blob: 3399701c680e392f46e21829cee8da0bf5482303 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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>
Randy.Dunlapc59ede72006-01-11 12:17:46 -080030#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031#include <linux/completion.h>
32#include <linux/kernel_stat.h>
Ingo Molnar9a11b49a2006-07-03 00:24:33 -070033#include <linux/debug_locks.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <linux/security.h>
35#include <linux/notifier.h>
36#include <linux/profile.h>
37#include <linux/suspend.h>
akpm@osdl.org198e2f12006-01-12 01:05:30 -080038#include <linux/vmalloc.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <linux/blkdev.h>
40#include <linux/delay.h>
41#include <linux/smp.h>
42#include <linux/threads.h>
43#include <linux/timer.h>
44#include <linux/rcupdate.h>
45#include <linux/cpu.h>
46#include <linux/cpuset.h>
47#include <linux/percpu.h>
48#include <linux/kthread.h>
49#include <linux/seq_file.h>
50#include <linux/syscalls.h>
51#include <linux/times.h>
Jay Lan8f0ab512006-09-30 23:28:59 -070052#include <linux/tsacct_kern.h>
bibo maoc6fd91f2006-03-26 01:38:20 -080053#include <linux/kprobes.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070054#include <linux/delayacct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070055#include <asm/tlb.h>
56
57#include <asm/unistd.h>
58
59/*
60 * Convert user-nice values [ -20 ... 0 ... 19 ]
61 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
62 * and back.
63 */
64#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
65#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
66#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
67
68/*
69 * 'User priority' is the nice value converted to something we
70 * can work with better when scaling various scheduler parameters,
71 * it's a [ 0 ... 39 ] range.
72 */
73#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
74#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
75#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
76
77/*
78 * Some helpers for converting nanosecond timing to jiffy resolution
79 */
80#define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
81#define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
82
83/*
84 * These are the 'tuning knobs' of the scheduler:
85 *
86 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
87 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
88 * Timeslices get refilled after they expire.
89 */
90#define MIN_TIMESLICE max(5 * HZ / 1000, 1)
91#define DEF_TIMESLICE (100 * HZ / 1000)
92#define ON_RUNQUEUE_WEIGHT 30
93#define CHILD_PENALTY 95
94#define PARENT_PENALTY 100
95#define EXIT_WEIGHT 3
96#define PRIO_BONUS_RATIO 25
97#define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
98#define INTERACTIVE_DELTA 2
99#define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
100#define STARVATION_LIMIT (MAX_SLEEP_AVG)
101#define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
102
103/*
104 * If a task is 'interactive' then we reinsert it in the active
105 * array after it has expired its current timeslice. (it will not
106 * continue to run immediately, it will still roundrobin with
107 * other interactive tasks.)
108 *
109 * This part scales the interactivity limit depending on niceness.
110 *
111 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
112 * Here are a few examples of different nice levels:
113 *
114 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
115 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
116 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
117 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
118 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
119 *
120 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
121 * priority range a task can explore, a value of '1' means the
122 * task is rated interactive.)
123 *
124 * Ie. nice +19 tasks can never get 'interactive' enough to be
125 * reinserted into the active array. And only heavily CPU-hog nice -20
126 * tasks will be expired. Default nice 0 tasks are somewhere between,
127 * it takes some effort for them to get interactive, but it's not
128 * too hard.
129 */
130
131#define CURRENT_BONUS(p) \
132 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
133 MAX_SLEEP_AVG)
134
135#define GRANULARITY (10 * HZ / 1000 ? : 1)
136
137#ifdef CONFIG_SMP
138#define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
139 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
140 num_online_cpus())
141#else
142#define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
143 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
144#endif
145
146#define SCALE(v1,v1_max,v2_max) \
147 (v1) * (v2_max) / (v1_max)
148
149#define DELTA(p) \
Martin Andersson013d3862006-03-27 01:15:18 -0800150 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
151 INTERACTIVE_DELTA)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152
153#define TASK_INTERACTIVE(p) \
154 ((p)->prio <= (p)->static_prio - DELTA(p))
155
156#define INTERACTIVE_SLEEP(p) \
157 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
158 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
159
160#define TASK_PREEMPTS_CURR(p, rq) \
161 ((p)->prio < (rq)->curr->prio)
162
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163#define SCALE_PRIO(x, prio) \
Peter Williams2dd73a42006-06-27 02:54:34 -0700164 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165
Peter Williams2dd73a42006-06-27 02:54:34 -0700166static unsigned int static_prio_timeslice(int static_prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167{
Peter Williams2dd73a42006-06-27 02:54:34 -0700168 if (static_prio < NICE_TO_PRIO(0))
169 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 else
Peter Williams2dd73a42006-06-27 02:54:34 -0700171 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172}
Peter Williams2dd73a42006-06-27 02:54:34 -0700173
Borislav Petkov91fcdd42006-10-19 23:28:29 -0700174/*
175 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
176 * to time slice values: [800ms ... 100ms ... 5ms]
177 *
178 * The higher a thread's priority, the bigger timeslices
179 * it gets during one round of execution. But even the lowest
180 * priority thread gets MIN_TIMESLICE worth of execution time.
181 */
182
Ingo Molnar36c8b582006-07-03 00:25:41 -0700183static inline unsigned int task_timeslice(struct task_struct *p)
Peter Williams2dd73a42006-06-27 02:54:34 -0700184{
185 return static_prio_timeslice(p->static_prio);
186}
187
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188/*
189 * These are the runqueue data structures:
190 */
191
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192struct prio_array {
193 unsigned int nr_active;
Steven Rostedtd4448862006-06-27 02:54:29 -0700194 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195 struct list_head queue[MAX_PRIO];
196};
197
198/*
199 * This is the main, per-CPU runqueue data structure.
200 *
201 * Locking rule: those places that want to lock multiple runqueues
202 * (such as the load balancing or the thread migration code), lock
203 * acquire operations must be ordered by ascending &runqueue.
204 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700205struct rq {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206 spinlock_t lock;
207
208 /*
209 * nr_running and cpu_load should be in the same cacheline because
210 * remote CPUs use both these fields when doing load calculation.
211 */
212 unsigned long nr_running;
Peter Williams2dd73a42006-06-27 02:54:34 -0700213 unsigned long raw_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214#ifdef CONFIG_SMP
Nick Piggin78979862005-06-25 14:57:13 -0700215 unsigned long cpu_load[3];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216#endif
217 unsigned long long nr_switches;
218
219 /*
220 * This is part of a global counter where only the total sum
221 * over all CPUs matters. A task can increase this counter on
222 * one CPU and if it got migrated afterwards it may decrease
223 * it on another CPU. Always updated under the runqueue lock:
224 */
225 unsigned long nr_uninterruptible;
226
227 unsigned long expired_timestamp;
228 unsigned long long timestamp_last_tick;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700229 struct task_struct *curr, *idle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230 struct mm_struct *prev_mm;
Ingo Molnar70b97a72006-07-03 00:25:42 -0700231 struct prio_array *active, *expired, arrays[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 int best_expired_prio;
233 atomic_t nr_iowait;
234
235#ifdef CONFIG_SMP
236 struct sched_domain *sd;
237
238 /* For active balancing */
239 int active_balance;
240 int push_cpu;
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700241 int cpu; /* cpu of this runqueue */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Ingo Molnar36c8b582006-07-03 00:25:41 -0700243 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244 struct list_head migration_queue;
245#endif
246
247#ifdef CONFIG_SCHEDSTATS
248 /* latency stats */
249 struct sched_info rq_sched_info;
250
251 /* sys_sched_yield() stats */
252 unsigned long yld_exp_empty;
253 unsigned long yld_act_empty;
254 unsigned long yld_both_empty;
255 unsigned long yld_cnt;
256
257 /* schedule() stats */
258 unsigned long sched_switch;
259 unsigned long sched_cnt;
260 unsigned long sched_goidle;
261
262 /* try_to_wake_up() stats */
263 unsigned long ttwu_cnt;
264 unsigned long ttwu_local;
265#endif
Ingo Molnarfcb99372006-07-03 00:25:10 -0700266 struct lock_class_key rq_lock_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267};
268
Ingo Molnar70b97a72006-07-03 00:25:42 -0700269static DEFINE_PER_CPU(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700271static inline int cpu_of(struct rq *rq)
272{
273#ifdef CONFIG_SMP
274 return rq->cpu;
275#else
276 return 0;
277#endif
278}
279
Nick Piggin674311d2005-06-25 14:57:27 -0700280/*
281 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700282 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700283 *
284 * The domain tree of any CPU may only be accessed from within
285 * preempt-disabled sections.
286 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700287#define for_each_domain(cpu, __sd) \
288 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289
290#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
291#define this_rq() (&__get_cpu_var(runqueues))
292#define task_rq(p) cpu_rq(task_cpu(p))
293#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
294
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700296# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700298#ifndef finish_arch_switch
299# define finish_arch_switch(prev) do { } while (0)
300#endif
301
302#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700303static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700304{
305 return rq->curr == p;
306}
307
Ingo Molnar70b97a72006-07-03 00:25:42 -0700308static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700309{
310}
311
Ingo Molnar70b97a72006-07-03 00:25:42 -0700312static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700313{
Ingo Molnarda04c032005-09-13 11:17:59 +0200314#ifdef CONFIG_DEBUG_SPINLOCK
315 /* this is a valid case when another task releases the spinlock */
316 rq->lock.owner = current;
317#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700318 /*
319 * If we are tracking spinlock dependencies then we have to
320 * fix up the runqueue lock - which gets 'carried over' from
321 * prev into current:
322 */
323 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
324
Nick Piggin4866cde2005-06-25 14:57:23 -0700325 spin_unlock_irq(&rq->lock);
326}
327
328#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700329static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700330{
331#ifdef CONFIG_SMP
332 return p->oncpu;
333#else
334 return rq->curr == p;
335#endif
336}
337
Ingo Molnar70b97a72006-07-03 00:25:42 -0700338static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700339{
340#ifdef CONFIG_SMP
341 /*
342 * We can optimise this out completely for !SMP, because the
343 * SMP rebalancing from interrupt is the only thing that cares
344 * here.
345 */
346 next->oncpu = 1;
347#endif
348#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
349 spin_unlock_irq(&rq->lock);
350#else
351 spin_unlock(&rq->lock);
352#endif
353}
354
Ingo Molnar70b97a72006-07-03 00:25:42 -0700355static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700356{
357#ifdef CONFIG_SMP
358 /*
359 * After ->oncpu is cleared, the task can be moved to a different CPU.
360 * We must ensure this doesn't happen until the switch is completely
361 * finished.
362 */
363 smp_wmb();
364 prev->oncpu = 0;
365#endif
366#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
367 local_irq_enable();
368#endif
369}
370#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371
372/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700373 * __task_rq_lock - lock the runqueue a given task resides on.
374 * Must be called interrupts disabled.
375 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700376static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700377 __acquires(rq->lock)
378{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700379 struct rq *rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700380
381repeat_lock_task:
382 rq = task_rq(p);
383 spin_lock(&rq->lock);
384 if (unlikely(rq != task_rq(p))) {
385 spin_unlock(&rq->lock);
386 goto repeat_lock_task;
387 }
388 return rq;
389}
390
391/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 * task_rq_lock - lock the runqueue a given task resides on and disable
393 * interrupts. Note the ordering: we can safely lookup the task_rq without
394 * explicitly disabling preemption.
395 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700396static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 __acquires(rq->lock)
398{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700399 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
401repeat_lock_task:
402 local_irq_save(*flags);
403 rq = task_rq(p);
404 spin_lock(&rq->lock);
405 if (unlikely(rq != task_rq(p))) {
406 spin_unlock_irqrestore(&rq->lock, *flags);
407 goto repeat_lock_task;
408 }
409 return rq;
410}
411
Ingo Molnar70b97a72006-07-03 00:25:42 -0700412static inline void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700413 __releases(rq->lock)
414{
415 spin_unlock(&rq->lock);
416}
417
Ingo Molnar70b97a72006-07-03 00:25:42 -0700418static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 __releases(rq->lock)
420{
421 spin_unlock_irqrestore(&rq->lock, *flags);
422}
423
424#ifdef CONFIG_SCHEDSTATS
425/*
426 * bump this up when changing the output format or the meaning of an existing
427 * format, so that tools can adapt (or abort)
428 */
Nick Piggin68767a02005-06-25 14:57:20 -0700429#define SCHEDSTAT_VERSION 12
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430
431static int show_schedstat(struct seq_file *seq, void *v)
432{
433 int cpu;
434
435 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
436 seq_printf(seq, "timestamp %lu\n", jiffies);
437 for_each_online_cpu(cpu) {
Ingo Molnar70b97a72006-07-03 00:25:42 -0700438 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439#ifdef CONFIG_SMP
440 struct sched_domain *sd;
441 int dcnt = 0;
442#endif
443
444 /* runqueue-specific stats */
445 seq_printf(seq,
446 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
447 cpu, rq->yld_both_empty,
448 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
449 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
450 rq->ttwu_cnt, rq->ttwu_local,
451 rq->rq_sched_info.cpu_time,
452 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
453
454 seq_printf(seq, "\n");
455
456#ifdef CONFIG_SMP
457 /* domain-specific stats */
Nick Piggin674311d2005-06-25 14:57:27 -0700458 preempt_disable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 for_each_domain(cpu, sd) {
460 enum idle_type itype;
461 char mask_str[NR_CPUS];
462
463 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
464 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
465 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
466 itype++) {
467 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
468 sd->lb_cnt[itype],
469 sd->lb_balanced[itype],
470 sd->lb_failed[itype],
471 sd->lb_imbalance[itype],
472 sd->lb_gained[itype],
473 sd->lb_hot_gained[itype],
474 sd->lb_nobusyq[itype],
475 sd->lb_nobusyg[itype]);
476 }
Nick Piggin68767a02005-06-25 14:57:20 -0700477 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
Nick Piggin68767a02005-06-25 14:57:20 -0700479 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
480 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
482 }
Nick Piggin674311d2005-06-25 14:57:27 -0700483 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484#endif
485 }
486 return 0;
487}
488
489static int schedstat_open(struct inode *inode, struct file *file)
490{
491 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
492 char *buf = kmalloc(size, GFP_KERNEL);
493 struct seq_file *m;
494 int res;
495
496 if (!buf)
497 return -ENOMEM;
498 res = single_open(file, show_schedstat, NULL);
499 if (!res) {
500 m = file->private_data;
501 m->buf = buf;
502 m->size = size;
503 } else
504 kfree(buf);
505 return res;
506}
507
508struct file_operations proc_schedstat_operations = {
509 .open = schedstat_open,
510 .read = seq_read,
511 .llseek = seq_lseek,
512 .release = single_release,
513};
514
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700515/*
516 * Expects runqueue lock to be held for atomicity of update
517 */
518static inline void
519rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
520{
521 if (rq) {
522 rq->rq_sched_info.run_delay += delta_jiffies;
523 rq->rq_sched_info.pcnt++;
524 }
525}
526
527/*
528 * Expects runqueue lock to be held for atomicity of update
529 */
530static inline void
531rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
532{
533 if (rq)
534 rq->rq_sched_info.cpu_time += delta_jiffies;
535}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536# define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
537# define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
538#else /* !CONFIG_SCHEDSTATS */
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700539static inline void
540rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
541{}
542static inline void
543rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
544{}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545# define schedstat_inc(rq, field) do { } while (0)
546# define schedstat_add(rq, field, amt) do { } while (0)
547#endif
548
549/*
550 * rq_lock - lock a given runqueue and disable interrupts.
551 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700552static inline struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 __acquires(rq->lock)
554{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700555 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556
557 local_irq_disable();
558 rq = this_rq();
559 spin_lock(&rq->lock);
560
561 return rq;
562}
563
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700564#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565/*
566 * Called when a process is dequeued from the active array and given
567 * the cpu. We should note that with the exception of interactive
568 * tasks, the expired queue will become the active queue after the active
569 * queue is empty, without explicitly dequeuing and requeuing tasks in the
570 * expired queue. (Interactive tasks may be requeued directly to the
571 * active queue, thus delaying tasks in the expired queue from running;
572 * see scheduler_tick()).
573 *
574 * This function is only called from sched_info_arrive(), rather than
575 * dequeue_task(). Even though a task may be queued and dequeued multiple
576 * times as it is shuffled about, we're really interested in knowing how
577 * long it was from the *first* time it was queued to the time that it
578 * finally hit a cpu.
579 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700580static inline void sched_info_dequeued(struct task_struct *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581{
582 t->sched_info.last_queued = 0;
583}
584
585/*
586 * Called when a task finally hits the cpu. We can now calculate how
587 * long it was waiting to run. We also note when it began so that we
588 * can keep stats on how long its timeslice is.
589 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700590static void sched_info_arrive(struct task_struct *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591{
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700592 unsigned long now = jiffies, delta_jiffies = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593
594 if (t->sched_info.last_queued)
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700595 delta_jiffies = now - t->sched_info.last_queued;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 sched_info_dequeued(t);
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700597 t->sched_info.run_delay += delta_jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 t->sched_info.last_arrival = now;
599 t->sched_info.pcnt++;
600
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700601 rq_sched_info_arrive(task_rq(t), delta_jiffies);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602}
603
604/*
605 * Called when a process is queued into either the active or expired
606 * array. The time is noted and later used to determine how long we
607 * had to wait for us to reach the cpu. Since the expired queue will
608 * become the active queue after active queue is empty, without dequeuing
609 * and requeuing any tasks, we are interested in queuing to either. It
610 * is unusual but not impossible for tasks to be dequeued and immediately
611 * requeued in the same or another array: this can happen in sched_yield(),
612 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
613 * to runqueue.
614 *
615 * This function is only called from enqueue_task(), but also only updates
616 * the timestamp if it is already not set. It's assumed that
617 * sched_info_dequeued() will clear that stamp when appropriate.
618 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700619static inline void sched_info_queued(struct task_struct *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620{
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700621 if (unlikely(sched_info_on()))
622 if (!t->sched_info.last_queued)
623 t->sched_info.last_queued = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624}
625
626/*
627 * Called when a process ceases being the active-running process, either
628 * voluntarily or involuntarily. Now we can calculate how long we ran.
629 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700630static inline void sched_info_depart(struct task_struct *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631{
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700632 unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700634 t->sched_info.cpu_time += delta_jiffies;
635 rq_sched_info_depart(task_rq(t), delta_jiffies);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636}
637
638/*
639 * Called when tasks are switched involuntarily due, typically, to expiring
640 * their time slice. (This may also be called when switching to or from
641 * the idle task.) We are only called when prev != next.
642 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700643static inline void
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700644__sched_info_switch(struct task_struct *prev, struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700646 struct rq *rq = task_rq(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647
648 /*
649 * prev now departs the cpu. It's not interesting to record
650 * stats about how efficient we were at scheduling the idle
651 * process, however.
652 */
653 if (prev != rq->idle)
654 sched_info_depart(prev);
655
656 if (next != rq->idle)
657 sched_info_arrive(next);
658}
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700659static inline void
660sched_info_switch(struct task_struct *prev, struct task_struct *next)
661{
662 if (unlikely(sched_info_on()))
663 __sched_info_switch(prev, next);
664}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665#else
666#define sched_info_queued(t) do { } while (0)
667#define sched_info_switch(t, next) do { } while (0)
Chandra Seetharaman52f17b62006-07-14 00:24:38 -0700668#endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
670/*
671 * Adding/removing a task to/from a priority array:
672 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700673static void dequeue_task(struct task_struct *p, struct prio_array *array)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674{
675 array->nr_active--;
676 list_del(&p->run_list);
677 if (list_empty(array->queue + p->prio))
678 __clear_bit(p->prio, array->bitmap);
679}
680
Ingo Molnar70b97a72006-07-03 00:25:42 -0700681static void enqueue_task(struct task_struct *p, struct prio_array *array)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682{
683 sched_info_queued(p);
684 list_add_tail(&p->run_list, array->queue + p->prio);
685 __set_bit(p->prio, array->bitmap);
686 array->nr_active++;
687 p->array = array;
688}
689
690/*
691 * Put task to the end of the run list without the overhead of dequeue
692 * followed by enqueue.
693 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700694static void requeue_task(struct task_struct *p, struct prio_array *array)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695{
696 list_move_tail(&p->run_list, array->queue + p->prio);
697}
698
Ingo Molnar70b97a72006-07-03 00:25:42 -0700699static inline void
700enqueue_task_head(struct task_struct *p, struct prio_array *array)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701{
702 list_add(&p->run_list, array->queue + p->prio);
703 __set_bit(p->prio, array->bitmap);
704 array->nr_active++;
705 p->array = array;
706}
707
708/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700709 * __normal_prio - return the priority that is based on the static
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 * priority but is modified by bonuses/penalties.
711 *
712 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
713 * into the -5 ... 0 ... +5 bonus/penalty range.
714 *
715 * We use 25% of the full 0...39 priority range so that:
716 *
717 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
718 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
719 *
720 * Both properties are important to certain workloads.
721 */
Ingo Molnarb29739f2006-06-27 02:54:51 -0700722
Ingo Molnar36c8b582006-07-03 00:25:41 -0700723static inline int __normal_prio(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724{
725 int bonus, prio;
726
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
728
729 prio = p->static_prio - bonus;
730 if (prio < MAX_RT_PRIO)
731 prio = MAX_RT_PRIO;
732 if (prio > MAX_PRIO-1)
733 prio = MAX_PRIO-1;
734 return prio;
735}
736
737/*
Peter Williams2dd73a42006-06-27 02:54:34 -0700738 * To aid in avoiding the subversion of "niceness" due to uneven distribution
739 * of tasks with abnormal "nice" values across CPUs the contribution that
740 * each task makes to its run queue's load is weighted according to its
741 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
742 * scaled version of the new time slice allocation that they receive on time
743 * slice expiry etc.
744 */
745
746/*
747 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
748 * If static_prio_timeslice() is ever changed to break this assumption then
749 * this code will need modification
750 */
751#define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
752#define LOAD_WEIGHT(lp) \
753 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
754#define PRIO_TO_LOAD_WEIGHT(prio) \
755 LOAD_WEIGHT(static_prio_timeslice(prio))
756#define RTPRIO_TO_LOAD_WEIGHT(rp) \
757 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
758
Ingo Molnar36c8b582006-07-03 00:25:41 -0700759static void set_load_weight(struct task_struct *p)
Peter Williams2dd73a42006-06-27 02:54:34 -0700760{
Ingo Molnarb29739f2006-06-27 02:54:51 -0700761 if (has_rt_policy(p)) {
Peter Williams2dd73a42006-06-27 02:54:34 -0700762#ifdef CONFIG_SMP
763 if (p == task_rq(p)->migration_thread)
764 /*
765 * The migration thread does the actual balancing.
766 * Giving its load any weight will skew balancing
767 * adversely.
768 */
769 p->load_weight = 0;
770 else
771#endif
772 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
773 } else
774 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
775}
776
Ingo Molnar36c8b582006-07-03 00:25:41 -0700777static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -0700778inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
Peter Williams2dd73a42006-06-27 02:54:34 -0700779{
780 rq->raw_weighted_load += p->load_weight;
781}
782
Ingo Molnar36c8b582006-07-03 00:25:41 -0700783static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -0700784dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
Peter Williams2dd73a42006-06-27 02:54:34 -0700785{
786 rq->raw_weighted_load -= p->load_weight;
787}
788
Ingo Molnar70b97a72006-07-03 00:25:42 -0700789static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
Peter Williams2dd73a42006-06-27 02:54:34 -0700790{
791 rq->nr_running++;
792 inc_raw_weighted_load(rq, p);
793}
794
Ingo Molnar70b97a72006-07-03 00:25:42 -0700795static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
Peter Williams2dd73a42006-06-27 02:54:34 -0700796{
797 rq->nr_running--;
798 dec_raw_weighted_load(rq, p);
799}
800
801/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700802 * Calculate the expected normal priority: i.e. priority
803 * without taking RT-inheritance into account. Might be
804 * boosted by interactivity modifiers. Changes upon fork,
805 * setprio syscalls, and whenever the interactivity
806 * estimator recalculates.
807 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700808static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700809{
810 int prio;
811
812 if (has_rt_policy(p))
813 prio = MAX_RT_PRIO-1 - p->rt_priority;
814 else
815 prio = __normal_prio(p);
816 return prio;
817}
818
819/*
820 * Calculate the current priority, i.e. the priority
821 * taken into account by the scheduler. This value might
822 * be boosted by RT tasks, or might be boosted by
823 * interactivity modifiers. Will be RT if the task got
824 * RT-boosted. If not then it returns p->normal_prio.
825 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700826static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700827{
828 p->normal_prio = normal_prio(p);
829 /*
830 * If we are RT tasks or we were boosted to RT priority,
831 * keep the priority unchanged. Otherwise, update priority
832 * to the normal priority:
833 */
834 if (!rt_prio(p->prio))
835 return p->normal_prio;
836 return p->prio;
837}
838
839/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840 * __activate_task - move a task to the runqueue.
841 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700842static void __activate_task(struct task_struct *p, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700844 struct prio_array *target = rq->active;
Con Kolivasd425b272006-03-31 02:31:29 -0800845
Linus Torvaldsf1adad72006-05-21 18:54:09 -0700846 if (batch_task(p))
Con Kolivasd425b272006-03-31 02:31:29 -0800847 target = rq->expired;
848 enqueue_task(p, target);
Peter Williams2dd73a42006-06-27 02:54:34 -0700849 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850}
851
852/*
853 * __activate_idle_task - move idle task to the _front_ of runqueue.
854 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700855static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856{
857 enqueue_task_head(p, rq->active);
Peter Williams2dd73a42006-06-27 02:54:34 -0700858 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859}
860
Ingo Molnarb29739f2006-06-27 02:54:51 -0700861/*
862 * Recalculate p->normal_prio and p->prio after having slept,
863 * updating the sleep-average too:
864 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700865static int recalc_task_prio(struct task_struct *p, unsigned long long now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866{
867 /* Caller must always ensure 'now >= p->timestamp' */
Con Kolivas72d28542006-06-27 02:54:30 -0700868 unsigned long sleep_time = now - p->timestamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869
Con Kolivasd425b272006-03-31 02:31:29 -0800870 if (batch_task(p))
Ingo Molnarb0a94992006-01-14 13:20:41 -0800871 sleep_time = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872
873 if (likely(sleep_time > 0)) {
874 /*
Con Kolivas72d28542006-06-27 02:54:30 -0700875 * This ceiling is set to the lowest priority that would allow
876 * a task to be reinserted into the active array on timeslice
877 * completion.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 */
Con Kolivas72d28542006-06-27 02:54:30 -0700879 unsigned long ceiling = INTERACTIVE_SLEEP(p);
Con Kolivase72ff0b2006-03-31 02:31:26 -0800880
Con Kolivas72d28542006-06-27 02:54:30 -0700881 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
882 /*
883 * Prevents user tasks from achieving best priority
884 * with one single large enough sleep.
885 */
886 p->sleep_avg = ceiling;
887 /*
888 * Using INTERACTIVE_SLEEP() as a ceiling places a
889 * nice(0) task 1ms sleep away from promotion, and
890 * gives it 700ms to round-robin with no chance of
891 * being demoted. This is more than generous, so
892 * mark this sleep as non-interactive to prevent the
893 * on-runqueue bonus logic from intervening should
894 * this task not receive cpu immediately.
895 */
896 p->sleep_type = SLEEP_NONINTERACTIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 } else {
898 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 * Tasks waking from uninterruptible sleep are
900 * limited in their sleep_avg rise as they
901 * are likely to be waiting on I/O
902 */
Con Kolivas3dee3862006-03-31 02:31:23 -0800903 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
Con Kolivas72d28542006-06-27 02:54:30 -0700904 if (p->sleep_avg >= ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 sleep_time = 0;
906 else if (p->sleep_avg + sleep_time >=
Con Kolivas72d28542006-06-27 02:54:30 -0700907 ceiling) {
908 p->sleep_avg = ceiling;
909 sleep_time = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 }
911 }
912
913 /*
914 * This code gives a bonus to interactive tasks.
