queue_t wait;                              \
3467         init_waitqueue_entry(&wait, current);
3468
3469 #define SLEEP_ON_HEAD                                   \
3470         spin_lock_irqsave(&q->lock,flags);              \
3471         __add_wait_queue(q, &wait);                     \
3472         spin_unlock(&q->lock);
3473
3474 #define SLEEP_ON_TAIL                                   \
3475         spin_lock_irq(&q->lock);                        \
3476         __remove_wait_queue(q, &wait);                  \
3477         spin_unlock_irqrestore(&q->lock, flags);
3478
3479 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3480 {
3481         SLEEP_ON_VAR
3482
3483         current->state = TASK_INTERRUPTIBLE;
3484
3485         SLEEP_ON_HEAD
3486         schedule();
3487         SLEEP_ON_TAIL
3488 }
3489
3490 EXPORT_SYMBOL(interruptible_sleep_on);
3491
3492 long fastcall __sched
3493 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3494 {
3495         SLEEP_ON_VAR
3496
3497         current->state = TASK_INTERRUPTIBLE;
3498
3499         SLEEP_ON_HEAD
3500         timeout = schedule_timeout(timeout);
3501         SLEEP_ON_TAIL
3502
3503         return timeout;
3504 }
3505
3506 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3507
3508 void fastcall __sched sleep_on(wait_queue_head_t *q)
3509 {
3510         SLEEP_ON_VAR
3511
3512         current->state = TASK_UNINTERRUPTIBLE;
3513
3514         SLEEP_ON_HEAD
3515         schedule();
3516         SLEEP_ON_TAIL
3517 }
3518
3519 EXPORT_SYMBOL(sleep_on);
3520
3521 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3522 {
3523         SLEEP_ON_VAR
3524
3525         current->state = TASK_UNINTERRUPTIBLE;
3526
3527         SLEEP_ON_HEAD
3528         timeout = schedule_timeout(timeout);
3529         SLEEP_ON_TAIL
3530
3531         return timeout;
3532 }
3533
3534 EXPORT_SYMBOL(sleep_on_timeout);
3535
3536 void set_user_nice(task_t *p, long nice)
3537 {
3538         unsigned long flags;
3539         prio_array_t *array;
3540         runqueue_t *rq;
3541         int old_prio, new_prio, delta;
3542
3543         if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3544                 return;
3545         /*
3546          * We have to be careful, if called from sys_setpriority(),
3547          * the task might be in the middle of scheduling on another CPU.
3548          */
3549         rq = task_rq_lock(p, &flags);
3550         /*
3551          * The RT priorities are set via sched_setscheduler(), but we still
3552          * allow the 'normal' nice value to be set - but as expected
3553          * it wont have any effect on scheduling until the task is
3554          * not SCHED_NORMAL:
3555          */
3556         if (rt_task(p)) {
3557                 p->static_prio = NICE_TO_PRIO(nice);
3558                 goto out_unlock;
3559         }
3560         array = p->array;
3561         if (array) {
3562                 dequeue_task(p, array);
3563                 dec_prio_bias(rq, p->static_prio);
3564         }
3565
3566         old_prio = p->prio;
3567         new_prio = NICE_TO_PRIO(nice);
3568         delta = new_prio - old_prio;
3569         p->static_prio = NICE_TO_PRIO(nice);
3570         p->prio += delta;
3571
3572         if (array) {
3573                 enqueue_task(p, array);
3574                 inc_prio_bias(rq, p->static_prio);
3575                 /*
3576                  * If the task increased its priority or is running and
3577                  * lowered its priority, then reschedule its CPU:
3578                  */
3579                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3580                         resched_task(rq->curr);
3581         }
3582 out_unlock:
3583         task_rq_unlock(rq, &flags);
3584 }
3585
3586 EXPORT_SYMBOL(set_user_nice);
3587
3588 /*
3589  * can_nice - check if a task can reduce its nice value
3590  * @p: task
3591  * @nice: nice value
3592  */
3593 int can_nice(const task_t *p, const int nice)
3594 {
3595         /* convert nice value [19,-20] to rlimit style value [1,40] */
3596         int nice_rlim = 20 - nice;
3597         return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3598                 capable(CAP_SYS_NICE));
3599 }
3600
3601 #ifdef __ARCH_WANT_SYS_NICE
3602
3603 /*
3604  * sys_nice - change the priority of the current process.
3605  * @increment: priority increment
3606  *
3607  * sys_setpriority is a more generic, but much slower function that
3608  * does similar things.
3609  */
3610 asmlinkage long sys_nice(int increment)
3611 {
3612         int retval;
3613         long nice;
3614
3615         /*
3616          * Setpriority might change our priority at the same moment.
3617          * We don't have to worry. Conceptually one call occurs first
3618          * and we have a single winner.
3619          */
3620         if (increment < -40)
3621                 increment = -40;
3622         if (increment > 40)
3623                 increment = 40;
3624
3625         nice = PRIO_TO_NICE(current->static_prio) + increment;
3626         if (nice < -20)
3627                 nice = -20;
3628         if (nice > 19)
3629                 nice = 19;
3630
3631         if (increment < 0 && !can_nice(current, nice))
3632                 return -EPERM;
3633
3634         retval = security_task_setnice(current, nice);
3635         if (retval)
3636                 return retval;
3637
3638         set_user_nice(current, nice);
3639         return 0;
3640 }
3641
3642 #endif
3643
3644 /**
3645  * task_prio - return the priority value of a given task.
3646  * @p: the task in question.
3647  *
3648  * This is the priority value as seen by users in /proc.
3649  * RT tasks are offset by -200. Normal tasks are centered
3650  * around 0, value goes from -16 to +15.
3651  */
3652 int task_prio(const task_t *p)
3653 {
3654         return p->prio - MAX_RT_PRIO;
3655 }
3656
3657 /**
3658  * task_nice - return the nice value of a given task.
3659  * @p: the task in question.
3660  */
3661 int task_nice(const task_t *p)
3662 {
3663         return TASK_NICE(p);
3664 }
3665 EXPORT_SYMBOL_GPL(task_nice);
3666
3667 /**
3668  * idle_cpu - is a given cpu idle currently?
3669  * @cpu: the processor in question.
3670  */
3671 int idle_cpu(int cpu)
3672 {
3673         return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3674 }
3675
3676 /**
3677  * idle_task - return the idle task for a given cpu.
3678  * @cpu: the processor in question.
3679  */
3680 task_t *idle_task(int cpu)
3681 {
3682         return cpu_rq(cpu)->idle;
3683 }
3684
3685 /**
3686  * find_process_by_pid - find a process with a matching PID value.
3687  * @pid: the pid in question.
3688  */
3689 static inline task_t *find_process_by_pid(pid_t pid)
3690 {
3691         return pid ? find_task_by_pid(pid) : current;
3692 }
3693
3694 /* Actually do priority change: must hold rq lock. */
3695 static void __setscheduler(struct task_struct *p, int policy, int prio)
3696 {
3697         BUG_ON(p->array);
3698         p->policy = policy;
3699         p->rt_priority = prio;
3700         if (policy != SCHED_NORMAL)
3701                 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3702         else
3703                 p->prio = p->static_prio;
3704 }
3705
3706 /**
3707  * sched_setscheduler - change the scheduling policy and/or RT priority of
3708  * a thread.
3709  * @p: the task in question.
3710  * @policy: new policy.
3711  * @param: structure containing the new RT priority.
3712  */
3713 int sched_setscheduler(struct task_struct *p, int policy,
3714                        struct sched_param *param)
3715 {
3716         int retval;
3717         int oldprio, oldpolicy = -1;
3718         prio_array_t *array;
3719         unsigned long flags;
3720         runqueue_t *rq;
3721
3722 recheck:
3723         /* double check policy once rq lock held */
3724         if (policy < 0)
3725                 policy = oldpolicy = p->policy;
3726         else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3727                                 policy != SCHED_NORMAL)
3728                         return -EINVAL;
3729         /*
3730          * Valid priorities for SCHED_FIFO and SCHED_RR are
3731          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3732          */
3733         if (param->sched_priority < 0 ||
3734             (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3735             (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3736                 return -EINVAL;
3737         if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3738                 return -EINVAL;
3739
3740         /*
3741          * Allow unprivileged RT tasks to decrease priority:
3742          */
3743         if (!capable(CAP_SYS_NICE)) {
3744                 /* can't change policy */
3745                 if (policy != p->policy &&
3746                         !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3747                         return -EPERM;
3748                 /* can't increase priority */
3749                 if (policy != SCHED_NORMAL &&
3750                     param->sched_priority > p->rt_priority &&
3751                     param->sched_priority >
3752                                 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3753                         return -EPERM;
3754                 /* can't change other user's priorities */
3755                 if ((current->euid != p->euid) &&
3756                     (current->euid != p->uid))
3757                         return -EPERM;
3758         }
3759
3760         retval = security_task_setscheduler(p, policy, param);
3761         if (retval)
3762                 return retval;
3763         /*
3764          * To be able to change p->policy safely, the apropriate
3765          * runqueue lock must be held.
