blob: 86e55a9c2de60d564f2d0ba5d676aed58c3270bd [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
Linus Torvalds1da177e2005-04-16 15:20:36 -070025 */
26
27#include <linux/mm.h>
28#include <linux/module.h>
29#include <linux/nmi.h>
30#include <linux/init.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020031#include <linux/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include <linux/highmem.h>
33#include <linux/smp_lock.h>
34#include <asm/mmu_context.h>
35#include <linux/interrupt.h>
Randy.Dunlapc59ede72006-01-11 12:17:46 -080036#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070037#include <linux/completion.h>
38#include <linux/kernel_stat.h>
Ingo Molnar9a11b49a2006-07-03 00:24:33 -070039#include <linux/debug_locks.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040#include <linux/security.h>
41#include <linux/notifier.h>
42#include <linux/profile.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080043#include <linux/freezer.h>
akpm@osdl.org198e2f12006-01-12 01:05:30 -080044#include <linux/vmalloc.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070045#include <linux/blkdev.h>
46#include <linux/delay.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070047#include <linux/pid_namespace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070048#include <linux/smp.h>
49#include <linux/threads.h>
50#include <linux/timer.h>
51#include <linux/rcupdate.h>
52#include <linux/cpu.h>
53#include <linux/cpuset.h>
54#include <linux/percpu.h>
55#include <linux/kthread.h>
56#include <linux/seq_file.h>
Nick Piggine692ab52007-07-26 13:40:43 +020057#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070058#include <linux/syscalls.h>
59#include <linux/times.h>
Jay Lan8f0ab512006-09-30 23:28:59 -070060#include <linux/tsacct_kern.h>
bibo maoc6fd91f2006-03-26 01:38:20 -080061#include <linux/kprobes.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070062#include <linux/delayacct.h>
Eric Dumazet5517d862007-05-08 00:32:57 -070063#include <linux/reciprocal_div.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020064#include <linux/unistd.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020065#include <linux/pagemap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070066
Eric Dumazet5517d862007-05-08 00:32:57 -070067#include <asm/tlb.h>
Satyam Sharma838225b2007-10-24 18:23:50 +020068#include <asm/irq_regs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070069
70/*
Alexey Dobriyanb035b6d2007-02-10 01:45:10 -080071 * Scheduler clock - returns current time in nanosec units.
72 * This is default implementation.
73 * Architectures and sub-architectures can override this.
74 */
75unsigned long long __attribute__((weak)) sched_clock(void)
76{
Eric Dumazetd6322fa2007-11-09 22:39:38 +010077 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
Alexey Dobriyanb035b6d2007-02-10 01:45:10 -080078}
79
80/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070081 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 * and back.
84 */
85#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88
89/*
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
93 */
94#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97
98/*
99 * Some helpers for converting nanosecond timing to jiffy resolution
100 */
Eric Dumazetd6322fa2007-11-09 22:39:38 +0100101#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102#define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200104#define NICE_0_LOAD SCHED_LOAD_SCALE
105#define NICE_0_SHIFT SCHED_LOAD_SHIFT
106
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107/*
108 * These are the 'tuning knobs' of the scheduler:
109 *
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +0200110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 * Timeslices get refilled after they expire.
