blob: 328494e28df25cf5a633eba587b59f17e2daf2e3 [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>
Reynes Philippe30914a52008-03-17 16:19:05 -070069#include <linux/tick.h>
Mike Travis434d53b2008-04-04 18:11:04 -070070#include <linux/bootmem.h>
Peter Zijlstraf00b45c2008-04-19 19:45:00 +020071#include <linux/debugfs.h>
72#include <linux/ctype.h>
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +020073#include <linux/ftrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
Eric Dumazet5517d862007-05-08 00:32:57 -070075#include <asm/tlb.h>
Satyam Sharma838225b2007-10-24 18:23:50 +020076#include <asm/irq_regs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
78/*
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
81 * and back.
82 */
83#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
86
87/*
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
91 */
92#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
95
96/*
Ingo Molnard7876a02008-01-25 21:08:19 +010097 * Helpers for converting nanosecond timing to jiffy resolution
Linus Torvalds1da177e2005-04-16 15:20:36 -070098 */
Eric Dumazetd6322fa2007-11-09 22:39:38 +010099#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200101#define NICE_0_LOAD SCHED_LOAD_SCALE
102#define NICE_0_SHIFT SCHED_LOAD_SHIFT
103
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104/*
105 * These are the 'tuning knobs' of the scheduler:
106 *
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +0200107 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 * Timeslices get refilled after they expire.
109 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110#define DEF_TIMESLICE (100 * HZ / 1000)
Peter Williams2dd73a42006-06-27 02:54:34 -0700111
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200112/*
113 * single value that denotes runtime == period, ie unlimited time.
114 */
115#define RUNTIME_INF ((u64)~0ULL)
116
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
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200158struct rt_bandwidth {
Ingo Molnarea736ed2008-03-25 13:51:45 +0100159 /* nests inside the rq lock: */
160 spinlock_t rt_runtime_lock;
161 ktime_t rt_period;
162 u64 rt_runtime;
163 struct hrtimer rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200164};
165
166static struct rt_bandwidth def_rt_bandwidth;
167
168static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
169
170static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
171{
172 struct rt_bandwidth *rt_b =
173 container_of(timer, struct rt_bandwidth, rt_period_timer);
174 ktime_t now;
175 int overrun;
176 int idle = 0;
177
178 for (;;) {
179 now = hrtimer_cb_get_time(timer);
180 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
181
182 if (!overrun)
183 break;
184
185 idle = do_sched_rt_period_timer(rt_b, overrun);
186 }
187
188 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
189}
190
191static
192void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
193{
194 rt_b->rt_period = ns_to_ktime(period);
195 rt_b->rt_runtime = runtime;
196
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200197 spin_lock_init(&rt_b->rt_runtime_lock);
198
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200199 hrtimer_init(&rt_b->rt_period_timer,
200 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
201 rt_b->rt_period_timer.function = sched_rt_period_timer;
202 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
203}
204
205static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
206{
207 ktime_t now;
208
209 if (rt_b->rt_runtime == RUNTIME_INF)
210 return;
211
212 if (hrtimer_active(&rt_b->rt_period_timer))
213 return;
214
215 spin_lock(&rt_b->rt_runtime_lock);
216 for (;;) {
217 if (hrtimer_active(&rt_b->rt_period_timer))
218 break;
219
220 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
221 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
222 hrtimer_start(&rt_b->rt_period_timer,
223 rt_b->rt_period_timer.expires,
224 HRTIMER_MODE_ABS);
225 }
226 spin_unlock(&rt_b->rt_runtime_lock);
227}
228
229#ifdef CONFIG_RT_GROUP_SCHED
230static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
231{
232 hrtimer_cancel(&rt_b->rt_period_timer);
233}
234#endif
235
Heiko Carstens712555e2008-04-28 11:33:07 +0200236/*
237 * sched_domains_mutex serializes calls to arch_init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
239 */
240static DEFINE_MUTEX(sched_domains_mutex);
241
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100242#ifdef CONFIG_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200243
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700244#include <linux/cgroup.h>
245
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200246struct cfs_rq;
247
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100248static LIST_HEAD(task_groups);
249
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200250/* task group related information */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200251struct task_group {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100252#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700253 struct cgroup_subsys_state css;
254#endif
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100255
256#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200257 /* schedulable entities of this group on each cpu */
258 struct sched_entity **se;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq **cfs_rq;
261 unsigned long shares;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100262#endif
263
264#ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
267
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200268 struct rt_bandwidth rt_bandwidth;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100269#endif
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100270
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +0100271 struct rcu_head rcu;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100272 struct list_head list;
Peter Zijlstraf473aa52008-04-19 19:45:00 +0200273
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200277};
278
Dhaval Giani354d60c2008-04-19 19:44:59 +0200279#ifdef CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200280
281/*
282 * Root task group.
283 * Every UID task group (including init_task_group aka UID-0) will
284 * be a child to this group.
285 */
286struct task_group root_task_group;
287
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100288#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200289/* Default task group's sched entity on each cpu */
290static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
291/* Default task group's cfs_rq on each cpu */
292static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100293#endif
294
295#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100296static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
297static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100298#endif
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200299#else
300#define root_task_group init_task_group
Dhaval Giani354d60c2008-04-19 19:44:59 +0200301#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100302
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100303/* task_group_lock serializes add/remove of task groups and also changes to
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100304 * a task group's cpu shares.
305 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100306static DEFINE_SPINLOCK(task_group_lock);
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100307
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100308#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100309#ifdef CONFIG_USER_SCHED
Ingo Molnar0eab9142008-01-25 21:08:19 +0100310# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200311#else
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100312# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200313#endif
314
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800315/*
316 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
317 * (The default weight is 1024 - so there's no practical
318 * limitation from this.)
319 */
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200320#define MIN_SHARES 2
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800321#define MAX_SHARES (ULONG_MAX - 1)
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200322
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100323static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100324#endif
325
326/* Default task group.
327 * Every task in system belong to this group at bootup.
328 */
Mike Travis434d53b2008-04-04 18:11:04 -0700329struct task_group init_task_group;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200330
331/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200332static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200333{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200334 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200335
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100336#ifdef CONFIG_USER_SCHED
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200337 tg = p->user->tg;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100338#elif defined(CONFIG_CGROUP_SCHED)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700339 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
340 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200341#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100342 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200343#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200344 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200345}
346
347/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100348static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200349{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100350#ifdef CONFIG_FAIR_GROUP_SCHED
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100351 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
352 p->se.parent = task_group(p)->se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100353#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100354
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100355#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100356 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
357 p->rt.parent = task_group(p)->rt_se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100358#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200359}
360
361#else
362
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100363static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200364
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100365#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200366
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200367/* CFS-related fields in a runqueue */
368struct cfs_rq {
369 struct load_weight load;
370 unsigned long nr_running;
371
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200372 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200373 u64 min_vruntime;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200374
375 struct rb_root tasks_timeline;
376 struct rb_node *rb_leftmost;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +0200377
378 struct list_head tasks;
379 struct list_head *balance_iterator;
380
381 /*
382 * 'curr' points to currently running entity on this cfs_rq.
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200383 * It is set to NULL otherwise (i.e when none are currently running).
384 */
Peter Zijlstraaa2ac252008-03-14 21:12:12 +0100385 struct sched_entity *curr, *next;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200386
387 unsigned long nr_spread_over;
388
Ingo Molnar62160e32007-10-15 17:00:03 +0200389#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200390 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
391
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100392 /*
393 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200394 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
395 * (like users, containers etc.)
396 *
397 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
398 * list is used during load balance.
399 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100400 struct list_head leaf_cfs_rq_list;
401 struct task_group *tg; /* group that "owns" this runqueue */
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200402
403#ifdef CONFIG_SMP
404 unsigned long task_weight;
405 unsigned long shares;
406 /*
407 * We need space to build a sched_domain wide view of the full task
408 * group tree, in order to avoid depending on dynamic memory allocation
409 * during the load balancing we place this in the per cpu task group
410 * hierarchy. This limits the load balancing to one instance per cpu,
411 * but more should not be needed anyway.
412 */
413 struct aggregate_struct {
414 /*
415 * load = weight(cpus) * f(tg)
416 *
417 * Where f(tg) is the recursive weight fraction assigned to
418 * this group.
419 */
420 unsigned long load;
421
422 /*
423 * part of the group weight distributed to this span.
424 */
425 unsigned long shares;
426
427 /*
428 * The sum of all runqueue weights within this span.
429 */
430 unsigned long rq_weight;
431
432 /*
433 * Weight contributed by tasks; this is the part we can
434 * influence by moving tasks around.
435 */
436 unsigned long task_weight;
437 } aggregate;
438#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200439#endif
440};
441
442/* Real-Time classes' related field in a runqueue: */
443struct rt_rq {
444 struct rt_prio_array active;
Steven Rostedt63489e42008-01-25 21:08:03 +0100445 unsigned long rt_nr_running;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100446#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100447 int highest_prio; /* highest queued rt task prio */
448#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100449#ifdef CONFIG_SMP
Gregory Haskins73fe6aa2008-01-25 21:08:07 +0100450 unsigned long rt_nr_migratory;
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +0100451 int overloaded;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100452#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100453 int rt_throttled;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100454 u64 rt_time;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200455 u64 rt_runtime;
Ingo Molnarea736ed2008-03-25 13:51:45 +0100456 /* Nests inside the rq lock: */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200457 spinlock_t rt_runtime_lock;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100458
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100459#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100460 unsigned long rt_nr_boosted;
461
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100462 struct rq *rq;
463 struct list_head leaf_rt_rq_list;
464 struct task_group *tg;
465 struct sched_rt_entity *rt_se;
466#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200467};
468
Gregory Haskins57d885f2008-01-25 21:08:18 +0100469#ifdef CONFIG_SMP
470
471/*
472 * We add the notion of a root-domain which will be used to define per-domain
Ingo Molnar0eab9142008-01-25 21:08:19 +0100473 * variables. Each exclusive cpuset essentially defines an island domain by
474 * fully partitioning the member cpus from any other cpuset. Whenever a new
Gregory Haskins57d885f2008-01-25 21:08:18 +0100475 * exclusive cpuset is created, we also create and attach a new root-domain
476 * object.
477 *
Gregory Haskins57d885f2008-01-25 21:08:18 +0100478 */
479struct root_domain {
480 atomic_t refcount;
481 cpumask_t span;
482 cpumask_t online;
Gregory Haskins637f5082008-01-25 21:08:18 +0100483
Ingo Molnar0eab9142008-01-25 21:08:19 +0100484 /*
Gregory Haskins637f5082008-01-25 21:08:18 +0100485 * The "RT overload" flag: it gets set if a CPU has more than
486 * one runnable RT task.
487 */
488 cpumask_t rto_mask;
Ingo Molnar0eab9142008-01-25 21:08:19 +0100489 atomic_t rto_count;
Gregory Haskins57d885f2008-01-25 21:08:18 +0100490};
491
Gregory Haskinsdc938522008-01-25 21:08:26 +0100492/*
493 * By default the system creates a single root-domain with all cpus as
494 * members (mimicking the global state we have today).
495 */
Gregory Haskins57d885f2008-01-25 21:08:18 +0100496static struct root_domain def_root_domain;
497
498#endif
499
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200500/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 * This is the main, per-CPU runqueue data structure.
502 *
503 * Locking rule: those places that want to lock multiple runqueues
504 * (such as the load balancing or the thread migration code), lock
505 * acquire operations must be ordered by ascending &runqueue.
506 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700507struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200508 /* runqueue lock: */
509 spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510
511 /*
512 * nr_running and cpu_load should be in the same cacheline because
513 * remote CPUs use both these fields when doing load calculation.
514 */
515 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200516 #define CPU_LOAD_IDX_MAX 5
517 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh Bbdecea32007-05-08 00:32:48 -0700518 unsigned char idle_at_tick;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700519#ifdef CONFIG_NO_HZ
Guillaume Chazarain15934a32008-04-19 19:44:57 +0200520 unsigned long last_tick_seen;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700521 unsigned char in_nohz_recently;
522#endif
Ingo Molnard8016492007-10-18 21:32:55 +0200523 /* capture load from *all* tasks on this cpu: */
524 struct load_weight load;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200525 unsigned long nr_load_updates;
526 u64 nr_switches;
527
528 struct cfs_rq cfs;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100529 struct rt_rq rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100530
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200531#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200532 /* list of leaf cfs_rq on this cpu: */
533 struct list_head leaf_cfs_rq_list;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100534#endif
535#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100536 struct list_head leaf_rt_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538
539 /*
540 * This is part of a global counter where only the total sum
541 * over all CPUs matters. A task can increase this counter on
542 * one CPU and if it got migrated afterwards it may decrease
543 * it on another CPU. Always updated under the runqueue lock:
544 */
545 unsigned long nr_uninterruptible;
546
Ingo Molnar36c8b582006-07-03 00:25:41 -0700547 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800548 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200550
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200551 u64 clock;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200552
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 atomic_t nr_iowait;
554
555#ifdef CONFIG_SMP
Ingo Molnar0eab9142008-01-25 21:08:19 +0100556 struct root_domain *rd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557 struct sched_domain *sd;
558
559 /* For active balancing */
560 int active_balance;
561 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200562 /* cpu of this runqueue: */
563 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
Ingo Molnar36c8b582006-07-03 00:25:41 -0700565 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 struct list_head migration_queue;
567#endif
568
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100569#ifdef CONFIG_SCHED_HRTICK
570 unsigned long hrtick_flags;
571 ktime_t hrtick_expire;
572 struct hrtimer hrtick_timer;
573#endif
574
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575#ifdef CONFIG_SCHEDSTATS
576 /* latency stats */
577 struct sched_info rq_sched_info;
578
579 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200580 unsigned int yld_exp_empty;
581 unsigned int yld_act_empty;
582 unsigned int yld_both_empty;
583 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584
585 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200586 unsigned int sched_switch;
587 unsigned int sched_count;
588 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589
590 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200591 unsigned int ttwu_count;
592 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200593
594 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200595 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596#endif
Ingo Molnarfcb99372006-07-03 00:25:10 -0700597 struct lock_class_key rq_lock_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598};
599
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700600static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601
Ingo Molnardd41f592007-07-09 18:51:59 +0200602static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
603{
604 rq->curr->sched_class->check_preempt_curr(rq, p);
605}
606
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700607static inline int cpu_of(struct rq *rq)
608{
609#ifdef CONFIG_SMP
610 return rq->cpu;
611#else
612 return 0;
613#endif
614}
615
Ingo Molnar20d315d2007-07-09 18:51:58 +0200616/*
Nick Piggin674311d2005-06-25 14:57:27 -0700617 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700618 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700619 *
620 * The domain tree of any CPU may only be accessed from within
621 * preempt-disabled sections.
622 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700623#define for_each_domain(cpu, __sd) \
624 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625
626#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
627#define this_rq() (&__get_cpu_var(runqueues))
628#define task_rq(p) cpu_rq(task_cpu(p))
629#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
630
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200631static inline void update_rq_clock(struct rq *rq)
632{
633 rq->clock = sched_clock_cpu(cpu_of(rq));
634}
635
Ingo Molnare436d802007-07-19 21:28:35 +0200636/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200637 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
638 */
639#ifdef CONFIG_SCHED_DEBUG
640# define const_debug __read_mostly
641#else
642# define const_debug static const
643#endif
644
645/*
646 * Debugging: various feature bits
647 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200648
649#define SCHED_FEAT(name, enabled) \
650 __SCHED_FEAT_##name ,
651
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200652enum {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200653#include "sched_features.h"
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200654};
655
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200656#undef SCHED_FEAT
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200657
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200658#define SCHED_FEAT(name, enabled) \
659 (1UL << __SCHED_FEAT_##name) * enabled |
660
661const_debug unsigned int sysctl_sched_features =
662#include "sched_features.h"
663 0;
664
665#undef SCHED_FEAT
666
667#ifdef CONFIG_SCHED_DEBUG
668#define SCHED_FEAT(name, enabled) \
669 #name ,
670
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700671static __read_mostly char *sched_feat_names[] = {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200672#include "sched_features.h"
673 NULL
674};
675
676#undef SCHED_FEAT
677
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700678static int sched_feat_open(struct inode *inode, struct file *filp)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200679{
680 filp->private_data = inode->i_private;
681 return 0;
682}
683
684static ssize_t
685sched_feat_read(struct file *filp, char __user *ubuf,
686 size_t cnt, loff_t *ppos)
687{
688 char *buf;
689 int r = 0;
690 int len = 0;
691 int i;
692
693 for (i = 0; sched_feat_names[i]; i++) {
694 len += strlen(sched_feat_names[i]);
695 len += 4;
696 }
697
698 buf = kmalloc(len + 2, GFP_KERNEL);
699 if (!buf)
700 return -ENOMEM;
701
702 for (i = 0; sched_feat_names[i]; i++) {
703 if (sysctl_sched_features & (1UL << i))
704 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
705 else
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200706 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200707 }
708
709 r += sprintf(buf + r, "\n");
710 WARN_ON(r >= len + 2);
711
712 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
713
714 kfree(buf);
715
716 return r;
717}
718
719static ssize_t
720sched_feat_write(struct file *filp, const char __user *ubuf,
721 size_t cnt, loff_t *ppos)
722{
723 char buf[64];
724 char *cmp = buf;
725 int neg = 0;
726 int i;
727
728 if (cnt > 63)
729 cnt = 63;
730
731 if (copy_from_user(&buf, ubuf, cnt))
732 return -EFAULT;
733
734 buf[cnt] = 0;
735
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200736 if (strncmp(buf, "NO_", 3) == 0) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200737 neg = 1;
738 cmp += 3;
739 }
740
741 for (i = 0; sched_feat_names[i]; i++) {
742 int len = strlen(sched_feat_names[i]);
743
744 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
745 if (neg)
746 sysctl_sched_features &= ~(1UL << i);
747 else
748 sysctl_sched_features |= (1UL << i);
749 break;
750 }
751 }
752
753 if (!sched_feat_names[i])
754 return -EINVAL;
755
756 filp->f_pos += cnt;
757
758 return cnt;
759}
760
761static struct file_operations sched_feat_fops = {
762 .open = sched_feat_open,
763 .read = sched_feat_read,
764 .write = sched_feat_write,
765};
766
767static __init int sched_init_debug(void)
768{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200769 debugfs_create_file("sched_features", 0644, NULL, NULL,
770 &sched_feat_fops);
771
772 return 0;
773}
774late_initcall(sched_init_debug);
775
776#endif
777
778#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200779
780/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100781 * Number of tasks to iterate in a single balance run.
782 * Limited because this is done with IRQs disabled.
783 */
784const_debug unsigned int sysctl_sched_nr_migrate = 32;
785
786/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100787 * period over which we measure -rt task cpu usage in us.
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100788 * default: 1s
789 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100790unsigned int sysctl_sched_rt_period = 1000000;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100791
Ingo Molnar6892b752008-02-13 14:02:36 +0100792static __read_mostly int scheduler_running;
793
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100794/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100795 * part of the period that we allow rt tasks to run in us.
796 * default: 0.95s
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100797 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100798int sysctl_sched_rt_runtime = 950000;
799
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200800static inline u64 global_rt_period(void)
801{
802 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
803}
804
805static inline u64 global_rt_runtime(void)
806{
807 if (sysctl_sched_rt_period < 0)
808 return RUNTIME_INF;
809
810 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
811}
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100812
Ingo Molnar690229a2008-04-23 09:31:35 +0200813unsigned long long time_sync_thresh = 100000;
Ingo Molnar27ec4402008-02-28 21:00:21 +0100814
815static DEFINE_PER_CPU(unsigned long long, time_offset);
816static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
817
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100818/*
Ingo Molnar27ec4402008-02-28 21:00:21 +0100819 * Global lock which we take every now and then to synchronize
820 * the CPUs time. This method is not warp-safe, but it's good
821 * enough to synchronize slowly diverging time sources and thus
822 * it's good enough for tracing:
Ingo Molnare436d802007-07-19 21:28:35 +0200823 */
Ingo Molnar27ec4402008-02-28 21:00:21 +0100824static DEFINE_SPINLOCK(time_sync_lock);
825static unsigned long long prev_global_time;
826
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200827static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
Ingo Molnar27ec4402008-02-28 21:00:21 +0100828{
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200829 /*
830 * We want this inlined, to not get tracer function calls
831 * in this critical section:
832 */
833 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
834 __raw_spin_lock(&time_sync_lock.raw_lock);
Ingo Molnar27ec4402008-02-28 21:00:21 +0100835
836 if (time < prev_global_time) {
837 per_cpu(time_offset, cpu) += prev_global_time - time;
838 time = prev_global_time;
839 } else {
840 prev_global_time = time;
841 }
842
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200843 __raw_spin_unlock(&time_sync_lock.raw_lock);
844 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
Ingo Molnar27ec4402008-02-28 21:00:21 +0100845
846 return time;
847}
848
849static unsigned long long __cpu_clock(int cpu)
Ingo Molnare436d802007-07-19 21:28:35 +0200850{
Ingo Molnare436d802007-07-19 21:28:35 +0200851 unsigned long long now;
Ingo Molnare436d802007-07-19 21:28:35 +0200852
Ingo Molnar8ced5f62007-12-07 19:02:47 +0100853 /*
854 * Only call sched_clock() if the scheduler has already been
855 * initialized (some code might call cpu_clock() very early):
856 */
Ingo Molnar6892b752008-02-13 14:02:36 +0100857 if (unlikely(!scheduler_running))
858 return 0;
859
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200860 now = sched_clock_cpu(cpu);
Ingo Molnare436d802007-07-19 21:28:35 +0200861
862 return now;
863}
Ingo Molnar27ec4402008-02-28 21:00:21 +0100864
865/*
866 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
867 * clock constructed from sched_clock():
868 */
869unsigned long long cpu_clock(int cpu)
870{
871 unsigned long long prev_cpu_time, time, delta_time;
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200872 unsigned long flags;
Ingo Molnar27ec4402008-02-28 21:00:21 +0100873
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200874 local_irq_save(flags);
Ingo Molnar27ec4402008-02-28 21:00:21 +0100875 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
876 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
877 delta_time = time-prev_cpu_time;
878
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200879 if (unlikely(delta_time > time_sync_thresh)) {
Ingo Molnar27ec4402008-02-28 21:00:21 +0100880 time = __sync_cpu_clock(time, cpu);
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200881 per_cpu(prev_cpu_time, cpu) = time;
882 }
883 local_irq_restore(flags);
Ingo Molnar27ec4402008-02-28 21:00:21 +0100884
885 return time;
886}
Paul E. McKenneya58f6f22007-10-15 17:00:14 +0200887EXPORT_SYMBOL_GPL(cpu_clock);
Ingo Molnare436d802007-07-19 21:28:35 +0200888
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700890# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700892#ifndef finish_arch_switch
893# define finish_arch_switch(prev) do { } while (0)
894#endif
895
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100896static inline int task_current(struct rq *rq, struct task_struct *p)
897{
898 return rq->curr == p;
899}
900
Nick Piggin4866cde2005-06-25 14:57:23 -0700901#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700902static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700903{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100904 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700905}
906
Ingo Molnar70b97a72006-07-03 00:25:42 -0700907static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700908{
909}
910
Ingo Molnar70b97a72006-07-03 00:25:42 -0700911static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700912{
Ingo Molnarda04c032005-09-13 11:17:59 +0200913#ifdef CONFIG_DEBUG_SPINLOCK
914 /* this is a valid case when another task releases the spinlock */
915 rq->lock.owner = current;
916#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700917 /*
918 * If we are tracking spinlock dependencies then we have to
919 * fix up the runqueue lock - which gets 'carried over' from
920 * prev into current:
921 */
922 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
923
Nick Piggin4866cde2005-06-25 14:57:23 -0700924 spin_unlock_irq(&rq->lock);
925}
926
927#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700928static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700929{
930#ifdef CONFIG_SMP
931 return p->oncpu;
932#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100933 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700934#endif
935}
936
Ingo Molnar70b97a72006-07-03 00:25:42 -0700937static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700938{
939#ifdef CONFIG_SMP
940 /*
941 * We can optimise this out completely for !SMP, because the
942 * SMP rebalancing from interrupt is the only thing that cares
943 * here.
944 */
945 next->oncpu = 1;
946#endif
947#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
948 spin_unlock_irq(&rq->lock);
949#else
950 spin_unlock(&rq->lock);
951#endif
952}
953
Ingo Molnar70b97a72006-07-03 00:25:42 -0700954static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700955{
956#ifdef CONFIG_SMP
957 /*
958 * After ->oncpu is cleared, the task can be moved to a different CPU.
959 * We must ensure this doesn't happen until the switch is completely
960 * finished.
961 */
962 smp_wmb();
963 prev->oncpu = 0;
964#endif
965#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
966 local_irq_enable();
967#endif
968}
969#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970
971/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700972 * __task_rq_lock - lock the runqueue a given task resides on.
973 * Must be called interrupts disabled.
974 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700975static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700976 __acquires(rq->lock)
977{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200978 for (;;) {
979 struct rq *rq = task_rq(p);
980 spin_lock(&rq->lock);
981 if (likely(rq == task_rq(p)))
982 return rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700983 spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700984 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700985}
986
987/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100989 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 * explicitly disabling preemption.
991 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700992static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 __acquires(rq->lock)
994{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700995 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996
Andi Kleen3a5c3592007-10-15 17:00:14 +0200997 for (;;) {
998 local_irq_save(*flags);
999 rq = task_rq(p);
1000 spin_lock(&rq->lock);
1001 if (likely(rq == task_rq(p)))
1002 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003 spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005}
1006
Alexey Dobriyana9957442007-10-15 17:00:13 +02001007static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001008 __releases(rq->lock)
1009{
1010 spin_unlock(&rq->lock);
1011}
1012
Ingo Molnar70b97a72006-07-03 00:25:42 -07001013static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 __releases(rq->lock)
1015{
1016 spin_unlock_irqrestore(&rq->lock, *flags);
1017}
1018
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -08001020 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001022static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 __acquires(rq->lock)
1024{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001025 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026
1027 local_irq_disable();
1028 rq = this_rq();
1029 spin_lock(&rq->lock);
1030
1031 return rq;
1032}
1033
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001034static void __resched_task(struct task_struct *p, int tif_bit);
1035
1036static inline void resched_task(struct task_struct *p)
1037{
1038 __resched_task(p, TIF_NEED_RESCHED);
1039}
1040
1041#ifdef CONFIG_SCHED_HRTICK
1042/*
1043 * Use HR-timers to deliver accurate preemption points.
1044 *
1045 * Its all a bit involved since we cannot program an hrt while holding the
1046 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1047 * reschedule event.
1048 *
1049 * When we get rescheduled we reprogram the hrtick_timer outside of the
1050 * rq->lock.
1051 */
1052static inline void resched_hrt(struct task_struct *p)
1053{
1054 __resched_task(p, TIF_HRTICK_RESCHED);
1055}
1056
1057static inline void resched_rq(struct rq *rq)
1058{
1059 unsigned long flags;
1060
1061 spin_lock_irqsave(&rq->lock, flags);
1062 resched_task(rq->curr);
1063 spin_unlock_irqrestore(&rq->lock, flags);
1064}
1065
1066enum {
1067 HRTICK_SET, /* re-programm hrtick_timer */
1068 HRTICK_RESET, /* not a new slice */
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001069 HRTICK_BLOCK, /* stop hrtick operations */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001070};
1071
1072/*
1073 * Use hrtick when:
1074 * - enabled by features
1075 * - hrtimer is actually high res
1076 */
1077static inline int hrtick_enabled(struct rq *rq)
1078{
1079 if (!sched_feat(HRTICK))
1080 return 0;
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001081 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1082 return 0;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001083 return hrtimer_is_hres_active(&rq->hrtick_timer);
1084}
1085
1086/*
1087 * Called to set the hrtick timer state.
1088 *
1089 * called with rq->lock held and irqs disabled
1090 */
1091static void hrtick_start(struct rq *rq, u64 delay, int reset)
1092{
1093 assert_spin_locked(&rq->lock);
1094
1095 /*
1096 * preempt at: now + delay
1097 */
1098 rq->hrtick_expire =
1099 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1100 /*
1101 * indicate we need to program the timer
1102 */
1103 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1104 if (reset)
1105 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1106
1107 /*
1108 * New slices are called from the schedule path and don't need a
1109 * forced reschedule.
1110 */
1111 if (reset)
1112 resched_hrt(rq->curr);
1113}
1114
1115static void hrtick_clear(struct rq *rq)
1116{
1117 if (hrtimer_active(&rq->hrtick_timer))
1118 hrtimer_cancel(&rq->hrtick_timer);
1119}
1120
1121/*
1122 * Update the timer from the possible pending state.
1123 */
1124static void hrtick_set(struct rq *rq)
1125{
1126 ktime_t time;
1127 int set, reset;
1128 unsigned long flags;
1129
1130 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1131
1132 spin_lock_irqsave(&rq->lock, flags);
1133 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1134 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1135 time = rq->hrtick_expire;
1136 clear_thread_flag(TIF_HRTICK_RESCHED);
1137 spin_unlock_irqrestore(&rq->lock, flags);
1138
1139 if (set) {
1140 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1141 if (reset && !hrtimer_active(&rq->hrtick_timer))
1142 resched_rq(rq);
1143 } else
1144 hrtick_clear(rq);
1145}
1146
1147/*
1148 * High-resolution timer tick.
1149 * Runs from hardirq context with interrupts disabled.
1150 */
1151static enum hrtimer_restart hrtick(struct hrtimer *timer)
1152{
1153 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1154
1155 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1156
1157 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02001158 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001159 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1160 spin_unlock(&rq->lock);
1161
1162 return HRTIMER_NORESTART;
1163}
1164
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001165static void hotplug_hrtick_disable(int cpu)
1166{
1167 struct rq *rq = cpu_rq(cpu);
1168 unsigned long flags;
1169
1170 spin_lock_irqsave(&rq->lock, flags);
1171 rq->hrtick_flags = 0;
1172 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1173 spin_unlock_irqrestore(&rq->lock, flags);
1174
1175 hrtick_clear(rq);
1176}
1177
1178static void hotplug_hrtick_enable(int cpu)
1179{
1180 struct rq *rq = cpu_rq(cpu);
1181 unsigned long flags;
1182
1183 spin_lock_irqsave(&rq->lock, flags);
1184 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1185 spin_unlock_irqrestore(&rq->lock, flags);
1186}
1187
1188static int
1189hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1190{
1191 int cpu = (int)(long)hcpu;
1192
1193 switch (action) {
1194 case CPU_UP_CANCELED:
1195 case CPU_UP_CANCELED_FROZEN:
1196 case CPU_DOWN_PREPARE:
1197 case CPU_DOWN_PREPARE_FROZEN:
1198 case CPU_DEAD:
1199 case CPU_DEAD_FROZEN:
1200 hotplug_hrtick_disable(cpu);
1201 return NOTIFY_OK;
1202
1203 case CPU_UP_PREPARE:
1204 case CPU_UP_PREPARE_FROZEN:
1205 case CPU_DOWN_FAILED:
1206 case CPU_DOWN_FAILED_FROZEN:
1207 case CPU_ONLINE:
1208 case CPU_ONLINE_FROZEN:
1209 hotplug_hrtick_enable(cpu);
1210 return NOTIFY_OK;
1211 }
1212
1213 return NOTIFY_DONE;
1214}
1215
1216static void init_hrtick(void)
1217{
1218 hotcpu_notifier(hotplug_hrtick, 0);
1219}
1220
1221static void init_rq_hrtick(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001222{
1223 rq->hrtick_flags = 0;
1224 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1225 rq->hrtick_timer.function = hrtick;
1226 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1227}
1228
1229void hrtick_resched(void)
1230{
1231 struct rq *rq;
1232 unsigned long flags;
1233
1234 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1235 return;
1236
1237 local_irq_save(flags);
1238 rq = cpu_rq(smp_processor_id());
1239 hrtick_set(rq);
1240 local_irq_restore(flags);
1241}
1242#else
1243static inline void hrtick_clear(struct rq *rq)
1244{
1245}
1246
1247static inline void hrtick_set(struct rq *rq)
1248{
1249}
1250
1251static inline void init_rq_hrtick(struct rq *rq)
1252{
1253}
1254
1255void hrtick_resched(void)
1256{
1257}
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001258
1259static inline void init_hrtick(void)
1260{
1261}
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001262#endif
1263
Ingo Molnar1b9f19c2007-07-09 18:51:59 +02001264/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001265 * resched_task - mark a task 'to be rescheduled now'.
1266 *
1267 * On UP this means the setting of the need_resched flag, on SMP it
1268 * might also involve a cross-CPU call to trigger the scheduler on
1269 * the target CPU.
1270 */
1271#ifdef CONFIG_SMP
1272
1273#ifndef tsk_is_polling
1274#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1275#endif
1276
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001277static void __resched_task(struct task_struct *p, int tif_bit)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001278{
1279 int cpu;
1280
1281 assert_spin_locked(&task_rq(p)->lock);
1282
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001283 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001284 return;
1285
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001286 set_tsk_thread_flag(p, tif_bit);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001287
1288 cpu = task_cpu(p);
1289 if (cpu == smp_processor_id())
1290 return;
1291
1292 /* NEED_RESCHED must be visible before we test polling */
1293 smp_mb();
1294 if (!tsk_is_polling(p))
1295 smp_send_reschedule(cpu);
1296}
1297
1298static void resched_cpu(int cpu)
1299{
1300 struct rq *rq = cpu_rq(cpu);
1301 unsigned long flags;
1302
1303 if (!spin_trylock_irqsave(&rq->lock, flags))
1304 return;
1305 resched_task(cpu_curr(cpu));
1306 spin_unlock_irqrestore(&rq->lock, flags);
1307}
Thomas Gleixner06d83082008-03-22 09:20:24 +01001308
1309#ifdef CONFIG_NO_HZ
1310/*
1311 * When add_timer_on() enqueues a timer into the timer wheel of an
1312 * idle CPU then this timer might expire before the next timer event
1313 * which is scheduled to wake up that CPU. In case of a completely
1314 * idle system the next event might even be infinite time into the
1315 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1316 * leaves the inner idle loop so the newly added timer is taken into
1317 * account when the CPU goes back to idle and evaluates the timer
1318 * wheel for the next timer event.
1319 */
1320void wake_up_idle_cpu(int cpu)
1321{
1322 struct rq *rq = cpu_rq(cpu);
1323
1324 if (cpu == smp_processor_id())
1325 return;
1326
1327 /*
1328 * This is safe, as this function is called with the timer
1329 * wheel base lock of (cpu) held. When the CPU is on the way
1330 * to idle and has not yet set rq->curr to idle then it will
1331 * be serialized on the timer wheel base lock and take the new
1332 * timer into account automatically.