3766          */
3767         rq = task_rq_lock(p, &flags);
3768         /* recheck policy now with rq lock held */
3769         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3770                 policy = oldpolicy = -1;
3771                 task_rq_unlock(rq, &flags);
3772                 goto recheck;
3773         }
3774         array = p->array;
3775         if (array)
3776                 deactivate_task(p, rq);
3777         oldprio = p->prio;
3778         __setscheduler(p, policy, param->sched_priority);
3779         if (array) {
3780                 __activate_task(p, rq);
3781                 /*
3782                  * Reschedule if we are currently running on this runqueue and
3783                  * our priority decreased, or if we are not currently running on
3784                  * this runqueue and our priority is higher than the current's
3785                  */
3786                 if (task_running(rq, p)) {
3787                         if (p->prio > oldprio)
3788                                 resched_task(rq->curr);
3789                 } else if (TASK_PREEMPTS_CURR(p, rq))
3790                         resched_task(rq->curr);
3791         }
3792         task_rq_unlock(rq, &flags);
3793         return 0;
3794 }
3795 EXPORT_SYMBOL_GPL(sched_setscheduler);
3796
3797 static int
3798 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3799 {
3800         int retval;
3801         struct sched_param lparam;
3802         struct task_struct *p;
3803
3804         if (!param || pid < 0)
3805                 return -EINVAL;
3806         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3807                 return -EFAULT;
3808         read_lock_irq(&tasklist_lock);
3809         p = find_process_by_pid(pid);
3810         if (!p) {
3811                 read_unlock_irq(&tasklist_lock);
3812                 return -ESRCH;
3813         }
3814         retval = sched_setscheduler(p, policy, &lparam);
3815         read_unlock_irq(&tasklist_lock);
3816         return retval;
3817 }
3818
3819 /**
3820  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3821  * @pid: the pid in question.
3822  * @policy: new policy.
3823  * @param: structure containing the new RT priority.
3824  */
3825 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3826                                        struct sched_param __user *param)
3827 {
3828         return do_sched_setscheduler(pid, policy, param);
3829 }
3830
3831 /**
3832  * sys_sched_setparam - set/change the RT priority of a thread
3833  * @pid: the pid in question.
3834  * @param: structure containing the new RT priority.
3835  */
3836 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3837 {
3838         return do_sched_setscheduler(pid, -1, param);
3839 }
3840
3841 /**
3842  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3843  * @pid: the pid in question.
3844  */
3845 asmlinkage long sys_sched_getscheduler(pid_t pid)
3846 {
3847         int retval = -EINVAL;
3848         task_t *p;
3849
3850         if (pid < 0)
3851                 goto out_nounlock;
3852
3853         retval = -ESRCH;
3854         read_lock(&tasklist_lock);
3855         p = find_process_by_pid(pid);
3856         if (p) {
3857                 retval = security_task_getscheduler(p);
3858                 if (!retval)
3859                         retval = p->policy;
3860         }
3861         read_unlock(&tasklist_lock);
3862
3863 out_nounlock:
3864         return retval;
3865 }
3866
3867 /**
3868  * sys_sched_getscheduler - get the RT priority of a thread
3869  * @pid: the pid in question.
3870  * @param: structure containing the RT priority.
3871  */
3872 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3873 {
3874         struct sched_param lp;
3875         int retval = -EINVAL;
3876         task_t *p;
3877
3878         if (!param || pid < 0)
3879                 goto out_nounlock;
3880
3881         read_lock(&tasklist_lock);
3882         p = find_process_by_pid(pid);
3883         retval = -ESRCH;
3884         if (!p)
3885                 goto out_unlock;
3886
3887         retval = security_task_getscheduler(p);
3888         if (retval)
3889                 goto out_unlock;
3890
3891         lp.sched_priority = p->rt_priority;
3892         read_unlock(&tasklist_lock);
3893
3894         /*
3895          * This one might sleep, we cannot do it with a spinlock held ...
3896          */
3897         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3898
3899 out_nounlock:
3900         return retval;
3901
3902 out_unlock:
3903         read_unlock(&tasklist_lock);
3904         return retval;
3905 }
3906
3907 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3908 {
3909         task_t *p;
3910         int retval;
3911         cpumask_t cpus_allowed;
3912
3913         lock_cpu_hotplug();
3914         read_lock(&tasklist_lock);
3915
3916         p = find_process_by_pid(pid);
3917         if (!p) {
3918                 read_unlock(&tasklist_lock);
3919                 unlock_cpu_hotplug();
3920                 return -ESRCH;
3921         }
3922
3923         /*
3924          * It is not safe to call set_cpus_allowed with the
3925          * tasklist_lock held.  We will bump the task_struct's
3926          * usage count and then drop tasklist_lock.
3927          */
3928         get_task_struct(p);
3929         read_unlock(&tasklist_lock);
3930
3931         retval = -EPERM;
3932         if ((current->euid != p->euid) && (current->euid != p->uid) &&
3933                         !capable(CAP_SYS_NICE))
3934                 goto out_unlock;
3935
3936         cpus_allowed = cpuset_cpus_allowed(p);
3937         cpus_and(new_mask, new_mask, cpus_allowed);
3938         retval = set_cpus_allowed(p, new_mask);
3939
3940 out_unlock:
3941         put_task_struct(p);
3942         unlock_cpu_hotplug();
3943         return retval;
3944 }
3945
3946 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3947                              cpumask_t *new_mask)
3948 {
3949         if (len < sizeof(cpumask_t)) {
3950                 memset(new_mask, 0, sizeof(cpumask_t));
3951         } else if (len > sizeof(cpumask_t)) {
3952                 len = sizeof(cpumask_t);
3953         }
3954         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3955 }
3956
3957 /**
3958  * sys_sched_setaffinity - set the cpu affinity of a process
3959  * @pid: pid of the process
3960  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3961  * @user_mask_ptr: user-space pointer to the new cpu mask
3962  */
3963 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3964                                       unsigned long __user *user_mask_ptr)
3965 {
3966         cpumask_t new_mask;
3967         int retval;
3968
3969         retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3970         if (retval)
3971                 return retval;
3972
3973         return sched_setaffinity(pid, new_mask);
3974 }
3975
3976 /*
3977  * Represents all cpu's present in the system
3978  * In systems capable of hotplug, this map could dynamically grow
3979  * as new cpu's are detected in the system via any platform specific
3980  * method, such as ACPI for e.g.
3981  */
3982
3983 cpumask_t cpu_present_map __read_mostly;
3984 EXPORT_SYMBOL(cpu_present_map);
3985
3986 #ifndef CONFIG_SMP
3987 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
3988 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
3989 #endif
3990
3991 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3992 {
3993         int retval;
3994         task_t *p;
3995
3996         lock_cpu_hotplug();
3997         read_lock(&tasklist_lock);
3998
3999         retval = -ESRCH;
4000         p = find_process_by_pid(pid);
4001         if (!p)
4002                 goto out_unlock;
4003
4004         retval = 0;
4005         cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
4006
4007 out_unlock:
4008         read_unlock(&tasklist_lock);
4009         unlock_cpu_hotplug();
4010         if (retval)
4011                 return retval;
4012
4013         return 0;
4014 }
4015
4016 /**
4017  * sys_sched_getaffinity - get the cpu affinity of a process
4018  * @pid: pid of the process
4019  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4020  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4021  */
4022 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4023                                       unsigned long __user *user_mask_ptr)
4024 {
4025         int ret;
4026         cpumask_t mask;
4027
4028         if (len < sizeof(cpumask_t))
4029                 return -EINVAL;
4030
4031         ret = sched_getaffinity(pid, &mask);
4032         if (ret < 0)
4033                 return ret;
4034
4035         if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4036                 return -EFAULT;
4037
4038         return sizeof(cpumask_t);
4039 }
4040
4041 /**
4042  * sys_sched_yield - yield the current processor to other threads.
4043  *
4044  * this function yields the current CPU by moving the calling thread
4045  * to the expired array. If there are no other threads running on this
4046  * CPU then this function will return.
4047  */
4048 asmlinkage long sys_sched_yield(void)
4049 {
4050         runqueue_t *rq = this_rq_lock();
4051         prio_array_t *array = current->array;
4052         prio_array_t *target = rq->expired;
4053
4054         schedstat_inc(rq, yld_cnt);
4055         /*
4056          * We implement yielding by moving the task into the expired
4057          * queue.
4058          *
4059          * (special rule: RT tasks will just roundrobin in the active
4060          *  array.)
4061          */
4062         if (rt_task(current))
4063                 target = rq->active;
4064
4065         if (array->nr_active == 1) {
4066                 schedstat_inc(rq, yld_act_empty);
4067                 if (!rq->expired->nr_active)
4068                         schedstat_inc(rq, yld_both_empty);
4069         } else if (!rq->expired->nr_active)
4070                 schedstat_inc(rq, yld_exp_empty);
4071
4072         if (array != target) {
4073                 dequeue_task(current, array);
4074                 enqueue_task(current, target);
4075         } else
4076                 /*
4077                  * requeue_task is cheaper so perform that if possible.
4078                  */
4079                 requeue_task(current, array);
4080
4081         /*
4082          * Since we are going to call schedule() anyway, there's
4083          * no need to preempt or enable interrupts:
4084          */
4085         __release(rq->lock);
4086         _raw_spin_unlock(&rq->lock);
4087         preempt_enable_no_resched();
4088
4089         schedule();
4090
4091         return 0;
4092 }
4093
4094 static inline void __cond_resched(void)
4095 {
4096         /*
4097          * The BKS might be reacquired before we have dropped
4098          * PREEMPT_ACTIVE, which could trigger a second
4099          * cond_resched() call.