112 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113#define DEF_TIMESLICE (100 * HZ / 1000)
Peter Williams2dd73a42006-06-27 02:54:34 -0700114
Eric Dumazet5517d862007-05-08 00:32:57 -0700115#ifdef CONFIG_SMP
116/*
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
119 */
120static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
121{
122 return reciprocal_divide(load, sg->reciprocal_cpu_power);
123}
124
125/*
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
128 */
129static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
130{
131 sg->__cpu_power += val;
132 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
133}
134#endif
135
Ingo Molnare05606d2007-07-09 18:51:59 +0200136static inline int rt_policy(int policy)
137{
138 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
139 return 1;
140 return 0;
141}
142
143static inline int task_has_rt_policy(struct task_struct *p)
144{
145 return rt_policy(p->policy);
146}
147
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148/*
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200149 * This is the priority-queue data structure of the RT scheduling class:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200151struct rt_prio_array {
152 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
153 struct list_head queue[MAX_RT_PRIO];
154};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200156#ifdef CONFIG_FAIR_GROUP_SCHED
157
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700158#include <linux/cgroup.h>
159
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200160struct cfs_rq;
161
162/* task group related information */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200163struct task_group {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700164#ifdef CONFIG_FAIR_CGROUP_SCHED
165 struct cgroup_subsys_state css;
166#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200167 /* schedulable entities of this group on each cpu */
168 struct sched_entity **se;
169 /* runqueue "owned" by this group on each cpu */
170 struct cfs_rq **cfs_rq;
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100171
172 /*
173 * shares assigned to a task group governs how much of cpu bandwidth
174 * is allocated to the group. The more shares a group has, the more is
175 * the cpu bandwidth allocated to it.
176 *
177 * For ex, lets say that there are three task groups, A, B and C which
178 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
179 * cpu bandwidth allocated by the scheduler to task groups A, B and C
180 * should be:
181 *
182 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
183 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
184 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
185 *
186 * The weight assigned to a task group's schedulable entities on every
187 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
188 * group's shares. For ex: lets say that task group A has been
189 * assigned shares of 1000 and there are two CPUs in a system. Then,
190 *
191 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
192 *
193 * Note: It's not necessary that each of a task's group schedulable
194 * entity have the same weight on all CPUs. If the group
195 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
196 * better distribution of weight could be:
197 *
198 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
199 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
200 *
201 * rebalance_shares() is responsible for distributing the shares of a
202 * task groups like this among the group's schedulable entities across
203 * cpus.
204 *
205 */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200206 unsigned long shares;
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100207
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +0100208 struct rcu_head rcu;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200209};
210
211/* Default task group's sched entity on each cpu */
212static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
213/* Default task group's cfs_rq on each cpu */
214static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
215
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200216static struct sched_entity *init_sched_entity_p[NR_CPUS];
217static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200218
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100219/* task_group_mutex serializes add/remove of task groups and also changes to
220 * a task group's cpu shares.
221 */
222static DEFINE_MUTEX(task_group_mutex);
223
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +0100224/* doms_cur_mutex serializes access to doms_cur[] array */
225static DEFINE_MUTEX(doms_cur_mutex);
226
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100227#ifdef CONFIG_SMP
228/* kernel thread that runs rebalance_shares() periodically */
229static struct task_struct *lb_monitor_task;
230static int load_balance_monitor(void *unused);
231#endif
232
233static void set_se_shares(struct sched_entity *se, unsigned long shares);
234
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200235/* Default task group.
Ingo Molnar3a252012007-10-15 17:00:12 +0200236 * Every task in system belong to this group at bootup.
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200237 */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200238struct task_group init_task_group = {
Ingo Molnar3a252012007-10-15 17:00:12 +0200239 .se = init_sched_entity_p,
240 .cfs_rq = init_cfs_rq_p,
241};
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200242
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200243#ifdef CONFIG_FAIR_USER_SCHED
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100244# define INIT_TASK_GROUP_LOAD 2*NICE_0_LOAD
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200245#else
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100246# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200247#endif
248
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100249#define MIN_GROUP_SHARES 2
250
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100251static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200252
253/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200254static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200255{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200256 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200257
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200258#ifdef CONFIG_FAIR_USER_SCHED
259 tg = p->user->tg;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700260#elif defined(CONFIG_FAIR_CGROUP_SCHED)
261 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
262 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200263#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100264 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200265#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200266 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200267}
268
269/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100270static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200271{
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100272 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
273 p->se.parent = task_group(p)->se[cpu];
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200274}
275
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100276static inline void lock_task_group_list(void)
277{
278 mutex_lock(&task_group_mutex);
279}
280
281static inline void unlock_task_group_list(void)
282{
283 mutex_unlock(&task_group_mutex);
284}
285
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +0100286static inline void lock_doms_cur(void)
287{
288 mutex_lock(&doms_cur_mutex);
289}
290
291static inline void unlock_doms_cur(void)
292{
293 mutex_unlock(&doms_cur_mutex);
294}
295
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200296#else
297
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100298static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { }
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100299static inline void lock_task_group_list(void) { }
300static inline void unlock_task_group_list(void) { }
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +0100301static inline void lock_doms_cur(void) { }
302static inline void unlock_doms_cur(void) { }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200303
304#endif /* CONFIG_FAIR_GROUP_SCHED */
305
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200306/* CFS-related fields in a runqueue */
307struct cfs_rq {
308 struct load_weight load;
309 unsigned long nr_running;
310
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200311 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200312 u64 min_vruntime;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200313
314 struct rb_root tasks_timeline;
315 struct rb_node *rb_leftmost;
316 struct rb_node *rb_load_balance_curr;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200317 /* 'curr' points to currently running entity on this cfs_rq.