915 *
916 * The boost works by updating the 'average sleep time'
917 * value here, based on ->timestamp. The more time a
918 * task spends sleeping, the higher the average gets -
919 * and the higher the priority boost gets as well.
920 */
921 p->sleep_avg += sleep_time;
922
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 }
Con Kolivas72d28542006-06-27 02:54:30 -0700924 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
925 p->sleep_avg = NS_MAX_SLEEP_AVG;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 }
927
Chen Shanga3464a12005-06-25 14:57:31 -0700928 return effective_prio(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929}
930
931/*
932 * activate_task - move a task to the runqueue and do priority recalculation
933 *
934 * Update all the scheduling statistics stuff. (sleep average
935 * calculation, priority modifiers, etc.)
936 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700937static void activate_task(struct task_struct *p, struct rq *rq, int local)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938{
939 unsigned long long now;
940
941 now = sched_clock();
942#ifdef CONFIG_SMP
943 if (!local) {
944 /* Compensate for drifting sched_clock */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700945 struct rq *this_rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 now = (now - this_rq->timestamp_last_tick)
947 + rq->timestamp_last_tick;
948 }
949#endif
950
Chen, Kenneth Wa47ab932005-11-09 15:45:29 -0800951 if (!rt_task(p))
952 p->prio = recalc_task_prio(p, now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953
954 /*
955 * This checks to make sure it's not an uninterruptible task
956 * that is now waking up.
957 */
Con Kolivas3dee3862006-03-31 02:31:23 -0800958 if (p->sleep_type == SLEEP_NORMAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 /*
960 * Tasks which were woken up by interrupts (ie. hw events)
961 * are most likely of interactive nature. So we give them
962 * the credit of extending their sleep time to the period
963 * of time they spend on the runqueue, waiting for execution
964 * on a CPU, first time around:
965 */
966 if (in_interrupt())
Con Kolivas3dee3862006-03-31 02:31:23 -0800967 p->sleep_type = SLEEP_INTERRUPTED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 else {
969 /*
970 * Normal first-time wakeups get a credit too for
971 * on-runqueue time, but it will be weighted down:
972 */
Con Kolivas3dee3862006-03-31 02:31:23 -0800973 p->sleep_type = SLEEP_INTERACTIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 }
975 }
976 p->timestamp = now;
977
978 __activate_task(p, rq);
979}
980
981/*
982 * deactivate_task - remove a task from the runqueue.
983 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700984static void deactivate_task(struct task_struct *p, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985{
Peter Williams2dd73a42006-06-27 02:54:34 -0700986 dec_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 dequeue_task(p, p->array);
988 p->array = NULL;
989}
990
991/*
992 * resched_task - mark a task 'to be rescheduled now'.
993 *
994 * On UP this means the setting of the need_resched flag, on SMP it
995 * might also involve a cross-CPU call to trigger the scheduler on
996 * the target CPU.
997 */
998#ifdef CONFIG_SMP
Andi Kleen495ab9c2006-06-26 13:59:11 +0200999
1000#ifndef tsk_is_polling
1001#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1002#endif
1003
Ingo Molnar36c8b582006-07-03 00:25:41 -07001004static void resched_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005{
Nick Piggin64c7c8f2005-11-08 21:39:04 -08001006 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007
1008 assert_spin_locked(&task_rq(p)->lock);
1009
Nick Piggin64c7c8f2005-11-08 21:39:04 -08001010 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1011 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012
Nick Piggin64c7c8f2005-11-08 21:39:04 -08001013 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1014
1015 cpu = task_cpu(p);
1016 if (cpu == smp_processor_id())
1017 return;
1018
Andi Kleen495ab9c2006-06-26 13:59:11 +02001019 /* NEED_RESCHED must be visible before we test polling */
Nick Piggin64c7c8f2005-11-08 21:39:04 -08001020 smp_mb();
Andi Kleen495ab9c2006-06-26 13:59:11 +02001021 if (!tsk_is_polling(p))
Nick Piggin64c7c8f2005-11-08 21:39:04 -08001022 smp_send_reschedule(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023}
1024#else
Ingo Molnar36c8b582006-07-03 00:25:41 -07001025static inline void resched_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026{
Nick Piggin64c7c8f2005-11-08 21:39:04 -08001027 assert_spin_locked(&task_rq(p)->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 set_tsk_need_resched(p);
1029}
1030#endif
1031
1032/**
1033 * task_curr - is this task currently executing on a CPU?
1034 * @p: the task in question.
1035 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001036inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037{
1038 return cpu_curr(task_cpu(p)) == p;
1039}
1040
Peter Williams2dd73a42006-06-27 02:54:34 -07001041/* Used instead of source_load when we know the type == 0 */
1042unsigned long weighted_cpuload(const int cpu)
1043{
1044 return cpu_rq(cpu)->raw_weighted_load;
1045}
1046
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047#ifdef CONFIG_SMP
Ingo Molnar70b97a72006-07-03 00:25:42 -07001048struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050
Ingo Molnar36c8b582006-07-03 00:25:41 -07001051 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 int dest_cpu;
1053
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001055};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056
1057/*
1058 * The task's runqueue lock must be held.
1059 * Returns true if you have to wait for migration thread.
1060 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001061static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07001062migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001064 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065
1066 /*
1067 * If the task is not on a runqueue (and not running), then
1068 * it is sufficient to simply update the task's cpu field.
1069 */
1070 if (!p->array && !task_running(rq, p)) {
1071 set_task_cpu(p, dest_cpu);
1072 return 0;
1073 }
1074
1075 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 req->task = p;
1077 req->dest_cpu = dest_cpu;
1078 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07001079
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 return 1;
1081}
1082
1083/*
1084 * wait_task_inactive - wait for a thread to unschedule.
1085 *
1086 * The caller must ensure that the task *will* unschedule sometime soon,
1087 * else this function might spin for a *long* time. This function can't
1088 * be called with interrupts off, or it may introduce deadlock with
1089 * smp_call_function() if an IPI is sent by the same process we are
1090 * waiting to become inactive.
1091 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001092void wait_task_inactive(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093{
1094 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001095 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 int preempted;
1097
1098repeat:
1099 rq = task_rq_lock(p, &flags);
1100 /* Must be off runqueue entirely, not preempted. */
1101 if (unlikely(p->array || task_running(rq, p))) {
1102 /* If it's preempted, we yield. It could be a while. */
1103 preempted = !task_running(rq, p);
1104 task_rq_unlock(rq, &flags);
1105 cpu_relax();
1106 if (preempted)
1107 yield();
1108 goto repeat;
1109 }
1110 task_rq_unlock(rq, &flags);
1111}
1112
1113/***
1114 * kick_process - kick a running thread to enter/exit the kernel
1115 * @p: the to-be-kicked thread
1116 *
1117 * Cause a process which is running on another CPU to enter
1118 * kernel-mode, without any delay. (to get signals handled.)
1119 *
1120 * NOTE: this function doesnt have to take the runqueue lock,
1121 * because all it wants to ensure is that the remote task enters
1122 * the kernel. If the IPI races and the task has been migrated
1123 * to another CPU then no harm is done and the purpose has been
1124 * achieved as well.
1125 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001126void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127{
1128 int cpu;
1129
1130 preempt_disable();
1131 cpu = task_cpu(p);
1132 if ((cpu != smp_processor_id()) && task_curr(p))
1133 smp_send_reschedule(cpu);
1134 preempt_enable();
1135}
1136
1137/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001138 * Return a low guess at the load of a migration-source cpu weighted
1139 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 *
1141 * We want to under-estimate the load of migration sources, to
1142 * balance conservatively.
1143 */
Con Kolivasb9104722005-11-08 21:38:55 -08001144static inline unsigned long source_load(int cpu, int type)
1145{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001146 struct rq *rq = cpu_rq(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001147
Peter Williams2dd73a42006-06-27 02:54:34 -07001148 if (type == 0)
1149 return rq->raw_weighted_load;
1150
1151 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152}
1153
1154/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001155 * Return a high guess at the load of a migration-target cpu weighted
1156 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 */
Con Kolivasb9104722005-11-08 21:38:55 -08001158static inline unsigned long target_load(int cpu, int type)
1159{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001160 struct rq *rq = cpu_rq(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001161
Peter Williams2dd73a42006-06-27 02:54:34 -07001162 if (type == 0)
1163 return rq->raw_weighted_load;
1164
1165 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1166}
1167
1168/*
1169 * Return the average load per task on the cpu's run queue
1170 */
1171static inline unsigned long cpu_avg_load_per_task(int cpu)
1172{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001173 struct rq *rq = cpu_rq(cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07001174 unsigned long n = rq->nr_running;
1175
Ingo Molnar48f24c42006-07-03 00:25:40 -07001176 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177}
1178
Nick Piggin147cbb42005-06-25 14:57:19 -07001179/*
1180 * find_idlest_group finds and returns the least busy CPU group within the
1181 * domain.
1182 */
1183static struct sched_group *
1184find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1185{
1186 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1187 unsigned long min_load = ULONG_MAX, this_load = 0;
1188 int load_idx = sd->forkexec_idx;
1189 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1190
1191 do {
1192 unsigned long load, avg_load;
1193 int local_group;
1194 int i;
1195
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001196 /* Skip over this group if it has no CPUs allowed */
1197 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1198 goto nextgroup;
1199
Nick Piggin147cbb42005-06-25 14:57:19 -07001200 local_group = cpu_isset(this_cpu, group->cpumask);
Nick Piggin147cbb42005-06-25 14:57:19 -07001201
1202 /* Tally up the load of all CPUs in the group */
1203 avg_load = 0;
1204
1205 for_each_cpu_mask(i, group->cpumask) {
1206 /* Bias balancing toward cpus of our domain */
1207 if (local_group)
1208 load = source_load(i, load_idx);
1209 else
1210 load = target_load(i, load_idx);
1211
1212 avg_load += load;
1213 }
1214
1215 /* Adjust by relative CPU power of the group */
1216 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1217
1218 if (local_group) {
1219 this_load = avg_load;
1220 this = group;
1221 } else if (avg_load < min_load) {
1222 min_load = avg_load;
1223 idlest = group;
1224 }
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001225nextgroup:
Nick Piggin147cbb42005-06-25 14:57:19 -07001226 group = group->next;
1227 } while (group != sd->groups);
1228
1229 if (!idlest || 100*this_load < imbalance*min_load)
1230 return NULL;
1231 return idlest;
1232}
1233
1234/*
Satoru Takeuchi0feaece2006-10-03 01:14:10 -07001235 * find_idlest_cpu - find the idlest cpu among the cpus in group.
Nick Piggin147cbb42005-06-25 14:57:19 -07001236 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07001237static int
1238find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
Nick Piggin147cbb42005-06-25 14:57:19 -07001239{
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001240 cpumask_t tmp;
Nick Piggin147cbb42005-06-25 14:57:19 -07001241 unsigned long load, min_load = ULONG_MAX;
1242 int idlest = -1;
1243 int i;
1244
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001245 /* Traverse only the allowed CPUs */
1246 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1247
1248 for_each_cpu_mask(i, tmp) {
Peter Williams2dd73a42006-06-27 02:54:34 -07001249 load = weighted_cpuload(i);
Nick Piggin147cbb42005-06-25 14:57:19 -07001250
1251 if (load < min_load || (load == min_load && i == this_cpu)) {
1252 min_load = load;
1253 idlest = i;
1254 }
1255 }
1256
1257 return idlest;
1258}
1259
Nick Piggin476d1392005-06-25 14:57:29 -07001260/*
1261 * sched_balance_self: balance the current task (running on cpu) in domains
1262 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1263 * SD_BALANCE_EXEC.
1264 *
1265 * Balance, ie. select the least loaded group.
1266 *
1267 * Returns the target CPU number, or the same CPU if no balancing is needed.
1268 *
1269 * preempt must be disabled.
1270 */
1271static int sched_balance_self(int cpu, int flag)
1272{
1273 struct task_struct *t = current;
1274 struct sched_domain *tmp, *sd = NULL;
Nick Piggin147cbb42005-06-25 14:57:19 -07001275
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001276 for_each_domain(cpu, tmp) {
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07001277 /*
1278 * If power savings logic is enabled for a domain, stop there.
1279 */
1280 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1281 break;
Nick Piggin476d1392005-06-25 14:57:29 -07001282 if (tmp->flags & flag)
1283 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001284 }
Nick Piggin476d1392005-06-25 14:57:29 -07001285
1286 while (sd) {
1287 cpumask_t span;
1288 struct sched_group *group;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001289 int new_cpu, weight;
1290
1291 if (!(sd->flags & flag)) {
1292 sd = sd->child;
1293 continue;
1294 }
Nick Piggin476d1392005-06-25 14:57:29 -07001295
1296 span = sd->span;
1297 group = find_idlest_group(sd, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001298 if (!group) {
1299 sd = sd->child;
1300 continue;
1301 }
Nick Piggin476d1392005-06-25 14:57:29 -07001302
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001303 new_cpu = find_idlest_cpu(group, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001304 if (new_cpu == -1 || new_cpu == cpu) {
1305 /* Now try balancing at a lower domain level of cpu */
1306 sd = sd->child;
1307 continue;
1308 }
Nick Piggin476d1392005-06-25 14:57:29 -07001309
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001310 /* Now try balancing at a lower domain level of new_cpu */
Nick Piggin476d1392005-06-25 14:57:29 -07001311 cpu = new_cpu;
Nick Piggin476d1392005-06-25 14:57:29 -07001312 sd = NULL;
1313 weight = cpus_weight(span);
1314 for_each_domain(cpu, tmp) {
1315 if (weight <= cpus_weight(tmp->span))
1316 break;
1317 if (tmp->flags & flag)
1318 sd = tmp;
1319 }
1320 /* while loop will break here if sd == NULL */
1321 }
1322
1323 return cpu;
1324}
1325
1326#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327
1328/*
1329 * wake_idle() will wake a task on an idle cpu if task->cpu is
1330 * not idle and an idle cpu is available. The span of cpus to
1331 * search starts with cpus closest then further out as needed,
1332 * so we always favor a closer, idle cpu.
1333 *
1334 * Returns the CPU we should wake onto.
1335 */
1336#if defined(ARCH_HAS_SCHED_WAKE_IDLE)
Ingo Molnar36c8b582006-07-03 00:25:41 -07001337static int wake_idle(int cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338{
1339 cpumask_t tmp;
1340 struct sched_domain *sd;
1341 int i;
1342
1343 if (idle_cpu(cpu))
1344 return cpu;
1345
1346 for_each_domain(cpu, sd) {
1347 if (sd->flags & SD_WAKE_IDLE) {
Nick Piggine0f364f2005-06-25 14:57:06 -07001348 cpus_and(tmp, sd->span, p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 for_each_cpu_mask(i, tmp) {
1350 if (idle_cpu(i))
1351 return i;
1352 }
1353 }
Nick Piggine0f364f2005-06-25 14:57:06 -07001354 else
1355 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 }
1357 return cpu;
1358}
1359#else
Ingo Molnar36c8b582006-07-03 00:25:41 -07001360static inline int wake_idle(int cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361{
1362 return cpu;
1363}
1364#endif
1365
1366/***
1367 * try_to_wake_up - wake up a thread
1368 * @p: the to-be-woken-up thread
1369 * @state: the mask of task states that can be woken
1370 * @sync: do a synchronous wakeup?
1371 *
1372 * Put it on the run-queue if it's not already there. The "current"
1373 * thread is always on the run-queue (except when the actual
1374 * re-schedule is in progress), and as such you're allowed to do
1375 * the simpler "current->state = TASK_RUNNING" to mark yourself
1376 * runnable without the overhead of this.
1377 *
1378 * returns failure only if the task is already active.
1379 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001380static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381{
1382 int cpu, this_cpu, success = 0;
1383 unsigned long flags;
1384 long old_state;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001385 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386#ifdef CONFIG_SMP
Nick Piggin78979862005-06-25 14:57:13 -07001387 struct sched_domain *sd, *this_sd = NULL;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001388 unsigned long load, this_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 int new_cpu;
1390#endif
1391
1392 rq = task_rq_lock(p, &flags);
1393 old_state = p->state;
1394 if (!(old_state & state))
1395 goto out;
1396
1397 if (p->array)
1398 goto out_running;
1399
1400 cpu = task_cpu(p);
1401 this_cpu = smp_processor_id();
1402
1403#ifdef CONFIG_SMP
1404 if (unlikely(task_running(rq, p)))
1405 goto out_activate;
1406
Nick Piggin78979862005-06-25 14:57:13 -07001407 new_cpu = cpu;
1408
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 schedstat_inc(rq, ttwu_cnt);
1410 if (cpu == this_cpu) {
1411 schedstat_inc(rq, ttwu_local);
Nick Piggin78979862005-06-25 14:57:13 -07001412 goto out_set_cpu;
1413 }
1414
1415 for_each_domain(this_cpu, sd) {
1416 if (cpu_isset(cpu, sd->span)) {
1417 schedstat_inc(sd, ttwu_wake_remote);
1418 this_sd = sd;
1419 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 }
1421 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422
Nick Piggin78979862005-06-25 14:57:13 -07001423 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424 goto out_set_cpu;
1425
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 /*
Nick Piggin78979862005-06-25 14:57:13 -07001427 * Check for affine wakeup and passive balancing possibilities.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 */
Nick Piggin78979862005-06-25 14:57:13 -07001429 if (this_sd) {
1430 int idx = this_sd->wake_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 unsigned int imbalance;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432
Nick Piggina3f21bc2005-06-25 14:57:15 -07001433 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1434
Nick Piggin78979862005-06-25 14:57:13 -07001435 load = source_load(cpu, idx);
1436 this_load = target_load(this_cpu, idx);
1437
Nick Piggin78979862005-06-25 14:57:13 -07001438 new_cpu = this_cpu; /* Wake to this CPU if we can */
1439
Nick Piggina3f21bc2005-06-25 14:57:15 -07001440 if (this_sd->flags & SD_WAKE_AFFINE) {
1441 unsigned long tl = this_load;
Peter Williams2dd73a42006-06-27 02:54:34 -07001442 unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu);
1443
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 /*
Nick Piggina3f21bc2005-06-25 14:57:15 -07001445 * If sync wakeup then subtract the (maximum possible)
1446 * effect of the currently running task from the load
1447 * of the current CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448 */
Nick Piggina3f21bc2005-06-25 14:57:15 -07001449 if (sync)
Peter Williams2dd73a42006-06-27 02:54:34 -07001450 tl -= current->load_weight;
Nick Piggina3f21bc2005-06-25 14:57:15 -07001451
1452 if ((tl <= load &&
Peter Williams2dd73a42006-06-27 02:54:34 -07001453 tl + target_load(cpu, idx) <= tl_per_task) ||
1454 100*(tl + p->load_weight) <= imbalance*load) {
Nick Piggina3f21bc2005-06-25 14:57:15 -07001455 /*
1456 * This domain has SD_WAKE_AFFINE and
1457 * p is cache cold in this domain, and
1458 * there is no bad imbalance.
1459 */
1460 schedstat_inc(this_sd, ttwu_move_affine);
1461 goto out_set_cpu;
1462 }
1463 }
1464
1465 /*
1466 * Start passive balancing when half the imbalance_pct
1467 * limit is reached.
1468 */
1469 if (this_sd->flags & SD_WAKE_BALANCE) {
1470 if (imbalance*this_load <= 100*load) {
1471 schedstat_inc(this_sd, ttwu_move_balance);
1472 goto out_set_cpu;
1473 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 }
1475 }
1476
1477 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1478out_set_cpu:
1479 new_cpu = wake_idle(new_cpu, p);
1480 if (new_cpu != cpu) {
1481 set_task_cpu(p, new_cpu);
1482 task_rq_unlock(rq, &flags);
1483 /* might preempt at this point */
1484 rq = task_rq_lock(p, &flags);
1485 old_state = p->state;
1486 if (!(old_state & state))
1487 goto out;
1488 if (p->array)
1489 goto out_running;
1490
1491 this_cpu = smp_processor_id();
1492 cpu = task_cpu(p);
1493 }
1494
1495out_activate:
1496#endif /* CONFIG_SMP */
1497 if (old_state == TASK_UNINTERRUPTIBLE) {
1498 rq->nr_uninterruptible--;
1499 /*
1500 * Tasks on involuntary sleep don't earn
1501 * sleep_avg beyond just interactive state.
1502 */
Con Kolivas3dee3862006-03-31 02:31:23 -08001503 p->sleep_type = SLEEP_NONINTERACTIVE;
Con Kolivase7c38cb2006-03-31 02:31:25 -08001504 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505
1506 /*
Ingo Molnard79fc0f2005-09-10 00:26:12 -07001507 * Tasks that have marked their sleep as noninteractive get
Con Kolivase7c38cb2006-03-31 02:31:25 -08001508 * woken up with their sleep average not weighted in an
1509 * interactive way.
Ingo Molnard79fc0f2005-09-10 00:26:12 -07001510 */
Con Kolivase7c38cb2006-03-31 02:31:25 -08001511 if (old_state & TASK_NONINTERACTIVE)
1512 p->sleep_type = SLEEP_NONINTERACTIVE;
1513
1514
1515 activate_task(p, rq, cpu == this_cpu);
Ingo Molnard79fc0f2005-09-10 00:26:12 -07001516 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 * Sync wakeups (i.e. those types of wakeups where the waker
1518 * has indicated that it will leave the CPU in short order)
1519 * don't trigger a preemption, if the woken up task will run on
1520 * this cpu. (in this case the 'I will reschedule' promise of
1521 * the waker guarantees that the freshly woken up task is going
1522 * to be considered on this CPU.)
1523 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 if (!sync || cpu != this_cpu) {
1525 if (TASK_PREEMPTS_CURR(p, rq))
1526 resched_task(rq->curr);
1527 }
1528 success = 1;
1529
1530out_running:
1531 p->state = TASK_RUNNING;
1532out:
1533 task_rq_unlock(rq, &flags);
1534
1535 return success;
1536}
1537
Ingo Molnar36c8b582006-07-03 00:25:41 -07001538int fastcall wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539{
1540 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1541 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1542}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543EXPORT_SYMBOL(wake_up_process);
1544
Ingo Molnar36c8b582006-07-03 00:25:41 -07001545int fastcall wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546{
1547 return try_to_wake_up(p, state, 0);
1548}
1549
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550/*
1551 * Perform scheduler related setup for a newly forked process p.
1552 * p is forked by current.
1553 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001554void fastcall sched_fork(struct task_struct *p, int clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555{
Nick Piggin476d1392005-06-25 14:57:29 -07001556 int cpu = get_cpu();
1557
1558#ifdef CONFIG_SMP
1559 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1560#endif
1561 set_task_cpu(p, cpu);
1562
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 /*
1564 * We mark the process as running here, but have not actually
1565 * inserted it onto the runqueue yet. This guarantees that
1566 * nobody will actually run it, and a signal or other external
1567 * event cannot wake it up and insert it on the runqueue either.
1568 */
1569 p->state = TASK_RUNNING;
Ingo Molnarb29739f2006-06-27 02:54:51 -07001570
1571 /*
1572 * Make sure we do not leak PI boosting priority to the child:
1573 */
1574 p->prio = current->normal_prio;
1575
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576 INIT_LIST_HEAD(&p->run_list);
1577 p->array = NULL;
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07001578#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1579 if (unlikely(sched_info_on()))
1580 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08001582#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07001583 p->oncpu = 0;
1584#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07001586 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08001587 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588#endif
1589 /*
1590 * Share the timeslice between parent and child, thus the
1591 * total amount of pending timeslices in the system doesn't change,
1592 * resulting in more scheduling fairness.
1593 */
1594 local_irq_disable();
1595 p->time_slice = (current->time_slice + 1) >> 1;
1596 /*
1597 * The remainder of the first timeslice might be recovered by
1598 * the parent if the child exits early enough.
1599 */
1600 p->first_time_slice = 1;
1601 current->time_slice >>= 1;
1602 p->timestamp = sched_clock();
1603 if (unlikely(!current->time_slice)) {
1604 /*
1605 * This case is rare, it happens when the parent has only
1606 * a single jiffy left from its timeslice. Taking the
1607 * runqueue lock is not a problem.
1608 */
1609 current->time_slice = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 scheduler_tick();
Nick Piggin476d1392005-06-25 14:57:29 -07001611 }
1612 local_irq_enable();
1613 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614}
1615
1616/*
1617 * wake_up_new_task - wake up a newly created task for the first time.
1618 *
1619 * This function will do some initial scheduler statistics housekeeping
1620 * that must be done for every newly created context, then puts the task
1621 * on the runqueue and wakes it.
1622 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001623void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001625 struct rq *rq, *this_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 unsigned long flags;
1627 int this_cpu, cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628
1629 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 BUG_ON(p->state != TASK_RUNNING);
Nick Piggin147cbb42005-06-25 14:57:19 -07001631 this_cpu = smp_processor_id();
1632 cpu = task_cpu(p);
1633
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634 /*
1635 * We decrease the sleep average of forking parents
1636 * and children as well, to keep max-interactive tasks
1637 * from forking tasks that are max-interactive. The parent
1638 * (current) is done further down, under its lock.
1639 */
1640 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1641 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1642
1643 p->prio = effective_prio(p);
1644
1645 if (likely(cpu == this_cpu)) {
1646 if (!(clone_flags & CLONE_VM)) {
1647 /*
1648 * The VM isn't cloned, so we're in a good position to
1649 * do child-runs-first in anticipation of an exec. This
1650 * usually avoids a lot of COW overhead.
1651 */
1652 if (unlikely(!current->array))
1653 __activate_task(p, rq);
1654 else {
1655 p->prio = current->prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07001656 p->normal_prio = current->normal_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 list_add_tail(&p->run_list, &current->run_list);
1658 p->array = current->array;
1659 p->array->nr_active++;
Peter Williams2dd73a42006-06-27 02:54:34 -07001660 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 }
1662 set_need_resched();
1663 } else
1664 /* Run child last */
1665 __activate_task(p, rq);
1666 /*
1667 * We skip the following code due to cpu == this_cpu
1668 *
1669 * task_rq_unlock(rq, &flags);
1670 * this_rq = task_rq_lock(current, &flags);
1671 */
1672 this_rq = rq;
1673 } else {
1674 this_rq = cpu_rq(this_cpu);
1675
1676 /*
1677 * Not the local CPU - must adjust timestamp. This should
1678 * get optimised away in the !CONFIG_SMP case.
1679 */
1680 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1681 + rq->timestamp_last_tick;
1682 __activate_task(p, rq);
1683 if (TASK_PREEMPTS_CURR(p, rq))
1684 resched_task(rq->curr);
1685
1686 /*
1687 * Parent and child are on different CPUs, now get the
1688 * parent runqueue to update the parent's ->sleep_avg:
1689 */
1690 task_rq_unlock(rq, &flags);
1691 this_rq = task_rq_lock(current, &flags);
1692 }
1693 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1694 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1695 task_rq_unlock(this_rq, &flags);
1696}
1697
1698/*
1699 * Potentially available exiting-child timeslices are
1700 * retrieved here - this way the parent does not get
1701 * penalized for creating too many threads.
1702 *
1703 * (this cannot be used to 'generate' timeslices
1704 * artificially, because any timeslice recovered here
1705 * was given away by the parent in the first place.)
1706 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001707void fastcall sched_exit(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708{
1709 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001710 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711
1712 /*
1713 * If the child was a (relative-) CPU hog then decrease
1714 * the sleep_avg of the parent as well.
1715 */
1716 rq = task_rq_lock(p->parent, &flags);
Oleg Nesterov889dfaf2005-11-04 18:54:30 +03001717 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 p->parent->time_slice += p->time_slice;
1719 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1720 p->parent->time_slice = task_timeslice(p);
1721 }
1722 if (p->sleep_avg < p->parent->sleep_avg)
1723 p->parent->sleep_avg = p->parent->sleep_avg /
1724 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1725 (EXIT_WEIGHT + 1);
1726 task_rq_unlock(rq, &flags);
1727}
1728
1729/**
Nick Piggin4866cde2005-06-25 14:57:23 -07001730 * prepare_task_switch - prepare to switch tasks
1731 * @rq: the runqueue preparing to switch
1732 * @next: the task we are going to switch to.
