blob: dc0be113f41d71a58f6e58f6ca9c8f93df534fa5 [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070073
Eric Dumazet5517d862007-05-08 00:32:57 -070074#include <asm/tlb.h>
Satyam Sharma838225b2007-10-24 18:23:50 +020075#include <asm/irq_regs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070076
Gregory Haskins6e0534f2008-05-12 21:21:01 +020077#include "sched_cpupri.h"
78
Linus Torvalds1da177e2005-04-16 15:20:36 -070079/*
80 * Convert user-nice values [ -20 ... 0 ... 19 ]
81 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
82 * and back.
83 */
84#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
85#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
86#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
87
88/*
89 * 'User priority' is the nice value converted to something we
90 * can work with better when scaling various scheduler parameters,
91 * it's a [ 0 ... 39 ] range.
92 */
93#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
94#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
95#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
96
97/*
Ingo Molnard7876a02008-01-25 21:08:19 +010098 * Helpers for converting nanosecond timing to jiffy resolution
Linus Torvalds1da177e2005-04-16 15:20:36 -070099 */
Eric Dumazetd6322fa2007-11-09 22:39:38 +0100100#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200102#define NICE_0_LOAD SCHED_LOAD_SCALE
103#define NICE_0_SHIFT SCHED_LOAD_SHIFT
104
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105/*
106 * These are the 'tuning knobs' of the scheduler:
107 *
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +0200108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109 * Timeslices get refilled after they expire.
110 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111#define DEF_TIMESLICE (100 * HZ / 1000)
Peter Williams2dd73a42006-06-27 02:54:34 -0700112
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200113/*
114 * single value that denotes runtime == period, ie unlimited time.
115 */
116#define RUNTIME_INF ((u64)~0ULL)
117
Eric Dumazet5517d862007-05-08 00:32:57 -0700118#ifdef CONFIG_SMP
119/*
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 */
123static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
124{
125 return reciprocal_divide(load, sg->reciprocal_cpu_power);
126}
127
128/*
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
131 */
132static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
133{
134 sg->__cpu_power += val;
135 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136}
137#endif
138
Ingo Molnare05606d2007-07-09 18:51:59 +0200139static inline int rt_policy(int policy)
140{
Roel Kluin3f33a7c2008-05-13 23:44:11 +0200141 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
Ingo Molnare05606d2007-07-09 18:51:59 +0200142 return 1;
143 return 0;
144}
145
146static inline int task_has_rt_policy(struct task_struct *p)
147{
148 return rt_policy(p->policy);
149}
150
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151/*
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200152 * This is the priority-queue data structure of the RT scheduling class:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200154struct rt_prio_array {
155 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
Gregory Haskins45c01e82008-05-12 21:20:41 +0200156 struct list_head xqueue[MAX_RT_PRIO]; /* exclusive queue */
157 struct list_head squeue[MAX_RT_PRIO]; /* shared queue */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200158};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200160struct rt_bandwidth {
Ingo Molnarea736ed2008-03-25 13:51:45 +0100161 /* nests inside the rq lock: */
162 spinlock_t rt_runtime_lock;
163 ktime_t rt_period;
164 u64 rt_runtime;
165 struct hrtimer rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200166};
167
168static struct rt_bandwidth def_rt_bandwidth;
169
170static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
171
172static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
173{
174 struct rt_bandwidth *rt_b =
175 container_of(timer, struct rt_bandwidth, rt_period_timer);
176 ktime_t now;
177 int overrun;
178 int idle = 0;
179
180 for (;;) {
181 now = hrtimer_cb_get_time(timer);
182 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
183
184 if (!overrun)
185 break;
186
187 idle = do_sched_rt_period_timer(rt_b, overrun);
188 }
189
190 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
191}
192
193static
194void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
195{
196 rt_b->rt_period = ns_to_ktime(period);
197 rt_b->rt_runtime = runtime;
198
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200199 spin_lock_init(&rt_b->rt_runtime_lock);
200
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200201 hrtimer_init(&rt_b->rt_period_timer,
202 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
203 rt_b->rt_period_timer.function = sched_rt_period_timer;
204 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
205}
206
207static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
208{
209 ktime_t now;
210
211 if (rt_b->rt_runtime == RUNTIME_INF)
212 return;
213
214 if (hrtimer_active(&rt_b->rt_period_timer))
215 return;
216
217 spin_lock(&rt_b->rt_runtime_lock);
218 for (;;) {
219 if (hrtimer_active(&rt_b->rt_period_timer))
220 break;
221
222 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
223 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
224 hrtimer_start(&rt_b->rt_period_timer,
225 rt_b->rt_period_timer.expires,
226 HRTIMER_MODE_ABS);
227 }
228 spin_unlock(&rt_b->rt_runtime_lock);
229}
230
231#ifdef CONFIG_RT_GROUP_SCHED
232static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
233{
234 hrtimer_cancel(&rt_b->rt_period_timer);
235}
236#endif
237
Heiko Carstens712555e2008-04-28 11:33:07 +0200238/*
239 * sched_domains_mutex serializes calls to arch_init_sched_domains,
240 * detach_destroy_domains and partition_sched_domains.
241 */
242static DEFINE_MUTEX(sched_domains_mutex);
243
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100244#ifdef CONFIG_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200245
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700246#include <linux/cgroup.h>
247
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200248struct cfs_rq;
249
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100250static LIST_HEAD(task_groups);
251
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200252/* task group related information */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200253struct task_group {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100254#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700255 struct cgroup_subsys_state css;
256#endif
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100257
258#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200259 /* schedulable entities of this group on each cpu */
260 struct sched_entity **se;
261 /* runqueue "owned" by this group on each cpu */
262 struct cfs_rq **cfs_rq;
263 unsigned long shares;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100264#endif
265
266#ifdef CONFIG_RT_GROUP_SCHED
267 struct sched_rt_entity **rt_se;
268 struct rt_rq **rt_rq;
269
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200270 struct rt_bandwidth rt_bandwidth;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100271#endif
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100272
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +0100273 struct rcu_head rcu;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100274 struct list_head list;
Peter Zijlstraf473aa52008-04-19 19:45:00 +0200275
276 struct task_group *parent;
277 struct list_head siblings;
278 struct list_head children;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200279};
280
Dhaval Giani354d60c2008-04-19 19:44:59 +0200281#ifdef CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200282
283/*
284 * Root task group.
285 * Every UID task group (including init_task_group aka UID-0) will
286 * be a child to this group.
287 */
288struct task_group root_task_group;
289
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100290#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200291/* Default task group's sched entity on each cpu */
292static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
293/* Default task group's cfs_rq on each cpu */
294static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200295#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100296
297#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100298static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
299static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200300#endif /* CONFIG_RT_GROUP_SCHED */
301#else /* !CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200302#define root_task_group init_task_group
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200303#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100304
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100305/* task_group_lock serializes add/remove of task groups and also changes to
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100306 * a task group's cpu shares.
307 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100308static DEFINE_SPINLOCK(task_group_lock);
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100309
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100310#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100311#ifdef CONFIG_USER_SCHED
Ingo Molnar0eab9142008-01-25 21:08:19 +0100312# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200313#else /* !CONFIG_USER_SCHED */
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100314# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200315#endif /* CONFIG_USER_SCHED */
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200316
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800317/*
318 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
321 */
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200322#define MIN_SHARES 2
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800323#define MAX_SHARES (ULONG_MAX - 1)
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200324
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100325static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100326#endif
327
328/* Default task group.
329 * Every task in system belong to this group at bootup.
330 */
Mike Travis434d53b2008-04-04 18:11:04 -0700331struct task_group init_task_group;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200332
333/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200334static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200335{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200336 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200337
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100338#ifdef CONFIG_USER_SCHED
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200339 tg = p->user->tg;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100340#elif defined(CONFIG_CGROUP_SCHED)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700341 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
342 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200343#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100344 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200345#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200346 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200347}
348
349/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100350static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200351{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100352#ifdef CONFIG_FAIR_GROUP_SCHED
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100353 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
354 p->se.parent = task_group(p)->se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100355#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100356
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100357#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100358 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
359 p->rt.parent = task_group(p)->rt_se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100360#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200361}
362
363#else
364
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100365static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200366
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100367#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200368
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200369/* CFS-related fields in a runqueue */
370struct cfs_rq {
371 struct load_weight load;
372 unsigned long nr_running;
373
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200374 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200375 u64 min_vruntime;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200376
377 struct rb_root tasks_timeline;
378 struct rb_node *rb_leftmost;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +0200379
380 struct list_head tasks;
381 struct list_head *balance_iterator;
382
383 /*
384 * 'curr' points to currently running entity on this cfs_rq.
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200385 * It is set to NULL otherwise (i.e when none are currently running).
386 */
Peter Zijlstraaa2ac252008-03-14 21:12:12 +0100387 struct sched_entity *curr, *next;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200388
389 unsigned long nr_spread_over;
390
Ingo Molnar62160e3f2007-10-15 17:00:03 +0200391#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200392 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
393
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100394 /*
395 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200396 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
397 * (like users, containers etc.)
398 *
399 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
400 * list is used during load balance.
401 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100402 struct list_head leaf_cfs_rq_list;
403 struct task_group *tg; /* group that "owns" this runqueue */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200404#endif
405};
406
407/* Real-Time classes' related field in a runqueue: */
408struct rt_rq {
409 struct rt_prio_array active;
Steven Rostedt63489e42008-01-25 21:08:03 +0100410 unsigned long rt_nr_running;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100411#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100412 int highest_prio; /* highest queued rt task prio */
413#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100414#ifdef CONFIG_SMP
Gregory Haskins73fe6aa2008-01-25 21:08:07 +0100415 unsigned long rt_nr_migratory;
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +0100416 int overloaded;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100417#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100418 int rt_throttled;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100419 u64 rt_time;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200420 u64 rt_runtime;
Ingo Molnarea736ed2008-03-25 13:51:45 +0100421 /* Nests inside the rq lock: */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200422 spinlock_t rt_runtime_lock;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100423
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100424#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100425 unsigned long rt_nr_boosted;
426
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100427 struct rq *rq;
428 struct list_head leaf_rt_rq_list;
429 struct task_group *tg;
430 struct sched_rt_entity *rt_se;
431#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200432};
433
Gregory Haskins57d885f2008-01-25 21:08:18 +0100434#ifdef CONFIG_SMP
435
436/*
437 * We add the notion of a root-domain which will be used to define per-domain
Ingo Molnar0eab9142008-01-25 21:08:19 +0100438 * variables. Each exclusive cpuset essentially defines an island domain by
439 * fully partitioning the member cpus from any other cpuset. Whenever a new
Gregory Haskins57d885f2008-01-25 21:08:18 +0100440 * exclusive cpuset is created, we also create and attach a new root-domain
441 * object.
442 *
Gregory Haskins57d885f2008-01-25 21:08:18 +0100443 */
444struct root_domain {
445 atomic_t refcount;
446 cpumask_t span;
447 cpumask_t online;
Gregory Haskins637f5082008-01-25 21:08:18 +0100448
Ingo Molnar0eab9142008-01-25 21:08:19 +0100449 /*
Gregory Haskins637f5082008-01-25 21:08:18 +0100450 * The "RT overload" flag: it gets set if a CPU has more than
451 * one runnable RT task.
452 */
453 cpumask_t rto_mask;
Ingo Molnar0eab9142008-01-25 21:08:19 +0100454 atomic_t rto_count;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200455#ifdef CONFIG_SMP
456 struct cpupri cpupri;
457#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +0100458};
459
Gregory Haskinsdc938522008-01-25 21:08:26 +0100460/*
461 * By default the system creates a single root-domain with all cpus as
462 * members (mimicking the global state we have today).
463 */
Gregory Haskins57d885f2008-01-25 21:08:18 +0100464static struct root_domain def_root_domain;
465
466#endif
467
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200468/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 * This is the main, per-CPU runqueue data structure.
470 *
471 * Locking rule: those places that want to lock multiple runqueues
472 * (such as the load balancing or the thread migration code), lock
473 * acquire operations must be ordered by ascending &runqueue.
474 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700475struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200476 /* runqueue lock: */
477 spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
479 /*
480 * nr_running and cpu_load should be in the same cacheline because
481 * remote CPUs use both these fields when doing load calculation.
482 */
483 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200484 #define CPU_LOAD_IDX_MAX 5
485 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh Bbdecea32007-05-08 00:32:48 -0700486 unsigned char idle_at_tick;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700487#ifdef CONFIG_NO_HZ
Guillaume Chazarain15934a32008-04-19 19:44:57 +0200488 unsigned long last_tick_seen;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700489 unsigned char in_nohz_recently;
490#endif
Ingo Molnard8016492007-10-18 21:32:55 +0200491 /* capture load from *all* tasks on this cpu: */
492 struct load_weight load;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200493 unsigned long nr_load_updates;
494 u64 nr_switches;
495
496 struct cfs_rq cfs;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100497 struct rt_rq rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100498
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200499#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200500 /* list of leaf cfs_rq on this cpu: */
501 struct list_head leaf_cfs_rq_list;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100502#endif
503#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100504 struct list_head leaf_rt_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506
507 /*
508 * This is part of a global counter where only the total sum
509 * over all CPUs matters. A task can increase this counter on
510 * one CPU and if it got migrated afterwards it may decrease
511 * it on another CPU. Always updated under the runqueue lock:
512 */
513 unsigned long nr_uninterruptible;
514
Ingo Molnar36c8b582006-07-03 00:25:41 -0700515 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800516 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200518
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200519 u64 clock;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200520
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 atomic_t nr_iowait;
522
523#ifdef CONFIG_SMP
Ingo Molnar0eab9142008-01-25 21:08:19 +0100524 struct root_domain *rd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 struct sched_domain *sd;
526
527 /* For active balancing */
528 int active_balance;
529 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200530 /* cpu of this runqueue: */
531 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532
Ingo Molnar36c8b582006-07-03 00:25:41 -0700533 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 struct list_head migration_queue;
535#endif
536
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100537#ifdef CONFIG_SCHED_HRTICK
538 unsigned long hrtick_flags;
539 ktime_t hrtick_expire;
540 struct hrtimer hrtick_timer;
541#endif
542
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543#ifdef CONFIG_SCHEDSTATS
544 /* latency stats */
545 struct sched_info rq_sched_info;
546
547 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200548 unsigned int yld_exp_empty;
549 unsigned int yld_act_empty;
550 unsigned int yld_both_empty;
551 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552
553 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200554 unsigned int sched_switch;
555 unsigned int sched_count;
556 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557
558 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200559 unsigned int ttwu_count;
560 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200561
562 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200563 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564#endif
Ingo Molnarfcb99372006-07-03 00:25:10 -0700565 struct lock_class_key rq_lock_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566};
567
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700568static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569
Ingo Molnardd41f592007-07-09 18:51:59 +0200570static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
571{
572 rq->curr->sched_class->check_preempt_curr(rq, p);
573}
574
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700575static inline int cpu_of(struct rq *rq)
576{
577#ifdef CONFIG_SMP
578 return rq->cpu;
579#else
580 return 0;
581#endif
582}
583
Ingo Molnar20d315d2007-07-09 18:51:58 +0200584/*
Nick Piggin674311d2005-06-25 14:57:27 -0700585 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700586 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700587 *
588 * The domain tree of any CPU may only be accessed from within
589 * preempt-disabled sections.
590 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700591#define for_each_domain(cpu, __sd) \
592 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593
594#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
595#define this_rq() (&__get_cpu_var(runqueues))
596#define task_rq(p) cpu_rq(task_cpu(p))
597#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
598
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200599static inline void update_rq_clock(struct rq *rq)
600{
601 rq->clock = sched_clock_cpu(cpu_of(rq));
602}
603
Ingo Molnare436d802007-07-19 21:28:35 +0200604/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200605 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
606 */
607#ifdef CONFIG_SCHED_DEBUG
608# define const_debug __read_mostly
609#else
610# define const_debug static const
611#endif
612
613/*
614 * Debugging: various feature bits
615 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200616
617#define SCHED_FEAT(name, enabled) \
618 __SCHED_FEAT_##name ,
619
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200620enum {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200621#include "sched_features.h"
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200622};
623
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200624#undef SCHED_FEAT
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200625
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200626#define SCHED_FEAT(name, enabled) \
627 (1UL << __SCHED_FEAT_##name) * enabled |
628
629const_debug unsigned int sysctl_sched_features =
630#include "sched_features.h"
631 0;
632
633#undef SCHED_FEAT
634
635#ifdef CONFIG_SCHED_DEBUG
636#define SCHED_FEAT(name, enabled) \
637 #name ,
638
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700639static __read_mostly char *sched_feat_names[] = {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200640#include "sched_features.h"
641 NULL
642};
643
644#undef SCHED_FEAT
645
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700646static int sched_feat_open(struct inode *inode, struct file *filp)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200647{
648 filp->private_data = inode->i_private;
649 return 0;
650}
651
652static ssize_t
653sched_feat_read(struct file *filp, char __user *ubuf,
654 size_t cnt, loff_t *ppos)
655{
656 char *buf;
657 int r = 0;
658 int len = 0;
659 int i;
660
661 for (i = 0; sched_feat_names[i]; i++) {
662 len += strlen(sched_feat_names[i]);
663 len += 4;
664 }
665
666 buf = kmalloc(len + 2, GFP_KERNEL);
667 if (!buf)
668 return -ENOMEM;
669
670 for (i = 0; sched_feat_names[i]; i++) {
671 if (sysctl_sched_features & (1UL << i))
672 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
673 else
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200674 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200675 }
676
677 r += sprintf(buf + r, "\n");
678 WARN_ON(r >= len + 2);
679
680 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
681
682 kfree(buf);
683
684 return r;
685}
686
687static ssize_t
688sched_feat_write(struct file *filp, const char __user *ubuf,
689 size_t cnt, loff_t *ppos)
690{
691 char buf[64];
692 char *cmp = buf;
693 int neg = 0;
694 int i;
695
696 if (cnt > 63)
697 cnt = 63;
698
699 if (copy_from_user(&buf, ubuf, cnt))
700 return -EFAULT;
701
702 buf[cnt] = 0;
703
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200704 if (strncmp(buf, "NO_", 3) == 0) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200705 neg = 1;
706 cmp += 3;
707 }
708
709 for (i = 0; sched_feat_names[i]; i++) {
710 int len = strlen(sched_feat_names[i]);
711
712 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
713 if (neg)
714 sysctl_sched_features &= ~(1UL << i);
715 else
716 sysctl_sched_features |= (1UL << i);
717 break;
718 }
719 }
720
721 if (!sched_feat_names[i])
722 return -EINVAL;
723
724 filp->f_pos += cnt;
725
726 return cnt;
727}
728
729static struct file_operations sched_feat_fops = {
730 .open = sched_feat_open,
731 .read = sched_feat_read,
732 .write = sched_feat_write,
733};
734
735static __init int sched_init_debug(void)
736{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200737 debugfs_create_file("sched_features", 0644, NULL, NULL,
738 &sched_feat_fops);
739
740 return 0;
741}
742late_initcall(sched_init_debug);
743
744#endif
745
746#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200747
748/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100749 * Number of tasks to iterate in a single balance run.
750 * Limited because this is done with IRQs disabled.
751 */
752const_debug unsigned int sysctl_sched_nr_migrate = 32;
753
754/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100755 * period over which we measure -rt task cpu usage in us.
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100756 * default: 1s
757 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100758unsigned int sysctl_sched_rt_period = 1000000;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100759
Ingo Molnar6892b752008-02-13 14:02:36 +0100760static __read_mostly int scheduler_running;
761
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100762/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100763 * part of the period that we allow rt tasks to run in us.
764 * default: 0.95s
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100765 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100766int sysctl_sched_rt_runtime = 950000;
767
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200768static inline u64 global_rt_period(void)
769{
770 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
771}
772
773static inline u64 global_rt_runtime(void)
774{
775 if (sysctl_sched_rt_period < 0)
776 return RUNTIME_INF;
777
778 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
779}
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100780
Ingo Molnar690229a2008-04-23 09:31:35 +0200781unsigned long long time_sync_thresh = 100000;
Ingo Molnar27ec4402008-02-28 21:00:21 +0100782
783static DEFINE_PER_CPU(unsigned long long, time_offset);
784static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
785
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100786/*
Ingo Molnar27ec4402008-02-28 21:00:21 +0100787 * Global lock which we take every now and then to synchronize
788 * the CPUs time. This method is not warp-safe, but it's good
789 * enough to synchronize slowly diverging time sources and thus
790 * it's good enough for tracing:
Ingo Molnare436d802007-07-19 21:28:35 +0200791 */
Ingo Molnar27ec4402008-02-28 21:00:21 +0100792static DEFINE_SPINLOCK(time_sync_lock);
793static unsigned long long prev_global_time;
794
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200795static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
Ingo Molnar27ec4402008-02-28 21:00:21 +0100796{
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200797 /*
798 * We want this inlined, to not get tracer function calls
799 * in this critical section:
800 */
801 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
802 __raw_spin_lock(&time_sync_lock.raw_lock);
Ingo Molnar27ec4402008-02-28 21:00:21 +0100803
804 if (time < prev_global_time) {
805 per_cpu(time_offset, cpu) += prev_global_time - time;
806 time = prev_global_time;
807 } else {
808 prev_global_time = time;
809 }
810
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200811 __raw_spin_unlock(&time_sync_lock.raw_lock);
812 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
Ingo Molnar27ec4402008-02-28 21:00:21 +0100813
814 return time;
815}
816
817static unsigned long long __cpu_clock(int cpu)
Ingo Molnare436d802007-07-19 21:28:35 +0200818{
Ingo Molnare436d802007-07-19 21:28:35 +0200819 unsigned long long now;
Ingo Molnare436d802007-07-19 21:28:35 +0200820
Ingo Molnar8ced5f62007-12-07 19:02:47 +0100821 /*
822 * Only call sched_clock() if the scheduler has already been
823 * initialized (some code might call cpu_clock() very early):
824 */
Ingo Molnar6892b752008-02-13 14:02:36 +0100825 if (unlikely(!scheduler_running))
826 return 0;
827
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200828 now = sched_clock_cpu(cpu);
Ingo Molnare436d802007-07-19 21:28:35 +0200829
830 return now;
831}
Ingo Molnar27ec4402008-02-28 21:00:21 +0100832
833/*
834 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
835 * clock constructed from sched_clock():
836 */
837unsigned long long cpu_clock(int cpu)
838{
839 unsigned long long prev_cpu_time, time, delta_time;
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200840 unsigned long flags;
Ingo Molnar27ec4402008-02-28 21:00:21 +0100841
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200842 local_irq_save(flags);
Ingo Molnar27ec4402008-02-28 21:00:21 +0100843 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
844 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
845 delta_time = time-prev_cpu_time;
846
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200847 if (unlikely(delta_time > time_sync_thresh)) {
Ingo Molnar27ec4402008-02-28 21:00:21 +0100848 time = __sync_cpu_clock(time, cpu);
Ingo Molnardfbf4a12008-04-23 09:24:06 +0200849 per_cpu(prev_cpu_time, cpu) = time;
850 }
851 local_irq_restore(flags);
Ingo Molnar27ec4402008-02-28 21:00:21 +0100852
853 return time;
854}
Paul E. McKenneya58f6f22007-10-15 17:00:14 +0200855EXPORT_SYMBOL_GPL(cpu_clock);
Ingo Molnare436d802007-07-19 21:28:35 +0200856
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700858# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700860#ifndef finish_arch_switch
861# define finish_arch_switch(prev) do { } while (0)
862#endif
863
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100864static inline int task_current(struct rq *rq, struct task_struct *p)
865{
866 return rq->curr == p;
867}
868
Nick Piggin4866cde2005-06-25 14:57:23 -0700869#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700870static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700871{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100872 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700873}
874
Ingo Molnar70b97a72006-07-03 00:25:42 -0700875static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700876{
877}
878
Ingo Molnar70b97a72006-07-03 00:25:42 -0700879static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700880{
Ingo Molnarda04c032005-09-13 11:17:59 +0200881#ifdef CONFIG_DEBUG_SPINLOCK
882 /* this is a valid case when another task releases the spinlock */
883 rq->lock.owner = current;
884#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700885 /*
886 * If we are tracking spinlock dependencies then we have to
887 * fix up the runqueue lock - which gets 'carried over' from
888 * prev into current:
889 */
890 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
891
Nick Piggin4866cde2005-06-25 14:57:23 -0700892 spin_unlock_irq(&rq->lock);
893}
894
895#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700896static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700897{
898#ifdef CONFIG_SMP
899 return p->oncpu;
900#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100901 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700902#endif
903}
904
Ingo Molnar70b97a72006-07-03 00:25:42 -0700905static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700906{
907#ifdef CONFIG_SMP
908 /*
909 * We can optimise this out completely for !SMP, because the
910 * SMP rebalancing from interrupt is the only thing that cares
911 * here.
912 */
913 next->oncpu = 1;
914#endif
915#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
916 spin_unlock_irq(&rq->lock);
917#else
918 spin_unlock(&rq->lock);
919#endif
920}
921
Ingo Molnar70b97a72006-07-03 00:25:42 -0700922static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700923{
924#ifdef CONFIG_SMP
925 /*
926 * After ->oncpu is cleared, the task can be moved to a different CPU.
927 * We must ensure this doesn't happen until the switch is completely
928 * finished.
929 */
930 smp_wmb();
931 prev->oncpu = 0;
932#endif
933#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
934 local_irq_enable();
935#endif
936}
937#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938
939/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700940 * __task_rq_lock - lock the runqueue a given task resides on.
941 * Must be called interrupts disabled.
942 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700943static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700944 __acquires(rq->lock)
945{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200946 for (;;) {
947 struct rq *rq = task_rq(p);
948 spin_lock(&rq->lock);
949 if (likely(rq == task_rq(p)))
950 return rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700951 spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700952 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700953}
954
955/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100957 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 * explicitly disabling preemption.
959 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700960static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 __acquires(rq->lock)
962{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700963 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964
Andi Kleen3a5c3592007-10-15 17:00:14 +0200965 for (;;) {
966 local_irq_save(*flags);
967 rq = task_rq(p);
968 spin_lock(&rq->lock);
969 if (likely(rq == task_rq(p)))
970 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973}
974
Alexey Dobriyana9957442007-10-15 17:00:13 +0200975static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700976 __releases(rq->lock)
977{
978 spin_unlock(&rq->lock);
979}
980
Ingo Molnar70b97a72006-07-03 00:25:42 -0700981static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 __releases(rq->lock)
983{
984 spin_unlock_irqrestore(&rq->lock, *flags);
985}
986
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -0800988 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 */
Alexey Dobriyana9957442007-10-15 17:00:13 +0200990static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 __acquires(rq->lock)
992{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700993 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994
995 local_irq_disable();
996 rq = this_rq();
997 spin_lock(&rq->lock);
998
999 return rq;
1000}
1001
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001002static void __resched_task(struct task_struct *p, int tif_bit);
1003
1004static inline void resched_task(struct task_struct *p)
1005{
1006 __resched_task(p, TIF_NEED_RESCHED);
1007}
1008
1009#ifdef CONFIG_SCHED_HRTICK
1010/*
1011 * Use HR-timers to deliver accurate preemption points.
1012 *
1013 * Its all a bit involved since we cannot program an hrt while holding the
1014 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1015 * reschedule event.
1016 *
1017 * When we get rescheduled we reprogram the hrtick_timer outside of the
1018 * rq->lock.
1019 */
1020static inline void resched_hrt(struct task_struct *p)
1021{
1022 __resched_task(p, TIF_HRTICK_RESCHED);
1023}
1024
1025static inline void resched_rq(struct rq *rq)
1026{
1027 unsigned long flags;
1028
1029 spin_lock_irqsave(&rq->lock, flags);
1030 resched_task(rq->curr);
1031 spin_unlock_irqrestore(&rq->lock, flags);
1032}
1033
1034enum {
1035 HRTICK_SET, /* re-programm hrtick_timer */
1036 HRTICK_RESET, /* not a new slice */
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001037 HRTICK_BLOCK, /* stop hrtick operations */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001038};
1039
1040/*
1041 * Use hrtick when:
1042 * - enabled by features
1043 * - hrtimer is actually high res
1044 */
1045static inline int hrtick_enabled(struct rq *rq)
1046{
1047 if (!sched_feat(HRTICK))
1048 return 0;
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001049 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1050 return 0;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001051 return hrtimer_is_hres_active(&rq->hrtick_timer);
1052}
1053
1054/*
1055 * Called to set the hrtick timer state.
1056 *
1057 * called with rq->lock held and irqs disabled
1058 */
1059static void hrtick_start(struct rq *rq, u64 delay, int reset)
1060{
1061 assert_spin_locked(&rq->lock);
1062
1063 /*
1064 * preempt at: now + delay
1065 */
1066 rq->hrtick_expire =
1067 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1068 /*
1069 * indicate we need to program the timer
1070 */
1071 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1072 if (reset)
1073 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1074
1075 /*
1076 * New slices are called from the schedule path and don't need a
1077 * forced reschedule.
1078 */
1079 if (reset)
1080 resched_hrt(rq->curr);
1081}
1082
1083static void hrtick_clear(struct rq *rq)
1084{
1085 if (hrtimer_active(&rq->hrtick_timer))
1086 hrtimer_cancel(&rq->hrtick_timer);
1087}
1088
1089/*
1090 * Update the timer from the possible pending state.
1091 */
1092static void hrtick_set(struct rq *rq)
1093{
1094 ktime_t time;
1095 int set, reset;
1096 unsigned long flags;
1097
1098 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1099
1100 spin_lock_irqsave(&rq->lock, flags);
1101 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1102 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1103 time = rq->hrtick_expire;
1104 clear_thread_flag(TIF_HRTICK_RESCHED);
1105 spin_unlock_irqrestore(&rq->lock, flags);
1106
1107 if (set) {
1108 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1109 if (reset && !hrtimer_active(&rq->hrtick_timer))
1110 resched_rq(rq);
1111 } else
1112 hrtick_clear(rq);
1113}
1114
1115/*
1116 * High-resolution timer tick.
1117 * Runs from hardirq context with interrupts disabled.
1118 */
1119static enum hrtimer_restart hrtick(struct hrtimer *timer)
1120{
1121 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1122
1123 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1124
1125 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02001126 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001127 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1128 spin_unlock(&rq->lock);
1129
1130 return HRTIMER_NORESTART;
1131}
1132
Rabin Vincent81d41d72008-05-11 05:55:33 +05301133#ifdef CONFIG_SMP
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001134static void hotplug_hrtick_disable(int cpu)
1135{
1136 struct rq *rq = cpu_rq(cpu);
1137 unsigned long flags;
1138
1139 spin_lock_irqsave(&rq->lock, flags);
1140 rq->hrtick_flags = 0;
1141 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1142 spin_unlock_irqrestore(&rq->lock, flags);
1143
1144 hrtick_clear(rq);
1145}
1146
1147static void hotplug_hrtick_enable(int cpu)
1148{
1149 struct rq *rq = cpu_rq(cpu);
1150 unsigned long flags;
1151
1152 spin_lock_irqsave(&rq->lock, flags);
1153 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1154 spin_unlock_irqrestore(&rq->lock, flags);
1155}
1156
1157static int
1158hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1159{
1160 int cpu = (int)(long)hcpu;
1161
1162 switch (action) {
1163 case CPU_UP_CANCELED:
1164 case CPU_UP_CANCELED_FROZEN:
1165 case CPU_DOWN_PREPARE:
1166 case CPU_DOWN_PREPARE_FROZEN:
1167 case CPU_DEAD:
1168 case CPU_DEAD_FROZEN:
1169 hotplug_hrtick_disable(cpu);
1170 return NOTIFY_OK;
1171
1172 case CPU_UP_PREPARE:
1173 case CPU_UP_PREPARE_FROZEN:
1174 case CPU_DOWN_FAILED:
1175 case CPU_DOWN_FAILED_FROZEN:
1176 case CPU_ONLINE:
1177 case CPU_ONLINE_FROZEN:
1178 hotplug_hrtick_enable(cpu);
1179 return NOTIFY_OK;
1180 }
1181
1182 return NOTIFY_DONE;
1183}
1184
1185static void init_hrtick(void)
1186{
1187 hotcpu_notifier(hotplug_hrtick, 0);
1188}
Rabin Vincent81d41d72008-05-11 05:55:33 +05301189#endif /* CONFIG_SMP */
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001190
1191static void init_rq_hrtick(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001192{
1193 rq->hrtick_flags = 0;
1194 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1195 rq->hrtick_timer.function = hrtick;
1196 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1197}
1198
1199void hrtick_resched(void)
1200{
1201 struct rq *rq;
1202 unsigned long flags;
1203
1204 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1205 return;
1206
1207 local_irq_save(flags);
1208 rq = cpu_rq(smp_processor_id());
1209 hrtick_set(rq);
1210 local_irq_restore(flags);
1211}
1212#else
1213static inline void hrtick_clear(struct rq *rq)
1214{
1215}
1216
1217static inline void hrtick_set(struct rq *rq)
1218{
1219}
1220
1221static inline void init_rq_hrtick(struct rq *rq)
1222{
1223}
1224
1225void hrtick_resched(void)
1226{
1227}
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001228
1229static inline void init_hrtick(void)
1230{
1231}
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001232#endif
1233
Ingo Molnar1b9f19c2007-07-09 18:51:59 +02001234/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001235 * resched_task - mark a task 'to be rescheduled now'.
1236 *
1237 * On UP this means the setting of the need_resched flag, on SMP it
1238 * might also involve a cross-CPU call to trigger the scheduler on
1239 * the target CPU.
1240 */
1241#ifdef CONFIG_SMP
1242
1243#ifndef tsk_is_polling
1244#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1245#endif
1246
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001247static void __resched_task(struct task_struct *p, int tif_bit)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001248{
1249 int cpu;
1250
1251 assert_spin_locked(&task_rq(p)->lock);
1252
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001253 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001254 return;
1255
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001256 set_tsk_thread_flag(p, tif_bit);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001257
1258 cpu = task_cpu(p);
1259 if (cpu == smp_processor_id())
1260 return;
1261
1262 /* NEED_RESCHED must be visible before we test polling */
1263 smp_mb();
1264 if (!tsk_is_polling(p))
1265 smp_send_reschedule(cpu);
1266}
1267
1268static void resched_cpu(int cpu)
1269{
1270 struct rq *rq = cpu_rq(cpu);
1271 unsigned long flags;
1272
1273 if (!spin_trylock_irqsave(&rq->lock, flags))
1274 return;
1275 resched_task(cpu_curr(cpu));
1276 spin_unlock_irqrestore(&rq->lock, flags);
1277}
Thomas Gleixner06d83082008-03-22 09:20:24 +01001278
1279#ifdef CONFIG_NO_HZ
1280/*
1281 * When add_timer_on() enqueues a timer into the timer wheel of an
1282 * idle CPU then this timer might expire before the next timer event
1283 * which is scheduled to wake up that CPU. In case of a completely
1284 * idle system the next event might even be infinite time into the
1285 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1286 * leaves the inner idle loop so the newly added timer is taken into
1287 * account when the CPU goes back to idle and evaluates the timer
1288 * wheel for the next timer event.