1333 */
1334 if (rq->curr != rq->idle)
1335 return;
1336
1337 /*
1338 * We can set TIF_RESCHED on the idle task of the other CPU
1339 * lockless. The worst case is that the other CPU runs the
1340 * idle task through an additional NOOP schedule()
1341 */
1342 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1343
1344 /* NEED_RESCHED must be visible before we test polling */
1345 smp_mb();
1346 if (!tsk_is_polling(rq->idle))
1347 smp_send_reschedule(cpu);
1348}
1349#endif
1350
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001351#else
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001352static void __resched_task(struct task_struct *p, int tif_bit)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001353{
1354 assert_spin_locked(&task_rq(p)->lock);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001355 set_tsk_thread_flag(p, tif_bit);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001356}
1357#endif
1358
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001359#if BITS_PER_LONG == 32
1360# define WMULT_CONST (~0UL)
1361#else
1362# define WMULT_CONST (1UL << 32)
1363#endif
1364
1365#define WMULT_SHIFT 32
1366
Ingo Molnar194081e2007-08-09 11:16:51 +02001367/*
1368 * Shift right and round:
1369 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001370#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +02001371
Peter Zijlstra8f1bc382008-04-19 19:45:00 +02001372/*
1373 * delta *= weight / lw
1374 */
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +02001375static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001376calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1377 struct load_weight *lw)
1378{
1379 u64 tmp;
1380
Peter Zijlstrae05510d2008-05-05 23:56:17 +02001381 if (!lw->inv_weight)
1382 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001383
1384 tmp = (u64)delta_exec * weight;
1385 /*
1386 * Check whether we'd overflow the 64-bit multiplication:
1387 */
Ingo Molnar194081e2007-08-09 11:16:51 +02001388 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001389 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +02001390 WMULT_SHIFT/2);
1391 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001392 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001393
Ingo Molnarecf691d2007-08-02 17:41:40 +02001394 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001395}
1396
Ingo Molnar10919852007-10-15 17:00:04 +02001397static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001398{
1399 lw->weight += inc;
Ingo Molnare89996a2008-03-14 23:48:28 +01001400 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001401}
1402
Ingo Molnar10919852007-10-15 17:00:04 +02001403static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001404{
1405 lw->weight -= dec;
Ingo Molnare89996a2008-03-14 23:48:28 +01001406 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001407}
1408
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001410 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1411 * of tasks with abnormal "nice" values across CPUs the contribution that
1412 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001413 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -07001414 * scaled version of the new time slice allocation that they receive on time
1415 * slice expiry etc.
1416 */
1417
Ingo Molnardd41f592007-07-09 18:51:59 +02001418#define WEIGHT_IDLEPRIO 2
1419#define WMULT_IDLEPRIO (1 << 31)
1420
1421/*
1422 * Nice levels are multiplicative, with a gentle 10% change for every
1423 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1424 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1425 * that remained on nice 0.
1426 *
1427 * The "10% effect" is relative and cumulative: from _any_ nice level,
1428 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +02001429 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1430 * If a task goes up by ~10% and another task goes down by ~10% then
1431 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +02001432 */
1433static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001434 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1435 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1436 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1437 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1438 /* 0 */ 1024, 820, 655, 526, 423,
1439 /* 5 */ 335, 272, 215, 172, 137,
1440 /* 10 */ 110, 87, 70, 56, 45,
1441 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +02001442};
1443
Ingo Molnar5714d2d2007-07-16 09:46:31 +02001444/*
1445 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1446 *
1447 * In cases where the weight does not change often, we can use the
1448 * precalculated inverse to speed up arithmetics by turning divisions
1449 * into multiplications:
1450 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001451static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001452 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1453 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1454 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1455 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1456 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1457 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1458 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1459 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +02001460};
Peter Williams2dd73a42006-06-27 02:54:34 -07001461
Ingo Molnardd41f592007-07-09 18:51:59 +02001462static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1463
1464/*
1465 * runqueue iterator, to support SMP load-balancing between different
1466 * scheduling classes, without having to expose their internal data
1467 * structures to the load-balancing proper:
1468 */
1469struct rq_iterator {
1470 void *arg;
1471 struct task_struct *(*start)(void *);
1472 struct task_struct *(*next)(void *);
1473};
1474
Peter Williamse1d14842007-10-24 18:23:51 +02001475#ifdef CONFIG_SMP
1476static unsigned long
1477balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1478 unsigned long max_load_move, struct sched_domain *sd,
1479 enum cpu_idle_type idle, int *all_pinned,
1480 int *this_best_prio, struct rq_iterator *iterator);
1481
1482static int
1483iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1484 struct sched_domain *sd, enum cpu_idle_type idle,
1485 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +02001486#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001487
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001488#ifdef CONFIG_CGROUP_CPUACCT
1489static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1490#else
1491static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1492#endif
1493
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001494static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1495{
1496 update_load_add(&rq->load, load);
1497}
1498
1499static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1500{
1501 update_load_sub(&rq->load, load);
1502}
1503
Gregory Haskinse7693a32008-01-25 21:08:09 +01001504#ifdef CONFIG_SMP
1505static unsigned long source_load(int cpu, int type);
1506static unsigned long target_load(int cpu, int type);
1507static unsigned long cpu_avg_load_per_task(int cpu);
1508static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001509
1510#ifdef CONFIG_FAIR_GROUP_SCHED
1511
1512/*
1513 * Group load balancing.
1514 *
1515 * We calculate a few balance domain wide aggregate numbers; load and weight.
1516 * Given the pictures below, and assuming each item has equal weight:
1517 *
1518 * root 1 - thread
1519 * / | \ A - group
1520 * A 1 B
1521 * /|\ / \
1522 * C 2 D 3 4
1523 * | |
1524 * 5 6
1525 *
1526 * load:
1527 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1528 * which equals 1/9-th of the total load.
1529 *
1530 * shares:
1531 * The weight of this group on the selected cpus.
1532 *
1533 * rq_weight:
1534 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1535 * B would get 2.
1536 *
1537 * task_weight:
1538 * Part of the rq_weight contributed by tasks; all groups except B would
1539 * get 1, B gets 2.
1540 */
1541
1542static inline struct aggregate_struct *
1543aggregate(struct task_group *tg, struct sched_domain *sd)
1544{
1545 return &tg->cfs_rq[sd->first_cpu]->aggregate;
1546}
1547
1548typedef void (*aggregate_func)(struct task_group *, struct sched_domain *);
1549
1550/*
1551 * Iterate the full tree, calling @down when first entering a node and @up when
1552 * leaving it for the final time.
1553 */
1554static
1555void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1556 struct sched_domain *sd)
1557{
1558 struct task_group *parent, *child;
1559
1560 rcu_read_lock();
1561 parent = &root_task_group;
1562down:
1563 (*down)(parent, sd);
1564 list_for_each_entry_rcu(child, &parent->children, siblings) {
1565 parent = child;
1566 goto down;
1567
1568up:
1569 continue;
1570 }
1571 (*up)(parent, sd);
1572
1573 child = parent;
1574 parent = parent->parent;
1575 if (parent)
1576 goto up;
1577 rcu_read_unlock();
1578}
1579
1580/*
1581 * Calculate the aggregate runqueue weight.
1582 */
1583static
1584void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd)
1585{
1586 unsigned long rq_weight = 0;
1587 unsigned long task_weight = 0;
1588 int i;
1589
1590 for_each_cpu_mask(i, sd->span) {
1591 rq_weight += tg->cfs_rq[i]->load.weight;
1592 task_weight += tg->cfs_rq[i]->task_weight;
1593 }
1594
1595 aggregate(tg, sd)->rq_weight = rq_weight;
1596 aggregate(tg, sd)->task_weight = task_weight;
1597}
1598
1599/*
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001600 * Compute the weight of this group on the given cpus.
1601 */
1602static
1603void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd)
1604{
1605 unsigned long shares = 0;
1606 int i;
1607
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001608 for_each_cpu_mask(i, sd->span)
1609 shares += tg->cfs_rq[i]->shares;
1610
Peter Zijlstra3f5087a2008-04-25 00:25:08 +02001611 if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares)
1612 shares = tg->shares;
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001613
1614 aggregate(tg, sd)->shares = shares;
1615}
1616
1617/*
1618 * Compute the load fraction assigned to this group, relies on the aggregate
1619 * weight and this group's parent's load, i.e. top-down.
1620 */
1621static
1622void aggregate_group_load(struct task_group *tg, struct sched_domain *sd)
1623{
1624 unsigned long load;
1625
1626 if (!tg->parent) {
1627 int i;
1628
1629 load = 0;
1630 for_each_cpu_mask(i, sd->span)
1631 load += cpu_rq(i)->load.weight;
1632
1633 } else {
1634 load = aggregate(tg->parent, sd)->load;
1635
1636 /*
1637 * shares is our weight in the parent's rq so
1638 * shares/parent->rq_weight gives our fraction of the load
1639 */
1640 load *= aggregate(tg, sd)->shares;
1641 load /= aggregate(tg->parent, sd)->rq_weight + 1;
1642 }
1643
1644 aggregate(tg, sd)->load = load;
1645}
1646
1647static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1648
1649/*
1650 * Calculate and set the cpu's group shares.
1651 */
1652static void
1653__update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd,
1654 int tcpu)
1655{
1656 int boost = 0;
1657 unsigned long shares;
1658 unsigned long rq_weight;
1659
1660 if (!tg->se[tcpu])
1661 return;
1662
1663 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1664
1665 /*
1666 * If there are currently no tasks on the cpu pretend there is one of
1667 * average load so that when a new task gets to run here it will not
1668 * get delayed by group starvation.
1669 */
1670 if (!rq_weight) {
1671 boost = 1;
1672 rq_weight = NICE_0_LOAD;
1673 }
1674
1675 /*
1676 * \Sum shares * rq_weight
1677 * shares = -----------------------
1678 * \Sum rq_weight
1679 *
1680 */
1681 shares = aggregate(tg, sd)->shares * rq_weight;
1682 shares /= aggregate(tg, sd)->rq_weight + 1;
1683
1684 /*
1685 * record the actual number of shares, not the boosted amount.
1686 */
1687 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1688
1689 if (shares < MIN_SHARES)
1690 shares = MIN_SHARES;
Miao Xiecb4ad1f2008-04-28 12:54:56 +08001691 else if (shares > MAX_SHARES)
1692 shares = MAX_SHARES;
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001693
1694 __set_se_shares(tg->se[tcpu], shares);
1695}
1696
1697/*
1698 * Re-adjust the weights on the cpu the task came from and on the cpu the
1699 * task went to.
1700 */
1701static void
1702__move_group_shares(struct task_group *tg, struct sched_domain *sd,
1703 int scpu, int dcpu)
1704{
1705 unsigned long shares;
1706
1707 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1708
1709 __update_group_shares_cpu(tg, sd, scpu);
1710 __update_group_shares_cpu(tg, sd, dcpu);
1711
1712 /*
1713 * ensure we never loose shares due to rounding errors in the
1714 * above redistribution.
1715 */
1716 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1717 if (shares)
1718 tg->cfs_rq[dcpu]->shares += shares;
1719}
1720
1721/*
1722 * Because changing a group's shares changes the weight of the super-group
1723 * we need to walk up the tree and change all shares until we hit the root.
1724 */
1725static void
1726move_group_shares(struct task_group *tg, struct sched_domain *sd,
1727 int scpu, int dcpu)
1728{
1729 while (tg) {
1730 __move_group_shares(tg, sd, scpu, dcpu);
1731 tg = tg->parent;
1732 }
1733}
1734
1735static
1736void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd)
1737{
1738 unsigned long shares = aggregate(tg, sd)->shares;
1739 int i;
1740
1741 for_each_cpu_mask(i, sd->span) {
1742 struct rq *rq = cpu_rq(i);
1743 unsigned long flags;
1744
1745 spin_lock_irqsave(&rq->lock, flags);
1746 __update_group_shares_cpu(tg, sd, i);
1747 spin_unlock_irqrestore(&rq->lock, flags);
1748 }
1749
1750 aggregate_group_shares(tg, sd);
1751
1752 /*
1753 * ensure we never loose shares due to rounding errors in the
1754 * above redistribution.
1755 */
1756 shares -= aggregate(tg, sd)->shares;
1757 if (shares) {
1758 tg->cfs_rq[sd->first_cpu]->shares += shares;
1759 aggregate(tg, sd)->shares += shares;
1760 }
1761}
1762
1763/*
1764 * Calculate the accumulative weight and recursive load of each task group
1765 * while walking down the tree.
1766 */
1767static
1768void aggregate_get_down(struct task_group *tg, struct sched_domain *sd)
1769{
1770 aggregate_group_weight(tg, sd);
1771 aggregate_group_shares(tg, sd);
1772 aggregate_group_load(tg, sd);
1773}
1774
1775/*
1776 * Rebalance the cpu shares while walking back up the tree.
1777 */
1778static
1779void aggregate_get_up(struct task_group *tg, struct sched_domain *sd)
1780{
1781 aggregate_group_set_shares(tg, sd);
1782}
1783
1784static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1785
1786static void __init init_aggregate(void)
1787{
1788 int i;
1789
1790 for_each_possible_cpu(i)
1791 spin_lock_init(&per_cpu(aggregate_lock, i));
1792}
1793
1794static int get_aggregate(struct sched_domain *sd)
1795{
1796 if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu)))
1797 return 0;
1798
1799 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd);
1800 return 1;
1801}
1802
1803static void put_aggregate(struct sched_domain *sd)
1804{
1805 spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu));
1806}
1807
1808static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1809{
1810 cfs_rq->shares = shares;
1811}
1812
1813#else
1814
1815static inline void init_aggregate(void)
1816{
1817}
1818
1819static inline int get_aggregate(struct sched_domain *sd)
1820{
1821 return 0;
1822}
1823
1824static inline void put_aggregate(struct sched_domain *sd)
1825{
1826}
1827#endif
1828
1829#else /* CONFIG_SMP */
1830
1831#ifdef CONFIG_FAIR_GROUP_SCHED
1832static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1833{
1834}
1835#endif
1836
Gregory Haskinse7693a32008-01-25 21:08:09 +01001837#endif /* CONFIG_SMP */
1838
Ingo Molnardd41f592007-07-09 18:51:59 +02001839#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +02001840#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001841#include "sched_fair.c"
1842#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +02001843#ifdef CONFIG_SCHED_DEBUG
1844# include "sched_debug.c"
1845#endif
1846
1847#define sched_class_highest (&rt_sched_class)
1848
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001849static void inc_nr_running(struct rq *rq)
Peter Zijlstra62fb1852008-02-25 17:34:02 +01001850{
1851 rq->nr_running++;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01001852}
1853
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001854static void dec_nr_running(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001855{
1856 rq->nr_running--;
Ingo Molnar9c217242007-08-02 17:41:40 +02001857}
1858
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001859static void set_load_weight(struct task_struct *p)
1860{
1861 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001862 p->se.load.weight = prio_to_weight[0] * 2;
1863 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1864 return;
1865 }
1866
1867 /*
1868 * SCHED_IDLE tasks get minimal weight:
1869 */
1870 if (p->policy == SCHED_IDLE) {
1871 p->se.load.weight = WEIGHT_IDLEPRIO;
1872 p->se.load.inv_weight = WMULT_IDLEPRIO;
1873 return;
1874 }
1875
1876 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1877 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001878}
1879
Ingo Molnar8159f872007-08-09 11:16:49 +02001880static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001881{
1882 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001883 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001884 p->se.on_rq = 1;
1885}
1886
Ingo Molnar69be72c2007-08-09 11:16:49 +02001887static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001888{
Ingo Molnarf02231e2007-08-09 11:16:48 +02001889 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001890 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001891}
1892
1893/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001894 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001895 */
Ingo Molnar14531182007-07-09 18:51:59 +02001896static inline int __normal_prio(struct task_struct *p)
1897{
Ingo Molnardd41f592007-07-09 18:51:59 +02001898 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001899}
1900
1901/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001902 * Calculate the expected normal priority: i.e. priority
1903 * without taking RT-inheritance into account. Might be
1904 * boosted by interactivity modifiers. Changes upon fork,
1905 * setprio syscalls, and whenever the interactivity
1906 * estimator recalculates.
1907 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001908static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001909{
1910 int prio;
1911
Ingo Molnare05606d2007-07-09 18:51:59 +02001912 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001913 prio = MAX_RT_PRIO-1 - p->rt_priority;
1914 else
1915 prio = __normal_prio(p);
1916 return prio;
1917}
1918
1919/*
1920 * Calculate the current priority, i.e. the priority
1921 * taken into account by the scheduler. This value might
1922 * be boosted by RT tasks, or might be boosted by
1923 * interactivity modifiers. Will be RT if the task got
1924 * RT-boosted. If not then it returns p->normal_prio.
1925 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001926static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001927{
1928 p->normal_prio = normal_prio(p);
1929 /*
1930 * If we are RT tasks or we were boosted to RT priority,
1931 * keep the priority unchanged. Otherwise, update priority
1932 * to the normal priority:
1933 */
1934 if (!rt_prio(p->prio))
1935 return p->normal_prio;
1936 return p->prio;
1937}
1938
1939/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001940 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001942static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001944 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001945 rq->nr_uninterruptible--;
1946
Ingo Molnar8159f872007-08-09 11:16:49 +02001947 enqueue_task(rq, p, wakeup);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001948 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949}
1950
1951/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 * deactivate_task - remove a task from the runqueue.
1953 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001954static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001956 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001957 rq->nr_uninterruptible++;
1958
Ingo Molnar69be72c2007-08-09 11:16:49 +02001959 dequeue_task(rq, p, sleep);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001960 dec_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961}
1962
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963/**
1964 * task_curr - is this task currently executing on a CPU?
1965 * @p: the task in question.
1966 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001967inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968{
1969 return cpu_curr(task_cpu(p)) == p;
1970}
1971
Peter Williams2dd73a42006-06-27 02:54:34 -07001972/* Used instead of source_load when we know the type == 0 */
1973unsigned long weighted_cpuload(const int cpu)
1974{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02001975 return cpu_rq(cpu)->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02001976}
1977
1978static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1979{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001980 set_task_rq(p, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001981#ifdef CONFIG_SMP
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001982 /*
1983 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1984 * successfuly executed on another CPU. We must ensure that updates of
1985 * per-task data have been completed by this moment.
1986 */
1987 smp_wmb();
Ingo Molnardd41f592007-07-09 18:51:59 +02001988 task_thread_info(p)->cpu = cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02001989#endif
Peter Williams2dd73a42006-06-27 02:54:34 -07001990}
1991
Steven Rostedtcb469842008-01-25 21:08:22 +01001992static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1993 const struct sched_class *prev_class,
1994 int oldprio, int running)
1995{
1996 if (prev_class != p->sched_class) {
1997 if (prev_class->switched_from)
1998 prev_class->switched_from(rq, p, running);
1999 p->sched_class->switched_to(rq, p, running);
2000 } else
2001 p->sched_class->prio_changed(rq, p, oldprio, running);
2002}
2003
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004#ifdef CONFIG_SMP
Ingo Molnarc65cc872007-07-09 18:51:58 +02002005
Ingo Molnarcc367732007-10-15 17:00:18 +02002006/*
2007 * Is this task likely cache-hot:
2008 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002009static int
Ingo Molnarcc367732007-10-15 17:00:18 +02002010task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2011{
2012 s64 delta;
2013
Ingo Molnarf540a602008-03-15 17:10:34 +01002014 /*
2015 * Buddy candidates are cache hot:
2016 */
Ingo Molnard25ce4c2008-03-17 09:36:53 +01002017 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
Ingo Molnarf540a602008-03-15 17:10:34 +01002018 return 1;
2019
Ingo Molnarcc367732007-10-15 17:00:18 +02002020 if (p->sched_class != &fair_sched_class)
2021 return 0;
2022
Ingo Molnar6bc16652007-10-15 17:00:18 +02002023 if (sysctl_sched_migration_cost == -1)
2024 return 1;
2025 if (sysctl_sched_migration_cost == 0)
2026 return 0;
2027
Ingo Molnarcc367732007-10-15 17:00:18 +02002028 delta = now - p->se.exec_start;
2029
2030 return delta < (s64)sysctl_sched_migration_cost;
2031}
2032
2033
Ingo Molnardd41f592007-07-09 18:51:59 +02002034void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02002035{
Ingo Molnardd41f592007-07-09 18:51:59 +02002036 int old_cpu = task_cpu(p);
2037 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02002038 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2039 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnarbbdba7c2007-10-15 17:00:06 +02002040 u64 clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02002041
2042 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002043
2044#ifdef CONFIG_SCHEDSTATS
2045 if (p->se.wait_start)
2046 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02002047 if (p->se.sleep_start)
2048 p->se.sleep_start -= clock_offset;
2049 if (p->se.block_start)
2050 p->se.block_start -= clock_offset;
Ingo Molnarcc367732007-10-15 17:00:18 +02002051 if (old_cpu != new_cpu) {
2052 schedstat_inc(p, se.nr_migrations);
2053 if (task_hot(p, old_rq->clock, NULL))
2054 schedstat_inc(p, se.nr_forced2_migrations);
2055 }
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002056#endif
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02002057 p->se.vruntime -= old_cfsrq->min_vruntime -
2058 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02002059
2060 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02002061}
2062
Ingo Molnar70b97a72006-07-03 00:25:42 -07002063struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065
Ingo Molnar36c8b582006-07-03 00:25:41 -07002066 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 int dest_cpu;
2068
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002070};
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071
2072/*
2073 * The task's runqueue lock must be held.
2074 * Returns true if you have to wait for migration thread.
2075 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002076static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07002077migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002079 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080
2081 /*
2082 * If the task is not on a runqueue (and not running), then
2083 * it is sufficient to simply update the task's cpu field.
2084 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002085 if (!p->se.on_rq && !task_running(rq, p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 set_task_cpu(p, dest_cpu);
2087 return 0;
2088 }
2089
2090 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 req->task = p;
2092 req->dest_cpu = dest_cpu;
2093 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002094
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 return 1;
2096}
2097
2098/*
2099 * wait_task_inactive - wait for a thread to unschedule.
2100 *
2101 * The caller must ensure that the task *will* unschedule sometime soon,
2102 * else this function might spin for a *long* time. This function can't
2103 * be called with interrupts off, or it may introduce deadlock with
2104 * smp_call_function() if an IPI is sent by the same process we are
2105 * waiting to become inactive.
2106 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002107void wait_task_inactive(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108{
2109 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002110 int running, on_rq;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002111 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112
Andi Kleen3a5c3592007-10-15 17:00:14 +02002113 for (;;) {
2114 /*
2115 * We do the initial early heuristics without holding
2116 * any task-queue locks at all. We'll only try to get
2117 * the runqueue lock when things look like they will
2118 * work out!
2119 */
2120 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002121
Andi Kleen3a5c3592007-10-15 17:00:14 +02002122 /*
2123 * If the task is actively running on another CPU
2124 * still, just relax and busy-wait without holding
2125 * any locks.
2126 *
2127 * NOTE! Since we don't hold any locks, it's not
2128 * even sure that "rq" stays as the right runqueue!
2129 * But we don't care, since "task_running()" will
2130 * return false if the runqueue has changed and p
2131 * is actually now running somewhere else!
2132 */
2133 while (task_running(rq, p))
2134 cpu_relax();
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002135
Andi Kleen3a5c3592007-10-15 17:00:14 +02002136 /*
2137 * Ok, time to look more closely! We need the rq
2138 * lock now, to be *sure*. If we're wrong, we'll
2139 * just go back and repeat.
2140 */
2141 rq = task_rq_lock(p, &flags);
2142 running = task_running(rq, p);
2143 on_rq = p->se.on_rq;
2144 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002145
Andi Kleen3a5c3592007-10-15 17:00:14 +02002146 /*
2147 * Was it really running after all now that we
2148 * checked with the proper locks actually held?
2149 *
2150 * Oops. Go back and try again..
2151 */
2152 if (unlikely(running)) {
2153 cpu_relax();
2154 continue;
2155 }
2156
2157 /*
2158 * It's not enough that it's not actively running,
2159 * it must be off the runqueue _entirely_, and not
2160 * preempted!
2161 *
2162 * So if it wa still runnable (but just not actively
2163 * running right now), it's preempted, and we should
2164 * yield - it could be a while.
2165 */
2166 if (unlikely(on_rq)) {
2167 schedule_timeout_uninterruptible(1);
2168 continue;
2169 }
2170
2171 /*
2172 * Ahh, all good. It wasn't running, and it wasn't
2173 * runnable, which means that it will never become
2174 * running in the future either. We're all done!
2175 */
2176 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178}
2179
2180/***
2181 * kick_process - kick a running thread to enter/exit the kernel
2182 * @p: the to-be-kicked thread
2183 *
2184 * Cause a process which is running on another CPU to enter
2185 * kernel-mode, without any delay. (to get signals handled.)
2186 *
2187 * NOTE: this function doesnt have to take the runqueue lock,
2188 * because all it wants to ensure is that the remote task enters
2189 * the kernel. If the IPI races and the task has been migrated
2190 * to another CPU then no harm is done and the purpose has been
2191 * achieved as well.
2192 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002193void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194{
2195 int cpu;
2196
2197 preempt_disable();
2198 cpu = task_cpu(p);
2199 if ((cpu != smp_processor_id()) && task_curr(p))
2200 smp_send_reschedule(cpu);
2201 preempt_enable();
2202}
2203
2204/*
Peter Williams2dd73a42006-06-27 02:54:34 -07002205 * Return a low guess at the load of a migration-source cpu weighted
2206 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 *
2208 * We want to under-estimate the load of migration sources, to
2209 * balance conservatively.
2210 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002211static unsigned long source_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08002212{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002213 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002214 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08002215
Peter Williams2dd73a42006-06-27 02:54:34 -07002216 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02002217 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07002218
Ingo Molnardd41f592007-07-09 18:51:59 +02002219 return min(rq->cpu_load[type-1], total);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220}
2221
2222/*
Peter Williams2dd73a42006-06-27 02:54:34 -07002223 * Return a high guess at the load of a migration-target cpu weighted
2224 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002226static unsigned long target_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08002227{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002228 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002229 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08002230
Peter Williams2dd73a42006-06-27 02:54:34 -07002231 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02002232 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07002233
Ingo Molnardd41f592007-07-09 18:51:59 +02002234 return max(rq->cpu_load[type-1], total);
Peter Williams2dd73a42006-06-27 02:54:34 -07002235}
2236
2237/*
2238 * Return the average load per task on the cpu's run queue
2239 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002240static unsigned long cpu_avg_load_per_task(int cpu)
Peter Williams2dd73a42006-06-27 02:54:34 -07002241{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002242 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002243 unsigned long total = weighted_cpuload(cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07002244 unsigned long n = rq->nr_running;
2245
Ingo Molnardd41f592007-07-09 18:51:59 +02002246 return n ? total / n : SCHED_LOAD_SCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247}
2248
Nick Piggin147cbb42005-06-25 14:57:19 -07002249/*
2250 * find_idlest_group finds and returns the least busy CPU group within the
2251 * domain.
2252 */
2253static struct sched_group *
2254find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2255{
2256 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2257 unsigned long min_load = ULONG_MAX, this_load = 0;
2258 int load_idx = sd->forkexec_idx;
2259 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2260
2261 do {
2262 unsigned long load, avg_load;
2263 int local_group;
2264 int i;
2265
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002266 /* Skip over this group if it has no CPUs allowed */
2267 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
Andi Kleen3a5c3592007-10-15 17:00:14 +02002268 continue;
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002269
Nick Piggin147cbb42005-06-25 14:57:19 -07002270 local_group = cpu_isset(this_cpu, group->cpumask);
Nick Piggin147cbb42005-06-25 14:57:19 -07002271
2272 /* Tally up the load of all CPUs in the group */
2273 avg_load = 0;
2274
2275 for_each_cpu_mask(i, group->cpumask) {
2276 /* Bias balancing toward cpus of our domain */
2277 if (local_group)
2278 load = source_load(i, load_idx);
2279 else
2280 load = target_load(i, load_idx);
2281
2282 avg_load += load;
2283 }
2284
2285 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07002286 avg_load = sg_div_cpu_power(group,
2287 avg_load * SCHED_LOAD_SCALE);
Nick Piggin147cbb42005-06-25 14:57:19 -07002288
2289 if (local_group) {
2290 this_load = avg_load;
2291 this = group;
2292 } else if (avg_load < min_load) {
2293 min_load = avg_load;
2294 idlest = group;
2295 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02002296 } while (group = group->next, group != sd->groups);
Nick Piggin147cbb42005-06-25 14:57:19 -07002297
2298 if (!idlest || 100*this_load < imbalance*min_load)
2299 return NULL;
2300 return idlest;
2301}
2302
2303/*
Satoru Takeuchi0feaece2006-10-03 01:14:10 -07002304 * find_idlest_cpu - find the idlest cpu among the cpus in group.
Nick Piggin147cbb42005-06-25 14:57:19 -07002305 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002306static int
Mike Travis7c16ec52008-04-04 18:11:11 -07002307find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2308 cpumask_t *tmp)
Nick Piggin147cbb42005-06-25 14:57:19 -07002309{
2310 unsigned long load, min_load = ULONG_MAX;
2311 int idlest = -1;
2312 int i;
2313
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002314 /* Traverse only the allowed CPUs */
Mike Travis7c16ec52008-04-04 18:11:11 -07002315 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002316
Mike Travis7c16ec52008-04-04 18:11:11 -07002317 for_each_cpu_mask(i, *tmp) {
Peter Williams2dd73a42006-06-27 02:54:34 -07002318 load = weighted_cpuload(i);
Nick Piggin147cbb42005-06-25 14:57:19 -07002319
2320 if (load < min_load || (load == min_load && i == this_cpu)) {
2321 min_load = load;
2322 idlest = i;
2323 }
2324 }
2325
2326 return idlest;
2327}
2328
Nick Piggin476d1392005-06-25 14:57:29 -07002329/*
2330 * sched_balance_self: balance the current task (running on cpu) in domains
2331 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2332 * SD_BALANCE_EXEC.
2333 *
2334 * Balance, ie. select the least loaded group.
2335 *
2336 * Returns the target CPU number, or the same CPU if no balancing is needed.
2337 *
2338 * preempt must be disabled.
2339 */
2340static int sched_balance_self(int cpu, int flag)
2341{
2342 struct task_struct *t = current;
2343 struct sched_domain *tmp, *sd = NULL;
Nick Piggin147cbb42005-06-25 14:57:19 -07002344
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002345 for_each_domain(cpu, tmp) {
Ingo Molnar9761eea2007-07-09 18:52:00 +02002346 /*
2347 * If power savings logic is enabled for a domain, stop there.
2348 */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002349 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2350 break;
Nick Piggin476d1392005-06-25 14:57:29 -07002351 if (tmp->flags & flag)
2352 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002353 }
Nick Piggin476d1392005-06-25 14:57:29 -07002354
2355 while (sd) {
Mike Travis7c16ec52008-04-04 18:11:11 -07002356 cpumask_t span, tmpmask;
Nick Piggin476d1392005-06-25 14:57:29 -07002357 struct sched_group *group;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002358 int new_cpu, weight;
2359
2360 if (!(sd->flags & flag)) {
2361 sd = sd->child;
2362 continue;
2363 }
Nick Piggin476d1392005-06-25 14:57:29 -07002364
2365 span = sd->span;
2366 group = find_idlest_group(sd, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002367 if (!group) {
2368 sd = sd->child;
2369 continue;
2370 }
Nick Piggin476d1392005-06-25 14:57:29 -07002371
Mike Travis7c16ec52008-04-04 18:11:11 -07002372 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002373 if (new_cpu == -1 || new_cpu == cpu) {
2374 /* Now try balancing at a lower domain level of cpu */
2375 sd = sd->child;
2376 continue;
2377 }
Nick Piggin476d1392005-06-25 14:57:29 -07002378
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002379 /* Now try balancing at a lower domain level of new_cpu */
Nick Piggin476d1392005-06-25 14:57:29 -07002380 cpu = new_cpu;
Nick Piggin476d1392005-06-25 14:57:29 -07002381 sd = NULL;
2382 weight = cpus_weight(span);
2383 for_each_domain(cpu, tmp) {
2384 if (weight <= cpus_weight(tmp->span))
2385 break;
2386 if (tmp->flags & flag)
2387 sd = tmp;
2388 }
2389 /* while loop will break here if sd == NULL */
2390 }
2391
2392 return cpu;
2393}
2394
2395#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397/***
2398 * try_to_wake_up - wake up a thread
2399 * @p: the to-be-woken-up thread
2400 * @state: the mask of task states that can be woken
2401 * @sync: do a synchronous wakeup?
2402 *
2403 * Put it on the run-queue if it's not already there. The "current"
2404 * thread is always on the run-queue (except when the actual
2405 * re-schedule is in progress), and as such you're allowed to do
2406 * the simpler "current->state = TASK_RUNNING" to mark yourself
2407 * runnable without the overhead of this.
2408 *
2409 * returns failure only if the task is already active.
2410 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002411static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412{
Ingo Molnarcc367732007-10-15 17:00:18 +02002413 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414 unsigned long flags;
2415 long old_state;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002416 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417
Ingo Molnarb85d0662008-03-16 20:03:22 +01002418 if (!sched_feat(SYNC_WAKEUPS))
2419 sync = 0;
2420
Linus Torvalds04e2f172008-02-23 18:05:03 -08002421 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422 rq = task_rq_lock(p, &flags);
2423 old_state = p->state;
2424 if (!(old_state & state))
2425 goto out;
2426
Ingo Molnardd41f592007-07-09 18:51:59 +02002427 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 goto out_running;
2429
2430 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02002431 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 this_cpu = smp_processor_id();
2433
2434#ifdef CONFIG_SMP
2435 if (unlikely(task_running(rq, p)))
2436 goto out_activate;
2437
Dmitry Adamushko5d2f5a62008-01-25 21:08:21 +01002438 cpu = p->sched_class->select_task_rq(p, sync);
2439 if (cpu != orig_cpu) {
2440 set_task_cpu(p, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 task_rq_unlock(rq, &flags);
2442 /* might preempt at this point */
2443 rq = task_rq_lock(p, &flags);
2444 old_state = p->state;
2445 if (!(old_state & state))
2446 goto out;
Ingo Molnardd41f592007-07-09 18:51:59 +02002447 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 goto out_running;
2449
2450 this_cpu = smp_processor_id();
2451 cpu = task_cpu(p);
2452 }
2453
Gregory Haskinse7693a32008-01-25 21:08:09 +01002454#ifdef CONFIG_SCHEDSTATS
2455 schedstat_inc(rq, ttwu_count);
2456 if (cpu == this_cpu)
2457 schedstat_inc(rq, ttwu_local);
2458 else {
2459 struct sched_domain *sd;
2460 for_each_domain(this_cpu, sd) {
2461 if (cpu_isset(cpu, sd->span)) {
2462 schedstat_inc(sd, ttwu_wake_remote);
2463 break;
2464 }
2465 }
2466 }
Gregory Haskinse7693a32008-01-25 21:08:09 +01002467#endif
2468
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469out_activate:
2470#endif /* CONFIG_SMP */
Ingo Molnarbd3bff92008-05-12 21:20:41 +02002471 ftrace_wake_up_task(p, rq->curr);
Ingo Molnarcc367732007-10-15 17:00:18 +02002472 schedstat_inc(p, se.nr_wakeups);
2473 if (sync)
2474 schedstat_inc(p, se.nr_wakeups_sync);
2475 if (orig_cpu != cpu)
2476 schedstat_inc(p, se.nr_wakeups_migrate);
2477 if (cpu == this_cpu)
2478 schedstat_inc(p, se.nr_wakeups_local);
2479 else
2480 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnar2daa3572007-08-09 11:16:51 +02002481 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02002482 activate_task(rq, p, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 success = 1;
2484
2485out_running:
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002486 check_preempt_curr(rq, p);
2487
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 p->state = TASK_RUNNING;
Steven Rostedt9a897c52008-01-25 21:08:22 +01002489#ifdef CONFIG_SMP
2490 if (p->sched_class->task_wake_up)
2491 p->sched_class->task_wake_up(rq, p);
2492#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493out:
2494 task_rq_unlock(rq, &flags);
2495
2496 return success;
2497}
2498
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002499int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05002501 return try_to_wake_up(p, TASK_ALL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503EXPORT_SYMBOL(wake_up_process);
2504
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002505int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506{
2507 return try_to_wake_up(p, state, 0);
2508}
2509
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510/*
2511 * Perform scheduler related setup for a newly forked process p.
