f41fa94d2070b5f525f59775dff0fd9b67e567f1
[sfrench/cifs-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  * @rq: runqueue associated with task-switch
1482  * @prev: the thread we just switched away from.
1483  *
1484  * finish_task_switch must be called after the context switch, paired
1485  * with a prepare_task_switch call before the context switch.
1486  * finish_task_switch will reconcile locking set up by prepare_task_switch,
1487  * and do any other architecture-specific cleanup actions.
1488  *
1489  * Note that we may have delayed dropping an mm in context_switch(). If
1490  * so, we finish that here outside of the runqueue lock.  (Doing it
1491  * with the lock held can cause deadlocks; see schedule() for
1492  * details.)
1493  */
1494 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1495         __releases(rq->lock)
1496 {
1497         struct mm_struct *mm = rq->prev_mm;
1498         unsigned long prev_task_flags;
1499
1500         rq->prev_mm = NULL;
1501
1502         /*
1503          * A task struct has one reference for the use as "current".
1504          * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1505          * calls schedule one last time. The schedule call will never return,
1506          * and the scheduled task must drop that reference.
1507          * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1508          * still held, otherwise prev could be scheduled on another cpu, die
1509          * there before we look at prev->state, and then the reference would
1510          * be dropped twice.
1511          *              Manfred Spraul <manfred@colorfullife.com>
1512          */
1513         prev_task_flags = prev->flags;
1514         finish_arch_switch(prev);
1515         finish_lock_switch(rq, prev);
1516         if (mm)
1517                 mmdrop(mm);
1518         if (unlikely(prev_task_flags & PF_DEAD))
1519                 put_task_struct(prev);
1520 }
1521
1522 /**
1523  * schedule_tail - first thing a freshly forked thread must call.
1524  * @prev: the thread we just switched away from.
1525  */
1526 asmlinkage void schedule_tail(task_t *prev)
1527         __releases(rq->lock)
1528 {
1529         runqueue_t *rq = this_rq();
1530         finish_task_switch(rq, prev);
1531 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1532         /* In this case, finish_task_switch does not reenable preemption */
1533         preempt_enable();
1534 #endif
1535         if (current->set_child_tid)
1536                 put_user(current->pid, current->set_child_tid);
1537 }
1538
1539 /*
1540  * context_switch - switch to the new MM and the new
1541  * thread's register state.
1542  */
1543 static inline
1544 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1545 {
1546         struct mm_struct *mm = next->mm;
1547         struct mm_struct *oldmm = prev->active_mm;
1548
1549         if (unlikely(!mm)) {
1550                 next->active_mm = oldmm;
1551                 atomic_inc(&oldmm->mm_count);
1552                 enter_lazy_tlb(oldmm, next);
1553         } else
1554                 switch_mm(oldmm, mm, next);
1555
1556         if (unlikely(!prev->mm)) {
1557                 prev->active_mm = NULL;
1558                 WARN_ON(rq->prev_mm);
1559                 rq->prev_mm = oldmm;
1560         }
1561
1562         /* Here we just switch the register state and the stack. */
1563         switch_to(prev, next, prev);
1564
1565         return prev;
1566 }
1567
1568 /*
1569  * nr_running, nr_uninterruptible and nr_context_switches:
1570  *
1571  * externally visible scheduler statistics: current number of runnable
1572  * threads, current number of uninterruptible-sleeping threads, total
1573  * number of context switches performed since bootup.
1574  */
1575 unsigned long nr_running(void)
1576 {
1577         unsigned long i, sum = 0;
1578
1579         for_each_online_cpu(i)
1580                 sum += cpu_rq(i)->nr_running;
1581
1582         return sum;
1583 }
1584
1585 unsigned long nr_uninterruptible(void)
1586 {
1587         unsigned long i, sum = 0;
1588
1589         for_each_cpu(i)
1590                 sum += cpu_rq(i)->nr_uninterruptible;
1591
1592         /*
1593          * Since we read the counters lockless, it might be slightly
1594          * inaccurate. Do not allow it to go below zero though:
1595          */
1596         if (unlikely((long)sum < 0))
1597                 sum = 0;
1598
1599         return sum;
1600 }
1601
1602 unsigned long long nr_context_switches(void)
1603 {
1604         unsigned long long i, sum = 0;
1605
1606         for_each_cpu(i)
1607                 sum += cpu_rq(i)->nr_switches;
1608
1609         return sum;
1610 }
1611
1612 unsigned long nr_iowait(void)
1613 {
1614         unsigned long i, sum = 0;
1615
1616         for_each_cpu(i)
1617                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1618
1619         return sum;
1620 }
1621
1622 #ifdef CONFIG_SMP
1623
1624 /*
1625  * double_rq_lock - safely lock two runqueues
1626  *
1627  * Note this does not disable interrupts like task_rq_lock,
1628  * you need to do so manually before calling.
1629  */
1630 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1631         __acquires(rq1->lock)
1632         __acquires(rq2->lock)
1633 {
1634         if (rq1 == rq2) {
1635                 spin_lock(&rq1->lock);
1636                 __acquire(rq2->lock);   /* Fake it out ;) */
1637         } else {
1638                 if (rq1 < rq2) {
1639                         spin_lock(&rq1->lock);
1640                         spin_lock(&rq2->lock);
1641                 } else {
1642                         spin_lock(&rq2->lock);
1643                         spin_lock(&rq1->lock);
1644                 }
1645         }
1646 }
1647
1648 /*
1649  * double_rq_unlock - safely unlock two runqueues
1650  *
1651  * Note this does not restore interrupts like task_rq_unlock,
1652  * you need to do so manually after calling.
1653  */
1654 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1655         __releases(rq1->lock)
1656         __releases(rq2->lock)
1657 {
1658         spin_unlock(&rq1->lock);
1659         if (rq1 != rq2)
1660                 spin_unlock(&rq2->lock);
1661         else
1662                 __release(rq2->lock);
1663 }
1664
1665 /*
1666  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1667  */
1668 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1669         __releases(this_rq->lock)
1670         __acquires(busiest->lock)
1671         __acquires(this_rq->lock)
1672 {
1673         if (unlikely(!spin_trylock(&busiest->lock))) {
1674                 if (busiest < this_rq) {
1675                         spin_unlock(&this_rq->lock);
1676                         spin_lock(&busiest->lock);
1677                         spin_lock(&this_rq->lock);
1678                 } else
1679                         spin_lock(&busiest->lock);
1680         }
1681 }
1682
1683 /*
1684  * If dest_cpu is allowed for this process, migrate the task to it.
1685  * This is accomplished by forcing the cpu_allowed mask to only
1686  * allow dest_cpu, which will force the cpu onto dest_cpu.  Then
1687  * the cpu_allowed mask is restored.
1688  */
1689 static void sched_migrate_task(task_t *p, int dest_cpu)
1690 {
1691         migration_req_t req;
1692         runqueue_t *rq;
1693         unsigned long flags;
1694
1695         rq = task_rq_lock(p, &flags);
1696         if (!cpu_isset(dest_cpu, p->cpus_allowed)
1697             || unlikely(cpu_is_offline(dest_cpu)))
1698                 goto out;
1699
1700         /* force the process onto the specified CPU */
1701         if (migrate_task(p, dest_cpu, &req)) {
1702                 /* Need to wait for migration thread (might exit: take ref). */
1703                 struct task_struct *mt = rq->migration_thread;
1704                 get_task_struct(mt);
1705                 task_rq_unlock(rq, &flags);
1706                 wake_up_process(mt);
1707                 put_task_struct(mt);
1708                 wait_for_completion(&req.done);
1709                 return;
1710         }
1711 out:
1712         task_rq_unlock(rq, &flags);
1713 }
1714
1715 /*
1716  * sched_exec - execve() is a valuable balancing opportunity, because at
1717  * this point the task has the smallest effective memory and cache footprint.