1289 */
1290void wake_up_idle_cpu(int cpu)
1291{
1292 struct rq *rq = cpu_rq(cpu);
1293
1294 if (cpu == smp_processor_id())
1295 return;
1296
1297 /*
1298 * This is safe, as this function is called with the timer
1299 * wheel base lock of (cpu) held. When the CPU is on the way
1300 * to idle and has not yet set rq->curr to idle then it will
1301 * be serialized on the timer wheel base lock and take the new
1302 * timer into account automatically.
1303 */
1304 if (rq->curr != rq->idle)
1305 return;
1306
1307 /*
1308 * We can set TIF_RESCHED on the idle task of the other CPU
1309 * lockless. The worst case is that the other CPU runs the
1310 * idle task through an additional NOOP schedule()
1311 */
1312 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1313
1314 /* NEED_RESCHED must be visible before we test polling */
1315 smp_mb();
1316 if (!tsk_is_polling(rq->idle))
1317 smp_send_reschedule(cpu);
1318}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001319#endif /* CONFIG_NO_HZ */
Thomas Gleixner06d83082008-03-22 09:20:24 +01001320
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001321#else /* !CONFIG_SMP */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001322static void __resched_task(struct task_struct *p, int tif_bit)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001323{
1324 assert_spin_locked(&task_rq(p)->lock);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001325 set_tsk_thread_flag(p, tif_bit);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001326}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001327#endif /* CONFIG_SMP */
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001328
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001329#if BITS_PER_LONG == 32
1330# define WMULT_CONST (~0UL)
1331#else
1332# define WMULT_CONST (1UL << 32)
1333#endif
1334
1335#define WMULT_SHIFT 32
1336
Ingo Molnar194081e2007-08-09 11:16:51 +02001337/*
1338 * Shift right and round:
1339 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001340#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +02001341
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +02001342static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001343calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1344 struct load_weight *lw)
1345{
1346 u64 tmp;
1347
Peter Zijlstrae05510d2008-05-05 23:56:17 +02001348 if (!lw->inv_weight)
1349 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001350
1351 tmp = (u64)delta_exec * weight;
1352 /*
1353 * Check whether we'd overflow the 64-bit multiplication:
1354 */
Ingo Molnar194081e2007-08-09 11:16:51 +02001355 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001356 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +02001357 WMULT_SHIFT/2);
1358 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001359 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001360
Ingo Molnarecf691d2007-08-02 17:41:40 +02001361 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001362}
1363
Ingo Molnarf9305d42008-05-29 11:23:17 +02001364static inline unsigned long
1365calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1366{
1367 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1368}
1369
Ingo Molnar10919852007-10-15 17:00:04 +02001370static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001371{
1372 lw->weight += inc;
Ingo Molnare89996a2008-03-14 23:48:28 +01001373 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001374}
1375
Ingo Molnar10919852007-10-15 17:00:04 +02001376static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001377{
1378 lw->weight -= dec;
Ingo Molnare89996a2008-03-14 23:48:28 +01001379 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001380}
1381
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001383 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1384 * of tasks with abnormal "nice" values across CPUs the contribution that
1385 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001386 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -07001387 * scaled version of the new time slice allocation that they receive on time
1388 * slice expiry etc.
1389 */
1390
Ingo Molnardd41f592007-07-09 18:51:59 +02001391#define WEIGHT_IDLEPRIO 2
1392#define WMULT_IDLEPRIO (1 << 31)
1393
1394/*
1395 * Nice levels are multiplicative, with a gentle 10% change for every
1396 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1397 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1398 * that remained on nice 0.
1399 *
1400 * The "10% effect" is relative and cumulative: from _any_ nice level,
1401 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +02001402 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1403 * If a task goes up by ~10% and another task goes down by ~10% then
1404 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +02001405 */
1406static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001407 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1408 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1409 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1410 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1411 /* 0 */ 1024, 820, 655, 526, 423,
1412 /* 5 */ 335, 272, 215, 172, 137,
1413 /* 10 */ 110, 87, 70, 56, 45,
1414 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +02001415};
1416
Ingo Molnar5714d2d2007-07-16 09:46:31 +02001417/*
1418 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1419 *
1420 * In cases where the weight does not change often, we can use the
1421 * precalculated inverse to speed up arithmetics by turning divisions
1422 * into multiplications:
1423 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001424static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001425 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1426 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1427 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1428 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1429 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1430 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1431 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1432 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +02001433};
Peter Williams2dd73a42006-06-27 02:54:34 -07001434
Ingo Molnardd41f592007-07-09 18:51:59 +02001435static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1436
1437/*
1438 * runqueue iterator, to support SMP load-balancing between different
1439 * scheduling classes, without having to expose their internal data
1440 * structures to the load-balancing proper:
1441 */
1442struct rq_iterator {
1443 void *arg;
1444 struct task_struct *(*start)(void *);
1445 struct task_struct *(*next)(void *);
1446};
1447
Peter Williamse1d14842007-10-24 18:23:51 +02001448#ifdef CONFIG_SMP
1449static unsigned long
1450balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1451 unsigned long max_load_move, struct sched_domain *sd,
1452 enum cpu_idle_type idle, int *all_pinned,
1453 int *this_best_prio, struct rq_iterator *iterator);
1454
1455static int
1456iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1457 struct sched_domain *sd, enum cpu_idle_type idle,
1458 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +02001459#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001460
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001461#ifdef CONFIG_CGROUP_CPUACCT
1462static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1463#else
1464static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1465#endif
1466
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001467static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1468{
1469 update_load_add(&rq->load, load);
1470}
1471
1472static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1473{
1474 update_load_sub(&rq->load, load);
1475}
1476
Gregory Haskinse7693a32008-01-25 21:08:09 +01001477#ifdef CONFIG_SMP
1478static unsigned long source_load(int cpu, int type);
1479static unsigned long target_load(int cpu, int type);
1480static unsigned long cpu_avg_load_per_task(int cpu);
1481static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001482#else /* CONFIG_SMP */
1483
1484#ifdef CONFIG_FAIR_GROUP_SCHED
1485static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1486{
1487}
1488#endif
1489
Gregory Haskinse7693a32008-01-25 21:08:09 +01001490#endif /* CONFIG_SMP */
1491
Ingo Molnardd41f592007-07-09 18:51:59 +02001492#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +02001493#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001494#include "sched_fair.c"
1495#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +02001496#ifdef CONFIG_SCHED_DEBUG
1497# include "sched_debug.c"
1498#endif
1499
1500#define sched_class_highest (&rt_sched_class)
1501
Ingo Molnar6363ca52008-05-29 11:28:57 +02001502static inline void inc_load(struct rq *rq, const struct task_struct *p)
Peter Zijlstra62fb1852008-02-25 17:34:02 +01001503{
Ingo Molnar6363ca52008-05-29 11:28:57 +02001504 update_load_add(&rq->load, p->se.load.weight);
Peter Zijlstra62fb1852008-02-25 17:34:02 +01001505}
1506
Ingo Molnar6363ca52008-05-29 11:28:57 +02001507static inline void dec_load(struct rq *rq, const struct task_struct *p)
1508{
1509 update_load_sub(&rq->load, p->se.load.weight);
1510}
1511
1512static void inc_nr_running(struct task_struct *p, struct rq *rq)
1513{
1514 rq->nr_running++;
1515 inc_load(rq, p);
1516}
1517
1518static void dec_nr_running(struct task_struct *p, struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001519{
1520 rq->nr_running--;
Ingo Molnar6363ca52008-05-29 11:28:57 +02001521 dec_load(rq, p);
Ingo Molnar9c217242007-08-02 17:41:40 +02001522}
1523
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001524static void set_load_weight(struct task_struct *p)
1525{
1526 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001527 p->se.load.weight = prio_to_weight[0] * 2;
1528 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1529 return;
1530 }
1531
1532 /*
1533 * SCHED_IDLE tasks get minimal weight:
1534 */
1535 if (p->policy == SCHED_IDLE) {
1536 p->se.load.weight = WEIGHT_IDLEPRIO;
1537 p->se.load.inv_weight = WMULT_IDLEPRIO;
1538 return;
1539 }
1540
1541 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1542 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001543}
1544
Ingo Molnar8159f872007-08-09 11:16:49 +02001545static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001546{
1547 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001548 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001549 p->se.on_rq = 1;
1550}
1551
Ingo Molnar69be72c2007-08-09 11:16:49 +02001552static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001553{
Ingo Molnarf02231e2007-08-09 11:16:48 +02001554 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001555 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001556}
1557
1558/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001559 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001560 */
Ingo Molnar14531182007-07-09 18:51:59 +02001561static inline int __normal_prio(struct task_struct *p)
1562{
Ingo Molnardd41f592007-07-09 18:51:59 +02001563 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001564}
1565
1566/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001567 * Calculate the expected normal priority: i.e. priority
1568 * without taking RT-inheritance into account. Might be
1569 * boosted by interactivity modifiers. Changes upon fork,
1570 * setprio syscalls, and whenever the interactivity
1571 * estimator recalculates.
1572 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001573static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001574{
1575 int prio;
1576
Ingo Molnare05606d2007-07-09 18:51:59 +02001577 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001578 prio = MAX_RT_PRIO-1 - p->rt_priority;
1579 else
1580 prio = __normal_prio(p);
1581 return prio;
1582}
1583
1584/*
1585 * Calculate the current priority, i.e. the priority
1586 * taken into account by the scheduler. This value might
1587 * be boosted by RT tasks, or might be boosted by
1588 * interactivity modifiers. Will be RT if the task got
1589 * RT-boosted. If not then it returns p->normal_prio.
1590 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001591static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001592{
1593 p->normal_prio = normal_prio(p);
1594 /*
1595 * If we are RT tasks or we were boosted to RT priority,
1596 * keep the priority unchanged. Otherwise, update priority
1597 * to the normal priority:
1598 */
1599 if (!rt_prio(p->prio))
1600 return p->normal_prio;
1601 return p->prio;
1602}
1603
1604/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001605 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001607static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001609 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001610 rq->nr_uninterruptible--;
1611
Ingo Molnar8159f872007-08-09 11:16:49 +02001612 enqueue_task(rq, p, wakeup);
Ingo Molnar6363ca52008-05-29 11:28:57 +02001613 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614}
1615
1616/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 * deactivate_task - remove a task from the runqueue.
1618 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001619static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001621 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001622 rq->nr_uninterruptible++;
1623
Ingo Molnar69be72c2007-08-09 11:16:49 +02001624 dequeue_task(rq, p, sleep);
Ingo Molnar6363ca52008-05-29 11:28:57 +02001625 dec_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626}
1627
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628/**
1629 * task_curr - is this task currently executing on a CPU?
1630 * @p: the task in question.
1631 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001632inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633{
1634 return cpu_curr(task_cpu(p)) == p;
1635}
1636
Peter Williams2dd73a42006-06-27 02:54:34 -07001637/* Used instead of source_load when we know the type == 0 */
Thomas Gleixnerf7dcd802008-05-24 23:20:38 +02001638static unsigned long weighted_cpuload(const int cpu)
Peter Williams2dd73a42006-06-27 02:54:34 -07001639{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02001640 return cpu_rq(cpu)->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02001641}
1642
1643static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1644{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001645 set_task_rq(p, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001646#ifdef CONFIG_SMP
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001647 /*
1648 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1649 * successfuly executed on another CPU. We must ensure that updates of
1650 * per-task data have been completed by this moment.
1651 */
1652 smp_wmb();
Ingo Molnardd41f592007-07-09 18:51:59 +02001653 task_thread_info(p)->cpu = cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02001654#endif
Peter Williams2dd73a42006-06-27 02:54:34 -07001655}
1656
Steven Rostedtcb469842008-01-25 21:08:22 +01001657static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1658 const struct sched_class *prev_class,
1659 int oldprio, int running)
1660{
1661 if (prev_class != p->sched_class) {
1662 if (prev_class->switched_from)
1663 prev_class->switched_from(rq, p, running);
1664 p->sched_class->switched_to(rq, p, running);
1665 } else
1666 p->sched_class->prio_changed(rq, p, oldprio, running);
1667}
1668
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669#ifdef CONFIG_SMP
Ingo Molnarc65cc872007-07-09 18:51:58 +02001670
Ingo Molnarcc367732007-10-15 17:00:18 +02001671/*
1672 * Is this task likely cache-hot:
1673 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01001674static int
Ingo Molnarcc367732007-10-15 17:00:18 +02001675task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1676{
1677 s64 delta;
1678
Ingo Molnarf540a602008-03-15 17:10:34 +01001679 /*
1680 * Buddy candidates are cache hot:
1681 */
Ingo Molnard25ce4c2008-03-17 09:36:53 +01001682 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
Ingo Molnarf540a602008-03-15 17:10:34 +01001683 return 1;
1684
Ingo Molnarcc367732007-10-15 17:00:18 +02001685 if (p->sched_class != &fair_sched_class)
1686 return 0;
1687
Ingo Molnar6bc16652007-10-15 17:00:18 +02001688 if (sysctl_sched_migration_cost == -1)
1689 return 1;
1690 if (sysctl_sched_migration_cost == 0)
1691 return 0;
1692
Ingo Molnarcc367732007-10-15 17:00:18 +02001693 delta = now - p->se.exec_start;
1694
1695 return delta < (s64)sysctl_sched_migration_cost;
1696}
1697
1698
Ingo Molnardd41f592007-07-09 18:51:59 +02001699void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02001700{
Ingo Molnardd41f592007-07-09 18:51:59 +02001701 int old_cpu = task_cpu(p);
1702 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001703 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1704 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnarbbdba7c2007-10-15 17:00:06 +02001705 u64 clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001706
1707 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001708
1709#ifdef CONFIG_SCHEDSTATS
1710 if (p->se.wait_start)
1711 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001712 if (p->se.sleep_start)
1713 p->se.sleep_start -= clock_offset;
1714 if (p->se.block_start)
1715 p->se.block_start -= clock_offset;
Ingo Molnarcc367732007-10-15 17:00:18 +02001716 if (old_cpu != new_cpu) {
1717 schedstat_inc(p, se.nr_migrations);
1718 if (task_hot(p, old_rq->clock, NULL))
1719 schedstat_inc(p, se.nr_forced2_migrations);
1720 }
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001721#endif
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001722 p->se.vruntime -= old_cfsrq->min_vruntime -
1723 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02001724
1725 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02001726}
1727
Ingo Molnar70b97a72006-07-03 00:25:42 -07001728struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730
Ingo Molnar36c8b582006-07-03 00:25:41 -07001731 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 int dest_cpu;
1733
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001735};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736
1737/*
1738 * The task's runqueue lock must be held.
1739 * Returns true if you have to wait for migration thread.
1740 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001741static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07001742migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001744 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745
1746 /*
1747 * If the task is not on a runqueue (and not running), then
1748 * it is sufficient to simply update the task's cpu field.
1749 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001750 if (!p->se.on_rq && !task_running(rq, p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 set_task_cpu(p, dest_cpu);
1752 return 0;
1753 }
1754
1755 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 req->task = p;
1757 req->dest_cpu = dest_cpu;
1758 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07001759
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 return 1;
1761}
1762
1763/*
1764 * wait_task_inactive - wait for a thread to unschedule.
1765 *
1766 * The caller must ensure that the task *will* unschedule sometime soon,
1767 * else this function might spin for a *long* time. This function can't
1768 * be called with interrupts off, or it may introduce deadlock with
1769 * smp_call_function() if an IPI is sent by the same process we are
1770 * waiting to become inactive.
1771 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001772void wait_task_inactive(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773{
1774 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02001775 int running, on_rq;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001776 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777
Andi Kleen3a5c3592007-10-15 17:00:14 +02001778 for (;;) {
1779 /*
1780 * We do the initial early heuristics without holding
1781 * any task-queue locks at all. We'll only try to get
1782 * the runqueue lock when things look like they will
1783 * work out!
1784 */
1785 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001786
Andi Kleen3a5c3592007-10-15 17:00:14 +02001787 /*
1788 * If the task is actively running on another CPU
1789 * still, just relax and busy-wait without holding
1790 * any locks.
1791 *
1792 * NOTE! Since we don't hold any locks, it's not
1793 * even sure that "rq" stays as the right runqueue!
1794 * But we don't care, since "task_running()" will
1795 * return false if the runqueue has changed and p
1796 * is actually now running somewhere else!
1797 */
1798 while (task_running(rq, p))
1799 cpu_relax();
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001800
Andi Kleen3a5c3592007-10-15 17:00:14 +02001801 /*
1802 * Ok, time to look more closely! We need the rq
1803 * lock now, to be *sure*. If we're wrong, we'll
1804 * just go back and repeat.
1805 */
1806 rq = task_rq_lock(p, &flags);
1807 running = task_running(rq, p);
1808 on_rq = p->se.on_rq;
1809 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001810
Andi Kleen3a5c3592007-10-15 17:00:14 +02001811 /*
1812 * Was it really running after all now that we
1813 * checked with the proper locks actually held?
1814 *
1815 * Oops. Go back and try again..
1816 */
1817 if (unlikely(running)) {
1818 cpu_relax();
1819 continue;
1820 }
1821
1822 /*
1823 * It's not enough that it's not actively running,
1824 * it must be off the runqueue _entirely_, and not
1825 * preempted!
1826 *
1827 * So if it wa still runnable (but just not actively
1828 * running right now), it's preempted, and we should
1829 * yield - it could be a while.
1830 */
1831 if (unlikely(on_rq)) {
1832 schedule_timeout_uninterruptible(1);
1833 continue;
1834 }
1835
1836 /*
1837 * Ahh, all good. It wasn't running, and it wasn't
1838 * runnable, which means that it will never become
1839 * running in the future either. We're all done!
1840 */
1841 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843}
1844
1845/***
1846 * kick_process - kick a running thread to enter/exit the kernel
1847 * @p: the to-be-kicked thread
1848 *
1849 * Cause a process which is running on another CPU to enter
1850 * kernel-mode, without any delay. (to get signals handled.)
1851 *
1852 * NOTE: this function doesnt have to take the runqueue lock,
1853 * because all it wants to ensure is that the remote task enters
1854 * the kernel. If the IPI races and the task has been migrated
1855 * to another CPU then no harm is done and the purpose has been
1856 * achieved as well.
1857 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001858void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859{
1860 int cpu;
1861
1862 preempt_disable();
1863 cpu = task_cpu(p);
1864 if ((cpu != smp_processor_id()) && task_curr(p))
1865 smp_send_reschedule(cpu);
1866 preempt_enable();
1867}
1868
1869/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001870 * Return a low guess at the load of a migration-source cpu weighted
1871 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 *
1873 * We want to under-estimate the load of migration sources, to
1874 * balance conservatively.
1875 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001876static unsigned long source_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08001877{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001878 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001879 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001880
Peter Williams2dd73a42006-06-27 02:54:34 -07001881 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02001882 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07001883
Ingo Molnardd41f592007-07-09 18:51:59 +02001884 return min(rq->cpu_load[type-1], total);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885}
1886
1887/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001888 * Return a high guess at the load of a migration-target cpu weighted
1889 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001891static unsigned long target_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08001892{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001893 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001894 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001895
Peter Williams2dd73a42006-06-27 02:54:34 -07001896 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02001897 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07001898
Ingo Molnardd41f592007-07-09 18:51:59 +02001899 return max(rq->cpu_load[type-1], total);
Peter Williams2dd73a42006-06-27 02:54:34 -07001900}
1901
1902/*
1903 * Return the average load per task on the cpu's run queue
1904 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01001905static unsigned long cpu_avg_load_per_task(int cpu)
Peter Williams2dd73a42006-06-27 02:54:34 -07001906{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001907 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001908 unsigned long total = weighted_cpuload(cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07001909 unsigned long n = rq->nr_running;
1910
Ingo Molnardd41f592007-07-09 18:51:59 +02001911 return n ? total / n : SCHED_LOAD_SCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912}
1913
Nick Piggin147cbb42005-06-25 14:57:19 -07001914/*
1915 * find_idlest_group finds and returns the least busy CPU group within the
1916 * domain.
1917 */
1918static struct sched_group *
1919find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1920{
1921 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1922 unsigned long min_load = ULONG_MAX, this_load = 0;
1923 int load_idx = sd->forkexec_idx;
1924 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1925
1926 do {
1927 unsigned long load, avg_load;
1928 int local_group;
1929 int i;
1930
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001931 /* Skip over this group if it has no CPUs allowed */
1932 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
Andi Kleen3a5c3592007-10-15 17:00:14 +02001933 continue;
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001934
Nick Piggin147cbb42005-06-25 14:57:19 -07001935 local_group = cpu_isset(this_cpu, group->cpumask);
Nick Piggin147cbb42005-06-25 14:57:19 -07001936
1937 /* Tally up the load of all CPUs in the group */
1938 avg_load = 0;
1939
1940 for_each_cpu_mask(i, group->cpumask) {
1941 /* Bias balancing toward cpus of our domain */
1942 if (local_group)
1943 load = source_load(i, load_idx);
1944 else
1945 load = target_load(i, load_idx);
1946
1947 avg_load += load;
1948 }
1949
1950 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07001951 avg_load = sg_div_cpu_power(group,
1952 avg_load * SCHED_LOAD_SCALE);
Nick Piggin147cbb42005-06-25 14:57:19 -07001953
1954 if (local_group) {
1955 this_load = avg_load;
1956 this = group;
1957 } else if (avg_load < min_load) {
1958 min_load = avg_load;
1959 idlest = group;
1960 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02001961 } while (group = group->next, group != sd->groups);
Nick Piggin147cbb42005-06-25 14:57:19 -07001962
1963 if (!idlest || 100*this_load < imbalance*min_load)
1964 return NULL;
1965 return idlest;
1966}
1967
1968/*
Satoru Takeuchi0feaece2006-10-03 01:14:10 -07001969 * find_idlest_cpu - find the idlest cpu among the cpus in group.
Nick Piggin147cbb42005-06-25 14:57:19 -07001970 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07001971static int
Mike Travis7c16ec52008-04-04 18:11:11 -07001972find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1973 cpumask_t *tmp)
Nick Piggin147cbb42005-06-25 14:57:19 -07001974{
1975 unsigned long load, min_load = ULONG_MAX;
1976 int idlest = -1;
1977 int i;
1978
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001979 /* Traverse only the allowed CPUs */
Mike Travis7c16ec52008-04-04 18:11:11 -07001980 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001981
Mike Travis7c16ec52008-04-04 18:11:11 -07001982 for_each_cpu_mask(i, *tmp) {
Peter Williams2dd73a42006-06-27 02:54:34 -07001983 load = weighted_cpuload(i);
Nick Piggin147cbb42005-06-25 14:57:19 -07001984
1985 if (load < min_load || (load == min_load && i == this_cpu)) {
1986 min_load = load;
1987 idlest = i;
1988 }
1989 }
1990
1991 return idlest;
1992}
1993
Nick Piggin476d1392005-06-25 14:57:29 -07001994/*
1995 * sched_balance_self: balance the current task (running on cpu) in domains
1996 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1997 * SD_BALANCE_EXEC.
1998 *
1999 * Balance, ie. select the least loaded group.
2000 *
2001 * Returns the target CPU number, or the same CPU if no balancing is needed.
2002 *
2003 * preempt must be disabled.
2004 */
2005static int sched_balance_self(int cpu, int flag)
2006{
2007 struct task_struct *t = current;
2008 struct sched_domain *tmp, *sd = NULL;
Nick Piggin147cbb42005-06-25 14:57:19 -07002009
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002010 for_each_domain(cpu, tmp) {
Ingo Molnar9761eea2007-07-09 18:52:00 +02002011 /*
2012 * If power savings logic is enabled for a domain, stop there.
2013 */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002014 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2015 break;
Nick Piggin476d1392005-06-25 14:57:29 -07002016 if (tmp->flags & flag)
2017 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002018 }
Nick Piggin476d1392005-06-25 14:57:29 -07002019
2020 while (sd) {
Mike Travis7c16ec52008-04-04 18:11:11 -07002021 cpumask_t span, tmpmask;
Nick Piggin476d1392005-06-25 14:57:29 -07002022 struct sched_group *group;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002023 int new_cpu, weight;
2024
2025 if (!(sd->flags & flag)) {
2026 sd = sd->child;
2027 continue;
2028 }
Nick Piggin476d1392005-06-25 14:57:29 -07002029
2030 span = sd->span;
2031 group = find_idlest_group(sd, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002032 if (!group) {
2033 sd = sd->child;
2034 continue;
2035 }
Nick Piggin476d1392005-06-25 14:57:29 -07002036
Mike Travis7c16ec52008-04-04 18:11:11 -07002037 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002038 if (new_cpu == -1 || new_cpu == cpu) {
2039 /* Now try balancing at a lower domain level of cpu */
2040 sd = sd->child;
2041 continue;
2042 }
Nick Piggin476d1392005-06-25 14:57:29 -07002043
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002044 /* Now try balancing at a lower domain level of new_cpu */
Nick Piggin476d1392005-06-25 14:57:29 -07002045 cpu = new_cpu;
Nick Piggin476d1392005-06-25 14:57:29 -07002046 sd = NULL;
2047 weight = cpus_weight(span);
2048 for_each_domain(cpu, tmp) {
2049 if (weight <= cpus_weight(tmp->span))
2050 break;
2051 if (tmp->flags & flag)
2052 sd = tmp;
2053 }
2054 /* while loop will break here if sd == NULL */
2055 }
2056
2057 return cpu;
2058}
2059
2060#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062/***
2063 * try_to_wake_up - wake up a thread
2064 * @p: the to-be-woken-up thread
2065 * @state: the mask of task states that can be woken
2066 * @sync: do a synchronous wakeup?
2067 *
2068 * Put it on the run-queue if it's not already there. The "current"
2069 * thread is always on the run-queue (except when the actual
2070 * re-schedule is in progress), and as such you're allowed to do
2071 * the simpler "current->state = TASK_RUNNING" to mark yourself
2072 * runnable without the overhead of this.
2073 *
2074 * returns failure only if the task is already active.
2075 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002076static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077{
Ingo Molnarcc367732007-10-15 17:00:18 +02002078 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 unsigned long flags;
2080 long old_state;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002081 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082
Ingo Molnarb85d0662008-03-16 20:03:22 +01002083 if (!sched_feat(SYNC_WAKEUPS))
2084 sync = 0;
2085
Linus Torvalds04e2f172008-02-23 18:05:03 -08002086 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 rq = task_rq_lock(p, &flags);
2088 old_state = p->state;
2089 if (!(old_state & state))
2090 goto out;
2091
Ingo Molnardd41f592007-07-09 18:51:59 +02002092 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093 goto out_running;
2094
2095 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02002096 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 this_cpu = smp_processor_id();
2098
2099#ifdef CONFIG_SMP
2100 if (unlikely(task_running(rq, p)))
2101 goto out_activate;
2102
Dmitry Adamushko5d2f5a62008-01-25 21:08:21 +01002103 cpu = p->sched_class->select_task_rq(p, sync);
2104 if (cpu != orig_cpu) {
2105 set_task_cpu(p, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 task_rq_unlock(rq, &flags);
2107 /* might preempt at this point */
2108 rq = task_rq_lock(p, &flags);
2109 old_state = p->state;
2110 if (!(old_state & state))
2111 goto out;
Ingo Molnardd41f592007-07-09 18:51:59 +02002112 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 goto out_running;
2114
2115 this_cpu = smp_processor_id();
2116 cpu = task_cpu(p);
2117 }
2118
Gregory Haskinse7693a32008-01-25 21:08:09 +01002119#ifdef CONFIG_SCHEDSTATS
2120 schedstat_inc(rq, ttwu_count);
2121 if (cpu == this_cpu)
2122 schedstat_inc(rq, ttwu_local);
2123 else {
2124 struct sched_domain *sd;
2125 for_each_domain(this_cpu, sd) {
2126 if (cpu_isset(cpu, sd->span)) {
2127 schedstat_inc(sd, ttwu_wake_remote);
2128 break;
2129 }
2130 }
2131 }
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002132#endif /* CONFIG_SCHEDSTATS */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002133
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134out_activate:
2135#endif /* CONFIG_SMP */
Ingo Molnarcc367732007-10-15 17:00:18 +02002136 schedstat_inc(p, se.nr_wakeups);
2137 if (sync)
2138 schedstat_inc(p, se.nr_wakeups_sync);
2139 if (orig_cpu != cpu)
2140 schedstat_inc(p, se.nr_wakeups_migrate);
2141 if (cpu == this_cpu)
2142 schedstat_inc(p, se.nr_wakeups_local);
2143 else
2144 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnar2daa3572007-08-09 11:16:51 +02002145 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02002146 activate_task(rq, p, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 success = 1;
2148
2149out_running:
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002150 check_preempt_curr(rq, p);
2151
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 p->state = TASK_RUNNING;
Steven Rostedt9a897c52008-01-25 21:08:22 +01002153#ifdef CONFIG_SMP
2154 if (p->sched_class->task_wake_up)
2155 p->sched_class->task_wake_up(rq, p);
2156#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157out:
2158 task_rq_unlock(rq, &flags);
2159
2160 return success;
2161}
2162
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002163int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05002165 return try_to_wake_up(p, TASK_ALL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167EXPORT_SYMBOL(wake_up_process);
2168
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002169int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170{
2171 return try_to_wake_up(p, state, 0);
2172}
2173
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174/*
2175 * Perform scheduler related setup for a newly forked process p.
2176 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02002177 *
2178 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002180static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181{
Ingo Molnardd41f592007-07-09 18:51:59 +02002182 p->se.exec_start = 0;
2183 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02002184 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002185 p->se.last_wakeup = 0;
2186 p->se.avg_overlap = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002187
2188#ifdef CONFIG_SCHEDSTATS
2189 p->se.wait_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002190 p->se.sum_sleep_runtime = 0;
2191 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002192 p->se.block_start = 0;
2193 p->se.sleep_max = 0;
2194 p->se.block_max = 0;
2195 p->se.exec_max = 0;
Ingo Molnareba1ed42007-10-15 17:00:02 +02002196 p->se.slice_max = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002197 p->se.wait_max = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002198#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002199
Peter Zijlstrafa717062008-01-25 21:08:27 +01002200 INIT_LIST_HEAD(&p->rt.run_list);
Ingo Molnardd41f592007-07-09 18:51:59 +02002201 p->se.on_rq = 0;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02002202 INIT_LIST_HEAD(&p->se.group_node);
Nick Piggin476d1392005-06-25 14:57:29 -07002203
Avi Kivitye107be32007-07-26 13:40:43 +02002204#ifdef CONFIG_PREEMPT_NOTIFIERS
2205 INIT_HLIST_HEAD(&p->preempt_notifiers);
2206#endif
2207
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 /*
2209 * We mark the process as running here, but have not actually
2210 * inserted it onto the runqueue yet. This guarantees that
2211 * nobody will actually run it, and a signal or other external
2212 * event cannot wake it up and insert it on the runqueue either.
2213 */
2214 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02002215}
2216
2217/*
2218 * fork()/clone()-time setup:
2219 */
2220void sched_fork(struct task_struct *p, int clone_flags)
2221{
2222 int cpu = get_cpu();
2223
2224 __sched_fork(p);
2225
2226#ifdef CONFIG_SMP
2227 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2228#endif
Ingo Molnar02e4bac2007-10-15 17:00:11 +02002229 set_task_cpu(p, cpu);
Ingo Molnarb29739f2006-06-27 02:54:51 -07002230
2231 /*
2232 * Make sure we do not leak PI boosting priority to the child:
2233 */
2234 p->prio = current->normal_prio;
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02002235 if (!rt_prio(p->prio))
2236 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07002237
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002238#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02002239 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002240 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08002242#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07002243 p->oncpu = 0;
2244#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07002246 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08002247 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002249 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250}
2251
2252/*
2253 * wake_up_new_task - wake up a newly created task for the first time.
2254 *
2255 * This function will do some initial scheduler statistics housekeeping
2256 * that must be done for every newly created context, then puts the task
2257 * on the runqueue and wakes it.
2258 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002259void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260{
2261 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002262 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263
2264 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02002266 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267
2268 p->prio = effective_prio(p);
2269
Srivatsa Vaddagirib9dca1e2007-10-17 16:55:11 +02002270 if (!p->sched_class->task_new || !current->se.on_rq) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002271 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02002274 * Let the scheduling class do new task startup
2275 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02002277 p->sched_class->task_new(rq, p);
Ingo Molnar6363ca52008-05-29 11:28:57 +02002278 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279 }
Ingo Molnardd41f592007-07-09 18:51:59 +02002280 check_preempt_curr(rq, p);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002281#ifdef CONFIG_SMP
2282 if (p->sched_class->task_wake_up)
2283 p->sched_class->task_wake_up(rq, p);
2284#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02002285 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286}
2287
Avi Kivitye107be32007-07-26 13:40:43 +02002288#ifdef CONFIG_PREEMPT_NOTIFIERS
2289
2290/**
Randy Dunlap421cee22007-07-31 00:37:50 -07002291 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2292 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002293 */
2294void preempt_notifier_register(struct preempt_notifier *notifier)
2295{
2296 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2297}
2298EXPORT_SYMBOL_GPL(preempt_notifier_register);
2299
2300/**
2301 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002302 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002303 *
2304 * This is safe to call from within a preemption notifier.
2305 */
2306void preempt_notifier_unregister(struct preempt_notifier *notifier)
2307{
2308 hlist_del(&notifier->link);
2309}
2310EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2311
2312static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2313{
2314 struct preempt_notifier *notifier;
2315 struct hlist_node *node;
2316
2317 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2318 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2319}
2320
2321static void
2322fire_sched_out_preempt_notifiers(struct task_struct *curr,
2323 struct task_struct *next)
2324{
2325 struct preempt_notifier *notifier;
2326 struct hlist_node *node;
2327
2328 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2329 notifier->ops->sched_out(notifier, next);
2330}
2331
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002332#else /* !CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002333
2334static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2335{
2336}
2337
2338static void
2339fire_sched_out_preempt_notifiers(struct task_struct *curr,
2340 struct task_struct *next)
2341{
2342}
2343
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002344#endif /* CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002345
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002347 * prepare_task_switch - prepare to switch tasks
2348 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002349 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002350 * @next: the task we are going to switch to.