1733 *
1734 * This is called with the rq lock held and interrupts off. It must
1735 * be paired with a subsequent finish_task_switch after the context
1736 * switch.
1737 *
1738 * prepare_task_switch sets up locking and calls architecture specific
1739 * hooks.
1740 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07001741static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07001742{
1743 prepare_lock_switch(rq, next);
1744 prepare_arch_switch(next);
1745}
1746
1747/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04001749 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 * @prev: the thread we just switched away from.
1751 *
Nick Piggin4866cde2005-06-25 14:57:23 -07001752 * finish_task_switch must be called after the context switch, paired
1753 * with a prepare_task_switch call before the context switch.
1754 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1755 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 *
1757 * Note that we may have delayed dropping an mm in context_switch(). If
1758 * so, we finish that here outside of the runqueue lock. (Doing it
1759 * with the lock held can cause deadlocks; see schedule() for
1760 * details.)
1761 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07001762static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 __releases(rq->lock)
1764{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07001766 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767
1768 rq->prev_mm = NULL;
1769
1770 /*
1771 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07001772 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07001773 * schedule one last time. The schedule call will never return, and
1774 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07001775 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 * still held, otherwise prev could be scheduled on another cpu, die
1777 * there before we look at prev->state, and then the reference would
1778 * be dropped twice.
1779 * Manfred Spraul <manfred@colorfullife.com>
1780 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07001781 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07001782 finish_arch_switch(prev);
1783 finish_lock_switch(rq, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 if (mm)
1785 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07001786 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08001787 /*
1788 * Remove function-return probe instances associated with this
1789 * task and put them back on the free list.
1790 */
1791 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08001793 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794}
1795
1796/**
1797 * schedule_tail - first thing a freshly forked thread must call.
1798 * @prev: the thread we just switched away from.
1799 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001800asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801 __releases(rq->lock)
1802{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001803 struct rq *rq = this_rq();
1804
Nick Piggin4866cde2005-06-25 14:57:23 -07001805 finish_task_switch(rq, prev);
1806#ifdef __ARCH_WANT_UNLOCKED_CTXSW
1807 /* In this case, finish_task_switch does not reenable preemption */
1808 preempt_enable();
1809#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 if (current->set_child_tid)
1811 put_user(current->pid, current->set_child_tid);
1812}
1813
1814/*
1815 * context_switch - switch to the new MM and the new
1816 * thread's register state.
1817 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001818static inline struct task_struct *
Ingo Molnar70b97a72006-07-03 00:25:42 -07001819context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07001820 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821{
1822 struct mm_struct *mm = next->mm;
1823 struct mm_struct *oldmm = prev->active_mm;
1824
Nick Pigginbeed33a2006-10-11 01:21:52 -07001825 if (!mm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826 next->active_mm = oldmm;
1827 atomic_inc(&oldmm->mm_count);
1828 enter_lazy_tlb(oldmm, next);
1829 } else
1830 switch_mm(oldmm, mm, next);
1831
Nick Pigginbeed33a2006-10-11 01:21:52 -07001832 if (!prev->mm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 prev->active_mm = NULL;
1834 WARN_ON(rq->prev_mm);
1835 rq->prev_mm = oldmm;
1836 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07001837 /*
1838 * Since the runqueue lock will be released by the next
1839 * task (which is an invalid locking op but in the case
1840 * of the scheduler it's an obvious special-case), so we
1841 * do an early lockdep release here:
1842 */
1843#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07001844 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07001845#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846
1847 /* Here we just switch the register state and the stack. */
1848 switch_to(prev, next, prev);
1849
1850 return prev;
1851}
1852
1853/*
1854 * nr_running, nr_uninterruptible and nr_context_switches:
1855 *
1856 * externally visible scheduler statistics: current number of runnable
1857 * threads, current number of uninterruptible-sleeping threads, total
1858 * number of context switches performed since bootup.
1859 */
1860unsigned long nr_running(void)
1861{
1862 unsigned long i, sum = 0;
1863
1864 for_each_online_cpu(i)
1865 sum += cpu_rq(i)->nr_running;
1866
1867 return sum;
1868}
1869
1870unsigned long nr_uninterruptible(void)
1871{
1872 unsigned long i, sum = 0;
1873
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08001874 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 sum += cpu_rq(i)->nr_uninterruptible;
1876
1877 /*
1878 * Since we read the counters lockless, it might be slightly
1879 * inaccurate. Do not allow it to go below zero though:
1880 */
1881 if (unlikely((long)sum < 0))
1882 sum = 0;
1883
1884 return sum;
1885}
1886
1887unsigned long long nr_context_switches(void)
1888{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07001889 int i;
1890 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08001892 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 sum += cpu_rq(i)->nr_switches;
1894
1895 return sum;
1896}
1897
1898unsigned long nr_iowait(void)
1899{
1900 unsigned long i, sum = 0;
1901
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08001902 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1904
1905 return sum;
1906}
1907
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08001908unsigned long nr_active(void)
1909{
1910 unsigned long i, running = 0, uninterruptible = 0;
1911
1912 for_each_online_cpu(i) {
1913 running += cpu_rq(i)->nr_running;
1914 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1915 }
1916
1917 if (unlikely((long)uninterruptible < 0))
1918 uninterruptible = 0;
1919
1920 return running + uninterruptible;
1921}
1922
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923#ifdef CONFIG_SMP
1924
1925/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07001926 * Is this task likely cache-hot:
1927 */
1928static inline int
1929task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1930{
1931 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1932}
1933
1934/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 * double_rq_lock - safely lock two runqueues
1936 *
1937 * Note this does not disable interrupts like task_rq_lock,
1938 * you need to do so manually before calling.
1939 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07001940static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941 __acquires(rq1->lock)
1942 __acquires(rq2->lock)
1943{
1944 if (rq1 == rq2) {
1945 spin_lock(&rq1->lock);
1946 __acquire(rq2->lock); /* Fake it out ;) */
1947 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001948 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 spin_lock(&rq1->lock);
1950 spin_lock(&rq2->lock);
1951 } else {
1952 spin_lock(&rq2->lock);
1953 spin_lock(&rq1->lock);
1954 }
1955 }
1956}
1957
1958/*
1959 * double_rq_unlock - safely unlock two runqueues
1960 *
1961 * Note this does not restore interrupts like task_rq_unlock,
1962 * you need to do so manually after calling.
1963 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07001964static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 __releases(rq1->lock)
1966 __releases(rq2->lock)
1967{
1968 spin_unlock(&rq1->lock);
1969 if (rq1 != rq2)
1970 spin_unlock(&rq2->lock);
1971 else
1972 __release(rq2->lock);
1973}
1974
1975/*
1976 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1977 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07001978static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 __releases(this_rq->lock)
1980 __acquires(busiest->lock)
1981 __acquires(this_rq->lock)
1982{
1983 if (unlikely(!spin_trylock(&busiest->lock))) {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001984 if (busiest < this_rq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 spin_unlock(&this_rq->lock);
1986 spin_lock(&busiest->lock);
1987 spin_lock(&this_rq->lock);
1988 } else
1989 spin_lock(&busiest->lock);
1990 }
1991}
1992
1993/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 * If dest_cpu is allowed for this process, migrate the task to it.
1995 * This is accomplished by forcing the cpu_allowed mask to only
1996 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1997 * the cpu_allowed mask is restored.
1998 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001999static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002001 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002003 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004
2005 rq = task_rq_lock(p, &flags);
2006 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2007 || unlikely(cpu_is_offline(dest_cpu)))
2008 goto out;
2009
2010 /* force the process onto the specified CPU */
2011 if (migrate_task(p, dest_cpu, &req)) {
2012 /* Need to wait for migration thread (might exit: take ref). */
2013 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07002014
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 get_task_struct(mt);
2016 task_rq_unlock(rq, &flags);
2017 wake_up_process(mt);
2018 put_task_struct(mt);
2019 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07002020
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021 return;
2022 }
2023out:
2024 task_rq_unlock(rq, &flags);
2025}
2026
2027/*
Nick Piggin476d1392005-06-25 14:57:29 -07002028 * sched_exec - execve() is a valuable balancing opportunity, because at
2029 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 */
2031void sched_exec(void)
2032{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 int new_cpu, this_cpu = get_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002034 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002036 if (new_cpu != this_cpu)
2037 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038}
2039
2040/*
2041 * pull_task - move a task from a remote runqueue to the local runqueue.
2042 * Both runqueues must be locked.
2043 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002044static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2045 struct task_struct *p, struct rq *this_rq,
2046 struct prio_array *this_array, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047{
2048 dequeue_task(p, src_array);
Peter Williams2dd73a42006-06-27 02:54:34 -07002049 dec_nr_running(p, src_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 set_task_cpu(p, this_cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07002051 inc_nr_running(p, this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 enqueue_task(p, this_array);
2053 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
2054 + this_rq->timestamp_last_tick;
2055 /*
2056 * Note that idle threads have a prio of MAX_PRIO, for this test
2057 * to be always true for them.
2058 */
2059 if (TASK_PREEMPTS_CURR(p, this_rq))
2060 resched_task(this_rq->curr);
2061}
2062
2063/*
2064 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2065 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08002066static
Ingo Molnar70b97a72006-07-03 00:25:42 -07002067int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002068 struct sched_domain *sd, enum idle_type idle,
2069 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070{
2071 /*
2072 * We do not migrate tasks that are:
2073 * 1) running (obviously), or
2074 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2075 * 3) are cache-hot on their current CPU.
2076 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 if (!cpu_isset(this_cpu, p->cpus_allowed))
2078 return 0;
Nick Piggin81026792005-06-25 14:57:07 -07002079 *all_pinned = 0;
2080
2081 if (task_running(rq, p))
2082 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083
2084 /*
2085 * Aggressive migration if:
Nick Piggincafb20c2005-06-25 14:57:17 -07002086 * 1) task is cache cold, or
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 * 2) too many balance attempts have failed.
2088 */
2089
Nick Piggincafb20c2005-06-25 14:57:17 -07002090 if (sd->nr_balance_failed > sd->cache_nice_tries)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 return 1;
2092
2093 if (task_hot(p, rq->timestamp_last_tick, sd))
Nick Piggin81026792005-06-25 14:57:07 -07002094 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 return 1;
2096}
2097
Peter Williams615052d2006-06-27 02:54:37 -07002098#define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002099
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100/*
Peter Williams2dd73a42006-06-27 02:54:34 -07002101 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2102 * load from busiest to this_rq, as part of a balancing operation within
2103 * "domain". Returns the number of tasks moved.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 *
2105 * Called with both runqueues locked.
2106 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002107static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams2dd73a42006-06-27 02:54:34 -07002108 unsigned long max_nr_move, unsigned long max_load_move,
2109 struct sched_domain *sd, enum idle_type idle,
2110 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111{
Ingo Molnar48f24c42006-07-03 00:25:40 -07002112 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2113 best_prio_seen, skip_for_load;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002114 struct prio_array *array, *dst_array;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 struct list_head *head, *curr;
Ingo Molnar36c8b582006-07-03 00:25:41 -07002116 struct task_struct *tmp;
Peter Williams2dd73a42006-06-27 02:54:34 -07002117 long rem_load_move;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118
Peter Williams2dd73a42006-06-27 02:54:34 -07002119 if (max_nr_move == 0 || max_load_move == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 goto out;
2121
Peter Williams2dd73a42006-06-27 02:54:34 -07002122 rem_load_move = max_load_move;
Nick Piggin81026792005-06-25 14:57:07 -07002123 pinned = 1;
Peter Williams615052d2006-06-27 02:54:37 -07002124 this_best_prio = rq_best_prio(this_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002125 best_prio = rq_best_prio(busiest);
Peter Williams615052d2006-06-27 02:54:37 -07002126 /*
2127 * Enable handling of the case where there is more than one task
2128 * with the best priority. If the current running task is one
Ingo Molnar48f24c42006-07-03 00:25:40 -07002129 * of those with prio==best_prio we know it won't be moved
Peter Williams615052d2006-06-27 02:54:37 -07002130 * and therefore it's safe to override the skip (based on load) of
2131 * any task we find with that prio.
2132 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07002133 best_prio_seen = best_prio == busiest->curr->prio;
Nick Piggin81026792005-06-25 14:57:07 -07002134
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135 /*
2136 * We first consider expired tasks. Those will likely not be
2137 * executed in the near future, and they are most likely to
2138 * be cache-cold, thus switching CPUs has the least effect
2139 * on them.
2140 */
2141 if (busiest->expired->nr_active) {
2142 array = busiest->expired;
2143 dst_array = this_rq->expired;
2144 } else {
2145 array = busiest->active;
2146 dst_array = this_rq->active;
2147 }
2148
2149new_array:
2150 /* Start searching at priority 0: */
2151 idx = 0;
2152skip_bitmap:
2153 if (!idx)
2154 idx = sched_find_first_bit(array->bitmap);
2155 else
2156 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2157 if (idx >= MAX_PRIO) {
2158 if (array == busiest->expired && busiest->active->nr_active) {
2159 array = busiest->active;
2160 dst_array = this_rq->active;
2161 goto new_array;
2162 }
2163 goto out;
2164 }
2165
2166 head = array->queue + idx;
2167 curr = head->prev;
2168skip_queue:
Ingo Molnar36c8b582006-07-03 00:25:41 -07002169 tmp = list_entry(curr, struct task_struct, run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170
2171 curr = curr->prev;
2172
Peter Williams50ddd962006-06-27 02:54:36 -07002173 /*
2174 * To help distribute high priority tasks accross CPUs we don't
2175 * skip a task if it will be the highest priority task (i.e. smallest
2176 * prio value) on its new queue regardless of its load weight
2177 */
Peter Williams615052d2006-06-27 02:54:37 -07002178 skip_for_load = tmp->load_weight > rem_load_move;
2179 if (skip_for_load && idx < this_best_prio)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002180 skip_for_load = !best_prio_seen && idx == best_prio;
Peter Williams615052d2006-06-27 02:54:37 -07002181 if (skip_for_load ||
Peter Williams2dd73a42006-06-27 02:54:34 -07002182 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07002183
2184 best_prio_seen |= idx == best_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 if (curr != head)
2186 goto skip_queue;
2187 idx++;
2188 goto skip_bitmap;
2189 }
2190
2191#ifdef CONFIG_SCHEDSTATS
2192 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
2193 schedstat_inc(sd, lb_hot_gained[idle]);
2194#endif
2195
2196 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2197 pulled++;
Peter Williams2dd73a42006-06-27 02:54:34 -07002198 rem_load_move -= tmp->load_weight;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199
Peter Williams2dd73a42006-06-27 02:54:34 -07002200 /*
2201 * We only want to steal up to the prescribed number of tasks
2202 * and the prescribed amount of weighted load.
2203 */
2204 if (pulled < max_nr_move && rem_load_move > 0) {
Peter Williams615052d2006-06-27 02:54:37 -07002205 if (idx < this_best_prio)
2206 this_best_prio = idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 if (curr != head)
2208 goto skip_queue;
2209 idx++;
2210 goto skip_bitmap;
2211 }
2212out:
2213 /*
2214 * Right now, this is the only place pull_task() is called,
2215 * so we can safely collect pull_task() stats here rather than
2216 * inside pull_task().
2217 */
2218 schedstat_add(sd, lb_gained[idle], pulled);
Nick Piggin81026792005-06-25 14:57:07 -07002219
2220 if (all_pinned)
2221 *all_pinned = pinned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 return pulled;
2223}
2224
2225/*
2226 * find_busiest_group finds and returns the busiest CPU group within the
Ingo Molnar48f24c42006-07-03 00:25:40 -07002227 * domain. It calculates and returns the amount of weighted load which
2228 * should be moved to restore balance via the imbalance parameter.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 */
2230static struct sched_group *
2231find_busiest_group(struct sched_domain *sd, int this_cpu,
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002232 unsigned long *imbalance, enum idle_type idle, int *sd_idle,
2233 cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234{
2235 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2236 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002237 unsigned long max_pull;
Peter Williams2dd73a42006-06-27 02:54:34 -07002238 unsigned long busiest_load_per_task, busiest_nr_running;
2239 unsigned long this_load_per_task, this_nr_running;
Nick Piggin78979862005-06-25 14:57:13 -07002240 int load_idx;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002241#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2242 int power_savings_balance = 1;
2243 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2244 unsigned long min_nr_running = ULONG_MAX;
2245 struct sched_group *group_min = NULL, *group_leader = NULL;
2246#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247
2248 max_load = this_load = total_load = total_pwr = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07002249 busiest_load_per_task = busiest_nr_running = 0;
2250 this_load_per_task = this_nr_running = 0;
Nick Piggin78979862005-06-25 14:57:13 -07002251 if (idle == NOT_IDLE)
2252 load_idx = sd->busy_idx;
2253 else if (idle == NEWLY_IDLE)
2254 load_idx = sd->newidle_idx;
2255 else
2256 load_idx = sd->idle_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257
2258 do {
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002259 unsigned long load, group_capacity;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 int local_group;
2261 int i;
Peter Williams2dd73a42006-06-27 02:54:34 -07002262 unsigned long sum_nr_running, sum_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263
2264 local_group = cpu_isset(this_cpu, group->cpumask);
2265
2266 /* Tally up the load of all CPUs in the group */
Peter Williams2dd73a42006-06-27 02:54:34 -07002267 sum_weighted_load = sum_nr_running = avg_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268
2269 for_each_cpu_mask(i, group->cpumask) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002270 struct rq *rq;
2271
2272 if (!cpu_isset(i, *cpus))
2273 continue;
2274
2275 rq = cpu_rq(i);
Peter Williams2dd73a42006-06-27 02:54:34 -07002276
Nick Piggin5969fe02005-09-10 00:26:19 -07002277 if (*sd_idle && !idle_cpu(i))
2278 *sd_idle = 0;
2279
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 /* Bias balancing toward cpus of our domain */
2281 if (local_group)
Nick Piggina2000572006-02-10 01:51:02 -08002282 load = target_load(i, load_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 else
Nick Piggina2000572006-02-10 01:51:02 -08002284 load = source_load(i, load_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285
2286 avg_load += load;
Peter Williams2dd73a42006-06-27 02:54:34 -07002287 sum_nr_running += rq->nr_running;
2288 sum_weighted_load += rq->raw_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 }
2290
2291 total_load += avg_load;
2292 total_pwr += group->cpu_power;
2293
2294 /* Adjust by relative CPU power of the group */
2295 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2296
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002297 group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
2298
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 if (local_group) {
2300 this_load = avg_load;
2301 this = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07002302 this_nr_running = sum_nr_running;
2303 this_load_per_task = sum_weighted_load;
2304 } else if (avg_load > max_load &&
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002305 sum_nr_running > group_capacity) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306 max_load = avg_load;
2307 busiest = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07002308 busiest_nr_running = sum_nr_running;
2309 busiest_load_per_task = sum_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002311
2312#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2313 /*
2314 * Busy processors will not participate in power savings
2315 * balance.
2316 */
2317 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2318 goto group_next;
2319
2320 /*
2321 * If the local group is idle or completely loaded
2322 * no need to do power savings balance at this domain
2323 */
2324 if (local_group && (this_nr_running >= group_capacity ||
2325 !this_nr_running))
2326 power_savings_balance = 0;
2327
2328 /*
2329 * If a group is already running at full capacity or idle,
2330 * don't include that group in power savings calculations
2331 */
2332 if (!power_savings_balance || sum_nr_running >= group_capacity
2333 || !sum_nr_running)
2334 goto group_next;
2335
2336 /*
2337 * Calculate the group which has the least non-idle load.
2338 * This is the group from where we need to pick up the load
2339 * for saving power
2340 */
2341 if ((sum_nr_running < min_nr_running) ||
2342 (sum_nr_running == min_nr_running &&
2343 first_cpu(group->cpumask) <
2344 first_cpu(group_min->cpumask))) {
2345 group_min = group;
2346 min_nr_running = sum_nr_running;
2347 min_load_per_task = sum_weighted_load /
2348 sum_nr_running;
2349 }
2350
2351 /*
2352 * Calculate the group which is almost near its
2353 * capacity but still has some space to pick up some load
2354 * from other group and save more power
2355 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07002356 if (sum_nr_running <= group_capacity - 1) {
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002357 if (sum_nr_running > leader_nr_running ||
2358 (sum_nr_running == leader_nr_running &&
2359 first_cpu(group->cpumask) >
2360 first_cpu(group_leader->cpumask))) {
2361 group_leader = group;
2362 leader_nr_running = sum_nr_running;
2363 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002364 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002365group_next:
2366#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 group = group->next;
2368 } while (group != sd->groups);
2369
Peter Williams2dd73a42006-06-27 02:54:34 -07002370 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 goto out_balanced;
2372
2373 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2374
2375 if (this_load >= avg_load ||
2376 100*max_load <= sd->imbalance_pct*this_load)
2377 goto out_balanced;
2378
Peter Williams2dd73a42006-06-27 02:54:34 -07002379 busiest_load_per_task /= busiest_nr_running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 /*
2381 * We're trying to get all the cpus to the average_load, so we don't
2382 * want to push ourselves above the average load, nor do we wish to
2383 * reduce the max loaded cpu below the average load, as either of these
2384 * actions would just result in more rebalancing later, and ping-pong
2385 * tasks around. Thus we look for the minimum possible imbalance.
2386 * Negative imbalances (*we* are more loaded than anyone else) will
2387 * be counted as no imbalance for these purposes -- we can't fix that
2388 * by pulling tasks to us. Be careful of negative numbers as they'll
2389 * appear as very large values with unsigned longs.
2390 */
Peter Williams2dd73a42006-06-27 02:54:34 -07002391 if (max_load <= busiest_load_per_task)
2392 goto out_balanced;
2393
2394 /*
2395 * In the presence of smp nice balancing, certain scenarios can have
2396 * max load less than avg load(as we skip the groups at or below
2397 * its cpu_power, while calculating max_load..)
2398 */
2399 if (max_load < avg_load) {
2400 *imbalance = 0;
2401 goto small_imbalance;
2402 }
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002403
2404 /* Don't want to pull so many tasks that a group would go idle */
Peter Williams2dd73a42006-06-27 02:54:34 -07002405 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002406
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 /* How much load to actually move to equalise the imbalance */
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002408 *imbalance = min(max_pull * busiest->cpu_power,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 (avg_load - this_load) * this->cpu_power)
2410 / SCHED_LOAD_SCALE;
2411
Peter Williams2dd73a42006-06-27 02:54:34 -07002412 /*
2413 * if *imbalance is less than the average load per runnable task
2414 * there is no gaurantee that any tasks will be moved so we'll have
2415 * a think about bumping its value to force at least one task to be
2416 * moved
2417 */
2418 if (*imbalance < busiest_load_per_task) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07002419 unsigned long tmp, pwr_now, pwr_move;
Peter Williams2dd73a42006-06-27 02:54:34 -07002420 unsigned int imbn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421
Peter Williams2dd73a42006-06-27 02:54:34 -07002422small_imbalance:
2423 pwr_move = pwr_now = 0;
2424 imbn = 2;
2425 if (this_nr_running) {
2426 this_load_per_task /= this_nr_running;
2427 if (busiest_load_per_task > this_load_per_task)
2428 imbn = 1;
2429 } else
2430 this_load_per_task = SCHED_LOAD_SCALE;
2431
2432 if (max_load - this_load >= busiest_load_per_task * imbn) {
2433 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 return busiest;
2435 }
2436
2437 /*
2438 * OK, we don't have enough imbalance to justify moving tasks,
2439 * however we may be able to increase total CPU power used by
2440 * moving them.
2441 */
2442
Peter Williams2dd73a42006-06-27 02:54:34 -07002443 pwr_now += busiest->cpu_power *
2444 min(busiest_load_per_task, max_load);
2445 pwr_now += this->cpu_power *
2446 min(this_load_per_task, this_load);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 pwr_now /= SCHED_LOAD_SCALE;
2448
2449 /* Amount of load we'd subtract */
Peter Williams2dd73a42006-06-27 02:54:34 -07002450 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 if (max_load > tmp)
Peter Williams2dd73a42006-06-27 02:54:34 -07002452 pwr_move += busiest->cpu_power *
2453 min(busiest_load_per_task, max_load - tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454
2455 /* Amount of load we'd add */
2456 if (max_load*busiest->cpu_power <
Peter Williams2dd73a42006-06-27 02:54:34 -07002457 busiest_load_per_task*SCHED_LOAD_SCALE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 tmp = max_load*busiest->cpu_power/this->cpu_power;
2459 else
Peter Williams2dd73a42006-06-27 02:54:34 -07002460 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power;
2461 pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 pwr_move /= SCHED_LOAD_SCALE;
2463
2464 /* Move if we gain throughput */
2465 if (pwr_move <= pwr_now)
2466 goto out_balanced;
2467
Peter Williams2dd73a42006-06-27 02:54:34 -07002468 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469 }
2470
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471 return busiest;
2472
2473out_balanced:
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002474#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2475 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2476 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002478 if (this == group_leader && group_leader != group_min) {
2479 *imbalance = min_load_per_task;
2480 return group_min;
2481 }
2482ret:
2483#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 *imbalance = 0;
2485 return NULL;
2486}
2487
2488/*
2489 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2490 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002491static struct rq *
Ingo Molnar48f24c42006-07-03 00:25:40 -07002492find_busiest_queue(struct sched_group *group, enum idle_type idle,
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002493 unsigned long imbalance, cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002495 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07002496 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 int i;
2498
2499 for_each_cpu_mask(i, group->cpumask) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002500
2501 if (!cpu_isset(i, *cpus))
2502 continue;
2503
Ingo Molnar48f24c42006-07-03 00:25:40 -07002504 rq = cpu_rq(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505
Ingo Molnar48f24c42006-07-03 00:25:40 -07002506 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07002507 continue;
2508
Ingo Molnar48f24c42006-07-03 00:25:40 -07002509 if (rq->raw_weighted_load > max_load) {
2510 max_load = rq->raw_weighted_load;
2511 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 }
2513 }
2514
2515 return busiest;
2516}
2517
2518/*
Nick Piggin77391d72005-06-25 14:57:30 -07002519 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2520 * so long as it is large enough.
2521 */
2522#define MAX_PINNED_INTERVAL 512
2523
Ingo Molnar48f24c42006-07-03 00:25:40 -07002524static inline unsigned long minus_1_or_zero(unsigned long n)
2525{
2526 return n > 0 ? n - 1 : 0;
2527}
2528
Nick Piggin77391d72005-06-25 14:57:30 -07002529/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2531 * tasks if there is an imbalance.
2532 *
2533 * Called with this_rq unlocked.
2534 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002535static int load_balance(int this_cpu, struct rq *this_rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536 struct sched_domain *sd, enum idle_type idle)
2537{
Ingo Molnar48f24c42006-07-03 00:25:40 -07002538 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002541 struct rq *busiest;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002542 cpumask_t cpus = CPU_MASK_ALL;
Nick Piggin5969fe02005-09-10 00:26:19 -07002543
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002544 /*
2545 * When power savings policy is enabled for the parent domain, idle
2546 * sibling can pick up load irrespective of busy siblings. In this case,
2547 * let the state of idle sibling percolate up as IDLE, instead of
2548 * portraying it as NOT_IDLE.
2549 */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002550 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002551 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002552 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 schedstat_inc(sd, lb_cnt[idle]);
2555
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002556redo:
2557 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2558 &cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 if (!group) {
2560 schedstat_inc(sd, lb_nobusyg[idle]);
2561 goto out_balanced;
2562 }
2563
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002564 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 if (!busiest) {
2566 schedstat_inc(sd, lb_nobusyq[idle]);
2567 goto out_balanced;
2568 }
2569
Nick Piggindb935db2005-06-25 14:57:11 -07002570 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571
2572 schedstat_add(sd, lb_imbalance[idle], imbalance);
2573
2574 nr_moved = 0;
2575 if (busiest->nr_running > 1) {
2576 /*
2577 * Attempt to move tasks. If find_busiest_group has found
2578 * an imbalance but busiest->nr_running <= 1, the group is
2579 * still unbalanced. nr_moved simply stays zero, so it is
2580 * correctly treated as an imbalance.