1718  */
1719 void sched_exec(void)
1720 {
1721         int new_cpu, this_cpu = get_cpu();
1722         new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1723         put_cpu();
1724         if (new_cpu != this_cpu)
1725                 sched_migrate_task(current, new_cpu);
1726 }
1727
1728 /*
1729  * pull_task - move a task from a remote runqueue to the local runqueue.
1730  * Both runqueues must be locked.
1731  */
1732 static inline
1733 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1734                runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1735 {
1736         dequeue_task(p, src_array);
1737         src_rq->nr_running--;
1738         set_task_cpu(p, this_cpu);
1739         this_rq->nr_running++;
1740         enqueue_task(p, this_array);
1741         p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1742                                 + this_rq->timestamp_last_tick;
1743         /*
1744          * Note that idle threads have a prio of MAX_PRIO, for this test
1745          * to be always true for them.
1746          */
1747         if (TASK_PREEMPTS_CURR(p, this_rq))
1748                 resched_task(this_rq->curr);
1749 }
1750
1751 /*
1752  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1753  */
1754 static inline
1755 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1756              struct sched_domain *sd, enum idle_type idle, int *all_pinned)
1757 {
1758         /*
1759          * We do not migrate tasks that are:
1760          * 1) running (obviously), or
1761          * 2) cannot be migrated to this CPU due to cpus_allowed, or
1762          * 3) are cache-hot on their current CPU.
1763          */
1764         if (!cpu_isset(this_cpu, p->cpus_allowed))
1765                 return 0;
1766         *all_pinned = 0;
1767
1768         if (task_running(rq, p))
1769                 return 0;
1770
1771         /*
1772          * Aggressive migration if:
1773          * 1) task is cache cold, or
1774          * 2) too many balance attempts have failed.
1775          */
1776
1777         if (sd->nr_balance_failed > sd->cache_nice_tries)
1778                 return 1;
1779
1780         if (task_hot(p, rq->timestamp_last_tick, sd))
1781                 return 0;
1782         return 1;
1783 }
1784
1785 /*
1786  * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1787  * as part of a balancing operation within "domain". Returns the number of
1788  * tasks moved.
1789  *
1790  * Called with both runqueues locked.
1791  */
1792 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1793                       unsigned long max_nr_move, struct sched_domain *sd,
1794                       enum idle_type idle, int *all_pinned)
1795 {
1796         prio_array_t *array, *dst_array;
1797         struct list_head *head, *curr;
1798         int idx, pulled = 0, pinned = 0;
1799         task_t *tmp;
1800
1801         if (max_nr_move == 0)
1802                 goto out;
1803
1804         pinned = 1;
1805
1806         /*
1807          * We first consider expired tasks. Those will likely not be
1808          * executed in the near future, and they are most likely to
1809          * be cache-cold, thus switching CPUs has the least effect
1810          * on them.
1811          */
1812         if (busiest->expired->nr_active) {
1813                 array = busiest->expired;
1814                 dst_array = this_rq->expired;
1815         } else {
1816                 array = busiest->active;
1817                 dst_array = this_rq->active;
1818         }
1819
1820 new_array:
1821         /* Start searching at priority 0: */
1822         idx = 0;
1823 skip_bitmap:
1824         if (!idx)
1825                 idx = sched_find_first_bit(array->bitmap);
1826         else
1827                 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1828         if (idx >= MAX_PRIO) {
1829                 if (array == busiest->expired && busiest->active->nr_active) {
1830                         array = busiest->active;
1831                         dst_array = this_rq->active;
1832                         goto new_array;
1833                 }
1834                 goto out;
1835         }
1836
1837         head = array->queue + idx;
1838         curr = head->prev;
1839 skip_queue:
1840         tmp = list_entry(curr, task_t, run_list);
1841
1842         curr = curr->prev;
1843
1844         if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1845                 if (curr != head)
1846                         goto skip_queue;
1847                 idx++;
1848                 goto skip_bitmap;
1849         }
1850
1851 #ifdef CONFIG_SCHEDSTATS
1852         if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1853                 schedstat_inc(sd, lb_hot_gained[idle]);
1854 #endif
1855
1856         pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1857         pulled++;
1858
1859         /* We only want to steal up to the prescribed number of tasks. */
1860         if (pulled < max_nr_move) {
1861                 if (curr != head)
1862                         goto skip_queue;
1863                 idx++;
1864                 goto skip_bitmap;
1865         }
1866 out:
1867         /*
1868          * Right now, this is the only place pull_task() is called,
1869          * so we can safely collect pull_task() stats here rather than
1870          * inside pull_task().
1871          */
1872         schedstat_add(sd, lb_gained[idle], pulled);
1873
1874         if (all_pinned)
1875                 *all_pinned = pinned;
1876         return pulled;
1877 }
1878
1879 /*
1880  * find_busiest_group finds and returns the busiest CPU group within the
1881  * domain. It calculates and returns the number of tasks which should be
1882  * moved to restore balance via the imbalance parameter.
1883  */
1884 static struct sched_group *
1885 find_busiest_group(struct sched_domain *sd, int this_cpu,
1886                    unsigned long *imbalance, enum idle_type idle)
1887 {
1888         struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1889         unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1890         int load_idx;
1891
1892         max_load = this_load = total_load = total_pwr = 0;
1893         if (idle == NOT_IDLE)
1894                 load_idx = sd->busy_idx;
1895         else if (idle == NEWLY_IDLE)
1896                 load_idx = sd->newidle_idx;
1897         else
1898                 load_idx = sd->idle_idx;
1899
1900         do {
1901                 unsigned long load;
1902                 int local_group;
1903                 int i;
1904
1905                 local_group = cpu_isset(this_cpu, group->cpumask);
1906
1907                 /* Tally up the load of all CPUs in the group */
1908                 avg_load = 0;
1909
1910                 for_each_cpu_mask(i, group->cpumask) {
1911                         /* Bias balancing toward cpus of our domain */
1912                         if (local_group)
1913                                 load = target_load(i, load_idx);
1914                         else
1915                                 load = source_load(i, load_idx);
1916
1917                         avg_load += load;
1918                 }
1919
1920                 total_load += avg_load;
1921                 total_pwr += group->cpu_power;
1922
1923                 /* Adjust by relative CPU power of the group */
1924                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1925
1926                 if (local_group) {
1927                         this_load = avg_load;
1928                         this = group;
1929                 } else if (avg_load > max_load) {
1930                         max_load = avg_load;
1931                         busiest = group;
1932                 }
1933                 group = group->next;
1934         } while (group != sd->groups);
1935
1936         if (!busiest || this_load >= max_load)
1937                 goto out_balanced;
1938
1939         avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1940
1941         if (this_load >= avg_load ||
1942                         100*max_load <= sd->imbalance_pct*this_load)
1943                 goto out_balanced;
1944
1945         /*
1946          * We're trying to get all the cpus to the average_load, so we don't
1947          * want to push ourselves above the average load, nor do we wish to
1948          * reduce the max loaded cpu below the average load, as either of these
1949          * actions would just result in more rebalancing later, and ping-pong
1950          * tasks around. Thus we look for the minimum possible imbalance.
1951          * Negative imbalances (*we* are more loaded than anyone else) will
1952          * be counted as no imbalance for these purposes -- we can't fix that
1953          * by pulling tasks to us.  Be careful of negative numbers as they'll
1954          * appear as very large values with unsigned longs.
1955          */
1956         /* How much load to actually move to equalise the imbalance */
1957         *imbalance = min((max_load - avg_load) * busiest->cpu_power,
1958                                 (avg_load - this_load) * this->cpu_power)
1959                         / SCHED_LOAD_SCALE;
1960
1961         if (*imbalance < SCHED_LOAD_SCALE) {
1962                 unsigned long pwr_now = 0, pwr_move = 0;
1963                 unsigned long tmp;
1964
1965                 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1966                         *imbalance = 1;
1967                         return busiest;
1968                 }
1969
1970                 /*
1971                  * OK, we don't have enough imbalance to justify moving tasks,
1972                  * however we may be able to increase total CPU power used by
1973                  * moving them.