2512 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02002513 *
2514 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002516static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517{
Ingo Molnardd41f592007-07-09 18:51:59 +02002518 p->se.exec_start = 0;
2519 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02002520 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002521 p->se.last_wakeup = 0;
2522 p->se.avg_overlap = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002523
2524#ifdef CONFIG_SCHEDSTATS
2525 p->se.wait_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002526 p->se.sum_sleep_runtime = 0;
2527 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002528 p->se.block_start = 0;
2529 p->se.sleep_max = 0;
2530 p->se.block_max = 0;
2531 p->se.exec_max = 0;
Ingo Molnareba1ed42007-10-15 17:00:02 +02002532 p->se.slice_max = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002533 p->se.wait_max = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002534#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002535
Peter Zijlstrafa717062008-01-25 21:08:27 +01002536 INIT_LIST_HEAD(&p->rt.run_list);
Ingo Molnardd41f592007-07-09 18:51:59 +02002537 p->se.on_rq = 0;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02002538 INIT_LIST_HEAD(&p->se.group_node);
Nick Piggin476d1392005-06-25 14:57:29 -07002539
Avi Kivitye107be32007-07-26 13:40:43 +02002540#ifdef CONFIG_PREEMPT_NOTIFIERS
2541 INIT_HLIST_HEAD(&p->preempt_notifiers);
2542#endif
2543
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 /*
2545 * We mark the process as running here, but have not actually
2546 * inserted it onto the runqueue yet. This guarantees that
2547 * nobody will actually run it, and a signal or other external
2548 * event cannot wake it up and insert it on the runqueue either.
2549 */
2550 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02002551}
2552
2553/*
2554 * fork()/clone()-time setup:
2555 */
2556void sched_fork(struct task_struct *p, int clone_flags)
2557{
2558 int cpu = get_cpu();
2559
2560 __sched_fork(p);
2561
2562#ifdef CONFIG_SMP
2563 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2564#endif
Ingo Molnar02e4bac2007-10-15 17:00:11 +02002565 set_task_cpu(p, cpu);
Ingo Molnarb29739f2006-06-27 02:54:51 -07002566
2567 /*
2568 * Make sure we do not leak PI boosting priority to the child:
2569 */
2570 p->prio = current->normal_prio;
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02002571 if (!rt_prio(p->prio))
2572 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07002573
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002574#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02002575 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002576 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08002578#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07002579 p->oncpu = 0;
2580#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07002582 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08002583 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002585 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586}
2587
2588/*
2589 * wake_up_new_task - wake up a newly created task for the first time.
2590 *
2591 * This function will do some initial scheduler statistics housekeeping
2592 * that must be done for every newly created context, then puts the task
2593 * on the runqueue and wakes it.
2594 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002595void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596{
2597 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002598 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599
2600 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02002602 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603
2604 p->prio = effective_prio(p);
2605
Srivatsa Vaddagirib9dca1e2007-10-17 16:55:11 +02002606 if (!p->sched_class->task_new || !current->se.on_rq) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002607 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02002610 * Let the scheduling class do new task startup
2611 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02002613 p->sched_class->task_new(rq, p);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02002614 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615 }
Ingo Molnarbd3bff92008-05-12 21:20:41 +02002616 ftrace_wake_up_new_task(p, rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02002617 check_preempt_curr(rq, p);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002618#ifdef CONFIG_SMP
2619 if (p->sched_class->task_wake_up)
2620 p->sched_class->task_wake_up(rq, p);
2621#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02002622 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623}
2624
Avi Kivitye107be32007-07-26 13:40:43 +02002625#ifdef CONFIG_PREEMPT_NOTIFIERS
2626
2627/**
Randy Dunlap421cee22007-07-31 00:37:50 -07002628 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2629 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002630 */
2631void preempt_notifier_register(struct preempt_notifier *notifier)
2632{
2633 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2634}
2635EXPORT_SYMBOL_GPL(preempt_notifier_register);
2636
2637/**
2638 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002639 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002640 *
2641 * This is safe to call from within a preemption notifier.
2642 */
2643void preempt_notifier_unregister(struct preempt_notifier *notifier)
2644{
2645 hlist_del(&notifier->link);
2646}
2647EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2648
2649static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2650{
2651 struct preempt_notifier *notifier;
2652 struct hlist_node *node;
2653
2654 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2655 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2656}
2657
2658static void
2659fire_sched_out_preempt_notifiers(struct task_struct *curr,
2660 struct task_struct *next)
2661{
2662 struct preempt_notifier *notifier;
2663 struct hlist_node *node;
2664
2665 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2666 notifier->ops->sched_out(notifier, next);
2667}
2668
2669#else
2670
2671static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2672{
2673}
2674
2675static void
2676fire_sched_out_preempt_notifiers(struct task_struct *curr,
2677 struct task_struct *next)
2678{
2679}
2680
2681#endif
2682
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002684 * prepare_task_switch - prepare to switch tasks
2685 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002686 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002687 * @next: the task we are going to switch to.
2688 *
2689 * This is called with the rq lock held and interrupts off. It must
2690 * be paired with a subsequent finish_task_switch after the context
2691 * switch.
2692 *
2693 * prepare_task_switch sets up locking and calls architecture specific
2694 * hooks.
2695 */
Avi Kivitye107be32007-07-26 13:40:43 +02002696static inline void
2697prepare_task_switch(struct rq *rq, struct task_struct *prev,
2698 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002699{
Avi Kivitye107be32007-07-26 13:40:43 +02002700 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002701 prepare_lock_switch(rq, next);
2702 prepare_arch_switch(next);
2703}
2704
2705/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04002707 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 * @prev: the thread we just switched away from.
2709 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002710 * finish_task_switch must be called after the context switch, paired
2711 * with a prepare_task_switch call before the context switch.
2712 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2713 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 *
2715 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002716 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 * with the lock held can cause deadlocks; see schedule() for
2718 * details.)
2719 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002720static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 __releases(rq->lock)
2722{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002724 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725
2726 rq->prev_mm = NULL;
2727
2728 /*
2729 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002730 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002731 * schedule one last time. The schedule call will never return, and
2732 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002733 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734 * still held, otherwise prev could be scheduled on another cpu, die
2735 * there before we look at prev->state, and then the reference would
2736 * be dropped twice.
2737 * Manfred Spraul <manfred@colorfullife.com>
2738 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002739 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07002740 finish_arch_switch(prev);
2741 finish_lock_switch(rq, prev);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002742#ifdef CONFIG_SMP
2743 if (current->sched_class->post_schedule)
2744 current->sched_class->post_schedule(rq);
2745#endif
Steven Rostedte8fa1362008-01-25 21:08:05 +01002746
Avi Kivitye107be32007-07-26 13:40:43 +02002747 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 if (mm)
2749 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002750 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08002751 /*
2752 * Remove function-return probe instances associated with this
2753 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002754 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002755 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002757 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758}
2759
2760/**
2761 * schedule_tail - first thing a freshly forked thread must call.
2762 * @prev: the thread we just switched away from.
2763 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002764asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 __releases(rq->lock)
2766{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002767 struct rq *rq = this_rq();
2768
Nick Piggin4866cde2005-06-25 14:57:23 -07002769 finish_task_switch(rq, prev);
2770#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2771 /* In this case, finish_task_switch does not reenable preemption */
2772 preempt_enable();
2773#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002775 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776}
2777
2778/*
2779 * context_switch - switch to the new MM and the new
2780 * thread's register state.
2781 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002782static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002783context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002784 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785{
Ingo Molnardd41f592007-07-09 18:51:59 +02002786 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787
Avi Kivitye107be32007-07-26 13:40:43 +02002788 prepare_task_switch(rq, prev, next);
Ingo Molnarbd3bff92008-05-12 21:20:41 +02002789 ftrace_ctx_switch(prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02002790 mm = next->mm;
2791 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002792 /*
2793 * For paravirt, this is coupled with an exit in switch_to to
2794 * combine the page table reload and the switch backend into
2795 * one hypercall.
2796 */
2797 arch_enter_lazy_cpu_mode();
2798
Ingo Molnardd41f592007-07-09 18:51:59 +02002799 if (unlikely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 next->active_mm = oldmm;
2801 atomic_inc(&oldmm->mm_count);
2802 enter_lazy_tlb(oldmm, next);
2803 } else
2804 switch_mm(oldmm, mm, next);
2805
Ingo Molnardd41f592007-07-09 18:51:59 +02002806 if (unlikely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 rq->prev_mm = oldmm;
2809 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002810 /*
2811 * Since the runqueue lock will be released by the next
2812 * task (which is an invalid locking op but in the case
2813 * of the scheduler it's an obvious special-case), so we
2814 * do an early lockdep release here:
2815 */
2816#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002817 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002818#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819
2820 /* Here we just switch the register state and the stack. */
2821 switch_to(prev, next, prev);
2822
Ingo Molnardd41f592007-07-09 18:51:59 +02002823 barrier();
2824 /*
2825 * this_rq must be evaluated again because prev may have moved
2826 * CPUs since it called schedule(), thus the 'rq' on its stack
2827 * frame will be invalid.
2828 */
2829 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830}
2831
2832/*
2833 * nr_running, nr_uninterruptible and nr_context_switches:
2834 *
2835 * externally visible scheduler statistics: current number of runnable
2836 * threads, current number of uninterruptible-sleeping threads, total
2837 * number of context switches performed since bootup.
2838 */
2839unsigned long nr_running(void)
2840{
2841 unsigned long i, sum = 0;
2842
2843 for_each_online_cpu(i)
2844 sum += cpu_rq(i)->nr_running;
2845
2846 return sum;
2847}
2848
2849unsigned long nr_uninterruptible(void)
2850{
2851 unsigned long i, sum = 0;
2852
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002853 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 sum += cpu_rq(i)->nr_uninterruptible;
2855
2856 /*
2857 * Since we read the counters lockless, it might be slightly
2858 * inaccurate. Do not allow it to go below zero though:
2859 */
2860 if (unlikely((long)sum < 0))
2861 sum = 0;
2862
2863 return sum;
2864}
2865
2866unsigned long long nr_context_switches(void)
2867{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002868 int i;
2869 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002871 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 sum += cpu_rq(i)->nr_switches;
2873
2874 return sum;
2875}
2876
2877unsigned long nr_iowait(void)
2878{
2879 unsigned long i, sum = 0;
2880
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002881 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2883
2884 return sum;
2885}
2886
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002887unsigned long nr_active(void)
2888{
2889 unsigned long i, running = 0, uninterruptible = 0;
2890
2891 for_each_online_cpu(i) {
2892 running += cpu_rq(i)->nr_running;
2893 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2894 }
2895
2896 if (unlikely((long)uninterruptible < 0))
2897 uninterruptible = 0;
2898
2899 return running + uninterruptible;
2900}
2901
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902/*
Ingo Molnardd41f592007-07-09 18:51:59 +02002903 * Update rq->cpu_load[] statistics. This function is usually called every
2904 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07002905 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002906static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002907{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02002908 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02002909 int i, scale;
2910
2911 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02002912
2913 /* Update our load: */
2914 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2915 unsigned long old_load, new_load;
2916
2917 /* scale is effectively 1 << i now, and >> i divides by scale */
2918
2919 old_load = this_rq->cpu_load[i];
2920 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02002921 /*
2922 * Round up the averaging division if load is increasing. This
2923 * prevents us from getting stuck on 9 if the load is 10, for
2924 * example.
2925 */
2926 if (new_load > old_load)
2927 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02002928 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2929 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002930}
2931
Ingo Molnardd41f592007-07-09 18:51:59 +02002932#ifdef CONFIG_SMP
2933
Ingo Molnar48f24c42006-07-03 00:25:40 -07002934/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 * double_rq_lock - safely lock two runqueues
2936 *
2937 * Note this does not disable interrupts like task_rq_lock,
2938 * you need to do so manually before calling.
2939 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002940static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 __acquires(rq1->lock)
2942 __acquires(rq2->lock)
2943{
Kirill Korotaev054b9102006-12-10 02:20:11 -08002944 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 if (rq1 == rq2) {
2946 spin_lock(&rq1->lock);
2947 __acquire(rq2->lock); /* Fake it out ;) */
2948 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002949 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 spin_lock(&rq1->lock);
2951 spin_lock(&rq2->lock);
2952 } else {
2953 spin_lock(&rq2->lock);
2954 spin_lock(&rq1->lock);
2955 }
2956 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002957 update_rq_clock(rq1);
2958 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959}
2960
2961/*
2962 * double_rq_unlock - safely unlock two runqueues
2963 *
2964 * Note this does not restore interrupts like task_rq_unlock,
2965 * you need to do so manually after calling.
2966 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002967static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002968 __releases(rq1->lock)
2969 __releases(rq2->lock)
2970{
2971 spin_unlock(&rq1->lock);
2972 if (rq1 != rq2)
2973 spin_unlock(&rq2->lock);
2974 else
2975 __release(rq2->lock);
2976}
2977
2978/*
2979 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2980 */
Steven Rostedte8fa1362008-01-25 21:08:05 +01002981static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 __releases(this_rq->lock)
2983 __acquires(busiest->lock)
2984 __acquires(this_rq->lock)
2985{
Steven Rostedte8fa1362008-01-25 21:08:05 +01002986 int ret = 0;
2987
Kirill Korotaev054b9102006-12-10 02:20:11 -08002988 if (unlikely(!irqs_disabled())) {
2989 /* printk() doesn't work good under rq->lock */
2990 spin_unlock(&this_rq->lock);
2991 BUG_ON(1);
2992 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 if (unlikely(!spin_trylock(&busiest->lock))) {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002994 if (busiest < this_rq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 spin_unlock(&this_rq->lock);
2996 spin_lock(&busiest->lock);
2997 spin_lock(&this_rq->lock);
Steven Rostedte8fa1362008-01-25 21:08:05 +01002998 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 } else
3000 spin_lock(&busiest->lock);
3001 }
Steven Rostedte8fa1362008-01-25 21:08:05 +01003002 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003}
3004
3005/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 * If dest_cpu is allowed for this process, migrate the task to it.
3007 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003008 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 * the cpu_allowed mask is restored.
3010 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003011static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003013 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003015 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016
3017 rq = task_rq_lock(p, &flags);
3018 if (!cpu_isset(dest_cpu, p->cpus_allowed)
3019 || unlikely(cpu_is_offline(dest_cpu)))
3020 goto out;
3021
3022 /* force the process onto the specified CPU */
3023 if (migrate_task(p, dest_cpu, &req)) {
3024 /* Need to wait for migration thread (might exit: take ref). */
3025 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07003026
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 get_task_struct(mt);
3028 task_rq_unlock(rq, &flags);
3029 wake_up_process(mt);
3030 put_task_struct(mt);
3031 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07003032
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033 return;
3034 }
3035out:
3036 task_rq_unlock(rq, &flags);
3037}
3038
3039/*
Nick Piggin476d1392005-06-25 14:57:29 -07003040 * sched_exec - execve() is a valuable balancing opportunity, because at
3041 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042 */
3043void sched_exec(void)
3044{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 int new_cpu, this_cpu = get_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07003046 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07003048 if (new_cpu != this_cpu)
3049 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050}
3051
3052/*
3053 * pull_task - move a task from a remote runqueue to the local runqueue.
3054 * Both runqueues must be locked.
3055 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003056static void pull_task(struct rq *src_rq, struct task_struct *p,
3057 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02003059 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003061 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062 /*
3063 * Note that idle threads have a prio of MAX_PRIO, for this test
3064 * to be always true for them.
3065 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003066 check_preempt_curr(this_rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067}
3068
3069/*
3070 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3071 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08003072static
Ingo Molnar70b97a72006-07-03 00:25:42 -07003073int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003074 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003075 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076{
3077 /*
3078 * We do not migrate tasks that are:
3079 * 1) running (obviously), or
3080 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3081 * 3) are cache-hot on their current CPU.
3082 */
Ingo Molnarcc367732007-10-15 17:00:18 +02003083 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3084 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003086 }
Nick Piggin81026792005-06-25 14:57:07 -07003087 *all_pinned = 0;
3088
Ingo Molnarcc367732007-10-15 17:00:18 +02003089 if (task_running(rq, p)) {
3090 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07003091 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003092 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093
Ingo Molnarda84d962007-10-15 17:00:18 +02003094 /*
3095 * Aggressive migration if:
3096 * 1) task is cache cold, or
3097 * 2) too many balance attempts have failed.
3098 */
3099
Ingo Molnar6bc16652007-10-15 17:00:18 +02003100 if (!task_hot(p, rq->clock, sd) ||
3101 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003102#ifdef CONFIG_SCHEDSTATS
Ingo Molnarcc367732007-10-15 17:00:18 +02003103 if (task_hot(p, rq->clock, sd)) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003104 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02003105 schedstat_inc(p, se.nr_forced_migrations);
3106 }
Ingo Molnarda84d962007-10-15 17:00:18 +02003107#endif
3108 return 1;
3109 }
3110
Ingo Molnarcc367732007-10-15 17:00:18 +02003111 if (task_hot(p, rq->clock, sd)) {
3112 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02003113 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003114 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115 return 1;
3116}
3117
Peter Williamse1d14842007-10-24 18:23:51 +02003118static unsigned long
3119balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3120 unsigned long max_load_move, struct sched_domain *sd,
3121 enum cpu_idle_type idle, int *all_pinned,
3122 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02003123{
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003124 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
Ingo Molnardd41f592007-07-09 18:51:59 +02003125 struct task_struct *p;
3126 long rem_load_move = max_load_move;
3127
Peter Williamse1d14842007-10-24 18:23:51 +02003128 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02003129 goto out;
3130
3131 pinned = 1;
3132
3133 /*
3134 * Start the load-balancing iterator:
3135 */
3136 p = iterator->start(iterator->arg);
3137next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003138 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02003139 goto out;
3140 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003141 * To help distribute high priority tasks across CPUs we don't
Ingo Molnardd41f592007-07-09 18:51:59 +02003142 * skip a task if it will be the highest priority task (i.e. smallest
3143 * prio value) on its new queue regardless of its load weight
3144 */
3145 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3146 SCHED_LOAD_SCALE_FUZZ;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003147 if ((skip_for_load && p->prio >= *this_best_prio) ||
Ingo Molnardd41f592007-07-09 18:51:59 +02003148 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003149 p = iterator->next(iterator->arg);
3150 goto next;
3151 }
3152
3153 pull_task(busiest, p, this_rq, this_cpu);
3154 pulled++;
3155 rem_load_move -= p->se.load.weight;
3156
3157 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003158 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02003159 */
Peter Williamse1d14842007-10-24 18:23:51 +02003160 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003161 if (p->prio < *this_best_prio)
3162 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02003163 p = iterator->next(iterator->arg);
3164 goto next;
3165 }
3166out:
3167 /*
Peter Williamse1d14842007-10-24 18:23:51 +02003168 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02003169 * so we can safely collect pull_task() stats here rather than
3170 * inside pull_task().
3171 */
3172 schedstat_add(sd, lb_gained[idle], pulled);
3173
3174 if (all_pinned)
3175 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02003176
3177 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02003178}
Ingo Molnar48f24c42006-07-03 00:25:40 -07003179
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180/*
Peter Williams43010652007-08-09 11:16:46 +02003181 * move_tasks tries to move up to max_load_move weighted load from busiest to
3182 * this_rq, as part of a balancing operation within domain "sd".
3183 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184 *
3185 * Called with both runqueues locked.
3186 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003187static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02003188 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003189 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07003190 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003191{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003192 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02003193 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003194 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195
Ingo Molnardd41f592007-07-09 18:51:59 +02003196 do {
Peter Williams43010652007-08-09 11:16:46 +02003197 total_load_moved +=
3198 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02003199 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003200 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02003201 class = class->next;
Peter Williams43010652007-08-09 11:16:46 +02003202 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203
Peter Williams43010652007-08-09 11:16:46 +02003204 return total_load_moved > 0;
3205}
3206
Peter Williamse1d14842007-10-24 18:23:51 +02003207static int
3208iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3209 struct sched_domain *sd, enum cpu_idle_type idle,
3210 struct rq_iterator *iterator)
3211{
3212 struct task_struct *p = iterator->start(iterator->arg);
3213 int pinned = 0;
3214
3215 while (p) {
3216 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3217 pull_task(busiest, p, this_rq, this_cpu);
3218 /*
3219 * Right now, this is only the second place pull_task()
3220 * is called, so we can safely collect pull_task()
3221 * stats here rather than inside pull_task().
3222 */
3223 schedstat_inc(sd, lb_gained[idle]);
3224
3225 return 1;
3226 }
3227 p = iterator->next(iterator->arg);
3228 }
3229
3230 return 0;
3231}
3232
Peter Williams43010652007-08-09 11:16:46 +02003233/*
3234 * move_one_task tries to move exactly one task from busiest to this_rq, as
3235 * part of active balancing operations within "domain".
3236 * Returns 1 if successful and 0 otherwise.
3237 *
3238 * Called with both runqueues locked.
3239 */
3240static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3241 struct sched_domain *sd, enum cpu_idle_type idle)
3242{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003243 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02003244
3245 for (class = sched_class_highest; class; class = class->next)
Peter Williamse1d14842007-10-24 18:23:51 +02003246 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02003247 return 1;
3248
3249 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250}
3251
3252/*
3253 * find_busiest_group finds and returns the busiest CPU group within the
Ingo Molnar48f24c42006-07-03 00:25:40 -07003254 * domain. It calculates and returns the amount of weighted load which
3255 * should be moved to restore balance via the imbalance parameter.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 */
3257static struct sched_group *
3258find_busiest_group(struct sched_domain *sd, int this_cpu,
Ingo Molnardd41f592007-07-09 18:51:59 +02003259 unsigned long *imbalance, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003260 int *sd_idle, const cpumask_t *cpus, int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261{
3262 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3263 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003264 unsigned long max_pull;
Peter Williams2dd73a42006-06-27 02:54:34 -07003265 unsigned long busiest_load_per_task, busiest_nr_running;
3266 unsigned long this_load_per_task, this_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02003267 int load_idx, group_imb = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003268#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3269 int power_savings_balance = 1;
3270 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3271 unsigned long min_nr_running = ULONG_MAX;
3272 struct sched_group *group_min = NULL, *group_leader = NULL;
3273#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274
3275 max_load = this_load = total_load = total_pwr = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07003276 busiest_load_per_task = busiest_nr_running = 0;
3277 this_load_per_task = this_nr_running = 0;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003278 if (idle == CPU_NOT_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07003279 load_idx = sd->busy_idx;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003280 else if (idle == CPU_NEWLY_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07003281 load_idx = sd->newidle_idx;
3282 else
3283 load_idx = sd->idle_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284
3285 do {
Ken Chen908a7c12007-10-17 16:55:11 +02003286 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003287 int local_group;
3288 int i;
Ken Chen908a7c12007-10-17 16:55:11 +02003289 int __group_imb = 0;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003290 unsigned int balance_cpu = -1, first_idle_cpu = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07003291 unsigned long sum_nr_running, sum_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292
3293 local_group = cpu_isset(this_cpu, group->cpumask);
3294
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003295 if (local_group)
3296 balance_cpu = first_cpu(group->cpumask);
3297
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 /* Tally up the load of all CPUs in the group */
Peter Williams2dd73a42006-06-27 02:54:34 -07003299 sum_weighted_load = sum_nr_running = avg_load = 0;
Ken Chen908a7c12007-10-17 16:55:11 +02003300 max_cpu_load = 0;
3301 min_cpu_load = ~0UL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302
3303 for_each_cpu_mask(i, group->cpumask) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003304 struct rq *rq;
3305
3306 if (!cpu_isset(i, *cpus))
3307 continue;
3308
3309 rq = cpu_rq(i);
Peter Williams2dd73a42006-06-27 02:54:34 -07003310
Suresh Siddha9439aab2007-07-19 21:28:35 +02003311 if (*sd_idle && rq->nr_running)
Nick Piggin5969fe02005-09-10 00:26:19 -07003312 *sd_idle = 0;
3313
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 /* Bias balancing toward cpus of our domain */
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003315 if (local_group) {
3316 if (idle_cpu(i) && !first_idle_cpu) {
3317 first_idle_cpu = 1;
3318 balance_cpu = i;
3319 }
3320
Nick Piggina2000572006-02-10 01:51:02 -08003321 load = target_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02003322 } else {
Nick Piggina2000572006-02-10 01:51:02 -08003323 load = source_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02003324 if (load > max_cpu_load)
3325 max_cpu_load = load;
3326 if (min_cpu_load > load)
3327 min_cpu_load = load;
3328 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329
3330 avg_load += load;
Peter Williams2dd73a42006-06-27 02:54:34 -07003331 sum_nr_running += rq->nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02003332 sum_weighted_load += weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333 }
3334
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003335 /*
3336 * First idle cpu or the first cpu(busiest) in this sched group
3337 * is eligible for doing load balancing at this and above
Suresh Siddha9439aab2007-07-19 21:28:35 +02003338 * domains. In the newly idle case, we will allow all the cpu's
3339 * to do the newly idle load balance.
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003340 */
Suresh Siddha9439aab2007-07-19 21:28:35 +02003341 if (idle != CPU_NEWLY_IDLE && local_group &&
3342 balance_cpu != this_cpu && balance) {
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003343 *balance = 0;
3344 goto ret;
3345 }
3346
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347 total_load += avg_load;
Eric Dumazet5517d862007-05-08 00:32:57 -07003348 total_pwr += group->__cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349
3350 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07003351 avg_load = sg_div_cpu_power(group,
3352 avg_load * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353
Ken Chen908a7c12007-10-17 16:55:11 +02003354 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3355 __group_imb = 1;
3356
Eric Dumazet5517d862007-05-08 00:32:57 -07003357 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003358
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359 if (local_group) {
3360 this_load = avg_load;
3361 this = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07003362 this_nr_running = sum_nr_running;
3363 this_load_per_task = sum_weighted_load;
3364 } else if (avg_load > max_load &&
Ken Chen908a7c12007-10-17 16:55:11 +02003365 (sum_nr_running > group_capacity || __group_imb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366 max_load = avg_load;
3367 busiest = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07003368 busiest_nr_running = sum_nr_running;
3369 busiest_load_per_task = sum_weighted_load;
Ken Chen908a7c12007-10-17 16:55:11 +02003370 group_imb = __group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003372
3373#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3374 /*
3375 * Busy processors will not participate in power savings
3376 * balance.
3377 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003378 if (idle == CPU_NOT_IDLE ||
3379 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3380 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003381
3382 /*
3383 * If the local group is idle or completely loaded
3384 * no need to do power savings balance at this domain
3385 */
3386 if (local_group && (this_nr_running >= group_capacity ||
3387 !this_nr_running))
3388 power_savings_balance = 0;
3389
Ingo Molnardd41f592007-07-09 18:51:59 +02003390 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003391 * If a group is already running at full capacity or idle,
3392 * don't include that group in power savings calculations
Ingo Molnardd41f592007-07-09 18:51:59 +02003393 */
3394 if (!power_savings_balance || sum_nr_running >= group_capacity
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003395 || !sum_nr_running)
Ingo Molnardd41f592007-07-09 18:51:59 +02003396 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003397
Ingo Molnardd41f592007-07-09 18:51:59 +02003398 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003399 * Calculate the group which has the least non-idle load.
Ingo Molnardd41f592007-07-09 18:51:59 +02003400 * This is the group from where we need to pick up the load
3401 * for saving power
3402 */
3403 if ((sum_nr_running < min_nr_running) ||
3404 (sum_nr_running == min_nr_running &&
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003405 first_cpu(group->cpumask) <
3406 first_cpu(group_min->cpumask))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003407 group_min = group;
3408 min_nr_running = sum_nr_running;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003409 min_load_per_task = sum_weighted_load /
3410 sum_nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02003411 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003412
Ingo Molnardd41f592007-07-09 18:51:59 +02003413 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003414 * Calculate the group which is almost near its
Ingo Molnardd41f592007-07-09 18:51:59 +02003415 * capacity but still has some space to pick up some load
3416 * from other group and save more power
3417 */
3418 if (sum_nr_running <= group_capacity - 1) {
3419 if (sum_nr_running > leader_nr_running ||
3420 (sum_nr_running == leader_nr_running &&
3421 first_cpu(group->cpumask) >
3422 first_cpu(group_leader->cpumask))) {
3423 group_leader = group;
3424 leader_nr_running = sum_nr_running;
3425 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07003426 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003427group_next:
3428#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429 group = group->next;
3430 } while (group != sd->groups);
3431
Peter Williams2dd73a42006-06-27 02:54:34 -07003432 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433 goto out_balanced;
3434
3435 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3436
3437 if (this_load >= avg_load ||
3438 100*max_load <= sd->imbalance_pct*this_load)
3439 goto out_balanced;
3440
Peter Williams2dd73a42006-06-27 02:54:34 -07003441 busiest_load_per_task /= busiest_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02003442 if (group_imb)
3443 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3444
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445 /*
3446 * We're trying to get all the cpus to the average_load, so we don't
3447 * want to push ourselves above the average load, nor do we wish to
3448 * reduce the max loaded cpu below the average load, as either of these
3449 * actions would just result in more rebalancing later, and ping-pong
3450 * tasks around. Thus we look for the minimum possible imbalance.
3451 * Negative imbalances (*we* are more loaded than anyone else) will
3452 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003453 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454 * appear as very large values with unsigned longs.
3455 */
Peter Williams2dd73a42006-06-27 02:54:34 -07003456 if (max_load <= busiest_load_per_task)
3457 goto out_balanced;
3458
3459 /*
3460 * In the presence of smp nice balancing, certain scenarios can have
3461 * max load less than avg load(as we skip the groups at or below
3462 * its cpu_power, while calculating max_load..)
3463 */
3464 if (max_load < avg_load) {
3465 *imbalance = 0;
3466 goto small_imbalance;
3467 }
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003468
3469 /* Don't want to pull so many tasks that a group would go idle */
Peter Williams2dd73a42006-06-27 02:54:34 -07003470 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003471
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472 /* How much load to actually move to equalise the imbalance */
Eric Dumazet5517d862007-05-08 00:32:57 -07003473 *imbalance = min(max_pull * busiest->__cpu_power,
3474 (avg_load - this_load) * this->__cpu_power)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475 / SCHED_LOAD_SCALE;
3476
Peter Williams2dd73a42006-06-27 02:54:34 -07003477 /*
3478 * if *imbalance is less than the average load per runnable task
3479 * there is no gaurantee that any tasks will be moved so we'll have
3480 * a think about bumping its value to force at least one task to be
3481 * moved
3482 */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02003483 if (*imbalance < busiest_load_per_task) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07003484 unsigned long tmp, pwr_now, pwr_move;
Peter Williams2dd73a42006-06-27 02:54:34 -07003485 unsigned int imbn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486
Peter Williams2dd73a42006-06-27 02:54:34 -07003487small_imbalance:
3488 pwr_move = pwr_now = 0;
3489 imbn = 2;
3490 if (this_nr_running) {
3491 this_load_per_task /= this_nr_running;
3492 if (busiest_load_per_task > this_load_per_task)
3493 imbn = 1;
3494 } else
3495 this_load_per_task = SCHED_LOAD_SCALE;
3496
Ingo Molnardd41f592007-07-09 18:51:59 +02003497 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3498 busiest_load_per_task * imbn) {
Peter Williams2dd73a42006-06-27 02:54:34 -07003499 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 return busiest;
3501 }
3502
3503 /*
3504 * OK, we don't have enough imbalance to justify moving tasks,
3505 * however we may be able to increase total CPU power used by
3506 * moving them.
3507 */
3508
Eric Dumazet5517d862007-05-08 00:32:57 -07003509 pwr_now += busiest->__cpu_power *
3510 min(busiest_load_per_task, max_load);
3511 pwr_now += this->__cpu_power *
3512 min(this_load_per_task, this_load);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513 pwr_now /= SCHED_LOAD_SCALE;
3514
3515 /* Amount of load we'd subtract */
Eric Dumazet5517d862007-05-08 00:32:57 -07003516 tmp = sg_div_cpu_power(busiest,
3517 busiest_load_per_task * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518 if (max_load > tmp)
Eric Dumazet5517d862007-05-08 00:32:57 -07003519 pwr_move += busiest->__cpu_power *
Peter Williams2dd73a42006-06-27 02:54:34 -07003520 min(busiest_load_per_task, max_load - tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521
3522 /* Amount of load we'd add */
Eric Dumazet5517d862007-05-08 00:32:57 -07003523 if (max_load * busiest->__cpu_power <
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08003524 busiest_load_per_task * SCHED_LOAD_SCALE)
Eric Dumazet5517d862007-05-08 00:32:57 -07003525 tmp = sg_div_cpu_power(this,
3526 max_load * busiest->__cpu_power);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 else
Eric Dumazet5517d862007-05-08 00:32:57 -07003528 tmp = sg_div_cpu_power(this,
3529 busiest_load_per_task * SCHED_LOAD_SCALE);
3530 pwr_move += this->__cpu_power *
3531 min(this_load_per_task, this_load + tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532 pwr_move /= SCHED_LOAD_SCALE;
3533
3534 /* Move if we gain throughput */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02003535 if (pwr_move > pwr_now)
3536 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537 }
3538
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 return busiest;
3540
3541out_balanced:
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003542#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003543 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003544 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003546 if (this == group_leader && group_leader != group_min) {
3547 *imbalance = min_load_per_task;
3548 return group_min;
3549 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003550#endif
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003551ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552 *imbalance = 0;
3553 return NULL;
3554}
3555
3556/*
3557 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3558 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003559static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003560find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003561 unsigned long imbalance, const cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003563 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07003564 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565 int i;
3566
3567 for_each_cpu_mask(i, group->cpumask) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003568 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003569
3570 if (!cpu_isset(i, *cpus))
3571 continue;
3572
Ingo Molnar48f24c42006-07-03 00:25:40 -07003573 rq = cpu_rq(i);
Ingo Molnardd41f592007-07-09 18:51:59 +02003574 wl = weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575
Ingo Molnardd41f592007-07-09 18:51:59 +02003576 if (rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07003577 continue;
3578
Ingo Molnardd41f592007-07-09 18:51:59 +02003579 if (wl > max_load) {
3580 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003581 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582 }
3583 }
3584
3585 return busiest;
3586}
3587
3588/*
Nick Piggin77391d72005-06-25 14:57:30 -07003589 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3590 * so long as it is large enough.
