blob: 9474b23c28bf41f5989df3b94c5e810b0f1e971a [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
Ingo Molnarc31f2e82007-07-09 18:52:01 +020019 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
Ingo Molnarb9131762008-01-25 21:08:19 +010025 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
Linus Torvalds1da177e2005-04-16 15:20:36 -070027 */
28
29#include <linux/mm.h>
30#include <linux/module.h>
31#include <linux/nmi.h>
32#include <linux/init.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020033#include <linux/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <linux/highmem.h>
35#include <linux/smp_lock.h>
36#include <asm/mmu_context.h>
37#include <linux/interrupt.h>
Randy.Dunlapc59ede72006-01-11 12:17:46 -080038#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <linux/completion.h>
40#include <linux/kernel_stat.h>
Ingo Molnar9a11b49a2006-07-03 00:24:33 -070041#include <linux/debug_locks.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070042#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/profile.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080045#include <linux/freezer.h>
akpm@osdl.org198e2f12006-01-12 01:05:30 -080046#include <linux/vmalloc.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070047#include <linux/blkdev.h>
48#include <linux/delay.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070049#include <linux/pid_namespace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070050#include <linux/smp.h>
51#include <linux/threads.h>
52#include <linux/timer.h>
53#include <linux/rcupdate.h>
54#include <linux/cpu.h>
55#include <linux/cpuset.h>
56#include <linux/percpu.h>
57#include <linux/kthread.h>
58#include <linux/seq_file.h>
Nick Piggine692ab52007-07-26 13:40:43 +020059#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060#include <linux/syscalls.h>
61#include <linux/times.h>
Jay Lan8f0ab512006-09-30 23:28:59 -070062#include <linux/tsacct_kern.h>
bibo maoc6fd91f2006-03-26 01:38:20 -080063#include <linux/kprobes.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070064#include <linux/delayacct.h>
Eric Dumazet5517d862007-05-08 00:32:57 -070065#include <linux/reciprocal_div.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020066#include <linux/unistd.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020067#include <linux/pagemap.h>
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +010068#include <linux/hrtimer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070069
Eric Dumazet5517d862007-05-08 00:32:57 -070070#include <asm/tlb.h>
Satyam Sharma838225b2007-10-24 18:23:50 +020071#include <asm/irq_regs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070072
73/*
Alexey Dobriyanb035b6d2007-02-10 01:45:10 -080074 * Scheduler clock - returns current time in nanosec units.
75 * This is default implementation.
76 * Architectures and sub-architectures can override this.
77 */
78unsigned long long __attribute__((weak)) sched_clock(void)
79{
Eric Dumazetd6322fa2007-11-09 22:39:38 +010080 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
Alexey Dobriyanb035b6d2007-02-10 01:45:10 -080081}
82
83/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
86 * and back.
87 */
88#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
91
92/*
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
96 */
97#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
100
101/*
Ingo Molnard7876a02008-01-25 21:08:19 +0100102 * Helpers for converting nanosecond timing to jiffy resolution
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 */
Eric Dumazetd6322fa2007-11-09 22:39:38 +0100104#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200106#define NICE_0_LOAD SCHED_LOAD_SCALE
107#define NICE_0_SHIFT SCHED_LOAD_SHIFT
108
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109/*
110 * These are the 'tuning knobs' of the scheduler:
111 *
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +0200112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 * Timeslices get refilled after they expire.
114 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115#define DEF_TIMESLICE (100 * HZ / 1000)
Peter Williams2dd73a42006-06-27 02:54:34 -0700116
Eric Dumazet5517d862007-05-08 00:32:57 -0700117#ifdef CONFIG_SMP
118/*
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
121 */
122static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
123{
124 return reciprocal_divide(load, sg->reciprocal_cpu_power);
125}
126
127/*
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
130 */
131static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
132{
133 sg->__cpu_power += val;
134 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
135}
136#endif
137
Ingo Molnare05606d2007-07-09 18:51:59 +0200138static inline int rt_policy(int policy)
139{
140 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
141 return 1;
142 return 0;
143}
144
145static inline int task_has_rt_policy(struct task_struct *p)
146{
147 return rt_policy(p->policy);
148}
149
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150/*
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200151 * This is the priority-queue data structure of the RT scheduling class:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200153struct rt_prio_array {
154 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
155 struct list_head queue[MAX_RT_PRIO];
156};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200158#ifdef CONFIG_FAIR_GROUP_SCHED
159
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700160#include <linux/cgroup.h>
161
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200162struct cfs_rq;
163
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100164static LIST_HEAD(task_groups);
165
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200166/* task group related information */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200167struct task_group {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700168#ifdef CONFIG_FAIR_CGROUP_SCHED
169 struct cgroup_subsys_state css;
170#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200171 /* schedulable entities of this group on each cpu */
172 struct sched_entity **se;
173 /* runqueue "owned" by this group on each cpu */
174 struct cfs_rq **cfs_rq;
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100175
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100176 struct sched_rt_entity **rt_se;
177 struct rt_rq **rt_rq;
178
179 unsigned int rt_ratio;
180
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100181 /*
182 * shares assigned to a task group governs how much of cpu bandwidth
183 * is allocated to the group. The more shares a group has, the more is
184 * the cpu bandwidth allocated to it.
185 *
186 * For ex, lets say that there are three task groups, A, B and C which
187 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
188 * cpu bandwidth allocated by the scheduler to task groups A, B and C
189 * should be:
190 *
191 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
192 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
Ingo Molnar03319ec2008-01-25 21:08:28 +0100193 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100194 *
195 * The weight assigned to a task group's schedulable entities on every
196 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
197 * group's shares. For ex: lets say that task group A has been
198 * assigned shares of 1000 and there are two CPUs in a system. Then,
199 *
200 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
201 *
202 * Note: It's not necessary that each of a task's group schedulable
Ingo Molnar03319ec2008-01-25 21:08:28 +0100203 * entity have the same weight on all CPUs. If the group
204 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
205 * better distribution of weight could be:
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100206 *
207 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
208 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
209 *
210 * rebalance_shares() is responsible for distributing the shares of a
211 * task groups like this among the group's schedulable entities across
212 * cpus.
213 *
214 */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200215 unsigned long shares;
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100216
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +0100217 struct rcu_head rcu;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100218 struct list_head list;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200219};
220
221/* Default task group's sched entity on each cpu */
222static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
223/* Default task group's cfs_rq on each cpu */
224static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
225
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100226static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
227static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
228
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200229static struct sched_entity *init_sched_entity_p[NR_CPUS];
230static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200231
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100232static struct sched_rt_entity *init_sched_rt_entity_p[NR_CPUS];
233static struct rt_rq *init_rt_rq_p[NR_CPUS];
234
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100235/* task_group_mutex serializes add/remove of task groups and also changes to
236 * a task group's cpu shares.
237 */
238static DEFINE_MUTEX(task_group_mutex);
239
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +0100240/* doms_cur_mutex serializes access to doms_cur[] array */
241static DEFINE_MUTEX(doms_cur_mutex);
242
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100243#ifdef CONFIG_SMP
244/* kernel thread that runs rebalance_shares() periodically */
245static struct task_struct *lb_monitor_task;
246static int load_balance_monitor(void *unused);
247#endif
248
249static void set_se_shares(struct sched_entity *se, unsigned long shares);
250
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200251/* Default task group.
Ingo Molnar3a252012007-10-15 17:00:12 +0200252 * Every task in system belong to this group at bootup.
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200253 */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200254struct task_group init_task_group = {
Ingo Molnar0eab9142008-01-25 21:08:19 +0100255 .se = init_sched_entity_p,
Ingo Molnar3a252012007-10-15 17:00:12 +0200256 .cfs_rq = init_cfs_rq_p,
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100257
258 .rt_se = init_sched_rt_entity_p,
259 .rt_rq = init_rt_rq_p,
Ingo Molnar3a252012007-10-15 17:00:12 +0200260};
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200261
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200262#ifdef CONFIG_FAIR_USER_SCHED
Ingo Molnar0eab9142008-01-25 21:08:19 +0100263# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200264#else
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100265# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200266#endif
267
Ingo Molnar0eab9142008-01-25 21:08:19 +0100268#define MIN_GROUP_SHARES 2
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100269
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100270static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200271
272/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200273static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200274{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200275 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200276
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200277#ifdef CONFIG_FAIR_USER_SCHED
278 tg = p->user->tg;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700279#elif defined(CONFIG_FAIR_CGROUP_SCHED)
280 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
281 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200282#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100283 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200284#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200285 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200286}
287
288/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100289static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200290{
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100291 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
292 p->se.parent = task_group(p)->se[cpu];
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100293
294 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
295 p->rt.parent = task_group(p)->rt_se[cpu];
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200296}
297
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100298static inline void lock_task_group_list(void)
299{
300 mutex_lock(&task_group_mutex);
301}
302
303static inline void unlock_task_group_list(void)
304{
305 mutex_unlock(&task_group_mutex);
306}
307
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +0100308static inline void lock_doms_cur(void)
309{
310 mutex_lock(&doms_cur_mutex);
311}
312
313static inline void unlock_doms_cur(void)
314{
315 mutex_unlock(&doms_cur_mutex);
316}
317
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200318#else
319
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100320static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100321static inline void lock_task_group_list(void) { }
322static inline void unlock_task_group_list(void) { }
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +0100323static inline void lock_doms_cur(void) { }
324static inline void unlock_doms_cur(void) { }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200325
326#endif /* CONFIG_FAIR_GROUP_SCHED */
327
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200328/* CFS-related fields in a runqueue */
329struct cfs_rq {
330 struct load_weight load;
331 unsigned long nr_running;
332
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200333 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200334 u64 min_vruntime;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200335
336 struct rb_root tasks_timeline;
337 struct rb_node *rb_leftmost;
338 struct rb_node *rb_load_balance_curr;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200339 /* 'curr' points to currently running entity on this cfs_rq.
340 * It is set to NULL otherwise (i.e when none are currently running).
341 */
342 struct sched_entity *curr;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200343
344 unsigned long nr_spread_over;
345
Ingo Molnar62160e32007-10-15 17:00:03 +0200346#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200347 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
348
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100349 /*
350 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200351 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
352 * (like users, containers etc.)
353 *
354 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
355 * list is used during load balance.
356 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100357 struct list_head leaf_cfs_rq_list;
358 struct task_group *tg; /* group that "owns" this runqueue */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200359#endif
360};
361
362/* Real-Time classes' related field in a runqueue: */
363struct rt_rq {
364 struct rt_prio_array active;
Steven Rostedt63489e42008-01-25 21:08:03 +0100365 unsigned long rt_nr_running;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100366#if defined CONFIG_SMP || defined CONFIG_FAIR_GROUP_SCHED
367 int highest_prio; /* highest queued rt task prio */
368#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100369#ifdef CONFIG_SMP
Gregory Haskins73fe6aa2008-01-25 21:08:07 +0100370 unsigned long rt_nr_migratory;
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +0100371 int overloaded;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100372#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100373 int rt_throttled;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100374 u64 rt_time;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100375
376#ifdef CONFIG_FAIR_GROUP_SCHED
377 struct rq *rq;
378 struct list_head leaf_rt_rq_list;
379 struct task_group *tg;
380 struct sched_rt_entity *rt_se;
381#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200382};
383
Gregory Haskins57d885f2008-01-25 21:08:18 +0100384#ifdef CONFIG_SMP
385
386/*
387 * We add the notion of a root-domain which will be used to define per-domain
Ingo Molnar0eab9142008-01-25 21:08:19 +0100388 * variables. Each exclusive cpuset essentially defines an island domain by
389 * fully partitioning the member cpus from any other cpuset. Whenever a new
Gregory Haskins57d885f2008-01-25 21:08:18 +0100390 * exclusive cpuset is created, we also create and attach a new root-domain
391 * object.
392 *
Gregory Haskins57d885f2008-01-25 21:08:18 +0100393 */
394struct root_domain {
395 atomic_t refcount;
396 cpumask_t span;
397 cpumask_t online;
Gregory Haskins637f5082008-01-25 21:08:18 +0100398
Ingo Molnar0eab9142008-01-25 21:08:19 +0100399 /*
Gregory Haskins637f5082008-01-25 21:08:18 +0100400 * The "RT overload" flag: it gets set if a CPU has more than
401 * one runnable RT task.
402 */
403 cpumask_t rto_mask;
Ingo Molnar0eab9142008-01-25 21:08:19 +0100404 atomic_t rto_count;
Gregory Haskins57d885f2008-01-25 21:08:18 +0100405};
406
Gregory Haskinsdc938522008-01-25 21:08:26 +0100407/*
408 * By default the system creates a single root-domain with all cpus as
409 * members (mimicking the global state we have today).
410 */
Gregory Haskins57d885f2008-01-25 21:08:18 +0100411static struct root_domain def_root_domain;
412
413#endif
414
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200415/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 * This is the main, per-CPU runqueue data structure.
417 *
418 * Locking rule: those places that want to lock multiple runqueues
419 * (such as the load balancing or the thread migration code), lock
420 * acquire operations must be ordered by ascending &runqueue.
421 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700422struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200423 /* runqueue lock: */
424 spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425
426 /*
427 * nr_running and cpu_load should be in the same cacheline because
428 * remote CPUs use both these fields when doing load calculation.
429 */
430 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200431 #define CPU_LOAD_IDX_MAX 5
432 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh Bbdecea32007-05-08 00:32:48 -0700433 unsigned char idle_at_tick;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700434#ifdef CONFIG_NO_HZ
435 unsigned char in_nohz_recently;
436#endif
Ingo Molnard8016492007-10-18 21:32:55 +0200437 /* capture load from *all* tasks on this cpu: */
438 struct load_weight load;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200439 unsigned long nr_load_updates;
440 u64 nr_switches;
441
442 struct cfs_rq cfs;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100443 struct rt_rq rt;
444 u64 rt_period_expire;
Peter Zijlstra48d5e252008-01-25 21:08:31 +0100445 int rt_throttled;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100446
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200447#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200448 /* list of leaf cfs_rq on this cpu: */
449 struct list_head leaf_cfs_rq_list;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100450 struct list_head leaf_rt_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
453 /*
454 * This is part of a global counter where only the total sum
455 * over all CPUs matters. A task can increase this counter on
456 * one CPU and if it got migrated afterwards it may decrease
457 * it on another CPU. Always updated under the runqueue lock:
458 */
459 unsigned long nr_uninterruptible;
460
Ingo Molnar36c8b582006-07-03 00:25:41 -0700461 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800462 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200464
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200465 u64 clock, prev_clock_raw;
466 s64 clock_max_delta;
467
Guillaume Chazaraincc203d22008-01-25 21:08:34 +0100468 unsigned int clock_warps, clock_overflows, clock_underflows;
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200469 u64 idle_clock;
470 unsigned int clock_deep_idle_events;
Ingo Molnar529c7722007-08-10 23:05:11 +0200471 u64 tick_timestamp;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200472
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 atomic_t nr_iowait;
474
475#ifdef CONFIG_SMP
Ingo Molnar0eab9142008-01-25 21:08:19 +0100476 struct root_domain *rd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 struct sched_domain *sd;
478
479 /* For active balancing */
480 int active_balance;
481 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200482 /* cpu of this runqueue: */
483 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484
Ingo Molnar36c8b582006-07-03 00:25:41 -0700485 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 struct list_head migration_queue;
487#endif
488
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100489#ifdef CONFIG_SCHED_HRTICK
490 unsigned long hrtick_flags;
491 ktime_t hrtick_expire;
492 struct hrtimer hrtick_timer;
493#endif
494
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495#ifdef CONFIG_SCHEDSTATS
496 /* latency stats */
497 struct sched_info rq_sched_info;
498
499 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200500 unsigned int yld_exp_empty;
501 unsigned int yld_act_empty;
502 unsigned int yld_both_empty;
503 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504
505 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200506 unsigned int sched_switch;
507 unsigned int sched_count;
508 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509
510 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200511 unsigned int ttwu_count;
512 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200513
514 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200515 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516#endif
Ingo Molnarfcb99372006-07-03 00:25:10 -0700517 struct lock_class_key rq_lock_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518};
519
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700520static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521
Ingo Molnardd41f592007-07-09 18:51:59 +0200522static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
523{
524 rq->curr->sched_class->check_preempt_curr(rq, p);
525}
526
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700527static inline int cpu_of(struct rq *rq)
528{
529#ifdef CONFIG_SMP
530 return rq->cpu;
531#else
532 return 0;
533#endif
534}
535
Nick Piggin674311d2005-06-25 14:57:27 -0700536/*
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200537 * Update the per-runqueue clock, as finegrained as the platform can give
538 * us, but without assuming monotonicity, etc.:
Ingo Molnar20d315d2007-07-09 18:51:58 +0200539 */
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200540static void __update_rq_clock(struct rq *rq)
Ingo Molnar20d315d2007-07-09 18:51:58 +0200541{
542 u64 prev_raw = rq->prev_clock_raw;
543 u64 now = sched_clock();
544 s64 delta = now - prev_raw;
545 u64 clock = rq->clock;
546
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200547#ifdef CONFIG_SCHED_DEBUG
548 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
549#endif
Ingo Molnar20d315d2007-07-09 18:51:58 +0200550 /*
551 * Protect against sched_clock() occasionally going backwards:
552 */
553 if (unlikely(delta < 0)) {
554 clock++;
555 rq->clock_warps++;
556 } else {
557 /*
558 * Catch too large forward jumps too:
559 */
Ingo Molnar529c7722007-08-10 23:05:11 +0200560 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
561 if (clock < rq->tick_timestamp + TICK_NSEC)
562 clock = rq->tick_timestamp + TICK_NSEC;
563 else
564 clock++;
Ingo Molnar20d315d2007-07-09 18:51:58 +0200565 rq->clock_overflows++;
566 } else {
567 if (unlikely(delta > rq->clock_max_delta))
568 rq->clock_max_delta = delta;
569 clock += delta;
570 }
571 }
572
573 rq->prev_clock_raw = now;
574 rq->clock = clock;
Ingo Molnar20d315d2007-07-09 18:51:58 +0200575}
576
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200577static void update_rq_clock(struct rq *rq)
Ingo Molnar20d315d2007-07-09 18:51:58 +0200578{
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200579 if (likely(smp_processor_id() == cpu_of(rq)))
580 __update_rq_clock(rq);
581}
Ingo Molnar20d315d2007-07-09 18:51:58 +0200582
Ingo Molnar20d315d2007-07-09 18:51:58 +0200583/*
Nick Piggin674311d2005-06-25 14:57:27 -0700584 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700585 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700586 *
587 * The domain tree of any CPU may only be accessed from within
588 * preempt-disabled sections.
589 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700590#define for_each_domain(cpu, __sd) \
591 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592
593#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
594#define this_rq() (&__get_cpu_var(runqueues))
595#define task_rq(p) cpu_rq(task_cpu(p))
596#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
597
Peter Zijlstra48d5e252008-01-25 21:08:31 +0100598unsigned long rt_needs_cpu(int cpu)
599{
600 struct rq *rq = cpu_rq(cpu);
601 u64 delta;
602
603 if (!rq->rt_throttled)
604 return 0;
605
606 if (rq->clock > rq->rt_period_expire)
607 return 1;
608
609 delta = rq->rt_period_expire - rq->clock;
610 do_div(delta, NSEC_PER_SEC / HZ);
611
612 return (unsigned long)delta;
613}
614
Ingo Molnare436d802007-07-19 21:28:35 +0200615/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200616 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
617 */
618#ifdef CONFIG_SCHED_DEBUG
619# define const_debug __read_mostly
620#else
621# define const_debug static const
622#endif
623
624/*
625 * Debugging: various feature bits
626 */
627enum {
Ingo Molnarbbdba7c2007-10-15 17:00:06 +0200628 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
Ingo Molnar96126332007-11-15 20:57:40 +0100629 SCHED_FEAT_WAKEUP_PREEMPT = 2,
630 SCHED_FEAT_START_DEBIT = 4,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100631 SCHED_FEAT_TREE_AVG = 8,
632 SCHED_FEAT_APPROX_AVG = 16,
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100633 SCHED_FEAT_HRTICK = 32,
634 SCHED_FEAT_DOUBLE_TICK = 64,
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200635};
636
637const_debug unsigned int sysctl_sched_features =
Ingo Molnar8401f772007-10-18 21:32:55 +0200638 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
Ingo Molnar96126332007-11-15 20:57:40 +0100639 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
Ingo Molnar8401f772007-10-18 21:32:55 +0200640 SCHED_FEAT_START_DEBIT * 1 |
641 SCHED_FEAT_TREE_AVG * 0 |
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100642 SCHED_FEAT_APPROX_AVG * 0 |
643 SCHED_FEAT_HRTICK * 1 |
644 SCHED_FEAT_DOUBLE_TICK * 0;
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200645
646#define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
647
648/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100649 * Number of tasks to iterate in a single balance run.
650 * Limited because this is done with IRQs disabled.
651 */
652const_debug unsigned int sysctl_sched_nr_migrate = 32;
653
654/*
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100655 * period over which we measure -rt task cpu usage in ms.
656 * default: 1s
657 */
658const_debug unsigned int sysctl_sched_rt_period = 1000;
659
660#define SCHED_RT_FRAC_SHIFT 16
661#define SCHED_RT_FRAC (1UL << SCHED_RT_FRAC_SHIFT)
662
663/*
664 * ratio of time -rt tasks may consume.
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100665 * default: 95%
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100666 */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100667const_debug unsigned int sysctl_sched_rt_ratio = 62259;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100668
669/*
Ingo Molnare436d802007-07-19 21:28:35 +0200670 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
671 * clock constructed from sched_clock():
672 */
673unsigned long long cpu_clock(int cpu)
674{
Ingo Molnare436d802007-07-19 21:28:35 +0200675 unsigned long long now;
676 unsigned long flags;
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200677 struct rq *rq;
Ingo Molnare436d802007-07-19 21:28:35 +0200678
Ingo Molnar2cd4d0e2007-07-26 13:40:43 +0200679 local_irq_save(flags);
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200680 rq = cpu_rq(cpu);
Ingo Molnar8ced5f62007-12-07 19:02:47 +0100681 /*
682 * Only call sched_clock() if the scheduler has already been
683 * initialized (some code might call cpu_clock() very early):
684 */
685 if (rq->idle)
686 update_rq_clock(rq);
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200687 now = rq->clock;
Ingo Molnar2cd4d0e2007-07-26 13:40:43 +0200688 local_irq_restore(flags);
Ingo Molnare436d802007-07-19 21:28:35 +0200689
690 return now;
691}
Paul E. McKenneya58f6f22007-10-15 17:00:14 +0200692EXPORT_SYMBOL_GPL(cpu_clock);
Ingo Molnare436d802007-07-19 21:28:35 +0200693
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700695# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700697#ifndef finish_arch_switch
698# define finish_arch_switch(prev) do { } while (0)
699#endif
700
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100701static inline int task_current(struct rq *rq, struct task_struct *p)
702{
703 return rq->curr == p;
704}
705
Nick Piggin4866cde2005-06-25 14:57:23 -0700706#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700707static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700708{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100709 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700710}
711
Ingo Molnar70b97a72006-07-03 00:25:42 -0700712static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700713{
714}
715
Ingo Molnar70b97a72006-07-03 00:25:42 -0700716static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700717{
Ingo Molnarda04c032005-09-13 11:17:59 +0200718#ifdef CONFIG_DEBUG_SPINLOCK
719 /* this is a valid case when another task releases the spinlock */
720 rq->lock.owner = current;
721#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700722 /*
723 * If we are tracking spinlock dependencies then we have to
724 * fix up the runqueue lock - which gets 'carried over' from
725 * prev into current:
726 */
727 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
728
Nick Piggin4866cde2005-06-25 14:57:23 -0700729 spin_unlock_irq(&rq->lock);
730}
731
732#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700733static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700734{
735#ifdef CONFIG_SMP
736 return p->oncpu;
737#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100738 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700739#endif
740}
741
Ingo Molnar70b97a72006-07-03 00:25:42 -0700742static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700743{
744#ifdef CONFIG_SMP
745 /*
746 * We can optimise this out completely for !SMP, because the
747 * SMP rebalancing from interrupt is the only thing that cares
748 * here.
749 */
750 next->oncpu = 1;
751#endif
752#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
753 spin_unlock_irq(&rq->lock);
754#else
755 spin_unlock(&rq->lock);
756#endif
757}
758
Ingo Molnar70b97a72006-07-03 00:25:42 -0700759static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700760{
761#ifdef CONFIG_SMP
762 /*
763 * After ->oncpu is cleared, the task can be moved to a different CPU.
764 * We must ensure this doesn't happen until the switch is completely
765 * finished.
766 */
767 smp_wmb();
768 prev->oncpu = 0;
769#endif
770#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
771 local_irq_enable();
772#endif
773}
774#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775
776/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700777 * __task_rq_lock - lock the runqueue a given task resides on.
778 * Must be called interrupts disabled.
779 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700780static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700781 __acquires(rq->lock)
782{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200783 for (;;) {
784 struct rq *rq = task_rq(p);
785 spin_lock(&rq->lock);
786 if (likely(rq == task_rq(p)))
787 return rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700788 spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700789 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700790}
791
792/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100794 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 * explicitly disabling preemption.
796 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700797static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 __acquires(rq->lock)
799{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700800 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801
Andi Kleen3a5c3592007-10-15 17:00:14 +0200802 for (;;) {
803 local_irq_save(*flags);
804 rq = task_rq(p);
805 spin_lock(&rq->lock);
806 if (likely(rq == task_rq(p)))
807 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810}
811
Alexey Dobriyana9957442007-10-15 17:00:13 +0200812static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700813 __releases(rq->lock)
814{
815 spin_unlock(&rq->lock);
816}
817
Ingo Molnar70b97a72006-07-03 00:25:42 -0700818static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 __releases(rq->lock)
820{
821 spin_unlock_irqrestore(&rq->lock, *flags);
822}
823
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -0800825 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 */
Alexey Dobriyana9957442007-10-15 17:00:13 +0200827static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 __acquires(rq->lock)
829{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700830 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831
832 local_irq_disable();
833 rq = this_rq();
834 spin_lock(&rq->lock);
835
836 return rq;
837}
838
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200839/*
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200840 * We are going deep-idle (irqs are disabled):
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200841 */
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200842void sched_clock_idle_sleep_event(void)
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200843{
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200844 struct rq *rq = cpu_rq(smp_processor_id());
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200845
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200846 spin_lock(&rq->lock);
847 __update_rq_clock(rq);
848 spin_unlock(&rq->lock);
849 rq->clock_deep_idle_events++;
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200850}
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200851EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
852
853/*
854 * We just idled delta nanoseconds (called with irqs disabled):
855 */
856void sched_clock_idle_wakeup_event(u64 delta_ns)
857{
858 struct rq *rq = cpu_rq(smp_processor_id());
859 u64 now = sched_clock();
860
861 rq->idle_clock += delta_ns;
862 /*
863 * Override the previous timestamp and ignore all
864 * sched_clock() deltas that occured while we idled,
865 * and use the PM-provided delta_ns to advance the
866 * rq clock:
867 */
868 spin_lock(&rq->lock);
869 rq->prev_clock_raw = now;
870 rq->clock += delta_ns;
871 spin_unlock(&rq->lock);
Guillaume Chazarain782daee2008-01-25 21:08:33 +0100872 touch_softlockup_watchdog();
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200873}
874EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200875
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100876static void __resched_task(struct task_struct *p, int tif_bit);
877
878static inline void resched_task(struct task_struct *p)
879{
880 __resched_task(p, TIF_NEED_RESCHED);
881}
882
883#ifdef CONFIG_SCHED_HRTICK
884/*
885 * Use HR-timers to deliver accurate preemption points.
886 *
887 * Its all a bit involved since we cannot program an hrt while holding the
888 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
889 * reschedule event.
890 *
891 * When we get rescheduled we reprogram the hrtick_timer outside of the
892 * rq->lock.
893 */
894static inline void resched_hrt(struct task_struct *p)
895{
896 __resched_task(p, TIF_HRTICK_RESCHED);
897}
898
899static inline void resched_rq(struct rq *rq)
900{
901 unsigned long flags;
902
903 spin_lock_irqsave(&rq->lock, flags);
904 resched_task(rq->curr);
905 spin_unlock_irqrestore(&rq->lock, flags);
906}
907
908enum {
909 HRTICK_SET, /* re-programm hrtick_timer */
910 HRTICK_RESET, /* not a new slice */
911};
912
913/*
914 * Use hrtick when:
915 * - enabled by features
916 * - hrtimer is actually high res
917 */
918static inline int hrtick_enabled(struct rq *rq)
919{
920 if (!sched_feat(HRTICK))
921 return 0;
922 return hrtimer_is_hres_active(&rq->hrtick_timer);
923}
924
925/*
926 * Called to set the hrtick timer state.
927 *
928 * called with rq->lock held and irqs disabled
929 */
930static void hrtick_start(struct rq *rq, u64 delay, int reset)
931{
932 assert_spin_locked(&rq->lock);
933
934 /*
935 * preempt at: now + delay
936 */
937 rq->hrtick_expire =
938 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
939 /*
940 * indicate we need to program the timer
941 */
942 __set_bit(HRTICK_SET, &rq->hrtick_flags);
943 if (reset)
944 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
945
946 /*
947 * New slices are called from the schedule path and don't need a
948 * forced reschedule.
949 */
950 if (reset)
951 resched_hrt(rq->curr);
952}
953
954static void hrtick_clear(struct rq *rq)
955{
956 if (hrtimer_active(&rq->hrtick_timer))
957 hrtimer_cancel(&rq->hrtick_timer);
958}
959
960/*
961 * Update the timer from the possible pending state.
962 */
963static void hrtick_set(struct rq *rq)
964{
965 ktime_t time;
966 int set, reset;
967 unsigned long flags;
968
969 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
970
971 spin_lock_irqsave(&rq->lock, flags);
972 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
973 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
974 time = rq->hrtick_expire;
975 clear_thread_flag(TIF_HRTICK_RESCHED);
976 spin_unlock_irqrestore(&rq->lock, flags);
977
978 if (set) {
979 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
980 if (reset && !hrtimer_active(&rq->hrtick_timer))
981 resched_rq(rq);
982 } else
983 hrtick_clear(rq);
984}
985
986/*
987 * High-resolution timer tick.
988 * Runs from hardirq context with interrupts disabled.
989 */
990static enum hrtimer_restart hrtick(struct hrtimer *timer)
991{
992 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
993
994 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
995
996 spin_lock(&rq->lock);
997 __update_rq_clock(rq);
998 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
999 spin_unlock(&rq->lock);
1000
1001 return HRTIMER_NORESTART;
1002}
1003
1004static inline void init_rq_hrtick(struct rq *rq)
1005{
1006 rq->hrtick_flags = 0;
1007 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1008 rq->hrtick_timer.function = hrtick;
1009 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1010}
1011
1012void hrtick_resched(void)
1013{
1014 struct rq *rq;
1015 unsigned long flags;
1016
1017 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1018 return;
1019
1020 local_irq_save(flags);
1021 rq = cpu_rq(smp_processor_id());
1022 hrtick_set(rq);
1023 local_irq_restore(flags);
1024}
1025#else
1026static inline void hrtick_clear(struct rq *rq)
1027{
1028}
1029
1030static inline void hrtick_set(struct rq *rq)
1031{
1032}
1033
1034static inline void init_rq_hrtick(struct rq *rq)
1035{
1036}
1037
1038void hrtick_resched(void)
1039{
1040}
1041#endif
1042
Ingo Molnar1b9f19c2007-07-09 18:51:59 +02001043/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001044 * resched_task - mark a task 'to be rescheduled now'.
