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