318 * It is set to NULL otherwise (i.e when none are currently running).
319 */
320 struct sched_entity *curr;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200321
322 unsigned long nr_spread_over;
323
Ingo Molnar62160e3f2007-10-15 17:00:03 +0200324#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200325 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
326
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100327 /*
328 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200329 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
330 * (like users, containers etc.)
331 *
332 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
333 * list is used during load balance.
334 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100335 struct list_head leaf_cfs_rq_list;
336 struct task_group *tg; /* group that "owns" this runqueue */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200337#endif
338};
339
340/* Real-Time classes' related field in a runqueue: */
341struct rt_rq {
342 struct rt_prio_array active;
343 int rt_load_balance_idx;
344 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
345};
346
347/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 * This is the main, per-CPU runqueue data structure.
349 *
350 * Locking rule: those places that want to lock multiple runqueues
351 * (such as the load balancing or the thread migration code), lock
352 * acquire operations must be ordered by ascending &runqueue.
353 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700354struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200355 /* runqueue lock: */
356 spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357
358 /*
359 * nr_running and cpu_load should be in the same cacheline because
360 * remote CPUs use both these fields when doing load calculation.
361 */
362 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200363 #define CPU_LOAD_IDX_MAX 5
364 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh Bbdecea32007-05-08 00:32:48 -0700365 unsigned char idle_at_tick;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700366#ifdef CONFIG_NO_HZ
367 unsigned char in_nohz_recently;
368#endif
Ingo Molnard8016492007-10-18 21:32:55 +0200369 /* capture load from *all* tasks on this cpu: */
370 struct load_weight load;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200371 unsigned long nr_load_updates;
372 u64 nr_switches;
373
374 struct cfs_rq cfs;
375#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200376 /* list of leaf cfs_rq on this cpu: */
377 struct list_head leaf_cfs_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378#endif
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100379 struct rt_rq rt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380
381 /*
382 * This is part of a global counter where only the total sum
383 * over all CPUs matters. A task can increase this counter on
384 * one CPU and if it got migrated afterwards it may decrease
385 * it on another CPU. Always updated under the runqueue lock:
386 */
387 unsigned long nr_uninterruptible;
388
Ingo Molnar36c8b582006-07-03 00:25:41 -0700389 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800390 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200392
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200393 u64 clock, prev_clock_raw;
394 s64 clock_max_delta;
395
396 unsigned int clock_warps, clock_overflows;
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200397 u64 idle_clock;
398 unsigned int clock_deep_idle_events;
Ingo Molnar529c7722007-08-10 23:05:11 +0200399 u64 tick_timestamp;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200400
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 atomic_t nr_iowait;
402
403#ifdef CONFIG_SMP
404 struct sched_domain *sd;
405
406 /* For active balancing */
407 int active_balance;
408 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200409 /* cpu of this runqueue: */
410 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Ingo Molnar36c8b582006-07-03 00:25:41 -0700412 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 struct list_head migration_queue;
414#endif
415
416#ifdef CONFIG_SCHEDSTATS
417 /* latency stats */
418 struct sched_info rq_sched_info;
419
420 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200421 unsigned int yld_exp_empty;
422 unsigned int yld_act_empty;
423 unsigned int yld_both_empty;
424 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425
426 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200427 unsigned int sched_switch;
428 unsigned int sched_count;
429 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430
431 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200432 unsigned int ttwu_count;
433 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200434