2581 */
Nick Piggine17224b2005-09-10 00:26:18 -07002582 double_rq_lock(this_rq, busiest);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583 nr_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07002584 minus_1_or_zero(busiest->nr_running),
2585 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07002586 double_rq_unlock(this_rq, busiest);
Nick Piggin81026792005-06-25 14:57:07 -07002587
2588 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002589 if (unlikely(all_pinned)) {
2590 cpu_clear(cpu_of(busiest), cpus);
2591 if (!cpus_empty(cpus))
2592 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07002593 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002594 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 }
Nick Piggin81026792005-06-25 14:57:07 -07002596
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 if (!nr_moved) {
2598 schedstat_inc(sd, lb_failed[idle]);
2599 sd->nr_balance_failed++;
2600
2601 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602
2603 spin_lock(&busiest->lock);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07002604
2605 /* don't kick the migration_thread, if the curr
2606 * task on busiest cpu can't be moved to this_cpu
2607 */
2608 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2609 spin_unlock(&busiest->lock);
2610 all_pinned = 1;
2611 goto out_one_pinned;
2612 }
2613
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 if (!busiest->active_balance) {
2615 busiest->active_balance = 1;
2616 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07002617 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002618 }
2619 spin_unlock(&busiest->lock);
Nick Piggin81026792005-06-25 14:57:07 -07002620 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 wake_up_process(busiest->migration_thread);
2622
2623 /*
2624 * We've kicked active balancing, reset the failure
2625 * counter.
2626 */
Nick Piggin39507452005-06-25 14:57:09 -07002627 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 }
Nick Piggin81026792005-06-25 14:57:07 -07002629 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630 sd->nr_balance_failed = 0;
2631
Nick Piggin81026792005-06-25 14:57:07 -07002632 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 /* We were unbalanced, so reset the balancing interval */
2634 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07002635 } else {
2636 /*
2637 * If we've begun active balancing, start to back off. This
2638 * case may not be covered by the all_pinned logic if there
2639 * is only 1 task on the busy runqueue (because we don't call
2640 * move_tasks).
2641 */
2642 if (sd->balance_interval < sd->max_interval)
2643 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 }
2645
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002646 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002647 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002648 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 return nr_moved;
2650
2651out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 schedstat_inc(sd, lb_balanced[idle]);
2653
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002654 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07002655
2656out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07002658 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2659 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 sd->balance_interval *= 2;
2661
Ingo Molnar48f24c42006-07-03 00:25:40 -07002662 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002663 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002664 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 return 0;
2666}
2667
2668/*
2669 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2670 * tasks if there is an imbalance.
2671 *
2672 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2673 * this_rq is locked.
2674 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07002675static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07002676load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677{
2678 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002679 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 unsigned long imbalance;
2681 int nr_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07002682 int sd_idle = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002683 cpumask_t cpus = CPU_MASK_ALL;
Nick Piggin5969fe02005-09-10 00:26:19 -07002684
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002685 /*
2686 * When power savings policy is enabled for the parent domain, idle
2687 * sibling can pick up load irrespective of busy siblings. In this case,
2688 * let the state of idle sibling percolate up as IDLE, instead of
2689 * portraying it as NOT_IDLE.
2690 */
2691 if (sd->flags & SD_SHARE_CPUPOWER &&
2692 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002693 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694
2695 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002696redo:
2697 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2698 &sd_idle, &cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699 if (!group) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002701 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702 }
2703
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002704 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2705 &cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07002706 if (!busiest) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002708 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 }
2710
Nick Piggindb935db2005-06-25 14:57:11 -07002711 BUG_ON(busiest == this_rq);
2712
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07002714
2715 nr_moved = 0;
2716 if (busiest->nr_running > 1) {
2717 /* Attempt to move tasks */
2718 double_lock_balance(this_rq, busiest);
2719 nr_moved = move_tasks(this_rq, this_cpu, busiest,
Peter Williams2dd73a42006-06-27 02:54:34 -07002720 minus_1_or_zero(busiest->nr_running),
Nick Piggin81026792005-06-25 14:57:07 -07002721 imbalance, sd, NEWLY_IDLE, NULL);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07002722 spin_unlock(&busiest->lock);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002723
2724 if (!nr_moved) {
2725 cpu_clear(cpu_of(busiest), cpus);
2726 if (!cpus_empty(cpus))
2727 goto redo;
2728 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07002729 }
2730
Nick Piggin5969fe02005-09-10 00:26:19 -07002731 if (!nr_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002733 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2734 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002735 return -1;
2736 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002737 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739 return nr_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002740
2741out_balanced:
2742 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002743 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002744 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002745 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002746 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002747
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002748 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749}
2750
2751/*
2752 * idle_balance is called by schedule() if this_cpu is about to become
2753 * idle. Attempts to pull tasks from other CPUs.
2754 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002755static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756{
2757 struct sched_domain *sd;
2758
2759 for_each_domain(this_cpu, sd) {
2760 if (sd->flags & SD_BALANCE_NEWIDLE) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07002761 /* If we've pulled tasks over stop searching: */
2762 if (load_balance_newidle(this_cpu, this_rq, sd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 }
2765 }
2766}
2767
2768/*
2769 * active_load_balance is run by migration threads. It pushes running tasks
2770 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2771 * running on each physical CPU where possible, and avoids physical /
2772 * logical imbalances.
2773 *
2774 * Called with busiest_rq locked.
2775 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002776static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777{
Nick Piggin39507452005-06-25 14:57:09 -07002778 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002779 struct sched_domain *sd;
2780 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07002781
Ingo Molnar48f24c42006-07-03 00:25:40 -07002782 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07002783 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07002784 return;
2785
2786 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787
2788 /*
Nick Piggin39507452005-06-25 14:57:09 -07002789 * This condition is "impossible", if it occurs
2790 * we need to fix it. Originally reported by
2791 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 */
Nick Piggin39507452005-06-25 14:57:09 -07002793 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794
Nick Piggin39507452005-06-25 14:57:09 -07002795 /* move a task from busiest_rq to target_rq */
2796 double_lock_balance(busiest_rq, target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797
Nick Piggin39507452005-06-25 14:57:09 -07002798 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002799 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07002800 if ((sd->flags & SD_LOAD_BALANCE) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07002801 cpu_isset(busiest_cpu, sd->span))
Nick Piggin39507452005-06-25 14:57:09 -07002802 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002803 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804
Ingo Molnar48f24c42006-07-03 00:25:40 -07002805 if (likely(sd)) {
2806 schedstat_inc(sd, alb_cnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807
Ingo Molnar48f24c42006-07-03 00:25:40 -07002808 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2809 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2810 NULL))
2811 schedstat_inc(sd, alb_pushed);
2812 else
2813 schedstat_inc(sd, alb_failed);
2814 }
Nick Piggin39507452005-06-25 14:57:09 -07002815 spin_unlock(&target_rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816}
2817
2818/*
2819 * rebalance_tick will get called every timer tick, on every CPU.
2820 *
2821 * It checks each scheduling domain to see if it is due to be balanced,
2822 * and initiates a balancing operation if so.
2823 *
2824 * Balancing parameters are set up in arch_init_sched_domains.
2825 */
2826
Ingo Molnar48f24c42006-07-03 00:25:40 -07002827/* Don't have all balancing operations going off at once: */
2828static inline unsigned long cpu_offset(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829{
Ingo Molnar48f24c42006-07-03 00:25:40 -07002830 return jiffies + cpu * HZ / NR_CPUS;
2831}
2832
2833static void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002834rebalance_tick(int this_cpu, struct rq *this_rq, enum idle_type idle)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002835{
2836 unsigned long this_load, interval, j = cpu_offset(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837 struct sched_domain *sd;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002838 int i, scale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839
Peter Williams2dd73a42006-06-27 02:54:34 -07002840 this_load = this_rq->raw_weighted_load;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002841
2842 /* Update our load: */
2843 for (i = 0, scale = 1; i < 3; i++, scale <<= 1) {
2844 unsigned long old_load, new_load;
2845
Nick Piggin78979862005-06-25 14:57:13 -07002846 old_load = this_rq->cpu_load[i];
Ingo Molnar48f24c42006-07-03 00:25:40 -07002847 new_load = this_load;
Nick Piggin78979862005-06-25 14:57:13 -07002848 /*
2849 * Round up the averaging division if load is increasing. This
2850 * prevents us from getting stuck on 9 if the load is 10, for
2851 * example.
2852 */
2853 if (new_load > old_load)
2854 new_load += scale-1;
2855 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2856 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857
2858 for_each_domain(this_cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 if (!(sd->flags & SD_LOAD_BALANCE))
2860 continue;
2861
2862 interval = sd->balance_interval;
2863 if (idle != SCHED_IDLE)
2864 interval *= sd->busy_factor;
2865
2866 /* scale ms to jiffies */
2867 interval = msecs_to_jiffies(interval);
2868 if (unlikely(!interval))
2869 interval = 1;
2870
2871 if (j - sd->last_balance >= interval) {
2872 if (load_balance(this_cpu, this_rq, sd, idle)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07002873 /*
2874 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07002875 * longer idle, or one of our SMT siblings is
2876 * not idle.
2877 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 idle = NOT_IDLE;
2879 }
2880 sd->last_balance += interval;
2881 }
2882 }
2883}
2884#else
2885/*
2886 * on UP we do not need to balance between CPUs:
2887 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002888static inline void rebalance_tick(int cpu, struct rq *rq, enum idle_type idle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889{
2890}
Ingo Molnar70b97a72006-07-03 00:25:42 -07002891static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892{
2893}
2894#endif
2895
Ingo Molnar70b97a72006-07-03 00:25:42 -07002896static inline int wake_priority_sleeper(struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897{
2898 int ret = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002899
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900#ifdef CONFIG_SCHED_SMT
2901 spin_lock(&rq->lock);
2902 /*
2903 * If an SMT sibling task has been put to sleep for priority
2904 * reasons reschedule the idle task to see if it can now run.
2905 */
2906 if (rq->nr_running) {
2907 resched_task(rq->idle);
2908 ret = 1;
2909 }
2910 spin_unlock(&rq->lock);
2911#endif
2912 return ret;
2913}
2914
2915DEFINE_PER_CPU(struct kernel_stat, kstat);
2916
2917EXPORT_PER_CPU_SYMBOL(kstat);
2918
2919/*
2920 * This is called on clock ticks and on context switches.
2921 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2922 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07002923static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002924update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925{
Ingo Molnar48f24c42006-07-03 00:25:40 -07002926 p->sched_time += now - max(p->timestamp, rq->timestamp_last_tick);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927}
2928
2929/*
2930 * Return current->sched_time plus any more ns on the sched_clock
2931 * that have not yet been banked.
2932 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002933unsigned long long current_sched_time(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934{
2935 unsigned long long ns;
2936 unsigned long flags;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002937
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 local_irq_save(flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002939 ns = max(p->timestamp, task_rq(p)->timestamp_last_tick);
2940 ns = p->sched_time + sched_clock() - ns;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 local_irq_restore(flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002942
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 return ns;
2944}
2945
2946/*
Linus Torvaldsf1adad72006-05-21 18:54:09 -07002947 * We place interactive tasks back into the active array, if possible.
2948 *
2949 * To guarantee that this does not starve expired tasks we ignore the
2950 * interactivity of a task if the first expired task had to wait more
2951 * than a 'reasonable' amount of time. This deadline timeout is
2952 * load-dependent, as the frequency of array switched decreases with
2953 * increasing number of running tasks. We also ignore the interactivity
2954 * if a better static_prio task has expired:
2955 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002956static inline int expired_starving(struct rq *rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002957{
2958 if (rq->curr->static_prio > rq->best_expired_prio)
2959 return 1;
2960 if (!STARVATION_LIMIT || !rq->expired_timestamp)
2961 return 0;
2962 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
2963 return 1;
2964 return 0;
2965}
Linus Torvaldsf1adad72006-05-21 18:54:09 -07002966
2967/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002968 * Account user cpu time to a process.
2969 * @p: the process that the cpu time gets accounted to
2970 * @hardirq_offset: the offset to subtract from hardirq_count()
2971 * @cputime: the cpu time spent in user space since the last update
2972 */
2973void account_user_time(struct task_struct *p, cputime_t cputime)
2974{
2975 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2976 cputime64_t tmp;
2977
2978 p->utime = cputime_add(p->utime, cputime);
2979
2980 /* Add user time to cpustat. */
2981 tmp = cputime_to_cputime64(cputime);
2982 if (TASK_NICE(p) > 0)
2983 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2984 else
2985 cpustat->user = cputime64_add(cpustat->user, tmp);
2986}
2987
2988/*
2989 * Account system cpu time to a process.
2990 * @p: the process that the cpu time gets accounted to
2991 * @hardirq_offset: the offset to subtract from hardirq_count()
2992 * @cputime: the cpu time spent in kernel space since the last update
2993 */
2994void account_system_time(struct task_struct *p, int hardirq_offset,
2995 cputime_t cputime)
2996{
2997 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002998 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 cputime64_t tmp;
3000
3001 p->stime = cputime_add(p->stime, cputime);
3002
3003 /* Add system time to cpustat. */
3004 tmp = cputime_to_cputime64(cputime);
3005 if (hardirq_count() - hardirq_offset)
3006 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3007 else if (softirq_count())
3008 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3009 else if (p != rq->idle)
3010 cpustat->system = cputime64_add(cpustat->system, tmp);
3011 else if (atomic_read(&rq->nr_iowait) > 0)
3012 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3013 else
3014 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3015 /* Account for system time used */
3016 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017}
3018
3019/*
3020 * Account for involuntary wait time.
3021 * @p: the process from which the cpu time has been stolen
3022 * @steal: the cpu time spent in involuntary wait
3023 */
3024void account_steal_time(struct task_struct *p, cputime_t steal)
3025{
3026 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3027 cputime64_t tmp = cputime_to_cputime64(steal);
Ingo Molnar70b97a72006-07-03 00:25:42 -07003028 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029
3030 if (p == rq->idle) {
3031 p->stime = cputime_add(p->stime, steal);
3032 if (atomic_read(&rq->nr_iowait) > 0)
3033 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3034 else
3035 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3036 } else
3037 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3038}
3039
3040/*
3041 * This function gets called by the timer code, with HZ frequency.
3042 * We call it with interrupts disabled.
3043 *
3044 * It also gets called by the fork code, when changing the parent's
3045 * timeslices.
3046 */
3047void scheduler_tick(void)
3048{
Ingo Molnar48f24c42006-07-03 00:25:40 -07003049 unsigned long long now = sched_clock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07003050 struct task_struct *p = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 int cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07003052 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053
3054 update_cpu_clock(p, rq, now);
3055
3056 rq->timestamp_last_tick = now;
3057
3058 if (p == rq->idle) {
3059 if (wake_priority_sleeper(rq))
3060 goto out;
3061 rebalance_tick(cpu, rq, SCHED_IDLE);
3062 return;
3063 }
3064
3065 /* Task might have expired already, but not scheduled off yet */
3066 if (p->array != rq->active) {
3067 set_tsk_need_resched(p);
3068 goto out;
3069 }
3070 spin_lock(&rq->lock);
3071 /*
3072 * The task was running during this tick - update the
3073 * time slice counter. Note: we do not update a thread's
3074 * priority until it either goes to sleep or uses up its
3075 * timeslice. This makes it possible for interactive tasks
3076 * to use up their timeslices at their highest priority levels.
3077 */
3078 if (rt_task(p)) {
3079 /*
3080 * RR tasks need a special form of timeslice management.
3081 * FIFO tasks have no timeslices.
3082 */
3083 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3084 p->time_slice = task_timeslice(p);
3085 p->first_time_slice = 0;
3086 set_tsk_need_resched(p);
3087
3088 /* put it at the end of the queue: */
3089 requeue_task(p, rq->active);
3090 }
3091 goto out_unlock;
3092 }
3093 if (!--p->time_slice) {
3094 dequeue_task(p, rq->active);
3095 set_tsk_need_resched(p);
3096 p->prio = effective_prio(p);
3097 p->time_slice = task_timeslice(p);
3098 p->first_time_slice = 0;
3099
3100 if (!rq->expired_timestamp)
3101 rq->expired_timestamp = jiffies;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003102 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103 enqueue_task(p, rq->expired);
3104 if (p->static_prio < rq->best_expired_prio)
3105 rq->best_expired_prio = p->static_prio;
3106 } else
3107 enqueue_task(p, rq->active);
3108 } else {
3109 /*
3110 * Prevent a too long timeslice allowing a task to monopolize
3111 * the CPU. We do this by splitting up the timeslice into
3112 * smaller pieces.
3113 *
3114 * Note: this does not mean the task's timeslices expire or
3115 * get lost in any way, they just might be preempted by
3116 * another task of equal priority. (one with higher
3117 * priority would have preempted this task already.) We
3118 * requeue this task to the end of the list on this priority
3119 * level, which is in essence a round-robin of tasks with
3120 * equal priority.
3121 *
3122 * This only applies to tasks in the interactive
3123 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3124 */
3125 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3126 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3127 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3128 (p->array == rq->active)) {
3129
3130 requeue_task(p, rq->active);
3131 set_tsk_need_resched(p);
3132 }
3133 }
3134out_unlock:
3135 spin_unlock(&rq->lock);
3136out:
3137 rebalance_tick(cpu, rq, NOT_IDLE);
3138}
3139
3140#ifdef CONFIG_SCHED_SMT
Ingo Molnar70b97a72006-07-03 00:25:42 -07003141static inline void wakeup_busy_runqueue(struct rq *rq)
Con Kolivasfc38ed72005-09-10 00:26:08 -07003142{
3143 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3144 if (rq->curr == rq->idle && rq->nr_running)
3145 resched_task(rq->idle);
3146}
3147
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003148/*
3149 * Called with interrupt disabled and this_rq's runqueue locked.
3150 */
3151static void wake_sleeping_dependent(int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152{
Nick Piggin41c7ce92005-06-25 14:57:24 -07003153 struct sched_domain *tmp, *sd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154 int i;
3155
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003156 for_each_domain(this_cpu, tmp) {
3157 if (tmp->flags & SD_SHARE_CPUPOWER) {
Nick Piggin41c7ce92005-06-25 14:57:24 -07003158 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003159 break;
3160 }
3161 }
Nick Piggin41c7ce92005-06-25 14:57:24 -07003162
3163 if (!sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164 return;
3165
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003166 for_each_cpu_mask(i, sd->span) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07003167 struct rq *smt_rq = cpu_rq(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003169 if (i == this_cpu)
3170 continue;
3171 if (unlikely(!spin_trylock(&smt_rq->lock)))
3172 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003174 wakeup_busy_runqueue(smt_rq);
3175 spin_unlock(&smt_rq->lock);
3176 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177}
3178
Ingo Molnar67f9a612005-09-10 00:26:16 -07003179/*
3180 * number of 'lost' timeslices this task wont be able to fully
3181 * utilize, if another task runs on a sibling. This models the
3182 * slowdown effect of other tasks running on siblings:
3183 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003184static inline unsigned long
3185smt_slice(struct task_struct *p, struct sched_domain *sd)
Ingo Molnar67f9a612005-09-10 00:26:16 -07003186{
3187 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
3188}
3189
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003190/*
3191 * To minimise lock contention and not have to drop this_rq's runlock we only
3192 * trylock the sibling runqueues and bypass those runqueues if we fail to
3193 * acquire their lock. As we only trylock the normal locking order does not
3194 * need to be obeyed.
3195 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003196static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07003197dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198{
Nick Piggin41c7ce92005-06-25 14:57:24 -07003199 struct sched_domain *tmp, *sd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200 int ret = 0, i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003202 /* kernel/rt threads do not participate in dependent sleeping */
3203 if (!p->mm || rt_task(p))
3204 return 0;
3205
3206 for_each_domain(this_cpu, tmp) {
3207 if (tmp->flags & SD_SHARE_CPUPOWER) {
Nick Piggin41c7ce92005-06-25 14:57:24 -07003208 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003209 break;
3210 }
3211 }
Nick Piggin41c7ce92005-06-25 14:57:24 -07003212
3213 if (!sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214 return 0;
3215
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003216 for_each_cpu_mask(i, sd->span) {
Ingo Molnar36c8b582006-07-03 00:25:41 -07003217 struct task_struct *smt_curr;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003218 struct rq *smt_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003220 if (i == this_cpu)
3221 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003222
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003223 smt_rq = cpu_rq(i);
3224 if (unlikely(!spin_trylock(&smt_rq->lock)))
3225 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003227 smt_curr = smt_rq->curr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003229 if (!smt_curr->mm)
3230 goto unlock;
Con Kolivasfc38ed72005-09-10 00:26:08 -07003231
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232 /*
3233 * If a user task with lower static priority than the
3234 * running task on the SMT sibling is trying to schedule,
3235 * delay it till there is proportionately less timeslice
3236 * left of the sibling task to prevent a lower priority
3237 * task from using an unfair proportion of the
3238 * physical cpu's resources. -ck
3239 */
Con Kolivasfc38ed72005-09-10 00:26:08 -07003240 if (rt_task(smt_curr)) {
3241 /*
3242 * With real time tasks we run non-rt tasks only
3243 * per_cpu_gain% of the time.
3244 */
3245 if ((jiffies % DEF_TIMESLICE) >
3246 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
3247 ret = 1;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003248 } else {
Ingo Molnar67f9a612005-09-10 00:26:16 -07003249 if (smt_curr->static_prio < p->static_prio &&
3250 !TASK_PREEMPTS_CURR(p, smt_rq) &&
3251 smt_slice(smt_curr, sd) > task_timeslice(p))
Con Kolivasfc38ed72005-09-10 00:26:08 -07003252 ret = 1;
Con Kolivasfc38ed72005-09-10 00:26:08 -07003253 }
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003254unlock:
3255 spin_unlock(&smt_rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257 return ret;
3258}
3259#else
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003260static inline void wake_sleeping_dependent(int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261{
3262}
Ingo Molnar48f24c42006-07-03 00:25:40 -07003263static inline int
Ingo Molnar70b97a72006-07-03 00:25:42 -07003264dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265{
3266 return 0;
3267}
3268#endif
3269
3270#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3271
3272void fastcall add_preempt_count(int val)
3273{
3274 /*
3275 * Underflow?
3276 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003277 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3278 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003279 preempt_count() += val;
3280 /*
3281 * Spinlock count overflowing soon?
3282 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003283 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284}
3285EXPORT_SYMBOL(add_preempt_count);
3286
3287void fastcall sub_preempt_count(int val)
3288{
3289 /*
3290 * Underflow?
3291 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003292 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3293 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 /*
3295 * Is the spinlock portion underflowing?
3296 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003297 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3298 !(preempt_count() & PREEMPT_MASK)))
3299 return;
3300
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301 preempt_count() -= val;
3302}
3303EXPORT_SYMBOL(sub_preempt_count);
3304
3305#endif
3306
Con Kolivas3dee3862006-03-31 02:31:23 -08003307static inline int interactive_sleep(enum sleep_type sleep_type)
3308{
3309 return (sleep_type == SLEEP_INTERACTIVE ||
3310 sleep_type == SLEEP_INTERRUPTED);
3311}
3312
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313/*
3314 * schedule() is the main scheduler function.
3315 */
3316asmlinkage void __sched schedule(void)
3317{
Ingo Molnar36c8b582006-07-03 00:25:41 -07003318 struct task_struct *prev, *next;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003319 struct prio_array *array;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320 struct list_head *queue;
3321 unsigned long long now;
3322 unsigned long run_time;
Chen Shanga3464a12005-06-25 14:57:31 -07003323 int cpu, idx, new_prio;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003324 long *switch_count;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003325 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326
3327 /*
3328 * Test if we are atomic. Since do_exit() needs to call into
3329 * schedule() atomically, we ignore that path for now.
3330 * Otherwise, whine if we are scheduling when we should not be.
3331 */
Andreas Mohr77e4bfb2006-03-27 01:15:20 -08003332 if (unlikely(in_atomic() && !current->exit_state)) {
3333 printk(KERN_ERR "BUG: scheduling while atomic: "
3334 "%s/0x%08x/%d\n",
3335 current->comm, preempt_count(), current->pid);
3336 dump_stack();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 }
3338 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3339
3340need_resched:
3341 preempt_disable();
3342 prev = current;
3343 release_kernel_lock(prev);
3344need_resched_nonpreemptible:
3345 rq = this_rq();
3346
3347 /*
3348 * The idle thread is not allowed to schedule!
3349 * Remove this check after it has been exercised a bit.
3350 */
3351 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3352 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3353 dump_stack();
3354 }
3355
3356 schedstat_inc(rq, sched_cnt);
3357 now = sched_clock();
Ingo Molnar238628e2005-04-18 10:58:36 -07003358 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359 run_time = now - prev->timestamp;
Ingo Molnar238628e2005-04-18 10:58:36 -07003360 if (unlikely((long long)(now - prev->timestamp) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 run_time = 0;
3362 } else
3363 run_time = NS_MAX_SLEEP_AVG;
3364
3365 /*
3366 * Tasks charged proportionately less run_time at high sleep_avg to
3367 * delay them losing their interactive status
3368 */
3369 run_time /= (CURRENT_BONUS(prev) ? : 1);
3370
3371 spin_lock_irq(&rq->lock);
3372
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 switch_count = &prev->nivcsw;
3374 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3375 switch_count = &prev->nvcsw;
3376 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3377 unlikely(signal_pending(prev))))
3378 prev->state = TASK_RUNNING;
3379 else {
3380 if (prev->state == TASK_UNINTERRUPTIBLE)
3381 rq->nr_uninterruptible++;
3382 deactivate_task(prev, rq);
3383 }
3384 }
3385
3386 cpu = smp_processor_id();
3387 if (unlikely(!rq->nr_running)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388 idle_balance(cpu, rq);
3389 if (!rq->nr_running) {
3390 next = rq->idle;
3391 rq->expired_timestamp = 0;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003392 wake_sleeping_dependent(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393 goto switch_tasks;
3394 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003395 }
3396
3397 array = rq->active;
3398 if (unlikely(!array->nr_active)) {
3399 /*
3400 * Switch the active and expired arrays.
3401 */
3402 schedstat_inc(rq, sched_switch);
3403 rq->active = rq->expired;
3404 rq->expired = array;
3405 array = rq->active;
3406 rq->expired_timestamp = 0;
3407 rq->best_expired_prio = MAX_PRIO;
3408 }
3409
3410 idx = sched_find_first_bit(array->bitmap);
3411 queue = array->queue + idx;
Ingo Molnar36c8b582006-07-03 00:25:41 -07003412 next = list_entry(queue->next, struct task_struct, run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413
Con Kolivas3dee3862006-03-31 02:31:23 -08003414 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415 unsigned long long delta = now - next->timestamp;
Ingo Molnar238628e2005-04-18 10:58:36 -07003416 if (unlikely((long long)(now - next->timestamp) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417 delta = 0;
3418
Con Kolivas3dee3862006-03-31 02:31:23 -08003419 if (next->sleep_type == SLEEP_INTERACTIVE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3421
3422 array = next->array;
Chen Shanga3464a12005-06-25 14:57:31 -07003423 new_prio = recalc_task_prio(next, next->timestamp + delta);
3424
3425 if (unlikely(next->prio != new_prio)) {
3426 dequeue_task(next, array);
3427 next->prio = new_prio;
3428 enqueue_task(next, array);
Con Kolivas7c4bb1f2006-03-31 02:31:29 -08003429 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 }
Con Kolivas3dee3862006-03-31 02:31:23 -08003431 next->sleep_type = SLEEP_NORMAL;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003432 if (dependent_sleeper(cpu, rq, next))
3433 next = rq->idle;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434switch_tasks:
3435 if (next == rq->idle)
3436 schedstat_inc(rq, sched_goidle);
3437 prefetch(next);
Chen, Kenneth W383f2832005-09-09 13:02:02 -07003438 prefetch_stack(next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439 clear_tsk_need_resched(prev);
3440 rcu_qsctr_inc(task_cpu(prev));
3441
3442 update_cpu_clock(prev, rq, now);
3443
3444 prev->sleep_avg -= run_time;
3445 if ((long)prev->sleep_avg <= 0)
3446 prev->sleep_avg = 0;
3447 prev->timestamp = prev->last_ran = now;
3448
3449 sched_info_switch(prev, next);
3450 if (likely(prev != next)) {
3451 next->timestamp = now;
3452 rq->nr_switches++;
3453 rq->curr = next;
3454 ++*switch_count;
3455
Nick Piggin4866cde2005-06-25 14:57:23 -07003456 prepare_task_switch(rq, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457 prev = context_switch(rq, prev, next);
3458 barrier();
Nick Piggin4866cde2005-06-25 14:57:23 -07003459 /*
3460 * this_rq must be evaluated again because prev may have moved
3461 * CPUs since it called schedule(), thus the 'rq' on its stack
3462 * frame will be invalid.