1045 *
1046 * On UP this means the setting of the need_resched flag, on SMP it
1047 * might also involve a cross-CPU call to trigger the scheduler on
1048 * the target CPU.
1049 */
1050#ifdef CONFIG_SMP
1051
1052#ifndef tsk_is_polling
1053#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1054#endif
1055
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001056static void __resched_task(struct task_struct *p, int tif_bit)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001057{
1058 int cpu;
1059
1060 assert_spin_locked(&task_rq(p)->lock);
1061
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001062 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001063 return;
1064
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001065 set_tsk_thread_flag(p, tif_bit);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001066
1067 cpu = task_cpu(p);
1068 if (cpu == smp_processor_id())
1069 return;
1070
1071 /* NEED_RESCHED must be visible before we test polling */
1072 smp_mb();
1073 if (!tsk_is_polling(p))
1074 smp_send_reschedule(cpu);
1075}
1076
1077static void resched_cpu(int cpu)
1078{
1079 struct rq *rq = cpu_rq(cpu);
1080 unsigned long flags;
1081
1082 if (!spin_trylock_irqsave(&rq->lock, flags))
1083 return;
1084 resched_task(cpu_curr(cpu));
1085 spin_unlock_irqrestore(&rq->lock, flags);
1086}
1087#else
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001088static void __resched_task(struct task_struct *p, int tif_bit)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001089{
1090 assert_spin_locked(&task_rq(p)->lock);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001091 set_tsk_thread_flag(p, tif_bit);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001092}
1093#endif
1094
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001095#if BITS_PER_LONG == 32
1096# define WMULT_CONST (~0UL)
1097#else
1098# define WMULT_CONST (1UL << 32)
1099#endif
1100
1101#define WMULT_SHIFT 32
1102
Ingo Molnar194081e2007-08-09 11:16:51 +02001103/*
1104 * Shift right and round:
1105 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001106#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +02001107
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +02001108static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001109calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1110 struct load_weight *lw)
1111{
1112 u64 tmp;
1113
1114 if (unlikely(!lw->inv_weight))
Ingo Molnar194081e2007-08-09 11:16:51 +02001115 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001116
1117 tmp = (u64)delta_exec * weight;
1118 /*
1119 * Check whether we'd overflow the 64-bit multiplication:
1120 */
Ingo Molnar194081e2007-08-09 11:16:51 +02001121 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001122 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +02001123 WMULT_SHIFT/2);
1124 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001125 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001126
Ingo Molnarecf691d2007-08-02 17:41:40 +02001127 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001128}
1129
1130static inline unsigned long
1131calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1132{
1133 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1134}
1135
Ingo Molnar10919852007-10-15 17:00:04 +02001136static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001137{
1138 lw->weight += inc;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001139}
1140
Ingo Molnar10919852007-10-15 17:00:04 +02001141static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001142{
1143 lw->weight -= dec;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001144}
1145
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001147 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1148 * of tasks with abnormal "nice" values across CPUs the contribution that
1149 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001150 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -07001151 * scaled version of the new time slice allocation that they receive on time
1152 * slice expiry etc.
1153 */
1154
Ingo Molnardd41f592007-07-09 18:51:59 +02001155#define WEIGHT_IDLEPRIO 2
1156#define WMULT_IDLEPRIO (1 << 31)
1157
1158/*
1159 * Nice levels are multiplicative, with a gentle 10% change for every
1160 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1161 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1162 * that remained on nice 0.
1163 *
1164 * The "10% effect" is relative and cumulative: from _any_ nice level,
1165 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +02001166 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1167 * If a task goes up by ~10% and another task goes down by ~10% then
1168 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +02001169 */
1170static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001171 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1172 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1173 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1174 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1175 /* 0 */ 1024, 820, 655, 526, 423,
1176 /* 5 */ 335, 272, 215, 172, 137,
1177 /* 10 */ 110, 87, 70, 56, 45,
1178 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +02001179};
1180
Ingo Molnar5714d2d2007-07-16 09:46:31 +02001181/*
1182 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1183 *
1184 * In cases where the weight does not change often, we can use the
1185 * precalculated inverse to speed up arithmetics by turning divisions
1186 * into multiplications:
1187 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001188static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001189 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1190 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1191 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1192 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1193 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1194 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1195 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1196 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +02001197};
Peter Williams2dd73a42006-06-27 02:54:34 -07001198
Ingo Molnardd41f592007-07-09 18:51:59 +02001199static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1200
1201/*
1202 * runqueue iterator, to support SMP load-balancing between different
1203 * scheduling classes, without having to expose their internal data
1204 * structures to the load-balancing proper:
1205 */
1206struct rq_iterator {
1207 void *arg;
1208 struct task_struct *(*start)(void *);
1209 struct task_struct *(*next)(void *);
1210};
1211
Peter Williamse1d14842007-10-24 18:23:51 +02001212#ifdef CONFIG_SMP
1213static unsigned long
1214balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1215 unsigned long max_load_move, struct sched_domain *sd,
1216 enum cpu_idle_type idle, int *all_pinned,
1217 int *this_best_prio, struct rq_iterator *iterator);
1218
1219static int
1220iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1221 struct sched_domain *sd, enum cpu_idle_type idle,
1222 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +02001223#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001224
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001225#ifdef CONFIG_CGROUP_CPUACCT
1226static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1227#else
1228static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1229#endif
1230
Srivatsa Vaddagiri58e2d4c2008-01-25 21:08:00 +01001231static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1232{
1233 update_load_add(&rq->load, load);
1234}
1235
1236static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1237{
1238 update_load_sub(&rq->load, load);
1239}
1240
Gregory Haskinse7693a32008-01-25 21:08:09 +01001241#ifdef CONFIG_SMP
1242static unsigned long source_load(int cpu, int type);
1243static unsigned long target_load(int cpu, int type);
1244static unsigned long cpu_avg_load_per_task(int cpu);
1245static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1246#endif /* CONFIG_SMP */
1247
Ingo Molnardd41f592007-07-09 18:51:59 +02001248#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +02001249#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001250#include "sched_fair.c"
1251#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +02001252#ifdef CONFIG_SCHED_DEBUG
1253# include "sched_debug.c"
1254#endif
1255
1256#define sched_class_highest (&rt_sched_class)
1257
Gerald Stralko5aff0532008-01-31 22:45:23 +01001258static void inc_nr_running(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001259{
1260 rq->nr_running++;
Ingo Molnar9c217242007-08-02 17:41:40 +02001261}
1262
Gerald Stralko5aff0532008-01-31 22:45:23 +01001263static void dec_nr_running(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001264{
1265 rq->nr_running--;
Ingo Molnar9c217242007-08-02 17:41:40 +02001266}
1267
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001268static void set_load_weight(struct task_struct *p)
1269{
1270 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001271 p->se.load.weight = prio_to_weight[0] * 2;
1272 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1273 return;
1274 }
1275
1276 /*
1277 * SCHED_IDLE tasks get minimal weight:
1278 */
1279 if (p->policy == SCHED_IDLE) {
1280 p->se.load.weight = WEIGHT_IDLEPRIO;
1281 p->se.load.inv_weight = WMULT_IDLEPRIO;
1282 return;
1283 }
1284
1285 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1286 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001287}
1288
Ingo Molnar8159f872007-08-09 11:16:49 +02001289static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001290{
1291 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001292 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001293 p->se.on_rq = 1;
1294}
1295
Ingo Molnar69be72c2007-08-09 11:16:49 +02001296static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001297{
Ingo Molnarf02231e2007-08-09 11:16:48 +02001298 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001299 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001300}
1301
1302/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001303 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001304 */
Ingo Molnar14531182007-07-09 18:51:59 +02001305static inline int __normal_prio(struct task_struct *p)
1306{
Ingo Molnardd41f592007-07-09 18:51:59 +02001307 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001308}
1309
1310/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001311 * Calculate the expected normal priority: i.e. priority
1312 * without taking RT-inheritance into account. Might be
1313 * boosted by interactivity modifiers. Changes upon fork,
1314 * setprio syscalls, and whenever the interactivity
1315 * estimator recalculates.
1316 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001317static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001318{
1319 int prio;
1320
Ingo Molnare05606d2007-07-09 18:51:59 +02001321 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001322 prio = MAX_RT_PRIO-1 - p->rt_priority;
1323 else
1324 prio = __normal_prio(p);
1325 return prio;
1326}
1327
1328/*
1329 * Calculate the current priority, i.e. the priority
1330 * taken into account by the scheduler. This value might
1331 * be boosted by RT tasks, or might be boosted by
1332 * interactivity modifiers. Will be RT if the task got
1333 * RT-boosted. If not then it returns p->normal_prio.
1334 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001335static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001336{
1337 p->normal_prio = normal_prio(p);
1338 /*
1339 * If we are RT tasks or we were boosted to RT priority,
1340 * keep the priority unchanged. Otherwise, update priority
1341 * to the normal priority:
1342 */
1343 if (!rt_prio(p->prio))
1344 return p->normal_prio;
1345 return p->prio;
1346}
1347
1348/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001349 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001351static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001353 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001354 rq->nr_uninterruptible--;
1355
Ingo Molnar8159f872007-08-09 11:16:49 +02001356 enqueue_task(rq, p, wakeup);
Gerald Stralko5aff0532008-01-31 22:45:23 +01001357 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358}
1359
1360/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 * deactivate_task - remove a task from the runqueue.
1362 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001363static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001365 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001366 rq->nr_uninterruptible++;
1367
Ingo Molnar69be72c2007-08-09 11:16:49 +02001368 dequeue_task(rq, p, sleep);
Gerald Stralko5aff0532008-01-31 22:45:23 +01001369 dec_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370}
1371
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372/**
1373 * task_curr - is this task currently executing on a CPU?
1374 * @p: the task in question.
1375 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001376inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377{
1378 return cpu_curr(task_cpu(p)) == p;
1379}
1380
Peter Williams2dd73a42006-06-27 02:54:34 -07001381/* Used instead of source_load when we know the type == 0 */
1382unsigned long weighted_cpuload(const int cpu)
1383{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02001384 return cpu_rq(cpu)->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02001385}
1386
1387static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1388{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001389 set_task_rq(p, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001390#ifdef CONFIG_SMP
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001391 /*
1392 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1393 * successfuly executed on another CPU. We must ensure that updates of
1394 * per-task data have been completed by this moment.
1395 */
1396 smp_wmb();
Ingo Molnardd41f592007-07-09 18:51:59 +02001397 task_thread_info(p)->cpu = cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02001398#endif
Peter Williams2dd73a42006-06-27 02:54:34 -07001399}
1400
Steven Rostedtcb469842008-01-25 21:08:22 +01001401static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1402 const struct sched_class *prev_class,
1403 int oldprio, int running)
1404{
1405 if (prev_class != p->sched_class) {
1406 if (prev_class->switched_from)
1407 prev_class->switched_from(rq, p, running);
1408 p->sched_class->switched_to(rq, p, running);
1409 } else
1410 p->sched_class->prio_changed(rq, p, oldprio, running);
1411}
1412
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413#ifdef CONFIG_SMP
Ingo Molnarc65cc872007-07-09 18:51:58 +02001414
Ingo Molnarcc367732007-10-15 17:00:18 +02001415/*
1416 * Is this task likely cache-hot:
1417 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01001418static int
Ingo Molnarcc367732007-10-15 17:00:18 +02001419task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1420{
1421 s64 delta;
1422
1423 if (p->sched_class != &fair_sched_class)
1424 return 0;
1425
Ingo Molnar6bc16652007-10-15 17:00:18 +02001426 if (sysctl_sched_migration_cost == -1)
1427 return 1;
1428 if (sysctl_sched_migration_cost == 0)
1429 return 0;
1430
Ingo Molnarcc367732007-10-15 17:00:18 +02001431 delta = now - p->se.exec_start;
1432
1433 return delta < (s64)sysctl_sched_migration_cost;
1434}
1435
1436
Ingo Molnardd41f592007-07-09 18:51:59 +02001437void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02001438{
Ingo Molnardd41f592007-07-09 18:51:59 +02001439 int old_cpu = task_cpu(p);
1440 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001441 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1442 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnarbbdba7c2007-10-15 17:00:06 +02001443 u64 clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001444
1445 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001446
1447#ifdef CONFIG_SCHEDSTATS
1448 if (p->se.wait_start)
1449 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001450 if (p->se.sleep_start)
1451 p->se.sleep_start -= clock_offset;
1452 if (p->se.block_start)
1453 p->se.block_start -= clock_offset;
Ingo Molnarcc367732007-10-15 17:00:18 +02001454 if (old_cpu != new_cpu) {
1455 schedstat_inc(p, se.nr_migrations);
1456 if (task_hot(p, old_rq->clock, NULL))
1457 schedstat_inc(p, se.nr_forced2_migrations);
1458 }
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001459#endif
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001460 p->se.vruntime -= old_cfsrq->min_vruntime -
1461 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02001462
1463 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02001464}
1465
Ingo Molnar70b97a72006-07-03 00:25:42 -07001466struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468
Ingo Molnar36c8b582006-07-03 00:25:41 -07001469 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 int dest_cpu;
1471
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001473};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474
1475/*
1476 * The task's runqueue lock must be held.
1477 * Returns true if you have to wait for migration thread.
1478 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001479static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07001480migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001482 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483
1484 /*
1485 * If the task is not on a runqueue (and not running), then
1486 * it is sufficient to simply update the task's cpu field.
1487 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001488 if (!p->se.on_rq && !task_running(rq, p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 set_task_cpu(p, dest_cpu);
1490 return 0;
1491 }
1492
1493 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 req->task = p;
1495 req->dest_cpu = dest_cpu;
1496 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07001497
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498 return 1;
1499}
1500
1501/*
1502 * wait_task_inactive - wait for a thread to unschedule.
1503 *
1504 * The caller must ensure that the task *will* unschedule sometime soon,
1505 * else this function might spin for a *long* time. This function can't
1506 * be called with interrupts off, or it may introduce deadlock with
1507 * smp_call_function() if an IPI is sent by the same process we are
1508 * waiting to become inactive.
1509 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001510void wait_task_inactive(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511{
1512 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02001513 int running, on_rq;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001514 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515
Andi Kleen3a5c3592007-10-15 17:00:14 +02001516 for (;;) {
1517 /*
1518 * We do the initial early heuristics without holding
1519 * any task-queue locks at all. We'll only try to get
1520 * the runqueue lock when things look like they will
1521 * work out!
1522 */
1523 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001524
Andi Kleen3a5c3592007-10-15 17:00:14 +02001525 /*
1526 * If the task is actively running on another CPU
1527 * still, just relax and busy-wait without holding
1528 * any locks.
1529 *
1530 * NOTE! Since we don't hold any locks, it's not
1531 * even sure that "rq" stays as the right runqueue!
1532 * But we don't care, since "task_running()" will
1533 * return false if the runqueue has changed and p
1534 * is actually now running somewhere else!
1535 */
1536 while (task_running(rq, p))
1537 cpu_relax();
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001538
Andi Kleen3a5c3592007-10-15 17:00:14 +02001539 /*
1540 * Ok, time to look more closely! We need the rq
1541 * lock now, to be *sure*. If we're wrong, we'll
1542 * just go back and repeat.
1543 */
1544 rq = task_rq_lock(p, &flags);
1545 running = task_running(rq, p);
1546 on_rq = p->se.on_rq;
1547 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001548
Andi Kleen3a5c3592007-10-15 17:00:14 +02001549 /*
1550 * Was it really running after all now that we
1551 * checked with the proper locks actually held?
1552 *
1553 * Oops. Go back and try again..
1554 */
1555 if (unlikely(running)) {
1556 cpu_relax();
1557 continue;
1558 }
1559
1560 /*
1561 * It's not enough that it's not actively running,
1562 * it must be off the runqueue _entirely_, and not
1563 * preempted!
1564 *
1565 * So if it wa still runnable (but just not actively
1566 * running right now), it's preempted, and we should
1567 * yield - it could be a while.
1568 */
1569 if (unlikely(on_rq)) {
1570 schedule_timeout_uninterruptible(1);
1571 continue;
1572 }
1573
1574 /*
1575 * Ahh, all good. It wasn't running, and it wasn't
1576 * runnable, which means that it will never become
1577 * running in the future either. We're all done!
1578 */
1579 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581}
1582
1583/***
1584 * kick_process - kick a running thread to enter/exit the kernel
1585 * @p: the to-be-kicked thread
1586 *
1587 * Cause a process which is running on another CPU to enter
1588 * kernel-mode, without any delay. (to get signals handled.)
1589 *
1590 * NOTE: this function doesnt have to take the runqueue lock,
1591 * because all it wants to ensure is that the remote task enters
1592 * the kernel. If the IPI races and the task has been migrated
1593 * to another CPU then no harm is done and the purpose has been
1594 * achieved as well.
1595 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001596void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597{
1598 int cpu;
1599
1600 preempt_disable();
1601 cpu = task_cpu(p);
1602 if ((cpu != smp_processor_id()) && task_curr(p))
1603 smp_send_reschedule(cpu);
1604 preempt_enable();
1605}
1606
1607/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001608 * Return a low guess at the load of a migration-source cpu weighted
1609 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 *
1611 * We want to under-estimate the load of migration sources, to
1612 * balance conservatively.
1613 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001614static unsigned long source_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08001615{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001616 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001617 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001618
Peter Williams2dd73a42006-06-27 02:54:34 -07001619 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02001620 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07001621
Ingo Molnardd41f592007-07-09 18:51:59 +02001622 return min(rq->cpu_load[type-1], total);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623}
1624
1625/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001626 * Return a high guess at the load of a migration-target cpu weighted
1627 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001629static unsigned long target_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08001630{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001631 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001632 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001633
Peter Williams2dd73a42006-06-27 02:54:34 -07001634 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02001635 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07001636
Ingo Molnardd41f592007-07-09 18:51:59 +02001637 return max(rq->cpu_load[type-1], total);
Peter Williams2dd73a42006-06-27 02:54:34 -07001638}
1639
1640/*
1641 * Return the average load per task on the cpu's run queue
1642 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01001643static unsigned long cpu_avg_load_per_task(int cpu)
Peter Williams2dd73a42006-06-27 02:54:34 -07001644{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001645 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001646 unsigned long total = weighted_cpuload(cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07001647 unsigned long n = rq->nr_running;
1648
Ingo Molnardd41f592007-07-09 18:51:59 +02001649 return n ? total / n : SCHED_LOAD_SCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650}
1651
Nick Piggin147cbb42005-06-25 14:57:19 -07001652/*
1653 * find_idlest_group finds and returns the least busy CPU group within the
1654 * domain.
1655 */
1656static struct sched_group *
1657find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1658{
1659 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1660 unsigned long min_load = ULONG_MAX, this_load = 0;
1661 int load_idx = sd->forkexec_idx;
1662 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1663
1664 do {
1665 unsigned long load, avg_load;
1666 int local_group;
1667 int i;
1668
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001669 /* Skip over this group if it has no CPUs allowed */
1670 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
Andi Kleen3a5c3592007-10-15 17:00:14 +02001671 continue;
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001672
Nick Piggin147cbb42005-06-25 14:57:19 -07001673 local_group = cpu_isset(this_cpu, group->cpumask);
Nick Piggin147cbb42005-06-25 14:57:19 -07001674
1675 /* Tally up the load of all CPUs in the group */
1676 avg_load = 0;
1677
1678 for_each_cpu_mask(i, group->cpumask) {
1679 /* Bias balancing toward cpus of our domain */
1680 if (local_group)
1681 load = source_load(i, load_idx);
1682 else
1683 load = target_load(i, load_idx);
1684
1685 avg_load += load;
1686 }
1687
1688 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07001689 avg_load = sg_div_cpu_power(group,
1690 avg_load * SCHED_LOAD_SCALE);
Nick Piggin147cbb42005-06-25 14:57:19 -07001691
1692 if (local_group) {
1693 this_load = avg_load;
1694 this = group;
1695 } else if (avg_load < min_load) {
1696 min_load = avg_load;
1697 idlest = group;
1698 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02001699 } while (group = group->next, group != sd->groups);
Nick Piggin147cbb42005-06-25 14:57:19 -07001700
1701 if (!idlest || 100*this_load < imbalance*min_load)
1702 return NULL;
1703 return idlest;
1704}
1705
1706/*
Satoru Takeuchi0feaece2006-10-03 01:14:10 -07001707 * find_idlest_cpu - find the idlest cpu among the cpus in group.
Nick Piggin147cbb42005-06-25 14:57:19 -07001708 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07001709static int
1710find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
Nick Piggin147cbb42005-06-25 14:57:19 -07001711{
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001712 cpumask_t tmp;
Nick Piggin147cbb42005-06-25 14:57:19 -07001713 unsigned long load, min_load = ULONG_MAX;
1714 int idlest = -1;
1715 int i;
1716
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001717 /* Traverse only the allowed CPUs */
1718 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1719
1720 for_each_cpu_mask(i, tmp) {
Peter Williams2dd73a42006-06-27 02:54:34 -07001721 load = weighted_cpuload(i);
Nick Piggin147cbb42005-06-25 14:57:19 -07001722
1723 if (load < min_load || (load == min_load && i == this_cpu)) {
1724 min_load = load;
1725 idlest = i;
1726 }
1727 }
1728
1729 return idlest;
1730}
1731
Nick Piggin476d1392005-06-25 14:57:29 -07001732/*
1733 * sched_balance_self: balance the current task (running on cpu) in domains
1734 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1735 * SD_BALANCE_EXEC.
1736 *
1737 * Balance, ie. select the least loaded group.
1738 *
1739 * Returns the target CPU number, or the same CPU if no balancing is needed.
1740 *
1741 * preempt must be disabled.
1742 */
1743static int sched_balance_self(int cpu, int flag)
1744{
1745 struct task_struct *t = current;
1746 struct sched_domain *tmp, *sd = NULL;
Nick Piggin147cbb42005-06-25 14:57:19 -07001747
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001748 for_each_domain(cpu, tmp) {
Ingo Molnar9761eea2007-07-09 18:52:00 +02001749 /*
1750 * If power savings logic is enabled for a domain, stop there.
1751 */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07001752 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1753 break;
Nick Piggin476d1392005-06-25 14:57:29 -07001754 if (tmp->flags & flag)
1755 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001756 }
Nick Piggin476d1392005-06-25 14:57:29 -07001757
1758 while (sd) {
1759 cpumask_t span;
1760 struct sched_group *group;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001761 int new_cpu, weight;
1762
1763 if (!(sd->flags & flag)) {
1764 sd = sd->child;
1765 continue;
1766 }
Nick Piggin476d1392005-06-25 14:57:29 -07001767
1768 span = sd->span;
1769 group = find_idlest_group(sd, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001770 if (!group) {
1771 sd = sd->child;
1772 continue;
1773 }
Nick Piggin476d1392005-06-25 14:57:29 -07001774
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001775 new_cpu = find_idlest_cpu(group, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001776 if (new_cpu == -1 || new_cpu == cpu) {
1777 /* Now try balancing at a lower domain level of cpu */
1778 sd = sd->child;
1779 continue;
1780 }
Nick Piggin476d1392005-06-25 14:57:29 -07001781
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001782 /* Now try balancing at a lower domain level of new_cpu */
Nick Piggin476d1392005-06-25 14:57:29 -07001783 cpu = new_cpu;
Nick Piggin476d1392005-06-25 14:57:29 -07001784 sd = NULL;
1785 weight = cpus_weight(span);
1786 for_each_domain(cpu, tmp) {
1787 if (weight <= cpus_weight(tmp->span))
1788 break;
1789 if (tmp->flags & flag)
1790 sd = tmp;
1791 }
1792 /* while loop will break here if sd == NULL */
1793 }
1794
1795 return cpu;
1796}
1797
1798#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800/***
1801 * try_to_wake_up - wake up a thread
1802 * @p: the to-be-woken-up thread
1803 * @state: the mask of task states that can be woken
1804 * @sync: do a synchronous wakeup?
1805 *
1806 * Put it on the run-queue if it's not already there. The "current"
1807 * thread is always on the run-queue (except when the actual
1808 * re-schedule is in progress), and as such you're allowed to do
1809 * the simpler "current->state = TASK_RUNNING" to mark yourself
1810 * runnable without the overhead of this.
1811 *
1812 * returns failure only if the task is already active.
1813 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001814static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815{
Ingo Molnarcc367732007-10-15 17:00:18 +02001816 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 unsigned long flags;
1818 long old_state;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001819 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820
1821 rq = task_rq_lock(p, &flags);
1822 old_state = p->state;
1823 if (!(old_state & state))
1824 goto out;
1825
Ingo Molnardd41f592007-07-09 18:51:59 +02001826 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 goto out_running;
1828
1829 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02001830 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 this_cpu = smp_processor_id();
1832
1833#ifdef CONFIG_SMP
1834 if (unlikely(task_running(rq, p)))
1835 goto out_activate;
1836
Dmitry Adamushko5d2f5a62008-01-25 21:08:21 +01001837 cpu = p->sched_class->select_task_rq(p, sync);
1838 if (cpu != orig_cpu) {
1839 set_task_cpu(p, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840 task_rq_unlock(rq, &flags);
1841 /* might preempt at this point */
1842 rq = task_rq_lock(p, &flags);
1843 old_state = p->state;
1844 if (!(old_state & state))
1845 goto out;
Ingo Molnardd41f592007-07-09 18:51:59 +02001846 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 goto out_running;
1848
1849 this_cpu = smp_processor_id();
1850 cpu = task_cpu(p);
1851 }
1852
Gregory Haskinse7693a32008-01-25 21:08:09 +01001853#ifdef CONFIG_SCHEDSTATS
1854 schedstat_inc(rq, ttwu_count);
1855 if (cpu == this_cpu)
1856 schedstat_inc(rq, ttwu_local);
1857 else {
1858 struct sched_domain *sd;
1859 for_each_domain(this_cpu, sd) {
1860 if (cpu_isset(cpu, sd->span)) {
1861 schedstat_inc(sd, ttwu_wake_remote);
1862 break;
1863 }
1864 }
1865 }
Gregory Haskinse7693a32008-01-25 21:08:09 +01001866#endif
1867
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868out_activate:
1869#endif /* CONFIG_SMP */
Ingo Molnarcc367732007-10-15 17:00:18 +02001870 schedstat_inc(p, se.nr_wakeups);
1871 if (sync)
1872 schedstat_inc(p, se.nr_wakeups_sync);
1873 if (orig_cpu != cpu)
1874 schedstat_inc(p, se.nr_wakeups_migrate);
1875 if (cpu == this_cpu)
1876 schedstat_inc(p, se.nr_wakeups_local);
1877 else
1878 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnar2daa3572007-08-09 11:16:51 +02001879 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02001880 activate_task(rq, p, 1);
Ingo Molnar9c63d9c2007-10-15 17:00:20 +02001881 check_preempt_curr(rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 success = 1;
1883
1884out_running:
1885 p->state = TASK_RUNNING;
Steven Rostedt9a897c52008-01-25 21:08:22 +01001886#ifdef CONFIG_SMP
1887 if (p->sched_class->task_wake_up)
1888 p->sched_class->task_wake_up(rq, p);
1889#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890out:
1891 task_rq_unlock(rq, &flags);
1892
1893 return success;
1894}
1895
Ingo Molnar36c8b582006-07-03 00:25:41 -07001896int fastcall wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001898 return try_to_wake_up(p, TASK_ALL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900EXPORT_SYMBOL(wake_up_process);
1901
Ingo Molnar36c8b582006-07-03 00:25:41 -07001902int fastcall wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903{
1904 return try_to_wake_up(p, state, 0);
1905}
1906
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907/*
1908 * Perform scheduler related setup for a newly forked process p.
1909 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02001910 *
1911 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001913static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914{
Ingo Molnardd41f592007-07-09 18:51:59 +02001915 p->se.exec_start = 0;
1916 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02001917 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001918
1919#ifdef CONFIG_SCHEDSTATS
1920 p->se.wait_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02001921 p->se.sum_sleep_runtime = 0;
1922 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02001923 p->se.block_start = 0;
1924 p->se.sleep_max = 0;
1925 p->se.block_max = 0;
1926 p->se.exec_max = 0;
Ingo Molnareba1ed42007-10-15 17:00:02 +02001927 p->se.slice_max = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02001928 p->se.wait_max = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001929#endif
Nick Piggin476d1392005-06-25 14:57:29 -07001930
Peter Zijlstrafa717062008-01-25 21:08:27 +01001931 INIT_LIST_HEAD(&p->rt.run_list);
Ingo Molnardd41f592007-07-09 18:51:59 +02001932 p->se.on_rq = 0;
Nick Piggin476d1392005-06-25 14:57:29 -07001933
Avi Kivitye107be32007-07-26 13:40:43 +02001934#ifdef CONFIG_PREEMPT_NOTIFIERS
1935 INIT_HLIST_HEAD(&p->preempt_notifiers);
1936#endif
1937
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 /*
1939 * We mark the process as running here, but have not actually
1940 * inserted it onto the runqueue yet. This guarantees that
1941 * nobody will actually run it, and a signal or other external
1942 * event cannot wake it up and insert it on the runqueue either.
1943 */
1944 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02001945}
1946
1947/*
1948 * fork()/clone()-time setup:
1949 */
1950void sched_fork(struct task_struct *p, int clone_flags)
1951{
1952 int cpu = get_cpu();
1953
1954 __sched_fork(p);
1955
1956#ifdef CONFIG_SMP
1957 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1958#endif
Ingo Molnar02e4bac2007-10-15 17:00:11 +02001959 set_task_cpu(p, cpu);
Ingo Molnarb29739f2006-06-27 02:54:51 -07001960
1961 /*
1962 * Make sure we do not leak PI boosting priority to the child:
1963 */
1964 p->prio = current->normal_prio;
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02001965 if (!rt_prio(p->prio))
1966 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07001967
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07001968#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02001969 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07001970 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08001972#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07001973 p->oncpu = 0;
1974#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07001976 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08001977 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978#endif
Nick Piggin476d1392005-06-25 14:57:29 -07001979 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980}
1981
1982/*
1983 * wake_up_new_task - wake up a newly created task for the first time.
1984 *
1985 * This function will do some initial scheduler statistics housekeeping
1986 * that must be done for every newly created context, then puts the task
1987 * on the runqueue and wakes it.
1988 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001989void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990{
1991 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02001992 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993
1994 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02001996 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997
1998 p->prio = effective_prio(p);
1999
Srivatsa Vaddagirib9dca1e2007-10-17 16:55:11 +02002000 if (!p->sched_class->task_new || !current->se.on_rq) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002001 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02002004 * Let the scheduling class do new task startup
2005 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02002007 p->sched_class->task_new(rq, p);
Gerald Stralko5aff0532008-01-31 22:45:23 +01002008 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009 }
Ingo Molnardd41f592007-07-09 18:51:59 +02002010 check_preempt_curr(rq, p);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002011#ifdef CONFIG_SMP
2012 if (p->sched_class->task_wake_up)
2013 p->sched_class->task_wake_up(rq, p);
2014#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02002015 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016}
2017
Avi Kivitye107be32007-07-26 13:40:43 +02002018#ifdef CONFIG_PREEMPT_NOTIFIERS
2019
2020/**
Randy Dunlap421cee22007-07-31 00:37:50 -07002021 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2022 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002023 */
2024void preempt_notifier_register(struct preempt_notifier *notifier)
2025{
2026 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2027}
2028EXPORT_SYMBOL_GPL(preempt_notifier_register);
2029
2030/**
2031 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002032 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002033 *
2034 * This is safe to call from within a preemption notifier.