435 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200436 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437#endif
Ingo Molnarfcb99372006-07-03 00:25:10 -0700438 struct lock_class_key rq_lock_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439};
440
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700441static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Gautham R Shenoy5be93612007-05-09 02:34:04 -0700442static DEFINE_MUTEX(sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
Ingo Molnardd41f592007-07-09 18:51:59 +0200444static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
445{
446 rq->curr->sched_class->check_preempt_curr(rq, p);
447}
448
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700449static inline int cpu_of(struct rq *rq)
450{
451#ifdef CONFIG_SMP
452 return rq->cpu;
453#else
454 return 0;
455#endif
456}
457
Nick Piggin674311d2005-06-25 14:57:27 -0700458/*
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200459 * Update the per-runqueue clock, as finegrained as the platform can give
460 * us, but without assuming monotonicity, etc.:
Ingo Molnar20d315d2007-07-09 18:51:58 +0200461 */
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200462static void __update_rq_clock(struct rq *rq)
Ingo Molnar20d315d2007-07-09 18:51:58 +0200463{
464 u64 prev_raw = rq->prev_clock_raw;
465 u64 now = sched_clock();
466 s64 delta = now - prev_raw;
467 u64 clock = rq->clock;
468
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200469#ifdef CONFIG_SCHED_DEBUG
470 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
471#endif
Ingo Molnar20d315d2007-07-09 18:51:58 +0200472 /*
473 * Protect against sched_clock() occasionally going backwards:
474 */
475 if (unlikely(delta < 0)) {
476 clock++;
477 rq->clock_warps++;
478 } else {
479 /*
480 * Catch too large forward jumps too:
481 */
Ingo Molnar529c7722007-08-10 23:05:11 +0200482 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
483 if (clock < rq->tick_timestamp + TICK_NSEC)
484 clock = rq->tick_timestamp + TICK_NSEC;
485 else
486 clock++;
Ingo Molnar20d315d2007-07-09 18:51:58 +0200487 rq->clock_overflows++;
488 } else {
489 if (unlikely(delta > rq->clock_max_delta))
490 rq->clock_max_delta = delta;
491 clock += delta;
492 }
493 }
494
495 rq->prev_clock_raw = now;
496 rq->clock = clock;
Ingo Molnar20d315d2007-07-09 18:51:58 +0200497}
498
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200499static void update_rq_clock(struct rq *rq)
Ingo Molnar20d315d2007-07-09 18:51:58 +0200500{
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200501 if (likely(smp_processor_id() == cpu_of(rq)))
502 __update_rq_clock(rq);
503}
Ingo Molnar20d315d2007-07-09 18:51:58 +0200504
Ingo Molnar20d315d2007-07-09 18:51:58 +0200505/*
Nick Piggin674311d2005-06-25 14:57:27 -0700506 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700507 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700508 *
509 * The domain tree of any CPU may only be accessed from within
510 * preempt-disabled sections.
511 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700512#define for_each_domain(cpu, __sd) \
513 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514
515#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
516#define this_rq() (&__get_cpu_var(runqueues))
517#define task_rq(p) cpu_rq(task_cpu(p))
518#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
519
Ingo Molnare436d802007-07-19 21:28:35 +0200520/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200521 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
522 */
523#ifdef CONFIG_SCHED_DEBUG
524# define const_debug __read_mostly
525#else
526# define const_debug static const
527#endif
528
529/*
530 * Debugging: various feature bits
531 */
532enum {
Ingo Molnarbbdba7c2007-10-15 17:00:06 +0200533 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
Ingo Molnar96126332007-11-15 20:57:40 +0100534 SCHED_FEAT_WAKEUP_PREEMPT = 2,
535 SCHED_FEAT_START_DEBIT = 4,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100536 SCHED_FEAT_TREE_AVG = 8,
537 SCHED_FEAT_APPROX_AVG = 16,
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200538};
539
540const_debug unsigned int sysctl_sched_features =
Ingo Molnar8401f772007-10-18 21:32:55 +0200541 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
Ingo Molnar96126332007-11-15 20:57:40 +0100542 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
Ingo Molnar8401f772007-10-18 21:32:55 +0200543 SCHED_FEAT_START_DEBIT * 1 |
544 SCHED_FEAT_TREE_AVG * 0 |
Ingo Molnar96126332007-11-15 20:57:40 +0100545 SCHED_FEAT_APPROX_AVG * 0;
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200546
547#define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
548
549/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100550 * Number of tasks to iterate in a single balance run.