3591 */
3592#define MAX_PINNED_INTERVAL 512
3593
3594/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3596 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003598static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003599 struct sched_domain *sd, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003600 int *balance, cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601{
Peter Williams43010652007-08-09 11:16:46 +02003602 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003605 struct rq *busiest;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003606 unsigned long flags;
Peter Zijlstra18d95a22008-04-19 19:45:00 +02003607 int unlock_aggregate;
Nick Piggin5969fe02005-09-10 00:26:19 -07003608
Mike Travis7c16ec52008-04-04 18:11:11 -07003609 cpus_setall(*cpus);
3610
Peter Zijlstra18d95a22008-04-19 19:45:00 +02003611 unlock_aggregate = get_aggregate(sd);
3612
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003613 /*
3614 * When power savings policy is enabled for the parent domain, idle
3615 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02003616 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003617 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003618 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003619 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003620 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003621 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622
Ingo Molnar2d723762007-10-15 17:00:12 +02003623 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003625redo:
3626 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003627 cpus, balance);
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003628
Chen, Kenneth W06066712006-12-10 02:20:35 -08003629 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003630 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003631
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632 if (!group) {
3633 schedstat_inc(sd, lb_nobusyg[idle]);
3634 goto out_balanced;
3635 }
3636
Mike Travis7c16ec52008-04-04 18:11:11 -07003637 busiest = find_busiest_queue(group, idle, imbalance, cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638 if (!busiest) {
3639 schedstat_inc(sd, lb_nobusyq[idle]);
3640 goto out_balanced;
3641 }
3642
Nick Piggindb935db2005-06-25 14:57:11 -07003643 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644
3645 schedstat_add(sd, lb_imbalance[idle], imbalance);
3646
Peter Williams43010652007-08-09 11:16:46 +02003647 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648 if (busiest->nr_running > 1) {
3649 /*
3650 * Attempt to move tasks. If find_busiest_group has found
3651 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02003652 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653 * correctly treated as an imbalance.
3654 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003655 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07003656 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02003657 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07003658 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07003659 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003660 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07003661
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003662 /*
3663 * some other cpu did the load balance for us.
3664 */
Peter Williams43010652007-08-09 11:16:46 +02003665 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003666 resched_cpu(this_cpu);
3667
Nick Piggin81026792005-06-25 14:57:07 -07003668 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003669 if (unlikely(all_pinned)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003670 cpu_clear(cpu_of(busiest), *cpus);
3671 if (!cpus_empty(*cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003672 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07003673 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003674 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675 }
Nick Piggin81026792005-06-25 14:57:07 -07003676
Peter Williams43010652007-08-09 11:16:46 +02003677 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 schedstat_inc(sd, lb_failed[idle]);
3679 sd->nr_balance_failed++;
3680
3681 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003683 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003684
3685 /* don't kick the migration_thread, if the curr
3686 * task on busiest cpu can't be moved to this_cpu
3687 */
3688 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003689 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003690 all_pinned = 1;
3691 goto out_one_pinned;
3692 }
3693
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694 if (!busiest->active_balance) {
3695 busiest->active_balance = 1;
3696 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07003697 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003699 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07003700 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701 wake_up_process(busiest->migration_thread);
3702
3703 /*
3704 * We've kicked active balancing, reset the failure
3705 * counter.
3706 */
Nick Piggin39507452005-06-25 14:57:09 -07003707 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708 }
Nick Piggin81026792005-06-25 14:57:07 -07003709 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710 sd->nr_balance_failed = 0;
3711
Nick Piggin81026792005-06-25 14:57:07 -07003712 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713 /* We were unbalanced, so reset the balancing interval */
3714 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07003715 } else {
3716 /*
3717 * If we've begun active balancing, start to back off. This
3718 * case may not be covered by the all_pinned logic if there
3719 * is only 1 task on the busy runqueue (because we don't call
3720 * move_tasks).
3721 */
3722 if (sd->balance_interval < sd->max_interval)
3723 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724 }
3725
Peter Williams43010652007-08-09 11:16:46 +02003726 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003727 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstra18d95a22008-04-19 19:45:00 +02003728 ld_moved = -1;
3729
3730 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731
3732out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003733 schedstat_inc(sd, lb_balanced[idle]);
3734
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003735 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003736
3737out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07003739 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3740 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003741 sd->balance_interval *= 2;
3742
Ingo Molnar48f24c42006-07-03 00:25:40 -07003743 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003744 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstra18d95a22008-04-19 19:45:00 +02003745 ld_moved = -1;
3746 else
3747 ld_moved = 0;
3748out:
3749 if (unlock_aggregate)
3750 put_aggregate(sd);
3751 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752}
3753
3754/*
3755 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3756 * tasks if there is an imbalance.
3757 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003758 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759 * this_rq is locked.
3760 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07003761static int
Mike Travis7c16ec52008-04-04 18:11:11 -07003762load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3763 cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764{
3765 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003766 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02003768 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07003769 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003770 int all_pinned = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07003771
3772 cpus_setall(*cpus);
Nick Piggin5969fe02005-09-10 00:26:19 -07003773
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003774 /*
3775 * When power savings policy is enabled for the parent domain, idle
3776 * sibling can pick up load irrespective of busy siblings. In this case,
3777 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003778 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003779 */
3780 if (sd->flags & SD_SHARE_CPUPOWER &&
3781 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003782 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783
Ingo Molnar2d723762007-10-15 17:00:12 +02003784 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003785redo:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003786 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Mike Travis7c16ec52008-04-04 18:11:11 -07003787 &sd_idle, cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003789 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003790 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791 }
3792
Mike Travis7c16ec52008-04-04 18:11:11 -07003793 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07003794 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003795 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003796 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797 }
3798
Nick Piggindb935db2005-06-25 14:57:11 -07003799 BUG_ON(busiest == this_rq);
3800
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003801 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003802
Peter Williams43010652007-08-09 11:16:46 +02003803 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003804 if (busiest->nr_running > 1) {
3805 /* Attempt to move tasks */
3806 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003807 /* this_rq->clock is already updated */
3808 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02003809 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003810 imbalance, sd, CPU_NEWLY_IDLE,
3811 &all_pinned);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003812 spin_unlock(&busiest->lock);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003813
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003814 if (unlikely(all_pinned)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003815 cpu_clear(cpu_of(busiest), *cpus);
3816 if (!cpus_empty(*cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003817 goto redo;
3818 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003819 }
3820
Peter Williams43010652007-08-09 11:16:46 +02003821 if (!ld_moved) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003822 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003823 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3824 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003825 return -1;
3826 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003827 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828
Peter Williams43010652007-08-09 11:16:46 +02003829 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003830
3831out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003832 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003833 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003834 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003835 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003836 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003837
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003838 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839}
3840
3841/*
3842 * idle_balance is called by schedule() if this_cpu is about to become
3843 * idle. Attempts to pull tasks from other CPUs.
3844 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003845static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003846{
3847 struct sched_domain *sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02003848 int pulled_task = -1;
3849 unsigned long next_balance = jiffies + HZ;
Mike Travis7c16ec52008-04-04 18:11:11 -07003850 cpumask_t tmpmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851
3852 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003853 unsigned long interval;
3854
3855 if (!(sd->flags & SD_LOAD_BALANCE))
3856 continue;
3857
3858 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07003859 /* If we've pulled tasks over stop searching: */
Mike Travis7c16ec52008-04-04 18:11:11 -07003860 pulled_task = load_balance_newidle(this_cpu, this_rq,
3861 sd, &tmpmask);
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003862
3863 interval = msecs_to_jiffies(sd->balance_interval);
3864 if (time_after(next_balance, sd->last_balance + interval))
3865 next_balance = sd->last_balance + interval;
3866 if (pulled_task)
3867 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003868 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003869 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003870 /*
3871 * We are going idle. next_balance may be set based on
3872 * a busy processor. So reset next_balance.
3873 */
3874 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02003875 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876}
3877
3878/*
3879 * active_load_balance is run by migration threads. It pushes running tasks
3880 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3881 * running on each physical CPU where possible, and avoids physical /
3882 * logical imbalances.
3883 *
3884 * Called with busiest_rq locked.
3885 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003886static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887{
Nick Piggin39507452005-06-25 14:57:09 -07003888 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003889 struct sched_domain *sd;
3890 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07003891
Ingo Molnar48f24c42006-07-03 00:25:40 -07003892 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07003893 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07003894 return;
3895
3896 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897
3898 /*
Nick Piggin39507452005-06-25 14:57:09 -07003899 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003900 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07003901 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902 */
Nick Piggin39507452005-06-25 14:57:09 -07003903 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904
Nick Piggin39507452005-06-25 14:57:09 -07003905 /* move a task from busiest_rq to target_rq */
3906 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003907 update_rq_clock(busiest_rq);
3908 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909
Nick Piggin39507452005-06-25 14:57:09 -07003910 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003911 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07003912 if ((sd->flags & SD_LOAD_BALANCE) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07003913 cpu_isset(busiest_cpu, sd->span))
Nick Piggin39507452005-06-25 14:57:09 -07003914 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003915 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916
Ingo Molnar48f24c42006-07-03 00:25:40 -07003917 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02003918 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919
Peter Williams43010652007-08-09 11:16:46 +02003920 if (move_one_task(target_rq, target_cpu, busiest_rq,
3921 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07003922 schedstat_inc(sd, alb_pushed);
3923 else
3924 schedstat_inc(sd, alb_failed);
3925 }
Nick Piggin39507452005-06-25 14:57:09 -07003926 spin_unlock(&target_rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003927}
3928
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003929#ifdef CONFIG_NO_HZ
3930static struct {
3931 atomic_t load_balancer;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003932 cpumask_t cpu_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003933} nohz ____cacheline_aligned = {
3934 .load_balancer = ATOMIC_INIT(-1),
3935 .cpu_mask = CPU_MASK_NONE,
3936};
3937
Christoph Lameter7835b982006-12-10 02:20:22 -08003938/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003939 * This routine will try to nominate the ilb (idle load balancing)
3940 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3941 * load balancing on behalf of all those cpus. If all the cpus in the system
3942 * go into this tickless mode, then there will be no ilb owner (as there is
3943 * no need for one) and all the cpus will sleep till the next wakeup event
3944 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08003945 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003946 * For the ilb owner, tick is not stopped. And this tick will be used
3947 * for idle load balancing. ilb owner will still be part of
3948 * nohz.cpu_mask..
3949 *
3950 * While stopping the tick, this cpu will become the ilb owner if there
3951 * is no other owner. And will be the owner till that cpu becomes busy
3952 * or if all cpus in the system stop their ticks at which point
3953 * there is no need for ilb owner.
3954 *
3955 * When the ilb owner becomes busy, it nominates another owner, during the
3956 * next busy scheduler_tick()
3957 */
3958int select_nohz_load_balancer(int stop_tick)
3959{
3960 int cpu = smp_processor_id();
3961
3962 if (stop_tick) {
3963 cpu_set(cpu, nohz.cpu_mask);
3964 cpu_rq(cpu)->in_nohz_recently = 1;
3965
3966 /*
3967 * If we are going offline and still the leader, give up!
3968 */
3969 if (cpu_is_offline(cpu) &&
3970 atomic_read(&nohz.load_balancer) == cpu) {
3971 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3972 BUG();
3973 return 0;
3974 }
3975
3976 /* time for ilb owner also to sleep */
3977 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3978 if (atomic_read(&nohz.load_balancer) == cpu)
3979 atomic_set(&nohz.load_balancer, -1);
3980 return 0;
3981 }
3982
3983 if (atomic_read(&nohz.load_balancer) == -1) {
3984 /* make me the ilb owner */
3985 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3986 return 1;
3987 } else if (atomic_read(&nohz.load_balancer) == cpu)
3988 return 1;
3989 } else {
3990 if (!cpu_isset(cpu, nohz.cpu_mask))
3991 return 0;
3992
3993 cpu_clear(cpu, nohz.cpu_mask);
3994
3995 if (atomic_read(&nohz.load_balancer) == cpu)
3996 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3997 BUG();
3998 }
3999 return 0;
4000}
4001#endif
4002
4003static DEFINE_SPINLOCK(balancing);
4004
4005/*
Christoph Lameter7835b982006-12-10 02:20:22 -08004006 * It checks each scheduling domain to see if it is due to be balanced,
4007 * and initiates a balancing operation if so.
4008 *
4009 * Balancing parameters are set up in arch_init_sched_domains.
4010 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02004011static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08004012{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004013 int balance = 1;
4014 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08004015 unsigned long interval;
4016 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004017 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08004018 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004019 int update_next_balance = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07004020 cpumask_t tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004022 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023 if (!(sd->flags & SD_LOAD_BALANCE))
4024 continue;
4025
4026 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004027 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028 interval *= sd->busy_factor;
4029
4030 /* scale ms to jiffies */
4031 interval = msecs_to_jiffies(interval);
4032 if (unlikely(!interval))
4033 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02004034 if (interval > HZ*NR_CPUS/10)
4035 interval = HZ*NR_CPUS/10;
4036
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037
Christoph Lameter08c183f2006-12-10 02:20:29 -08004038 if (sd->flags & SD_SERIALIZE) {
4039 if (!spin_trylock(&balancing))
4040 goto out;
4041 }
4042
Christoph Lameterc9819f42006-12-10 02:20:25 -08004043 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07004044 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004045 /*
4046 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07004047 * longer idle, or one of our SMT siblings is
4048 * not idle.
4049 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004050 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08004052 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053 }
Christoph Lameter08c183f2006-12-10 02:20:29 -08004054 if (sd->flags & SD_SERIALIZE)
4055 spin_unlock(&balancing);
4056out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02004057 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08004058 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004059 update_next_balance = 1;
4060 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004061
4062 /*
4063 * Stop the load balance at this level. There is another
4064 * CPU in our sched group which is doing load balancing more
4065 * actively.
4066 */
4067 if (!balance)
4068 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004069 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02004070
4071 /*
4072 * next_balance will be updated only when there is a need.
4073 * When the cpu is attached to null domain for ex, it will not be
4074 * updated.
4075 */
4076 if (likely(update_next_balance))
4077 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004078}
4079
4080/*
4081 * run_rebalance_domains is triggered when needed from the scheduler tick.
4082 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4083 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4084 */
4085static void run_rebalance_domains(struct softirq_action *h)
4086{
Ingo Molnardd41f592007-07-09 18:51:59 +02004087 int this_cpu = smp_processor_id();
4088 struct rq *this_rq = cpu_rq(this_cpu);
4089 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4090 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004091
Ingo Molnardd41f592007-07-09 18:51:59 +02004092 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004093
4094#ifdef CONFIG_NO_HZ
4095 /*
4096 * If this cpu is the owner for idle load balancing, then do the
4097 * balancing on behalf of the other idle cpus whose ticks are
4098 * stopped.
4099 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004100 if (this_rq->idle_at_tick &&
4101 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004102 cpumask_t cpus = nohz.cpu_mask;
4103 struct rq *rq;
4104 int balance_cpu;
4105
Ingo Molnardd41f592007-07-09 18:51:59 +02004106 cpu_clear(this_cpu, cpus);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004107 for_each_cpu_mask(balance_cpu, cpus) {
4108 /*
4109 * If this cpu gets work to do, stop the load balancing
4110 * work being done for other cpus. Next load
4111 * balancing owner will pick it up.
4112 */
4113 if (need_resched())
4114 break;
4115
Oleg Nesterovde0cf892007-08-12 18:08:19 +02004116 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004117
4118 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004119 if (time_after(this_rq->next_balance, rq->next_balance))
4120 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004121 }
4122 }
4123#endif
4124}
4125
4126/*
4127 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4128 *
4129 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4130 * idle load balancing owner or decide to stop the periodic load balancing,
4131 * if the whole system is idle.
4132 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004133static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004134{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004135#ifdef CONFIG_NO_HZ
4136 /*
4137 * If we were in the nohz mode recently and busy at the current
4138 * scheduler tick, then check if we need to nominate new idle
4139 * load balancer.
4140 */
4141 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4142 rq->in_nohz_recently = 0;
4143
4144 if (atomic_read(&nohz.load_balancer) == cpu) {
4145 cpu_clear(cpu, nohz.cpu_mask);
4146 atomic_set(&nohz.load_balancer, -1);
4147 }
4148
4149 if (atomic_read(&nohz.load_balancer) == -1) {
4150 /*
4151 * simple selection for now: Nominate the
4152 * first cpu in the nohz list to be the next
4153 * ilb owner.
4154 *
4155 * TBD: Traverse the sched domains and nominate
4156 * the nearest cpu in the nohz.cpu_mask.
4157 */
4158 int ilb = first_cpu(nohz.cpu_mask);
4159
Mike Travis434d53b2008-04-04 18:11:04 -07004160 if (ilb < nr_cpu_ids)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004161 resched_cpu(ilb);
4162 }
4163 }
4164
4165 /*
4166 * If this cpu is idle and doing idle load balancing for all the
4167 * cpus with ticks stopped, is it time for that to stop?
4168 */
4169 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4170 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4171 resched_cpu(cpu);
4172 return;
4173 }
4174
4175 /*
4176 * If this cpu is idle and the idle load balancing is done by
4177 * someone else, then no need raise the SCHED_SOFTIRQ
4178 */
4179 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4180 cpu_isset(cpu, nohz.cpu_mask))
4181 return;
4182#endif
4183 if (time_after_eq(jiffies, rq->next_balance))
4184 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185}
Ingo Molnardd41f592007-07-09 18:51:59 +02004186
4187#else /* CONFIG_SMP */
4188
Linus Torvalds1da177e2005-04-16 15:20:36 -07004189/*
4190 * on UP we do not need to balance between CPUs:
4191 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004192static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193{
4194}
Ingo Molnardd41f592007-07-09 18:51:59 +02004195
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196#endif
4197
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198DEFINE_PER_CPU(struct kernel_stat, kstat);
4199
4200EXPORT_PER_CPU_SYMBOL(kstat);
4201
4202/*
Ingo Molnar41b86e92007-07-09 18:51:58 +02004203 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4204 * that have not yet been banked in case the task is currently running.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 */
Ingo Molnar41b86e92007-07-09 18:51:58 +02004206unsigned long long task_sched_runtime(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02004209 u64 ns, delta_exec;
4210 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004211
Ingo Molnar41b86e92007-07-09 18:51:58 +02004212 rq = task_rq_lock(p, &flags);
4213 ns = p->se.sum_exec_runtime;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004214 if (task_current(rq, p)) {
Ingo Molnara8e504d2007-08-09 11:16:47 +02004215 update_rq_clock(rq);
4216 delta_exec = rq->clock - p->se.exec_start;
Ingo Molnar41b86e92007-07-09 18:51:58 +02004217 if ((s64)delta_exec > 0)
4218 ns += delta_exec;
4219 }
4220 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004221
Linus Torvalds1da177e2005-04-16 15:20:36 -07004222 return ns;
4223}
4224
4225/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226 * Account user cpu time to a process.
4227 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07004228 * @cputime: the cpu time spent in user space since the last update
4229 */
4230void account_user_time(struct task_struct *p, cputime_t cputime)
4231{
4232 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4233 cputime64_t tmp;
4234
4235 p->utime = cputime_add(p->utime, cputime);
4236
4237 /* Add user time to cpustat. */
4238 tmp = cputime_to_cputime64(cputime);
4239 if (TASK_NICE(p) > 0)
4240 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4241 else
4242 cpustat->user = cputime64_add(cpustat->user, tmp);
4243}
4244
4245/*
Laurent Vivier94886b82007-10-15 17:00:19 +02004246 * Account guest cpu time to a process.
4247 * @p: the process that the cpu time gets accounted to
4248 * @cputime: the cpu time spent in virtual machine since the last update
4249 */
Adrian Bunkf7402e02007-10-29 21:18:10 +01004250static void account_guest_time(struct task_struct *p, cputime_t cputime)
Laurent Vivier94886b82007-10-15 17:00:19 +02004251{
4252 cputime64_t tmp;
4253 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4254
4255 tmp = cputime_to_cputime64(cputime);
4256
4257 p->utime = cputime_add(p->utime, cputime);
4258 p->gtime = cputime_add(p->gtime, cputime);
4259
4260 cpustat->user = cputime64_add(cpustat->user, tmp);
4261 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4262}
4263
4264/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07004265 * Account scaled user cpu time to a process.
4266 * @p: the process that the cpu time gets accounted to
4267 * @cputime: the cpu time spent in user space since the last update
4268 */
4269void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4270{
4271 p->utimescaled = cputime_add(p->utimescaled, cputime);
4272}
4273
4274/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004275 * Account system cpu time to a process.
4276 * @p: the process that the cpu time gets accounted to
4277 * @hardirq_offset: the offset to subtract from hardirq_count()
4278 * @cputime: the cpu time spent in kernel space since the last update
4279 */
4280void account_system_time(struct task_struct *p, int hardirq_offset,
4281 cputime_t cputime)
4282{
4283 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004284 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004285 cputime64_t tmp;
4286
Harvey Harrison983ed7a2008-04-24 18:17:55 -07004287 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4288 account_guest_time(p, cputime);
4289 return;
4290 }
Laurent Vivier94886b82007-10-15 17:00:19 +02004291
Linus Torvalds1da177e2005-04-16 15:20:36 -07004292 p->stime = cputime_add(p->stime, cputime);
4293
4294 /* Add system time to cpustat. */
4295 tmp = cputime_to_cputime64(cputime);
4296 if (hardirq_count() - hardirq_offset)
4297 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4298 else if (softirq_count())
4299 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08004300 else if (p != rq->idle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004301 cpustat->system = cputime64_add(cpustat->system, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08004302 else if (atomic_read(&rq->nr_iowait) > 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004303 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4304 else
4305 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4306 /* Account for system time used */
4307 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308}
4309
4310/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07004311 * Account scaled system cpu time to a process.
4312 * @p: the process that the cpu time gets accounted to
4313 * @hardirq_offset: the offset to subtract from hardirq_count()
4314 * @cputime: the cpu time spent in kernel space since the last update
4315 */
4316void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4317{
4318 p->stimescaled = cputime_add(p->stimescaled, cputime);
4319}
4320
4321/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004322 * Account for involuntary wait time.
4323 * @p: the process from which the cpu time has been stolen
4324 * @steal: the cpu time spent in involuntary wait
4325 */
4326void account_steal_time(struct task_struct *p, cputime_t steal)
4327{
4328 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4329 cputime64_t tmp = cputime_to_cputime64(steal);
Ingo Molnar70b97a72006-07-03 00:25:42 -07004330 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004331
4332 if (p == rq->idle) {
4333 p->stime = cputime_add(p->stime, steal);
4334 if (atomic_read(&rq->nr_iowait) > 0)
4335 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4336 else
4337 cpustat->idle = cputime64_add(cpustat->idle, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08004338 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07004339 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4340}
4341
Christoph Lameter7835b982006-12-10 02:20:22 -08004342/*
4343 * This function gets called by the timer code, with HZ frequency.
4344 * We call it with interrupts disabled.
4345 *
4346 * It also gets called by the fork code, when changing the parent's
4347 * timeslices.
4348 */
4349void scheduler_tick(void)
4350{
Christoph Lameter7835b982006-12-10 02:20:22 -08004351 int cpu = smp_processor_id();
4352 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004353 struct task_struct *curr = rq->curr;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004354
4355 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08004356
Ingo Molnardd41f592007-07-09 18:51:59 +02004357 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004358 update_rq_clock(rq);
Ingo Molnarf1a438d2007-08-09 11:16:45 +02004359 update_cpu_load(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01004360 curr->sched_class->task_tick(rq, curr, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02004361 spin_unlock(&rq->lock);
4362
Christoph Lametere418e1c2006-12-10 02:20:23 -08004363#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02004364 rq->idle_at_tick = idle_cpu(cpu);
4365 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08004366#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004367}
4368
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004369#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4370 defined(CONFIG_PREEMPT_TRACER))
4371
4372static inline unsigned long get_parent_ip(unsigned long addr)
4373{
4374 if (in_lock_functions(addr)) {
4375 addr = CALLER_ADDR2;
4376 if (in_lock_functions(addr))
4377 addr = CALLER_ADDR3;
4378 }
4379 return addr;
4380}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381
Srinivasa Ds43627582008-02-23 15:24:04 -08004382void __kprobes add_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004383{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004384#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07004385 /*
4386 * Underflow?
4387 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004388 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4389 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004390#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004391 preempt_count() += val;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004392#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07004393 /*
4394 * Spinlock count overflowing soon?
4395 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08004396 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4397 PREEMPT_MASK - 10);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004398#endif
4399 if (preempt_count() == val)
4400 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004401}
4402EXPORT_SYMBOL(add_preempt_count);
4403
Srinivasa Ds43627582008-02-23 15:24:04 -08004404void __kprobes sub_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004405{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004406#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07004407 /*
4408 * Underflow?
4409 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004410 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4411 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004412 /*
4413 * Is the spinlock portion underflowing?
4414 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004415 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4416 !(preempt_count() & PREEMPT_MASK)))
4417 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004418#endif
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004419
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004420 if (preempt_count() == val)
4421 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004422 preempt_count() -= val;
4423}
4424EXPORT_SYMBOL(sub_preempt_count);
4425
4426#endif
4427
4428/*
Ingo Molnardd41f592007-07-09 18:51:59 +02004429 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004430 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004431static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004432{
Satyam Sharma838225b2007-10-24 18:23:50 +02004433 struct pt_regs *regs = get_irq_regs();
4434
4435 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4436 prev->comm, prev->pid, preempt_count());
4437
Ingo Molnardd41f592007-07-09 18:51:59 +02004438 debug_show_held_locks(prev);
4439 if (irqs_disabled())
4440 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02004441
4442 if (regs)
4443 show_regs(regs);
4444 else
4445 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02004446}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004447
Ingo Molnardd41f592007-07-09 18:51:59 +02004448/*
4449 * Various schedule()-time debugging checks and statistics:
4450 */
4451static inline void schedule_debug(struct task_struct *prev)
4452{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004453 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004454 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07004455 * schedule() atomically, we ignore that path for now.
4456 * Otherwise, whine if we are scheduling when we should not be.
4457 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004458 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4459 __schedule_bug(prev);
4460
Linus Torvalds1da177e2005-04-16 15:20:36 -07004461 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4462
Ingo Molnar2d723762007-10-15 17:00:12 +02004463 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02004464#ifdef CONFIG_SCHEDSTATS
4465 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02004466 schedstat_inc(this_rq(), bkl_count);
4467 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02004468 }
4469#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02004470}
4471
4472/*
4473 * Pick up the highest-prio task:
4474 */
4475static inline struct task_struct *
Ingo Molnarff95f3d2007-08-09 11:16:49 +02004476pick_next_task(struct rq *rq, struct task_struct *prev)
Ingo Molnardd41f592007-07-09 18:51:59 +02004477{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02004478 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02004479 struct task_struct *p;
4480
4481 /*
4482 * Optimization: we know that if all tasks are in
4483 * the fair class we can call that function directly:
4484 */
4485 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02004486 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004487 if (likely(p))
4488 return p;
4489 }
4490
4491 class = sched_class_highest;
4492 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02004493 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004494 if (p)
4495 return p;
4496 /*
4497 * Will never be NULL as the idle class always
4498 * returns a non-NULL p:
4499 */
4500 class = class->next;
4501 }
4502}
4503
4504/*
4505 * schedule() is the main scheduler function.
4506 */
4507asmlinkage void __sched schedule(void)
4508{
4509 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08004510 unsigned long *switch_count;
Ingo Molnardd41f592007-07-09 18:51:59 +02004511 struct rq *rq;
Ingo Molnardd41f592007-07-09 18:51:59 +02004512 int cpu;
4513
Linus Torvalds1da177e2005-04-16 15:20:36 -07004514need_resched:
4515 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02004516 cpu = smp_processor_id();
4517 rq = cpu_rq(cpu);
4518 rcu_qsctr_inc(cpu);
4519 prev = rq->curr;
4520 switch_count = &prev->nivcsw;
4521
Linus Torvalds1da177e2005-04-16 15:20:36 -07004522 release_kernel_lock(prev);
4523need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004524
Ingo Molnardd41f592007-07-09 18:51:59 +02004525 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004526
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004527 hrtick_clear(rq);
4528
Ingo Molnar1e819952007-10-15 17:00:13 +02004529 /*
4530 * Do the rq-clock update outside the rq lock:
4531 */
4532 local_irq_disable();
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004533 update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02004534 spin_lock(&rq->lock);
4535 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004536
Ingo Molnardd41f592007-07-09 18:51:59 +02004537 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4538 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
Roel Kluin23e3c3c2008-03-13 17:41:59 +01004539 signal_pending(prev))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02004540 prev->state = TASK_RUNNING;
4541 } else {
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004542 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02004543 }
4544 switch_count = &prev->nvcsw;
4545 }
4546
Steven Rostedt9a897c52008-01-25 21:08:22 +01004547#ifdef CONFIG_SMP
4548 if (prev->sched_class->pre_schedule)
4549 prev->sched_class->pre_schedule(rq, prev);
4550#endif
Steven Rostedtf65eda42008-01-25 21:08:07 +01004551
Ingo Molnardd41f592007-07-09 18:51:59 +02004552 if (unlikely(!rq->nr_running))
4553 idle_balance(cpu, rq);
4554
Ingo Molnar31ee5292007-08-09 11:16:49 +02004555 prev->sched_class->put_prev_task(rq, prev);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02004556 next = pick_next_task(rq, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004557
Linus Torvalds1da177e2005-04-16 15:20:36 -07004558 if (likely(prev != next)) {
David Simner673a90a2008-04-29 10:08:59 +01004559 sched_info_switch(prev, next);
4560
Linus Torvalds1da177e2005-04-16 15:20:36 -07004561 rq->nr_switches++;
4562 rq->curr = next;
4563 ++*switch_count;
4564
Ingo Molnardd41f592007-07-09 18:51:59 +02004565 context_switch(rq, prev, next); /* unlocks the rq */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004566 /*
4567 * the context switch might have flipped the stack from under
4568 * us, hence refresh the local variables.
4569 */
4570 cpu = smp_processor_id();
4571 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004572 } else
4573 spin_unlock_irq(&rq->lock);
4574
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004575 hrtick_set(rq);
4576
4577 if (unlikely(reacquire_kernel_lock(current) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004578 goto need_resched_nonpreemptible;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004579
Linus Torvalds1da177e2005-04-16 15:20:36 -07004580 preempt_enable_no_resched();
4581 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4582 goto need_resched;
4583}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004584EXPORT_SYMBOL(schedule);
4585
4586#ifdef CONFIG_PREEMPT
4587/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004588 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004589 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07004590 * occur there and call schedule directly.
4591 */
4592asmlinkage void __sched preempt_schedule(void)
4593{
4594 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01004595
Linus Torvalds1da177e2005-04-16 15:20:36 -07004596 /*
4597 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004598 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07004599 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07004600 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004601 return;
4602
Andi Kleen3a5c3592007-10-15 17:00:14 +02004603 do {
4604 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02004605 schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02004606 sub_preempt_count(PREEMPT_ACTIVE);
4607
4608 /*
4609 * Check again in case we missed a preemption opportunity
4610 * between schedule and now.
4611 */
4612 barrier();
4613 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004614}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004615EXPORT_SYMBOL(preempt_schedule);
4616
4617/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004618 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07004619 * off of irq context.
4620 * Note, that this is called and return with irqs disabled. This will
4621 * protect us against recursive calling from irq.
4622 */
4623asmlinkage void __sched preempt_schedule_irq(void)
4624{
4625 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01004626
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004627 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004628 BUG_ON(ti->preempt_count || !irqs_disabled());
4629
Andi Kleen3a5c3592007-10-15 17:00:14 +02004630 do {
4631 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02004632 local_irq_enable();
4633 schedule();
4634 local_irq_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02004635 sub_preempt_count(PREEMPT_ACTIVE);
4636
4637 /*
4638 * Check again in case we missed a preemption opportunity
4639 * between schedule and now.
4640 */
4641 barrier();
4642 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004643}
4644
4645#endif /* CONFIG_PREEMPT */
4646
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004647int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4648 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004649{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004650 return try_to_wake_up(curr->private, mode, sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004651}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004652EXPORT_SYMBOL(default_wake_function);
4653
4654/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004655 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4656 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07004657 * number) then we wake all the non-exclusive tasks and one exclusive task.
4658 *
4659 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004660 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07004661 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4662 */
4663static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4664 int nr_exclusive, int sync, void *key)
4665{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004666 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004667
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004668 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07004669 unsigned flags = curr->flags;
4670
Linus Torvalds1da177e2005-04-16 15:20:36 -07004671 if (curr->func(curr, mode, sync, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07004672 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004673 break;
4674 }
4675}
4676
4677/**
4678 * __wake_up - wake up threads blocked on a waitqueue.
4679 * @q: the waitqueue
4680 * @mode: which threads
4681 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07004682 * @key: is directly passed to the wakeup function
Linus Torvalds1da177e2005-04-16 15:20:36 -07004683 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004684void __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004685 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004686{
4687 unsigned long flags;
4688
4689 spin_lock_irqsave(&q->lock, flags);
4690 __wake_up_common(q, mode, nr_exclusive, 0, key);
4691 spin_unlock_irqrestore(&q->lock, flags);
4692}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004693EXPORT_SYMBOL(__wake_up);
4694
4695/*
4696 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4697 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004698void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004699{
4700 __wake_up_common(q, mode, 1, 0, NULL);
4701}
4702
4703/**
Martin Waitz67be2dd2005-05-01 08:59:26 -07004704 * __wake_up_sync - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004705 * @q: the waitqueue
4706 * @mode: which threads
4707 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4708 *
4709 * The sync wakeup differs that the waker knows that it will schedule
4710 * away soon, so while the target thread will be woken up, it will not
4711 * be migrated to another CPU - ie. the two threads are 'synchronized'
4712 * with each other. This can prevent needless bouncing between CPUs.
4713 *
4714 * On UP it can prevent extra preemption.