4100          */
4101         if (unlikely(preempt_count()))
4102                 return;
4103         do {
4104                 add_preempt_count(PREEMPT_ACTIVE);
4105                 schedule();
4106                 sub_preempt_count(PREEMPT_ACTIVE);
4107         } while (need_resched());
4108 }
4109
4110 int __sched cond_resched(void)
4111 {
4112         if (need_resched()) {
4113                 __cond_resched();
4114                 return 1;
4115         }
4116         return 0;
4117 }
4118
4119 EXPORT_SYMBOL(cond_resched);
4120
4121 /*
4122  * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4123  * call schedule, and on return reacquire the lock.
4124  *
4125  * This works OK both with and without CONFIG_PREEMPT.  We do strange low-level
4126  * operations here to prevent schedule() from being called twice (once via
4127  * spin_unlock(), once by hand).
4128  */
4129 int cond_resched_lock(spinlock_t *lock)
4130 {
4131         int ret = 0;
4132
4133         if (need_lockbreak(lock)) {
4134                 spin_unlock(lock);
4135                 cpu_relax();
4136                 ret = 1;
4137                 spin_lock(lock);
4138         }
4139         if (need_resched()) {
4140                 _raw_spin_unlock(lock);
4141                 preempt_enable_no_resched();
4142                 __cond_resched();
4143                 ret = 1;
4144                 spin_lock(lock);
4145         }
4146         return ret;
4147 }
4148
4149 EXPORT_SYMBOL(cond_resched_lock);
4150
4151 int __sched cond_resched_softirq(void)
4152 {
4153         BUG_ON(!in_softirq());
4154
4155         if (need_resched()) {
4156                 __local_bh_enable();
4157                 __cond_resched();
4158                 local_bh_disable();
4159                 return 1;
4160         }
4161         return 0;
4162 }
4163
4164 EXPORT_SYMBOL(cond_resched_softirq);
4165
4166
4167 /**
4168  * yield - yield the current processor to other threads.
4169  *
4170  * this is a shortcut for kernel-space yielding - it marks the
4171  * thread runnable and calls sys_sched_yield().
4172  */
4173 void __sched yield(void)
4174 {
4175         set_current_state(TASK_RUNNING);
4176         sys_sched_yield();
4177 }
4178
4179 EXPORT_SYMBOL(yield);
4180
4181 /*
4182  * This task is about to go to sleep on IO.  Increment rq->nr_iowait so
4183  * that process accounting knows that this is a task in IO wait state.
4184  *
4185  * But don't do that if it is a deliberate, throttling IO wait (this task
4186  * has set its backing_dev_info: the queue against which it should throttle)
4187  */
4188 void __sched io_schedule(void)
4189 {
4190         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4191
4192         atomic_inc(&rq->nr_iowait);
4193         schedule();
4194         atomic_dec(&rq->nr_iowait);
4195 }
4196
4197 EXPORT_SYMBOL(io_schedule);
4198
4199 long __sched io_schedule_timeout(long timeout)
4200 {
4201         struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4202         long ret;
4203
4204         atomic_inc(&rq->nr_iowait);
4205         ret = schedule_timeout(timeout);
4206         atomic_dec(&rq->nr_iowait);
4207         return ret;
4208 }
4209
4210 /**
4211  * sys_sched_get_priority_max - return maximum RT priority.
4212  * @policy: scheduling class.
4213  *
4214  * this syscall returns the maximum rt_priority that can be used
4215  * by a given scheduling class.
4216  */
4217 asmlinkage long sys_sched_get_priority_max(int policy)
4218 {
4219         int ret = -EINVAL;
4220
4221         switch (policy) {
4222         case SCHED_FIFO:
4223         case SCHED_RR:
4224                 ret = MAX_USER_RT_PRIO-1;
4225                 break;
4226         case SCHED_NORMAL:
4227                 ret = 0;
4228                 break;
4229         }
4230         return ret;
4231 }
4232
4233 /**
4234  * sys_sched_get_priority_min - return minimum RT priority.
4235  * @policy: scheduling class.
4236  *
4237  * this syscall returns the minimum rt_priority that can be used
4238  * by a given scheduling class.
4239  */
4240 asmlinkage long sys_sched_get_priority_min(int policy)
4241 {
4242         int ret = -EINVAL;
4243
4244         switch (policy) {
4245         case SCHED_FIFO:
4246         case SCHED_RR:
4247                 ret = 1;
4248                 break;
4249         case SCHED_NORMAL:
4250                 ret = 0;
4251         }
4252         return ret;
4253 }
4254
4255 /**
4256  * sys_sched_rr_get_interval - return the default timeslice of a process.
4257  * @pid: pid of the process.
4258  * @interval: userspace pointer to the timeslice value.
4259  *
4260  * this syscall writes the default timeslice value of a given process
4261  * into the user-space timespec buffer. A value of '0' means infinity.
4262  */
4263 asmlinkage
4264 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4265 {
4266         int retval = -EINVAL;
4267         struct timespec t;
4268         task_t *p;
4269
4270         if (pid < 0)
4271                 goto out_nounlock;
4272
4273         retval = -ESRCH;
4274         read_lock(&tasklist_lock);
4275         p = find_process_by_pid(pid);
4276         if (!p)
4277                 goto out_unlock;
4278
4279         retval = security_task_getscheduler(p);
4280         if (retval)
4281                 goto out_unlock;
4282
4283         jiffies_to_timespec(p->policy & SCHED_FIFO ?
4284                                 0 : task_timeslice(p), &t);
4285         read_unlock(&tasklist_lock);
4286         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4287 out_nounlock:
4288         return retval;
4289 out_unlock:
4290         read_unlock(&tasklist_lock);
4291         return retval;
4292 }
4293
4294 static inline struct task_struct *eldest_child(struct task_struct *p)
4295 {
4296         if (list_empty(&p->children)) return NULL;
4297         return list_entry(p->children.next,struct task_struct,sibling);
4298 }
4299
4300 static inline struct task_struct *older_sibling(struct task_struct *p)
4301 {
4302         if (p->sibling.prev==&p->parent->children) return NULL;
4303         return list_entry(p->sibling.prev,struct task_struct,sibling);
4304 }
4305
4306 static inline struct task_struct *younger_sibling(struct task_struct *p)
4307 {
4308         if (p->sibling.next==&p->parent->children) return NULL;
4309         return list_entry(p->sibling.next,struct task_struct,sibling);
4310 }
4311
4312 static void show_task(task_t *p)
4313 {
4314         task_t *relative;
4315         unsigned state;
4316         unsigned long free = 0;
4317         static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4318
4319         printk("%-13.13s ", p->comm);
4320         state = p->state ? __ffs(p->state) + 1 : 0;
4321         if (state < ARRAY_SIZE(stat_nam))
4322                 printk(stat_nam[state]);
4323         else
4324                 printk("?");
4325 #if (BITS_PER_LONG == 32)
4326         if (state == TASK_RUNNING)
4327                 printk(" running ");
4328         else
4329                 printk(" %08lX ", thread_saved_pc(p));
4330 #else
4331         if (state == TASK_RUNNING)
4332                 printk("  running task   ");
4333         else
4334                 printk(" %016lx ", thread_saved_pc(p));
4335 #endif
4336 #ifdef CONFIG_DEBUG_STACK_USAGE
4337         {
4338                 unsigned long *n = end_of_stack(p);
4339                 while (!*n)
4340                         n++;
4341                 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4342         }
4343 #endif
4344         printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4345         if ((relative = eldest_child(p)))
4346                 printk("%5d ", relative->pid);
4347         else
4348                 printk("      ");
4349         if ((relative = younger_sibling(p)))
4350                 printk("%7d", relative->pid);
4351         else
4352                 printk("       ");
4353         if ((relative = older_sibling(p)))
4354                 printk(" %5d", relative->pid);
4355         else
4356                 printk("      ");
4357         if (!p->mm)
4358                 printk(" (L-TLB)\n");
4359         else
4360                 printk(" (NOTLB)\n");
4361
4362         if (state != TASK_RUNNING)
4363                 show_stack(p, NULL);
4364 }
4365
4366 void show_state(void)
4367 {
4368         task_t *g, *p;
4369
4370 #if (BITS_PER_LONG == 32)
4371         printk("\n"
4372                "                                               sibling\n");
4373         printk("  task             PC      pid father child younger older\n");
4374 #else
4375         printk("\n"
4376                "                                                       sibling\n");
4377         printk("  task                 PC          pid father child younger older\n");
4378 #endif
4379         read_lock(&tasklist_lock);
4380         do_each_thread(g, p) {
4381                 /*
4382                  * reset the NMI-timeout, listing all files on a slow
4383                  * console might take alot of time:
4384                  */
4385                 touch_nmi_watchdog();
4386                 show_task(p);
4387         } while_each_thread(g, p);
4388
4389         read_unlock(&tasklist_lock);
4390         mutex_debug_show_all_locks();
4391 }
4392
4393 /**
4394  * init_idle - set up an idle thread for a given CPU
4395  * @idle: task in question
4396  * @cpu: cpu the idle task belongs to
4397  *
4398  * NOTE: this function does not set the idle thread's NEED_RESCHED
4399  * flag, to make booting more robust.