2035 */
2036void preempt_notifier_unregister(struct preempt_notifier *notifier)
2037{
2038 hlist_del(&notifier->link);
2039}
2040EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2041
2042static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2043{
2044 struct preempt_notifier *notifier;
2045 struct hlist_node *node;
2046
2047 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2048 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2049}
2050
2051static void
2052fire_sched_out_preempt_notifiers(struct task_struct *curr,
2053 struct task_struct *next)
2054{
2055 struct preempt_notifier *notifier;
2056 struct hlist_node *node;
2057
2058 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2059 notifier->ops->sched_out(notifier, next);
2060}
2061
2062#else
2063
2064static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2065{
2066}
2067
2068static void
2069fire_sched_out_preempt_notifiers(struct task_struct *curr,
2070 struct task_struct *next)
2071{
2072}
2073
2074#endif
2075
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002077 * prepare_task_switch - prepare to switch tasks
2078 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002079 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002080 * @next: the task we are going to switch to.
2081 *
2082 * This is called with the rq lock held and interrupts off. It must
2083 * be paired with a subsequent finish_task_switch after the context
2084 * switch.
2085 *
2086 * prepare_task_switch sets up locking and calls architecture specific
2087 * hooks.
2088 */
Avi Kivitye107be32007-07-26 13:40:43 +02002089static inline void
2090prepare_task_switch(struct rq *rq, struct task_struct *prev,
2091 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002092{
Avi Kivitye107be32007-07-26 13:40:43 +02002093 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002094 prepare_lock_switch(rq, next);
2095 prepare_arch_switch(next);
2096}
2097
2098/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04002100 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 * @prev: the thread we just switched away from.
2102 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002103 * finish_task_switch must be called after the context switch, paired
2104 * with a prepare_task_switch call before the context switch.
2105 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2106 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 *
2108 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002109 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 * with the lock held can cause deadlocks; see schedule() for
2111 * details.)
2112 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002113static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 __releases(rq->lock)
2115{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002117 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118
2119 rq->prev_mm = NULL;
2120
2121 /*
2122 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002123 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002124 * schedule one last time. The schedule call will never return, and
2125 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002126 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 * still held, otherwise prev could be scheduled on another cpu, die
2128 * there before we look at prev->state, and then the reference would
2129 * be dropped twice.
2130 * Manfred Spraul <manfred@colorfullife.com>
2131 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002132 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07002133 finish_arch_switch(prev);
2134 finish_lock_switch(rq, prev);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002135#ifdef CONFIG_SMP
2136 if (current->sched_class->post_schedule)
2137 current->sched_class->post_schedule(rq);
2138#endif
Steven Rostedte8fa1362008-01-25 21:08:05 +01002139
Avi Kivitye107be32007-07-26 13:40:43 +02002140 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 if (mm)
2142 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002143 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08002144 /*
2145 * Remove function-return probe instances associated with this
2146 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002147 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002148 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002150 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151}
2152
2153/**
2154 * schedule_tail - first thing a freshly forked thread must call.
2155 * @prev: the thread we just switched away from.
2156 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002157asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 __releases(rq->lock)
2159{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002160 struct rq *rq = this_rq();
2161
Nick Piggin4866cde2005-06-25 14:57:23 -07002162 finish_task_switch(rq, prev);
2163#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2164 /* In this case, finish_task_switch does not reenable preemption */
2165 preempt_enable();
2166#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002168 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169}
2170
2171/*
2172 * context_switch - switch to the new MM and the new
2173 * thread's register state.
2174 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002175static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002176context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002177 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178{
Ingo Molnardd41f592007-07-09 18:51:59 +02002179 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180
Avi Kivitye107be32007-07-26 13:40:43 +02002181 prepare_task_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02002182 mm = next->mm;
2183 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002184 /*
2185 * For paravirt, this is coupled with an exit in switch_to to
2186 * combine the page table reload and the switch backend into
2187 * one hypercall.
2188 */
2189 arch_enter_lazy_cpu_mode();
2190
Ingo Molnardd41f592007-07-09 18:51:59 +02002191 if (unlikely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192 next->active_mm = oldmm;
2193 atomic_inc(&oldmm->mm_count);
2194 enter_lazy_tlb(oldmm, next);
2195 } else
2196 switch_mm(oldmm, mm, next);
2197
Ingo Molnardd41f592007-07-09 18:51:59 +02002198 if (unlikely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 rq->prev_mm = oldmm;
2201 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002202 /*
2203 * Since the runqueue lock will be released by the next
2204 * task (which is an invalid locking op but in the case
2205 * of the scheduler it's an obvious special-case), so we
2206 * do an early lockdep release here:
2207 */
2208#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002209 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002210#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211
2212 /* Here we just switch the register state and the stack. */
2213 switch_to(prev, next, prev);
2214
Ingo Molnardd41f592007-07-09 18:51:59 +02002215 barrier();
2216 /*
2217 * this_rq must be evaluated again because prev may have moved
2218 * CPUs since it called schedule(), thus the 'rq' on its stack
2219 * frame will be invalid.
2220 */
2221 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222}
2223
2224/*
2225 * nr_running, nr_uninterruptible and nr_context_switches:
2226 *
2227 * externally visible scheduler statistics: current number of runnable
2228 * threads, current number of uninterruptible-sleeping threads, total
2229 * number of context switches performed since bootup.
2230 */
2231unsigned long nr_running(void)
2232{
2233 unsigned long i, sum = 0;
2234
2235 for_each_online_cpu(i)
2236 sum += cpu_rq(i)->nr_running;
2237
2238 return sum;
2239}
2240
2241unsigned long nr_uninterruptible(void)
2242{
2243 unsigned long i, sum = 0;
2244
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002245 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 sum += cpu_rq(i)->nr_uninterruptible;
2247
2248 /*
2249 * Since we read the counters lockless, it might be slightly
2250 * inaccurate. Do not allow it to go below zero though:
2251 */
2252 if (unlikely((long)sum < 0))
2253 sum = 0;
2254
2255 return sum;
2256}
2257
2258unsigned long long nr_context_switches(void)
2259{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002260 int i;
2261 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002263 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264 sum += cpu_rq(i)->nr_switches;
2265
2266 return sum;
2267}
2268
2269unsigned long nr_iowait(void)
2270{
2271 unsigned long i, sum = 0;
2272
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002273 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2275
2276 return sum;
2277}
2278
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002279unsigned long nr_active(void)
2280{
2281 unsigned long i, running = 0, uninterruptible = 0;
2282
2283 for_each_online_cpu(i) {
2284 running += cpu_rq(i)->nr_running;
2285 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2286 }
2287
2288 if (unlikely((long)uninterruptible < 0))
2289 uninterruptible = 0;
2290
2291 return running + uninterruptible;
2292}
2293
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294/*
Ingo Molnardd41f592007-07-09 18:51:59 +02002295 * Update rq->cpu_load[] statistics. This function is usually called every
2296 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07002297 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002298static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002299{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02002300 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02002301 int i, scale;
2302
2303 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02002304
2305 /* Update our load: */
2306 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2307 unsigned long old_load, new_load;
2308
2309 /* scale is effectively 1 << i now, and >> i divides by scale */
2310
2311 old_load = this_rq->cpu_load[i];
2312 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02002313 /*
2314 * Round up the averaging division if load is increasing. This
2315 * prevents us from getting stuck on 9 if the load is 10, for
2316 * example.
2317 */
2318 if (new_load > old_load)
2319 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02002320 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2321 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002322}
2323
Ingo Molnardd41f592007-07-09 18:51:59 +02002324#ifdef CONFIG_SMP
2325
Ingo Molnar48f24c42006-07-03 00:25:40 -07002326/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 * double_rq_lock - safely lock two runqueues
2328 *
2329 * Note this does not disable interrupts like task_rq_lock,
2330 * you need to do so manually before calling.
2331 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002332static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 __acquires(rq1->lock)
2334 __acquires(rq2->lock)
2335{
Kirill Korotaev054b9102006-12-10 02:20:11 -08002336 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 if (rq1 == rq2) {
2338 spin_lock(&rq1->lock);
2339 __acquire(rq2->lock); /* Fake it out ;) */
2340 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002341 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 spin_lock(&rq1->lock);
2343 spin_lock(&rq2->lock);
2344 } else {
2345 spin_lock(&rq2->lock);
2346 spin_lock(&rq1->lock);
2347 }
2348 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002349 update_rq_clock(rq1);
2350 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351}
2352
2353/*
2354 * double_rq_unlock - safely unlock two runqueues
2355 *
2356 * Note this does not restore interrupts like task_rq_unlock,
2357 * you need to do so manually after calling.
2358 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002359static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 __releases(rq1->lock)
2361 __releases(rq2->lock)
2362{
2363 spin_unlock(&rq1->lock);
2364 if (rq1 != rq2)
2365 spin_unlock(&rq2->lock);
2366 else
2367 __release(rq2->lock);
2368}
2369
2370/*
2371 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2372 */
Steven Rostedte8fa1362008-01-25 21:08:05 +01002373static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 __releases(this_rq->lock)
2375 __acquires(busiest->lock)
2376 __acquires(this_rq->lock)
2377{
Steven Rostedte8fa1362008-01-25 21:08:05 +01002378 int ret = 0;
2379
Kirill Korotaev054b9102006-12-10 02:20:11 -08002380 if (unlikely(!irqs_disabled())) {
2381 /* printk() doesn't work good under rq->lock */
2382 spin_unlock(&this_rq->lock);
2383 BUG_ON(1);
2384 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385 if (unlikely(!spin_trylock(&busiest->lock))) {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002386 if (busiest < this_rq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387 spin_unlock(&this_rq->lock);
2388 spin_lock(&busiest->lock);
2389 spin_lock(&this_rq->lock);
Steven Rostedte8fa1362008-01-25 21:08:05 +01002390 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 } else
2392 spin_lock(&busiest->lock);
2393 }
Steven Rostedte8fa1362008-01-25 21:08:05 +01002394 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395}
2396
2397/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 * If dest_cpu is allowed for this process, migrate the task to it.
2399 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002400 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 * the cpu_allowed mask is restored.
2402 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002403static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002405 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002407 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408
2409 rq = task_rq_lock(p, &flags);
2410 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2411 || unlikely(cpu_is_offline(dest_cpu)))
2412 goto out;
2413
2414 /* force the process onto the specified CPU */
2415 if (migrate_task(p, dest_cpu, &req)) {
2416 /* Need to wait for migration thread (might exit: take ref). */
2417 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07002418
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 get_task_struct(mt);
2420 task_rq_unlock(rq, &flags);
2421 wake_up_process(mt);
2422 put_task_struct(mt);
2423 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07002424
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425 return;
2426 }
2427out:
2428 task_rq_unlock(rq, &flags);
2429}
2430
2431/*
Nick Piggin476d1392005-06-25 14:57:29 -07002432 * sched_exec - execve() is a valuable balancing opportunity, because at
2433 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 */
2435void sched_exec(void)
2436{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 int new_cpu, this_cpu = get_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002438 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002440 if (new_cpu != this_cpu)
2441 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442}
2443
2444/*
2445 * pull_task - move a task from a remote runqueue to the local runqueue.
2446 * Both runqueues must be locked.
2447 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002448static void pull_task(struct rq *src_rq, struct task_struct *p,
2449 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02002451 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002453 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454 /*
2455 * Note that idle threads have a prio of MAX_PRIO, for this test
2456 * to be always true for them.
2457 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002458 check_preempt_curr(this_rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459}
2460
2461/*
2462 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2463 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08002464static
Ingo Molnar70b97a72006-07-03 00:25:42 -07002465int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002466 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002467 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468{
2469 /*
2470 * We do not migrate tasks that are:
2471 * 1) running (obviously), or
2472 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2473 * 3) are cache-hot on their current CPU.
2474 */
Ingo Molnarcc367732007-10-15 17:00:18 +02002475 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2476 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002478 }
Nick Piggin81026792005-06-25 14:57:07 -07002479 *all_pinned = 0;
2480
Ingo Molnarcc367732007-10-15 17:00:18 +02002481 if (task_running(rq, p)) {
2482 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07002483 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002484 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485
Ingo Molnarda84d962007-10-15 17:00:18 +02002486 /*
2487 * Aggressive migration if:
2488 * 1) task is cache cold, or
2489 * 2) too many balance attempts have failed.
2490 */
2491
Ingo Molnar6bc16652007-10-15 17:00:18 +02002492 if (!task_hot(p, rq->clock, sd) ||
2493 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002494#ifdef CONFIG_SCHEDSTATS
Ingo Molnarcc367732007-10-15 17:00:18 +02002495 if (task_hot(p, rq->clock, sd)) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002496 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02002497 schedstat_inc(p, se.nr_forced_migrations);
2498 }
Ingo Molnarda84d962007-10-15 17:00:18 +02002499#endif
2500 return 1;
2501 }
2502
Ingo Molnarcc367732007-10-15 17:00:18 +02002503 if (task_hot(p, rq->clock, sd)) {
2504 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02002505 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002506 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 return 1;
2508}
2509
Peter Williamse1d14842007-10-24 18:23:51 +02002510static unsigned long
2511balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2512 unsigned long max_load_move, struct sched_domain *sd,
2513 enum cpu_idle_type idle, int *all_pinned,
2514 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02002515{
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002516 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
Ingo Molnardd41f592007-07-09 18:51:59 +02002517 struct task_struct *p;
2518 long rem_load_move = max_load_move;
2519
Peter Williamse1d14842007-10-24 18:23:51 +02002520 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02002521 goto out;
2522
2523 pinned = 1;
2524
2525 /*
2526 * Start the load-balancing iterator:
2527 */
2528 p = iterator->start(iterator->arg);
2529next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002530 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02002531 goto out;
2532 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002533 * To help distribute high priority tasks across CPUs we don't
Ingo Molnardd41f592007-07-09 18:51:59 +02002534 * skip a task if it will be the highest priority task (i.e. smallest
2535 * prio value) on its new queue regardless of its load weight
2536 */
2537 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2538 SCHED_LOAD_SCALE_FUZZ;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002539 if ((skip_for_load && p->prio >= *this_best_prio) ||
Ingo Molnardd41f592007-07-09 18:51:59 +02002540 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002541 p = iterator->next(iterator->arg);
2542 goto next;
2543 }
2544
2545 pull_task(busiest, p, this_rq, this_cpu);
2546 pulled++;
2547 rem_load_move -= p->se.load.weight;
2548
2549 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002550 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02002551 */
Peter Williamse1d14842007-10-24 18:23:51 +02002552 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002553 if (p->prio < *this_best_prio)
2554 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02002555 p = iterator->next(iterator->arg);
2556 goto next;
2557 }
2558out:
2559 /*
Peter Williamse1d14842007-10-24 18:23:51 +02002560 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02002561 * so we can safely collect pull_task() stats here rather than
2562 * inside pull_task().
2563 */
2564 schedstat_add(sd, lb_gained[idle], pulled);
2565
2566 if (all_pinned)
2567 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02002568
2569 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02002570}
Ingo Molnar48f24c42006-07-03 00:25:40 -07002571
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572/*
Peter Williams43010652007-08-09 11:16:46 +02002573 * move_tasks tries to move up to max_load_move weighted load from busiest to
2574 * this_rq, as part of a balancing operation within domain "sd".
2575 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 *
2577 * Called with both runqueues locked.
2578 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002579static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02002580 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002581 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07002582 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02002584 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02002585 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002586 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587
Ingo Molnardd41f592007-07-09 18:51:59 +02002588 do {
Peter Williams43010652007-08-09 11:16:46 +02002589 total_load_moved +=
2590 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02002591 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002592 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02002593 class = class->next;
Peter Williams43010652007-08-09 11:16:46 +02002594 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595
Peter Williams43010652007-08-09 11:16:46 +02002596 return total_load_moved > 0;
2597}
2598
Peter Williamse1d14842007-10-24 18:23:51 +02002599static int
2600iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2601 struct sched_domain *sd, enum cpu_idle_type idle,
2602 struct rq_iterator *iterator)
2603{
2604 struct task_struct *p = iterator->start(iterator->arg);
2605 int pinned = 0;
2606
2607 while (p) {
2608 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2609 pull_task(busiest, p, this_rq, this_cpu);
2610 /*
2611 * Right now, this is only the second place pull_task()
2612 * is called, so we can safely collect pull_task()
2613 * stats here rather than inside pull_task().
2614 */
2615 schedstat_inc(sd, lb_gained[idle]);
2616
2617 return 1;
2618 }
2619 p = iterator->next(iterator->arg);
2620 }
2621
2622 return 0;
2623}
2624
Peter Williams43010652007-08-09 11:16:46 +02002625/*
2626 * move_one_task tries to move exactly one task from busiest to this_rq, as
2627 * part of active balancing operations within "domain".
2628 * Returns 1 if successful and 0 otherwise.
2629 *
2630 * Called with both runqueues locked.
2631 */
2632static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2633 struct sched_domain *sd, enum cpu_idle_type idle)
2634{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02002635 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02002636
2637 for (class = sched_class_highest; class; class = class->next)
Peter Williamse1d14842007-10-24 18:23:51 +02002638 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02002639 return 1;
2640
2641 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642}
2643
2644/*
2645 * find_busiest_group finds and returns the busiest CPU group within the
Ingo Molnar48f24c42006-07-03 00:25:40 -07002646 * domain. It calculates and returns the amount of weighted load which
2647 * should be moved to restore balance via the imbalance parameter.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 */
2649static struct sched_group *
2650find_busiest_group(struct sched_domain *sd, int this_cpu,
Ingo Molnardd41f592007-07-09 18:51:59 +02002651 unsigned long *imbalance, enum cpu_idle_type idle,
2652 int *sd_idle, cpumask_t *cpus, int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653{
2654 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2655 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002656 unsigned long max_pull;
Peter Williams2dd73a42006-06-27 02:54:34 -07002657 unsigned long busiest_load_per_task, busiest_nr_running;
2658 unsigned long this_load_per_task, this_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02002659 int load_idx, group_imb = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002660#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2661 int power_savings_balance = 1;
2662 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2663 unsigned long min_nr_running = ULONG_MAX;
2664 struct sched_group *group_min = NULL, *group_leader = NULL;
2665#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666
2667 max_load = this_load = total_load = total_pwr = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07002668 busiest_load_per_task = busiest_nr_running = 0;
2669 this_load_per_task = this_nr_running = 0;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002670 if (idle == CPU_NOT_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07002671 load_idx = sd->busy_idx;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002672 else if (idle == CPU_NEWLY_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07002673 load_idx = sd->newidle_idx;
2674 else
2675 load_idx = sd->idle_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676
2677 do {
Ken Chen908a7c12007-10-17 16:55:11 +02002678 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 int local_group;
2680 int i;
Ken Chen908a7c12007-10-17 16:55:11 +02002681 int __group_imb = 0;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002682 unsigned int balance_cpu = -1, first_idle_cpu = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07002683 unsigned long sum_nr_running, sum_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684
2685 local_group = cpu_isset(this_cpu, group->cpumask);
2686
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002687 if (local_group)
2688 balance_cpu = first_cpu(group->cpumask);
2689
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 /* Tally up the load of all CPUs in the group */
Peter Williams2dd73a42006-06-27 02:54:34 -07002691 sum_weighted_load = sum_nr_running = avg_load = 0;
Ken Chen908a7c12007-10-17 16:55:11 +02002692 max_cpu_load = 0;
2693 min_cpu_load = ~0UL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694
2695 for_each_cpu_mask(i, group->cpumask) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002696 struct rq *rq;
2697
2698 if (!cpu_isset(i, *cpus))
2699 continue;
2700
2701 rq = cpu_rq(i);
Peter Williams2dd73a42006-06-27 02:54:34 -07002702
Suresh Siddha9439aab2007-07-19 21:28:35 +02002703 if (*sd_idle && rq->nr_running)
Nick Piggin5969fe02005-09-10 00:26:19 -07002704 *sd_idle = 0;
2705
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 /* Bias balancing toward cpus of our domain */
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002707 if (local_group) {
2708 if (idle_cpu(i) && !first_idle_cpu) {
2709 first_idle_cpu = 1;
2710 balance_cpu = i;
2711 }
2712
Nick Piggina2000572006-02-10 01:51:02 -08002713 load = target_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02002714 } else {
Nick Piggina2000572006-02-10 01:51:02 -08002715 load = source_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02002716 if (load > max_cpu_load)
2717 max_cpu_load = load;
2718 if (min_cpu_load > load)
2719 min_cpu_load = load;
2720 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721
2722 avg_load += load;
Peter Williams2dd73a42006-06-27 02:54:34 -07002723 sum_nr_running += rq->nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02002724 sum_weighted_load += weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 }
2726
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002727 /*
2728 * First idle cpu or the first cpu(busiest) in this sched group
2729 * is eligible for doing load balancing at this and above
Suresh Siddha9439aab2007-07-19 21:28:35 +02002730 * domains. In the newly idle case, we will allow all the cpu's
2731 * to do the newly idle load balance.
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002732 */
Suresh Siddha9439aab2007-07-19 21:28:35 +02002733 if (idle != CPU_NEWLY_IDLE && local_group &&
2734 balance_cpu != this_cpu && balance) {
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002735 *balance = 0;
2736 goto ret;
2737 }
2738
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739 total_load += avg_load;
Eric Dumazet5517d862007-05-08 00:32:57 -07002740 total_pwr += group->__cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741
2742 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07002743 avg_load = sg_div_cpu_power(group,
2744 avg_load * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745
Ken Chen908a7c12007-10-17 16:55:11 +02002746 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2747 __group_imb = 1;
2748
Eric Dumazet5517d862007-05-08 00:32:57 -07002749 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002750
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751 if (local_group) {
2752 this_load = avg_load;
2753 this = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07002754 this_nr_running = sum_nr_running;
2755 this_load_per_task = sum_weighted_load;
2756 } else if (avg_load > max_load &&
Ken Chen908a7c12007-10-17 16:55:11 +02002757 (sum_nr_running > group_capacity || __group_imb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 max_load = avg_load;
2759 busiest = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07002760 busiest_nr_running = sum_nr_running;
2761 busiest_load_per_task = sum_weighted_load;
Ken Chen908a7c12007-10-17 16:55:11 +02002762 group_imb = __group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002764
2765#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2766 /*
2767 * Busy processors will not participate in power savings
2768 * balance.
2769 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002770 if (idle == CPU_NOT_IDLE ||
2771 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2772 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002773
2774 /*
2775 * If the local group is idle or completely loaded
2776 * no need to do power savings balance at this domain
2777 */
2778 if (local_group && (this_nr_running >= group_capacity ||
2779 !this_nr_running))
2780 power_savings_balance = 0;
2781
Ingo Molnardd41f592007-07-09 18:51:59 +02002782 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002783 * If a group is already running at full capacity or idle,
2784 * don't include that group in power savings calculations
Ingo Molnardd41f592007-07-09 18:51:59 +02002785 */
2786 if (!power_savings_balance || sum_nr_running >= group_capacity
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002787 || !sum_nr_running)
Ingo Molnardd41f592007-07-09 18:51:59 +02002788 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002789
Ingo Molnardd41f592007-07-09 18:51:59 +02002790 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002791 * Calculate the group which has the least non-idle load.
Ingo Molnardd41f592007-07-09 18:51:59 +02002792 * This is the group from where we need to pick up the load
2793 * for saving power
2794 */
2795 if ((sum_nr_running < min_nr_running) ||
2796 (sum_nr_running == min_nr_running &&
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002797 first_cpu(group->cpumask) <
2798 first_cpu(group_min->cpumask))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002799 group_min = group;
2800 min_nr_running = sum_nr_running;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002801 min_load_per_task = sum_weighted_load /
2802 sum_nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02002803 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002804
Ingo Molnardd41f592007-07-09 18:51:59 +02002805 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002806 * Calculate the group which is almost near its
Ingo Molnardd41f592007-07-09 18:51:59 +02002807 * capacity but still has some space to pick up some load
2808 * from other group and save more power
2809 */
2810 if (sum_nr_running <= group_capacity - 1) {
2811 if (sum_nr_running > leader_nr_running ||
2812 (sum_nr_running == leader_nr_running &&
2813 first_cpu(group->cpumask) >
2814 first_cpu(group_leader->cpumask))) {
2815 group_leader = group;
2816 leader_nr_running = sum_nr_running;
2817 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002818 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002819group_next:
2820#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 group = group->next;
2822 } while (group != sd->groups);
2823
Peter Williams2dd73a42006-06-27 02:54:34 -07002824 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825 goto out_balanced;
2826
2827 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2828
2829 if (this_load >= avg_load ||
2830 100*max_load <= sd->imbalance_pct*this_load)
2831 goto out_balanced;
2832
Peter Williams2dd73a42006-06-27 02:54:34 -07002833 busiest_load_per_task /= busiest_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02002834 if (group_imb)
2835 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2836
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837 /*
2838 * We're trying to get all the cpus to the average_load, so we don't
2839 * want to push ourselves above the average load, nor do we wish to
2840 * reduce the max loaded cpu below the average load, as either of these
2841 * actions would just result in more rebalancing later, and ping-pong
2842 * tasks around. Thus we look for the minimum possible imbalance.
2843 * Negative imbalances (*we* are more loaded than anyone else) will
2844 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002845 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 * appear as very large values with unsigned longs.
2847 */
Peter Williams2dd73a42006-06-27 02:54:34 -07002848 if (max_load <= busiest_load_per_task)
2849 goto out_balanced;
2850
2851 /*
2852 * In the presence of smp nice balancing, certain scenarios can have
2853 * max load less than avg load(as we skip the groups at or below
2854 * its cpu_power, while calculating max_load..)
2855 */
2856 if (max_load < avg_load) {
2857 *imbalance = 0;
2858 goto small_imbalance;
2859 }
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002860
2861 /* Don't want to pull so many tasks that a group would go idle */
Peter Williams2dd73a42006-06-27 02:54:34 -07002862 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002863
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 /* How much load to actually move to equalise the imbalance */
Eric Dumazet5517d862007-05-08 00:32:57 -07002865 *imbalance = min(max_pull * busiest->__cpu_power,
2866 (avg_load - this_load) * this->__cpu_power)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 / SCHED_LOAD_SCALE;
2868
Peter Williams2dd73a42006-06-27 02:54:34 -07002869 /*
2870 * if *imbalance is less than the average load per runnable task
2871 * there is no gaurantee that any tasks will be moved so we'll have
2872 * a think about bumping its value to force at least one task to be
2873 * moved
2874 */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02002875 if (*imbalance < busiest_load_per_task) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07002876 unsigned long tmp, pwr_now, pwr_move;
Peter Williams2dd73a42006-06-27 02:54:34 -07002877 unsigned int imbn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878
Peter Williams2dd73a42006-06-27 02:54:34 -07002879small_imbalance:
2880 pwr_move = pwr_now = 0;
2881 imbn = 2;
2882 if (this_nr_running) {
2883 this_load_per_task /= this_nr_running;
2884 if (busiest_load_per_task > this_load_per_task)
2885 imbn = 1;
2886 } else
2887 this_load_per_task = SCHED_LOAD_SCALE;
2888
Ingo Molnardd41f592007-07-09 18:51:59 +02002889 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2890 busiest_load_per_task * imbn) {
Peter Williams2dd73a42006-06-27 02:54:34 -07002891 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 return busiest;
2893 }
2894
2895 /*
2896 * OK, we don't have enough imbalance to justify moving tasks,
2897 * however we may be able to increase total CPU power used by
2898 * moving them.
2899 */
2900
Eric Dumazet5517d862007-05-08 00:32:57 -07002901 pwr_now += busiest->__cpu_power *
2902 min(busiest_load_per_task, max_load);
2903 pwr_now += this->__cpu_power *
2904 min(this_load_per_task, this_load);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 pwr_now /= SCHED_LOAD_SCALE;
2906
2907 /* Amount of load we'd subtract */
Eric Dumazet5517d862007-05-08 00:32:57 -07002908 tmp = sg_div_cpu_power(busiest,
2909 busiest_load_per_task * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 if (max_load > tmp)
Eric Dumazet5517d862007-05-08 00:32:57 -07002911 pwr_move += busiest->__cpu_power *
Peter Williams2dd73a42006-06-27 02:54:34 -07002912 min(busiest_load_per_task, max_load - tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913
2914 /* Amount of load we'd add */
Eric Dumazet5517d862007-05-08 00:32:57 -07002915 if (max_load * busiest->__cpu_power <
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08002916 busiest_load_per_task * SCHED_LOAD_SCALE)
Eric Dumazet5517d862007-05-08 00:32:57 -07002917 tmp = sg_div_cpu_power(this,
2918 max_load * busiest->__cpu_power);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 else
Eric Dumazet5517d862007-05-08 00:32:57 -07002920 tmp = sg_div_cpu_power(this,
2921 busiest_load_per_task * SCHED_LOAD_SCALE);
2922 pwr_move += this->__cpu_power *
2923 min(this_load_per_task, this_load + tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 pwr_move /= SCHED_LOAD_SCALE;
2925
2926 /* Move if we gain throughput */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02002927 if (pwr_move > pwr_now)
2928 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 }
2930
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 return busiest;
2932
2933out_balanced:
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002934#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002935 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002936 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002938 if (this == group_leader && group_leader != group_min) {
2939 *imbalance = min_load_per_task;
2940 return group_min;
2941 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002942#endif
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002943ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944 *imbalance = 0;
2945 return NULL;
2946}
2947
2948/*
2949 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2950 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002951static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002952find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002953 unsigned long imbalance, cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002955 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07002956 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 int i;
2958
2959 for_each_cpu_mask(i, group->cpumask) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002960 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002961
2962 if (!cpu_isset(i, *cpus))
2963 continue;
2964
Ingo Molnar48f24c42006-07-03 00:25:40 -07002965 rq = cpu_rq(i);
Ingo Molnardd41f592007-07-09 18:51:59 +02002966 wl = weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967
Ingo Molnardd41f592007-07-09 18:51:59 +02002968 if (rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07002969 continue;
2970
Ingo Molnardd41f592007-07-09 18:51:59 +02002971 if (wl > max_load) {
2972 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002973 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974 }
2975 }
2976
2977 return busiest;
2978}
2979
2980/*
Nick Piggin77391d72005-06-25 14:57:30 -07002981 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2982 * so long as it is large enough.
