blob: cdd09462fc98e8305e07d7422f64939afe2012cc [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 */
Dmitry Adamushko20b63312008-06-11 00:58:30 +0200156 struct list_head queue[MAX_RT_PRIO];
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200157};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200159struct rt_bandwidth {
Ingo Molnarea736ed2008-03-25 13:51:45 +0100160 /* nests inside the rq lock: */
161 spinlock_t rt_runtime_lock;
162 ktime_t rt_period;
163 u64 rt_runtime;
164 struct hrtimer rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200165};
166
167static struct rt_bandwidth def_rt_bandwidth;
168
169static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
170
171static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
172{
173 struct rt_bandwidth *rt_b =
174 container_of(timer, struct rt_bandwidth, rt_period_timer);
175 ktime_t now;
176 int overrun;
177 int idle = 0;
178
179 for (;;) {
180 now = hrtimer_cb_get_time(timer);
181 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
182
183 if (!overrun)
184 break;
185
186 idle = do_sched_rt_period_timer(rt_b, overrun);
187 }
188
189 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
190}
191
192static
193void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
194{
195 rt_b->rt_period = ns_to_ktime(period);
196 rt_b->rt_runtime = runtime;
197
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200198 spin_lock_init(&rt_b->rt_runtime_lock);
199
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200200 hrtimer_init(&rt_b->rt_period_timer,
201 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
202 rt_b->rt_period_timer.function = sched_rt_period_timer;
203 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
204}
205
206static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
207{
208 ktime_t now;
209
210 if (rt_b->rt_runtime == RUNTIME_INF)
211 return;
212
213 if (hrtimer_active(&rt_b->rt_period_timer))
214 return;
215
216 spin_lock(&rt_b->rt_runtime_lock);
217 for (;;) {
218 if (hrtimer_active(&rt_b->rt_period_timer))
219 break;
220
221 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
222 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
223 hrtimer_start(&rt_b->rt_period_timer,
224 rt_b->rt_period_timer.expires,
225 HRTIMER_MODE_ABS);
226 }
227 spin_unlock(&rt_b->rt_runtime_lock);
228}
229
230#ifdef CONFIG_RT_GROUP_SCHED
231static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
232{
233 hrtimer_cancel(&rt_b->rt_period_timer);
234}
235#endif
236
Heiko Carstens712555e2008-04-28 11:33:07 +0200237/*
238 * sched_domains_mutex serializes calls to arch_init_sched_domains,
239 * detach_destroy_domains and partition_sched_domains.
240 */
241static DEFINE_MUTEX(sched_domains_mutex);
242
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100243#ifdef CONFIG_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200244
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700245#include <linux/cgroup.h>
246
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200247struct cfs_rq;
248
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100249static LIST_HEAD(task_groups);
250
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200251/* task group related information */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200252struct task_group {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100253#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700254 struct cgroup_subsys_state css;
255#endif
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100256
257#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200258 /* schedulable entities of this group on each cpu */
259 struct sched_entity **se;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq **cfs_rq;
262 unsigned long shares;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100263#endif
264
265#ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity **rt_se;
267 struct rt_rq **rt_rq;
268
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200269 struct rt_bandwidth rt_bandwidth;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100270#endif
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100271
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +0100272 struct rcu_head rcu;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100273 struct list_head list;
Peter Zijlstraf473aa52008-04-19 19:45:00 +0200274
275 struct task_group *parent;
276 struct list_head siblings;
277 struct list_head children;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200278};
279
Dhaval Giani354d60c2008-04-19 19:44:59 +0200280#ifdef CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200281
282/*
283 * Root task group.
284 * Every UID task group (including init_task_group aka UID-0) will
285 * be a child to this group.
286 */
287struct task_group root_task_group;
288
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100289#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200290/* Default task group's sched entity on each cpu */
291static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
292/* Default task group's cfs_rq on each cpu */
293static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200294#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100295
296#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100297static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
298static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200299#endif /* CONFIG_RT_GROUP_SCHED */
300#else /* !CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200301#define root_task_group init_task_group
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200302#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100303
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100304/* task_group_lock serializes add/remove of task groups and also changes to
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100305 * a task group's cpu shares.
306 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100307static DEFINE_SPINLOCK(task_group_lock);
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100308
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100309#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100310#ifdef CONFIG_USER_SCHED
Ingo Molnar0eab9142008-01-25 21:08:19 +0100311# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200312#else /* !CONFIG_USER_SCHED */
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100313# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200314#endif /* CONFIG_USER_SCHED */
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200315
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800316/*
Lai Jiangshan2e084782008-06-12 16:42:58 +0800317 * A weight of 0 or 1 can cause arithmetics problems.
318 * A weight of a cfs_rq is the sum of weights of which entities
319 * are queued on this cfs_rq, so a weight of a entity should not be
320 * too large, so as the shares value of a task group.
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800321 * (The default weight is 1024 - so there's no practical
322 * limitation from this.)
323 */
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200324#define MIN_SHARES 2
Lai Jiangshan2e084782008-06-12 16:42:58 +0800325#define MAX_SHARES (1UL << 18)
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200326
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100327static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100328#endif
329
330/* Default task group.
331 * Every task in system belong to this group at bootup.
332 */
Mike Travis434d53b2008-04-04 18:11:04 -0700333struct task_group init_task_group;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200334
335/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200336static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200337{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200338 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200339
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100340#ifdef CONFIG_USER_SCHED
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200341 tg = p->user->tg;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100342#elif defined(CONFIG_CGROUP_SCHED)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700343 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
344 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200345#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100346 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200347#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200348 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200349}
350
351/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100352static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200353{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100354#ifdef CONFIG_FAIR_GROUP_SCHED
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100355 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
356 p->se.parent = task_group(p)->se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100357#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100358
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100359#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100360 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
361 p->rt.parent = task_group(p)->rt_se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100362#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200363}
364
365#else
366
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100367static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200368
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100369#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200370
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200371/* CFS-related fields in a runqueue */
372struct cfs_rq {
373 struct load_weight load;
374 unsigned long nr_running;
375
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200376 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200377 u64 min_vruntime;
Peter Zijlstra103638d92008-06-27 13:41:16 +0200378 u64 pair_start;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200379
380 struct rb_root tasks_timeline;
381 struct rb_node *rb_leftmost;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +0200382
383 struct list_head tasks;
384 struct list_head *balance_iterator;
385
386 /*
387 * 'curr' points to currently running entity on this cfs_rq.
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200388 * It is set to NULL otherwise (i.e when none are currently running).
389 */
Peter Zijlstraaa2ac252008-03-14 21:12:12 +0100390 struct sched_entity *curr, *next;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200391
392 unsigned long nr_spread_over;
393
Ingo Molnar62160e3f2007-10-15 17:00:03 +0200394#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200395 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
396
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100397 /*
398 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200399 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
400 * (like users, containers etc.)
401 *
402 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
403 * list is used during load balance.
404 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100405 struct list_head leaf_cfs_rq_list;
406 struct task_group *tg; /* group that "owns" this runqueue */
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200407
408#ifdef CONFIG_SMP
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200409 /*
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200410 * the part of load.weight contributed by tasks
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200411 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200412 unsigned long task_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200413
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200414 /*
415 * h_load = weight * f(tg)
416 *
417 * Where f(tg) is the recursive weight fraction assigned to
418 * this group.
419 */
420 unsigned long h_load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200421
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200422 /*
423 * this cpu's part of tg->shares
424 */
425 unsigned long shares;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200426#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200427#endif
428};
429
430/* Real-Time classes' related field in a runqueue: */
431struct rt_rq {
432 struct rt_prio_array active;
Steven Rostedt63489e42008-01-25 21:08:03 +0100433 unsigned long rt_nr_running;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100434#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100435 int highest_prio; /* highest queued rt task prio */
436#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100437#ifdef CONFIG_SMP
Gregory Haskins73fe6aa2008-01-25 21:08:07 +0100438 unsigned long rt_nr_migratory;
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +0100439 int overloaded;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100440#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100441 int rt_throttled;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100442 u64 rt_time;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200443 u64 rt_runtime;
Ingo Molnarea736ed2008-03-25 13:51:45 +0100444 /* Nests inside the rq lock: */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200445 spinlock_t rt_runtime_lock;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100446
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100447#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100448 unsigned long rt_nr_boosted;
449
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100450 struct rq *rq;
451 struct list_head leaf_rt_rq_list;
452 struct task_group *tg;
453 struct sched_rt_entity *rt_se;
454#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200455};
456
Gregory Haskins57d885f2008-01-25 21:08:18 +0100457#ifdef CONFIG_SMP
458
459/*
460 * We add the notion of a root-domain which will be used to define per-domain
Ingo Molnar0eab9142008-01-25 21:08:19 +0100461 * variables. Each exclusive cpuset essentially defines an island domain by
462 * fully partitioning the member cpus from any other cpuset. Whenever a new
Gregory Haskins57d885f2008-01-25 21:08:18 +0100463 * exclusive cpuset is created, we also create and attach a new root-domain
464 * object.
465 *
Gregory Haskins57d885f2008-01-25 21:08:18 +0100466 */
467struct root_domain {
468 atomic_t refcount;
469 cpumask_t span;
470 cpumask_t online;
Gregory Haskins637f5082008-01-25 21:08:18 +0100471
Ingo Molnar0eab9142008-01-25 21:08:19 +0100472 /*
Gregory Haskins637f5082008-01-25 21:08:18 +0100473 * The "RT overload" flag: it gets set if a CPU has more than
474 * one runnable RT task.
475 */
476 cpumask_t rto_mask;
Ingo Molnar0eab9142008-01-25 21:08:19 +0100477 atomic_t rto_count;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200478#ifdef CONFIG_SMP
479 struct cpupri cpupri;
480#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +0100481};
482
Gregory Haskinsdc938522008-01-25 21:08:26 +0100483/*
484 * By default the system creates a single root-domain with all cpus as
485 * members (mimicking the global state we have today).
486 */
Gregory Haskins57d885f2008-01-25 21:08:18 +0100487static struct root_domain def_root_domain;
488
489#endif
490
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200491/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 * This is the main, per-CPU runqueue data structure.
493 *
494 * Locking rule: those places that want to lock multiple runqueues
495 * (such as the load balancing or the thread migration code), lock
496 * acquire operations must be ordered by ascending &runqueue.
497 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700498struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200499 /* runqueue lock: */
500 spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501
502 /*
503 * nr_running and cpu_load should be in the same cacheline because
504 * remote CPUs use both these fields when doing load calculation.
505 */
506 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200507 #define CPU_LOAD_IDX_MAX 5
508 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh Bbdecea32007-05-08 00:32:48 -0700509 unsigned char idle_at_tick;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700510#ifdef CONFIG_NO_HZ
Guillaume Chazarain15934a32008-04-19 19:44:57 +0200511 unsigned long last_tick_seen;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700512 unsigned char in_nohz_recently;
513#endif
Ingo Molnard8016492007-10-18 21:32:55 +0200514 /* capture load from *all* tasks on this cpu: */
515 struct load_weight load;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200516 unsigned long nr_load_updates;
517 u64 nr_switches;
518
519 struct cfs_rq cfs;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100520 struct rt_rq rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100521
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200522#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200523 /* list of leaf cfs_rq on this cpu: */
524 struct list_head leaf_cfs_rq_list;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100525#endif
526#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100527 struct list_head leaf_rt_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529
530 /*
531 * This is part of a global counter where only the total sum
532 * over all CPUs matters. A task can increase this counter on
533 * one CPU and if it got migrated afterwards it may decrease
534 * it on another CPU. Always updated under the runqueue lock:
535 */
536 unsigned long nr_uninterruptible;
537
Ingo Molnar36c8b582006-07-03 00:25:41 -0700538 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800539 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200541
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200542 u64 clock;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200543
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 atomic_t nr_iowait;
545
546#ifdef CONFIG_SMP
Ingo Molnar0eab9142008-01-25 21:08:19 +0100547 struct root_domain *rd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 struct sched_domain *sd;
549
550 /* For active balancing */
551 int active_balance;
552 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200553 /* cpu of this runqueue: */
554 int cpu;
Gregory Haskins1f11eb62008-06-04 15:04:05 -0400555 int online;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556
Ingo Molnar36c8b582006-07-03 00:25:41 -0700557 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 struct list_head migration_queue;
559#endif
560
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100561#ifdef CONFIG_SCHED_HRTICK
562 unsigned long hrtick_flags;
563 ktime_t hrtick_expire;
564 struct hrtimer hrtick_timer;
565#endif
566
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567#ifdef CONFIG_SCHEDSTATS
568 /* latency stats */
569 struct sched_info rq_sched_info;
570
571 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200572 unsigned int yld_exp_empty;
573 unsigned int yld_act_empty;
574 unsigned int yld_both_empty;
575 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576
577 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200578 unsigned int sched_switch;
579 unsigned int sched_count;
580 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581
582 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200583 unsigned int ttwu_count;
584 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200585
586 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200587 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588#endif
Ingo Molnarfcb99372006-07-03 00:25:10 -0700589 struct lock_class_key rq_lock_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590};
591
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700592static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593
Ingo Molnardd41f592007-07-09 18:51:59 +0200594static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
595{
596 rq->curr->sched_class->check_preempt_curr(rq, p);
597}
598
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700599static inline int cpu_of(struct rq *rq)
600{
601#ifdef CONFIG_SMP
602 return rq->cpu;
603#else
604 return 0;
605#endif
606}
607
Ingo Molnar20d315d2007-07-09 18:51:58 +0200608/*
Nick Piggin674311d2005-06-25 14:57:27 -0700609 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700610 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700611 *
612 * The domain tree of any CPU may only be accessed from within
613 * preempt-disabled sections.
614 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700615#define for_each_domain(cpu, __sd) \
616 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617
618#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
619#define this_rq() (&__get_cpu_var(runqueues))
620#define task_rq(p) cpu_rq(task_cpu(p))
621#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
622
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200623static inline void update_rq_clock(struct rq *rq)
624{
625 rq->clock = sched_clock_cpu(cpu_of(rq));
626}
627
Ingo Molnare436d802007-07-19 21:28:35 +0200628/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200629 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
630 */
631#ifdef CONFIG_SCHED_DEBUG
632# define const_debug __read_mostly
633#else
634# define const_debug static const
635#endif
636
637/*
638 * Debugging: various feature bits
639 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200640
641#define SCHED_FEAT(name, enabled) \
642 __SCHED_FEAT_##name ,
643
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200644enum {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200645#include "sched_features.h"
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200646};
647
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200648#undef SCHED_FEAT
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200649
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200650#define SCHED_FEAT(name, enabled) \
651 (1UL << __SCHED_FEAT_##name) * enabled |
652
653const_debug unsigned int sysctl_sched_features =
654#include "sched_features.h"
655 0;
656
657#undef SCHED_FEAT
658
659#ifdef CONFIG_SCHED_DEBUG
660#define SCHED_FEAT(name, enabled) \
661 #name ,
662
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700663static __read_mostly char *sched_feat_names[] = {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200664#include "sched_features.h"
665 NULL
666};
667
668#undef SCHED_FEAT
669
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700670static int sched_feat_open(struct inode *inode, struct file *filp)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200671{
672 filp->private_data = inode->i_private;
673 return 0;
674}
675
676static ssize_t
677sched_feat_read(struct file *filp, char __user *ubuf,
678 size_t cnt, loff_t *ppos)
679{
680 char *buf;
681 int r = 0;
682 int len = 0;
683 int i;
684
685 for (i = 0; sched_feat_names[i]; i++) {
686 len += strlen(sched_feat_names[i]);
687 len += 4;
688 }
689
690 buf = kmalloc(len + 2, GFP_KERNEL);
691 if (!buf)
692 return -ENOMEM;
693
694 for (i = 0; sched_feat_names[i]; i++) {
695 if (sysctl_sched_features & (1UL << i))
696 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
697 else
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200698 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200699 }
700
701 r += sprintf(buf + r, "\n");
702 WARN_ON(r >= len + 2);
703
704 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
705
706 kfree(buf);
707
708 return r;
709}
710
711static ssize_t
712sched_feat_write(struct file *filp, const char __user *ubuf,
713 size_t cnt, loff_t *ppos)
714{
715 char buf[64];
716 char *cmp = buf;
717 int neg = 0;
718 int i;
719
720 if (cnt > 63)
721 cnt = 63;
722
723 if (copy_from_user(&buf, ubuf, cnt))
724 return -EFAULT;
725
726 buf[cnt] = 0;
727
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200728 if (strncmp(buf, "NO_", 3) == 0) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200729 neg = 1;
730 cmp += 3;
731 }
732
733 for (i = 0; sched_feat_names[i]; i++) {
734 int len = strlen(sched_feat_names[i]);
735
736 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
737 if (neg)
738 sysctl_sched_features &= ~(1UL << i);
739 else
740 sysctl_sched_features |= (1UL << i);
741 break;
742 }
743 }
744
745 if (!sched_feat_names[i])
746 return -EINVAL;
747
748 filp->f_pos += cnt;
749
750 return cnt;
751}
752
753static struct file_operations sched_feat_fops = {
754 .open = sched_feat_open,
755 .read = sched_feat_read,
756 .write = sched_feat_write,
757};
758
759static __init int sched_init_debug(void)
760{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200761 debugfs_create_file("sched_features", 0644, NULL, NULL,
762 &sched_feat_fops);
763
764 return 0;
765}
766late_initcall(sched_init_debug);
767
768#endif
769
770#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200771
772/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100773 * Number of tasks to iterate in a single balance run.
774 * Limited because this is done with IRQs disabled.
775 */
776const_debug unsigned int sysctl_sched_nr_migrate = 32;
777
778/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100779 * period over which we measure -rt task cpu usage in us.
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100780 * default: 1s
781 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100782unsigned int sysctl_sched_rt_period = 1000000;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100783
Ingo Molnar6892b752008-02-13 14:02:36 +0100784static __read_mostly int scheduler_running;
785
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100786/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100787 * part of the period that we allow rt tasks to run in us.
788 * default: 0.95s
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100789 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100790int sysctl_sched_rt_runtime = 950000;
791
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200792static inline u64 global_rt_period(void)
793{
794 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
795}
796
797static inline u64 global_rt_runtime(void)
798{
799 if (sysctl_sched_rt_period < 0)
800 return RUNTIME_INF;
801
802 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
803}
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100804
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700806# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700808#ifndef finish_arch_switch
809# define finish_arch_switch(prev) do { } while (0)
810#endif
811
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100812static inline int task_current(struct rq *rq, struct task_struct *p)
813{
814 return rq->curr == p;
815}
816
Nick Piggin4866cde2005-06-25 14:57:23 -0700817#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700818static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700819{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100820 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700821}
822
Ingo Molnar70b97a72006-07-03 00:25:42 -0700823static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700824{
825}
826
Ingo Molnar70b97a72006-07-03 00:25:42 -0700827static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700828{
Ingo Molnarda04c032005-09-13 11:17:59 +0200829#ifdef CONFIG_DEBUG_SPINLOCK
830 /* this is a valid case when another task releases the spinlock */
831 rq->lock.owner = current;
832#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700833 /*
834 * If we are tracking spinlock dependencies then we have to
835 * fix up the runqueue lock - which gets 'carried over' from
836 * prev into current:
837 */
838 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
839
Nick Piggin4866cde2005-06-25 14:57:23 -0700840 spin_unlock_irq(&rq->lock);
841}
842
843#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700844static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700845{
846#ifdef CONFIG_SMP
847 return p->oncpu;
848#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100849 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700850#endif
851}
852
Ingo Molnar70b97a72006-07-03 00:25:42 -0700853static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700854{
855#ifdef CONFIG_SMP
856 /*
857 * We can optimise this out completely for !SMP, because the
858 * SMP rebalancing from interrupt is the only thing that cares
859 * here.
860 */
861 next->oncpu = 1;
862#endif
863#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
864 spin_unlock_irq(&rq->lock);
865#else
866 spin_unlock(&rq->lock);
867#endif
868}
869
Ingo Molnar70b97a72006-07-03 00:25:42 -0700870static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700871{
872#ifdef CONFIG_SMP
873 /*
874 * After ->oncpu is cleared, the task can be moved to a different CPU.
875 * We must ensure this doesn't happen until the switch is completely
876 * finished.
877 */
878 smp_wmb();
879 prev->oncpu = 0;
880#endif
881#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
882 local_irq_enable();
883#endif
884}
885#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886
887/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700888 * __task_rq_lock - lock the runqueue a given task resides on.
889 * Must be called interrupts disabled.
890 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700891static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700892 __acquires(rq->lock)
893{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200894 for (;;) {
895 struct rq *rq = task_rq(p);
896 spin_lock(&rq->lock);
897 if (likely(rq == task_rq(p)))
898 return rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700899 spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700900 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700901}
902
903/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100905 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 * explicitly disabling preemption.
907 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700908static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 __acquires(rq->lock)
910{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700911 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912
Andi Kleen3a5c3592007-10-15 17:00:14 +0200913 for (;;) {
914 local_irq_save(*flags);
915 rq = task_rq(p);
916 spin_lock(&rq->lock);
917 if (likely(rq == task_rq(p)))
918 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919 spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921}
922
Alexey Dobriyana9957442007-10-15 17:00:13 +0200923static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700924 __releases(rq->lock)
925{
926 spin_unlock(&rq->lock);
927}
928
Ingo Molnar70b97a72006-07-03 00:25:42 -0700929static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 __releases(rq->lock)
931{
932 spin_unlock_irqrestore(&rq->lock, *flags);
933}
934
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -0800936 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 */
Alexey Dobriyana9957442007-10-15 17:00:13 +0200938static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 __acquires(rq->lock)
940{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700941 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942
943 local_irq_disable();
944 rq = this_rq();
945 spin_lock(&rq->lock);
946
947 return rq;
948}
949
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100950static void __resched_task(struct task_struct *p, int tif_bit);
951
952static inline void resched_task(struct task_struct *p)
953{
954 __resched_task(p, TIF_NEED_RESCHED);
955}
956
957#ifdef CONFIG_SCHED_HRTICK
958/*
959 * Use HR-timers to deliver accurate preemption points.
960 *
961 * Its all a bit involved since we cannot program an hrt while holding the
962 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
963 * reschedule event.
964 *
965 * When we get rescheduled we reprogram the hrtick_timer outside of the
966 * rq->lock.
967 */
968static inline void resched_hrt(struct task_struct *p)
969{
970 __resched_task(p, TIF_HRTICK_RESCHED);
971}
972
973static inline void resched_rq(struct rq *rq)
974{
975 unsigned long flags;
976
977 spin_lock_irqsave(&rq->lock, flags);
978 resched_task(rq->curr);
979 spin_unlock_irqrestore(&rq->lock, flags);
980}
981
982enum {
983 HRTICK_SET, /* re-programm hrtick_timer */
984 HRTICK_RESET, /* not a new slice */
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200985 HRTICK_BLOCK, /* stop hrtick operations */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100986};
987
988/*
989 * Use hrtick when:
990 * - enabled by features
991 * - hrtimer is actually high res
992 */
993static inline int hrtick_enabled(struct rq *rq)
994{
995 if (!sched_feat(HRTICK))
996 return 0;
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200997 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
998 return 0;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100999 return hrtimer_is_hres_active(&rq->hrtick_timer);
1000}
1001
1002/*
1003 * Called to set the hrtick timer state.
1004 *
1005 * called with rq->lock held and irqs disabled
1006 */
1007static void hrtick_start(struct rq *rq, u64 delay, int reset)
1008{
1009 assert_spin_locked(&rq->lock);
1010
1011 /*
1012 * preempt at: now + delay
1013 */
1014 rq->hrtick_expire =
1015 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1016 /*
1017 * indicate we need to program the timer
1018 */
1019 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1020 if (reset)
1021 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1022
1023 /*
1024 * New slices are called from the schedule path and don't need a
1025 * forced reschedule.
1026 */
1027 if (reset)
1028 resched_hrt(rq->curr);
1029}
1030
1031static void hrtick_clear(struct rq *rq)
1032{
1033 if (hrtimer_active(&rq->hrtick_timer))
1034 hrtimer_cancel(&rq->hrtick_timer);
1035}
1036
1037/*
1038 * Update the timer from the possible pending state.
1039 */
1040static void hrtick_set(struct rq *rq)
1041{
1042 ktime_t time;
1043 int set, reset;
1044 unsigned long flags;
1045
1046 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1047
1048 spin_lock_irqsave(&rq->lock, flags);
1049 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1050 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1051 time = rq->hrtick_expire;
1052 clear_thread_flag(TIF_HRTICK_RESCHED);
1053 spin_unlock_irqrestore(&rq->lock, flags);
1054
1055 if (set) {
1056 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1057 if (reset && !hrtimer_active(&rq->hrtick_timer))
1058 resched_rq(rq);
1059 } else
1060 hrtick_clear(rq);
1061}
1062
1063/*
1064 * High-resolution timer tick.
1065 * Runs from hardirq context with interrupts disabled.
1066 */
1067static enum hrtimer_restart hrtick(struct hrtimer *timer)
1068{
1069 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1070
1071 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1072
1073 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02001074 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001075 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1076 spin_unlock(&rq->lock);
1077
1078 return HRTIMER_NORESTART;
1079}
1080
Rabin Vincent81d41d72008-05-11 05:55:33 +05301081#ifdef CONFIG_SMP
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001082static void hotplug_hrtick_disable(int cpu)
1083{
1084 struct rq *rq = cpu_rq(cpu);
1085 unsigned long flags;
1086
1087 spin_lock_irqsave(&rq->lock, flags);
1088 rq->hrtick_flags = 0;
1089 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1090 spin_unlock_irqrestore(&rq->lock, flags);
1091
1092 hrtick_clear(rq);
1093}
1094
1095static void hotplug_hrtick_enable(int cpu)
1096{
1097 struct rq *rq = cpu_rq(cpu);
1098 unsigned long flags;
1099
1100 spin_lock_irqsave(&rq->lock, flags);
1101 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1102 spin_unlock_irqrestore(&rq->lock, flags);
1103}
1104
1105static int
1106hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1107{
1108 int cpu = (int)(long)hcpu;
1109
1110 switch (action) {
1111 case CPU_UP_CANCELED:
1112 case CPU_UP_CANCELED_FROZEN:
1113 case CPU_DOWN_PREPARE:
1114 case CPU_DOWN_PREPARE_FROZEN:
1115 case CPU_DEAD:
1116 case CPU_DEAD_FROZEN:
1117 hotplug_hrtick_disable(cpu);
1118 return NOTIFY_OK;
1119
1120 case CPU_UP_PREPARE:
1121 case CPU_UP_PREPARE_FROZEN:
1122 case CPU_DOWN_FAILED:
1123 case CPU_DOWN_FAILED_FROZEN:
1124 case CPU_ONLINE:
1125 case CPU_ONLINE_FROZEN:
1126 hotplug_hrtick_enable(cpu);
1127 return NOTIFY_OK;
1128 }
1129
1130 return NOTIFY_DONE;
1131}
1132
1133static void init_hrtick(void)
1134{
1135 hotcpu_notifier(hotplug_hrtick, 0);
1136}
Rabin Vincent81d41d72008-05-11 05:55:33 +05301137#endif /* CONFIG_SMP */
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001138
1139static void init_rq_hrtick(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001140{
1141 rq->hrtick_flags = 0;
1142 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1143 rq->hrtick_timer.function = hrtick;
1144 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1145}
1146
1147void hrtick_resched(void)
1148{
1149 struct rq *rq;
1150 unsigned long flags;
1151
1152 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1153 return;
1154
1155 local_irq_save(flags);
1156 rq = cpu_rq(smp_processor_id());
1157 hrtick_set(rq);
1158 local_irq_restore(flags);
1159}
1160#else
1161static inline void hrtick_clear(struct rq *rq)
1162{
1163}
1164
1165static inline void hrtick_set(struct rq *rq)
1166{
1167}
1168
1169static inline void init_rq_hrtick(struct rq *rq)
1170{
1171}
1172
1173void hrtick_resched(void)
1174{
1175}
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001176
1177static inline void init_hrtick(void)
1178{
1179}
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001180#endif
1181
Ingo Molnar1b9f19c2007-07-09 18:51:59 +02001182/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001183 * resched_task - mark a task 'to be rescheduled now'.
1184 *
1185 * On UP this means the setting of the need_resched flag, on SMP it
1186 * might also involve a cross-CPU call to trigger the scheduler on
1187 * the target CPU.
1188 */
1189#ifdef CONFIG_SMP
1190
1191#ifndef tsk_is_polling
1192#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1193#endif
1194
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001195static void __resched_task(struct task_struct *p, int tif_bit)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001196{
1197 int cpu;
1198
1199 assert_spin_locked(&task_rq(p)->lock);
1200
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001201 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001202 return;
1203
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001204 set_tsk_thread_flag(p, tif_bit);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001205
1206 cpu = task_cpu(p);
1207 if (cpu == smp_processor_id())
1208 return;
1209
1210 /* NEED_RESCHED must be visible before we test polling */
1211 smp_mb();
1212 if (!tsk_is_polling(p))
1213 smp_send_reschedule(cpu);
1214}
1215
1216static void resched_cpu(int cpu)
1217{
1218 struct rq *rq = cpu_rq(cpu);
1219 unsigned long flags;
1220
1221 if (!spin_trylock_irqsave(&rq->lock, flags))
1222 return;
1223 resched_task(cpu_curr(cpu));
1224 spin_unlock_irqrestore(&rq->lock, flags);
1225}
Thomas Gleixner06d83082008-03-22 09:20:24 +01001226
1227#ifdef CONFIG_NO_HZ
1228/*
1229 * When add_timer_on() enqueues a timer into the timer wheel of an
1230 * idle CPU then this timer might expire before the next timer event
1231 * which is scheduled to wake up that CPU. In case of a completely
1232 * idle system the next event might even be infinite time into the
1233 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1234 * leaves the inner idle loop so the newly added timer is taken into
1235 * account when the CPU goes back to idle and evaluates the timer
1236 * wheel for the next timer event.
1237 */
1238void wake_up_idle_cpu(int cpu)
1239{
1240 struct rq *rq = cpu_rq(cpu);
1241
1242 if (cpu == smp_processor_id())
1243 return;
1244
1245 /*
1246 * This is safe, as this function is called with the timer
1247 * wheel base lock of (cpu) held. When the CPU is on the way
1248 * to idle and has not yet set rq->curr to idle then it will
1249 * be serialized on the timer wheel base lock and take the new
1250 * timer into account automatically.
1251 */
1252 if (rq->curr != rq->idle)
1253 return;
1254
1255 /*
1256 * We can set TIF_RESCHED on the idle task of the other CPU
1257 * lockless. The worst case is that the other CPU runs the
1258 * idle task through an additional NOOP schedule()
1259 */
1260 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1261
1262 /* NEED_RESCHED must be visible before we test polling */
1263 smp_mb();
1264 if (!tsk_is_polling(rq->idle))
1265 smp_send_reschedule(cpu);
1266}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001267#endif /* CONFIG_NO_HZ */
Thomas Gleixner06d83082008-03-22 09:20:24 +01001268
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001269#else /* !CONFIG_SMP */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001270static void __resched_task(struct task_struct *p, int tif_bit)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001271{
1272 assert_spin_locked(&task_rq(p)->lock);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001273 set_tsk_thread_flag(p, tif_bit);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001274}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001275#endif /* CONFIG_SMP */
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001276
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001277#if BITS_PER_LONG == 32
1278# define WMULT_CONST (~0UL)
1279#else
1280# define WMULT_CONST (1UL << 32)
1281#endif
1282
1283#define WMULT_SHIFT 32
1284
Ingo Molnar194081e2007-08-09 11:16:51 +02001285/*
1286 * Shift right and round:
1287 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001288#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +02001289
Peter Zijlstraa7be37a2008-06-27 13:41:11 +02001290/*
1291 * delta *= weight / lw
1292 */
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +02001293static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001294calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1295 struct load_weight *lw)
1296{
1297 u64 tmp;
1298
Lai Jiangshan7a232e02008-06-12 16:43:07 +08001299 if (!lw->inv_weight) {
1300 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1301 lw->inv_weight = 1;
1302 else
1303 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1304 / (lw->weight+1);
1305 }
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001306
1307 tmp = (u64)delta_exec * weight;
1308 /*
1309 * Check whether we'd overflow the 64-bit multiplication:
1310 */
Ingo Molnar194081e2007-08-09 11:16:51 +02001311 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001312 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +02001313 WMULT_SHIFT/2);
1314 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001315 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001316
Ingo Molnarecf691d2007-08-02 17:41:40 +02001317 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001318}
1319
Ingo Molnar10919852007-10-15 17:00:04 +02001320static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001321{
1322 lw->weight += inc;
Ingo Molnare89996a2008-03-14 23:48:28 +01001323 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001324}
1325
Ingo Molnar10919852007-10-15 17:00:04 +02001326static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001327{
1328 lw->weight -= dec;
Ingo Molnare89996a2008-03-14 23:48:28 +01001329 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001330}
1331
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001333 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1334 * of tasks with abnormal "nice" values across CPUs the contribution that
1335 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001336 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -07001337 * scaled version of the new time slice allocation that they receive on time
1338 * slice expiry etc.
1339 */
1340
Ingo Molnardd41f592007-07-09 18:51:59 +02001341#define WEIGHT_IDLEPRIO 2
1342#define WMULT_IDLEPRIO (1 << 31)
1343
1344/*
1345 * Nice levels are multiplicative, with a gentle 10% change for every
1346 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1347 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1348 * that remained on nice 0.