2351 *
2352 * This is called with the rq lock held and interrupts off. It must
2353 * be paired with a subsequent finish_task_switch after the context
2354 * switch.
2355 *
2356 * prepare_task_switch sets up locking and calls architecture specific
2357 * hooks.
2358 */
Avi Kivitye107be32007-07-26 13:40:43 +02002359static inline void
2360prepare_task_switch(struct rq *rq, struct task_struct *prev,
2361 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002362{
Avi Kivitye107be32007-07-26 13:40:43 +02002363 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002364 prepare_lock_switch(rq, next);
2365 prepare_arch_switch(next);
2366}
2367
2368/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04002370 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 * @prev: the thread we just switched away from.
2372 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002373 * finish_task_switch must be called after the context switch, paired
2374 * with a prepare_task_switch call before the context switch.
2375 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2376 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377 *
2378 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002379 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 * with the lock held can cause deadlocks; see schedule() for
2381 * details.)
2382 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002383static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 __releases(rq->lock)
2385{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002387 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388
2389 rq->prev_mm = NULL;
2390
2391 /*
2392 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002393 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002394 * schedule one last time. The schedule call will never return, and
2395 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002396 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397 * still held, otherwise prev could be scheduled on another cpu, die
2398 * there before we look at prev->state, and then the reference would
2399 * be dropped twice.
2400 * Manfred Spraul <manfred@colorfullife.com>
2401 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002402 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07002403 finish_arch_switch(prev);
2404 finish_lock_switch(rq, prev);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002405#ifdef CONFIG_SMP
2406 if (current->sched_class->post_schedule)
2407 current->sched_class->post_schedule(rq);
2408#endif
Steven Rostedte8fa1362008-01-25 21:08:05 +01002409
Avi Kivitye107be32007-07-26 13:40:43 +02002410 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411 if (mm)
2412 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002413 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08002414 /*
2415 * Remove function-return probe instances associated with this
2416 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002417 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002418 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002420 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421}
2422
2423/**
2424 * schedule_tail - first thing a freshly forked thread must call.
2425 * @prev: the thread we just switched away from.
2426 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002427asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 __releases(rq->lock)
2429{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002430 struct rq *rq = this_rq();
2431
Nick Piggin4866cde2005-06-25 14:57:23 -07002432 finish_task_switch(rq, prev);
2433#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2434 /* In this case, finish_task_switch does not reenable preemption */
2435 preempt_enable();
2436#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002438 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439}
2440
2441/*
2442 * context_switch - switch to the new MM and the new
2443 * thread's register state.
2444 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002445static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002446context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002447 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448{
Ingo Molnardd41f592007-07-09 18:51:59 +02002449 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450
Avi Kivitye107be32007-07-26 13:40:43 +02002451 prepare_task_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02002452 mm = next->mm;
2453 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002454 /*
2455 * For paravirt, this is coupled with an exit in switch_to to
2456 * combine the page table reload and the switch backend into
2457 * one hypercall.
2458 */
2459 arch_enter_lazy_cpu_mode();
2460
Ingo Molnardd41f592007-07-09 18:51:59 +02002461 if (unlikely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 next->active_mm = oldmm;
2463 atomic_inc(&oldmm->mm_count);
2464 enter_lazy_tlb(oldmm, next);
2465 } else
2466 switch_mm(oldmm, mm, next);
2467
Ingo Molnardd41f592007-07-09 18:51:59 +02002468 if (unlikely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 rq->prev_mm = oldmm;
2471 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002472 /*
2473 * Since the runqueue lock will be released by the next
2474 * task (which is an invalid locking op but in the case
2475 * of the scheduler it's an obvious special-case), so we
2476 * do an early lockdep release here:
2477 */
2478#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002479 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002480#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481
2482 /* Here we just switch the register state and the stack. */
2483 switch_to(prev, next, prev);
2484
Ingo Molnardd41f592007-07-09 18:51:59 +02002485 barrier();
2486 /*
2487 * this_rq must be evaluated again because prev may have moved
2488 * CPUs since it called schedule(), thus the 'rq' on its stack
2489 * frame will be invalid.
2490 */
2491 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492}
2493
2494/*
2495 * nr_running, nr_uninterruptible and nr_context_switches:
2496 *
2497 * externally visible scheduler statistics: current number of runnable
2498 * threads, current number of uninterruptible-sleeping threads, total
2499 * number of context switches performed since bootup.
2500 */
2501unsigned long nr_running(void)
2502{
2503 unsigned long i, sum = 0;
2504
2505 for_each_online_cpu(i)
2506 sum += cpu_rq(i)->nr_running;
2507
2508 return sum;
2509}
2510
2511unsigned long nr_uninterruptible(void)
2512{
2513 unsigned long i, sum = 0;
2514
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002515 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 sum += cpu_rq(i)->nr_uninterruptible;
2517
2518 /*
2519 * Since we read the counters lockless, it might be slightly
2520 * inaccurate. Do not allow it to go below zero though:
2521 */
2522 if (unlikely((long)sum < 0))
2523 sum = 0;
2524
2525 return sum;
2526}
2527
2528unsigned long long nr_context_switches(void)
2529{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002530 int i;
2531 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002533 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 sum += cpu_rq(i)->nr_switches;
2535
2536 return sum;
2537}
2538
2539unsigned long nr_iowait(void)
2540{
2541 unsigned long i, sum = 0;
2542
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002543 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2545
2546 return sum;
2547}
2548
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002549unsigned long nr_active(void)
2550{
2551 unsigned long i, running = 0, uninterruptible = 0;
2552
2553 for_each_online_cpu(i) {
2554 running += cpu_rq(i)->nr_running;
2555 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2556 }
2557
2558 if (unlikely((long)uninterruptible < 0))
2559 uninterruptible = 0;
2560
2561 return running + uninterruptible;
2562}
2563
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564/*
Ingo Molnardd41f592007-07-09 18:51:59 +02002565 * Update rq->cpu_load[] statistics. This function is usually called every
2566 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07002567 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002568static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002569{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02002570 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02002571 int i, scale;
2572
2573 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02002574
2575 /* Update our load: */
2576 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2577 unsigned long old_load, new_load;
2578
2579 /* scale is effectively 1 << i now, and >> i divides by scale */
2580
2581 old_load = this_rq->cpu_load[i];
2582 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02002583 /*
2584 * Round up the averaging division if load is increasing. This
2585 * prevents us from getting stuck on 9 if the load is 10, for
2586 * example.
2587 */
2588 if (new_load > old_load)
2589 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02002590 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2591 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002592}
2593
Ingo Molnardd41f592007-07-09 18:51:59 +02002594#ifdef CONFIG_SMP
2595
Ingo Molnar48f24c42006-07-03 00:25:40 -07002596/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 * double_rq_lock - safely lock two runqueues
2598 *
2599 * Note this does not disable interrupts like task_rq_lock,
2600 * you need to do so manually before calling.
2601 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002602static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 __acquires(rq1->lock)
2604 __acquires(rq2->lock)
2605{
Kirill Korotaev054b9102006-12-10 02:20:11 -08002606 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 if (rq1 == rq2) {
2608 spin_lock(&rq1->lock);
2609 __acquire(rq2->lock); /* Fake it out ;) */
2610 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002611 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 spin_lock(&rq1->lock);
2613 spin_lock(&rq2->lock);
2614 } else {
2615 spin_lock(&rq2->lock);
2616 spin_lock(&rq1->lock);
2617 }
2618 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002619 update_rq_clock(rq1);
2620 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621}
2622
2623/*
2624 * double_rq_unlock - safely unlock two runqueues
2625 *
2626 * Note this does not restore interrupts like task_rq_unlock,
2627 * you need to do so manually after calling.
2628 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002629static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630 __releases(rq1->lock)
2631 __releases(rq2->lock)
2632{
2633 spin_unlock(&rq1->lock);
2634 if (rq1 != rq2)
2635 spin_unlock(&rq2->lock);
2636 else
2637 __release(rq2->lock);
2638}
2639
2640/*
2641 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2642 */
Steven Rostedte8fa1362008-01-25 21:08:05 +01002643static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 __releases(this_rq->lock)
2645 __acquires(busiest->lock)
2646 __acquires(this_rq->lock)
2647{
Steven Rostedte8fa1362008-01-25 21:08:05 +01002648 int ret = 0;
2649
Kirill Korotaev054b9102006-12-10 02:20:11 -08002650 if (unlikely(!irqs_disabled())) {
2651 /* printk() doesn't work good under rq->lock */
2652 spin_unlock(&this_rq->lock);
2653 BUG_ON(1);
2654 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 if (unlikely(!spin_trylock(&busiest->lock))) {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002656 if (busiest < this_rq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 spin_unlock(&this_rq->lock);
2658 spin_lock(&busiest->lock);
2659 spin_lock(&this_rq->lock);
Steven Rostedte8fa1362008-01-25 21:08:05 +01002660 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661 } else
2662 spin_lock(&busiest->lock);
2663 }
Steven Rostedte8fa1362008-01-25 21:08:05 +01002664 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665}
2666
2667/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668 * If dest_cpu is allowed for this process, migrate the task to it.
2669 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002670 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 * the cpu_allowed mask is restored.
2672 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002673static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002675 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002677 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678
2679 rq = task_rq_lock(p, &flags);
2680 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2681 || unlikely(cpu_is_offline(dest_cpu)))
2682 goto out;
2683
2684 /* force the process onto the specified CPU */
2685 if (migrate_task(p, dest_cpu, &req)) {
2686 /* Need to wait for migration thread (might exit: take ref). */
2687 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07002688
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689 get_task_struct(mt);
2690 task_rq_unlock(rq, &flags);
2691 wake_up_process(mt);
2692 put_task_struct(mt);
2693 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07002694
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695 return;
2696 }
2697out:
2698 task_rq_unlock(rq, &flags);
2699}
2700
2701/*
Nick Piggin476d1392005-06-25 14:57:29 -07002702 * sched_exec - execve() is a valuable balancing opportunity, because at
2703 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 */
2705void sched_exec(void)
2706{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707 int new_cpu, this_cpu = get_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002708 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002710 if (new_cpu != this_cpu)
2711 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712}
2713
2714/*
2715 * pull_task - move a task from a remote runqueue to the local runqueue.
2716 * Both runqueues must be locked.
2717 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002718static void pull_task(struct rq *src_rq, struct task_struct *p,
2719 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02002721 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002723 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724 /*
2725 * Note that idle threads have a prio of MAX_PRIO, for this test
2726 * to be always true for them.
2727 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002728 check_preempt_curr(this_rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729}
2730
2731/*
2732 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2733 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08002734static
Ingo Molnar70b97a72006-07-03 00:25:42 -07002735int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002736 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002737 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738{
2739 /*
2740 * We do not migrate tasks that are:
2741 * 1) running (obviously), or
2742 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2743 * 3) are cache-hot on their current CPU.
2744 */
Ingo Molnarcc367732007-10-15 17:00:18 +02002745 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2746 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002748 }
Nick Piggin81026792005-06-25 14:57:07 -07002749 *all_pinned = 0;
2750
Ingo Molnarcc367732007-10-15 17:00:18 +02002751 if (task_running(rq, p)) {
2752 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07002753 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002754 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755
Ingo Molnarda84d962007-10-15 17:00:18 +02002756 /*
2757 * Aggressive migration if:
2758 * 1) task is cache cold, or
2759 * 2) too many balance attempts have failed.
2760 */
2761
Ingo Molnar6bc16652007-10-15 17:00:18 +02002762 if (!task_hot(p, rq->clock, sd) ||
2763 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002764#ifdef CONFIG_SCHEDSTATS
Ingo Molnarcc367732007-10-15 17:00:18 +02002765 if (task_hot(p, rq->clock, sd)) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002766 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02002767 schedstat_inc(p, se.nr_forced_migrations);
2768 }
Ingo Molnarda84d962007-10-15 17:00:18 +02002769#endif
2770 return 1;
2771 }
2772
Ingo Molnarcc367732007-10-15 17:00:18 +02002773 if (task_hot(p, rq->clock, sd)) {
2774 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02002775 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002776 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777 return 1;
2778}
2779
Peter Williamse1d14842007-10-24 18:23:51 +02002780static unsigned long
2781balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2782 unsigned long max_load_move, struct sched_domain *sd,
2783 enum cpu_idle_type idle, int *all_pinned,
2784 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02002785{
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002786 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
Ingo Molnardd41f592007-07-09 18:51:59 +02002787 struct task_struct *p;
2788 long rem_load_move = max_load_move;
2789
Peter Williamse1d14842007-10-24 18:23:51 +02002790 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02002791 goto out;
2792
2793 pinned = 1;
2794
2795 /*
2796 * Start the load-balancing iterator:
2797 */
2798 p = iterator->start(iterator->arg);
2799next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002800 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02002801 goto out;
2802 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002803 * To help distribute high priority tasks across CPUs we don't
Ingo Molnardd41f592007-07-09 18:51:59 +02002804 * skip a task if it will be the highest priority task (i.e. smallest
2805 * prio value) on its new queue regardless of its load weight
2806 */
2807 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2808 SCHED_LOAD_SCALE_FUZZ;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002809 if ((skip_for_load && p->prio >= *this_best_prio) ||
Ingo Molnardd41f592007-07-09 18:51:59 +02002810 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002811 p = iterator->next(iterator->arg);
2812 goto next;
2813 }
2814
2815 pull_task(busiest, p, this_rq, this_cpu);
2816 pulled++;
2817 rem_load_move -= p->se.load.weight;
2818
2819 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002820 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02002821 */
Peter Williamse1d14842007-10-24 18:23:51 +02002822 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002823 if (p->prio < *this_best_prio)
2824 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02002825 p = iterator->next(iterator->arg);
2826 goto next;
2827 }
2828out:
2829 /*
Peter Williamse1d14842007-10-24 18:23:51 +02002830 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02002831 * so we can safely collect pull_task() stats here rather than
2832 * inside pull_task().
2833 */
2834 schedstat_add(sd, lb_gained[idle], pulled);
2835
2836 if (all_pinned)
2837 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02002838
2839 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02002840}
Ingo Molnar48f24c42006-07-03 00:25:40 -07002841
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842/*
Peter Williams43010652007-08-09 11:16:46 +02002843 * move_tasks tries to move up to max_load_move weighted load from busiest to
2844 * this_rq, as part of a balancing operation within domain "sd".
2845 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 *
2847 * Called with both runqueues locked.
2848 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002849static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02002850 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002851 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07002852 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02002854 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02002855 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002856 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857
Ingo Molnardd41f592007-07-09 18:51:59 +02002858 do {
Peter Williams43010652007-08-09 11:16:46 +02002859 total_load_moved +=
2860 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02002861 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002862 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02002863 class = class->next;
Peter Williams43010652007-08-09 11:16:46 +02002864 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865
Peter Williams43010652007-08-09 11:16:46 +02002866 return total_load_moved > 0;
2867}
2868
Peter Williamse1d14842007-10-24 18:23:51 +02002869static int
2870iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2871 struct sched_domain *sd, enum cpu_idle_type idle,
2872 struct rq_iterator *iterator)
2873{
2874 struct task_struct *p = iterator->start(iterator->arg);
2875 int pinned = 0;
2876
2877 while (p) {
2878 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2879 pull_task(busiest, p, this_rq, this_cpu);
2880 /*
2881 * Right now, this is only the second place pull_task()
2882 * is called, so we can safely collect pull_task()
2883 * stats here rather than inside pull_task().
2884 */
2885 schedstat_inc(sd, lb_gained[idle]);
2886
2887 return 1;
2888 }
2889 p = iterator->next(iterator->arg);
2890 }
2891
2892 return 0;
2893}
2894
Peter Williams43010652007-08-09 11:16:46 +02002895/*
2896 * move_one_task tries to move exactly one task from busiest to this_rq, as
2897 * part of active balancing operations within "domain".
2898 * Returns 1 if successful and 0 otherwise.
2899 *
2900 * Called with both runqueues locked.
2901 */
2902static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2903 struct sched_domain *sd, enum cpu_idle_type idle)
2904{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02002905 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02002906
2907 for (class = sched_class_highest; class; class = class->next)
Peter Williamse1d14842007-10-24 18:23:51 +02002908 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02002909 return 1;
2910
2911 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912}
2913
2914/*
2915 * find_busiest_group finds and returns the busiest CPU group within the
Ingo Molnar48f24c42006-07-03 00:25:40 -07002916 * domain. It calculates and returns the amount of weighted load which
2917 * should be moved to restore balance via the imbalance parameter.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 */
2919static struct sched_group *
2920find_busiest_group(struct sched_domain *sd, int this_cpu,
Ingo Molnardd41f592007-07-09 18:51:59 +02002921 unsigned long *imbalance, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07002922 int *sd_idle, const cpumask_t *cpus, int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923{
2924 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2925 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002926 unsigned long max_pull;
Peter Williams2dd73a42006-06-27 02:54:34 -07002927 unsigned long busiest_load_per_task, busiest_nr_running;
2928 unsigned long this_load_per_task, this_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02002929 int load_idx, group_imb = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002930#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2931 int power_savings_balance = 1;
2932 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2933 unsigned long min_nr_running = ULONG_MAX;
2934 struct sched_group *group_min = NULL, *group_leader = NULL;
2935#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936
2937 max_load = this_load = total_load = total_pwr = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07002938 busiest_load_per_task = busiest_nr_running = 0;
2939 this_load_per_task = this_nr_running = 0;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002940 if (idle == CPU_NOT_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07002941 load_idx = sd->busy_idx;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002942 else if (idle == CPU_NEWLY_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07002943 load_idx = sd->newidle_idx;
2944 else
2945 load_idx = sd->idle_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946
2947 do {
Ken Chen908a7c12007-10-17 16:55:11 +02002948 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 int local_group;
2950 int i;
Ken Chen908a7c12007-10-17 16:55:11 +02002951 int __group_imb = 0;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002952 unsigned int balance_cpu = -1, first_idle_cpu = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07002953 unsigned long sum_nr_running, sum_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954
2955 local_group = cpu_isset(this_cpu, group->cpumask);
2956
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002957 if (local_group)
2958 balance_cpu = first_cpu(group->cpumask);
2959
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 /* Tally up the load of all CPUs in the group */
Peter Williams2dd73a42006-06-27 02:54:34 -07002961 sum_weighted_load = sum_nr_running = avg_load = 0;
Ken Chen908a7c12007-10-17 16:55:11 +02002962 max_cpu_load = 0;
2963 min_cpu_load = ~0UL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964
2965 for_each_cpu_mask(i, group->cpumask) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002966 struct rq *rq;
2967
2968 if (!cpu_isset(i, *cpus))
2969 continue;
2970
2971 rq = cpu_rq(i);
Peter Williams2dd73a42006-06-27 02:54:34 -07002972
Suresh Siddha9439aab2007-07-19 21:28:35 +02002973 if (*sd_idle && rq->nr_running)
Nick Piggin5969fe02005-09-10 00:26:19 -07002974 *sd_idle = 0;
2975
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 /* Bias balancing toward cpus of our domain */
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002977 if (local_group) {
2978 if (idle_cpu(i) && !first_idle_cpu) {
2979 first_idle_cpu = 1;
2980 balance_cpu = i;
2981 }
2982
Nick Piggina2000572006-02-10 01:51:02 -08002983 load = target_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02002984 } else {
Nick Piggina2000572006-02-10 01:51:02 -08002985 load = source_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02002986 if (load > max_cpu_load)
2987 max_cpu_load = load;
2988 if (min_cpu_load > load)
2989 min_cpu_load = load;
2990 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991
2992 avg_load += load;
Peter Williams2dd73a42006-06-27 02:54:34 -07002993 sum_nr_running += rq->nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02002994 sum_weighted_load += weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 }
2996
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002997 /*
2998 * First idle cpu or the first cpu(busiest) in this sched group
2999 * is eligible for doing load balancing at this and above
Suresh Siddha9439aab2007-07-19 21:28:35 +02003000 * domains. In the newly idle case, we will allow all the cpu's
3001 * to do the newly idle load balance.
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003002 */
Suresh Siddha9439aab2007-07-19 21:28:35 +02003003 if (idle != CPU_NEWLY_IDLE && local_group &&
3004 balance_cpu != this_cpu && balance) {
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003005 *balance = 0;
3006 goto ret;
3007 }
3008
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 total_load += avg_load;
Eric Dumazet5517d862007-05-08 00:32:57 -07003010 total_pwr += group->__cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011
3012 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07003013 avg_load = sg_div_cpu_power(group,
3014 avg_load * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015
Ken Chen908a7c12007-10-17 16:55:11 +02003016 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3017 __group_imb = 1;
3018
Eric Dumazet5517d862007-05-08 00:32:57 -07003019 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003020
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021 if (local_group) {
3022 this_load = avg_load;
3023 this = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07003024 this_nr_running = sum_nr_running;
3025 this_load_per_task = sum_weighted_load;
3026 } else if (avg_load > max_load &&
Ken Chen908a7c12007-10-17 16:55:11 +02003027 (sum_nr_running > group_capacity || __group_imb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 max_load = avg_load;
3029 busiest = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07003030 busiest_nr_running = sum_nr_running;
3031 busiest_load_per_task = sum_weighted_load;
Ken Chen908a7c12007-10-17 16:55:11 +02003032 group_imb = __group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003034
3035#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3036 /*
3037 * Busy processors will not participate in power savings
3038 * balance.
3039 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003040 if (idle == CPU_NOT_IDLE ||
3041 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3042 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003043
3044 /*
3045 * If the local group is idle or completely loaded
3046 * no need to do power savings balance at this domain
3047 */
3048 if (local_group && (this_nr_running >= group_capacity ||
3049 !this_nr_running))
3050 power_savings_balance = 0;
3051
Ingo Molnardd41f592007-07-09 18:51:59 +02003052 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003053 * If a group is already running at full capacity or idle,
3054 * don't include that group in power savings calculations
Ingo Molnardd41f592007-07-09 18:51:59 +02003055 */
3056 if (!power_savings_balance || sum_nr_running >= group_capacity
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003057 || !sum_nr_running)
Ingo Molnardd41f592007-07-09 18:51:59 +02003058 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003059
Ingo Molnardd41f592007-07-09 18:51:59 +02003060 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003061 * Calculate the group which has the least non-idle load.
Ingo Molnardd41f592007-07-09 18:51:59 +02003062 * This is the group from where we need to pick up the load
3063 * for saving power
3064 */
3065 if ((sum_nr_running < min_nr_running) ||
3066 (sum_nr_running == min_nr_running &&
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003067 first_cpu(group->cpumask) <
3068 first_cpu(group_min->cpumask))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003069 group_min = group;
3070 min_nr_running = sum_nr_running;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003071 min_load_per_task = sum_weighted_load /
3072 sum_nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02003073 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003074
Ingo Molnardd41f592007-07-09 18:51:59 +02003075 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003076 * Calculate the group which is almost near its
Ingo Molnardd41f592007-07-09 18:51:59 +02003077 * capacity but still has some space to pick up some load
3078 * from other group and save more power
3079 */
3080 if (sum_nr_running <= group_capacity - 1) {
3081 if (sum_nr_running > leader_nr_running ||
3082 (sum_nr_running == leader_nr_running &&
3083 first_cpu(group->cpumask) >
3084 first_cpu(group_leader->cpumask))) {
3085 group_leader = group;
3086 leader_nr_running = sum_nr_running;
3087 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07003088 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003089group_next:
3090#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 group = group->next;
3092 } while (group != sd->groups);
3093
Peter Williams2dd73a42006-06-27 02:54:34 -07003094 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003095 goto out_balanced;
3096
3097 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3098
3099 if (this_load >= avg_load ||
3100 100*max_load <= sd->imbalance_pct*this_load)
3101 goto out_balanced;
3102
Peter Williams2dd73a42006-06-27 02:54:34 -07003103 busiest_load_per_task /= busiest_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02003104 if (group_imb)
3105 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3106
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107 /*
3108 * We're trying to get all the cpus to the average_load, so we don't
3109 * want to push ourselves above the average load, nor do we wish to
3110 * reduce the max loaded cpu below the average load, as either of these
3111 * actions would just result in more rebalancing later, and ping-pong
3112 * tasks around. Thus we look for the minimum possible imbalance.
3113 * Negative imbalances (*we* are more loaded than anyone else) will
3114 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003115 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116 * appear as very large values with unsigned longs.
3117 */
Peter Williams2dd73a42006-06-27 02:54:34 -07003118 if (max_load <= busiest_load_per_task)
3119 goto out_balanced;
3120
3121 /*
3122 * In the presence of smp nice balancing, certain scenarios can have
3123 * max load less than avg load(as we skip the groups at or below
3124 * its cpu_power, while calculating max_load..)
3125 */
3126 if (max_load < avg_load) {
3127 *imbalance = 0;
3128 goto small_imbalance;
3129 }
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003130
3131 /* Don't want to pull so many tasks that a group would go idle */
Peter Williams2dd73a42006-06-27 02:54:34 -07003132 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003133
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134 /* How much load to actually move to equalise the imbalance */
Eric Dumazet5517d862007-05-08 00:32:57 -07003135 *imbalance = min(max_pull * busiest->__cpu_power,
3136 (avg_load - this_load) * this->__cpu_power)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 / SCHED_LOAD_SCALE;
3138
Peter Williams2dd73a42006-06-27 02:54:34 -07003139 /*
3140 * if *imbalance is less than the average load per runnable task
3141 * there is no gaurantee that any tasks will be moved so we'll have
3142 * a think about bumping its value to force at least one task to be
3143 * moved
3144 */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02003145 if (*imbalance < busiest_load_per_task) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07003146 unsigned long tmp, pwr_now, pwr_move;
Peter Williams2dd73a42006-06-27 02:54:34 -07003147 unsigned int imbn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148
Peter Williams2dd73a42006-06-27 02:54:34 -07003149small_imbalance:
3150 pwr_move = pwr_now = 0;
3151 imbn = 2;
3152 if (this_nr_running) {
3153 this_load_per_task /= this_nr_running;
3154 if (busiest_load_per_task > this_load_per_task)
3155 imbn = 1;
3156 } else
3157 this_load_per_task = SCHED_LOAD_SCALE;
3158
Ingo Molnardd41f592007-07-09 18:51:59 +02003159 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3160 busiest_load_per_task * imbn) {
Peter Williams2dd73a42006-06-27 02:54:34 -07003161 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162 return busiest;
3163 }
3164
3165 /*
3166 * OK, we don't have enough imbalance to justify moving tasks,
3167 * however we may be able to increase total CPU power used by
3168 * moving them.
3169 */
3170
Eric Dumazet5517d862007-05-08 00:32:57 -07003171 pwr_now += busiest->__cpu_power *
3172 min(busiest_load_per_task, max_load);
3173 pwr_now += this->__cpu_power *
3174 min(this_load_per_task, this_load);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175 pwr_now /= SCHED_LOAD_SCALE;
3176
3177 /* Amount of load we'd subtract */
Eric Dumazet5517d862007-05-08 00:32:57 -07003178 tmp = sg_div_cpu_power(busiest,
3179 busiest_load_per_task * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 if (max_load > tmp)
Eric Dumazet5517d862007-05-08 00:32:57 -07003181 pwr_move += busiest->__cpu_power *
Peter Williams2dd73a42006-06-27 02:54:34 -07003182 min(busiest_load_per_task, max_load - tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183
3184 /* Amount of load we'd add */
Eric Dumazet5517d862007-05-08 00:32:57 -07003185 if (max_load * busiest->__cpu_power <
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08003186 busiest_load_per_task * SCHED_LOAD_SCALE)
Eric Dumazet5517d862007-05-08 00:32:57 -07003187 tmp = sg_div_cpu_power(this,
3188 max_load * busiest->__cpu_power);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 else
Eric Dumazet5517d862007-05-08 00:32:57 -07003190 tmp = sg_div_cpu_power(this,
3191 busiest_load_per_task * SCHED_LOAD_SCALE);
3192 pwr_move += this->__cpu_power *
3193 min(this_load_per_task, this_load + tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194 pwr_move /= SCHED_LOAD_SCALE;
3195
3196 /* Move if we gain throughput */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02003197 if (pwr_move > pwr_now)
3198 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199 }
3200
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201 return busiest;
3202
3203out_balanced:
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003204#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003205 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003206 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003208 if (this == group_leader && group_leader != group_min) {
3209 *imbalance = min_load_per_task;
3210 return group_min;
3211 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003212#endif
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003213ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214 *imbalance = 0;
3215 return NULL;
3216}
3217
3218/*
3219 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3220 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003221static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003222find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003223 unsigned long imbalance, const cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003224{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003225 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07003226 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227 int i;
3228
3229 for_each_cpu_mask(i, group->cpumask) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003230 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003231
3232 if (!cpu_isset(i, *cpus))
3233 continue;
3234
Ingo Molnar48f24c42006-07-03 00:25:40 -07003235 rq = cpu_rq(i);
Ingo Molnardd41f592007-07-09 18:51:59 +02003236 wl = weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237
Ingo Molnardd41f592007-07-09 18:51:59 +02003238 if (rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07003239 continue;
3240
Ingo Molnardd41f592007-07-09 18:51:59 +02003241 if (wl > max_load) {
3242 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003243 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003244 }
3245 }
3246
3247 return busiest;
3248}
3249
3250/*
Nick Piggin77391d72005-06-25 14:57:30 -07003251 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3252 * so long as it is large enough.
3253 */
3254#define MAX_PINNED_INTERVAL 512
3255
3256/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3258 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003259 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003260static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003261 struct sched_domain *sd, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003262 int *balance, cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263{
Peter Williams43010652007-08-09 11:16:46 +02003264 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003266 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003267 struct rq *busiest;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003268 unsigned long flags;
Nick Piggin5969fe02005-09-10 00:26:19 -07003269
Mike Travis7c16ec52008-04-04 18:11:11 -07003270 cpus_setall(*cpus);
3271
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003272 /*
3273 * When power savings policy is enabled for the parent domain, idle
3274 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02003275 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003276 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003277 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003278 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003279 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003280 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281
Ingo Molnar2d723762007-10-15 17:00:12 +02003282 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003284redo:
3285 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003286 cpus, balance);
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003287
Chen, Kenneth W06066712006-12-10 02:20:35 -08003288 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003289 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003290
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291 if (!group) {
3292 schedstat_inc(sd, lb_nobusyg[idle]);
3293 goto out_balanced;
3294 }
3295
Mike Travis7c16ec52008-04-04 18:11:11 -07003296 busiest = find_busiest_queue(group, idle, imbalance, cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297 if (!busiest) {
3298 schedstat_inc(sd, lb_nobusyq[idle]);
3299 goto out_balanced;
3300 }
3301
Nick Piggindb935db2005-06-25 14:57:11 -07003302 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303
3304 schedstat_add(sd, lb_imbalance[idle], imbalance);
3305
Peter Williams43010652007-08-09 11:16:46 +02003306 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003307 if (busiest->nr_running > 1) {
3308 /*
3309 * Attempt to move tasks. If find_busiest_group has found
3310 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02003311 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 * correctly treated as an imbalance.
3313 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003314 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07003315 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02003316 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07003317 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07003318 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003319 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07003320
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003321 /*
3322 * some other cpu did the load balance for us.
3323 */
Peter Williams43010652007-08-09 11:16:46 +02003324 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003325 resched_cpu(this_cpu);
3326
Nick Piggin81026792005-06-25 14:57:07 -07003327 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003328 if (unlikely(all_pinned)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003329 cpu_clear(cpu_of(busiest), *cpus);
3330 if (!cpus_empty(*cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003331 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07003332 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003333 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334 }
Nick Piggin81026792005-06-25 14:57:07 -07003335
Peter Williams43010652007-08-09 11:16:46 +02003336 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 schedstat_inc(sd, lb_failed[idle]);
3338 sd->nr_balance_failed++;
3339
3340 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003341
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003342 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003343
3344 /* don't kick the migration_thread, if the curr
3345 * task on busiest cpu can't be moved to this_cpu
3346 */
3347 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003348 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003349 all_pinned = 1;
3350 goto out_one_pinned;
3351 }
3352
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353 if (!busiest->active_balance) {
3354 busiest->active_balance = 1;
3355 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07003356 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003358 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07003359 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360 wake_up_process(busiest->migration_thread);
3361
3362 /*
3363 * We've kicked active balancing, reset the failure
3364 * counter.
3365 */
Nick Piggin39507452005-06-25 14:57:09 -07003366 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367 }
Nick Piggin81026792005-06-25 14:57:07 -07003368 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07003369 sd->nr_balance_failed = 0;
3370
Nick Piggin81026792005-06-25 14:57:07 -07003371 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003372 /* We were unbalanced, so reset the balancing interval */
3373 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07003374 } else {
3375 /*
3376 * If we've begun active balancing, start to back off. This
3377 * case may not be covered by the all_pinned logic if there
3378 * is only 1 task on the busy runqueue (because we don't call
3379 * move_tasks).
3380 */
3381 if (sd->balance_interval < sd->max_interval)
3382 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383 }
3384
Peter Williams43010652007-08-09 11:16:46 +02003385 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003386 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Ingo Molnar6363ca52008-05-29 11:28:57 +02003387 return -1;
3388 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389
3390out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391 schedstat_inc(sd, lb_balanced[idle]);
3392
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003393 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003394
3395out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07003397 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3398 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399 sd->balance_interval *= 2;
3400
Ingo Molnar48f24c42006-07-03 00:25:40 -07003401 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003402 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Ingo Molnar6363ca52008-05-29 11:28:57 +02003403 return -1;
3404 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405}
3406
3407/*
3408 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3409 * tasks if there is an imbalance.
3410 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003411 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412 * this_rq is locked.