2983 */
2984#define MAX_PINNED_INTERVAL 512
2985
2986/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2988 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002990static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002991 struct sched_domain *sd, enum cpu_idle_type idle,
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002992 int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993{
Peter Williams43010652007-08-09 11:16:46 +02002994 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002997 struct rq *busiest;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002998 cpumask_t cpus = CPU_MASK_ALL;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002999 unsigned long flags;
Nick Piggin5969fe02005-09-10 00:26:19 -07003000
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003001 /*
3002 * When power savings policy is enabled for the parent domain, idle
3003 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02003004 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003005 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003006 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003007 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003008 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003009 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010
Ingo Molnar2d723762007-10-15 17:00:12 +02003011 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003013redo:
3014 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003015 &cpus, balance);
3016
Chen, Kenneth W06066712006-12-10 02:20:35 -08003017 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003018 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003019
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 if (!group) {
3021 schedstat_inc(sd, lb_nobusyg[idle]);
3022 goto out_balanced;
3023 }
3024
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003025 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 if (!busiest) {
3027 schedstat_inc(sd, lb_nobusyq[idle]);
3028 goto out_balanced;
3029 }
3030
Nick Piggindb935db2005-06-25 14:57:11 -07003031 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032
3033 schedstat_add(sd, lb_imbalance[idle], imbalance);
3034
Peter Williams43010652007-08-09 11:16:46 +02003035 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 if (busiest->nr_running > 1) {
3037 /*
3038 * Attempt to move tasks. If find_busiest_group has found
3039 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02003040 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 * correctly treated as an imbalance.
3042 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003043 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07003044 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02003045 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07003046 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07003047 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003048 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07003049
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003050 /*
3051 * some other cpu did the load balance for us.
3052 */
Peter Williams43010652007-08-09 11:16:46 +02003053 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003054 resched_cpu(this_cpu);
3055
Nick Piggin81026792005-06-25 14:57:07 -07003056 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003057 if (unlikely(all_pinned)) {
3058 cpu_clear(cpu_of(busiest), cpus);
3059 if (!cpus_empty(cpus))
3060 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07003061 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003062 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003063 }
Nick Piggin81026792005-06-25 14:57:07 -07003064
Peter Williams43010652007-08-09 11:16:46 +02003065 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 schedstat_inc(sd, lb_failed[idle]);
3067 sd->nr_balance_failed++;
3068
3069 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003071 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003072
3073 /* don't kick the migration_thread, if the curr
3074 * task on busiest cpu can't be moved to this_cpu
3075 */
3076 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003077 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003078 all_pinned = 1;
3079 goto out_one_pinned;
3080 }
3081
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082 if (!busiest->active_balance) {
3083 busiest->active_balance = 1;
3084 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07003085 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003087 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07003088 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089 wake_up_process(busiest->migration_thread);
3090
3091 /*
3092 * We've kicked active balancing, reset the failure
3093 * counter.
3094 */
Nick Piggin39507452005-06-25 14:57:09 -07003095 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096 }
Nick Piggin81026792005-06-25 14:57:07 -07003097 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098 sd->nr_balance_failed = 0;
3099
Nick Piggin81026792005-06-25 14:57:07 -07003100 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003101 /* We were unbalanced, so reset the balancing interval */
3102 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07003103 } else {
3104 /*
3105 * If we've begun active balancing, start to back off. This
3106 * case may not be covered by the all_pinned logic if there
3107 * is only 1 task on the busy runqueue (because we don't call
3108 * move_tasks).
3109 */
3110 if (sd->balance_interval < sd->max_interval)
3111 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 }
3113
Peter Williams43010652007-08-09 11:16:46 +02003114 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003115 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003116 return -1;
Peter Williams43010652007-08-09 11:16:46 +02003117 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118
3119out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120 schedstat_inc(sd, lb_balanced[idle]);
3121
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003122 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003123
3124out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07003126 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3127 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128 sd->balance_interval *= 2;
3129
Ingo Molnar48f24c42006-07-03 00:25:40 -07003130 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003131 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003132 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133 return 0;
3134}
3135
3136/*
3137 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3138 * tasks if there is an imbalance.
3139 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003140 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 * this_rq is locked.
3142 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07003143static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07003144load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145{
3146 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003147 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02003149 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07003150 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003151 int all_pinned = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003152 cpumask_t cpus = CPU_MASK_ALL;
Nick Piggin5969fe02005-09-10 00:26:19 -07003153
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003154 /*
3155 * When power savings policy is enabled for the parent domain, idle
3156 * sibling can pick up load irrespective of busy siblings. In this case,
3157 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003158 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003159 */
3160 if (sd->flags & SD_SHARE_CPUPOWER &&
3161 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003162 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163
Ingo Molnar2d723762007-10-15 17:00:12 +02003164 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003165redo:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003166 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003167 &sd_idle, &cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003169 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003170 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003171 }
3172
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003173 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003174 &cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07003175 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003176 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003177 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178 }
3179
Nick Piggindb935db2005-06-25 14:57:11 -07003180 BUG_ON(busiest == this_rq);
3181
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003182 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003183
Peter Williams43010652007-08-09 11:16:46 +02003184 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003185 if (busiest->nr_running > 1) {
3186 /* Attempt to move tasks */
3187 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003188 /* this_rq->clock is already updated */
3189 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02003190 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003191 imbalance, sd, CPU_NEWLY_IDLE,
3192 &all_pinned);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003193 spin_unlock(&busiest->lock);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003194
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003195 if (unlikely(all_pinned)) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003196 cpu_clear(cpu_of(busiest), cpus);
3197 if (!cpus_empty(cpus))
3198 goto redo;
3199 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003200 }
3201
Peter Williams43010652007-08-09 11:16:46 +02003202 if (!ld_moved) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003203 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003204 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3205 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003206 return -1;
3207 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003208 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209
Peter Williams43010652007-08-09 11:16:46 +02003210 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003211
3212out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003213 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003214 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003215 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003216 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003217 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003218
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003219 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220}
3221
3222/*
3223 * idle_balance is called by schedule() if this_cpu is about to become
3224 * idle. Attempts to pull tasks from other CPUs.
3225 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003226static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227{
3228 struct sched_domain *sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02003229 int pulled_task = -1;
3230 unsigned long next_balance = jiffies + HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231
3232 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003233 unsigned long interval;
3234
3235 if (!(sd->flags & SD_LOAD_BALANCE))
3236 continue;
3237
3238 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07003239 /* If we've pulled tasks over stop searching: */
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003240 pulled_task = load_balance_newidle(this_cpu,
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003241 this_rq, sd);
3242
3243 interval = msecs_to_jiffies(sd->balance_interval);
3244 if (time_after(next_balance, sd->last_balance + interval))
3245 next_balance = sd->last_balance + interval;
3246 if (pulled_task)
3247 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003249 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003250 /*
3251 * We are going idle. next_balance may be set based on
3252 * a busy processor. So reset next_balance.
3253 */
3254 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02003255 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256}
3257
3258/*
3259 * active_load_balance is run by migration threads. It pushes running tasks
3260 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3261 * running on each physical CPU where possible, and avoids physical /
3262 * logical imbalances.
3263 *
3264 * Called with busiest_rq locked.
3265 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003266static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267{
Nick Piggin39507452005-06-25 14:57:09 -07003268 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003269 struct sched_domain *sd;
3270 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07003271
Ingo Molnar48f24c42006-07-03 00:25:40 -07003272 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07003273 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07003274 return;
3275
3276 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003277
3278 /*
Nick Piggin39507452005-06-25 14:57:09 -07003279 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003280 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07003281 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282 */
Nick Piggin39507452005-06-25 14:57:09 -07003283 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284
Nick Piggin39507452005-06-25 14:57:09 -07003285 /* move a task from busiest_rq to target_rq */
3286 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003287 update_rq_clock(busiest_rq);
3288 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003289
Nick Piggin39507452005-06-25 14:57:09 -07003290 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003291 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07003292 if ((sd->flags & SD_LOAD_BALANCE) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07003293 cpu_isset(busiest_cpu, sd->span))
Nick Piggin39507452005-06-25 14:57:09 -07003294 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003295 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296
Ingo Molnar48f24c42006-07-03 00:25:40 -07003297 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02003298 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299
Peter Williams43010652007-08-09 11:16:46 +02003300 if (move_one_task(target_rq, target_cpu, busiest_rq,
3301 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07003302 schedstat_inc(sd, alb_pushed);
3303 else
3304 schedstat_inc(sd, alb_failed);
3305 }
Nick Piggin39507452005-06-25 14:57:09 -07003306 spin_unlock(&target_rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003307}
3308
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003309#ifdef CONFIG_NO_HZ
3310static struct {
3311 atomic_t load_balancer;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003312 cpumask_t cpu_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003313} nohz ____cacheline_aligned = {
3314 .load_balancer = ATOMIC_INIT(-1),
3315 .cpu_mask = CPU_MASK_NONE,
3316};
3317
Christoph Lameter7835b982006-12-10 02:20:22 -08003318/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003319 * This routine will try to nominate the ilb (idle load balancing)
3320 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3321 * load balancing on behalf of all those cpus. If all the cpus in the system
3322 * go into this tickless mode, then there will be no ilb owner (as there is
3323 * no need for one) and all the cpus will sleep till the next wakeup event
3324 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08003325 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003326 * For the ilb owner, tick is not stopped. And this tick will be used
3327 * for idle load balancing. ilb owner will still be part of
3328 * nohz.cpu_mask..
3329 *
3330 * While stopping the tick, this cpu will become the ilb owner if there
3331 * is no other owner. And will be the owner till that cpu becomes busy
3332 * or if all cpus in the system stop their ticks at which point
3333 * there is no need for ilb owner.
3334 *
3335 * When the ilb owner becomes busy, it nominates another owner, during the
3336 * next busy scheduler_tick()
3337 */
3338int select_nohz_load_balancer(int stop_tick)
3339{
3340 int cpu = smp_processor_id();
3341
3342 if (stop_tick) {
3343 cpu_set(cpu, nohz.cpu_mask);
3344 cpu_rq(cpu)->in_nohz_recently = 1;
3345
3346 /*
3347 * If we are going offline and still the leader, give up!
3348 */
3349 if (cpu_is_offline(cpu) &&
3350 atomic_read(&nohz.load_balancer) == cpu) {
3351 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3352 BUG();
3353 return 0;
3354 }
3355
3356 /* time for ilb owner also to sleep */
3357 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3358 if (atomic_read(&nohz.load_balancer) == cpu)
3359 atomic_set(&nohz.load_balancer, -1);
3360 return 0;
3361 }
3362
3363 if (atomic_read(&nohz.load_balancer) == -1) {
3364 /* make me the ilb owner */
3365 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3366 return 1;
3367 } else if (atomic_read(&nohz.load_balancer) == cpu)
3368 return 1;
3369 } else {
3370 if (!cpu_isset(cpu, nohz.cpu_mask))
3371 return 0;
3372
3373 cpu_clear(cpu, nohz.cpu_mask);
3374
3375 if (atomic_read(&nohz.load_balancer) == cpu)
3376 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3377 BUG();
3378 }
3379 return 0;
3380}
3381#endif
3382
3383static DEFINE_SPINLOCK(balancing);
3384
3385/*
Christoph Lameter7835b982006-12-10 02:20:22 -08003386 * It checks each scheduling domain to see if it is due to be balanced,
3387 * and initiates a balancing operation if so.
3388 *
3389 * Balancing parameters are set up in arch_init_sched_domains.
3390 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02003391static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08003392{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003393 int balance = 1;
3394 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08003395 unsigned long interval;
3396 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003397 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08003398 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003399 int update_next_balance = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003401 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402 if (!(sd->flags & SD_LOAD_BALANCE))
3403 continue;
3404
3405 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003406 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 interval *= sd->busy_factor;
3408
3409 /* scale ms to jiffies */
3410 interval = msecs_to_jiffies(interval);
3411 if (unlikely(!interval))
3412 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003413 if (interval > HZ*NR_CPUS/10)
3414 interval = HZ*NR_CPUS/10;
3415
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416
Christoph Lameter08c183f2006-12-10 02:20:29 -08003417 if (sd->flags & SD_SERIALIZE) {
3418 if (!spin_trylock(&balancing))
3419 goto out;
3420 }
3421
Christoph Lameterc9819f42006-12-10 02:20:25 -08003422 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003423 if (load_balance(cpu, rq, sd, idle, &balance)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003424 /*
3425 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07003426 * longer idle, or one of our SMT siblings is
3427 * not idle.
3428 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003429 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003431 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432 }
Christoph Lameter08c183f2006-12-10 02:20:29 -08003433 if (sd->flags & SD_SERIALIZE)
3434 spin_unlock(&balancing);
3435out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02003436 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08003437 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003438 update_next_balance = 1;
3439 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003440
3441 /*
3442 * Stop the load balance at this level. There is another
3443 * CPU in our sched group which is doing load balancing more
3444 * actively.
3445 */
3446 if (!balance)
3447 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02003449
3450 /*
3451 * next_balance will be updated only when there is a need.
3452 * When the cpu is attached to null domain for ex, it will not be
3453 * updated.
3454 */
3455 if (likely(update_next_balance))
3456 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003457}
3458
3459/*
3460 * run_rebalance_domains is triggered when needed from the scheduler tick.
3461 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3462 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3463 */
3464static void run_rebalance_domains(struct softirq_action *h)
3465{
Ingo Molnardd41f592007-07-09 18:51:59 +02003466 int this_cpu = smp_processor_id();
3467 struct rq *this_rq = cpu_rq(this_cpu);
3468 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3469 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003470
Ingo Molnardd41f592007-07-09 18:51:59 +02003471 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003472
3473#ifdef CONFIG_NO_HZ
3474 /*
3475 * If this cpu is the owner for idle load balancing, then do the
3476 * balancing on behalf of the other idle cpus whose ticks are
3477 * stopped.
3478 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003479 if (this_rq->idle_at_tick &&
3480 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003481 cpumask_t cpus = nohz.cpu_mask;
3482 struct rq *rq;
3483 int balance_cpu;
3484
Ingo Molnardd41f592007-07-09 18:51:59 +02003485 cpu_clear(this_cpu, cpus);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003486 for_each_cpu_mask(balance_cpu, cpus) {
3487 /*
3488 * If this cpu gets work to do, stop the load balancing
3489 * work being done for other cpus. Next load
3490 * balancing owner will pick it up.
3491 */
3492 if (need_resched())
3493 break;
3494
Oleg Nesterovde0cf892007-08-12 18:08:19 +02003495 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003496
3497 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003498 if (time_after(this_rq->next_balance, rq->next_balance))
3499 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003500 }
3501 }
3502#endif
3503}
3504
3505/*
3506 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3507 *
3508 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3509 * idle load balancing owner or decide to stop the periodic load balancing,
3510 * if the whole system is idle.
3511 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003512static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003513{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003514#ifdef CONFIG_NO_HZ
3515 /*
3516 * If we were in the nohz mode recently and busy at the current
3517 * scheduler tick, then check if we need to nominate new idle
3518 * load balancer.
3519 */
3520 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3521 rq->in_nohz_recently = 0;
3522
3523 if (atomic_read(&nohz.load_balancer) == cpu) {
3524 cpu_clear(cpu, nohz.cpu_mask);
3525 atomic_set(&nohz.load_balancer, -1);
3526 }
3527
3528 if (atomic_read(&nohz.load_balancer) == -1) {
3529 /*
3530 * simple selection for now: Nominate the
3531 * first cpu in the nohz list to be the next
3532 * ilb owner.
3533 *
3534 * TBD: Traverse the sched domains and nominate
3535 * the nearest cpu in the nohz.cpu_mask.
3536 */
3537 int ilb = first_cpu(nohz.cpu_mask);
3538
3539 if (ilb != NR_CPUS)
3540 resched_cpu(ilb);
3541 }
3542 }
3543
3544 /*
3545 * If this cpu is idle and doing idle load balancing for all the
3546 * cpus with ticks stopped, is it time for that to stop?
3547 */
3548 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3549 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3550 resched_cpu(cpu);
3551 return;
3552 }
3553
3554 /*
3555 * If this cpu is idle and the idle load balancing is done by
3556 * someone else, then no need raise the SCHED_SOFTIRQ
3557 */
3558 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3559 cpu_isset(cpu, nohz.cpu_mask))
3560 return;
3561#endif
3562 if (time_after_eq(jiffies, rq->next_balance))
3563 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564}
Ingo Molnardd41f592007-07-09 18:51:59 +02003565
3566#else /* CONFIG_SMP */
3567
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568/*
3569 * on UP we do not need to balance between CPUs:
3570 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003571static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572{
3573}
Ingo Molnardd41f592007-07-09 18:51:59 +02003574
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575#endif
3576
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577DEFINE_PER_CPU(struct kernel_stat, kstat);
3578
3579EXPORT_PER_CPU_SYMBOL(kstat);
3580
3581/*
Ingo Molnar41b86e92007-07-09 18:51:58 +02003582 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3583 * that have not yet been banked in case the task is currently running.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584 */
Ingo Molnar41b86e92007-07-09 18:51:58 +02003585unsigned long long task_sched_runtime(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003588 u64 ns, delta_exec;
3589 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003590
Ingo Molnar41b86e92007-07-09 18:51:58 +02003591 rq = task_rq_lock(p, &flags);
3592 ns = p->se.sum_exec_runtime;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01003593 if (task_current(rq, p)) {
Ingo Molnara8e504d2007-08-09 11:16:47 +02003594 update_rq_clock(rq);
3595 delta_exec = rq->clock - p->se.exec_start;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003596 if ((s64)delta_exec > 0)
3597 ns += delta_exec;
3598 }
3599 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003600
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601 return ns;
3602}
3603
3604/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605 * Account user cpu time to a process.
3606 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607 * @cputime: the cpu time spent in user space since the last update
3608 */
3609void account_user_time(struct task_struct *p, cputime_t cputime)
3610{
3611 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3612 cputime64_t tmp;
3613
3614 p->utime = cputime_add(p->utime, cputime);
3615
3616 /* Add user time to cpustat. */
3617 tmp = cputime_to_cputime64(cputime);
3618 if (TASK_NICE(p) > 0)
3619 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3620 else
3621 cpustat->user = cputime64_add(cpustat->user, tmp);
3622}
3623
3624/*
Laurent Vivier94886b82007-10-15 17:00:19 +02003625 * Account guest cpu time to a process.
3626 * @p: the process that the cpu time gets accounted to
3627 * @cputime: the cpu time spent in virtual machine since the last update
3628 */
Adrian Bunkf7402e02007-10-29 21:18:10 +01003629static void account_guest_time(struct task_struct *p, cputime_t cputime)
Laurent Vivier94886b82007-10-15 17:00:19 +02003630{
3631 cputime64_t tmp;
3632 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3633
3634 tmp = cputime_to_cputime64(cputime);
3635
3636 p->utime = cputime_add(p->utime, cputime);
3637 p->gtime = cputime_add(p->gtime, cputime);
3638
3639 cpustat->user = cputime64_add(cpustat->user, tmp);
3640 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3641}
3642
3643/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07003644 * Account scaled user cpu time to a process.
3645 * @p: the process that the cpu time gets accounted to
3646 * @cputime: the cpu time spent in user space since the last update
3647 */
3648void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3649{
3650 p->utimescaled = cputime_add(p->utimescaled, cputime);
3651}
3652
3653/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654 * Account system cpu time to a process.
3655 * @p: the process that the cpu time gets accounted to
3656 * @hardirq_offset: the offset to subtract from hardirq_count()
3657 * @cputime: the cpu time spent in kernel space since the last update
3658 */
3659void account_system_time(struct task_struct *p, int hardirq_offset,
3660 cputime_t cputime)
3661{
3662 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003663 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 cputime64_t tmp;
3665
Christian Borntraeger97783852007-11-15 20:57:39 +01003666 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3667 return account_guest_time(p, cputime);
Laurent Vivier94886b82007-10-15 17:00:19 +02003668
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669 p->stime = cputime_add(p->stime, cputime);
3670
3671 /* Add system time to cpustat. */
3672 tmp = cputime_to_cputime64(cputime);
3673 if (hardirq_count() - hardirq_offset)
3674 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3675 else if (softirq_count())
3676 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08003677 else if (p != rq->idle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 cpustat->system = cputime64_add(cpustat->system, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08003679 else if (atomic_read(&rq->nr_iowait) > 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3681 else
3682 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3683 /* Account for system time used */
3684 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685}
3686
3687/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07003688 * Account scaled system cpu time to a process.
3689 * @p: the process that the cpu time gets accounted to
3690 * @hardirq_offset: the offset to subtract from hardirq_count()
3691 * @cputime: the cpu time spent in kernel space since the last update
3692 */
3693void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3694{
3695 p->stimescaled = cputime_add(p->stimescaled, cputime);
3696}
3697
3698/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003699 * Account for involuntary wait time.
3700 * @p: the process from which the cpu time has been stolen
3701 * @steal: the cpu time spent in involuntary wait
3702 */
3703void account_steal_time(struct task_struct *p, cputime_t steal)
3704{
3705 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3706 cputime64_t tmp = cputime_to_cputime64(steal);
Ingo Molnar70b97a72006-07-03 00:25:42 -07003707 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708
3709 if (p == rq->idle) {
3710 p->stime = cputime_add(p->stime, steal);
3711 if (atomic_read(&rq->nr_iowait) > 0)
3712 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3713 else
3714 cpustat->idle = cputime64_add(cpustat->idle, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08003715 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3717}
3718
Christoph Lameter7835b982006-12-10 02:20:22 -08003719/*
3720 * This function gets called by the timer code, with HZ frequency.
3721 * We call it with interrupts disabled.
3722 *
3723 * It also gets called by the fork code, when changing the parent's
3724 * timeslices.
3725 */
3726void scheduler_tick(void)
3727{
Christoph Lameter7835b982006-12-10 02:20:22 -08003728 int cpu = smp_processor_id();
3729 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003730 struct task_struct *curr = rq->curr;
Ingo Molnar529c7722007-08-10 23:05:11 +02003731 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
Christoph Lameter7835b982006-12-10 02:20:22 -08003732
Ingo Molnardd41f592007-07-09 18:51:59 +02003733 spin_lock(&rq->lock);
Ingo Molnar546fe3c2007-08-09 11:16:51 +02003734 __update_rq_clock(rq);
Ingo Molnar529c7722007-08-10 23:05:11 +02003735 /*
3736 * Let rq->clock advance by at least TICK_NSEC:
3737 */
Guillaume Chazaraincc203d22008-01-25 21:08:34 +01003738 if (unlikely(rq->clock < next_tick)) {
Ingo Molnar529c7722007-08-10 23:05:11 +02003739 rq->clock = next_tick;
Guillaume Chazaraincc203d22008-01-25 21:08:34 +01003740 rq->clock_underflows++;
3741 }
Ingo Molnar529c7722007-08-10 23:05:11 +02003742 rq->tick_timestamp = rq->clock;
Ingo Molnarf1a438d2007-08-09 11:16:45 +02003743 update_cpu_load(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01003744 curr->sched_class->task_tick(rq, curr, 0);
3745 update_sched_rt_period(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02003746 spin_unlock(&rq->lock);
3747
Christoph Lametere418e1c2006-12-10 02:20:23 -08003748#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02003749 rq->idle_at_tick = idle_cpu(cpu);
3750 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08003751#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752}
3753
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3755
3756void fastcall add_preempt_count(int val)
3757{
3758 /*
3759 * Underflow?
3760 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003761 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3762 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003763 preempt_count() += val;
3764 /*
3765 * Spinlock count overflowing soon?
3766 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08003767 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3768 PREEMPT_MASK - 10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769}
3770EXPORT_SYMBOL(add_preempt_count);
3771
3772void fastcall sub_preempt_count(int val)
3773{
3774 /*
3775 * Underflow?
3776 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003777 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3778 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779 /*
3780 * Is the spinlock portion underflowing?
3781 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003782 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3783 !(preempt_count() & PREEMPT_MASK)))
3784 return;
3785
Linus Torvalds1da177e2005-04-16 15:20:36 -07003786 preempt_count() -= val;
3787}
3788EXPORT_SYMBOL(sub_preempt_count);
3789
3790#endif
3791
3792/*
Ingo Molnardd41f592007-07-09 18:51:59 +02003793 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003795static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796{
Satyam Sharma838225b2007-10-24 18:23:50 +02003797 struct pt_regs *regs = get_irq_regs();
3798
3799 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3800 prev->comm, prev->pid, preempt_count());
3801
Ingo Molnardd41f592007-07-09 18:51:59 +02003802 debug_show_held_locks(prev);
3803 if (irqs_disabled())
3804 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02003805
3806 if (regs)
3807 show_regs(regs);
3808 else
3809 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02003810}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003811
Ingo Molnardd41f592007-07-09 18:51:59 +02003812/*
3813 * Various schedule()-time debugging checks and statistics:
3814 */
3815static inline void schedule_debug(struct task_struct *prev)
3816{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003818 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819 * schedule() atomically, we ignore that path for now.
3820 * Otherwise, whine if we are scheduling when we should not be.
3821 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003822 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3823 __schedule_bug(prev);
3824
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3826
Ingo Molnar2d723762007-10-15 17:00:12 +02003827 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02003828#ifdef CONFIG_SCHEDSTATS
3829 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02003830 schedstat_inc(this_rq(), bkl_count);
3831 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02003832 }
3833#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02003834}
3835
3836/*
3837 * Pick up the highest-prio task:
3838 */
3839static inline struct task_struct *
Ingo Molnarff95f3d2007-08-09 11:16:49 +02003840pick_next_task(struct rq *rq, struct task_struct *prev)
Ingo Molnardd41f592007-07-09 18:51:59 +02003841{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003842 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02003843 struct task_struct *p;
3844
3845 /*
3846 * Optimization: we know that if all tasks are in
3847 * the fair class we can call that function directly:
3848 */
3849 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02003850 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02003851 if (likely(p))
3852 return p;
3853 }
3854
3855 class = sched_class_highest;
3856 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02003857 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02003858 if (p)
3859 return p;
3860 /*
3861 * Will never be NULL as the idle class always
3862 * returns a non-NULL p:
3863 */
3864 class = class->next;
3865 }
3866}
3867
3868/*
3869 * schedule() is the main scheduler function.
3870 */
3871asmlinkage void __sched schedule(void)
3872{
3873 struct task_struct *prev, *next;
3874 long *switch_count;
3875 struct rq *rq;
Ingo Molnardd41f592007-07-09 18:51:59 +02003876 int cpu;
3877
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878need_resched:
3879 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02003880 cpu = smp_processor_id();
3881 rq = cpu_rq(cpu);
3882 rcu_qsctr_inc(cpu);
3883 prev = rq->curr;
3884 switch_count = &prev->nivcsw;
3885
Linus Torvalds1da177e2005-04-16 15:20:36 -07003886 release_kernel_lock(prev);
3887need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888
Ingo Molnardd41f592007-07-09 18:51:59 +02003889 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01003891 hrtick_clear(rq);
3892
Ingo Molnar1e819952007-10-15 17:00:13 +02003893 /*
3894 * Do the rq-clock update outside the rq lock:
3895 */
3896 local_irq_disable();
Ingo Molnarc1b3da32007-08-09 11:16:47 +02003897 __update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02003898 spin_lock(&rq->lock);
3899 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900
Ingo Molnardd41f592007-07-09 18:51:59 +02003901 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3902 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3903 unlikely(signal_pending(prev)))) {
3904 prev->state = TASK_RUNNING;
3905 } else {
Ingo Molnar2e1cb742007-08-09 11:16:49 +02003906 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02003907 }
3908 switch_count = &prev->nvcsw;
3909 }
3910
Steven Rostedt9a897c52008-01-25 21:08:22 +01003911#ifdef CONFIG_SMP
3912 if (prev->sched_class->pre_schedule)
3913 prev->sched_class->pre_schedule(rq, prev);
3914#endif
Steven Rostedtf65eda42008-01-25 21:08:07 +01003915
Ingo Molnardd41f592007-07-09 18:51:59 +02003916 if (unlikely(!rq->nr_running))
3917 idle_balance(cpu, rq);
3918
Ingo Molnar31ee5292007-08-09 11:16:49 +02003919 prev->sched_class->put_prev_task(rq, prev);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02003920 next = pick_next_task(rq, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921
3922 sched_info_switch(prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02003923
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924 if (likely(prev != next)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925 rq->nr_switches++;
3926 rq->curr = next;
3927 ++*switch_count;
3928
Ingo Molnardd41f592007-07-09 18:51:59 +02003929 context_switch(rq, prev, next); /* unlocks the rq */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01003930 /*
3931 * the context switch might have flipped the stack from under
3932 * us, hence refresh the local variables.
3933 */
3934 cpu = smp_processor_id();
3935 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936 } else
3937 spin_unlock_irq(&rq->lock);
3938
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01003939 hrtick_set(rq);
3940
3941 if (unlikely(reacquire_kernel_lock(current) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942 goto need_resched_nonpreemptible;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01003943
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 preempt_enable_no_resched();
3945 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3946 goto need_resched;
3947}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948EXPORT_SYMBOL(schedule);
3949
3950#ifdef CONFIG_PREEMPT
3951/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003952 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003953 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954 * occur there and call schedule directly.
3955 */
3956asmlinkage void __sched preempt_schedule(void)
3957{
3958 struct thread_info *ti = current_thread_info();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959 struct task_struct *task = current;
3960 int saved_lock_depth;
Ingo Molnar6478d882008-01-25 21:08:33 +01003961
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962 /*
3963 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003964 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07003966 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003967 return;
3968
Andi Kleen3a5c3592007-10-15 17:00:14 +02003969 do {
3970 add_preempt_count(PREEMPT_ACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971
Andi Kleen3a5c3592007-10-15 17:00:14 +02003972 /*
3973 * We keep the big kernel semaphore locked, but we
3974 * clear ->lock_depth so that schedule() doesnt
3975 * auto-release the semaphore:
3976 */
Andi Kleen3a5c3592007-10-15 17:00:14 +02003977 saved_lock_depth = task->lock_depth;
3978 task->lock_depth = -1;
Andi Kleen3a5c3592007-10-15 17:00:14 +02003979 schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02003980 task->lock_depth = saved_lock_depth;
Andi Kleen3a5c3592007-10-15 17:00:14 +02003981 sub_preempt_count(PREEMPT_ACTIVE);
3982
3983 /*
3984 * Check again in case we missed a preemption opportunity
3985 * between schedule and now.
3986 */
3987 barrier();
3988 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990EXPORT_SYMBOL(preempt_schedule);
3991
3992/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003993 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994 * off of irq context.
3995 * Note, that this is called and return with irqs disabled. This will
3996 * protect us against recursive calling from irq.
3997 */
3998asmlinkage void __sched preempt_schedule_irq(void)
3999{
4000 struct thread_info *ti = current_thread_info();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001 struct task_struct *task = current;
4002 int saved_lock_depth;
Ingo Molnar6478d882008-01-25 21:08:33 +01004003
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004004 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 BUG_ON(ti->preempt_count || !irqs_disabled());
4006
Andi Kleen3a5c3592007-10-15 17:00:14 +02004007 do {
4008 add_preempt_count(PREEMPT_ACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009
Andi Kleen3a5c3592007-10-15 17:00:14 +02004010 /*
4011 * We keep the big kernel semaphore locked, but we
4012 * clear ->lock_depth so that schedule() doesnt
4013 * auto-release the semaphore:
4014 */
Andi Kleen3a5c3592007-10-15 17:00:14 +02004015 saved_lock_depth = task->lock_depth;
4016 task->lock_depth = -1;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004017 local_irq_enable();
4018 schedule();
4019 local_irq_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02004020 task->lock_depth = saved_lock_depth;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004021 sub_preempt_count(PREEMPT_ACTIVE);
4022
4023 /*
4024 * Check again in case we missed a preemption opportunity
4025 * between schedule and now.
4026 */
4027 barrier();
4028 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029}
4030
4031#endif /* CONFIG_PREEMPT */
4032
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004033int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4034 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004036 return try_to_wake_up(curr->private, mode, sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038EXPORT_SYMBOL(default_wake_function);
4039
4040/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004041 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4042 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043 * number) then we wake all the non-exclusive tasks and one exclusive task.