1974                  */
1975
1976                 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
1977                 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
1978                 pwr_now /= SCHED_LOAD_SCALE;
1979
1980                 /* Amount of load we'd subtract */
1981                 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
1982                 if (max_load > tmp)
1983                         pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
1984                                                         max_load - tmp);
1985
1986                 /* Amount of load we'd add */
1987                 if (max_load*busiest->cpu_power <
1988                                 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
1989                         tmp = max_load*busiest->cpu_power/this->cpu_power;
1990                 else
1991                         tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
1992                 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
1993                 pwr_move /= SCHED_LOAD_SCALE;
1994
1995                 /* Move if we gain throughput */
1996                 if (pwr_move <= pwr_now)
1997                         goto out_balanced;
1998
1999                 *imbalance = 1;
2000                 return busiest;
2001         }
2002
2003         /* Get rid of the scaling factor, rounding down as we divide */
2004         *imbalance = *imbalance / SCHED_LOAD_SCALE;
2005         return busiest;
2006
2007 out_balanced:
2008
2009         *imbalance = 0;
2010         return NULL;
2011 }
2012
2013 /*
2014  * find_busiest_queue - find the busiest runqueue among the cpus in group.
2015  */
2016 static runqueue_t *find_busiest_queue(struct sched_group *group)
2017 {
2018         unsigned long load, max_load = 0;
2019         runqueue_t *busiest = NULL;
2020         int i;
2021
2022         for_each_cpu_mask(i, group->cpumask) {
2023                 load = source_load(i, 0);
2024
2025                 if (load > max_load) {
2026                         max_load = load;
2027                         busiest = cpu_rq(i);
2028                 }
2029         }
2030
2031         return busiest;
2032 }
2033
2034 /*
2035  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2036  * so long as it is large enough.
2037  */
2038 #define MAX_PINNED_INTERVAL     512
2039
2040 /*
2041  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2042  * tasks if there is an imbalance.
2043  *
2044  * Called with this_rq unlocked.
2045  */
2046 static int load_balance(int this_cpu, runqueue_t *this_rq,
2047                         struct sched_domain *sd, enum idle_type idle)
2048 {
2049         struct sched_group *group;
2050         runqueue_t *busiest;
2051         unsigned long imbalance;
2052         int nr_moved, all_pinned = 0;
2053         int active_balance = 0;
2054
2055         spin_lock(&this_rq->lock);
2056         schedstat_inc(sd, lb_cnt[idle]);
2057
2058         group = find_busiest_group(sd, this_cpu, &imbalance, idle);
2059         if (!group) {
2060                 schedstat_inc(sd, lb_nobusyg[idle]);
2061                 goto out_balanced;
2062         }
2063
2064         busiest = find_busiest_queue(group);
2065         if (!busiest) {
2066                 schedstat_inc(sd, lb_nobusyq[idle]);
2067                 goto out_balanced;
2068         }
2069
2070         BUG_ON(busiest == this_rq);
2071
2072         schedstat_add(sd, lb_imbalance[idle], imbalance);
2073
2074         nr_moved = 0;
2075         if (busiest->nr_running > 1) {
2076                 /*
2077                  * Attempt to move tasks. If find_busiest_group has found
2078                  * an imbalance but busiest->nr_running <= 1, the group is
2079                  * still unbalanced. nr_moved simply stays zero, so it is
2080                  * correctly treated as an imbalance.
2081                  */
2082                 double_lock_balance(this_rq, busiest);
2083                 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2084                                                 imbalance, sd, idle,
2085                                                 &all_pinned);
2086                 spin_unlock(&busiest->lock);
2087
2088                 /* All tasks on this runqueue were pinned by CPU affinity */
2089                 if (unlikely(all_pinned))
2090                         goto out_balanced;
2091         }
2092
2093         spin_unlock(&this_rq->lock);
2094
2095         if (!nr_moved) {
2096                 schedstat_inc(sd, lb_failed[idle]);
2097                 sd->nr_balance_failed++;
2098
2099                 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2100
2101                         spin_lock(&busiest->lock);
2102                         if (!busiest->active_balance) {
2103                                 busiest->active_balance = 1;
2104                                 busiest->push_cpu = this_cpu;
2105                                 active_balance = 1;
2106                         }
2107                         spin_unlock(&busiest->lock);
2108                         if (active_balance)
2109                                 wake_up_process(busiest->migration_thread);
2110
2111                         /*
2112                          * We've kicked active balancing, reset the failure
2113                          * counter.
2114                          */
2115                         sd->nr_balance_failed = sd->cache_nice_tries+1;
2116                 }
2117         } else
2118                 sd->nr_balance_failed = 0;
2119
2120         if (likely(!active_balance)) {
2121                 /* We were unbalanced, so reset the balancing interval */
2122                 sd->balance_interval = sd->min_interval;
2123         } else {
2124                 /*
2125                  * If we've begun active balancing, start to back off. This
2126                  * case may not be covered by the all_pinned logic if there
2127                  * is only 1 task on the busy runqueue (because we don't call
2128                  * move_tasks).
2129                  */
2130                 if (sd->balance_interval < sd->max_interval)
2131                         sd->balance_interval *= 2;
2132         }
2133
2134         return nr_moved;
2135
2136 out_balanced:
2137         spin_unlock(&this_rq->lock);
2138
2139         schedstat_inc(sd, lb_balanced[idle]);
2140
2141         sd->nr_balance_failed = 0;
2142         /* tune up the balancing interval */
2143         if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2144                         (sd->balance_interval < sd->max_interval))
2145                 sd->balance_interval *= 2;
2146
2147         return 0;
2148 }
2149
2150 /*
2151  * Check this_cpu to ensure it is balanced within domain. Attempt to move
2152  * tasks if there is an imbalance.
2153  *
2154  * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2155  * this_rq is locked.
2156  */
2157 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2158                                 struct sched_domain *sd)
2159 {
2160         struct sched_group *group;
2161         runqueue_t *busiest = NULL;
2162         unsigned long imbalance;
2163         int nr_moved = 0;
2164
2165         schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2166         group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);
2167         if (!group) {
2168                 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2169                 goto out_balanced;
2170         }
2171
2172         busiest = find_busiest_queue(group);
2173         if (!busiest) {
2174                 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2175                 goto out_balanced;
2176         }
2177
2178         BUG_ON(busiest == this_rq);
2179
2180         /* Attempt to move tasks */
2181         double_lock_balance(this_rq, busiest);
2182
2183         schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2184         nr_moved = move_tasks(this_rq, this_cpu, busiest,
2185                                         imbalance, sd, NEWLY_IDLE, NULL);
2186         if (!nr_moved)
2187                 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2188         else
2189                 sd->nr_balance_failed = 0;
2190
2191         spin_unlock(&busiest->lock);
2192         return nr_moved;
2193
2194 out_balanced:
2195         schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2196         sd->nr_balance_failed = 0;
2197         return 0;
2198 }
2199
2200 /*
2201  * idle_balance is called by schedule() if this_cpu is about to become
2202  * idle. Attempts to pull tasks from other CPUs.
2203  */
2204 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2205 {
2206         struct sched_domain *sd;
2207
2208         for_each_domain(this_cpu, sd) {
2209                 if (sd->flags & SD_BALANCE_NEWIDLE) {
2210                         if (load_balance_newidle(this_cpu, this_rq, sd)) {
2211                                 /* We've pulled tasks over so stop searching */
2212                                 break;
2213                         }
2214                 }
2215         }
2216 }
2217
2218 /*
2219  * active_load_balance is run by migration threads. It pushes running tasks
2220  * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2221  * running on each physical CPU where possible, and avoids physical /
2222  * logical imbalances.
2223  *
2224  * Called with busiest_rq locked.