4715 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004716void
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004717__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004718{
4719 unsigned long flags;
4720 int sync = 1;
4721
4722 if (unlikely(!q))
4723 return;
4724
4725 if (unlikely(!nr_exclusive))
4726 sync = 0;
4727
4728 spin_lock_irqsave(&q->lock, flags);
4729 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4730 spin_unlock_irqrestore(&q->lock, flags);
4731}
4732EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4733
Ingo Molnarb15136e2007-10-24 18:23:48 +02004734void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004735{
4736 unsigned long flags;
4737
4738 spin_lock_irqsave(&x->wait.lock, flags);
4739 x->done++;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004740 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004741 spin_unlock_irqrestore(&x->wait.lock, flags);
4742}
4743EXPORT_SYMBOL(complete);
4744
Ingo Molnarb15136e2007-10-24 18:23:48 +02004745void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004746{
4747 unsigned long flags;
4748
4749 spin_lock_irqsave(&x->wait.lock, flags);
4750 x->done += UINT_MAX/2;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004751 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004752 spin_unlock_irqrestore(&x->wait.lock, flags);
4753}
4754EXPORT_SYMBOL(complete_all);
4755
Andi Kleen8cbbe862007-10-15 17:00:14 +02004756static inline long __sched
4757do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004758{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004759 if (!x->done) {
4760 DECLARE_WAITQUEUE(wait, current);
4761
4762 wait.flags |= WQ_FLAG_EXCLUSIVE;
4763 __add_wait_queue_tail(&x->wait, &wait);
4764 do {
Matthew Wilcox009e5772007-12-06 12:29:54 -05004765 if ((state == TASK_INTERRUPTIBLE &&
4766 signal_pending(current)) ||
4767 (state == TASK_KILLABLE &&
4768 fatal_signal_pending(current))) {
Andi Kleen8cbbe862007-10-15 17:00:14 +02004769 __remove_wait_queue(&x->wait, &wait);
4770 return -ERESTARTSYS;
4771 }
4772 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004773 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004774 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004775 spin_lock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004776 if (!timeout) {
4777 __remove_wait_queue(&x->wait, &wait);
4778 return timeout;
4779 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004780 } while (!x->done);
4781 __remove_wait_queue(&x->wait, &wait);
4782 }
4783 x->done--;
Andi Kleen8cbbe862007-10-15 17:00:14 +02004784 return timeout;
4785}
4786
4787static long __sched
4788wait_for_common(struct completion *x, long timeout, int state)
4789{
4790 might_sleep();
4791
4792 spin_lock_irq(&x->wait.lock);
4793 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004794 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004795 return timeout;
4796}
4797
Ingo Molnarb15136e2007-10-24 18:23:48 +02004798void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02004799{
4800 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004801}
4802EXPORT_SYMBOL(wait_for_completion);
4803
Ingo Molnarb15136e2007-10-24 18:23:48 +02004804unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004805wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4806{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004807 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004808}
4809EXPORT_SYMBOL(wait_for_completion_timeout);
4810
Andi Kleen8cbbe862007-10-15 17:00:14 +02004811int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004812{
Andi Kleen51e97992007-10-18 21:32:55 +02004813 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4814 if (t == -ERESTARTSYS)
4815 return t;
4816 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004817}
4818EXPORT_SYMBOL(wait_for_completion_interruptible);
4819
Ingo Molnarb15136e2007-10-24 18:23:48 +02004820unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004821wait_for_completion_interruptible_timeout(struct completion *x,
4822 unsigned long timeout)
4823{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004824 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004825}
4826EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4827
Matthew Wilcox009e5772007-12-06 12:29:54 -05004828int __sched wait_for_completion_killable(struct completion *x)
4829{
4830 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4831 if (t == -ERESTARTSYS)
4832 return t;
4833 return 0;
4834}
4835EXPORT_SYMBOL(wait_for_completion_killable);
4836
Andi Kleen8cbbe862007-10-15 17:00:14 +02004837static long __sched
4838sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02004839{
4840 unsigned long flags;
4841 wait_queue_t wait;
4842
4843 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004844
Andi Kleen8cbbe862007-10-15 17:00:14 +02004845 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004846
Andi Kleen8cbbe862007-10-15 17:00:14 +02004847 spin_lock_irqsave(&q->lock, flags);
4848 __add_wait_queue(q, &wait);
4849 spin_unlock(&q->lock);
4850 timeout = schedule_timeout(timeout);
4851 spin_lock_irq(&q->lock);
4852 __remove_wait_queue(q, &wait);
4853 spin_unlock_irqrestore(&q->lock, flags);
4854
4855 return timeout;
4856}
4857
4858void __sched interruptible_sleep_on(wait_queue_head_t *q)
4859{
4860 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004861}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004862EXPORT_SYMBOL(interruptible_sleep_on);
4863
Ingo Molnar0fec1712007-07-09 18:52:01 +02004864long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004865interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004866{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004867 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004868}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004869EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4870
Ingo Molnar0fec1712007-07-09 18:52:01 +02004871void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004872{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004873 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004874}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004875EXPORT_SYMBOL(sleep_on);
4876
Ingo Molnar0fec1712007-07-09 18:52:01 +02004877long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004878{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004879 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004880}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004881EXPORT_SYMBOL(sleep_on_timeout);
4882
Ingo Molnarb29739f2006-06-27 02:54:51 -07004883#ifdef CONFIG_RT_MUTEXES
4884
4885/*
4886 * rt_mutex_setprio - set the current priority of a task
4887 * @p: task
4888 * @prio: prio value (kernel-internal form)
4889 *
4890 * This function changes the 'effective' priority of a task. It does
4891 * not touch ->normal_prio like __setscheduler().
4892 *
4893 * Used by the rt_mutex code to implement priority inheritance logic.
4894 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004895void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07004896{
4897 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004898 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004899 struct rq *rq;
Steven Rostedtcb469842008-01-25 21:08:22 +01004900 const struct sched_class *prev_class = p->sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004901
4902 BUG_ON(prio < 0 || prio > MAX_PRIO);
4903
4904 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004905 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004906
Andrew Mortond5f9f942007-05-08 20:27:06 -07004907 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004908 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004909 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004910 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02004911 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004912 if (running)
4913 p->sched_class->put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02004914
4915 if (rt_prio(prio))
4916 p->sched_class = &rt_sched_class;
4917 else
4918 p->sched_class = &fair_sched_class;
4919
Ingo Molnarb29739f2006-06-27 02:54:51 -07004920 p->prio = prio;
4921
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004922 if (running)
4923 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004924 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004925 enqueue_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004926
4927 check_class_changed(rq, p, prev_class, oldprio, running);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004928 }
4929 task_rq_unlock(rq, &flags);
4930}
4931
4932#endif
4933
Ingo Molnar36c8b582006-07-03 00:25:41 -07004934void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004935{
Ingo Molnardd41f592007-07-09 18:51:59 +02004936 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004937 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004938 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004939
4940 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4941 return;
4942 /*
4943 * We have to be careful, if called from sys_setpriority(),
4944 * the task might be in the middle of scheduling on another CPU.
4945 */
4946 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004947 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004948 /*
4949 * The RT priorities are set via sched_setscheduler(), but we still
4950 * allow the 'normal' nice value to be set - but as expected
4951 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02004952 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004953 */
Ingo Molnare05606d2007-07-09 18:51:59 +02004954 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004955 p->static_prio = NICE_TO_PRIO(nice);
4956 goto out_unlock;
4957 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004958 on_rq = p->se.on_rq;
Peter Zijlstra18d95a22008-04-19 19:45:00 +02004959 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02004960 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004961
Linus Torvalds1da177e2005-04-16 15:20:36 -07004962 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07004963 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004964 old_prio = p->prio;
4965 p->prio = effective_prio(p);
4966 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004967
Ingo Molnardd41f592007-07-09 18:51:59 +02004968 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004969 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004970 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07004971 * If the task increased its priority or is running and
4972 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004973 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07004974 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004975 resched_task(rq->curr);
4976 }
4977out_unlock:
4978 task_rq_unlock(rq, &flags);
4979}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004980EXPORT_SYMBOL(set_user_nice);
4981
Matt Mackalle43379f2005-05-01 08:59:00 -07004982/*
4983 * can_nice - check if a task can reduce its nice value
4984 * @p: task
4985 * @nice: nice value
4986 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004987int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07004988{
Matt Mackall024f4742005-08-18 11:24:19 -07004989 /* convert nice value [19,-20] to rlimit style value [1,40] */
4990 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004991
Matt Mackalle43379f2005-05-01 08:59:00 -07004992 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4993 capable(CAP_SYS_NICE));
4994}
4995
Linus Torvalds1da177e2005-04-16 15:20:36 -07004996#ifdef __ARCH_WANT_SYS_NICE
4997
4998/*
4999 * sys_nice - change the priority of the current process.
5000 * @increment: priority increment
5001 *
5002 * sys_setpriority is a more generic, but much slower function that
5003 * does similar things.
5004 */
5005asmlinkage long sys_nice(int increment)
5006{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005007 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005008
5009 /*
5010 * Setpriority might change our priority at the same moment.
5011 * We don't have to worry. Conceptually one call occurs first
5012 * and we have a single winner.
5013 */
Matt Mackalle43379f2005-05-01 08:59:00 -07005014 if (increment < -40)
5015 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005016 if (increment > 40)
5017 increment = 40;
5018
5019 nice = PRIO_TO_NICE(current->static_prio) + increment;
5020 if (nice < -20)
5021 nice = -20;
5022 if (nice > 19)
5023 nice = 19;
5024
Matt Mackalle43379f2005-05-01 08:59:00 -07005025 if (increment < 0 && !can_nice(current, nice))
5026 return -EPERM;
5027
Linus Torvalds1da177e2005-04-16 15:20:36 -07005028 retval = security_task_setnice(current, nice);
5029 if (retval)
5030 return retval;
5031
5032 set_user_nice(current, nice);
5033 return 0;
5034}
5035
5036#endif
5037
5038/**
5039 * task_prio - return the priority value of a given task.
5040 * @p: the task in question.
5041 *
5042 * This is the priority value as seen by users in /proc.
5043 * RT tasks are offset by -200. Normal tasks are centered
5044 * around 0, value goes from -16 to +15.
5045 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005046int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005047{
5048 return p->prio - MAX_RT_PRIO;
5049}
5050
5051/**
5052 * task_nice - return the nice value of a given task.
5053 * @p: the task in question.
5054 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005055int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005056{
5057 return TASK_NICE(p);
5058}
Pavel Roskin150d8be2008-03-05 16:56:37 -05005059EXPORT_SYMBOL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005060
5061/**
5062 * idle_cpu - is a given cpu idle currently?
5063 * @cpu: the processor in question.
5064 */
5065int idle_cpu(int cpu)
5066{
5067 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5068}
5069
Linus Torvalds1da177e2005-04-16 15:20:36 -07005070/**
5071 * idle_task - return the idle task for a given cpu.
5072 * @cpu: the processor in question.
5073 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005074struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005075{
5076 return cpu_rq(cpu)->idle;
5077}
5078
5079/**
5080 * find_process_by_pid - find a process with a matching PID value.
5081 * @pid: the pid in question.
5082 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02005083static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005084{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07005085 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005086}
5087
5088/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02005089static void
5090__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005091{
Ingo Molnardd41f592007-07-09 18:51:59 +02005092 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005093
Linus Torvalds1da177e2005-04-16 15:20:36 -07005094 p->policy = policy;
Ingo Molnardd41f592007-07-09 18:51:59 +02005095 switch (p->policy) {
5096 case SCHED_NORMAL:
5097 case SCHED_BATCH:
5098 case SCHED_IDLE:
5099 p->sched_class = &fair_sched_class;
5100 break;
5101 case SCHED_FIFO:
5102 case SCHED_RR:
5103 p->sched_class = &rt_sched_class;
5104 break;
5105 }
5106
Linus Torvalds1da177e2005-04-16 15:20:36 -07005107 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07005108 p->normal_prio = normal_prio(p);
5109 /* we are holding p->pi_lock already */
5110 p->prio = rt_mutex_getprio(p);
Peter Williams2dd73a42006-06-27 02:54:34 -07005111 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005112}
5113
5114/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08005115 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005116 * @p: the task in question.
5117 * @policy: new policy.
5118 * @param: structure containing the new RT priority.
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005119 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08005120 * NOTE that the task may be already dead.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005121 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005122int sched_setscheduler(struct task_struct *p, int policy,
5123 struct sched_param *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005124{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02005125 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005126 unsigned long flags;
Steven Rostedtcb469842008-01-25 21:08:22 +01005127 const struct sched_class *prev_class = p->sched_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005128 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005129
Steven Rostedt66e53932006-06-27 02:54:44 -07005130 /* may grab non-irq protected spin_locks */
5131 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005132recheck:
5133 /* double check policy once rq lock held */
5134 if (policy < 0)
5135 policy = oldpolicy = p->policy;
5136 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
Ingo Molnardd41f592007-07-09 18:51:59 +02005137 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5138 policy != SCHED_IDLE)
Ingo Molnarb0a94992006-01-14 13:20:41 -08005139 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005140 /*
5141 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02005142 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5143 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005144 */
5145 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005146 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04005147 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005148 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02005149 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005150 return -EINVAL;
5151
Olivier Croquette37e4ab32005-06-25 14:57:32 -07005152 /*
5153 * Allow unprivileged RT tasks to decrease priority:
5154 */
5155 if (!capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02005156 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07005157 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005158
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07005159 if (!lock_task_sighand(p, &flags))
5160 return -ESRCH;
5161 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5162 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005163
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07005164 /* can't set/change the rt policy */
5165 if (policy != p->policy && !rlim_rtprio)
5166 return -EPERM;
5167
5168 /* can't increase priority */
5169 if (param->sched_priority > p->rt_priority &&
5170 param->sched_priority > rlim_rtprio)
5171 return -EPERM;
5172 }
Ingo Molnardd41f592007-07-09 18:51:59 +02005173 /*
5174 * Like positive nice levels, dont allow tasks to
5175 * move out of SCHED_IDLE either:
5176 */
5177 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5178 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07005179
Olivier Croquette37e4ab32005-06-25 14:57:32 -07005180 /* can't change other user's priorities */
5181 if ((current->euid != p->euid) &&
5182 (current->euid != p->uid))
5183 return -EPERM;
5184 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005185
Peter Zijlstrab68aa232008-02-13 15:45:40 +01005186#ifdef CONFIG_RT_GROUP_SCHED
5187 /*
5188 * Do not allow realtime tasks into groups that have no runtime
5189 * assigned.
5190 */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02005191 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
Peter Zijlstrab68aa232008-02-13 15:45:40 +01005192 return -EPERM;
5193#endif
5194
Linus Torvalds1da177e2005-04-16 15:20:36 -07005195 retval = security_task_setscheduler(p, policy, param);
5196 if (retval)
5197 return retval;
5198 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07005199 * make sure no PI-waiters arrive (or leave) while we are
5200 * changing the priority of the task:
5201 */
5202 spin_lock_irqsave(&p->pi_lock, flags);
5203 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005204 * To be able to change p->policy safely, the apropriate
5205 * runqueue lock must be held.
5206 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07005207 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005208 /* recheck policy now with rq lock held */
5209 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5210 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07005211 __task_rq_unlock(rq);
5212 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005213 goto recheck;
5214 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02005215 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005216 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01005217 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005218 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005219 deactivate_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005220 if (running)
5221 p->sched_class->put_prev_task(rq, p);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02005222
Linus Torvalds1da177e2005-04-16 15:20:36 -07005223 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02005224 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02005225
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005226 if (running)
5227 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005228 if (on_rq) {
5229 activate_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01005230
5231 check_class_changed(rq, p, prev_class, oldprio, running);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005232 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07005233 __task_rq_unlock(rq);
5234 spin_unlock_irqrestore(&p->pi_lock, flags);
5235
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07005236 rt_mutex_adjust_pi(p);
5237
Linus Torvalds1da177e2005-04-16 15:20:36 -07005238 return 0;
5239}
5240EXPORT_SYMBOL_GPL(sched_setscheduler);
5241
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005242static int
5243do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005244{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005245 struct sched_param lparam;
5246 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005247 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005248
5249 if (!param || pid < 0)
5250 return -EINVAL;
5251 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5252 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005253
5254 rcu_read_lock();
5255 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005256 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005257 if (p != NULL)
5258 retval = sched_setscheduler(p, policy, &lparam);
5259 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07005260
Linus Torvalds1da177e2005-04-16 15:20:36 -07005261 return retval;
5262}
5263
5264/**
5265 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5266 * @pid: the pid in question.
5267 * @policy: new policy.
5268 * @param: structure containing the new RT priority.
5269 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005270asmlinkage long
5271sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005272{
Jason Baronc21761f2006-01-18 17:43:03 -08005273 /* negative values for policy are not valid */
5274 if (policy < 0)
5275 return -EINVAL;
5276
Linus Torvalds1da177e2005-04-16 15:20:36 -07005277 return do_sched_setscheduler(pid, policy, param);
5278}
5279
5280/**
5281 * sys_sched_setparam - set/change the RT priority of a thread
5282 * @pid: the pid in question.
5283 * @param: structure containing the new RT priority.
5284 */
5285asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5286{
5287 return do_sched_setscheduler(pid, -1, param);
5288}
5289
5290/**
5291 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5292 * @pid: the pid in question.
5293 */
5294asmlinkage long sys_sched_getscheduler(pid_t pid)
5295{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005296 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005297 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005298
5299 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005300 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005301
5302 retval = -ESRCH;
5303 read_lock(&tasklist_lock);
5304 p = find_process_by_pid(pid);
5305 if (p) {
5306 retval = security_task_getscheduler(p);
5307 if (!retval)
5308 retval = p->policy;
5309 }
5310 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005311 return retval;
5312}
5313
5314/**
5315 * sys_sched_getscheduler - get the RT priority of a thread
5316 * @pid: the pid in question.
5317 * @param: structure containing the RT priority.
5318 */
5319asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5320{
5321 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005322 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005323 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005324
5325 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005326 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005327
5328 read_lock(&tasklist_lock);
5329 p = find_process_by_pid(pid);
5330 retval = -ESRCH;
5331 if (!p)
5332 goto out_unlock;
5333
5334 retval = security_task_getscheduler(p);
5335 if (retval)
5336 goto out_unlock;
5337
5338 lp.sched_priority = p->rt_priority;
5339 read_unlock(&tasklist_lock);
5340
5341 /*
5342 * This one might sleep, we cannot do it with a spinlock held ...
5343 */
5344 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5345
Linus Torvalds1da177e2005-04-16 15:20:36 -07005346 return retval;
5347
5348out_unlock:
5349 read_unlock(&tasklist_lock);
5350 return retval;
5351}
5352
Mike Travisb53e9212008-04-04 18:11:08 -07005353long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005354{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005355 cpumask_t cpus_allowed;
Mike Travisb53e9212008-04-04 18:11:08 -07005356 cpumask_t new_mask = *in_mask;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005357 struct task_struct *p;
5358 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005359
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005360 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005361 read_lock(&tasklist_lock);
5362
5363 p = find_process_by_pid(pid);
5364 if (!p) {
5365 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005366 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005367 return -ESRCH;
5368 }
5369
5370 /*
5371 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005372 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005373 * usage count and then drop tasklist_lock.
5374 */
5375 get_task_struct(p);
5376 read_unlock(&tasklist_lock);
5377
5378 retval = -EPERM;
5379 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5380 !capable(CAP_SYS_NICE))
5381 goto out_unlock;
5382
David Quigleye7834f82006-06-23 02:03:59 -07005383 retval = security_task_setscheduler(p, 0, NULL);
5384 if (retval)
5385 goto out_unlock;
5386
Mike Travisf9a86fc2008-04-04 18:11:07 -07005387 cpuset_cpus_allowed(p, &cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005388 cpus_and(new_mask, new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07005389 again:
Mike Travis7c16ec52008-04-04 18:11:11 -07005390 retval = set_cpus_allowed_ptr(p, &new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005391
Paul Menage8707d8b2007-10-18 23:40:22 -07005392 if (!retval) {
Mike Travisf9a86fc2008-04-04 18:11:07 -07005393 cpuset_cpus_allowed(p, &cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07005394 if (!cpus_subset(new_mask, cpus_allowed)) {
5395 /*
5396 * We must have raced with a concurrent cpuset
5397 * update. Just reset the cpus_allowed to the
5398 * cpuset's cpus_allowed
5399 */
5400 new_mask = cpus_allowed;
5401 goto again;
5402 }
5403 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005404out_unlock:
5405 put_task_struct(p);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005406 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005407 return retval;
5408}
5409
5410static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5411 cpumask_t *new_mask)
5412{
5413 if (len < sizeof(cpumask_t)) {
5414 memset(new_mask, 0, sizeof(cpumask_t));
5415 } else if (len > sizeof(cpumask_t)) {
5416 len = sizeof(cpumask_t);
5417 }
5418 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5419}
5420
5421/**
5422 * sys_sched_setaffinity - set the cpu affinity of a process
5423 * @pid: pid of the process
5424 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5425 * @user_mask_ptr: user-space pointer to the new cpu mask
5426 */
5427asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5428 unsigned long __user *user_mask_ptr)
5429{
5430 cpumask_t new_mask;
5431 int retval;
5432
5433 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5434 if (retval)
5435 return retval;
5436
Mike Travisb53e9212008-04-04 18:11:08 -07005437 return sched_setaffinity(pid, &new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005438}
5439
5440/*
5441 * Represents all cpu's present in the system
5442 * In systems capable of hotplug, this map could dynamically grow
5443 * as new cpu's are detected in the system via any platform specific
5444 * method, such as ACPI for e.g.
5445 */
5446
Andi Kleen4cef0c62006-01-11 22:44:57 +01005447cpumask_t cpu_present_map __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005448EXPORT_SYMBOL(cpu_present_map);
5449
5450#ifndef CONFIG_SMP
Andi Kleen4cef0c62006-01-11 22:44:57 +01005451cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07005452EXPORT_SYMBOL(cpu_online_map);
5453
Andi Kleen4cef0c62006-01-11 22:44:57 +01005454cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07005455EXPORT_SYMBOL(cpu_possible_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005456#endif
5457
5458long sched_getaffinity(pid_t pid, cpumask_t *mask)
5459{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005460 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005461 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005462
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005463 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005464 read_lock(&tasklist_lock);
5465
5466 retval = -ESRCH;
5467 p = find_process_by_pid(pid);
5468 if (!p)
5469 goto out_unlock;
5470
David Quigleye7834f82006-06-23 02:03:59 -07005471 retval = security_task_getscheduler(p);
5472 if (retval)
5473 goto out_unlock;
5474
Jack Steiner2f7016d2006-02-01 03:05:18 -08005475 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005476
5477out_unlock:
5478 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005479 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005480
Ulrich Drepper9531b622007-08-09 11:16:46 +02005481 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005482}
5483
5484/**
5485 * sys_sched_getaffinity - get the cpu affinity of a process
5486 * @pid: pid of the process
5487 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5488 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5489 */
5490asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5491 unsigned long __user *user_mask_ptr)
5492{
5493 int ret;
5494 cpumask_t mask;
5495
5496 if (len < sizeof(cpumask_t))
5497 return -EINVAL;
5498
5499 ret = sched_getaffinity(pid, &mask);
5500 if (ret < 0)
5501 return ret;
5502
5503 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5504 return -EFAULT;
5505
5506 return sizeof(cpumask_t);
5507}
5508
5509/**
5510 * sys_sched_yield - yield the current processor to other threads.
5511 *
Ingo Molnardd41f592007-07-09 18:51:59 +02005512 * This function yields the current CPU to other tasks. If there are no
5513 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005514 */
5515asmlinkage long sys_sched_yield(void)
5516{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005517 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005518
Ingo Molnar2d723762007-10-15 17:00:12 +02005519 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02005520 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005521
5522 /*
5523 * Since we are going to call schedule() anyway, there's
5524 * no need to preempt or enable interrupts:
5525 */
5526 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07005527 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005528 _raw_spin_unlock(&rq->lock);
5529 preempt_enable_no_resched();
5530
5531 schedule();
5532
5533 return 0;
5534}
5535
Andrew Mortone7b38402006-06-30 01:56:00 -07005536static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005537{
Ingo Molnar8e0a43d2006-06-23 02:05:23 -07005538#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5539 __might_sleep(__FILE__, __LINE__);
5540#endif
Ingo Molnar5bbcfd92005-07-07 17:57:04 -07005541 /*
5542 * The BKS might be reacquired before we have dropped
5543 * PREEMPT_ACTIVE, which could trigger a second
5544 * cond_resched() call.
5545 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005546 do {
5547 add_preempt_count(PREEMPT_ACTIVE);
5548 schedule();
5549 sub_preempt_count(PREEMPT_ACTIVE);
5550 } while (need_resched());
5551}
5552
Herbert Xu02b67cc2008-01-25 21:08:28 +01005553int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005554{
Ingo Molnar94142322006-12-29 16:48:13 -08005555 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5556 system_state == SYSTEM_RUNNING) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005557 __cond_resched();
5558 return 1;
5559 }
5560 return 0;
5561}
Herbert Xu02b67cc2008-01-25 21:08:28 +01005562EXPORT_SYMBOL(_cond_resched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005563
5564/*
5565 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5566 * call schedule, and on return reacquire the lock.
5567 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005568 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07005569 * operations here to prevent schedule() from being called twice (once via
5570 * spin_unlock(), once by hand).
5571 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005572int cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005573{
Nick Piggin95c354f2008-01-30 13:31:20 +01005574 int resched = need_resched() && system_state == SYSTEM_RUNNING;
Jan Kara6df3cec2005-06-13 15:52:32 -07005575 int ret = 0;
5576
Nick Piggin95c354f2008-01-30 13:31:20 +01005577 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005578 spin_unlock(lock);
Nick Piggin95c354f2008-01-30 13:31:20 +01005579 if (resched && need_resched())
5580 __cond_resched();
5581 else
5582 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07005583 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005584 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005585 }
Jan Kara6df3cec2005-06-13 15:52:32 -07005586 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005587}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005588EXPORT_SYMBOL(cond_resched_lock);
5589
5590int __sched cond_resched_softirq(void)
5591{
5592 BUG_ON(!in_softirq());
5593
Ingo Molnar94142322006-12-29 16:48:13 -08005594 if (need_resched() && system_state == SYSTEM_RUNNING) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07005595 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005596 __cond_resched();
5597 local_bh_disable();
5598 return 1;
5599 }
5600 return 0;
5601}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005602EXPORT_SYMBOL(cond_resched_softirq);
5603
Linus Torvalds1da177e2005-04-16 15:20:36 -07005604/**
5605 * yield - yield the current processor to other threads.
5606 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08005607 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005608 * thread runnable and calls sys_sched_yield().
5609 */
5610void __sched yield(void)
5611{
5612 set_current_state(TASK_RUNNING);
5613 sys_sched_yield();
5614}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005615EXPORT_SYMBOL(yield);
5616
5617/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005618 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07005619 * that process accounting knows that this is a task in IO wait state.
5620 *
5621 * But don't do that if it is a deliberate, throttling IO wait (this task
5622 * has set its backing_dev_info: the queue against which it should throttle)
5623 */
5624void __sched io_schedule(void)
5625{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005626 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005627
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005628 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005629 atomic_inc(&rq->nr_iowait);
5630 schedule();
5631 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005632 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005633}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005634EXPORT_SYMBOL(io_schedule);
5635
5636long __sched io_schedule_timeout(long timeout)
5637{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005638 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005639 long ret;
5640
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005641 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005642 atomic_inc(&rq->nr_iowait);
5643 ret = schedule_timeout(timeout);
5644 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005645 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005646 return ret;
5647}
5648
5649/**
5650 * sys_sched_get_priority_max - return maximum RT priority.
5651 * @policy: scheduling class.
5652 *
5653 * this syscall returns the maximum rt_priority that can be used
5654 * by a given scheduling class.
5655 */
5656asmlinkage long sys_sched_get_priority_max(int policy)
5657{
5658 int ret = -EINVAL;
5659
5660 switch (policy) {
5661 case SCHED_FIFO:
5662 case SCHED_RR:
5663 ret = MAX_USER_RT_PRIO-1;
5664 break;
5665 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005666 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005667 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005668 ret = 0;
5669 break;
5670 }
5671 return ret;
5672}
5673
5674/**
5675 * sys_sched_get_priority_min - return minimum RT priority.
5676 * @policy: scheduling class.
5677 *
5678 * this syscall returns the minimum rt_priority that can be used
5679 * by a given scheduling class.
5680 */
5681asmlinkage long sys_sched_get_priority_min(int policy)
5682{
5683 int ret = -EINVAL;
5684
5685 switch (policy) {
5686 case SCHED_FIFO:
5687 case SCHED_RR:
5688 ret = 1;
5689 break;
5690 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005691 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005692 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005693 ret = 0;
5694 }
5695 return ret;
5696}
5697
5698/**
5699 * sys_sched_rr_get_interval - return the default timeslice of a process.
5700 * @pid: pid of the process.
5701 * @interval: userspace pointer to the timeslice value.
5702 *
5703 * this syscall writes the default timeslice value of a given process
5704 * into the user-space timespec buffer. A value of '0' means infinity.
5705 */
5706asmlinkage
5707long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5708{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005709 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005710 unsigned int time_slice;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005711 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005712 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005713
5714 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005715 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005716
5717 retval = -ESRCH;
5718 read_lock(&tasklist_lock);
5719 p = find_process_by_pid(pid);
5720 if (!p)
5721 goto out_unlock;
5722
5723 retval = security_task_getscheduler(p);
5724 if (retval)
5725 goto out_unlock;
5726
Ingo Molnar77034932007-12-04 17:04:39 +01005727 /*
5728 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5729 * tasks that are on an otherwise idle runqueue:
5730 */
5731 time_slice = 0;
5732 if (p->policy == SCHED_RR) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005733 time_slice = DEF_TIMESLICE;
Miao Xie1868f952008-03-07 09:35:06 +08005734 } else if (p->policy != SCHED_FIFO) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005735 struct sched_entity *se = &p->se;
5736 unsigned long flags;
5737 struct rq *rq;
5738
5739 rq = task_rq_lock(p, &flags);
Ingo Molnar77034932007-12-04 17:04:39 +01005740 if (rq->cfs.load.weight)
5741 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005742 task_rq_unlock(rq, &flags);
5743 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005744 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005745 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005746 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005747 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005748
Linus Torvalds1da177e2005-04-16 15:20:36 -07005749out_unlock:
5750 read_unlock(&tasklist_lock);
5751 return retval;
5752}
5753
Steven Rostedt7c731e02008-05-12 21:20:41 +02005754static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005755
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005756void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005757{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005758 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005759 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005760
Linus Torvalds1da177e2005-04-16 15:20:36 -07005761 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005762 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005763 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02005764#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07005765 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005766 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005767 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005768 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005769#else
5770 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005771 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005772 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005773 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005774#endif
5775#ifdef CONFIG_DEBUG_STACK_USAGE
5776 {
Al Viro10ebffd2005-11-13 16:06:56 -08005777 unsigned long *n = end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005778 while (!*n)
5779 n++;
Al Viro10ebffd2005-11-13 16:06:56 -08005780 free = (unsigned long)n - (unsigned long)end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005781 }
5782#endif
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -07005783 printk(KERN_CONT "%5lu %5d %6d\n", free,
Roland McGrathfcfd50a2008-01-09 00:03:23 -08005784 task_pid_nr(p), task_pid_nr(p->real_parent));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005785
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01005786 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005787}
5788
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005789void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005790{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005791 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005792
Ingo Molnar4bd77322007-07-11 21:21:47 +02005793#if BITS_PER_LONG == 32
5794 printk(KERN_INFO
5795 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005796#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02005797 printk(KERN_INFO
5798 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005799#endif
5800 read_lock(&tasklist_lock);
5801 do_each_thread(g, p) {
5802 /*
5803 * reset the NMI-timeout, listing all files on a slow
5804 * console might take alot of time:
5805 */
5806 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07005807 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005808 sched_show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005809 } while_each_thread(g, p);
5810
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07005811 touch_all_softlockup_watchdogs();
5812
Ingo Molnardd41f592007-07-09 18:51:59 +02005813#ifdef CONFIG_SCHED_DEBUG
5814 sysrq_sched_debug_show();
5815#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005816 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005817 /*
5818 * Only show locks if all tasks are dumped:
5819 */
5820 if (state_filter == -1)
5821 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005822}
5823
Ingo Molnar1df21052007-07-09 18:51:58 +02005824void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5825{
Ingo Molnardd41f592007-07-09 18:51:59 +02005826 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02005827}
5828
Ingo Molnarf340c0d2005-06-28 16:40:42 +02005829/**
5830 * init_idle - set up an idle thread for a given CPU
5831 * @idle: task in question
5832 * @cpu: cpu the idle task belongs to
5833 *
5834 * NOTE: this function does not set the idle thread's NEED_RESCHED
5835 * flag, to make booting more robust.
5836 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07005837void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005838{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005839 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005840 unsigned long flags;
5841
Ingo Molnardd41f592007-07-09 18:51:59 +02005842 __sched_fork(idle);
5843 idle->se.exec_start = sched_clock();
5844
Ingo Molnarb29739f2006-06-27 02:54:51 -07005845 idle->prio = idle->normal_prio = MAX_PRIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005846 idle->cpus_allowed = cpumask_of_cpu(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005847 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005848
5849 spin_lock_irqsave(&rq->lock, flags);
5850 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07005851#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5852 idle->oncpu = 1;
5853#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005854 spin_unlock_irqrestore(&rq->lock, flags);
5855
5856 /* Set the preempt count _outside_ the spinlocks! */
Linus Torvalds8e3e0762008-05-10 20:58:02 -07005857#if defined(CONFIG_PREEMPT)
5858 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5859#else
Al Viroa1261f52005-11-13 16:06:55 -08005860 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds8e3e0762008-05-10 20:58:02 -07005861#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02005862 /*
5863 * The idle tasks have their own, simple scheduling class:
5864 */
5865 idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005866}
5867
5868/*
5869 * In a system that switches off the HZ timer nohz_cpu_mask
5870 * indicates which cpus entered this state. This is used
5871 * in the rcu update to wait only for active cpus. For system
5872 * which do not switch off the HZ timer nohz_cpu_mask should
5873 * always be CPU_MASK_NONE.
5874 */
5875cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5876
Ingo Molnar19978ca2007-11-09 22:39:38 +01005877/*
5878 * Increase the granularity value when there are more CPUs,
5879 * because with more CPUs the 'effective latency' as visible
5880 * to users decreases. But the relationship is not linear,
5881 * so pick a second-best guess by going with the log2 of the
5882 * number of CPUs.
5883 *
5884 * This idea comes from the SD scheduler of Con Kolivas:
5885 */
5886static inline void sched_init_granularity(void)
5887{
5888 unsigned int factor = 1 + ilog2(num_online_cpus());
5889 const unsigned long limit = 200000000;
5890
5891 sysctl_sched_min_granularity *= factor;
5892 if (sysctl_sched_min_granularity > limit)
5893 sysctl_sched_min_granularity = limit;
5894
5895 sysctl_sched_latency *= factor;
5896 if (sysctl_sched_latency > limit)
5897 sysctl_sched_latency = limit;
5898
5899 sysctl_sched_wakeup_granularity *= factor;
Ingo Molnar19978ca2007-11-09 22:39:38 +01005900}
5901
Linus Torvalds1da177e2005-04-16 15:20:36 -07005902#ifdef CONFIG_SMP
5903/*
5904 * This is how migration works:
5905 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07005906 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005907 * runqueue and wake up that CPU's migration thread.
5908 * 2) we down() the locked semaphore => thread blocks.
5909 * 3) migration thread wakes up (implicitly it forces the migrated
5910 * thread off the CPU)
5911 * 4) it gets the migration request and checks whether the migrated
5912 * task is still in the wrong runqueue.
5913 * 5) if it's in the wrong runqueue then the migration thread removes
5914 * it and puts it into the right queue.
5915 * 6) migration thread up()s the semaphore.
5916 * 7) we wake up and the migration is done.
5917 */
5918
5919/*
5920 * Change a given task's CPU affinity. Migrate the thread to a
5921 * proper CPU and schedule it away if the CPU it's executing on
5922 * is removed from the allowed bitmask.
5923 *
5924 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005925 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07005926 * call is not atomic; no spinlocks may be held.
5927 */
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005928int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005929{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005930 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005931 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005932 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005933 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005934
5935 rq = task_rq_lock(p, &flags);
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005936 if (!cpus_intersects(*new_mask, cpu_online_map)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005937 ret = -EINVAL;
5938 goto out;
5939 }
5940
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005941 if (p->sched_class->set_cpus_allowed)
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005942 p->sched_class->set_cpus_allowed(p, new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005943 else {
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005944 p->cpus_allowed = *new_mask;
5945 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005946 }
5947
Linus Torvalds1da177e2005-04-16 15:20:36 -07005948 /* Can the task run on the task's current CPU? If so, we're done */
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005949 if (cpu_isset(task_cpu(p), *new_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005950 goto out;
5951
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005952 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005953 /* Need help from migration thread: drop lock and wait. */
5954 task_rq_unlock(rq, &flags);
5955 wake_up_process(rq->migration_thread);
5956 wait_for_completion(&req.done);
5957 tlb_migrate_finish(p->mm);
5958 return 0;
5959 }
5960out:
5961 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005962
Linus Torvalds1da177e2005-04-16 15:20:36 -07005963 return ret;
5964}
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005965EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005966
5967/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005968 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07005969 * this because either it can't run here any more (set_cpus_allowed()
5970 * away from this CPU, or CPU going down), or because we're
5971 * attempting to rebalance this task on exec (sched_exec).