1349 *
1350 * The "10% effect" is relative and cumulative: from _any_ nice level,
1351 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +02001352 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1353 * If a task goes up by ~10% and another task goes down by ~10% then
1354 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +02001355 */
1356static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001357 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1358 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1359 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1360 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1361 /* 0 */ 1024, 820, 655, 526, 423,
1362 /* 5 */ 335, 272, 215, 172, 137,
1363 /* 10 */ 110, 87, 70, 56, 45,
1364 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +02001365};
1366
Ingo Molnar5714d2d2007-07-16 09:46:31 +02001367/*
1368 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1369 *
1370 * In cases where the weight does not change often, we can use the
1371 * precalculated inverse to speed up arithmetics by turning divisions
1372 * into multiplications:
1373 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001374static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001375 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1376 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1377 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1378 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1379 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1380 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1381 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1382 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +02001383};
Peter Williams2dd73a42006-06-27 02:54:34 -07001384
Ingo Molnardd41f592007-07-09 18:51:59 +02001385static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1386
1387/*
1388 * runqueue iterator, to support SMP load-balancing between different
1389 * scheduling classes, without having to expose their internal data
1390 * structures to the load-balancing proper:
1391 */
1392struct rq_iterator {
1393 void *arg;
1394 struct task_struct *(*start)(void *);
1395 struct task_struct *(*next)(void *);
1396};
1397
Peter Williamse1d14842007-10-24 18:23:51 +02001398#ifdef CONFIG_SMP
1399static unsigned long
1400balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1401 unsigned long max_load_move, struct sched_domain *sd,
1402 enum cpu_idle_type idle, int *all_pinned,
1403 int *this_best_prio, struct rq_iterator *iterator);
1404
1405static int
1406iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1407 struct sched_domain *sd, enum cpu_idle_type idle,
1408 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +02001409#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001410
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001411#ifdef CONFIG_CGROUP_CPUACCT
1412static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1413#else
1414static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1415#endif
1416
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001417static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1418{
1419 update_load_add(&rq->load, load);
1420}
1421
1422static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1423{
1424 update_load_sub(&rq->load, load);
1425}
1426
Gregory Haskinse7693a32008-01-25 21:08:09 +01001427#ifdef CONFIG_SMP
1428static unsigned long source_load(int cpu, int type);
1429static unsigned long target_load(int cpu, int type);
1430static unsigned long cpu_avg_load_per_task(int cpu);
1431static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001432
1433#ifdef CONFIG_FAIR_GROUP_SCHED
1434
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001435typedef void (*tg_visitor)(struct task_group *, int, struct sched_domain *);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001436
1437/*
1438 * Iterate the full tree, calling @down when first entering a node and @up when
1439 * leaving it for the final time.
1440 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001441static void
1442walk_tg_tree(tg_visitor down, tg_visitor up, int cpu, struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001443{
1444 struct task_group *parent, *child;
1445
1446 rcu_read_lock();
1447 parent = &root_task_group;
1448down:
Peter Zijlstrab6a86c72008-06-27 13:41:18 +02001449 (*down)(parent, cpu, sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001450 list_for_each_entry_rcu(child, &parent->children, siblings) {
1451 parent = child;
1452 goto down;
1453
1454up:
1455 continue;
1456 }
Peter Zijlstrab6a86c72008-06-27 13:41:18 +02001457 (*up)(parent, cpu, sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001458
1459 child = parent;
1460 parent = parent->parent;
1461 if (parent)
1462 goto up;
1463 rcu_read_unlock();
1464}
1465
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001466static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1467
1468/*
1469 * Calculate and set the cpu's group shares.
1470 */
1471static void
Peter Zijlstrab6a86c72008-06-27 13:41:18 +02001472__update_group_shares_cpu(struct task_group *tg, int cpu,
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001473 unsigned long sd_shares, unsigned long sd_rq_weight)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001474{
1475 int boost = 0;
1476 unsigned long shares;
1477 unsigned long rq_weight;
1478
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001479 if (!tg->se[cpu])
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001480 return;
1481
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001482 rq_weight = tg->cfs_rq[cpu]->load.weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001483
1484 /*
1485 * If there are currently no tasks on the cpu pretend there is one of
1486 * average load so that when a new task gets to run here it will not
1487 * get delayed by group starvation.
1488 */
1489 if (!rq_weight) {
1490 boost = 1;
1491 rq_weight = NICE_0_LOAD;
1492 }
1493
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001494 if (unlikely(rq_weight > sd_rq_weight))
1495 rq_weight = sd_rq_weight;
1496
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001497 /*
1498 * \Sum shares * rq_weight
1499 * shares = -----------------------
1500 * \Sum rq_weight
1501 *
1502 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001503 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001504
1505 /*
1506 * record the actual number of shares, not the boosted amount.
1507 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001508 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001509
1510 if (shares < MIN_SHARES)
1511 shares = MIN_SHARES;
1512 else if (shares > MAX_SHARES)
1513 shares = MAX_SHARES;
1514
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001515 __set_se_shares(tg->se[cpu], shares);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001516}
1517
1518/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001519 * Re-compute the task group their per cpu shares over the given domain.
1520 * This needs to be done in a bottom-up fashion because the rq weight of a
1521 * parent group depends on the shares of its child groups.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001522 */
1523static void
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001524tg_shares_up(struct task_group *tg, int cpu, struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001525{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001526 unsigned long rq_weight = 0;
1527 unsigned long shares = 0;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001528 int i;
1529
1530 for_each_cpu_mask(i, sd->span) {
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001531 rq_weight += tg->cfs_rq[i]->load.weight;
1532 shares += tg->cfs_rq[i]->shares;
1533 }
1534
1535 if ((!shares && rq_weight) || shares > tg->shares)
1536 shares = tg->shares;
1537
1538 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1539 shares = tg->shares;
1540
1541 for_each_cpu_mask(i, sd->span) {
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001542 struct rq *rq = cpu_rq(i);
1543 unsigned long flags;
1544
1545 spin_lock_irqsave(&rq->lock, flags);
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001546 __update_group_shares_cpu(tg, i, shares, rq_weight);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001547 spin_unlock_irqrestore(&rq->lock, flags);
1548 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001549}
1550
1551/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001552 * Compute the cpu's hierarchical load factor for each task group.
1553 * This needs to be done in a top-down fashion because the load of a child
1554 * group is a fraction of its parents load.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001555 */
Peter Zijlstrab6a86c72008-06-27 13:41:18 +02001556static void
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001557tg_load_down(struct task_group *tg, int cpu, struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001558{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001559 unsigned long load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001560
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001561 if (!tg->parent) {
1562 load = cpu_rq(cpu)->load.weight;
1563 } else {
1564 load = tg->parent->cfs_rq[cpu]->h_load;
1565 load *= tg->cfs_rq[cpu]->shares;
1566 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1567 }
1568
1569 tg->cfs_rq[cpu]->h_load = load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001570}
1571
Peter Zijlstra4d8d5952008-06-27 13:41:19 +02001572static void
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001573tg_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
Peter Zijlstra4d8d5952008-06-27 13:41:19 +02001574{
1575}
1576
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001577static void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001578{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001579 walk_tg_tree(tg_nop, tg_shares_up, 0, sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001580}
1581
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001582static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1583{
1584 spin_unlock(&rq->lock);
1585 update_shares(sd);
1586 spin_lock(&rq->lock);
1587}
1588
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001589static void update_h_load(int cpu)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001590{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001591 walk_tg_tree(tg_load_down, tg_nop, cpu, NULL);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001592}
1593
1594static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1595{
1596 cfs_rq->shares = shares;
1597}
1598
1599#else
1600
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001601static inline void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001602{
1603}
1604
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001605static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1606{
1607}
1608
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001609#endif
1610
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001611#endif
1612
Ingo Molnardd41f592007-07-09 18:51:59 +02001613#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +02001614#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001615#include "sched_fair.c"
1616#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +02001617#ifdef CONFIG_SCHED_DEBUG
1618# include "sched_debug.c"
1619#endif
1620
1621#define sched_class_highest (&rt_sched_class)
Gregory Haskins1f11eb62008-06-04 15:04:05 -04001622#define for_each_class(class) \
1623 for (class = sched_class_highest; class; class = class->next)
Ingo Molnardd41f592007-07-09 18:51:59 +02001624
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001625static void inc_nr_running(struct rq *rq)
Ingo Molnar6363ca52008-05-29 11:28:57 +02001626{
1627 rq->nr_running++;
Ingo Molnar6363ca52008-05-29 11:28:57 +02001628}
1629
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001630static void dec_nr_running(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001631{
1632 rq->nr_running--;
Ingo Molnar9c217242007-08-02 17:41:40 +02001633}
1634
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001635static void set_load_weight(struct task_struct *p)
1636{
1637 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001638 p->se.load.weight = prio_to_weight[0] * 2;
1639 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1640 return;
1641 }
1642
1643 /*
1644 * SCHED_IDLE tasks get minimal weight:
1645 */
1646 if (p->policy == SCHED_IDLE) {
1647 p->se.load.weight = WEIGHT_IDLEPRIO;
1648 p->se.load.inv_weight = WMULT_IDLEPRIO;
1649 return;
1650 }
1651
1652 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1653 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001654}
1655
Ingo Molnar8159f872007-08-09 11:16:49 +02001656static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001657{
1658 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001659 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001660 p->se.on_rq = 1;
1661}
1662
Ingo Molnar69be72c2007-08-09 11:16:49 +02001663static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001664{
Ingo Molnarf02231e2007-08-09 11:16:48 +02001665 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001666 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001667}
1668
1669/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001670 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001671 */
Ingo Molnar14531182007-07-09 18:51:59 +02001672static inline int __normal_prio(struct task_struct *p)
1673{
Ingo Molnardd41f592007-07-09 18:51:59 +02001674 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001675}
1676
1677/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001678 * Calculate the expected normal priority: i.e. priority
1679 * without taking RT-inheritance into account. Might be
1680 * boosted by interactivity modifiers. Changes upon fork,
1681 * setprio syscalls, and whenever the interactivity
1682 * estimator recalculates.
1683 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001684static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001685{
1686 int prio;
1687
Ingo Molnare05606d2007-07-09 18:51:59 +02001688 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001689 prio = MAX_RT_PRIO-1 - p->rt_priority;
1690 else
1691 prio = __normal_prio(p);
1692 return prio;
1693}
1694
1695/*
1696 * Calculate the current priority, i.e. the priority
1697 * taken into account by the scheduler. This value might
1698 * be boosted by RT tasks, or might be boosted by
1699 * interactivity modifiers. Will be RT if the task got
1700 * RT-boosted. If not then it returns p->normal_prio.
1701 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001702static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001703{
1704 p->normal_prio = normal_prio(p);
1705 /*
1706 * If we are RT tasks or we were boosted to RT priority,
1707 * keep the priority unchanged. Otherwise, update priority
1708 * to the normal priority:
1709 */
1710 if (!rt_prio(p->prio))
1711 return p->normal_prio;
1712 return p->prio;
1713}
1714
1715/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001716 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001718static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001720 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001721 rq->nr_uninterruptible--;
1722
Ingo Molnar8159f872007-08-09 11:16:49 +02001723 enqueue_task(rq, p, wakeup);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001724 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725}
1726
1727/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 * deactivate_task - remove a task from the runqueue.
1729 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001730static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001732 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001733 rq->nr_uninterruptible++;
1734
Ingo Molnar69be72c2007-08-09 11:16:49 +02001735 dequeue_task(rq, p, sleep);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001736 dec_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737}
1738
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739/**
1740 * task_curr - is this task currently executing on a CPU?
1741 * @p: the task in question.
1742 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001743inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744{
1745 return cpu_curr(task_cpu(p)) == p;
1746}
1747
Ingo Molnardd41f592007-07-09 18:51:59 +02001748static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1749{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001750 set_task_rq(p, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001751#ifdef CONFIG_SMP
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001752 /*
1753 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1754 * successfuly executed on another CPU. We must ensure that updates of
1755 * per-task data have been completed by this moment.
1756 */
1757 smp_wmb();
Ingo Molnardd41f592007-07-09 18:51:59 +02001758 task_thread_info(p)->cpu = cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02001759#endif
Peter Williams2dd73a42006-06-27 02:54:34 -07001760}
1761
Steven Rostedtcb469842008-01-25 21:08:22 +01001762static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1763 const struct sched_class *prev_class,
1764 int oldprio, int running)
1765{
1766 if (prev_class != p->sched_class) {
1767 if (prev_class->switched_from)
1768 prev_class->switched_from(rq, p, running);
1769 p->sched_class->switched_to(rq, p, running);
1770 } else
1771 p->sched_class->prio_changed(rq, p, oldprio, running);
1772}
1773
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774#ifdef CONFIG_SMP
Ingo Molnarc65cc872007-07-09 18:51:58 +02001775
Thomas Gleixnere958b362008-06-04 23:22:32 +02001776/* Used instead of source_load when we know the type == 0 */
1777static unsigned long weighted_cpuload(const int cpu)
1778{
1779 return cpu_rq(cpu)->load.weight;
1780}
1781
Ingo Molnarcc367732007-10-15 17:00:18 +02001782/*
1783 * Is this task likely cache-hot:
1784 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01001785static int
Ingo Molnarcc367732007-10-15 17:00:18 +02001786task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1787{
1788 s64 delta;
1789
Ingo Molnarf540a602008-03-15 17:10:34 +01001790 /*
1791 * Buddy candidates are cache hot:
1792 */
Ingo Molnard25ce4c2008-03-17 09:36:53 +01001793 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
Ingo Molnarf540a602008-03-15 17:10:34 +01001794 return 1;
1795
Ingo Molnarcc367732007-10-15 17:00:18 +02001796 if (p->sched_class != &fair_sched_class)
1797 return 0;
1798
Ingo Molnar6bc16652007-10-15 17:00:18 +02001799 if (sysctl_sched_migration_cost == -1)
1800 return 1;
1801 if (sysctl_sched_migration_cost == 0)
1802 return 0;
1803
Ingo Molnarcc367732007-10-15 17:00:18 +02001804 delta = now - p->se.exec_start;
1805
1806 return delta < (s64)sysctl_sched_migration_cost;
1807}
1808
1809
Ingo Molnardd41f592007-07-09 18:51:59 +02001810void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02001811{
Ingo Molnardd41f592007-07-09 18:51:59 +02001812 int old_cpu = task_cpu(p);
1813 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001814 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1815 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnarbbdba7c2007-10-15 17:00:06 +02001816 u64 clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001817
1818 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001819
1820#ifdef CONFIG_SCHEDSTATS
1821 if (p->se.wait_start)
1822 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001823 if (p->se.sleep_start)
1824 p->se.sleep_start -= clock_offset;
1825 if (p->se.block_start)
1826 p->se.block_start -= clock_offset;
Ingo Molnarcc367732007-10-15 17:00:18 +02001827 if (old_cpu != new_cpu) {
1828 schedstat_inc(p, se.nr_migrations);
1829 if (task_hot(p, old_rq->clock, NULL))
1830 schedstat_inc(p, se.nr_forced2_migrations);
1831 }
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001832#endif
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001833 p->se.vruntime -= old_cfsrq->min_vruntime -
1834 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02001835
1836 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02001837}
1838
Ingo Molnar70b97a72006-07-03 00:25:42 -07001839struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841
Ingo Molnar36c8b582006-07-03 00:25:41 -07001842 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 int dest_cpu;
1844
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001846};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847
1848/*
1849 * The task's runqueue lock must be held.
1850 * Returns true if you have to wait for migration thread.
1851 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001852static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07001853migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001855 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856
1857 /*
1858 * If the task is not on a runqueue (and not running), then
1859 * it is sufficient to simply update the task's cpu field.
1860 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001861 if (!p->se.on_rq && !task_running(rq, p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 set_task_cpu(p, dest_cpu);
1863 return 0;
1864 }
1865
1866 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 req->task = p;
1868 req->dest_cpu = dest_cpu;
1869 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07001870
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 return 1;
1872}
1873
1874/*
1875 * wait_task_inactive - wait for a thread to unschedule.
1876 *
1877 * The caller must ensure that the task *will* unschedule sometime soon,
1878 * else this function might spin for a *long* time. This function can't
1879 * be called with interrupts off, or it may introduce deadlock with
1880 * smp_call_function() if an IPI is sent by the same process we are
1881 * waiting to become inactive.
1882 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001883void wait_task_inactive(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884{
1885 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02001886 int running, on_rq;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001887 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888
Andi Kleen3a5c3592007-10-15 17:00:14 +02001889 for (;;) {
1890 /*
1891 * We do the initial early heuristics without holding
1892 * any task-queue locks at all. We'll only try to get
1893 * the runqueue lock when things look like they will
1894 * work out!
1895 */
1896 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001897
Andi Kleen3a5c3592007-10-15 17:00:14 +02001898 /*
1899 * If the task is actively running on another CPU
1900 * still, just relax and busy-wait without holding
1901 * any locks.
1902 *
1903 * NOTE! Since we don't hold any locks, it's not
1904 * even sure that "rq" stays as the right runqueue!
1905 * But we don't care, since "task_running()" will
1906 * return false if the runqueue has changed and p
1907 * is actually now running somewhere else!
1908 */
1909 while (task_running(rq, p))
1910 cpu_relax();
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001911
Andi Kleen3a5c3592007-10-15 17:00:14 +02001912 /*
1913 * Ok, time to look more closely! We need the rq
1914 * lock now, to be *sure*. If we're wrong, we'll
1915 * just go back and repeat.
1916 */
1917 rq = task_rq_lock(p, &flags);
1918 running = task_running(rq, p);
1919 on_rq = p->se.on_rq;
1920 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001921
Andi Kleen3a5c3592007-10-15 17:00:14 +02001922 /*
1923 * Was it really running after all now that we
1924 * checked with the proper locks actually held?
1925 *
1926 * Oops. Go back and try again..
1927 */
1928 if (unlikely(running)) {
1929 cpu_relax();
1930 continue;
1931 }
1932
1933 /*
1934 * It's not enough that it's not actively running,
1935 * it must be off the runqueue _entirely_, and not
1936 * preempted!
1937 *
1938 * So if it wa still runnable (but just not actively
1939 * running right now), it's preempted, and we should
1940 * yield - it could be a while.
1941 */
1942 if (unlikely(on_rq)) {
1943 schedule_timeout_uninterruptible(1);
1944 continue;
1945 }
1946
1947 /*
1948 * Ahh, all good. It wasn't running, and it wasn't
1949 * runnable, which means that it will never become
1950 * running in the future either. We're all done!
1951 */
1952 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954}
1955
1956/***
1957 * kick_process - kick a running thread to enter/exit the kernel
1958 * @p: the to-be-kicked thread
1959 *
1960 * Cause a process which is running on another CPU to enter
1961 * kernel-mode, without any delay. (to get signals handled.)
1962 *
1963 * NOTE: this function doesnt have to take the runqueue lock,
1964 * because all it wants to ensure is that the remote task enters
1965 * the kernel. If the IPI races and the task has been migrated
1966 * to another CPU then no harm is done and the purpose has been
1967 * achieved as well.
1968 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001969void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970{
1971 int cpu;
1972
1973 preempt_disable();
1974 cpu = task_cpu(p);
1975 if ((cpu != smp_processor_id()) && task_curr(p))
1976 smp_send_reschedule(cpu);
1977 preempt_enable();
1978}
1979
1980/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001981 * Return a low guess at the load of a migration-source cpu weighted
1982 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 *
1984 * We want to under-estimate the load of migration sources, to
1985 * balance conservatively.
1986 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001987static unsigned long source_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08001988{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001989 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001990 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001991
Peter Williams2dd73a42006-06-27 02:54:34 -07001992 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02001993 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07001994
Ingo Molnardd41f592007-07-09 18:51:59 +02001995 return min(rq->cpu_load[type-1], total);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996}
1997
1998/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001999 * Return a high guess at the load of a migration-target cpu weighted
2000 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002002static unsigned long target_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08002003{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002004 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002005 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08002006
Peter Williams2dd73a42006-06-27 02:54:34 -07002007 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02002008 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07002009
Ingo Molnardd41f592007-07-09 18:51:59 +02002010 return max(rq->cpu_load[type-1], total);
Peter Williams2dd73a42006-06-27 02:54:34 -07002011}
2012
2013/*
2014 * Return the average load per task on the cpu's run queue
2015 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002016static unsigned long cpu_avg_load_per_task(int cpu)
Peter Williams2dd73a42006-06-27 02:54:34 -07002017{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002018 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002019 unsigned long total = weighted_cpuload(cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07002020 unsigned long n = rq->nr_running;
2021
Ingo Molnardd41f592007-07-09 18:51:59 +02002022 return n ? total / n : SCHED_LOAD_SCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023}
2024
Nick Piggin147cbb42005-06-25 14:57:19 -07002025/*
2026 * find_idlest_group finds and returns the least busy CPU group within the
2027 * domain.
2028 */
2029static struct sched_group *
2030find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2031{
2032 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2033 unsigned long min_load = ULONG_MAX, this_load = 0;
2034 int load_idx = sd->forkexec_idx;
2035 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2036
2037 do {
2038 unsigned long load, avg_load;
2039 int local_group;
2040 int i;
2041
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002042 /* Skip over this group if it has no CPUs allowed */
2043 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
Andi Kleen3a5c3592007-10-15 17:00:14 +02002044 continue;
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002045
Nick Piggin147cbb42005-06-25 14:57:19 -07002046 local_group = cpu_isset(this_cpu, group->cpumask);
Nick Piggin147cbb42005-06-25 14:57:19 -07002047
2048 /* Tally up the load of all CPUs in the group */
2049 avg_load = 0;
2050
2051 for_each_cpu_mask(i, group->cpumask) {
2052 /* Bias balancing toward cpus of our domain */
2053 if (local_group)
2054 load = source_load(i, load_idx);
2055 else
2056 load = target_load(i, load_idx);
2057
2058 avg_load += load;
2059 }
2060
2061 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07002062 avg_load = sg_div_cpu_power(group,
2063 avg_load * SCHED_LOAD_SCALE);
Nick Piggin147cbb42005-06-25 14:57:19 -07002064
2065 if (local_group) {
2066 this_load = avg_load;
2067 this = group;
2068 } else if (avg_load < min_load) {
2069 min_load = avg_load;
2070 idlest = group;
2071 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02002072 } while (group = group->next, group != sd->groups);
Nick Piggin147cbb42005-06-25 14:57:19 -07002073
2074 if (!idlest || 100*this_load < imbalance*min_load)
2075 return NULL;
2076 return idlest;
2077}
2078
2079/*
Satoru Takeuchi0feaece2006-10-03 01:14:10 -07002080 * find_idlest_cpu - find the idlest cpu among the cpus in group.
Nick Piggin147cbb42005-06-25 14:57:19 -07002081 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002082static int
Mike Travis7c16ec52008-04-04 18:11:11 -07002083find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2084 cpumask_t *tmp)
Nick Piggin147cbb42005-06-25 14:57:19 -07002085{
2086 unsigned long load, min_load = ULONG_MAX;
2087 int idlest = -1;
2088 int i;
2089
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002090 /* Traverse only the allowed CPUs */
Mike Travis7c16ec52008-04-04 18:11:11 -07002091 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002092
Mike Travis7c16ec52008-04-04 18:11:11 -07002093 for_each_cpu_mask(i, *tmp) {
Peter Williams2dd73a42006-06-27 02:54:34 -07002094 load = weighted_cpuload(i);
Nick Piggin147cbb42005-06-25 14:57:19 -07002095
2096 if (load < min_load || (load == min_load && i == this_cpu)) {
2097 min_load = load;
2098 idlest = i;
2099 }
2100 }
2101
2102 return idlest;
2103}
2104
Nick Piggin476d1392005-06-25 14:57:29 -07002105/*
2106 * sched_balance_self: balance the current task (running on cpu) in domains
2107 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2108 * SD_BALANCE_EXEC.
2109 *
2110 * Balance, ie. select the least loaded group.
2111 *
2112 * Returns the target CPU number, or the same CPU if no balancing is needed.
2113 *
2114 * preempt must be disabled.
2115 */
2116static int sched_balance_self(int cpu, int flag)
2117{
2118 struct task_struct *t = current;
2119 struct sched_domain *tmp, *sd = NULL;
Nick Piggin147cbb42005-06-25 14:57:19 -07002120
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002121 for_each_domain(cpu, tmp) {
Ingo Molnar9761eea2007-07-09 18:52:00 +02002122 /*
2123 * If power savings logic is enabled for a domain, stop there.
2124 */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002125 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2126 break;
Nick Piggin476d1392005-06-25 14:57:29 -07002127 if (tmp->flags & flag)
2128 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002129 }
Nick Piggin476d1392005-06-25 14:57:29 -07002130
2131 while (sd) {
Mike Travis7c16ec52008-04-04 18:11:11 -07002132 cpumask_t span, tmpmask;
Nick Piggin476d1392005-06-25 14:57:29 -07002133 struct sched_group *group;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002134 int new_cpu, weight;
2135
2136 if (!(sd->flags & flag)) {
2137 sd = sd->child;
2138 continue;
2139 }
Nick Piggin476d1392005-06-25 14:57:29 -07002140
2141 span = sd->span;
2142 group = find_idlest_group(sd, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002143 if (!group) {
2144 sd = sd->child;
2145 continue;
2146 }
Nick Piggin476d1392005-06-25 14:57:29 -07002147
Mike Travis7c16ec52008-04-04 18:11:11 -07002148 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002149 if (new_cpu == -1 || new_cpu == cpu) {
2150 /* Now try balancing at a lower domain level of cpu */
2151 sd = sd->child;
2152 continue;
2153 }
Nick Piggin476d1392005-06-25 14:57:29 -07002154
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002155 /* Now try balancing at a lower domain level of new_cpu */
Nick Piggin476d1392005-06-25 14:57:29 -07002156 cpu = new_cpu;
Nick Piggin476d1392005-06-25 14:57:29 -07002157 sd = NULL;
2158 weight = cpus_weight(span);
2159 for_each_domain(cpu, tmp) {
2160 if (weight <= cpus_weight(tmp->span))
2161 break;
2162 if (tmp->flags & flag)
2163 sd = tmp;
2164 }
2165 /* while loop will break here if sd == NULL */
2166 }
2167
2168 return cpu;
2169}
2170
2171#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173/***
2174 * try_to_wake_up - wake up a thread
2175 * @p: the to-be-woken-up thread
2176 * @state: the mask of task states that can be woken
2177 * @sync: do a synchronous wakeup?
2178 *
2179 * Put it on the run-queue if it's not already there. The "current"
2180 * thread is always on the run-queue (except when the actual
2181 * re-schedule is in progress), and as such you're allowed to do
2182 * the simpler "current->state = TASK_RUNNING" to mark yourself
2183 * runnable without the overhead of this.
2184 *
2185 * returns failure only if the task is already active.
2186 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002187static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188{
Ingo Molnarcc367732007-10-15 17:00:18 +02002189 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 unsigned long flags;
2191 long old_state;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002192 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193
Ingo Molnarb85d0662008-03-16 20:03:22 +01002194 if (!sched_feat(SYNC_WAKEUPS))
2195 sync = 0;
2196
Linus Torvalds04e2f172008-02-23 18:05:03 -08002197 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 rq = task_rq_lock(p, &flags);
2199 old_state = p->state;
2200 if (!(old_state & state))
2201 goto out;
2202
Ingo Molnardd41f592007-07-09 18:51:59 +02002203 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 goto out_running;
2205
2206 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02002207 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 this_cpu = smp_processor_id();
2209
2210#ifdef CONFIG_SMP
2211 if (unlikely(task_running(rq, p)))
2212 goto out_activate;
2213
Dmitry Adamushko5d2f5a62008-01-25 21:08:21 +01002214 cpu = p->sched_class->select_task_rq(p, sync);
2215 if (cpu != orig_cpu) {
2216 set_task_cpu(p, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 task_rq_unlock(rq, &flags);
2218 /* might preempt at this point */
2219 rq = task_rq_lock(p, &flags);
2220 old_state = p->state;
2221 if (!(old_state & state))
2222 goto out;
Ingo Molnardd41f592007-07-09 18:51:59 +02002223 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 goto out_running;
2225
2226 this_cpu = smp_processor_id();
2227 cpu = task_cpu(p);
2228 }
2229
Gregory Haskinse7693a32008-01-25 21:08:09 +01002230#ifdef CONFIG_SCHEDSTATS
2231 schedstat_inc(rq, ttwu_count);
2232 if (cpu == this_cpu)
2233 schedstat_inc(rq, ttwu_local);
2234 else {
2235 struct sched_domain *sd;
2236 for_each_domain(this_cpu, sd) {
2237 if (cpu_isset(cpu, sd->span)) {
2238 schedstat_inc(sd, ttwu_wake_remote);
2239 break;
2240 }
2241 }
2242 }
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002243#endif /* CONFIG_SCHEDSTATS */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002244
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245out_activate:
2246#endif /* CONFIG_SMP */
Ingo Molnarcc367732007-10-15 17:00:18 +02002247 schedstat_inc(p, se.nr_wakeups);
2248 if (sync)
2249 schedstat_inc(p, se.nr_wakeups_sync);
2250 if (orig_cpu != cpu)
2251 schedstat_inc(p, se.nr_wakeups_migrate);
2252 if (cpu == this_cpu)
2253 schedstat_inc(p, se.nr_wakeups_local);
2254 else
2255 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnar2daa3572007-08-09 11:16:51 +02002256 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02002257 activate_task(rq, p, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 success = 1;
2259
2260out_running:
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002261 check_preempt_curr(rq, p);
2262
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 p->state = TASK_RUNNING;
Steven Rostedt9a897c52008-01-25 21:08:22 +01002264#ifdef CONFIG_SMP
2265 if (p->sched_class->task_wake_up)
2266 p->sched_class->task_wake_up(rq, p);
2267#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268out:
2269 task_rq_unlock(rq, &flags);
2270
2271 return success;
2272}
2273
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002274int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05002276 return try_to_wake_up(p, TASK_ALL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278EXPORT_SYMBOL(wake_up_process);
2279
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002280int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281{
2282 return try_to_wake_up(p, state, 0);
2283}
2284
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285/*
2286 * Perform scheduler related setup for a newly forked process p.
2287 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02002288 *
2289 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002291static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292{
Ingo Molnardd41f592007-07-09 18:51:59 +02002293 p->se.exec_start = 0;
2294 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02002295 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002296 p->se.last_wakeup = 0;
2297 p->se.avg_overlap = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002298
2299#ifdef CONFIG_SCHEDSTATS
2300 p->se.wait_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002301 p->se.sum_sleep_runtime = 0;
2302 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002303 p->se.block_start = 0;
2304 p->se.sleep_max = 0;
2305 p->se.block_max = 0;
2306 p->se.exec_max = 0;
Ingo Molnareba1ed42007-10-15 17:00:02 +02002307 p->se.slice_max = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002308 p->se.wait_max = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002309#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002310
Peter Zijlstrafa717062008-01-25 21:08:27 +01002311 INIT_LIST_HEAD(&p->rt.run_list);
Ingo Molnardd41f592007-07-09 18:51:59 +02002312 p->se.on_rq = 0;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02002313 INIT_LIST_HEAD(&p->se.group_node);
Nick Piggin476d1392005-06-25 14:57:29 -07002314
Avi Kivitye107be32007-07-26 13:40:43 +02002315#ifdef CONFIG_PREEMPT_NOTIFIERS
2316 INIT_HLIST_HEAD(&p->preempt_notifiers);
2317#endif
2318
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 /*
2320 * We mark the process as running here, but have not actually
2321 * inserted it onto the runqueue yet. This guarantees that
2322 * nobody will actually run it, and a signal or other external
2323 * event cannot wake it up and insert it on the runqueue either.
2324 */
2325 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02002326}
2327
2328/*
2329 * fork()/clone()-time setup:
2330 */
2331void sched_fork(struct task_struct *p, int clone_flags)
2332{
2333 int cpu = get_cpu();
2334
2335 __sched_fork(p);
2336
2337#ifdef CONFIG_SMP
2338 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2339#endif
Ingo Molnar02e4bac2007-10-15 17:00:11 +02002340 set_task_cpu(p, cpu);
Ingo Molnarb29739f2006-06-27 02:54:51 -07002341
2342 /*
2343 * Make sure we do not leak PI boosting priority to the child:
2344 */
2345 p->prio = current->normal_prio;
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02002346 if (!rt_prio(p->prio))
2347 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07002348
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002349#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02002350 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002351 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08002353#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07002354 p->oncpu = 0;
2355#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07002357 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08002358 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002360 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361}
2362
2363/*
2364 * wake_up_new_task - wake up a newly created task for the first time.
2365 *
2366 * This function will do some initial scheduler statistics housekeeping
2367 * that must be done for every newly created context, then puts the task
2368 * on the runqueue and wakes it.
2369 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002370void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371{
2372 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002373 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374
2375 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02002377 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378
2379 p->prio = effective_prio(p);
2380
Srivatsa Vaddagirib9dca1e2007-10-17 16:55:11 +02002381 if (!p->sched_class->task_new || !current->se.on_rq) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002382 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02002385 * Let the scheduling class do new task startup
2386 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02002388 p->sched_class->task_new(rq, p);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02002389 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 }
Ingo Molnardd41f592007-07-09 18:51:59 +02002391 check_preempt_curr(rq, p);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002392#ifdef CONFIG_SMP
2393 if (p->sched_class->task_wake_up)
2394 p->sched_class->task_wake_up(rq, p);
2395#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02002396 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397}
2398
Avi Kivitye107be32007-07-26 13:40:43 +02002399#ifdef CONFIG_PREEMPT_NOTIFIERS
2400
2401/**
Randy Dunlap421cee22007-07-31 00:37:50 -07002402 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2403 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002404 */
2405void preempt_notifier_register(struct preempt_notifier *notifier)
2406{
2407 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2408}
2409EXPORT_SYMBOL_GPL(preempt_notifier_register);
2410
2411/**
2412 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002413 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002414 *
2415 * This is safe to call from within a preemption notifier.
2416 */
2417void preempt_notifier_unregister(struct preempt_notifier *notifier)
2418{
2419 hlist_del(&notifier->link);
2420}
2421EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2422
2423static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2424{
2425 struct preempt_notifier *notifier;
2426 struct hlist_node *node;
2427
2428 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2429 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2430}
2431
2432static void
2433fire_sched_out_preempt_notifiers(struct task_struct *curr,
2434 struct task_struct *next)
2435{
2436 struct preempt_notifier *notifier;
2437 struct hlist_node *node;
2438
2439 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2440 notifier->ops->sched_out(notifier, next);
2441}
2442
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002443#else /* !CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002444
2445static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2446{
2447}
2448
2449static void
2450fire_sched_out_preempt_notifiers(struct task_struct *curr,
2451 struct task_struct *next)
2452{
2453}
2454
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002455#endif /* CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002456
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002458 * prepare_task_switch - prepare to switch tasks
2459 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002460 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002461 * @next: the task we are going to switch to.
2462 *
2463 * This is called with the rq lock held and interrupts off. It must
2464 * be paired with a subsequent finish_task_switch after the context
2465 * switch.
2466 *
2467 * prepare_task_switch sets up locking and calls architecture specific
2468 * hooks.
2469 */
Avi Kivitye107be32007-07-26 13:40:43 +02002470static inline void
2471prepare_task_switch(struct rq *rq, struct task_struct *prev,
2472 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002473{
Avi Kivitye107be32007-07-26 13:40:43 +02002474 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002475 prepare_lock_switch(rq, next);
2476 prepare_arch_switch(next);
2477}
2478
2479/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04002481 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 * @prev: the thread we just switched away from.
2483 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002484 * finish_task_switch must be called after the context switch, paired
2485 * with a prepare_task_switch call before the context switch.
2486 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2487 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 *
2489 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002490 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 * with the lock held can cause deadlocks; see schedule() for
2492 * details.)