3463 */
3464 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465 } else
3466 spin_unlock_irq(&rq->lock);
3467
3468 prev = current;
3469 if (unlikely(reacquire_kernel_lock(prev) < 0))
3470 goto need_resched_nonpreemptible;
3471 preempt_enable_no_resched();
3472 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3473 goto need_resched;
3474}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475EXPORT_SYMBOL(schedule);
3476
3477#ifdef CONFIG_PREEMPT
3478/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003479 * this is the entry point to schedule() from in-kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480 * off of preempt_enable. Kernel preemptions off return from interrupt
3481 * occur there and call schedule directly.
3482 */
3483asmlinkage void __sched preempt_schedule(void)
3484{
3485 struct thread_info *ti = current_thread_info();
3486#ifdef CONFIG_PREEMPT_BKL
3487 struct task_struct *task = current;
3488 int saved_lock_depth;
3489#endif
3490 /*
3491 * If there is a non-zero preempt_count or interrupts are disabled,
3492 * we do not want to preempt the current task. Just return..
3493 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07003494 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 return;
3496
3497need_resched:
3498 add_preempt_count(PREEMPT_ACTIVE);
3499 /*
3500 * We keep the big kernel semaphore locked, but we
3501 * clear ->lock_depth so that schedule() doesnt
3502 * auto-release the semaphore:
3503 */
3504#ifdef CONFIG_PREEMPT_BKL
3505 saved_lock_depth = task->lock_depth;
3506 task->lock_depth = -1;
3507#endif
3508 schedule();
3509#ifdef CONFIG_PREEMPT_BKL
3510 task->lock_depth = saved_lock_depth;
3511#endif
3512 sub_preempt_count(PREEMPT_ACTIVE);
3513
3514 /* we could miss a preemption opportunity between schedule and now */
3515 barrier();
3516 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3517 goto need_resched;
3518}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519EXPORT_SYMBOL(preempt_schedule);
3520
3521/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003522 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523 * off of irq context.
3524 * Note, that this is called and return with irqs disabled. This will
3525 * protect us against recursive calling from irq.
3526 */
3527asmlinkage void __sched preempt_schedule_irq(void)
3528{
3529 struct thread_info *ti = current_thread_info();
3530#ifdef CONFIG_PREEMPT_BKL
3531 struct task_struct *task = current;
3532 int saved_lock_depth;
3533#endif
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003534 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535 BUG_ON(ti->preempt_count || !irqs_disabled());
3536
3537need_resched:
3538 add_preempt_count(PREEMPT_ACTIVE);
3539 /*
3540 * We keep the big kernel semaphore locked, but we
3541 * clear ->lock_depth so that schedule() doesnt
3542 * auto-release the semaphore:
3543 */
3544#ifdef CONFIG_PREEMPT_BKL
3545 saved_lock_depth = task->lock_depth;
3546 task->lock_depth = -1;
3547#endif
3548 local_irq_enable();
3549 schedule();
3550 local_irq_disable();
3551#ifdef CONFIG_PREEMPT_BKL
3552 task->lock_depth = saved_lock_depth;
3553#endif
3554 sub_preempt_count(PREEMPT_ACTIVE);
3555
3556 /* we could miss a preemption opportunity between schedule and now */
3557 barrier();
3558 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3559 goto need_resched;
3560}
3561
3562#endif /* CONFIG_PREEMPT */
3563
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003564int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3565 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566{
Ingo Molnar48f24c42006-07-03 00:25:40 -07003567 return try_to_wake_up(curr->private, mode, sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569EXPORT_SYMBOL(default_wake_function);
3570
3571/*
3572 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3573 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3574 * number) then we wake all the non-exclusive tasks and one exclusive task.
3575 *
3576 * There are circumstances in which we can try to wake a task which has already
3577 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3578 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3579 */
3580static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3581 int nr_exclusive, int sync, void *key)
3582{
3583 struct list_head *tmp, *next;
3584
3585 list_for_each_safe(tmp, next, &q->task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07003586 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3587 unsigned flags = curr->flags;
3588
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 if (curr->func(curr, mode, sync, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07003590 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591 break;
3592 }
3593}
3594
3595/**
3596 * __wake_up - wake up threads blocked on a waitqueue.
3597 * @q: the waitqueue
3598 * @mode: which threads
3599 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07003600 * @key: is directly passed to the wakeup function
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601 */
3602void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003603 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604{
3605 unsigned long flags;
3606
3607 spin_lock_irqsave(&q->lock, flags);
3608 __wake_up_common(q, mode, nr_exclusive, 0, key);
3609 spin_unlock_irqrestore(&q->lock, flags);
3610}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611EXPORT_SYMBOL(__wake_up);
3612
3613/*
3614 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3615 */
3616void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3617{
3618 __wake_up_common(q, mode, 1, 0, NULL);
3619}
3620
3621/**
Martin Waitz67be2dd2005-05-01 08:59:26 -07003622 * __wake_up_sync - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623 * @q: the waitqueue
3624 * @mode: which threads
3625 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3626 *
3627 * The sync wakeup differs that the waker knows that it will schedule
3628 * away soon, so while the target thread will be woken up, it will not
3629 * be migrated to another CPU - ie. the two threads are 'synchronized'
3630 * with each other. This can prevent needless bouncing between CPUs.
3631 *
3632 * On UP it can prevent extra preemption.
3633 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003634void fastcall
3635__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636{
3637 unsigned long flags;
3638 int sync = 1;
3639
3640 if (unlikely(!q))
3641 return;
3642
3643 if (unlikely(!nr_exclusive))
3644 sync = 0;
3645
3646 spin_lock_irqsave(&q->lock, flags);
3647 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3648 spin_unlock_irqrestore(&q->lock, flags);
3649}
3650EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3651
3652void fastcall complete(struct completion *x)
3653{
3654 unsigned long flags;
3655
3656 spin_lock_irqsave(&x->wait.lock, flags);
3657 x->done++;
3658 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3659 1, 0, NULL);
3660 spin_unlock_irqrestore(&x->wait.lock, flags);
3661}
3662EXPORT_SYMBOL(complete);
3663
3664void fastcall complete_all(struct completion *x)
3665{
3666 unsigned long flags;
3667
3668 spin_lock_irqsave(&x->wait.lock, flags);
3669 x->done += UINT_MAX/2;
3670 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3671 0, 0, NULL);
3672 spin_unlock_irqrestore(&x->wait.lock, flags);
3673}
3674EXPORT_SYMBOL(complete_all);
3675
3676void fastcall __sched wait_for_completion(struct completion *x)
3677{
3678 might_sleep();
Ingo Molnar48f24c42006-07-03 00:25:40 -07003679
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680 spin_lock_irq(&x->wait.lock);
3681 if (!x->done) {
3682 DECLARE_WAITQUEUE(wait, current);
3683
3684 wait.flags |= WQ_FLAG_EXCLUSIVE;
3685 __add_wait_queue_tail(&x->wait, &wait);
3686 do {
3687 __set_current_state(TASK_UNINTERRUPTIBLE);
3688 spin_unlock_irq(&x->wait.lock);
3689 schedule();
3690 spin_lock_irq(&x->wait.lock);
3691 } while (!x->done);
3692 __remove_wait_queue(&x->wait, &wait);
3693 }
3694 x->done--;
3695 spin_unlock_irq(&x->wait.lock);
3696}
3697EXPORT_SYMBOL(wait_for_completion);
3698
3699unsigned long fastcall __sched
3700wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3701{
3702 might_sleep();
3703
3704 spin_lock_irq(&x->wait.lock);
3705 if (!x->done) {
3706 DECLARE_WAITQUEUE(wait, current);
3707
3708 wait.flags |= WQ_FLAG_EXCLUSIVE;
3709 __add_wait_queue_tail(&x->wait, &wait);
3710 do {
3711 __set_current_state(TASK_UNINTERRUPTIBLE);
3712 spin_unlock_irq(&x->wait.lock);
3713 timeout = schedule_timeout(timeout);
3714 spin_lock_irq(&x->wait.lock);
3715 if (!timeout) {
3716 __remove_wait_queue(&x->wait, &wait);
3717 goto out;
3718 }
3719 } while (!x->done);
3720 __remove_wait_queue(&x->wait, &wait);
3721 }
3722 x->done--;
3723out:
3724 spin_unlock_irq(&x->wait.lock);
3725 return timeout;
3726}
3727EXPORT_SYMBOL(wait_for_completion_timeout);
3728
3729int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3730{
3731 int ret = 0;
3732
3733 might_sleep();
3734
3735 spin_lock_irq(&x->wait.lock);
3736 if (!x->done) {
3737 DECLARE_WAITQUEUE(wait, current);
3738
3739 wait.flags |= WQ_FLAG_EXCLUSIVE;
3740 __add_wait_queue_tail(&x->wait, &wait);
3741 do {
3742 if (signal_pending(current)) {
3743 ret = -ERESTARTSYS;
3744 __remove_wait_queue(&x->wait, &wait);
3745 goto out;
3746 }
3747 __set_current_state(TASK_INTERRUPTIBLE);
3748 spin_unlock_irq(&x->wait.lock);
3749 schedule();
3750 spin_lock_irq(&x->wait.lock);
3751 } while (!x->done);
3752 __remove_wait_queue(&x->wait, &wait);
3753 }
3754 x->done--;
3755out:
3756 spin_unlock_irq(&x->wait.lock);
3757
3758 return ret;
3759}
3760EXPORT_SYMBOL(wait_for_completion_interruptible);
3761
3762unsigned long fastcall __sched
3763wait_for_completion_interruptible_timeout(struct completion *x,
3764 unsigned long timeout)
3765{
3766 might_sleep();
3767
3768 spin_lock_irq(&x->wait.lock);
3769 if (!x->done) {
3770 DECLARE_WAITQUEUE(wait, current);
3771
3772 wait.flags |= WQ_FLAG_EXCLUSIVE;
3773 __add_wait_queue_tail(&x->wait, &wait);
3774 do {
3775 if (signal_pending(current)) {
3776 timeout = -ERESTARTSYS;
3777 __remove_wait_queue(&x->wait, &wait);
3778 goto out;
3779 }
3780 __set_current_state(TASK_INTERRUPTIBLE);
3781 spin_unlock_irq(&x->wait.lock);
3782 timeout = schedule_timeout(timeout);
3783 spin_lock_irq(&x->wait.lock);
3784 if (!timeout) {
3785 __remove_wait_queue(&x->wait, &wait);
3786 goto out;
3787 }
3788 } while (!x->done);
3789 __remove_wait_queue(&x->wait, &wait);
3790 }
3791 x->done--;
3792out:
3793 spin_unlock_irq(&x->wait.lock);
3794 return timeout;
3795}
3796EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3797
3798
3799#define SLEEP_ON_VAR \
3800 unsigned long flags; \
3801 wait_queue_t wait; \
3802 init_waitqueue_entry(&wait, current);
3803
3804#define SLEEP_ON_HEAD \
3805 spin_lock_irqsave(&q->lock,flags); \
3806 __add_wait_queue(q, &wait); \
3807 spin_unlock(&q->lock);
3808
3809#define SLEEP_ON_TAIL \
3810 spin_lock_irq(&q->lock); \
3811 __remove_wait_queue(q, &wait); \
3812 spin_unlock_irqrestore(&q->lock, flags);
3813
3814void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3815{
3816 SLEEP_ON_VAR
3817
3818 current->state = TASK_INTERRUPTIBLE;
3819
3820 SLEEP_ON_HEAD
3821 schedule();
3822 SLEEP_ON_TAIL
3823}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824EXPORT_SYMBOL(interruptible_sleep_on);
3825
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003826long fastcall __sched
3827interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828{
3829 SLEEP_ON_VAR
3830
3831 current->state = TASK_INTERRUPTIBLE;
3832
3833 SLEEP_ON_HEAD
3834 timeout = schedule_timeout(timeout);
3835 SLEEP_ON_TAIL
3836
3837 return timeout;
3838}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3840
3841void fastcall __sched sleep_on(wait_queue_head_t *q)
3842{
3843 SLEEP_ON_VAR
3844
3845 current->state = TASK_UNINTERRUPTIBLE;
3846
3847 SLEEP_ON_HEAD
3848 schedule();
3849 SLEEP_ON_TAIL
3850}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851EXPORT_SYMBOL(sleep_on);
3852
3853long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3854{
3855 SLEEP_ON_VAR
3856
3857 current->state = TASK_UNINTERRUPTIBLE;
3858
3859 SLEEP_ON_HEAD
3860 timeout = schedule_timeout(timeout);
3861 SLEEP_ON_TAIL
3862
3863 return timeout;
3864}
3865
3866EXPORT_SYMBOL(sleep_on_timeout);
3867
Ingo Molnarb29739f2006-06-27 02:54:51 -07003868#ifdef CONFIG_RT_MUTEXES
3869
3870/*
3871 * rt_mutex_setprio - set the current priority of a task
3872 * @p: task
3873 * @prio: prio value (kernel-internal form)
3874 *
3875 * This function changes the 'effective' priority of a task. It does
3876 * not touch ->normal_prio like __setscheduler().
3877 *
3878 * Used by the rt_mutex code to implement priority inheritance logic.
3879 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003880void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07003881{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003882 struct prio_array *array;
Ingo Molnarb29739f2006-06-27 02:54:51 -07003883 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003884 struct rq *rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -07003885 int oldprio;
3886
3887 BUG_ON(prio < 0 || prio > MAX_PRIO);
3888
3889 rq = task_rq_lock(p, &flags);
3890
3891 oldprio = p->prio;
3892 array = p->array;
3893 if (array)
3894 dequeue_task(p, array);
3895 p->prio = prio;
3896
3897 if (array) {
3898 /*
3899 * If changing to an RT priority then queue it
3900 * in the active array!
3901 */
3902 if (rt_task(p))
3903 array = rq->active;
3904 enqueue_task(p, array);
3905 /*
3906 * Reschedule if we are currently running on this runqueue and
3907 * our priority decreased, or if we are not currently running on
3908 * this runqueue and our priority is higher than the current's
3909 */
3910 if (task_running(rq, p)) {
3911 if (p->prio > oldprio)
3912 resched_task(rq->curr);
3913 } else if (TASK_PREEMPTS_CURR(p, rq))
3914 resched_task(rq->curr);
3915 }
3916 task_rq_unlock(rq, &flags);
3917}
3918
3919#endif
3920
Ingo Molnar36c8b582006-07-03 00:25:41 -07003921void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003923 struct prio_array *array;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003924 int old_prio, delta;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003926 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003927
3928 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3929 return;
3930 /*
3931 * We have to be careful, if called from sys_setpriority(),
3932 * the task might be in the middle of scheduling on another CPU.
3933 */
3934 rq = task_rq_lock(p, &flags);
3935 /*
3936 * The RT priorities are set via sched_setscheduler(), but we still
3937 * allow the 'normal' nice value to be set - but as expected
3938 * it wont have any effect on scheduling until the task is
Ingo Molnarb0a94992006-01-14 13:20:41 -08003939 * not SCHED_NORMAL/SCHED_BATCH:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07003941 if (has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942 p->static_prio = NICE_TO_PRIO(nice);
3943 goto out_unlock;
3944 }
3945 array = p->array;
Peter Williams2dd73a42006-06-27 02:54:34 -07003946 if (array) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003947 dequeue_task(p, array);
Peter Williams2dd73a42006-06-27 02:54:34 -07003948 dec_raw_weighted_load(rq, p);
3949 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07003952 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07003953 old_prio = p->prio;
3954 p->prio = effective_prio(p);
3955 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956
3957 if (array) {
3958 enqueue_task(p, array);
Peter Williams2dd73a42006-06-27 02:54:34 -07003959 inc_raw_weighted_load(rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960 /*
3961 * If the task increased its priority or is running and
3962 * lowered its priority, then reschedule its CPU:
3963 */
3964 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3965 resched_task(rq->curr);
3966 }
3967out_unlock:
3968 task_rq_unlock(rq, &flags);
3969}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970EXPORT_SYMBOL(set_user_nice);
3971
Matt Mackalle43379f2005-05-01 08:59:00 -07003972/*
3973 * can_nice - check if a task can reduce its nice value
3974 * @p: task
3975 * @nice: nice value
3976 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003977int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07003978{
Matt Mackall024f4742005-08-18 11:24:19 -07003979 /* convert nice value [19,-20] to rlimit style value [1,40] */
3980 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003981
Matt Mackalle43379f2005-05-01 08:59:00 -07003982 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3983 capable(CAP_SYS_NICE));
3984}
3985
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986#ifdef __ARCH_WANT_SYS_NICE
3987
3988/*
3989 * sys_nice - change the priority of the current process.
3990 * @increment: priority increment
3991 *
3992 * sys_setpriority is a more generic, but much slower function that
3993 * does similar things.
3994 */
3995asmlinkage long sys_nice(int increment)
3996{
Ingo Molnar48f24c42006-07-03 00:25:40 -07003997 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998
3999 /*
4000 * Setpriority might change our priority at the same moment.
4001 * We don't have to worry. Conceptually one call occurs first
4002 * and we have a single winner.
4003 */
Matt Mackalle43379f2005-05-01 08:59:00 -07004004 if (increment < -40)
4005 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006 if (increment > 40)
4007 increment = 40;
4008
4009 nice = PRIO_TO_NICE(current->static_prio) + increment;
4010 if (nice < -20)
4011 nice = -20;
4012 if (nice > 19)
4013 nice = 19;
4014
Matt Mackalle43379f2005-05-01 08:59:00 -07004015 if (increment < 0 && !can_nice(current, nice))
4016 return -EPERM;
4017
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018 retval = security_task_setnice(current, nice);
4019 if (retval)
4020 return retval;
4021
4022 set_user_nice(current, nice);
4023 return 0;
4024}
4025
4026#endif
4027
4028/**
4029 * task_prio - return the priority value of a given task.
4030 * @p: the task in question.
4031 *
4032 * This is the priority value as seen by users in /proc.
4033 * RT tasks are offset by -200. Normal tasks are centered
4034 * around 0, value goes from -16 to +15.
4035 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004036int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037{
4038 return p->prio - MAX_RT_PRIO;
4039}
4040
4041/**
4042 * task_nice - return the nice value of a given task.
4043 * @p: the task in question.
4044 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004045int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046{
4047 return TASK_NICE(p);
4048}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049EXPORT_SYMBOL_GPL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050
4051/**
4052 * idle_cpu - is a given cpu idle currently?
4053 * @cpu: the processor in question.
4054 */
4055int idle_cpu(int cpu)
4056{
4057 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4058}
4059
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060/**
4061 * idle_task - return the idle task for a given cpu.
4062 * @cpu: the processor in question.
4063 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004064struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065{
4066 return cpu_rq(cpu)->idle;
4067}
4068
4069/**
4070 * find_process_by_pid - find a process with a matching PID value.
4071 * @pid: the pid in question.
4072 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004073static inline struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074{
4075 return pid ? find_task_by_pid(pid) : current;
4076}
4077
4078/* Actually do priority change: must hold rq lock. */
4079static void __setscheduler(struct task_struct *p, int policy, int prio)
4080{
4081 BUG_ON(p->array);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004082
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083 p->policy = policy;
4084 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004085 p->normal_prio = normal_prio(p);
4086 /* we are holding p->pi_lock already */
4087 p->prio = rt_mutex_getprio(p);
4088 /*
4089 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4090 */
4091 if (policy == SCHED_BATCH)
4092 p->sleep_avg = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07004093 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094}
4095
4096/**
4097 * sched_setscheduler - change the scheduling policy and/or RT priority of
4098 * a thread.
4099 * @p: the task in question.
4100 * @policy: new policy.
4101 * @param: structure containing the new RT priority.
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004102 *
4103 * NOTE: the task may be already dead
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004105int sched_setscheduler(struct task_struct *p, int policy,
4106 struct sched_param *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004108 int retval, oldprio, oldpolicy = -1;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004109 struct prio_array *array;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004111 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112
Steven Rostedt66e53932006-06-27 02:54:44 -07004113 /* may grab non-irq protected spin_locks */
4114 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115recheck:
4116 /* double check policy once rq lock held */
4117 if (policy < 0)
4118 policy = oldpolicy = p->policy;
4119 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
Ingo Molnarb0a94992006-01-14 13:20:41 -08004120 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4121 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004122 /*
4123 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnarb0a94992006-01-14 13:20:41 -08004124 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4125 * SCHED_BATCH is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004126 */
4127 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004128 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04004129 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130 return -EINVAL;
Oleg Nesterov57a6f512006-09-29 02:00:49 -07004131 if (is_rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132 return -EINVAL;
4133
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004134 /*
4135 * Allow unprivileged RT tasks to decrease priority:
4136 */
4137 if (!capable(CAP_SYS_NICE)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004138 if (is_rt_policy(policy)) {
4139 unsigned long rlim_rtprio;
4140 unsigned long flags;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004141
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004142 if (!lock_task_sighand(p, &flags))
4143 return -ESRCH;
4144 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4145 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004146
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004147 /* can't set/change the rt policy */
4148 if (policy != p->policy && !rlim_rtprio)
4149 return -EPERM;
4150
4151 /* can't increase priority */
4152 if (param->sched_priority > p->rt_priority &&
4153 param->sched_priority > rlim_rtprio)
4154 return -EPERM;
4155 }
4156
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004157 /* can't change other user's priorities */
4158 if ((current->euid != p->euid) &&
4159 (current->euid != p->uid))
4160 return -EPERM;
4161 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162
4163 retval = security_task_setscheduler(p, policy, param);
4164 if (retval)
4165 return retval;
4166 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07004167 * make sure no PI-waiters arrive (or leave) while we are
4168 * changing the priority of the task:
4169 */
4170 spin_lock_irqsave(&p->pi_lock, flags);
4171 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172 * To be able to change p->policy safely, the apropriate
4173 * runqueue lock must be held.
4174 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07004175 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004176 /* recheck policy now with rq lock held */
4177 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4178 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004179 __task_rq_unlock(rq);
4180 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181 goto recheck;
4182 }
4183 array = p->array;
4184 if (array)
4185 deactivate_task(p, rq);
4186 oldprio = p->prio;
4187 __setscheduler(p, policy, param->sched_priority);
4188 if (array) {
4189 __activate_task(p, rq);
4190 /*
4191 * Reschedule if we are currently running on this runqueue and
4192 * our priority decreased, or if we are not currently running on
4193 * this runqueue and our priority is higher than the current's
4194 */
4195 if (task_running(rq, p)) {
4196 if (p->prio > oldprio)
4197 resched_task(rq->curr);
4198 } else if (TASK_PREEMPTS_CURR(p, rq))
4199 resched_task(rq->curr);
4200 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07004201 __task_rq_unlock(rq);
4202 spin_unlock_irqrestore(&p->pi_lock, flags);
4203
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07004204 rt_mutex_adjust_pi(p);
4205
Linus Torvalds1da177e2005-04-16 15:20:36 -07004206 return 0;
4207}
4208EXPORT_SYMBOL_GPL(sched_setscheduler);
4209
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004210static int
4211do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213 struct sched_param lparam;
4214 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004215 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004216
4217 if (!param || pid < 0)
4218 return -EINVAL;
4219 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4220 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004221
4222 rcu_read_lock();
4223 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004224 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004225 if (p != NULL)
4226 retval = sched_setscheduler(p, policy, &lparam);
4227 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07004228
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229 return retval;
4230}
4231
4232/**
4233 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4234 * @pid: the pid in question.
4235 * @policy: new policy.
4236 * @param: structure containing the new RT priority.
4237 */
4238asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4239 struct sched_param __user *param)
4240{
Jason Baronc21761f2006-01-18 17:43:03 -08004241 /* negative values for policy are not valid */
4242 if (policy < 0)
4243 return -EINVAL;
4244
Linus Torvalds1da177e2005-04-16 15:20:36 -07004245 return do_sched_setscheduler(pid, policy, param);
4246}
4247
4248/**
4249 * sys_sched_setparam - set/change the RT priority of a thread
4250 * @pid: the pid in question.
4251 * @param: structure containing the new RT priority.
4252 */
4253asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4254{
4255 return do_sched_setscheduler(pid, -1, param);
4256}
4257
4258/**
4259 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4260 * @pid: the pid in question.
4261 */
4262asmlinkage long sys_sched_getscheduler(pid_t pid)
4263{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004264 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004265 int retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004266
4267 if (pid < 0)
4268 goto out_nounlock;
4269
4270 retval = -ESRCH;
4271 read_lock(&tasklist_lock);
4272 p = find_process_by_pid(pid);
4273 if (p) {
4274 retval = security_task_getscheduler(p);
4275 if (!retval)
4276 retval = p->policy;
4277 }
4278 read_unlock(&tasklist_lock);
4279
4280out_nounlock:
4281 return retval;
4282}
4283
4284/**
4285 * sys_sched_getscheduler - get the RT priority of a thread
4286 * @pid: the pid in question.
4287 * @param: structure containing the RT priority.
4288 */
4289asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4290{
4291 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004292 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004293 int retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004294
4295 if (!param || pid < 0)
4296 goto out_nounlock;
4297
4298 read_lock(&tasklist_lock);
4299 p = find_process_by_pid(pid);
4300 retval = -ESRCH;
4301 if (!p)
4302 goto out_unlock;
4303
4304 retval = security_task_getscheduler(p);
4305 if (retval)
4306 goto out_unlock;
4307
4308 lp.sched_priority = p->rt_priority;
4309 read_unlock(&tasklist_lock);
4310
4311 /*
4312 * This one might sleep, we cannot do it with a spinlock held ...
4313 */
4314 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4315
4316out_nounlock:
4317 return retval;
4318
4319out_unlock:
4320 read_unlock(&tasklist_lock);
4321 return retval;
4322}
4323
4324long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4325{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004326 cpumask_t cpus_allowed;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004327 struct task_struct *p;
4328 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004329
4330 lock_cpu_hotplug();
4331 read_lock(&tasklist_lock);
4332
4333 p = find_process_by_pid(pid);
4334 if (!p) {
4335 read_unlock(&tasklist_lock);
4336 unlock_cpu_hotplug();
4337 return -ESRCH;
4338 }
4339
4340 /*
4341 * It is not safe to call set_cpus_allowed with the
4342 * tasklist_lock held. We will bump the task_struct's
4343 * usage count and then drop tasklist_lock.
4344 */
4345 get_task_struct(p);
4346 read_unlock(&tasklist_lock);
4347
4348 retval = -EPERM;
4349 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4350 !capable(CAP_SYS_NICE))
4351 goto out_unlock;
4352
David Quigleye7834f82006-06-23 02:03:59 -07004353 retval = security_task_setscheduler(p, 0, NULL);
4354 if (retval)
4355 goto out_unlock;
4356
Linus Torvalds1da177e2005-04-16 15:20:36 -07004357 cpus_allowed = cpuset_cpus_allowed(p);
4358 cpus_and(new_mask, new_mask, cpus_allowed);
4359 retval = set_cpus_allowed(p, new_mask);
4360
4361out_unlock:
4362 put_task_struct(p);
4363 unlock_cpu_hotplug();
4364 return retval;
4365}
4366
4367static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4368 cpumask_t *new_mask)
4369{
4370 if (len < sizeof(cpumask_t)) {
4371 memset(new_mask, 0, sizeof(cpumask_t));
4372 } else if (len > sizeof(cpumask_t)) {
4373 len = sizeof(cpumask_t);
4374 }
4375 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4376}
4377
4378/**
4379 * sys_sched_setaffinity - set the cpu affinity of a process
4380 * @pid: pid of the process
4381 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4382 * @user_mask_ptr: user-space pointer to the new cpu mask
4383 */
4384asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4385 unsigned long __user *user_mask_ptr)
4386{
4387 cpumask_t new_mask;
4388 int retval;
4389
4390 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4391 if (retval)
4392 return retval;
4393
4394 return sched_setaffinity(pid, new_mask);
4395}
4396
4397/*
4398 * Represents all cpu's present in the system
4399 * In systems capable of hotplug, this map could dynamically grow
4400 * as new cpu's are detected in the system via any platform specific
4401 * method, such as ACPI for e.g.