4044 *
4045 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004046 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4048 */
4049static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4050 int nr_exclusive, int sync, void *key)
4051{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004052 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004054 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07004055 unsigned flags = curr->flags;
4056
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057 if (curr->func(curr, mode, sync, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07004058 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059 break;
4060 }
4061}
4062
4063/**
4064 * __wake_up - wake up threads blocked on a waitqueue.
4065 * @q: the waitqueue
4066 * @mode: which threads
4067 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07004068 * @key: is directly passed to the wakeup function
Linus Torvalds1da177e2005-04-16 15:20:36 -07004069 */
4070void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004071 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004072{
4073 unsigned long flags;
4074
4075 spin_lock_irqsave(&q->lock, flags);
4076 __wake_up_common(q, mode, nr_exclusive, 0, key);
4077 spin_unlock_irqrestore(&q->lock, flags);
4078}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004079EXPORT_SYMBOL(__wake_up);
4080
4081/*
4082 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4083 */
4084void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4085{
4086 __wake_up_common(q, mode, 1, 0, NULL);
4087}
4088
4089/**
Martin Waitz67be2dd2005-05-01 08:59:26 -07004090 * __wake_up_sync - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091 * @q: the waitqueue
4092 * @mode: which threads
4093 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4094 *
4095 * The sync wakeup differs that the waker knows that it will schedule
4096 * away soon, so while the target thread will be woken up, it will not
4097 * be migrated to another CPU - ie. the two threads are 'synchronized'
4098 * with each other. This can prevent needless bouncing between CPUs.
4099 *
4100 * On UP it can prevent extra preemption.
4101 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004102void fastcall
4103__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104{
4105 unsigned long flags;
4106 int sync = 1;
4107
4108 if (unlikely(!q))
4109 return;
4110
4111 if (unlikely(!nr_exclusive))
4112 sync = 0;
4113
4114 spin_lock_irqsave(&q->lock, flags);
4115 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4116 spin_unlock_irqrestore(&q->lock, flags);
4117}
4118EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4119
Ingo Molnarb15136e2007-10-24 18:23:48 +02004120void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121{
4122 unsigned long flags;
4123
4124 spin_lock_irqsave(&x->wait.lock, flags);
4125 x->done++;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004126 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127 spin_unlock_irqrestore(&x->wait.lock, flags);
4128}
4129EXPORT_SYMBOL(complete);
4130
Ingo Molnarb15136e2007-10-24 18:23:48 +02004131void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132{
4133 unsigned long flags;
4134
4135 spin_lock_irqsave(&x->wait.lock, flags);
4136 x->done += UINT_MAX/2;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004137 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138 spin_unlock_irqrestore(&x->wait.lock, flags);
4139}
4140EXPORT_SYMBOL(complete_all);
4141
Andi Kleen8cbbe862007-10-15 17:00:14 +02004142static inline long __sched
4143do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145 if (!x->done) {
4146 DECLARE_WAITQUEUE(wait, current);
4147
4148 wait.flags |= WQ_FLAG_EXCLUSIVE;
4149 __add_wait_queue_tail(&x->wait, &wait);
4150 do {
Matthew Wilcox009e5772007-12-06 12:29:54 -05004151 if ((state == TASK_INTERRUPTIBLE &&
4152 signal_pending(current)) ||
4153 (state == TASK_KILLABLE &&
4154 fatal_signal_pending(current))) {
Andi Kleen8cbbe862007-10-15 17:00:14 +02004155 __remove_wait_queue(&x->wait, &wait);
4156 return -ERESTARTSYS;
4157 }
4158 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004160 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161 spin_lock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004162 if (!timeout) {
4163 __remove_wait_queue(&x->wait, &wait);
4164 return timeout;
4165 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004166 } while (!x->done);
4167 __remove_wait_queue(&x->wait, &wait);
4168 }
4169 x->done--;
Andi Kleen8cbbe862007-10-15 17:00:14 +02004170 return timeout;
4171}
4172
4173static long __sched
4174wait_for_common(struct completion *x, long timeout, int state)
4175{
4176 might_sleep();
4177
4178 spin_lock_irq(&x->wait.lock);
4179 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004181 return timeout;
4182}
4183
Ingo Molnarb15136e2007-10-24 18:23:48 +02004184void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02004185{
4186 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187}
4188EXPORT_SYMBOL(wait_for_completion);
4189
Ingo Molnarb15136e2007-10-24 18:23:48 +02004190unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004191wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4192{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004193 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004194}
4195EXPORT_SYMBOL(wait_for_completion_timeout);
4196
Andi Kleen8cbbe862007-10-15 17:00:14 +02004197int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198{
Andi Kleen51e97992007-10-18 21:32:55 +02004199 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4200 if (t == -ERESTARTSYS)
4201 return t;
4202 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004203}
4204EXPORT_SYMBOL(wait_for_completion_interruptible);
4205
Ingo Molnarb15136e2007-10-24 18:23:48 +02004206unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207wait_for_completion_interruptible_timeout(struct completion *x,
4208 unsigned long timeout)
4209{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004210 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211}
4212EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4213
Matthew Wilcox009e5772007-12-06 12:29:54 -05004214int __sched wait_for_completion_killable(struct completion *x)
4215{
4216 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4217 if (t == -ERESTARTSYS)
4218 return t;
4219 return 0;
4220}
4221EXPORT_SYMBOL(wait_for_completion_killable);
4222
Andi Kleen8cbbe862007-10-15 17:00:14 +02004223static long __sched
4224sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02004225{
4226 unsigned long flags;
4227 wait_queue_t wait;
4228
4229 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230
Andi Kleen8cbbe862007-10-15 17:00:14 +02004231 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004232
Andi Kleen8cbbe862007-10-15 17:00:14 +02004233 spin_lock_irqsave(&q->lock, flags);
4234 __add_wait_queue(q, &wait);
4235 spin_unlock(&q->lock);
4236 timeout = schedule_timeout(timeout);
4237 spin_lock_irq(&q->lock);
4238 __remove_wait_queue(q, &wait);
4239 spin_unlock_irqrestore(&q->lock, flags);
4240
4241 return timeout;
4242}
4243
4244void __sched interruptible_sleep_on(wait_queue_head_t *q)
4245{
4246 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004247}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248EXPORT_SYMBOL(interruptible_sleep_on);
4249
Ingo Molnar0fec1712007-07-09 18:52:01 +02004250long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004251interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004252{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004253 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004255EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4256
Ingo Molnar0fec1712007-07-09 18:52:01 +02004257void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004258{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004259 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004261EXPORT_SYMBOL(sleep_on);
4262
Ingo Molnar0fec1712007-07-09 18:52:01 +02004263long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004265 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004266}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267EXPORT_SYMBOL(sleep_on_timeout);
4268
Ingo Molnarb29739f2006-06-27 02:54:51 -07004269#ifdef CONFIG_RT_MUTEXES
4270
4271/*
4272 * rt_mutex_setprio - set the current priority of a task
4273 * @p: task
4274 * @prio: prio value (kernel-internal form)
4275 *
4276 * This function changes the 'effective' priority of a task. It does
4277 * not touch ->normal_prio like __setscheduler().
4278 *
4279 * Used by the rt_mutex code to implement priority inheritance logic.
4280 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004281void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07004282{
4283 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004284 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004285 struct rq *rq;
Steven Rostedtcb469842008-01-25 21:08:22 +01004286 const struct sched_class *prev_class = p->sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004287
4288 BUG_ON(prio < 0 || prio > MAX_PRIO);
4289
4290 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004291 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004292
Andrew Mortond5f9f942007-05-08 20:27:06 -07004293 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004294 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004295 running = task_current(rq, p);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004296 if (on_rq) {
Ingo Molnar69be72c2007-08-09 11:16:49 +02004297 dequeue_task(rq, p, 0);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004298 if (running)
4299 p->sched_class->put_prev_task(rq, p);
4300 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004301
4302 if (rt_prio(prio))
4303 p->sched_class = &rt_sched_class;
4304 else
4305 p->sched_class = &fair_sched_class;
4306
Ingo Molnarb29739f2006-06-27 02:54:51 -07004307 p->prio = prio;
4308
Ingo Molnardd41f592007-07-09 18:51:59 +02004309 if (on_rq) {
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004310 if (running)
4311 p->sched_class->set_curr_task(rq);
Steven Rostedtcb469842008-01-25 21:08:22 +01004312
Ingo Molnar8159f872007-08-09 11:16:49 +02004313 enqueue_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004314
4315 check_class_changed(rq, p, prev_class, oldprio, running);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004316 }
4317 task_rq_unlock(rq, &flags);
4318}
4319
4320#endif
4321
Ingo Molnar36c8b582006-07-03 00:25:41 -07004322void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004323{
Ingo Molnardd41f592007-07-09 18:51:59 +02004324 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004325 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004326 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004327
4328 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4329 return;
4330 /*
4331 * We have to be careful, if called from sys_setpriority(),
4332 * the task might be in the middle of scheduling on another CPU.
4333 */
4334 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004335 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004336 /*
4337 * The RT priorities are set via sched_setscheduler(), but we still
4338 * allow the 'normal' nice value to be set - but as expected
4339 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02004340 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004341 */
Ingo Molnare05606d2007-07-09 18:51:59 +02004342 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004343 p->static_prio = NICE_TO_PRIO(nice);
4344 goto out_unlock;
4345 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004346 on_rq = p->se.on_rq;
Srivatsa Vaddagiri58e2d4c2008-01-25 21:08:00 +01004347 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02004348 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004349
Linus Torvalds1da177e2005-04-16 15:20:36 -07004350 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07004351 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004352 old_prio = p->prio;
4353 p->prio = effective_prio(p);
4354 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004355
Ingo Molnardd41f592007-07-09 18:51:59 +02004356 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004357 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004358 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07004359 * If the task increased its priority or is running and
4360 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004361 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07004362 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004363 resched_task(rq->curr);
4364 }
4365out_unlock:
4366 task_rq_unlock(rq, &flags);
4367}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004368EXPORT_SYMBOL(set_user_nice);
4369
Matt Mackalle43379f2005-05-01 08:59:00 -07004370/*
4371 * can_nice - check if a task can reduce its nice value
4372 * @p: task
4373 * @nice: nice value
4374 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004375int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07004376{
Matt Mackall024f4742005-08-18 11:24:19 -07004377 /* convert nice value [19,-20] to rlimit style value [1,40] */
4378 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004379
Matt Mackalle43379f2005-05-01 08:59:00 -07004380 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4381 capable(CAP_SYS_NICE));
4382}
4383
Linus Torvalds1da177e2005-04-16 15:20:36 -07004384#ifdef __ARCH_WANT_SYS_NICE
4385
4386/*
4387 * sys_nice - change the priority of the current process.
4388 * @increment: priority increment
4389 *
4390 * sys_setpriority is a more generic, but much slower function that
4391 * does similar things.
4392 */
4393asmlinkage long sys_nice(int increment)
4394{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004395 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004396
4397 /*
4398 * Setpriority might change our priority at the same moment.
4399 * We don't have to worry. Conceptually one call occurs first
4400 * and we have a single winner.
4401 */
Matt Mackalle43379f2005-05-01 08:59:00 -07004402 if (increment < -40)
4403 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004404 if (increment > 40)
4405 increment = 40;
4406
4407 nice = PRIO_TO_NICE(current->static_prio) + increment;
4408 if (nice < -20)
4409 nice = -20;
4410 if (nice > 19)
4411 nice = 19;
4412
Matt Mackalle43379f2005-05-01 08:59:00 -07004413 if (increment < 0 && !can_nice(current, nice))
4414 return -EPERM;
4415
Linus Torvalds1da177e2005-04-16 15:20:36 -07004416 retval = security_task_setnice(current, nice);
4417 if (retval)
4418 return retval;
4419
4420 set_user_nice(current, nice);
4421 return 0;
4422}
4423
4424#endif
4425
4426/**
4427 * task_prio - return the priority value of a given task.
4428 * @p: the task in question.
4429 *
4430 * This is the priority value as seen by users in /proc.
4431 * RT tasks are offset by -200. Normal tasks are centered
4432 * around 0, value goes from -16 to +15.
4433 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004434int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004435{
4436 return p->prio - MAX_RT_PRIO;
4437}
4438
4439/**
4440 * task_nice - return the nice value of a given task.
4441 * @p: the task in question.
4442 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004443int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004444{
4445 return TASK_NICE(p);
4446}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004447EXPORT_SYMBOL_GPL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004448
4449/**
4450 * idle_cpu - is a given cpu idle currently?
4451 * @cpu: the processor in question.
4452 */
4453int idle_cpu(int cpu)
4454{
4455 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4456}
4457
Linus Torvalds1da177e2005-04-16 15:20:36 -07004458/**
4459 * idle_task - return the idle task for a given cpu.
4460 * @cpu: the processor in question.
4461 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004462struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004463{
4464 return cpu_rq(cpu)->idle;
4465}
4466
4467/**
4468 * find_process_by_pid - find a process with a matching PID value.
4469 * @pid: the pid in question.
4470 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02004471static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004472{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07004473 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004474}
4475
4476/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02004477static void
4478__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479{
Ingo Molnardd41f592007-07-09 18:51:59 +02004480 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004481
Linus Torvalds1da177e2005-04-16 15:20:36 -07004482 p->policy = policy;
Ingo Molnardd41f592007-07-09 18:51:59 +02004483 switch (p->policy) {
4484 case SCHED_NORMAL:
4485 case SCHED_BATCH:
4486 case SCHED_IDLE:
4487 p->sched_class = &fair_sched_class;
4488 break;
4489 case SCHED_FIFO:
4490 case SCHED_RR:
4491 p->sched_class = &rt_sched_class;
4492 break;
4493 }
4494
Linus Torvalds1da177e2005-04-16 15:20:36 -07004495 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004496 p->normal_prio = normal_prio(p);
4497 /* we are holding p->pi_lock already */
4498 p->prio = rt_mutex_getprio(p);
Peter Williams2dd73a42006-06-27 02:54:34 -07004499 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004500}
4501
4502/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004503 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004504 * @p: the task in question.
4505 * @policy: new policy.
4506 * @param: structure containing the new RT priority.
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004507 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004508 * NOTE that the task may be already dead.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004509 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004510int sched_setscheduler(struct task_struct *p, int policy,
4511 struct sched_param *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004512{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004513 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004514 unsigned long flags;
Steven Rostedtcb469842008-01-25 21:08:22 +01004515 const struct sched_class *prev_class = p->sched_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004516 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004517
Steven Rostedt66e53932006-06-27 02:54:44 -07004518 /* may grab non-irq protected spin_locks */
4519 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07004520recheck:
4521 /* double check policy once rq lock held */
4522 if (policy < 0)
4523 policy = oldpolicy = p->policy;
4524 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
Ingo Molnardd41f592007-07-09 18:51:59 +02004525 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4526 policy != SCHED_IDLE)
Ingo Molnarb0a94992006-01-14 13:20:41 -08004527 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004528 /*
4529 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02004530 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4531 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004532 */
4533 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004534 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04004535 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004536 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02004537 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004538 return -EINVAL;
4539
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004540 /*
4541 * Allow unprivileged RT tasks to decrease priority:
4542 */
4543 if (!capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02004544 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004545 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004546
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004547 if (!lock_task_sighand(p, &flags))
4548 return -ESRCH;
4549 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4550 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004551
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004552 /* can't set/change the rt policy */
4553 if (policy != p->policy && !rlim_rtprio)
4554 return -EPERM;
4555
4556 /* can't increase priority */
4557 if (param->sched_priority > p->rt_priority &&
4558 param->sched_priority > rlim_rtprio)
4559 return -EPERM;
4560 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004561 /*
4562 * Like positive nice levels, dont allow tasks to
4563 * move out of SCHED_IDLE either:
4564 */
4565 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4566 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004567
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004568 /* can't change other user's priorities */
4569 if ((current->euid != p->euid) &&
4570 (current->euid != p->uid))
4571 return -EPERM;
4572 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004573
4574 retval = security_task_setscheduler(p, policy, param);
4575 if (retval)
4576 return retval;
4577 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07004578 * make sure no PI-waiters arrive (or leave) while we are
4579 * changing the priority of the task:
4580 */
4581 spin_lock_irqsave(&p->pi_lock, flags);
4582 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004583 * To be able to change p->policy safely, the apropriate
4584 * runqueue lock must be held.
4585 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07004586 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004587 /* recheck policy now with rq lock held */
4588 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4589 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004590 __task_rq_unlock(rq);
4591 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004592 goto recheck;
4593 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02004594 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004595 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004596 running = task_current(rq, p);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004597 if (on_rq) {
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004598 deactivate_task(rq, p, 0);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004599 if (running)
4600 p->sched_class->put_prev_task(rq, p);
4601 }
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02004602
Linus Torvalds1da177e2005-04-16 15:20:36 -07004603 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004604 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02004605
Ingo Molnardd41f592007-07-09 18:51:59 +02004606 if (on_rq) {
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004607 if (running)
4608 p->sched_class->set_curr_task(rq);
Steven Rostedtcb469842008-01-25 21:08:22 +01004609
Ingo Molnardd41f592007-07-09 18:51:59 +02004610 activate_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004611
4612 check_class_changed(rq, p, prev_class, oldprio, running);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004613 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07004614 __task_rq_unlock(rq);
4615 spin_unlock_irqrestore(&p->pi_lock, flags);
4616
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07004617 rt_mutex_adjust_pi(p);
4618
Linus Torvalds1da177e2005-04-16 15:20:36 -07004619 return 0;
4620}
4621EXPORT_SYMBOL_GPL(sched_setscheduler);
4622
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004623static int
4624do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004625{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004626 struct sched_param lparam;
4627 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004628 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004629
4630 if (!param || pid < 0)
4631 return -EINVAL;
4632 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4633 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004634
4635 rcu_read_lock();
4636 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004637 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004638 if (p != NULL)
4639 retval = sched_setscheduler(p, policy, &lparam);
4640 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07004641
Linus Torvalds1da177e2005-04-16 15:20:36 -07004642 return retval;
4643}
4644
4645/**
4646 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4647 * @pid: the pid in question.
4648 * @policy: new policy.
4649 * @param: structure containing the new RT priority.
4650 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004651asmlinkage long
4652sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004653{
Jason Baronc21761f2006-01-18 17:43:03 -08004654 /* negative values for policy are not valid */
4655 if (policy < 0)
4656 return -EINVAL;
4657
Linus Torvalds1da177e2005-04-16 15:20:36 -07004658 return do_sched_setscheduler(pid, policy, param);
4659}
4660
4661/**
4662 * sys_sched_setparam - set/change the RT priority of a thread
4663 * @pid: the pid in question.
4664 * @param: structure containing the new RT priority.
4665 */
4666asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4667{
4668 return do_sched_setscheduler(pid, -1, param);
4669}
4670
4671/**
4672 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4673 * @pid: the pid in question.
4674 */
4675asmlinkage long sys_sched_getscheduler(pid_t pid)
4676{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004677 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004678 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004679
4680 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004681 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004682
4683 retval = -ESRCH;
4684 read_lock(&tasklist_lock);
4685 p = find_process_by_pid(pid);
4686 if (p) {
4687 retval = security_task_getscheduler(p);
4688 if (!retval)
4689 retval = p->policy;
4690 }
4691 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004692 return retval;
4693}
4694
4695/**
4696 * sys_sched_getscheduler - get the RT priority of a thread
4697 * @pid: the pid in question.
4698 * @param: structure containing the RT priority.
4699 */
4700asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4701{
4702 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004703 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004704 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004705
4706 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004707 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004708
4709 read_lock(&tasklist_lock);
4710 p = find_process_by_pid(pid);
4711 retval = -ESRCH;
4712 if (!p)
4713 goto out_unlock;
4714
4715 retval = security_task_getscheduler(p);
4716 if (retval)
4717 goto out_unlock;
4718
4719 lp.sched_priority = p->rt_priority;
4720 read_unlock(&tasklist_lock);
4721
4722 /*
4723 * This one might sleep, we cannot do it with a spinlock held ...
4724 */
4725 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4726
Linus Torvalds1da177e2005-04-16 15:20:36 -07004727 return retval;
4728
4729out_unlock:
4730 read_unlock(&tasklist_lock);
4731 return retval;
4732}
4733
4734long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4735{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004736 cpumask_t cpus_allowed;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004737 struct task_struct *p;
4738 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004739
Gautham R Shenoy95402b32008-01-25 21:08:02 +01004740 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004741 read_lock(&tasklist_lock);
4742
4743 p = find_process_by_pid(pid);
4744 if (!p) {
4745 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01004746 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004747 return -ESRCH;
4748 }
4749
4750 /*
4751 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004752 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07004753 * usage count and then drop tasklist_lock.
4754 */
4755 get_task_struct(p);
4756 read_unlock(&tasklist_lock);
4757
4758 retval = -EPERM;
4759 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4760 !capable(CAP_SYS_NICE))
4761 goto out_unlock;
4762
David Quigleye7834f82006-06-23 02:03:59 -07004763 retval = security_task_setscheduler(p, 0, NULL);
4764 if (retval)
4765 goto out_unlock;
4766
Linus Torvalds1da177e2005-04-16 15:20:36 -07004767 cpus_allowed = cpuset_cpus_allowed(p);
4768 cpus_and(new_mask, new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07004769 again:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004770 retval = set_cpus_allowed(p, new_mask);
4771
Paul Menage8707d8b2007-10-18 23:40:22 -07004772 if (!retval) {
4773 cpus_allowed = cpuset_cpus_allowed(p);
4774 if (!cpus_subset(new_mask, cpus_allowed)) {
4775 /*
4776 * We must have raced with a concurrent cpuset
4777 * update. Just reset the cpus_allowed to the
4778 * cpuset's cpus_allowed
4779 */
4780 new_mask = cpus_allowed;
4781 goto again;
4782 }
4783 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004784out_unlock:
4785 put_task_struct(p);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01004786 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004787 return retval;
4788}
4789
4790static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4791 cpumask_t *new_mask)
4792{
4793 if (len < sizeof(cpumask_t)) {
4794 memset(new_mask, 0, sizeof(cpumask_t));
4795 } else if (len > sizeof(cpumask_t)) {
4796 len = sizeof(cpumask_t);
4797 }
4798 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4799}
4800
4801/**
4802 * sys_sched_setaffinity - set the cpu affinity of a process
4803 * @pid: pid of the process
4804 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4805 * @user_mask_ptr: user-space pointer to the new cpu mask
4806 */
4807asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4808 unsigned long __user *user_mask_ptr)
4809{
4810 cpumask_t new_mask;
4811 int retval;
4812
4813 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4814 if (retval)
4815 return retval;
4816
4817 return sched_setaffinity(pid, new_mask);
4818}
4819
4820/*
4821 * Represents all cpu's present in the system
4822 * In systems capable of hotplug, this map could dynamically grow
4823 * as new cpu's are detected in the system via any platform specific
4824 * method, such as ACPI for e.g.
4825 */
4826
Andi Kleen4cef0c62006-01-11 22:44:57 +01004827cpumask_t cpu_present_map __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004828EXPORT_SYMBOL(cpu_present_map);
4829
4830#ifndef CONFIG_SMP
Andi Kleen4cef0c62006-01-11 22:44:57 +01004831cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07004832EXPORT_SYMBOL(cpu_online_map);
4833
Andi Kleen4cef0c62006-01-11 22:44:57 +01004834cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07004835EXPORT_SYMBOL(cpu_possible_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004836#endif
4837
4838long sched_getaffinity(pid_t pid, cpumask_t *mask)
4839{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004840 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004841 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004842
Gautham R Shenoy95402b32008-01-25 21:08:02 +01004843 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004844 read_lock(&tasklist_lock);
4845
4846 retval = -ESRCH;
4847 p = find_process_by_pid(pid);
4848 if (!p)
4849 goto out_unlock;
4850
David Quigleye7834f82006-06-23 02:03:59 -07004851 retval = security_task_getscheduler(p);
4852 if (retval)
4853 goto out_unlock;
4854
Jack Steiner2f7016d2006-02-01 03:05:18 -08004855 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004856
4857out_unlock:
4858 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01004859 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004860
Ulrich Drepper9531b622007-08-09 11:16:46 +02004861 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004862}
4863
4864/**
4865 * sys_sched_getaffinity - get the cpu affinity of a process
4866 * @pid: pid of the process
4867 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4868 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4869 */
4870asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4871 unsigned long __user *user_mask_ptr)
4872{
4873 int ret;
4874 cpumask_t mask;
4875
4876 if (len < sizeof(cpumask_t))
4877 return -EINVAL;
4878
4879 ret = sched_getaffinity(pid, &mask);
4880 if (ret < 0)
4881 return ret;
4882
4883 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4884 return -EFAULT;
4885
4886 return sizeof(cpumask_t);
4887}
4888
4889/**
4890 * sys_sched_yield - yield the current processor to other threads.
4891 *
Ingo Molnardd41f592007-07-09 18:51:59 +02004892 * This function yields the current CPU to other tasks. If there are no
4893 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004894 */
4895asmlinkage long sys_sched_yield(void)
4896{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004897 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004898
Ingo Molnar2d723762007-10-15 17:00:12 +02004899 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02004900 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004901
4902 /*
4903 * Since we are going to call schedule() anyway, there's
4904 * no need to preempt or enable interrupts:
4905 */
4906 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07004907 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004908 _raw_spin_unlock(&rq->lock);
4909 preempt_enable_no_resched();
4910
4911 schedule();
4912
4913 return 0;
4914}
4915
Andrew Mortone7b38402006-06-30 01:56:00 -07004916static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004917{
Ingo Molnar8e0a43d2006-06-23 02:05:23 -07004918#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4919 __might_sleep(__FILE__, __LINE__);
4920#endif
Ingo Molnar5bbcfd92005-07-07 17:57:04 -07004921 /*
4922 * The BKS might be reacquired before we have dropped
4923 * PREEMPT_ACTIVE, which could trigger a second
4924 * cond_resched() call.
4925 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004926 do {
4927 add_preempt_count(PREEMPT_ACTIVE);
4928 schedule();
4929 sub_preempt_count(PREEMPT_ACTIVE);
4930 } while (need_resched());
4931}
4932
Herbert Xu02b67cc2008-01-25 21:08:28 +01004933#if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
4934int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004935{
Ingo Molnar94142322006-12-29 16:48:13 -08004936 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4937 system_state == SYSTEM_RUNNING) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004938 __cond_resched();
4939 return 1;
4940 }
4941 return 0;
4942}
Herbert Xu02b67cc2008-01-25 21:08:28 +01004943EXPORT_SYMBOL(_cond_resched);
4944#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004945
4946/*
4947 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4948 * call schedule, and on return reacquire the lock.
4949 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004950 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07004951 * operations here to prevent schedule() from being called twice (once via
4952 * spin_unlock(), once by hand).
4953 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004954int cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004955{
Nick Piggin95c354f2008-01-30 13:31:20 +01004956 int resched = need_resched() && system_state == SYSTEM_RUNNING;
Jan Kara6df3cec2005-06-13 15:52:32 -07004957 int ret = 0;
4958
Nick Piggin95c354f2008-01-30 13:31:20 +01004959 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004960 spin_unlock(lock);
Nick Piggin95c354f2008-01-30 13:31:20 +01004961 if (resched && need_resched())
4962 __cond_resched();
4963 else
4964 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07004965 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004966 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004967 }
Jan Kara6df3cec2005-06-13 15:52:32 -07004968 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004969}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004970EXPORT_SYMBOL(cond_resched_lock);
4971
4972int __sched cond_resched_softirq(void)
4973{
4974 BUG_ON(!in_softirq());
4975
Ingo Molnar94142322006-12-29 16:48:13 -08004976 if (need_resched() && system_state == SYSTEM_RUNNING) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07004977 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004978 __cond_resched();
4979 local_bh_disable();
4980 return 1;
4981 }
4982 return 0;
4983}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004984EXPORT_SYMBOL(cond_resched_softirq);
4985
Linus Torvalds1da177e2005-04-16 15:20:36 -07004986/**
4987 * yield - yield the current processor to other threads.
4988 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004989 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07004990 * thread runnable and calls sys_sched_yield().
4991 */
4992void __sched yield(void)
4993{
4994 set_current_state(TASK_RUNNING);
4995 sys_sched_yield();
4996}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004997EXPORT_SYMBOL(yield);
4998
4999/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005000 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07005001 * that process accounting knows that this is a task in IO wait state.
5002 *
5003 * But don't do that if it is a deliberate, throttling IO wait (this task
5004 * has set its backing_dev_info: the queue against which it should throttle)
5005 */
5006void __sched io_schedule(void)
5007{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005008 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005009
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005010 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005011 atomic_inc(&rq->nr_iowait);
5012 schedule();
5013 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005014 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005015}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005016EXPORT_SYMBOL(io_schedule);
5017
5018long __sched io_schedule_timeout(long timeout)
5019{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005020 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005021 long ret;
5022
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005023 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005024 atomic_inc(&rq->nr_iowait);
5025 ret = schedule_timeout(timeout);
5026 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005027 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005028 return ret;
5029}
5030
5031/**
5032 * sys_sched_get_priority_max - return maximum RT priority.
5033 * @policy: scheduling class.
5034 *
5035 * this syscall returns the maximum rt_priority that can be used
5036 * by a given scheduling class.
5037 */
5038asmlinkage long sys_sched_get_priority_max(int policy)
5039{
5040 int ret = -EINVAL;
5041
5042 switch (policy) {
5043 case SCHED_FIFO:
5044 case SCHED_RR:
5045 ret = MAX_USER_RT_PRIO-1;
5046 break;
5047 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005048 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005049 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005050 ret = 0;
5051 break;
5052 }
5053 return ret;
5054}
5055
5056/**
5057 * sys_sched_get_priority_min - return minimum RT priority.
5058 * @policy: scheduling class.
5059 *
5060 * this syscall returns the minimum rt_priority that can be used
5061 * by a given scheduling class.
5062 */
5063asmlinkage long sys_sched_get_priority_min(int policy)
5064{
5065 int ret = -EINVAL;
5066
5067 switch (policy) {
5068 case SCHED_FIFO:
5069 case SCHED_RR:
5070 ret = 1;
5071 break;
5072 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005073 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005074 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005075 ret = 0;
5076 }
5077 return ret;
5078}
5079
5080/**
5081 * sys_sched_rr_get_interval - return the default timeslice of a process.
5082 * @pid: pid of the process.
5083 * @interval: userspace pointer to the timeslice value.
5084 *
5085 * this syscall writes the default timeslice value of a given process
5086 * into the user-space timespec buffer. A value of '0' means infinity.