2493 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002494static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495 __releases(rq->lock)
2496{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002498 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499
2500 rq->prev_mm = NULL;
2501
2502 /*
2503 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002504 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002505 * schedule one last time. The schedule call will never return, and
2506 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002507 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 * still held, otherwise prev could be scheduled on another cpu, die
2509 * there before we look at prev->state, and then the reference would
2510 * be dropped twice.
2511 * Manfred Spraul <manfred@colorfullife.com>
2512 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002513 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07002514 finish_arch_switch(prev);
2515 finish_lock_switch(rq, prev);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002516#ifdef CONFIG_SMP
2517 if (current->sched_class->post_schedule)
2518 current->sched_class->post_schedule(rq);
2519#endif
Steven Rostedte8fa1362008-01-25 21:08:05 +01002520
Avi Kivitye107be32007-07-26 13:40:43 +02002521 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 if (mm)
2523 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002524 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08002525 /*
2526 * Remove function-return probe instances associated with this
2527 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002528 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002529 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002531 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532}
2533
2534/**
2535 * schedule_tail - first thing a freshly forked thread must call.
2536 * @prev: the thread we just switched away from.
2537 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002538asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 __releases(rq->lock)
2540{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002541 struct rq *rq = this_rq();
2542
Nick Piggin4866cde2005-06-25 14:57:23 -07002543 finish_task_switch(rq, prev);
2544#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2545 /* In this case, finish_task_switch does not reenable preemption */
2546 preempt_enable();
2547#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002549 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550}
2551
2552/*
2553 * context_switch - switch to the new MM and the new
2554 * thread's register state.
2555 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002556static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002557context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002558 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559{
Ingo Molnardd41f592007-07-09 18:51:59 +02002560 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561
Avi Kivitye107be32007-07-26 13:40:43 +02002562 prepare_task_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02002563 mm = next->mm;
2564 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002565 /*
2566 * For paravirt, this is coupled with an exit in switch_to to
2567 * combine the page table reload and the switch backend into
2568 * one hypercall.
2569 */
2570 arch_enter_lazy_cpu_mode();
2571
Ingo Molnardd41f592007-07-09 18:51:59 +02002572 if (unlikely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 next->active_mm = oldmm;
2574 atomic_inc(&oldmm->mm_count);
2575 enter_lazy_tlb(oldmm, next);
2576 } else
2577 switch_mm(oldmm, mm, next);
2578
Ingo Molnardd41f592007-07-09 18:51:59 +02002579 if (unlikely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581 rq->prev_mm = oldmm;
2582 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002583 /*
2584 * Since the runqueue lock will be released by the next
2585 * task (which is an invalid locking op but in the case
2586 * of the scheduler it's an obvious special-case), so we
2587 * do an early lockdep release here:
2588 */
2589#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002590 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002591#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592
2593 /* Here we just switch the register state and the stack. */
2594 switch_to(prev, next, prev);
2595
Ingo Molnardd41f592007-07-09 18:51:59 +02002596 barrier();
2597 /*
2598 * this_rq must be evaluated again because prev may have moved
2599 * CPUs since it called schedule(), thus the 'rq' on its stack
2600 * frame will be invalid.
2601 */
2602 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603}
2604
2605/*
2606 * nr_running, nr_uninterruptible and nr_context_switches:
2607 *
2608 * externally visible scheduler statistics: current number of runnable
2609 * threads, current number of uninterruptible-sleeping threads, total
2610 * number of context switches performed since bootup.
2611 */
2612unsigned long nr_running(void)
2613{
2614 unsigned long i, sum = 0;
2615
2616 for_each_online_cpu(i)
2617 sum += cpu_rq(i)->nr_running;
2618
2619 return sum;
2620}
2621
2622unsigned long nr_uninterruptible(void)
2623{
2624 unsigned long i, sum = 0;
2625
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002626 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 sum += cpu_rq(i)->nr_uninterruptible;
2628
2629 /*
2630 * Since we read the counters lockless, it might be slightly
2631 * inaccurate. Do not allow it to go below zero though:
2632 */
2633 if (unlikely((long)sum < 0))
2634 sum = 0;
2635
2636 return sum;
2637}
2638
2639unsigned long long nr_context_switches(void)
2640{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002641 int i;
2642 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002644 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645 sum += cpu_rq(i)->nr_switches;
2646
2647 return sum;
2648}
2649
2650unsigned long nr_iowait(void)
2651{
2652 unsigned long i, sum = 0;
2653
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002654 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2656
2657 return sum;
2658}
2659
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002660unsigned long nr_active(void)
2661{
2662 unsigned long i, running = 0, uninterruptible = 0;
2663
2664 for_each_online_cpu(i) {
2665 running += cpu_rq(i)->nr_running;
2666 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2667 }
2668
2669 if (unlikely((long)uninterruptible < 0))
2670 uninterruptible = 0;
2671
2672 return running + uninterruptible;
2673}
2674
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675/*
Ingo Molnardd41f592007-07-09 18:51:59 +02002676 * Update rq->cpu_load[] statistics. This function is usually called every
2677 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07002678 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002679static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002680{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02002681 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02002682 int i, scale;
2683
2684 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02002685
2686 /* Update our load: */
2687 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2688 unsigned long old_load, new_load;
2689
2690 /* scale is effectively 1 << i now, and >> i divides by scale */
2691
2692 old_load = this_rq->cpu_load[i];
2693 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02002694 /*
2695 * Round up the averaging division if load is increasing. This
2696 * prevents us from getting stuck on 9 if the load is 10, for
2697 * example.
2698 */
2699 if (new_load > old_load)
2700 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02002701 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2702 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002703}
2704
Ingo Molnardd41f592007-07-09 18:51:59 +02002705#ifdef CONFIG_SMP
2706
Ingo Molnar48f24c42006-07-03 00:25:40 -07002707/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 * double_rq_lock - safely lock two runqueues
2709 *
2710 * Note this does not disable interrupts like task_rq_lock,
2711 * you need to do so manually before calling.
2712 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002713static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 __acquires(rq1->lock)
2715 __acquires(rq2->lock)
2716{
Kirill Korotaev054b9102006-12-10 02:20:11 -08002717 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 if (rq1 == rq2) {
2719 spin_lock(&rq1->lock);
2720 __acquire(rq2->lock); /* Fake it out ;) */
2721 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002722 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 spin_lock(&rq1->lock);
2724 spin_lock(&rq2->lock);
2725 } else {
2726 spin_lock(&rq2->lock);
2727 spin_lock(&rq1->lock);
2728 }
2729 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002730 update_rq_clock(rq1);
2731 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732}
2733
2734/*
2735 * double_rq_unlock - safely unlock two runqueues
2736 *
2737 * Note this does not restore interrupts like task_rq_unlock,
2738 * you need to do so manually after calling.
2739 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002740static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741 __releases(rq1->lock)
2742 __releases(rq2->lock)
2743{
2744 spin_unlock(&rq1->lock);
2745 if (rq1 != rq2)
2746 spin_unlock(&rq2->lock);
2747 else
2748 __release(rq2->lock);
2749}
2750
2751/*
2752 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2753 */
Steven Rostedte8fa1362008-01-25 21:08:05 +01002754static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 __releases(this_rq->lock)
2756 __acquires(busiest->lock)
2757 __acquires(this_rq->lock)
2758{
Steven Rostedte8fa1362008-01-25 21:08:05 +01002759 int ret = 0;
2760
Kirill Korotaev054b9102006-12-10 02:20:11 -08002761 if (unlikely(!irqs_disabled())) {
2762 /* printk() doesn't work good under rq->lock */
2763 spin_unlock(&this_rq->lock);
2764 BUG_ON(1);
2765 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 if (unlikely(!spin_trylock(&busiest->lock))) {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002767 if (busiest < this_rq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 spin_unlock(&this_rq->lock);
2769 spin_lock(&busiest->lock);
2770 spin_lock(&this_rq->lock);
Steven Rostedte8fa1362008-01-25 21:08:05 +01002771 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 } else
2773 spin_lock(&busiest->lock);
2774 }
Steven Rostedte8fa1362008-01-25 21:08:05 +01002775 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776}
2777
2778/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 * If dest_cpu is allowed for this process, migrate the task to it.
2780 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002781 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 * the cpu_allowed mask is restored.
2783 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002784static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002786 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002788 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789
2790 rq = task_rq_lock(p, &flags);
2791 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2792 || unlikely(cpu_is_offline(dest_cpu)))
2793 goto out;
2794
2795 /* force the process onto the specified CPU */
2796 if (migrate_task(p, dest_cpu, &req)) {
2797 /* Need to wait for migration thread (might exit: take ref). */
2798 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07002799
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 get_task_struct(mt);
2801 task_rq_unlock(rq, &flags);
2802 wake_up_process(mt);
2803 put_task_struct(mt);
2804 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07002805
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 return;
2807 }
2808out:
2809 task_rq_unlock(rq, &flags);
2810}
2811
2812/*
Nick Piggin476d1392005-06-25 14:57:29 -07002813 * sched_exec - execve() is a valuable balancing opportunity, because at
2814 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815 */
2816void sched_exec(void)
2817{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818 int new_cpu, this_cpu = get_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002819 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002821 if (new_cpu != this_cpu)
2822 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823}
2824
2825/*
2826 * pull_task - move a task from a remote runqueue to the local runqueue.
2827 * Both runqueues must be locked.
2828 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002829static void pull_task(struct rq *src_rq, struct task_struct *p,
2830 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02002832 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002834 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 /*
2836 * Note that idle threads have a prio of MAX_PRIO, for this test
2837 * to be always true for them.
2838 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002839 check_preempt_curr(this_rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840}
2841
2842/*
2843 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2844 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08002845static
Ingo Molnar70b97a72006-07-03 00:25:42 -07002846int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002847 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002848 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849{
2850 /*
2851 * We do not migrate tasks that are:
2852 * 1) running (obviously), or
2853 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2854 * 3) are cache-hot on their current CPU.
2855 */
Ingo Molnarcc367732007-10-15 17:00:18 +02002856 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2857 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002859 }
Nick Piggin81026792005-06-25 14:57:07 -07002860 *all_pinned = 0;
2861
Ingo Molnarcc367732007-10-15 17:00:18 +02002862 if (task_running(rq, p)) {
2863 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07002864 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002865 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866
Ingo Molnarda84d962007-10-15 17:00:18 +02002867 /*
2868 * Aggressive migration if:
2869 * 1) task is cache cold, or
2870 * 2) too many balance attempts have failed.
2871 */
2872
Ingo Molnar6bc16652007-10-15 17:00:18 +02002873 if (!task_hot(p, rq->clock, sd) ||
2874 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002875#ifdef CONFIG_SCHEDSTATS
Ingo Molnarcc367732007-10-15 17:00:18 +02002876 if (task_hot(p, rq->clock, sd)) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002877 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02002878 schedstat_inc(p, se.nr_forced_migrations);
2879 }
Ingo Molnarda84d962007-10-15 17:00:18 +02002880#endif
2881 return 1;
2882 }
2883
Ingo Molnarcc367732007-10-15 17:00:18 +02002884 if (task_hot(p, rq->clock, sd)) {
2885 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02002886 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002887 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888 return 1;
2889}
2890
Peter Williamse1d14842007-10-24 18:23:51 +02002891static unsigned long
2892balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2893 unsigned long max_load_move, struct sched_domain *sd,
2894 enum cpu_idle_type idle, int *all_pinned,
2895 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02002896{
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002897 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
Ingo Molnardd41f592007-07-09 18:51:59 +02002898 struct task_struct *p;
2899 long rem_load_move = max_load_move;
2900
Peter Williamse1d14842007-10-24 18:23:51 +02002901 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02002902 goto out;
2903
2904 pinned = 1;
2905
2906 /*
2907 * Start the load-balancing iterator:
2908 */
2909 p = iterator->start(iterator->arg);
2910next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002911 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02002912 goto out;
2913 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002914 * To help distribute high priority tasks across CPUs we don't
Ingo Molnardd41f592007-07-09 18:51:59 +02002915 * skip a task if it will be the highest priority task (i.e. smallest
2916 * prio value) on its new queue regardless of its load weight
2917 */
2918 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2919 SCHED_LOAD_SCALE_FUZZ;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002920 if ((skip_for_load && p->prio >= *this_best_prio) ||
Ingo Molnardd41f592007-07-09 18:51:59 +02002921 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002922 p = iterator->next(iterator->arg);
2923 goto next;
2924 }
2925
2926 pull_task(busiest, p, this_rq, this_cpu);
2927 pulled++;
2928 rem_load_move -= p->se.load.weight;
2929
2930 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002931 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02002932 */
Peter Williamse1d14842007-10-24 18:23:51 +02002933 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002934 if (p->prio < *this_best_prio)
2935 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02002936 p = iterator->next(iterator->arg);
2937 goto next;
2938 }
2939out:
2940 /*
Peter Williamse1d14842007-10-24 18:23:51 +02002941 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02002942 * so we can safely collect pull_task() stats here rather than
2943 * inside pull_task().
2944 */
2945 schedstat_add(sd, lb_gained[idle], pulled);
2946
2947 if (all_pinned)
2948 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02002949
2950 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02002951}
Ingo Molnar48f24c42006-07-03 00:25:40 -07002952
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953/*
Peter Williams43010652007-08-09 11:16:46 +02002954 * move_tasks tries to move up to max_load_move weighted load from busiest to
2955 * this_rq, as part of a balancing operation within domain "sd".
2956 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 *
2958 * Called with both runqueues locked.
2959 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002960static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02002961 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002962 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07002963 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02002965 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02002966 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002967 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002968
Ingo Molnardd41f592007-07-09 18:51:59 +02002969 do {
Peter Williams43010652007-08-09 11:16:46 +02002970 total_load_moved +=
2971 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02002972 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002973 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02002974 class = class->next;
Peter Williams43010652007-08-09 11:16:46 +02002975 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976
Peter Williams43010652007-08-09 11:16:46 +02002977 return total_load_moved > 0;
2978}
2979
Peter Williamse1d14842007-10-24 18:23:51 +02002980static int
2981iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2982 struct sched_domain *sd, enum cpu_idle_type idle,
2983 struct rq_iterator *iterator)
2984{
2985 struct task_struct *p = iterator->start(iterator->arg);
2986 int pinned = 0;
2987
2988 while (p) {
2989 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2990 pull_task(busiest, p, this_rq, this_cpu);
2991 /*
2992 * Right now, this is only the second place pull_task()
2993 * is called, so we can safely collect pull_task()
2994 * stats here rather than inside pull_task().
2995 */
2996 schedstat_inc(sd, lb_gained[idle]);
2997
2998 return 1;
2999 }
3000 p = iterator->next(iterator->arg);
3001 }
3002
3003 return 0;
3004}
3005
Peter Williams43010652007-08-09 11:16:46 +02003006/*
3007 * move_one_task tries to move exactly one task from busiest to this_rq, as
3008 * part of active balancing operations within "domain".
3009 * Returns 1 if successful and 0 otherwise.
3010 *
3011 * Called with both runqueues locked.
3012 */
3013static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3014 struct sched_domain *sd, enum cpu_idle_type idle)
3015{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003016 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02003017
3018 for (class = sched_class_highest; class; class = class->next)
Peter Williamse1d14842007-10-24 18:23:51 +02003019 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02003020 return 1;
3021
3022 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023}
3024
3025/*
3026 * find_busiest_group finds and returns the busiest CPU group within the
Ingo Molnar48f24c42006-07-03 00:25:40 -07003027 * domain. It calculates and returns the amount of weighted load which
3028 * should be moved to restore balance via the imbalance parameter.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 */
3030static struct sched_group *
3031find_busiest_group(struct sched_domain *sd, int this_cpu,
Ingo Molnardd41f592007-07-09 18:51:59 +02003032 unsigned long *imbalance, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003033 int *sd_idle, const cpumask_t *cpus, int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034{
3035 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3036 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003037 unsigned long max_pull;
Peter Williams2dd73a42006-06-27 02:54:34 -07003038 unsigned long busiest_load_per_task, busiest_nr_running;
3039 unsigned long this_load_per_task, this_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02003040 int load_idx, group_imb = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003041#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3042 int power_savings_balance = 1;
3043 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3044 unsigned long min_nr_running = ULONG_MAX;
3045 struct sched_group *group_min = NULL, *group_leader = NULL;
3046#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047
3048 max_load = this_load = total_load = total_pwr = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07003049 busiest_load_per_task = busiest_nr_running = 0;
3050 this_load_per_task = this_nr_running = 0;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003051 if (idle == CPU_NOT_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07003052 load_idx = sd->busy_idx;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003053 else if (idle == CPU_NEWLY_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07003054 load_idx = sd->newidle_idx;
3055 else
3056 load_idx = sd->idle_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057
3058 do {
Ken Chen908a7c12007-10-17 16:55:11 +02003059 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 int local_group;
3061 int i;
Ken Chen908a7c12007-10-17 16:55:11 +02003062 int __group_imb = 0;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003063 unsigned int balance_cpu = -1, first_idle_cpu = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07003064 unsigned long sum_nr_running, sum_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065
3066 local_group = cpu_isset(this_cpu, group->cpumask);
3067
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003068 if (local_group)
3069 balance_cpu = first_cpu(group->cpumask);
3070
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 /* Tally up the load of all CPUs in the group */
Peter Williams2dd73a42006-06-27 02:54:34 -07003072 sum_weighted_load = sum_nr_running = avg_load = 0;
Ken Chen908a7c12007-10-17 16:55:11 +02003073 max_cpu_load = 0;
3074 min_cpu_load = ~0UL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075
3076 for_each_cpu_mask(i, group->cpumask) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003077 struct rq *rq;
3078
3079 if (!cpu_isset(i, *cpus))
3080 continue;
3081
3082 rq = cpu_rq(i);
Peter Williams2dd73a42006-06-27 02:54:34 -07003083
Suresh Siddha9439aab2007-07-19 21:28:35 +02003084 if (*sd_idle && rq->nr_running)
Nick Piggin5969fe02005-09-10 00:26:19 -07003085 *sd_idle = 0;
3086
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087 /* Bias balancing toward cpus of our domain */
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003088 if (local_group) {
3089 if (idle_cpu(i) && !first_idle_cpu) {
3090 first_idle_cpu = 1;
3091 balance_cpu = i;
3092 }
3093
Nick Piggina2000572006-02-10 01:51:02 -08003094 load = target_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02003095 } else {
Nick Piggina2000572006-02-10 01:51:02 -08003096 load = source_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02003097 if (load > max_cpu_load)
3098 max_cpu_load = load;
3099 if (min_cpu_load > load)
3100 min_cpu_load = load;
3101 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003102
3103 avg_load += load;
Peter Williams2dd73a42006-06-27 02:54:34 -07003104 sum_nr_running += rq->nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02003105 sum_weighted_load += weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106 }
3107
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003108 /*
3109 * First idle cpu or the first cpu(busiest) in this sched group
3110 * is eligible for doing load balancing at this and above
Suresh Siddha9439aab2007-07-19 21:28:35 +02003111 * domains. In the newly idle case, we will allow all the cpu's
3112 * to do the newly idle load balance.
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003113 */
Suresh Siddha9439aab2007-07-19 21:28:35 +02003114 if (idle != CPU_NEWLY_IDLE && local_group &&
3115 balance_cpu != this_cpu && balance) {
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003116 *balance = 0;
3117 goto ret;
3118 }
3119
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120 total_load += avg_load;
Eric Dumazet5517d862007-05-08 00:32:57 -07003121 total_pwr += group->__cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122
3123 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07003124 avg_load = sg_div_cpu_power(group,
3125 avg_load * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126
Ken Chen908a7c12007-10-17 16:55:11 +02003127 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3128 __group_imb = 1;
3129
Eric Dumazet5517d862007-05-08 00:32:57 -07003130 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003131
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132 if (local_group) {
3133 this_load = avg_load;
3134 this = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07003135 this_nr_running = sum_nr_running;
3136 this_load_per_task = sum_weighted_load;
3137 } else if (avg_load > max_load &&
Ken Chen908a7c12007-10-17 16:55:11 +02003138 (sum_nr_running > group_capacity || __group_imb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139 max_load = avg_load;
3140 busiest = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07003141 busiest_nr_running = sum_nr_running;
3142 busiest_load_per_task = sum_weighted_load;
Ken Chen908a7c12007-10-17 16:55:11 +02003143 group_imb = __group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003145
3146#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3147 /*
3148 * Busy processors will not participate in power savings
3149 * balance.
3150 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003151 if (idle == CPU_NOT_IDLE ||
3152 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3153 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003154
3155 /*
3156 * If the local group is idle or completely loaded
3157 * no need to do power savings balance at this domain
3158 */
3159 if (local_group && (this_nr_running >= group_capacity ||
3160 !this_nr_running))
3161 power_savings_balance = 0;
3162
Ingo Molnardd41f592007-07-09 18:51:59 +02003163 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003164 * If a group is already running at full capacity or idle,
3165 * don't include that group in power savings calculations
Ingo Molnardd41f592007-07-09 18:51:59 +02003166 */
3167 if (!power_savings_balance || sum_nr_running >= group_capacity
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003168 || !sum_nr_running)
Ingo Molnardd41f592007-07-09 18:51:59 +02003169 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003170
Ingo Molnardd41f592007-07-09 18:51:59 +02003171 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003172 * Calculate the group which has the least non-idle load.
Ingo Molnardd41f592007-07-09 18:51:59 +02003173 * This is the group from where we need to pick up the load
3174 * for saving power
3175 */
3176 if ((sum_nr_running < min_nr_running) ||
3177 (sum_nr_running == min_nr_running &&
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003178 first_cpu(group->cpumask) <
3179 first_cpu(group_min->cpumask))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003180 group_min = group;
3181 min_nr_running = sum_nr_running;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003182 min_load_per_task = sum_weighted_load /
3183 sum_nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02003184 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003185
Ingo Molnardd41f592007-07-09 18:51:59 +02003186 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003187 * Calculate the group which is almost near its
Ingo Molnardd41f592007-07-09 18:51:59 +02003188 * capacity but still has some space to pick up some load
3189 * from other group and save more power
3190 */
3191 if (sum_nr_running <= group_capacity - 1) {
3192 if (sum_nr_running > leader_nr_running ||
3193 (sum_nr_running == leader_nr_running &&
3194 first_cpu(group->cpumask) >
3195 first_cpu(group_leader->cpumask))) {
3196 group_leader = group;
3197 leader_nr_running = sum_nr_running;
3198 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07003199 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003200group_next:
3201#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202 group = group->next;
3203 } while (group != sd->groups);
3204
Peter Williams2dd73a42006-06-27 02:54:34 -07003205 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206 goto out_balanced;
3207
3208 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3209
3210 if (this_load >= avg_load ||
3211 100*max_load <= sd->imbalance_pct*this_load)
3212 goto out_balanced;
3213
Peter Williams2dd73a42006-06-27 02:54:34 -07003214 busiest_load_per_task /= busiest_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02003215 if (group_imb)
3216 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3217
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 /*
3219 * We're trying to get all the cpus to the average_load, so we don't
3220 * want to push ourselves above the average load, nor do we wish to
3221 * reduce the max loaded cpu below the average load, as either of these
3222 * actions would just result in more rebalancing later, and ping-pong
3223 * tasks around. Thus we look for the minimum possible imbalance.
3224 * Negative imbalances (*we* are more loaded than anyone else) will
3225 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003226 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227 * appear as very large values with unsigned longs.
3228 */
Peter Williams2dd73a42006-06-27 02:54:34 -07003229 if (max_load <= busiest_load_per_task)
3230 goto out_balanced;
3231
3232 /*
3233 * In the presence of smp nice balancing, certain scenarios can have
3234 * max load less than avg load(as we skip the groups at or below
3235 * its cpu_power, while calculating max_load..)
3236 */
3237 if (max_load < avg_load) {
3238 *imbalance = 0;
3239 goto small_imbalance;
3240 }
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003241
3242 /* Don't want to pull so many tasks that a group would go idle */
Peter Williams2dd73a42006-06-27 02:54:34 -07003243 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003244
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245 /* How much load to actually move to equalise the imbalance */
Eric Dumazet5517d862007-05-08 00:32:57 -07003246 *imbalance = min(max_pull * busiest->__cpu_power,
3247 (avg_load - this_load) * this->__cpu_power)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248 / SCHED_LOAD_SCALE;
3249
Peter Williams2dd73a42006-06-27 02:54:34 -07003250 /*
3251 * if *imbalance is less than the average load per runnable task
3252 * there is no gaurantee that any tasks will be moved so we'll have
3253 * a think about bumping its value to force at least one task to be
3254 * moved
3255 */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02003256 if (*imbalance < busiest_load_per_task) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07003257 unsigned long tmp, pwr_now, pwr_move;
Peter Williams2dd73a42006-06-27 02:54:34 -07003258 unsigned int imbn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003259
Peter Williams2dd73a42006-06-27 02:54:34 -07003260small_imbalance:
3261 pwr_move = pwr_now = 0;
3262 imbn = 2;
3263 if (this_nr_running) {
3264 this_load_per_task /= this_nr_running;
3265 if (busiest_load_per_task > this_load_per_task)
3266 imbn = 1;
3267 } else
3268 this_load_per_task = SCHED_LOAD_SCALE;
3269
Ingo Molnardd41f592007-07-09 18:51:59 +02003270 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3271 busiest_load_per_task * imbn) {
Peter Williams2dd73a42006-06-27 02:54:34 -07003272 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273 return busiest;
3274 }
3275
3276 /*
3277 * OK, we don't have enough imbalance to justify moving tasks,
3278 * however we may be able to increase total CPU power used by
3279 * moving them.
3280 */
3281
Eric Dumazet5517d862007-05-08 00:32:57 -07003282 pwr_now += busiest->__cpu_power *
3283 min(busiest_load_per_task, max_load);
3284 pwr_now += this->__cpu_power *
3285 min(this_load_per_task, this_load);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286 pwr_now /= SCHED_LOAD_SCALE;
3287
3288 /* Amount of load we'd subtract */
Eric Dumazet5517d862007-05-08 00:32:57 -07003289 tmp = sg_div_cpu_power(busiest,
3290 busiest_load_per_task * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291 if (max_load > tmp)
Eric Dumazet5517d862007-05-08 00:32:57 -07003292 pwr_move += busiest->__cpu_power *
Peter Williams2dd73a42006-06-27 02:54:34 -07003293 min(busiest_load_per_task, max_load - tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294
3295 /* Amount of load we'd add */
Eric Dumazet5517d862007-05-08 00:32:57 -07003296 if (max_load * busiest->__cpu_power <
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08003297 busiest_load_per_task * SCHED_LOAD_SCALE)
Eric Dumazet5517d862007-05-08 00:32:57 -07003298 tmp = sg_div_cpu_power(this,
3299 max_load * busiest->__cpu_power);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300 else
Eric Dumazet5517d862007-05-08 00:32:57 -07003301 tmp = sg_div_cpu_power(this,
3302 busiest_load_per_task * SCHED_LOAD_SCALE);
3303 pwr_move += this->__cpu_power *
3304 min(this_load_per_task, this_load + tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305 pwr_move /= SCHED_LOAD_SCALE;
3306
3307 /* Move if we gain throughput */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02003308 if (pwr_move > pwr_now)
3309 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 }
3311
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 return busiest;
3313
3314out_balanced:
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003315#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003316 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003317 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003319 if (this == group_leader && group_leader != group_min) {
3320 *imbalance = min_load_per_task;
3321 return group_min;
3322 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003323#endif
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003324ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325 *imbalance = 0;
3326 return NULL;
3327}
3328
3329/*
3330 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3331 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003332static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003333find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003334 unsigned long imbalance, const cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003336 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07003337 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338 int i;
3339
3340 for_each_cpu_mask(i, group->cpumask) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003341 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003342
3343 if (!cpu_isset(i, *cpus))
3344 continue;
3345
Ingo Molnar48f24c42006-07-03 00:25:40 -07003346 rq = cpu_rq(i);
Ingo Molnardd41f592007-07-09 18:51:59 +02003347 wl = weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348
Ingo Molnardd41f592007-07-09 18:51:59 +02003349 if (rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07003350 continue;
3351
Ingo Molnardd41f592007-07-09 18:51:59 +02003352 if (wl > max_load) {
3353 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003354 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003355 }
3356 }
3357
3358 return busiest;
3359}
3360
3361/*
Nick Piggin77391d72005-06-25 14:57:30 -07003362 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3363 * so long as it is large enough.
3364 */
3365#define MAX_PINNED_INTERVAL 512
3366
3367/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3369 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003371static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003372 struct sched_domain *sd, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003373 int *balance, cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003374{
Peter Williams43010652007-08-09 11:16:46 +02003375 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003378 struct rq *busiest;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003379 unsigned long flags;
Nick Piggin5969fe02005-09-10 00:26:19 -07003380
Mike Travis7c16ec52008-04-04 18:11:11 -07003381 cpus_setall(*cpus);
3382
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003383 /*
3384 * When power savings policy is enabled for the parent domain, idle
3385 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02003386 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003387 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003388 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003389 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003390 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003391 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392
Ingo Molnar2d723762007-10-15 17:00:12 +02003393 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003395redo:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02003396 update_shares(sd);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003397 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003398 cpus, balance);
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003399
Chen, Kenneth W06066712006-12-10 02:20:35 -08003400 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003401 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003402
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403 if (!group) {
3404 schedstat_inc(sd, lb_nobusyg[idle]);
3405 goto out_balanced;
3406 }
3407
Mike Travis7c16ec52008-04-04 18:11:11 -07003408 busiest = find_busiest_queue(group, idle, imbalance, cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 if (!busiest) {
3410 schedstat_inc(sd, lb_nobusyq[idle]);
3411 goto out_balanced;
3412 }
3413
Nick Piggindb935db2005-06-25 14:57:11 -07003414 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415
3416 schedstat_add(sd, lb_imbalance[idle], imbalance);
3417
Peter Williams43010652007-08-09 11:16:46 +02003418 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419 if (busiest->nr_running > 1) {
3420 /*
3421 * Attempt to move tasks. If find_busiest_group has found
3422 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02003423 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 * correctly treated as an imbalance.
3425 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003426 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07003427 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02003428 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07003429 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07003430 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003431 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07003432
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003433 /*
3434 * some other cpu did the load balance for us.
3435 */
Peter Williams43010652007-08-09 11:16:46 +02003436 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003437 resched_cpu(this_cpu);
3438
Nick Piggin81026792005-06-25 14:57:07 -07003439 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003440 if (unlikely(all_pinned)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003441 cpu_clear(cpu_of(busiest), *cpus);
3442 if (!cpus_empty(*cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003443 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07003444 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003445 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 }
Nick Piggin81026792005-06-25 14:57:07 -07003447
Peter Williams43010652007-08-09 11:16:46 +02003448 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449 schedstat_inc(sd, lb_failed[idle]);
3450 sd->nr_balance_failed++;
3451
3452 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003454 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003455
3456 /* don't kick the migration_thread, if the curr
3457 * task on busiest cpu can't be moved to this_cpu
3458 */
3459 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003460 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003461 all_pinned = 1;
3462 goto out_one_pinned;
3463 }
3464
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465 if (!busiest->active_balance) {
3466 busiest->active_balance = 1;
3467 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07003468 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003470 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07003471 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472 wake_up_process(busiest->migration_thread);
3473
3474 /*
3475 * We've kicked active balancing, reset the failure
3476 * counter.
3477 */
Nick Piggin39507452005-06-25 14:57:09 -07003478 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479 }
Nick Piggin81026792005-06-25 14:57:07 -07003480 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481 sd->nr_balance_failed = 0;
3482
Nick Piggin81026792005-06-25 14:57:07 -07003483 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484 /* We were unbalanced, so reset the balancing interval */
3485 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07003486 } else {
3487 /*
3488 * If we've begun active balancing, start to back off. This
3489 * case may not be covered by the all_pinned logic if there
3490 * is only 1 task on the busy runqueue (because we don't call
3491 * move_tasks).
3492 */
3493 if (sd->balance_interval < sd->max_interval)
3494 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 }
3496
Peter Williams43010652007-08-09 11:16:46 +02003497 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003498 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02003499 ld_moved = -1;
3500
3501 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502
3503out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504 schedstat_inc(sd, lb_balanced[idle]);
3505
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003506 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003507
3508out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07003510 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3511 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512 sd->balance_interval *= 2;
3513
Ingo Molnar48f24c42006-07-03 00:25:40 -07003514 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003515 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02003516 ld_moved = -1;
3517 else
3518 ld_moved = 0;
3519out:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02003520 if (ld_moved)
3521 update_shares(sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02003522 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523}
3524
3525/*
3526 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3527 * tasks if there is an imbalance.
3528 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003529 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530 * this_rq is locked.
3531 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07003532static int
Mike Travis7c16ec52008-04-04 18:11:11 -07003533load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3534 cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535{
3536 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003537 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02003539 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07003540 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003541 int all_pinned = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07003542
3543 cpus_setall(*cpus);
Nick Piggin5969fe02005-09-10 00:26:19 -07003544
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003545 /*
3546 * When power savings policy is enabled for the parent domain, idle
3547 * sibling can pick up load irrespective of busy siblings. In this case,
3548 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003549 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003550 */
3551 if (sd->flags & SD_SHARE_CPUPOWER &&
3552 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003553 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554
Ingo Molnar2d723762007-10-15 17:00:12 +02003555 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003556redo:
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02003557 update_shares_locked(this_rq, sd);
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003558 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Mike Travis7c16ec52008-04-04 18:11:11 -07003559 &sd_idle, cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003561 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003562 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563 }
3564
Mike Travis7c16ec52008-04-04 18:11:11 -07003565 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07003566 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003567 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003568 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569 }
3570
Nick Piggindb935db2005-06-25 14:57:11 -07003571 BUG_ON(busiest == this_rq);
3572
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003573 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003574
Peter Williams43010652007-08-09 11:16:46 +02003575 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003576 if (busiest->nr_running > 1) {
3577 /* Attempt to move tasks */
3578 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003579 /* this_rq->clock is already updated */
3580 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02003581 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003582 imbalance, sd, CPU_NEWLY_IDLE,
3583 &all_pinned);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003584 spin_unlock(&busiest->lock);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003585
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003586 if (unlikely(all_pinned)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003587 cpu_clear(cpu_of(busiest), *cpus);
3588 if (!cpus_empty(*cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003589 goto redo;
3590 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003591 }
3592
Peter Williams43010652007-08-09 11:16:46 +02003593 if (!ld_moved) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003594 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003595 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3596 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003597 return -1;
3598 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003599 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02003601 update_shares_locked(this_rq, sd);
Peter Williams43010652007-08-09 11:16:46 +02003602 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003603
3604out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003605 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003606 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003607 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003608 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003609 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003610
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003611 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612}
3613
3614/*
3615 * idle_balance is called by schedule() if this_cpu is about to become
3616 * idle. Attempts to pull tasks from other CPUs.