3413 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07003414static int
Mike Travis7c16ec52008-04-04 18:11:11 -07003415load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3416 cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417{
3418 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003419 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02003421 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07003422 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003423 int all_pinned = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07003424
3425 cpus_setall(*cpus);
Nick Piggin5969fe02005-09-10 00:26:19 -07003426
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003427 /*
3428 * When power savings policy is enabled for the parent domain, idle
3429 * sibling can pick up load irrespective of busy siblings. In this case,
3430 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003431 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003432 */
3433 if (sd->flags & SD_SHARE_CPUPOWER &&
3434 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003435 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436
Ingo Molnar2d723762007-10-15 17:00:12 +02003437 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003438redo:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003439 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Mike Travis7c16ec52008-04-04 18:11:11 -07003440 &sd_idle, cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003442 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003443 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444 }
3445
Mike Travis7c16ec52008-04-04 18:11:11 -07003446 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07003447 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003448 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003449 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450 }
3451
Nick Piggindb935db2005-06-25 14:57:11 -07003452 BUG_ON(busiest == this_rq);
3453
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003454 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003455
Peter Williams43010652007-08-09 11:16:46 +02003456 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003457 if (busiest->nr_running > 1) {
3458 /* Attempt to move tasks */
3459 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003460 /* this_rq->clock is already updated */
3461 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02003462 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003463 imbalance, sd, CPU_NEWLY_IDLE,
3464 &all_pinned);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003465 spin_unlock(&busiest->lock);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003466
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003467 if (unlikely(all_pinned)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003468 cpu_clear(cpu_of(busiest), *cpus);
3469 if (!cpus_empty(*cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003470 goto redo;
3471 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003472 }
3473
Peter Williams43010652007-08-09 11:16:46 +02003474 if (!ld_moved) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003475 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003476 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3477 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003478 return -1;
3479 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003480 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481
Peter Williams43010652007-08-09 11:16:46 +02003482 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003483
3484out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003485 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003486 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003487 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003488 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003489 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003490
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003491 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492}
3493
3494/*
3495 * idle_balance is called by schedule() if this_cpu is about to become
3496 * idle. Attempts to pull tasks from other CPUs.
3497 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003498static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499{
3500 struct sched_domain *sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02003501 int pulled_task = -1;
3502 unsigned long next_balance = jiffies + HZ;
Mike Travis7c16ec52008-04-04 18:11:11 -07003503 cpumask_t tmpmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504
3505 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003506 unsigned long interval;
3507
3508 if (!(sd->flags & SD_LOAD_BALANCE))
3509 continue;
3510
3511 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07003512 /* If we've pulled tasks over stop searching: */
Mike Travis7c16ec52008-04-04 18:11:11 -07003513 pulled_task = load_balance_newidle(this_cpu, this_rq,
3514 sd, &tmpmask);
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003515
3516 interval = msecs_to_jiffies(sd->balance_interval);
3517 if (time_after(next_balance, sd->last_balance + interval))
3518 next_balance = sd->last_balance + interval;
3519 if (pulled_task)
3520 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003522 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003523 /*
3524 * We are going idle. next_balance may be set based on
3525 * a busy processor. So reset next_balance.
3526 */
3527 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02003528 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529}
3530
3531/*
3532 * active_load_balance is run by migration threads. It pushes running tasks
3533 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3534 * running on each physical CPU where possible, and avoids physical /
3535 * logical imbalances.
3536 *
3537 * Called with busiest_rq locked.
3538 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003539static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540{
Nick Piggin39507452005-06-25 14:57:09 -07003541 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003542 struct sched_domain *sd;
3543 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07003544
Ingo Molnar48f24c42006-07-03 00:25:40 -07003545 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07003546 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07003547 return;
3548
3549 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550
3551 /*
Nick Piggin39507452005-06-25 14:57:09 -07003552 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003553 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07003554 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555 */
Nick Piggin39507452005-06-25 14:57:09 -07003556 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557
Nick Piggin39507452005-06-25 14:57:09 -07003558 /* move a task from busiest_rq to target_rq */
3559 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003560 update_rq_clock(busiest_rq);
3561 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562
Nick Piggin39507452005-06-25 14:57:09 -07003563 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003564 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07003565 if ((sd->flags & SD_LOAD_BALANCE) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07003566 cpu_isset(busiest_cpu, sd->span))
Nick Piggin39507452005-06-25 14:57:09 -07003567 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003568 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569
Ingo Molnar48f24c42006-07-03 00:25:40 -07003570 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02003571 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572
Peter Williams43010652007-08-09 11:16:46 +02003573 if (move_one_task(target_rq, target_cpu, busiest_rq,
3574 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07003575 schedstat_inc(sd, alb_pushed);
3576 else
3577 schedstat_inc(sd, alb_failed);
3578 }
Nick Piggin39507452005-06-25 14:57:09 -07003579 spin_unlock(&target_rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580}
3581
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003582#ifdef CONFIG_NO_HZ
3583static struct {
3584 atomic_t load_balancer;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003585 cpumask_t cpu_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003586} nohz ____cacheline_aligned = {
3587 .load_balancer = ATOMIC_INIT(-1),
3588 .cpu_mask = CPU_MASK_NONE,
3589};
3590
Christoph Lameter7835b982006-12-10 02:20:22 -08003591/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003592 * This routine will try to nominate the ilb (idle load balancing)
3593 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3594 * load balancing on behalf of all those cpus. If all the cpus in the system
3595 * go into this tickless mode, then there will be no ilb owner (as there is
3596 * no need for one) and all the cpus will sleep till the next wakeup event
3597 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08003598 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003599 * For the ilb owner, tick is not stopped. And this tick will be used
3600 * for idle load balancing. ilb owner will still be part of
3601 * nohz.cpu_mask..
3602 *
3603 * While stopping the tick, this cpu will become the ilb owner if there
3604 * is no other owner. And will be the owner till that cpu becomes busy
3605 * or if all cpus in the system stop their ticks at which point
3606 * there is no need for ilb owner.
3607 *
3608 * When the ilb owner becomes busy, it nominates another owner, during the
3609 * next busy scheduler_tick()
3610 */
3611int select_nohz_load_balancer(int stop_tick)
3612{
3613 int cpu = smp_processor_id();
3614
3615 if (stop_tick) {
3616 cpu_set(cpu, nohz.cpu_mask);
3617 cpu_rq(cpu)->in_nohz_recently = 1;
3618
3619 /*
3620 * If we are going offline and still the leader, give up!
3621 */
3622 if (cpu_is_offline(cpu) &&
3623 atomic_read(&nohz.load_balancer) == cpu) {
3624 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3625 BUG();
3626 return 0;
3627 }
3628
3629 /* time for ilb owner also to sleep */
3630 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3631 if (atomic_read(&nohz.load_balancer) == cpu)
3632 atomic_set(&nohz.load_balancer, -1);
3633 return 0;
3634 }
3635
3636 if (atomic_read(&nohz.load_balancer) == -1) {
3637 /* make me the ilb owner */
3638 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3639 return 1;
3640 } else if (atomic_read(&nohz.load_balancer) == cpu)
3641 return 1;
3642 } else {
3643 if (!cpu_isset(cpu, nohz.cpu_mask))
3644 return 0;
3645
3646 cpu_clear(cpu, nohz.cpu_mask);
3647
3648 if (atomic_read(&nohz.load_balancer) == cpu)
3649 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3650 BUG();
3651 }
3652 return 0;
3653}
3654#endif
3655
3656static DEFINE_SPINLOCK(balancing);
3657
3658/*
Christoph Lameter7835b982006-12-10 02:20:22 -08003659 * It checks each scheduling domain to see if it is due to be balanced,
3660 * and initiates a balancing operation if so.
3661 *
3662 * Balancing parameters are set up in arch_init_sched_domains.
3663 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02003664static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08003665{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003666 int balance = 1;
3667 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08003668 unsigned long interval;
3669 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003670 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08003671 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003672 int update_next_balance = 0;
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003673 int need_serialize;
Mike Travis7c16ec52008-04-04 18:11:11 -07003674 cpumask_t tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003676 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677 if (!(sd->flags & SD_LOAD_BALANCE))
3678 continue;
3679
3680 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003681 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682 interval *= sd->busy_factor;
3683
3684 /* scale ms to jiffies */
3685 interval = msecs_to_jiffies(interval);
3686 if (unlikely(!interval))
3687 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003688 if (interval > HZ*NR_CPUS/10)
3689 interval = HZ*NR_CPUS/10;
3690
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003691 need_serialize = sd->flags & SD_SERIALIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003692
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003693 if (need_serialize) {
Christoph Lameter08c183f2006-12-10 02:20:29 -08003694 if (!spin_trylock(&balancing))
3695 goto out;
3696 }
3697
Christoph Lameterc9819f42006-12-10 02:20:25 -08003698 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003699 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003700 /*
3701 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07003702 * longer idle, or one of our SMT siblings is
3703 * not idle.
3704 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003705 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003707 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708 }
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003709 if (need_serialize)
Christoph Lameter08c183f2006-12-10 02:20:29 -08003710 spin_unlock(&balancing);
3711out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02003712 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08003713 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003714 update_next_balance = 1;
3715 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003716
3717 /*
3718 * Stop the load balance at this level. There is another
3719 * CPU in our sched group which is doing load balancing more
3720 * actively.
3721 */
3722 if (!balance)
3723 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02003725
3726 /*
3727 * next_balance will be updated only when there is a need.
3728 * When the cpu is attached to null domain for ex, it will not be
3729 * updated.
3730 */
3731 if (likely(update_next_balance))
3732 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003733}
3734
3735/*
3736 * run_rebalance_domains is triggered when needed from the scheduler tick.
3737 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3738 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3739 */
3740static void run_rebalance_domains(struct softirq_action *h)
3741{
Ingo Molnardd41f592007-07-09 18:51:59 +02003742 int this_cpu = smp_processor_id();
3743 struct rq *this_rq = cpu_rq(this_cpu);
3744 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3745 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003746
Ingo Molnardd41f592007-07-09 18:51:59 +02003747 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003748
3749#ifdef CONFIG_NO_HZ
3750 /*
3751 * If this cpu is the owner for idle load balancing, then do the
3752 * balancing on behalf of the other idle cpus whose ticks are
3753 * stopped.
3754 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003755 if (this_rq->idle_at_tick &&
3756 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003757 cpumask_t cpus = nohz.cpu_mask;
3758 struct rq *rq;
3759 int balance_cpu;
3760
Ingo Molnardd41f592007-07-09 18:51:59 +02003761 cpu_clear(this_cpu, cpus);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003762 for_each_cpu_mask(balance_cpu, cpus) {
3763 /*
3764 * If this cpu gets work to do, stop the load balancing
3765 * work being done for other cpus. Next load
3766 * balancing owner will pick it up.
3767 */
3768 if (need_resched())
3769 break;
3770
Oleg Nesterovde0cf892007-08-12 18:08:19 +02003771 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003772
3773 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003774 if (time_after(this_rq->next_balance, rq->next_balance))
3775 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003776 }
3777 }
3778#endif
3779}
3780
3781/*
3782 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3783 *
3784 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3785 * idle load balancing owner or decide to stop the periodic load balancing,
3786 * if the whole system is idle.
3787 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003788static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003789{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003790#ifdef CONFIG_NO_HZ
3791 /*
3792 * If we were in the nohz mode recently and busy at the current
3793 * scheduler tick, then check if we need to nominate new idle
3794 * load balancer.
3795 */
3796 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3797 rq->in_nohz_recently = 0;
3798
3799 if (atomic_read(&nohz.load_balancer) == cpu) {
3800 cpu_clear(cpu, nohz.cpu_mask);
3801 atomic_set(&nohz.load_balancer, -1);
3802 }
3803
3804 if (atomic_read(&nohz.load_balancer) == -1) {
3805 /*
3806 * simple selection for now: Nominate the
3807 * first cpu in the nohz list to be the next
3808 * ilb owner.
3809 *
3810 * TBD: Traverse the sched domains and nominate
3811 * the nearest cpu in the nohz.cpu_mask.
3812 */
3813 int ilb = first_cpu(nohz.cpu_mask);
3814
Mike Travis434d53b2008-04-04 18:11:04 -07003815 if (ilb < nr_cpu_ids)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003816 resched_cpu(ilb);
3817 }
3818 }
3819
3820 /*
3821 * If this cpu is idle and doing idle load balancing for all the
3822 * cpus with ticks stopped, is it time for that to stop?
3823 */
3824 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3825 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3826 resched_cpu(cpu);
3827 return;
3828 }
3829
3830 /*
3831 * If this cpu is idle and the idle load balancing is done by
3832 * someone else, then no need raise the SCHED_SOFTIRQ
3833 */
3834 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3835 cpu_isset(cpu, nohz.cpu_mask))
3836 return;
3837#endif
3838 if (time_after_eq(jiffies, rq->next_balance))
3839 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840}
Ingo Molnardd41f592007-07-09 18:51:59 +02003841
3842#else /* CONFIG_SMP */
3843
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844/*
3845 * on UP we do not need to balance between CPUs:
3846 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003847static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848{
3849}
Ingo Molnardd41f592007-07-09 18:51:59 +02003850
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851#endif
3852
Linus Torvalds1da177e2005-04-16 15:20:36 -07003853DEFINE_PER_CPU(struct kernel_stat, kstat);
3854
3855EXPORT_PER_CPU_SYMBOL(kstat);
3856
3857/*
Ingo Molnar41b86e92007-07-09 18:51:58 +02003858 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3859 * that have not yet been banked in case the task is currently running.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860 */
Ingo Molnar41b86e92007-07-09 18:51:58 +02003861unsigned long long task_sched_runtime(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003864 u64 ns, delta_exec;
3865 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003866
Ingo Molnar41b86e92007-07-09 18:51:58 +02003867 rq = task_rq_lock(p, &flags);
3868 ns = p->se.sum_exec_runtime;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01003869 if (task_current(rq, p)) {
Ingo Molnara8e504d2007-08-09 11:16:47 +02003870 update_rq_clock(rq);
3871 delta_exec = rq->clock - p->se.exec_start;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003872 if ((s64)delta_exec > 0)
3873 ns += delta_exec;
3874 }
3875 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003876
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877 return ns;
3878}
3879
3880/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003881 * Account user cpu time to a process.
3882 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883 * @cputime: the cpu time spent in user space since the last update
3884 */
3885void account_user_time(struct task_struct *p, cputime_t cputime)
3886{
3887 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3888 cputime64_t tmp;
3889
3890 p->utime = cputime_add(p->utime, cputime);
3891
3892 /* Add user time to cpustat. */
3893 tmp = cputime_to_cputime64(cputime);
3894 if (TASK_NICE(p) > 0)
3895 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3896 else
3897 cpustat->user = cputime64_add(cpustat->user, tmp);
3898}
3899
3900/*
Laurent Vivier94886b82007-10-15 17:00:19 +02003901 * Account guest cpu time to a process.
3902 * @p: the process that the cpu time gets accounted to
3903 * @cputime: the cpu time spent in virtual machine since the last update
3904 */
Adrian Bunkf7402e02007-10-29 21:18:10 +01003905static void account_guest_time(struct task_struct *p, cputime_t cputime)
Laurent Vivier94886b82007-10-15 17:00:19 +02003906{
3907 cputime64_t tmp;
3908 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3909
3910 tmp = cputime_to_cputime64(cputime);
3911
3912 p->utime = cputime_add(p->utime, cputime);
3913 p->gtime = cputime_add(p->gtime, cputime);
3914
3915 cpustat->user = cputime64_add(cpustat->user, tmp);
3916 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3917}
3918
3919/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07003920 * Account scaled user cpu time to a process.
3921 * @p: the process that the cpu time gets accounted to
3922 * @cputime: the cpu time spent in user space since the last update
3923 */
3924void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3925{
3926 p->utimescaled = cputime_add(p->utimescaled, cputime);
3927}
3928
3929/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 * Account system cpu time to a process.
3931 * @p: the process that the cpu time gets accounted to
3932 * @hardirq_offset: the offset to subtract from hardirq_count()
3933 * @cputime: the cpu time spent in kernel space since the last update
3934 */
3935void account_system_time(struct task_struct *p, int hardirq_offset,
3936 cputime_t cputime)
3937{
3938 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003939 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 cputime64_t tmp;
3941
Harvey Harrison983ed7a2008-04-24 18:17:55 -07003942 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3943 account_guest_time(p, cputime);
3944 return;
3945 }
Laurent Vivier94886b82007-10-15 17:00:19 +02003946
Linus Torvalds1da177e2005-04-16 15:20:36 -07003947 p->stime = cputime_add(p->stime, cputime);
3948
3949 /* Add system time to cpustat. */
3950 tmp = cputime_to_cputime64(cputime);
3951 if (hardirq_count() - hardirq_offset)
3952 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3953 else if (softirq_count())
3954 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08003955 else if (p != rq->idle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 cpustat->system = cputime64_add(cpustat->system, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08003957 else if (atomic_read(&rq->nr_iowait) > 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3959 else
3960 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3961 /* Account for system time used */
3962 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003963}
3964
3965/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07003966 * Account scaled system cpu time to a process.
3967 * @p: the process that the cpu time gets accounted to
3968 * @hardirq_offset: the offset to subtract from hardirq_count()
3969 * @cputime: the cpu time spent in kernel space since the last update
3970 */
3971void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3972{
3973 p->stimescaled = cputime_add(p->stimescaled, cputime);
3974}
3975
3976/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977 * Account for involuntary wait time.
3978 * @p: the process from which the cpu time has been stolen
3979 * @steal: the cpu time spent in involuntary wait
3980 */
3981void account_steal_time(struct task_struct *p, cputime_t steal)
3982{
3983 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3984 cputime64_t tmp = cputime_to_cputime64(steal);
Ingo Molnar70b97a72006-07-03 00:25:42 -07003985 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986
3987 if (p == rq->idle) {
3988 p->stime = cputime_add(p->stime, steal);
3989 if (atomic_read(&rq->nr_iowait) > 0)
3990 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3991 else
3992 cpustat->idle = cputime64_add(cpustat->idle, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08003993 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3995}
3996
Christoph Lameter7835b982006-12-10 02:20:22 -08003997/*
3998 * This function gets called by the timer code, with HZ frequency.
3999 * We call it with interrupts disabled.
4000 *
4001 * It also gets called by the fork code, when changing the parent's
4002 * timeslices.
4003 */
4004void scheduler_tick(void)
4005{
Christoph Lameter7835b982006-12-10 02:20:22 -08004006 int cpu = smp_processor_id();
4007 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004008 struct task_struct *curr = rq->curr;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004009
4010 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08004011
Ingo Molnardd41f592007-07-09 18:51:59 +02004012 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004013 update_rq_clock(rq);
Ingo Molnarf1a438d2007-08-09 11:16:45 +02004014 update_cpu_load(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01004015 curr->sched_class->task_tick(rq, curr, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02004016 spin_unlock(&rq->lock);
4017
Christoph Lametere418e1c2006-12-10 02:20:23 -08004018#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02004019 rq->idle_at_tick = idle_cpu(cpu);
4020 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08004021#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022}
4023
Linus Torvalds1da177e2005-04-16 15:20:36 -07004024#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4025
Srinivasa Ds43627582008-02-23 15:24:04 -08004026void __kprobes add_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027{
4028 /*
4029 * Underflow?
4030 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004031 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4032 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033 preempt_count() += val;
4034 /*
4035 * Spinlock count overflowing soon?
4036 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08004037 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4038 PREEMPT_MASK - 10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039}
4040EXPORT_SYMBOL(add_preempt_count);
4041
Srinivasa Ds43627582008-02-23 15:24:04 -08004042void __kprobes sub_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043{
4044 /*
4045 * Underflow?
4046 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004047 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4048 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049 /*
4050 * Is the spinlock portion underflowing?
4051 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004052 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4053 !(preempt_count() & PREEMPT_MASK)))
4054 return;
4055
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056 preempt_count() -= val;
4057}
4058EXPORT_SYMBOL(sub_preempt_count);
4059
4060#endif
4061
4062/*
Ingo Molnardd41f592007-07-09 18:51:59 +02004063 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004064 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004065static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066{
Satyam Sharma838225b2007-10-24 18:23:50 +02004067 struct pt_regs *regs = get_irq_regs();
4068
4069 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4070 prev->comm, prev->pid, preempt_count());
4071
Ingo Molnardd41f592007-07-09 18:51:59 +02004072 debug_show_held_locks(prev);
Arjan van de Vene21f5b12008-05-23 09:05:58 -07004073 print_modules();
Ingo Molnardd41f592007-07-09 18:51:59 +02004074 if (irqs_disabled())
4075 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02004076
4077 if (regs)
4078 show_regs(regs);
4079 else
4080 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02004081}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082
Ingo Molnardd41f592007-07-09 18:51:59 +02004083/*
4084 * Various schedule()-time debugging checks and statistics:
4085 */
4086static inline void schedule_debug(struct task_struct *prev)
4087{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004089 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090 * schedule() atomically, we ignore that path for now.
4091 * Otherwise, whine if we are scheduling when we should not be.
4092 */
Roel Kluin3f33a7c2008-05-13 23:44:11 +02004093 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
Ingo Molnardd41f592007-07-09 18:51:59 +02004094 __schedule_bug(prev);
4095
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4097
Ingo Molnar2d723762007-10-15 17:00:12 +02004098 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02004099#ifdef CONFIG_SCHEDSTATS
4100 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02004101 schedstat_inc(this_rq(), bkl_count);
4102 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02004103 }
4104#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02004105}
4106
4107/*
4108 * Pick up the highest-prio task:
4109 */
4110static inline struct task_struct *
Ingo Molnarff95f3d2007-08-09 11:16:49 +02004111pick_next_task(struct rq *rq, struct task_struct *prev)
Ingo Molnardd41f592007-07-09 18:51:59 +02004112{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02004113 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02004114 struct task_struct *p;
4115
4116 /*
4117 * Optimization: we know that if all tasks are in
4118 * the fair class we can call that function directly:
4119 */
4120 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02004121 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004122 if (likely(p))
4123 return p;
4124 }
4125
4126 class = sched_class_highest;
4127 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02004128 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004129 if (p)
4130 return p;
4131 /*
4132 * Will never be NULL as the idle class always
4133 * returns a non-NULL p:
4134 */
4135 class = class->next;
4136 }
4137}
4138
4139/*
4140 * schedule() is the main scheduler function.
4141 */
4142asmlinkage void __sched schedule(void)
4143{
4144 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08004145 unsigned long *switch_count;
Ingo Molnardd41f592007-07-09 18:51:59 +02004146 struct rq *rq;
Mike Galbraithf333fdc2008-05-12 21:20:55 +02004147 int cpu, hrtick = sched_feat(HRTICK);
Ingo Molnardd41f592007-07-09 18:51:59 +02004148
Linus Torvalds1da177e2005-04-16 15:20:36 -07004149need_resched:
4150 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02004151 cpu = smp_processor_id();
4152 rq = cpu_rq(cpu);
4153 rcu_qsctr_inc(cpu);
4154 prev = rq->curr;
4155 switch_count = &prev->nivcsw;
4156
Linus Torvalds1da177e2005-04-16 15:20:36 -07004157 release_kernel_lock(prev);
4158need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159
Ingo Molnardd41f592007-07-09 18:51:59 +02004160 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161
Mike Galbraithf333fdc2008-05-12 21:20:55 +02004162 if (hrtick)
4163 hrtick_clear(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004164
Ingo Molnar1e819952007-10-15 17:00:13 +02004165 /*
4166 * Do the rq-clock update outside the rq lock:
4167 */
4168 local_irq_disable();
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004169 update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02004170 spin_lock(&rq->lock);
4171 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172
Ingo Molnardd41f592007-07-09 18:51:59 +02004173 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4174 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
Roel Kluin23e3c3c2008-03-13 17:41:59 +01004175 signal_pending(prev))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02004176 prev->state = TASK_RUNNING;
4177 } else {
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004178 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02004179 }
4180 switch_count = &prev->nvcsw;
4181 }
4182
Steven Rostedt9a897c52008-01-25 21:08:22 +01004183#ifdef CONFIG_SMP
4184 if (prev->sched_class->pre_schedule)
4185 prev->sched_class->pre_schedule(rq, prev);
4186#endif
Steven Rostedtf65eda42008-01-25 21:08:07 +01004187
Ingo Molnardd41f592007-07-09 18:51:59 +02004188 if (unlikely(!rq->nr_running))
4189 idle_balance(cpu, rq);
4190
Ingo Molnar31ee5292007-08-09 11:16:49 +02004191 prev->sched_class->put_prev_task(rq, prev);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02004192 next = pick_next_task(rq, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193
Linus Torvalds1da177e2005-04-16 15:20:36 -07004194 if (likely(prev != next)) {
David Simner673a90a2008-04-29 10:08:59 +01004195 sched_info_switch(prev, next);
4196
Linus Torvalds1da177e2005-04-16 15:20:36 -07004197 rq->nr_switches++;
4198 rq->curr = next;
4199 ++*switch_count;
4200
Ingo Molnardd41f592007-07-09 18:51:59 +02004201 context_switch(rq, prev, next); /* unlocks the rq */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004202 /*
4203 * the context switch might have flipped the stack from under
4204 * us, hence refresh the local variables.
4205 */
4206 cpu = smp_processor_id();
4207 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208 } else
4209 spin_unlock_irq(&rq->lock);
4210
Mike Galbraithf333fdc2008-05-12 21:20:55 +02004211 if (hrtick)
4212 hrtick_set(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004213
4214 if (unlikely(reacquire_kernel_lock(current) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 goto need_resched_nonpreemptible;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004216
Linus Torvalds1da177e2005-04-16 15:20:36 -07004217 preempt_enable_no_resched();
4218 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4219 goto need_resched;
4220}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221EXPORT_SYMBOL(schedule);
4222
4223#ifdef CONFIG_PREEMPT
4224/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004225 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004226 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07004227 * occur there and call schedule directly.
4228 */
4229asmlinkage void __sched preempt_schedule(void)
4230{
4231 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01004232
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233 /*
4234 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004235 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07004237 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238 return;
4239
Andi Kleen3a5c3592007-10-15 17:00:14 +02004240 do {
4241 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02004242 schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02004243 sub_preempt_count(PREEMPT_ACTIVE);
4244
4245 /*
4246 * Check again in case we missed a preemption opportunity
4247 * between schedule and now.
4248 */
4249 barrier();
4250 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004252EXPORT_SYMBOL(preempt_schedule);
4253
4254/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004255 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07004256 * off of irq context.
4257 * Note, that this is called and return with irqs disabled. This will
4258 * protect us against recursive calling from irq.
4259 */
4260asmlinkage void __sched preempt_schedule_irq(void)
4261{
4262 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01004263
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004264 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004265 BUG_ON(ti->preempt_count || !irqs_disabled());
4266
Andi Kleen3a5c3592007-10-15 17:00:14 +02004267 do {
4268 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02004269 local_irq_enable();
4270 schedule();
4271 local_irq_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02004272 sub_preempt_count(PREEMPT_ACTIVE);
4273
4274 /*
4275 * Check again in case we missed a preemption opportunity
4276 * between schedule and now.
4277 */
4278 barrier();
4279 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004280}
4281
4282#endif /* CONFIG_PREEMPT */
4283
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004284int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4285 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004287 return try_to_wake_up(curr->private, mode, sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004288}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289EXPORT_SYMBOL(default_wake_function);
4290
4291/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004292 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4293 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07004294 * number) then we wake all the non-exclusive tasks and one exclusive task.
4295 *
4296 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004297 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07004298 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4299 */
4300static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4301 int nr_exclusive, int sync, void *key)
4302{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004303 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004304
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004305 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07004306 unsigned flags = curr->flags;
4307
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308 if (curr->func(curr, mode, sync, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07004309 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004310 break;
4311 }
4312}
4313
4314/**
4315 * __wake_up - wake up threads blocked on a waitqueue.
4316 * @q: the waitqueue
4317 * @mode: which threads
4318 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07004319 * @key: is directly passed to the wakeup function
Linus Torvalds1da177e2005-04-16 15:20:36 -07004320 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004321void __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004322 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004323{
4324 unsigned long flags;
4325
4326 spin_lock_irqsave(&q->lock, flags);
4327 __wake_up_common(q, mode, nr_exclusive, 0, key);
4328 spin_unlock_irqrestore(&q->lock, flags);
4329}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004330EXPORT_SYMBOL(__wake_up);
4331
4332/*
4333 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4334 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004335void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004336{
4337 __wake_up_common(q, mode, 1, 0, NULL);
4338}
4339
4340/**
Martin Waitz67be2dd2005-05-01 08:59:26 -07004341 * __wake_up_sync - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004342 * @q: the waitqueue
4343 * @mode: which threads
4344 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4345 *
4346 * The sync wakeup differs that the waker knows that it will schedule
4347 * away soon, so while the target thread will be woken up, it will not
4348 * be migrated to another CPU - ie. the two threads are 'synchronized'
4349 * with each other. This can prevent needless bouncing between CPUs.
4350 *
4351 * On UP it can prevent extra preemption.
4352 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004353void
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004354__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004355{
4356 unsigned long flags;
4357 int sync = 1;
4358
4359 if (unlikely(!q))
4360 return;
4361
4362 if (unlikely(!nr_exclusive))
4363 sync = 0;
4364
4365 spin_lock_irqsave(&q->lock, flags);
4366 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4367 spin_unlock_irqrestore(&q->lock, flags);
4368}
4369EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4370
Ingo Molnarb15136e2007-10-24 18:23:48 +02004371void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004372{
4373 unsigned long flags;
4374
4375 spin_lock_irqsave(&x->wait.lock, flags);
4376 x->done++;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004377 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004378 spin_unlock_irqrestore(&x->wait.lock, flags);
4379}
4380EXPORT_SYMBOL(complete);
4381
Ingo Molnarb15136e2007-10-24 18:23:48 +02004382void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004383{
4384 unsigned long flags;
4385
4386 spin_lock_irqsave(&x->wait.lock, flags);
4387 x->done += UINT_MAX/2;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004388 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004389 spin_unlock_irqrestore(&x->wait.lock, flags);
4390}
4391EXPORT_SYMBOL(complete_all);
4392
Andi Kleen8cbbe862007-10-15 17:00:14 +02004393static inline long __sched
4394do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004395{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004396 if (!x->done) {
4397 DECLARE_WAITQUEUE(wait, current);
4398
4399 wait.flags |= WQ_FLAG_EXCLUSIVE;
4400 __add_wait_queue_tail(&x->wait, &wait);
4401 do {
Matthew Wilcox009e5772007-12-06 12:29:54 -05004402 if ((state == TASK_INTERRUPTIBLE &&
4403 signal_pending(current)) ||
4404 (state == TASK_KILLABLE &&
4405 fatal_signal_pending(current))) {
Andi Kleen8cbbe862007-10-15 17:00:14 +02004406 __remove_wait_queue(&x->wait, &wait);
4407 return -ERESTARTSYS;
4408 }
4409 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004410 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004411 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004412 spin_lock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004413 if (!timeout) {
4414 __remove_wait_queue(&x->wait, &wait);
4415 return timeout;
4416 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004417 } while (!x->done);
4418 __remove_wait_queue(&x->wait, &wait);
4419 }
4420 x->done--;
Andi Kleen8cbbe862007-10-15 17:00:14 +02004421 return timeout;
4422}
4423
4424static long __sched
4425wait_for_common(struct completion *x, long timeout, int state)
4426{
4427 might_sleep();
4428
4429 spin_lock_irq(&x->wait.lock);
4430 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004431 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004432 return timeout;
4433}
4434
Ingo Molnarb15136e2007-10-24 18:23:48 +02004435void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02004436{
4437 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004438}
4439EXPORT_SYMBOL(wait_for_completion);
4440
Ingo Molnarb15136e2007-10-24 18:23:48 +02004441unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004442wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4443{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004444 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004445}
4446EXPORT_SYMBOL(wait_for_completion_timeout);
4447
Andi Kleen8cbbe862007-10-15 17:00:14 +02004448int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004449{
Andi Kleen51e97992007-10-18 21:32:55 +02004450 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4451 if (t == -ERESTARTSYS)
4452 return t;
4453 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004454}
4455EXPORT_SYMBOL(wait_for_completion_interruptible);
4456
Ingo Molnarb15136e2007-10-24 18:23:48 +02004457unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004458wait_for_completion_interruptible_timeout(struct completion *x,
4459 unsigned long timeout)
4460{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004461 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004462}
4463EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4464
Matthew Wilcox009e5772007-12-06 12:29:54 -05004465int __sched wait_for_completion_killable(struct completion *x)
4466{
4467 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4468 if (t == -ERESTARTSYS)
4469 return t;
4470 return 0;
4471}
4472EXPORT_SYMBOL(wait_for_completion_killable);
4473
Andi Kleen8cbbe862007-10-15 17:00:14 +02004474static long __sched
4475sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02004476{
4477 unsigned long flags;
4478 wait_queue_t wait;
4479
4480 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481
Andi Kleen8cbbe862007-10-15 17:00:14 +02004482 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004483
Andi Kleen8cbbe862007-10-15 17:00:14 +02004484 spin_lock_irqsave(&q->lock, flags);
4485 __add_wait_queue(q, &wait);
4486 spin_unlock(&q->lock);
4487 timeout = schedule_timeout(timeout);
4488 spin_lock_irq(&q->lock);
4489 __remove_wait_queue(q, &wait);
4490 spin_unlock_irqrestore(&q->lock, flags);
4491
4492 return timeout;
4493}
4494
4495void __sched interruptible_sleep_on(wait_queue_head_t *q)
4496{
4497 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004498}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004499EXPORT_SYMBOL(interruptible_sleep_on);
4500
Ingo Molnar0fec1712007-07-09 18:52:01 +02004501long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004502interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004503{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004504 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004505}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004506EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4507
Ingo Molnar0fec1712007-07-09 18:52:01 +02004508void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004509{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004510 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004511}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004512EXPORT_SYMBOL(sleep_on);
4513
Ingo Molnar0fec1712007-07-09 18:52:01 +02004514long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004515{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004516 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004517}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004518EXPORT_SYMBOL(sleep_on_timeout);
4519
Ingo Molnarb29739f2006-06-27 02:54:51 -07004520#ifdef CONFIG_RT_MUTEXES
4521
4522/*
4523 * rt_mutex_setprio - set the current priority of a task
4524 * @p: task
4525 * @prio: prio value (kernel-internal form)
4526 *
4527 * This function changes the 'effective' priority of a task. It does
4528 * not touch ->normal_prio like __setscheduler().
4529 *
4530 * Used by the rt_mutex code to implement priority inheritance logic.