2225  */
2226 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2227 {
2228         struct sched_domain *sd;
2229         runqueue_t *target_rq;
2230         int target_cpu = busiest_rq->push_cpu;
2231
2232         if (busiest_rq->nr_running <= 1)
2233                 /* no task to move */
2234                 return;
2235
2236         target_rq = cpu_rq(target_cpu);
2237
2238         /*
2239          * This condition is "impossible", if it occurs
2240          * we need to fix it.  Originally reported by
2241          * Bjorn Helgaas on a 128-cpu setup.
2242          */
2243         BUG_ON(busiest_rq == target_rq);
2244
2245         /* move a task from busiest_rq to target_rq */
2246         double_lock_balance(busiest_rq, target_rq);
2247
2248         /* Search for an sd spanning us and the target CPU. */
2249         for_each_domain(target_cpu, sd)
2250                 if ((sd->flags & SD_LOAD_BALANCE) &&
2251                         cpu_isset(busiest_cpu, sd->span))
2252                                 break;
2253
2254         if (unlikely(sd == NULL))
2255                 goto out;
2256
2257         schedstat_inc(sd, alb_cnt);
2258
2259         if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2260                 schedstat_inc(sd, alb_pushed);
2261         else
2262                 schedstat_inc(sd, alb_failed);
2263 out:
2264         spin_unlock(&target_rq->lock);
2265 }
2266
2267 /*
2268  * rebalance_tick will get called every timer tick, on every CPU.
2269  *
2270  * It checks each scheduling domain to see if it is due to be balanced,
2271  * and initiates a balancing operation if so.
2272  *
2273  * Balancing parameters are set up in arch_init_sched_domains.
2274  */
2275
2276 /* Don't have all balancing operations going off at once */
2277 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2278
2279 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2280                            enum idle_type idle)
2281 {
2282         unsigned long old_load, this_load;
2283         unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2284         struct sched_domain *sd;
2285         int i;
2286
2287         this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2288         /* Update our load */
2289         for (i = 0; i < 3; i++) {
2290                 unsigned long new_load = this_load;
2291                 int scale = 1 << i;
2292                 old_load = this_rq->cpu_load[i];
2293                 /*
2294                  * Round up the averaging division if load is increasing. This
2295                  * prevents us from getting stuck on 9 if the load is 10, for
2296                  * example.
2297                  */
2298                 if (new_load > old_load)
2299                         new_load += scale-1;
2300                 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2301         }
2302
2303         for_each_domain(this_cpu, sd) {
2304                 unsigned long interval;
2305
2306                 if (!(sd->flags & SD_LOAD_BALANCE))
2307                         continue;
2308
2309                 interval = sd->balance_interval;
2310                 if (idle != SCHED_IDLE)
2311                         interval *= sd->busy_factor;
2312
2313                 /* scale ms to jiffies */
2314                 interval = msecs_to_jiffies(interval);
2315                 if (unlikely(!interval))
2316                         interval = 1;
2317
2318                 if (j - sd->last_balance >= interval) {
2319                         if (load_balance(this_cpu, this_rq, sd, idle)) {
2320                                 /* We've pulled tasks over so no longer idle */
2321                                 idle = NOT_IDLE;
2322                         }
2323                         sd->last_balance += interval;
2324                 }
2325         }
2326 }
2327 #else
2328 /*
2329  * on UP we do not need to balance between CPUs:
2330  */
2331 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2332 {
2333 }
2334 static inline void idle_balance(int cpu, runqueue_t *rq)
2335 {
2336 }
2337 #endif
2338
2339 static inline int wake_priority_sleeper(runqueue_t *rq)
2340 {
2341         int ret = 0;
2342 #ifdef CONFIG_SCHED_SMT
2343         spin_lock(&rq->lock);
2344         /*
2345          * If an SMT sibling task has been put to sleep for priority
2346          * reasons reschedule the idle task to see if it can now run.
2347          */
2348         if (rq->nr_running) {
2349                 resched_task(rq->idle);
2350                 ret = 1;
2351         }
2352         spin_unlock(&rq->lock);
2353 #endif
2354         return ret;
2355 }
2356
2357 DEFINE_PER_CPU(struct kernel_stat, kstat);
2358
2359 EXPORT_PER_CPU_SYMBOL(kstat);
2360
2361 /*
2362  * This is called on clock ticks and on context switches.
2363  * Bank in p->sched_time the ns elapsed since the last tick or switch.
2364  */
2365 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2366                                     unsigned long long now)
2367 {
2368         unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2369         p->sched_time += now - last;
2370 }
2371
2372 /*
2373  * Return current->sched_time plus any more ns on the sched_clock
2374  * that have not yet been banked.
2375  */
2376 unsigned long long current_sched_time(const task_t *tsk)
2377 {
2378         unsigned long long ns;
2379         unsigned long flags;
2380         local_irq_save(flags);
2381         ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2382         ns = tsk->sched_time + (sched_clock() - ns);
2383         local_irq_restore(flags);
2384         return ns;
2385 }
2386
2387 /*
2388  * We place interactive tasks back into the active array, if possible.
2389  *
2390  * To guarantee that this does not starve expired tasks we ignore the
2391  * interactivity of a task if the first expired task had to wait more
2392  * than a 'reasonable' amount of time. This deadline timeout is
2393  * load-dependent, as the frequency of array switched decreases with
2394  * increasing number of running tasks. We also ignore the interactivity
2395  * if a better static_prio task has expired:
2396  */
2397 #define EXPIRED_STARVING(rq) \
2398         ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2399                 (jiffies - (rq)->expired_timestamp >= \
2400                         STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2401                         ((rq)->curr->static_prio > (rq)->best_expired_prio))
2402
2403 /*
2404  * Account user cpu time to a process.
2405  * @p: the process that the cpu time gets accounted to
2406  * @hardirq_offset: the offset to subtract from hardirq_count()
2407  * @cputime: the cpu time spent in user space since the last update
2408  */
2409 void account_user_time(struct task_struct *p, cputime_t cputime)
2410 {
2411         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2412         cputime64_t tmp;
2413
2414         p->utime = cputime_add(p->utime, cputime);
2415
2416         /* Add user time to cpustat. */
2417         tmp = cputime_to_cputime64(cputime);
2418         if (TASK_NICE(p) > 0)
2419                 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2420         else
2421                 cpustat->user = cputime64_add(cpustat->user, tmp);
2422 }
2423
2424 /*
2425  * Account system cpu time to a process.
2426  * @p: the process that the cpu time gets accounted to
2427  * @hardirq_offset: the offset to subtract from hardirq_count()
2428  * @cputime: the cpu time spent in kernel space since the last update
2429  */
2430 void account_system_time(struct task_struct *p, int hardirq_offset,
2431                          cputime_t cputime)
2432 {
2433         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2434         runqueue_t *rq = this_rq();
2435         cputime64_t tmp;
2436
2437         p->stime = cputime_add(p->stime, cputime);
2438
2439         /* Add system time to cpustat. */
2440         tmp = cputime_to_cputime64(cputime);
2441         if (hardirq_count() - hardirq_offset)
2442                 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2443         else if (softirq_count())
2444                 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2445         else if (p != rq->idle)
2446                 cpustat->system = cputime64_add(cpustat->system, tmp);
2447         else if (atomic_read(&rq->nr_iowait) > 0)
2448                 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2449         else
2450                 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2451         /* Account for system time used */
2452         acct_update_integrals(p);
2453         /* Update rss highwater mark */
2454         update_mem_hiwater(p);
2455 }
2456
2457 /*
2458  * Account for involuntary wait time.
2459  * @p: the process from which the cpu time has been stolen
2460  * @steal: the cpu time spent in involuntary wait
2461  */
2462 void account_steal_time(struct task_struct *p, cputime_t steal)
2463 {
2464         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2465         cputime64_t tmp = cputime_to_cputime64(steal);
2466         runqueue_t *rq = this_rq();
2467
2468         if (p == rq->idle) {
2469                 p->stime = cputime_add(p->stime, steal);
2470                 if (atomic_read(&rq->nr_iowait) > 0)
2471                         cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2472                 else
2473                         cpustat->idle = cputime64_add(cpustat->idle, tmp);
2474         } else
2475                 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2476 }
2477
2478 /*
2479  * This function gets called by the timer code, with HZ frequency.