4402 */
4403
Andi Kleen4cef0c62006-01-11 22:44:57 +01004404cpumask_t cpu_present_map __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004405EXPORT_SYMBOL(cpu_present_map);
4406
4407#ifndef CONFIG_SMP
Andi Kleen4cef0c62006-01-11 22:44:57 +01004408cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07004409EXPORT_SYMBOL(cpu_online_map);
4410
Andi Kleen4cef0c62006-01-11 22:44:57 +01004411cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07004412EXPORT_SYMBOL(cpu_possible_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004413#endif
4414
4415long sched_getaffinity(pid_t pid, cpumask_t *mask)
4416{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004417 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004418 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004419
4420 lock_cpu_hotplug();
4421 read_lock(&tasklist_lock);
4422
4423 retval = -ESRCH;
4424 p = find_process_by_pid(pid);
4425 if (!p)
4426 goto out_unlock;
4427
David Quigleye7834f82006-06-23 02:03:59 -07004428 retval = security_task_getscheduler(p);
4429 if (retval)
4430 goto out_unlock;
4431
Jack Steiner2f7016d2006-02-01 03:05:18 -08004432 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004433
4434out_unlock:
4435 read_unlock(&tasklist_lock);
4436 unlock_cpu_hotplug();
4437 if (retval)
4438 return retval;
4439
4440 return 0;
4441}
4442
4443/**
4444 * sys_sched_getaffinity - get the cpu affinity of a process
4445 * @pid: pid of the process
4446 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4447 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4448 */
4449asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4450 unsigned long __user *user_mask_ptr)
4451{
4452 int ret;
4453 cpumask_t mask;
4454
4455 if (len < sizeof(cpumask_t))
4456 return -EINVAL;
4457
4458 ret = sched_getaffinity(pid, &mask);
4459 if (ret < 0)
4460 return ret;
4461
4462 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4463 return -EFAULT;
4464
4465 return sizeof(cpumask_t);
4466}
4467
4468/**
4469 * sys_sched_yield - yield the current processor to other threads.
4470 *
4471 * this function yields the current CPU by moving the calling thread
4472 * to the expired array. If there are no other threads running on this
4473 * CPU then this function will return.
4474 */
4475asmlinkage long sys_sched_yield(void)
4476{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004477 struct rq *rq = this_rq_lock();
4478 struct prio_array *array = current->array, *target = rq->expired;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479
4480 schedstat_inc(rq, yld_cnt);
4481 /*
4482 * We implement yielding by moving the task into the expired
4483 * queue.
4484 *
4485 * (special rule: RT tasks will just roundrobin in the active
4486 * array.)
4487 */
4488 if (rt_task(current))
4489 target = rq->active;
4490
Renaud Lienhart5927ad72005-09-10 00:26:20 -07004491 if (array->nr_active == 1) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004492 schedstat_inc(rq, yld_act_empty);
4493 if (!rq->expired->nr_active)
4494 schedstat_inc(rq, yld_both_empty);
4495 } else if (!rq->expired->nr_active)
4496 schedstat_inc(rq, yld_exp_empty);
4497
4498 if (array != target) {
4499 dequeue_task(current, array);
4500 enqueue_task(current, target);
4501 } else
4502 /*
4503 * requeue_task is cheaper so perform that if possible.
4504 */
4505 requeue_task(current, array);
4506
4507 /*
4508 * Since we are going to call schedule() anyway, there's
4509 * no need to preempt or enable interrupts:
4510 */
4511 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07004512 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004513 _raw_spin_unlock(&rq->lock);
4514 preempt_enable_no_resched();
4515
4516 schedule();
4517
4518 return 0;
4519}
4520
Jim Houston2d7d2532006-07-30 03:03:39 -07004521static inline int __resched_legal(int expected_preempt_count)
Andrew Mortone7b38402006-06-30 01:56:00 -07004522{
Jim Houston2d7d2532006-07-30 03:03:39 -07004523 if (unlikely(preempt_count() != expected_preempt_count))
Andrew Mortone7b38402006-06-30 01:56:00 -07004524 return 0;
4525 if (unlikely(system_state != SYSTEM_RUNNING))
4526 return 0;
4527 return 1;
4528}
4529
4530static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004531{
Ingo Molnar8e0a43d2006-06-23 02:05:23 -07004532#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4533 __might_sleep(__FILE__, __LINE__);
4534#endif
Ingo Molnar5bbcfd92005-07-07 17:57:04 -07004535 /*
4536 * The BKS might be reacquired before we have dropped
4537 * PREEMPT_ACTIVE, which could trigger a second
4538 * cond_resched() call.
4539 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004540 do {
4541 add_preempt_count(PREEMPT_ACTIVE);
4542 schedule();
4543 sub_preempt_count(PREEMPT_ACTIVE);
4544 } while (need_resched());
4545}
4546
4547int __sched cond_resched(void)
4548{
Jim Houston2d7d2532006-07-30 03:03:39 -07004549 if (need_resched() && __resched_legal(0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004550 __cond_resched();
4551 return 1;
4552 }
4553 return 0;
4554}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004555EXPORT_SYMBOL(cond_resched);
4556
4557/*
4558 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4559 * call schedule, and on return reacquire the lock.
4560 *
4561 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4562 * operations here to prevent schedule() from being called twice (once via
4563 * spin_unlock(), once by hand).
4564 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004565int cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004566{
Jan Kara6df3cec2005-06-13 15:52:32 -07004567 int ret = 0;
4568
Linus Torvalds1da177e2005-04-16 15:20:36 -07004569 if (need_lockbreak(lock)) {
4570 spin_unlock(lock);
4571 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07004572 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004573 spin_lock(lock);
4574 }
Jim Houston2d7d2532006-07-30 03:03:39 -07004575 if (need_resched() && __resched_legal(1)) {
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07004576 spin_release(&lock->dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004577 _raw_spin_unlock(lock);
4578 preempt_enable_no_resched();
4579 __cond_resched();
Jan Kara6df3cec2005-06-13 15:52:32 -07004580 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004581 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004582 }
Jan Kara6df3cec2005-06-13 15:52:32 -07004583 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004584}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004585EXPORT_SYMBOL(cond_resched_lock);
4586
4587int __sched cond_resched_softirq(void)
4588{
4589 BUG_ON(!in_softirq());
4590
Jim Houston2d7d2532006-07-30 03:03:39 -07004591 if (need_resched() && __resched_legal(0)) {
Ingo Molnarde30a2b2006-07-03 00:24:42 -07004592 raw_local_irq_disable();
4593 _local_bh_enable();
4594 raw_local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004595 __cond_resched();
4596 local_bh_disable();
4597 return 1;
4598 }
4599 return 0;
4600}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004601EXPORT_SYMBOL(cond_resched_softirq);
4602
Linus Torvalds1da177e2005-04-16 15:20:36 -07004603/**
4604 * yield - yield the current processor to other threads.
4605 *
4606 * this is a shortcut for kernel-space yielding - it marks the
4607 * thread runnable and calls sys_sched_yield().
4608 */
4609void __sched yield(void)
4610{
4611 set_current_state(TASK_RUNNING);
4612 sys_sched_yield();
4613}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004614EXPORT_SYMBOL(yield);
4615
4616/*
4617 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4618 * that process accounting knows that this is a task in IO wait state.
4619 *
4620 * But don't do that if it is a deliberate, throttling IO wait (this task
4621 * has set its backing_dev_info: the queue against which it should throttle)
4622 */
4623void __sched io_schedule(void)
4624{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004625 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004626
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004627 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004628 atomic_inc(&rq->nr_iowait);
4629 schedule();
4630 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004631 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004632}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004633EXPORT_SYMBOL(io_schedule);
4634
4635long __sched io_schedule_timeout(long timeout)
4636{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004637 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004638 long ret;
4639
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004640 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004641 atomic_inc(&rq->nr_iowait);
4642 ret = schedule_timeout(timeout);
4643 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004644 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004645 return ret;
4646}
4647
4648/**
4649 * sys_sched_get_priority_max - return maximum RT priority.
4650 * @policy: scheduling class.
4651 *
4652 * this syscall returns the maximum rt_priority that can be used
4653 * by a given scheduling class.
4654 */
4655asmlinkage long sys_sched_get_priority_max(int policy)
4656{
4657 int ret = -EINVAL;
4658
4659 switch (policy) {
4660 case SCHED_FIFO:
4661 case SCHED_RR:
4662 ret = MAX_USER_RT_PRIO-1;
4663 break;
4664 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08004665 case SCHED_BATCH:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004666 ret = 0;
4667 break;
4668 }
4669 return ret;
4670}
4671
4672/**
4673 * sys_sched_get_priority_min - return minimum RT priority.
4674 * @policy: scheduling class.
4675 *
4676 * this syscall returns the minimum rt_priority that can be used
4677 * by a given scheduling class.
4678 */
4679asmlinkage long sys_sched_get_priority_min(int policy)
4680{
4681 int ret = -EINVAL;
4682
4683 switch (policy) {
4684 case SCHED_FIFO:
4685 case SCHED_RR:
4686 ret = 1;
4687 break;
4688 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08004689 case SCHED_BATCH:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004690 ret = 0;
4691 }
4692 return ret;
4693}
4694
4695/**
4696 * sys_sched_rr_get_interval - return the default timeslice of a process.
4697 * @pid: pid of the process.
4698 * @interval: userspace pointer to the timeslice value.
4699 *
4700 * this syscall writes the default timeslice value of a given process
4701 * into the user-space timespec buffer. A value of '0' means infinity.
4702 */
4703asmlinkage
4704long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4705{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004706 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004707 int retval = -EINVAL;
4708 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004709
4710 if (pid < 0)
4711 goto out_nounlock;
4712
4713 retval = -ESRCH;
4714 read_lock(&tasklist_lock);
4715 p = find_process_by_pid(pid);
4716 if (!p)
4717 goto out_unlock;
4718
4719 retval = security_task_getscheduler(p);
4720 if (retval)
4721 goto out_unlock;
4722
Peter Williamsb78709c2006-06-26 16:58:00 +10004723 jiffies_to_timespec(p->policy == SCHED_FIFO ?
Linus Torvalds1da177e2005-04-16 15:20:36 -07004724 0 : task_timeslice(p), &t);
4725 read_unlock(&tasklist_lock);
4726 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4727out_nounlock:
4728 return retval;
4729out_unlock:
4730 read_unlock(&tasklist_lock);
4731 return retval;
4732}
4733
4734static inline struct task_struct *eldest_child(struct task_struct *p)
4735{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004736 if (list_empty(&p->children))
4737 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004738 return list_entry(p->children.next,struct task_struct,sibling);
4739}
4740
4741static inline struct task_struct *older_sibling(struct task_struct *p)
4742{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004743 if (p->sibling.prev==&p->parent->children)
4744 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004745 return list_entry(p->sibling.prev,struct task_struct,sibling);
4746}
4747
4748static inline struct task_struct *younger_sibling(struct task_struct *p)
4749{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004750 if (p->sibling.next==&p->parent->children)
4751 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004752 return list_entry(p->sibling.next,struct task_struct,sibling);
4753}
4754
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004755static const char stat_nam[] = "RSDTtZX";
Ingo Molnar36c8b582006-07-03 00:25:41 -07004756
4757static void show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004758{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004759 struct task_struct *relative;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004760 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004761 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004762
Linus Torvalds1da177e2005-04-16 15:20:36 -07004763 state = p->state ? __ffs(p->state) + 1 : 0;
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004764 printk("%-13.13s %c", p->comm,
4765 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004766#if (BITS_PER_LONG == 32)
4767 if (state == TASK_RUNNING)
4768 printk(" running ");
4769 else
4770 printk(" %08lX ", thread_saved_pc(p));
4771#else
4772 if (state == TASK_RUNNING)
4773 printk(" running task ");
4774 else
4775 printk(" %016lx ", thread_saved_pc(p));
4776#endif
4777#ifdef CONFIG_DEBUG_STACK_USAGE
4778 {
Al Viro10ebffd2005-11-13 16:06:56 -08004779 unsigned long *n = end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004780 while (!*n)
4781 n++;
Al Viro10ebffd2005-11-13 16:06:56 -08004782 free = (unsigned long)n - (unsigned long)end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004783 }
4784#endif
4785 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4786 if ((relative = eldest_child(p)))
4787 printk("%5d ", relative->pid);
4788 else
4789 printk(" ");
4790 if ((relative = younger_sibling(p)))
4791 printk("%7d", relative->pid);
4792 else
4793 printk(" ");
4794 if ((relative = older_sibling(p)))
4795 printk(" %5d", relative->pid);
4796 else
4797 printk(" ");
4798 if (!p->mm)
4799 printk(" (L-TLB)\n");
4800 else
4801 printk(" (NOTLB)\n");
4802
4803 if (state != TASK_RUNNING)
4804 show_stack(p, NULL);
4805}
4806
4807void show_state(void)
4808{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004809 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004810
4811#if (BITS_PER_LONG == 32)
4812 printk("\n"
4813 " sibling\n");
4814 printk(" task PC pid father child younger older\n");
4815#else
4816 printk("\n"
4817 " sibling\n");
4818 printk(" task PC pid father child younger older\n");
4819#endif
4820 read_lock(&tasklist_lock);
4821 do_each_thread(g, p) {
4822 /*
4823 * reset the NMI-timeout, listing all files on a slow
4824 * console might take alot of time:
4825 */
4826 touch_nmi_watchdog();
4827 show_task(p);
4828 } while_each_thread(g, p);
4829
4830 read_unlock(&tasklist_lock);
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004831 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004832}
4833
Ingo Molnarf340c0d2005-06-28 16:40:42 +02004834/**
4835 * init_idle - set up an idle thread for a given CPU
4836 * @idle: task in question
4837 * @cpu: cpu the idle task belongs to
4838 *
4839 * NOTE: this function does not set the idle thread's NEED_RESCHED
4840 * flag, to make booting more robust.
4841 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07004842void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004843{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004844 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004845 unsigned long flags;
4846
Ingo Molnar81c29a82006-03-07 21:55:27 -08004847 idle->timestamp = sched_clock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004848 idle->sleep_avg = 0;
4849 idle->array = NULL;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004850 idle->prio = idle->normal_prio = MAX_PRIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004851 idle->state = TASK_RUNNING;
4852 idle->cpus_allowed = cpumask_of_cpu(cpu);
4853 set_task_cpu(idle, cpu);
4854
4855 spin_lock_irqsave(&rq->lock, flags);
4856 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07004857#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4858 idle->oncpu = 1;
4859#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004860 spin_unlock_irqrestore(&rq->lock, flags);
4861
4862 /* Set the preempt count _outside_ the spinlocks! */
4863#if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
Al Viroa1261f52005-11-13 16:06:55 -08004864 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004865#else
Al Viroa1261f52005-11-13 16:06:55 -08004866 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004867#endif
4868}
4869
4870/*
4871 * In a system that switches off the HZ timer nohz_cpu_mask
4872 * indicates which cpus entered this state. This is used
4873 * in the rcu update to wait only for active cpus. For system
4874 * which do not switch off the HZ timer nohz_cpu_mask should
4875 * always be CPU_MASK_NONE.
4876 */
4877cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4878
4879#ifdef CONFIG_SMP
4880/*
4881 * This is how migration works:
4882 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07004883 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07004884 * runqueue and wake up that CPU's migration thread.
4885 * 2) we down() the locked semaphore => thread blocks.
4886 * 3) migration thread wakes up (implicitly it forces the migrated
4887 * thread off the CPU)
4888 * 4) it gets the migration request and checks whether the migrated
4889 * task is still in the wrong runqueue.
4890 * 5) if it's in the wrong runqueue then the migration thread removes
4891 * it and puts it into the right queue.
4892 * 6) migration thread up()s the semaphore.
4893 * 7) we wake up and the migration is done.
4894 */
4895
4896/*
4897 * Change a given task's CPU affinity. Migrate the thread to a
4898 * proper CPU and schedule it away if the CPU it's executing on
4899 * is removed from the allowed bitmask.
4900 *
4901 * NOTE: the caller must have a valid reference to the task, the
4902 * task must not exit() & deallocate itself prematurely. The
4903 * call is not atomic; no spinlocks may be held.
4904 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004905int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004906{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004907 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004908 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004909 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004910 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004911
4912 rq = task_rq_lock(p, &flags);
4913 if (!cpus_intersects(new_mask, cpu_online_map)) {
4914 ret = -EINVAL;
4915 goto out;
4916 }
4917
4918 p->cpus_allowed = new_mask;
4919 /* Can the task run on the task's current CPU? If so, we're done */
4920 if (cpu_isset(task_cpu(p), new_mask))
4921 goto out;
4922
4923 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4924 /* Need help from migration thread: drop lock and wait. */
4925 task_rq_unlock(rq, &flags);
4926 wake_up_process(rq->migration_thread);
4927 wait_for_completion(&req.done);
4928 tlb_migrate_finish(p->mm);
4929 return 0;
4930 }
4931out:
4932 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004933
Linus Torvalds1da177e2005-04-16 15:20:36 -07004934 return ret;
4935}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004936EXPORT_SYMBOL_GPL(set_cpus_allowed);
4937
4938/*
4939 * Move (not current) task off this cpu, onto dest cpu. We're doing
4940 * this because either it can't run here any more (set_cpus_allowed()
4941 * away from this CPU, or CPU going down), or because we're
4942 * attempting to rebalance this task on exec (sched_exec).
4943 *
4944 * So we race with normal scheduler movements, but that's OK, as long
4945 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07004946 *
4947 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004948 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07004949static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004950{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004951 struct rq *rq_dest, *rq_src;
Kirill Korotaevefc30812006-06-27 02:54:32 -07004952 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004953
4954 if (unlikely(cpu_is_offline(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07004955 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004956
4957 rq_src = cpu_rq(src_cpu);
4958 rq_dest = cpu_rq(dest_cpu);
4959
4960 double_rq_lock(rq_src, rq_dest);
4961 /* Already moved. */
4962 if (task_cpu(p) != src_cpu)
4963 goto out;
4964 /* Affinity changed (again). */
4965 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4966 goto out;
4967
4968 set_task_cpu(p, dest_cpu);
4969 if (p->array) {
4970 /*
4971 * Sync timestamp with rq_dest's before activating.
4972 * The same thing could be achieved by doing this step
4973 * afterwards, and pretending it was a local activate.
4974 * This way is cleaner and logically correct.
4975 */
4976 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4977 + rq_dest->timestamp_last_tick;
4978 deactivate_task(p, rq_src);
Peter Williams0a565f72006-07-10 04:43:51 -07004979 __activate_task(p, rq_dest);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004980 if (TASK_PREEMPTS_CURR(p, rq_dest))
4981 resched_task(rq_dest->curr);
4982 }
Kirill Korotaevefc30812006-06-27 02:54:32 -07004983 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004984out:
4985 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07004986 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004987}
4988
4989/*
4990 * migration_thread - this is a highprio system thread that performs
4991 * thread migration by bumping thread off CPU then 'pushing' onto
4992 * another runqueue.
4993 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004994static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004995{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004996 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004997 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004998
4999 rq = cpu_rq(cpu);
5000 BUG_ON(rq->migration_thread != current);
5001
5002 set_current_state(TASK_INTERRUPTIBLE);
5003 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005004 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005005 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005006
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07005007 try_to_freeze();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005008
5009 spin_lock_irq(&rq->lock);
5010
5011 if (cpu_is_offline(cpu)) {
5012 spin_unlock_irq(&rq->lock);
5013 goto wait_to_die;
5014 }
5015
5016 if (rq->active_balance) {
5017 active_load_balance(rq, cpu);
5018 rq->active_balance = 0;
5019 }
5020
5021 head = &rq->migration_queue;
5022
5023 if (list_empty(head)) {
5024 spin_unlock_irq(&rq->lock);
5025 schedule();
5026 set_current_state(TASK_INTERRUPTIBLE);
5027 continue;
5028 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07005029 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005030 list_del_init(head->next);
5031
Nick Piggin674311d2005-06-25 14:57:27 -07005032 spin_unlock(&rq->lock);
5033 __migrate_task(req->task, cpu, req->dest_cpu);
5034 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005035
5036 complete(&req->done);
5037 }
5038 __set_current_state(TASK_RUNNING);
5039 return 0;
5040
5041wait_to_die:
5042 /* Wait for kthread_stop */
5043 set_current_state(TASK_INTERRUPTIBLE);
5044 while (!kthread_should_stop()) {
5045 schedule();
5046 set_current_state(TASK_INTERRUPTIBLE);
5047 }
5048 __set_current_state(TASK_RUNNING);
5049 return 0;
5050}
5051
5052#ifdef CONFIG_HOTPLUG_CPU
5053/* Figure out where task on dead CPU should go, use force if neccessary. */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005054static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005055{
Kirill Korotaevefc30812006-06-27 02:54:32 -07005056 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005057 cpumask_t mask;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005058 struct rq *rq;
5059 int dest_cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005060
Kirill Korotaevefc30812006-06-27 02:54:32 -07005061restart:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005062 /* On same node? */
5063 mask = node_to_cpumask(cpu_to_node(dead_cpu));
Ingo Molnar48f24c42006-07-03 00:25:40 -07005064 cpus_and(mask, mask, p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005065 dest_cpu = any_online_cpu(mask);
5066
5067 /* On any allowed CPU? */
5068 if (dest_cpu == NR_CPUS)
Ingo Molnar48f24c42006-07-03 00:25:40 -07005069 dest_cpu = any_online_cpu(p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005070
5071 /* No more Mr. Nice Guy. */
5072 if (dest_cpu == NR_CPUS) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07005073 rq = task_rq_lock(p, &flags);
5074 cpus_setall(p->cpus_allowed);
5075 dest_cpu = any_online_cpu(p->cpus_allowed);
Kirill Korotaevefc30812006-06-27 02:54:32 -07005076 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005077
5078 /*
5079 * Don't tell them about moving exiting tasks or
5080 * kernel threads (both mm NULL), since they never
5081 * leave kernel.
5082 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005083 if (p->mm && printk_ratelimit())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005084 printk(KERN_INFO "process %d (%s) no "
5085 "longer affine to cpu%d\n",
Ingo Molnar48f24c42006-07-03 00:25:40 -07005086 p->pid, p->comm, dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005087 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07005088 if (!__migrate_task(p, dead_cpu, dest_cpu))
Kirill Korotaevefc30812006-06-27 02:54:32 -07005089 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005090}
5091
5092/*
5093 * While a dead CPU has no uninterruptible tasks queued at this point,
5094 * it might still have a nonzero ->nr_uninterruptible counter, because
5095 * for performance reasons the counter is not stricly tracking tasks to
5096 * their home CPUs. So we just add the counter to another CPU's counter,
5097 * to keep the global sum constant after CPU-down:
5098 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07005099static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005100{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005101 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005102 unsigned long flags;
5103
5104 local_irq_save(flags);
5105 double_rq_lock(rq_src, rq_dest);
5106 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5107 rq_src->nr_uninterruptible = 0;
5108 double_rq_unlock(rq_src, rq_dest);
5109 local_irq_restore(flags);
5110}
5111
5112/* Run through task list and migrate tasks from the dead cpu. */
5113static void migrate_live_tasks(int src_cpu)
5114{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005115 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005116
5117 write_lock_irq(&tasklist_lock);
5118
Ingo Molnar48f24c42006-07-03 00:25:40 -07005119 do_each_thread(t, p) {
5120 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005121 continue;
5122
Ingo Molnar48f24c42006-07-03 00:25:40 -07005123 if (task_cpu(p) == src_cpu)
5124 move_task_off_dead_cpu(src_cpu, p);
5125 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005126
5127 write_unlock_irq(&tasklist_lock);
5128}
5129
5130/* Schedules idle task to be the next runnable task on current CPU.
5131 * It does so by boosting its priority to highest possible and adding it to
Ingo Molnar48f24c42006-07-03 00:25:40 -07005132 * the _front_ of the runqueue. Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005133 */
5134void sched_idle_next(void)
5135{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005136 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07005137 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005138 struct task_struct *p = rq->idle;
5139 unsigned long flags;
5140
5141 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005142 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005143
Ingo Molnar48f24c42006-07-03 00:25:40 -07005144 /*
5145 * Strictly not necessary since rest of the CPUs are stopped by now
5146 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005147 */
5148 spin_lock_irqsave(&rq->lock, flags);
5149
5150 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005151
5152 /* Add idle task to the _front_ of its priority queue: */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005153 __activate_idle_task(p, rq);
5154
5155 spin_unlock_irqrestore(&rq->lock, flags);
5156}
5157
Ingo Molnar48f24c42006-07-03 00:25:40 -07005158/*
5159 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07005160 * offline.
5161 */
5162void idle_task_exit(void)
5163{
5164 struct mm_struct *mm = current->active_mm;
5165
5166 BUG_ON(cpu_online(smp_processor_id()));
5167
5168 if (mm != &init_mm)
5169 switch_mm(mm, &init_mm, current);
5170 mmdrop(mm);
5171}
5172
Ingo Molnar36c8b582006-07-03 00:25:41 -07005173static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005174{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005175 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005176
5177 /* Must be exiting, otherwise would be on tasklist. */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005178 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005179
5180 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07005181 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005182
Ingo Molnar48f24c42006-07-03 00:25:40 -07005183 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005184
5185 /*
5186 * Drop lock around migration; if someone else moves it,
5187 * that's OK. No task can be added to this CPU, so iteration is
5188 * fine.
5189 */
5190 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005191 move_task_off_dead_cpu(dead_cpu, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005192 spin_lock_irq(&rq->lock);
5193
Ingo Molnar48f24c42006-07-03 00:25:40 -07005194 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005195}
5196
5197/* release_task() removes task from tasklist, so we won't find dead tasks. */
5198static void migrate_dead_tasks(unsigned int dead_cpu)
5199{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005200 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005201 unsigned int arr, i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005202
5203 for (arr = 0; arr < 2; arr++) {
5204 for (i = 0; i < MAX_PRIO; i++) {
5205 struct list_head *list = &rq->arrays[arr].queue[i];
Ingo Molnar48f24c42006-07-03 00:25:40 -07005206
Linus Torvalds1da177e2005-04-16 15:20:36 -07005207 while (!list_empty(list))
Ingo Molnar36c8b582006-07-03 00:25:41 -07005208 migrate_dead(dead_cpu, list_entry(list->next,
5209 struct task_struct, run_list));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005210 }
5211 }
5212}
5213#endif /* CONFIG_HOTPLUG_CPU */
5214
5215/*
5216 * migration_call - callback that gets triggered when a CPU is added.
5217 * Here we can start up the necessary migration thread for the new CPU.