5972 *
5973 * So we race with normal scheduler movements, but that's OK, as long
5974 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07005975 *
5976 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005977 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07005978static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005979{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005980 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02005981 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005982
5983 if (unlikely(cpu_is_offline(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07005984 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005985
5986 rq_src = cpu_rq(src_cpu);
5987 rq_dest = cpu_rq(dest_cpu);
5988
5989 double_rq_lock(rq_src, rq_dest);
5990 /* Already moved. */
5991 if (task_cpu(p) != src_cpu)
5992 goto out;
5993 /* Affinity changed (again). */
5994 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5995 goto out;
5996
Ingo Molnardd41f592007-07-09 18:51:59 +02005997 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02005998 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005999 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02006000
Linus Torvalds1da177e2005-04-16 15:20:36 -07006001 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02006002 if (on_rq) {
6003 activate_task(rq_dest, p, 0);
6004 check_preempt_curr(rq_dest, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006005 }
Kirill Korotaevefc30812006-06-27 02:54:32 -07006006 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006007out:
6008 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07006009 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006010}
6011
6012/*
6013 * migration_thread - this is a highprio system thread that performs
6014 * thread migration by bumping thread off CPU then 'pushing' onto
6015 * another runqueue.
6016 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006017static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006018{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006019 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006020 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006021
6022 rq = cpu_rq(cpu);
6023 BUG_ON(rq->migration_thread != current);
6024
6025 set_current_state(TASK_INTERRUPTIBLE);
6026 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07006027 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006028 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006029
Linus Torvalds1da177e2005-04-16 15:20:36 -07006030 spin_lock_irq(&rq->lock);
6031
6032 if (cpu_is_offline(cpu)) {
6033 spin_unlock_irq(&rq->lock);
6034 goto wait_to_die;
6035 }
6036
6037 if (rq->active_balance) {
6038 active_load_balance(rq, cpu);
6039 rq->active_balance = 0;
6040 }
6041
6042 head = &rq->migration_queue;
6043
6044 if (list_empty(head)) {
6045 spin_unlock_irq(&rq->lock);
6046 schedule();
6047 set_current_state(TASK_INTERRUPTIBLE);
6048 continue;
6049 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07006050 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006051 list_del_init(head->next);
6052
Nick Piggin674311d2005-06-25 14:57:27 -07006053 spin_unlock(&rq->lock);
6054 __migrate_task(req->task, cpu, req->dest_cpu);
6055 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006056
6057 complete(&req->done);
6058 }
6059 __set_current_state(TASK_RUNNING);
6060 return 0;
6061
6062wait_to_die:
6063 /* Wait for kthread_stop */
6064 set_current_state(TASK_INTERRUPTIBLE);
6065 while (!kthread_should_stop()) {
6066 schedule();
6067 set_current_state(TASK_INTERRUPTIBLE);
6068 }
6069 __set_current_state(TASK_RUNNING);
6070 return 0;
6071}
6072
6073#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006074
6075static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6076{
6077 int ret;
6078
6079 local_irq_disable();
6080 ret = __migrate_task(p, src_cpu, dest_cpu);
6081 local_irq_enable();
6082 return ret;
6083}
6084
Kirill Korotaev054b9102006-12-10 02:20:11 -08006085/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02006086 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08006087 * NOTE: interrupts should be disabled by the caller
6088 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006089static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006090{
Kirill Korotaevefc30812006-06-27 02:54:32 -07006091 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006092 cpumask_t mask;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006093 struct rq *rq;
6094 int dest_cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006095
Andi Kleen3a5c3592007-10-15 17:00:14 +02006096 do {
6097 /* On same node? */
6098 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6099 cpus_and(mask, mask, p->cpus_allowed);
6100 dest_cpu = any_online_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006101
Andi Kleen3a5c3592007-10-15 17:00:14 +02006102 /* On any allowed CPU? */
Mike Travis434d53b2008-04-04 18:11:04 -07006103 if (dest_cpu >= nr_cpu_ids)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006104 dest_cpu = any_online_cpu(p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006105
Andi Kleen3a5c3592007-10-15 17:00:14 +02006106 /* No more Mr. Nice Guy. */
Mike Travis434d53b2008-04-04 18:11:04 -07006107 if (dest_cpu >= nr_cpu_ids) {
Mike Travisf9a86fc2008-04-04 18:11:07 -07006108 cpumask_t cpus_allowed;
6109
6110 cpuset_cpus_allowed_locked(p, &cpus_allowed);
Cliff Wickman470fd642007-10-18 23:40:46 -07006111 /*
6112 * Try to stay on the same cpuset, where the
6113 * current cpuset may be a subset of all cpus.
6114 * The cpuset_cpus_allowed_locked() variant of
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006115 * cpuset_cpus_allowed() will not block. It must be
Cliff Wickman470fd642007-10-18 23:40:46 -07006116 * called within calls to cpuset_lock/cpuset_unlock.
6117 */
Andi Kleen3a5c3592007-10-15 17:00:14 +02006118 rq = task_rq_lock(p, &flags);
Cliff Wickman470fd642007-10-18 23:40:46 -07006119 p->cpus_allowed = cpus_allowed;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006120 dest_cpu = any_online_cpu(p->cpus_allowed);
6121 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006122
Andi Kleen3a5c3592007-10-15 17:00:14 +02006123 /*
6124 * Don't tell them about moving exiting tasks or
6125 * kernel threads (both mm NULL), since they never
6126 * leave kernel.
6127 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006128 if (p->mm && printk_ratelimit()) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02006129 printk(KERN_INFO "process %d (%s) no "
6130 "longer affine to cpu%d\n",
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006131 task_pid_nr(p), p->comm, dead_cpu);
6132 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02006133 }
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006134 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006135}
6136
6137/*
6138 * While a dead CPU has no uninterruptible tasks queued at this point,
6139 * it might still have a nonzero ->nr_uninterruptible counter, because
6140 * for performance reasons the counter is not stricly tracking tasks to
6141 * their home CPUs. So we just add the counter to another CPU's counter,
6142 * to keep the global sum constant after CPU-down:
6143 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07006144static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006145{
Mike Travis7c16ec52008-04-04 18:11:11 -07006146 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006147 unsigned long flags;
6148
6149 local_irq_save(flags);
6150 double_rq_lock(rq_src, rq_dest);
6151 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6152 rq_src->nr_uninterruptible = 0;
6153 double_rq_unlock(rq_src, rq_dest);
6154 local_irq_restore(flags);
6155}
6156
6157/* Run through task list and migrate tasks from the dead cpu. */
6158static void migrate_live_tasks(int src_cpu)
6159{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006160 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006161
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006162 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006163
Ingo Molnar48f24c42006-07-03 00:25:40 -07006164 do_each_thread(t, p) {
6165 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006166 continue;
6167
Ingo Molnar48f24c42006-07-03 00:25:40 -07006168 if (task_cpu(p) == src_cpu)
6169 move_task_off_dead_cpu(src_cpu, p);
6170 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006171
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006172 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006173}
6174
Ingo Molnardd41f592007-07-09 18:51:59 +02006175/*
6176 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01006177 * It does so by boosting its priority to highest possible.
6178 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006179 */
6180void sched_idle_next(void)
6181{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006182 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07006183 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006184 struct task_struct *p = rq->idle;
6185 unsigned long flags;
6186
6187 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006188 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006189
Ingo Molnar48f24c42006-07-03 00:25:40 -07006190 /*
6191 * Strictly not necessary since rest of the CPUs are stopped by now
6192 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006193 */
6194 spin_lock_irqsave(&rq->lock, flags);
6195
Ingo Molnardd41f592007-07-09 18:51:59 +02006196 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006197
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01006198 update_rq_clock(rq);
6199 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006200
6201 spin_unlock_irqrestore(&rq->lock, flags);
6202}
6203
Ingo Molnar48f24c42006-07-03 00:25:40 -07006204/*
6205 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07006206 * offline.
6207 */
6208void idle_task_exit(void)
6209{
6210 struct mm_struct *mm = current->active_mm;
6211
6212 BUG_ON(cpu_online(smp_processor_id()));
6213
6214 if (mm != &init_mm)
6215 switch_mm(mm, &init_mm, current);
6216 mmdrop(mm);
6217}
6218
Kirill Korotaev054b9102006-12-10 02:20:11 -08006219/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006220static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006221{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006222 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006223
6224 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07006225 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006226
6227 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07006228 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006229
Ingo Molnar48f24c42006-07-03 00:25:40 -07006230 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006231
6232 /*
6233 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006234 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07006235 * fine.
6236 */
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006237 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006238 move_task_off_dead_cpu(dead_cpu, p);
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006239 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006240
Ingo Molnar48f24c42006-07-03 00:25:40 -07006241 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006242}
6243
6244/* release_task() removes task from tasklist, so we won't find dead tasks. */
6245static void migrate_dead_tasks(unsigned int dead_cpu)
6246{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006247 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02006248 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006249
Ingo Molnardd41f592007-07-09 18:51:59 +02006250 for ( ; ; ) {
6251 if (!rq->nr_running)
6252 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02006253 update_rq_clock(rq);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02006254 next = pick_next_task(rq, rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02006255 if (!next)
6256 break;
6257 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02006258
Linus Torvalds1da177e2005-04-16 15:20:36 -07006259 }
6260}
6261#endif /* CONFIG_HOTPLUG_CPU */
6262
Nick Piggine692ab52007-07-26 13:40:43 +02006263#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6264
6265static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02006266 {
6267 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006268 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02006269 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01006270 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02006271};
6272
6273static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02006274 {
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006275 .ctl_name = CTL_KERN,
Alexey Dobriyane0361852007-08-09 11:16:46 +02006276 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006277 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02006278 .child = sd_ctl_dir,
6279 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01006280 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02006281};
6282
6283static struct ctl_table *sd_alloc_ctl_entry(int n)
6284{
6285 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02006286 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02006287
Nick Piggine692ab52007-07-26 13:40:43 +02006288 return entry;
6289}
6290
Milton Miller6382bc92007-10-15 17:00:19 +02006291static void sd_free_ctl_entry(struct ctl_table **tablep)
6292{
Milton Millercd790072007-10-17 16:55:11 +02006293 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02006294
Milton Millercd790072007-10-17 16:55:11 +02006295 /*
6296 * In the intermediate directories, both the child directory and
6297 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006298 * will always be set. In the lowest directory the names are
Milton Millercd790072007-10-17 16:55:11 +02006299 * static strings and all have proc handlers.
6300 */
6301 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02006302 if (entry->child)
6303 sd_free_ctl_entry(&entry->child);
Milton Millercd790072007-10-17 16:55:11 +02006304 if (entry->proc_handler == NULL)
6305 kfree(entry->procname);
6306 }
Milton Miller6382bc92007-10-15 17:00:19 +02006307
6308 kfree(*tablep);
6309 *tablep = NULL;
6310}
6311
Nick Piggine692ab52007-07-26 13:40:43 +02006312static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02006313set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02006314 const char *procname, void *data, int maxlen,
6315 mode_t mode, proc_handler *proc_handler)
6316{
Nick Piggine692ab52007-07-26 13:40:43 +02006317 entry->procname = procname;
6318 entry->data = data;
6319 entry->maxlen = maxlen;
6320 entry->mode = mode;
6321 entry->proc_handler = proc_handler;
6322}
6323
6324static struct ctl_table *
6325sd_alloc_ctl_domain_table(struct sched_domain *sd)
6326{
Zou Nan haiace8b3d2007-10-15 17:00:14 +02006327 struct ctl_table *table = sd_alloc_ctl_entry(12);
Nick Piggine692ab52007-07-26 13:40:43 +02006328
Milton Millerad1cdc12007-10-15 17:00:19 +02006329 if (table == NULL)
6330 return NULL;
6331
Alexey Dobriyane0361852007-08-09 11:16:46 +02006332 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02006333 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006334 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02006335 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006336 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006337 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006338 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006339 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006340 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006341 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006342 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006343 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006344 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006345 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006346 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02006347 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006348 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02006349 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02006350 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02006351 &sd->cache_nice_tries,
6352 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02006353 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02006354 sizeof(int), 0644, proc_dointvec_minmax);
Milton Miller6323469f2007-10-15 17:00:19 +02006355 /* &table[11] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02006356
6357 return table;
6358}
6359
Ingo Molnar9a4e7152007-11-28 15:52:56 +01006360static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02006361{
6362 struct ctl_table *entry, *table;
6363 struct sched_domain *sd;
6364 int domain_num = 0, i;
6365 char buf[32];
6366
6367 for_each_domain(cpu, sd)
6368 domain_num++;
6369 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02006370 if (table == NULL)
6371 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02006372
6373 i = 0;
6374 for_each_domain(cpu, sd) {
6375 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02006376 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006377 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02006378 entry->child = sd_alloc_ctl_domain_table(sd);
6379 entry++;
6380 i++;
6381 }
6382 return table;
6383}
6384
6385static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02006386static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02006387{
6388 int i, cpu_num = num_online_cpus();
6389 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6390 char buf[32];
6391
Milton Miller73785472007-10-24 18:23:48 +02006392 WARN_ON(sd_ctl_dir[0].child);
6393 sd_ctl_dir[0].child = entry;
6394
Milton Millerad1cdc12007-10-15 17:00:19 +02006395 if (entry == NULL)
6396 return;
6397
Milton Miller97b6ea72007-10-15 17:00:19 +02006398 for_each_online_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02006399 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02006400 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006401 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02006402 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02006403 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02006404 }
Milton Miller73785472007-10-24 18:23:48 +02006405
6406 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02006407 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6408}
Milton Miller6382bc92007-10-15 17:00:19 +02006409
Milton Miller73785472007-10-24 18:23:48 +02006410/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02006411static void unregister_sched_domain_sysctl(void)
6412{
Milton Miller73785472007-10-24 18:23:48 +02006413 if (sd_sysctl_header)
6414 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02006415 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02006416 if (sd_ctl_dir[0].child)
6417 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02006418}
Nick Piggine692ab52007-07-26 13:40:43 +02006419#else
Milton Miller6382bc92007-10-15 17:00:19 +02006420static void register_sched_domain_sysctl(void)
6421{
6422}
6423static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02006424{
6425}
6426#endif
6427
Linus Torvalds1da177e2005-04-16 15:20:36 -07006428/*
6429 * migration_call - callback that gets triggered when a CPU is added.
6430 * Here we can start up the necessary migration thread for the new CPU.
6431 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006432static int __cpuinit
6433migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006434{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006435 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006436 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006437 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006438 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006439
6440 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006441
Linus Torvalds1da177e2005-04-16 15:20:36 -07006442 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006443 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02006444 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006445 if (IS_ERR(p))
6446 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006447 kthread_bind(p, cpu);
6448 /* Must be high prio: stop_machine expects to yield to it. */
6449 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02006450 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006451 task_rq_unlock(rq, &flags);
6452 cpu_rq(cpu)->migration_thread = p;
6453 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006454
Linus Torvalds1da177e2005-04-16 15:20:36 -07006455 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006456 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02006457 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006458 wake_up_process(cpu_rq(cpu)->migration_thread);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006459
6460 /* Update our root-domain */
6461 rq = cpu_rq(cpu);
6462 spin_lock_irqsave(&rq->lock, flags);
6463 if (rq->rd) {
6464 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6465 cpu_set(cpu, rq->rd->online);
6466 }
6467 spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006468 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006469
Linus Torvalds1da177e2005-04-16 15:20:36 -07006470#ifdef CONFIG_HOTPLUG_CPU
6471 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006472 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07006473 if (!cpu_rq(cpu)->migration_thread)
6474 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006475 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08006476 kthread_bind(cpu_rq(cpu)->migration_thread,
6477 any_online_cpu(cpu_online_map));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006478 kthread_stop(cpu_rq(cpu)->migration_thread);
6479 cpu_rq(cpu)->migration_thread = NULL;
6480 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006481
Linus Torvalds1da177e2005-04-16 15:20:36 -07006482 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006483 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07006484 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006485 migrate_live_tasks(cpu);
6486 rq = cpu_rq(cpu);
6487 kthread_stop(rq->migration_thread);
6488 rq->migration_thread = NULL;
6489 /* Idle task back to normal (off runqueue, low prio) */
Oleg Nesterovd2da2722007-10-16 23:30:56 -07006490 spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02006491 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02006492 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006493 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02006494 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6495 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006496 migrate_dead_tasks(cpu);
Oleg Nesterovd2da2722007-10-16 23:30:56 -07006497 spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07006498 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006499 migrate_nr_uninterruptible(rq);
6500 BUG_ON(rq->nr_running != 0);
6501
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006502 /*
6503 * No need to migrate the tasks: it was best-effort if
6504 * they didn't take sched_hotcpu_mutex. Just wake up
6505 * the requestors.
6506 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006507 spin_lock_irq(&rq->lock);
6508 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07006509 struct migration_req *req;
6510
Linus Torvalds1da177e2005-04-16 15:20:36 -07006511 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07006512 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006513 list_del_init(&req->list);
6514 complete(&req->done);
6515 }
6516 spin_unlock_irq(&rq->lock);
6517 break;
Gregory Haskins57d885f2008-01-25 21:08:18 +01006518
Gregory Haskins08f503b2008-03-10 17:59:11 -04006519 case CPU_DYING:
6520 case CPU_DYING_FROZEN:
Gregory Haskins57d885f2008-01-25 21:08:18 +01006521 /* Update our root-domain */
6522 rq = cpu_rq(cpu);
6523 spin_lock_irqsave(&rq->lock, flags);
6524 if (rq->rd) {
6525 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6526 cpu_clear(cpu, rq->rd->online);
6527 }
6528 spin_unlock_irqrestore(&rq->lock, flags);
6529 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006530#endif
6531 }
6532 return NOTIFY_OK;
6533}
6534
6535/* Register at highest priority so that task migration (migrate_all_tasks)
6536 * happens before everything else.
6537 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07006538static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006539 .notifier_call = migration_call,
6540 .priority = 10
6541};
6542
Adrian Bunke6fe6642007-11-09 22:39:39 +01006543void __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006544{
6545 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07006546 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006547
6548 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07006549 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6550 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006551 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6552 register_cpu_notifier(&migration_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006553}
6554#endif
6555
6556#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07006557
Ingo Molnar3e9830d2007-10-15 17:00:13 +02006558#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006559
Mike Travis7c16ec52008-04-04 18:11:11 -07006560static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6561 cpumask_t *groupmask)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006562{
6563 struct sched_group *group = sd->groups;
Mike Travis434d53b2008-04-04 18:11:04 -07006564 char str[256];
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006565
Mike Travis434d53b2008-04-04 18:11:04 -07006566 cpulist_scnprintf(str, sizeof(str), sd->span);
Mike Travis7c16ec52008-04-04 18:11:11 -07006567 cpus_clear(*groupmask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006568
6569 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6570
6571 if (!(sd->flags & SD_LOAD_BALANCE)) {
6572 printk("does not load-balance\n");
6573 if (sd->parent)
6574 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6575 " has parent");
6576 return -1;
6577 }
6578
6579 printk(KERN_CONT "span %s\n", str);
6580
6581 if (!cpu_isset(cpu, sd->span)) {
6582 printk(KERN_ERR "ERROR: domain->span does not contain "
6583 "CPU%d\n", cpu);
6584 }
6585 if (!cpu_isset(cpu, group->cpumask)) {
6586 printk(KERN_ERR "ERROR: domain->groups does not contain"
6587 " CPU%d\n", cpu);
6588 }
6589
6590 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6591 do {
6592 if (!group) {
6593 printk("\n");
6594 printk(KERN_ERR "ERROR: group is NULL\n");
6595 break;
6596 }
6597
6598 if (!group->__cpu_power) {
6599 printk(KERN_CONT "\n");
6600 printk(KERN_ERR "ERROR: domain->cpu_power not "
6601 "set\n");
6602 break;
6603 }
6604
6605 if (!cpus_weight(group->cpumask)) {
6606 printk(KERN_CONT "\n");
6607 printk(KERN_ERR "ERROR: empty group\n");
6608 break;
6609 }
6610
Mike Travis7c16ec52008-04-04 18:11:11 -07006611 if (cpus_intersects(*groupmask, group->cpumask)) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006612 printk(KERN_CONT "\n");
6613 printk(KERN_ERR "ERROR: repeated CPUs\n");
6614 break;
6615 }
6616
Mike Travis7c16ec52008-04-04 18:11:11 -07006617 cpus_or(*groupmask, *groupmask, group->cpumask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006618
Mike Travis434d53b2008-04-04 18:11:04 -07006619 cpulist_scnprintf(str, sizeof(str), group->cpumask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006620 printk(KERN_CONT " %s", str);
6621
6622 group = group->next;
6623 } while (group != sd->groups);
6624 printk(KERN_CONT "\n");
6625
Mike Travis7c16ec52008-04-04 18:11:11 -07006626 if (!cpus_equal(sd->span, *groupmask))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006627 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6628
Mike Travis7c16ec52008-04-04 18:11:11 -07006629 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006630 printk(KERN_ERR "ERROR: parent span is not a superset "
6631 "of domain->span\n");
6632 return 0;
6633}
6634
Linus Torvalds1da177e2005-04-16 15:20:36 -07006635static void sched_domain_debug(struct sched_domain *sd, int cpu)
6636{
Mike Travis7c16ec52008-04-04 18:11:11 -07006637 cpumask_t *groupmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006638 int level = 0;
6639
Nick Piggin41c7ce92005-06-25 14:57:24 -07006640 if (!sd) {
6641 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6642 return;
6643 }
6644
Linus Torvalds1da177e2005-04-16 15:20:36 -07006645 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6646
Mike Travis7c16ec52008-04-04 18:11:11 -07006647 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6648 if (!groupmask) {
6649 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6650 return;
6651 }
6652
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006653 for (;;) {
Mike Travis7c16ec52008-04-04 18:11:11 -07006654 if (sched_domain_debug_one(sd, cpu, level, groupmask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006655 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006656 level++;
6657 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08006658 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006659 break;
6660 }
Mike Travis7c16ec52008-04-04 18:11:11 -07006661 kfree(groupmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006662}
6663#else
Ingo Molnar48f24c42006-07-03 00:25:40 -07006664# define sched_domain_debug(sd, cpu) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006665#endif
6666
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006667static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006668{
6669 if (cpus_weight(sd->span) == 1)
6670 return 1;
6671
6672 /* Following flags need at least 2 groups */
6673 if (sd->flags & (SD_LOAD_BALANCE |
6674 SD_BALANCE_NEWIDLE |
6675 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006676 SD_BALANCE_EXEC |
6677 SD_SHARE_CPUPOWER |
6678 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006679 if (sd->groups != sd->groups->next)
6680 return 0;
6681 }
6682
6683 /* Following flags don't use groups */
6684 if (sd->flags & (SD_WAKE_IDLE |
6685 SD_WAKE_AFFINE |
6686 SD_WAKE_BALANCE))
6687 return 0;
6688
6689 return 1;
6690}
6691
Ingo Molnar48f24c42006-07-03 00:25:40 -07006692static int
6693sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006694{
6695 unsigned long cflags = sd->flags, pflags = parent->flags;
6696
6697 if (sd_degenerate(parent))
6698 return 1;
6699
6700 if (!cpus_equal(sd->span, parent->span))
6701 return 0;
6702
6703 /* Does parent contain flags not in child? */
6704 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6705 if (cflags & SD_WAKE_AFFINE)
6706 pflags &= ~SD_WAKE_BALANCE;
6707 /* Flags needing groups don't count if only 1 group in parent */
6708 if (parent->groups == parent->groups->next) {
6709 pflags &= ~(SD_LOAD_BALANCE |
6710 SD_BALANCE_NEWIDLE |
6711 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006712 SD_BALANCE_EXEC |
6713 SD_SHARE_CPUPOWER |
6714 SD_SHARE_PKG_RESOURCES);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006715 }
6716 if (~cflags & pflags)
6717 return 0;
6718
6719 return 1;
6720}
6721
Gregory Haskins57d885f2008-01-25 21:08:18 +01006722static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6723{
6724 unsigned long flags;
6725 const struct sched_class *class;
6726
6727 spin_lock_irqsave(&rq->lock, flags);
6728
6729 if (rq->rd) {
6730 struct root_domain *old_rd = rq->rd;
6731
Ingo Molnar0eab9142008-01-25 21:08:19 +01006732 for (class = sched_class_highest; class; class = class->next) {
Gregory Haskins57d885f2008-01-25 21:08:18 +01006733 if (class->leave_domain)
6734 class->leave_domain(rq);
Ingo Molnar0eab9142008-01-25 21:08:19 +01006735 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01006736
Gregory Haskinsdc938522008-01-25 21:08:26 +01006737 cpu_clear(rq->cpu, old_rd->span);
6738 cpu_clear(rq->cpu, old_rd->online);
6739
Gregory Haskins57d885f2008-01-25 21:08:18 +01006740 if (atomic_dec_and_test(&old_rd->refcount))
6741 kfree(old_rd);
6742 }
6743
6744 atomic_inc(&rd->refcount);
6745 rq->rd = rd;
6746
Gregory Haskinsdc938522008-01-25 21:08:26 +01006747 cpu_set(rq->cpu, rd->span);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006748 if (cpu_isset(rq->cpu, cpu_online_map))
6749 cpu_set(rq->cpu, rd->online);
Gregory Haskinsdc938522008-01-25 21:08:26 +01006750
Ingo Molnar0eab9142008-01-25 21:08:19 +01006751 for (class = sched_class_highest; class; class = class->next) {
Gregory Haskins57d885f2008-01-25 21:08:18 +01006752 if (class->join_domain)
6753 class->join_domain(rq);
Ingo Molnar0eab9142008-01-25 21:08:19 +01006754 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01006755
6756 spin_unlock_irqrestore(&rq->lock, flags);
6757}
6758
Gregory Haskinsdc938522008-01-25 21:08:26 +01006759static void init_rootdomain(struct root_domain *rd)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006760{
6761 memset(rd, 0, sizeof(*rd));
6762
Gregory Haskinsdc938522008-01-25 21:08:26 +01006763 cpus_clear(rd->span);
6764 cpus_clear(rd->online);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006765}
6766
6767static void init_defrootdomain(void)
6768{
Gregory Haskinsdc938522008-01-25 21:08:26 +01006769 init_rootdomain(&def_root_domain);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006770 atomic_set(&def_root_domain.refcount, 1);
6771}
6772
Gregory Haskinsdc938522008-01-25 21:08:26 +01006773static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006774{
6775 struct root_domain *rd;
6776
6777 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6778 if (!rd)
6779 return NULL;
6780
Gregory Haskinsdc938522008-01-25 21:08:26 +01006781 init_rootdomain(rd);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006782
6783 return rd;
6784}
6785
Linus Torvalds1da177e2005-04-16 15:20:36 -07006786/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01006787 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07006788 * hold the hotplug lock.
6789 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01006790static void
6791cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006792{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006793 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006794 struct sched_domain *tmp;
6795
6796 /* Remove the sched domains which do not contribute to scheduling. */
6797 for (tmp = sd; tmp; tmp = tmp->parent) {
6798 struct sched_domain *parent = tmp->parent;
6799 if (!parent)
6800 break;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006801 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006802 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006803 if (parent->parent)
6804 parent->parent->child = tmp;
6805 }
Suresh Siddha245af2c2005-06-25 14:57:25 -07006806 }
6807
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006808 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006809 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006810 if (sd)
6811 sd->child = NULL;
6812 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006813
6814 sched_domain_debug(sd, cpu);
6815
Gregory Haskins57d885f2008-01-25 21:08:18 +01006816 rq_attach_root(rq, rd);
Nick Piggin674311d2005-06-25 14:57:27 -07006817 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006818}
6819
6820/* cpus with isolated domains */
Tim Chen67af63a2006-12-22 01:07:50 -08006821static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006822
6823/* Setup the mask of cpus configured for isolated domains */
6824static int __init isolated_cpu_setup(char *str)
6825{
6826 int ints[NR_CPUS], i;
6827
6828 str = get_options(str, ARRAY_SIZE(ints), ints);
6829 cpus_clear(cpu_isolated_map);
6830 for (i = 1; i <= ints[0]; i++)
6831 if (ints[i] < NR_CPUS)
6832 cpu_set(ints[i], cpu_isolated_map);
6833 return 1;
6834}
6835
Ingo Molnar8927f492007-10-15 17:00:13 +02006836__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006837
6838/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006839 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6840 * to a function which identifies what group(along with sched group) a CPU
6841 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6842 * (due to the fact that we keep track of groups covered with a cpumask_t).
Linus Torvalds1da177e2005-04-16 15:20:36 -07006843 *
6844 * init_sched_build_groups will build a circular linked list of the groups
6845 * covered by the given span, and will set each group's ->cpumask correctly,
6846 * and ->cpu_power to 0.
6847 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006848static void
Mike Travis7c16ec52008-04-04 18:11:11 -07006849init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006850 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07006851 struct sched_group **sg,
6852 cpumask_t *tmpmask),
6853 cpumask_t *covered, cpumask_t *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006854{
6855 struct sched_group *first = NULL, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006856 int i;
6857
Mike Travis7c16ec52008-04-04 18:11:11 -07006858 cpus_clear(*covered);
6859
6860 for_each_cpu_mask(i, *span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006861 struct sched_group *sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07006862 int group = group_fn(i, cpu_map, &sg, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006863 int j;
6864
Mike Travis7c16ec52008-04-04 18:11:11 -07006865 if (cpu_isset(i, *covered))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006866 continue;
6867
Mike Travis7c16ec52008-04-04 18:11:11 -07006868 cpus_clear(sg->cpumask);
Eric Dumazet5517d862007-05-08 00:32:57 -07006869 sg->__cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006870
Mike Travis7c16ec52008-04-04 18:11:11 -07006871 for_each_cpu_mask(j, *span) {
6872 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006873 continue;
6874
Mike Travis7c16ec52008-04-04 18:11:11 -07006875 cpu_set(j, *covered);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006876 cpu_set(j, sg->cpumask);
6877 }
6878 if (!first)
6879 first = sg;
6880 if (last)
6881 last->next = sg;
6882 last = sg;
6883 }
6884 last->next = first;
6885}
6886
John Hawkes9c1cfda2005-09-06 15:18:14 -07006887#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07006888
John Hawkes9c1cfda2005-09-06 15:18:14 -07006889#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08006890
John Hawkes9c1cfda2005-09-06 15:18:14 -07006891/**
6892 * find_next_best_node - find the next node to include in a sched_domain
6893 * @node: node whose sched_domain we're building
6894 * @used_nodes: nodes already in the sched_domain
6895 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006896 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07006897 * finds the closest node not already in the @used_nodes map.
6898 *
6899 * Should use nodemask_t.
6900 */
Mike Travisc5f59f02008-04-04 18:11:10 -07006901static int find_next_best_node(int node, nodemask_t *used_nodes)
John Hawkes9c1cfda2005-09-06 15:18:14 -07006902{
6903 int i, n, val, min_val, best_node = 0;
6904
6905 min_val = INT_MAX;
6906
6907 for (i = 0; i < MAX_NUMNODES; i++) {
6908 /* Start at @node */
6909 n = (node + i) % MAX_NUMNODES;
6910
6911 if (!nr_cpus_node(n))
6912 continue;
6913
6914 /* Skip already used nodes */
Mike Travisc5f59f02008-04-04 18:11:10 -07006915 if (node_isset(n, *used_nodes))
John Hawkes9c1cfda2005-09-06 15:18:14 -07006916 continue;
6917
6918 /* Simple min distance search */
6919 val = node_distance(node, n);
6920
6921 if (val < min_val) {
6922 min_val = val;
6923 best_node = n;
6924 }
6925 }
6926
Mike Travisc5f59f02008-04-04 18:11:10 -07006927 node_set(best_node, *used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006928 return best_node;
6929}
6930
6931/**
6932 * sched_domain_node_span - get a cpumask for a node's sched_domain
6933 * @node: node whose cpumask we're constructing
Randy Dunlap73486722008-04-22 10:07:22 -07006934 * @span: resulting cpumask
John Hawkes9c1cfda2005-09-06 15:18:14 -07006935 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006936 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07006937 * should be one that prevents unnecessary balancing, but also spreads tasks
6938 * out optimally.
6939 */
Mike Travis4bdbaad32008-04-15 16:35:52 -07006940static void sched_domain_node_span(int node, cpumask_t *span)
John Hawkes9c1cfda2005-09-06 15:18:14 -07006941{
Mike Travisc5f59f02008-04-04 18:11:10 -07006942 nodemask_t used_nodes;
Mike Travisc5f59f02008-04-04 18:11:10 -07006943 node_to_cpumask_ptr(nodemask, node);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006944 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006945
Mike Travis4bdbaad32008-04-15 16:35:52 -07006946 cpus_clear(*span);
Mike Travisc5f59f02008-04-04 18:11:10 -07006947 nodes_clear(used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006948
Mike Travis4bdbaad32008-04-15 16:35:52 -07006949 cpus_or(*span, *span, *nodemask);
Mike Travisc5f59f02008-04-04 18:11:10 -07006950 node_set(node, used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006951
6952 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
Mike Travisc5f59f02008-04-04 18:11:10 -07006953 int next_node = find_next_best_node(node, &used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006954
Mike Travisc5f59f02008-04-04 18:11:10 -07006955 node_to_cpumask_ptr_next(nodemask, next_node);
Mike Travis4bdbaad32008-04-15 16:35:52 -07006956 cpus_or(*span, *span, *nodemask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006957 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006958}
6959#endif
6960
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006961int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006962
John Hawkes9c1cfda2005-09-06 15:18:14 -07006963/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07006964 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07006965 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006966#ifdef CONFIG_SCHED_SMT
6967static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006968static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006969
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006970static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006971cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6972 cpumask_t *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006973{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006974 if (sg)
6975 *sg = &per_cpu(sched_group_cpus, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006976 return cpu;
6977}
6978#endif
6979
Ingo Molnar48f24c42006-07-03 00:25:40 -07006980/*
6981 * multi-core sched-domains:
6982 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006983#ifdef CONFIG_SCHED_MC
6984static DEFINE_PER_CPU(struct sched_domain, core_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006985static DEFINE_PER_CPU(struct sched_group, sched_group_core);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006986#endif
6987
6988#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006989static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006990cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6991 cpumask_t *mask)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006992{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006993 int group;
Mike Travis7c16ec52008-04-04 18:11:11 -07006994
6995 *mask = per_cpu(cpu_sibling_map, cpu);
6996 cpus_and(*mask, *mask, *cpu_map);
6997 group = first_cpu(*mask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006998 if (sg)
6999 *sg = &per_cpu(sched_group_core, group);
7000 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007001}
7002#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007003static int
Mike Travis7c16ec52008-04-04 18:11:11 -07007004cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7005 cpumask_t *unused)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007006{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007007 if (sg)
7008 *sg = &per_cpu(sched_group_core, cpu);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007009 return cpu;
7010}
7011#endif
7012
Linus Torvalds1da177e2005-04-16 15:20:36 -07007013static DEFINE_PER_CPU(struct sched_domain, phys_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007014static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007015
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007016static int
Mike Travis7c16ec52008-04-04 18:11:11 -07007017cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7018 cpumask_t *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007019{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007020 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007021#ifdef CONFIG_SCHED_MC
Mike Travis7c16ec52008-04-04 18:11:11 -07007022 *mask = cpu_coregroup_map(cpu);
7023 cpus_and(*mask, *mask, *cpu_map);
7024 group = first_cpu(*mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007025#elif defined(CONFIG_SCHED_SMT)
Mike Travis7c16ec52008-04-04 18:11:11 -07007026 *mask = per_cpu(cpu_sibling_map, cpu);
7027 cpus_and(*mask, *mask, *cpu_map);
7028 group = first_cpu(*mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007029#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007030 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007031#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007032 if (sg)
7033 *sg = &per_cpu(sched_group_phys, group);
7034 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007035}
7036
7037#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07007038/*
7039 * The init_sched_build_groups can't handle what we want to do with node
7040 * groups, so roll our own. Now each node has its own list of groups which
7041 * gets dynamically allocated.