3617 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003618static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619{
3620 struct sched_domain *sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02003621 int pulled_task = -1;
3622 unsigned long next_balance = jiffies + HZ;
Mike Travis7c16ec52008-04-04 18:11:11 -07003623 cpumask_t tmpmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624
3625 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003626 unsigned long interval;
3627
3628 if (!(sd->flags & SD_LOAD_BALANCE))
3629 continue;
3630
3631 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07003632 /* If we've pulled tasks over stop searching: */
Mike Travis7c16ec52008-04-04 18:11:11 -07003633 pulled_task = load_balance_newidle(this_cpu, this_rq,
3634 sd, &tmpmask);
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003635
3636 interval = msecs_to_jiffies(sd->balance_interval);
3637 if (time_after(next_balance, sd->last_balance + interval))
3638 next_balance = sd->last_balance + interval;
3639 if (pulled_task)
3640 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003641 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003642 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003643 /*
3644 * We are going idle. next_balance may be set based on
3645 * a busy processor. So reset next_balance.
3646 */
3647 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02003648 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649}
3650
3651/*
3652 * active_load_balance is run by migration threads. It pushes running tasks
3653 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3654 * running on each physical CPU where possible, and avoids physical /
3655 * logical imbalances.
3656 *
3657 * Called with busiest_rq locked.
3658 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003659static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660{
Nick Piggin39507452005-06-25 14:57:09 -07003661 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003662 struct sched_domain *sd;
3663 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07003664
Ingo Molnar48f24c42006-07-03 00:25:40 -07003665 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07003666 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07003667 return;
3668
3669 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670
3671 /*
Nick Piggin39507452005-06-25 14:57:09 -07003672 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003673 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07003674 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675 */
Nick Piggin39507452005-06-25 14:57:09 -07003676 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677
Nick Piggin39507452005-06-25 14:57:09 -07003678 /* move a task from busiest_rq to target_rq */
3679 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003680 update_rq_clock(busiest_rq);
3681 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682
Nick Piggin39507452005-06-25 14:57:09 -07003683 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003684 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07003685 if ((sd->flags & SD_LOAD_BALANCE) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07003686 cpu_isset(busiest_cpu, sd->span))
Nick Piggin39507452005-06-25 14:57:09 -07003687 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003688 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003689
Ingo Molnar48f24c42006-07-03 00:25:40 -07003690 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02003691 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003692
Peter Williams43010652007-08-09 11:16:46 +02003693 if (move_one_task(target_rq, target_cpu, busiest_rq,
3694 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07003695 schedstat_inc(sd, alb_pushed);
3696 else
3697 schedstat_inc(sd, alb_failed);
3698 }
Nick Piggin39507452005-06-25 14:57:09 -07003699 spin_unlock(&target_rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700}
3701
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003702#ifdef CONFIG_NO_HZ
3703static struct {
3704 atomic_t load_balancer;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003705 cpumask_t cpu_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003706} nohz ____cacheline_aligned = {
3707 .load_balancer = ATOMIC_INIT(-1),
3708 .cpu_mask = CPU_MASK_NONE,
3709};
3710
Christoph Lameter7835b982006-12-10 02:20:22 -08003711/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003712 * This routine will try to nominate the ilb (idle load balancing)
3713 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3714 * load balancing on behalf of all those cpus. If all the cpus in the system
3715 * go into this tickless mode, then there will be no ilb owner (as there is
3716 * no need for one) and all the cpus will sleep till the next wakeup event
3717 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08003718 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003719 * For the ilb owner, tick is not stopped. And this tick will be used
3720 * for idle load balancing. ilb owner will still be part of
3721 * nohz.cpu_mask..
3722 *
3723 * While stopping the tick, this cpu will become the ilb owner if there
3724 * is no other owner. And will be the owner till that cpu becomes busy
3725 * or if all cpus in the system stop their ticks at which point
3726 * there is no need for ilb owner.
3727 *
3728 * When the ilb owner becomes busy, it nominates another owner, during the
3729 * next busy scheduler_tick()
3730 */
3731int select_nohz_load_balancer(int stop_tick)
3732{
3733 int cpu = smp_processor_id();
3734
3735 if (stop_tick) {
3736 cpu_set(cpu, nohz.cpu_mask);
3737 cpu_rq(cpu)->in_nohz_recently = 1;
3738
3739 /*
3740 * If we are going offline and still the leader, give up!
3741 */
3742 if (cpu_is_offline(cpu) &&
3743 atomic_read(&nohz.load_balancer) == cpu) {
3744 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3745 BUG();
3746 return 0;
3747 }
3748
3749 /* time for ilb owner also to sleep */
3750 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3751 if (atomic_read(&nohz.load_balancer) == cpu)
3752 atomic_set(&nohz.load_balancer, -1);
3753 return 0;
3754 }
3755
3756 if (atomic_read(&nohz.load_balancer) == -1) {
3757 /* make me the ilb owner */
3758 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3759 return 1;
3760 } else if (atomic_read(&nohz.load_balancer) == cpu)
3761 return 1;
3762 } else {
3763 if (!cpu_isset(cpu, nohz.cpu_mask))
3764 return 0;
3765
3766 cpu_clear(cpu, nohz.cpu_mask);
3767
3768 if (atomic_read(&nohz.load_balancer) == cpu)
3769 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3770 BUG();
3771 }
3772 return 0;
3773}
3774#endif
3775
3776static DEFINE_SPINLOCK(balancing);
3777
3778/*
Christoph Lameter7835b982006-12-10 02:20:22 -08003779 * It checks each scheduling domain to see if it is due to be balanced,
3780 * and initiates a balancing operation if so.
3781 *
3782 * Balancing parameters are set up in arch_init_sched_domains.
3783 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02003784static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08003785{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003786 int balance = 1;
3787 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08003788 unsigned long interval;
3789 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003790 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08003791 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003792 int update_next_balance = 0;
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003793 int need_serialize;
Mike Travis7c16ec52008-04-04 18:11:11 -07003794 cpumask_t tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003796 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797 if (!(sd->flags & SD_LOAD_BALANCE))
3798 continue;
3799
3800 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003801 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802 interval *= sd->busy_factor;
3803
3804 /* scale ms to jiffies */
3805 interval = msecs_to_jiffies(interval);
3806 if (unlikely(!interval))
3807 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003808 if (interval > HZ*NR_CPUS/10)
3809 interval = HZ*NR_CPUS/10;
3810
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003811 need_serialize = sd->flags & SD_SERIALIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003812
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003813 if (need_serialize) {
Christoph Lameter08c183f2006-12-10 02:20:29 -08003814 if (!spin_trylock(&balancing))
3815 goto out;
3816 }
3817
Christoph Lameterc9819f42006-12-10 02:20:25 -08003818 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003819 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003820 /*
3821 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07003822 * longer idle, or one of our SMT siblings is
3823 * not idle.
3824 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003825 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003827 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828 }
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003829 if (need_serialize)
Christoph Lameter08c183f2006-12-10 02:20:29 -08003830 spin_unlock(&balancing);
3831out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02003832 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08003833 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003834 update_next_balance = 1;
3835 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003836
3837 /*
3838 * Stop the load balance at this level. There is another
3839 * CPU in our sched group which is doing load balancing more
3840 * actively.
3841 */
3842 if (!balance)
3843 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02003845
3846 /*
3847 * next_balance will be updated only when there is a need.
3848 * When the cpu is attached to null domain for ex, it will not be
3849 * updated.
3850 */
3851 if (likely(update_next_balance))
3852 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003853}
3854
3855/*
3856 * run_rebalance_domains is triggered when needed from the scheduler tick.
3857 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3858 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3859 */
3860static void run_rebalance_domains(struct softirq_action *h)
3861{
Ingo Molnardd41f592007-07-09 18:51:59 +02003862 int this_cpu = smp_processor_id();
3863 struct rq *this_rq = cpu_rq(this_cpu);
3864 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3865 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003866
Ingo Molnardd41f592007-07-09 18:51:59 +02003867 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003868
3869#ifdef CONFIG_NO_HZ
3870 /*
3871 * If this cpu is the owner for idle load balancing, then do the
3872 * balancing on behalf of the other idle cpus whose ticks are
3873 * stopped.
3874 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003875 if (this_rq->idle_at_tick &&
3876 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003877 cpumask_t cpus = nohz.cpu_mask;
3878 struct rq *rq;
3879 int balance_cpu;
3880
Ingo Molnardd41f592007-07-09 18:51:59 +02003881 cpu_clear(this_cpu, cpus);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003882 for_each_cpu_mask(balance_cpu, cpus) {
3883 /*
3884 * If this cpu gets work to do, stop the load balancing
3885 * work being done for other cpus. Next load
3886 * balancing owner will pick it up.
3887 */
3888 if (need_resched())
3889 break;
3890
Oleg Nesterovde0cf892007-08-12 18:08:19 +02003891 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003892
3893 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003894 if (time_after(this_rq->next_balance, rq->next_balance))
3895 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003896 }
3897 }
3898#endif
3899}
3900
3901/*
3902 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3903 *
3904 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3905 * idle load balancing owner or decide to stop the periodic load balancing,
3906 * if the whole system is idle.
3907 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003908static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003909{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003910#ifdef CONFIG_NO_HZ
3911 /*
3912 * If we were in the nohz mode recently and busy at the current
3913 * scheduler tick, then check if we need to nominate new idle
3914 * load balancer.
3915 */
3916 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3917 rq->in_nohz_recently = 0;
3918
3919 if (atomic_read(&nohz.load_balancer) == cpu) {
3920 cpu_clear(cpu, nohz.cpu_mask);
3921 atomic_set(&nohz.load_balancer, -1);
3922 }
3923
3924 if (atomic_read(&nohz.load_balancer) == -1) {
3925 /*
3926 * simple selection for now: Nominate the
3927 * first cpu in the nohz list to be the next
3928 * ilb owner.
3929 *
3930 * TBD: Traverse the sched domains and nominate
3931 * the nearest cpu in the nohz.cpu_mask.
3932 */
3933 int ilb = first_cpu(nohz.cpu_mask);
3934
Mike Travis434d53b2008-04-04 18:11:04 -07003935 if (ilb < nr_cpu_ids)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003936 resched_cpu(ilb);
3937 }
3938 }
3939
3940 /*
3941 * If this cpu is idle and doing idle load balancing for all the
3942 * cpus with ticks stopped, is it time for that to stop?
3943 */
3944 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3945 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3946 resched_cpu(cpu);
3947 return;
3948 }
3949
3950 /*
3951 * If this cpu is idle and the idle load balancing is done by
3952 * someone else, then no need raise the SCHED_SOFTIRQ
3953 */
3954 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3955 cpu_isset(cpu, nohz.cpu_mask))
3956 return;
3957#endif
3958 if (time_after_eq(jiffies, rq->next_balance))
3959 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960}
Ingo Molnardd41f592007-07-09 18:51:59 +02003961
3962#else /* CONFIG_SMP */
3963
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964/*
3965 * on UP we do not need to balance between CPUs:
3966 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003967static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968{
3969}
Ingo Molnardd41f592007-07-09 18:51:59 +02003970
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971#endif
3972
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973DEFINE_PER_CPU(struct kernel_stat, kstat);
3974
3975EXPORT_PER_CPU_SYMBOL(kstat);
3976
3977/*
Ingo Molnar41b86e92007-07-09 18:51:58 +02003978 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3979 * that have not yet been banked in case the task is currently running.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980 */
Ingo Molnar41b86e92007-07-09 18:51:58 +02003981unsigned long long task_sched_runtime(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003984 u64 ns, delta_exec;
3985 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003986
Ingo Molnar41b86e92007-07-09 18:51:58 +02003987 rq = task_rq_lock(p, &flags);
3988 ns = p->se.sum_exec_runtime;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01003989 if (task_current(rq, p)) {
Ingo Molnara8e504d2007-08-09 11:16:47 +02003990 update_rq_clock(rq);
3991 delta_exec = rq->clock - p->se.exec_start;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003992 if ((s64)delta_exec > 0)
3993 ns += delta_exec;
3994 }
3995 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003996
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997 return ns;
3998}
3999
4000/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001 * Account user cpu time to a process.
4002 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003 * @cputime: the cpu time spent in user space since the last update
4004 */
4005void account_user_time(struct task_struct *p, cputime_t cputime)
4006{
4007 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4008 cputime64_t tmp;
4009
4010 p->utime = cputime_add(p->utime, cputime);
4011
4012 /* Add user time to cpustat. */
4013 tmp = cputime_to_cputime64(cputime);
4014 if (TASK_NICE(p) > 0)
4015 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4016 else
4017 cpustat->user = cputime64_add(cpustat->user, tmp);
4018}
4019
4020/*
Laurent Vivier94886b82007-10-15 17:00:19 +02004021 * Account guest cpu time to a process.
4022 * @p: the process that the cpu time gets accounted to
4023 * @cputime: the cpu time spent in virtual machine since the last update
4024 */
Adrian Bunkf7402e02007-10-29 21:18:10 +01004025static void account_guest_time(struct task_struct *p, cputime_t cputime)
Laurent Vivier94886b82007-10-15 17:00:19 +02004026{
4027 cputime64_t tmp;
4028 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4029
4030 tmp = cputime_to_cputime64(cputime);
4031
4032 p->utime = cputime_add(p->utime, cputime);
4033 p->gtime = cputime_add(p->gtime, cputime);
4034
4035 cpustat->user = cputime64_add(cpustat->user, tmp);
4036 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4037}
4038
4039/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07004040 * Account scaled user cpu time to a process.
4041 * @p: the process that the cpu time gets accounted to
4042 * @cputime: the cpu time spent in user space since the last update
4043 */
4044void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4045{
4046 p->utimescaled = cputime_add(p->utimescaled, cputime);
4047}
4048
4049/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050 * Account system cpu time to a process.
4051 * @p: the process that the cpu time gets accounted to
4052 * @hardirq_offset: the offset to subtract from hardirq_count()
4053 * @cputime: the cpu time spent in kernel space since the last update
4054 */
4055void account_system_time(struct task_struct *p, int hardirq_offset,
4056 cputime_t cputime)
4057{
4058 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004059 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060 cputime64_t tmp;
4061
Harvey Harrison983ed7a2008-04-24 18:17:55 -07004062 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4063 account_guest_time(p, cputime);
4064 return;
4065 }
Laurent Vivier94886b82007-10-15 17:00:19 +02004066
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067 p->stime = cputime_add(p->stime, cputime);
4068
4069 /* Add system time to cpustat. */
4070 tmp = cputime_to_cputime64(cputime);
4071 if (hardirq_count() - hardirq_offset)
4072 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4073 else if (softirq_count())
4074 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08004075 else if (p != rq->idle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076 cpustat->system = cputime64_add(cpustat->system, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08004077 else if (atomic_read(&rq->nr_iowait) > 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4079 else
4080 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4081 /* Account for system time used */
4082 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083}
4084
4085/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07004086 * Account scaled system cpu time to a process.
4087 * @p: the process that the cpu time gets accounted to
4088 * @hardirq_offset: the offset to subtract from hardirq_count()
4089 * @cputime: the cpu time spent in kernel space since the last update
4090 */
4091void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4092{
4093 p->stimescaled = cputime_add(p->stimescaled, cputime);
4094}
4095
4096/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097 * Account for involuntary wait time.
4098 * @p: the process from which the cpu time has been stolen
4099 * @steal: the cpu time spent in involuntary wait
4100 */
4101void account_steal_time(struct task_struct *p, cputime_t steal)
4102{
4103 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4104 cputime64_t tmp = cputime_to_cputime64(steal);
Ingo Molnar70b97a72006-07-03 00:25:42 -07004105 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106
4107 if (p == rq->idle) {
4108 p->stime = cputime_add(p->stime, steal);
4109 if (atomic_read(&rq->nr_iowait) > 0)
4110 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4111 else
4112 cpustat->idle = cputime64_add(cpustat->idle, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08004113 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4115}
4116
Christoph Lameter7835b982006-12-10 02:20:22 -08004117/*
4118 * This function gets called by the timer code, with HZ frequency.
4119 * We call it with interrupts disabled.
4120 *
4121 * It also gets called by the fork code, when changing the parent's
4122 * timeslices.
4123 */
4124void scheduler_tick(void)
4125{
Christoph Lameter7835b982006-12-10 02:20:22 -08004126 int cpu = smp_processor_id();
4127 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004128 struct task_struct *curr = rq->curr;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004129
4130 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08004131
Ingo Molnardd41f592007-07-09 18:51:59 +02004132 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004133 update_rq_clock(rq);
Ingo Molnarf1a438d2007-08-09 11:16:45 +02004134 update_cpu_load(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01004135 curr->sched_class->task_tick(rq, curr, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02004136 spin_unlock(&rq->lock);
4137
Christoph Lametere418e1c2006-12-10 02:20:23 -08004138#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02004139 rq->idle_at_tick = idle_cpu(cpu);
4140 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08004141#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142}
4143
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4145
Srinivasa Ds43627582008-02-23 15:24:04 -08004146void __kprobes add_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147{
4148 /*
4149 * Underflow?
4150 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004151 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4152 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153 preempt_count() += val;
4154 /*
4155 * Spinlock count overflowing soon?
4156 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08004157 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4158 PREEMPT_MASK - 10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159}
4160EXPORT_SYMBOL(add_preempt_count);
4161
Srinivasa Ds43627582008-02-23 15:24:04 -08004162void __kprobes sub_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163{
4164 /*
4165 * Underflow?
4166 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004167 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4168 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169 /*
4170 * Is the spinlock portion underflowing?
4171 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004172 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4173 !(preempt_count() & PREEMPT_MASK)))
4174 return;
4175
Linus Torvalds1da177e2005-04-16 15:20:36 -07004176 preempt_count() -= val;
4177}
4178EXPORT_SYMBOL(sub_preempt_count);
4179
4180#endif
4181
4182/*
Ingo Molnardd41f592007-07-09 18:51:59 +02004183 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004184 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004185static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186{
Satyam Sharma838225b2007-10-24 18:23:50 +02004187 struct pt_regs *regs = get_irq_regs();
4188
4189 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4190 prev->comm, prev->pid, preempt_count());
4191
Ingo Molnardd41f592007-07-09 18:51:59 +02004192 debug_show_held_locks(prev);
Arjan van de Vene21f5b12008-05-23 09:05:58 -07004193 print_modules();
Ingo Molnardd41f592007-07-09 18:51:59 +02004194 if (irqs_disabled())
4195 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02004196
4197 if (regs)
4198 show_regs(regs);
4199 else
4200 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02004201}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202
Ingo Molnardd41f592007-07-09 18:51:59 +02004203/*
4204 * Various schedule()-time debugging checks and statistics:
4205 */
4206static inline void schedule_debug(struct task_struct *prev)
4207{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004209 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210 * schedule() atomically, we ignore that path for now.
4211 * Otherwise, whine if we are scheduling when we should not be.
4212 */
Roel Kluin3f33a7c2008-05-13 23:44:11 +02004213 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
Ingo Molnardd41f592007-07-09 18:51:59 +02004214 __schedule_bug(prev);
4215
Linus Torvalds1da177e2005-04-16 15:20:36 -07004216 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4217
Ingo Molnar2d723762007-10-15 17:00:12 +02004218 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02004219#ifdef CONFIG_SCHEDSTATS
4220 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02004221 schedstat_inc(this_rq(), bkl_count);
4222 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02004223 }
4224#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02004225}
4226
4227/*
4228 * Pick up the highest-prio task:
4229 */
4230static inline struct task_struct *
Ingo Molnarff95f3d2007-08-09 11:16:49 +02004231pick_next_task(struct rq *rq, struct task_struct *prev)
Ingo Molnardd41f592007-07-09 18:51:59 +02004232{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02004233 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02004234 struct task_struct *p;
4235
4236 /*
4237 * Optimization: we know that if all tasks are in
4238 * the fair class we can call that function directly:
4239 */
4240 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02004241 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004242 if (likely(p))
4243 return p;
4244 }
4245
4246 class = sched_class_highest;
4247 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02004248 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004249 if (p)
4250 return p;
4251 /*
4252 * Will never be NULL as the idle class always
4253 * returns a non-NULL p:
4254 */
4255 class = class->next;
4256 }
4257}
4258
4259/*
4260 * schedule() is the main scheduler function.
4261 */
4262asmlinkage void __sched schedule(void)
4263{
4264 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08004265 unsigned long *switch_count;
Ingo Molnardd41f592007-07-09 18:51:59 +02004266 struct rq *rq;
Mike Galbraithf333fdc2008-05-12 21:20:55 +02004267 int cpu, hrtick = sched_feat(HRTICK);
Ingo Molnardd41f592007-07-09 18:51:59 +02004268
Linus Torvalds1da177e2005-04-16 15:20:36 -07004269need_resched:
4270 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02004271 cpu = smp_processor_id();
4272 rq = cpu_rq(cpu);
4273 rcu_qsctr_inc(cpu);
4274 prev = rq->curr;
4275 switch_count = &prev->nivcsw;
4276
Linus Torvalds1da177e2005-04-16 15:20:36 -07004277 release_kernel_lock(prev);
4278need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004279
Ingo Molnardd41f592007-07-09 18:51:59 +02004280 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281
Mike Galbraithf333fdc2008-05-12 21:20:55 +02004282 if (hrtick)
4283 hrtick_clear(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004284
Ingo Molnar1e819952007-10-15 17:00:13 +02004285 /*
4286 * Do the rq-clock update outside the rq lock:
4287 */
4288 local_irq_disable();
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004289 update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02004290 spin_lock(&rq->lock);
4291 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004292
Ingo Molnardd41f592007-07-09 18:51:59 +02004293 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
Oleg Nesterov16882c12008-06-08 21:20:41 +04004294 if (unlikely(signal_pending_state(prev->state, prev)))
Ingo Molnardd41f592007-07-09 18:51:59 +02004295 prev->state = TASK_RUNNING;
Oleg Nesterov16882c12008-06-08 21:20:41 +04004296 else
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004297 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02004298 switch_count = &prev->nvcsw;
4299 }
4300
Steven Rostedt9a897c52008-01-25 21:08:22 +01004301#ifdef CONFIG_SMP
4302 if (prev->sched_class->pre_schedule)
4303 prev->sched_class->pre_schedule(rq, prev);
4304#endif
Steven Rostedtf65eda42008-01-25 21:08:07 +01004305
Ingo Molnardd41f592007-07-09 18:51:59 +02004306 if (unlikely(!rq->nr_running))
4307 idle_balance(cpu, rq);
4308
Ingo Molnar31ee5292007-08-09 11:16:49 +02004309 prev->sched_class->put_prev_task(rq, prev);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02004310 next = pick_next_task(rq, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004311
Linus Torvalds1da177e2005-04-16 15:20:36 -07004312 if (likely(prev != next)) {
David Simner673a90a2008-04-29 10:08:59 +01004313 sched_info_switch(prev, next);
4314
Linus Torvalds1da177e2005-04-16 15:20:36 -07004315 rq->nr_switches++;
4316 rq->curr = next;
4317 ++*switch_count;
4318
Ingo Molnardd41f592007-07-09 18:51:59 +02004319 context_switch(rq, prev, next); /* unlocks the rq */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004320 /*
4321 * the context switch might have flipped the stack from under
4322 * us, hence refresh the local variables.
4323 */
4324 cpu = smp_processor_id();
4325 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004326 } else
4327 spin_unlock_irq(&rq->lock);
4328
Mike Galbraithf333fdc2008-05-12 21:20:55 +02004329 if (hrtick)
4330 hrtick_set(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004331
4332 if (unlikely(reacquire_kernel_lock(current) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004333 goto need_resched_nonpreemptible;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004334
Linus Torvalds1da177e2005-04-16 15:20:36 -07004335 preempt_enable_no_resched();
4336 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4337 goto need_resched;
4338}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004339EXPORT_SYMBOL(schedule);
4340
4341#ifdef CONFIG_PREEMPT
4342/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004343 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004344 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07004345 * occur there and call schedule directly.
4346 */
4347asmlinkage void __sched preempt_schedule(void)
4348{
4349 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01004350
Linus Torvalds1da177e2005-04-16 15:20:36 -07004351 /*
4352 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004353 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07004354 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07004355 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004356 return;
4357
Andi Kleen3a5c3592007-10-15 17:00:14 +02004358 do {
4359 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02004360 schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02004361 sub_preempt_count(PREEMPT_ACTIVE);
4362
4363 /*
4364 * Check again in case we missed a preemption opportunity
4365 * between schedule and now.
4366 */
4367 barrier();
4368 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004369}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004370EXPORT_SYMBOL(preempt_schedule);
4371
4372/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004373 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07004374 * off of irq context.
4375 * Note, that this is called and return with irqs disabled. This will
4376 * protect us against recursive calling from irq.
4377 */
4378asmlinkage void __sched preempt_schedule_irq(void)
4379{
4380 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01004381
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004382 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004383 BUG_ON(ti->preempt_count || !irqs_disabled());
4384
Andi Kleen3a5c3592007-10-15 17:00:14 +02004385 do {
4386 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02004387 local_irq_enable();
4388 schedule();
4389 local_irq_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02004390 sub_preempt_count(PREEMPT_ACTIVE);
4391
4392 /*
4393 * Check again in case we missed a preemption opportunity
4394 * between schedule and now.
4395 */
4396 barrier();
4397 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004398}
4399
4400#endif /* CONFIG_PREEMPT */
4401
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004402int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4403 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004404{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004405 return try_to_wake_up(curr->private, mode, sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004406}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004407EXPORT_SYMBOL(default_wake_function);
4408
4409/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004410 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4411 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07004412 * number) then we wake all the non-exclusive tasks and one exclusive task.
4413 *
4414 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004415 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07004416 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4417 */
4418static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4419 int nr_exclusive, int sync, void *key)
4420{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004421 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004422
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004423 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07004424 unsigned flags = curr->flags;
4425
Linus Torvalds1da177e2005-04-16 15:20:36 -07004426 if (curr->func(curr, mode, sync, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07004427 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004428 break;
4429 }
4430}
4431
4432/**
4433 * __wake_up - wake up threads blocked on a waitqueue.
4434 * @q: the waitqueue
4435 * @mode: which threads
4436 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07004437 * @key: is directly passed to the wakeup function
Linus Torvalds1da177e2005-04-16 15:20:36 -07004438 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004439void __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004440 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004441{
4442 unsigned long flags;
4443
4444 spin_lock_irqsave(&q->lock, flags);
4445 __wake_up_common(q, mode, nr_exclusive, 0, key);
4446 spin_unlock_irqrestore(&q->lock, flags);
4447}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004448EXPORT_SYMBOL(__wake_up);
4449
4450/*
4451 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4452 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004453void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004454{
4455 __wake_up_common(q, mode, 1, 0, NULL);
4456}
4457
4458/**
Martin Waitz67be2dd2005-05-01 08:59:26 -07004459 * __wake_up_sync - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004460 * @q: the waitqueue
4461 * @mode: which threads
4462 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4463 *
4464 * The sync wakeup differs that the waker knows that it will schedule
4465 * away soon, so while the target thread will be woken up, it will not
4466 * be migrated to another CPU - ie. the two threads are 'synchronized'
4467 * with each other. This can prevent needless bouncing between CPUs.
4468 *
4469 * On UP it can prevent extra preemption.
4470 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004471void
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004472__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004473{
4474 unsigned long flags;
4475 int sync = 1;
4476
4477 if (unlikely(!q))
4478 return;
4479
4480 if (unlikely(!nr_exclusive))
4481 sync = 0;
4482
4483 spin_lock_irqsave(&q->lock, flags);
4484 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4485 spin_unlock_irqrestore(&q->lock, flags);
4486}
4487EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4488
Ingo Molnarb15136e2007-10-24 18:23:48 +02004489void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004490{
4491 unsigned long flags;
4492
4493 spin_lock_irqsave(&x->wait.lock, flags);
4494 x->done++;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004495 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004496 spin_unlock_irqrestore(&x->wait.lock, flags);
4497}
4498EXPORT_SYMBOL(complete);
4499
Ingo Molnarb15136e2007-10-24 18:23:48 +02004500void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004501{
4502 unsigned long flags;
4503
4504 spin_lock_irqsave(&x->wait.lock, flags);
4505 x->done += UINT_MAX/2;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004506 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004507 spin_unlock_irqrestore(&x->wait.lock, flags);
4508}
4509EXPORT_SYMBOL(complete_all);
4510
Andi Kleen8cbbe862007-10-15 17:00:14 +02004511static inline long __sched
4512do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004513{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004514 if (!x->done) {
4515 DECLARE_WAITQUEUE(wait, current);
4516
4517 wait.flags |= WQ_FLAG_EXCLUSIVE;
4518 __add_wait_queue_tail(&x->wait, &wait);
4519 do {
Matthew Wilcox009e5772007-12-06 12:29:54 -05004520 if ((state == TASK_INTERRUPTIBLE &&
4521 signal_pending(current)) ||
4522 (state == TASK_KILLABLE &&
4523 fatal_signal_pending(current))) {
Oleg Nesterovea71a542008-06-20 18:32:20 +04004524 timeout = -ERESTARTSYS;
4525 break;
Andi Kleen8cbbe862007-10-15 17:00:14 +02004526 }
4527 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004528 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004529 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004530 spin_lock_irq(&x->wait.lock);
Oleg Nesterovea71a542008-06-20 18:32:20 +04004531 } while (!x->done && timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004532 __remove_wait_queue(&x->wait, &wait);
Oleg Nesterovea71a542008-06-20 18:32:20 +04004533 if (!x->done)
4534 return timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004535 }
4536 x->done--;
Oleg Nesterovea71a542008-06-20 18:32:20 +04004537 return timeout ?: 1;
Andi Kleen8cbbe862007-10-15 17:00:14 +02004538}
4539
4540static long __sched
4541wait_for_common(struct completion *x, long timeout, int state)
4542{
4543 might_sleep();
4544
4545 spin_lock_irq(&x->wait.lock);
4546 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004547 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004548 return timeout;
4549}
4550
Ingo Molnarb15136e2007-10-24 18:23:48 +02004551void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02004552{
4553 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004554}
4555EXPORT_SYMBOL(wait_for_completion);
4556
Ingo Molnarb15136e2007-10-24 18:23:48 +02004557unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004558wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4559{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004560 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004561}
4562EXPORT_SYMBOL(wait_for_completion_timeout);
4563
Andi Kleen8cbbe862007-10-15 17:00:14 +02004564int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004565{
Andi Kleen51e97992007-10-18 21:32:55 +02004566 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4567 if (t == -ERESTARTSYS)
4568 return t;
4569 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004570}
4571EXPORT_SYMBOL(wait_for_completion_interruptible);
4572
Ingo Molnarb15136e2007-10-24 18:23:48 +02004573unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004574wait_for_completion_interruptible_timeout(struct completion *x,
4575 unsigned long timeout)
4576{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004577 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004578}
4579EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4580
Matthew Wilcox009e5772007-12-06 12:29:54 -05004581int __sched wait_for_completion_killable(struct completion *x)
4582{
4583 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4584 if (t == -ERESTARTSYS)
4585 return t;
4586 return 0;
4587}
4588EXPORT_SYMBOL(wait_for_completion_killable);
4589
Andi Kleen8cbbe862007-10-15 17:00:14 +02004590static long __sched
4591sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02004592{
4593 unsigned long flags;
4594 wait_queue_t wait;
4595
4596 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004597
Andi Kleen8cbbe862007-10-15 17:00:14 +02004598 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004599
Andi Kleen8cbbe862007-10-15 17:00:14 +02004600 spin_lock_irqsave(&q->lock, flags);
4601 __add_wait_queue(q, &wait);
4602 spin_unlock(&q->lock);
4603 timeout = schedule_timeout(timeout);
4604 spin_lock_irq(&q->lock);
4605 __remove_wait_queue(q, &wait);
4606 spin_unlock_irqrestore(&q->lock, flags);
4607
4608 return timeout;
4609}
4610
4611void __sched interruptible_sleep_on(wait_queue_head_t *q)
4612{
4613 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004614}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004615EXPORT_SYMBOL(interruptible_sleep_on);
4616
Ingo Molnar0fec1712007-07-09 18:52:01 +02004617long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004618interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004619{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004620 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004621}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004622EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4623
Ingo Molnar0fec1712007-07-09 18:52:01 +02004624void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004625{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004626 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004627}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004628EXPORT_SYMBOL(sleep_on);
4629
Ingo Molnar0fec1712007-07-09 18:52:01 +02004630long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004631{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004632 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004633}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004634EXPORT_SYMBOL(sleep_on_timeout);
4635
Ingo Molnarb29739f2006-06-27 02:54:51 -07004636#ifdef CONFIG_RT_MUTEXES
4637
4638/*
4639 * rt_mutex_setprio - set the current priority of a task
4640 * @p: task
4641 * @prio: prio value (kernel-internal form)
4642 *
4643 * This function changes the 'effective' priority of a task. It does
4644 * not touch ->normal_prio like __setscheduler().
4645 *
4646 * Used by the rt_mutex code to implement priority inheritance logic.
4647 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004648void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07004649{
4650 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004651 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004652 struct rq *rq;
Steven Rostedtcb469842008-01-25 21:08:22 +01004653 const struct sched_class *prev_class = p->sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004654
4655 BUG_ON(prio < 0 || prio > MAX_PRIO);
4656
4657 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004658 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004659
Andrew Mortond5f9f942007-05-08 20:27:06 -07004660 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004661 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004662 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004663 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02004664 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004665 if (running)
4666 p->sched_class->put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02004667
4668 if (rt_prio(prio))
4669 p->sched_class = &rt_sched_class;
4670 else
4671 p->sched_class = &fair_sched_class;
4672
Ingo Molnarb29739f2006-06-27 02:54:51 -07004673 p->prio = prio;
4674
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004675 if (running)
4676 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004677 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004678 enqueue_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004679
4680 check_class_changed(rq, p, prev_class, oldprio, running);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004681 }
4682 task_rq_unlock(rq, &flags);
4683}
4684
4685#endif
4686
Ingo Molnar36c8b582006-07-03 00:25:41 -07004687void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004688{
Ingo Molnardd41f592007-07-09 18:51:59 +02004689 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004690 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004691 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004692
4693 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4694 return;
4695 /*
4696 * We have to be careful, if called from sys_setpriority(),
4697 * the task might be in the middle of scheduling on another CPU.