2480  * We call it with interrupts disabled.
2481  *
2482  * It also gets called by the fork code, when changing the parent's
2483  * timeslices.
2484  */
2485 void scheduler_tick(void)
2486 {
2487         int cpu = smp_processor_id();
2488         runqueue_t *rq = this_rq();
2489         task_t *p = current;
2490         unsigned long long now = sched_clock();
2491
2492         update_cpu_clock(p, rq, now);
2493
2494         rq->timestamp_last_tick = now;
2495
2496         if (p == rq->idle) {
2497                 if (wake_priority_sleeper(rq))
2498                         goto out;
2499                 rebalance_tick(cpu, rq, SCHED_IDLE);
2500                 return;
2501         }
2502
2503         /* Task might have expired already, but not scheduled off yet */
2504         if (p->array != rq->active) {
2505                 set_tsk_need_resched(p);
2506                 goto out;
2507         }
2508         spin_lock(&rq->lock);
2509         /*
2510          * The task was running during this tick - update the
2511          * time slice counter. Note: we do not update a thread's
2512          * priority until it either goes to sleep or uses up its
2513          * timeslice. This makes it possible for interactive tasks
2514          * to use up their timeslices at their highest priority levels.
2515          */
2516         if (rt_task(p)) {
2517                 /*
2518                  * RR tasks need a special form of timeslice management.
2519                  * FIFO tasks have no timeslices.
2520                  */
2521                 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2522                         p->time_slice = task_timeslice(p);
2523                         p->first_time_slice = 0;
2524                         set_tsk_need_resched(p);
2525
2526                         /* put it at the end of the queue: */
2527                         requeue_task(p, rq->active);
2528                 }
2529                 goto out_unlock;
2530         }
2531         if (!--p->time_slice) {
2532                 dequeue_task(p, rq->active);
2533                 set_tsk_need_resched(p);
2534                 p->prio = effective_prio(p);
2535                 p->time_slice = task_timeslice(p);
2536                 p->first_time_slice = 0;
2537
2538                 if (!rq->expired_timestamp)
2539                         rq->expired_timestamp = jiffies;
2540                 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2541                         enqueue_task(p, rq->expired);
2542                         if (p->static_prio < rq->best_expired_prio)
2543                                 rq->best_expired_prio = p->static_prio;
2544                 } else
2545                         enqueue_task(p, rq->active);
2546         } else {
2547                 /*
2548                  * Prevent a too long timeslice allowing a task to monopolize
2549                  * the CPU. We do this by splitting up the timeslice into
2550                  * smaller pieces.
2551                  *
2552                  * Note: this does not mean the task's timeslices expire or
2553                  * get lost in any way, they just might be preempted by
2554                  * another task of equal priority. (one with higher
2555                  * priority would have preempted this task already.) We
2556                  * requeue this task to the end of the list on this priority
2557                  * level, which is in essence a round-robin of tasks with
2558                  * equal priority.
2559                  *
2560                  * This only applies to tasks in the interactive
2561                  * delta range with at least TIMESLICE_GRANULARITY to requeue.
2562                  */
2563                 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2564                         p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2565                         (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2566                         (p->array == rq->active)) {
2567
2568                         requeue_task(p, rq->active);
2569                         set_tsk_need_resched(p);
2570                 }
2571         }
2572 out_unlock:
2573         spin_unlock(&rq->lock);
2574 out:
2575         rebalance_tick(cpu, rq, NOT_IDLE);
2576 }
2577
2578 #ifdef CONFIG_SCHED_SMT
2579 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2580 {
2581         struct sched_domain *tmp, *sd = NULL;
2582         cpumask_t sibling_map;
2583         int i;
2584
2585         for_each_domain(this_cpu, tmp)
2586                 if (tmp->flags & SD_SHARE_CPUPOWER)
2587                         sd = tmp;
2588
2589         if (!sd)
2590                 return;
2591
2592         /*
2593          * Unlock the current runqueue because we have to lock in
2594          * CPU order to avoid deadlocks. Caller knows that we might
2595          * unlock. We keep IRQs disabled.
2596          */
2597         spin_unlock(&this_rq->lock);
2598
2599         sibling_map = sd->span;
2600
2601         for_each_cpu_mask(i, sibling_map)
2602                 spin_lock(&cpu_rq(i)->lock);
2603         /*
2604          * We clear this CPU from the mask. This both simplifies the
2605          * inner loop and keps this_rq locked when we exit:
2606          */
2607         cpu_clear(this_cpu, sibling_map);
2608
2609         for_each_cpu_mask(i, sibling_map) {
2610                 runqueue_t *smt_rq = cpu_rq(i);
2611
2612                 /*
2613                  * If an SMT sibling task is sleeping due to priority
2614                  * reasons wake it up now.
2615                  */
2616                 if (smt_rq->curr == smt_rq->idle && smt_rq->nr_running)
2617                         resched_task(smt_rq->idle);
2618         }
2619
2620         for_each_cpu_mask(i, sibling_map)
2621                 spin_unlock(&cpu_rq(i)->lock);
2622         /*
2623          * We exit with this_cpu's rq still held and IRQs
2624          * still disabled:
2625          */
2626 }
2627
2628 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2629 {
2630         struct sched_domain *tmp, *sd = NULL;
2631         cpumask_t sibling_map;
2632         prio_array_t *array;
2633         int ret = 0, i;
2634         task_t *p;
2635
2636         for_each_domain(this_cpu, tmp)
2637                 if (tmp->flags & SD_SHARE_CPUPOWER)
2638                         sd = tmp;
2639
2640         if (!sd)
2641                 return 0;
2642
2643         /*
2644          * The same locking rules and details apply as for
2645          * wake_sleeping_dependent():
2646          */
2647         spin_unlock(&this_rq->lock);
2648         sibling_map = sd->span;
2649         for_each_cpu_mask(i, sibling_map)
2650                 spin_lock(&cpu_rq(i)->lock);
2651         cpu_clear(this_cpu, sibling_map);
2652
2653         /*
2654          * Establish next task to be run - it might have gone away because
2655          * we released the runqueue lock above:
2656          */
2657         if (!this_rq->nr_running)
2658                 goto out_unlock;
2659         array = this_rq->active;
2660         if (!array->nr_active)
2661                 array = this_rq->expired;
2662         BUG_ON(!array->nr_active);
2663
2664         p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2665                 task_t, run_list);
2666
2667         for_each_cpu_mask(i, sibling_map) {
2668                 runqueue_t *smt_rq = cpu_rq(i);
2669                 task_t *smt_curr = smt_rq->curr;
2670
2671                 /*
2672                  * If a user task with lower static priority than the
2673                  * running task on the SMT sibling is trying to schedule,
2674                  * delay it till there is proportionately less timeslice
2675                  * left of the sibling task to prevent a lower priority
2676                  * task from using an unfair proportion of the
2677                  * physical cpu's resources. -ck
2678                  */
2679                 if (((smt_curr->time_slice * (100 - sd->per_cpu_gain) / 100) >
2680                         task_timeslice(p) || rt_task(smt_curr)) &&
2681                         p->mm && smt_curr->mm && !rt_task(p))
2682                                 ret = 1;
2683
2684                 /*
2685                  * Reschedule a lower priority task on the SMT sibling,
2686                  * or wake it up if it has been put to sleep for priority
2687                  * reasons.
2688                  */
2689                 if ((((p->time_slice * (100 - sd->per_cpu_gain) / 100) >
2690                         task_timeslice(smt_curr) || rt_task(p)) &&
2691                         smt_curr->mm && p->mm && !rt_task(smt_curr)) ||
2692                         (smt_curr == smt_rq->idle && smt_rq->nr_running))
2693                                 resched_task(smt_curr);
2694         }
2695 out_unlock:
2696         for_each_cpu_mask(i, sibling_map)
2697                 spin_unlock(&cpu_rq(i)->lock);
2698         return ret;
2699 }
2700 #else
2701 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2702 {
2703 }
2704
2705 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2706 {
2707         return 0;
2708 }
2709 #endif
2710
2711 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2712
2713 void fastcall add_preempt_count(int val)
2714 {
2715         /*
2716          * Underflow?