4531 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004532void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07004533{
4534 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004535 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004536 struct rq *rq;
Steven Rostedtcb469842008-01-25 21:08:22 +01004537 const struct sched_class *prev_class = p->sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004538
4539 BUG_ON(prio < 0 || prio > MAX_PRIO);
4540
4541 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004542 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004543
Andrew Mortond5f9f942007-05-08 20:27:06 -07004544 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004545 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004546 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004547 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02004548 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004549 if (running)
4550 p->sched_class->put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02004551
4552 if (rt_prio(prio))
4553 p->sched_class = &rt_sched_class;
4554 else
4555 p->sched_class = &fair_sched_class;
4556
Ingo Molnarb29739f2006-06-27 02:54:51 -07004557 p->prio = prio;
4558
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004559 if (running)
4560 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004561 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004562 enqueue_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004563
4564 check_class_changed(rq, p, prev_class, oldprio, running);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004565 }
4566 task_rq_unlock(rq, &flags);
4567}
4568
4569#endif
4570
Ingo Molnar36c8b582006-07-03 00:25:41 -07004571void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004572{
Ingo Molnardd41f592007-07-09 18:51:59 +02004573 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004574 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004575 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004576
4577 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4578 return;
4579 /*
4580 * We have to be careful, if called from sys_setpriority(),
4581 * the task might be in the middle of scheduling on another CPU.
4582 */
4583 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004584 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004585 /*
4586 * The RT priorities are set via sched_setscheduler(), but we still
4587 * allow the 'normal' nice value to be set - but as expected
4588 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02004589 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004590 */
Ingo Molnare05606d2007-07-09 18:51:59 +02004591 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004592 p->static_prio = NICE_TO_PRIO(nice);
4593 goto out_unlock;
4594 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004595 on_rq = p->se.on_rq;
Ingo Molnar6363ca52008-05-29 11:28:57 +02004596 if (on_rq) {
Ingo Molnar69be72c2007-08-09 11:16:49 +02004597 dequeue_task(rq, p, 0);
Ingo Molnar6363ca52008-05-29 11:28:57 +02004598 dec_load(rq, p);
4599 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004600
Linus Torvalds1da177e2005-04-16 15:20:36 -07004601 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07004602 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004603 old_prio = p->prio;
4604 p->prio = effective_prio(p);
4605 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004606
Ingo Molnardd41f592007-07-09 18:51:59 +02004607 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004608 enqueue_task(rq, p, 0);
Ingo Molnar6363ca52008-05-29 11:28:57 +02004609 inc_load(rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004610 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07004611 * If the task increased its priority or is running and
4612 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004613 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07004614 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004615 resched_task(rq->curr);
4616 }
4617out_unlock:
4618 task_rq_unlock(rq, &flags);
4619}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004620EXPORT_SYMBOL(set_user_nice);
4621
Matt Mackalle43379f2005-05-01 08:59:00 -07004622/*
4623 * can_nice - check if a task can reduce its nice value
4624 * @p: task
4625 * @nice: nice value
4626 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004627int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07004628{
Matt Mackall024f4742005-08-18 11:24:19 -07004629 /* convert nice value [19,-20] to rlimit style value [1,40] */
4630 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004631
Matt Mackalle43379f2005-05-01 08:59:00 -07004632 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4633 capable(CAP_SYS_NICE));
4634}
4635
Linus Torvalds1da177e2005-04-16 15:20:36 -07004636#ifdef __ARCH_WANT_SYS_NICE
4637
4638/*
4639 * sys_nice - change the priority of the current process.
4640 * @increment: priority increment
4641 *
4642 * sys_setpriority is a more generic, but much slower function that
4643 * does similar things.
4644 */
4645asmlinkage long sys_nice(int increment)
4646{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004647 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004648
4649 /*
4650 * Setpriority might change our priority at the same moment.
4651 * We don't have to worry. Conceptually one call occurs first
4652 * and we have a single winner.
4653 */
Matt Mackalle43379f2005-05-01 08:59:00 -07004654 if (increment < -40)
4655 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004656 if (increment > 40)
4657 increment = 40;
4658
4659 nice = PRIO_TO_NICE(current->static_prio) + increment;
4660 if (nice < -20)
4661 nice = -20;
4662 if (nice > 19)
4663 nice = 19;
4664
Matt Mackalle43379f2005-05-01 08:59:00 -07004665 if (increment < 0 && !can_nice(current, nice))
4666 return -EPERM;
4667
Linus Torvalds1da177e2005-04-16 15:20:36 -07004668 retval = security_task_setnice(current, nice);
4669 if (retval)
4670 return retval;
4671
4672 set_user_nice(current, nice);
4673 return 0;
4674}
4675
4676#endif
4677
4678/**
4679 * task_prio - return the priority value of a given task.
4680 * @p: the task in question.
4681 *
4682 * This is the priority value as seen by users in /proc.
4683 * RT tasks are offset by -200. Normal tasks are centered
4684 * around 0, value goes from -16 to +15.
4685 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004686int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004687{
4688 return p->prio - MAX_RT_PRIO;
4689}
4690
4691/**
4692 * task_nice - return the nice value of a given task.
4693 * @p: the task in question.
4694 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004695int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004696{
4697 return TASK_NICE(p);
4698}
Pavel Roskin150d8be2008-03-05 16:56:37 -05004699EXPORT_SYMBOL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004700
4701/**
4702 * idle_cpu - is a given cpu idle currently?
4703 * @cpu: the processor in question.
4704 */
4705int idle_cpu(int cpu)
4706{
4707 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4708}
4709
Linus Torvalds1da177e2005-04-16 15:20:36 -07004710/**
4711 * idle_task - return the idle task for a given cpu.
4712 * @cpu: the processor in question.
4713 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004714struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004715{
4716 return cpu_rq(cpu)->idle;
4717}
4718
4719/**
4720 * find_process_by_pid - find a process with a matching PID value.
4721 * @pid: the pid in question.
4722 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02004723static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004724{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07004725 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004726}
4727
4728/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02004729static void
4730__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004731{
Ingo Molnardd41f592007-07-09 18:51:59 +02004732 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004733
Linus Torvalds1da177e2005-04-16 15:20:36 -07004734 p->policy = policy;
Ingo Molnardd41f592007-07-09 18:51:59 +02004735 switch (p->policy) {
4736 case SCHED_NORMAL:
4737 case SCHED_BATCH:
4738 case SCHED_IDLE:
4739 p->sched_class = &fair_sched_class;
4740 break;
4741 case SCHED_FIFO:
4742 case SCHED_RR:
4743 p->sched_class = &rt_sched_class;
4744 break;
4745 }
4746
Linus Torvalds1da177e2005-04-16 15:20:36 -07004747 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004748 p->normal_prio = normal_prio(p);
4749 /* we are holding p->pi_lock already */
4750 p->prio = rt_mutex_getprio(p);
Peter Williams2dd73a42006-06-27 02:54:34 -07004751 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004752}
4753
4754/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004755 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004756 * @p: the task in question.
4757 * @policy: new policy.
4758 * @param: structure containing the new RT priority.
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004759 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004760 * NOTE that the task may be already dead.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004761 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004762int sched_setscheduler(struct task_struct *p, int policy,
4763 struct sched_param *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004764{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004765 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004766 unsigned long flags;
Steven Rostedtcb469842008-01-25 21:08:22 +01004767 const struct sched_class *prev_class = p->sched_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004768 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004769
Steven Rostedt66e53932006-06-27 02:54:44 -07004770 /* may grab non-irq protected spin_locks */
4771 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07004772recheck:
4773 /* double check policy once rq lock held */
4774 if (policy < 0)
4775 policy = oldpolicy = p->policy;
4776 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
Ingo Molnardd41f592007-07-09 18:51:59 +02004777 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4778 policy != SCHED_IDLE)
Ingo Molnarb0a94992006-01-14 13:20:41 -08004779 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004780 /*
4781 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02004782 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4783 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004784 */
4785 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004786 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04004787 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004788 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02004789 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004790 return -EINVAL;
4791
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004792 /*
4793 * Allow unprivileged RT tasks to decrease priority:
4794 */
4795 if (!capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02004796 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004797 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004798
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004799 if (!lock_task_sighand(p, &flags))
4800 return -ESRCH;
4801 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4802 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004803
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004804 /* can't set/change the rt policy */
4805 if (policy != p->policy && !rlim_rtprio)
4806 return -EPERM;
4807
4808 /* can't increase priority */
4809 if (param->sched_priority > p->rt_priority &&
4810 param->sched_priority > rlim_rtprio)
4811 return -EPERM;
4812 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004813 /*
4814 * Like positive nice levels, dont allow tasks to
4815 * move out of SCHED_IDLE either:
4816 */
4817 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4818 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004819
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004820 /* can't change other user's priorities */
4821 if ((current->euid != p->euid) &&
4822 (current->euid != p->uid))
4823 return -EPERM;
4824 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004825
Peter Zijlstrab68aa232008-02-13 15:45:40 +01004826#ifdef CONFIG_RT_GROUP_SCHED
4827 /*
4828 * Do not allow realtime tasks into groups that have no runtime
4829 * assigned.
4830 */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02004831 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
Peter Zijlstrab68aa232008-02-13 15:45:40 +01004832 return -EPERM;
4833#endif
4834
Linus Torvalds1da177e2005-04-16 15:20:36 -07004835 retval = security_task_setscheduler(p, policy, param);
4836 if (retval)
4837 return retval;
4838 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07004839 * make sure no PI-waiters arrive (or leave) while we are
4840 * changing the priority of the task:
4841 */
4842 spin_lock_irqsave(&p->pi_lock, flags);
4843 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004844 * To be able to change p->policy safely, the apropriate
4845 * runqueue lock must be held.
4846 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07004847 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004848 /* recheck policy now with rq lock held */
4849 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4850 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004851 __task_rq_unlock(rq);
4852 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004853 goto recheck;
4854 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02004855 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004856 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004857 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004858 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004859 deactivate_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004860 if (running)
4861 p->sched_class->put_prev_task(rq, p);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02004862
Linus Torvalds1da177e2005-04-16 15:20:36 -07004863 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004864 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02004865
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004866 if (running)
4867 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004868 if (on_rq) {
4869 activate_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004870
4871 check_class_changed(rq, p, prev_class, oldprio, running);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004872 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07004873 __task_rq_unlock(rq);
4874 spin_unlock_irqrestore(&p->pi_lock, flags);
4875
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07004876 rt_mutex_adjust_pi(p);
4877
Linus Torvalds1da177e2005-04-16 15:20:36 -07004878 return 0;
4879}
4880EXPORT_SYMBOL_GPL(sched_setscheduler);
4881
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004882static int
4883do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004884{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004885 struct sched_param lparam;
4886 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004887 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004888
4889 if (!param || pid < 0)
4890 return -EINVAL;
4891 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4892 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004893
4894 rcu_read_lock();
4895 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004896 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004897 if (p != NULL)
4898 retval = sched_setscheduler(p, policy, &lparam);
4899 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07004900
Linus Torvalds1da177e2005-04-16 15:20:36 -07004901 return retval;
4902}
4903
4904/**
4905 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4906 * @pid: the pid in question.
4907 * @policy: new policy.
4908 * @param: structure containing the new RT priority.
4909 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004910asmlinkage long
4911sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004912{
Jason Baronc21761f2006-01-18 17:43:03 -08004913 /* negative values for policy are not valid */
4914 if (policy < 0)
4915 return -EINVAL;
4916
Linus Torvalds1da177e2005-04-16 15:20:36 -07004917 return do_sched_setscheduler(pid, policy, param);
4918}
4919
4920/**
4921 * sys_sched_setparam - set/change the RT priority of a thread
4922 * @pid: the pid in question.
4923 * @param: structure containing the new RT priority.
4924 */
4925asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4926{
4927 return do_sched_setscheduler(pid, -1, param);
4928}
4929
4930/**
4931 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4932 * @pid: the pid in question.
4933 */
4934asmlinkage long sys_sched_getscheduler(pid_t pid)
4935{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004936 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004937 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004938
4939 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004940 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004941
4942 retval = -ESRCH;
4943 read_lock(&tasklist_lock);
4944 p = find_process_by_pid(pid);
4945 if (p) {
4946 retval = security_task_getscheduler(p);
4947 if (!retval)
4948 retval = p->policy;
4949 }
4950 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004951 return retval;
4952}
4953
4954/**
4955 * sys_sched_getscheduler - get the RT priority of a thread
4956 * @pid: the pid in question.
4957 * @param: structure containing the RT priority.
4958 */
4959asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4960{
4961 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004962 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004963 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004964
4965 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004966 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004967
4968 read_lock(&tasklist_lock);
4969 p = find_process_by_pid(pid);
4970 retval = -ESRCH;
4971 if (!p)
4972 goto out_unlock;
4973
4974 retval = security_task_getscheduler(p);
4975 if (retval)
4976 goto out_unlock;
4977
4978 lp.sched_priority = p->rt_priority;
4979 read_unlock(&tasklist_lock);
4980
4981 /*
4982 * This one might sleep, we cannot do it with a spinlock held ...
4983 */
4984 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4985
Linus Torvalds1da177e2005-04-16 15:20:36 -07004986 return retval;
4987
4988out_unlock:
4989 read_unlock(&tasklist_lock);
4990 return retval;
4991}
4992
Mike Travisb53e9212008-04-04 18:11:08 -07004993long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004994{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004995 cpumask_t cpus_allowed;
Mike Travisb53e9212008-04-04 18:11:08 -07004996 cpumask_t new_mask = *in_mask;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004997 struct task_struct *p;
4998 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004999
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005000 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005001 read_lock(&tasklist_lock);
5002
5003 p = find_process_by_pid(pid);
5004 if (!p) {
5005 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005006 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005007 return -ESRCH;
5008 }
5009
5010 /*
5011 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005012 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005013 * usage count and then drop tasklist_lock.
5014 */
5015 get_task_struct(p);
5016 read_unlock(&tasklist_lock);
5017
5018 retval = -EPERM;
5019 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5020 !capable(CAP_SYS_NICE))
5021 goto out_unlock;
5022
David Quigleye7834f82006-06-23 02:03:59 -07005023 retval = security_task_setscheduler(p, 0, NULL);
5024 if (retval)
5025 goto out_unlock;
5026
Mike Travisf9a86fc2008-04-04 18:11:07 -07005027 cpuset_cpus_allowed(p, &cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005028 cpus_and(new_mask, new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07005029 again:
Mike Travis7c16ec52008-04-04 18:11:11 -07005030 retval = set_cpus_allowed_ptr(p, &new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005031
Paul Menage8707d8b2007-10-18 23:40:22 -07005032 if (!retval) {
Mike Travisf9a86fc2008-04-04 18:11:07 -07005033 cpuset_cpus_allowed(p, &cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07005034 if (!cpus_subset(new_mask, cpus_allowed)) {
5035 /*
5036 * We must have raced with a concurrent cpuset
5037 * update. Just reset the cpus_allowed to the
5038 * cpuset's cpus_allowed
5039 */
5040 new_mask = cpus_allowed;
5041 goto again;
5042 }
5043 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005044out_unlock:
5045 put_task_struct(p);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005046 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005047 return retval;
5048}
5049
5050static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5051 cpumask_t *new_mask)
5052{
5053 if (len < sizeof(cpumask_t)) {
5054 memset(new_mask, 0, sizeof(cpumask_t));
5055 } else if (len > sizeof(cpumask_t)) {
5056 len = sizeof(cpumask_t);
5057 }
5058 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5059}
5060
5061/**
5062 * sys_sched_setaffinity - set the cpu affinity of a process
5063 * @pid: pid of the process
5064 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5065 * @user_mask_ptr: user-space pointer to the new cpu mask
5066 */
5067asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5068 unsigned long __user *user_mask_ptr)
5069{
5070 cpumask_t new_mask;
5071 int retval;
5072
5073 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5074 if (retval)
5075 return retval;
5076
Mike Travisb53e9212008-04-04 18:11:08 -07005077 return sched_setaffinity(pid, &new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005078}
5079
5080/*
5081 * Represents all cpu's present in the system
5082 * In systems capable of hotplug, this map could dynamically grow
5083 * as new cpu's are detected in the system via any platform specific
5084 * method, such as ACPI for e.g.
5085 */
5086
Andi Kleen4cef0c62006-01-11 22:44:57 +01005087cpumask_t cpu_present_map __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005088EXPORT_SYMBOL(cpu_present_map);
5089
5090#ifndef CONFIG_SMP
Andi Kleen4cef0c62006-01-11 22:44:57 +01005091cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07005092EXPORT_SYMBOL(cpu_online_map);
5093
Andi Kleen4cef0c62006-01-11 22:44:57 +01005094cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07005095EXPORT_SYMBOL(cpu_possible_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005096#endif
5097
5098long sched_getaffinity(pid_t pid, cpumask_t *mask)
5099{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005100 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005101 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005102
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005103 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005104 read_lock(&tasklist_lock);
5105
5106 retval = -ESRCH;
5107 p = find_process_by_pid(pid);
5108 if (!p)
5109 goto out_unlock;
5110
David Quigleye7834f82006-06-23 02:03:59 -07005111 retval = security_task_getscheduler(p);
5112 if (retval)
5113 goto out_unlock;
5114
Jack Steiner2f7016d2006-02-01 03:05:18 -08005115 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005116
5117out_unlock:
5118 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005119 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005120
Ulrich Drepper9531b622007-08-09 11:16:46 +02005121 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005122}
5123
5124/**
5125 * sys_sched_getaffinity - get the cpu affinity of a process
5126 * @pid: pid of the process
5127 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5128 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5129 */
5130asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5131 unsigned long __user *user_mask_ptr)
5132{
5133 int ret;
5134 cpumask_t mask;
5135
5136 if (len < sizeof(cpumask_t))
5137 return -EINVAL;
5138
5139 ret = sched_getaffinity(pid, &mask);
5140 if (ret < 0)
5141 return ret;
5142
5143 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5144 return -EFAULT;
5145
5146 return sizeof(cpumask_t);
5147}
5148
5149/**
5150 * sys_sched_yield - yield the current processor to other threads.
5151 *
Ingo Molnardd41f592007-07-09 18:51:59 +02005152 * This function yields the current CPU to other tasks. If there are no
5153 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005154 */
5155asmlinkage long sys_sched_yield(void)
5156{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005157 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005158
Ingo Molnar2d723762007-10-15 17:00:12 +02005159 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02005160 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005161
5162 /*
5163 * Since we are going to call schedule() anyway, there's
5164 * no need to preempt or enable interrupts:
5165 */
5166 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07005167 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005168 _raw_spin_unlock(&rq->lock);
5169 preempt_enable_no_resched();
5170
5171 schedule();
5172
5173 return 0;
5174}
5175
Andrew Mortone7b38402006-06-30 01:56:00 -07005176static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005177{
Ingo Molnar8e0a43d2006-06-23 02:05:23 -07005178#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5179 __might_sleep(__FILE__, __LINE__);
5180#endif
Ingo Molnar5bbcfd92005-07-07 17:57:04 -07005181 /*
5182 * The BKS might be reacquired before we have dropped
5183 * PREEMPT_ACTIVE, which could trigger a second
5184 * cond_resched() call.
5185 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005186 do {
5187 add_preempt_count(PREEMPT_ACTIVE);
5188 schedule();
5189 sub_preempt_count(PREEMPT_ACTIVE);
5190 } while (need_resched());
5191}
5192
Herbert Xu02b67cc32008-01-25 21:08:28 +01005193int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005194{
Ingo Molnar94142322006-12-29 16:48:13 -08005195 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5196 system_state == SYSTEM_RUNNING) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005197 __cond_resched();
5198 return 1;
5199 }
5200 return 0;
5201}
Herbert Xu02b67cc32008-01-25 21:08:28 +01005202EXPORT_SYMBOL(_cond_resched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005203
5204/*
5205 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5206 * call schedule, and on return reacquire the lock.
5207 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005208 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07005209 * operations here to prevent schedule() from being called twice (once via
5210 * spin_unlock(), once by hand).
5211 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005212int cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005213{
Nick Piggin95c354f2008-01-30 13:31:20 +01005214 int resched = need_resched() && system_state == SYSTEM_RUNNING;
Jan Kara6df3cec2005-06-13 15:52:32 -07005215 int ret = 0;
5216
Nick Piggin95c354f2008-01-30 13:31:20 +01005217 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005218 spin_unlock(lock);
Nick Piggin95c354f2008-01-30 13:31:20 +01005219 if (resched && need_resched())
5220 __cond_resched();
5221 else
5222 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07005223 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005224 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005225 }
Jan Kara6df3cec2005-06-13 15:52:32 -07005226 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005227}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005228EXPORT_SYMBOL(cond_resched_lock);
5229
5230int __sched cond_resched_softirq(void)
5231{
5232 BUG_ON(!in_softirq());
5233
Ingo Molnar94142322006-12-29 16:48:13 -08005234 if (need_resched() && system_state == SYSTEM_RUNNING) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07005235 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005236 __cond_resched();
5237 local_bh_disable();
5238 return 1;
5239 }
5240 return 0;
5241}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005242EXPORT_SYMBOL(cond_resched_softirq);
5243
Linus Torvalds1da177e2005-04-16 15:20:36 -07005244/**
5245 * yield - yield the current processor to other threads.
5246 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08005247 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005248 * thread runnable and calls sys_sched_yield().
5249 */
5250void __sched yield(void)
5251{
5252 set_current_state(TASK_RUNNING);
5253 sys_sched_yield();
5254}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005255EXPORT_SYMBOL(yield);
5256
5257/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005258 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07005259 * that process accounting knows that this is a task in IO wait state.
5260 *
5261 * But don't do that if it is a deliberate, throttling IO wait (this task
5262 * has set its backing_dev_info: the queue against which it should throttle)
5263 */
5264void __sched io_schedule(void)
5265{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005266 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005267
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005268 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005269 atomic_inc(&rq->nr_iowait);
5270 schedule();
5271 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005272 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005273}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005274EXPORT_SYMBOL(io_schedule);
5275
5276long __sched io_schedule_timeout(long timeout)
5277{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005278 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005279 long ret;
5280
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005281 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005282 atomic_inc(&rq->nr_iowait);
5283 ret = schedule_timeout(timeout);
5284 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005285 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005286 return ret;
5287}
5288
5289/**
5290 * sys_sched_get_priority_max - return maximum RT priority.
5291 * @policy: scheduling class.
5292 *
5293 * this syscall returns the maximum rt_priority that can be used
5294 * by a given scheduling class.
5295 */
5296asmlinkage long sys_sched_get_priority_max(int policy)
5297{
5298 int ret = -EINVAL;
5299
5300 switch (policy) {
5301 case SCHED_FIFO:
5302 case SCHED_RR:
5303 ret = MAX_USER_RT_PRIO-1;
5304 break;
5305 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005306 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005307 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005308 ret = 0;
5309 break;
5310 }
5311 return ret;
5312}
5313
5314/**
5315 * sys_sched_get_priority_min - return minimum RT priority.
5316 * @policy: scheduling class.
5317 *
5318 * this syscall returns the minimum rt_priority that can be used
5319 * by a given scheduling class.
5320 */
5321asmlinkage long sys_sched_get_priority_min(int policy)
5322{
5323 int ret = -EINVAL;
5324
5325 switch (policy) {
5326 case SCHED_FIFO:
5327 case SCHED_RR:
5328 ret = 1;
5329 break;
5330 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005331 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005332 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005333 ret = 0;
5334 }
5335 return ret;
5336}
5337
5338/**
5339 * sys_sched_rr_get_interval - return the default timeslice of a process.
5340 * @pid: pid of the process.
5341 * @interval: userspace pointer to the timeslice value.
5342 *
5343 * this syscall writes the default timeslice value of a given process
5344 * into the user-space timespec buffer. A value of '0' means infinity.
5345 */
5346asmlinkage
5347long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5348{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005349 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005350 unsigned int time_slice;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005351 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005352 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005353
5354 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005355 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005356
5357 retval = -ESRCH;
5358 read_lock(&tasklist_lock);
5359 p = find_process_by_pid(pid);
5360 if (!p)
5361 goto out_unlock;
5362
5363 retval = security_task_getscheduler(p);
5364 if (retval)
5365 goto out_unlock;
5366
Ingo Molnar77034932007-12-04 17:04:39 +01005367 /*
5368 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5369 * tasks that are on an otherwise idle runqueue:
5370 */
5371 time_slice = 0;
5372 if (p->policy == SCHED_RR) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005373 time_slice = DEF_TIMESLICE;
Miao Xie1868f952008-03-07 09:35:06 +08005374 } else if (p->policy != SCHED_FIFO) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005375 struct sched_entity *se = &p->se;
5376 unsigned long flags;
5377 struct rq *rq;
5378
5379 rq = task_rq_lock(p, &flags);
Ingo Molnar77034932007-12-04 17:04:39 +01005380 if (rq->cfs.load.weight)
5381 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005382 task_rq_unlock(rq, &flags);
5383 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005384 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005385 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005386 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005387 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005388
Linus Torvalds1da177e2005-04-16 15:20:36 -07005389out_unlock:
5390 read_unlock(&tasklist_lock);
5391 return retval;
5392}
5393
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005394static const char stat_nam[] = "RSDTtZX";
Ingo Molnar36c8b582006-07-03 00:25:41 -07005395
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005396void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005397{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005398 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005399 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005400
Linus Torvalds1da177e2005-04-16 15:20:36 -07005401 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005402 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005403 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02005404#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07005405 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005406 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005407 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005408 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005409#else
5410 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005411 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005412 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005413 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005414#endif
5415#ifdef CONFIG_DEBUG_STACK_USAGE
5416 {
Al Viro10ebffd2005-11-13 16:06:56 -08005417 unsigned long *n = end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005418 while (!*n)
5419 n++;
Al Viro10ebffd2005-11-13 16:06:56 -08005420 free = (unsigned long)n - (unsigned long)end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005421 }
5422#endif
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -07005423 printk(KERN_CONT "%5lu %5d %6d\n", free,
Roland McGrathfcfd50a2008-01-09 00:03:23 -08005424 task_pid_nr(p), task_pid_nr(p->real_parent));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005425
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01005426 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005427}
5428
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005429void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005430{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005431 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005432
Ingo Molnar4bd77322007-07-11 21:21:47 +02005433#if BITS_PER_LONG == 32
5434 printk(KERN_INFO
5435 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005436#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02005437 printk(KERN_INFO
5438 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005439#endif
5440 read_lock(&tasklist_lock);
5441 do_each_thread(g, p) {
5442 /*
5443 * reset the NMI-timeout, listing all files on a slow
5444 * console might take alot of time:
5445 */
5446 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07005447 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005448 sched_show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005449 } while_each_thread(g, p);
5450
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07005451 touch_all_softlockup_watchdogs();
5452
Ingo Molnardd41f592007-07-09 18:51:59 +02005453#ifdef CONFIG_SCHED_DEBUG
5454 sysrq_sched_debug_show();
5455#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005456 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005457 /*
5458 * Only show locks if all tasks are dumped:
5459 */
5460 if (state_filter == -1)
5461 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005462}
5463
Ingo Molnar1df21052007-07-09 18:51:58 +02005464void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5465{
Ingo Molnardd41f592007-07-09 18:51:59 +02005466 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02005467}
5468
Ingo Molnarf340c0d2005-06-28 16:40:42 +02005469/**
5470 * init_idle - set up an idle thread for a given CPU
5471 * @idle: task in question
5472 * @cpu: cpu the idle task belongs to
5473 *
5474 * NOTE: this function does not set the idle thread's NEED_RESCHED
5475 * flag, to make booting more robust.
5476 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07005477void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005478{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005479 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005480 unsigned long flags;
5481
Ingo Molnardd41f592007-07-09 18:51:59 +02005482 __sched_fork(idle);
5483 idle->se.exec_start = sched_clock();
5484
Ingo Molnarb29739f2006-06-27 02:54:51 -07005485 idle->prio = idle->normal_prio = MAX_PRIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005486 idle->cpus_allowed = cpumask_of_cpu(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005487 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005488
5489 spin_lock_irqsave(&rq->lock, flags);
5490 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07005491#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5492 idle->oncpu = 1;
5493#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005494 spin_unlock_irqrestore(&rq->lock, flags);
5495
5496 /* Set the preempt count _outside_ the spinlocks! */
Linus Torvalds8e3e0762008-05-10 20:58:02 -07005497#if defined(CONFIG_PREEMPT)
5498 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5499#else
Al Viroa1261f52005-11-13 16:06:55 -08005500 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds8e3e0762008-05-10 20:58:02 -07005501#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02005502 /*
5503 * The idle tasks have their own, simple scheduling class:
5504 */
5505 idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005506}
5507
5508/*
5509 * In a system that switches off the HZ timer nohz_cpu_mask
5510 * indicates which cpus entered this state. This is used
5511 * in the rcu update to wait only for active cpus. For system
5512 * which do not switch off the HZ timer nohz_cpu_mask should
5513 * always be CPU_MASK_NONE.
5514 */
5515cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5516
Ingo Molnar19978ca2007-11-09 22:39:38 +01005517/*
5518 * Increase the granularity value when there are more CPUs,
5519 * because with more CPUs the 'effective latency' as visible
5520 * to users decreases. But the relationship is not linear,
5521 * so pick a second-best guess by going with the log2 of the
5522 * number of CPUs.
5523 *
5524 * This idea comes from the SD scheduler of Con Kolivas:
5525 */
5526static inline void sched_init_granularity(void)
5527{
5528 unsigned int factor = 1 + ilog2(num_online_cpus());
5529 const unsigned long limit = 200000000;
5530
5531 sysctl_sched_min_granularity *= factor;
5532 if (sysctl_sched_min_granularity > limit)
5533 sysctl_sched_min_granularity = limit;
5534
5535 sysctl_sched_latency *= factor;
5536 if (sysctl_sched_latency > limit)
5537 sysctl_sched_latency = limit;
5538
5539 sysctl_sched_wakeup_granularity *= factor;
Ingo Molnar19978ca2007-11-09 22:39:38 +01005540}
5541
Linus Torvalds1da177e2005-04-16 15:20:36 -07005542#ifdef CONFIG_SMP
5543/*
5544 * This is how migration works:
5545 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07005546 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005547 * runqueue and wake up that CPU's migration thread.
5548 * 2) we down() the locked semaphore => thread blocks.
5549 * 3) migration thread wakes up (implicitly it forces the migrated
5550 * thread off the CPU)
5551 * 4) it gets the migration request and checks whether the migrated
5552 * task is still in the wrong runqueue.
5553 * 5) if it's in the wrong runqueue then the migration thread removes
5554 * it and puts it into the right queue.
5555 * 6) migration thread up()s the semaphore.
5556 * 7) we wake up and the migration is done.
5557 */
5558
5559/*
5560 * Change a given task's CPU affinity. Migrate the thread to a
5561 * proper CPU and schedule it away if the CPU it's executing on
5562 * is removed from the allowed bitmask.
5563 *
5564 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005565 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07005566 * call is not atomic; no spinlocks may be held.
5567 */
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005568int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005569{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005570 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005571 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005572 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005573 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005574
5575 rq = task_rq_lock(p, &flags);
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005576 if (!cpus_intersects(*new_mask, cpu_online_map)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005577 ret = -EINVAL;
5578 goto out;
5579 }
5580
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005581 if (p->sched_class->set_cpus_allowed)
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005582 p->sched_class->set_cpus_allowed(p, new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005583 else {
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005584 p->cpus_allowed = *new_mask;
5585 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005586 }
5587
Linus Torvalds1da177e2005-04-16 15:20:36 -07005588 /* Can the task run on the task's current CPU? If so, we're done */
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005589 if (cpu_isset(task_cpu(p), *new_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005590 goto out;
5591
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005592 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005593 /* Need help from migration thread: drop lock and wait. */
5594 task_rq_unlock(rq, &flags);
5595 wake_up_process(rq->migration_thread);
5596 wait_for_completion(&req.done);
5597 tlb_migrate_finish(p->mm);
5598 return 0;
5599 }
5600out:
5601 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005602
Linus Torvalds1da177e2005-04-16 15:20:36 -07005603 return ret;
5604}
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005605EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005606
5607/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005608 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07005609 * this because either it can't run here any more (set_cpus_allowed()
5610 * away from this CPU, or CPU going down), or because we're
5611 * attempting to rebalance this task on exec (sched_exec).
5612 *
5613 * So we race with normal scheduler movements, but that's OK, as long
5614 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07005615 *
5616 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005617 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07005618static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005619{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005620 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02005621 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005622
5623 if (unlikely(cpu_is_offline(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07005624 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005625
5626 rq_src = cpu_rq(src_cpu);
5627 rq_dest = cpu_rq(dest_cpu);
5628
5629 double_rq_lock(rq_src, rq_dest);
5630 /* Already moved. */
5631 if (task_cpu(p) != src_cpu)
5632 goto out;
5633 /* Affinity changed (again). */
5634 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5635 goto out;
5636
Ingo Molnardd41f592007-07-09 18:51:59 +02005637 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02005638 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005639 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02005640
Linus Torvalds1da177e2005-04-16 15:20:36 -07005641 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005642 if (on_rq) {
5643 activate_task(rq_dest, p, 0);
5644 check_preempt_curr(rq_dest, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005645 }
Kirill Korotaevefc30812006-06-27 02:54:32 -07005646 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005647out:
5648 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07005649 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005650}
5651
5652/*
5653 * migration_thread - this is a highprio system thread that performs
5654 * thread migration by bumping thread off CPU then 'pushing' onto
5655 * another runqueue.