4698 */
4699 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004700 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004701 /*
4702 * The RT priorities are set via sched_setscheduler(), but we still
4703 * allow the 'normal' nice value to be set - but as expected
4704 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02004705 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004706 */
Ingo Molnare05606d2007-07-09 18:51:59 +02004707 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004708 p->static_prio = NICE_TO_PRIO(nice);
4709 goto out_unlock;
4710 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004711 on_rq = p->se.on_rq;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004712 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02004713 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004714
Linus Torvalds1da177e2005-04-16 15:20:36 -07004715 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07004716 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004717 old_prio = p->prio;
4718 p->prio = effective_prio(p);
4719 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004720
Ingo Molnardd41f592007-07-09 18:51:59 +02004721 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004722 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004723 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07004724 * If the task increased its priority or is running and
4725 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004726 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07004727 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004728 resched_task(rq->curr);
4729 }
4730out_unlock:
4731 task_rq_unlock(rq, &flags);
4732}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004733EXPORT_SYMBOL(set_user_nice);
4734
Matt Mackalle43379f2005-05-01 08:59:00 -07004735/*
4736 * can_nice - check if a task can reduce its nice value
4737 * @p: task
4738 * @nice: nice value
4739 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004740int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07004741{
Matt Mackall024f4742005-08-18 11:24:19 -07004742 /* convert nice value [19,-20] to rlimit style value [1,40] */
4743 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004744
Matt Mackalle43379f2005-05-01 08:59:00 -07004745 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4746 capable(CAP_SYS_NICE));
4747}
4748
Linus Torvalds1da177e2005-04-16 15:20:36 -07004749#ifdef __ARCH_WANT_SYS_NICE
4750
4751/*
4752 * sys_nice - change the priority of the current process.
4753 * @increment: priority increment
4754 *
4755 * sys_setpriority is a more generic, but much slower function that
4756 * does similar things.
4757 */
4758asmlinkage long sys_nice(int increment)
4759{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004760 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004761
4762 /*
4763 * Setpriority might change our priority at the same moment.
4764 * We don't have to worry. Conceptually one call occurs first
4765 * and we have a single winner.
4766 */
Matt Mackalle43379f2005-05-01 08:59:00 -07004767 if (increment < -40)
4768 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004769 if (increment > 40)
4770 increment = 40;
4771
4772 nice = PRIO_TO_NICE(current->static_prio) + increment;
4773 if (nice < -20)
4774 nice = -20;
4775 if (nice > 19)
4776 nice = 19;
4777
Matt Mackalle43379f2005-05-01 08:59:00 -07004778 if (increment < 0 && !can_nice(current, nice))
4779 return -EPERM;
4780
Linus Torvalds1da177e2005-04-16 15:20:36 -07004781 retval = security_task_setnice(current, nice);
4782 if (retval)
4783 return retval;
4784
4785 set_user_nice(current, nice);
4786 return 0;
4787}
4788
4789#endif
4790
4791/**
4792 * task_prio - return the priority value of a given task.
4793 * @p: the task in question.
4794 *
4795 * This is the priority value as seen by users in /proc.
4796 * RT tasks are offset by -200. Normal tasks are centered
4797 * around 0, value goes from -16 to +15.
4798 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004799int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004800{
4801 return p->prio - MAX_RT_PRIO;
4802}
4803
4804/**
4805 * task_nice - return the nice value of a given task.
4806 * @p: the task in question.
4807 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004808int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004809{
4810 return TASK_NICE(p);
4811}
Pavel Roskin150d8be2008-03-05 16:56:37 -05004812EXPORT_SYMBOL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004813
4814/**
4815 * idle_cpu - is a given cpu idle currently?
4816 * @cpu: the processor in question.
4817 */
4818int idle_cpu(int cpu)
4819{
4820 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4821}
4822
Linus Torvalds1da177e2005-04-16 15:20:36 -07004823/**
4824 * idle_task - return the idle task for a given cpu.
4825 * @cpu: the processor in question.
4826 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004827struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004828{
4829 return cpu_rq(cpu)->idle;
4830}
4831
4832/**
4833 * find_process_by_pid - find a process with a matching PID value.
4834 * @pid: the pid in question.
4835 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02004836static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004837{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07004838 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004839}
4840
4841/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02004842static void
4843__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004844{
Ingo Molnardd41f592007-07-09 18:51:59 +02004845 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004846
Linus Torvalds1da177e2005-04-16 15:20:36 -07004847 p->policy = policy;
Ingo Molnardd41f592007-07-09 18:51:59 +02004848 switch (p->policy) {
4849 case SCHED_NORMAL:
4850 case SCHED_BATCH:
4851 case SCHED_IDLE:
4852 p->sched_class = &fair_sched_class;
4853 break;
4854 case SCHED_FIFO:
4855 case SCHED_RR:
4856 p->sched_class = &rt_sched_class;
4857 break;
4858 }
4859
Linus Torvalds1da177e2005-04-16 15:20:36 -07004860 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004861 p->normal_prio = normal_prio(p);
4862 /* we are holding p->pi_lock already */
4863 p->prio = rt_mutex_getprio(p);
Peter Williams2dd73a42006-06-27 02:54:34 -07004864 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004865}
4866
4867/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004868 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004869 * @p: the task in question.
4870 * @policy: new policy.
4871 * @param: structure containing the new RT priority.
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004872 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004873 * NOTE that the task may be already dead.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004874 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004875int sched_setscheduler(struct task_struct *p, int policy,
4876 struct sched_param *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004877{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004878 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004879 unsigned long flags;
Steven Rostedtcb469842008-01-25 21:08:22 +01004880 const struct sched_class *prev_class = p->sched_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004881 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004882
Steven Rostedt66e53932006-06-27 02:54:44 -07004883 /* may grab non-irq protected spin_locks */
4884 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07004885recheck:
4886 /* double check policy once rq lock held */
4887 if (policy < 0)
4888 policy = oldpolicy = p->policy;
4889 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
Ingo Molnardd41f592007-07-09 18:51:59 +02004890 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4891 policy != SCHED_IDLE)
Ingo Molnarb0a94992006-01-14 13:20:41 -08004892 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004893 /*
4894 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02004895 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4896 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004897 */
4898 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004899 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04004900 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004901 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02004902 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004903 return -EINVAL;
4904
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004905 /*
4906 * Allow unprivileged RT tasks to decrease priority:
4907 */
4908 if (!capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02004909 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004910 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004911
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004912 if (!lock_task_sighand(p, &flags))
4913 return -ESRCH;
4914 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4915 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004916
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004917 /* can't set/change the rt policy */
4918 if (policy != p->policy && !rlim_rtprio)
4919 return -EPERM;
4920
4921 /* can't increase priority */
4922 if (param->sched_priority > p->rt_priority &&
4923 param->sched_priority > rlim_rtprio)
4924 return -EPERM;
4925 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004926 /*
4927 * Like positive nice levels, dont allow tasks to
4928 * move out of SCHED_IDLE either:
4929 */
4930 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4931 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004932
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004933 /* can't change other user's priorities */
4934 if ((current->euid != p->euid) &&
4935 (current->euid != p->uid))
4936 return -EPERM;
4937 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004938
Peter Zijlstrab68aa232008-02-13 15:45:40 +01004939#ifdef CONFIG_RT_GROUP_SCHED
4940 /*
4941 * Do not allow realtime tasks into groups that have no runtime
4942 * assigned.
4943 */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02004944 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
Peter Zijlstrab68aa232008-02-13 15:45:40 +01004945 return -EPERM;
4946#endif
4947
Linus Torvalds1da177e2005-04-16 15:20:36 -07004948 retval = security_task_setscheduler(p, policy, param);
4949 if (retval)
4950 return retval;
4951 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07004952 * make sure no PI-waiters arrive (or leave) while we are
4953 * changing the priority of the task:
4954 */
4955 spin_lock_irqsave(&p->pi_lock, flags);
4956 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004957 * To be able to change p->policy safely, the apropriate
4958 * runqueue lock must be held.
4959 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07004960 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004961 /* recheck policy now with rq lock held */
4962 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4963 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004964 __task_rq_unlock(rq);
4965 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004966 goto recheck;
4967 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02004968 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004969 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004970 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004971 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004972 deactivate_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004973 if (running)
4974 p->sched_class->put_prev_task(rq, p);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02004975
Linus Torvalds1da177e2005-04-16 15:20:36 -07004976 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004977 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02004978
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004979 if (running)
4980 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004981 if (on_rq) {
4982 activate_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004983
4984 check_class_changed(rq, p, prev_class, oldprio, running);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004985 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07004986 __task_rq_unlock(rq);
4987 spin_unlock_irqrestore(&p->pi_lock, flags);
4988
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07004989 rt_mutex_adjust_pi(p);
4990
Linus Torvalds1da177e2005-04-16 15:20:36 -07004991 return 0;
4992}
4993EXPORT_SYMBOL_GPL(sched_setscheduler);
4994
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004995static int
4996do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004997{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004998 struct sched_param lparam;
4999 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005000 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005001
5002 if (!param || pid < 0)
5003 return -EINVAL;
5004 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5005 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005006
5007 rcu_read_lock();
5008 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005009 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005010 if (p != NULL)
5011 retval = sched_setscheduler(p, policy, &lparam);
5012 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07005013
Linus Torvalds1da177e2005-04-16 15:20:36 -07005014 return retval;
5015}
5016
5017/**
5018 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5019 * @pid: the pid in question.
5020 * @policy: new policy.
5021 * @param: structure containing the new RT priority.
5022 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005023asmlinkage long
5024sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005025{
Jason Baronc21761f2006-01-18 17:43:03 -08005026 /* negative values for policy are not valid */
5027 if (policy < 0)
5028 return -EINVAL;
5029
Linus Torvalds1da177e2005-04-16 15:20:36 -07005030 return do_sched_setscheduler(pid, policy, param);
5031}
5032
5033/**
5034 * sys_sched_setparam - set/change the RT priority of a thread
5035 * @pid: the pid in question.
5036 * @param: structure containing the new RT priority.
5037 */
5038asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5039{
5040 return do_sched_setscheduler(pid, -1, param);
5041}
5042
5043/**
5044 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5045 * @pid: the pid in question.
5046 */
5047asmlinkage long sys_sched_getscheduler(pid_t pid)
5048{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005049 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005050 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005051
5052 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005053 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005054
5055 retval = -ESRCH;
5056 read_lock(&tasklist_lock);
5057 p = find_process_by_pid(pid);
5058 if (p) {
5059 retval = security_task_getscheduler(p);
5060 if (!retval)
5061 retval = p->policy;
5062 }
5063 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005064 return retval;
5065}
5066
5067/**
5068 * sys_sched_getscheduler - get the RT priority of a thread
5069 * @pid: the pid in question.
5070 * @param: structure containing the RT priority.
5071 */
5072asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5073{
5074 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005075 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005076 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005077
5078 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005079 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005080
5081 read_lock(&tasklist_lock);
5082 p = find_process_by_pid(pid);
5083 retval = -ESRCH;
5084 if (!p)
5085 goto out_unlock;
5086
5087 retval = security_task_getscheduler(p);
5088 if (retval)
5089 goto out_unlock;
5090
5091 lp.sched_priority = p->rt_priority;
5092 read_unlock(&tasklist_lock);
5093
5094 /*
5095 * This one might sleep, we cannot do it with a spinlock held ...
5096 */
5097 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5098
Linus Torvalds1da177e2005-04-16 15:20:36 -07005099 return retval;
5100
5101out_unlock:
5102 read_unlock(&tasklist_lock);
5103 return retval;
5104}
5105
Mike Travisb53e9212008-04-04 18:11:08 -07005106long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005107{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005108 cpumask_t cpus_allowed;
Mike Travisb53e9212008-04-04 18:11:08 -07005109 cpumask_t new_mask = *in_mask;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005110 struct task_struct *p;
5111 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005112
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005113 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005114 read_lock(&tasklist_lock);
5115
5116 p = find_process_by_pid(pid);
5117 if (!p) {
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 return -ESRCH;
5121 }
5122
5123 /*
5124 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005125 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005126 * usage count and then drop tasklist_lock.
5127 */
5128 get_task_struct(p);
5129 read_unlock(&tasklist_lock);
5130
5131 retval = -EPERM;
5132 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5133 !capable(CAP_SYS_NICE))
5134 goto out_unlock;
5135
David Quigleye7834f82006-06-23 02:03:59 -07005136 retval = security_task_setscheduler(p, 0, NULL);
5137 if (retval)
5138 goto out_unlock;
5139
Mike Travisf9a86fc2008-04-04 18:11:07 -07005140 cpuset_cpus_allowed(p, &cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005141 cpus_and(new_mask, new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07005142 again:
Mike Travis7c16ec52008-04-04 18:11:11 -07005143 retval = set_cpus_allowed_ptr(p, &new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005144
Paul Menage8707d8b2007-10-18 23:40:22 -07005145 if (!retval) {
Mike Travisf9a86fc2008-04-04 18:11:07 -07005146 cpuset_cpus_allowed(p, &cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07005147 if (!cpus_subset(new_mask, cpus_allowed)) {
5148 /*
5149 * We must have raced with a concurrent cpuset
5150 * update. Just reset the cpus_allowed to the
5151 * cpuset's cpus_allowed
5152 */
5153 new_mask = cpus_allowed;
5154 goto again;
5155 }
5156 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005157out_unlock:
5158 put_task_struct(p);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005159 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005160 return retval;
5161}
5162
5163static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5164 cpumask_t *new_mask)
5165{
5166 if (len < sizeof(cpumask_t)) {
5167 memset(new_mask, 0, sizeof(cpumask_t));
5168 } else if (len > sizeof(cpumask_t)) {
5169 len = sizeof(cpumask_t);
5170 }
5171 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5172}
5173
5174/**
5175 * sys_sched_setaffinity - set the cpu affinity of a process
5176 * @pid: pid of the process
5177 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5178 * @user_mask_ptr: user-space pointer to the new cpu mask
5179 */
5180asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5181 unsigned long __user *user_mask_ptr)
5182{
5183 cpumask_t new_mask;
5184 int retval;
5185
5186 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5187 if (retval)
5188 return retval;
5189
Mike Travisb53e9212008-04-04 18:11:08 -07005190 return sched_setaffinity(pid, &new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005191}
5192
Linus Torvalds1da177e2005-04-16 15:20:36 -07005193long sched_getaffinity(pid_t pid, cpumask_t *mask)
5194{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005195 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005196 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005197
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005198 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005199 read_lock(&tasklist_lock);
5200
5201 retval = -ESRCH;
5202 p = find_process_by_pid(pid);
5203 if (!p)
5204 goto out_unlock;
5205
David Quigleye7834f82006-06-23 02:03:59 -07005206 retval = security_task_getscheduler(p);
5207 if (retval)
5208 goto out_unlock;
5209
Jack Steiner2f7016d2006-02-01 03:05:18 -08005210 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005211
5212out_unlock:
5213 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005214 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005215
Ulrich Drepper9531b622007-08-09 11:16:46 +02005216 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005217}
5218
5219/**
5220 * sys_sched_getaffinity - get the cpu affinity of a process
5221 * @pid: pid of the process
5222 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5223 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5224 */
5225asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5226 unsigned long __user *user_mask_ptr)
5227{
5228 int ret;
5229 cpumask_t mask;
5230
5231 if (len < sizeof(cpumask_t))
5232 return -EINVAL;
5233
5234 ret = sched_getaffinity(pid, &mask);
5235 if (ret < 0)
5236 return ret;
5237
5238 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5239 return -EFAULT;
5240
5241 return sizeof(cpumask_t);
5242}
5243
5244/**
5245 * sys_sched_yield - yield the current processor to other threads.
5246 *
Ingo Molnardd41f592007-07-09 18:51:59 +02005247 * This function yields the current CPU to other tasks. If there are no
5248 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005249 */
5250asmlinkage long sys_sched_yield(void)
5251{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005252 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005253
Ingo Molnar2d723762007-10-15 17:00:12 +02005254 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02005255 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005256
5257 /*
5258 * Since we are going to call schedule() anyway, there's
5259 * no need to preempt or enable interrupts:
5260 */
5261 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07005262 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005263 _raw_spin_unlock(&rq->lock);
5264 preempt_enable_no_resched();
5265
5266 schedule();
5267
5268 return 0;
5269}
5270
Andrew Mortone7b38402006-06-30 01:56:00 -07005271static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005272{
Ingo Molnar8e0a43d2006-06-23 02:05:23 -07005273#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5274 __might_sleep(__FILE__, __LINE__);
5275#endif
Ingo Molnar5bbcfd92005-07-07 17:57:04 -07005276 /*
5277 * The BKS might be reacquired before we have dropped
5278 * PREEMPT_ACTIVE, which could trigger a second
5279 * cond_resched() call.
5280 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005281 do {
5282 add_preempt_count(PREEMPT_ACTIVE);
5283 schedule();
5284 sub_preempt_count(PREEMPT_ACTIVE);
5285 } while (need_resched());
5286}
5287
Herbert Xu02b67cc32008-01-25 21:08:28 +01005288int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005289{
Ingo Molnar94142322006-12-29 16:48:13 -08005290 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5291 system_state == SYSTEM_RUNNING) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005292 __cond_resched();
5293 return 1;
5294 }
5295 return 0;
5296}
Herbert Xu02b67cc32008-01-25 21:08:28 +01005297EXPORT_SYMBOL(_cond_resched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005298
5299/*
5300 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5301 * call schedule, and on return reacquire the lock.
5302 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005303 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07005304 * operations here to prevent schedule() from being called twice (once via
5305 * spin_unlock(), once by hand).
5306 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005307int cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005308{
Nick Piggin95c354f2008-01-30 13:31:20 +01005309 int resched = need_resched() && system_state == SYSTEM_RUNNING;
Jan Kara6df3cec2005-06-13 15:52:32 -07005310 int ret = 0;
5311
Nick Piggin95c354f2008-01-30 13:31:20 +01005312 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005313 spin_unlock(lock);
Nick Piggin95c354f2008-01-30 13:31:20 +01005314 if (resched && need_resched())
5315 __cond_resched();
5316 else
5317 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07005318 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005319 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005320 }
Jan Kara6df3cec2005-06-13 15:52:32 -07005321 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005322}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005323EXPORT_SYMBOL(cond_resched_lock);
5324
5325int __sched cond_resched_softirq(void)
5326{
5327 BUG_ON(!in_softirq());
5328
Ingo Molnar94142322006-12-29 16:48:13 -08005329 if (need_resched() && system_state == SYSTEM_RUNNING) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07005330 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005331 __cond_resched();
5332 local_bh_disable();
5333 return 1;
5334 }
5335 return 0;
5336}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005337EXPORT_SYMBOL(cond_resched_softirq);
5338
Linus Torvalds1da177e2005-04-16 15:20:36 -07005339/**
5340 * yield - yield the current processor to other threads.
5341 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08005342 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005343 * thread runnable and calls sys_sched_yield().
5344 */
5345void __sched yield(void)
5346{
5347 set_current_state(TASK_RUNNING);
5348 sys_sched_yield();
5349}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005350EXPORT_SYMBOL(yield);
5351
5352/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005353 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07005354 * that process accounting knows that this is a task in IO wait state.
5355 *
5356 * But don't do that if it is a deliberate, throttling IO wait (this task
5357 * has set its backing_dev_info: the queue against which it should throttle)
5358 */
5359void __sched io_schedule(void)
5360{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005361 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005362
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005363 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005364 atomic_inc(&rq->nr_iowait);
5365 schedule();
5366 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005367 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005368}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005369EXPORT_SYMBOL(io_schedule);
5370
5371long __sched io_schedule_timeout(long timeout)
5372{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005373 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005374 long ret;
5375
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005376 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005377 atomic_inc(&rq->nr_iowait);
5378 ret = schedule_timeout(timeout);
5379 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005380 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005381 return ret;
5382}
5383
5384/**
5385 * sys_sched_get_priority_max - return maximum RT priority.
5386 * @policy: scheduling class.
5387 *
5388 * this syscall returns the maximum rt_priority that can be used
5389 * by a given scheduling class.
5390 */
5391asmlinkage long sys_sched_get_priority_max(int policy)
5392{
5393 int ret = -EINVAL;
5394
5395 switch (policy) {
5396 case SCHED_FIFO:
5397 case SCHED_RR:
5398 ret = MAX_USER_RT_PRIO-1;
5399 break;
5400 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005401 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005402 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005403 ret = 0;
5404 break;
5405 }
5406 return ret;
5407}
5408
5409/**
5410 * sys_sched_get_priority_min - return minimum RT priority.
5411 * @policy: scheduling class.
5412 *
5413 * this syscall returns the minimum rt_priority that can be used
5414 * by a given scheduling class.
5415 */
5416asmlinkage long sys_sched_get_priority_min(int policy)
5417{
5418 int ret = -EINVAL;
5419
5420 switch (policy) {
5421 case SCHED_FIFO:
5422 case SCHED_RR:
5423 ret = 1;
5424 break;
5425 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005426 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005427 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005428 ret = 0;
5429 }
5430 return ret;
5431}
5432
5433/**
5434 * sys_sched_rr_get_interval - return the default timeslice of a process.
5435 * @pid: pid of the process.
5436 * @interval: userspace pointer to the timeslice value.
5437 *
5438 * this syscall writes the default timeslice value of a given process
5439 * into the user-space timespec buffer. A value of '0' means infinity.
5440 */
5441asmlinkage
5442long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5443{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005444 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005445 unsigned int time_slice;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005446 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005447 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005448
5449 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005450 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005451
5452 retval = -ESRCH;
5453 read_lock(&tasklist_lock);
5454 p = find_process_by_pid(pid);
5455 if (!p)
5456 goto out_unlock;
5457
5458 retval = security_task_getscheduler(p);
5459 if (retval)
5460 goto out_unlock;
5461
Ingo Molnar77034932007-12-04 17:04:39 +01005462 /*
5463 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5464 * tasks that are on an otherwise idle runqueue:
5465 */
5466 time_slice = 0;
5467 if (p->policy == SCHED_RR) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005468 time_slice = DEF_TIMESLICE;
Miao Xie1868f952008-03-07 09:35:06 +08005469 } else if (p->policy != SCHED_FIFO) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005470 struct sched_entity *se = &p->se;
5471 unsigned long flags;
5472 struct rq *rq;
5473
5474 rq = task_rq_lock(p, &flags);
Ingo Molnar77034932007-12-04 17:04:39 +01005475 if (rq->cfs.load.weight)
5476 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005477 task_rq_unlock(rq, &flags);
5478 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005479 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005480 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005481 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005482 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005483
Linus Torvalds1da177e2005-04-16 15:20:36 -07005484out_unlock:
5485 read_unlock(&tasklist_lock);
5486 return retval;
5487}
5488
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005489static const char stat_nam[] = "RSDTtZX";
Ingo Molnar36c8b582006-07-03 00:25:41 -07005490
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005491void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005492{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005493 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005494 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005495
Linus Torvalds1da177e2005-04-16 15:20:36 -07005496 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005497 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005498 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02005499#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07005500 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005501 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005502 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005503 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005504#else
5505 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005506 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005507 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005508 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005509#endif
5510#ifdef CONFIG_DEBUG_STACK_USAGE
5511 {
Al Viro10ebffd2005-11-13 16:06:56 -08005512 unsigned long *n = end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005513 while (!*n)
5514 n++;
Al Viro10ebffd2005-11-13 16:06:56 -08005515 free = (unsigned long)n - (unsigned long)end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005516 }
5517#endif
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -07005518 printk(KERN_CONT "%5lu %5d %6d\n", free,
Roland McGrathfcfd50a2008-01-09 00:03:23 -08005519 task_pid_nr(p), task_pid_nr(p->real_parent));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005520
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01005521 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005522}
5523
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005524void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005525{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005526 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005527
Ingo Molnar4bd77322007-07-11 21:21:47 +02005528#if BITS_PER_LONG == 32
5529 printk(KERN_INFO
5530 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005531#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02005532 printk(KERN_INFO
5533 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005534#endif
5535 read_lock(&tasklist_lock);
5536 do_each_thread(g, p) {
5537 /*
5538 * reset the NMI-timeout, listing all files on a slow
5539 * console might take alot of time:
5540 */
5541 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07005542 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005543 sched_show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005544 } while_each_thread(g, p);
5545
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07005546 touch_all_softlockup_watchdogs();
5547
Ingo Molnardd41f592007-07-09 18:51:59 +02005548#ifdef CONFIG_SCHED_DEBUG
5549 sysrq_sched_debug_show();
5550#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005551 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005552 /*
5553 * Only show locks if all tasks are dumped:
5554 */
5555 if (state_filter == -1)
5556 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005557}
5558
Ingo Molnar1df21052007-07-09 18:51:58 +02005559void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5560{
Ingo Molnardd41f592007-07-09 18:51:59 +02005561 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02005562}
5563
Ingo Molnarf340c0d2005-06-28 16:40:42 +02005564/**
5565 * init_idle - set up an idle thread for a given CPU
5566 * @idle: task in question
5567 * @cpu: cpu the idle task belongs to
5568 *
5569 * NOTE: this function does not set the idle thread's NEED_RESCHED
5570 * flag, to make booting more robust.
5571 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07005572void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005573{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005574 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005575 unsigned long flags;
5576
Ingo Molnardd41f592007-07-09 18:51:59 +02005577 __sched_fork(idle);
5578 idle->se.exec_start = sched_clock();
5579
Ingo Molnarb29739f2006-06-27 02:54:51 -07005580 idle->prio = idle->normal_prio = MAX_PRIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005581 idle->cpus_allowed = cpumask_of_cpu(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005582 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005583
5584 spin_lock_irqsave(&rq->lock, flags);
5585 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07005586#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5587 idle->oncpu = 1;
5588#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005589 spin_unlock_irqrestore(&rq->lock, flags);
5590
5591 /* Set the preempt count _outside_ the spinlocks! */
Linus Torvalds8e3e0762008-05-10 20:58:02 -07005592#if defined(CONFIG_PREEMPT)
5593 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5594#else
Al Viroa1261f52005-11-13 16:06:55 -08005595 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds8e3e0762008-05-10 20:58:02 -07005596#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02005597 /*
5598 * The idle tasks have their own, simple scheduling class:
5599 */
5600 idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005601}
5602
5603/*
5604 * In a system that switches off the HZ timer nohz_cpu_mask
5605 * indicates which cpus entered this state. This is used
5606 * in the rcu update to wait only for active cpus. For system
5607 * which do not switch off the HZ timer nohz_cpu_mask should
5608 * always be CPU_MASK_NONE.
5609 */
5610cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5611
Ingo Molnar19978ca2007-11-09 22:39:38 +01005612/*
5613 * Increase the granularity value when there are more CPUs,
5614 * because with more CPUs the 'effective latency' as visible
5615 * to users decreases. But the relationship is not linear,
5616 * so pick a second-best guess by going with the log2 of the
5617 * number of CPUs.
5618 *
5619 * This idea comes from the SD scheduler of Con Kolivas:
5620 */
5621static inline void sched_init_granularity(void)
5622{
5623 unsigned int factor = 1 + ilog2(num_online_cpus());
5624 const unsigned long limit = 200000000;
5625
5626 sysctl_sched_min_granularity *= factor;
5627 if (sysctl_sched_min_granularity > limit)
5628 sysctl_sched_min_granularity = limit;
5629
5630 sysctl_sched_latency *= factor;
5631 if (sysctl_sched_latency > limit)
5632 sysctl_sched_latency = limit;
5633
5634 sysctl_sched_wakeup_granularity *= factor;
Ingo Molnar19978ca2007-11-09 22:39:38 +01005635}
5636
Linus Torvalds1da177e2005-04-16 15:20:36 -07005637#ifdef CONFIG_SMP
5638/*
5639 * This is how migration works:
5640 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07005641 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005642 * runqueue and wake up that CPU's migration thread.
5643 * 2) we down() the locked semaphore => thread blocks.
5644 * 3) migration thread wakes up (implicitly it forces the migrated
5645 * thread off the CPU)
5646 * 4) it gets the migration request and checks whether the migrated
5647 * task is still in the wrong runqueue.
5648 * 5) if it's in the wrong runqueue then the migration thread removes
5649 * it and puts it into the right queue.
5650 * 6) migration thread up()s the semaphore.
5651 * 7) we wake up and the migration is done.
5652 */
5653
5654/*
5655 * Change a given task's CPU affinity. Migrate the thread to a
5656 * proper CPU and schedule it away if the CPU it's executing on
5657 * is removed from the allowed bitmask.
5658 *
5659 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005660 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07005661 * call is not atomic; no spinlocks may be held.
5662 */
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005663int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005664{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005665 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005666 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005667 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005668 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005669
5670 rq = task_rq_lock(p, &flags);
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005671 if (!cpus_intersects(*new_mask, cpu_online_map)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005672 ret = -EINVAL;
5673 goto out;
5674 }
5675
David Rientjes9985b0b2008-06-05 12:57:11 -07005676 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5677 !cpus_equal(p->cpus_allowed, *new_mask))) {
5678 ret = -EINVAL;
5679 goto out;
5680 }
5681
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005682 if (p->sched_class->set_cpus_allowed)
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005683 p->sched_class->set_cpus_allowed(p, new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005684 else {
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005685 p->cpus_allowed = *new_mask;
5686 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005687 }
5688
Linus Torvalds1da177e2005-04-16 15:20:36 -07005689 /* Can the task run on the task's current CPU? If so, we're done */
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005690 if (cpu_isset(task_cpu(p), *new_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005691 goto out;
5692
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005693 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005694 /* Need help from migration thread: drop lock and wait. */
5695 task_rq_unlock(rq, &flags);
5696 wake_up_process(rq->migration_thread);
5697 wait_for_completion(&req.done);
5698 tlb_migrate_finish(p->mm);
5699 return 0;
5700 }
5701out:
5702 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005703
Linus Torvalds1da177e2005-04-16 15:20:36 -07005704 return ret;
5705}
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005706EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005707
5708/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005709 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07005710 * this because either it can't run here any more (set_cpus_allowed()
5711 * away from this CPU, or CPU going down), or because we're
5712 * attempting to rebalance this task on exec (sched_exec).
5713 *
5714 * So we race with normal scheduler movements, but that's OK, as long
5715 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07005716 *
5717 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005718 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07005719static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005720{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005721 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02005722 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005723
5724 if (unlikely(cpu_is_offline(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07005725 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005726
5727 rq_src = cpu_rq(src_cpu);
5728 rq_dest = cpu_rq(dest_cpu);
5729
5730 double_rq_lock(rq_src, rq_dest);
5731 /* Already moved. */
5732 if (task_cpu(p) != src_cpu)
5733 goto out;
5734 /* Affinity changed (again). */
5735 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5736 goto out;
5737
Ingo Molnardd41f592007-07-09 18:51:59 +02005738 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02005739 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005740 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02005741
Linus Torvalds1da177e2005-04-16 15:20:36 -07005742 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005743 if (on_rq) {
5744 activate_task(rq_dest, p, 0);
5745 check_preempt_curr(rq_dest, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005746 }
Kirill Korotaevefc30812006-06-27 02:54:32 -07005747 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005748out:
5749 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07005750 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005751}
5752
5753/*
5754 * migration_thread - this is a highprio system thread that performs
5755 * thread migration by bumping thread off CPU then 'pushing' onto
5756 * another runqueue.
5757 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005758static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005759{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005760 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005761 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005762
5763 rq = cpu_rq(cpu);
5764 BUG_ON(rq->migration_thread != current);
5765
5766 set_current_state(TASK_INTERRUPTIBLE);
5767 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005768 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005769 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005770
Linus Torvalds1da177e2005-04-16 15:20:36 -07005771 spin_lock_irq(&rq->lock);
5772
5773 if (cpu_is_offline(cpu)) {
5774 spin_unlock_irq(&rq->lock);
5775 goto wait_to_die;
5776 }
5777
5778 if (rq->active_balance) {
5779 active_load_balance(rq, cpu);
5780 rq->active_balance = 0;
5781 }
5782
5783 head = &rq->migration_queue;
5784
5785 if (list_empty(head)) {
5786 spin_unlock_irq(&rq->lock);
5787 schedule();
5788 set_current_state(TASK_INTERRUPTIBLE);
5789 continue;
5790 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07005791 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005792 list_del_init(head->next);
5793
Nick Piggin674311d2005-06-25 14:57:27 -07005794 spin_unlock(&rq->lock);
5795 __migrate_task(req->task, cpu, req->dest_cpu);
5796 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005797
5798 complete(&req->done);
5799 }
5800 __set_current_state(TASK_RUNNING);
5801 return 0;
5802
5803wait_to_die:
5804 /* Wait for kthread_stop */
5805 set_current_state(TASK_INTERRUPTIBLE);
5806 while (!kthread_should_stop()) {
5807 schedule();
5808 set_current_state(TASK_INTERRUPTIBLE);
5809 }
5810 __set_current_state(TASK_RUNNING);
5811 return 0;
5812}
5813
5814#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005815
5816static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5817{
5818 int ret;
5819
5820 local_irq_disable();
5821 ret = __migrate_task(p, src_cpu, dest_cpu);
5822 local_irq_enable();
5823 return ret;
5824}
5825
Kirill Korotaev054b9102006-12-10 02:20:11 -08005826/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02005827 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08005828 * NOTE: interrupts should be disabled by the caller
5829 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005830static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005831{
Kirill Korotaevefc30812006-06-27 02:54:32 -07005832 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005833 cpumask_t mask;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005834 struct rq *rq;
5835 int dest_cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005836
Andi Kleen3a5c3592007-10-15 17:00:14 +02005837 do {
5838 /* On same node? */
5839 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5840 cpus_and(mask, mask, p->cpus_allowed);
5841 dest_cpu = any_online_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005842
Andi Kleen3a5c3592007-10-15 17:00:14 +02005843 /* On any allowed CPU? */
Mike Travis434d53b2008-04-04 18:11:04 -07005844 if (dest_cpu >= nr_cpu_ids)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005845 dest_cpu = any_online_cpu(p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005846
Andi Kleen3a5c3592007-10-15 17:00:14 +02005847 /* No more Mr. Nice Guy. */
Mike Travis434d53b2008-04-04 18:11:04 -07005848 if (dest_cpu >= nr_cpu_ids) {
Mike Travisf9a86fc2008-04-04 18:11:07 -07005849 cpumask_t cpus_allowed;
5850
5851 cpuset_cpus_allowed_locked(p, &cpus_allowed);
Cliff Wickman470fd642007-10-18 23:40:46 -07005852 /*
5853 * Try to stay on the same cpuset, where the
5854 * current cpuset may be a subset of all cpus.
5855 * The cpuset_cpus_allowed_locked() variant of
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005856 * cpuset_cpus_allowed() will not block. It must be
Cliff Wickman470fd642007-10-18 23:40:46 -07005857 * called within calls to cpuset_lock/cpuset_unlock.