2717          */
2718         BUG_ON((preempt_count() < 0));
2719         preempt_count() += val;
2720         /*
2721          * Spinlock count overflowing soon?
2722          */
2723         BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2724 }
2725 EXPORT_SYMBOL(add_preempt_count);
2726
2727 void fastcall sub_preempt_count(int val)
2728 {
2729         /*
2730          * Underflow?
2731          */
2732         BUG_ON(val > preempt_count());
2733         /*
2734          * Is the spinlock portion underflowing?
2735          */
2736         BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2737         preempt_count() -= val;
2738 }
2739 EXPORT_SYMBOL(sub_preempt_count);
2740
2741 #endif
2742
2743 /*
2744  * schedule() is the main scheduler function.
2745  */
2746 asmlinkage void __sched schedule(void)
2747 {
2748         long *switch_count;
2749         task_t *prev, *next;
2750         runqueue_t *rq;
2751         prio_array_t *array;
2752         struct list_head *queue;
2753         unsigned long long now;
2754         unsigned long run_time;
2755         int cpu, idx, new_prio;
2756
2757         /*
2758          * Test if we are atomic.  Since do_exit() needs to call into
2759          * schedule() atomically, we ignore that path for now.
2760          * Otherwise, whine if we are scheduling when we should not be.
2761          */
2762         if (likely(!current->exit_state)) {
2763                 if (unlikely(in_atomic())) {
2764                         printk(KERN_ERR "scheduling while atomic: "
2765                                 "%s/0x%08x/%d\n",
2766                                 current->comm, preempt_count(), current->pid);
2767                         dump_stack();
2768                 }
2769         }
2770         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2771
2772 need_resched:
2773         preempt_disable();
2774         prev = current;
2775         release_kernel_lock(prev);
2776 need_resched_nonpreemptible:
2777         rq = this_rq();
2778
2779         /*
2780          * The idle thread is not allowed to schedule!
2781          * Remove this check after it has been exercised a bit.
2782          */
2783         if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2784                 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2785                 dump_stack();
2786         }
2787
2788         schedstat_inc(rq, sched_cnt);
2789         now = sched_clock();
2790         if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2791                 run_time = now - prev->timestamp;
2792                 if (unlikely((long long)(now - prev->timestamp) < 0))
2793                         run_time = 0;
2794         } else
2795                 run_time = NS_MAX_SLEEP_AVG;
2796
2797         /*
2798          * Tasks charged proportionately less run_time at high sleep_avg to
2799          * delay them losing their interactive status
2800          */
2801         run_time /= (CURRENT_BONUS(prev) ? : 1);
2802
2803         spin_lock_irq(&rq->lock);
2804
2805         if (unlikely(prev->flags & PF_DEAD))
2806                 prev->state = EXIT_DEAD;
2807
2808         switch_count = &prev->nivcsw;
2809         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2810                 switch_count = &prev->nvcsw;
2811                 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2812                                 unlikely(signal_pending(prev))))
2813                         prev->state = TASK_RUNNING;
2814                 else {
2815                         if (prev->state == TASK_UNINTERRUPTIBLE)
2816                                 rq->nr_uninterruptible++;
2817                         deactivate_task(prev, rq);
2818                 }
2819         }
2820
2821         cpu = smp_processor_id();
2822         if (unlikely(!rq->nr_running)) {
2823 go_idle:
2824                 idle_balance(cpu, rq);
2825                 if (!rq->nr_running) {
2826                         next = rq->idle;
2827                         rq->expired_timestamp = 0;
2828                         wake_sleeping_dependent(cpu, rq);
2829                         /*
2830                          * wake_sleeping_dependent() might have released
2831                          * the runqueue, so break out if we got new
2832                          * tasks meanwhile:
2833                          */
2834                         if (!rq->nr_running)
2835                                 goto switch_tasks;
2836                 }
2837         } else {
2838                 if (dependent_sleeper(cpu, rq)) {
2839                         next = rq->idle;
2840                         goto switch_tasks;
2841                 }
2842                 /*
2843                  * dependent_sleeper() releases and reacquires the runqueue
2844                  * lock, hence go into the idle loop if the rq went
2845                  * empty meanwhile:
2846                  */
2847                 if (unlikely(!rq->nr_running))
2848                         goto go_idle;
2849         }
2850
2851         array = rq->active;
2852         if (unlikely(!array->nr_active)) {
2853                 /*
2854                  * Switch the active and expired arrays.
2855                  */
2856                 schedstat_inc(rq, sched_switch);
2857                 rq->active = rq->expired;
2858                 rq->expired = array;
2859                 array = rq->active;
2860                 rq->expired_timestamp = 0;
2861                 rq->best_expired_prio = MAX_PRIO;
2862         }
2863
2864         idx = sched_find_first_bit(array->bitmap);
2865         queue = array->queue + idx;
2866         next = list_entry(queue->next, task_t, run_list);
2867
2868         if (!rt_task(next) && next->activated > 0) {
2869                 unsigned long long delta = now - next->timestamp;
2870                 if (unlikely((long long)(now - next->timestamp) < 0))
2871                         delta = 0;
2872
2873                 if (next->activated == 1)
2874                         delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2875
2876                 array = next->array;
2877                 new_prio = recalc_task_prio(next, next->timestamp + delta);
2878
2879                 if (unlikely(next->prio != new_prio)) {
2880                         dequeue_task(next, array);
2881                         next->prio = new_prio;
2882                         enqueue_task(next, array);
2883                 } else
2884                         requeue_task(next, array);
2885         }
2886         next->activated = 0;
2887 switch_tasks:
2888         if (next == rq->idle)
2889                 schedstat_inc(rq, sched_goidle);
2890         prefetch(next);
2891         clear_tsk_need_resched(prev);
2892         rcu_qsctr_inc(task_cpu(prev));
2893
2894         update_cpu_clock(prev, rq, now);
2895
2896         prev->sleep_avg -= run_time;
2897         if ((long)prev->sleep_avg <= 0)
2898                 prev->sleep_avg = 0;
2899         prev->timestamp = prev->last_ran = now;
2900
2901         sched_info_switch(prev, next);
2902         if (likely(prev != next)) {
2903                 next->timestamp = now;
2904                 rq->nr_switches++;
2905                 rq->curr = next;
2906                 ++*switch_count;
2907
2908                 prepare_task_switch(rq, next);
2909                 prev = context_switch(rq, prev, next);
2910                 barrier();
2911                 /*
2912                  * this_rq must be evaluated again because prev may have moved
2913                  * CPUs since it called schedule(), thus the 'rq' on its stack
2914                  * frame will be invalid.
2915                  */
2916                 finish_task_switch(this_rq(), prev);
2917         } else
2918                 spin_unlock_irq(&rq->lock);
2919
2920         prev = current;
2921         if (unlikely(reacquire_kernel_lock(prev) < 0))
2922                 goto need_resched_nonpreemptible;
2923         preempt_enable_no_resched();
2924         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2925                 goto need_resched;
2926 }
2927
2928 EXPORT_SYMBOL(schedule);
2929
2930 #ifdef CONFIG_PREEMPT
2931 /*
2932  * this is is the entry point to schedule() from in-kernel preemption
2933  * off of preempt_enable.  Kernel preemptions off return from interrupt
2934  * occur there and call schedule directly.
2935  */
2936 asmlinkage void __sched preempt_schedule(void)
2937 {
2938         struct thread_info *ti = current_thread_info();
2939 #ifdef CONFIG_PREEMPT_BKL
2940         struct task_struct *task = current;
2941         int saved_lock_depth;
2942 #endif
2943         /*
2944          * If there is a non-zero preempt_count or interrupts are disabled,
2945          * we do not want to preempt the current task.  Just return..