551 * Limited because this is done with IRQs disabled.
552 */
553const_debug unsigned int sysctl_sched_nr_migrate = 32;
554
555/*
Ingo Molnare436d802007-07-19 21:28:35 +0200556 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
557 * clock constructed from sched_clock():
558 */
559unsigned long long cpu_clock(int cpu)
560{
Ingo Molnare436d802007-07-19 21:28:35 +0200561 unsigned long long now;
562 unsigned long flags;
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200563 struct rq *rq;
Ingo Molnare436d802007-07-19 21:28:35 +0200564
Ingo Molnar2cd4d0e2007-07-26 13:40:43 +0200565 local_irq_save(flags);
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200566 rq = cpu_rq(cpu);
Ingo Molnar8ced5f62007-12-07 19:02:47 +0100567 /*
568 * Only call sched_clock() if the scheduler has already been
569 * initialized (some code might call cpu_clock() very early):
570 */
571 if (rq->idle)
572 update_rq_clock(rq);
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200573 now = rq->clock;
Ingo Molnar2cd4d0e2007-07-26 13:40:43 +0200574 local_irq_restore(flags);
Ingo Molnare436d802007-07-19 21:28:35 +0200575
576 return now;
577}
Paul E. McKenneya58f6f22007-10-15 17:00:14 +0200578EXPORT_SYMBOL_GPL(cpu_clock);
Ingo Molnare436d802007-07-19 21:28:35 +0200579
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700581# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700583#ifndef finish_arch_switch
584# define finish_arch_switch(prev) do { } while (0)
585#endif
586
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100587static inline int task_current(struct rq *rq, struct task_struct *p)
588{
589 return rq->curr == p;
590}
591
Nick Piggin4866cde2005-06-25 14:57:23 -0700592#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700593static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700594{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100595 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700596}
597
Ingo Molnar70b97a72006-07-03 00:25:42 -0700598static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700599{
600}
601
Ingo Molnar70b97a72006-07-03 00:25:42 -0700602static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700603{
Ingo Molnarda04c032005-09-13 11:17:59 +0200604#ifdef CONFIG_DEBUG_SPINLOCK
605 /* this is a valid case when another task releases the spinlock */
606 rq->lock.owner = current;
607#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700608 /*
609 * If we are tracking spinlock dependencies then we have to
610 * fix up the runqueue lock - which gets 'carried over' from
611 * prev into current:
612 */
613 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
614
Nick Piggin4866cde2005-06-25 14:57:23 -0700615 spin_unlock_irq(&rq->lock);
616}
617
618#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700619static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700620{
621#ifdef CONFIG_SMP
622 return p->oncpu;
623#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100624 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700625#endif
626}
627
Ingo Molnar70b97a72006-07-03 00:25:42 -0700628static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700629{
630#ifdef CONFIG_SMP
631 /*
632 * We can optimise this out completely for !SMP, because the
633 * SMP rebalancing from interrupt is the only thing that cares
634 * here.
635 */
636 next->oncpu = 1;
637#endif
638#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
639 spin_unlock_irq(&rq->lock);
640#else
641 spin_unlock(&rq->lock);
642#endif
643}
644
Ingo Molnar70b97a72006-07-03 00:25:42 -0700645static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700646{
647#ifdef CONFIG_SMP
648 /*
649 * After ->oncpu is cleared, the task can be moved to a different CPU.