5656 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005657static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005658{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005659 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005660 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005661
5662 rq = cpu_rq(cpu);
5663 BUG_ON(rq->migration_thread != current);
5664
5665 set_current_state(TASK_INTERRUPTIBLE);
5666 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005667 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005668 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005669
Linus Torvalds1da177e2005-04-16 15:20:36 -07005670 spin_lock_irq(&rq->lock);
5671
5672 if (cpu_is_offline(cpu)) {
5673 spin_unlock_irq(&rq->lock);
5674 goto wait_to_die;
5675 }
5676
5677 if (rq->active_balance) {
5678 active_load_balance(rq, cpu);
5679 rq->active_balance = 0;
5680 }
5681
5682 head = &rq->migration_queue;
5683
5684 if (list_empty(head)) {
5685 spin_unlock_irq(&rq->lock);
5686 schedule();
5687 set_current_state(TASK_INTERRUPTIBLE);
5688 continue;
5689 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07005690 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005691 list_del_init(head->next);
5692
Nick Piggin674311d2005-06-25 14:57:27 -07005693 spin_unlock(&rq->lock);
5694 __migrate_task(req->task, cpu, req->dest_cpu);
5695 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005696
5697 complete(&req->done);
5698 }
5699 __set_current_state(TASK_RUNNING);
5700 return 0;
5701
5702wait_to_die:
5703 /* Wait for kthread_stop */
5704 set_current_state(TASK_INTERRUPTIBLE);
5705 while (!kthread_should_stop()) {
5706 schedule();
5707 set_current_state(TASK_INTERRUPTIBLE);
5708 }
5709 __set_current_state(TASK_RUNNING);
5710 return 0;
5711}
5712
5713#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005714
5715static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5716{
5717 int ret;
5718
5719 local_irq_disable();
5720 ret = __migrate_task(p, src_cpu, dest_cpu);
5721 local_irq_enable();
5722 return ret;
5723}
5724
Kirill Korotaev054b9102006-12-10 02:20:11 -08005725/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02005726 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08005727 * NOTE: interrupts should be disabled by the caller
5728 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005729static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005730{
Kirill Korotaevefc30812006-06-27 02:54:32 -07005731 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005732 cpumask_t mask;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005733 struct rq *rq;
5734 int dest_cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005735
Andi Kleen3a5c3592007-10-15 17:00:14 +02005736 do {
5737 /* On same node? */
5738 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5739 cpus_and(mask, mask, p->cpus_allowed);
5740 dest_cpu = any_online_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005741
Andi Kleen3a5c3592007-10-15 17:00:14 +02005742 /* On any allowed CPU? */
Mike Travis434d53b2008-04-04 18:11:04 -07005743 if (dest_cpu >= nr_cpu_ids)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005744 dest_cpu = any_online_cpu(p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005745
Andi Kleen3a5c3592007-10-15 17:00:14 +02005746 /* No more Mr. Nice Guy. */
Mike Travis434d53b2008-04-04 18:11:04 -07005747 if (dest_cpu >= nr_cpu_ids) {
Mike Travisf9a86fc2008-04-04 18:11:07 -07005748 cpumask_t cpus_allowed;
5749
5750 cpuset_cpus_allowed_locked(p, &cpus_allowed);
Cliff Wickman470fd642007-10-18 23:40:46 -07005751 /*
5752 * Try to stay on the same cpuset, where the
5753 * current cpuset may be a subset of all cpus.
5754 * The cpuset_cpus_allowed_locked() variant of
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005755 * cpuset_cpus_allowed() will not block. It must be
Cliff Wickman470fd642007-10-18 23:40:46 -07005756 * called within calls to cpuset_lock/cpuset_unlock.
5757 */
Andi Kleen3a5c3592007-10-15 17:00:14 +02005758 rq = task_rq_lock(p, &flags);
Cliff Wickman470fd642007-10-18 23:40:46 -07005759 p->cpus_allowed = cpus_allowed;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005760 dest_cpu = any_online_cpu(p->cpus_allowed);
5761 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005762
Andi Kleen3a5c3592007-10-15 17:00:14 +02005763 /*
5764 * Don't tell them about moving exiting tasks or
5765 * kernel threads (both mm NULL), since they never
5766 * leave kernel.
5767 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005768 if (p->mm && printk_ratelimit()) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02005769 printk(KERN_INFO "process %d (%s) no "
5770 "longer affine to cpu%d\n",
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005771 task_pid_nr(p), p->comm, dead_cpu);
5772 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02005773 }
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005774 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005775}
5776
5777/*
5778 * While a dead CPU has no uninterruptible tasks queued at this point,
5779 * it might still have a nonzero ->nr_uninterruptible counter, because
5780 * for performance reasons the counter is not stricly tracking tasks to
5781 * their home CPUs. So we just add the counter to another CPU's counter,
5782 * to keep the global sum constant after CPU-down:
5783 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07005784static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005785{
Mike Travis7c16ec52008-04-04 18:11:11 -07005786 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005787 unsigned long flags;
5788
5789 local_irq_save(flags);
5790 double_rq_lock(rq_src, rq_dest);
5791 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5792 rq_src->nr_uninterruptible = 0;
5793 double_rq_unlock(rq_src, rq_dest);
5794 local_irq_restore(flags);
5795}
5796
5797/* Run through task list and migrate tasks from the dead cpu. */
5798static void migrate_live_tasks(int src_cpu)
5799{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005800 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005801
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005802 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005803
Ingo Molnar48f24c42006-07-03 00:25:40 -07005804 do_each_thread(t, p) {
5805 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005806 continue;
5807
Ingo Molnar48f24c42006-07-03 00:25:40 -07005808 if (task_cpu(p) == src_cpu)
5809 move_task_off_dead_cpu(src_cpu, p);
5810 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005811
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005812 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005813}
5814
Ingo Molnardd41f592007-07-09 18:51:59 +02005815/*
5816 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01005817 * It does so by boosting its priority to highest possible.
5818 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005819 */
5820void sched_idle_next(void)
5821{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005822 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07005823 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005824 struct task_struct *p = rq->idle;
5825 unsigned long flags;
5826
5827 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005828 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005829
Ingo Molnar48f24c42006-07-03 00:25:40 -07005830 /*
5831 * Strictly not necessary since rest of the CPUs are stopped by now
5832 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005833 */
5834 spin_lock_irqsave(&rq->lock, flags);
5835
Ingo Molnardd41f592007-07-09 18:51:59 +02005836 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005837
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01005838 update_rq_clock(rq);
5839 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005840
5841 spin_unlock_irqrestore(&rq->lock, flags);
5842}
5843
Ingo Molnar48f24c42006-07-03 00:25:40 -07005844/*
5845 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07005846 * offline.
5847 */
5848void idle_task_exit(void)
5849{
5850 struct mm_struct *mm = current->active_mm;
5851
5852 BUG_ON(cpu_online(smp_processor_id()));
5853
5854 if (mm != &init_mm)
5855 switch_mm(mm, &init_mm, current);
5856 mmdrop(mm);
5857}
5858
Kirill Korotaev054b9102006-12-10 02:20:11 -08005859/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005860static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005861{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005862 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005863
5864 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07005865 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005866
5867 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07005868 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005869
Ingo Molnar48f24c42006-07-03 00:25:40 -07005870 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005871
5872 /*
5873 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005874 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005875 * fine.
5876 */
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005877 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005878 move_task_off_dead_cpu(dead_cpu, p);
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005879 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005880
Ingo Molnar48f24c42006-07-03 00:25:40 -07005881 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005882}
5883
5884/* release_task() removes task from tasklist, so we won't find dead tasks. */
5885static void migrate_dead_tasks(unsigned int dead_cpu)
5886{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005887 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005888 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005889
Ingo Molnardd41f592007-07-09 18:51:59 +02005890 for ( ; ; ) {
5891 if (!rq->nr_running)
5892 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02005893 update_rq_clock(rq);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02005894 next = pick_next_task(rq, rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02005895 if (!next)
5896 break;
5897 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02005898
Linus Torvalds1da177e2005-04-16 15:20:36 -07005899 }
5900}
5901#endif /* CONFIG_HOTPLUG_CPU */
5902
Nick Piggine692ab52007-07-26 13:40:43 +02005903#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5904
5905static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005906 {
5907 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005908 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005909 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01005910 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02005911};
5912
5913static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005914 {
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005915 .ctl_name = CTL_KERN,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005916 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005917 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005918 .child = sd_ctl_dir,
5919 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01005920 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02005921};
5922
5923static struct ctl_table *sd_alloc_ctl_entry(int n)
5924{
5925 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02005926 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02005927
Nick Piggine692ab52007-07-26 13:40:43 +02005928 return entry;
5929}
5930
Milton Miller6382bc92007-10-15 17:00:19 +02005931static void sd_free_ctl_entry(struct ctl_table **tablep)
5932{
Milton Millercd7900762007-10-17 16:55:11 +02005933 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02005934
Milton Millercd7900762007-10-17 16:55:11 +02005935 /*
5936 * In the intermediate directories, both the child directory and
5937 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005938 * will always be set. In the lowest directory the names are
Milton Millercd7900762007-10-17 16:55:11 +02005939 * static strings and all have proc handlers.
5940 */
5941 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02005942 if (entry->child)
5943 sd_free_ctl_entry(&entry->child);
Milton Millercd7900762007-10-17 16:55:11 +02005944 if (entry->proc_handler == NULL)
5945 kfree(entry->procname);
5946 }
Milton Miller6382bc92007-10-15 17:00:19 +02005947
5948 kfree(*tablep);
5949 *tablep = NULL;
5950}
5951
Nick Piggine692ab52007-07-26 13:40:43 +02005952static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02005953set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02005954 const char *procname, void *data, int maxlen,
5955 mode_t mode, proc_handler *proc_handler)
5956{
Nick Piggine692ab52007-07-26 13:40:43 +02005957 entry->procname = procname;
5958 entry->data = data;
5959 entry->maxlen = maxlen;
5960 entry->mode = mode;
5961 entry->proc_handler = proc_handler;
5962}
5963
5964static struct ctl_table *
5965sd_alloc_ctl_domain_table(struct sched_domain *sd)
5966{
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005967 struct ctl_table *table = sd_alloc_ctl_entry(12);
Nick Piggine692ab52007-07-26 13:40:43 +02005968
Milton Millerad1cdc12007-10-15 17:00:19 +02005969 if (table == NULL)
5970 return NULL;
5971
Alexey Dobriyane0361852007-08-09 11:16:46 +02005972 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02005973 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005974 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02005975 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005976 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005977 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005978 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005979 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005980 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005981 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005982 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005983 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005984 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005985 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005986 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02005987 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005988 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02005989 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005990 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02005991 &sd->cache_nice_tries,
5992 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005993 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02005994 sizeof(int), 0644, proc_dointvec_minmax);
Milton Miller6323469f2007-10-15 17:00:19 +02005995 /* &table[11] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02005996
5997 return table;
5998}
5999
Ingo Molnar9a4e7152007-11-28 15:52:56 +01006000static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02006001{
6002 struct ctl_table *entry, *table;
6003 struct sched_domain *sd;
6004 int domain_num = 0, i;
6005 char buf[32];
6006
6007 for_each_domain(cpu, sd)
6008 domain_num++;
6009 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02006010 if (table == NULL)
6011 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02006012
6013 i = 0;
6014 for_each_domain(cpu, sd) {
6015 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02006016 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006017 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02006018 entry->child = sd_alloc_ctl_domain_table(sd);
6019 entry++;
6020 i++;
6021 }
6022 return table;
6023}
6024
6025static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02006026static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02006027{
6028 int i, cpu_num = num_online_cpus();
6029 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6030 char buf[32];
6031
Milton Miller73785472007-10-24 18:23:48 +02006032 WARN_ON(sd_ctl_dir[0].child);
6033 sd_ctl_dir[0].child = entry;
6034
Milton Millerad1cdc12007-10-15 17:00:19 +02006035 if (entry == NULL)
6036 return;
6037
Milton Miller97b6ea72007-10-15 17:00:19 +02006038 for_each_online_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02006039 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02006040 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006041 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02006042 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02006043 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02006044 }
Milton Miller73785472007-10-24 18:23:48 +02006045
6046 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02006047 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6048}
Milton Miller6382bc92007-10-15 17:00:19 +02006049
Milton Miller73785472007-10-24 18:23:48 +02006050/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02006051static void unregister_sched_domain_sysctl(void)
6052{
Milton Miller73785472007-10-24 18:23:48 +02006053 if (sd_sysctl_header)
6054 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02006055 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02006056 if (sd_ctl_dir[0].child)
6057 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02006058}
Nick Piggine692ab52007-07-26 13:40:43 +02006059#else
Milton Miller6382bc92007-10-15 17:00:19 +02006060static void register_sched_domain_sysctl(void)
6061{
6062}
6063static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02006064{
6065}
6066#endif
6067
Linus Torvalds1da177e2005-04-16 15:20:36 -07006068/*
6069 * migration_call - callback that gets triggered when a CPU is added.
6070 * Here we can start up the necessary migration thread for the new CPU.
6071 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006072static int __cpuinit
6073migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006074{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006075 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006076 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006077 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006078 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006079
6080 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006081
Linus Torvalds1da177e2005-04-16 15:20:36 -07006082 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006083 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02006084 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006085 if (IS_ERR(p))
6086 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006087 kthread_bind(p, cpu);
6088 /* Must be high prio: stop_machine expects to yield to it. */
6089 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02006090 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006091 task_rq_unlock(rq, &flags);
6092 cpu_rq(cpu)->migration_thread = p;
6093 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006094
Linus Torvalds1da177e2005-04-16 15:20:36 -07006095 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006096 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02006097 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006098 wake_up_process(cpu_rq(cpu)->migration_thread);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006099
6100 /* Update our root-domain */
6101 rq = cpu_rq(cpu);
6102 spin_lock_irqsave(&rq->lock, flags);
6103 if (rq->rd) {
6104 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6105 cpu_set(cpu, rq->rd->online);
6106 }
6107 spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006108 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006109
Linus Torvalds1da177e2005-04-16 15:20:36 -07006110#ifdef CONFIG_HOTPLUG_CPU
6111 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006112 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07006113 if (!cpu_rq(cpu)->migration_thread)
6114 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006115 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08006116 kthread_bind(cpu_rq(cpu)->migration_thread,
6117 any_online_cpu(cpu_online_map));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006118 kthread_stop(cpu_rq(cpu)->migration_thread);
6119 cpu_rq(cpu)->migration_thread = NULL;
6120 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006121
Linus Torvalds1da177e2005-04-16 15:20:36 -07006122 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006123 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07006124 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006125 migrate_live_tasks(cpu);
6126 rq = cpu_rq(cpu);
6127 kthread_stop(rq->migration_thread);
6128 rq->migration_thread = NULL;
6129 /* Idle task back to normal (off runqueue, low prio) */
Oleg Nesterovd2da2722007-10-16 23:30:56 -07006130 spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02006131 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02006132 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006133 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02006134 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6135 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006136 migrate_dead_tasks(cpu);
Oleg Nesterovd2da2722007-10-16 23:30:56 -07006137 spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07006138 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006139 migrate_nr_uninterruptible(rq);
6140 BUG_ON(rq->nr_running != 0);
6141
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006142 /*
6143 * No need to migrate the tasks: it was best-effort if
6144 * they didn't take sched_hotcpu_mutex. Just wake up
6145 * the requestors.
6146 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006147 spin_lock_irq(&rq->lock);
6148 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07006149 struct migration_req *req;
6150
Linus Torvalds1da177e2005-04-16 15:20:36 -07006151 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07006152 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006153 list_del_init(&req->list);
6154 complete(&req->done);
6155 }
6156 spin_unlock_irq(&rq->lock);
6157 break;
Gregory Haskins57d885f2008-01-25 21:08:18 +01006158
Gregory Haskins08f503b2008-03-10 17:59:11 -04006159 case CPU_DYING:
6160 case CPU_DYING_FROZEN:
Gregory Haskins57d885f2008-01-25 21:08:18 +01006161 /* Update our root-domain */
6162 rq = cpu_rq(cpu);
6163 spin_lock_irqsave(&rq->lock, flags);
6164 if (rq->rd) {
6165 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6166 cpu_clear(cpu, rq->rd->online);
6167 }
6168 spin_unlock_irqrestore(&rq->lock, flags);
6169 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006170#endif
6171 }
6172 return NOTIFY_OK;
6173}
6174
6175/* Register at highest priority so that task migration (migrate_all_tasks)
6176 * happens before everything else.
6177 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07006178static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006179 .notifier_call = migration_call,
6180 .priority = 10
6181};
6182
Adrian Bunke6fe6642007-11-09 22:39:39 +01006183void __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006184{
6185 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07006186 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006187
6188 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07006189 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6190 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006191 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6192 register_cpu_notifier(&migration_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006193}
6194#endif
6195
6196#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07006197
Ingo Molnar3e9830d2007-10-15 17:00:13 +02006198#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006199
Gautham R Shenoy099f98c2008-05-29 20:56:32 +05306200static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6201{
6202 switch (lvl) {
6203 case SD_LV_NONE:
6204 return "NONE";
6205 case SD_LV_SIBLING:
6206 return "SIBLING";
6207 case SD_LV_MC:
6208 return "MC";
6209 case SD_LV_CPU:
6210 return "CPU";
6211 case SD_LV_NODE:
6212 return "NODE";
6213 case SD_LV_ALLNODES:
6214 return "ALLNODES";
6215 case SD_LV_MAX:
6216 return "MAX";
6217
6218 }
6219 return "MAX";
6220}
6221
Mike Travis7c16ec52008-04-04 18:11:11 -07006222static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6223 cpumask_t *groupmask)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006224{
6225 struct sched_group *group = sd->groups;
Mike Travis434d53b2008-04-04 18:11:04 -07006226 char str[256];
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006227
Mike Travis434d53b2008-04-04 18:11:04 -07006228 cpulist_scnprintf(str, sizeof(str), sd->span);
Mike Travis7c16ec52008-04-04 18:11:11 -07006229 cpus_clear(*groupmask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006230
6231 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6232
6233 if (!(sd->flags & SD_LOAD_BALANCE)) {
6234 printk("does not load-balance\n");
6235 if (sd->parent)
6236 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6237 " has parent");
6238 return -1;
6239 }
6240
Gautham R Shenoy099f98c2008-05-29 20:56:32 +05306241 printk(KERN_CONT "span %s level %s\n",
6242 str, sd_level_to_string(sd->level));
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006243
6244 if (!cpu_isset(cpu, sd->span)) {
6245 printk(KERN_ERR "ERROR: domain->span does not contain "
6246 "CPU%d\n", cpu);
6247 }
6248 if (!cpu_isset(cpu, group->cpumask)) {
6249 printk(KERN_ERR "ERROR: domain->groups does not contain"
6250 " CPU%d\n", cpu);
6251 }
6252
6253 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6254 do {
6255 if (!group) {
6256 printk("\n");
6257 printk(KERN_ERR "ERROR: group is NULL\n");
6258 break;
6259 }
6260
6261 if (!group->__cpu_power) {
6262 printk(KERN_CONT "\n");
6263 printk(KERN_ERR "ERROR: domain->cpu_power not "
6264 "set\n");
6265 break;
6266 }
6267
6268 if (!cpus_weight(group->cpumask)) {
6269 printk(KERN_CONT "\n");
6270 printk(KERN_ERR "ERROR: empty group\n");
6271 break;
6272 }
6273
Mike Travis7c16ec52008-04-04 18:11:11 -07006274 if (cpus_intersects(*groupmask, group->cpumask)) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006275 printk(KERN_CONT "\n");
6276 printk(KERN_ERR "ERROR: repeated CPUs\n");
6277 break;
6278 }
6279
Mike Travis7c16ec52008-04-04 18:11:11 -07006280 cpus_or(*groupmask, *groupmask, group->cpumask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006281
Mike Travis434d53b2008-04-04 18:11:04 -07006282 cpulist_scnprintf(str, sizeof(str), group->cpumask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006283 printk(KERN_CONT " %s", str);
6284
6285 group = group->next;
6286 } while (group != sd->groups);
6287 printk(KERN_CONT "\n");
6288
Mike Travis7c16ec52008-04-04 18:11:11 -07006289 if (!cpus_equal(sd->span, *groupmask))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006290 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6291
Mike Travis7c16ec52008-04-04 18:11:11 -07006292 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006293 printk(KERN_ERR "ERROR: parent span is not a superset "
6294 "of domain->span\n");
6295 return 0;
6296}
6297
Linus Torvalds1da177e2005-04-16 15:20:36 -07006298static void sched_domain_debug(struct sched_domain *sd, int cpu)
6299{
Mike Travis7c16ec52008-04-04 18:11:11 -07006300 cpumask_t *groupmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006301 int level = 0;
6302
Nick Piggin41c7ce92005-06-25 14:57:24 -07006303 if (!sd) {
6304 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6305 return;
6306 }
6307
Linus Torvalds1da177e2005-04-16 15:20:36 -07006308 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6309
Mike Travis7c16ec52008-04-04 18:11:11 -07006310 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6311 if (!groupmask) {
6312 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6313 return;
6314 }
6315
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006316 for (;;) {
Mike Travis7c16ec52008-04-04 18:11:11 -07006317 if (sched_domain_debug_one(sd, cpu, level, groupmask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006318 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006319 level++;
6320 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08006321 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006322 break;
6323 }
Mike Travis7c16ec52008-04-04 18:11:11 -07006324 kfree(groupmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006325}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006326#else /* !CONFIG_SCHED_DEBUG */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006327# define sched_domain_debug(sd, cpu) do { } while (0)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006328#endif /* CONFIG_SCHED_DEBUG */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006329
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006330static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006331{
6332 if (cpus_weight(sd->span) == 1)
6333 return 1;
6334
6335 /* Following flags need at least 2 groups */
6336 if (sd->flags & (SD_LOAD_BALANCE |
6337 SD_BALANCE_NEWIDLE |
6338 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006339 SD_BALANCE_EXEC |
6340 SD_SHARE_CPUPOWER |
6341 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006342 if (sd->groups != sd->groups->next)
6343 return 0;
6344 }
6345
6346 /* Following flags don't use groups */
6347 if (sd->flags & (SD_WAKE_IDLE |
6348 SD_WAKE_AFFINE |
6349 SD_WAKE_BALANCE))
6350 return 0;
6351
6352 return 1;
6353}
6354
Ingo Molnar48f24c42006-07-03 00:25:40 -07006355static int
6356sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006357{
6358 unsigned long cflags = sd->flags, pflags = parent->flags;
6359
6360 if (sd_degenerate(parent))
6361 return 1;
6362
6363 if (!cpus_equal(sd->span, parent->span))
6364 return 0;
6365
6366 /* Does parent contain flags not in child? */
6367 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6368 if (cflags & SD_WAKE_AFFINE)
6369 pflags &= ~SD_WAKE_BALANCE;
6370 /* Flags needing groups don't count if only 1 group in parent */
6371 if (parent->groups == parent->groups->next) {
6372 pflags &= ~(SD_LOAD_BALANCE |
6373 SD_BALANCE_NEWIDLE |
6374 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006375 SD_BALANCE_EXEC |
6376 SD_SHARE_CPUPOWER |
6377 SD_SHARE_PKG_RESOURCES);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006378 }
6379 if (~cflags & pflags)
6380 return 0;
6381
6382 return 1;
6383}
6384
Gregory Haskins57d885f2008-01-25 21:08:18 +01006385static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6386{
6387 unsigned long flags;
6388 const struct sched_class *class;
6389
6390 spin_lock_irqsave(&rq->lock, flags);
6391
6392 if (rq->rd) {
6393 struct root_domain *old_rd = rq->rd;
6394
Ingo Molnar0eab9142008-01-25 21:08:19 +01006395 for (class = sched_class_highest; class; class = class->next) {
Gregory Haskins57d885f2008-01-25 21:08:18 +01006396 if (class->leave_domain)
6397 class->leave_domain(rq);
Ingo Molnar0eab9142008-01-25 21:08:19 +01006398 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01006399
Gregory Haskinsdc938522008-01-25 21:08:26 +01006400 cpu_clear(rq->cpu, old_rd->span);
6401 cpu_clear(rq->cpu, old_rd->online);
6402
Gregory Haskins57d885f2008-01-25 21:08:18 +01006403 if (atomic_dec_and_test(&old_rd->refcount))
6404 kfree(old_rd);
6405 }
6406
6407 atomic_inc(&rd->refcount);
6408 rq->rd = rd;
6409
Gregory Haskinsdc938522008-01-25 21:08:26 +01006410 cpu_set(rq->cpu, rd->span);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006411 if (cpu_isset(rq->cpu, cpu_online_map))
6412 cpu_set(rq->cpu, rd->online);
Gregory Haskinsdc938522008-01-25 21:08:26 +01006413
Ingo Molnar0eab9142008-01-25 21:08:19 +01006414 for (class = sched_class_highest; class; class = class->next) {
Gregory Haskins57d885f2008-01-25 21:08:18 +01006415 if (class->join_domain)
6416 class->join_domain(rq);
Ingo Molnar0eab9142008-01-25 21:08:19 +01006417 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01006418
6419 spin_unlock_irqrestore(&rq->lock, flags);
6420}
6421
Gregory Haskinsdc938522008-01-25 21:08:26 +01006422static void init_rootdomain(struct root_domain *rd)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006423{
6424 memset(rd, 0, sizeof(*rd));
6425
Gregory Haskinsdc938522008-01-25 21:08:26 +01006426 cpus_clear(rd->span);
6427 cpus_clear(rd->online);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02006428
6429 cpupri_init(&rd->cpupri);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006430}
6431
6432static void init_defrootdomain(void)
6433{
Gregory Haskinsdc938522008-01-25 21:08:26 +01006434 init_rootdomain(&def_root_domain);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006435 atomic_set(&def_root_domain.refcount, 1);
6436}
6437
Gregory Haskinsdc938522008-01-25 21:08:26 +01006438static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006439{
6440 struct root_domain *rd;
6441
6442 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6443 if (!rd)
6444 return NULL;
6445
Gregory Haskinsdc938522008-01-25 21:08:26 +01006446 init_rootdomain(rd);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006447
6448 return rd;
6449}
6450
Linus Torvalds1da177e2005-04-16 15:20:36 -07006451/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01006452 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07006453 * hold the hotplug lock.
6454 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01006455static void
6456cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006457{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006458 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006459 struct sched_domain *tmp;
6460
6461 /* Remove the sched domains which do not contribute to scheduling. */
6462 for (tmp = sd; tmp; tmp = tmp->parent) {
6463 struct sched_domain *parent = tmp->parent;
6464 if (!parent)
6465 break;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006466 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006467 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006468 if (parent->parent)
6469 parent->parent->child = tmp;
6470 }
Suresh Siddha245af2c2005-06-25 14:57:25 -07006471 }
6472
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006473 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006474 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006475 if (sd)
6476 sd->child = NULL;
6477 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006478
6479 sched_domain_debug(sd, cpu);
6480
Gregory Haskins57d885f2008-01-25 21:08:18 +01006481 rq_attach_root(rq, rd);
Nick Piggin674311d2005-06-25 14:57:27 -07006482 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006483}
6484
6485/* cpus with isolated domains */
Tim Chen67af63a2006-12-22 01:07:50 -08006486static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006487
6488/* Setup the mask of cpus configured for isolated domains */
6489static int __init isolated_cpu_setup(char *str)
6490{
6491 int ints[NR_CPUS], i;
6492
6493 str = get_options(str, ARRAY_SIZE(ints), ints);
6494 cpus_clear(cpu_isolated_map);
6495 for (i = 1; i <= ints[0]; i++)
6496 if (ints[i] < NR_CPUS)
6497 cpu_set(ints[i], cpu_isolated_map);
6498 return 1;
6499}
6500
Ingo Molnar8927f492007-10-15 17:00:13 +02006501__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006502
6503/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006504 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6505 * to a function which identifies what group(along with sched group) a CPU
6506 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6507 * (due to the fact that we keep track of groups covered with a cpumask_t).
Linus Torvalds1da177e2005-04-16 15:20:36 -07006508 *
6509 * init_sched_build_groups will build a circular linked list of the groups
6510 * covered by the given span, and will set each group's ->cpumask correctly,
6511 * and ->cpu_power to 0.
6512 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006513static void
Mike Travis7c16ec52008-04-04 18:11:11 -07006514init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006515 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07006516 struct sched_group **sg,
6517 cpumask_t *tmpmask),
6518 cpumask_t *covered, cpumask_t *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006519{
6520 struct sched_group *first = NULL, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006521 int i;
6522
Mike Travis7c16ec52008-04-04 18:11:11 -07006523 cpus_clear(*covered);
6524
6525 for_each_cpu_mask(i, *span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006526 struct sched_group *sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07006527 int group = group_fn(i, cpu_map, &sg, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006528 int j;
6529
Mike Travis7c16ec52008-04-04 18:11:11 -07006530 if (cpu_isset(i, *covered))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006531 continue;
6532
Mike Travis7c16ec52008-04-04 18:11:11 -07006533 cpus_clear(sg->cpumask);
Eric Dumazet5517d862007-05-08 00:32:57 -07006534 sg->__cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006535
Mike Travis7c16ec52008-04-04 18:11:11 -07006536 for_each_cpu_mask(j, *span) {
6537 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006538 continue;
6539
Mike Travis7c16ec52008-04-04 18:11:11 -07006540 cpu_set(j, *covered);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006541 cpu_set(j, sg->cpumask);
6542 }
6543 if (!first)
6544 first = sg;
6545 if (last)
6546 last->next = sg;
6547 last = sg;
6548 }
6549 last->next = first;
6550}
6551
John Hawkes9c1cfda2005-09-06 15:18:14 -07006552#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07006553
John Hawkes9c1cfda2005-09-06 15:18:14 -07006554#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08006555
John Hawkes9c1cfda2005-09-06 15:18:14 -07006556/**
6557 * find_next_best_node - find the next node to include in a sched_domain
6558 * @node: node whose sched_domain we're building
6559 * @used_nodes: nodes already in the sched_domain
6560 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006561 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07006562 * finds the closest node not already in the @used_nodes map.
6563 *
6564 * Should use nodemask_t.
6565 */
Mike Travisc5f59f02008-04-04 18:11:10 -07006566static int find_next_best_node(int node, nodemask_t *used_nodes)
John Hawkes9c1cfda2005-09-06 15:18:14 -07006567{
6568 int i, n, val, min_val, best_node = 0;
6569
6570 min_val = INT_MAX;
6571
6572 for (i = 0; i < MAX_NUMNODES; i++) {
6573 /* Start at @node */
6574 n = (node + i) % MAX_NUMNODES;
6575
6576 if (!nr_cpus_node(n))
6577 continue;
6578
6579 /* Skip already used nodes */
Mike Travisc5f59f02008-04-04 18:11:10 -07006580 if (node_isset(n, *used_nodes))
John Hawkes9c1cfda2005-09-06 15:18:14 -07006581 continue;
6582
6583 /* Simple min distance search */
6584 val = node_distance(node, n);
6585
6586 if (val < min_val) {
6587 min_val = val;
6588 best_node = n;
6589 }
6590 }
6591
Mike Travisc5f59f02008-04-04 18:11:10 -07006592 node_set(best_node, *used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006593 return best_node;
6594}
6595
6596/**
6597 * sched_domain_node_span - get a cpumask for a node's sched_domain
6598 * @node: node whose cpumask we're constructing
Randy Dunlap73486722008-04-22 10:07:22 -07006599 * @span: resulting cpumask
John Hawkes9c1cfda2005-09-06 15:18:14 -07006600 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006601 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07006602 * should be one that prevents unnecessary balancing, but also spreads tasks
6603 * out optimally.
6604 */
Mike Travis4bdbaad32008-04-15 16:35:52 -07006605static void sched_domain_node_span(int node, cpumask_t *span)
John Hawkes9c1cfda2005-09-06 15:18:14 -07006606{
Mike Travisc5f59f02008-04-04 18:11:10 -07006607 nodemask_t used_nodes;
Mike Travisc5f59f02008-04-04 18:11:10 -07006608 node_to_cpumask_ptr(nodemask, node);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006609 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006610
Mike Travis4bdbaad32008-04-15 16:35:52 -07006611 cpus_clear(*span);
Mike Travisc5f59f02008-04-04 18:11:10 -07006612 nodes_clear(used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006613
Mike Travis4bdbaad32008-04-15 16:35:52 -07006614 cpus_or(*span, *span, *nodemask);
Mike Travisc5f59f02008-04-04 18:11:10 -07006615 node_set(node, used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006616
6617 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
Mike Travisc5f59f02008-04-04 18:11:10 -07006618 int next_node = find_next_best_node(node, &used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006619
Mike Travisc5f59f02008-04-04 18:11:10 -07006620 node_to_cpumask_ptr_next(nodemask, next_node);
Mike Travis4bdbaad32008-04-15 16:35:52 -07006621 cpus_or(*span, *span, *nodemask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006622 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006623}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006624#endif /* CONFIG_NUMA */
John Hawkes9c1cfda2005-09-06 15:18:14 -07006625
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006626int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006627
John Hawkes9c1cfda2005-09-06 15:18:14 -07006628/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07006629 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07006630 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006631#ifdef CONFIG_SCHED_SMT
6632static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006633static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006634
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006635static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006636cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6637 cpumask_t *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006638{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006639 if (sg)
6640 *sg = &per_cpu(sched_group_cpus, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006641 return cpu;
6642}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006643#endif /* CONFIG_SCHED_SMT */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006644
Ingo Molnar48f24c42006-07-03 00:25:40 -07006645/*
6646 * multi-core sched-domains:
6647 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006648#ifdef CONFIG_SCHED_MC
6649static DEFINE_PER_CPU(struct sched_domain, core_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006650static DEFINE_PER_CPU(struct sched_group, sched_group_core);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006651#endif /* CONFIG_SCHED_MC */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006652
6653#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006654static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006655cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6656 cpumask_t *mask)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006657{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006658 int group;
Mike Travis7c16ec52008-04-04 18:11:11 -07006659
6660 *mask = per_cpu(cpu_sibling_map, cpu);
6661 cpus_and(*mask, *mask, *cpu_map);
6662 group = first_cpu(*mask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006663 if (sg)
6664 *sg = &per_cpu(sched_group_core, group);
6665 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006666}
6667#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006668static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006669cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6670 cpumask_t *unused)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006671{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006672 if (sg)
6673 *sg = &per_cpu(sched_group_core, cpu);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006674 return cpu;
6675}
6676#endif
6677
Linus Torvalds1da177e2005-04-16 15:20:36 -07006678static DEFINE_PER_CPU(struct sched_domain, phys_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006679static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006680
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006681static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006682cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6683 cpumask_t *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006684{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006685 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006686#ifdef CONFIG_SCHED_MC
Mike Travis7c16ec52008-04-04 18:11:11 -07006687 *mask = cpu_coregroup_map(cpu);
6688 cpus_and(*mask, *mask, *cpu_map);
6689 group = first_cpu(*mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006690#elif defined(CONFIG_SCHED_SMT)
Mike Travis7c16ec52008-04-04 18:11:11 -07006691 *mask = per_cpu(cpu_sibling_map, cpu);
6692 cpus_and(*mask, *mask, *cpu_map);
6693 group = first_cpu(*mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006694#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006695 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006696#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006697 if (sg)
6698 *sg = &per_cpu(sched_group_phys, group);
6699 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006700}
6701
6702#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07006703/*
6704 * The init_sched_build_groups can't handle what we want to do with node
6705 * groups, so roll our own. Now each node has its own list of groups which
6706 * gets dynamically allocated.