7042 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007043static DEFINE_PER_CPU(struct sched_domain, node_domains);
Mike Travis434d53b2008-04-04 18:11:04 -07007044static struct sched_group ***sched_group_nodes_bycpu;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007045
7046static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007047static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007048
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007049static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007050 struct sched_group **sg, cpumask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007051{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007052 int group;
7053
Mike Travis7c16ec52008-04-04 18:11:11 -07007054 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7055 cpus_and(*nodemask, *nodemask, *cpu_map);
7056 group = first_cpu(*nodemask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007057
7058 if (sg)
7059 *sg = &per_cpu(sched_group_allnodes, group);
7060 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007061}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007062
Siddha, Suresh B08069032006-03-27 01:15:23 -08007063static void init_numa_sched_groups_power(struct sched_group *group_head)
7064{
7065 struct sched_group *sg = group_head;
7066 int j;
7067
7068 if (!sg)
7069 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02007070 do {
7071 for_each_cpu_mask(j, sg->cpumask) {
7072 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08007073
Andi Kleen3a5c3592007-10-15 17:00:14 +02007074 sd = &per_cpu(phys_domains, j);
7075 if (j != first_cpu(sd->groups->cpumask)) {
7076 /*
7077 * Only add "power" once for each
7078 * physical package.
7079 */
7080 continue;
7081 }
7082
7083 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
Siddha, Suresh B08069032006-03-27 01:15:23 -08007084 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02007085 sg = sg->next;
7086 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08007087}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007088#endif
7089
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007090#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007091/* Free memory allocated for various sched_group structures */
Mike Travis7c16ec52008-04-04 18:11:11 -07007092static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007093{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007094 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007095
7096 for_each_cpu_mask(cpu, *cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007097 struct sched_group **sched_group_nodes
7098 = sched_group_nodes_bycpu[cpu];
7099
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007100 if (!sched_group_nodes)
7101 continue;
7102
7103 for (i = 0; i < MAX_NUMNODES; i++) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007104 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7105
Mike Travis7c16ec52008-04-04 18:11:11 -07007106 *nodemask = node_to_cpumask(i);
7107 cpus_and(*nodemask, *nodemask, *cpu_map);
7108 if (cpus_empty(*nodemask))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007109 continue;
7110
7111 if (sg == NULL)
7112 continue;
7113 sg = sg->next;
7114next_sg:
7115 oldsg = sg;
7116 sg = sg->next;
7117 kfree(oldsg);
7118 if (oldsg != sched_group_nodes[i])
7119 goto next_sg;
7120 }
7121 kfree(sched_group_nodes);
7122 sched_group_nodes_bycpu[cpu] = NULL;
7123 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007124}
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007125#else
Mike Travis7c16ec52008-04-04 18:11:11 -07007126static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007127{
7128}
7129#endif
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007130
Linus Torvalds1da177e2005-04-16 15:20:36 -07007131/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007132 * Initialize sched groups cpu_power.
7133 *
7134 * cpu_power indicates the capacity of sched group, which is used while
7135 * distributing the load between different sched groups in a sched domain.
7136 * Typically cpu_power for all the groups in a sched domain will be same unless
7137 * there are asymmetries in the topology. If there are asymmetries, group
7138 * having more cpu_power will pickup more load compared to the group having
7139 * less cpu_power.
7140 *
7141 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7142 * the maximum number of tasks a group can handle in the presence of other idle
7143 * or lightly loaded groups in the same sched domain.
7144 */
7145static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7146{
7147 struct sched_domain *child;
7148 struct sched_group *group;
7149
7150 WARN_ON(!sd || !sd->groups);
7151
7152 if (cpu != first_cpu(sd->groups->cpumask))
7153 return;
7154
7155 child = sd->child;
7156
Eric Dumazet5517d862007-05-08 00:32:57 -07007157 sd->groups->__cpu_power = 0;
7158
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007159 /*
7160 * For perf policy, if the groups in child domain share resources
7161 * (for example cores sharing some portions of the cache hierarchy
7162 * or SMT), then set this domain groups cpu_power such that each group
7163 * can handle only one task, when there are other idle groups in the
7164 * same sched domain.
7165 */
7166 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7167 (child->flags &
7168 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
Eric Dumazet5517d862007-05-08 00:32:57 -07007169 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007170 return;
7171 }
7172
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007173 /*
7174 * add cpu_power of each child group to this groups cpu_power
7175 */
7176 group = child->groups;
7177 do {
Eric Dumazet5517d862007-05-08 00:32:57 -07007178 sg_inc_cpu_power(sd->groups, group->__cpu_power);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007179 group = group->next;
7180 } while (group != child->groups);
7181}
7182
7183/*
Mike Travis7c16ec52008-04-04 18:11:11 -07007184 * Initializers for schedule domains
7185 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7186 */
7187
7188#define SD_INIT(sd, type) sd_init_##type(sd)
7189#define SD_INIT_FUNC(type) \
7190static noinline void sd_init_##type(struct sched_domain *sd) \
7191{ \
7192 memset(sd, 0, sizeof(*sd)); \
7193 *sd = SD_##type##_INIT; \
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007194 sd->level = SD_LV_##type; \
Mike Travis7c16ec52008-04-04 18:11:11 -07007195}
7196
7197SD_INIT_FUNC(CPU)
7198#ifdef CONFIG_NUMA
7199 SD_INIT_FUNC(ALLNODES)
7200 SD_INIT_FUNC(NODE)
7201#endif
7202#ifdef CONFIG_SCHED_SMT
7203 SD_INIT_FUNC(SIBLING)
7204#endif
7205#ifdef CONFIG_SCHED_MC
7206 SD_INIT_FUNC(MC)
7207#endif
7208
7209/*
7210 * To minimize stack usage kmalloc room for cpumasks and share the
7211 * space as the usage in build_sched_domains() dictates. Used only
7212 * if the amount of space is significant.
7213 */
7214struct allmasks {
7215 cpumask_t tmpmask; /* make this one first */
7216 union {
7217 cpumask_t nodemask;
7218 cpumask_t this_sibling_map;
7219 cpumask_t this_core_map;
7220 };
7221 cpumask_t send_covered;
7222
7223#ifdef CONFIG_NUMA
7224 cpumask_t domainspan;
7225 cpumask_t covered;
7226 cpumask_t notcovered;
7227#endif
7228};
7229
7230#if NR_CPUS > 128
7231#define SCHED_CPUMASK_ALLOC 1
7232#define SCHED_CPUMASK_FREE(v) kfree(v)
7233#define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7234#else
7235#define SCHED_CPUMASK_ALLOC 0
7236#define SCHED_CPUMASK_FREE(v)
7237#define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7238#endif
7239
7240#define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7241 ((unsigned long)(a) + offsetof(struct allmasks, v))
7242
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007243static int default_relax_domain_level = -1;
7244
7245static int __init setup_relax_domain_level(char *str)
7246{
7247 default_relax_domain_level = simple_strtoul(str, NULL, 0);
7248 return 1;
7249}
7250__setup("relax_domain_level=", setup_relax_domain_level);
7251
7252static void set_domain_attribute(struct sched_domain *sd,
7253 struct sched_domain_attr *attr)
7254{
7255 int request;
7256
7257 if (!attr || attr->relax_domain_level < 0) {
7258 if (default_relax_domain_level < 0)
7259 return;
7260 else
7261 request = default_relax_domain_level;
7262 } else
7263 request = attr->relax_domain_level;
7264 if (request < sd->level) {
7265 /* turn off idle balance on this domain */
7266 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7267 } else {
7268 /* turn on idle balance on this domain */
7269 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7270 }
7271}
7272
Mike Travis7c16ec52008-04-04 18:11:11 -07007273/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007274 * Build sched domains for a given set of cpus and attach the sched domains
7275 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07007276 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007277static int __build_sched_domains(const cpumask_t *cpu_map,
7278 struct sched_domain_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007279{
7280 int i;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007281 struct root_domain *rd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007282 SCHED_CPUMASK_DECLARE(allmasks);
7283 cpumask_t *tmpmask;
John Hawkesd1b55132005-09-06 15:18:14 -07007284#ifdef CONFIG_NUMA
7285 struct sched_group **sched_group_nodes = NULL;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007286 int sd_allnodes = 0;
John Hawkesd1b55132005-09-06 15:18:14 -07007287
7288 /*
7289 * Allocate the per-node list of sched groups
7290 */
Milton Miller5cf9f062007-10-15 17:00:19 +02007291 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007292 GFP_KERNEL);
John Hawkesd1b55132005-09-06 15:18:14 -07007293 if (!sched_group_nodes) {
7294 printk(KERN_WARNING "Can not alloc sched group node list\n");
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007295 return -ENOMEM;
John Hawkesd1b55132005-09-06 15:18:14 -07007296 }
John Hawkesd1b55132005-09-06 15:18:14 -07007297#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007298
Gregory Haskinsdc938522008-01-25 21:08:26 +01007299 rd = alloc_rootdomain();
Gregory Haskins57d885f2008-01-25 21:08:18 +01007300 if (!rd) {
7301 printk(KERN_WARNING "Cannot alloc root domain\n");
Mike Travis7c16ec52008-04-04 18:11:11 -07007302#ifdef CONFIG_NUMA
7303 kfree(sched_group_nodes);
7304#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01007305 return -ENOMEM;
7306 }
7307
Mike Travis7c16ec52008-04-04 18:11:11 -07007308#if SCHED_CPUMASK_ALLOC
7309 /* get space for all scratch cpumask variables */
7310 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7311 if (!allmasks) {
7312 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7313 kfree(rd);
7314#ifdef CONFIG_NUMA
7315 kfree(sched_group_nodes);
7316#endif
7317 return -ENOMEM;
7318 }
7319#endif
7320 tmpmask = (cpumask_t *)allmasks;
7321
7322
7323#ifdef CONFIG_NUMA
7324 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7325#endif
7326
Linus Torvalds1da177e2005-04-16 15:20:36 -07007327 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007328 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007329 */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007330 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007331 struct sched_domain *sd = NULL, *p;
Mike Travis7c16ec52008-04-04 18:11:11 -07007332 SCHED_CPUMASK_VAR(nodemask, allmasks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007333
Mike Travis7c16ec52008-04-04 18:11:11 -07007334 *nodemask = node_to_cpumask(cpu_to_node(i));
7335 cpus_and(*nodemask, *nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007336
7337#ifdef CONFIG_NUMA
Ingo Molnardd41f592007-07-09 18:51:59 +02007338 if (cpus_weight(*cpu_map) >
Mike Travis7c16ec52008-04-04 18:11:11 -07007339 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07007340 sd = &per_cpu(allnodes_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007341 SD_INIT(sd, ALLNODES);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007342 set_domain_attribute(sd, attr);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007343 sd->span = *cpu_map;
Peter Zijlstra18d95a22008-04-19 19:45:00 +02007344 sd->first_cpu = first_cpu(sd->span);
Mike Travis7c16ec52008-04-04 18:11:11 -07007345 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007346 p = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007347 sd_allnodes = 1;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007348 } else
7349 p = NULL;
7350
Linus Torvalds1da177e2005-04-16 15:20:36 -07007351 sd = &per_cpu(node_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007352 SD_INIT(sd, NODE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007353 set_domain_attribute(sd, attr);
Mike Travis4bdbaad32008-04-15 16:35:52 -07007354 sched_domain_node_span(cpu_to_node(i), &sd->span);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02007355 sd->first_cpu = first_cpu(sd->span);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007356 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007357 if (p)
7358 p->child = sd;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007359 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007360#endif
7361
7362 p = sd;
7363 sd = &per_cpu(phys_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007364 SD_INIT(sd, CPU);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007365 set_domain_attribute(sd, attr);
Mike Travis7c16ec52008-04-04 18:11:11 -07007366 sd->span = *nodemask;
Peter Zijlstra18d95a22008-04-19 19:45:00 +02007367 sd->first_cpu = first_cpu(sd->span);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007368 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007369 if (p)
7370 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007371 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007372
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007373#ifdef CONFIG_SCHED_MC
7374 p = sd;
7375 sd = &per_cpu(core_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007376 SD_INIT(sd, MC);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007377 set_domain_attribute(sd, attr);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007378 sd->span = cpu_coregroup_map(i);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02007379 sd->first_cpu = first_cpu(sd->span);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007380 cpus_and(sd->span, sd->span, *cpu_map);
7381 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007382 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007383 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007384#endif
7385
Linus Torvalds1da177e2005-04-16 15:20:36 -07007386#ifdef CONFIG_SCHED_SMT
7387 p = sd;
7388 sd = &per_cpu(cpu_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007389 SD_INIT(sd, SIBLING);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007390 set_domain_attribute(sd, attr);
Mike Travisd5a74302007-10-16 01:24:05 -07007391 sd->span = per_cpu(cpu_sibling_map, i);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02007392 sd->first_cpu = first_cpu(sd->span);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007393 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007394 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007395 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007396 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007397#endif
7398 }
7399
7400#ifdef CONFIG_SCHED_SMT
7401 /* Set up CPU (sibling) groups */
John Hawkes9c1cfda2005-09-06 15:18:14 -07007402 for_each_cpu_mask(i, *cpu_map) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007403 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7404 SCHED_CPUMASK_VAR(send_covered, allmasks);
7405
7406 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7407 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7408 if (i != first_cpu(*this_sibling_map))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007409 continue;
7410
Ingo Molnardd41f592007-07-09 18:51:59 +02007411 init_sched_build_groups(this_sibling_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007412 &cpu_to_cpu_group,
7413 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007414 }
7415#endif
7416
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007417#ifdef CONFIG_SCHED_MC
7418 /* Set up multi-core groups */
7419 for_each_cpu_mask(i, *cpu_map) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007420 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7421 SCHED_CPUMASK_VAR(send_covered, allmasks);
7422
7423 *this_core_map = cpu_coregroup_map(i);
7424 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7425 if (i != first_cpu(*this_core_map))
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007426 continue;
Mike Travis7c16ec52008-04-04 18:11:11 -07007427
Ingo Molnardd41f592007-07-09 18:51:59 +02007428 init_sched_build_groups(this_core_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007429 &cpu_to_core_group,
7430 send_covered, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007431 }
7432#endif
7433
Linus Torvalds1da177e2005-04-16 15:20:36 -07007434 /* Set up physical groups */
7435 for (i = 0; i < MAX_NUMNODES; i++) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007436 SCHED_CPUMASK_VAR(nodemask, allmasks);
7437 SCHED_CPUMASK_VAR(send_covered, allmasks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007438
Mike Travis7c16ec52008-04-04 18:11:11 -07007439 *nodemask = node_to_cpumask(i);
7440 cpus_and(*nodemask, *nodemask, *cpu_map);
7441 if (cpus_empty(*nodemask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007442 continue;
7443
Mike Travis7c16ec52008-04-04 18:11:11 -07007444 init_sched_build_groups(nodemask, cpu_map,
7445 &cpu_to_phys_group,
7446 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007447 }
7448
7449#ifdef CONFIG_NUMA
7450 /* Set up node groups */
Mike Travis7c16ec52008-04-04 18:11:11 -07007451 if (sd_allnodes) {
7452 SCHED_CPUMASK_VAR(send_covered, allmasks);
7453
7454 init_sched_build_groups(cpu_map, cpu_map,
7455 &cpu_to_allnodes_group,
7456 send_covered, tmpmask);
7457 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007458
7459 for (i = 0; i < MAX_NUMNODES; i++) {
7460 /* Set up node groups */
7461 struct sched_group *sg, *prev;
Mike Travis7c16ec52008-04-04 18:11:11 -07007462 SCHED_CPUMASK_VAR(nodemask, allmasks);
7463 SCHED_CPUMASK_VAR(domainspan, allmasks);
7464 SCHED_CPUMASK_VAR(covered, allmasks);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007465 int j;
7466
Mike Travis7c16ec52008-04-04 18:11:11 -07007467 *nodemask = node_to_cpumask(i);
7468 cpus_clear(*covered);
7469
7470 cpus_and(*nodemask, *nodemask, *cpu_map);
7471 if (cpus_empty(*nodemask)) {
John Hawkesd1b55132005-09-06 15:18:14 -07007472 sched_group_nodes[i] = NULL;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007473 continue;
John Hawkesd1b55132005-09-06 15:18:14 -07007474 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007475
Mike Travis4bdbaad32008-04-15 16:35:52 -07007476 sched_domain_node_span(i, domainspan);
Mike Travis7c16ec52008-04-04 18:11:11 -07007477 cpus_and(*domainspan, *domainspan, *cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007478
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07007479 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007480 if (!sg) {
7481 printk(KERN_WARNING "Can not alloc domain group for "
7482 "node %d\n", i);
7483 goto error;
7484 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007485 sched_group_nodes[i] = sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07007486 for_each_cpu_mask(j, *nodemask) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07007487 struct sched_domain *sd;
Ingo Molnar9761eea2007-07-09 18:52:00 +02007488
John Hawkes9c1cfda2005-09-06 15:18:14 -07007489 sd = &per_cpu(node_domains, j);
7490 sd->groups = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007491 }
Eric Dumazet5517d862007-05-08 00:32:57 -07007492 sg->__cpu_power = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07007493 sg->cpumask = *nodemask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007494 sg->next = sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07007495 cpus_or(*covered, *covered, *nodemask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007496 prev = sg;
7497
7498 for (j = 0; j < MAX_NUMNODES; j++) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007499 SCHED_CPUMASK_VAR(notcovered, allmasks);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007500 int n = (i + j) % MAX_NUMNODES;
Mike Travisc5f59f02008-04-04 18:11:10 -07007501 node_to_cpumask_ptr(pnodemask, n);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007502
Mike Travis7c16ec52008-04-04 18:11:11 -07007503 cpus_complement(*notcovered, *covered);
7504 cpus_and(*tmpmask, *notcovered, *cpu_map);
7505 cpus_and(*tmpmask, *tmpmask, *domainspan);
7506 if (cpus_empty(*tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07007507 break;
7508
Mike Travis7c16ec52008-04-04 18:11:11 -07007509 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7510 if (cpus_empty(*tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07007511 continue;
7512
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07007513 sg = kmalloc_node(sizeof(struct sched_group),
7514 GFP_KERNEL, i);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007515 if (!sg) {
7516 printk(KERN_WARNING
7517 "Can not alloc domain group for node %d\n", j);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007518 goto error;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007519 }
Eric Dumazet5517d862007-05-08 00:32:57 -07007520 sg->__cpu_power = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07007521 sg->cpumask = *tmpmask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007522 sg->next = prev->next;
Mike Travis7c16ec52008-04-04 18:11:11 -07007523 cpus_or(*covered, *covered, *tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007524 prev->next = sg;
7525 prev = sg;
7526 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007527 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007528#endif
7529
7530 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007531#ifdef CONFIG_SCHED_SMT
7532 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007533 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7534
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007535 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007536 }
7537#endif
7538#ifdef CONFIG_SCHED_MC
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007539 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007540 struct sched_domain *sd = &per_cpu(core_domains, i);
7541
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007542 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007543 }
7544#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007545
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007546 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007547 struct sched_domain *sd = &per_cpu(phys_domains, i);
7548
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007549 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007550 }
7551
John Hawkes9c1cfda2005-09-06 15:18:14 -07007552#ifdef CONFIG_NUMA
Siddha, Suresh B08069032006-03-27 01:15:23 -08007553 for (i = 0; i < MAX_NUMNODES; i++)
7554 init_numa_sched_groups_power(sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007555
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007556 if (sd_allnodes) {
7557 struct sched_group *sg;
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07007558
Mike Travis7c16ec52008-04-04 18:11:11 -07007559 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7560 tmpmask);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07007561 init_numa_sched_groups_power(sg);
7562 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007563#endif
7564
Linus Torvalds1da177e2005-04-16 15:20:36 -07007565 /* Attach the domains */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007566 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007567 struct sched_domain *sd;
7568#ifdef CONFIG_SCHED_SMT
7569 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007570#elif defined(CONFIG_SCHED_MC)
7571 sd = &per_cpu(core_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007572#else
7573 sd = &per_cpu(phys_domains, i);
7574#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01007575 cpu_attach_domain(sd, rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007576 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007577
Mike Travis7c16ec52008-04-04 18:11:11 -07007578 SCHED_CPUMASK_FREE((void *)allmasks);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007579 return 0;
7580
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007581#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007582error:
Mike Travis7c16ec52008-04-04 18:11:11 -07007583 free_sched_groups(cpu_map, tmpmask);
7584 SCHED_CPUMASK_FREE((void *)allmasks);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007585 return -ENOMEM;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007586#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007587}
Paul Jackson029190c2007-10-18 23:40:20 -07007588
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007589static int build_sched_domains(const cpumask_t *cpu_map)
7590{
7591 return __build_sched_domains(cpu_map, NULL);
7592}
7593
Paul Jackson029190c2007-10-18 23:40:20 -07007594static cpumask_t *doms_cur; /* current sched domains */
7595static int ndoms_cur; /* number of sched domains in 'doms_cur' */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007596static struct sched_domain_attr *dattr_cur; /* attribues of custom domains
7597 in 'doms_cur' */
Paul Jackson029190c2007-10-18 23:40:20 -07007598
7599/*
7600 * Special case: If a kmalloc of a doms_cur partition (array of
7601 * cpumask_t) fails, then fallback to a single sched domain,
7602 * as determined by the single cpumask_t fallback_doms.
7603 */
7604static cpumask_t fallback_doms;
7605
Heiko Carstens22e52b02008-03-12 18:31:59 +01007606void __attribute__((weak)) arch_update_cpu_topology(void)
7607{
7608}
7609
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007610/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007611 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07007612 * For now this just excludes isolated cpus, but could be used to
7613 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007614 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007615static int arch_init_sched_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007616{
Milton Miller73785472007-10-24 18:23:48 +02007617 int err;
7618
Heiko Carstens22e52b02008-03-12 18:31:59 +01007619 arch_update_cpu_topology();
Paul Jackson029190c2007-10-18 23:40:20 -07007620 ndoms_cur = 1;
7621 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7622 if (!doms_cur)
7623 doms_cur = &fallback_doms;
7624 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007625 dattr_cur = NULL;
Milton Miller73785472007-10-24 18:23:48 +02007626 err = build_sched_domains(doms_cur);
Milton Miller6382bc92007-10-15 17:00:19 +02007627 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02007628
7629 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007630}
7631
Mike Travis7c16ec52008-04-04 18:11:11 -07007632static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7633 cpumask_t *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007634{
Mike Travis7c16ec52008-04-04 18:11:11 -07007635 free_sched_groups(cpu_map, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007636}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007637
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007638/*
7639 * Detach sched domains from a group of cpus specified in cpu_map
7640 * These cpus will now be attached to the NULL domain
7641 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08007642static void detach_destroy_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007643{
Mike Travis7c16ec52008-04-04 18:11:11 -07007644 cpumask_t tmpmask;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007645 int i;
7646
Milton Miller6382bc92007-10-15 17:00:19 +02007647 unregister_sched_domain_sysctl();
7648
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007649 for_each_cpu_mask(i, *cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01007650 cpu_attach_domain(NULL, &def_root_domain, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007651 synchronize_sched();
Mike Travis7c16ec52008-04-04 18:11:11 -07007652 arch_destroy_sched_domains(cpu_map, &tmpmask);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007653}
7654
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007655/* handle null as "default" */
7656static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7657 struct sched_domain_attr *new, int idx_new)
7658{
7659 struct sched_domain_attr tmp;
7660
7661 /* fast path */
7662 if (!new && !cur)
7663 return 1;
7664
7665 tmp = SD_ATTR_INIT;
7666 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7667 new ? (new + idx_new) : &tmp,
7668 sizeof(struct sched_domain_attr));
7669}
7670
Paul Jackson029190c2007-10-18 23:40:20 -07007671/*
7672 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007673 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07007674 * doms_new[] to the current sched domain partitioning, doms_cur[].
7675 * It destroys each deleted domain and builds each new domain.
7676 *
7677 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007678 * The masks don't intersect (don't overlap.) We should setup one
7679 * sched domain for each mask. CPUs not in any of the cpumasks will
7680 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07007681 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7682 * it as it is.
7683 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007684 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7685 * ownership of it and will kfree it when done with it. If the caller
Paul Jackson029190c2007-10-18 23:40:20 -07007686 * failed the kmalloc call, then it can pass in doms_new == NULL,
7687 * and partition_sched_domains() will fallback to the single partition
7688 * 'fallback_doms'.
7689 *
7690 * Call with hotplug lock held
7691 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007692void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7693 struct sched_domain_attr *dattr_new)
Paul Jackson029190c2007-10-18 23:40:20 -07007694{
7695 int i, j;
7696
Heiko Carstens712555e2008-04-28 11:33:07 +02007697 mutex_lock(&sched_domains_mutex);
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007698
Milton Miller73785472007-10-24 18:23:48 +02007699 /* always unregister in case we don't destroy any domains */
7700 unregister_sched_domain_sysctl();
7701
Paul Jackson029190c2007-10-18 23:40:20 -07007702 if (doms_new == NULL) {
7703 ndoms_new = 1;
7704 doms_new = &fallback_doms;
7705 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007706 dattr_new = NULL;
Paul Jackson029190c2007-10-18 23:40:20 -07007707 }
7708
7709 /* Destroy deleted domains */
7710 for (i = 0; i < ndoms_cur; i++) {
7711 for (j = 0; j < ndoms_new; j++) {
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007712 if (cpus_equal(doms_cur[i], doms_new[j])
7713 && dattrs_equal(dattr_cur, i, dattr_new, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007714 goto match1;
7715 }
7716 /* no match - a current sched domain not in new doms_new[] */
7717 detach_destroy_domains(doms_cur + i);
7718match1:
7719 ;
7720 }
7721
7722 /* Build new domains */
7723 for (i = 0; i < ndoms_new; i++) {
7724 for (j = 0; j < ndoms_cur; j++) {
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007725 if (cpus_equal(doms_new[i], doms_cur[j])
7726 && dattrs_equal(dattr_new, i, dattr_cur, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007727 goto match2;
7728 }
7729 /* no match - add a new doms_new */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007730 __build_sched_domains(doms_new + i,
7731 dattr_new ? dattr_new + i : NULL);
Paul Jackson029190c2007-10-18 23:40:20 -07007732match2:
7733 ;
7734 }
7735
7736 /* Remember the new sched domains */
7737 if (doms_cur != &fallback_doms)
7738 kfree(doms_cur);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007739 kfree(dattr_cur); /* kfree(NULL) is safe */
Paul Jackson029190c2007-10-18 23:40:20 -07007740 doms_cur = doms_new;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007741 dattr_cur = dattr_new;
Paul Jackson029190c2007-10-18 23:40:20 -07007742 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02007743
7744 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007745
Heiko Carstens712555e2008-04-28 11:33:07 +02007746 mutex_unlock(&sched_domains_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -07007747}
7748
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007749#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Heiko Carstens9aefd0a2008-03-12 18:31:58 +01007750int arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007751{
7752 int err;
7753
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007754 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02007755 mutex_lock(&sched_domains_mutex);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007756 detach_destroy_domains(&cpu_online_map);
7757 err = arch_init_sched_domains(&cpu_online_map);
Heiko Carstens712555e2008-04-28 11:33:07 +02007758 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007759 put_online_cpus();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007760
7761 return err;
7762}
7763
7764static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7765{
7766 int ret;
7767
7768 if (buf[0] != '0' && buf[0] != '1')
7769 return -EINVAL;
7770
7771 if (smt)
7772 sched_smt_power_savings = (buf[0] == '1');
7773 else
7774 sched_mc_power_savings = (buf[0] == '1');
7775
7776 ret = arch_reinit_sched_domains();
7777
7778 return ret ? ret : count;
7779}
7780
Adrian Bunk6707de002007-08-12 18:08:19 +02007781#ifdef CONFIG_SCHED_MC
7782static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7783{
7784 return sprintf(page, "%u\n", sched_mc_power_savings);
7785}
7786static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7787 const char *buf, size_t count)
7788{
7789 return sched_power_savings_store(buf, count, 0);
7790}
7791static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7792 sched_mc_power_savings_store);
7793#endif
7794
7795#ifdef CONFIG_SCHED_SMT
7796static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7797{
7798 return sprintf(page, "%u\n", sched_smt_power_savings);
7799}
7800static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7801 const char *buf, size_t count)
7802{
7803 return sched_power_savings_store(buf, count, 1);
7804}
7805static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7806 sched_smt_power_savings_store);
7807#endif
7808
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007809int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7810{
7811 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007812
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007813#ifdef CONFIG_SCHED_SMT
7814 if (smt_capable())
7815 err = sysfs_create_file(&cls->kset.kobj,
7816 &attr_sched_smt_power_savings.attr);
7817#endif
7818#ifdef CONFIG_SCHED_MC
7819 if (!err && mc_capable())
7820 err = sysfs_create_file(&cls->kset.kobj,
7821 &attr_sched_mc_power_savings.attr);
7822#endif
7823 return err;
7824}
7825#endif
7826
Linus Torvalds1da177e2005-04-16 15:20:36 -07007827/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007828 * Force a reinitialization of the sched domains hierarchy. The domains
Linus Torvalds1da177e2005-04-16 15:20:36 -07007829 * and groups cannot be updated in place without racing with the balancing
Nick Piggin41c7ce92005-06-25 14:57:24 -07007830 * code, so we temporarily attach all running cpus to the NULL domain
Linus Torvalds1da177e2005-04-16 15:20:36 -07007831 * which will prevent rebalancing while the sched domains are recalculated.
7832 */
7833static int update_sched_domains(struct notifier_block *nfb,
7834 unsigned long action, void *hcpu)
7835{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007836 switch (action) {
7837 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007838 case CPU_UP_PREPARE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007839 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007840 case CPU_DOWN_PREPARE_FROZEN:
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007841 detach_destroy_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007842 return NOTIFY_OK;
7843
7844 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007845 case CPU_UP_CANCELED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007846 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007847 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007848 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007849 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007850 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007851 case CPU_DEAD_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007852 /*
7853 * Fall through and re-initialise the domains.
7854 */
7855 break;
7856 default:
7857 return NOTIFY_DONE;
7858 }
7859
7860 /* The hotplug lock is already held by cpu_up/cpu_down */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007861 arch_init_sched_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007862
7863 return NOTIFY_OK;
7864}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007865
7866void __init sched_init_smp(void)
7867{
Nick Piggin5c1e1762006-10-03 01:14:04 -07007868 cpumask_t non_isolated_cpus;
7869
Mike Travis434d53b2008-04-04 18:11:04 -07007870#if defined(CONFIG_NUMA)
7871 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7872 GFP_KERNEL);
7873 BUG_ON(sched_group_nodes_bycpu == NULL);
7874#endif
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007875 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02007876 mutex_lock(&sched_domains_mutex);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007877 arch_init_sched_domains(&cpu_online_map);
Nathan Lynche5e56732007-01-10 23:15:28 -08007878 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
Nick Piggin5c1e1762006-10-03 01:14:04 -07007879 if (cpus_empty(non_isolated_cpus))
7880 cpu_set(smp_processor_id(), non_isolated_cpus);
Heiko Carstens712555e2008-04-28 11:33:07 +02007881 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007882 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007883 /* XXX: Theoretical race here - CPU may be hotplugged now */
7884 hotcpu_notifier(update_sched_domains, 0);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02007885 init_hrtick();
Nick Piggin5c1e1762006-10-03 01:14:04 -07007886
7887 /* Move init over to a non-isolated CPU */
Mike Travis7c16ec52008-04-04 18:11:11 -07007888 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
Nick Piggin5c1e1762006-10-03 01:14:04 -07007889 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01007890 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007891}
7892#else
7893void __init sched_init_smp(void)
7894{
Ingo Molnar19978ca2007-11-09 22:39:38 +01007895 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007896}
7897#endif /* CONFIG_SMP */
7898
7899int in_sched_functions(unsigned long addr)
7900{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007901 return in_lock_functions(addr) ||
7902 (addr >= (unsigned long)__sched_text_start
7903 && addr < (unsigned long)__sched_text_end);
7904}
7905
Alexey Dobriyana9957442007-10-15 17:00:13 +02007906static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02007907{
7908 cfs_rq->tasks_timeline = RB_ROOT;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02007909 INIT_LIST_HEAD(&cfs_rq->tasks);
Ingo Molnardd41f592007-07-09 18:51:59 +02007910#ifdef CONFIG_FAIR_GROUP_SCHED
7911 cfs_rq->rq = rq;
7912#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02007913 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02007914}
7915
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007916static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7917{
7918 struct rt_prio_array *array;
7919 int i;
7920
7921 array = &rt_rq->active;
7922 for (i = 0; i < MAX_RT_PRIO; i++) {
7923 INIT_LIST_HEAD(array->queue + i);
7924 __clear_bit(i, array->bitmap);
7925 }
7926 /* delimiter for bitsearch: */
7927 __set_bit(MAX_RT_PRIO, array->bitmap);
7928
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007929#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Peter Zijlstra48d5e252008-01-25 21:08:31 +01007930 rt_rq->highest_prio = MAX_RT_PRIO;
7931#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007932#ifdef CONFIG_SMP
7933 rt_rq->rt_nr_migratory = 0;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007934 rt_rq->overloaded = 0;
7935#endif
7936
7937 rt_rq->rt_time = 0;
7938 rt_rq->rt_throttled = 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007939 rt_rq->rt_runtime = 0;
7940 spin_lock_init(&rt_rq->rt_runtime_lock);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007941
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007942#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01007943 rt_rq->rt_nr_boosted = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007944 rt_rq->rq = rq;
7945#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007946}
7947
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007948#ifdef CONFIG_FAIR_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007949static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7950 struct sched_entity *se, int cpu, int add,
7951 struct sched_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007952{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007953 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007954 tg->cfs_rq[cpu] = cfs_rq;
7955 init_cfs_rq(cfs_rq, rq);
7956 cfs_rq->tg = tg;
7957 if (add)
7958 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7959
7960 tg->se[cpu] = se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02007961 /* se could be NULL for init_task_group */
7962 if (!se)
7963 return;
7964
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007965 if (!parent)
7966 se->cfs_rq = &rq->cfs;
7967 else
7968 se->cfs_rq = parent->my_q;
7969
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007970 se->my_q = cfs_rq;
7971 se->load.weight = tg->shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02007972 se->load.inv_weight = 0;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007973 se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007974}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007975#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007976
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007977#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007978static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7979 struct sched_rt_entity *rt_se, int cpu, int add,
7980 struct sched_rt_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007981{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007982 struct rq *rq = cpu_rq(cpu);
7983
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007984 tg->rt_rq[cpu] = rt_rq;
7985 init_rt_rq(rt_rq, rq);
7986 rt_rq->tg = tg;
7987 rt_rq->rt_se = rt_se;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007988 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007989 if (add)
7990 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7991
7992 tg->rt_se[cpu] = rt_se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02007993 if (!rt_se)
7994 return;
7995
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007996 if (!parent)
7997 rt_se->rt_rq = &rq->rt;
7998 else
7999 rt_se->rt_rq = parent->my_q;
8000
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008001 rt_se->rt_rq = &rq->rt;
8002 rt_se->my_q = rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008003 rt_se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008004 INIT_LIST_HEAD(&rt_se->run_list);
8005}
8006#endif
8007
Linus Torvalds1da177e2005-04-16 15:20:36 -07008008void __init sched_init(void)
8009{
Ingo Molnardd41f592007-07-09 18:51:59 +02008010 int i, j;
Mike Travis434d53b2008-04-04 18:11:04 -07008011 unsigned long alloc_size = 0, ptr;
8012
8013#ifdef CONFIG_FAIR_GROUP_SCHED
8014 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8015#endif
8016#ifdef CONFIG_RT_GROUP_SCHED
8017 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8018#endif
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008019#ifdef CONFIG_USER_SCHED
8020 alloc_size *= 2;
8021#endif
Mike Travis434d53b2008-04-04 18:11:04 -07008022 /*
8023 * As sched_init() is called before page_alloc is setup,
8024 * we use alloc_bootmem().