4400  */
4401 void __devinit init_idle(task_t *idle, int cpu)
4402 {
4403         runqueue_t *rq = cpu_rq(cpu);
4404         unsigned long flags;
4405
4406         idle->sleep_avg = 0;
4407         idle->array = NULL;
4408         idle->prio = MAX_PRIO;
4409         idle->state = TASK_RUNNING;
4410         idle->cpus_allowed = cpumask_of_cpu(cpu);
4411         set_task_cpu(idle, cpu);
4412
4413         spin_lock_irqsave(&rq->lock, flags);
4414         rq->curr = rq->idle = idle;
4415 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4416         idle->oncpu = 1;
4417 #endif
4418         spin_unlock_irqrestore(&rq->lock, flags);
4419
4420         /* Set the preempt count _outside_ the spinlocks! */
4421 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4422         task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4423 #else
4424         task_thread_info(idle)->preempt_count = 0;
4425 #endif
4426 }
4427
4428 /*
4429  * In a system that switches off the HZ timer nohz_cpu_mask
4430  * indicates which cpus entered this state. This is used
4431  * in the rcu update to wait only for active cpus. For system
4432  * which do not switch off the HZ timer nohz_cpu_mask should
4433  * always be CPU_MASK_NONE.
4434  */
4435 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4436
4437 #ifdef CONFIG_SMP
4438 /*
4439  * This is how migration works:
4440  *
4441  * 1) we queue a migration_req_t structure in the source CPU's
4442  *    runqueue and wake up that CPU's migration thread.
4443  * 2) we down() the locked semaphore => thread blocks.
4444  * 3) migration thread wakes up (implicitly it forces the migrated
4445  *    thread off the CPU)
4446  * 4) it gets the migration request and checks whether the migrated
4447  *    task is still in the wrong runqueue.
4448  * 5) if it's in the wrong runqueue then the migration thread removes
4449  *    it and puts it into the right queue.
4450  * 6) migration thread up()s the semaphore.
4451  * 7) we wake up and the migration is done.
4452  */
4453
4454 /*
4455  * Change a given task's CPU affinity. Migrate the thread to a
4456  * proper CPU and schedule it away if the CPU it's executing on
4457  * is removed from the allowed bitmask.
4458  *
4459  * NOTE: the caller must have a valid reference to the task, the
4460  * task must not exit() & deallocate itself prematurely.  The
4461  * call is not atomic; no spinlocks may be held.
4462  */
4463 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4464 {
4465         unsigned long flags;
4466         int ret = 0;
4467         migration_req_t req;
4468         runqueue_t *rq;
4469
4470         rq = task_rq_lock(p, &flags);
4471         if (!cpus_intersects(new_mask, cpu_online_map)) {
4472                 ret = -EINVAL;
4473                 goto out;
4474         }
4475
4476         p->cpus_allowed = new_mask;
4477         /* Can the task run on the task's current CPU? If so, we're done */
4478         if (cpu_isset(task_cpu(p), new_mask))
4479                 goto out;
4480
4481         if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4482                 /* Need help from migration thread: drop lock and wait. */
4483                 task_rq_unlock(rq, &flags);
4484                 wake_up_process(rq->migration_thread);
4485                 wait_for_completion(&req.done);
4486                 tlb_migrate_finish(p->mm);
4487                 return 0;
4488         }
4489 out:
4490         task_rq_unlock(rq, &flags);
4491         return ret;
4492 }
4493
4494 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4495
4496 /*
4497  * Move (not current) task off this cpu, onto dest cpu.  We're doing
4498  * this because either it can't run here any more (set_cpus_allowed()
4499  * away from this CPU, or CPU going down), or because we're
4500  * attempting to rebalance this task on exec (sched_exec).
4501  *
4502  * So we race with normal scheduler movements, but that's OK, as long
4503  * as the task is no longer on this CPU.
4504  */
4505 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4506 {
4507         runqueue_t *rq_dest, *rq_src;
4508
4509         if (unlikely(cpu_is_offline(dest_cpu)))
4510                 return;
4511
4512         rq_src = cpu_rq(src_cpu);
4513         rq_dest = cpu_rq(dest_cpu);
4514
4515         double_rq_lock(rq_src, rq_dest);
4516         /* Already moved. */
4517         if (task_cpu(p) != src_cpu)
4518                 goto out;
4519         /* Affinity changed (again). */
4520         if (!cpu_isset(dest_cpu, p->cpus_allowed))
4521                 goto out;
4522
4523         set_task_cpu(p, dest_cpu);
4524         if (p->array) {
4525                 /*
4526                  * Sync timestamp with rq_dest's before activating.
4527                  * The same thing could be achieved by doing this step
4528                  * afterwards, and pretending it was a local activate.
4529                  * This way is cleaner and logically correct.
4530                  */
4531                 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4532                                 + rq_dest->timestamp_last_tick;
4533                 deactivate_task(p, rq_src);
4534                 activate_task(p, rq_dest, 0);
4535                 if (TASK_PREEMPTS_CURR(p, rq_dest))
4536                         resched_task(rq_dest->curr);
4537         }
4538
4539 out:
4540         double_rq_unlock(rq_src, rq_dest);
4541 }
4542
4543 /*
4544  * migration_thread - this is a highprio system thread that performs
4545  * thread migration by bumping thread off CPU then 'pushing' onto
4546  * another runqueue.
4547  */
4548 static int migration_thread(void *data)
4549 {
4550         runqueue_t *rq;
4551         int cpu = (long)data;
4552
4553         rq = cpu_rq(cpu);
4554         BUG_ON(rq->migration_thread != current);
4555
4556         set_current_state(TASK_INTERRUPTIBLE);
4557         while (!kthread_should_stop()) {
4558                 struct list_head *head;
4559                 migration_req_t *req;
4560
4561                 try_to_freeze();
4562
4563                 spin_lock_irq(&rq->lock);
4564
4565                 if (cpu_is_offline(cpu)) {
4566                         spin_unlock_irq(&rq->lock);
4567                         goto wait_to_die;
4568                 }
4569
4570                 if (rq->active_balance) {
4571                         active_load_balance(rq, cpu);
4572                         rq->active_balance = 0;
4573                 }
4574
4575                 head = &rq->migration_queue;
4576
4577                 if (list_empty(head)) {
4578                         spin_unlock_irq(&rq->lock);
4579                         schedule();
4580                         set_current_state(TASK_INTERRUPTIBLE);
4581                         continue;
4582                 }
4583                 req = list_entry(head->next, migration_req_t, list);
4584                 list_del_init(head->next);
4585
4586                 spin_unlock(&rq->lock);
4587                 __migrate_task(req->task, cpu, req->dest_cpu);
4588                 local_irq_enable();
4589
4590                 complete(&req->done);
4591         }
4592         __set_current_state(TASK_RUNNING);
4593         return 0;
4594
4595 wait_to_die:
4596         /* Wait for kthread_stop */
4597         set_current_state(TASK_INTERRUPTIBLE);
4598         while (!kthread_should_stop()) {
4599                 schedule();
4600                 set_current_state(TASK_INTERRUPTIBLE);
4601         }
4602         __set_current_state(TASK_RUNNING);
4603         return 0;
4604 }
4605
4606 #ifdef CONFIG_HOTPLUG_CPU
4607 /* Figure out where task on dead CPU should go, use force if neccessary. */
4608 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4609 {
4610         int dest_cpu;
4611         cpumask_t mask;
4612
4613         /* On same node? */
4614         mask = node_to_cpumask(cpu_to_node(dead_cpu));
4615         cpus_and(mask, mask, tsk->cpus_allowed);
4616         dest_cpu = any_online_cpu(mask);
4617
4618         /* On any allowed CPU? */
4619         if (dest_cpu == NR_CPUS)
4620                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4621
4622         /* No more Mr. Nice Guy. */
4623         if (dest_cpu == NR_CPUS) {
4624                 cpus_setall(tsk->cpus_allowed);
4625                 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4626
4627                 /*
4628                  * Don't tell them about moving exiting tasks or
4629                  * kernel threads (both mm NULL), since they never
4630                  * leave kernel.
4631                  */
4632                 if (tsk->mm && printk_ratelimit())
4633                         printk(KERN_INFO "process %d (%s) no "
4634                                "longer affine to cpu%d\n",
4635                                tsk->pid, tsk->comm, dead_cpu);
4636         }
4637         __migrate_task(tsk, dead_cpu, dest_cpu);
4638 }
4639
4640 /*
4641  * While a dead CPU has no uninterruptible tasks queued at this point,
4642  * it might still have a nonzero ->nr_uninterruptible counter, because
4643  * for performance reasons the counter is not stricly tracking tasks to
4644  * their home CPUs. So we just add the counter to another CPU's counter,
4645  * to keep the global sum constant after CPU-down:
4646  */
4647 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4648 {
4649         runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4650         unsigned long flags;
4651
4652         local_irq_save(flags);
4653         double_rq_lock(rq_src, rq_dest);
4654         rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4655         rq_src->nr_uninterruptible = 0;
4656         double_rq_unlock(rq_src, rq_dest);
4657         local_irq_restore(flags);
4658 }
4659
4660 /* Run through task list and migrate tasks from the dead cpu. */
4661 static void migrate_live_tasks(int src_cpu)
4662 {
4663         struct task_struct *tsk, *t;
4664
4665         write_lock_irq(&tasklist_lock);
4666
4667         do_each_thread(t, tsk) {
4668                 if (tsk == current)
4669                         continue;
4670
4671                 if (task_cpu(tsk) == src_cpu)
4672                         move_task_off_dead_cpu(src_cpu, tsk);
4673         } while_each_thread(t, tsk);
4674
4675         write_unlock_irq(&tasklist_lock);
4676 }
4677
4678 /* Schedules idle task to be the next runnable task on current CPU.