5087 */
5088asmlinkage
5089long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5090{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005091 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005092 unsigned int time_slice;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005093 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005094 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005095
5096 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005097 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005098
5099 retval = -ESRCH;
5100 read_lock(&tasklist_lock);
5101 p = find_process_by_pid(pid);
5102 if (!p)
5103 goto out_unlock;
5104
5105 retval = security_task_getscheduler(p);
5106 if (retval)
5107 goto out_unlock;
5108
Ingo Molnar77034932007-12-04 17:04:39 +01005109 /*
5110 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5111 * tasks that are on an otherwise idle runqueue:
5112 */
5113 time_slice = 0;
5114 if (p->policy == SCHED_RR) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005115 time_slice = DEF_TIMESLICE;
Ingo Molnar77034932007-12-04 17:04:39 +01005116 } else {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005117 struct sched_entity *se = &p->se;
5118 unsigned long flags;
5119 struct rq *rq;
5120
5121 rq = task_rq_lock(p, &flags);
Ingo Molnar77034932007-12-04 17:04:39 +01005122 if (rq->cfs.load.weight)
5123 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005124 task_rq_unlock(rq, &flags);
5125 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005126 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005127 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005128 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005129 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005130
Linus Torvalds1da177e2005-04-16 15:20:36 -07005131out_unlock:
5132 read_unlock(&tasklist_lock);
5133 return retval;
5134}
5135
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005136static const char stat_nam[] = "RSDTtZX";
Ingo Molnar36c8b582006-07-03 00:25:41 -07005137
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005138void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005139{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005140 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005141 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005142
Linus Torvalds1da177e2005-04-16 15:20:36 -07005143 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005144 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005145 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02005146#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07005147 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005148 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005149 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005150 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005151#else
5152 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005153 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005154 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005155 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005156#endif
5157#ifdef CONFIG_DEBUG_STACK_USAGE
5158 {
Al Viro10ebffd2005-11-13 16:06:56 -08005159 unsigned long *n = end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005160 while (!*n)
5161 n++;
Al Viro10ebffd2005-11-13 16:06:56 -08005162 free = (unsigned long)n - (unsigned long)end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005163 }
5164#endif
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -07005165 printk(KERN_CONT "%5lu %5d %6d\n", free,
Roland McGrathfcfd50a2008-01-09 00:03:23 -08005166 task_pid_nr(p), task_pid_nr(p->real_parent));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005167
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01005168 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005169}
5170
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005171void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005172{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005173 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005174
Ingo Molnar4bd77322007-07-11 21:21:47 +02005175#if BITS_PER_LONG == 32
5176 printk(KERN_INFO
5177 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005178#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02005179 printk(KERN_INFO
5180 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005181#endif
5182 read_lock(&tasklist_lock);
5183 do_each_thread(g, p) {
5184 /*
5185 * reset the NMI-timeout, listing all files on a slow
5186 * console might take alot of time:
5187 */
5188 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07005189 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005190 sched_show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005191 } while_each_thread(g, p);
5192
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07005193 touch_all_softlockup_watchdogs();
5194
Ingo Molnardd41f592007-07-09 18:51:59 +02005195#ifdef CONFIG_SCHED_DEBUG
5196 sysrq_sched_debug_show();
5197#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005198 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005199 /*
5200 * Only show locks if all tasks are dumped:
5201 */
5202 if (state_filter == -1)
5203 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005204}
5205
Ingo Molnar1df21052007-07-09 18:51:58 +02005206void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5207{
Ingo Molnardd41f592007-07-09 18:51:59 +02005208 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02005209}
5210
Ingo Molnarf340c0d2005-06-28 16:40:42 +02005211/**
5212 * init_idle - set up an idle thread for a given CPU
5213 * @idle: task in question
5214 * @cpu: cpu the idle task belongs to
5215 *
5216 * NOTE: this function does not set the idle thread's NEED_RESCHED
5217 * flag, to make booting more robust.
5218 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07005219void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005220{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005221 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005222 unsigned long flags;
5223
Ingo Molnardd41f592007-07-09 18:51:59 +02005224 __sched_fork(idle);
5225 idle->se.exec_start = sched_clock();
5226
Ingo Molnarb29739f2006-06-27 02:54:51 -07005227 idle->prio = idle->normal_prio = MAX_PRIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005228 idle->cpus_allowed = cpumask_of_cpu(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005229 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005230
5231 spin_lock_irqsave(&rq->lock, flags);
5232 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07005233#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5234 idle->oncpu = 1;
5235#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005236 spin_unlock_irqrestore(&rq->lock, flags);
5237
5238 /* Set the preempt count _outside_ the spinlocks! */
Al Viroa1261f52005-11-13 16:06:55 -08005239 task_thread_info(idle)->preempt_count = 0;
Ingo Molnar6478d882008-01-25 21:08:33 +01005240
Ingo Molnardd41f592007-07-09 18:51:59 +02005241 /*
5242 * The idle tasks have their own, simple scheduling class:
5243 */
5244 idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005245}
5246
5247/*
5248 * In a system that switches off the HZ timer nohz_cpu_mask
5249 * indicates which cpus entered this state. This is used
5250 * in the rcu update to wait only for active cpus. For system
5251 * which do not switch off the HZ timer nohz_cpu_mask should
5252 * always be CPU_MASK_NONE.
5253 */
5254cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5255
Ingo Molnar19978ca2007-11-09 22:39:38 +01005256/*
5257 * Increase the granularity value when there are more CPUs,
5258 * because with more CPUs the 'effective latency' as visible
5259 * to users decreases. But the relationship is not linear,
5260 * so pick a second-best guess by going with the log2 of the
5261 * number of CPUs.
5262 *
5263 * This idea comes from the SD scheduler of Con Kolivas:
5264 */
5265static inline void sched_init_granularity(void)
5266{
5267 unsigned int factor = 1 + ilog2(num_online_cpus());
5268 const unsigned long limit = 200000000;
5269
5270 sysctl_sched_min_granularity *= factor;
5271 if (sysctl_sched_min_granularity > limit)
5272 sysctl_sched_min_granularity = limit;
5273
5274 sysctl_sched_latency *= factor;
5275 if (sysctl_sched_latency > limit)
5276 sysctl_sched_latency = limit;
5277
5278 sysctl_sched_wakeup_granularity *= factor;
5279 sysctl_sched_batch_wakeup_granularity *= factor;
5280}
5281
Linus Torvalds1da177e2005-04-16 15:20:36 -07005282#ifdef CONFIG_SMP
5283/*
5284 * This is how migration works:
5285 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07005286 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005287 * runqueue and wake up that CPU's migration thread.
5288 * 2) we down() the locked semaphore => thread blocks.
5289 * 3) migration thread wakes up (implicitly it forces the migrated
5290 * thread off the CPU)
5291 * 4) it gets the migration request and checks whether the migrated
5292 * task is still in the wrong runqueue.
5293 * 5) if it's in the wrong runqueue then the migration thread removes
5294 * it and puts it into the right queue.
5295 * 6) migration thread up()s the semaphore.
5296 * 7) we wake up and the migration is done.
5297 */
5298
5299/*
5300 * Change a given task's CPU affinity. Migrate the thread to a
5301 * proper CPU and schedule it away if the CPU it's executing on
5302 * is removed from the allowed bitmask.
5303 *
5304 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005305 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07005306 * call is not atomic; no spinlocks may be held.
5307 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005308int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005309{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005310 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005311 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005312 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005313 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005314
5315 rq = task_rq_lock(p, &flags);
5316 if (!cpus_intersects(new_mask, cpu_online_map)) {
5317 ret = -EINVAL;
5318 goto out;
5319 }
5320
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005321 if (p->sched_class->set_cpus_allowed)
5322 p->sched_class->set_cpus_allowed(p, &new_mask);
5323 else {
Ingo Molnar0eab9142008-01-25 21:08:19 +01005324 p->cpus_allowed = new_mask;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01005325 p->rt.nr_cpus_allowed = cpus_weight(new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005326 }
5327
Linus Torvalds1da177e2005-04-16 15:20:36 -07005328 /* Can the task run on the task's current CPU? If so, we're done */
5329 if (cpu_isset(task_cpu(p), new_mask))
5330 goto out;
5331
5332 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5333 /* Need help from migration thread: drop lock and wait. */
5334 task_rq_unlock(rq, &flags);
5335 wake_up_process(rq->migration_thread);
5336 wait_for_completion(&req.done);
5337 tlb_migrate_finish(p->mm);
5338 return 0;
5339 }
5340out:
5341 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005342
Linus Torvalds1da177e2005-04-16 15:20:36 -07005343 return ret;
5344}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005345EXPORT_SYMBOL_GPL(set_cpus_allowed);
5346
5347/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005348 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07005349 * this because either it can't run here any more (set_cpus_allowed()
5350 * away from this CPU, or CPU going down), or because we're
5351 * attempting to rebalance this task on exec (sched_exec).
5352 *
5353 * So we race with normal scheduler movements, but that's OK, as long
5354 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07005355 *
5356 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005357 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07005358static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005359{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005360 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02005361 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005362
5363 if (unlikely(cpu_is_offline(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07005364 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005365
5366 rq_src = cpu_rq(src_cpu);
5367 rq_dest = cpu_rq(dest_cpu);
5368
5369 double_rq_lock(rq_src, rq_dest);
5370 /* Already moved. */
5371 if (task_cpu(p) != src_cpu)
5372 goto out;
5373 /* Affinity changed (again). */
5374 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5375 goto out;
5376
Ingo Molnardd41f592007-07-09 18:51:59 +02005377 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02005378 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005379 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02005380
Linus Torvalds1da177e2005-04-16 15:20:36 -07005381 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005382 if (on_rq) {
5383 activate_task(rq_dest, p, 0);
5384 check_preempt_curr(rq_dest, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005385 }
Kirill Korotaevefc30812006-06-27 02:54:32 -07005386 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005387out:
5388 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07005389 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005390}
5391
5392/*
5393 * migration_thread - this is a highprio system thread that performs
5394 * thread migration by bumping thread off CPU then 'pushing' onto
5395 * another runqueue.
5396 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005397static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005398{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005399 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005400 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005401
5402 rq = cpu_rq(cpu);
5403 BUG_ON(rq->migration_thread != current);
5404
5405 set_current_state(TASK_INTERRUPTIBLE);
5406 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005407 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005408 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005409
Linus Torvalds1da177e2005-04-16 15:20:36 -07005410 spin_lock_irq(&rq->lock);
5411
5412 if (cpu_is_offline(cpu)) {
5413 spin_unlock_irq(&rq->lock);
5414 goto wait_to_die;
5415 }
5416
5417 if (rq->active_balance) {
5418 active_load_balance(rq, cpu);
5419 rq->active_balance = 0;
5420 }
5421
5422 head = &rq->migration_queue;
5423
5424 if (list_empty(head)) {
5425 spin_unlock_irq(&rq->lock);
5426 schedule();
5427 set_current_state(TASK_INTERRUPTIBLE);
5428 continue;
5429 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07005430 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005431 list_del_init(head->next);
5432
Nick Piggin674311d2005-06-25 14:57:27 -07005433 spin_unlock(&rq->lock);
5434 __migrate_task(req->task, cpu, req->dest_cpu);
5435 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005436
5437 complete(&req->done);
5438 }
5439 __set_current_state(TASK_RUNNING);
5440 return 0;
5441
5442wait_to_die:
5443 /* Wait for kthread_stop */
5444 set_current_state(TASK_INTERRUPTIBLE);
5445 while (!kthread_should_stop()) {
5446 schedule();
5447 set_current_state(TASK_INTERRUPTIBLE);
5448 }
5449 __set_current_state(TASK_RUNNING);
5450 return 0;
5451}
5452
5453#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005454
5455static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5456{
5457 int ret;
5458
5459 local_irq_disable();
5460 ret = __migrate_task(p, src_cpu, dest_cpu);
5461 local_irq_enable();
5462 return ret;
5463}
5464
Kirill Korotaev054b9102006-12-10 02:20:11 -08005465/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02005466 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08005467 * NOTE: interrupts should be disabled by the caller
5468 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005469static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005470{
Kirill Korotaevefc30812006-06-27 02:54:32 -07005471 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005472 cpumask_t mask;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005473 struct rq *rq;
5474 int dest_cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005475
Andi Kleen3a5c3592007-10-15 17:00:14 +02005476 do {
5477 /* On same node? */
5478 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5479 cpus_and(mask, mask, p->cpus_allowed);
5480 dest_cpu = any_online_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005481
Andi Kleen3a5c3592007-10-15 17:00:14 +02005482 /* On any allowed CPU? */
5483 if (dest_cpu == NR_CPUS)
5484 dest_cpu = any_online_cpu(p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005485
Andi Kleen3a5c3592007-10-15 17:00:14 +02005486 /* No more Mr. Nice Guy. */
5487 if (dest_cpu == NR_CPUS) {
Cliff Wickman470fd642007-10-18 23:40:46 -07005488 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5489 /*
5490 * Try to stay on the same cpuset, where the
5491 * current cpuset may be a subset of all cpus.
5492 * The cpuset_cpus_allowed_locked() variant of
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005493 * cpuset_cpus_allowed() will not block. It must be
Cliff Wickman470fd642007-10-18 23:40:46 -07005494 * called within calls to cpuset_lock/cpuset_unlock.
5495 */
Andi Kleen3a5c3592007-10-15 17:00:14 +02005496 rq = task_rq_lock(p, &flags);
Cliff Wickman470fd642007-10-18 23:40:46 -07005497 p->cpus_allowed = cpus_allowed;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005498 dest_cpu = any_online_cpu(p->cpus_allowed);
5499 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005500
Andi Kleen3a5c3592007-10-15 17:00:14 +02005501 /*
5502 * Don't tell them about moving exiting tasks or
5503 * kernel threads (both mm NULL), since they never
5504 * leave kernel.
5505 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005506 if (p->mm && printk_ratelimit()) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02005507 printk(KERN_INFO "process %d (%s) no "
5508 "longer affine to cpu%d\n",
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005509 task_pid_nr(p), p->comm, dead_cpu);
5510 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02005511 }
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005512 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005513}
5514
5515/*
5516 * While a dead CPU has no uninterruptible tasks queued at this point,
5517 * it might still have a nonzero ->nr_uninterruptible counter, because
5518 * for performance reasons the counter is not stricly tracking tasks to
5519 * their home CPUs. So we just add the counter to another CPU's counter,
5520 * to keep the global sum constant after CPU-down:
5521 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07005522static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005523{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005524 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005525 unsigned long flags;
5526
5527 local_irq_save(flags);
5528 double_rq_lock(rq_src, rq_dest);
5529 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5530 rq_src->nr_uninterruptible = 0;
5531 double_rq_unlock(rq_src, rq_dest);
5532 local_irq_restore(flags);
5533}
5534
5535/* Run through task list and migrate tasks from the dead cpu. */
5536static void migrate_live_tasks(int src_cpu)
5537{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005538 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005539
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005540 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005541
Ingo Molnar48f24c42006-07-03 00:25:40 -07005542 do_each_thread(t, p) {
5543 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005544 continue;
5545
Ingo Molnar48f24c42006-07-03 00:25:40 -07005546 if (task_cpu(p) == src_cpu)
5547 move_task_off_dead_cpu(src_cpu, p);
5548 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005549
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005550 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005551}
5552
Ingo Molnardd41f592007-07-09 18:51:59 +02005553/*
5554 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01005555 * It does so by boosting its priority to highest possible.
5556 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005557 */
5558void sched_idle_next(void)
5559{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005560 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07005561 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005562 struct task_struct *p = rq->idle;
5563 unsigned long flags;
5564
5565 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005566 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005567
Ingo Molnar48f24c42006-07-03 00:25:40 -07005568 /*
5569 * Strictly not necessary since rest of the CPUs are stopped by now
5570 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005571 */
5572 spin_lock_irqsave(&rq->lock, flags);
5573
Ingo Molnardd41f592007-07-09 18:51:59 +02005574 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005575
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01005576 update_rq_clock(rq);
5577 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005578
5579 spin_unlock_irqrestore(&rq->lock, flags);
5580}
5581
Ingo Molnar48f24c42006-07-03 00:25:40 -07005582/*
5583 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07005584 * offline.
5585 */
5586void idle_task_exit(void)
5587{
5588 struct mm_struct *mm = current->active_mm;
5589
5590 BUG_ON(cpu_online(smp_processor_id()));
5591
5592 if (mm != &init_mm)
5593 switch_mm(mm, &init_mm, current);
5594 mmdrop(mm);
5595}
5596
Kirill Korotaev054b9102006-12-10 02:20:11 -08005597/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005598static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005599{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005600 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005601
5602 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07005603 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005604
5605 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07005606 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005607
Ingo Molnar48f24c42006-07-03 00:25:40 -07005608 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005609
5610 /*
5611 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005612 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005613 * fine.
5614 */
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005615 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005616 move_task_off_dead_cpu(dead_cpu, p);
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005617 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005618
Ingo Molnar48f24c42006-07-03 00:25:40 -07005619 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005620}
5621
5622/* release_task() removes task from tasklist, so we won't find dead tasks. */
5623static void migrate_dead_tasks(unsigned int dead_cpu)
5624{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005625 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005626 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005627
Ingo Molnardd41f592007-07-09 18:51:59 +02005628 for ( ; ; ) {
5629 if (!rq->nr_running)
5630 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02005631 update_rq_clock(rq);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02005632 next = pick_next_task(rq, rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02005633 if (!next)
5634 break;
5635 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02005636
Linus Torvalds1da177e2005-04-16 15:20:36 -07005637 }
5638}
5639#endif /* CONFIG_HOTPLUG_CPU */
5640
Nick Piggine692ab52007-07-26 13:40:43 +02005641#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5642
5643static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005644 {
5645 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005646 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005647 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01005648 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02005649};
5650
5651static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005652 {
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005653 .ctl_name = CTL_KERN,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005654 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005655 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005656 .child = sd_ctl_dir,
5657 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01005658 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02005659};
5660
5661static struct ctl_table *sd_alloc_ctl_entry(int n)
5662{
5663 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02005664 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02005665
Nick Piggine692ab52007-07-26 13:40:43 +02005666 return entry;
5667}
5668
Milton Miller6382bc92007-10-15 17:00:19 +02005669static void sd_free_ctl_entry(struct ctl_table **tablep)
5670{
Milton Millercd790072007-10-17 16:55:11 +02005671 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02005672
Milton Millercd790072007-10-17 16:55:11 +02005673 /*
5674 * In the intermediate directories, both the child directory and
5675 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005676 * will always be set. In the lowest directory the names are
Milton Millercd790072007-10-17 16:55:11 +02005677 * static strings and all have proc handlers.
5678 */
5679 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02005680 if (entry->child)
5681 sd_free_ctl_entry(&entry->child);
Milton Millercd790072007-10-17 16:55:11 +02005682 if (entry->proc_handler == NULL)
5683 kfree(entry->procname);
5684 }
Milton Miller6382bc92007-10-15 17:00:19 +02005685
5686 kfree(*tablep);
5687 *tablep = NULL;
5688}
5689
Nick Piggine692ab52007-07-26 13:40:43 +02005690static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02005691set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02005692 const char *procname, void *data, int maxlen,
5693 mode_t mode, proc_handler *proc_handler)
5694{
Nick Piggine692ab52007-07-26 13:40:43 +02005695 entry->procname = procname;
5696 entry->data = data;
5697 entry->maxlen = maxlen;
5698 entry->mode = mode;
5699 entry->proc_handler = proc_handler;
5700}
5701
5702static struct ctl_table *
5703sd_alloc_ctl_domain_table(struct sched_domain *sd)
5704{
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005705 struct ctl_table *table = sd_alloc_ctl_entry(12);
Nick Piggine692ab52007-07-26 13:40:43 +02005706
Milton Millerad1cdc12007-10-15 17:00:19 +02005707 if (table == NULL)
5708 return NULL;
5709
Alexey Dobriyane0361852007-08-09 11:16:46 +02005710 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02005711 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005712 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02005713 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005714 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005715 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005716 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005717 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005718 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005719 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005720 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005721 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005722 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005723 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005724 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02005725 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005726 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02005727 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005728 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02005729 &sd->cache_nice_tries,
5730 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005731 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02005732 sizeof(int), 0644, proc_dointvec_minmax);
Milton Miller6323469f2007-10-15 17:00:19 +02005733 /* &table[11] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02005734
5735 return table;
5736}
5737
Ingo Molnar9a4e7152007-11-28 15:52:56 +01005738static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02005739{
5740 struct ctl_table *entry, *table;
5741 struct sched_domain *sd;
5742 int domain_num = 0, i;
5743 char buf[32];
5744
5745 for_each_domain(cpu, sd)
5746 domain_num++;
5747 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02005748 if (table == NULL)
5749 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02005750
5751 i = 0;
5752 for_each_domain(cpu, sd) {
5753 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02005754 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005755 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02005756 entry->child = sd_alloc_ctl_domain_table(sd);
5757 entry++;
5758 i++;
5759 }
5760 return table;
5761}
5762
5763static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02005764static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02005765{
5766 int i, cpu_num = num_online_cpus();
5767 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5768 char buf[32];
5769
Milton Miller73785472007-10-24 18:23:48 +02005770 WARN_ON(sd_ctl_dir[0].child);
5771 sd_ctl_dir[0].child = entry;
5772
Milton Millerad1cdc12007-10-15 17:00:19 +02005773 if (entry == NULL)
5774 return;
5775
Milton Miller97b6ea72007-10-15 17:00:19 +02005776 for_each_online_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02005777 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02005778 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005779 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02005780 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02005781 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02005782 }
Milton Miller73785472007-10-24 18:23:48 +02005783
5784 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02005785 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5786}
Milton Miller6382bc92007-10-15 17:00:19 +02005787
Milton Miller73785472007-10-24 18:23:48 +02005788/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02005789static void unregister_sched_domain_sysctl(void)
5790{
Milton Miller73785472007-10-24 18:23:48 +02005791 if (sd_sysctl_header)
5792 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02005793 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02005794 if (sd_ctl_dir[0].child)
5795 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02005796}
Nick Piggine692ab52007-07-26 13:40:43 +02005797#else
Milton Miller6382bc92007-10-15 17:00:19 +02005798static void register_sched_domain_sysctl(void)
5799{
5800}
5801static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02005802{
5803}
5804#endif
5805
Linus Torvalds1da177e2005-04-16 15:20:36 -07005806/*
5807 * migration_call - callback that gets triggered when a CPU is added.
5808 * Here we can start up the necessary migration thread for the new CPU.
5809 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005810static int __cpuinit
5811migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005812{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005813 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005814 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005815 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005816 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005817
5818 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07005819
Linus Torvalds1da177e2005-04-16 15:20:36 -07005820 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005821 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02005822 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005823 if (IS_ERR(p))
5824 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005825 kthread_bind(p, cpu);
5826 /* Must be high prio: stop_machine expects to yield to it. */
5827 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02005828 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005829 task_rq_unlock(rq, &flags);
5830 cpu_rq(cpu)->migration_thread = p;
5831 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005832
Linus Torvalds1da177e2005-04-16 15:20:36 -07005833 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005834 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02005835 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005836 wake_up_process(cpu_rq(cpu)->migration_thread);
Gregory Haskins57d885f2008-01-25 21:08:18 +01005837
5838 /* Update our root-domain */
5839 rq = cpu_rq(cpu);
5840 spin_lock_irqsave(&rq->lock, flags);
5841 if (rq->rd) {
5842 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5843 cpu_set(cpu, rq->rd->online);
5844 }
5845 spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005846 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005847
Linus Torvalds1da177e2005-04-16 15:20:36 -07005848#ifdef CONFIG_HOTPLUG_CPU
5849 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005850 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07005851 if (!cpu_rq(cpu)->migration_thread)
5852 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005853 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08005854 kthread_bind(cpu_rq(cpu)->migration_thread,
5855 any_online_cpu(cpu_online_map));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005856 kthread_stop(cpu_rq(cpu)->migration_thread);
5857 cpu_rq(cpu)->migration_thread = NULL;
5858 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005859
Linus Torvalds1da177e2005-04-16 15:20:36 -07005860 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005861 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07005862 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005863 migrate_live_tasks(cpu);
5864 rq = cpu_rq(cpu);
5865 kthread_stop(rq->migration_thread);
5866 rq->migration_thread = NULL;
5867 /* Idle task back to normal (off runqueue, low prio) */
Oleg Nesterovd2da2722007-10-16 23:30:56 -07005868 spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02005869 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005870 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005871 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02005872 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5873 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005874 migrate_dead_tasks(cpu);
Oleg Nesterovd2da2722007-10-16 23:30:56 -07005875 spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07005876 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005877 migrate_nr_uninterruptible(rq);
5878 BUG_ON(rq->nr_running != 0);
5879
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005880 /*
5881 * No need to migrate the tasks: it was best-effort if
5882 * they didn't take sched_hotcpu_mutex. Just wake up
5883 * the requestors.
5884 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005885 spin_lock_irq(&rq->lock);
5886 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005887 struct migration_req *req;
5888
Linus Torvalds1da177e2005-04-16 15:20:36 -07005889 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07005890 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005891 list_del_init(&req->list);
5892 complete(&req->done);
5893 }
5894 spin_unlock_irq(&rq->lock);
5895 break;
Gregory Haskins57d885f2008-01-25 21:08:18 +01005896
5897 case CPU_DOWN_PREPARE:
5898 /* Update our root-domain */
5899 rq = cpu_rq(cpu);
5900 spin_lock_irqsave(&rq->lock, flags);
5901 if (rq->rd) {
5902 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5903 cpu_clear(cpu, rq->rd->online);
5904 }
5905 spin_unlock_irqrestore(&rq->lock, flags);
5906 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005907#endif
5908 }
5909 return NOTIFY_OK;
5910}
5911
5912/* Register at highest priority so that task migration (migrate_all_tasks)
5913 * happens before everything else.
5914 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07005915static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005916 .notifier_call = migration_call,
5917 .priority = 10
5918};
5919
Adrian Bunke6fe6642007-11-09 22:39:39 +01005920void __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005921{
5922 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07005923 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005924
5925 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07005926 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5927 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005928 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5929 register_cpu_notifier(&migration_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005930}
5931#endif
5932
5933#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07005934
5935/* Number of possible processor ids */
5936int nr_cpu_ids __read_mostly = NR_CPUS;
5937EXPORT_SYMBOL(nr_cpu_ids);
5938
Ingo Molnar3e9830d2007-10-15 17:00:13 +02005939#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005940
5941static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5942{
5943 struct sched_group *group = sd->groups;
5944 cpumask_t groupmask;
5945 char str[NR_CPUS];
5946
5947 cpumask_scnprintf(str, NR_CPUS, sd->span);
5948 cpus_clear(groupmask);
5949
5950 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5951
5952 if (!(sd->flags & SD_LOAD_BALANCE)) {
5953 printk("does not load-balance\n");
5954 if (sd->parent)
5955 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5956 " has parent");
5957 return -1;
5958 }
5959
5960 printk(KERN_CONT "span %s\n", str);
5961
5962 if (!cpu_isset(cpu, sd->span)) {
5963 printk(KERN_ERR "ERROR: domain->span does not contain "
5964 "CPU%d\n", cpu);
5965 }
5966 if (!cpu_isset(cpu, group->cpumask)) {
5967 printk(KERN_ERR "ERROR: domain->groups does not contain"
5968 " CPU%d\n", cpu);
5969 }
5970
5971 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5972 do {
5973 if (!group) {
5974 printk("\n");
5975 printk(KERN_ERR "ERROR: group is NULL\n");
5976 break;
5977 }
5978
5979 if (!group->__cpu_power) {
5980 printk(KERN_CONT "\n");
5981 printk(KERN_ERR "ERROR: domain->cpu_power not "
5982 "set\n");
5983 break;
5984 }
5985
5986 if (!cpus_weight(group->cpumask)) {
5987 printk(KERN_CONT "\n");
5988 printk(KERN_ERR "ERROR: empty group\n");
5989 break;
5990 }
5991
5992 if (cpus_intersects(groupmask, group->cpumask)) {
5993 printk(KERN_CONT "\n");
5994 printk(KERN_ERR "ERROR: repeated CPUs\n");
5995 break;
5996 }
5997
5998 cpus_or(groupmask, groupmask, group->cpumask);
5999
6000 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
6001 printk(KERN_CONT " %s", str);
6002
6003 group = group->next;
6004 } while (group != sd->groups);
6005 printk(KERN_CONT "\n");
6006
6007 if (!cpus_equal(sd->span, groupmask))
6008 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6009
6010 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
6011 printk(KERN_ERR "ERROR: parent span is not a superset "
6012 "of domain->span\n");
6013 return 0;
6014}
6015
Linus Torvalds1da177e2005-04-16 15:20:36 -07006016static void sched_domain_debug(struct sched_domain *sd, int cpu)
6017{
6018 int level = 0;
6019
Nick Piggin41c7ce92005-06-25 14:57:24 -07006020 if (!sd) {
6021 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6022 return;
6023 }
6024
Linus Torvalds1da177e2005-04-16 15:20:36 -07006025 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6026
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006027 for (;;) {
6028 if (sched_domain_debug_one(sd, cpu, level))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006029 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006030 level++;
6031 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08006032 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006033 break;
6034 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006035}
6036#else
Ingo Molnar48f24c42006-07-03 00:25:40 -07006037# define sched_domain_debug(sd, cpu) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006038#endif
6039
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006040static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006041{
6042 if (cpus_weight(sd->span) == 1)
6043 return 1;
6044
6045 /* Following flags need at least 2 groups */
6046 if (sd->flags & (SD_LOAD_BALANCE |
6047 SD_BALANCE_NEWIDLE |
6048 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006049 SD_BALANCE_EXEC |
6050 SD_SHARE_CPUPOWER |
6051 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006052 if (sd->groups != sd->groups->next)
6053 return 0;
6054 }
6055
6056 /* Following flags don't use groups */
6057 if (sd->flags & (SD_WAKE_IDLE |
6058 SD_WAKE_AFFINE |
6059 SD_WAKE_BALANCE))
6060 return 0;
6061
6062 return 1;
6063}
6064
Ingo Molnar48f24c42006-07-03 00:25:40 -07006065static int
6066sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006067{
6068 unsigned long cflags = sd->flags, pflags = parent->flags;
6069
6070 if (sd_degenerate(parent))
6071 return 1;
6072
6073 if (!cpus_equal(sd->span, parent->span))
6074 return 0;
6075
6076 /* Does parent contain flags not in child? */
6077 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6078 if (cflags & SD_WAKE_AFFINE)
6079 pflags &= ~SD_WAKE_BALANCE;
6080 /* Flags needing groups don't count if only 1 group in parent */
6081 if (parent->groups == parent->groups->next) {
6082 pflags &= ~(SD_LOAD_BALANCE |
6083 SD_BALANCE_NEWIDLE |
6084 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006085 SD_BALANCE_EXEC |
6086 SD_SHARE_CPUPOWER |
6087 SD_SHARE_PKG_RESOURCES);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006088 }
6089 if (~cflags & pflags)
6090 return 0;
6091
6092 return 1;
6093}
6094
Gregory Haskins57d885f2008-01-25 21:08:18 +01006095static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6096{
6097 unsigned long flags;
6098 const struct sched_class *class;
6099
6100 spin_lock_irqsave(&rq->lock, flags);
6101
6102 if (rq->rd) {
6103 struct root_domain *old_rd = rq->rd;
6104
Ingo Molnar0eab9142008-01-25 21:08:19 +01006105 for (class = sched_class_highest; class; class = class->next) {
Gregory Haskins57d885f2008-01-25 21:08:18 +01006106 if (class->leave_domain)
6107 class->leave_domain(rq);
Ingo Molnar0eab9142008-01-25 21:08:19 +01006108 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01006109
Gregory Haskinsdc938522008-01-25 21:08:26 +01006110 cpu_clear(rq->cpu, old_rd->span);
6111 cpu_clear(rq->cpu, old_rd->online);
6112
Gregory Haskins57d885f2008-01-25 21:08:18 +01006113 if (atomic_dec_and_test(&old_rd->refcount))
6114 kfree(old_rd);
6115 }
6116
6117 atomic_inc(&rd->refcount);
6118 rq->rd = rd;
6119
Gregory Haskinsdc938522008-01-25 21:08:26 +01006120 cpu_set(rq->cpu, rd->span);
6121 if (cpu_isset(rq->cpu, cpu_online_map))
6122 cpu_set(rq->cpu, rd->online);
6123
Ingo Molnar0eab9142008-01-25 21:08:19 +01006124 for (class = sched_class_highest; class; class = class->next) {
Gregory Haskins57d885f2008-01-25 21:08:18 +01006125 if (class->join_domain)
6126 class->join_domain(rq);
Ingo Molnar0eab9142008-01-25 21:08:19 +01006127 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01006128
6129 spin_unlock_irqrestore(&rq->lock, flags);
6130}
6131
Gregory Haskinsdc938522008-01-25 21:08:26 +01006132static void init_rootdomain(struct root_domain *rd)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006133{
6134 memset(rd, 0, sizeof(*rd));
6135
Gregory Haskinsdc938522008-01-25 21:08:26 +01006136 cpus_clear(rd->span);
6137 cpus_clear(rd->online);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006138}
6139
6140static void init_defrootdomain(void)
6141{
Gregory Haskinsdc938522008-01-25 21:08:26 +01006142 init_rootdomain(&def_root_domain);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006143 atomic_set(&def_root_domain.refcount, 1);
6144}
6145
Gregory Haskinsdc938522008-01-25 21:08:26 +01006146static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006147{
6148 struct root_domain *rd;
6149
6150 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6151 if (!rd)
6152 return NULL;
6153
Gregory Haskinsdc938522008-01-25 21:08:26 +01006154 init_rootdomain(rd);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006155
6156 return rd;
6157}
6158
Linus Torvalds1da177e2005-04-16 15:20:36 -07006159/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01006160 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07006161 * hold the hotplug lock.