2946          */
2947         if (unlikely(ti->preempt_count || irqs_disabled()))
2948                 return;
2949
2950 need_resched:
2951         add_preempt_count(PREEMPT_ACTIVE);
2952         /*
2953          * We keep the big kernel semaphore locked, but we
2954          * clear ->lock_depth so that schedule() doesnt
2955          * auto-release the semaphore:
2956          */
2957 #ifdef CONFIG_PREEMPT_BKL
2958         saved_lock_depth = task->lock_depth;
2959         task->lock_depth = -1;
2960 #endif
2961         schedule();
2962 #ifdef CONFIG_PREEMPT_BKL
2963         task->lock_depth = saved_lock_depth;
2964 #endif
2965         sub_preempt_count(PREEMPT_ACTIVE);
2966
2967         /* we could miss a preemption opportunity between schedule and now */
2968         barrier();
2969         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2970                 goto need_resched;
2971 }
2972
2973 EXPORT_SYMBOL(preempt_schedule);
2974
2975 /*
2976  * this is is the entry point to schedule() from kernel preemption
2977  * off of irq context.
2978  * Note, that this is called and return with irqs disabled. This will
2979  * protect us against recursive calling from irq.
2980  */
2981 asmlinkage void __sched preempt_schedule_irq(void)
2982 {
2983         struct thread_info *ti = current_thread_info();
2984 #ifdef CONFIG_PREEMPT_BKL
2985         struct task_struct *task = current;
2986         int saved_lock_depth;
2987 #endif
2988         /* Catch callers which need to be fixed*/
2989         BUG_ON(ti->preempt_count || !irqs_disabled());
2990
2991 need_resched:
2992         add_preempt_count(PREEMPT_ACTIVE);
2993         /*
2994          * We keep the big kernel semaphore locked, but we
2995          * clear ->lock_depth so that schedule() doesnt
2996          * auto-release the semaphore:
2997          */
2998 #ifdef CONFIG_PREEMPT_BKL
2999         saved_lock_depth = task->lock_depth;
3000         task->lock_depth = -1;
3001 #endif
3002         local_irq_enable();
3003         schedule();
3004         local_irq_disable();
3005 #ifdef CONFIG_PREEMPT_BKL
3006         task->lock_depth = saved_lock_depth;
3007 #endif
3008         sub_preempt_count(PREEMPT_ACTIVE);
3009
3010         /* we could miss a preemption opportunity between schedule and now */
3011         barrier();
3012         if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3013                 goto need_resched;
3014 }
3015
3016 #endif /* CONFIG_PREEMPT */
3017
3018 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, void *key)
3019 {
3020         task_t *p = curr->private;
3021         return try_to_wake_up(p, mode, sync);
3022 }
3023
3024 EXPORT_SYMBOL(default_wake_function);
3025
3026 /*
3027  * The core wakeup function.  Non-exclusive wakeups (nr_exclusive == 0) just
3028  * wake everything up.  If it's an exclusive wakeup (nr_exclusive == small +ve
3029  * number) then we wake all the non-exclusive tasks and one exclusive task.
3030  *
3031  * There are circumstances in which we can try to wake a task which has already
3032  * started to run but is not in state TASK_RUNNING.  try_to_wake_up() returns
3033  * zero in this (rare) case, and we handle it by continuing to scan the queue.
3034  */
3035 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3036                              int nr_exclusive, int sync, void *key)
3037 {
3038         struct list_head *tmp, *next;
3039
3040         list_for_each_safe(tmp, next, &q->task_list) {
3041                 wait_queue_t *curr;
3042                 unsigned flags;
3043                 curr = list_entry(tmp, wait_queue_t, task_list);
3044                 flags = curr->flags;
3045                 if (curr->func(curr, mode, sync, key) &&
3046                     (flags & WQ_FLAG_EXCLUSIVE) &&
3047                     !--nr_exclusive)
3048                         break;
3049         }
3050 }
3051
3052 /**
3053  * __wake_up - wake up threads blocked on a waitqueue.
3054  * @q: the waitqueue
3055  * @mode: which threads
3056  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3057  * @key: is directly passed to the wakeup function
3058  */
3059 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3060                                 int nr_exclusive, void *key)
3061 {
3062         unsigned long flags;
3063
3064         spin_lock_irqsave(&q->lock, flags);
3065         __wake_up_common(q, mode, nr_exclusive, 0, key);
3066         spin_unlock_irqrestore(&q->lock, flags);
3067 }
3068
3069 EXPORT_SYMBOL(__wake_up);
3070
3071 /*
3072  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3073  */
3074 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3075 {
3076         __wake_up_common(q, mode, 1, 0, NULL);
3077 }
3078
3079 /**
3080  * __wake_up_sync - wake up threads blocked on a waitqueue.
3081  * @q: the waitqueue
3082  * @mode: which threads
3083  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3084  *
3085  * The sync wakeup differs that the waker knows that it will schedule
3086  * away soon, so while the target thread will be woken up, it will not
3087  * be migrated to another CPU - ie. the two threads are 'synchronized'
3088  * with each other. This can prevent needless bouncing between CPUs.
3089  *
3090  * On UP it can prevent extra preemption.
3091  */
3092 void fastcall __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3093 {
3094         unsigned long flags;
3095         int sync = 1;
3096
3097         if (unlikely(!q))
3098                 return;
3099
3100         if (unlikely(!nr_exclusive))
3101                 sync = 0;
3102
3103         spin_lock_irqsave(&q->lock, flags);
3104         __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3105         spin_unlock_irqrestore(&q->lock, flags);
3106 }
3107 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3108
3109 void fastcall complete(struct completion *x)
3110 {
3111         unsigned long flags;
3112
3113         spin_lock_irqsave(&x->wait.lock, flags);
3114         x->done++;
3115         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3116                          1, 0, NULL);
3117         spin_unlock_irqrestore(&x->wait.lock, flags);
3118 }
3119 EXPORT_SYMBOL(complete);
3120
3121 void fastcall complete_all(struct completion *x)
3122 {
3123         unsigned long flags;
3124
3125         spin_lock_irqsave(&x->wait.lock, flags);
3126         x->done += UINT_MAX/2;
3127         __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3128                          0, 0, NULL);
3129         spin_unlock_irqrestore(&x->wait.lock, flags);
3130 }
3131 EXPORT_SYMBOL(complete_all);
3132
3133 void fastcall __sched wait_for_completion(struct completion *x)
3134 {
3135         might_sleep();
3136         spin_lock_irq(&x->wait.lock);
3137         if (!x->done) {
3138                 DECLARE_WAITQUEUE(wait, current);
3139
3140                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3141                 __add_wait_queue_tail(&x->wait, &wait);
3142                 do {
3143                         __set_current_state(TASK_UNINTERRUPTIBLE);
3144                         spin_unlock_irq(&x->wait.lock);
3145                         schedule();
3146                         spin_lock_irq(&x->wait.lock);
3147                 } while (!x->done);
3148                 __remove_wait_queue(&x->wait, &wait);
3149         }
3150         x->done--;
3151         spin_unlock_irq(&x->wait.lock);
3152 }
3153 EXPORT_SYMBOL(wait_for_completion);
3154
3155 unsigned long fastcall __sched
3156 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3157 {
3158         might_sleep();
3159
3160         spin_lock_irq(&x->wait.lock);
3161         if (!x->done) {
3162                 DECLARE_WAITQUEUE(wait, current);
3163
3164                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3165                 __add_wait_queue_tail(&x->wait, &wait);
3166                 do {
3167                         __set_current_state(TASK_UNINTERRUPTIBLE);
3168                         spin_unlock_irq(&x->wait.lock);
3169                         timeout = schedule_timeout(timeout);
3170                         spin_lock_irq(&x->wait.lock);
3171                         if (!timeout) {
3172                                 __remove_wait_queue(&x->wait, &wait);
3173                                 goto out;
3174                         }