4679  * It does so by boosting its priority to highest possible and adding it to
4680  * the _front_ of runqueue. Used by CPU offline code.
4681  */
4682 void sched_idle_next(void)
4683 {
4684         int cpu = smp_processor_id();
4685         runqueue_t *rq = this_rq();
4686         struct task_struct *p = rq->idle;
4687         unsigned long flags;
4688
4689         /* cpu has to be offline */
4690         BUG_ON(cpu_online(cpu));
4691
4692         /* Strictly not necessary since rest of the CPUs are stopped by now
4693          * and interrupts disabled on current cpu.
4694          */
4695         spin_lock_irqsave(&rq->lock, flags);
4696
4697         __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4698         /* Add idle task to _front_ of it's priority queue */
4699         __activate_idle_task(p, rq);
4700
4701         spin_unlock_irqrestore(&rq->lock, flags);
4702 }
4703
4704 /* Ensures that the idle task is using init_mm right before its cpu goes
4705  * offline.
4706  */
4707 void idle_task_exit(void)
4708 {
4709         struct mm_struct *mm = current->active_mm;
4710
4711         BUG_ON(cpu_online(smp_processor_id()));
4712
4713         if (mm != &init_mm)
4714                 switch_mm(mm, &init_mm, current);
4715         mmdrop(mm);
4716 }
4717
4718 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4719 {
4720         struct runqueue *rq = cpu_rq(dead_cpu);
4721
4722         /* Must be exiting, otherwise would be on tasklist. */
4723         BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4724
4725         /* Cannot have done final schedule yet: would have vanished. */
4726         BUG_ON(tsk->flags & PF_DEAD);
4727
4728         get_task_struct(tsk);
4729
4730         /*
4731          * Drop lock around migration; if someone else moves it,
4732          * that's OK.  No task can be added to this CPU, so iteration is
4733          * fine.
4734          */
4735         spin_unlock_irq(&rq->lock);
4736         move_task_off_dead_cpu(dead_cpu, tsk);
4737         spin_lock_irq(&rq->lock);
4738
4739         put_task_struct(tsk);
4740 }
4741
4742 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4743 static void migrate_dead_tasks(unsigned int dead_cpu)
4744 {
4745         unsigned arr, i;
4746         struct runqueue *rq = cpu_rq(dead_cpu);
4747
4748         for (arr = 0; arr < 2; arr++) {
4749                 for (i = 0; i < MAX_PRIO; i++) {
4750                         struct list_head *list = &rq->arrays[arr].queue[i];
4751                         while (!list_empty(list))
4752                                 migrate_dead(dead_cpu,
4753                                              list_entry(list->next, task_t,
4754                                                         run_list));
4755                 }
4756         }
4757 }
4758 #endif /* CONFIG_HOTPLUG_CPU */
4759
4760 /*
4761  * migration_call - callback that gets triggered when a CPU is added.
4762  * Here we can start up the necessary migration thread for the new CPU.
4763  */
4764 static int migration_call(struct notifier_block *nfb, unsigned long action,
4765                           void *hcpu)
4766 {
4767         int cpu = (long)hcpu;
4768         struct task_struct *p;
4769         struct runqueue *rq;
4770         unsigned long flags;
4771
4772         switch (action) {
4773         case CPU_UP_PREPARE:
4774                 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4775                 if (IS_ERR(p))
4776                         return NOTIFY_BAD;
4777                 p->flags |= PF_NOFREEZE;
4778                 kthread_bind(p, cpu);
4779                 /* Must be high prio: stop_machine expects to yield to it. */
4780                 rq = task_rq_lock(p, &flags);
4781                 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4782                 task_rq_unlock(rq, &flags);
4783                 cpu_rq(cpu)->migration_thread = p;
4784                 break;
4785         case CPU_ONLINE:
4786                 /* Strictly unneccessary, as first user will wake it. */
4787                 wake_up_process(cpu_rq(cpu)->migration_thread);
4788                 break;
4789 #ifdef CONFIG_HOTPLUG_CPU
4790         case CPU_UP_CANCELED:
4791                 /* Unbind it from offline cpu so it can run.  Fall thru. */
4792                 kthread_bind(cpu_rq(cpu)->migration_thread,
4793                              any_online_cpu(cpu_online_map));
4794                 kthread_stop(cpu_rq(cpu)->migration_thread);
4795                 cpu_rq(cpu)->migration_thread = NULL;
4796                 break;
4797         case CPU_DEAD:
4798                 migrate_live_tasks(cpu);
4799                 rq = cpu_rq(cpu);
4800                 kthread_stop(rq->migration_thread);
4801                 rq->migration_thread = NULL;
4802                 /* Idle task back to normal (off runqueue, low prio) */
4803                 rq = task_rq_lock(rq->idle, &flags);
4804                 deactivate_task(rq->idle, rq);
4805                 rq->idle->static_prio = MAX_PRIO;
4806                 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4807                 migrate_dead_tasks(cpu);
4808                 task_rq_unlock(rq, &flags);
4809                 migrate_nr_uninterruptible(rq);
4810                 BUG_ON(rq->nr_running != 0);
4811
4812                 /* No need to migrate the tasks: it was best-effort if
4813                  * they didn't do lock_cpu_hotplug().  Just wake up
4814                  * the requestors. */
4815                 spin_lock_irq(&rq->lock);
4816                 while (!list_empty(&rq->migration_queue)) {
4817                         migration_req_t *req;
4818                         req = list_entry(rq->migration_queue.next,
4819                                          migration_req_t, list);
4820                         list_del_init(&req->list);
4821                         complete(&req->done);
4822                 }
4823                 spin_unlock_irq(&rq->lock);
4824                 break;
4825 #endif
4826         }
4827         return NOTIFY_OK;
4828 }
4829
4830 /* Register at highest priority so that task migration (migrate_all_tasks)
4831  * happens before everything else.
4832  */
4833 static struct notifier_block __devinitdata migration_notifier = {
4834         .notifier_call = migration_call,
4835         .priority = 10
4836 };
4837
4838 int __init migration_init(void)
4839 {
4840         void *cpu = (void *)(long)smp_processor_id();
4841         /* Start one for boot CPU. */
4842         migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4843         migration_call(&migration_notifier, CPU_ONLINE, cpu);
4844         register_cpu_notifier(&migration_notifier);
4845         return 0;
4846 }
4847 #endif
4848
4849 #ifdef CONFIG_SMP
4850 #undef SCHED_DOMAIN_DEBUG
4851 #ifdef SCHED_DOMAIN_DEBUG
4852 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4853 {
4854         int level = 0;
4855
4856         if (!sd) {
4857                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4858                 return;
4859         }
4860
4861         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4862
4863         do {
4864                 int i;
4865                 char str[NR_CPUS];
4866                 struct sched_group *group = sd->groups;
4867                 cpumask_t groupmask;
4868
4869                 cpumask_scnprintf(str, NR_CPUS, sd->span);
4870                 cpus_clear(groupmask);
4871
4872                 printk(KERN_DEBUG);
4873                 for (i = 0; i < level + 1; i++)
4874                         printk(" ");
4875                 printk("domain %d: ", level);
4876
4877                 if (!(sd->flags & SD_LOAD_BALANCE)) {
4878                         printk("does not load-balance\n");
4879                         if (sd->parent)
4880                                 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4881                         break;
4882                 }
4883
4884                 printk("span %s\n", str);
4885
4886                 if (!cpu_isset(cpu, sd->span))
4887                         printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4888                 if (!cpu_isset(cpu, group->cpumask))
4889                         printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4890
4891                 printk(KERN_DEBUG);
4892                 for (i = 0; i < level + 2; i++)
4893                         printk(" ");
4894                 printk("groups:");
4895                 do {
4896                         if (!group) {
4897                                 printk("\n");
4898                                 printk(KERN_ERR "ERROR: group is NULL\n");
4899                                 break;
4900                         }
4901
4902                         if (!group->cpu_power) {
4903                                 printk("\n");
4904                                 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4905                         }
4906
4907                         if (!cpus_weight(group->cpumask)) {
4908                                 printk("\n");
4909                                 printk(KERN_ERR "ERROR: empty group\n");
4910                         }
4911
4912                         if (cpus_intersects(groupmask, group->cpumask)) {
4913                                 printk("\n");
4914                                 printk(KERN_ERR "ERROR: repeated CPUs\n");
4915                         }
4916
4917                         cpus_or(groupmask, groupmask, group->cpumask);
4918
4919                         cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4920                         printk(" %s", str);
4921
4922                         group = group->next;
4923                 } while (group != sd->groups);
4924                 printk("\n");
4925
4926                 if (!cpus_equal(sd->span, groupmask))
4927                         printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4928
4929                 level++;
4930                 sd = sd->parent;
4931
4932                 if (sd) {
4933                         if (!cpus_subset(groupmask, sd->span))
4934                                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4935                 }
4936
4937         } while (sd);
4938 }
4939 #else
4940 #define sched_domain_debug(sd, cpu) {}
4941 #endif
4942
4943 static int sd_degenerate(struct sched_domain *sd)
4944 {
4945         if (cpus_weight(sd->span) == 1)
4946                 return 1;
4947
4948         /* Following flags need at least 2 groups */
4949         if (sd->flags & (SD_LOAD_BALANCE |
4950                          SD_BALANCE_NEWIDLE |
4951                          SD_BALANCE_FORK |
4952                          SD_BALANCE_EXEC)) {
4953                 if (sd->groups != sd->groups->next)
4954                         return 0;
4955         }
4956
4957         /* Following flags don't use groups */
4958         if (sd->flags & (SD_WAKE_IDLE |
4959                          SD_WAKE_AFFINE |
4960                          SD_WAKE_BALANCE))
4961                 return 0;
4962
4963         return 1;
4964 }
4965
4966 static int sd_parent_degenerate(struct sched_domain *sd,
4967                                                 struct sched_domain *parent)
4968 {
4969         unsigned long cflags = sd->flags, pflags = parent->flags;
4970
4971         if (sd_degenerate(parent))
4972                 return 1;
4973
4974         if (!cpus_equal(sd->span, parent->span))
4975                 return 0;
4976
4977         /* Does parent contain flags not in child? */
4978         /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4979         if (cflags & SD_WAKE_AFFINE)
4980                 pflags &= ~SD_WAKE_BALANCE;
4981         /* Flags needing groups don't count if only 1 group in parent */
4982         if (parent->groups == parent->groups->next) {
4983                 pflags &= ~(SD_LOAD_BALANCE |
4984                                 SD_BALANCE_NEWIDLE |
4985                                 SD_BALANCE_FORK |
4986                                 SD_BALANCE_EXEC);
4987         }
4988         if (~cflags & pflags)
4989                 return 0;
4990
4991         return 1;
4992 }
4993
4994 /*
4995  * Attach the domain 'sd' to 'cpu' as its base domain.  Callers must
4996  * hold the hotplug lock.