6162 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01006163static void
6164cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006165{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006166 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006167 struct sched_domain *tmp;
6168
6169 /* Remove the sched domains which do not contribute to scheduling. */
6170 for (tmp = sd; tmp; tmp = tmp->parent) {
6171 struct sched_domain *parent = tmp->parent;
6172 if (!parent)
6173 break;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006174 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006175 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006176 if (parent->parent)
6177 parent->parent->child = tmp;
6178 }
Suresh Siddha245af2c2005-06-25 14:57:25 -07006179 }
6180
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006181 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006182 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006183 if (sd)
6184 sd->child = NULL;
6185 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006186
6187 sched_domain_debug(sd, cpu);
6188
Gregory Haskins57d885f2008-01-25 21:08:18 +01006189 rq_attach_root(rq, rd);
Nick Piggin674311d2005-06-25 14:57:27 -07006190 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006191}
6192
6193/* cpus with isolated domains */
Tim Chen67af63a2006-12-22 01:07:50 -08006194static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006195
6196/* Setup the mask of cpus configured for isolated domains */
6197static int __init isolated_cpu_setup(char *str)
6198{
6199 int ints[NR_CPUS], i;
6200
6201 str = get_options(str, ARRAY_SIZE(ints), ints);
6202 cpus_clear(cpu_isolated_map);
6203 for (i = 1; i <= ints[0]; i++)
6204 if (ints[i] < NR_CPUS)
6205 cpu_set(ints[i], cpu_isolated_map);
6206 return 1;
6207}
6208
Ingo Molnar8927f492007-10-15 17:00:13 +02006209__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006210
6211/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006212 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6213 * to a function which identifies what group(along with sched group) a CPU
6214 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6215 * (due to the fact that we keep track of groups covered with a cpumask_t).
Linus Torvalds1da177e2005-04-16 15:20:36 -07006216 *
6217 * init_sched_build_groups will build a circular linked list of the groups
6218 * covered by the given span, and will set each group's ->cpumask correctly,
6219 * and ->cpu_power to 0.
6220 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006221static void
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006222init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
6223 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6224 struct sched_group **sg))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006225{
6226 struct sched_group *first = NULL, *last = NULL;
6227 cpumask_t covered = CPU_MASK_NONE;
6228 int i;
6229
6230 for_each_cpu_mask(i, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006231 struct sched_group *sg;
6232 int group = group_fn(i, cpu_map, &sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006233 int j;
6234
6235 if (cpu_isset(i, covered))
6236 continue;
6237
6238 sg->cpumask = CPU_MASK_NONE;
Eric Dumazet5517d862007-05-08 00:32:57 -07006239 sg->__cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006240
6241 for_each_cpu_mask(j, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006242 if (group_fn(j, cpu_map, NULL) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006243 continue;
6244
6245 cpu_set(j, covered);
6246 cpu_set(j, sg->cpumask);
6247 }
6248 if (!first)
6249 first = sg;
6250 if (last)
6251 last->next = sg;
6252 last = sg;
6253 }
6254 last->next = first;
6255}
6256
John Hawkes9c1cfda2005-09-06 15:18:14 -07006257#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07006258
John Hawkes9c1cfda2005-09-06 15:18:14 -07006259#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08006260
John Hawkes9c1cfda2005-09-06 15:18:14 -07006261/**
6262 * find_next_best_node - find the next node to include in a sched_domain
6263 * @node: node whose sched_domain we're building
6264 * @used_nodes: nodes already in the sched_domain
6265 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006266 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07006267 * finds the closest node not already in the @used_nodes map.
6268 *
6269 * Should use nodemask_t.
6270 */
6271static int find_next_best_node(int node, unsigned long *used_nodes)
6272{
6273 int i, n, val, min_val, best_node = 0;
6274
6275 min_val = INT_MAX;
6276
6277 for (i = 0; i < MAX_NUMNODES; i++) {
6278 /* Start at @node */
6279 n = (node + i) % MAX_NUMNODES;
6280
6281 if (!nr_cpus_node(n))
6282 continue;
6283
6284 /* Skip already used nodes */
6285 if (test_bit(n, used_nodes))
6286 continue;
6287
6288 /* Simple min distance search */
6289 val = node_distance(node, n);
6290
6291 if (val < min_val) {
6292 min_val = val;
6293 best_node = n;
6294 }
6295 }
6296
6297 set_bit(best_node, used_nodes);
6298 return best_node;
6299}
6300
6301/**
6302 * sched_domain_node_span - get a cpumask for a node's sched_domain
6303 * @node: node whose cpumask we're constructing
6304 * @size: number of nodes to include in this span
6305 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006306 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07006307 * should be one that prevents unnecessary balancing, but also spreads tasks
6308 * out optimally.
6309 */
6310static cpumask_t sched_domain_node_span(int node)
6311{
John Hawkes9c1cfda2005-09-06 15:18:14 -07006312 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006313 cpumask_t span, nodemask;
6314 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006315
6316 cpus_clear(span);
6317 bitmap_zero(used_nodes, MAX_NUMNODES);
6318
6319 nodemask = node_to_cpumask(node);
6320 cpus_or(span, span, nodemask);
6321 set_bit(node, used_nodes);
6322
6323 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6324 int next_node = find_next_best_node(node, used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006325
John Hawkes9c1cfda2005-09-06 15:18:14 -07006326 nodemask = node_to_cpumask(next_node);
6327 cpus_or(span, span, nodemask);
6328 }
6329
6330 return span;
6331}
6332#endif
6333
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006334int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006335
John Hawkes9c1cfda2005-09-06 15:18:14 -07006336/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07006337 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07006338 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006339#ifdef CONFIG_SCHED_SMT
6340static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006341static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006342
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006343static int
6344cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006345{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006346 if (sg)
6347 *sg = &per_cpu(sched_group_cpus, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006348 return cpu;
6349}
6350#endif
6351
Ingo Molnar48f24c42006-07-03 00:25:40 -07006352/*
6353 * multi-core sched-domains:
6354 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006355#ifdef CONFIG_SCHED_MC
6356static DEFINE_PER_CPU(struct sched_domain, core_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006357static DEFINE_PER_CPU(struct sched_group, sched_group_core);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006358#endif
6359
6360#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006361static int
6362cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006363{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006364 int group;
Mike Travisd5a74302007-10-16 01:24:05 -07006365 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006366 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006367 group = first_cpu(mask);
6368 if (sg)
6369 *sg = &per_cpu(sched_group_core, group);
6370 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006371}
6372#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006373static int
6374cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006375{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006376 if (sg)
6377 *sg = &per_cpu(sched_group_core, cpu);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006378 return cpu;
6379}
6380#endif
6381
Linus Torvalds1da177e2005-04-16 15:20:36 -07006382static DEFINE_PER_CPU(struct sched_domain, phys_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006383static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006384
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006385static int
6386cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006387{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006388 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006389#ifdef CONFIG_SCHED_MC
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006390 cpumask_t mask = cpu_coregroup_map(cpu);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006391 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006392 group = first_cpu(mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006393#elif defined(CONFIG_SCHED_SMT)
Mike Travisd5a74302007-10-16 01:24:05 -07006394 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006395 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006396 group = first_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006397#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006398 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006399#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006400 if (sg)
6401 *sg = &per_cpu(sched_group_phys, group);
6402 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006403}
6404
6405#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07006406/*
6407 * The init_sched_build_groups can't handle what we want to do with node
6408 * groups, so roll our own. Now each node has its own list of groups which
6409 * gets dynamically allocated.
6410 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006411static DEFINE_PER_CPU(struct sched_domain, node_domains);
John Hawkesd1b55132005-09-06 15:18:14 -07006412static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
John Hawkes9c1cfda2005-09-06 15:18:14 -07006413
6414static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006415static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006416
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006417static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6418 struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006419{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006420 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6421 int group;
6422
6423 cpus_and(nodemask, nodemask, *cpu_map);
6424 group = first_cpu(nodemask);
6425
6426 if (sg)
6427 *sg = &per_cpu(sched_group_allnodes, group);
6428 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006429}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006430
Siddha, Suresh B08069032006-03-27 01:15:23 -08006431static void init_numa_sched_groups_power(struct sched_group *group_head)
6432{
6433 struct sched_group *sg = group_head;
6434 int j;
6435
6436 if (!sg)
6437 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006438 do {
6439 for_each_cpu_mask(j, sg->cpumask) {
6440 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08006441
Andi Kleen3a5c3592007-10-15 17:00:14 +02006442 sd = &per_cpu(phys_domains, j);
6443 if (j != first_cpu(sd->groups->cpumask)) {
6444 /*
6445 * Only add "power" once for each
6446 * physical package.
6447 */
6448 continue;
6449 }
6450
6451 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
Siddha, Suresh B08069032006-03-27 01:15:23 -08006452 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02006453 sg = sg->next;
6454 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08006455}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006456#endif
6457
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006458#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006459/* Free memory allocated for various sched_group structures */
6460static void free_sched_groups(const cpumask_t *cpu_map)
6461{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006462 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006463
6464 for_each_cpu_mask(cpu, *cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006465 struct sched_group **sched_group_nodes
6466 = sched_group_nodes_bycpu[cpu];
6467
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006468 if (!sched_group_nodes)
6469 continue;
6470
6471 for (i = 0; i < MAX_NUMNODES; i++) {
6472 cpumask_t nodemask = node_to_cpumask(i);
6473 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6474
6475 cpus_and(nodemask, nodemask, *cpu_map);
6476 if (cpus_empty(nodemask))
6477 continue;
6478
6479 if (sg == NULL)
6480 continue;
6481 sg = sg->next;
6482next_sg:
6483 oldsg = sg;
6484 sg = sg->next;
6485 kfree(oldsg);
6486 if (oldsg != sched_group_nodes[i])
6487 goto next_sg;
6488 }
6489 kfree(sched_group_nodes);
6490 sched_group_nodes_bycpu[cpu] = NULL;
6491 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006492}
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006493#else
6494static void free_sched_groups(const cpumask_t *cpu_map)
6495{
6496}
6497#endif
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006498
Linus Torvalds1da177e2005-04-16 15:20:36 -07006499/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006500 * Initialize sched groups cpu_power.
6501 *
6502 * cpu_power indicates the capacity of sched group, which is used while
6503 * distributing the load between different sched groups in a sched domain.
6504 * Typically cpu_power for all the groups in a sched domain will be same unless
6505 * there are asymmetries in the topology. If there are asymmetries, group
6506 * having more cpu_power will pickup more load compared to the group having
6507 * less cpu_power.
6508 *
6509 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6510 * the maximum number of tasks a group can handle in the presence of other idle
6511 * or lightly loaded groups in the same sched domain.
6512 */
6513static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6514{
6515 struct sched_domain *child;
6516 struct sched_group *group;
6517
6518 WARN_ON(!sd || !sd->groups);
6519
6520 if (cpu != first_cpu(sd->groups->cpumask))
6521 return;
6522
6523 child = sd->child;
6524
Eric Dumazet5517d862007-05-08 00:32:57 -07006525 sd->groups->__cpu_power = 0;
6526
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006527 /*
6528 * For perf policy, if the groups in child domain share resources
6529 * (for example cores sharing some portions of the cache hierarchy
6530 * or SMT), then set this domain groups cpu_power such that each group
6531 * can handle only one task, when there are other idle groups in the
6532 * same sched domain.
6533 */
6534 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6535 (child->flags &
6536 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
Eric Dumazet5517d862007-05-08 00:32:57 -07006537 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006538 return;
6539 }
6540
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006541 /*
6542 * add cpu_power of each child group to this groups cpu_power
6543 */
6544 group = child->groups;
6545 do {
Eric Dumazet5517d862007-05-08 00:32:57 -07006546 sg_inc_cpu_power(sd->groups, group->__cpu_power);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006547 group = group->next;
6548 } while (group != child->groups);
6549}
6550
6551/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006552 * Build sched domains for a given set of cpus and attach the sched domains
6553 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07006554 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006555static int build_sched_domains(const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006556{
6557 int i;
Gregory Haskins57d885f2008-01-25 21:08:18 +01006558 struct root_domain *rd;
John Hawkesd1b55132005-09-06 15:18:14 -07006559#ifdef CONFIG_NUMA
6560 struct sched_group **sched_group_nodes = NULL;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006561 int sd_allnodes = 0;
John Hawkesd1b55132005-09-06 15:18:14 -07006562
6563 /*
6564 * Allocate the per-node list of sched groups
6565 */
Milton Miller5cf9f062007-10-15 17:00:19 +02006566 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006567 GFP_KERNEL);
John Hawkesd1b55132005-09-06 15:18:14 -07006568 if (!sched_group_nodes) {
6569 printk(KERN_WARNING "Can not alloc sched group node list\n");
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006570 return -ENOMEM;
John Hawkesd1b55132005-09-06 15:18:14 -07006571 }
6572 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6573#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006574
Gregory Haskinsdc938522008-01-25 21:08:26 +01006575 rd = alloc_rootdomain();
Gregory Haskins57d885f2008-01-25 21:08:18 +01006576 if (!rd) {
6577 printk(KERN_WARNING "Cannot alloc root domain\n");
6578 return -ENOMEM;
6579 }
6580
Linus Torvalds1da177e2005-04-16 15:20:36 -07006581 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006582 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006583 */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006584 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006585 struct sched_domain *sd = NULL, *p;
6586 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6587
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006588 cpus_and(nodemask, nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006589
6590#ifdef CONFIG_NUMA
Ingo Molnardd41f592007-07-09 18:51:59 +02006591 if (cpus_weight(*cpu_map) >
6592 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07006593 sd = &per_cpu(allnodes_domains, i);
6594 *sd = SD_ALLNODES_INIT;
6595 sd->span = *cpu_map;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006596 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006597 p = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006598 sd_allnodes = 1;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006599 } else
6600 p = NULL;
6601
Linus Torvalds1da177e2005-04-16 15:20:36 -07006602 sd = &per_cpu(node_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006603 *sd = SD_NODE_INIT;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006604 sd->span = sched_domain_node_span(cpu_to_node(i));
6605 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006606 if (p)
6607 p->child = sd;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006608 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006609#endif
6610
6611 p = sd;
6612 sd = &per_cpu(phys_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006613 *sd = SD_CPU_INIT;
6614 sd->span = nodemask;
6615 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006616 if (p)
6617 p->child = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006618 cpu_to_phys_group(i, cpu_map, &sd->groups);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006619
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006620#ifdef CONFIG_SCHED_MC
6621 p = sd;
6622 sd = &per_cpu(core_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006623 *sd = SD_MC_INIT;
6624 sd->span = cpu_coregroup_map(i);
6625 cpus_and(sd->span, sd->span, *cpu_map);
6626 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006627 p->child = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006628 cpu_to_core_group(i, cpu_map, &sd->groups);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006629#endif
6630
Linus Torvalds1da177e2005-04-16 15:20:36 -07006631#ifdef CONFIG_SCHED_SMT
6632 p = sd;
6633 sd = &per_cpu(cpu_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006634 *sd = SD_SIBLING_INIT;
Mike Travisd5a74302007-10-16 01:24:05 -07006635 sd->span = per_cpu(cpu_sibling_map, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006636 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006637 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006638 p->child = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006639 cpu_to_cpu_group(i, cpu_map, &sd->groups);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006640#endif
6641 }
6642
6643#ifdef CONFIG_SCHED_SMT
6644 /* Set up CPU (sibling) groups */
John Hawkes9c1cfda2005-09-06 15:18:14 -07006645 for_each_cpu_mask(i, *cpu_map) {
Mike Travisd5a74302007-10-16 01:24:05 -07006646 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006647 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006648 if (i != first_cpu(this_sibling_map))
6649 continue;
6650
Ingo Molnardd41f592007-07-09 18:51:59 +02006651 init_sched_build_groups(this_sibling_map, cpu_map,
6652 &cpu_to_cpu_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006653 }
6654#endif
6655
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006656#ifdef CONFIG_SCHED_MC
6657 /* Set up multi-core groups */
6658 for_each_cpu_mask(i, *cpu_map) {
6659 cpumask_t this_core_map = cpu_coregroup_map(i);
6660 cpus_and(this_core_map, this_core_map, *cpu_map);
6661 if (i != first_cpu(this_core_map))
6662 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02006663 init_sched_build_groups(this_core_map, cpu_map,
6664 &cpu_to_core_group);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006665 }
6666#endif
6667
Linus Torvalds1da177e2005-04-16 15:20:36 -07006668 /* Set up physical groups */
6669 for (i = 0; i < MAX_NUMNODES; i++) {
6670 cpumask_t nodemask = node_to_cpumask(i);
6671
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006672 cpus_and(nodemask, nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006673 if (cpus_empty(nodemask))
6674 continue;
6675
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006676 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006677 }
6678
6679#ifdef CONFIG_NUMA
6680 /* Set up node groups */
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006681 if (sd_allnodes)
Ingo Molnardd41f592007-07-09 18:51:59 +02006682 init_sched_build_groups(*cpu_map, cpu_map,
6683 &cpu_to_allnodes_group);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006684
6685 for (i = 0; i < MAX_NUMNODES; i++) {
6686 /* Set up node groups */
6687 struct sched_group *sg, *prev;
6688 cpumask_t nodemask = node_to_cpumask(i);
6689 cpumask_t domainspan;
6690 cpumask_t covered = CPU_MASK_NONE;
6691 int j;
6692
6693 cpus_and(nodemask, nodemask, *cpu_map);
John Hawkesd1b55132005-09-06 15:18:14 -07006694 if (cpus_empty(nodemask)) {
6695 sched_group_nodes[i] = NULL;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006696 continue;
John Hawkesd1b55132005-09-06 15:18:14 -07006697 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006698
6699 domainspan = sched_domain_node_span(i);
6700 cpus_and(domainspan, domainspan, *cpu_map);
6701
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07006702 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006703 if (!sg) {
6704 printk(KERN_WARNING "Can not alloc domain group for "
6705 "node %d\n", i);
6706 goto error;
6707 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006708 sched_group_nodes[i] = sg;
6709 for_each_cpu_mask(j, nodemask) {
6710 struct sched_domain *sd;
Ingo Molnar9761eea2007-07-09 18:52:00 +02006711
John Hawkes9c1cfda2005-09-06 15:18:14 -07006712 sd = &per_cpu(node_domains, j);
6713 sd->groups = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006714 }
Eric Dumazet5517d862007-05-08 00:32:57 -07006715 sg->__cpu_power = 0;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006716 sg->cpumask = nodemask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006717 sg->next = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006718 cpus_or(covered, covered, nodemask);
6719 prev = sg;
6720
6721 for (j = 0; j < MAX_NUMNODES; j++) {
6722 cpumask_t tmp, notcovered;
6723 int n = (i + j) % MAX_NUMNODES;
6724
6725 cpus_complement(notcovered, covered);
6726 cpus_and(tmp, notcovered, *cpu_map);
6727 cpus_and(tmp, tmp, domainspan);
6728 if (cpus_empty(tmp))
6729 break;
6730
6731 nodemask = node_to_cpumask(n);
6732 cpus_and(tmp, tmp, nodemask);
6733 if (cpus_empty(tmp))
6734 continue;
6735
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07006736 sg = kmalloc_node(sizeof(struct sched_group),
6737 GFP_KERNEL, i);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006738 if (!sg) {
6739 printk(KERN_WARNING
6740 "Can not alloc domain group for node %d\n", j);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006741 goto error;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006742 }
Eric Dumazet5517d862007-05-08 00:32:57 -07006743 sg->__cpu_power = 0;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006744 sg->cpumask = tmp;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006745 sg->next = prev->next;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006746 cpus_or(covered, covered, tmp);
6747 prev->next = sg;
6748 prev = sg;
6749 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006750 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006751#endif
6752
6753 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006754#ifdef CONFIG_SCHED_SMT
6755 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006756 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6757
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006758 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006759 }
6760#endif
6761#ifdef CONFIG_SCHED_MC
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006762 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006763 struct sched_domain *sd = &per_cpu(core_domains, i);
6764
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006765 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006766 }
6767#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006768
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006769 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006770 struct sched_domain *sd = &per_cpu(phys_domains, i);
6771
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006772 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006773 }
6774
John Hawkes9c1cfda2005-09-06 15:18:14 -07006775#ifdef CONFIG_NUMA
Siddha, Suresh B08069032006-03-27 01:15:23 -08006776 for (i = 0; i < MAX_NUMNODES; i++)
6777 init_numa_sched_groups_power(sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006778
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006779 if (sd_allnodes) {
6780 struct sched_group *sg;
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07006781
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006782 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07006783 init_numa_sched_groups_power(sg);
6784 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006785#endif
6786
Linus Torvalds1da177e2005-04-16 15:20:36 -07006787 /* Attach the domains */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006788 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006789 struct sched_domain *sd;
6790#ifdef CONFIG_SCHED_SMT
6791 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006792#elif defined(CONFIG_SCHED_MC)
6793 sd = &per_cpu(core_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006794#else
6795 sd = &per_cpu(phys_domains, i);
6796#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01006797 cpu_attach_domain(sd, rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006798 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006799
6800 return 0;
6801
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006802#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006803error:
6804 free_sched_groups(cpu_map);
6805 return -ENOMEM;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006806#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006807}
Paul Jackson029190c2007-10-18 23:40:20 -07006808
6809static cpumask_t *doms_cur; /* current sched domains */
6810static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6811
6812/*
6813 * Special case: If a kmalloc of a doms_cur partition (array of
6814 * cpumask_t) fails, then fallback to a single sched domain,
6815 * as determined by the single cpumask_t fallback_doms.
6816 */
6817static cpumask_t fallback_doms;
6818
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006819/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006820 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07006821 * For now this just excludes isolated cpus, but could be used to
6822 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006823 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006824static int arch_init_sched_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006825{
Milton Miller73785472007-10-24 18:23:48 +02006826 int err;
6827
Paul Jackson029190c2007-10-18 23:40:20 -07006828 ndoms_cur = 1;
6829 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6830 if (!doms_cur)
6831 doms_cur = &fallback_doms;
6832 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
Milton Miller73785472007-10-24 18:23:48 +02006833 err = build_sched_domains(doms_cur);
Milton Miller6382bc92007-10-15 17:00:19 +02006834 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02006835
6836 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006837}
6838
6839static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006840{
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006841 free_sched_groups(cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006842}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006843
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006844/*
6845 * Detach sched domains from a group of cpus specified in cpu_map
6846 * These cpus will now be attached to the NULL domain
6847 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08006848static void detach_destroy_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006849{
6850 int i;
6851
Milton Miller6382bc92007-10-15 17:00:19 +02006852 unregister_sched_domain_sysctl();
6853
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006854 for_each_cpu_mask(i, *cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006855 cpu_attach_domain(NULL, &def_root_domain, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006856 synchronize_sched();
6857 arch_destroy_sched_domains(cpu_map);
6858}
6859
Paul Jackson029190c2007-10-18 23:40:20 -07006860/*
6861 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006862 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07006863 * doms_new[] to the current sched domain partitioning, doms_cur[].
6864 * It destroys each deleted domain and builds each new domain.
6865 *
6866 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006867 * The masks don't intersect (don't overlap.) We should setup one
6868 * sched domain for each mask. CPUs not in any of the cpumasks will
6869 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07006870 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6871 * it as it is.
6872 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006873 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6874 * ownership of it and will kfree it when done with it. If the caller
Paul Jackson029190c2007-10-18 23:40:20 -07006875 * failed the kmalloc call, then it can pass in doms_new == NULL,
6876 * and partition_sched_domains() will fallback to the single partition
6877 * 'fallback_doms'.
6878 *
6879 * Call with hotplug lock held
6880 */
6881void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6882{
6883 int i, j;
6884
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01006885 lock_doms_cur();
6886
Milton Miller73785472007-10-24 18:23:48 +02006887 /* always unregister in case we don't destroy any domains */
6888 unregister_sched_domain_sysctl();
6889
Paul Jackson029190c2007-10-18 23:40:20 -07006890 if (doms_new == NULL) {
6891 ndoms_new = 1;
6892 doms_new = &fallback_doms;
6893 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6894 }
6895
6896 /* Destroy deleted domains */
6897 for (i = 0; i < ndoms_cur; i++) {
6898 for (j = 0; j < ndoms_new; j++) {
6899 if (cpus_equal(doms_cur[i], doms_new[j]))
6900 goto match1;
6901 }
6902 /* no match - a current sched domain not in new doms_new[] */
6903 detach_destroy_domains(doms_cur + i);
6904match1:
6905 ;
6906 }
6907
6908 /* Build new domains */
6909 for (i = 0; i < ndoms_new; i++) {
6910 for (j = 0; j < ndoms_cur; j++) {
6911 if (cpus_equal(doms_new[i], doms_cur[j]))
6912 goto match2;
6913 }
6914 /* no match - add a new doms_new */
6915 build_sched_domains(doms_new + i);
6916match2:
6917 ;
6918 }
6919
6920 /* Remember the new sched domains */
6921 if (doms_cur != &fallback_doms)
6922 kfree(doms_cur);
6923 doms_cur = doms_new;
6924 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02006925
6926 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01006927
6928 unlock_doms_cur();
Paul Jackson029190c2007-10-18 23:40:20 -07006929}
6930
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006931#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Adrian Bunk6707de002007-08-12 18:08:19 +02006932static int arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006933{
6934 int err;
6935
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006936 get_online_cpus();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006937 detach_destroy_domains(&cpu_online_map);
6938 err = arch_init_sched_domains(&cpu_online_map);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006939 put_online_cpus();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006940
6941 return err;
6942}
6943
6944static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6945{
6946 int ret;
6947
6948 if (buf[0] != '0' && buf[0] != '1')
6949 return -EINVAL;
6950
6951 if (smt)
6952 sched_smt_power_savings = (buf[0] == '1');
6953 else
6954 sched_mc_power_savings = (buf[0] == '1');
6955
6956 ret = arch_reinit_sched_domains();
6957
6958 return ret ? ret : count;
6959}
6960
Adrian Bunk6707de002007-08-12 18:08:19 +02006961#ifdef CONFIG_SCHED_MC
6962static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6963{
6964 return sprintf(page, "%u\n", sched_mc_power_savings);
6965}
6966static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6967 const char *buf, size_t count)
6968{
6969 return sched_power_savings_store(buf, count, 0);
6970}
6971static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6972 sched_mc_power_savings_store);
6973#endif
6974
6975#ifdef CONFIG_SCHED_SMT
6976static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6977{
6978 return sprintf(page, "%u\n", sched_smt_power_savings);
6979}
6980static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6981 const char *buf, size_t count)
6982{
6983 return sched_power_savings_store(buf, count, 1);
6984}
6985static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6986 sched_smt_power_savings_store);
6987#endif
6988
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006989int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6990{
6991 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006992
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006993#ifdef CONFIG_SCHED_SMT
6994 if (smt_capable())
6995 err = sysfs_create_file(&cls->kset.kobj,
6996 &attr_sched_smt_power_savings.attr);
6997#endif
6998#ifdef CONFIG_SCHED_MC
6999 if (!err && mc_capable())
7000 err = sysfs_create_file(&cls->kset.kobj,
7001 &attr_sched_mc_power_savings.attr);
7002#endif
7003 return err;
7004}
7005#endif
7006
Linus Torvalds1da177e2005-04-16 15:20:36 -07007007/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007008 * Force a reinitialization of the sched domains hierarchy. The domains
Linus Torvalds1da177e2005-04-16 15:20:36 -07007009 * and groups cannot be updated in place without racing with the balancing
Nick Piggin41c7ce92005-06-25 14:57:24 -07007010 * code, so we temporarily attach all running cpus to the NULL domain
Linus Torvalds1da177e2005-04-16 15:20:36 -07007011 * which will prevent rebalancing while the sched domains are recalculated.
7012 */
7013static int update_sched_domains(struct notifier_block *nfb,
7014 unsigned long action, void *hcpu)
7015{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007016 switch (action) {
7017 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007018 case CPU_UP_PREPARE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007019 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007020 case CPU_DOWN_PREPARE_FROZEN:
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007021 detach_destroy_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007022 return NOTIFY_OK;
7023
7024 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007025 case CPU_UP_CANCELED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007026 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007027 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007028 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007029 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007030 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007031 case CPU_DEAD_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007032 /*
7033 * Fall through and re-initialise the domains.