5858 */
Andi Kleen3a5c3592007-10-15 17:00:14 +02005859 rq = task_rq_lock(p, &flags);
Cliff Wickman470fd642007-10-18 23:40:46 -07005860 p->cpus_allowed = cpus_allowed;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005861 dest_cpu = any_online_cpu(p->cpus_allowed);
5862 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005863
Andi Kleen3a5c3592007-10-15 17:00:14 +02005864 /*
5865 * Don't tell them about moving exiting tasks or
5866 * kernel threads (both mm NULL), since they never
5867 * leave kernel.
5868 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005869 if (p->mm && printk_ratelimit()) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02005870 printk(KERN_INFO "process %d (%s) no "
5871 "longer affine to cpu%d\n",
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005872 task_pid_nr(p), p->comm, dead_cpu);
5873 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02005874 }
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005875 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005876}
5877
5878/*
5879 * While a dead CPU has no uninterruptible tasks queued at this point,
5880 * it might still have a nonzero ->nr_uninterruptible counter, because
5881 * for performance reasons the counter is not stricly tracking tasks to
5882 * their home CPUs. So we just add the counter to another CPU's counter,
5883 * to keep the global sum constant after CPU-down:
5884 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07005885static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005886{
Mike Travis7c16ec52008-04-04 18:11:11 -07005887 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005888 unsigned long flags;
5889
5890 local_irq_save(flags);
5891 double_rq_lock(rq_src, rq_dest);
5892 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5893 rq_src->nr_uninterruptible = 0;
5894 double_rq_unlock(rq_src, rq_dest);
5895 local_irq_restore(flags);
5896}
5897
5898/* Run through task list and migrate tasks from the dead cpu. */
5899static void migrate_live_tasks(int src_cpu)
5900{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005901 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005902
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005903 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005904
Ingo Molnar48f24c42006-07-03 00:25:40 -07005905 do_each_thread(t, p) {
5906 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005907 continue;
5908
Ingo Molnar48f24c42006-07-03 00:25:40 -07005909 if (task_cpu(p) == src_cpu)
5910 move_task_off_dead_cpu(src_cpu, p);
5911 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005912
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005913 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005914}
5915
Ingo Molnardd41f592007-07-09 18:51:59 +02005916/*
5917 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01005918 * It does so by boosting its priority to highest possible.
5919 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005920 */
5921void sched_idle_next(void)
5922{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005923 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07005924 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005925 struct task_struct *p = rq->idle;
5926 unsigned long flags;
5927
5928 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005929 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005930
Ingo Molnar48f24c42006-07-03 00:25:40 -07005931 /*
5932 * Strictly not necessary since rest of the CPUs are stopped by now
5933 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005934 */
5935 spin_lock_irqsave(&rq->lock, flags);
5936
Ingo Molnardd41f592007-07-09 18:51:59 +02005937 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005938
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01005939 update_rq_clock(rq);
5940 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005941
5942 spin_unlock_irqrestore(&rq->lock, flags);
5943}
5944
Ingo Molnar48f24c42006-07-03 00:25:40 -07005945/*
5946 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07005947 * offline.
5948 */
5949void idle_task_exit(void)
5950{
5951 struct mm_struct *mm = current->active_mm;
5952
5953 BUG_ON(cpu_online(smp_processor_id()));
5954
5955 if (mm != &init_mm)
5956 switch_mm(mm, &init_mm, current);
5957 mmdrop(mm);
5958}
5959
Kirill Korotaev054b9102006-12-10 02:20:11 -08005960/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005961static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005962{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005963 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005964
5965 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07005966 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005967
5968 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07005969 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005970
Ingo Molnar48f24c42006-07-03 00:25:40 -07005971 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005972
5973 /*
5974 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005975 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005976 * fine.
5977 */
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005978 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005979 move_task_off_dead_cpu(dead_cpu, p);
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005980 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005981
Ingo Molnar48f24c42006-07-03 00:25:40 -07005982 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005983}
5984
5985/* release_task() removes task from tasklist, so we won't find dead tasks. */
5986static void migrate_dead_tasks(unsigned int dead_cpu)
5987{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005988 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005989 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005990
Ingo Molnardd41f592007-07-09 18:51:59 +02005991 for ( ; ; ) {
5992 if (!rq->nr_running)
5993 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02005994 update_rq_clock(rq);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02005995 next = pick_next_task(rq, rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02005996 if (!next)
5997 break;
5998 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02005999
Linus Torvalds1da177e2005-04-16 15:20:36 -07006000 }
6001}
6002#endif /* CONFIG_HOTPLUG_CPU */
6003
Nick Piggine692ab52007-07-26 13:40:43 +02006004#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6005
6006static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02006007 {
6008 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006009 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02006010 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01006011 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02006012};
6013
6014static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02006015 {
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006016 .ctl_name = CTL_KERN,
Alexey Dobriyane0361852007-08-09 11:16:46 +02006017 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006018 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02006019 .child = sd_ctl_dir,
6020 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01006021 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02006022};
6023
6024static struct ctl_table *sd_alloc_ctl_entry(int n)
6025{
6026 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02006027 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02006028
Nick Piggine692ab52007-07-26 13:40:43 +02006029 return entry;
6030}
6031
Milton Miller6382bc92007-10-15 17:00:19 +02006032static void sd_free_ctl_entry(struct ctl_table **tablep)
6033{
Milton Millercd7900762007-10-17 16:55:11 +02006034 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02006035
Milton Millercd7900762007-10-17 16:55:11 +02006036 /*
6037 * In the intermediate directories, both the child directory and
6038 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006039 * will always be set. In the lowest directory the names are
Milton Millercd7900762007-10-17 16:55:11 +02006040 * static strings and all have proc handlers.
6041 */
6042 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02006043 if (entry->child)
6044 sd_free_ctl_entry(&entry->child);
Milton Millercd7900762007-10-17 16:55:11 +02006045 if (entry->proc_handler == NULL)
6046 kfree(entry->procname);
6047 }
Milton Miller6382bc92007-10-15 17:00:19 +02006048
6049 kfree(*tablep);
6050 *tablep = NULL;
6051}
6052
Nick Piggine692ab52007-07-26 13:40:43 +02006053static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02006054set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02006055 const char *procname, void *data, int maxlen,
6056 mode_t mode, proc_handler *proc_handler)
6057{
Nick Piggine692ab52007-07-26 13:40:43 +02006058 entry->procname = procname;
6059 entry->data = data;
6060 entry->maxlen = maxlen;
6061 entry->mode = mode;
6062 entry->proc_handler = proc_handler;
6063}
6064
6065static struct ctl_table *
6066sd_alloc_ctl_domain_table(struct sched_domain *sd)
6067{
Zou Nan haiace8b3d2007-10-15 17:00:14 +02006068 struct ctl_table *table = sd_alloc_ctl_entry(12);
Nick Piggine692ab52007-07-26 13:40:43 +02006069
Milton Millerad1cdc12007-10-15 17:00:19 +02006070 if (table == NULL)
6071 return NULL;
6072
Alexey Dobriyane0361852007-08-09 11:16:46 +02006073 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02006074 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006075 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02006076 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006077 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006078 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006079 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006080 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006081 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006082 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006083 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006084 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006085 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006086 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006087 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02006088 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006089 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02006090 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02006091 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02006092 &sd->cache_nice_tries,
6093 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02006094 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02006095 sizeof(int), 0644, proc_dointvec_minmax);
Milton Miller6323469f2007-10-15 17:00:19 +02006096 /* &table[11] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02006097
6098 return table;
6099}
6100
Ingo Molnar9a4e7152007-11-28 15:52:56 +01006101static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02006102{
6103 struct ctl_table *entry, *table;
6104 struct sched_domain *sd;
6105 int domain_num = 0, i;
6106 char buf[32];
6107
6108 for_each_domain(cpu, sd)
6109 domain_num++;
6110 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02006111 if (table == NULL)
6112 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02006113
6114 i = 0;
6115 for_each_domain(cpu, sd) {
6116 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02006117 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006118 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02006119 entry->child = sd_alloc_ctl_domain_table(sd);
6120 entry++;
6121 i++;
6122 }
6123 return table;
6124}
6125
6126static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02006127static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02006128{
6129 int i, cpu_num = num_online_cpus();
6130 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6131 char buf[32];
6132
Milton Miller73785472007-10-24 18:23:48 +02006133 WARN_ON(sd_ctl_dir[0].child);
6134 sd_ctl_dir[0].child = entry;
6135
Milton Millerad1cdc12007-10-15 17:00:19 +02006136 if (entry == NULL)
6137 return;
6138
Milton Miller97b6ea72007-10-15 17:00:19 +02006139 for_each_online_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02006140 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02006141 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006142 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02006143 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02006144 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02006145 }
Milton Miller73785472007-10-24 18:23:48 +02006146
6147 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02006148 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6149}
Milton Miller6382bc92007-10-15 17:00:19 +02006150
Milton Miller73785472007-10-24 18:23:48 +02006151/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02006152static void unregister_sched_domain_sysctl(void)
6153{
Milton Miller73785472007-10-24 18:23:48 +02006154 if (sd_sysctl_header)
6155 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02006156 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02006157 if (sd_ctl_dir[0].child)
6158 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02006159}
Nick Piggine692ab52007-07-26 13:40:43 +02006160#else
Milton Miller6382bc92007-10-15 17:00:19 +02006161static void register_sched_domain_sysctl(void)
6162{
6163}
6164static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02006165{
6166}
6167#endif
6168
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006169static void set_rq_online(struct rq *rq)
6170{
6171 if (!rq->online) {
6172 const struct sched_class *class;
6173
6174 cpu_set(rq->cpu, rq->rd->online);
6175 rq->online = 1;
6176
6177 for_each_class(class) {
6178 if (class->rq_online)
6179 class->rq_online(rq);
6180 }
6181 }
6182}
6183
6184static void set_rq_offline(struct rq *rq)
6185{
6186 if (rq->online) {
6187 const struct sched_class *class;
6188
6189 for_each_class(class) {
6190 if (class->rq_offline)
6191 class->rq_offline(rq);
6192 }
6193
6194 cpu_clear(rq->cpu, rq->rd->online);
6195 rq->online = 0;
6196 }
6197}
6198
Linus Torvalds1da177e2005-04-16 15:20:36 -07006199/*
6200 * migration_call - callback that gets triggered when a CPU is added.
6201 * Here we can start up the necessary migration thread for the new CPU.
6202 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006203static int __cpuinit
6204migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006205{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006206 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006207 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006208 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006209 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006210
6211 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006212
Linus Torvalds1da177e2005-04-16 15:20:36 -07006213 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006214 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02006215 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006216 if (IS_ERR(p))
6217 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006218 kthread_bind(p, cpu);
6219 /* Must be high prio: stop_machine expects to yield to it. */
6220 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02006221 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006222 task_rq_unlock(rq, &flags);
6223 cpu_rq(cpu)->migration_thread = p;
6224 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006225
Linus Torvalds1da177e2005-04-16 15:20:36 -07006226 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006227 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02006228 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006229 wake_up_process(cpu_rq(cpu)->migration_thread);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006230
6231 /* Update our root-domain */
6232 rq = cpu_rq(cpu);
6233 spin_lock_irqsave(&rq->lock, flags);
6234 if (rq->rd) {
6235 BUG_ON(!cpu_isset(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006236
6237 set_rq_online(rq);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006238 }
6239 spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006240 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006241
Linus Torvalds1da177e2005-04-16 15:20:36 -07006242#ifdef CONFIG_HOTPLUG_CPU
6243 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006244 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07006245 if (!cpu_rq(cpu)->migration_thread)
6246 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006247 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08006248 kthread_bind(cpu_rq(cpu)->migration_thread,
6249 any_online_cpu(cpu_online_map));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006250 kthread_stop(cpu_rq(cpu)->migration_thread);
6251 cpu_rq(cpu)->migration_thread = NULL;
6252 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006253
Linus Torvalds1da177e2005-04-16 15:20:36 -07006254 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006255 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07006256 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006257 migrate_live_tasks(cpu);
6258 rq = cpu_rq(cpu);
6259 kthread_stop(rq->migration_thread);
6260 rq->migration_thread = NULL;
6261 /* Idle task back to normal (off runqueue, low prio) */
Oleg Nesterovd2da2722007-10-16 23:30:56 -07006262 spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02006263 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02006264 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006265 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02006266 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6267 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006268 migrate_dead_tasks(cpu);
Oleg Nesterovd2da2722007-10-16 23:30:56 -07006269 spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07006270 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006271 migrate_nr_uninterruptible(rq);
6272 BUG_ON(rq->nr_running != 0);
6273
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006274 /*
6275 * No need to migrate the tasks: it was best-effort if
6276 * they didn't take sched_hotcpu_mutex. Just wake up
6277 * the requestors.
6278 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006279 spin_lock_irq(&rq->lock);
6280 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07006281 struct migration_req *req;
6282
Linus Torvalds1da177e2005-04-16 15:20:36 -07006283 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07006284 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006285 list_del_init(&req->list);
6286 complete(&req->done);
6287 }
6288 spin_unlock_irq(&rq->lock);
6289 break;
Gregory Haskins57d885f2008-01-25 21:08:18 +01006290
Gregory Haskins08f503b2008-03-10 17:59:11 -04006291 case CPU_DYING:
6292 case CPU_DYING_FROZEN:
Gregory Haskins57d885f2008-01-25 21:08:18 +01006293 /* Update our root-domain */
6294 rq = cpu_rq(cpu);
6295 spin_lock_irqsave(&rq->lock, flags);
6296 if (rq->rd) {
6297 BUG_ON(!cpu_isset(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006298 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006299 }
6300 spin_unlock_irqrestore(&rq->lock, flags);
6301 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006302#endif
6303 }
6304 return NOTIFY_OK;
6305}
6306
6307/* Register at highest priority so that task migration (migrate_all_tasks)
6308 * happens before everything else.
6309 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07006310static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006311 .notifier_call = migration_call,
6312 .priority = 10
6313};
6314
Adrian Bunke6fe6642007-11-09 22:39:39 +01006315void __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006316{
6317 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07006318 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006319
6320 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07006321 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6322 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006323 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6324 register_cpu_notifier(&migration_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006325}
6326#endif
6327
6328#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07006329
Ingo Molnar3e9830d2007-10-15 17:00:13 +02006330#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006331
Gautham R Shenoy099f98c2008-05-29 20:56:32 +05306332static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6333{
6334 switch (lvl) {
6335 case SD_LV_NONE:
6336 return "NONE";
6337 case SD_LV_SIBLING:
6338 return "SIBLING";
6339 case SD_LV_MC:
6340 return "MC";
6341 case SD_LV_CPU:
6342 return "CPU";
6343 case SD_LV_NODE:
6344 return "NODE";
6345 case SD_LV_ALLNODES:
6346 return "ALLNODES";
6347 case SD_LV_MAX:
6348 return "MAX";
6349
6350 }
6351 return "MAX";
6352}
6353
Mike Travis7c16ec52008-04-04 18:11:11 -07006354static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6355 cpumask_t *groupmask)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006356{
6357 struct sched_group *group = sd->groups;
Mike Travis434d53b2008-04-04 18:11:04 -07006358 char str[256];
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006359
Mike Travis434d53b2008-04-04 18:11:04 -07006360 cpulist_scnprintf(str, sizeof(str), sd->span);
Mike Travis7c16ec52008-04-04 18:11:11 -07006361 cpus_clear(*groupmask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006362
6363 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6364
6365 if (!(sd->flags & SD_LOAD_BALANCE)) {
6366 printk("does not load-balance\n");
6367 if (sd->parent)
6368 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6369 " has parent");
6370 return -1;
6371 }
6372
Gautham R Shenoy099f98c2008-05-29 20:56:32 +05306373 printk(KERN_CONT "span %s level %s\n",
6374 str, sd_level_to_string(sd->level));
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006375
6376 if (!cpu_isset(cpu, sd->span)) {
6377 printk(KERN_ERR "ERROR: domain->span does not contain "
6378 "CPU%d\n", cpu);
6379 }
6380 if (!cpu_isset(cpu, group->cpumask)) {
6381 printk(KERN_ERR "ERROR: domain->groups does not contain"
6382 " CPU%d\n", cpu);
6383 }
6384
6385 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6386 do {
6387 if (!group) {
6388 printk("\n");
6389 printk(KERN_ERR "ERROR: group is NULL\n");
6390 break;
6391 }
6392
6393 if (!group->__cpu_power) {
6394 printk(KERN_CONT "\n");
6395 printk(KERN_ERR "ERROR: domain->cpu_power not "
6396 "set\n");
6397 break;
6398 }
6399
6400 if (!cpus_weight(group->cpumask)) {
6401 printk(KERN_CONT "\n");
6402 printk(KERN_ERR "ERROR: empty group\n");
6403 break;
6404 }
6405
Mike Travis7c16ec52008-04-04 18:11:11 -07006406 if (cpus_intersects(*groupmask, group->cpumask)) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006407 printk(KERN_CONT "\n");
6408 printk(KERN_ERR "ERROR: repeated CPUs\n");
6409 break;
6410 }
6411
Mike Travis7c16ec52008-04-04 18:11:11 -07006412 cpus_or(*groupmask, *groupmask, group->cpumask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006413
Mike Travis434d53b2008-04-04 18:11:04 -07006414 cpulist_scnprintf(str, sizeof(str), group->cpumask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006415 printk(KERN_CONT " %s", str);
6416
6417 group = group->next;
6418 } while (group != sd->groups);
6419 printk(KERN_CONT "\n");
6420
Mike Travis7c16ec52008-04-04 18:11:11 -07006421 if (!cpus_equal(sd->span, *groupmask))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006422 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6423
Mike Travis7c16ec52008-04-04 18:11:11 -07006424 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006425 printk(KERN_ERR "ERROR: parent span is not a superset "
6426 "of domain->span\n");
6427 return 0;
6428}
6429
Linus Torvalds1da177e2005-04-16 15:20:36 -07006430static void sched_domain_debug(struct sched_domain *sd, int cpu)
6431{
Mike Travis7c16ec52008-04-04 18:11:11 -07006432 cpumask_t *groupmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006433 int level = 0;
6434
Nick Piggin41c7ce92005-06-25 14:57:24 -07006435 if (!sd) {
6436 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6437 return;
6438 }
6439
Linus Torvalds1da177e2005-04-16 15:20:36 -07006440 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6441
Mike Travis7c16ec52008-04-04 18:11:11 -07006442 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6443 if (!groupmask) {
6444 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6445 return;
6446 }
6447
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006448 for (;;) {
Mike Travis7c16ec52008-04-04 18:11:11 -07006449 if (sched_domain_debug_one(sd, cpu, level, groupmask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006450 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006451 level++;
6452 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08006453 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006454 break;
6455 }
Mike Travis7c16ec52008-04-04 18:11:11 -07006456 kfree(groupmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006457}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006458#else /* !CONFIG_SCHED_DEBUG */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006459# define sched_domain_debug(sd, cpu) do { } while (0)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006460#endif /* CONFIG_SCHED_DEBUG */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006461
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006462static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006463{
6464 if (cpus_weight(sd->span) == 1)
6465 return 1;
6466
6467 /* Following flags need at least 2 groups */
6468 if (sd->flags & (SD_LOAD_BALANCE |
6469 SD_BALANCE_NEWIDLE |
6470 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006471 SD_BALANCE_EXEC |
6472 SD_SHARE_CPUPOWER |
6473 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006474 if (sd->groups != sd->groups->next)
6475 return 0;
6476 }
6477
6478 /* Following flags don't use groups */
6479 if (sd->flags & (SD_WAKE_IDLE |
6480 SD_WAKE_AFFINE |
6481 SD_WAKE_BALANCE))
6482 return 0;
6483
6484 return 1;
6485}
6486
Ingo Molnar48f24c42006-07-03 00:25:40 -07006487static int
6488sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006489{
6490 unsigned long cflags = sd->flags, pflags = parent->flags;
6491
6492 if (sd_degenerate(parent))
6493 return 1;
6494
6495 if (!cpus_equal(sd->span, parent->span))
6496 return 0;
6497
6498 /* Does parent contain flags not in child? */
6499 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6500 if (cflags & SD_WAKE_AFFINE)
6501 pflags &= ~SD_WAKE_BALANCE;
6502 /* Flags needing groups don't count if only 1 group in parent */
6503 if (parent->groups == parent->groups->next) {
6504 pflags &= ~(SD_LOAD_BALANCE |
6505 SD_BALANCE_NEWIDLE |
6506 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006507 SD_BALANCE_EXEC |
6508 SD_SHARE_CPUPOWER |
6509 SD_SHARE_PKG_RESOURCES);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006510 }
6511 if (~cflags & pflags)
6512 return 0;
6513
6514 return 1;
6515}
6516
Gregory Haskins57d885f2008-01-25 21:08:18 +01006517static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6518{
6519 unsigned long flags;
Gregory Haskins57d885f2008-01-25 21:08:18 +01006520
6521 spin_lock_irqsave(&rq->lock, flags);
6522
6523 if (rq->rd) {
6524 struct root_domain *old_rd = rq->rd;
6525
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006526 if (cpu_isset(rq->cpu, old_rd->online))
6527 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006528
Gregory Haskinsdc938522008-01-25 21:08:26 +01006529 cpu_clear(rq->cpu, old_rd->span);
Gregory Haskinsdc938522008-01-25 21:08:26 +01006530
Gregory Haskins57d885f2008-01-25 21:08:18 +01006531 if (atomic_dec_and_test(&old_rd->refcount))
6532 kfree(old_rd);
6533 }
6534
6535 atomic_inc(&rd->refcount);
6536 rq->rd = rd;
6537
Gregory Haskinsdc938522008-01-25 21:08:26 +01006538 cpu_set(rq->cpu, rd->span);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006539 if (cpu_isset(rq->cpu, cpu_online_map))
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006540 set_rq_online(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006541
6542 spin_unlock_irqrestore(&rq->lock, flags);
6543}
6544
Gregory Haskinsdc938522008-01-25 21:08:26 +01006545static void init_rootdomain(struct root_domain *rd)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006546{
6547 memset(rd, 0, sizeof(*rd));
6548
Gregory Haskinsdc938522008-01-25 21:08:26 +01006549 cpus_clear(rd->span);
6550 cpus_clear(rd->online);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02006551
6552 cpupri_init(&rd->cpupri);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006553}
6554
6555static void init_defrootdomain(void)
6556{
Gregory Haskinsdc938522008-01-25 21:08:26 +01006557 init_rootdomain(&def_root_domain);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006558 atomic_set(&def_root_domain.refcount, 1);
6559}
6560
Gregory Haskinsdc938522008-01-25 21:08:26 +01006561static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006562{
6563 struct root_domain *rd;
6564
6565 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6566 if (!rd)
6567 return NULL;
6568
Gregory Haskinsdc938522008-01-25 21:08:26 +01006569 init_rootdomain(rd);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006570
6571 return rd;
6572}
6573
Linus Torvalds1da177e2005-04-16 15:20:36 -07006574/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01006575 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07006576 * hold the hotplug lock.
6577 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01006578static void
6579cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006580{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006581 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006582 struct sched_domain *tmp;
6583
6584 /* Remove the sched domains which do not contribute to scheduling. */
6585 for (tmp = sd; tmp; tmp = tmp->parent) {
6586 struct sched_domain *parent = tmp->parent;
6587 if (!parent)
6588 break;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006589 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006590 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006591 if (parent->parent)
6592 parent->parent->child = tmp;
6593 }
Suresh Siddha245af2c2005-06-25 14:57:25 -07006594 }
6595
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006596 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006597 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006598 if (sd)
6599 sd->child = NULL;
6600 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006601
6602 sched_domain_debug(sd, cpu);
6603
Gregory Haskins57d885f2008-01-25 21:08:18 +01006604 rq_attach_root(rq, rd);
Nick Piggin674311d2005-06-25 14:57:27 -07006605 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006606}
6607
6608/* cpus with isolated domains */
Tim Chen67af63a2006-12-22 01:07:50 -08006609static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006610
6611/* Setup the mask of cpus configured for isolated domains */
6612static int __init isolated_cpu_setup(char *str)
6613{
6614 int ints[NR_CPUS], i;
6615
6616 str = get_options(str, ARRAY_SIZE(ints), ints);
6617 cpus_clear(cpu_isolated_map);
6618 for (i = 1; i <= ints[0]; i++)
6619 if (ints[i] < NR_CPUS)
6620 cpu_set(ints[i], cpu_isolated_map);
6621 return 1;
6622}
6623
Ingo Molnar8927f492007-10-15 17:00:13 +02006624__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006625
6626/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006627 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6628 * to a function which identifies what group(along with sched group) a CPU
6629 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6630 * (due to the fact that we keep track of groups covered with a cpumask_t).
Linus Torvalds1da177e2005-04-16 15:20:36 -07006631 *
6632 * init_sched_build_groups will build a circular linked list of the groups
6633 * covered by the given span, and will set each group's ->cpumask correctly,
6634 * and ->cpu_power to 0.
6635 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006636static void
Mike Travis7c16ec52008-04-04 18:11:11 -07006637init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006638 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07006639 struct sched_group **sg,
6640 cpumask_t *tmpmask),
6641 cpumask_t *covered, cpumask_t *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006642{
6643 struct sched_group *first = NULL, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006644 int i;
6645
Mike Travis7c16ec52008-04-04 18:11:11 -07006646 cpus_clear(*covered);
6647
6648 for_each_cpu_mask(i, *span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006649 struct sched_group *sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07006650 int group = group_fn(i, cpu_map, &sg, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006651 int j;
6652
Mike Travis7c16ec52008-04-04 18:11:11 -07006653 if (cpu_isset(i, *covered))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006654 continue;
6655
Mike Travis7c16ec52008-04-04 18:11:11 -07006656 cpus_clear(sg->cpumask);
Eric Dumazet5517d862007-05-08 00:32:57 -07006657 sg->__cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006658
Mike Travis7c16ec52008-04-04 18:11:11 -07006659 for_each_cpu_mask(j, *span) {
6660 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006661 continue;
6662
Mike Travis7c16ec52008-04-04 18:11:11 -07006663 cpu_set(j, *covered);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006664 cpu_set(j, sg->cpumask);
6665 }
6666 if (!first)
6667 first = sg;
6668 if (last)
6669 last->next = sg;
6670 last = sg;
6671 }
6672 last->next = first;
6673}
6674
John Hawkes9c1cfda2005-09-06 15:18:14 -07006675#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07006676
John Hawkes9c1cfda2005-09-06 15:18:14 -07006677#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08006678
John Hawkes9c1cfda2005-09-06 15:18:14 -07006679/**
6680 * find_next_best_node - find the next node to include in a sched_domain
6681 * @node: node whose sched_domain we're building
6682 * @used_nodes: nodes already in the sched_domain
6683 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006684 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07006685 * finds the closest node not already in the @used_nodes map.
6686 *
6687 * Should use nodemask_t.
6688 */
Mike Travisc5f59f02008-04-04 18:11:10 -07006689static int find_next_best_node(int node, nodemask_t *used_nodes)
John Hawkes9c1cfda2005-09-06 15:18:14 -07006690{
6691 int i, n, val, min_val, best_node = 0;
6692
6693 min_val = INT_MAX;
6694
6695 for (i = 0; i < MAX_NUMNODES; i++) {
6696 /* Start at @node */
6697 n = (node + i) % MAX_NUMNODES;
6698
6699 if (!nr_cpus_node(n))
6700 continue;
6701
6702 /* Skip already used nodes */
Mike Travisc5f59f02008-04-04 18:11:10 -07006703 if (node_isset(n, *used_nodes))
John Hawkes9c1cfda2005-09-06 15:18:14 -07006704 continue;
6705
6706 /* Simple min distance search */
6707 val = node_distance(node, n);
6708
6709 if (val < min_val) {
6710 min_val = val;
6711 best_node = n;
6712 }
6713 }
6714
Mike Travisc5f59f02008-04-04 18:11:10 -07006715 node_set(best_node, *used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006716 return best_node;
6717}
6718
6719/**
6720 * sched_domain_node_span - get a cpumask for a node's sched_domain
6721 * @node: node whose cpumask we're constructing
Randy Dunlap73486722008-04-22 10:07:22 -07006722 * @span: resulting cpumask
John Hawkes9c1cfda2005-09-06 15:18:14 -07006723 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006724 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07006725 * should be one that prevents unnecessary balancing, but also spreads tasks
6726 * out optimally.
6727 */
Mike Travis4bdbaad32008-04-15 16:35:52 -07006728static void sched_domain_node_span(int node, cpumask_t *span)
John Hawkes9c1cfda2005-09-06 15:18:14 -07006729{
Mike Travisc5f59f02008-04-04 18:11:10 -07006730 nodemask_t used_nodes;
Mike Travisc5f59f02008-04-04 18:11:10 -07006731 node_to_cpumask_ptr(nodemask, node);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006732 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006733
Mike Travis4bdbaad32008-04-15 16:35:52 -07006734 cpus_clear(*span);
Mike Travisc5f59f02008-04-04 18:11:10 -07006735 nodes_clear(used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006736
Mike Travis4bdbaad32008-04-15 16:35:52 -07006737 cpus_or(*span, *span, *nodemask);
Mike Travisc5f59f02008-04-04 18:11:10 -07006738 node_set(node, used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006739
6740 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
Mike Travisc5f59f02008-04-04 18:11:10 -07006741 int next_node = find_next_best_node(node, &used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006742
Mike Travisc5f59f02008-04-04 18:11:10 -07006743 node_to_cpumask_ptr_next(nodemask, next_node);
Mike Travis4bdbaad32008-04-15 16:35:52 -07006744 cpus_or(*span, *span, *nodemask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006745 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006746}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006747#endif /* CONFIG_NUMA */
John Hawkes9c1cfda2005-09-06 15:18:14 -07006748
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006749int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006750
John Hawkes9c1cfda2005-09-06 15:18:14 -07006751/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07006752 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07006753 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006754#ifdef CONFIG_SCHED_SMT
6755static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006756static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006757
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006758static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006759cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6760 cpumask_t *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006761{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006762 if (sg)
6763 *sg = &per_cpu(sched_group_cpus, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006764 return cpu;
6765}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006766#endif /* CONFIG_SCHED_SMT */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006767
Ingo Molnar48f24c42006-07-03 00:25:40 -07006768/*
6769 * multi-core sched-domains:
6770 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006771#ifdef CONFIG_SCHED_MC
6772static DEFINE_PER_CPU(struct sched_domain, core_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006773static DEFINE_PER_CPU(struct sched_group, sched_group_core);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006774#endif /* CONFIG_SCHED_MC */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006775
6776#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006777static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006778cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6779 cpumask_t *mask)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006780{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006781 int group;
Mike Travis7c16ec52008-04-04 18:11:11 -07006782
6783 *mask = per_cpu(cpu_sibling_map, cpu);
6784 cpus_and(*mask, *mask, *cpu_map);
6785 group = first_cpu(*mask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006786 if (sg)
6787 *sg = &per_cpu(sched_group_core, group);
6788 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006789}
6790#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006791static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006792cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6793 cpumask_t *unused)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006794{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006795 if (sg)
6796 *sg = &per_cpu(sched_group_core, cpu);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006797 return cpu;
6798}
6799#endif
6800
Linus Torvalds1da177e2005-04-16 15:20:36 -07006801static DEFINE_PER_CPU(struct sched_domain, phys_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006802static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006803
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006804static int
Mike Travis7c16ec52008-04-04 18:11:11 -07006805cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6806 cpumask_t *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006807{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006808 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006809#ifdef CONFIG_SCHED_MC
Mike Travis7c16ec52008-04-04 18:11:11 -07006810 *mask = cpu_coregroup_map(cpu);
6811 cpus_and(*mask, *mask, *cpu_map);
6812 group = first_cpu(*mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006813#elif defined(CONFIG_SCHED_SMT)
Mike Travis7c16ec52008-04-04 18:11:11 -07006814 *mask = per_cpu(cpu_sibling_map, cpu);
6815 cpus_and(*mask, *mask, *cpu_map);
6816 group = first_cpu(*mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006817#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006818 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006819#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006820 if (sg)
6821 *sg = &per_cpu(sched_group_phys, group);
6822 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006823}
6824
6825#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07006826/*
6827 * The init_sched_build_groups can't handle what we want to do with node
6828 * groups, so roll our own. Now each node has its own list of groups which
6829 * gets dynamically allocated.
6830 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006831static DEFINE_PER_CPU(struct sched_domain, node_domains);
Mike Travis434d53b2008-04-04 18:11:04 -07006832static struct sched_group ***sched_group_nodes_bycpu;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006833
6834static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006835static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006836
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006837static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07006838 struct sched_group **sg, cpumask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006839{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006840 int group;
6841
Mike Travis7c16ec52008-04-04 18:11:11 -07006842 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6843 cpus_and(*nodemask, *nodemask, *cpu_map);
6844 group = first_cpu(*nodemask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006845
6846 if (sg)
6847 *sg = &per_cpu(sched_group_allnodes, group);
6848 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006849}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006850
Siddha, Suresh B08069032006-03-27 01:15:23 -08006851static void init_numa_sched_groups_power(struct sched_group *group_head)
6852{
6853 struct sched_group *sg = group_head;
6854 int j;
6855
6856 if (!sg)
6857 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006858 do {
6859 for_each_cpu_mask(j, sg->cpumask) {
6860 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08006861
Andi Kleen3a5c3592007-10-15 17:00:14 +02006862 sd = &per_cpu(phys_domains, j);
6863 if (j != first_cpu(sd->groups->cpumask)) {
6864 /*
6865 * Only add "power" once for each
6866 * physical package.
6867 */
6868 continue;
6869 }
6870
6871 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
Siddha, Suresh B08069032006-03-27 01:15:23 -08006872 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02006873 sg = sg->next;
6874 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08006875}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006876#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006877
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006878#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006879/* Free memory allocated for various sched_group structures */
Mike Travis7c16ec52008-04-04 18:11:11 -07006880static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006881{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006882 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006883
6884 for_each_cpu_mask(cpu, *cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006885 struct sched_group **sched_group_nodes
6886 = sched_group_nodes_bycpu[cpu];
6887
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006888 if (!sched_group_nodes)
6889 continue;
6890
6891 for (i = 0; i < MAX_NUMNODES; i++) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006892 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6893
Mike Travis7c16ec52008-04-04 18:11:11 -07006894 *nodemask = node_to_cpumask(i);
6895 cpus_and(*nodemask, *nodemask, *cpu_map);
6896 if (cpus_empty(*nodemask))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006897 continue;
6898
6899 if (sg == NULL)
6900 continue;
6901 sg = sg->next;
6902next_sg:
6903 oldsg = sg;
6904 sg = sg->next;
6905 kfree(oldsg);
6906 if (oldsg != sched_group_nodes[i])
6907 goto next_sg;
6908 }
6909 kfree(sched_group_nodes);
6910 sched_group_nodes_bycpu[cpu] = NULL;
6911 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006912}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006913#else /* !CONFIG_NUMA */
Mike Travis7c16ec52008-04-04 18:11:11 -07006914static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006915{
6916}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006917#endif /* CONFIG_NUMA */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006918
Linus Torvalds1da177e2005-04-16 15:20:36 -07006919/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006920 * Initialize sched groups cpu_power.