4997  */
4998 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4999 {
5000         runqueue_t *rq = cpu_rq(cpu);
5001         struct sched_domain *tmp;
5002
5003         /* Remove the sched domains which do not contribute to scheduling. */
5004         for (tmp = sd; tmp; tmp = tmp->parent) {
5005                 struct sched_domain *parent = tmp->parent;
5006                 if (!parent)
5007                         break;
5008                 if (sd_parent_degenerate(tmp, parent))
5009                         tmp->parent = parent->parent;
5010         }
5011
5012         if (sd && sd_degenerate(sd))
5013                 sd = sd->parent;
5014
5015         sched_domain_debug(sd, cpu);
5016
5017         rcu_assign_pointer(rq->sd, sd);
5018 }
5019
5020 /* cpus with isolated domains */
5021 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
5022
5023 /* Setup the mask of cpus configured for isolated domains */
5024 static int __init isolated_cpu_setup(char *str)
5025 {
5026         int ints[NR_CPUS], i;
5027
5028         str = get_options(str, ARRAY_SIZE(ints), ints);
5029         cpus_clear(cpu_isolated_map);
5030         for (i = 1; i <= ints[0]; i++)
5031                 if (ints[i] < NR_CPUS)
5032                         cpu_set(ints[i], cpu_isolated_map);
5033         return 1;
5034 }
5035
5036 __setup ("isolcpus=", isolated_cpu_setup);
5037
5038 /*
5039  * init_sched_build_groups takes an array of groups, the cpumask we wish
5040  * to span, and a pointer to a function which identifies what group a CPU
5041  * belongs to. The return value of group_fn must be a valid index into the
5042  * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5043  * keep track of groups covered with a cpumask_t).
5044  *
5045  * init_sched_build_groups will build a circular linked list of the groups
5046  * covered by the given span, and will set each group's ->cpumask correctly,
5047  * and ->cpu_power to 0.
5048  */
5049 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5050                                     int (*group_fn)(int cpu))
5051 {
5052         struct sched_group *first = NULL, *last = NULL;
5053         cpumask_t covered = CPU_MASK_NONE;
5054         int i;
5055
5056         for_each_cpu_mask(i, span) {
5057                 int group = group_fn(i);
5058                 struct sched_group *sg = &groups[group];
5059                 int j;
5060
5061                 if (cpu_isset(i, covered))
5062                         continue;
5063
5064                 sg->cpumask = CPU_MASK_NONE;
5065                 sg->cpu_power = 0;
5066
5067                 for_each_cpu_mask(j, span) {
5068                         if (group_fn(j) != group)
5069                                 continue;
5070
5071                         cpu_set(j, covered);
5072                         cpu_set(j, sg->cpumask);
5073                 }
5074                 if (!first)
5075                         first = sg;
5076                 if (last)
5077                         last->next = sg;
5078                 last = sg;
5079         }
5080         last->next = first;
5081 }
5082
5083 #define SD_NODES_PER_DOMAIN 16
5084
5085 #ifdef CONFIG_NUMA
5086 /**
5087  * find_next_best_node - find the next node to include in a sched_domain
5088  * @node: node whose sched_domain we're building
5089  * @used_nodes: nodes already in the sched_domain
5090  *
5091  * Find the next node to include in a given scheduling domain.  Simply
5092  * finds the closest node not already in the @used_nodes map.
5093  *
5094  * Should use nodemask_t.
5095  */
5096 static int find_next_best_node(int node, unsigned long *used_nodes)
5097 {
5098         int i, n, val, min_val, best_node = 0;
5099
5100         min_val = INT_MAX;
5101
5102         for (i = 0; i < MAX_NUMNODES; i++) {
5103                 /* Start at @node */
5104                 n = (node + i) % MAX_NUMNODES;
5105
5106                 if (!nr_cpus_node(n))
5107                         continue;
5108
5109                 /* Skip already used nodes */
5110                 if (test_bit(n, used_nodes))
5111                         continue;
5112
5113                 /* Simple min distance search */
5114                 val = node_distance(node, n);
5115
5116                 if (val < min_val) {
5117                         min_val = val;
5118                         best_node = n;
5119                 }
5120         }
5121
5122         set_bit(best_node, used_nodes);
5123         return best_node;
5124 }
5125
5126 /**
5127  * sched_domain_node_span - get a cpumask for a node's sched_domain
5128  * @node: node whose cpumask we're constructing
5129  * @size: number of nodes to include in this span
5130  *
5131  * Given a node, construct a good cpumask for its sched_domain to span.  It
5132  * should be one that prevents unnecessary balancing, but also spreads tasks
5133  * out optimally.
5134  */
5135 static cpumask_t sched_domain_node_span(int node)
5136 {
5137         int i;
5138         cpumask_t span, nodemask;
5139         DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5140
5141         cpus_clear(span);
5142         bitmap_zero(used_nodes, MAX_NUMNODES);
5143
5144         nodemask = node_to_cpumask(node);
5145         cpus_or(span, span, nodemask);
5146         set_bit(node, used_nodes);
5147
5148         for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5149                 int next_node = find_next_best_node(node, used_nodes);
5150                 nodemask = node_to_cpumask(next_node);
5151                 cpus_or(span, span, nodemask);
5152         }
5153
5154         return span;
5155 }
5156 #endif
5157
5158 /*
5159  * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5160  * can switch it on easily if needed.
5161  */
5162 #ifdef CONFIG_SCHED_SMT
5163 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5164 static struct sched_group sched_group_cpus[NR_CPUS];
5165 static int cpu_to_cpu_group(int cpu)
5166 {
5167         return cpu;
5168 }
5169 #endif
5170
5171 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5172 static struct sched_group sched_group_phys[NR_CPUS];
5173 static int cpu_to_phys_group(int cpu)
5174 {
5175 #ifdef CONFIG_SCHED_SMT
5176         return first_cpu(cpu_sibling_map[cpu]);
5177 #else
5178         return cpu;
5179 #endif
5180 }
5181
5182 #ifdef CONFIG_NUMA
5183 /*
5184  * The init_sched_build_groups can't handle what we want to do with node
5185  * groups, so roll our own. Now each node has its own list of groups which
5186  * gets dynamically allocated.
5187  */
5188 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5189 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5190
5191 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5192 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5193
5194 static int cpu_to_allnodes_group(int cpu)
5195 {
5196         return cpu_to_node(cpu);
5197 }
5198 #endif
5199
5200 /*
5201  * Build sched domains for a given set of cpus and attach the sched domains
5202  * to the individual cpus
5203  */
5204 void build_sched_domains(const cpumask_t *cpu_map)
5205 {
5206         int i;
5207 #ifdef CONFIG_NUMA
5208         struct sched_group **sched_group_nodes = NULL;
5209         struct sched_group *sched_group_allnodes = NULL;
5210
5211         /*
5212          * Allocate the per-node list of sched groups
5213          */
5214         sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5215                                            GFP_ATOMIC);
5216         if (!sched_group_nodes) {
5217                 printk(KERN_WARNING "Can not alloc sched group node list\n");
5218                 return;
5219         }
5220         sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5221 #endif
5222
5223         /*
5224          * Set up domains for cpus specified by the cpu_map.