650 * We must ensure this doesn't happen until the switch is completely
651 * finished.
652 */
653 smp_wmb();
654 prev->oncpu = 0;
655#endif
656#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
657 local_irq_enable();
658#endif
659}
660#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661
662/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700663 * __task_rq_lock - lock the runqueue a given task resides on.
664 * Must be called interrupts disabled.
665 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700666static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700667 __acquires(rq->lock)
668{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200669 for (;;) {
670 struct rq *rq = task_rq(p);
671 spin_lock(&rq->lock);
672 if (likely(rq == task_rq(p)))
673 return rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700674 spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700675 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700676}
677
678/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100680 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 * explicitly disabling preemption.
682 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700683static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 __acquires(rq->lock)
685{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700686 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687
Andi Kleen3a5c3592007-10-15 17:00:14 +0200688 for (;;) {
689 local_irq_save(*flags);
690 rq = task_rq(p);
691 spin_lock(&rq->lock);
692 if (likely(rq == task_rq(p)))
693 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696}
697
Alexey Dobriyana9957442007-10-15 17:00:13 +0200698static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700699 __releases(rq->lock)
700{
701 spin_unlock(&rq->lock);
702}
703
Ingo Molnar70b97a72006-07-03 00:25:42 -0700704static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 __releases(rq->lock)
706{
707 spin_unlock_irqrestore(&rq->lock, *flags);
708}
709
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -0800711 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 */
Alexey Dobriyana9957442007-10-15 17:00:13 +0200713static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 __acquires(rq->lock)
715{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700716 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717
718 local_irq_disable();
719 rq = this_rq();
720 spin_lock(&rq->lock);
721
722 return rq;
723}
724
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200725/*
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200726 * We are going deep-idle (irqs are disabled):
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200727 */
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200728void sched_clock_idle_sleep_event(void)
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200729{
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200730 struct rq *rq = cpu_rq(smp_processor_id());
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200731
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200732 spin_lock(&rq->lock);
733 __update_rq_clock(rq);
734 spin_unlock(&rq->lock);
735 rq->clock_deep_idle_events++;
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200736}
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200737EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
738
739/*
740 * We just idled delta nanoseconds (called with irqs disabled):
741 */
742void sched_clock_idle_wakeup_event(u64 delta_ns)
743{
744 struct rq *rq = cpu_rq(smp_processor_id());
745 u64 now = sched_clock();
746
Ingo Molnar2bacec82007-12-18 15:21:13 +0100747 touch_softlockup_watchdog();
Ingo Molnar2aa44d02007-08-23 15:18:02 +0200748 rq->idle_clock += delta_ns;
749 /*
750 * Override the previous timestamp and ignore all
751 * sched_clock() deltas that occured while we idled,
752 * and use the PM-provided delta_ns to advance the
753 * rq clock:
754 */
755 spin_lock(&rq->lock);
756 rq->prev_clock_raw = now;
757 rq->clock += delta_ns;
758 spin_unlock(&rq->lock);
759}
760EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200761
762/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200763 * resched_task - mark a task 'to be rescheduled now'.
764 *
765 * On UP this means the setting of the need_resched flag, on SMP it
766 * might also involve a cross-CPU call to trigger the scheduler on
767 * the target CPU.