6921 *
6922 * cpu_power indicates the capacity of sched group, which is used while
6923 * distributing the load between different sched groups in a sched domain.
6924 * Typically cpu_power for all the groups in a sched domain will be same unless
6925 * there are asymmetries in the topology. If there are asymmetries, group
6926 * having more cpu_power will pickup more load compared to the group having
6927 * less cpu_power.
6928 *
6929 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6930 * the maximum number of tasks a group can handle in the presence of other idle
6931 * or lightly loaded groups in the same sched domain.
6932 */
6933static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6934{
6935 struct sched_domain *child;
6936 struct sched_group *group;
6937
6938 WARN_ON(!sd || !sd->groups);
6939
6940 if (cpu != first_cpu(sd->groups->cpumask))
6941 return;
6942
6943 child = sd->child;
6944
Eric Dumazet5517d862007-05-08 00:32:57 -07006945 sd->groups->__cpu_power = 0;
6946
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006947 /*
6948 * For perf policy, if the groups in child domain share resources
6949 * (for example cores sharing some portions of the cache hierarchy
6950 * or SMT), then set this domain groups cpu_power such that each group
6951 * can handle only one task, when there are other idle groups in the
6952 * same sched domain.
6953 */
6954 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6955 (child->flags &
6956 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
Eric Dumazet5517d862007-05-08 00:32:57 -07006957 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006958 return;
6959 }
6960
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006961 /*
6962 * add cpu_power of each child group to this groups cpu_power
6963 */
6964 group = child->groups;
6965 do {
Eric Dumazet5517d862007-05-08 00:32:57 -07006966 sg_inc_cpu_power(sd->groups, group->__cpu_power);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006967 group = group->next;
6968 } while (group != child->groups);
6969}
6970
6971/*
Mike Travis7c16ec52008-04-04 18:11:11 -07006972 * Initializers for schedule domains
6973 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6974 */
6975
6976#define SD_INIT(sd, type) sd_init_##type(sd)
6977#define SD_INIT_FUNC(type) \
6978static noinline void sd_init_##type(struct sched_domain *sd) \
6979{ \
6980 memset(sd, 0, sizeof(*sd)); \
6981 *sd = SD_##type##_INIT; \
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09006982 sd->level = SD_LV_##type; \
Mike Travis7c16ec52008-04-04 18:11:11 -07006983}
6984
6985SD_INIT_FUNC(CPU)
6986#ifdef CONFIG_NUMA
6987 SD_INIT_FUNC(ALLNODES)
6988 SD_INIT_FUNC(NODE)
6989#endif
6990#ifdef CONFIG_SCHED_SMT
6991 SD_INIT_FUNC(SIBLING)
6992#endif
6993#ifdef CONFIG_SCHED_MC
6994 SD_INIT_FUNC(MC)
6995#endif
6996
6997/*
6998 * To minimize stack usage kmalloc room for cpumasks and share the
6999 * space as the usage in build_sched_domains() dictates. Used only
7000 * if the amount of space is significant.
7001 */
7002struct allmasks {
7003 cpumask_t tmpmask; /* make this one first */
7004 union {
7005 cpumask_t nodemask;
7006 cpumask_t this_sibling_map;
7007 cpumask_t this_core_map;
7008 };
7009 cpumask_t send_covered;
7010
7011#ifdef CONFIG_NUMA
7012 cpumask_t domainspan;
7013 cpumask_t covered;
7014 cpumask_t notcovered;
7015#endif
7016};
7017
7018#if NR_CPUS > 128
7019#define SCHED_CPUMASK_ALLOC 1
7020#define SCHED_CPUMASK_FREE(v) kfree(v)
7021#define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7022#else
7023#define SCHED_CPUMASK_ALLOC 0
7024#define SCHED_CPUMASK_FREE(v)
7025#define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7026#endif
7027
7028#define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7029 ((unsigned long)(a) + offsetof(struct allmasks, v))
7030
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007031static int default_relax_domain_level = -1;
7032
7033static int __init setup_relax_domain_level(char *str)
7034{
Li Zefan30e0e172008-05-13 10:27:17 +08007035 unsigned long val;
7036
7037 val = simple_strtoul(str, NULL, 0);
7038 if (val < SD_LV_MAX)
7039 default_relax_domain_level = val;
7040
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007041 return 1;
7042}
7043__setup("relax_domain_level=", setup_relax_domain_level);
7044
7045static void set_domain_attribute(struct sched_domain *sd,
7046 struct sched_domain_attr *attr)
7047{
7048 int request;
7049
7050 if (!attr || attr->relax_domain_level < 0) {
7051 if (default_relax_domain_level < 0)
7052 return;
7053 else
7054 request = default_relax_domain_level;
7055 } else
7056 request = attr->relax_domain_level;
7057 if (request < sd->level) {
7058 /* turn off idle balance on this domain */
7059 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7060 } else {
7061 /* turn on idle balance on this domain */
7062 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7063 }
7064}
7065
Mike Travis7c16ec52008-04-04 18:11:11 -07007066/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007067 * Build sched domains for a given set of cpus and attach the sched domains
7068 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07007069 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007070static int __build_sched_domains(const cpumask_t *cpu_map,
7071 struct sched_domain_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007072{
7073 int i;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007074 struct root_domain *rd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007075 SCHED_CPUMASK_DECLARE(allmasks);
7076 cpumask_t *tmpmask;
John Hawkesd1b55132005-09-06 15:18:14 -07007077#ifdef CONFIG_NUMA
7078 struct sched_group **sched_group_nodes = NULL;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007079 int sd_allnodes = 0;
John Hawkesd1b55132005-09-06 15:18:14 -07007080
7081 /*
7082 * Allocate the per-node list of sched groups
7083 */
Milton Miller5cf9f062007-10-15 17:00:19 +02007084 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007085 GFP_KERNEL);
John Hawkesd1b55132005-09-06 15:18:14 -07007086 if (!sched_group_nodes) {
7087 printk(KERN_WARNING "Can not alloc sched group node list\n");
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007088 return -ENOMEM;
John Hawkesd1b55132005-09-06 15:18:14 -07007089 }
John Hawkesd1b55132005-09-06 15:18:14 -07007090#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007091
Gregory Haskinsdc938522008-01-25 21:08:26 +01007092 rd = alloc_rootdomain();
Gregory Haskins57d885f2008-01-25 21:08:18 +01007093 if (!rd) {
7094 printk(KERN_WARNING "Cannot alloc root domain\n");
Mike Travis7c16ec52008-04-04 18:11:11 -07007095#ifdef CONFIG_NUMA
7096 kfree(sched_group_nodes);
7097#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01007098 return -ENOMEM;
7099 }
7100
Mike Travis7c16ec52008-04-04 18:11:11 -07007101#if SCHED_CPUMASK_ALLOC
7102 /* get space for all scratch cpumask variables */
7103 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7104 if (!allmasks) {
7105 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7106 kfree(rd);
7107#ifdef CONFIG_NUMA
7108 kfree(sched_group_nodes);
7109#endif
7110 return -ENOMEM;
7111 }
7112#endif
7113 tmpmask = (cpumask_t *)allmasks;
7114
7115
7116#ifdef CONFIG_NUMA
7117 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7118#endif
7119
Linus Torvalds1da177e2005-04-16 15:20:36 -07007120 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007121 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007122 */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007123 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007124 struct sched_domain *sd = NULL, *p;
Mike Travis7c16ec52008-04-04 18:11:11 -07007125 SCHED_CPUMASK_VAR(nodemask, allmasks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007126
Mike Travis7c16ec52008-04-04 18:11:11 -07007127 *nodemask = node_to_cpumask(cpu_to_node(i));
7128 cpus_and(*nodemask, *nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007129
7130#ifdef CONFIG_NUMA
Ingo Molnardd41f592007-07-09 18:51:59 +02007131 if (cpus_weight(*cpu_map) >
Mike Travis7c16ec52008-04-04 18:11:11 -07007132 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07007133 sd = &per_cpu(allnodes_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007134 SD_INIT(sd, ALLNODES);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007135 set_domain_attribute(sd, attr);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007136 sd->span = *cpu_map;
Mike Travis7c16ec52008-04-04 18:11:11 -07007137 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007138 p = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007139 sd_allnodes = 1;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007140 } else
7141 p = NULL;
7142
Linus Torvalds1da177e2005-04-16 15:20:36 -07007143 sd = &per_cpu(node_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007144 SD_INIT(sd, NODE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007145 set_domain_attribute(sd, attr);
Mike Travis4bdbaad32008-04-15 16:35:52 -07007146 sched_domain_node_span(cpu_to_node(i), &sd->span);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007147 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007148 if (p)
7149 p->child = sd;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007150 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007151#endif
7152
7153 p = sd;
7154 sd = &per_cpu(phys_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007155 SD_INIT(sd, CPU);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007156 set_domain_attribute(sd, attr);
Mike Travis7c16ec52008-04-04 18:11:11 -07007157 sd->span = *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007158 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007159 if (p)
7160 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007161 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007162
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007163#ifdef CONFIG_SCHED_MC
7164 p = sd;
7165 sd = &per_cpu(core_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007166 SD_INIT(sd, MC);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007167 set_domain_attribute(sd, attr);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007168 sd->span = cpu_coregroup_map(i);
7169 cpus_and(sd->span, sd->span, *cpu_map);
7170 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007171 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007172 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007173#endif
7174
Linus Torvalds1da177e2005-04-16 15:20:36 -07007175#ifdef CONFIG_SCHED_SMT
7176 p = sd;
7177 sd = &per_cpu(cpu_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007178 SD_INIT(sd, SIBLING);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007179 set_domain_attribute(sd, attr);
Mike Travisd5a74302007-10-16 01:24:05 -07007180 sd->span = per_cpu(cpu_sibling_map, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007181 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007182 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007183 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007184 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007185#endif
7186 }
7187
7188#ifdef CONFIG_SCHED_SMT
7189 /* Set up CPU (sibling) groups */
John Hawkes9c1cfda2005-09-06 15:18:14 -07007190 for_each_cpu_mask(i, *cpu_map) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007191 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7192 SCHED_CPUMASK_VAR(send_covered, allmasks);
7193
7194 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7195 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7196 if (i != first_cpu(*this_sibling_map))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007197 continue;
7198
Ingo Molnardd41f592007-07-09 18:51:59 +02007199 init_sched_build_groups(this_sibling_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007200 &cpu_to_cpu_group,
7201 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007202 }
7203#endif
7204
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007205#ifdef CONFIG_SCHED_MC
7206 /* Set up multi-core groups */
7207 for_each_cpu_mask(i, *cpu_map) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007208 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7209 SCHED_CPUMASK_VAR(send_covered, allmasks);
7210
7211 *this_core_map = cpu_coregroup_map(i);
7212 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7213 if (i != first_cpu(*this_core_map))
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007214 continue;
Mike Travis7c16ec52008-04-04 18:11:11 -07007215
Ingo Molnardd41f592007-07-09 18:51:59 +02007216 init_sched_build_groups(this_core_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007217 &cpu_to_core_group,
7218 send_covered, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007219 }
7220#endif
7221
Linus Torvalds1da177e2005-04-16 15:20:36 -07007222 /* Set up physical groups */
7223 for (i = 0; i < MAX_NUMNODES; i++) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007224 SCHED_CPUMASK_VAR(nodemask, allmasks);
7225 SCHED_CPUMASK_VAR(send_covered, allmasks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007226
Mike Travis7c16ec52008-04-04 18:11:11 -07007227 *nodemask = node_to_cpumask(i);
7228 cpus_and(*nodemask, *nodemask, *cpu_map);
7229 if (cpus_empty(*nodemask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007230 continue;
7231
Mike Travis7c16ec52008-04-04 18:11:11 -07007232 init_sched_build_groups(nodemask, cpu_map,
7233 &cpu_to_phys_group,
7234 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007235 }
7236
7237#ifdef CONFIG_NUMA
7238 /* Set up node groups */
Mike Travis7c16ec52008-04-04 18:11:11 -07007239 if (sd_allnodes) {
7240 SCHED_CPUMASK_VAR(send_covered, allmasks);
7241
7242 init_sched_build_groups(cpu_map, cpu_map,
7243 &cpu_to_allnodes_group,
7244 send_covered, tmpmask);
7245 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007246
7247 for (i = 0; i < MAX_NUMNODES; i++) {
7248 /* Set up node groups */
7249 struct sched_group *sg, *prev;
Mike Travis7c16ec52008-04-04 18:11:11 -07007250 SCHED_CPUMASK_VAR(nodemask, allmasks);
7251 SCHED_CPUMASK_VAR(domainspan, allmasks);
7252 SCHED_CPUMASK_VAR(covered, allmasks);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007253 int j;
7254
Mike Travis7c16ec52008-04-04 18:11:11 -07007255 *nodemask = node_to_cpumask(i);
7256 cpus_clear(*covered);
7257
7258 cpus_and(*nodemask, *nodemask, *cpu_map);
7259 if (cpus_empty(*nodemask)) {
John Hawkesd1b55132005-09-06 15:18:14 -07007260 sched_group_nodes[i] = NULL;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007261 continue;
John Hawkesd1b55132005-09-06 15:18:14 -07007262 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007263
Mike Travis4bdbaad32008-04-15 16:35:52 -07007264 sched_domain_node_span(i, domainspan);
Mike Travis7c16ec52008-04-04 18:11:11 -07007265 cpus_and(*domainspan, *domainspan, *cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007266
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07007267 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007268 if (!sg) {
7269 printk(KERN_WARNING "Can not alloc domain group for "
7270 "node %d\n", i);
7271 goto error;
7272 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007273 sched_group_nodes[i] = sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07007274 for_each_cpu_mask(j, *nodemask) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07007275 struct sched_domain *sd;
Ingo Molnar9761eea2007-07-09 18:52:00 +02007276
John Hawkes9c1cfda2005-09-06 15:18:14 -07007277 sd = &per_cpu(node_domains, j);
7278 sd->groups = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007279 }
Eric Dumazet5517d862007-05-08 00:32:57 -07007280 sg->__cpu_power = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07007281 sg->cpumask = *nodemask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007282 sg->next = sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07007283 cpus_or(*covered, *covered, *nodemask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007284 prev = sg;
7285
7286 for (j = 0; j < MAX_NUMNODES; j++) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007287 SCHED_CPUMASK_VAR(notcovered, allmasks);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007288 int n = (i + j) % MAX_NUMNODES;
Mike Travisc5f59f02008-04-04 18:11:10 -07007289 node_to_cpumask_ptr(pnodemask, n);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007290
Mike Travis7c16ec52008-04-04 18:11:11 -07007291 cpus_complement(*notcovered, *covered);
7292 cpus_and(*tmpmask, *notcovered, *cpu_map);
7293 cpus_and(*tmpmask, *tmpmask, *domainspan);
7294 if (cpus_empty(*tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07007295 break;
7296
Mike Travis7c16ec52008-04-04 18:11:11 -07007297 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7298 if (cpus_empty(*tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07007299 continue;
7300
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07007301 sg = kmalloc_node(sizeof(struct sched_group),
7302 GFP_KERNEL, i);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007303 if (!sg) {
7304 printk(KERN_WARNING
7305 "Can not alloc domain group for node %d\n", j);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007306 goto error;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007307 }
Eric Dumazet5517d862007-05-08 00:32:57 -07007308 sg->__cpu_power = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07007309 sg->cpumask = *tmpmask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007310 sg->next = prev->next;
Mike Travis7c16ec52008-04-04 18:11:11 -07007311 cpus_or(*covered, *covered, *tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007312 prev->next = sg;
7313 prev = sg;
7314 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007315 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007316#endif
7317
7318 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007319#ifdef CONFIG_SCHED_SMT
7320 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007321 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7322
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007323 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007324 }
7325#endif
7326#ifdef CONFIG_SCHED_MC
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007327 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007328 struct sched_domain *sd = &per_cpu(core_domains, i);
7329
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007330 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007331 }
7332#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007333
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007334 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007335 struct sched_domain *sd = &per_cpu(phys_domains, i);
7336
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007337 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007338 }
7339
John Hawkes9c1cfda2005-09-06 15:18:14 -07007340#ifdef CONFIG_NUMA
Siddha, Suresh B08069032006-03-27 01:15:23 -08007341 for (i = 0; i < MAX_NUMNODES; i++)
7342 init_numa_sched_groups_power(sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007343
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007344 if (sd_allnodes) {
7345 struct sched_group *sg;
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07007346
Mike Travis7c16ec52008-04-04 18:11:11 -07007347 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7348 tmpmask);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07007349 init_numa_sched_groups_power(sg);
7350 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007351#endif
7352
Linus Torvalds1da177e2005-04-16 15:20:36 -07007353 /* Attach the domains */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007354 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007355 struct sched_domain *sd;
7356#ifdef CONFIG_SCHED_SMT
7357 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007358#elif defined(CONFIG_SCHED_MC)
7359 sd = &per_cpu(core_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007360#else
7361 sd = &per_cpu(phys_domains, i);
7362#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01007363 cpu_attach_domain(sd, rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007364 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007365
Mike Travis7c16ec52008-04-04 18:11:11 -07007366 SCHED_CPUMASK_FREE((void *)allmasks);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007367 return 0;
7368
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007369#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007370error:
Mike Travis7c16ec52008-04-04 18:11:11 -07007371 free_sched_groups(cpu_map, tmpmask);
7372 SCHED_CPUMASK_FREE((void *)allmasks);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007373 return -ENOMEM;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007374#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007375}
Paul Jackson029190c2007-10-18 23:40:20 -07007376
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007377static int build_sched_domains(const cpumask_t *cpu_map)
7378{
7379 return __build_sched_domains(cpu_map, NULL);
7380}
7381
Paul Jackson029190c2007-10-18 23:40:20 -07007382static cpumask_t *doms_cur; /* current sched domains */
7383static int ndoms_cur; /* number of sched domains in 'doms_cur' */
Ingo Molnar4285f5942008-05-16 17:47:14 +02007384static struct sched_domain_attr *dattr_cur;
7385 /* attribues of custom domains in 'doms_cur' */
Paul Jackson029190c2007-10-18 23:40:20 -07007386
7387/*
7388 * Special case: If a kmalloc of a doms_cur partition (array of
7389 * cpumask_t) fails, then fallback to a single sched domain,
7390 * as determined by the single cpumask_t fallback_doms.
7391 */
7392static cpumask_t fallback_doms;
7393
Heiko Carstens22e52b02008-03-12 18:31:59 +01007394void __attribute__((weak)) arch_update_cpu_topology(void)
7395{
7396}
7397
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007398/*
Max Krasnyansky5c8e1ed2008-05-29 11:17:01 -07007399 * Free current domain masks.
7400 * Called after all cpus are attached to NULL domain.
7401 */
7402static void free_sched_domains(void)
7403{
7404 ndoms_cur = 0;
7405 if (doms_cur != &fallback_doms)
7406 kfree(doms_cur);
7407 doms_cur = &fallback_doms;
7408}
7409
7410/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007411 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07007412 * For now this just excludes isolated cpus, but could be used to
7413 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007414 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007415static int arch_init_sched_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007416{
Milton Miller73785472007-10-24 18:23:48 +02007417 int err;
7418
Heiko Carstens22e52b02008-03-12 18:31:59 +01007419 arch_update_cpu_topology();
Paul Jackson029190c2007-10-18 23:40:20 -07007420 ndoms_cur = 1;
7421 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7422 if (!doms_cur)
7423 doms_cur = &fallback_doms;
7424 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007425 dattr_cur = NULL;
Milton Miller73785472007-10-24 18:23:48 +02007426 err = build_sched_domains(doms_cur);
Milton Miller6382bc92007-10-15 17:00:19 +02007427 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02007428
7429 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007430}
7431
Mike Travis7c16ec52008-04-04 18:11:11 -07007432static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7433 cpumask_t *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007434{
Mike Travis7c16ec52008-04-04 18:11:11 -07007435 free_sched_groups(cpu_map, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007436}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007437
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007438/*
7439 * Detach sched domains from a group of cpus specified in cpu_map
7440 * These cpus will now be attached to the NULL domain
7441 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08007442static void detach_destroy_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007443{
Mike Travis7c16ec52008-04-04 18:11:11 -07007444 cpumask_t tmpmask;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007445 int i;
7446
Milton Miller6382bc92007-10-15 17:00:19 +02007447 unregister_sched_domain_sysctl();
7448
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007449 for_each_cpu_mask(i, *cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01007450 cpu_attach_domain(NULL, &def_root_domain, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007451 synchronize_sched();
Mike Travis7c16ec52008-04-04 18:11:11 -07007452 arch_destroy_sched_domains(cpu_map, &tmpmask);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007453}
7454
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007455/* handle null as "default" */
7456static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7457 struct sched_domain_attr *new, int idx_new)
7458{
7459 struct sched_domain_attr tmp;
7460
7461 /* fast path */
7462 if (!new && !cur)
7463 return 1;
7464
7465 tmp = SD_ATTR_INIT;
7466 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7467 new ? (new + idx_new) : &tmp,
7468 sizeof(struct sched_domain_attr));
7469}
7470
Paul Jackson029190c2007-10-18 23:40:20 -07007471/*
7472 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007473 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07007474 * doms_new[] to the current sched domain partitioning, doms_cur[].
7475 * It destroys each deleted domain and builds each new domain.
7476 *
7477 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007478 * The masks don't intersect (don't overlap.) We should setup one
7479 * sched domain for each mask. CPUs not in any of the cpumasks will
7480 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07007481 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7482 * it as it is.
7483 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007484 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7485 * ownership of it and will kfree it when done with it. If the caller
Paul Jackson029190c2007-10-18 23:40:20 -07007486 * failed the kmalloc call, then it can pass in doms_new == NULL,
7487 * and partition_sched_domains() will fallback to the single partition
7488 * 'fallback_doms'.
7489 *
7490 * Call with hotplug lock held
7491 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007492void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7493 struct sched_domain_attr *dattr_new)
Paul Jackson029190c2007-10-18 23:40:20 -07007494{
7495 int i, j;
7496
Heiko Carstens712555e2008-04-28 11:33:07 +02007497 mutex_lock(&sched_domains_mutex);
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007498
Milton Miller73785472007-10-24 18:23:48 +02007499 /* always unregister in case we don't destroy any domains */
7500 unregister_sched_domain_sysctl();
7501
Paul Jackson029190c2007-10-18 23:40:20 -07007502 if (doms_new == NULL) {
7503 ndoms_new = 1;
7504 doms_new = &fallback_doms;
7505 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007506 dattr_new = NULL;
Paul Jackson029190c2007-10-18 23:40:20 -07007507 }
7508
7509 /* Destroy deleted domains */
7510 for (i = 0; i < ndoms_cur; i++) {
7511 for (j = 0; j < ndoms_new; j++) {
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007512 if (cpus_equal(doms_cur[i], doms_new[j])
7513 && dattrs_equal(dattr_cur, i, dattr_new, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007514 goto match1;
7515 }
7516 /* no match - a current sched domain not in new doms_new[] */
7517 detach_destroy_domains(doms_cur + i);
7518match1:
7519 ;
7520 }
7521
7522 /* Build new domains */
7523 for (i = 0; i < ndoms_new; i++) {
7524 for (j = 0; j < ndoms_cur; j++) {
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007525 if (cpus_equal(doms_new[i], doms_cur[j])
7526 && dattrs_equal(dattr_new, i, dattr_cur, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007527 goto match2;
7528 }
7529 /* no match - add a new doms_new */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007530 __build_sched_domains(doms_new + i,
7531 dattr_new ? dattr_new + i : NULL);
Paul Jackson029190c2007-10-18 23:40:20 -07007532match2:
7533 ;
7534 }
7535
7536 /* Remember the new sched domains */
7537 if (doms_cur != &fallback_doms)
7538 kfree(doms_cur);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007539 kfree(dattr_cur); /* kfree(NULL) is safe */
Paul Jackson029190c2007-10-18 23:40:20 -07007540 doms_cur = doms_new;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007541 dattr_cur = dattr_new;
Paul Jackson029190c2007-10-18 23:40:20 -07007542 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02007543
7544 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007545
Heiko Carstens712555e2008-04-28 11:33:07 +02007546 mutex_unlock(&sched_domains_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -07007547}
7548
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007549#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Heiko Carstens9aefd0a2008-03-12 18:31:58 +01007550int arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007551{
7552 int err;
7553
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007554 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02007555 mutex_lock(&sched_domains_mutex);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007556 detach_destroy_domains(&cpu_online_map);
Max Krasnyansky5c8e1ed2008-05-29 11:17:01 -07007557 free_sched_domains();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007558 err = arch_init_sched_domains(&cpu_online_map);
Heiko Carstens712555e2008-04-28 11:33:07 +02007559 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007560 put_online_cpus();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007561
7562 return err;
7563}
7564
7565static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7566{
7567 int ret;
7568
7569 if (buf[0] != '0' && buf[0] != '1')
7570 return -EINVAL;
7571
7572 if (smt)
7573 sched_smt_power_savings = (buf[0] == '1');
7574 else
7575 sched_mc_power_savings = (buf[0] == '1');
7576
7577 ret = arch_reinit_sched_domains();
7578
7579 return ret ? ret : count;
7580}
7581
Adrian Bunk6707de002007-08-12 18:08:19 +02007582#ifdef CONFIG_SCHED_MC
7583static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7584{
7585 return sprintf(page, "%u\n", sched_mc_power_savings);
7586}
7587static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7588 const char *buf, size_t count)
7589{
7590 return sched_power_savings_store(buf, count, 0);
7591}
7592static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7593 sched_mc_power_savings_store);
7594#endif
7595
7596#ifdef CONFIG_SCHED_SMT
7597static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7598{
7599 return sprintf(page, "%u\n", sched_smt_power_savings);
7600}
7601static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7602 const char *buf, size_t count)
7603{
7604 return sched_power_savings_store(buf, count, 1);
7605}
7606static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7607 sched_smt_power_savings_store);
7608#endif
7609
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007610int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7611{
7612 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007613
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007614#ifdef CONFIG_SCHED_SMT
7615 if (smt_capable())
7616 err = sysfs_create_file(&cls->kset.kobj,
7617 &attr_sched_smt_power_savings.attr);
7618#endif
7619#ifdef CONFIG_SCHED_MC
7620 if (!err && mc_capable())
7621 err = sysfs_create_file(&cls->kset.kobj,
7622 &attr_sched_mc_power_savings.attr);
7623#endif
7624 return err;
7625}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007626#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007627
Linus Torvalds1da177e2005-04-16 15:20:36 -07007628/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007629 * Force a reinitialization of the sched domains hierarchy. The domains
Linus Torvalds1da177e2005-04-16 15:20:36 -07007630 * and groups cannot be updated in place without racing with the balancing
Nick Piggin41c7ce92005-06-25 14:57:24 -07007631 * code, so we temporarily attach all running cpus to the NULL domain
Linus Torvalds1da177e2005-04-16 15:20:36 -07007632 * which will prevent rebalancing while the sched domains are recalculated.
7633 */
7634static int update_sched_domains(struct notifier_block *nfb,
7635 unsigned long action, void *hcpu)
7636{
Peter Zijlstra7def2be2008-06-05 14:49:58 +02007637 int cpu = (int)(long)hcpu;
7638
Linus Torvalds1da177e2005-04-16 15:20:36 -07007639 switch (action) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007640 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007641 case CPU_DOWN_PREPARE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02007642 disable_runtime(cpu_rq(cpu));
7643 /* fall-through */
7644 case CPU_UP_PREPARE:
7645 case CPU_UP_PREPARE_FROZEN:
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007646 detach_destroy_domains(&cpu_online_map);
Max Krasnyansky5c8e1ed2008-05-29 11:17:01 -07007647 free_sched_domains();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007648 return NOTIFY_OK;
7649
Peter Zijlstra7def2be2008-06-05 14:49:58 +02007650
Linus Torvalds1da177e2005-04-16 15:20:36 -07007651 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007652 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007653 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007654 case CPU_ONLINE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02007655 enable_runtime(cpu_rq(cpu));
7656 /* fall-through */
7657 case CPU_UP_CANCELED:
7658 case CPU_UP_CANCELED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007659 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007660 case CPU_DEAD_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007661 /*
7662 * Fall through and re-initialise the domains.
7663 */
7664 break;
7665 default:
7666 return NOTIFY_DONE;
7667 }
7668
Max Krasnyansky5c8e1ed2008-05-29 11:17:01 -07007669#ifndef CONFIG_CPUSETS
7670 /*
7671 * Create default domain partitioning if cpusets are disabled.
7672 * Otherwise we let cpusets rebuild the domains based on the
7673 * current setup.
7674 */
7675
Linus Torvalds1da177e2005-04-16 15:20:36 -07007676 /* The hotplug lock is already held by cpu_up/cpu_down */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007677 arch_init_sched_domains(&cpu_online_map);
Max Krasnyansky5c8e1ed2008-05-29 11:17:01 -07007678#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007679
7680 return NOTIFY_OK;
7681}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007682
7683void __init sched_init_smp(void)
7684{
Nick Piggin5c1e1762006-10-03 01:14:04 -07007685 cpumask_t non_isolated_cpus;
7686
Mike Travis434d53b2008-04-04 18:11:04 -07007687#if defined(CONFIG_NUMA)
7688 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7689 GFP_KERNEL);
7690 BUG_ON(sched_group_nodes_bycpu == NULL);
7691#endif
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007692 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02007693 mutex_lock(&sched_domains_mutex);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007694 arch_init_sched_domains(&cpu_online_map);
Nathan Lynche5e56732007-01-10 23:15:28 -08007695 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
Nick Piggin5c1e1762006-10-03 01:14:04 -07007696 if (cpus_empty(non_isolated_cpus))
7697 cpu_set(smp_processor_id(), non_isolated_cpus);
Heiko Carstens712555e2008-04-28 11:33:07 +02007698 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007699 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007700 /* XXX: Theoretical race here - CPU may be hotplugged now */
7701 hotcpu_notifier(update_sched_domains, 0);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02007702 init_hrtick();
Nick Piggin5c1e1762006-10-03 01:14:04 -07007703
7704 /* Move init over to a non-isolated CPU */
Mike Travis7c16ec52008-04-04 18:11:11 -07007705 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
Nick Piggin5c1e1762006-10-03 01:14:04 -07007706 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01007707 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007708}
7709#else
7710void __init sched_init_smp(void)
7711{
Ingo Molnar19978ca2007-11-09 22:39:38 +01007712 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007713}
7714#endif /* CONFIG_SMP */
7715
7716int in_sched_functions(unsigned long addr)
7717{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007718 return in_lock_functions(addr) ||
7719 (addr >= (unsigned long)__sched_text_start
7720 && addr < (unsigned long)__sched_text_end);
7721}
7722
Alexey Dobriyana9957442007-10-15 17:00:13 +02007723static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02007724{
7725 cfs_rq->tasks_timeline = RB_ROOT;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02007726 INIT_LIST_HEAD(&cfs_rq->tasks);
Ingo Molnardd41f592007-07-09 18:51:59 +02007727#ifdef CONFIG_FAIR_GROUP_SCHED
7728 cfs_rq->rq = rq;
7729#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02007730 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02007731}
7732
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007733static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7734{
7735 struct rt_prio_array *array;
7736 int i;
7737
7738 array = &rt_rq->active;
7739 for (i = 0; i < MAX_RT_PRIO; i++) {
Dmitry Adamushko20b63312008-06-11 00:58:30 +02007740 INIT_LIST_HEAD(array->queue + i);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007741 __clear_bit(i, array->bitmap);
7742 }
7743 /* delimiter for bitsearch: */
7744 __set_bit(MAX_RT_PRIO, array->bitmap);
7745
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007746#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Peter Zijlstra48d5e252008-01-25 21:08:31 +01007747 rt_rq->highest_prio = MAX_RT_PRIO;
7748#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007749#ifdef CONFIG_SMP
7750 rt_rq->rt_nr_migratory = 0;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007751 rt_rq->overloaded = 0;
7752#endif
7753
7754 rt_rq->rt_time = 0;
7755 rt_rq->rt_throttled = 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007756 rt_rq->rt_runtime = 0;
7757 spin_lock_init(&rt_rq->rt_runtime_lock);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007758
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007759#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01007760 rt_rq->rt_nr_boosted = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007761 rt_rq->rq = rq;
7762#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007763}
7764
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007765#ifdef CONFIG_FAIR_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007766static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7767 struct sched_entity *se, int cpu, int add,
7768 struct sched_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007769{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007770 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007771 tg->cfs_rq[cpu] = cfs_rq;
7772 init_cfs_rq(cfs_rq, rq);
7773 cfs_rq->tg = tg;
7774 if (add)
7775 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7776
7777 tg->se[cpu] = se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02007778 /* se could be NULL for init_task_group */
7779 if (!se)
7780 return;
7781
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007782 if (!parent)
7783 se->cfs_rq = &rq->cfs;
7784 else
7785 se->cfs_rq = parent->my_q;
7786
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007787 se->my_q = cfs_rq;
7788 se->load.weight = tg->shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02007789 se->load.inv_weight = 0;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007790 se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007791}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007792#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007793
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007794#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007795static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7796 struct sched_rt_entity *rt_se, int cpu, int add,
7797 struct sched_rt_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007798{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007799 struct rq *rq = cpu_rq(cpu);
7800
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007801 tg->rt_rq[cpu] = rt_rq;
7802 init_rt_rq(rt_rq, rq);
7803 rt_rq->tg = tg;
7804 rt_rq->rt_se = rt_se;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007805 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007806 if (add)
7807 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7808
7809 tg->rt_se[cpu] = rt_se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02007810 if (!rt_se)
7811 return;
7812
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007813 if (!parent)
7814 rt_se->rt_rq = &rq->rt;
7815 else
7816 rt_se->rt_rq = parent->my_q;
7817
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007818 rt_se->my_q = rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007819 rt_se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007820 INIT_LIST_HEAD(&rt_se->run_list);
7821}
7822#endif
7823
Linus Torvalds1da177e2005-04-16 15:20:36 -07007824void __init sched_init(void)
7825{
Ingo Molnardd41f592007-07-09 18:51:59 +02007826 int i, j;
Mike Travis434d53b2008-04-04 18:11:04 -07007827 unsigned long alloc_size = 0, ptr;
7828
7829#ifdef CONFIG_FAIR_GROUP_SCHED
7830 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7831#endif
7832#ifdef CONFIG_RT_GROUP_SCHED
7833 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7834#endif
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007835#ifdef CONFIG_USER_SCHED
7836 alloc_size *= 2;
7837#endif
Mike Travis434d53b2008-04-04 18:11:04 -07007838 /*
7839 * As sched_init() is called before page_alloc is setup,
7840 * we use alloc_bootmem().