5225          */
5226         for_each_cpu_mask(i, *cpu_map) {
5227                 int group;
5228                 struct sched_domain *sd = NULL, *p;
5229                 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5230
5231                 cpus_and(nodemask, nodemask, *cpu_map);
5232
5233 #ifdef CONFIG_NUMA
5234                 if (cpus_weight(*cpu_map)
5235                                 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5236                         if (!sched_group_allnodes) {
5237                                 sched_group_allnodes
5238                                         = kmalloc(sizeof(struct sched_group)
5239                                                         * MAX_NUMNODES,
5240                                                   GFP_KERNEL);
5241                                 if (!sched_group_allnodes) {
5242                                         printk(KERN_WARNING
5243                                         "Can not alloc allnodes sched group\n");
5244                                         break;
5245                                 }
5246                                 sched_group_allnodes_bycpu[i]
5247                                                 = sched_group_allnodes;
5248                         }
5249                         sd = &per_cpu(allnodes_domains, i);
5250                         *sd = SD_ALLNODES_INIT;
5251                         sd->span = *cpu_map;
5252                         group = cpu_to_allnodes_group(i);
5253                         sd->groups = &sched_group_allnodes[group];
5254                         p = sd;
5255                 } else
5256                         p = NULL;
5257
5258                 sd = &per_cpu(node_domains, i);
5259                 *sd = SD_NODE_INIT;
5260                 sd->span = sched_domain_node_span(cpu_to_node(i));
5261                 sd->parent = p;
5262                 cpus_and(sd->span, sd->span, *cpu_map);
5263 #endif
5264
5265                 p = sd;
5266                 sd = &per_cpu(phys_domains, i);
5267                 group = cpu_to_phys_group(i);
5268                 *sd = SD_CPU_INIT;
5269                 sd->span = nodemask;
5270                 sd->parent = p;
5271                 sd->groups = &sched_group_phys[group];
5272
5273 #ifdef CONFIG_SCHED_SMT
5274                 p = sd;
5275                 sd = &per_cpu(cpu_domains, i);
5276                 group = cpu_to_cpu_group(i);
5277                 *sd = SD_SIBLING_INIT;
5278                 sd->span = cpu_sibling_map[i];
5279                 cpus_and(sd->span, sd->span, *cpu_map);
5280                 sd->parent = p;
5281                 sd->groups = &sched_group_cpus[group];
5282 #endif
5283         }
5284
5285 #ifdef CONFIG_SCHED_SMT
5286         /* Set up CPU (sibling) groups */
5287         for_each_cpu_mask(i, *cpu_map) {
5288                 cpumask_t this_sibling_map = cpu_sibling_map[i];
5289                 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5290                 if (i != first_cpu(this_sibling_map))
5291                         continue;
5292
5293                 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5294                                                 &cpu_to_cpu_group);
5295         }
5296 #endif
5297
5298         /* Set up physical groups */
5299         for (i = 0; i < MAX_NUMNODES; i++) {
5300                 cpumask_t nodemask = node_to_cpumask(i);
5301
5302                 cpus_and(nodemask, nodemask, *cpu_map);
5303                 if (cpus_empty(nodemask))
5304                         continue;
5305
5306                 init_sched_build_groups(sched_group_phys, nodemask,
5307                                                 &cpu_to_phys_group);
5308         }
5309
5310 #ifdef CONFIG_NUMA
5311         /* Set up node groups */
5312         if (sched_group_allnodes)
5313                 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5314                                         &cpu_to_allnodes_group);
5315
5316         for (i = 0; i < MAX_NUMNODES; i++) {
5317                 /* Set up node groups */
5318                 struct sched_group *sg, *prev;
5319                 cpumask_t nodemask = node_to_cpumask(i);
5320                 cpumask_t domainspan;
5321                 cpumask_t covered = CPU_MASK_NONE;
5322                 int j;
5323
5324                 cpus_and(nodemask, nodemask, *cpu_map);
5325                 if (cpus_empty(nodemask)) {
5326                         sched_group_nodes[i] = NULL;
5327                         continue;
5328                 }
5329
5330                 domainspan = sched_domain_node_span(i);
5331                 cpus_and(domainspan, domainspan, *cpu_map);
5332
5333                 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5334                 sched_group_nodes[i] = sg;
5335                 for_each_cpu_mask(j, nodemask) {
5336                         struct sched_domain *sd;
5337                         sd = &per_cpu(node_domains, j);
5338                         sd->groups = sg;
5339                         if (sd->groups == NULL) {
5340                                 /* Turn off balancing if we have no groups */
5341                                 sd->flags = 0;
5342                         }
5343                 }
5344                 if (!sg) {
5345                         printk(KERN_WARNING
5346                         "Can not alloc domain group for node %d\n", i);
5347                         continue;
5348                 }
5349                 sg->cpu_power = 0;
5350                 sg->cpumask = nodemask;
5351                 cpus_or(covered, covered, nodemask);
5352                 prev = sg;
5353
5354                 for (j = 0; j < MAX_NUMNODES; j++) {
5355                         cpumask_t tmp, notcovered;
5356                         int n = (i + j) % MAX_NUMNODES;
5357
5358                         cpus_complement(notcovered, covered);
5359                         cpus_and(tmp, notcovered, *cpu_map);
5360                         cpus_and(tmp, tmp, domainspan);
5361                         if (cpus_empty(tmp))
5362                                 break;
5363
5364                         nodemask = node_to_cpumask(n);
5365                         cpus_and(tmp, tmp, nodemask);
5366                         if (cpus_empty(tmp))
5367                                 continue;
5368
5369                         sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5370                         if (!sg) {
5371                                 printk(KERN_WARNING
5372                                 "Can not alloc domain group for node %d\n", j);
5373                                 break;
5374                         }
5375                         sg->cpu_power = 0;
5376                         sg->cpumask = tmp;
5377                         cpus_or(covered, covered, tmp);
5378                         prev->next = sg;
5379                         prev = sg;
5380                 }
5381                 prev->next = sched_group_nodes[i];
5382         }
5383 #endif
5384
5385         /* Calculate CPU power for physical packages and nodes */
5386         for_each_cpu_mask(i, *cpu_map) {
5387                 int power;
5388                 struct sched_domain *sd;
5389 #ifdef CONFIG_SCHED_SMT
5390                 sd = &per_cpu(cpu_domains, i);
5391                 power = SCHED_LOAD_SCALE;
5392                 sd->groups->cpu_power = power;
5393 #endif
5394
5395                 sd = &per_cpu(phys_domains, i);
5396                 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5397                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5398                 sd->groups->cpu_power = power;
5399
5400 #ifdef CONFIG_NUMA
5401                 sd = &per_cpu(allnodes_domains, i);
5402                 if (sd->groups) {
5403                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5404                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5405                         sd->groups->cpu_power = power;
5406                 }
5407 #endif
5408         }
5409
5410 #ifdef CONFIG_NUMA
5411         for (i = 0; i < MAX_NUMNODES; i++) {
5412                 struct sched_group *sg = sched_group_nodes[i];
5413                 int j;
5414
5415                 if (sg == NULL)
5416                         continue;
5417 next_sg:
5418                 for_each_cpu_mask(j, sg->cpumask) {
5419                         struct sched_domain *sd;
5420                         int power;
5421
5422                         sd = &per_cpu(phys_domains, j);
5423                         if (j != first_cpu(sd->groups->cpumask)) {
5424                                 /*
5425                                  * Only add "power" once for each
5426                                  * physical package.
5427                                  */
5428                                 continue;
5429                         }
5430                         power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5431                                 (cpus_weight(sd->groups->cpumask)-1) / 10;
5432
5433                         sg->cpu_power += power;
5434                 }
5435                 sg = sg->next;
5436                 if (sg != sched_group_nodes[i])
5437                         goto next_sg;
5438         }
5439 #endif
5440
5441         /* Attach the domains */
5442         for_each_cpu_mask(i, *cpu_map) {
5443                 struct sched_domain *sd;
5444 #ifdef CONFIG_SCHED_SMT
5445                 sd = &per_cpu(cpu_domains, i);
5446 #else
5447                 sd = &per_cpu(phys_domains, i);
5448 #endif
5449                 cpu_attach_domain(sd, i);
5450         }
5451 }
5452 /*
5453  * Set up scheduler domains and groups.  Callers must hold the hotplug lock.
5454  */
5455 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5456 {
5457         cpumask_t cpu_default_map;
5458
5459         /*
5460          * Setup mask for cpus without special case scheduling requirements.
5461          * For now this just excludes isolated cpus, but could be used to
5462          * exclude other special cases in the future.