3175                 } while (!x->done);
3176                 __remove_wait_queue(&x->wait, &wait);
3177         }
3178         x->done--;
3179 out:
3180         spin_unlock_irq(&x->wait.lock);
3181         return timeout;
3182 }
3183 EXPORT_SYMBOL(wait_for_completion_timeout);
3184
3185 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3186 {
3187         int ret = 0;
3188
3189         might_sleep();
3190
3191         spin_lock_irq(&x->wait.lock);
3192         if (!x->done) {
3193                 DECLARE_WAITQUEUE(wait, current);
3194
3195                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3196                 __add_wait_queue_tail(&x->wait, &wait);
3197                 do {
3198                         if (signal_pending(current)) {
3199                                 ret = -ERESTARTSYS;
3200                                 __remove_wait_queue(&x->wait, &wait);
3201                                 goto out;
3202                         }
3203                         __set_current_state(TASK_INTERRUPTIBLE);
3204                         spin_unlock_irq(&x->wait.lock);
3205                         schedule();
3206                         spin_lock_irq(&x->wait.lock);
3207                 } while (!x->done);
3208                 __remove_wait_queue(&x->wait, &wait);
3209         }
3210         x->done--;
3211 out:
3212         spin_unlock_irq(&x->wait.lock);
3213
3214         return ret;
3215 }
3216 EXPORT_SYMBOL(wait_for_completion_interruptible);
3217
3218 unsigned long fastcall __sched
3219 wait_for_completion_interruptible_timeout(struct completion *x,
3220                                           unsigned long timeout)
3221 {
3222         might_sleep();
3223
3224         spin_lock_irq(&x->wait.lock);
3225         if (!x->done) {
3226                 DECLARE_WAITQUEUE(wait, current);
3227
3228                 wait.flags |= WQ_FLAG_EXCLUSIVE;
3229                 __add_wait_queue_tail(&x->wait, &wait);
3230                 do {
3231                         if (signal_pending(current)) {
3232                                 timeout = -ERESTARTSYS;
3233                                 __remove_wait_queue(&x->wait, &wait);
3234                                 goto out;
3235                         }
3236                         __set_current_state(TASK_INTERRUPTIBLE);
3237                         spin_unlock_irq(&x->wait.lock);
3238                         timeout = schedule_timeout(timeout);
3239                         spin_lock_irq(&x->wait.lock);
3240                         if (!timeout) {
3241                                 __remove_wait_queue(&x->wait, &wait);
3242                                 goto out;
3243                         }
3244                 } while (!x->done);
3245                 __remove_wait_queue(&x->wait, &wait);
3246         }
3247         x->done--;
3248 out:
3249         spin_unlock_irq(&x->wait.lock);
3250         return timeout;
3251 }
3252 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3253
3254
3255 #define SLEEP_ON_VAR                                    \
3256         unsigned long flags;                            \
3257         wait_queue_t wait;                              \
3258         init_waitqueue_entry(&wait, current);
3259
3260 #define SLEEP_ON_HEAD                                   \
3261         spin_lock_irqsave(&q->lock,flags);              \
3262         __add_wait_queue(q, &wait);                     \
3263         spin_unlock(&q->lock);
3264
3265 #define SLEEP_ON_TAIL                                   \
3266         spin_lock_irq(&q->lock);                        \
3267         __remove_wait_queue(q, &wait);                  \
3268         spin_unlock_irqrestore(&q->lock, flags);
3269
3270 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3271 {
3272         SLEEP_ON_VAR
3273
3274         current->state = TASK_INTERRUPTIBLE;
3275
3276         SLEEP_ON_HEAD
3277         schedule();
3278         SLEEP_ON_TAIL
3279 }
3280
3281 EXPORT_SYMBOL(interruptible_sleep_on);
3282
3283 long fastcall __sched interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3284 {
3285         SLEEP_ON_VAR
3286
3287         current->state = TASK_INTERRUPTIBLE;
3288
3289         SLEEP_ON_HEAD
3290         timeout = schedule_timeout(timeout);
3291         SLEEP_ON_TAIL
3292
3293         return timeout;
3294 }
3295
3296 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3297
3298 void fastcall __sched sleep_on(wait_queue_head_t *q)
3299 {
3300         SLEEP_ON_VAR
3301
3302         current->state = TASK_UNINTERRUPTIBLE;
3303
3304         SLEEP_ON_HEAD
3305         schedule();
3306         SLEEP_ON_TAIL
3307 }
3308
3309 EXPORT_SYMBOL(sleep_on);
3310
3311 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3312 {
3313         SLEEP_ON_VAR
3314
3315         current->state = TASK_UNINTERRUPTIBLE;
3316
3317         SLEEP_ON_HEAD
3318         timeout = schedule_timeout(timeout);
3319         SLEEP_ON_TAIL
3320
3321         return timeout;
3322 }
3323
3324 EXPORT_SYMBOL(sleep_on_timeout);
3325
3326 void set_user_nice(task_t *p, long nice)
3327 {
3328         unsigned long flags;
3329         prio_array_t *array;
3330         runqueue_t *rq;
3331         int old_prio, new_prio, delta;
3332
3333         if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3334                 return;
3335         /*
3336          * We have to be careful, if called from sys_setpriority(),
3337          * the task might be in the middle of scheduling on another CPU.
3338          */
3339         rq = task_rq_lock(p, &flags);
3340         /*
3341          * The RT priorities are set via sched_setscheduler(), but we still
3342          * allow the 'normal' nice value to be set - but as expected
3343          * it wont have any effect on scheduling until the task is
3344          * not SCHED_NORMAL:
3345          */
3346         if (rt_task(p)) {
3347                 p->static_prio = NICE_TO_PRIO(nice);
3348                 goto out_unlock;
3349         }
3350         array = p->array;
3351         if (array)
3352                 dequeue_task(p, array);
3353
3354         old_prio = p->prio;
3355         new_prio = NICE_TO_PRIO(nice);
3356         delta = new_prio - old_prio;
3357         p->static_prio = NICE_TO_PRIO(nice);
3358         p->prio += delta;
3359
3360         if (array) {
3361                 enqueue_task(p, array);
3362                 /*
3363                  * If the task increased its priority or is running and
3364                  * lowered its priority, then reschedule its CPU:
3365                  */
3366                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3367                         resched_task(rq->curr);
3368         }
3369 out_unlock:
3370         task_rq_unlock(rq, &flags);
3371 }
3372
3373 EXPORT_SYMBOL(set_user_nice);
3374
3375 /*
3376  * can_nice - check if a task can reduce its nice value
3377  * @p: task
3378  * @nice: nice value
3379  */
3380 int can_nice(const task_t *p, const int nice)
3381 {
3382         /* convert nice value [19,-20] to rlimit style value [1,40] */
3383         int nice_rlim = 20 - nice;
3384         return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3385                 capable(CAP_SYS_NICE));
3386 }
3387
3388 #ifdef __ARCH_WANT_SYS_NICE
3389
3390 /*
3391  * sys_nice - change the priority of the current process.
3392  * @increment: priority increment
3393  *
3394  * sys_setpriority is a more generic, but much slower function that
3395  * does similar things.
3396  */
3397 asmlinkage long sys_nice(int increment)
3398 {
3399         int retval;
3400         long nice;
3401
3402         /*
3403          * Setpriority might change our priority at the same moment.
3404          * We don't have to worry. Conceptually one call occurs first
3405          * and we have a single winner.
3406          */
3407         if (increment < -40)
3408                 increment = -40;
3409         if (increment > 40)
3410                 increment = 40;
3411
3412         nice = PRIO_TO_NICE(current->static_prio) + increment;
3413         if (nice < -20)
3414                 nice = -20;
3415         if (nice > 19)
3416                 nice = 19;
3417
3418         if (increment < 0 && !can_nice(current, nice))
3419                 return -EPERM;
3420
3421         retval = security_task_setnice(current, nice);
3422         if (retval)