5218 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005219static int __cpuinit
5220migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005221{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005222 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005223 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005224 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005225 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005226
5227 switch (action) {
5228 case CPU_UP_PREPARE:
5229 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5230 if (IS_ERR(p))
5231 return NOTIFY_BAD;
5232 p->flags |= PF_NOFREEZE;
5233 kthread_bind(p, cpu);
5234 /* Must be high prio: stop_machine expects to yield to it. */
5235 rq = task_rq_lock(p, &flags);
5236 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5237 task_rq_unlock(rq, &flags);
5238 cpu_rq(cpu)->migration_thread = p;
5239 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005240
Linus Torvalds1da177e2005-04-16 15:20:36 -07005241 case CPU_ONLINE:
5242 /* Strictly unneccessary, as first user will wake it. */
5243 wake_up_process(cpu_rq(cpu)->migration_thread);
5244 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005245
Linus Torvalds1da177e2005-04-16 15:20:36 -07005246#ifdef CONFIG_HOTPLUG_CPU
5247 case CPU_UP_CANCELED:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07005248 if (!cpu_rq(cpu)->migration_thread)
5249 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005250 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08005251 kthread_bind(cpu_rq(cpu)->migration_thread,
5252 any_online_cpu(cpu_online_map));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005253 kthread_stop(cpu_rq(cpu)->migration_thread);
5254 cpu_rq(cpu)->migration_thread = NULL;
5255 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005256
Linus Torvalds1da177e2005-04-16 15:20:36 -07005257 case CPU_DEAD:
5258 migrate_live_tasks(cpu);
5259 rq = cpu_rq(cpu);
5260 kthread_stop(rq->migration_thread);
5261 rq->migration_thread = NULL;
5262 /* Idle task back to normal (off runqueue, low prio) */
5263 rq = task_rq_lock(rq->idle, &flags);
5264 deactivate_task(rq->idle, rq);
5265 rq->idle->static_prio = MAX_PRIO;
5266 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5267 migrate_dead_tasks(cpu);
5268 task_rq_unlock(rq, &flags);
5269 migrate_nr_uninterruptible(rq);
5270 BUG_ON(rq->nr_running != 0);
5271
5272 /* No need to migrate the tasks: it was best-effort if
5273 * they didn't do lock_cpu_hotplug(). Just wake up
5274 * the requestors. */
5275 spin_lock_irq(&rq->lock);
5276 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005277 struct migration_req *req;
5278
Linus Torvalds1da177e2005-04-16 15:20:36 -07005279 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07005280 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005281 list_del_init(&req->list);
5282 complete(&req->done);
5283 }
5284 spin_unlock_irq(&rq->lock);
5285 break;
5286#endif
5287 }
5288 return NOTIFY_OK;
5289}
5290
5291/* Register at highest priority so that task migration (migrate_all_tasks)
5292 * happens before everything else.
5293 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07005294static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005295 .notifier_call = migration_call,
5296 .priority = 10
5297};
5298
5299int __init migration_init(void)
5300{
5301 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07005302 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005303
5304 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07005305 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5306 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005307 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5308 register_cpu_notifier(&migration_notifier);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005309
Linus Torvalds1da177e2005-04-16 15:20:36 -07005310 return 0;
5311}
5312#endif
5313
5314#ifdef CONFIG_SMP
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07005315#undef SCHED_DOMAIN_DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07005316#ifdef SCHED_DOMAIN_DEBUG
5317static void sched_domain_debug(struct sched_domain *sd, int cpu)
5318{
5319 int level = 0;
5320
Nick Piggin41c7ce92005-06-25 14:57:24 -07005321 if (!sd) {
5322 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5323 return;
5324 }
5325
Linus Torvalds1da177e2005-04-16 15:20:36 -07005326 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5327
5328 do {
5329 int i;
5330 char str[NR_CPUS];
5331 struct sched_group *group = sd->groups;
5332 cpumask_t groupmask;
5333
5334 cpumask_scnprintf(str, NR_CPUS, sd->span);
5335 cpus_clear(groupmask);
5336
5337 printk(KERN_DEBUG);
5338 for (i = 0; i < level + 1; i++)
5339 printk(" ");
5340 printk("domain %d: ", level);
5341
5342 if (!(sd->flags & SD_LOAD_BALANCE)) {
5343 printk("does not load-balance\n");
5344 if (sd->parent)
5345 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
5346 break;
5347 }
5348
5349 printk("span %s\n", str);
5350
5351 if (!cpu_isset(cpu, sd->span))
5352 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
5353 if (!cpu_isset(cpu, group->cpumask))
5354 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
5355
5356 printk(KERN_DEBUG);
5357 for (i = 0; i < level + 2; i++)
5358 printk(" ");
5359 printk("groups:");
5360 do {
5361 if (!group) {
5362 printk("\n");
5363 printk(KERN_ERR "ERROR: group is NULL\n");
5364 break;
5365 }
5366
5367 if (!group->cpu_power) {
5368 printk("\n");
5369 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
5370 }
5371
5372 if (!cpus_weight(group->cpumask)) {
5373 printk("\n");
5374 printk(KERN_ERR "ERROR: empty group\n");
5375 }
5376
5377 if (cpus_intersects(groupmask, group->cpumask)) {
5378 printk("\n");
5379 printk(KERN_ERR "ERROR: repeated CPUs\n");
5380 }
5381
5382 cpus_or(groupmask, groupmask, group->cpumask);
5383
5384 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5385 printk(" %s", str);
5386
5387 group = group->next;
5388 } while (group != sd->groups);
5389 printk("\n");
5390
5391 if (!cpus_equal(sd->span, groupmask))
5392 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5393
5394 level++;
5395 sd = sd->parent;
5396
5397 if (sd) {
5398 if (!cpus_subset(groupmask, sd->span))
5399 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
5400 }
5401
5402 } while (sd);
5403}
5404#else
Ingo Molnar48f24c42006-07-03 00:25:40 -07005405# define sched_domain_debug(sd, cpu) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005406#endif
5407
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07005408static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07005409{
5410 if (cpus_weight(sd->span) == 1)
5411 return 1;
5412
5413 /* Following flags need at least 2 groups */
5414 if (sd->flags & (SD_LOAD_BALANCE |
5415 SD_BALANCE_NEWIDLE |
5416 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005417 SD_BALANCE_EXEC |
5418 SD_SHARE_CPUPOWER |
5419 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005420 if (sd->groups != sd->groups->next)
5421 return 0;
5422 }
5423
5424 /* Following flags don't use groups */
5425 if (sd->flags & (SD_WAKE_IDLE |
5426 SD_WAKE_AFFINE |
5427 SD_WAKE_BALANCE))
5428 return 0;
5429
5430 return 1;
5431}
5432
Ingo Molnar48f24c42006-07-03 00:25:40 -07005433static int
5434sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07005435{
5436 unsigned long cflags = sd->flags, pflags = parent->flags;
5437
5438 if (sd_degenerate(parent))
5439 return 1;
5440
5441 if (!cpus_equal(sd->span, parent->span))
5442 return 0;
5443
5444 /* Does parent contain flags not in child? */
5445 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5446 if (cflags & SD_WAKE_AFFINE)
5447 pflags &= ~SD_WAKE_BALANCE;
5448 /* Flags needing groups don't count if only 1 group in parent */
5449 if (parent->groups == parent->groups->next) {
5450 pflags &= ~(SD_LOAD_BALANCE |
5451 SD_BALANCE_NEWIDLE |
5452 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005453 SD_BALANCE_EXEC |
5454 SD_SHARE_CPUPOWER |
5455 SD_SHARE_PKG_RESOURCES);
Suresh Siddha245af2c2005-06-25 14:57:25 -07005456 }
5457 if (~cflags & pflags)
5458 return 0;
5459
5460 return 1;
5461}
5462
Linus Torvalds1da177e2005-04-16 15:20:36 -07005463/*
5464 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5465 * hold the hotplug lock.
5466 */
John Hawkes9c1cfda2005-09-06 15:18:14 -07005467static void cpu_attach_domain(struct sched_domain *sd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005468{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005469 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07005470 struct sched_domain *tmp;
5471
5472 /* Remove the sched domains which do not contribute to scheduling. */
5473 for (tmp = sd; tmp; tmp = tmp->parent) {
5474 struct sched_domain *parent = tmp->parent;
5475 if (!parent)
5476 break;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005477 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005478 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005479 if (parent->parent)
5480 parent->parent->child = tmp;
5481 }
Suresh Siddha245af2c2005-06-25 14:57:25 -07005482 }
5483
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005484 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005485 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005486 if (sd)
5487 sd->child = NULL;
5488 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005489
5490 sched_domain_debug(sd, cpu);
5491
Nick Piggin674311d2005-06-25 14:57:27 -07005492 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005493}
5494
5495/* cpus with isolated domains */
Nick Piggin5c1e1762006-10-03 01:14:04 -07005496static cpumask_t __cpuinitdata cpu_isolated_map = CPU_MASK_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005497
5498/* Setup the mask of cpus configured for isolated domains */
5499static int __init isolated_cpu_setup(char *str)
5500{
5501 int ints[NR_CPUS], i;
5502
5503 str = get_options(str, ARRAY_SIZE(ints), ints);
5504 cpus_clear(cpu_isolated_map);
5505 for (i = 1; i <= ints[0]; i++)
5506 if (ints[i] < NR_CPUS)
5507 cpu_set(ints[i], cpu_isolated_map);
5508 return 1;
5509}
5510
5511__setup ("isolcpus=", isolated_cpu_setup);
5512
5513/*
5514 * init_sched_build_groups takes an array of groups, the cpumask we wish
5515 * to span, and a pointer to a function which identifies what group a CPU
5516 * belongs to. The return value of group_fn must be a valid index into the
5517 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5518 * keep track of groups covered with a cpumask_t).
5519 *
5520 * init_sched_build_groups will build a circular linked list of the groups
5521 * covered by the given span, and will set each group's ->cpumask correctly,
5522 * and ->cpu_power to 0.
5523 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005524static void
5525init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5526 const cpumask_t *cpu_map,
5527 int (*group_fn)(int cpu, const cpumask_t *cpu_map))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005528{
5529 struct sched_group *first = NULL, *last = NULL;
5530 cpumask_t covered = CPU_MASK_NONE;
5531 int i;
5532
5533 for_each_cpu_mask(i, span) {
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005534 int group = group_fn(i, cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005535 struct sched_group *sg = &groups[group];
5536 int j;
5537
5538 if (cpu_isset(i, covered))
5539 continue;
5540
5541 sg->cpumask = CPU_MASK_NONE;
5542 sg->cpu_power = 0;
5543
5544 for_each_cpu_mask(j, span) {
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005545 if (group_fn(j, cpu_map) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005546 continue;
5547
5548 cpu_set(j, covered);
5549 cpu_set(j, sg->cpumask);
5550 }
5551 if (!first)
5552 first = sg;
5553 if (last)
5554 last->next = sg;
5555 last = sg;
5556 }
5557 last->next = first;
5558}
5559
John Hawkes9c1cfda2005-09-06 15:18:14 -07005560#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07005561
akpm@osdl.org198e2f12006-01-12 01:05:30 -08005562/*
5563 * Self-tuning task migration cost measurement between source and target CPUs.
5564 *
5565 * This is done by measuring the cost of manipulating buffers of varying
5566 * sizes. For a given buffer-size here are the steps that are taken:
5567 *
5568 * 1) the source CPU reads+dirties a shared buffer
5569 * 2) the target CPU reads+dirties the same shared buffer
5570 *
5571 * We measure how long they take, in the following 4 scenarios:
5572 *
5573 * - source: CPU1, target: CPU2 | cost1
5574 * - source: CPU2, target: CPU1 | cost2
5575 * - source: CPU1, target: CPU1 | cost3
5576 * - source: CPU2, target: CPU2 | cost4
5577 *
5578 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5579 * the cost of migration.
5580 *
5581 * We then start off from a small buffer-size and iterate up to larger
5582 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5583 * doing a maximum search for the cost. (The maximum cost for a migration
5584 * normally occurs when the working set size is around the effective cache
5585 * size.)
5586 */
5587#define SEARCH_SCOPE 2
5588#define MIN_CACHE_SIZE (64*1024U)
5589#define DEFAULT_CACHE_SIZE (5*1024*1024U)
Ingo Molnar70b4d632006-01-30 20:24:38 +01005590#define ITERATIONS 1
akpm@osdl.org198e2f12006-01-12 01:05:30 -08005591#define SIZE_THRESH 130
5592#define COST_THRESH 130
5593
5594/*
5595 * The migration cost is a function of 'domain distance'. Domain
5596 * distance is the number of steps a CPU has to iterate down its
5597 * domain tree to share a domain with the other CPU. The farther
5598 * two CPUs are from each other, the larger the distance gets.
5599 *
5600 * Note that we use the distance only to cache measurement results,
5601 * the distance value is not used numerically otherwise. When two
5602 * CPUs have the same distance it is assumed that the migration
5603 * cost is the same. (this is a simplification but quite practical)
5604 */
5605#define MAX_DOMAIN_DISTANCE 32
5606
5607static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
Ingo Molnar4bbf39c2006-02-17 13:52:44 -08005608 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5609/*
5610 * Architectures may override the migration cost and thus avoid
5611 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5612 * virtualized hardware:
5613 */
5614#ifdef CONFIG_DEFAULT_MIGRATION_COST
5615 CONFIG_DEFAULT_MIGRATION_COST
5616#else
5617 -1LL
5618#endif
5619};
akpm@osdl.org198e2f12006-01-12 01:05:30 -08005620
5621/*
5622 * Allow override of migration cost - in units of microseconds.
5623 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5624 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5625 */
5626static int __init migration_cost_setup(char *str)
5627{
5628 int ints[MAX_DOMAIN_DISTANCE+1], i;
5629
5630 str = get_options(str, ARRAY_SIZE(ints), ints);
5631
5632 printk("#ints: %d\n", ints[0]);
5633 for (i = 1; i <= ints[0]; i++) {
5634 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5635 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5636 }
5637 return 1;
5638}
5639
5640__setup ("migration_cost=", migration_cost_setup);
5641
5642/*
5643 * Global multiplier (divisor) for migration-cutoff values,
5644 * in percentiles. E.g. use a value of 150 to get 1.5 times
5645 * longer cache-hot cutoff times.
5646 *
5647 * (We scale it from 100 to 128 to long long handling easier.)
5648 */
5649
5650#define MIGRATION_FACTOR_SCALE 128
5651
5652static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5653
5654static int __init setup_migration_factor(char *str)
5655{
5656 get_option(&str, &migration_factor);
5657 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5658 return 1;
5659}
5660
5661__setup("migration_factor=", setup_migration_factor);
5662
5663/*
5664 * Estimated distance of two CPUs, measured via the number of domains
5665 * we have to pass for the two CPUs to be in the same span:
5666 */
5667static unsigned long domain_distance(int cpu1, int cpu2)
5668{
5669 unsigned long distance = 0;
5670 struct sched_domain *sd;
5671
5672 for_each_domain(cpu1, sd) {
5673 WARN_ON(!cpu_isset(cpu1, sd->span));
5674 if (cpu_isset(cpu2, sd->span))
5675 return distance;
5676 distance++;
5677 }
5678 if (distance >= MAX_DOMAIN_DISTANCE) {
5679 WARN_ON(1);
5680 distance = MAX_DOMAIN_DISTANCE-1;
5681 }
5682
5683 return distance;
5684}
5685
5686static unsigned int migration_debug;
5687
5688static int __init setup_migration_debug(char *str)
5689{
5690 get_option(&str, &migration_debug);
5691 return 1;
5692}
5693
5694__setup("migration_debug=", setup_migration_debug);
5695
5696/*
5697 * Maximum cache-size that the scheduler should try to measure.
5698 * Architectures with larger caches should tune this up during
5699 * bootup. Gets used in the domain-setup code (i.e. during SMP
5700 * bootup).
5701 */
5702unsigned int max_cache_size;
5703
5704static int __init setup_max_cache_size(char *str)
5705{
5706 get_option(&str, &max_cache_size);
5707 return 1;
5708}
5709
5710__setup("max_cache_size=", setup_max_cache_size);
5711
5712/*
5713 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5714 * is the operation that is timed, so we try to generate unpredictable
5715 * cachemisses that still end up filling the L2 cache:
5716 */
5717static void touch_cache(void *__cache, unsigned long __size)
5718{
5719 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5720 chunk2 = 2*size/3;
5721 unsigned long *cache = __cache;
5722 int i;
5723
5724 for (i = 0; i < size/6; i += 8) {
5725 switch (i % 6) {
5726 case 0: cache[i]++;
5727 case 1: cache[size-1-i]++;
5728 case 2: cache[chunk1-i]++;
5729 case 3: cache[chunk1+i]++;
5730 case 4: cache[chunk2-i]++;
5731 case 5: cache[chunk2+i]++;
5732 }
5733 }
5734}
5735
5736/*
5737 * Measure the cache-cost of one task migration. Returns in units of nsec.
5738 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005739static unsigned long long
5740measure_one(void *cache, unsigned long size, int source, int target)
akpm@osdl.org198e2f12006-01-12 01:05:30 -08005741{
5742 cpumask_t mask, saved_mask;
5743 unsigned long long t0, t1, t2, t3, cost;
5744
5745 saved_mask = current->cpus_allowed;
5746
5747 /*
5748 * Flush source caches to RAM and invalidate them:
5749 */
5750 sched_cacheflush();
5751
5752 /*
5753 * Migrate to the source CPU:
5754 */
5755 mask = cpumask_of_cpu(source);
5756 set_cpus_allowed(current, mask);
5757 WARN_ON(smp_processor_id() != source);
5758
5759 /*
5760 * Dirty the working set:
5761 */
5762 t0 = sched_clock();
5763 touch_cache(cache, size);
5764 t1 = sched_clock();
5765
5766 /*
5767 * Migrate to the target CPU, dirty the L2 cache and access
5768 * the shared buffer. (which represents the working set
5769 * of a migrated task.)
5770 */
5771 mask = cpumask_of_cpu(target);
5772 set_cpus_allowed(current, mask);
5773 WARN_ON(smp_processor_id() != target);
5774
5775 t2 = sched_clock();
5776 touch_cache(cache, size);
5777 t3 = sched_clock();
5778
5779 cost = t1-t0 + t3-t2;
5780
5781 if (migration_debug >= 2)
5782 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5783 source, target, t1-t0, t1-t0, t3-t2, cost);
5784 /*
5785 * Flush target caches to RAM and invalidate them:
5786 */
5787 sched_cacheflush();
5788
5789 set_cpus_allowed(current, saved_mask);
5790
5791 return cost;
5792}
5793
5794/*
5795 * Measure a series of task migrations and return the average
5796 * result. Since this code runs early during bootup the system
5797 * is 'undisturbed' and the average latency makes sense.
5798 *
5799 * The algorithm in essence auto-detects the relevant cache-size,
5800 * so it will properly detect different cachesizes for different
5801 * cache-hierarchies, depending on how the CPUs are connected.
5802 *
5803 * Architectures can prime the upper limit of the search range via
5804 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5805 */
5806static unsigned long long
5807measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5808{
5809 unsigned long long cost1, cost2;
5810 int i;
5811
5812 /*
5813 * Measure the migration cost of 'size' bytes, over an
5814 * average of 10 runs:
5815 *
5816 * (We perturb the cache size by a small (0..4k)
5817 * value to compensate size/alignment related artifacts.
5818 * We also subtract the cost of the operation done on
5819 * the same CPU.)
5820 */
5821 cost1 = 0;
5822
5823 /*
5824 * dry run, to make sure we start off cache-cold on cpu1,
5825 * and to get any vmalloc pagefaults in advance:
5826 */
5827 measure_one(cache, size, cpu1, cpu2);
5828 for (i = 0; i < ITERATIONS; i++)
5829 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5830
5831 measure_one(cache, size, cpu2, cpu1);
5832 for (i = 0; i < ITERATIONS; i++)
5833 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5834
5835 /*
5836 * (We measure the non-migrating [cached] cost on both
5837 * cpu1 and cpu2, to handle CPUs with different speeds)
5838 */
5839 cost2 = 0;
5840
5841 measure_one(cache, size, cpu1, cpu1);
5842 for (i = 0; i < ITERATIONS; i++)
5843 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5844
5845 measure_one(cache, size, cpu2, cpu2);
5846 for (i = 0; i < ITERATIONS; i++)
5847 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5848
5849 /*
5850 * Get the per-iteration migration cost:
5851 */
5852 do_div(cost1, 2*ITERATIONS);
5853 do_div(cost2, 2*ITERATIONS);
5854
5855 return cost1 - cost2;
5856}
5857
5858static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5859{
5860 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5861 unsigned int max_size, size, size_found = 0;
5862 long long cost = 0, prev_cost;
5863 void *cache;
5864
5865 /*
5866 * Search from max_cache_size*5 down to 64K - the real relevant
5867 * cachesize has to lie somewhere inbetween.
5868 */
5869 if (max_cache_size) {
5870 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5871 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5872 } else {
5873 /*
5874 * Since we have no estimation about the relevant
5875 * search range
5876 */
5877 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5878 size = MIN_CACHE_SIZE;
5879 }
5880
5881 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5882 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5883 return 0;
5884 }
5885
5886 /*
5887 * Allocate the working set:
5888 */
5889 cache = vmalloc(max_size);
5890 if (!cache) {
5891 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005892 return 1000000; /* return 1 msec on very small boxen */
akpm@osdl.org198e2f12006-01-12 01:05:30 -08005893 }
5894
5895 while (size <= max_size) {
5896 prev_cost = cost;
5897 cost = measure_cost(cpu1, cpu2, cache, size);
5898
5899 /*
5900 * Update the max:
5901 */
5902 if (cost > 0) {
5903 if (max_cost < cost) {
5904 max_cost = cost;
5905 size_found = size;
5906 }
5907 }
5908 /*
5909 * Calculate average fluctuation, we use this to prevent
5910 * noise from triggering an early break out of the loop:
5911 */
5912 fluct = abs(cost - prev_cost);
5913 avg_fluct = (avg_fluct + fluct)/2;
5914
5915 if (migration_debug)
5916 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5917 cpu1, cpu2, size,
5918 (long)cost / 1000000,
5919 ((long)cost / 100000) % 10,
5920 (long)max_cost / 1000000,
5921 ((long)max_cost / 100000) % 10,
5922 domain_distance(cpu1, cpu2),
5923 cost, avg_fluct);
5924
5925 /*
5926 * If we iterated at least 20% past the previous maximum,
5927 * and the cost has dropped by more than 20% already,
5928 * (taking fluctuations into account) then we assume to
5929 * have found the maximum and break out of the loop early:
5930 */
5931 if (size_found && (size*100 > size_found*SIZE_THRESH))
5932 if (cost+avg_fluct <= 0 ||
5933 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5934
5935 if (migration_debug)
5936 printk("-> found max.\n");
5937 break;
5938 }
5939 /*
Ingo Molnar70b4d632006-01-30 20:24:38 +01005940 * Increase the cachesize in 10% steps:
akpm@osdl.org198e2f12006-01-12 01:05:30 -08005941 */
Ingo Molnar70b4d632006-01-30 20:24:38 +01005942 size = size * 10 / 9;
akpm@osdl.org198e2f12006-01-12 01:05:30 -08005943 }
5944
5945 if (migration_debug)
5946 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5947 cpu1, cpu2, size_found, max_cost);
5948
5949 vfree(cache);
5950
5951 /*
5952 * A task is considered 'cache cold' if at least 2 times
5953 * the worst-case cost of migration has passed.
5954 *
5955 * (this limit is only listened to if the load-balancing
5956 * situation is 'nice' - if there is a large imbalance we
5957 * ignore it for the sake of CPU utilization and
5958 * processing fairness.)
5959 */
5960 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5961}
5962
5963static void calibrate_migration_costs(const cpumask_t *cpu_map)
5964{
5965 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5966 unsigned long j0, j1, distance, max_distance = 0;
5967 struct sched_domain *sd;
5968
5969 j0 = jiffies;
5970
5971 /*
5972 * First pass - calculate the cacheflush times:
5973 */
5974 for_each_cpu_mask(cpu1, *cpu_map) {
5975 for_each_cpu_mask(cpu2, *cpu_map) {
5976 if (cpu1 == cpu2)
5977 continue;
5978 distance = domain_distance(cpu1, cpu2);
5979 max_distance = max(max_distance, distance);
5980 /*
5981 * No result cached yet?
5982 */
5983 if (migration_cost[distance] == -1LL)
5984 migration_cost[distance] =
5985 measure_migration_cost(cpu1, cpu2);
5986 }
5987 }
5988 /*
5989 * Second pass - update the sched domain hierarchy with
5990 * the new cache-hot-time estimations:
5991 */
5992 for_each_cpu_mask(cpu, *cpu_map) {
5993 distance = 0;
5994 for_each_domain(cpu, sd) {
5995 sd->cache_hot_time = migration_cost[distance];
5996 distance++;
5997 }
5998 }
5999 /*
6000 * Print the matrix:
6001 */
6002 if (migration_debug)
6003 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6004 max_cache_size,
6005#ifdef CONFIG_X86
6006 cpu_khz/1000
6007#else
6008 -1
6009#endif
6010 );
Chuck Ebbertbd576c92006-02-04 23:27:42 -08006011 if (system_state == SYSTEM_BOOTING) {
Dave Jones74732642006-10-03 01:14:07 -07006012 if (num_online_cpus() > 1) {
6013 printk("migration_cost=");
6014 for (distance = 0; distance <= max_distance; distance++) {
6015 if (distance)
6016 printk(",");
6017 printk("%ld", (long)migration_cost[distance] / 1000);
6018 }
6019 printk("\n");
Chuck Ebbertbd576c92006-02-04 23:27:42 -08006020 }
akpm@osdl.org198e2f12006-01-12 01:05:30 -08006021 }
akpm@osdl.org198e2f12006-01-12 01:05:30 -08006022 j1 = jiffies;
6023 if (migration_debug)
6024 printk("migration: %ld seconds\n", (j1-j0)/HZ);
6025
6026 /*
6027 * Move back to the original CPU. NUMA-Q gets confused
6028 * if we migrate to another quad during bootup.
6029 */
6030 if (raw_smp_processor_id() != orig_cpu) {
6031 cpumask_t mask = cpumask_of_cpu(orig_cpu),
6032 saved_mask = current->cpus_allowed;
6033
6034 set_cpus_allowed(current, mask);
6035 set_cpus_allowed(current, saved_mask);
6036 }
6037}
6038
John Hawkes9c1cfda2005-09-06 15:18:14 -07006039#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08006040
John Hawkes9c1cfda2005-09-06 15:18:14 -07006041/**
6042 * find_next_best_node - find the next node to include in a sched_domain
6043 * @node: node whose sched_domain we're building
6044 * @used_nodes: nodes already in the sched_domain
6045 *
6046 * Find the next node to include in a given scheduling domain. Simply
6047 * finds the closest node not already in the @used_nodes map.
6048 *
6049 * Should use nodemask_t.
6050 */
6051static int find_next_best_node(int node, unsigned long *used_nodes)
6052{
6053 int i, n, val, min_val, best_node = 0;
6054
6055 min_val = INT_MAX;
6056
6057 for (i = 0; i < MAX_NUMNODES; i++) {
6058 /* Start at @node */
6059 n = (node + i) % MAX_NUMNODES;
6060
6061 if (!nr_cpus_node(n))
6062 continue;
6063
6064 /* Skip already used nodes */
6065 if (test_bit(n, used_nodes))
6066 continue;
6067
6068 /* Simple min distance search */
6069 val = node_distance(node, n);
6070
6071 if (val < min_val) {
6072 min_val = val;
6073 best_node = n;
6074 }
6075 }
6076
6077 set_bit(best_node, used_nodes);
6078 return best_node;
6079}
6080
6081/**
6082 * sched_domain_node_span - get a cpumask for a node's sched_domain
6083 * @node: node whose cpumask we're constructing
6084 * @size: number of nodes to include in this span
6085 *
6086 * Given a node, construct a good cpumask for its sched_domain to span. It
6087 * should be one that prevents unnecessary balancing, but also spreads tasks
6088 * out optimally.