5463          */
5464         cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5465
5466         build_sched_domains(&cpu_default_map);
5467 }
5468
5469 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5470 {
5471 #ifdef CONFIG_NUMA
5472         int i;
5473         int cpu;
5474
5475         for_each_cpu_mask(cpu, *cpu_map) {
5476                 struct sched_group *sched_group_allnodes
5477                         = sched_group_allnodes_bycpu[cpu];
5478                 struct sched_group **sched_group_nodes
5479                         = sched_group_nodes_bycpu[cpu];
5480
5481                 if (sched_group_allnodes) {
5482                         kfree(sched_group_allnodes);
5483                         sched_group_allnodes_bycpu[cpu] = NULL;
5484                 }
5485
5486                 if (!sched_group_nodes)
5487                         continue;
5488
5489                 for (i = 0; i < MAX_NUMNODES; i++) {
5490                         cpumask_t nodemask = node_to_cpumask(i);
5491                         struct sched_group *oldsg, *sg = sched_group_nodes[i];
5492
5493                         cpus_and(nodemask, nodemask, *cpu_map);
5494                         if (cpus_empty(nodemask))
5495                                 continue;
5496
5497                         if (sg == NULL)
5498                                 continue;
5499                         sg = sg->next;
5500 next_sg:
5501                         oldsg = sg;
5502                         sg = sg->next;
5503                         kfree(oldsg);
5504                         if (oldsg != sched_group_nodes[i])
5505                                 goto next_sg;
5506                 }
5507                 kfree(sched_group_nodes);
5508                 sched_group_nodes_bycpu[cpu] = NULL;
5509         }
5510 #endif
5511 }
5512
5513 /*
5514  * Detach sched domains from a group of cpus specified in cpu_map
5515  * These cpus will now be attached to the NULL domain
5516  */
5517 static inline void detach_destroy_domains(const cpumask_t *cpu_map)
5518 {
5519         int i;
5520
5521         for_each_cpu_mask(i, *cpu_map)
5522                 cpu_attach_domain(NULL, i);
5523         synchronize_sched();
5524         arch_destroy_sched_domains(cpu_map);
5525 }
5526
5527 /*
5528  * Partition sched domains as specified by the cpumasks below.
5529  * This attaches all cpus from the cpumasks to the NULL domain,
5530  * waits for a RCU quiescent period, recalculates sched
5531  * domain information and then attaches them back to the
5532  * correct sched domains
5533  * Call with hotplug lock held
5534  */
5535 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
5536 {
5537         cpumask_t change_map;
5538
5539         cpus_and(*partition1, *partition1, cpu_online_map);
5540         cpus_and(*partition2, *partition2, cpu_online_map);
5541         cpus_or(change_map, *partition1, *partition2);
5542
5543         /* Detach sched domains from all of the affected cpus */
5544         detach_destroy_domains(&change_map);
5545         if (!cpus_empty(*partition1))
5546                 build_sched_domains(partition1);
5547         if (!cpus_empty(*partition2))
5548                 build_sched_domains(partition2);
5549 }
5550
5551 #ifdef CONFIG_HOTPLUG_CPU
5552 /*
5553  * Force a reinitialization of the sched domains hierarchy.  The domains
5554  * and groups cannot be updated in place without racing with the balancing
5555  * code, so we temporarily attach all running cpus to the NULL domain
5556  * which will prevent rebalancing while the sched domains are recalculated.
5557  */
5558 static int update_sched_domains(struct notifier_block *nfb,
5559                                 unsigned long action, void *hcpu)
5560 {
5561         switch (action) {
5562         case CPU_UP_PREPARE:
5563         case CPU_DOWN_PREPARE:
5564                 detach_destroy_domains(&cpu_online_map);
5565                 return NOTIFY_OK;
5566
5567         case CPU_UP_CANCELED:
5568         case CPU_DOWN_FAILED:
5569         case CPU_ONLINE:
5570         case CPU_DEAD:
5571                 /*
5572                  * Fall through and re-initialise the domains.
5573                  */
5574                 break;
5575         default:
5576                 return NOTIFY_DONE;
5577         }
5578
5579         /* The hotplug lock is already held by cpu_up/cpu_down */
5580         arch_init_sched_domains(&cpu_online_map);
5581
5582         return NOTIFY_OK;
5583 }
5584 #endif
5585
5586 void __init sched_init_smp(void)
5587 {
5588         lock_cpu_hotplug();
5589         arch_init_sched_domains(&cpu_online_map);
5590         unlock_cpu_hotplug();
5591         /* XXX: Theoretical race here - CPU may be hotplugged now */
5592         hotcpu_notifier(update_sched_domains, 0);
5593 }
5594 #else
5595 void __init sched_init_smp(void)
5596 {
5597 }
5598 #endif /* CONFIG_SMP */
5599
5600 int in_sched_functions(unsigned long addr)
5601 {
5602         /* Linker adds these: start and end of __sched functions */
5603         extern char __sched_text_start[], __sched_text_end[];
5604         return in_lock_functions(addr) ||
5605                 (addr >= (unsigned long)__sched_text_start
5606                 && addr < (unsigned long)__sched_text_end);
5607 }
5608
5609 void __init sched_init(void)
5610 {
5611         runqueue_t *rq;
5612         int i, j, k;
5613
5614         for (i = 0; i < NR_CPUS; i++) {
5615                 prio_array_t *array;
5616
5617                 rq = cpu_rq(i);
5618                 spin_lock_init(&rq->lock);
5619                 rq->nr_running = 0;
5620                 rq->active = rq->arrays;
5621                 rq->expired = rq->arrays + 1;
5622                 rq->best_expired_prio = MAX_PRIO;
5623
5624 #ifdef CONFIG_SMP
5625                 rq->sd = NULL;
5626                 for (j = 1; j < 3; j++)
5627                         rq->cpu_load[j] = 0;
5628                 rq->active_balance = 0;
5629                 rq->push_cpu = 0;
5630                 rq->migration_thread = NULL;
5631                 INIT_LIST_HEAD(&rq->migration_queue);
5632 #endif
5633                 atomic_set(&rq->nr_iowait, 0);
5634
5635                 for (j = 0; j < 2; j++) {
5636                         array = rq->arrays + j;
5637                         for (k = 0; k < MAX_PRIO; k++) {
5638                                 INIT_LIST_HEAD(array->queue + k);
5639                                 __clear_bit(k, array->bitmap);
5640                         }
5641                         // delimiter for bitsearch
5642                         __set_bit(MAX_PRIO, array->bitmap);
5643                 }
5644         }
5645
5646         /*
5647          * The boot idle thread does lazy MMU switching as well:
5648          */
5649         atomic_inc(&init_mm.mm_count);
5650         enter_lazy_tlb(&init_mm, current);
5651
5652         /*
5653          * Make us the idle thread. Technically, schedule() should not be
5654          * called from this thread, however somewhere below it might be,
5655          * but because we are the idle thread, we just pick up running again
5656          * when this runqueue becomes "idle".
5657          */
5658         init_idle(current, smp_processor_id());
5659 }
5660
5661 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5662 void __might_sleep(char *file, int line)
5663 {
5664 #if defined(in_atomic)
5665         static unsigned long prev_jiffy;        /* ratelimiting */
5666
5667         if ((in_atomic() || irqs_disabled()) &&
5668             system_state == SYSTEM_RUNNING && !oops_in_progress) {
5669                 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5670                         return;
5671                 prev_jiffy = jiffies;
5672                 printk(KERN_ERR "Debug: sleeping function called from invalid"
5673                                 " context at %s:%d\n", file, line);
5674                 printk("in_atomic():%d, irqs_disabled():%d\n",
5675                         in_atomic(), irqs_disabled());
5676                 dump_stack();
5677         }
5678 #endif
5679 }
5680 EXPORT_SYMBOL(__might_sleep);
5681 #endif
5682
5683 #ifdef CONFIG_MAGIC_SYSRQ
5684 void normalize_rt_tasks(void)
5685 {
5686         struct task_struct *p;
5687         prio_array_t *array;
5688         unsigned long flags;
5689         runqueue_t *rq;
5690
5691         read_lock_irq(&tasklist_lock);
5692         for_each_process (p) {
5693                 if (!rt_task(p))
5694                         continue;
5695
5696                 rq = task_rq_lock(p, &flags);
5697
5698                 array = p->array;
5699                 if (array)
5700                         deactivate_task(p, task_rq(p));
5701                 __setscheduler(p, SCHED_NORMAL, 0);
5702                 if (array) {
5703                         __activate_task(p, task_rq(p));
5704                         resched_task(rq->curr);
5705                 }
5706
5707                 task_rq_unlock(rq, &flags);
5708         }
5709         read_unlock_irq(&tasklist_lock);
5710 }
5711
5712 #endif /* CONFIG_MAGIC_SYSRQ */
5713
5714 #ifdef CONFIG_IA64
5715 /*
5716  * These functions are only useful for the IA64 MCA handling.
5717  *
5718  * They can only be called when the whole system has been
5719  * stopped - every CPU needs to be quiescent, and no scheduling
5720  * activity can take place. Using them for anything else would
5721  * be a serious bug, and as a result, they aren't even visible
5722  * under any other configuration.
5723  */
5724
5725 /**
5726  * curr_task - return the current task for a given cpu.
5727  * @cpu: the processor in question.
5728  *
5729  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5730  */
5731 task_t *curr_task(int cpu)
5732 {
5733         return cpu_curr(cpu);
5734 }
5735
5736 /**
5737  * set_curr_task - set the current task for a given cpu.
5738  * @cpu: the processor in question.
5739  * @p: the task pointer to set.
5740  *
5741  * Description: This function must only be used when non-maskable interrupts
5742  * are serviced on a separate stack.  It allows the architecture to switch the
5743  * notion of the current task on a cpu in a non-blocking manner.  This function
5744  * must be called with all CPU's synchronized, and interrupts disabled, the
5745  * and caller must save the original value of the current task (see
5746  * curr_task() above) and restore that value before reenabling interrupts and
5747  * re-starting the system.
5748  *
5749  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5750  */
5751 void set_curr_task(int cpu, task_t *p)
5752 {
5753         cpu_curr(cpu) = p;
5754 }
5755
5756 #endif