6707 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006708static DEFINE_PER_CPU(struct sched_domain, node_domains);
Mike Travis434d53b2008-04-04 18:11:04 -07006709static struct sched_group ***sched_group_nodes_bycpu;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006710
6711static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006712static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006713
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006714static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07006715 struct sched_group **sg, cpumask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006716{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006717 int group;
6718
Mike Travis7c16ec52008-04-04 18:11:11 -07006719 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6720 cpus_and(*nodemask, *nodemask, *cpu_map);
6721 group = first_cpu(*nodemask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006722
6723 if (sg)
6724 *sg = &per_cpu(sched_group_allnodes, group);
6725 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006726}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006727
Siddha, Suresh B08069032006-03-27 01:15:23 -08006728static void init_numa_sched_groups_power(struct sched_group *group_head)
6729{
6730 struct sched_group *sg = group_head;
6731 int j;
6732
6733 if (!sg)
6734 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006735 do {
6736 for_each_cpu_mask(j, sg->cpumask) {
6737 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08006738
Andi Kleen3a5c3592007-10-15 17:00:14 +02006739 sd = &per_cpu(phys_domains, j);
6740 if (j != first_cpu(sd->groups->cpumask)) {
6741 /*
6742 * Only add "power" once for each
6743 * physical package.
6744 */
6745 continue;
6746 }
6747
6748 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
Siddha, Suresh B08069032006-03-27 01:15:23 -08006749 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02006750 sg = sg->next;
6751 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08006752}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006753#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006754
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006755#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006756/* Free memory allocated for various sched_group structures */
Mike Travis7c16ec52008-04-04 18:11:11 -07006757static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006758{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006759 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006760
6761 for_each_cpu_mask(cpu, *cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006762 struct sched_group **sched_group_nodes
6763 = sched_group_nodes_bycpu[cpu];
6764
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006765 if (!sched_group_nodes)
6766 continue;
6767
6768 for (i = 0; i < MAX_NUMNODES; i++) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006769 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6770
Mike Travis7c16ec52008-04-04 18:11:11 -07006771 *nodemask = node_to_cpumask(i);
6772 cpus_and(*nodemask, *nodemask, *cpu_map);
6773 if (cpus_empty(*nodemask))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006774 continue;
6775
6776 if (sg == NULL)
6777 continue;
6778 sg = sg->next;
6779next_sg:
6780 oldsg = sg;
6781 sg = sg->next;
6782 kfree(oldsg);
6783 if (oldsg != sched_group_nodes[i])
6784 goto next_sg;
6785 }
6786 kfree(sched_group_nodes);
6787 sched_group_nodes_bycpu[cpu] = NULL;
6788 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006789}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006790#else /* !CONFIG_NUMA */
Mike Travis7c16ec52008-04-04 18:11:11 -07006791static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006792{
6793}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006794#endif /* CONFIG_NUMA */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006795
Linus Torvalds1da177e2005-04-16 15:20:36 -07006796/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006797 * Initialize sched groups cpu_power.
6798 *
6799 * cpu_power indicates the capacity of sched group, which is used while
6800 * distributing the load between different sched groups in a sched domain.
6801 * Typically cpu_power for all the groups in a sched domain will be same unless
6802 * there are asymmetries in the topology. If there are asymmetries, group
6803 * having more cpu_power will pickup more load compared to the group having
6804 * less cpu_power.
6805 *
6806 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6807 * the maximum number of tasks a group can handle in the presence of other idle
6808 * or lightly loaded groups in the same sched domain.
6809 */
6810static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6811{
6812 struct sched_domain *child;
6813 struct sched_group *group;
6814
6815 WARN_ON(!sd || !sd->groups);
6816
6817 if (cpu != first_cpu(sd->groups->cpumask))
6818 return;
6819
6820 child = sd->child;
6821
Eric Dumazet5517d862007-05-08 00:32:57 -07006822 sd->groups->__cpu_power = 0;
6823
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006824 /*
6825 * For perf policy, if the groups in child domain share resources
6826 * (for example cores sharing some portions of the cache hierarchy
6827 * or SMT), then set this domain groups cpu_power such that each group
6828 * can handle only one task, when there are other idle groups in the
6829 * same sched domain.
6830 */
6831 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6832 (child->flags &
6833 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
Eric Dumazet5517d862007-05-08 00:32:57 -07006834 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006835 return;
6836 }
6837
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006838 /*
6839 * add cpu_power of each child group to this groups cpu_power
6840 */
6841 group = child->groups;
6842 do {
Eric Dumazet5517d862007-05-08 00:32:57 -07006843 sg_inc_cpu_power(sd->groups, group->__cpu_power);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006844 group = group->next;
6845 } while (group != child->groups);
6846}
6847
6848/*
Mike Travis7c16ec52008-04-04 18:11:11 -07006849 * Initializers for schedule domains
6850 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6851 */
6852
6853#define SD_INIT(sd, type) sd_init_##type(sd)
6854#define SD_INIT_FUNC(type) \
6855static noinline void sd_init_##type(struct sched_domain *sd) \
6856{ \
6857 memset(sd, 0, sizeof(*sd)); \
6858 *sd = SD_##type##_INIT; \
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09006859 sd->level = SD_LV_##type; \
Mike Travis7c16ec52008-04-04 18:11:11 -07006860}
6861
6862SD_INIT_FUNC(CPU)
6863#ifdef CONFIG_NUMA
6864 SD_INIT_FUNC(ALLNODES)
6865 SD_INIT_FUNC(NODE)
6866#endif
6867#ifdef CONFIG_SCHED_SMT
6868 SD_INIT_FUNC(SIBLING)
6869#endif
6870#ifdef CONFIG_SCHED_MC
6871 SD_INIT_FUNC(MC)
6872#endif
6873
6874/*
6875 * To minimize stack usage kmalloc room for cpumasks and share the
6876 * space as the usage in build_sched_domains() dictates. Used only
6877 * if the amount of space is significant.
6878 */
6879struct allmasks {
6880 cpumask_t tmpmask; /* make this one first */
6881 union {
6882 cpumask_t nodemask;
6883 cpumask_t this_sibling_map;
6884 cpumask_t this_core_map;
6885 };
6886 cpumask_t send_covered;
6887
6888#ifdef CONFIG_NUMA
6889 cpumask_t domainspan;
6890 cpumask_t covered;
6891 cpumask_t notcovered;
6892#endif
6893};
6894
6895#if NR_CPUS > 128
6896#define SCHED_CPUMASK_ALLOC 1
6897#define SCHED_CPUMASK_FREE(v) kfree(v)
6898#define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6899#else
6900#define SCHED_CPUMASK_ALLOC 0
6901#define SCHED_CPUMASK_FREE(v)
6902#define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6903#endif
6904
6905#define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6906 ((unsigned long)(a) + offsetof(struct allmasks, v))
6907
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09006908static int default_relax_domain_level = -1;
6909
6910static int __init setup_relax_domain_level(char *str)
6911{
6912 default_relax_domain_level = simple_strtoul(str, NULL, 0);
6913 return 1;
6914}
6915__setup("relax_domain_level=", setup_relax_domain_level);
6916
6917static void set_domain_attribute(struct sched_domain *sd,
6918 struct sched_domain_attr *attr)
6919{
6920 int request;
6921
6922 if (!attr || attr->relax_domain_level < 0) {
6923 if (default_relax_domain_level < 0)
6924 return;
6925 else
6926 request = default_relax_domain_level;
6927 } else
6928 request = attr->relax_domain_level;
6929 if (request < sd->level) {
6930 /* turn off idle balance on this domain */
6931 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
6932 } else {
6933 /* turn on idle balance on this domain */
6934 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
6935 }
6936}
6937
Mike Travis7c16ec52008-04-04 18:11:11 -07006938/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006939 * Build sched domains for a given set of cpus and attach the sched domains
6940 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07006941 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09006942static int __build_sched_domains(const cpumask_t *cpu_map,
6943 struct sched_domain_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006944{
6945 int i;
Gregory Haskins57d885f2008-01-25 21:08:18 +01006946 struct root_domain *rd;
Mike Travis7c16ec52008-04-04 18:11:11 -07006947 SCHED_CPUMASK_DECLARE(allmasks);
6948 cpumask_t *tmpmask;
John Hawkesd1b55132005-09-06 15:18:14 -07006949#ifdef CONFIG_NUMA
6950 struct sched_group **sched_group_nodes = NULL;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006951 int sd_allnodes = 0;
John Hawkesd1b55132005-09-06 15:18:14 -07006952
6953 /*
6954 * Allocate the per-node list of sched groups
6955 */
Milton Miller5cf9f062007-10-15 17:00:19 +02006956 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006957 GFP_KERNEL);
John Hawkesd1b55132005-09-06 15:18:14 -07006958 if (!sched_group_nodes) {
6959 printk(KERN_WARNING "Can not alloc sched group node list\n");
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006960 return -ENOMEM;
John Hawkesd1b55132005-09-06 15:18:14 -07006961 }
John Hawkesd1b55132005-09-06 15:18:14 -07006962#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006963
Gregory Haskinsdc938522008-01-25 21:08:26 +01006964 rd = alloc_rootdomain();
Gregory Haskins57d885f2008-01-25 21:08:18 +01006965 if (!rd) {
6966 printk(KERN_WARNING "Cannot alloc root domain\n");
Mike Travis7c16ec52008-04-04 18:11:11 -07006967#ifdef CONFIG_NUMA
6968 kfree(sched_group_nodes);
6969#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01006970 return -ENOMEM;
6971 }
6972
Mike Travis7c16ec52008-04-04 18:11:11 -07006973#if SCHED_CPUMASK_ALLOC
6974 /* get space for all scratch cpumask variables */
6975 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6976 if (!allmasks) {
6977 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6978 kfree(rd);
6979#ifdef CONFIG_NUMA
6980 kfree(sched_group_nodes);
6981#endif
6982 return -ENOMEM;
6983 }
6984#endif
6985 tmpmask = (cpumask_t *)allmasks;
6986
6987
6988#ifdef CONFIG_NUMA
6989 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6990#endif
6991
Linus Torvalds1da177e2005-04-16 15:20:36 -07006992 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006993 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006994 */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006995 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006996 struct sched_domain *sd = NULL, *p;
Mike Travis7c16ec52008-04-04 18:11:11 -07006997 SCHED_CPUMASK_VAR(nodemask, allmasks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006998
Mike Travis7c16ec52008-04-04 18:11:11 -07006999 *nodemask = node_to_cpumask(cpu_to_node(i));
7000 cpus_and(*nodemask, *nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007001
7002#ifdef CONFIG_NUMA
Ingo Molnardd41f592007-07-09 18:51:59 +02007003 if (cpus_weight(*cpu_map) >
Mike Travis7c16ec52008-04-04 18:11:11 -07007004 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07007005 sd = &per_cpu(allnodes_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007006 SD_INIT(sd, ALLNODES);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007007 set_domain_attribute(sd, attr);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007008 sd->span = *cpu_map;
Mike Travis7c16ec52008-04-04 18:11:11 -07007009 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007010 p = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007011 sd_allnodes = 1;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007012 } else
7013 p = NULL;
7014
Linus Torvalds1da177e2005-04-16 15:20:36 -07007015 sd = &per_cpu(node_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007016 SD_INIT(sd, NODE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007017 set_domain_attribute(sd, attr);
Mike Travis4bdbaad32008-04-15 16:35:52 -07007018 sched_domain_node_span(cpu_to_node(i), &sd->span);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007019 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007020 if (p)
7021 p->child = sd;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007022 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007023#endif
7024
7025 p = sd;
7026 sd = &per_cpu(phys_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007027 SD_INIT(sd, CPU);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007028 set_domain_attribute(sd, attr);
Mike Travis7c16ec52008-04-04 18:11:11 -07007029 sd->span = *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007030 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007031 if (p)
7032 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007033 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007034
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007035#ifdef CONFIG_SCHED_MC
7036 p = sd;
7037 sd = &per_cpu(core_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007038 SD_INIT(sd, MC);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007039 set_domain_attribute(sd, attr);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007040 sd->span = cpu_coregroup_map(i);
7041 cpus_and(sd->span, sd->span, *cpu_map);
7042 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007043 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007044 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007045#endif
7046
Linus Torvalds1da177e2005-04-16 15:20:36 -07007047#ifdef CONFIG_SCHED_SMT
7048 p = sd;
7049 sd = &per_cpu(cpu_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007050 SD_INIT(sd, SIBLING);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007051 set_domain_attribute(sd, attr);
Mike Travisd5a74302007-10-16 01:24:05 -07007052 sd->span = per_cpu(cpu_sibling_map, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007053 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007054 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007055 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007056 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007057#endif
7058 }
7059
7060#ifdef CONFIG_SCHED_SMT
7061 /* Set up CPU (sibling) groups */
John Hawkes9c1cfda2005-09-06 15:18:14 -07007062 for_each_cpu_mask(i, *cpu_map) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007063 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7064 SCHED_CPUMASK_VAR(send_covered, allmasks);
7065
7066 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7067 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7068 if (i != first_cpu(*this_sibling_map))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007069 continue;
7070
Ingo Molnardd41f592007-07-09 18:51:59 +02007071 init_sched_build_groups(this_sibling_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007072 &cpu_to_cpu_group,
7073 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007074 }
7075#endif
7076
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007077#ifdef CONFIG_SCHED_MC
7078 /* Set up multi-core groups */
7079 for_each_cpu_mask(i, *cpu_map) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007080 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7081 SCHED_CPUMASK_VAR(send_covered, allmasks);
7082
7083 *this_core_map = cpu_coregroup_map(i);
7084 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7085 if (i != first_cpu(*this_core_map))
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007086 continue;
Mike Travis7c16ec52008-04-04 18:11:11 -07007087
Ingo Molnardd41f592007-07-09 18:51:59 +02007088 init_sched_build_groups(this_core_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007089 &cpu_to_core_group,
7090 send_covered, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007091 }
7092#endif
7093
Linus Torvalds1da177e2005-04-16 15:20:36 -07007094 /* Set up physical groups */
7095 for (i = 0; i < MAX_NUMNODES; i++) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007096 SCHED_CPUMASK_VAR(nodemask, allmasks);
7097 SCHED_CPUMASK_VAR(send_covered, allmasks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007098
Mike Travis7c16ec52008-04-04 18:11:11 -07007099 *nodemask = node_to_cpumask(i);
7100 cpus_and(*nodemask, *nodemask, *cpu_map);
7101 if (cpus_empty(*nodemask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007102 continue;
7103
Mike Travis7c16ec52008-04-04 18:11:11 -07007104 init_sched_build_groups(nodemask, cpu_map,
7105 &cpu_to_phys_group,
7106 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007107 }
7108
7109#ifdef CONFIG_NUMA
7110 /* Set up node groups */
Mike Travis7c16ec52008-04-04 18:11:11 -07007111 if (sd_allnodes) {
7112 SCHED_CPUMASK_VAR(send_covered, allmasks);
7113
7114 init_sched_build_groups(cpu_map, cpu_map,
7115 &cpu_to_allnodes_group,
7116 send_covered, tmpmask);
7117 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007118
7119 for (i = 0; i < MAX_NUMNODES; i++) {
7120 /* Set up node groups */
7121 struct sched_group *sg, *prev;
Mike Travis7c16ec52008-04-04 18:11:11 -07007122 SCHED_CPUMASK_VAR(nodemask, allmasks);
7123 SCHED_CPUMASK_VAR(domainspan, allmasks);
7124 SCHED_CPUMASK_VAR(covered, allmasks);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007125 int j;
7126
Mike Travis7c16ec52008-04-04 18:11:11 -07007127 *nodemask = node_to_cpumask(i);
7128 cpus_clear(*covered);
7129
7130 cpus_and(*nodemask, *nodemask, *cpu_map);
7131 if (cpus_empty(*nodemask)) {
John Hawkesd1b55132005-09-06 15:18:14 -07007132 sched_group_nodes[i] = NULL;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007133 continue;
John Hawkesd1b55132005-09-06 15:18:14 -07007134 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007135
Mike Travis4bdbaad32008-04-15 16:35:52 -07007136 sched_domain_node_span(i, domainspan);
Mike Travis7c16ec52008-04-04 18:11:11 -07007137 cpus_and(*domainspan, *domainspan, *cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007138
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07007139 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007140 if (!sg) {
7141 printk(KERN_WARNING "Can not alloc domain group for "
7142 "node %d\n", i);
7143 goto error;
7144 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007145 sched_group_nodes[i] = sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07007146 for_each_cpu_mask(j, *nodemask) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07007147 struct sched_domain *sd;
Ingo Molnar9761eea2007-07-09 18:52:00 +02007148
John Hawkes9c1cfda2005-09-06 15:18:14 -07007149 sd = &per_cpu(node_domains, j);
7150 sd->groups = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007151 }
Eric Dumazet5517d862007-05-08 00:32:57 -07007152 sg->__cpu_power = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07007153 sg->cpumask = *nodemask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007154 sg->next = sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07007155 cpus_or(*covered, *covered, *nodemask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007156 prev = sg;
7157
7158 for (j = 0; j < MAX_NUMNODES; j++) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007159 SCHED_CPUMASK_VAR(notcovered, allmasks);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007160 int n = (i + j) % MAX_NUMNODES;
Mike Travisc5f59f02008-04-04 18:11:10 -07007161 node_to_cpumask_ptr(pnodemask, n);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007162
Mike Travis7c16ec52008-04-04 18:11:11 -07007163 cpus_complement(*notcovered, *covered);
7164 cpus_and(*tmpmask, *notcovered, *cpu_map);
7165 cpus_and(*tmpmask, *tmpmask, *domainspan);
7166 if (cpus_empty(*tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07007167 break;
7168
Mike Travis7c16ec52008-04-04 18:11:11 -07007169 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7170 if (cpus_empty(*tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07007171 continue;
7172
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07007173 sg = kmalloc_node(sizeof(struct sched_group),
7174 GFP_KERNEL, i);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007175 if (!sg) {
7176 printk(KERN_WARNING
7177 "Can not alloc domain group for node %d\n", j);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007178 goto error;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007179 }
Eric Dumazet5517d862007-05-08 00:32:57 -07007180 sg->__cpu_power = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07007181 sg->cpumask = *tmpmask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007182 sg->next = prev->next;
Mike Travis7c16ec52008-04-04 18:11:11 -07007183 cpus_or(*covered, *covered, *tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007184 prev->next = sg;
7185 prev = sg;
7186 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007187 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007188#endif
7189
7190 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007191#ifdef CONFIG_SCHED_SMT
7192 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007193 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7194
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007195 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007196 }
7197#endif
7198#ifdef CONFIG_SCHED_MC
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007199 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007200 struct sched_domain *sd = &per_cpu(core_domains, i);
7201
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007202 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007203 }
7204#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007205
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007206 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007207 struct sched_domain *sd = &per_cpu(phys_domains, i);
7208
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007209 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007210 }
7211
John Hawkes9c1cfda2005-09-06 15:18:14 -07007212#ifdef CONFIG_NUMA
Siddha, Suresh B08069032006-03-27 01:15:23 -08007213 for (i = 0; i < MAX_NUMNODES; i++)
7214 init_numa_sched_groups_power(sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007215
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007216 if (sd_allnodes) {
7217 struct sched_group *sg;
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07007218
Mike Travis7c16ec52008-04-04 18:11:11 -07007219 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7220 tmpmask);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07007221 init_numa_sched_groups_power(sg);
7222 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007223#endif
7224
Linus Torvalds1da177e2005-04-16 15:20:36 -07007225 /* Attach the domains */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007226 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007227 struct sched_domain *sd;
7228#ifdef CONFIG_SCHED_SMT
7229 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007230#elif defined(CONFIG_SCHED_MC)
7231 sd = &per_cpu(core_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007232#else
7233 sd = &per_cpu(phys_domains, i);
7234#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01007235 cpu_attach_domain(sd, rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007236 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007237
Mike Travis7c16ec52008-04-04 18:11:11 -07007238 SCHED_CPUMASK_FREE((void *)allmasks);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007239 return 0;
7240
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007241#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007242error:
Mike Travis7c16ec52008-04-04 18:11:11 -07007243 free_sched_groups(cpu_map, tmpmask);
7244 SCHED_CPUMASK_FREE((void *)allmasks);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007245 return -ENOMEM;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007246#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007247}
Paul Jackson029190c2007-10-18 23:40:20 -07007248
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007249static int build_sched_domains(const cpumask_t *cpu_map)
7250{
7251 return __build_sched_domains(cpu_map, NULL);
7252}
7253
Paul Jackson029190c2007-10-18 23:40:20 -07007254static cpumask_t *doms_cur; /* current sched domains */
7255static int ndoms_cur; /* number of sched domains in 'doms_cur' */
Ingo Molnar4285f5942008-05-16 17:47:14 +02007256static struct sched_domain_attr *dattr_cur;
7257 /* attribues of custom domains in 'doms_cur' */
Paul Jackson029190c2007-10-18 23:40:20 -07007258
7259/*
7260 * Special case: If a kmalloc of a doms_cur partition (array of
7261 * cpumask_t) fails, then fallback to a single sched domain,
7262 * as determined by the single cpumask_t fallback_doms.
7263 */
7264static cpumask_t fallback_doms;
7265
Heiko Carstens22e52b02008-03-12 18:31:59 +01007266void __attribute__((weak)) arch_update_cpu_topology(void)
7267{
7268}
7269
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007270/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007271 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07007272 * For now this just excludes isolated cpus, but could be used to
7273 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007274 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007275static int arch_init_sched_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007276{
Milton Miller73785472007-10-24 18:23:48 +02007277 int err;
7278
Heiko Carstens22e52b02008-03-12 18:31:59 +01007279 arch_update_cpu_topology();
Paul Jackson029190c2007-10-18 23:40:20 -07007280 ndoms_cur = 1;
7281 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7282 if (!doms_cur)
7283 doms_cur = &fallback_doms;
7284 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007285 dattr_cur = NULL;
Milton Miller73785472007-10-24 18:23:48 +02007286 err = build_sched_domains(doms_cur);
Milton Miller6382bc92007-10-15 17:00:19 +02007287 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02007288
7289 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007290}
7291
Mike Travis7c16ec52008-04-04 18:11:11 -07007292static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7293 cpumask_t *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007294{
Mike Travis7c16ec52008-04-04 18:11:11 -07007295 free_sched_groups(cpu_map, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007296}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007297
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007298/*
7299 * Detach sched domains from a group of cpus specified in cpu_map
7300 * These cpus will now be attached to the NULL domain
7301 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08007302static void detach_destroy_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007303{
Mike Travis7c16ec52008-04-04 18:11:11 -07007304 cpumask_t tmpmask;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007305 int i;
7306
Milton Miller6382bc92007-10-15 17:00:19 +02007307 unregister_sched_domain_sysctl();
7308
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007309 for_each_cpu_mask(i, *cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01007310 cpu_attach_domain(NULL, &def_root_domain, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007311 synchronize_sched();
Mike Travis7c16ec52008-04-04 18:11:11 -07007312 arch_destroy_sched_domains(cpu_map, &tmpmask);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007313}
7314
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007315/* handle null as "default" */
7316static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7317 struct sched_domain_attr *new, int idx_new)
7318{
7319 struct sched_domain_attr tmp;
7320
7321 /* fast path */
7322 if (!new && !cur)
7323 return 1;
7324
7325 tmp = SD_ATTR_INIT;
7326 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7327 new ? (new + idx_new) : &tmp,
7328 sizeof(struct sched_domain_attr));
7329}
7330
Paul Jackson029190c2007-10-18 23:40:20 -07007331/*
7332 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007333 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07007334 * doms_new[] to the current sched domain partitioning, doms_cur[].
7335 * It destroys each deleted domain and builds each new domain.
7336 *
7337 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007338 * The masks don't intersect (don't overlap.) We should setup one
7339 * sched domain for each mask. CPUs not in any of the cpumasks will
7340 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07007341 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7342 * it as it is.
7343 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007344 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7345 * ownership of it and will kfree it when done with it. If the caller
Paul Jackson029190c2007-10-18 23:40:20 -07007346 * failed the kmalloc call, then it can pass in doms_new == NULL,
7347 * and partition_sched_domains() will fallback to the single partition
7348 * 'fallback_doms'.
7349 *
7350 * Call with hotplug lock held
7351 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007352void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7353 struct sched_domain_attr *dattr_new)
Paul Jackson029190c2007-10-18 23:40:20 -07007354{
7355 int i, j;
7356
Heiko Carstens712555e2008-04-28 11:33:07 +02007357 mutex_lock(&sched_domains_mutex);
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007358
Milton Miller73785472007-10-24 18:23:48 +02007359 /* always unregister in case we don't destroy any domains */
7360 unregister_sched_domain_sysctl();
7361
Paul Jackson029190c2007-10-18 23:40:20 -07007362 if (doms_new == NULL) {
7363 ndoms_new = 1;
7364 doms_new = &fallback_doms;
7365 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007366 dattr_new = NULL;
Paul Jackson029190c2007-10-18 23:40:20 -07007367 }
7368
7369 /* Destroy deleted domains */
7370 for (i = 0; i < ndoms_cur; i++) {
7371 for (j = 0; j < ndoms_new; j++) {
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007372 if (cpus_equal(doms_cur[i], doms_new[j])
7373 && dattrs_equal(dattr_cur, i, dattr_new, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007374 goto match1;
7375 }
7376 /* no match - a current sched domain not in new doms_new[] */
7377 detach_destroy_domains(doms_cur + i);
7378match1:
7379 ;
7380 }
7381
7382 /* Build new domains */
7383 for (i = 0; i < ndoms_new; i++) {
7384 for (j = 0; j < ndoms_cur; j++) {
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007385 if (cpus_equal(doms_new[i], doms_cur[j])
7386 && dattrs_equal(dattr_new, i, dattr_cur, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007387 goto match2;
7388 }
7389 /* no match - add a new doms_new */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007390 __build_sched_domains(doms_new + i,
7391 dattr_new ? dattr_new + i : NULL);
Paul Jackson029190c2007-10-18 23:40:20 -07007392match2:
7393 ;
7394 }
7395
7396 /* Remember the new sched domains */
7397 if (doms_cur != &fallback_doms)
7398 kfree(doms_cur);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007399 kfree(dattr_cur); /* kfree(NULL) is safe */
Paul Jackson029190c2007-10-18 23:40:20 -07007400 doms_cur = doms_new;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007401 dattr_cur = dattr_new;
Paul Jackson029190c2007-10-18 23:40:20 -07007402 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02007403
7404 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007405
Heiko Carstens712555e2008-04-28 11:33:07 +02007406 mutex_unlock(&sched_domains_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -07007407}
7408
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007409#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Heiko Carstens9aefd0a2008-03-12 18:31:58 +01007410int arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007411{
7412 int err;
7413
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007414 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02007415 mutex_lock(&sched_domains_mutex);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007416 detach_destroy_domains(&cpu_online_map);
7417 err = arch_init_sched_domains(&cpu_online_map);
Heiko Carstens712555e2008-04-28 11:33:07 +02007418 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007419 put_online_cpus();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007420
7421 return err;
7422}
7423
7424static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7425{
7426 int ret;
7427
7428 if (buf[0] != '0' && buf[0] != '1')
7429 return -EINVAL;
7430
7431 if (smt)
7432 sched_smt_power_savings = (buf[0] == '1');
7433 else
7434 sched_mc_power_savings = (buf[0] == '1');
7435
7436 ret = arch_reinit_sched_domains();
7437
7438 return ret ? ret : count;
7439}
7440
Adrian Bunk6707de002007-08-12 18:08:19 +02007441#ifdef CONFIG_SCHED_MC
7442static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7443{
7444 return sprintf(page, "%u\n", sched_mc_power_savings);
7445}
7446static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7447 const char *buf, size_t count)
7448{
7449 return sched_power_savings_store(buf, count, 0);
7450}
7451static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7452 sched_mc_power_savings_store);
7453#endif
7454
7455#ifdef CONFIG_SCHED_SMT
7456static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7457{
7458 return sprintf(page, "%u\n", sched_smt_power_savings);
7459}
7460static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7461 const char *buf, size_t count)
7462{
7463 return sched_power_savings_store(buf, count, 1);
7464}
7465static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7466 sched_smt_power_savings_store);
7467#endif
7468
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007469int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7470{
7471 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007472
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007473#ifdef CONFIG_SCHED_SMT
7474 if (smt_capable())
7475 err = sysfs_create_file(&cls->kset.kobj,
7476 &attr_sched_smt_power_savings.attr);
7477#endif
7478#ifdef CONFIG_SCHED_MC
7479 if (!err && mc_capable())
7480 err = sysfs_create_file(&cls->kset.kobj,
7481 &attr_sched_mc_power_savings.attr);
7482#endif
7483 return err;
7484}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007485#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007486
Linus Torvalds1da177e2005-04-16 15:20:36 -07007487/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007488 * Force a reinitialization of the sched domains hierarchy. The domains
Linus Torvalds1da177e2005-04-16 15:20:36 -07007489 * and groups cannot be updated in place without racing with the balancing
Nick Piggin41c7ce92005-06-25 14:57:24 -07007490 * code, so we temporarily attach all running cpus to the NULL domain
Linus Torvalds1da177e2005-04-16 15:20:36 -07007491 * which will prevent rebalancing while the sched domains are recalculated.
7492 */
7493static int update_sched_domains(struct notifier_block *nfb,
7494 unsigned long action, void *hcpu)
7495{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007496 switch (action) {
7497 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007498 case CPU_UP_PREPARE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007499 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007500 case CPU_DOWN_PREPARE_FROZEN:
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007501 detach_destroy_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007502 return NOTIFY_OK;
7503
7504 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007505 case CPU_UP_CANCELED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007506 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007507 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007508 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007509 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007510 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007511 case CPU_DEAD_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007512 /*
7513 * Fall through and re-initialise the domains.
7514 */
7515 break;
7516 default:
7517 return NOTIFY_DONE;
7518 }
7519
7520 /* The hotplug lock is already held by cpu_up/cpu_down */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007521 arch_init_sched_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007522
7523 return NOTIFY_OK;
7524}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007525
7526void __init sched_init_smp(void)
7527{
Nick Piggin5c1e1762006-10-03 01:14:04 -07007528 cpumask_t non_isolated_cpus;
7529
Mike Travis434d53b2008-04-04 18:11:04 -07007530#if defined(CONFIG_NUMA)
7531 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7532 GFP_KERNEL);
7533 BUG_ON(sched_group_nodes_bycpu == NULL);
7534#endif
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007535 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02007536 mutex_lock(&sched_domains_mutex);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007537 arch_init_sched_domains(&cpu_online_map);
Nathan Lynche5e56732007-01-10 23:15:28 -08007538 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
Nick Piggin5c1e1762006-10-03 01:14:04 -07007539 if (cpus_empty(non_isolated_cpus))
7540 cpu_set(smp_processor_id(), non_isolated_cpus);
Heiko Carstens712555e2008-04-28 11:33:07 +02007541 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007542 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007543 /* XXX: Theoretical race here - CPU may be hotplugged now */
7544 hotcpu_notifier(update_sched_domains, 0);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02007545 init_hrtick();
Nick Piggin5c1e1762006-10-03 01:14:04 -07007546
7547 /* Move init over to a non-isolated CPU */
Mike Travis7c16ec52008-04-04 18:11:11 -07007548 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
Nick Piggin5c1e1762006-10-03 01:14:04 -07007549 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01007550 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007551}
7552#else
7553void __init sched_init_smp(void)
7554{
Ingo Molnar19978ca2007-11-09 22:39:38 +01007555 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007556}
7557#endif /* CONFIG_SMP */
7558
7559int in_sched_functions(unsigned long addr)
7560{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007561 return in_lock_functions(addr) ||
7562 (addr >= (unsigned long)__sched_text_start
7563 && addr < (unsigned long)__sched_text_end);
7564}
7565
Alexey Dobriyana9957442007-10-15 17:00:13 +02007566static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02007567{
7568 cfs_rq->tasks_timeline = RB_ROOT;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02007569 INIT_LIST_HEAD(&cfs_rq->tasks);
Ingo Molnardd41f592007-07-09 18:51:59 +02007570#ifdef CONFIG_FAIR_GROUP_SCHED
7571 cfs_rq->rq = rq;
7572#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02007573 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02007574}
7575
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007576static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7577{
7578 struct rt_prio_array *array;
7579 int i;
7580
7581 array = &rt_rq->active;
7582 for (i = 0; i < MAX_RT_PRIO; i++) {
Gregory Haskins45c01e82008-05-12 21:20:41 +02007583 INIT_LIST_HEAD(array->xqueue + i);
7584 INIT_LIST_HEAD(array->squeue + i);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007585 __clear_bit(i, array->bitmap);
7586 }
7587 /* delimiter for bitsearch: */
7588 __set_bit(MAX_RT_PRIO, array->bitmap);
7589
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007590#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Peter Zijlstra48d5e252008-01-25 21:08:31 +01007591 rt_rq->highest_prio = MAX_RT_PRIO;
7592#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007593#ifdef CONFIG_SMP
7594 rt_rq->rt_nr_migratory = 0;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007595 rt_rq->overloaded = 0;
7596#endif
7597
7598 rt_rq->rt_time = 0;
7599 rt_rq->rt_throttled = 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007600 rt_rq->rt_runtime = 0;
7601 spin_lock_init(&rt_rq->rt_runtime_lock);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007602
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007603#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01007604 rt_rq->rt_nr_boosted = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007605 rt_rq->rq = rq;
7606#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007607}
7608
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007609#ifdef CONFIG_FAIR_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007610static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7611 struct sched_entity *se, int cpu, int add,
7612 struct sched_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007613{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007614 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007615 tg->cfs_rq[cpu] = cfs_rq;
7616 init_cfs_rq(cfs_rq, rq);
7617 cfs_rq->tg = tg;
7618 if (add)
7619 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7620
7621 tg->se[cpu] = se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02007622 /* se could be NULL for init_task_group */
7623 if (!se)
7624 return;
7625
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007626 if (!parent)
7627 se->cfs_rq = &rq->cfs;
7628 else
7629 se->cfs_rq = parent->my_q;
7630
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007631 se->my_q = cfs_rq;
7632 se->load.weight = tg->shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02007633 se->load.inv_weight = 0;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007634 se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007635}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007636#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007637
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007638#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007639static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7640 struct sched_rt_entity *rt_se, int cpu, int add,
7641 struct sched_rt_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007642{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007643 struct rq *rq = cpu_rq(cpu);
7644
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007645 tg->rt_rq[cpu] = rt_rq;
7646 init_rt_rq(rt_rq, rq);
7647 rt_rq->tg = tg;
7648 rt_rq->rt_se = rt_se;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007649 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007650 if (add)
7651 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7652
7653 tg->rt_se[cpu] = rt_se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02007654 if (!rt_se)
7655 return;
7656
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007657 if (!parent)
7658 rt_se->rt_rq = &rq->rt;
7659 else
7660 rt_se->rt_rq = parent->my_q;
7661
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007662 rt_se->rt_rq = &rq->rt;
7663 rt_se->my_q = rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007664 rt_se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007665 INIT_LIST_HEAD(&rt_se->run_list);
7666}
7667#endif
7668
Linus Torvalds1da177e2005-04-16 15:20:36 -07007669void __init sched_init(void)
7670{
Ingo Molnardd41f592007-07-09 18:51:59 +02007671 int i, j;
Mike Travis434d53b2008-04-04 18:11:04 -07007672 unsigned long alloc_size = 0, ptr;
7673
7674#ifdef CONFIG_FAIR_GROUP_SCHED
7675 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7676#endif
7677#ifdef CONFIG_RT_GROUP_SCHED
7678 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7679#endif
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007680#ifdef CONFIG_USER_SCHED
7681 alloc_size *= 2;
7682#endif
Mike Travis434d53b2008-04-04 18:11:04 -07007683 /*
7684 * As sched_init() is called before page_alloc is setup,
7685 * we use alloc_bootmem().