768 */
769#ifdef CONFIG_SMP
770
771#ifndef tsk_is_polling
772#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
773#endif
774
775static void resched_task(struct task_struct *p)
776{
777 int cpu;
778
779 assert_spin_locked(&task_rq(p)->lock);
780
781 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
782 return;
783
784 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
785
786 cpu = task_cpu(p);
787 if (cpu == smp_processor_id())
788 return;
789
790 /* NEED_RESCHED must be visible before we test polling */
791 smp_mb();
792 if (!tsk_is_polling(p))
793 smp_send_reschedule(cpu);
794}
795
796static void resched_cpu(int cpu)
797{
798 struct rq *rq = cpu_rq(cpu);
799 unsigned long flags;
800
801 if (!spin_trylock_irqsave(&rq->lock, flags))
802 return;
803 resched_task(cpu_curr(cpu));
804 spin_unlock_irqrestore(&rq->lock, flags);
805}
806#else
807static inline void resched_task(struct task_struct *p)
808{
809 assert_spin_locked(&task_rq(p)->lock);
810 set_tsk_need_resched(p);
811}
812#endif
813
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200814#if BITS_PER_LONG == 32
815# define WMULT_CONST (~0UL)
816#else
817# define WMULT_CONST (1UL << 32)
818#endif
819
820#define WMULT_SHIFT 32
821
Ingo Molnar194081e2007-08-09 11:16:51 +0200822/*
823 * Shift right and round:
824 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +0200825#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +0200826
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +0200827static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200828calc_delta_mine(unsigned long delta_exec, unsigned long weight,
829 struct load_weight *lw)
830{
831 u64 tmp;
832
833 if (unlikely(!lw->inv_weight))
Ingo Molnar194081e2007-08-09 11:16:51 +0200834 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200835
836 tmp = (u64)delta_exec * weight;
837 /*
838 * Check whether we'd overflow the 64-bit multiplication:
839 */
Ingo Molnar194081e2007-08-09 11:16:51 +0200840 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +0200841 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +0200842 WMULT_SHIFT/2);
843 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +0200844 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200845
Ingo Molnarecf691d2007-08-02 17:41:40 +0200846 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200847}
848
849static inline unsigned long
850calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
851{
852 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
853}
854
Ingo Molnar10919852007-10-15 17:00:04 +0200855static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200856{
857 lw->weight += inc;
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200858}
859
Ingo Molnar10919852007-10-15 17:00:04 +0200860static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200861{
862 lw->weight -= dec;
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200863}
864
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865/*
Peter Williams2dd73a42006-06-27 02:54:34 -0700866 * To aid in avoiding the subversion of "niceness" due to uneven distribution
867 * of tasks with abnormal "nice" values across CPUs the contribution that
868 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100869 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -0700870 * scaled version of the new time slice allocation that they receive on time
871 * slice expiry etc.
872 */
873
Ingo Molnardd41f592007-07-09 18:51:59 +0200874#define WEIGHT_IDLEPRIO 2
875#define WMULT_IDLEPRIO (1 << 31)
876
877/*
878 * Nice levels are multiplicative, with a gentle 10% change for every
879 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
880 * nice 1, it will get ~10% less CPU time than another CPU-bound task
881 * that remained on nice 0.
882 *
883 * The "10% effect" is relative and cumulative: from _any_ nice level,
884 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +0200885 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
886 * If a task goes up by ~10% and another task goes down by ~10% then
887 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +0200888 */
889static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +0200890 /* -20 */ 88761, 71755, 56483, 46273, 36291,
891 /* -15 */ 29154, 23254, 18705, 14949, 11916,
892 /* -10 */ 9548, 7620, 6100, 4904, 3906,
893 /* -5 */ 3121, 2501, 1991, 1586, 1277,
894 /* 0 */ 1024, 820, 655, 526, 423,
895 /* 5 */ 335, 272, 215, 172, 137,
896 /* 10 */ 110, 87, 70, 56, 45,
897 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +0200898};
899
Ingo Molnar5714d2d2007-07-16 09:46:31 +0200900/*
901 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
902 *
903 * In cases where the weight does not change often, we can use the
904 * precalculated inverse to speed up arithmetics by turning divisions
905 * into multiplications:
906 */
Ingo Molnardd41f592007-07-09 18:51:59 +0200907static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +0200908 /* -20 */ 48388, 59856, 76040, 92818, 118348,
909 /* -15 */ 147320, 184698, 229616, 287308, 360437,
910 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
911 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