6089 */
6090static cpumask_t sched_domain_node_span(int node)
6091{
John Hawkes9c1cfda2005-09-06 15:18:14 -07006092 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006093 cpumask_t span, nodemask;
6094 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006095
6096 cpus_clear(span);
6097 bitmap_zero(used_nodes, MAX_NUMNODES);
6098
6099 nodemask = node_to_cpumask(node);
6100 cpus_or(span, span, nodemask);
6101 set_bit(node, used_nodes);
6102
6103 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6104 int next_node = find_next_best_node(node, used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006105
John Hawkes9c1cfda2005-09-06 15:18:14 -07006106 nodemask = node_to_cpumask(next_node);
6107 cpus_or(span, span, nodemask);
6108 }
6109
6110 return span;
6111}
6112#endif
6113
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006114int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006115
John Hawkes9c1cfda2005-09-06 15:18:14 -07006116/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07006117 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07006118 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006119#ifdef CONFIG_SCHED_SMT
6120static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6121static struct sched_group sched_group_cpus[NR_CPUS];
Ingo Molnar48f24c42006-07-03 00:25:40 -07006122
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006123static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006124{
6125 return cpu;
6126}
6127#endif
6128
Ingo Molnar48f24c42006-07-03 00:25:40 -07006129/*
6130 * multi-core sched-domains:
6131 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006132#ifdef CONFIG_SCHED_MC
6133static DEFINE_PER_CPU(struct sched_domain, core_domains);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006134static struct sched_group sched_group_core[NR_CPUS];
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006135#endif
6136
6137#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006138static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006139{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006140 cpumask_t mask = cpu_sibling_map[cpu];
6141 cpus_and(mask, mask, *cpu_map);
6142 return first_cpu(mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006143}
6144#elif defined(CONFIG_SCHED_MC)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006145static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006146{
6147 return cpu;
6148}
6149#endif
6150
Linus Torvalds1da177e2005-04-16 15:20:36 -07006151static DEFINE_PER_CPU(struct sched_domain, phys_domains);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006152static struct sched_group sched_group_phys[NR_CPUS];
Ingo Molnar48f24c42006-07-03 00:25:40 -07006153
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006154static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006155{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006156#ifdef CONFIG_SCHED_MC
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006157 cpumask_t mask = cpu_coregroup_map(cpu);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006158 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006159 return first_cpu(mask);
6160#elif defined(CONFIG_SCHED_SMT)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006161 cpumask_t mask = cpu_sibling_map[cpu];
6162 cpus_and(mask, mask, *cpu_map);
6163 return first_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006164#else
6165 return cpu;
6166#endif
6167}
6168
6169#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07006170/*
6171 * The init_sched_build_groups can't handle what we want to do with node
6172 * groups, so roll our own. Now each node has its own list of groups which
6173 * gets dynamically allocated.
6174 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006175static DEFINE_PER_CPU(struct sched_domain, node_domains);
John Hawkesd1b55132005-09-06 15:18:14 -07006176static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
John Hawkes9c1cfda2005-09-06 15:18:14 -07006177
6178static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
John Hawkesd1b55132005-09-06 15:18:14 -07006179static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
John Hawkes9c1cfda2005-09-06 15:18:14 -07006180
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006181static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006182{
6183 return cpu_to_node(cpu);
6184}
Siddha, Suresh B08069032006-03-27 01:15:23 -08006185static void init_numa_sched_groups_power(struct sched_group *group_head)
6186{
6187 struct sched_group *sg = group_head;
6188 int j;
6189
6190 if (!sg)
6191 return;
6192next_sg:
6193 for_each_cpu_mask(j, sg->cpumask) {
6194 struct sched_domain *sd;
6195
6196 sd = &per_cpu(phys_domains, j);
6197 if (j != first_cpu(sd->groups->cpumask)) {
6198 /*
6199 * Only add "power" once for each
6200 * physical package.
6201 */
6202 continue;
6203 }
6204
6205 sg->cpu_power += sd->groups->cpu_power;
6206 }
6207 sg = sg->next;
6208 if (sg != group_head)
6209 goto next_sg;
6210}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006211#endif
6212
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006213#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006214/* Free memory allocated for various sched_group structures */
6215static void free_sched_groups(const cpumask_t *cpu_map)
6216{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006217 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006218
6219 for_each_cpu_mask(cpu, *cpu_map) {
6220 struct sched_group *sched_group_allnodes
6221 = sched_group_allnodes_bycpu[cpu];
6222 struct sched_group **sched_group_nodes
6223 = sched_group_nodes_bycpu[cpu];
6224
6225 if (sched_group_allnodes) {
6226 kfree(sched_group_allnodes);
6227 sched_group_allnodes_bycpu[cpu] = NULL;
6228 }
6229
6230 if (!sched_group_nodes)
6231 continue;
6232
6233 for (i = 0; i < MAX_NUMNODES; i++) {
6234 cpumask_t nodemask = node_to_cpumask(i);
6235 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6236
6237 cpus_and(nodemask, nodemask, *cpu_map);
6238 if (cpus_empty(nodemask))
6239 continue;
6240
6241 if (sg == NULL)
6242 continue;
6243 sg = sg->next;
6244next_sg:
6245 oldsg = sg;
6246 sg = sg->next;
6247 kfree(oldsg);
6248 if (oldsg != sched_group_nodes[i])
6249 goto next_sg;
6250 }
6251 kfree(sched_group_nodes);
6252 sched_group_nodes_bycpu[cpu] = NULL;
6253 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006254}
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006255#else
6256static void free_sched_groups(const cpumask_t *cpu_map)
6257{
6258}
6259#endif
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006260
Linus Torvalds1da177e2005-04-16 15:20:36 -07006261/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006262 * Initialize sched groups cpu_power.
6263 *
6264 * cpu_power indicates the capacity of sched group, which is used while
6265 * distributing the load between different sched groups in a sched domain.
6266 * Typically cpu_power for all the groups in a sched domain will be same unless
6267 * there are asymmetries in the topology. If there are asymmetries, group
6268 * having more cpu_power will pickup more load compared to the group having
6269 * less cpu_power.
6270 *
6271 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6272 * the maximum number of tasks a group can handle in the presence of other idle
6273 * or lightly loaded groups in the same sched domain.
6274 */
6275static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6276{
6277 struct sched_domain *child;
6278 struct sched_group *group;
6279
6280 WARN_ON(!sd || !sd->groups);
6281
6282 if (cpu != first_cpu(sd->groups->cpumask))
6283 return;
6284
6285 child = sd->child;
6286
6287 /*
6288 * For perf policy, if the groups in child domain share resources
6289 * (for example cores sharing some portions of the cache hierarchy
6290 * or SMT), then set this domain groups cpu_power such that each group
6291 * can handle only one task, when there are other idle groups in the
6292 * same sched domain.
6293 */
6294 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6295 (child->flags &
6296 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6297 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6298 return;
6299 }
6300
6301 sd->groups->cpu_power = 0;
6302
6303 /*
6304 * add cpu_power of each child group to this groups cpu_power
6305 */
6306 group = child->groups;
6307 do {
6308 sd->groups->cpu_power += group->cpu_power;
6309 group = group->next;
6310 } while (group != child->groups);
6311}
6312
6313/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006314 * Build sched domains for a given set of cpus and attach the sched domains
6315 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07006316 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006317static int build_sched_domains(const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006318{
6319 int i;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006320 struct sched_domain *sd;
John Hawkesd1b55132005-09-06 15:18:14 -07006321#ifdef CONFIG_NUMA
6322 struct sched_group **sched_group_nodes = NULL;
6323 struct sched_group *sched_group_allnodes = NULL;
6324
6325 /*
6326 * Allocate the per-node list of sched groups
6327 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006328 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
Srivatsa Vaddagirid3a5aa92006-06-27 02:54:39 -07006329 GFP_KERNEL);
John Hawkesd1b55132005-09-06 15:18:14 -07006330 if (!sched_group_nodes) {
6331 printk(KERN_WARNING "Can not alloc sched group node list\n");
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006332 return -ENOMEM;
John Hawkesd1b55132005-09-06 15:18:14 -07006333 }
6334 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6335#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006336
6337 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006338 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006339 */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006340 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006341 int group;
6342 struct sched_domain *sd = NULL, *p;
6343 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6344
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006345 cpus_and(nodemask, nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006346
6347#ifdef CONFIG_NUMA
John Hawkesd1b55132005-09-06 15:18:14 -07006348 if (cpus_weight(*cpu_map)
John Hawkes9c1cfda2005-09-06 15:18:14 -07006349 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
John Hawkesd1b55132005-09-06 15:18:14 -07006350 if (!sched_group_allnodes) {
6351 sched_group_allnodes
Christoph Lameterce164422006-10-03 01:14:11 -07006352 = kmalloc_node(sizeof(struct sched_group)
6353 * MAX_NUMNODES,
6354 GFP_KERNEL,
6355 cpu_to_node(i));
John Hawkesd1b55132005-09-06 15:18:14 -07006356 if (!sched_group_allnodes) {
6357 printk(KERN_WARNING
6358 "Can not alloc allnodes sched group\n");
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006359 goto error;
John Hawkesd1b55132005-09-06 15:18:14 -07006360 }
6361 sched_group_allnodes_bycpu[i]
6362 = sched_group_allnodes;
6363 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006364 sd = &per_cpu(allnodes_domains, i);
6365 *sd = SD_ALLNODES_INIT;
6366 sd->span = *cpu_map;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006367 group = cpu_to_allnodes_group(i, cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006368 sd->groups = &sched_group_allnodes[group];
6369 p = sd;
6370 } else
6371 p = NULL;
6372
Linus Torvalds1da177e2005-04-16 15:20:36 -07006373 sd = &per_cpu(node_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006374 *sd = SD_NODE_INIT;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006375 sd->span = sched_domain_node_span(cpu_to_node(i));
6376 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006377 if (p)
6378 p->child = sd;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006379 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006380#endif
6381
6382 p = sd;
6383 sd = &per_cpu(phys_domains, i);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006384 group = cpu_to_phys_group(i, cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006385 *sd = SD_CPU_INIT;
6386 sd->span = nodemask;
6387 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006388 if (p)
6389 p->child = sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006390 sd->groups = &sched_group_phys[group];
6391
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006392#ifdef CONFIG_SCHED_MC
6393 p = sd;
6394 sd = &per_cpu(core_domains, i);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006395 group = cpu_to_core_group(i, cpu_map);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006396 *sd = SD_MC_INIT;
6397 sd->span = cpu_coregroup_map(i);
6398 cpus_and(sd->span, sd->span, *cpu_map);
6399 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006400 p->child = sd;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006401 sd->groups = &sched_group_core[group];
6402#endif
6403
Linus Torvalds1da177e2005-04-16 15:20:36 -07006404#ifdef CONFIG_SCHED_SMT
6405 p = sd;
6406 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006407 group = cpu_to_cpu_group(i, cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006408 *sd = SD_SIBLING_INIT;
6409 sd->span = cpu_sibling_map[i];
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006410 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006411 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006412 p->child = sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006413 sd->groups = &sched_group_cpus[group];
6414#endif
6415 }
6416
6417#ifdef CONFIG_SCHED_SMT
6418 /* Set up CPU (sibling) groups */
John Hawkes9c1cfda2005-09-06 15:18:14 -07006419 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006420 cpumask_t this_sibling_map = cpu_sibling_map[i];
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006421 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006422 if (i != first_cpu(this_sibling_map))
6423 continue;
6424
6425 init_sched_build_groups(sched_group_cpus, this_sibling_map,
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006426 cpu_map, &cpu_to_cpu_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006427 }
6428#endif
6429
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006430#ifdef CONFIG_SCHED_MC
6431 /* Set up multi-core groups */
6432 for_each_cpu_mask(i, *cpu_map) {
6433 cpumask_t this_core_map = cpu_coregroup_map(i);
6434 cpus_and(this_core_map, this_core_map, *cpu_map);
6435 if (i != first_cpu(this_core_map))
6436 continue;
6437 init_sched_build_groups(sched_group_core, this_core_map,
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006438 cpu_map, &cpu_to_core_group);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006439 }
6440#endif
6441
6442
Linus Torvalds1da177e2005-04-16 15:20:36 -07006443 /* Set up physical groups */
6444 for (i = 0; i < MAX_NUMNODES; i++) {
6445 cpumask_t nodemask = node_to_cpumask(i);
6446
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006447 cpus_and(nodemask, nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006448 if (cpus_empty(nodemask))
6449 continue;
6450
6451 init_sched_build_groups(sched_group_phys, nodemask,
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006452 cpu_map, &cpu_to_phys_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006453 }
6454
6455#ifdef CONFIG_NUMA
6456 /* Set up node groups */
John Hawkesd1b55132005-09-06 15:18:14 -07006457 if (sched_group_allnodes)
6458 init_sched_build_groups(sched_group_allnodes, *cpu_map,
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006459 cpu_map, &cpu_to_allnodes_group);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006460
6461 for (i = 0; i < MAX_NUMNODES; i++) {
6462 /* Set up node groups */
6463 struct sched_group *sg, *prev;
6464 cpumask_t nodemask = node_to_cpumask(i);
6465 cpumask_t domainspan;
6466 cpumask_t covered = CPU_MASK_NONE;
6467 int j;
6468
6469 cpus_and(nodemask, nodemask, *cpu_map);
John Hawkesd1b55132005-09-06 15:18:14 -07006470 if (cpus_empty(nodemask)) {
6471 sched_group_nodes[i] = NULL;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006472 continue;
John Hawkesd1b55132005-09-06 15:18:14 -07006473 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006474
6475 domainspan = sched_domain_node_span(i);
6476 cpus_and(domainspan, domainspan, *cpu_map);
6477
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07006478 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006479 if (!sg) {
6480 printk(KERN_WARNING "Can not alloc domain group for "
6481 "node %d\n", i);
6482 goto error;
6483 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006484 sched_group_nodes[i] = sg;
6485 for_each_cpu_mask(j, nodemask) {
6486 struct sched_domain *sd;
6487 sd = &per_cpu(node_domains, j);
6488 sd->groups = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006489 }
6490 sg->cpu_power = 0;
6491 sg->cpumask = nodemask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006492 sg->next = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006493 cpus_or(covered, covered, nodemask);
6494 prev = sg;
6495
6496 for (j = 0; j < MAX_NUMNODES; j++) {
6497 cpumask_t tmp, notcovered;
6498 int n = (i + j) % MAX_NUMNODES;
6499
6500 cpus_complement(notcovered, covered);
6501 cpus_and(tmp, notcovered, *cpu_map);
6502 cpus_and(tmp, tmp, domainspan);
6503 if (cpus_empty(tmp))
6504 break;
6505
6506 nodemask = node_to_cpumask(n);
6507 cpus_and(tmp, tmp, nodemask);
6508 if (cpus_empty(tmp))
6509 continue;
6510
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07006511 sg = kmalloc_node(sizeof(struct sched_group),
6512 GFP_KERNEL, i);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006513 if (!sg) {
6514 printk(KERN_WARNING
6515 "Can not alloc domain group for node %d\n", j);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006516 goto error;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006517 }
6518 sg->cpu_power = 0;
6519 sg->cpumask = tmp;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006520 sg->next = prev->next;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006521 cpus_or(covered, covered, tmp);
6522 prev->next = sg;
6523 prev = sg;
6524 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006525 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006526#endif
6527
6528 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006529#ifdef CONFIG_SCHED_SMT
6530 for_each_cpu_mask(i, *cpu_map) {
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006531 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006532 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006533 }
6534#endif
6535#ifdef CONFIG_SCHED_MC
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006536 for_each_cpu_mask(i, *cpu_map) {
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006537 sd = &per_cpu(core_domains, i);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006538 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006539 }
6540#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006541
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006542 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006543 sd = &per_cpu(phys_domains, i);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006544 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006545 }
6546
John Hawkes9c1cfda2005-09-06 15:18:14 -07006547#ifdef CONFIG_NUMA
Siddha, Suresh B08069032006-03-27 01:15:23 -08006548 for (i = 0; i < MAX_NUMNODES; i++)
6549 init_numa_sched_groups_power(sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006550
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07006551 if (sched_group_allnodes) {
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006552 int group = cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07006553 struct sched_group *sg = &sched_group_allnodes[group];
6554
6555 init_numa_sched_groups_power(sg);
6556 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006557#endif
6558
Linus Torvalds1da177e2005-04-16 15:20:36 -07006559 /* Attach the domains */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006560 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006561 struct sched_domain *sd;
6562#ifdef CONFIG_SCHED_SMT
6563 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006564#elif defined(CONFIG_SCHED_MC)
6565 sd = &per_cpu(core_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006566#else
6567 sd = &per_cpu(phys_domains, i);
6568#endif
6569 cpu_attach_domain(sd, i);
6570 }
akpm@osdl.org198e2f12006-01-12 01:05:30 -08006571 /*
6572 * Tune cache-hot values:
6573 */
6574 calibrate_migration_costs(cpu_map);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006575
6576 return 0;
6577
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006578#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006579error:
6580 free_sched_groups(cpu_map);
6581 return -ENOMEM;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006582#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006583}
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006584/*
6585 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6586 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006587static int arch_init_sched_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006588{
6589 cpumask_t cpu_default_map;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006590 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006591
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006592 /*
6593 * Setup mask for cpus without special case scheduling requirements.
6594 * For now this just excludes isolated cpus, but could be used to
6595 * exclude other special cases in the future.
6596 */
6597 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6598
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006599 err = build_sched_domains(&cpu_default_map);
6600
6601 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006602}
6603
6604static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006605{
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006606 free_sched_groups(cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006607}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006608
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006609/*
6610 * Detach sched domains from a group of cpus specified in cpu_map
6611 * These cpus will now be attached to the NULL domain
6612 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08006613static void detach_destroy_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006614{
6615 int i;
6616
6617 for_each_cpu_mask(i, *cpu_map)
6618 cpu_attach_domain(NULL, i);
6619 synchronize_sched();
6620 arch_destroy_sched_domains(cpu_map);
6621}
6622
6623/*
6624 * Partition sched domains as specified by the cpumasks below.
6625 * This attaches all cpus from the cpumasks to the NULL domain,
6626 * waits for a RCU quiescent period, recalculates sched
6627 * domain information and then attaches them back to the
6628 * correct sched domains
6629 * Call with hotplug lock held
6630 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006631int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006632{
6633 cpumask_t change_map;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006634 int err = 0;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006635
6636 cpus_and(*partition1, *partition1, cpu_online_map);
6637 cpus_and(*partition2, *partition2, cpu_online_map);
6638 cpus_or(change_map, *partition1, *partition2);
6639
6640 /* Detach sched domains from all of the affected cpus */
6641 detach_destroy_domains(&change_map);
6642 if (!cpus_empty(*partition1))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006643 err = build_sched_domains(partition1);
6644 if (!err && !cpus_empty(*partition2))
6645 err = build_sched_domains(partition2);
6646
6647 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006648}
6649
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006650#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6651int arch_reinit_sched_domains(void)
6652{
6653 int err;
6654
6655 lock_cpu_hotplug();
6656 detach_destroy_domains(&cpu_online_map);
6657 err = arch_init_sched_domains(&cpu_online_map);
6658 unlock_cpu_hotplug();
6659
6660 return err;
6661}
6662
6663static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6664{
6665 int ret;
6666
6667 if (buf[0] != '0' && buf[0] != '1')
6668 return -EINVAL;
6669
6670 if (smt)
6671 sched_smt_power_savings = (buf[0] == '1');
6672 else
6673 sched_mc_power_savings = (buf[0] == '1');
6674
6675 ret = arch_reinit_sched_domains();
6676
6677 return ret ? ret : count;
6678}
6679
6680int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6681{
6682 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006683
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006684#ifdef CONFIG_SCHED_SMT
6685 if (smt_capable())
6686 err = sysfs_create_file(&cls->kset.kobj,
6687 &attr_sched_smt_power_savings.attr);
6688#endif
6689#ifdef CONFIG_SCHED_MC
6690 if (!err && mc_capable())
6691 err = sysfs_create_file(&cls->kset.kobj,
6692 &attr_sched_mc_power_savings.attr);
6693#endif
6694 return err;
6695}
6696#endif
6697
6698#ifdef CONFIG_SCHED_MC
6699static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6700{
6701 return sprintf(page, "%u\n", sched_mc_power_savings);
6702}
Ingo Molnar48f24c42006-07-03 00:25:40 -07006703static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6704 const char *buf, size_t count)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006705{
6706 return sched_power_savings_store(buf, count, 0);
6707}
6708SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6709 sched_mc_power_savings_store);
6710#endif
6711
6712#ifdef CONFIG_SCHED_SMT
6713static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6714{
6715 return sprintf(page, "%u\n", sched_smt_power_savings);
6716}
Ingo Molnar48f24c42006-07-03 00:25:40 -07006717static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6718 const char *buf, size_t count)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006719{
6720 return sched_power_savings_store(buf, count, 1);
6721}
6722SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6723 sched_smt_power_savings_store);
6724#endif
6725
6726
Linus Torvalds1da177e2005-04-16 15:20:36 -07006727#ifdef CONFIG_HOTPLUG_CPU
6728/*
6729 * Force a reinitialization of the sched domains hierarchy. The domains
6730 * and groups cannot be updated in place without racing with the balancing
Nick Piggin41c7ce92005-06-25 14:57:24 -07006731 * code, so we temporarily attach all running cpus to the NULL domain
Linus Torvalds1da177e2005-04-16 15:20:36 -07006732 * which will prevent rebalancing while the sched domains are recalculated.
6733 */
6734static int update_sched_domains(struct notifier_block *nfb,
6735 unsigned long action, void *hcpu)
6736{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006737 switch (action) {
6738 case CPU_UP_PREPARE:
6739 case CPU_DOWN_PREPARE:
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006740 detach_destroy_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006741 return NOTIFY_OK;
6742
6743 case CPU_UP_CANCELED:
6744 case CPU_DOWN_FAILED:
6745 case CPU_ONLINE:
6746 case CPU_DEAD:
6747 /*
6748 * Fall through and re-initialise the domains.
6749 */
6750 break;
6751 default:
6752 return NOTIFY_DONE;
6753 }
6754
6755 /* The hotplug lock is already held by cpu_up/cpu_down */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006756 arch_init_sched_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006757
6758 return NOTIFY_OK;
6759}
6760#endif
6761
6762void __init sched_init_smp(void)
6763{
Nick Piggin5c1e1762006-10-03 01:14:04 -07006764 cpumask_t non_isolated_cpus;
6765
Linus Torvalds1da177e2005-04-16 15:20:36 -07006766 lock_cpu_hotplug();
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006767 arch_init_sched_domains(&cpu_online_map);
Nick Piggin5c1e1762006-10-03 01:14:04 -07006768 cpus_andnot(non_isolated_cpus, cpu_online_map, cpu_isolated_map);
6769 if (cpus_empty(non_isolated_cpus))
6770 cpu_set(smp_processor_id(), non_isolated_cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006771 unlock_cpu_hotplug();
6772 /* XXX: Theoretical race here - CPU may be hotplugged now */
6773 hotcpu_notifier(update_sched_domains, 0);
Nick Piggin5c1e1762006-10-03 01:14:04 -07006774
6775 /* Move init over to a non-isolated CPU */
6776 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6777 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006778}
6779#else
6780void __init sched_init_smp(void)
6781{
6782}
6783#endif /* CONFIG_SMP */
6784
6785int in_sched_functions(unsigned long addr)
6786{
6787 /* Linker adds these: start and end of __sched functions */
6788 extern char __sched_text_start[], __sched_text_end[];
Ingo Molnar48f24c42006-07-03 00:25:40 -07006789
Linus Torvalds1da177e2005-04-16 15:20:36 -07006790 return in_lock_functions(addr) ||
6791 (addr >= (unsigned long)__sched_text_start
6792 && addr < (unsigned long)__sched_text_end);
6793}
6794
6795void __init sched_init(void)
6796{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006797 int i, j, k;
6798
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08006799 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07006800 struct prio_array *array;
6801 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006802
6803 rq = cpu_rq(i);
6804 spin_lock_init(&rq->lock);
Ingo Molnarfcb99372006-07-03 00:25:10 -07006805 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
Nick Piggin78979862005-06-25 14:57:13 -07006806 rq->nr_running = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006807 rq->active = rq->arrays;
6808 rq->expired = rq->arrays + 1;
6809 rq->best_expired_prio = MAX_PRIO;
6810
6811#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07006812 rq->sd = NULL;
Nick Piggin78979862005-06-25 14:57:13 -07006813 for (j = 1; j < 3; j++)
6814 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006815 rq->active_balance = 0;
6816 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07006817 rq->cpu = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006818 rq->migration_thread = NULL;
6819 INIT_LIST_HEAD(&rq->migration_queue);
6820#endif
6821 atomic_set(&rq->nr_iowait, 0);
6822
6823 for (j = 0; j < 2; j++) {
6824 array = rq->arrays + j;
6825 for (k = 0; k < MAX_PRIO; k++) {
6826 INIT_LIST_HEAD(array->queue + k);
6827 __clear_bit(k, array->bitmap);
6828 }
6829 // delimiter for bitsearch
6830 __set_bit(MAX_PRIO, array->bitmap);
6831 }
6832 }
6833
Peter Williams2dd73a42006-06-27 02:54:34 -07006834 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07006835
6836#ifdef CONFIG_RT_MUTEXES
6837 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6838#endif
6839
Linus Torvalds1da177e2005-04-16 15:20:36 -07006840 /*
6841 * The boot idle thread does lazy MMU switching as well:
6842 */
6843 atomic_inc(&init_mm.mm_count);
6844 enter_lazy_tlb(&init_mm, current);
6845
6846 /*
6847 * Make us the idle thread. Technically, schedule() should not be
6848 * called from this thread, however somewhere below it might be,
6849 * but because we are the idle thread, we just pick up running again
6850 * when this runqueue becomes "idle".
6851 */
6852 init_idle(current, smp_processor_id());
6853}
6854
6855#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6856void __might_sleep(char *file, int line)
6857{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006858#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07006859 static unsigned long prev_jiffy; /* ratelimiting */
6860
6861 if ((in_atomic() || irqs_disabled()) &&
6862 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6863 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6864 return;
6865 prev_jiffy = jiffies;
Ingo Molnar91368d72006-03-23 03:00:54 -08006866 printk(KERN_ERR "BUG: sleeping function called from invalid"
Linus Torvalds1da177e2005-04-16 15:20:36 -07006867 " context at %s:%d\n", file, line);
6868 printk("in_atomic():%d, irqs_disabled():%d\n",
6869 in_atomic(), irqs_disabled());
6870 dump_stack();
6871 }
6872#endif
6873}
6874EXPORT_SYMBOL(__might_sleep);
6875#endif
6876
6877#ifdef CONFIG_MAGIC_SYSRQ
6878void normalize_rt_tasks(void)
6879{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006880 struct prio_array *array;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006881 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006882 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006883 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006884
6885 read_lock_irq(&tasklist_lock);
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07006886 for_each_process(p) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006887 if (!rt_task(p))
6888 continue;
6889
Ingo Molnarb29739f2006-06-27 02:54:51 -07006890 spin_lock_irqsave(&p->pi_lock, flags);
6891 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006892
6893 array = p->array;
6894 if (array)
6895 deactivate_task(p, task_rq(p));
6896 __setscheduler(p, SCHED_NORMAL, 0);
6897 if (array) {
6898 __activate_task(p, task_rq(p));
6899 resched_task(rq->curr);
6900 }
6901
Ingo Molnarb29739f2006-06-27 02:54:51 -07006902 __task_rq_unlock(rq);
6903 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006904 }
6905 read_unlock_irq(&tasklist_lock);
6906}
6907
6908#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07006909
6910#ifdef CONFIG_IA64
6911/*
6912 * These functions are only useful for the IA64 MCA handling.
6913 *
6914 * They can only be called when the whole system has been
6915 * stopped - every CPU needs to be quiescent, and no scheduling
6916 * activity can take place. Using them for anything else would
6917 * be a serious bug, and as a result, they aren't even visible
6918 * under any other configuration.
6919 */
6920
6921/**
6922 * curr_task - return the current task for a given cpu.
6923 * @cpu: the processor in question.
6924 *
6925 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6926 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006927struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07006928{
6929 return cpu_curr(cpu);
6930}
6931
6932/**
6933 * set_curr_task - set the current task for a given cpu.
6934 * @cpu: the processor in question.
6935 * @p: the task pointer to set.
6936 *
6937 * Description: This function must only be used when non-maskable interrupts
6938 * are serviced on a separate stack. It allows the architecture to switch the
6939 * notion of the current task on a cpu in a non-blocking manner. This function
6940 * must be called with all CPU's synchronized, and interrupts disabled, the
6941 * and caller must save the original value of the current task (see
6942 * curr_task() above) and restore that value before reenabling interrupts and
6943 * re-starting the system.
6944 *
6945 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6946 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006947void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07006948{
6949 cpu_curr(cpu) = p;
6950}
6951
6952#endif