7034 */
7035 break;
7036 default:
7037 return NOTIFY_DONE;
7038 }
7039
7040 /* The hotplug lock is already held by cpu_up/cpu_down */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007041 arch_init_sched_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007042
7043 return NOTIFY_OK;
7044}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007045
7046void __init sched_init_smp(void)
7047{
Nick Piggin5c1e1762006-10-03 01:14:04 -07007048 cpumask_t non_isolated_cpus;
7049
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007050 get_online_cpus();
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007051 arch_init_sched_domains(&cpu_online_map);
Nathan Lynche5e56732007-01-10 23:15:28 -08007052 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
Nick Piggin5c1e1762006-10-03 01:14:04 -07007053 if (cpus_empty(non_isolated_cpus))
7054 cpu_set(smp_processor_id(), non_isolated_cpus);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007055 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007056 /* XXX: Theoretical race here - CPU may be hotplugged now */
7057 hotcpu_notifier(update_sched_domains, 0);
Nick Piggin5c1e1762006-10-03 01:14:04 -07007058
7059 /* Move init over to a non-isolated CPU */
7060 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
7061 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01007062 sched_init_granularity();
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007063
7064#ifdef CONFIG_FAIR_GROUP_SCHED
7065 if (nr_cpu_ids == 1)
7066 return;
7067
7068 lb_monitor_task = kthread_create(load_balance_monitor, NULL,
7069 "group_balance");
7070 if (!IS_ERR(lb_monitor_task)) {
7071 lb_monitor_task->flags |= PF_NOFREEZE;
7072 wake_up_process(lb_monitor_task);
7073 } else {
7074 printk(KERN_ERR "Could not create load balance monitor thread"
7075 "(error = %ld) \n", PTR_ERR(lb_monitor_task));
7076 }
7077#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007078}
7079#else
7080void __init sched_init_smp(void)
7081{
Ingo Molnar19978ca2007-11-09 22:39:38 +01007082 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007083}
7084#endif /* CONFIG_SMP */
7085
7086int in_sched_functions(unsigned long addr)
7087{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007088 return in_lock_functions(addr) ||
7089 (addr >= (unsigned long)__sched_text_start
7090 && addr < (unsigned long)__sched_text_end);
7091}
7092
Alexey Dobriyana9957442007-10-15 17:00:13 +02007093static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02007094{
7095 cfs_rq->tasks_timeline = RB_ROOT;
Ingo Molnardd41f592007-07-09 18:51:59 +02007096#ifdef CONFIG_FAIR_GROUP_SCHED
7097 cfs_rq->rq = rq;
7098#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02007099 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02007100}
7101
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007102static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7103{
7104 struct rt_prio_array *array;
7105 int i;
7106
7107 array = &rt_rq->active;
7108 for (i = 0; i < MAX_RT_PRIO; i++) {
7109 INIT_LIST_HEAD(array->queue + i);
7110 __clear_bit(i, array->bitmap);
7111 }
7112 /* delimiter for bitsearch: */
7113 __set_bit(MAX_RT_PRIO, array->bitmap);
7114
Peter Zijlstra48d5e252008-01-25 21:08:31 +01007115#if defined CONFIG_SMP || defined CONFIG_FAIR_GROUP_SCHED
7116 rt_rq->highest_prio = MAX_RT_PRIO;
7117#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007118#ifdef CONFIG_SMP
7119 rt_rq->rt_nr_migratory = 0;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007120 rt_rq->overloaded = 0;
7121#endif
7122
7123 rt_rq->rt_time = 0;
7124 rt_rq->rt_throttled = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007125
7126#ifdef CONFIG_FAIR_GROUP_SCHED
7127 rt_rq->rq = rq;
7128#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007129}
7130
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007131#ifdef CONFIG_FAIR_GROUP_SCHED
7132static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7133 struct cfs_rq *cfs_rq, struct sched_entity *se,
7134 int cpu, int add)
7135{
7136 tg->cfs_rq[cpu] = cfs_rq;
7137 init_cfs_rq(cfs_rq, rq);
7138 cfs_rq->tg = tg;
7139 if (add)
7140 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7141
7142 tg->se[cpu] = se;
7143 se->cfs_rq = &rq->cfs;
7144 se->my_q = cfs_rq;
7145 se->load.weight = tg->shares;
7146 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7147 se->parent = NULL;
7148}
7149
7150static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7151 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7152 int cpu, int add)
7153{
7154 tg->rt_rq[cpu] = rt_rq;
7155 init_rt_rq(rt_rq, rq);
7156 rt_rq->tg = tg;
7157 rt_rq->rt_se = rt_se;
7158 if (add)
7159 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7160
7161 tg->rt_se[cpu] = rt_se;
7162 rt_se->rt_rq = &rq->rt;
7163 rt_se->my_q = rt_rq;
7164 rt_se->parent = NULL;
7165 INIT_LIST_HEAD(&rt_se->run_list);
7166}
7167#endif
7168
Linus Torvalds1da177e2005-04-16 15:20:36 -07007169void __init sched_init(void)
7170{
Christoph Lameter476f3532007-05-06 14:48:58 -07007171 int highest_cpu = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007172 int i, j;
7173
Gregory Haskins57d885f2008-01-25 21:08:18 +01007174#ifdef CONFIG_SMP
7175 init_defrootdomain();
7176#endif
7177
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007178#ifdef CONFIG_FAIR_GROUP_SCHED
7179 list_add(&init_task_group.list, &task_groups);
7180#endif
7181
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08007182 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007183 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007184
7185 rq = cpu_rq(i);
7186 spin_lock_init(&rq->lock);
Ingo Molnarfcb99372006-07-03 00:25:10 -07007187 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
Nick Piggin78979862005-06-25 14:57:13 -07007188 rq->nr_running = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007189 rq->clock = 1;
7190 init_cfs_rq(&rq->cfs, rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007191 init_rt_rq(&rq->rt, rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007192#ifdef CONFIG_FAIR_GROUP_SCHED
7193 init_task_group.shares = init_task_group_load;
7194 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7195 init_tg_cfs_entry(rq, &init_task_group,
7196 &per_cpu(init_cfs_rq, i),
7197 &per_cpu(init_sched_entity, i), i, 1);
7198
7199 init_task_group.rt_ratio = sysctl_sched_rt_ratio; /* XXX */
7200 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7201 init_tg_rt_entry(rq, &init_task_group,
7202 &per_cpu(init_rt_rq, i),
7203 &per_cpu(init_sched_rt_entity, i), i, 1);
7204#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007205 rq->rt_period_expire = 0;
Peter Zijlstra48d5e252008-01-25 21:08:31 +01007206 rq->rt_throttled = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007207
Ingo Molnardd41f592007-07-09 18:51:59 +02007208 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7209 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007210#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07007211 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007212 rq->rd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007213 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007214 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007215 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07007216 rq->cpu = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007217 rq->migration_thread = NULL;
7218 INIT_LIST_HEAD(&rq->migration_queue);
Gregory Haskinsdc938522008-01-25 21:08:26 +01007219 rq_attach_root(rq, &def_root_domain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007220#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01007221 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007222 atomic_set(&rq->nr_iowait, 0);
Christoph Lameter476f3532007-05-06 14:48:58 -07007223 highest_cpu = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007224 }
7225
Peter Williams2dd73a42006-06-27 02:54:34 -07007226 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07007227
Avi Kivitye107be32007-07-26 13:40:43 +02007228#ifdef CONFIG_PREEMPT_NOTIFIERS
7229 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7230#endif
7231
Christoph Lameterc9819f42006-12-10 02:20:25 -08007232#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07007233 nr_cpu_ids = highest_cpu + 1;
Christoph Lameterc9819f42006-12-10 02:20:25 -08007234 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7235#endif
7236
Heiko Carstensb50f60c2006-07-30 03:03:52 -07007237#ifdef CONFIG_RT_MUTEXES
7238 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7239#endif
7240
Linus Torvalds1da177e2005-04-16 15:20:36 -07007241 /*
7242 * The boot idle thread does lazy MMU switching as well:
7243 */
7244 atomic_inc(&init_mm.mm_count);
7245 enter_lazy_tlb(&init_mm, current);
7246
7247 /*
7248 * Make us the idle thread. Technically, schedule() should not be
7249 * called from this thread, however somewhere below it might be,
7250 * but because we are the idle thread, we just pick up running again
7251 * when this runqueue becomes "idle".
7252 */
7253 init_idle(current, smp_processor_id());
Ingo Molnardd41f592007-07-09 18:51:59 +02007254 /*
7255 * During early bootup we pretend to be a normal task:
7256 */
7257 current->sched_class = &fair_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007258}
7259
7260#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7261void __might_sleep(char *file, int line)
7262{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007263#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07007264 static unsigned long prev_jiffy; /* ratelimiting */
7265
7266 if ((in_atomic() || irqs_disabled()) &&
7267 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7268 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7269 return;
7270 prev_jiffy = jiffies;
Ingo Molnar91368d72006-03-23 03:00:54 -08007271 printk(KERN_ERR "BUG: sleeping function called from invalid"
Linus Torvalds1da177e2005-04-16 15:20:36 -07007272 " context at %s:%d\n", file, line);
7273 printk("in_atomic():%d, irqs_disabled():%d\n",
7274 in_atomic(), irqs_disabled());
Peter Zijlstraa4c410f2006-12-06 20:37:21 -08007275 debug_show_held_locks(current);
Ingo Molnar3117df02006-12-13 00:34:43 -08007276 if (irqs_disabled())
7277 print_irqtrace_events(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007278 dump_stack();
7279 }
7280#endif
7281}
7282EXPORT_SYMBOL(__might_sleep);
7283#endif
7284
7285#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02007286static void normalize_task(struct rq *rq, struct task_struct *p)
7287{
7288 int on_rq;
7289 update_rq_clock(rq);
7290 on_rq = p->se.on_rq;
7291 if (on_rq)
7292 deactivate_task(rq, p, 0);
7293 __setscheduler(rq, p, SCHED_NORMAL, 0);
7294 if (on_rq) {
7295 activate_task(rq, p, 0);
7296 resched_task(rq->curr);
7297 }
7298}
7299
Linus Torvalds1da177e2005-04-16 15:20:36 -07007300void normalize_rt_tasks(void)
7301{
Ingo Molnara0f98a12007-06-17 18:37:45 +02007302 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007303 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007304 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007305
7306 read_lock_irq(&tasklist_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02007307 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02007308 /*
7309 * Only normalize user tasks:
7310 */
7311 if (!p->mm)
7312 continue;
7313
Ingo Molnardd41f592007-07-09 18:51:59 +02007314 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02007315#ifdef CONFIG_SCHEDSTATS
7316 p->se.wait_start = 0;
7317 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007318 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02007319#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02007320 task_rq(p)->clock = 0;
7321
7322 if (!rt_task(p)) {
7323 /*
7324 * Renice negative nice level userspace
7325 * tasks back to 0:
7326 */
7327 if (TASK_NICE(p) < 0 && p->mm)
7328 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007329 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02007330 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007331
Ingo Molnarb29739f2006-06-27 02:54:51 -07007332 spin_lock_irqsave(&p->pi_lock, flags);
7333 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007334
Ingo Molnar178be792007-10-15 17:00:18 +02007335 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02007336
Ingo Molnarb29739f2006-06-27 02:54:51 -07007337 __task_rq_unlock(rq);
7338 spin_unlock_irqrestore(&p->pi_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02007339 } while_each_thread(g, p);
7340
Linus Torvalds1da177e2005-04-16 15:20:36 -07007341 read_unlock_irq(&tasklist_lock);
7342}
7343
7344#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07007345
7346#ifdef CONFIG_IA64
7347/*
7348 * These functions are only useful for the IA64 MCA handling.
7349 *
7350 * They can only be called when the whole system has been
7351 * stopped - every CPU needs to be quiescent, and no scheduling
7352 * activity can take place. Using them for anything else would
7353 * be a serious bug, and as a result, they aren't even visible
7354 * under any other configuration.
7355 */
7356
7357/**
7358 * curr_task - return the current task for a given cpu.
7359 * @cpu: the processor in question.
7360 *
7361 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7362 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007363struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07007364{
7365 return cpu_curr(cpu);
7366}
7367
7368/**
7369 * set_curr_task - set the current task for a given cpu.
7370 * @cpu: the processor in question.
7371 * @p: the task pointer to set.
7372 *
7373 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007374 * are serviced on a separate stack. It allows the architecture to switch the
7375 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07007376 * must be called with all CPU's synchronized, and interrupts disabled, the
7377 * and caller must save the original value of the current task (see
7378 * curr_task() above) and restore that value before reenabling interrupts and
7379 * re-starting the system.
7380 *
7381 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7382 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007383void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07007384{
7385 cpu_curr(cpu) = p;
7386}
7387
7388#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007389
7390#ifdef CONFIG_FAIR_GROUP_SCHED
7391
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007392#ifdef CONFIG_SMP
7393/*
7394 * distribute shares of all task groups among their schedulable entities,
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007395 * to reflect load distribution across cpus.
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007396 */
7397static int rebalance_shares(struct sched_domain *sd, int this_cpu)
7398{
7399 struct cfs_rq *cfs_rq;
7400 struct rq *rq = cpu_rq(this_cpu);
7401 cpumask_t sdspan = sd->span;
7402 int balanced = 1;
7403
7404 /* Walk thr' all the task groups that we have */
7405 for_each_leaf_cfs_rq(rq, cfs_rq) {
7406 int i;
7407 unsigned long total_load = 0, total_shares;
7408 struct task_group *tg = cfs_rq->tg;
7409
7410 /* Gather total task load of this group across cpus */
7411 for_each_cpu_mask(i, sdspan)
7412 total_load += tg->cfs_rq[i]->load.weight;
7413
Ingo Molnar0eab9142008-01-25 21:08:19 +01007414 /* Nothing to do if this group has no load */
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007415 if (!total_load)
7416 continue;
7417
7418 /*
7419 * tg->shares represents the number of cpu shares the task group
7420 * is eligible to hold on a single cpu. On N cpus, it is
7421 * eligible to hold (N * tg->shares) number of cpu shares.
7422 */
7423 total_shares = tg->shares * cpus_weight(sdspan);
7424
7425 /*
7426 * redistribute total_shares across cpus as per the task load
7427 * distribution.
7428 */
7429 for_each_cpu_mask(i, sdspan) {
7430 unsigned long local_load, local_shares;
7431
7432 local_load = tg->cfs_rq[i]->load.weight;
7433 local_shares = (local_load * total_shares) / total_load;
7434 if (!local_shares)
7435 local_shares = MIN_GROUP_SHARES;
7436 if (local_shares == tg->se[i]->load.weight)
7437 continue;
7438
7439 spin_lock_irq(&cpu_rq(i)->lock);
7440 set_se_shares(tg->se[i], local_shares);
7441 spin_unlock_irq(&cpu_rq(i)->lock);
7442 balanced = 0;
7443 }
7444 }
7445
7446 return balanced;
7447}
7448
7449/*
7450 * How frequently should we rebalance_shares() across cpus?
7451 *
7452 * The more frequently we rebalance shares, the more accurate is the fairness
7453 * of cpu bandwidth distribution between task groups. However higher frequency
7454 * also implies increased scheduling overhead.
7455 *
7456 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7457 * consecutive calls to rebalance_shares() in the same sched domain.
7458 *
7459 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7460 * consecutive calls to rebalance_shares() in the same sched domain.
7461 *
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007462 * These settings allows for the appropriate trade-off between accuracy of
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007463 * fairness and the associated overhead.
7464 *
7465 */
7466
7467/* default: 8ms, units: milliseconds */
7468const_debug unsigned int sysctl_sched_min_bal_int_shares = 8;
7469
7470/* default: 128ms, units: milliseconds */
7471const_debug unsigned int sysctl_sched_max_bal_int_shares = 128;
7472
7473/* kernel thread that runs rebalance_shares() periodically */
7474static int load_balance_monitor(void *unused)
7475{
7476 unsigned int timeout = sysctl_sched_min_bal_int_shares;
7477 struct sched_param schedparm;
7478 int ret;
7479
7480 /*
7481 * We don't want this thread's execution to be limited by the shares
7482 * assigned to default group (init_task_group). Hence make it run
7483 * as a SCHED_RR RT task at the lowest priority.
7484 */
7485 schedparm.sched_priority = 1;
7486 ret = sched_setscheduler(current, SCHED_RR, &schedparm);
7487 if (ret)
7488 printk(KERN_ERR "Couldn't set SCHED_RR policy for load balance"
7489 " monitor thread (error = %d) \n", ret);
7490
7491 while (!kthread_should_stop()) {
7492 int i, cpu, balanced = 1;
7493
7494 /* Prevent cpus going down or coming up */
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +01007495 get_online_cpus();
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007496 /* lockout changes to doms_cur[] array */
7497 lock_doms_cur();
7498 /*
7499 * Enter a rcu read-side critical section to safely walk rq->sd
7500 * chain on various cpus and to walk task group list
7501 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7502 */
7503 rcu_read_lock();
7504
7505 for (i = 0; i < ndoms_cur; i++) {
7506 cpumask_t cpumap = doms_cur[i];
7507 struct sched_domain *sd = NULL, *sd_prev = NULL;
7508
7509 cpu = first_cpu(cpumap);
7510
7511 /* Find the highest domain at which to balance shares */
7512 for_each_domain(cpu, sd) {
7513 if (!(sd->flags & SD_LOAD_BALANCE))
7514 continue;
7515 sd_prev = sd;
7516 }
7517
7518 sd = sd_prev;
7519 /* sd == NULL? No load balance reqd in this domain */
7520 if (!sd)
7521 continue;
7522
7523 balanced &= rebalance_shares(sd, cpu);
7524 }
7525
7526 rcu_read_unlock();
7527
7528 unlock_doms_cur();
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +01007529 put_online_cpus();
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007530
7531 if (!balanced)
7532 timeout = sysctl_sched_min_bal_int_shares;
7533 else if (timeout < sysctl_sched_max_bal_int_shares)
7534 timeout *= 2;
7535
7536 msleep_interruptible(timeout);
7537 }
7538
7539 return 0;
7540}
7541#endif /* CONFIG_SMP */
7542
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007543static void free_sched_group(struct task_group *tg)
7544{
7545 int i;
7546
7547 for_each_possible_cpu(i) {
7548 if (tg->cfs_rq)
7549 kfree(tg->cfs_rq[i]);
7550 if (tg->se)
7551 kfree(tg->se[i]);
7552 if (tg->rt_rq)
7553 kfree(tg->rt_rq[i]);
7554 if (tg->rt_se)
7555 kfree(tg->rt_se[i]);
7556 }
7557
7558 kfree(tg->cfs_rq);
7559 kfree(tg->se);
7560 kfree(tg->rt_rq);
7561 kfree(tg->rt_se);
7562 kfree(tg);
7563}
7564
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007565/* allocate runqueue etc for a new task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02007566struct task_group *sched_create_group(void)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007567{
Ingo Molnar4cf86d72007-10-15 17:00:14 +02007568 struct task_group *tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007569 struct cfs_rq *cfs_rq;
7570 struct sched_entity *se;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007571 struct rt_rq *rt_rq;
7572 struct sched_rt_entity *rt_se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007573 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007574 int i;
7575
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007576 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7577 if (!tg)
7578 return ERR_PTR(-ENOMEM);
7579
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007580 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007581 if (!tg->cfs_rq)
7582 goto err;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007583 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007584 if (!tg->se)
7585 goto err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007586 tg->rt_rq = kzalloc(sizeof(rt_rq) * NR_CPUS, GFP_KERNEL);
7587 if (!tg->rt_rq)
7588 goto err;
7589 tg->rt_se = kzalloc(sizeof(rt_se) * NR_CPUS, GFP_KERNEL);
7590 if (!tg->rt_se)
7591 goto err;
7592
7593 tg->shares = NICE_0_LOAD;
7594 tg->rt_ratio = 0; /* XXX */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007595
7596 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007597 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007598
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007599 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7600 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007601 if (!cfs_rq)
7602 goto err;
7603
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007604 se = kmalloc_node(sizeof(struct sched_entity),
7605 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007606 if (!se)
7607 goto err;
7608
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007609 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7610 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7611 if (!rt_rq)
7612 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007613
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007614 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7615 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7616 if (!rt_se)
7617 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007618
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007619 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7620 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007621 }
7622
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007623 lock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007624 for_each_possible_cpu(i) {
7625 rq = cpu_rq(i);
7626 cfs_rq = tg->cfs_rq[i];
7627 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007628 rt_rq = tg->rt_rq[i];
7629 list_add_rcu(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007630 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007631 list_add_rcu(&tg->list, &task_groups);
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007632 unlock_task_group_list();
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007633
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007634 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007635
7636err:
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007637 free_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007638 return ERR_PTR(-ENOMEM);
7639}
7640
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007641/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007642static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007643{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007644 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007645 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007646}
7647
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007648/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02007649void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007650{
James Bottomley7bae49d2007-10-29 21:18:11 +01007651 struct cfs_rq *cfs_rq = NULL;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007652 struct rt_rq *rt_rq = NULL;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007653 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007654
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007655 lock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007656 for_each_possible_cpu(i) {
7657 cfs_rq = tg->cfs_rq[i];
7658 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007659 rt_rq = tg->rt_rq[i];
7660 list_del_rcu(&rt_rq->leaf_rt_rq_list);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007661 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007662 list_del_rcu(&tg->list);
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007663 unlock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007664
James Bottomley7bae49d2007-10-29 21:18:11 +01007665 BUG_ON(!cfs_rq);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007666
7667 /* wait for possible concurrent references to cfs_rqs complete */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007668 call_rcu(&tg->rcu, free_sched_group_rcu);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007669}
7670
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007671/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +02007672 * The caller of this function should have put the task in its new group
7673 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7674 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007675 */
7676void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007677{
7678 int on_rq, running;
7679 unsigned long flags;
7680 struct rq *rq;
7681
7682 rq = task_rq_lock(tsk, &flags);
7683
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007684 update_rq_clock(rq);
7685
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01007686 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007687 on_rq = tsk->se.on_rq;
7688
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02007689 if (on_rq) {
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007690 dequeue_task(rq, tsk, 0);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02007691 if (unlikely(running))
7692 tsk->sched_class->put_prev_task(rq, tsk);
7693 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007694
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007695 set_task_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007696
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02007697 if (on_rq) {
7698 if (unlikely(running))
7699 tsk->sched_class->set_curr_task(rq);
Dmitry Adamushko7074bad2007-10-15 17:00:07 +02007700 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02007701 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007702
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007703 task_rq_unlock(rq, &flags);
7704}
7705
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007706/* rq->lock to be locked by caller */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007707static void set_se_shares(struct sched_entity *se, unsigned long shares)
7708{
7709 struct cfs_rq *cfs_rq = se->cfs_rq;
7710 struct rq *rq = cfs_rq->rq;
7711 int on_rq;
7712
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007713 if (!shares)
7714 shares = MIN_GROUP_SHARES;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007715
7716 on_rq = se->on_rq;
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007717 if (on_rq) {
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007718 dequeue_entity(cfs_rq, se, 0);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007719 dec_cpu_load(rq, se->load.weight);
7720 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007721
7722 se->load.weight = shares;
7723 se->load.inv_weight = div64_64((1ULL<<32), shares);
7724
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007725 if (on_rq) {
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007726 enqueue_entity(cfs_rq, se, 0);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007727 inc_cpu_load(rq, se->load.weight);
7728 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007729}
7730
Ingo Molnar4cf86d72007-10-15 17:00:14 +02007731int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007732{
7733 int i;
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007734 struct cfs_rq *cfs_rq;
7735 struct rq *rq;
Ingo Molnarc61935f2008-01-22 11:24:58 +01007736
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007737 lock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007738 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +02007739 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007740
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007741 if (shares < MIN_GROUP_SHARES)
7742 shares = MIN_GROUP_SHARES;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007743
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007744 /*
7745 * Prevent any load balance activity (rebalance_shares,
7746 * load_balance_fair) from referring to this group first,
7747 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7748 */
7749 for_each_possible_cpu(i) {
7750 cfs_rq = tg->cfs_rq[i];
7751 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7752 }
7753
7754 /* wait for any ongoing reference to this group to finish */
7755 synchronize_sched();
7756
7757 /*
7758 * Now we are free to modify the group's share on each cpu
7759 * w/o tripping rebalance_share or load_balance_fair.
7760 */
7761 tg->shares = shares;
7762 for_each_possible_cpu(i) {
7763 spin_lock_irq(&cpu_rq(i)->lock);
7764 set_se_shares(tg->se[i], shares);
7765 spin_unlock_irq(&cpu_rq(i)->lock);
7766 }
7767
7768 /*
7769 * Enable load balance activity on this group, by inserting it back on
7770 * each cpu's rq->leaf_cfs_rq_list.
7771 */
7772 for_each_possible_cpu(i) {
7773 rq = cpu_rq(i);
7774 cfs_rq = tg->cfs_rq[i];
7775 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7776 }
Dhaval Giani5cb350b2007-10-15 17:00:14 +02007777done:
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007778 unlock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007779 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007780}
7781
Dhaval Giani5cb350b2007-10-15 17:00:14 +02007782unsigned long sched_group_shares(struct task_group *tg)
7783{
7784 return tg->shares;
7785}
7786
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007787/*
7788 * Ensure the total rt_ratio <= sysctl_sched_rt_ratio
7789 */
7790int sched_group_set_rt_ratio(struct task_group *tg, unsigned long rt_ratio)
7791{
7792 struct task_group *tgi;
7793 unsigned long total = 0;
7794
7795 rcu_read_lock();
7796 list_for_each_entry_rcu(tgi, &task_groups, list)
7797 total += tgi->rt_ratio;
7798 rcu_read_unlock();
7799
7800 if (total + rt_ratio - tg->rt_ratio > sysctl_sched_rt_ratio)
7801 return -EINVAL;
7802
7803 tg->rt_ratio = rt_ratio;
7804 return 0;
7805}
7806
7807unsigned long sched_group_rt_ratio(struct task_group *tg)
7808{
7809 return tg->rt_ratio;
7810}
7811
Ingo Molnar3a252012007-10-15 17:00:12 +02007812#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007813
7814#ifdef CONFIG_FAIR_CGROUP_SCHED
7815
7816/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02007817static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007818{
Paul Menage2b01dfe2007-10-24 18:23:50 +02007819 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7820 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007821}
7822
7823static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +02007824cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007825{
7826 struct task_group *tg;
7827
Paul Menage2b01dfe2007-10-24 18:23:50 +02007828 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007829 /* This is early initialization for the top cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02007830 init_task_group.css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007831 return &init_task_group.css;
7832 }
7833
7834 /* we support only 1-level deep hierarchical scheduler atm */
Paul Menage2b01dfe2007-10-24 18:23:50 +02007835 if (cgrp->parent->parent)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007836 return ERR_PTR(-EINVAL);
7837
7838 tg = sched_create_group();
7839 if (IS_ERR(tg))
7840 return ERR_PTR(-ENOMEM);
7841
7842 /* Bind the cgroup to task_group object we just created */
Paul Menage2b01dfe2007-10-24 18:23:50 +02007843 tg->css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007844
7845 return &tg->css;
7846}
7847
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007848static void
7849cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007850{
Paul Menage2b01dfe2007-10-24 18:23:50 +02007851 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007852
7853 sched_destroy_group(tg);
7854}
7855
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007856static int
7857cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7858 struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007859{
7860 /* We don't support RT-tasks being in separate groups */
7861 if (tsk->sched_class != &fair_sched_class)
7862 return -EINVAL;
7863
7864 return 0;
7865}
7866
7867static void
Paul Menage2b01dfe2007-10-24 18:23:50 +02007868cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007869 struct cgroup *old_cont, struct task_struct *tsk)
7870{
7871 sched_move_task(tsk);
7872}
7873
Paul Menage2b01dfe2007-10-24 18:23:50 +02007874static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7875 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007876{
Paul Menage2b01dfe2007-10-24 18:23:50 +02007877 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007878}
7879
Paul Menage2b01dfe2007-10-24 18:23:50 +02007880static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007881{
Paul Menage2b01dfe2007-10-24 18:23:50 +02007882 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007883
7884 return (u64) tg->shares;
7885}
7886
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007887static int cpu_rt_ratio_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7888 u64 rt_ratio_val)
7889{
7890 return sched_group_set_rt_ratio(cgroup_tg(cgrp), rt_ratio_val);
7891}
7892
7893static u64 cpu_rt_ratio_read_uint(struct cgroup *cgrp, struct cftype *cft)
7894{
7895 struct task_group *tg = cgroup_tg(cgrp);
7896
7897 return (u64) tg->rt_ratio;
7898}
7899
Paul Menagefe5c7cc2007-10-29 21:18:11 +01007900static struct cftype cpu_files[] = {
7901 {
7902 .name = "shares",
7903 .read_uint = cpu_shares_read_uint,
7904 .write_uint = cpu_shares_write_uint,
7905 },
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007906 {
7907 .name = "rt_ratio",
7908 .read_uint = cpu_rt_ratio_read_uint,
7909 .write_uint = cpu_rt_ratio_write_uint,
7910 },
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007911};
7912
7913static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7914{
Paul Menagefe5c7cc2007-10-29 21:18:11 +01007915 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007916}
7917
7918struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +01007919 .name = "cpu",
7920 .create = cpu_cgroup_create,
7921 .destroy = cpu_cgroup_destroy,
7922 .can_attach = cpu_cgroup_can_attach,
7923 .attach = cpu_cgroup_attach,
7924 .populate = cpu_cgroup_populate,
7925 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007926 .early_init = 1,
7927};
7928
7929#endif /* CONFIG_FAIR_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01007930
7931#ifdef CONFIG_CGROUP_CPUACCT
7932
7933/*
7934 * CPU accounting code for task groups.
7935 *
7936 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7937 * (balbir@in.ibm.com).
7938 */
7939
7940/* track cpu usage of a group of tasks */
7941struct cpuacct {
7942 struct cgroup_subsys_state css;
7943 /* cpuusage holds pointer to a u64-type object on every cpu */
7944 u64 *cpuusage;
7945};
7946
7947struct cgroup_subsys cpuacct_subsys;
7948
7949/* return cpu accounting group corresponding to this container */
7950static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7951{
7952 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7953 struct cpuacct, css);
7954}
7955
7956/* return cpu accounting group to which this task belongs */
7957static inline struct cpuacct *task_ca(struct task_struct *tsk)
7958{
7959 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7960 struct cpuacct, css);
7961}
7962
7963/* create a new cpu accounting group */
7964static struct cgroup_subsys_state *cpuacct_create(
7965 struct cgroup_subsys *ss, struct cgroup *cont)
7966{
7967 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7968
7969 if (!ca)
7970 return ERR_PTR(-ENOMEM);
7971
7972 ca->cpuusage = alloc_percpu(u64);
7973 if (!ca->cpuusage) {
7974 kfree(ca);
7975 return ERR_PTR(-ENOMEM);
7976 }
7977
7978 return &ca->css;
7979}
7980
7981/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007982static void
7983cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01007984{
7985 struct cpuacct *ca = cgroup_ca(cont);
7986
7987 free_percpu(ca->cpuusage);
7988 kfree(ca);
7989}
7990
7991/* return total cpu usage (in nanoseconds) of a group */
7992static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7993{
7994 struct cpuacct *ca = cgroup_ca(cont);
7995 u64 totalcpuusage = 0;
7996 int i;
7997
7998 for_each_possible_cpu(i) {
7999 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8000
8001 /*
8002 * Take rq->lock to make 64-bit addition safe on 32-bit
8003 * platforms.
8004 */
8005 spin_lock_irq(&cpu_rq(i)->lock);
8006 totalcpuusage += *cpuusage;
8007 spin_unlock_irq(&cpu_rq(i)->lock);
8008 }
8009
8010 return totalcpuusage;
8011}
8012
8013static struct cftype files[] = {
8014 {
8015 .name = "usage",
8016 .read_uint = cpuusage_read,
8017 },
8018};
8019
8020static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8021{
8022 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
8023}
8024
8025/*
8026 * charge this task's execution time to its accounting group.
8027 *
8028 * called with rq->lock held.
8029 */
8030static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8031{
8032 struct cpuacct *ca;
8033
8034 if (!cpuacct_subsys.active)
8035 return;
8036
8037 ca = task_ca(tsk);
8038 if (ca) {
8039 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8040
8041 *cpuusage += cputime;
8042 }
8043}
8044
8045struct cgroup_subsys cpuacct_subsys = {
8046 .name = "cpuacct",
8047 .create = cpuacct_create,
8048 .destroy = cpuacct_destroy,
8049 .populate = cpuacct_populate,
8050 .subsys_id = cpuacct_subsys_id,
8051};
8052#endif /* CONFIG_CGROUP_CPUACCT */