8025 */
8026 if (alloc_size) {
David Miller5a9d3222008-04-24 20:46:20 -07008027 ptr = (unsigned long)alloc_bootmem(alloc_size);
Mike Travis434d53b2008-04-04 18:11:04 -07008028
8029#ifdef CONFIG_FAIR_GROUP_SCHED
8030 init_task_group.se = (struct sched_entity **)ptr;
8031 ptr += nr_cpu_ids * sizeof(void **);
8032
8033 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8034 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008035
8036#ifdef CONFIG_USER_SCHED
8037 root_task_group.se = (struct sched_entity **)ptr;
8038 ptr += nr_cpu_ids * sizeof(void **);
8039
8040 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8041 ptr += nr_cpu_ids * sizeof(void **);
8042#endif
Mike Travis434d53b2008-04-04 18:11:04 -07008043#endif
8044#ifdef CONFIG_RT_GROUP_SCHED
8045 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8046 ptr += nr_cpu_ids * sizeof(void **);
8047
8048 init_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008049 ptr += nr_cpu_ids * sizeof(void **);
8050
8051#ifdef CONFIG_USER_SCHED
8052 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8053 ptr += nr_cpu_ids * sizeof(void **);
8054
8055 root_task_group.rt_rq = (struct rt_rq **)ptr;
8056 ptr += nr_cpu_ids * sizeof(void **);
8057#endif
Mike Travis434d53b2008-04-04 18:11:04 -07008058#endif
8059 }
Ingo Molnardd41f592007-07-09 18:51:59 +02008060
Gregory Haskins57d885f2008-01-25 21:08:18 +01008061#ifdef CONFIG_SMP
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008062 init_aggregate();
Gregory Haskins57d885f2008-01-25 21:08:18 +01008063 init_defrootdomain();
8064#endif
8065
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008066 init_rt_bandwidth(&def_rt_bandwidth,
8067 global_rt_period(), global_rt_runtime());
8068
8069#ifdef CONFIG_RT_GROUP_SCHED
8070 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8071 global_rt_period(), global_rt_runtime());
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008072#ifdef CONFIG_USER_SCHED
8073 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8074 global_rt_period(), RUNTIME_INF);
8075#endif
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008076#endif
8077
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008078#ifdef CONFIG_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008079 list_add(&init_task_group.list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008080 INIT_LIST_HEAD(&init_task_group.children);
8081
8082#ifdef CONFIG_USER_SCHED
8083 INIT_LIST_HEAD(&root_task_group.children);
8084 init_task_group.parent = &root_task_group;
8085 list_add(&init_task_group.siblings, &root_task_group.children);
8086#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008087#endif
8088
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08008089 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07008090 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008091
8092 rq = cpu_rq(i);
8093 spin_lock_init(&rq->lock);
Ingo Molnarfcb99372006-07-03 00:25:10 -07008094 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
Nick Piggin78979862005-06-25 14:57:13 -07008095 rq->nr_running = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02008096 init_cfs_rq(&rq->cfs, rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01008097 init_rt_rq(&rq->rt, rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008098#ifdef CONFIG_FAIR_GROUP_SCHED
8099 init_task_group.shares = init_task_group_load;
8100 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008101#ifdef CONFIG_CGROUP_SCHED
8102 /*
8103 * How much cpu bandwidth does init_task_group get?
8104 *
8105 * In case of task-groups formed thr' the cgroup filesystem, it
8106 * gets 100% of the cpu resources in the system. This overall
8107 * system cpu resource is divided among the tasks of
8108 * init_task_group and its child task-groups in a fair manner,
8109 * based on each entity's (task or task-group's) weight
8110 * (se->load.weight).
8111 *
8112 * In other words, if init_task_group has 10 tasks of weight
8113 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8114 * then A0's share of the cpu resource is:
8115 *
8116 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8117 *
8118 * We achieve this by letting init_task_group's tasks sit
8119 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8120 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008121 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008122#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008123 root_task_group.shares = NICE_0_LOAD;
8124 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008125 /*
8126 * In case of task-groups formed thr' the user id of tasks,
8127 * init_task_group represents tasks belonging to root user.
8128 * Hence it forms a sibling of all subsequent groups formed.
8129 * In this case, init_task_group gets only a fraction of overall
8130 * system cpu resource, based on the weight assigned to root
8131 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8132 * by letting tasks of init_task_group sit in a separate cfs_rq
8133 * (init_cfs_rq) and having one entity represent this group of
8134 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8135 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008136 init_tg_cfs_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008137 &per_cpu(init_cfs_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008138 &per_cpu(init_sched_entity, i), i, 1,
8139 root_task_group.se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008140
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008141#endif
Dhaval Giani354d60c2008-04-19 19:44:59 +02008142#endif /* CONFIG_FAIR_GROUP_SCHED */
8143
8144 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008145#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008146 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008147#ifdef CONFIG_CGROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008148 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008149#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008150 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008151 init_tg_rt_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008152 &per_cpu(init_rt_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008153 &per_cpu(init_sched_rt_entity, i), i, 1,
8154 root_task_group.rt_se[i]);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008155#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008156#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008157
Ingo Molnardd41f592007-07-09 18:51:59 +02008158 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8159 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008160#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07008161 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01008162 rq->rd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008163 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02008164 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008165 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07008166 rq->cpu = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008167 rq->migration_thread = NULL;
8168 INIT_LIST_HEAD(&rq->migration_queue);
Gregory Haskinsdc938522008-01-25 21:08:26 +01008169 rq_attach_root(rq, &def_root_domain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008170#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01008171 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008172 atomic_set(&rq->nr_iowait, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008173 }
8174
Peter Williams2dd73a42006-06-27 02:54:34 -07008175 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07008176
Avi Kivitye107be32007-07-26 13:40:43 +02008177#ifdef CONFIG_PREEMPT_NOTIFIERS
8178 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8179#endif
8180
Christoph Lameterc9819f42006-12-10 02:20:25 -08008181#ifdef CONFIG_SMP
8182 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8183#endif
8184
Heiko Carstensb50f60c2006-07-30 03:03:52 -07008185#ifdef CONFIG_RT_MUTEXES
8186 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8187#endif
8188
Linus Torvalds1da177e2005-04-16 15:20:36 -07008189 /*
8190 * The boot idle thread does lazy MMU switching as well:
8191 */
8192 atomic_inc(&init_mm.mm_count);
8193 enter_lazy_tlb(&init_mm, current);
8194
8195 /*
8196 * Make us the idle thread. Technically, schedule() should not be
8197 * called from this thread, however somewhere below it might be,
8198 * but because we are the idle thread, we just pick up running again
8199 * when this runqueue becomes "idle".
8200 */
8201 init_idle(current, smp_processor_id());
Ingo Molnardd41f592007-07-09 18:51:59 +02008202 /*
8203 * During early bootup we pretend to be a normal task:
8204 */
8205 current->sched_class = &fair_sched_class;
Ingo Molnar6892b752008-02-13 14:02:36 +01008206
8207 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008208}
8209
8210#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8211void __might_sleep(char *file, int line)
8212{
Ingo Molnar48f24c42006-07-03 00:25:40 -07008213#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07008214 static unsigned long prev_jiffy; /* ratelimiting */
8215
8216 if ((in_atomic() || irqs_disabled()) &&
8217 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8218 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8219 return;
8220 prev_jiffy = jiffies;
Ingo Molnar91368d72006-03-23 03:00:54 -08008221 printk(KERN_ERR "BUG: sleeping function called from invalid"
Linus Torvalds1da177e2005-04-16 15:20:36 -07008222 " context at %s:%d\n", file, line);
8223 printk("in_atomic():%d, irqs_disabled():%d\n",
8224 in_atomic(), irqs_disabled());
Peter Zijlstraa4c410f2006-12-06 20:37:21 -08008225 debug_show_held_locks(current);
Ingo Molnar3117df02006-12-13 00:34:43 -08008226 if (irqs_disabled())
8227 print_irqtrace_events(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008228 dump_stack();
8229 }
8230#endif
8231}
8232EXPORT_SYMBOL(__might_sleep);
8233#endif
8234
8235#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02008236static void normalize_task(struct rq *rq, struct task_struct *p)
8237{
8238 int on_rq;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02008239
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02008240 update_rq_clock(rq);
8241 on_rq = p->se.on_rq;
8242 if (on_rq)
8243 deactivate_task(rq, p, 0);
8244 __setscheduler(rq, p, SCHED_NORMAL, 0);
8245 if (on_rq) {
8246 activate_task(rq, p, 0);
8247 resched_task(rq->curr);
8248 }
8249}
8250
Linus Torvalds1da177e2005-04-16 15:20:36 -07008251void normalize_rt_tasks(void)
8252{
Ingo Molnara0f98a12007-06-17 18:37:45 +02008253 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008254 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07008255 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008256
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008257 read_lock_irqsave(&tasklist_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02008258 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02008259 /*
8260 * Only normalize user tasks:
8261 */
8262 if (!p->mm)
8263 continue;
8264
Ingo Molnardd41f592007-07-09 18:51:59 +02008265 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02008266#ifdef CONFIG_SCHEDSTATS
8267 p->se.wait_start = 0;
8268 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02008269 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02008270#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02008271
8272 if (!rt_task(p)) {
8273 /*
8274 * Renice negative nice level userspace
8275 * tasks back to 0:
8276 */
8277 if (TASK_NICE(p) < 0 && p->mm)
8278 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008279 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02008280 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008281
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008282 spin_lock(&p->pi_lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -07008283 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008284
Ingo Molnar178be792007-10-15 17:00:18 +02008285 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02008286
Ingo Molnarb29739f2006-06-27 02:54:51 -07008287 __task_rq_unlock(rq);
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008288 spin_unlock(&p->pi_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02008289 } while_each_thread(g, p);
8290
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008291 read_unlock_irqrestore(&tasklist_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008292}
8293
8294#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07008295
8296#ifdef CONFIG_IA64
8297/*
8298 * These functions are only useful for the IA64 MCA handling.
8299 *
8300 * They can only be called when the whole system has been
8301 * stopped - every CPU needs to be quiescent, and no scheduling
8302 * activity can take place. Using them for anything else would
8303 * be a serious bug, and as a result, they aren't even visible
8304 * under any other configuration.
8305 */
8306
8307/**
8308 * curr_task - return the current task for a given cpu.
8309 * @cpu: the processor in question.
8310 *
8311 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8312 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07008313struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07008314{
8315 return cpu_curr(cpu);
8316}
8317
8318/**
8319 * set_curr_task - set the current task for a given cpu.
8320 * @cpu: the processor in question.
8321 * @p: the task pointer to set.
8322 *
8323 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008324 * are serviced on a separate stack. It allows the architecture to switch the
8325 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07008326 * must be called with all CPU's synchronized, and interrupts disabled, the
8327 * and caller must save the original value of the current task (see
8328 * curr_task() above) and restore that value before reenabling interrupts and
8329 * re-starting the system.
8330 *
8331 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8332 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07008333void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07008334{
8335 cpu_curr(cpu) = p;
8336}
8337
8338#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008339
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008340#ifdef CONFIG_FAIR_GROUP_SCHED
8341static void free_fair_sched_group(struct task_group *tg)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008342{
8343 int i;
8344
8345 for_each_possible_cpu(i) {
8346 if (tg->cfs_rq)
8347 kfree(tg->cfs_rq[i]);
8348 if (tg->se)
8349 kfree(tg->se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008350 }
8351
8352 kfree(tg->cfs_rq);
8353 kfree(tg->se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008354}
8355
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008356static
8357int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008358{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008359 struct cfs_rq *cfs_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008360 struct sched_entity *se, *parent_se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008361 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008362 int i;
8363
Mike Travis434d53b2008-04-04 18:11:04 -07008364 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008365 if (!tg->cfs_rq)
8366 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07008367 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008368 if (!tg->se)
8369 goto err;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008370
8371 tg->shares = NICE_0_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008372
8373 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008374 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008375
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008376 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8377 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008378 if (!cfs_rq)
8379 goto err;
8380
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008381 se = kmalloc_node(sizeof(struct sched_entity),
8382 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008383 if (!se)
8384 goto err;
8385
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008386 parent_se = parent ? parent->se[i] : NULL;
8387 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008388 }
8389
8390 return 1;
8391
8392 err:
8393 return 0;
8394}
8395
8396static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8397{
8398 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8399 &cpu_rq(cpu)->leaf_cfs_rq_list);
8400}
8401
8402static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8403{
8404 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8405}
8406#else
8407static inline void free_fair_sched_group(struct task_group *tg)
8408{
8409}
8410
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008411static inline
8412int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008413{
8414 return 1;
8415}
8416
8417static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8418{
8419}
8420
8421static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8422{
8423}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008424#endif
8425
8426#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008427static void free_rt_sched_group(struct task_group *tg)
8428{
8429 int i;
8430
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008431 destroy_rt_bandwidth(&tg->rt_bandwidth);
8432
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008433 for_each_possible_cpu(i) {
8434 if (tg->rt_rq)
8435 kfree(tg->rt_rq[i]);
8436 if (tg->rt_se)
8437 kfree(tg->rt_se[i]);
8438 }
8439
8440 kfree(tg->rt_rq);
8441 kfree(tg->rt_se);
8442}
8443
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008444static
8445int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008446{
8447 struct rt_rq *rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008448 struct sched_rt_entity *rt_se, *parent_se;
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008449 struct rq *rq;
8450 int i;
8451
Mike Travis434d53b2008-04-04 18:11:04 -07008452 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008453 if (!tg->rt_rq)
8454 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07008455 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008456 if (!tg->rt_se)
8457 goto err;
8458
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008459 init_rt_bandwidth(&tg->rt_bandwidth,
8460 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008461
8462 for_each_possible_cpu(i) {
8463 rq = cpu_rq(i);
8464
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008465 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8466 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8467 if (!rt_rq)
8468 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008469
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008470 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8471 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8472 if (!rt_se)
8473 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008474
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008475 parent_se = parent ? parent->rt_se[i] : NULL;
8476 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008477 }
8478
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008479 return 1;
8480
8481 err:
8482 return 0;
8483}
8484
8485static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8486{
8487 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8488 &cpu_rq(cpu)->leaf_rt_rq_list);
8489}
8490
8491static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8492{
8493 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8494}
8495#else
8496static inline void free_rt_sched_group(struct task_group *tg)
8497{
8498}
8499
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008500static inline
8501int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008502{
8503 return 1;
8504}
8505
8506static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8507{
8508}
8509
8510static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8511{
8512}
8513#endif
8514
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008515#ifdef CONFIG_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008516static void free_sched_group(struct task_group *tg)
8517{
8518 free_fair_sched_group(tg);
8519 free_rt_sched_group(tg);
8520 kfree(tg);
8521}
8522
8523/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008524struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008525{
8526 struct task_group *tg;
8527 unsigned long flags;
8528 int i;
8529
8530 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8531 if (!tg)
8532 return ERR_PTR(-ENOMEM);
8533
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008534 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008535 goto err;
8536
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008537 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008538 goto err;
8539
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008540 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008541 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008542 register_fair_sched_group(tg, i);
8543 register_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008544 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008545 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008546
8547 WARN_ON(!parent); /* root should already exist */
8548
8549 tg->parent = parent;
8550 list_add_rcu(&tg->siblings, &parent->children);
8551 INIT_LIST_HEAD(&tg->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008552 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008553
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008554 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008555
8556err:
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008557 free_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008558 return ERR_PTR(-ENOMEM);
8559}
8560
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008561/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008562static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008563{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008564 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008565 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008566}
8567
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008568/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02008569void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008570{
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008571 unsigned long flags;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008572 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008573
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008574 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008575 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008576 unregister_fair_sched_group(tg, i);
8577 unregister_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008578 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008579 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008580 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008581 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008582
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008583 /* wait for possible concurrent references to cfs_rqs complete */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008584 call_rcu(&tg->rcu, free_sched_group_rcu);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008585}
8586
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008587/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +02008588 * The caller of this function should have put the task in its new group
8589 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8590 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008591 */
8592void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008593{
8594 int on_rq, running;
8595 unsigned long flags;
8596 struct rq *rq;
8597
8598 rq = task_rq_lock(tsk, &flags);
8599
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008600 update_rq_clock(rq);
8601
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01008602 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008603 on_rq = tsk->se.on_rq;
8604
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008605 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008606 dequeue_task(rq, tsk, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008607 if (unlikely(running))
8608 tsk->sched_class->put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008609
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008610 set_task_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008611
Peter Zijlstra810b3812008-02-29 15:21:01 -05008612#ifdef CONFIG_FAIR_GROUP_SCHED
8613 if (tsk->sched_class->moved_group)
8614 tsk->sched_class->moved_group(tsk);
8615#endif
8616
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008617 if (unlikely(running))
8618 tsk->sched_class->set_curr_task(rq);
8619 if (on_rq)
Dmitry Adamushko7074bad2007-10-15 17:00:07 +02008620 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008621
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008622 task_rq_unlock(rq, &flags);
8623}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008624#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008625
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008626#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008627static void __set_se_shares(struct sched_entity *se, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008628{
8629 struct cfs_rq *cfs_rq = se->cfs_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008630 int on_rq;
8631
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008632 on_rq = se->on_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008633 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008634 dequeue_entity(cfs_rq, se, 0);
8635
8636 se->load.weight = shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02008637 se->load.inv_weight = 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008638
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008639 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008640 enqueue_entity(cfs_rq, se, 0);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008641}
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008642
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008643static void set_se_shares(struct sched_entity *se, unsigned long shares)
8644{
8645 struct cfs_rq *cfs_rq = se->cfs_rq;
8646 struct rq *rq = cfs_rq->rq;
8647 unsigned long flags;
8648
8649 spin_lock_irqsave(&rq->lock, flags);
8650 __set_se_shares(se, shares);
8651 spin_unlock_irqrestore(&rq->lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008652}
8653
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008654static DEFINE_MUTEX(shares_mutex);
8655
Ingo Molnar4cf86d72007-10-15 17:00:14 +02008656int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008657{
8658 int i;
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008659 unsigned long flags;
Ingo Molnarc61935f2008-01-22 11:24:58 +01008660
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008661 /*
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008662 * We can't change the weight of the root cgroup.
8663 */
8664 if (!tg->se[0])
8665 return -EINVAL;
8666
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008667 if (shares < MIN_SHARES)
8668 shares = MIN_SHARES;
Miao Xiecb4ad1f2008-04-28 12:54:56 +08008669 else if (shares > MAX_SHARES)
8670 shares = MAX_SHARES;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008671
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008672 mutex_lock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008673 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008674 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008675
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008676 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008677 for_each_possible_cpu(i)
8678 unregister_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008679 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008680 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01008681
8682 /* wait for any ongoing reference to this group to finish */
8683 synchronize_sched();
8684
8685 /*
8686 * Now we are free to modify the group's share on each cpu
8687 * w/o tripping rebalance_share or load_balance_fair.
8688 */
8689 tg->shares = shares;
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008690 for_each_possible_cpu(i) {
8691 /*
8692 * force a rebalance
8693 */
8694 cfs_rq_set_shares(tg->cfs_rq[i], 0);
Miao Xiecb4ad1f2008-04-28 12:54:56 +08008695 set_se_shares(tg->se[i], shares);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008696 }
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01008697
8698 /*
8699 * Enable load balance activity on this group, by inserting it back on
8700 * each cpu's rq->leaf_cfs_rq_list.
8701 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008702 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008703 for_each_possible_cpu(i)
8704 register_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008705 list_add_rcu(&tg->siblings, &tg->parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008706 spin_unlock_irqrestore(&task_group_lock, flags);
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008707done:
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008708 mutex_unlock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008709 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008710}
8711
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008712unsigned long sched_group_shares(struct task_group *tg)
8713{
8714 return tg->shares;
8715}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008716#endif
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008717
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008718#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008719/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008720 * Ensure that the real time constraints are schedulable.
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008721 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008722static DEFINE_MUTEX(rt_constraints_mutex);
8723
8724static unsigned long to_ratio(u64 period, u64 runtime)
8725{
8726 if (runtime == RUNTIME_INF)
8727 return 1ULL << 16;
8728
Roman Zippel6f6d6a12008-05-01 04:34:28 -07008729 return div64_u64(runtime << 16, period);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008730}
8731
Peter Zijlstrab40b2e82008-04-19 19:45:00 +02008732#ifdef CONFIG_CGROUP_SCHED
8733static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8734{
8735 struct task_group *tgi, *parent = tg->parent;
8736 unsigned long total = 0;
8737
8738 if (!parent) {
8739 if (global_rt_period() < period)
8740 return 0;
8741
8742 return to_ratio(period, runtime) <
8743 to_ratio(global_rt_period(), global_rt_runtime());
8744 }
8745
8746 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8747 return 0;
8748
8749 rcu_read_lock();
8750 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8751 if (tgi == tg)
8752 continue;
8753
8754 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8755 tgi->rt_bandwidth.rt_runtime);
8756 }
8757 rcu_read_unlock();
8758
8759 return total + to_ratio(period, runtime) <
8760 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8761 parent->rt_bandwidth.rt_runtime);
8762}
8763#elif defined CONFIG_USER_SCHED
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008764static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008765{
8766 struct task_group *tgi;
8767 unsigned long total = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008768 unsigned long global_ratio =
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008769 to_ratio(global_rt_period(), global_rt_runtime());
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008770
8771 rcu_read_lock();
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008772 list_for_each_entry_rcu(tgi, &task_groups, list) {
8773 if (tgi == tg)
8774 continue;
8775
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008776 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8777 tgi->rt_bandwidth.rt_runtime);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008778 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008779 rcu_read_unlock();
8780
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008781 return total + to_ratio(period, runtime) < global_ratio;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008782}
Peter Zijlstrab40b2e82008-04-19 19:45:00 +02008783#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008784
Dhaval Giani521f1a242008-02-28 15:21:56 +05308785/* Must be called with tasklist_lock held */
8786static inline int tg_has_rt_tasks(struct task_group *tg)
8787{
8788 struct task_struct *g, *p;
8789 do_each_thread(g, p) {
8790 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8791 return 1;
8792 } while_each_thread(g, p);
8793 return 0;
8794}
8795
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008796static int tg_set_bandwidth(struct task_group *tg,
8797 u64 rt_period, u64 rt_runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008798{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008799 int i, err = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008800
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008801 mutex_lock(&rt_constraints_mutex);
Dhaval Giani521f1a242008-02-28 15:21:56 +05308802 read_lock(&tasklist_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008803 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
Dhaval Giani521f1a242008-02-28 15:21:56 +05308804 err = -EBUSY;
8805 goto unlock;
8806 }
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008807 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8808 err = -EINVAL;
8809 goto unlock;
8810 }
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008811
8812 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008813 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8814 tg->rt_bandwidth.rt_runtime = rt_runtime;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008815
8816 for_each_possible_cpu(i) {
8817 struct rt_rq *rt_rq = tg->rt_rq[i];
8818
8819 spin_lock(&rt_rq->rt_runtime_lock);
8820 rt_rq->rt_runtime = rt_runtime;
8821 spin_unlock(&rt_rq->rt_runtime_lock);
8822 }
8823 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008824 unlock:
Dhaval Giani521f1a242008-02-28 15:21:56 +05308825 read_unlock(&tasklist_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008826 mutex_unlock(&rt_constraints_mutex);
8827
8828 return err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008829}
8830
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008831int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8832{
8833 u64 rt_runtime, rt_period;
8834
8835 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8836 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8837 if (rt_runtime_us < 0)
8838 rt_runtime = RUNTIME_INF;
8839
8840 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8841}
8842
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008843long sched_group_rt_runtime(struct task_group *tg)
8844{
8845 u64 rt_runtime_us;
8846
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008847 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008848 return -1;
8849
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008850 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008851 do_div(rt_runtime_us, NSEC_PER_USEC);
8852 return rt_runtime_us;
8853}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008854
8855int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8856{
8857 u64 rt_runtime, rt_period;
8858
8859 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8860 rt_runtime = tg->rt_bandwidth.rt_runtime;
8861
8862 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8863}
8864
8865long sched_group_rt_period(struct task_group *tg)
8866{
8867 u64 rt_period_us;
8868
8869 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8870 do_div(rt_period_us, NSEC_PER_USEC);
8871 return rt_period_us;
8872}
8873
8874static int sched_rt_global_constraints(void)
8875{
8876 int ret = 0;
8877
8878 mutex_lock(&rt_constraints_mutex);
8879 if (!__rt_schedulable(NULL, 1, 0))
8880 ret = -EINVAL;
8881 mutex_unlock(&rt_constraints_mutex);
8882
8883 return ret;
8884}
8885#else
8886static int sched_rt_global_constraints(void)
8887{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008888 unsigned long flags;
8889 int i;
8890
8891 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8892 for_each_possible_cpu(i) {
8893 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8894
8895 spin_lock(&rt_rq->rt_runtime_lock);
8896 rt_rq->rt_runtime = global_rt_runtime();
8897 spin_unlock(&rt_rq->rt_runtime_lock);
8898 }
8899 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8900
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008901 return 0;
8902}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008903#endif
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008904
8905int sched_rt_handler(struct ctl_table *table, int write,
8906 struct file *filp, void __user *buffer, size_t *lenp,
8907 loff_t *ppos)
8908{
8909 int ret;
8910 int old_period, old_runtime;
8911 static DEFINE_MUTEX(mutex);
8912
8913 mutex_lock(&mutex);
8914 old_period = sysctl_sched_rt_period;
8915 old_runtime = sysctl_sched_rt_runtime;
8916
8917 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8918
8919 if (!ret && write) {
8920 ret = sched_rt_global_constraints();
8921 if (ret) {
8922 sysctl_sched_rt_period = old_period;
8923 sysctl_sched_rt_runtime = old_runtime;
8924 } else {
8925 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8926 def_rt_bandwidth.rt_period =
8927 ns_to_ktime(global_rt_period());
8928 }
8929 }
8930 mutex_unlock(&mutex);
8931
8932 return ret;
8933}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008934
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008935#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008936
8937/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02008938static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008939{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008940 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8941 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008942}
8943
8944static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +02008945cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008946{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008947 struct task_group *tg, *parent;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008948
Paul Menage2b01dfe2007-10-24 18:23:50 +02008949 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008950 /* This is early initialization for the top cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02008951 init_task_group.css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008952 return &init_task_group.css;
8953 }
8954
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008955 parent = cgroup_tg(cgrp->parent);
8956 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008957 if (IS_ERR(tg))
8958 return ERR_PTR(-ENOMEM);
8959
8960 /* Bind the cgroup to task_group object we just created */
Paul Menage2b01dfe2007-10-24 18:23:50 +02008961 tg->css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008962
8963 return &tg->css;
8964}
8965
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008966static void
8967cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008968{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008969 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008970
8971 sched_destroy_group(tg);
8972}
8973
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008974static int
8975cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8976 struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008977{
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008978#ifdef CONFIG_RT_GROUP_SCHED
8979 /* Don't accept realtime tasks when there is no way for them to run */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008980 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008981 return -EINVAL;
8982#else
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008983 /* We don't support RT-tasks being in separate groups */
8984 if (tsk->sched_class != &fair_sched_class)
8985 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008986#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008987
8988 return 0;
8989}
8990
8991static void
Paul Menage2b01dfe2007-10-24 18:23:50 +02008992cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008993 struct cgroup *old_cont, struct task_struct *tsk)
8994{
8995 sched_move_task(tsk);
8996}
8997
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008998#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagef4c753b2008-04-29 00:59:56 -07008999static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
Paul Menage2b01dfe2007-10-24 18:23:50 +02009000 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009001{
Paul Menage2b01dfe2007-10-24 18:23:50 +02009002 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009003}
9004
Paul Menagef4c753b2008-04-29 00:59:56 -07009005static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009006{
Paul Menage2b01dfe2007-10-24 18:23:50 +02009007 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009008
9009 return (u64) tg->shares;
9010}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009011#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009012
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009013#ifdef CONFIG_RT_GROUP_SCHED
Mirco Tischler0c708142008-05-14 16:05:46 -07009014static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
Paul Menage06ecb272008-04-29 01:00:06 -07009015 s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009016{
Paul Menage06ecb272008-04-29 01:00:06 -07009017 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009018}
9019
Paul Menage06ecb272008-04-29 01:00:06 -07009020static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009021{
Paul Menage06ecb272008-04-29 01:00:06 -07009022 return sched_group_rt_runtime(cgroup_tg(cgrp));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009023}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009024
9025static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9026 u64 rt_period_us)
9027{
9028 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9029}
9030
9031static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9032{
9033 return sched_group_rt_period(cgroup_tg(cgrp));
9034}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009035#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009036
Paul Menagefe5c7cc2007-10-29 21:18:11 +01009037static struct cftype cpu_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009038#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +01009039 {
9040 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -07009041 .read_u64 = cpu_shares_read_u64,
9042 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +01009043 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009044#endif
9045#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009046 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01009047 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -07009048 .read_s64 = cpu_rt_runtime_read,
9049 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009050 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009051 {
9052 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -07009053 .read_u64 = cpu_rt_period_read_uint,
9054 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009055 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009056#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009057};
9058
9059static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9060{
Paul Menagefe5c7cc2007-10-29 21:18:11 +01009061 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009062}
9063
9064struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +01009065 .name = "cpu",
9066 .create = cpu_cgroup_create,
9067 .destroy = cpu_cgroup_destroy,
9068 .can_attach = cpu_cgroup_can_attach,
9069 .attach = cpu_cgroup_attach,
9070 .populate = cpu_cgroup_populate,
9071 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009072 .early_init = 1,
9073};
9074
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009075#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009076
9077#ifdef CONFIG_CGROUP_CPUACCT
9078
9079/*
9080 * CPU accounting code for task groups.
9081 *
9082 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9083 * (balbir@in.ibm.com).
9084 */
9085
9086/* track cpu usage of a group of tasks */
9087struct cpuacct {
9088 struct cgroup_subsys_state css;
9089 /* cpuusage holds pointer to a u64-type object on every cpu */
9090 u64 *cpuusage;
9091};
9092
9093struct cgroup_subsys cpuacct_subsys;
9094
9095/* return cpu accounting group corresponding to this container */
Dhaval Giani32cd7562008-02-29 10:02:43 +05309096static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009097{
Dhaval Giani32cd7562008-02-29 10:02:43 +05309098 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009099 struct cpuacct, css);
9100}
9101
9102/* return cpu accounting group to which this task belongs */
9103static inline struct cpuacct *task_ca(struct task_struct *tsk)
9104{
9105 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9106 struct cpuacct, css);
9107}
9108
9109/* create a new cpu accounting group */
9110static struct cgroup_subsys_state *cpuacct_create(
Dhaval Giani32cd7562008-02-29 10:02:43 +05309111 struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009112{
9113 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9114
9115 if (!ca)
9116 return ERR_PTR(-ENOMEM);
9117
9118 ca->cpuusage = alloc_percpu(u64);
9119 if (!ca->cpuusage) {
9120 kfree(ca);
9121 return ERR_PTR(-ENOMEM);
9122 }
9123
9124 return &ca->css;
9125}
9126
9127/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009128static void
Dhaval Giani32cd7562008-02-29 10:02:43 +05309129cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009130{
Dhaval Giani32cd7562008-02-29 10:02:43 +05309131 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009132
9133 free_percpu(ca->cpuusage);
9134 kfree(ca);
9135}
9136
9137/* return total cpu usage (in nanoseconds) of a group */
Dhaval Giani32cd7562008-02-29 10:02:43 +05309138static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009139{
Dhaval Giani32cd7562008-02-29 10:02:43 +05309140 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009141 u64 totalcpuusage = 0;
9142 int i;
9143
9144 for_each_possible_cpu(i) {
9145 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9146
9147 /*
9148 * Take rq->lock to make 64-bit addition safe on 32-bit
9149 * platforms.
9150 */
9151 spin_lock_irq(&cpu_rq(i)->lock);
9152 totalcpuusage += *cpuusage;
9153 spin_unlock_irq(&cpu_rq(i)->lock);
9154 }
9155
9156 return totalcpuusage;
9157}
9158
Dhaval Giani0297b802008-02-29 10:02:44 +05309159static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9160 u64 reset)
9161{
9162 struct cpuacct *ca = cgroup_ca(cgrp);
9163 int err = 0;
9164 int i;
9165
9166 if (reset) {
9167 err = -EINVAL;
9168 goto out;
9169 }
9170
9171 for_each_possible_cpu(i) {
9172 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9173
9174 spin_lock_irq(&cpu_rq(i)->lock);
9175 *cpuusage = 0;
9176 spin_unlock_irq(&cpu_rq(i)->lock);
9177 }
9178out:
9179 return err;
9180}
9181
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009182static struct cftype files[] = {
9183 {
9184 .name = "usage",
Paul Menagef4c753b2008-04-29 00:59:56 -07009185 .read_u64 = cpuusage_read,
9186 .write_u64 = cpuusage_write,
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009187 },
9188};
9189
Dhaval Giani32cd7562008-02-29 10:02:43 +05309190static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009191{
Dhaval Giani32cd7562008-02-29 10:02:43 +05309192 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009193}
9194
9195/*
9196 * charge this task's execution time to its accounting group.
9197 *
9198 * called with rq->lock held.
9199 */
9200static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9201{
9202 struct cpuacct *ca;
9203
9204 if (!cpuacct_subsys.active)
9205 return;
9206
9207 ca = task_ca(tsk);
9208 if (ca) {
9209 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9210
9211 *cpuusage += cputime;
9212 }
9213}
9214
9215struct cgroup_subsys cpuacct_subsys = {
9216 .name = "cpuacct",
9217 .create = cpuacct_create,
9218 .destroy = cpuacct_destroy,
9219 .populate = cpuacct_populate,
9220 .subsys_id = cpuacct_subsys_id,
9221};
9222#endif /* CONFIG_CGROUP_CPUACCT */