7686 */
7687 if (alloc_size) {
David Miller5a9d3222008-04-24 20:46:20 -07007688 ptr = (unsigned long)alloc_bootmem(alloc_size);
Mike Travis434d53b2008-04-04 18:11:04 -07007689
7690#ifdef CONFIG_FAIR_GROUP_SCHED
7691 init_task_group.se = (struct sched_entity **)ptr;
7692 ptr += nr_cpu_ids * sizeof(void **);
7693
7694 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7695 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007696
7697#ifdef CONFIG_USER_SCHED
7698 root_task_group.se = (struct sched_entity **)ptr;
7699 ptr += nr_cpu_ids * sizeof(void **);
7700
7701 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7702 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007703#endif /* CONFIG_USER_SCHED */
7704#endif /* CONFIG_FAIR_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07007705#ifdef CONFIG_RT_GROUP_SCHED
7706 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7707 ptr += nr_cpu_ids * sizeof(void **);
7708
7709 init_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007710 ptr += nr_cpu_ids * sizeof(void **);
7711
7712#ifdef CONFIG_USER_SCHED
7713 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7714 ptr += nr_cpu_ids * sizeof(void **);
7715
7716 root_task_group.rt_rq = (struct rt_rq **)ptr;
7717 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007718#endif /* CONFIG_USER_SCHED */
7719#endif /* CONFIG_RT_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07007720 }
Ingo Molnardd41f592007-07-09 18:51:59 +02007721
Gregory Haskins57d885f2008-01-25 21:08:18 +01007722#ifdef CONFIG_SMP
7723 init_defrootdomain();
7724#endif
7725
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007726 init_rt_bandwidth(&def_rt_bandwidth,
7727 global_rt_period(), global_rt_runtime());
7728
7729#ifdef CONFIG_RT_GROUP_SCHED
7730 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7731 global_rt_period(), global_rt_runtime());
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007732#ifdef CONFIG_USER_SCHED
7733 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7734 global_rt_period(), RUNTIME_INF);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007735#endif /* CONFIG_USER_SCHED */
7736#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007737
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007738#ifdef CONFIG_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007739 list_add(&init_task_group.list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02007740 INIT_LIST_HEAD(&init_task_group.children);
7741
7742#ifdef CONFIG_USER_SCHED
7743 INIT_LIST_HEAD(&root_task_group.children);
7744 init_task_group.parent = &root_task_group;
7745 list_add(&init_task_group.siblings, &root_task_group.children);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007746#endif /* CONFIG_USER_SCHED */
7747#endif /* CONFIG_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007748
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08007749 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007750 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007751
7752 rq = cpu_rq(i);
7753 spin_lock_init(&rq->lock);
Ingo Molnarfcb99372006-07-03 00:25:10 -07007754 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
Nick Piggin78979862005-06-25 14:57:13 -07007755 rq->nr_running = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007756 init_cfs_rq(&rq->cfs, rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007757 init_rt_rq(&rq->rt, rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007758#ifdef CONFIG_FAIR_GROUP_SCHED
7759 init_task_group.shares = init_task_group_load;
7760 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007761#ifdef CONFIG_CGROUP_SCHED
7762 /*
7763 * How much cpu bandwidth does init_task_group get?
7764 *
7765 * In case of task-groups formed thr' the cgroup filesystem, it
7766 * gets 100% of the cpu resources in the system. This overall
7767 * system cpu resource is divided among the tasks of
7768 * init_task_group and its child task-groups in a fair manner,
7769 * based on each entity's (task or task-group's) weight
7770 * (se->load.weight).
7771 *
7772 * In other words, if init_task_group has 10 tasks of weight
7773 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7774 * then A0's share of the cpu resource is:
7775 *
7776 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7777 *
7778 * We achieve this by letting init_task_group's tasks sit
7779 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7780 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007781 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007782#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007783 root_task_group.shares = NICE_0_LOAD;
7784 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007785 /*
7786 * In case of task-groups formed thr' the user id of tasks,
7787 * init_task_group represents tasks belonging to root user.
7788 * Hence it forms a sibling of all subsequent groups formed.
7789 * In this case, init_task_group gets only a fraction of overall
7790 * system cpu resource, based on the weight assigned to root
7791 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7792 * by letting tasks of init_task_group sit in a separate cfs_rq
7793 * (init_cfs_rq) and having one entity represent this group of
7794 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7795 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007796 init_tg_cfs_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007797 &per_cpu(init_cfs_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007798 &per_cpu(init_sched_entity, i), i, 1,
7799 root_task_group.se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007800
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007801#endif
Dhaval Giani354d60c2008-04-19 19:44:59 +02007802#endif /* CONFIG_FAIR_GROUP_SCHED */
7803
7804 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007805#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007806 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007807#ifdef CONFIG_CGROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007808 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007809#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007810 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007811 init_tg_rt_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007812 &per_cpu(init_rt_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007813 &per_cpu(init_sched_rt_entity, i), i, 1,
7814 root_task_group.rt_se[i]);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007815#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007816#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007817
Ingo Molnardd41f592007-07-09 18:51:59 +02007818 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7819 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007820#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07007821 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007822 rq->rd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007823 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007824 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007825 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07007826 rq->cpu = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007827 rq->migration_thread = NULL;
7828 INIT_LIST_HEAD(&rq->migration_queue);
Gregory Haskinsdc938522008-01-25 21:08:26 +01007829 rq_attach_root(rq, &def_root_domain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007830#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01007831 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007832 atomic_set(&rq->nr_iowait, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007833 }
7834
Peter Williams2dd73a42006-06-27 02:54:34 -07007835 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07007836
Avi Kivitye107be32007-07-26 13:40:43 +02007837#ifdef CONFIG_PREEMPT_NOTIFIERS
7838 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7839#endif
7840
Christoph Lameterc9819f42006-12-10 02:20:25 -08007841#ifdef CONFIG_SMP
7842 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7843#endif
7844
Heiko Carstensb50f60c2006-07-30 03:03:52 -07007845#ifdef CONFIG_RT_MUTEXES
7846 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7847#endif
7848
Linus Torvalds1da177e2005-04-16 15:20:36 -07007849 /*
7850 * The boot idle thread does lazy MMU switching as well:
7851 */
7852 atomic_inc(&init_mm.mm_count);
7853 enter_lazy_tlb(&init_mm, current);
7854
7855 /*
7856 * Make us the idle thread. Technically, schedule() should not be
7857 * called from this thread, however somewhere below it might be,
7858 * but because we are the idle thread, we just pick up running again
7859 * when this runqueue becomes "idle".
7860 */
7861 init_idle(current, smp_processor_id());
Ingo Molnardd41f592007-07-09 18:51:59 +02007862 /*
7863 * During early bootup we pretend to be a normal task:
7864 */
7865 current->sched_class = &fair_sched_class;
Ingo Molnar6892b752008-02-13 14:02:36 +01007866
7867 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007868}
7869
7870#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7871void __might_sleep(char *file, int line)
7872{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007873#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07007874 static unsigned long prev_jiffy; /* ratelimiting */
7875
7876 if ((in_atomic() || irqs_disabled()) &&
7877 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7878 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7879 return;
7880 prev_jiffy = jiffies;
Ingo Molnar91368d72006-03-23 03:00:54 -08007881 printk(KERN_ERR "BUG: sleeping function called from invalid"
Linus Torvalds1da177e2005-04-16 15:20:36 -07007882 " context at %s:%d\n", file, line);
7883 printk("in_atomic():%d, irqs_disabled():%d\n",
7884 in_atomic(), irqs_disabled());
Peter Zijlstraa4c410f2006-12-06 20:37:21 -08007885 debug_show_held_locks(current);
Ingo Molnar3117df02006-12-13 00:34:43 -08007886 if (irqs_disabled())
7887 print_irqtrace_events(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007888 dump_stack();
7889 }
7890#endif
7891}
7892EXPORT_SYMBOL(__might_sleep);
7893#endif
7894
7895#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02007896static void normalize_task(struct rq *rq, struct task_struct *p)
7897{
7898 int on_rq;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02007899
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02007900 update_rq_clock(rq);
7901 on_rq = p->se.on_rq;
7902 if (on_rq)
7903 deactivate_task(rq, p, 0);
7904 __setscheduler(rq, p, SCHED_NORMAL, 0);
7905 if (on_rq) {
7906 activate_task(rq, p, 0);
7907 resched_task(rq->curr);
7908 }
7909}
7910
Linus Torvalds1da177e2005-04-16 15:20:36 -07007911void normalize_rt_tasks(void)
7912{
Ingo Molnara0f98a12007-06-17 18:37:45 +02007913 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007914 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007915 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007916
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01007917 read_lock_irqsave(&tasklist_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02007918 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02007919 /*
7920 * Only normalize user tasks:
7921 */
7922 if (!p->mm)
7923 continue;
7924
Ingo Molnardd41f592007-07-09 18:51:59 +02007925 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02007926#ifdef CONFIG_SCHEDSTATS
7927 p->se.wait_start = 0;
7928 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007929 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02007930#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02007931
7932 if (!rt_task(p)) {
7933 /*
7934 * Renice negative nice level userspace
7935 * tasks back to 0:
7936 */
7937 if (TASK_NICE(p) < 0 && p->mm)
7938 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007939 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02007940 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007941
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01007942 spin_lock(&p->pi_lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -07007943 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007944
Ingo Molnar178be792007-10-15 17:00:18 +02007945 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02007946
Ingo Molnarb29739f2006-06-27 02:54:51 -07007947 __task_rq_unlock(rq);
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01007948 spin_unlock(&p->pi_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02007949 } while_each_thread(g, p);
7950
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01007951 read_unlock_irqrestore(&tasklist_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007952}
7953
7954#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07007955
7956#ifdef CONFIG_IA64
7957/*
7958 * These functions are only useful for the IA64 MCA handling.
7959 *
7960 * They can only be called when the whole system has been
7961 * stopped - every CPU needs to be quiescent, and no scheduling
7962 * activity can take place. Using them for anything else would
7963 * be a serious bug, and as a result, they aren't even visible
7964 * under any other configuration.
7965 */
7966
7967/**
7968 * curr_task - return the current task for a given cpu.
7969 * @cpu: the processor in question.
7970 *
7971 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7972 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007973struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07007974{
7975 return cpu_curr(cpu);
7976}
7977
7978/**
7979 * set_curr_task - set the current task for a given cpu.
7980 * @cpu: the processor in question.
7981 * @p: the task pointer to set.
7982 *
7983 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007984 * are serviced on a separate stack. It allows the architecture to switch the
7985 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07007986 * must be called with all CPU's synchronized, and interrupts disabled, the
7987 * and caller must save the original value of the current task (see
7988 * curr_task() above) and restore that value before reenabling interrupts and
7989 * re-starting the system.
7990 *
7991 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7992 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007993void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07007994{
7995 cpu_curr(cpu) = p;
7996}
7997
7998#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007999
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008000#ifdef CONFIG_FAIR_GROUP_SCHED
8001static void free_fair_sched_group(struct task_group *tg)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008002{
8003 int i;
8004
8005 for_each_possible_cpu(i) {
8006 if (tg->cfs_rq)
8007 kfree(tg->cfs_rq[i]);
8008 if (tg->se)
8009 kfree(tg->se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008010 }
8011
8012 kfree(tg->cfs_rq);
8013 kfree(tg->se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008014}
8015
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008016static
8017int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008018{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008019 struct cfs_rq *cfs_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008020 struct sched_entity *se, *parent_se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008021 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008022 int i;
8023
Mike Travis434d53b2008-04-04 18:11:04 -07008024 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008025 if (!tg->cfs_rq)
8026 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07008027 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008028 if (!tg->se)
8029 goto err;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008030
8031 tg->shares = NICE_0_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008032
8033 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008034 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008035
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008036 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8037 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008038 if (!cfs_rq)
8039 goto err;
8040
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008041 se = kmalloc_node(sizeof(struct sched_entity),
8042 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008043 if (!se)
8044 goto err;
8045
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008046 parent_se = parent ? parent->se[i] : NULL;
8047 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008048 }
8049
8050 return 1;
8051
8052 err:
8053 return 0;
8054}
8055
8056static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8057{
8058 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8059 &cpu_rq(cpu)->leaf_cfs_rq_list);
8060}
8061
8062static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8063{
8064 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8065}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008066#else /* !CONFG_FAIR_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008067static inline void free_fair_sched_group(struct task_group *tg)
8068{
8069}
8070
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008071static inline
8072int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008073{
8074 return 1;
8075}
8076
8077static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8078{
8079}
8080
8081static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8082{
8083}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008084#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008085
8086#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008087static void free_rt_sched_group(struct task_group *tg)
8088{
8089 int i;
8090
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008091 destroy_rt_bandwidth(&tg->rt_bandwidth);
8092
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008093 for_each_possible_cpu(i) {
8094 if (tg->rt_rq)
8095 kfree(tg->rt_rq[i]);
8096 if (tg->rt_se)
8097 kfree(tg->rt_se[i]);
8098 }
8099
8100 kfree(tg->rt_rq);
8101 kfree(tg->rt_se);
8102}
8103
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008104static
8105int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008106{
8107 struct rt_rq *rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008108 struct sched_rt_entity *rt_se, *parent_se;
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008109 struct rq *rq;
8110 int i;
8111
Mike Travis434d53b2008-04-04 18:11:04 -07008112 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008113 if (!tg->rt_rq)
8114 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07008115 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008116 if (!tg->rt_se)
8117 goto err;
8118
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008119 init_rt_bandwidth(&tg->rt_bandwidth,
8120 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008121
8122 for_each_possible_cpu(i) {
8123 rq = cpu_rq(i);
8124
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008125 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8126 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8127 if (!rt_rq)
8128 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008129
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008130 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8131 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8132 if (!rt_se)
8133 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008134
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008135 parent_se = parent ? parent->rt_se[i] : NULL;
8136 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008137 }
8138
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008139 return 1;
8140
8141 err:
8142 return 0;
8143}
8144
8145static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8146{
8147 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8148 &cpu_rq(cpu)->leaf_rt_rq_list);
8149}
8150
8151static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8152{
8153 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8154}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008155#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008156static inline void free_rt_sched_group(struct task_group *tg)
8157{
8158}
8159
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008160static inline
8161int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008162{
8163 return 1;
8164}
8165
8166static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8167{
8168}
8169
8170static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8171{
8172}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008173#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008174
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008175#ifdef CONFIG_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008176static void free_sched_group(struct task_group *tg)
8177{
8178 free_fair_sched_group(tg);
8179 free_rt_sched_group(tg);
8180 kfree(tg);
8181}
8182
8183/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008184struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008185{
8186 struct task_group *tg;
8187 unsigned long flags;
8188 int i;
8189
8190 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8191 if (!tg)
8192 return ERR_PTR(-ENOMEM);
8193
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008194 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008195 goto err;
8196
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008197 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008198 goto err;
8199
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008200 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008201 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008202 register_fair_sched_group(tg, i);
8203 register_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008204 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008205 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008206
8207 WARN_ON(!parent); /* root should already exist */
8208
8209 tg->parent = parent;
8210 list_add_rcu(&tg->siblings, &parent->children);
8211 INIT_LIST_HEAD(&tg->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008212 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008213
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008214 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008215
8216err:
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008217 free_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008218 return ERR_PTR(-ENOMEM);
8219}
8220
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008221/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008222static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008223{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008224 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008225 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008226}
8227
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008228/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02008229void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008230{
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008231 unsigned long flags;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008232 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008233
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008234 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008235 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008236 unregister_fair_sched_group(tg, i);
8237 unregister_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008238 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008239 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008240 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008241 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008242
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008243 /* wait for possible concurrent references to cfs_rqs complete */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008244 call_rcu(&tg->rcu, free_sched_group_rcu);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008245}
8246
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008247/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +02008248 * The caller of this function should have put the task in its new group
8249 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8250 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008251 */
8252void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008253{
8254 int on_rq, running;
8255 unsigned long flags;
8256 struct rq *rq;
8257
8258 rq = task_rq_lock(tsk, &flags);
8259
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008260 update_rq_clock(rq);
8261
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01008262 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008263 on_rq = tsk->se.on_rq;
8264
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008265 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008266 dequeue_task(rq, tsk, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008267 if (unlikely(running))
8268 tsk->sched_class->put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008269
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008270 set_task_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008271
Peter Zijlstra810b3812008-02-29 15:21:01 -05008272#ifdef CONFIG_FAIR_GROUP_SCHED
8273 if (tsk->sched_class->moved_group)
8274 tsk->sched_class->moved_group(tsk);
8275#endif
8276
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008277 if (unlikely(running))
8278 tsk->sched_class->set_curr_task(rq);
8279 if (on_rq)
Dmitry Adamushko7074bad2007-10-15 17:00:07 +02008280 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008281
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008282 task_rq_unlock(rq, &flags);
8283}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008284#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008285
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008286#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6363ca52008-05-29 11:28:57 +02008287static void set_se_shares(struct sched_entity *se, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008288{
8289 struct cfs_rq *cfs_rq = se->cfs_rq;
Ingo Molnar6363ca52008-05-29 11:28:57 +02008290 struct rq *rq = cfs_rq->rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008291 int on_rq;
8292
Ingo Molnar6363ca52008-05-29 11:28:57 +02008293 spin_lock_irq(&rq->lock);
8294
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008295 on_rq = se->on_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008296 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008297 dequeue_entity(cfs_rq, se, 0);
8298
8299 se->load.weight = shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02008300 se->load.inv_weight = 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008301
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008302 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008303 enqueue_entity(cfs_rq, se, 0);
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008304
Ingo Molnar6363ca52008-05-29 11:28:57 +02008305 spin_unlock_irq(&rq->lock);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008306}
8307
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008308static DEFINE_MUTEX(shares_mutex);
8309
Ingo Molnar4cf86d72007-10-15 17:00:14 +02008310int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008311{
8312 int i;
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008313 unsigned long flags;
Ingo Molnarc61935f2008-01-22 11:24:58 +01008314
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008315 /*
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008316 * We can't change the weight of the root cgroup.
8317 */
8318 if (!tg->se[0])
8319 return -EINVAL;
8320
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008321 if (shares < MIN_SHARES)
8322 shares = MIN_SHARES;
Miao Xiecb4ad1f2008-04-28 12:54:56 +08008323 else if (shares > MAX_SHARES)
8324 shares = MAX_SHARES;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008325
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008326 mutex_lock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008327 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008328 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008329
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008330 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008331 for_each_possible_cpu(i)
8332 unregister_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008333 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008334 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01008335
8336 /* wait for any ongoing reference to this group to finish */
8337 synchronize_sched();
8338
8339 /*
8340 * Now we are free to modify the group's share on each cpu
8341 * w/o tripping rebalance_share or load_balance_fair.
8342 */
8343 tg->shares = shares;
Ingo Molnar6363ca52008-05-29 11:28:57 +02008344 for_each_possible_cpu(i)
Miao Xiecb4ad1f2008-04-28 12:54:56 +08008345 set_se_shares(tg->se[i], shares);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01008346
8347 /*
8348 * Enable load balance activity on this group, by inserting it back on
8349 * each cpu's rq->leaf_cfs_rq_list.
8350 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008351 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008352 for_each_possible_cpu(i)
8353 register_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008354 list_add_rcu(&tg->siblings, &tg->parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008355 spin_unlock_irqrestore(&task_group_lock, flags);
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008356done:
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008357 mutex_unlock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008358 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008359}
8360
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008361unsigned long sched_group_shares(struct task_group *tg)
8362{
8363 return tg->shares;
8364}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008365#endif
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008366
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008367#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008368/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008369 * Ensure that the real time constraints are schedulable.
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008370 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008371static DEFINE_MUTEX(rt_constraints_mutex);
8372
8373static unsigned long to_ratio(u64 period, u64 runtime)
8374{
8375 if (runtime == RUNTIME_INF)
8376 return 1ULL << 16;
8377
Roman Zippel6f6d6a12008-05-01 04:34:28 -07008378 return div64_u64(runtime << 16, period);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008379}
8380
Peter Zijlstrab40b2e82008-04-19 19:45:00 +02008381#ifdef CONFIG_CGROUP_SCHED
8382static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8383{
8384 struct task_group *tgi, *parent = tg->parent;
8385 unsigned long total = 0;
8386
8387 if (!parent) {
8388 if (global_rt_period() < period)
8389 return 0;
8390
8391 return to_ratio(period, runtime) <
8392 to_ratio(global_rt_period(), global_rt_runtime());
8393 }
8394
8395 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8396 return 0;
8397
8398 rcu_read_lock();
8399 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8400 if (tgi == tg)
8401 continue;
8402
8403 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8404 tgi->rt_bandwidth.rt_runtime);
8405 }
8406 rcu_read_unlock();
8407
8408 return total + to_ratio(period, runtime) <
8409 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8410 parent->rt_bandwidth.rt_runtime);
8411}
8412#elif defined CONFIG_USER_SCHED
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008413static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008414{
8415 struct task_group *tgi;
8416 unsigned long total = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008417 unsigned long global_ratio =
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008418 to_ratio(global_rt_period(), global_rt_runtime());
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008419
8420 rcu_read_lock();
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008421 list_for_each_entry_rcu(tgi, &task_groups, list) {
8422 if (tgi == tg)
8423 continue;
8424
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008425 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8426 tgi->rt_bandwidth.rt_runtime);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008427 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008428 rcu_read_unlock();
8429
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008430 return total + to_ratio(period, runtime) < global_ratio;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008431}
Peter Zijlstrab40b2e82008-04-19 19:45:00 +02008432#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008433
Dhaval Giani521f1a242008-02-28 15:21:56 +05308434/* Must be called with tasklist_lock held */
8435static inline int tg_has_rt_tasks(struct task_group *tg)
8436{
8437 struct task_struct *g, *p;
8438 do_each_thread(g, p) {
8439 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8440 return 1;
8441 } while_each_thread(g, p);
8442 return 0;
8443}
8444
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008445static int tg_set_bandwidth(struct task_group *tg,
8446 u64 rt_period, u64 rt_runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008447{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008448 int i, err = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008449
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008450 mutex_lock(&rt_constraints_mutex);
Dhaval Giani521f1a242008-02-28 15:21:56 +05308451 read_lock(&tasklist_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008452 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
Dhaval Giani521f1a242008-02-28 15:21:56 +05308453 err = -EBUSY;
8454 goto unlock;
8455 }
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008456 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8457 err = -EINVAL;
8458 goto unlock;
8459 }
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008460
8461 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008462 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8463 tg->rt_bandwidth.rt_runtime = rt_runtime;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008464
8465 for_each_possible_cpu(i) {
8466 struct rt_rq *rt_rq = tg->rt_rq[i];
8467
8468 spin_lock(&rt_rq->rt_runtime_lock);
8469 rt_rq->rt_runtime = rt_runtime;
8470 spin_unlock(&rt_rq->rt_runtime_lock);
8471 }
8472 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008473 unlock:
Dhaval Giani521f1a242008-02-28 15:21:56 +05308474 read_unlock(&tasklist_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008475 mutex_unlock(&rt_constraints_mutex);
8476
8477 return err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008478}
8479
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008480int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8481{
8482 u64 rt_runtime, rt_period;
8483
8484 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8485 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8486 if (rt_runtime_us < 0)
8487 rt_runtime = RUNTIME_INF;
8488
8489 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8490}
8491
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008492long sched_group_rt_runtime(struct task_group *tg)
8493{
8494 u64 rt_runtime_us;
8495
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008496 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008497 return -1;
8498
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008499 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008500 do_div(rt_runtime_us, NSEC_PER_USEC);
8501 return rt_runtime_us;
8502}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008503
8504int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8505{
8506 u64 rt_runtime, rt_period;
8507
8508 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8509 rt_runtime = tg->rt_bandwidth.rt_runtime;
8510
8511 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8512}
8513
8514long sched_group_rt_period(struct task_group *tg)
8515{
8516 u64 rt_period_us;
8517
8518 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8519 do_div(rt_period_us, NSEC_PER_USEC);
8520 return rt_period_us;
8521}
8522
8523static int sched_rt_global_constraints(void)
8524{
8525 int ret = 0;
8526
8527 mutex_lock(&rt_constraints_mutex);
8528 if (!__rt_schedulable(NULL, 1, 0))
8529 ret = -EINVAL;
8530 mutex_unlock(&rt_constraints_mutex);
8531
8532 return ret;
8533}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008534#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008535static int sched_rt_global_constraints(void)
8536{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008537 unsigned long flags;
8538 int i;
8539
8540 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8541 for_each_possible_cpu(i) {
8542 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8543
8544 spin_lock(&rt_rq->rt_runtime_lock);
8545 rt_rq->rt_runtime = global_rt_runtime();
8546 spin_unlock(&rt_rq->rt_runtime_lock);
8547 }
8548 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8549
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008550 return 0;
8551}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008552#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008553
8554int sched_rt_handler(struct ctl_table *table, int write,
8555 struct file *filp, void __user *buffer, size_t *lenp,
8556 loff_t *ppos)
8557{
8558 int ret;
8559 int old_period, old_runtime;
8560 static DEFINE_MUTEX(mutex);
8561
8562 mutex_lock(&mutex);
8563 old_period = sysctl_sched_rt_period;
8564 old_runtime = sysctl_sched_rt_runtime;
8565
8566 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8567
8568 if (!ret && write) {
8569 ret = sched_rt_global_constraints();
8570 if (ret) {
8571 sysctl_sched_rt_period = old_period;
8572 sysctl_sched_rt_runtime = old_runtime;
8573 } else {
8574 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8575 def_rt_bandwidth.rt_period =
8576 ns_to_ktime(global_rt_period());
8577 }
8578 }
8579 mutex_unlock(&mutex);
8580
8581 return ret;
8582}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008583
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008584#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008585
8586/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02008587static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008588{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008589 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8590 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008591}
8592
8593static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +02008594cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008595{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008596 struct task_group *tg, *parent;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008597
Paul Menage2b01dfe2007-10-24 18:23:50 +02008598 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008599 /* This is early initialization for the top cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02008600 init_task_group.css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008601 return &init_task_group.css;
8602 }
8603
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008604 parent = cgroup_tg(cgrp->parent);
8605 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008606 if (IS_ERR(tg))
8607 return ERR_PTR(-ENOMEM);
8608
8609 /* Bind the cgroup to task_group object we just created */
Paul Menage2b01dfe2007-10-24 18:23:50 +02008610 tg->css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008611
8612 return &tg->css;
8613}
8614
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008615static void
8616cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008617{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008618 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008619
8620 sched_destroy_group(tg);
8621}
8622
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008623static int
8624cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8625 struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008626{
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008627#ifdef CONFIG_RT_GROUP_SCHED
8628 /* Don't accept realtime tasks when there is no way for them to run */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008629 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008630 return -EINVAL;
8631#else
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008632 /* We don't support RT-tasks being in separate groups */
8633 if (tsk->sched_class != &fair_sched_class)
8634 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008635#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008636
8637 return 0;
8638}
8639
8640static void
Paul Menage2b01dfe2007-10-24 18:23:50 +02008641cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008642 struct cgroup *old_cont, struct task_struct *tsk)
8643{
8644 sched_move_task(tsk);
8645}
8646
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008647#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagef4c753b2008-04-29 00:59:56 -07008648static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
Paul Menage2b01dfe2007-10-24 18:23:50 +02008649 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008650{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008651 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008652}
8653
Paul Menagef4c753b2008-04-29 00:59:56 -07008654static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008655{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008656 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008657
8658 return (u64) tg->shares;
8659}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008660#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008661
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008662#ifdef CONFIG_RT_GROUP_SCHED
Mirco Tischler0c708142008-05-14 16:05:46 -07008663static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
Paul Menage06ecb272008-04-29 01:00:06 -07008664 s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008665{
Paul Menage06ecb272008-04-29 01:00:06 -07008666 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008667}
8668
Paul Menage06ecb272008-04-29 01:00:06 -07008669static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008670{
Paul Menage06ecb272008-04-29 01:00:06 -07008671 return sched_group_rt_runtime(cgroup_tg(cgrp));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008672}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008673
8674static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8675 u64 rt_period_us)
8676{
8677 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8678}
8679
8680static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8681{
8682 return sched_group_rt_period(cgroup_tg(cgrp));
8683}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008684#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008685
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008686static struct cftype cpu_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008687#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008688 {
8689 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -07008690 .read_u64 = cpu_shares_read_u64,
8691 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008692 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008693#endif
8694#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008695 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008696 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -07008697 .read_s64 = cpu_rt_runtime_read,
8698 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008699 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008700 {
8701 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -07008702 .read_u64 = cpu_rt_period_read_uint,
8703 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008704 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008705#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008706};
8707
8708static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8709{
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008710 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008711}
8712
8713struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +01008714 .name = "cpu",
8715 .create = cpu_cgroup_create,
8716 .destroy = cpu_cgroup_destroy,
8717 .can_attach = cpu_cgroup_can_attach,
8718 .attach = cpu_cgroup_attach,
8719 .populate = cpu_cgroup_populate,
8720 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008721 .early_init = 1,
8722};
8723
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008724#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008725
8726#ifdef CONFIG_CGROUP_CPUACCT
8727
8728/*
8729 * CPU accounting code for task groups.
8730 *
8731 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8732 * (balbir@in.ibm.com).
8733 */
8734
8735/* track cpu usage of a group of tasks */
8736struct cpuacct {
8737 struct cgroup_subsys_state css;
8738 /* cpuusage holds pointer to a u64-type object on every cpu */
8739 u64 *cpuusage;
8740};
8741
8742struct cgroup_subsys cpuacct_subsys;
8743
8744/* return cpu accounting group corresponding to this container */
Dhaval Giani32cd7562008-02-29 10:02:43 +05308745static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008746{
Dhaval Giani32cd7562008-02-29 10:02:43 +05308747 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008748 struct cpuacct, css);
8749}
8750
8751/* return cpu accounting group to which this task belongs */
8752static inline struct cpuacct *task_ca(struct task_struct *tsk)
8753{
8754 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8755 struct cpuacct, css);
8756}
8757
8758/* create a new cpu accounting group */
8759static struct cgroup_subsys_state *cpuacct_create(
Dhaval Giani32cd7562008-02-29 10:02:43 +05308760 struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008761{
8762 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8763
8764 if (!ca)
8765 return ERR_PTR(-ENOMEM);
8766
8767 ca->cpuusage = alloc_percpu(u64);
8768 if (!ca->cpuusage) {
8769 kfree(ca);
8770 return ERR_PTR(-ENOMEM);
8771 }
8772
8773 return &ca->css;
8774}
8775
8776/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008777static void
Dhaval Giani32cd7562008-02-29 10:02:43 +05308778cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008779{
Dhaval Giani32cd7562008-02-29 10:02:43 +05308780 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008781
8782 free_percpu(ca->cpuusage);
8783 kfree(ca);
8784}
8785
8786/* return total cpu usage (in nanoseconds) of a group */
Dhaval Giani32cd7562008-02-29 10:02:43 +05308787static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008788{
Dhaval Giani32cd7562008-02-29 10:02:43 +05308789 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008790 u64 totalcpuusage = 0;
8791 int i;
8792
8793 for_each_possible_cpu(i) {
8794 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8795
8796 /*
8797 * Take rq->lock to make 64-bit addition safe on 32-bit
8798 * platforms.
8799 */
8800 spin_lock_irq(&cpu_rq(i)->lock);
8801 totalcpuusage += *cpuusage;
8802 spin_unlock_irq(&cpu_rq(i)->lock);
8803 }
8804
8805 return totalcpuusage;
8806}
8807
Dhaval Giani0297b802008-02-29 10:02:44 +05308808static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8809 u64 reset)
8810{
8811 struct cpuacct *ca = cgroup_ca(cgrp);
8812 int err = 0;
8813 int i;
8814
8815 if (reset) {
8816 err = -EINVAL;
8817 goto out;
8818 }
8819
8820 for_each_possible_cpu(i) {
8821 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8822
8823 spin_lock_irq(&cpu_rq(i)->lock);
8824 *cpuusage = 0;
8825 spin_unlock_irq(&cpu_rq(i)->lock);
8826 }
8827out:
8828 return err;
8829}
8830
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008831static struct cftype files[] = {
8832 {
8833 .name = "usage",
Paul Menagef4c753b2008-04-29 00:59:56 -07008834 .read_u64 = cpuusage_read,
8835 .write_u64 = cpuusage_write,
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008836 },
8837};
8838
Dhaval Giani32cd7562008-02-29 10:02:43 +05308839static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008840{
Dhaval Giani32cd7562008-02-29 10:02:43 +05308841 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008842}
8843
8844/*
8845 * charge this task's execution time to its accounting group.
8846 *
8847 * called with rq->lock held.
8848 */
8849static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8850{
8851 struct cpuacct *ca;
8852
8853 if (!cpuacct_subsys.active)
8854 return;
8855
8856 ca = task_ca(tsk);
8857 if (ca) {
8858 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8859
8860 *cpuusage += cputime;
8861 }
8862}
8863
8864struct cgroup_subsys cpuacct_subsys = {
8865 .name = "cpuacct",
8866 .create = cpuacct_create,
8867 .destroy = cpuacct_destroy,
8868 .populate = cpuacct_populate,
8869 .subsys_id = cpuacct_subsys_id,
8870};
8871#endif /* CONFIG_CGROUP_CPUACCT */