7841 */
7842 if (alloc_size) {
David Miller5a9d3222008-04-24 20:46:20 -07007843 ptr = (unsigned long)alloc_bootmem(alloc_size);
Mike Travis434d53b2008-04-04 18:11:04 -07007844
7845#ifdef CONFIG_FAIR_GROUP_SCHED
7846 init_task_group.se = (struct sched_entity **)ptr;
7847 ptr += nr_cpu_ids * sizeof(void **);
7848
7849 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7850 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007851
7852#ifdef CONFIG_USER_SCHED
7853 root_task_group.se = (struct sched_entity **)ptr;
7854 ptr += nr_cpu_ids * sizeof(void **);
7855
7856 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7857 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007858#endif /* CONFIG_USER_SCHED */
7859#endif /* CONFIG_FAIR_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07007860#ifdef CONFIG_RT_GROUP_SCHED
7861 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7862 ptr += nr_cpu_ids * sizeof(void **);
7863
7864 init_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007865 ptr += nr_cpu_ids * sizeof(void **);
7866
7867#ifdef CONFIG_USER_SCHED
7868 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7869 ptr += nr_cpu_ids * sizeof(void **);
7870
7871 root_task_group.rt_rq = (struct rt_rq **)ptr;
7872 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007873#endif /* CONFIG_USER_SCHED */
7874#endif /* CONFIG_RT_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07007875 }
Ingo Molnardd41f592007-07-09 18:51:59 +02007876
Gregory Haskins57d885f2008-01-25 21:08:18 +01007877#ifdef CONFIG_SMP
7878 init_defrootdomain();
7879#endif
7880
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007881 init_rt_bandwidth(&def_rt_bandwidth,
7882 global_rt_period(), global_rt_runtime());
7883
7884#ifdef CONFIG_RT_GROUP_SCHED
7885 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7886 global_rt_period(), global_rt_runtime());
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007887#ifdef CONFIG_USER_SCHED
7888 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7889 global_rt_period(), RUNTIME_INF);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007890#endif /* CONFIG_USER_SCHED */
7891#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007892
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007893#ifdef CONFIG_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007894 list_add(&init_task_group.list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02007895 INIT_LIST_HEAD(&init_task_group.children);
7896
7897#ifdef CONFIG_USER_SCHED
7898 INIT_LIST_HEAD(&root_task_group.children);
7899 init_task_group.parent = &root_task_group;
7900 list_add(&init_task_group.siblings, &root_task_group.children);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007901#endif /* CONFIG_USER_SCHED */
7902#endif /* CONFIG_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007903
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08007904 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007905 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007906
7907 rq = cpu_rq(i);
7908 spin_lock_init(&rq->lock);
Ingo Molnarfcb99372006-07-03 00:25:10 -07007909 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
Nick Piggin78979862005-06-25 14:57:13 -07007910 rq->nr_running = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007911 init_cfs_rq(&rq->cfs, rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007912 init_rt_rq(&rq->rt, rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007913#ifdef CONFIG_FAIR_GROUP_SCHED
7914 init_task_group.shares = init_task_group_load;
7915 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007916#ifdef CONFIG_CGROUP_SCHED
7917 /*
7918 * How much cpu bandwidth does init_task_group get?
7919 *
7920 * In case of task-groups formed thr' the cgroup filesystem, it
7921 * gets 100% of the cpu resources in the system. This overall
7922 * system cpu resource is divided among the tasks of
7923 * init_task_group and its child task-groups in a fair manner,
7924 * based on each entity's (task or task-group's) weight
7925 * (se->load.weight).
7926 *
7927 * In other words, if init_task_group has 10 tasks of weight
7928 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7929 * then A0's share of the cpu resource is:
7930 *
7931 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7932 *
7933 * We achieve this by letting init_task_group's tasks sit
7934 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7935 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007936 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007937#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007938 root_task_group.shares = NICE_0_LOAD;
7939 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007940 /*
7941 * In case of task-groups formed thr' the user id of tasks,
7942 * init_task_group represents tasks belonging to root user.
7943 * Hence it forms a sibling of all subsequent groups formed.
7944 * In this case, init_task_group gets only a fraction of overall
7945 * system cpu resource, based on the weight assigned to root
7946 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7947 * by letting tasks of init_task_group sit in a separate cfs_rq
7948 * (init_cfs_rq) and having one entity represent this group of
7949 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7950 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007951 init_tg_cfs_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007952 &per_cpu(init_cfs_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007953 &per_cpu(init_sched_entity, i), i, 1,
7954 root_task_group.se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007955
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007956#endif
Dhaval Giani354d60c2008-04-19 19:44:59 +02007957#endif /* CONFIG_FAIR_GROUP_SCHED */
7958
7959 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007960#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007961 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007962#ifdef CONFIG_CGROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007963 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007964#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007965 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007966 init_tg_rt_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007967 &per_cpu(init_rt_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007968 &per_cpu(init_sched_rt_entity, i), i, 1,
7969 root_task_group.rt_se[i]);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007970#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007971#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007972
Ingo Molnardd41f592007-07-09 18:51:59 +02007973 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7974 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007975#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07007976 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007977 rq->rd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007978 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007979 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007980 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07007981 rq->cpu = i;
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007982 rq->online = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007983 rq->migration_thread = NULL;
7984 INIT_LIST_HEAD(&rq->migration_queue);
Gregory Haskinsdc938522008-01-25 21:08:26 +01007985 rq_attach_root(rq, &def_root_domain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007986#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01007987 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007988 atomic_set(&rq->nr_iowait, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007989 }
7990
Peter Williams2dd73a42006-06-27 02:54:34 -07007991 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07007992
Avi Kivitye107be32007-07-26 13:40:43 +02007993#ifdef CONFIG_PREEMPT_NOTIFIERS
7994 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7995#endif
7996
Christoph Lameterc9819f42006-12-10 02:20:25 -08007997#ifdef CONFIG_SMP
7998 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7999#endif
8000
Heiko Carstensb50f60c2006-07-30 03:03:52 -07008001#ifdef CONFIG_RT_MUTEXES
8002 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8003#endif
8004
Linus Torvalds1da177e2005-04-16 15:20:36 -07008005 /*
8006 * The boot idle thread does lazy MMU switching as well:
8007 */
8008 atomic_inc(&init_mm.mm_count);
8009 enter_lazy_tlb(&init_mm, current);
8010
8011 /*
8012 * Make us the idle thread. Technically, schedule() should not be
8013 * called from this thread, however somewhere below it might be,
8014 * but because we are the idle thread, we just pick up running again
8015 * when this runqueue becomes "idle".
8016 */
8017 init_idle(current, smp_processor_id());
Ingo Molnardd41f592007-07-09 18:51:59 +02008018 /*
8019 * During early bootup we pretend to be a normal task:
8020 */
8021 current->sched_class = &fair_sched_class;
Ingo Molnar6892b752008-02-13 14:02:36 +01008022
8023 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008024}
8025
8026#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8027void __might_sleep(char *file, int line)
8028{
Ingo Molnar48f24c42006-07-03 00:25:40 -07008029#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07008030 static unsigned long prev_jiffy; /* ratelimiting */
8031
8032 if ((in_atomic() || irqs_disabled()) &&
8033 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8034 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8035 return;
8036 prev_jiffy = jiffies;
Ingo Molnar91368d72006-03-23 03:00:54 -08008037 printk(KERN_ERR "BUG: sleeping function called from invalid"
Linus Torvalds1da177e2005-04-16 15:20:36 -07008038 " context at %s:%d\n", file, line);
8039 printk("in_atomic():%d, irqs_disabled():%d\n",
8040 in_atomic(), irqs_disabled());
Peter Zijlstraa4c410f2006-12-06 20:37:21 -08008041 debug_show_held_locks(current);
Ingo Molnar3117df02006-12-13 00:34:43 -08008042 if (irqs_disabled())
8043 print_irqtrace_events(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008044 dump_stack();
8045 }
8046#endif
8047}
8048EXPORT_SYMBOL(__might_sleep);
8049#endif
8050
8051#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02008052static void normalize_task(struct rq *rq, struct task_struct *p)
8053{
8054 int on_rq;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02008055
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02008056 update_rq_clock(rq);
8057 on_rq = p->se.on_rq;
8058 if (on_rq)
8059 deactivate_task(rq, p, 0);
8060 __setscheduler(rq, p, SCHED_NORMAL, 0);
8061 if (on_rq) {
8062 activate_task(rq, p, 0);
8063 resched_task(rq->curr);
8064 }
8065}
8066
Linus Torvalds1da177e2005-04-16 15:20:36 -07008067void normalize_rt_tasks(void)
8068{
Ingo Molnara0f98a12007-06-17 18:37:45 +02008069 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008070 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07008071 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008072
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008073 read_lock_irqsave(&tasklist_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02008074 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02008075 /*
8076 * Only normalize user tasks:
8077 */
8078 if (!p->mm)
8079 continue;
8080
Ingo Molnardd41f592007-07-09 18:51:59 +02008081 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02008082#ifdef CONFIG_SCHEDSTATS
8083 p->se.wait_start = 0;
8084 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02008085 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02008086#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02008087
8088 if (!rt_task(p)) {
8089 /*
8090 * Renice negative nice level userspace
8091 * tasks back to 0:
8092 */
8093 if (TASK_NICE(p) < 0 && p->mm)
8094 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008095 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02008096 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008097
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008098 spin_lock(&p->pi_lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -07008099 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008100
Ingo Molnar178be792007-10-15 17:00:18 +02008101 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02008102
Ingo Molnarb29739f2006-06-27 02:54:51 -07008103 __task_rq_unlock(rq);
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008104 spin_unlock(&p->pi_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02008105 } while_each_thread(g, p);
8106
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008107 read_unlock_irqrestore(&tasklist_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008108}
8109
8110#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07008111
8112#ifdef CONFIG_IA64
8113/*
8114 * These functions are only useful for the IA64 MCA handling.
8115 *
8116 * They can only be called when the whole system has been
8117 * stopped - every CPU needs to be quiescent, and no scheduling
8118 * activity can take place. Using them for anything else would
8119 * be a serious bug, and as a result, they aren't even visible
8120 * under any other configuration.
8121 */
8122
8123/**
8124 * curr_task - return the current task for a given cpu.
8125 * @cpu: the processor in question.
8126 *
8127 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8128 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07008129struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07008130{
8131 return cpu_curr(cpu);
8132}
8133
8134/**
8135 * set_curr_task - set the current task for a given cpu.
8136 * @cpu: the processor in question.
8137 * @p: the task pointer to set.
8138 *
8139 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008140 * are serviced on a separate stack. It allows the architecture to switch the
8141 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07008142 * must be called with all CPU's synchronized, and interrupts disabled, the
8143 * and caller must save the original value of the current task (see
8144 * curr_task() above) and restore that value before reenabling interrupts and
8145 * re-starting the system.
8146 *
8147 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8148 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07008149void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07008150{
8151 cpu_curr(cpu) = p;
8152}
8153
8154#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008155
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008156#ifdef CONFIG_FAIR_GROUP_SCHED
8157static void free_fair_sched_group(struct task_group *tg)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008158{
8159 int i;
8160
8161 for_each_possible_cpu(i) {
8162 if (tg->cfs_rq)
8163 kfree(tg->cfs_rq[i]);
8164 if (tg->se)
8165 kfree(tg->se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008166 }
8167
8168 kfree(tg->cfs_rq);
8169 kfree(tg->se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008170}
8171
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008172static
8173int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008174{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008175 struct cfs_rq *cfs_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008176 struct sched_entity *se, *parent_se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008177 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008178 int i;
8179
Mike Travis434d53b2008-04-04 18:11:04 -07008180 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008181 if (!tg->cfs_rq)
8182 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07008183 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008184 if (!tg->se)
8185 goto err;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008186
8187 tg->shares = NICE_0_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008188
8189 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008190 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008191
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008192 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8193 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008194 if (!cfs_rq)
8195 goto err;
8196
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008197 se = kmalloc_node(sizeof(struct sched_entity),
8198 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008199 if (!se)
8200 goto err;
8201
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008202 parent_se = parent ? parent->se[i] : NULL;
8203 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008204 }
8205
8206 return 1;
8207
8208 err:
8209 return 0;
8210}
8211
8212static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8213{
8214 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8215 &cpu_rq(cpu)->leaf_cfs_rq_list);
8216}
8217
8218static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8219{
8220 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8221}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008222#else /* !CONFG_FAIR_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008223static inline void free_fair_sched_group(struct task_group *tg)
8224{
8225}
8226
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008227static inline
8228int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008229{
8230 return 1;
8231}
8232
8233static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8234{
8235}
8236
8237static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8238{
8239}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008240#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008241
8242#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008243static void free_rt_sched_group(struct task_group *tg)
8244{
8245 int i;
8246
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008247 destroy_rt_bandwidth(&tg->rt_bandwidth);
8248
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008249 for_each_possible_cpu(i) {
8250 if (tg->rt_rq)
8251 kfree(tg->rt_rq[i]);
8252 if (tg->rt_se)
8253 kfree(tg->rt_se[i]);
8254 }
8255
8256 kfree(tg->rt_rq);
8257 kfree(tg->rt_se);
8258}
8259
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008260static
8261int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008262{
8263 struct rt_rq *rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008264 struct sched_rt_entity *rt_se, *parent_se;
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008265 struct rq *rq;
8266 int i;
8267
Mike Travis434d53b2008-04-04 18:11:04 -07008268 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008269 if (!tg->rt_rq)
8270 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07008271 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008272 if (!tg->rt_se)
8273 goto err;
8274
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008275 init_rt_bandwidth(&tg->rt_bandwidth,
8276 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008277
8278 for_each_possible_cpu(i) {
8279 rq = cpu_rq(i);
8280
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008281 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8282 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8283 if (!rt_rq)
8284 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008285
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008286 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8287 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8288 if (!rt_se)
8289 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008290
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008291 parent_se = parent ? parent->rt_se[i] : NULL;
8292 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008293 }
8294
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008295 return 1;
8296
8297 err:
8298 return 0;
8299}
8300
8301static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8302{
8303 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8304 &cpu_rq(cpu)->leaf_rt_rq_list);
8305}
8306
8307static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8308{
8309 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8310}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008311#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008312static inline void free_rt_sched_group(struct task_group *tg)
8313{
8314}
8315
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008316static inline
8317int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008318{
8319 return 1;
8320}
8321
8322static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8323{
8324}
8325
8326static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8327{
8328}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008329#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008330
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008331#ifdef CONFIG_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008332static void free_sched_group(struct task_group *tg)
8333{
8334 free_fair_sched_group(tg);
8335 free_rt_sched_group(tg);
8336 kfree(tg);
8337}
8338
8339/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008340struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008341{
8342 struct task_group *tg;
8343 unsigned long flags;
8344 int i;
8345
8346 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8347 if (!tg)
8348 return ERR_PTR(-ENOMEM);
8349
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008350 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008351 goto err;
8352
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008353 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008354 goto err;
8355
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008356 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008357 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008358 register_fair_sched_group(tg, i);
8359 register_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008360 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008361 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008362
8363 WARN_ON(!parent); /* root should already exist */
8364
8365 tg->parent = parent;
8366 list_add_rcu(&tg->siblings, &parent->children);
8367 INIT_LIST_HEAD(&tg->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008368 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008369
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008370 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008371
8372err:
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008373 free_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008374 return ERR_PTR(-ENOMEM);
8375}
8376
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008377/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008378static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008379{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008380 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008381 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008382}
8383
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008384/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02008385void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008386{
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008387 unsigned long flags;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008388 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008389
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008390 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008391 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008392 unregister_fair_sched_group(tg, i);
8393 unregister_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008394 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008395 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008396 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008397 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008398
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008399 /* wait for possible concurrent references to cfs_rqs complete */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008400 call_rcu(&tg->rcu, free_sched_group_rcu);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008401}
8402
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008403/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +02008404 * The caller of this function should have put the task in its new group
8405 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8406 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008407 */
8408void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008409{
8410 int on_rq, running;
8411 unsigned long flags;
8412 struct rq *rq;
8413
8414 rq = task_rq_lock(tsk, &flags);
8415
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008416 update_rq_clock(rq);
8417
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01008418 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008419 on_rq = tsk->se.on_rq;
8420
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008421 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008422 dequeue_task(rq, tsk, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008423 if (unlikely(running))
8424 tsk->sched_class->put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008425
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008426 set_task_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008427
Peter Zijlstra810b3812008-02-29 15:21:01 -05008428#ifdef CONFIG_FAIR_GROUP_SCHED
8429 if (tsk->sched_class->moved_group)
8430 tsk->sched_class->moved_group(tsk);
8431#endif
8432
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008433 if (unlikely(running))
8434 tsk->sched_class->set_curr_task(rq);
8435 if (on_rq)
Dmitry Adamushko7074bad2007-10-15 17:00:07 +02008436 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008437
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008438 task_rq_unlock(rq, &flags);
8439}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008440#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008441
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008442#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008443static void __set_se_shares(struct sched_entity *se, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008444{
8445 struct cfs_rq *cfs_rq = se->cfs_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008446 int on_rq;
8447
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008448 on_rq = se->on_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008449 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008450 dequeue_entity(cfs_rq, se, 0);
8451
8452 se->load.weight = shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02008453 se->load.inv_weight = 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008454
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008455 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008456 enqueue_entity(cfs_rq, se, 0);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008457}
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008458
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008459static void set_se_shares(struct sched_entity *se, unsigned long shares)
8460{
8461 struct cfs_rq *cfs_rq = se->cfs_rq;
8462 struct rq *rq = cfs_rq->rq;
8463 unsigned long flags;
8464
8465 spin_lock_irqsave(&rq->lock, flags);
8466 __set_se_shares(se, shares);
8467 spin_unlock_irqrestore(&rq->lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008468}
8469
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008470static DEFINE_MUTEX(shares_mutex);
8471
Ingo Molnar4cf86d72007-10-15 17:00:14 +02008472int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008473{
8474 int i;
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008475 unsigned long flags;
Ingo Molnarc61935f2008-01-22 11:24:58 +01008476
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008477 /*
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008478 * We can't change the weight of the root cgroup.
8479 */
8480 if (!tg->se[0])
8481 return -EINVAL;
8482
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008483 if (shares < MIN_SHARES)
8484 shares = MIN_SHARES;
Miao Xiecb4ad1f2008-04-28 12:54:56 +08008485 else if (shares > MAX_SHARES)
8486 shares = MAX_SHARES;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008487
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008488 mutex_lock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008489 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008490 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008491
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008492 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008493 for_each_possible_cpu(i)
8494 unregister_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008495 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008496 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01008497
8498 /* wait for any ongoing reference to this group to finish */
8499 synchronize_sched();
8500
8501 /*
8502 * Now we are free to modify the group's share on each cpu
8503 * w/o tripping rebalance_share or load_balance_fair.
8504 */
8505 tg->shares = shares;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008506 for_each_possible_cpu(i) {
8507 /*
8508 * force a rebalance
8509 */
8510 cfs_rq_set_shares(tg->cfs_rq[i], 0);
Miao Xiecb4ad1f2008-04-28 12:54:56 +08008511 set_se_shares(tg->se[i], shares);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008512 }
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01008513
8514 /*
8515 * Enable load balance activity on this group, by inserting it back on
8516 * each cpu's rq->leaf_cfs_rq_list.
8517 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008518 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008519 for_each_possible_cpu(i)
8520 register_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008521 list_add_rcu(&tg->siblings, &tg->parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008522 spin_unlock_irqrestore(&task_group_lock, flags);
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008523done:
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008524 mutex_unlock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008525 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008526}
8527
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008528unsigned long sched_group_shares(struct task_group *tg)
8529{
8530 return tg->shares;
8531}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008532#endif
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008533
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008534#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008535/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008536 * Ensure that the real time constraints are schedulable.
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008537 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008538static DEFINE_MUTEX(rt_constraints_mutex);
8539
8540static unsigned long to_ratio(u64 period, u64 runtime)
8541{
8542 if (runtime == RUNTIME_INF)
8543 return 1ULL << 16;
8544
Roman Zippel6f6d6a12008-05-01 04:34:28 -07008545 return div64_u64(runtime << 16, period);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008546}
8547
Peter Zijlstrab40b2e82008-04-19 19:45:00 +02008548#ifdef CONFIG_CGROUP_SCHED
8549static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8550{
Peter Zijlstra10b612f2008-06-19 14:22:27 +02008551 struct task_group *tgi, *parent = tg->parent;
Peter Zijlstrab40b2e82008-04-19 19:45:00 +02008552 unsigned long total = 0;
8553
8554 if (!parent) {
8555 if (global_rt_period() < period)
8556 return 0;
8557
8558 return to_ratio(period, runtime) <
8559 to_ratio(global_rt_period(), global_rt_runtime());
8560 }
8561
8562 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8563 return 0;
8564
8565 rcu_read_lock();
8566 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8567 if (tgi == tg)
8568 continue;
8569
8570 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8571 tgi->rt_bandwidth.rt_runtime);
8572 }
8573 rcu_read_unlock();
8574
Peter Zijlstra10b612f2008-06-19 14:22:27 +02008575 return total + to_ratio(period, runtime) <=
Peter Zijlstrab40b2e82008-04-19 19:45:00 +02008576 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8577 parent->rt_bandwidth.rt_runtime);
8578}
8579#elif defined CONFIG_USER_SCHED
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008580static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008581{
8582 struct task_group *tgi;
8583 unsigned long total = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008584 unsigned long global_ratio =
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008585 to_ratio(global_rt_period(), global_rt_runtime());
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008586
8587 rcu_read_lock();
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008588 list_for_each_entry_rcu(tgi, &task_groups, list) {
8589 if (tgi == tg)
8590 continue;
8591
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008592 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8593 tgi->rt_bandwidth.rt_runtime);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008594 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008595 rcu_read_unlock();
8596
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008597 return total + to_ratio(period, runtime) < global_ratio;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008598}
Peter Zijlstrab40b2e82008-04-19 19:45:00 +02008599#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008600
Dhaval Giani521f1a242008-02-28 15:21:56 +05308601/* Must be called with tasklist_lock held */
8602static inline int tg_has_rt_tasks(struct task_group *tg)
8603{
8604 struct task_struct *g, *p;
8605 do_each_thread(g, p) {
8606 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8607 return 1;
8608 } while_each_thread(g, p);
8609 return 0;
8610}
8611
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008612static int tg_set_bandwidth(struct task_group *tg,
8613 u64 rt_period, u64 rt_runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008614{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008615 int i, err = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008616
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008617 mutex_lock(&rt_constraints_mutex);
Dhaval Giani521f1a242008-02-28 15:21:56 +05308618 read_lock(&tasklist_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008619 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
Dhaval Giani521f1a242008-02-28 15:21:56 +05308620 err = -EBUSY;
8621 goto unlock;
8622 }
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008623 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8624 err = -EINVAL;
8625 goto unlock;
8626 }
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008627
8628 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008629 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8630 tg->rt_bandwidth.rt_runtime = rt_runtime;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008631
8632 for_each_possible_cpu(i) {
8633 struct rt_rq *rt_rq = tg->rt_rq[i];
8634
8635 spin_lock(&rt_rq->rt_runtime_lock);
8636 rt_rq->rt_runtime = rt_runtime;
8637 spin_unlock(&rt_rq->rt_runtime_lock);
8638 }
8639 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008640 unlock:
Dhaval Giani521f1a242008-02-28 15:21:56 +05308641 read_unlock(&tasklist_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008642 mutex_unlock(&rt_constraints_mutex);
8643
8644 return err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008645}
8646
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008647int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8648{
8649 u64 rt_runtime, rt_period;
8650
8651 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8652 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8653 if (rt_runtime_us < 0)
8654 rt_runtime = RUNTIME_INF;
8655
8656 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8657}
8658
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008659long sched_group_rt_runtime(struct task_group *tg)
8660{
8661 u64 rt_runtime_us;
8662
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008663 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008664 return -1;
8665
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008666 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008667 do_div(rt_runtime_us, NSEC_PER_USEC);
8668 return rt_runtime_us;
8669}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008670
8671int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8672{
8673 u64 rt_runtime, rt_period;
8674
8675 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8676 rt_runtime = tg->rt_bandwidth.rt_runtime;
8677
8678 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8679}
8680
8681long sched_group_rt_period(struct task_group *tg)
8682{
8683 u64 rt_period_us;
8684
8685 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8686 do_div(rt_period_us, NSEC_PER_USEC);
8687 return rt_period_us;
8688}
8689
8690static int sched_rt_global_constraints(void)
8691{
Peter Zijlstra10b612f2008-06-19 14:22:27 +02008692 struct task_group *tg = &root_task_group;
8693 u64 rt_runtime, rt_period;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008694 int ret = 0;
8695
Peter Zijlstra10b612f2008-06-19 14:22:27 +02008696 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8697 rt_runtime = tg->rt_bandwidth.rt_runtime;
8698
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008699 mutex_lock(&rt_constraints_mutex);
Peter Zijlstra10b612f2008-06-19 14:22:27 +02008700 if (!__rt_schedulable(tg, rt_period, rt_runtime))
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008701 ret = -EINVAL;
8702 mutex_unlock(&rt_constraints_mutex);
8703
8704 return ret;
8705}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008706#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008707static int sched_rt_global_constraints(void)
8708{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008709 unsigned long flags;
8710 int i;
8711
8712 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8713 for_each_possible_cpu(i) {
8714 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8715
8716 spin_lock(&rt_rq->rt_runtime_lock);
8717 rt_rq->rt_runtime = global_rt_runtime();
8718 spin_unlock(&rt_rq->rt_runtime_lock);
8719 }
8720 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8721
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008722 return 0;
8723}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008724#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008725
8726int sched_rt_handler(struct ctl_table *table, int write,
8727 struct file *filp, void __user *buffer, size_t *lenp,
8728 loff_t *ppos)
8729{
8730 int ret;
8731 int old_period, old_runtime;
8732 static DEFINE_MUTEX(mutex);
8733
8734 mutex_lock(&mutex);
8735 old_period = sysctl_sched_rt_period;
8736 old_runtime = sysctl_sched_rt_runtime;
8737
8738 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8739
8740 if (!ret && write) {
8741 ret = sched_rt_global_constraints();
8742 if (ret) {
8743 sysctl_sched_rt_period = old_period;
8744 sysctl_sched_rt_runtime = old_runtime;
8745 } else {
8746 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8747 def_rt_bandwidth.rt_period =
8748 ns_to_ktime(global_rt_period());
8749 }
8750 }
8751 mutex_unlock(&mutex);
8752
8753 return ret;
8754}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008755
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008756#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008757
8758/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02008759static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008760{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008761 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8762 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008763}
8764
8765static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +02008766cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008767{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008768 struct task_group *tg, *parent;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008769
Paul Menage2b01dfe2007-10-24 18:23:50 +02008770 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008771 /* This is early initialization for the top cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02008772 init_task_group.css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008773 return &init_task_group.css;
8774 }
8775
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008776 parent = cgroup_tg(cgrp->parent);
8777 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008778 if (IS_ERR(tg))
8779 return ERR_PTR(-ENOMEM);
8780
8781 /* Bind the cgroup to task_group object we just created */
Paul Menage2b01dfe2007-10-24 18:23:50 +02008782 tg->css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008783
8784 return &tg->css;
8785}
8786
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008787static void
8788cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008789{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008790 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008791
8792 sched_destroy_group(tg);
8793}
8794
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008795static int
8796cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8797 struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008798{
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008799#ifdef CONFIG_RT_GROUP_SCHED
8800 /* Don't accept realtime tasks when there is no way for them to run */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008801 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008802 return -EINVAL;
8803#else
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008804 /* We don't support RT-tasks being in separate groups */
8805 if (tsk->sched_class != &fair_sched_class)
8806 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008807#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008808
8809 return 0;
8810}
8811
8812static void
Paul Menage2b01dfe2007-10-24 18:23:50 +02008813cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008814 struct cgroup *old_cont, struct task_struct *tsk)
8815{
8816 sched_move_task(tsk);
8817}
8818
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008819#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagef4c753b2008-04-29 00:59:56 -07008820static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
Paul Menage2b01dfe2007-10-24 18:23:50 +02008821 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008822{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008823 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008824}
8825
Paul Menagef4c753b2008-04-29 00:59:56 -07008826static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008827{
Paul Menage2b01dfe2007-10-24 18:23:50 +02008828 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008829
8830 return (u64) tg->shares;
8831}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008832#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008833
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008834#ifdef CONFIG_RT_GROUP_SCHED
Mirco Tischler0c708142008-05-14 16:05:46 -07008835static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
Paul Menage06ecb272008-04-29 01:00:06 -07008836 s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008837{
Paul Menage06ecb272008-04-29 01:00:06 -07008838 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008839}
8840
Paul Menage06ecb272008-04-29 01:00:06 -07008841static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008842{
Paul Menage06ecb272008-04-29 01:00:06 -07008843 return sched_group_rt_runtime(cgroup_tg(cgrp));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008844}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008845
8846static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8847 u64 rt_period_us)
8848{
8849 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8850}
8851
8852static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8853{
8854 return sched_group_rt_period(cgroup_tg(cgrp));
8855}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008856#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008857
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008858static struct cftype cpu_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008859#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008860 {
8861 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -07008862 .read_u64 = cpu_shares_read_u64,
8863 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008864 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008865#endif
8866#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008867 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008868 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -07008869 .read_s64 = cpu_rt_runtime_read,
8870 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008871 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008872 {
8873 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -07008874 .read_u64 = cpu_rt_period_read_uint,
8875 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008876 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008877#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008878};
8879
8880static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8881{
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008882 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008883}
8884
8885struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +01008886 .name = "cpu",
8887 .create = cpu_cgroup_create,
8888 .destroy = cpu_cgroup_destroy,
8889 .can_attach = cpu_cgroup_can_attach,
8890 .attach = cpu_cgroup_attach,
8891 .populate = cpu_cgroup_populate,
8892 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008893 .early_init = 1,
8894};
8895
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008896#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008897
8898#ifdef CONFIG_CGROUP_CPUACCT
8899
8900/*
8901 * CPU accounting code for task groups.
8902 *
8903 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8904 * (balbir@in.ibm.com).
8905 */
8906
8907/* track cpu usage of a group of tasks */
8908struct cpuacct {
8909 struct cgroup_subsys_state css;
8910 /* cpuusage holds pointer to a u64-type object on every cpu */
8911 u64 *cpuusage;
8912};
8913
8914struct cgroup_subsys cpuacct_subsys;
8915
8916/* return cpu accounting group corresponding to this container */
Dhaval Giani32cd7562008-02-29 10:02:43 +05308917static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008918{
Dhaval Giani32cd7562008-02-29 10:02:43 +05308919 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008920 struct cpuacct, css);
8921}
8922
8923/* return cpu accounting group to which this task belongs */
8924static inline struct cpuacct *task_ca(struct task_struct *tsk)
8925{
8926 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8927 struct cpuacct, css);
8928}
8929
8930/* create a new cpu accounting group */
8931static struct cgroup_subsys_state *cpuacct_create(
Dhaval Giani32cd7562008-02-29 10:02:43 +05308932 struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008933{
8934 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8935
8936 if (!ca)
8937 return ERR_PTR(-ENOMEM);
8938
8939 ca->cpuusage = alloc_percpu(u64);
8940 if (!ca->cpuusage) {
8941 kfree(ca);
8942 return ERR_PTR(-ENOMEM);
8943 }
8944
8945 return &ca->css;
8946}
8947
8948/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008949static void
Dhaval Giani32cd7562008-02-29 10:02:43 +05308950cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008951{
Dhaval Giani32cd7562008-02-29 10:02:43 +05308952 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008953
8954 free_percpu(ca->cpuusage);
8955 kfree(ca);
8956}
8957
8958/* return total cpu usage (in nanoseconds) of a group */
Dhaval Giani32cd7562008-02-29 10:02:43 +05308959static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008960{
Dhaval Giani32cd7562008-02-29 10:02:43 +05308961 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008962 u64 totalcpuusage = 0;
8963 int i;
8964
8965 for_each_possible_cpu(i) {
8966 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8967
8968 /*
8969 * Take rq->lock to make 64-bit addition safe on 32-bit
8970 * platforms.
8971 */
8972 spin_lock_irq(&cpu_rq(i)->lock);
8973 totalcpuusage += *cpuusage;
8974 spin_unlock_irq(&cpu_rq(i)->lock);
8975 }
8976
8977 return totalcpuusage;
8978}
8979
Dhaval Giani0297b802008-02-29 10:02:44 +05308980static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8981 u64 reset)
8982{
8983 struct cpuacct *ca = cgroup_ca(cgrp);
8984 int err = 0;
8985 int i;
8986
8987 if (reset) {
8988 err = -EINVAL;
8989 goto out;
8990 }
8991
8992 for_each_possible_cpu(i) {
8993 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8994
8995 spin_lock_irq(&cpu_rq(i)->lock);
8996 *cpuusage = 0;
8997 spin_unlock_irq(&cpu_rq(i)->lock);
8998 }
8999out:
9000 return err;
9001}
9002
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009003static struct cftype files[] = {
9004 {
9005 .name = "usage",
Paul Menagef4c753b2008-04-29 00:59:56 -07009006 .read_u64 = cpuusage_read,
9007 .write_u64 = cpuusage_write,
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009008 },
9009};
9010
Dhaval Giani32cd7562008-02-29 10:02:43 +05309011static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009012{
Dhaval Giani32cd7562008-02-29 10:02:43 +05309013 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009014}
9015
9016/*
9017 * charge this task's execution time to its accounting group.
9018 *
9019 * called with rq->lock held.
9020 */
9021static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9022{
9023 struct cpuacct *ca;
9024
9025 if (!cpuacct_subsys.active)
9026 return;
9027
9028 ca = task_ca(tsk);
9029 if (ca) {
9030 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9031
9032 *cpuusage += cputime;
9033 }
9034}
9035
9036struct cgroup_subsys cpuacct_subsys = {
9037 .name = "cpuacct",
9038 .create = cpuacct_create,
9039 .destroy = cpuacct_destroy,
9040 .populate = cpuacct_populate,
9041 .subsys_id = cpuacct_subsys_id,
9042};
9043#endif /* CONFIG_CGROUP_CPUACCT */