912 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
913 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
914 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
915 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +0200916};
Peter Williams2dd73a42006-06-27 02:54:34 -0700917
Ingo Molnardd41f592007-07-09 18:51:59 +0200918static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
919
920/*
921 * runqueue iterator, to support SMP load-balancing between different
922 * scheduling classes, without having to expose their internal data
923 * structures to the load-balancing proper:
924 */
925struct rq_iterator {
926 void *arg;
927 struct task_struct *(*start)(void *);
928 struct task_struct *(*next)(void *);
929};
930
Peter Williamse1d14842007-10-24 18:23:51 +0200931#ifdef CONFIG_SMP
932static unsigned long
933balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
934 unsigned long max_load_move, struct sched_domain *sd,
935 enum cpu_idle_type idle, int *all_pinned,
936 int *this_best_prio, struct rq_iterator *iterator);
937
938static int
939iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
940 struct sched_domain *sd, enum cpu_idle_type idle,
941 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +0200942#endif
Ingo Molnardd41f592007-07-09 18:51:59 +0200943
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100944#ifdef CONFIG_CGROUP_CPUACCT
945static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
946#else
947static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
948#endif
949
Srivatsa Vaddagiri58e2d4c2008-01-25 21:08:00 +0100950static inline void inc_cpu_load(struct rq *rq, unsigned long load)
951{
952 update_load_add(&rq->load, load);
953}
954
955static inline void dec_cpu_load(struct rq *rq, unsigned long load)
956{
957 update_load_sub(&rq->load, load);
958}
959
Ingo Molnardd41f592007-07-09 18:51:59 +0200960#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +0200961#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +0200962#include "sched_fair.c"
963#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +0200964#ifdef CONFIG_SCHED_DEBUG
965# include "sched_debug.c"
966#endif
967
968#define sched_class_highest (&rt_sched_class)
969
Ingo Molnare5fa2232007-08-09 11:16:49 +0200970static void inc_nr_running(struct task_struct *p, struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +0200971{
972 rq->nr_running++;
Ingo Molnar9c217242007-08-02 17:41:40 +0200973}
974
Ingo Molnardb531812007-08-09 11:16:49 +0200975static void dec_nr_running(struct task_struct *p, struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +0200976{
977 rq->nr_running--;
Ingo Molnar9c217242007-08-02 17:41:40 +0200978}
979
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200980static void set_load_weight(struct task_struct *p)
981{
982 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +0200983 p->se.load.weight = prio_to_weight[0] * 2;
984 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
985 return;
986 }
987
988 /*
989 * SCHED_IDLE tasks get minimal weight:
990 */
991 if (p->policy == SCHED_IDLE) {
992 p->se.load.weight = WEIGHT_IDLEPRIO;
993 p->se.load.inv_weight = WMULT_IDLEPRIO;
994 return;
995 }
996
997 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
998 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200999}
1000
Ingo Molnar8159f872007-08-09 11:16:49 +02001001static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001002{
1003 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001004 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001005 p->se.on_rq = 1;
1006}
1007
Ingo Molnar69be72c2007-08-09 11:16:49 +02001008static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001009{
Ingo Molnarf02231e2007-08-09 11:16:48 +02001010 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001011 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001012}
1013
1014/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001015 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001016 */
Ingo Molnar14531182007-07-09 18:51:59 +02001017static inline int __normal_prio(struct task_struct *p)
1018{
Ingo Molnardd41f592007-07-09 18:51:59 +02001019 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001020}
1021
1022/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001023 * Calculate the expected normal priority: i.e. priority
1024 * without taking RT-inheritance into account. Might be
1025 * boosted by interactivity modifiers. Changes upon fork,
1026 * setprio syscalls, and whenever the interactivity
1027 * estimator recalculates.
1028 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001029static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001030{
1031 int prio;
1032
Ingo Molnare05606d2007-07-09 18:51:59 +02001033 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001034 prio = MAX_RT_PRIO-1 - p->rt_priority;
1035 else
1036 prio = __normal_prio(p);
1037 return prio;
1038}
1039
1040/*
1041 * Calculate the current priority, i.e. the priority
1042 * taken into account by the scheduler. This value might
1043 * be boosted by RT tasks, or might be boosted by
1044 * interactivity modifiers. Will be RT if the task got
1045 * RT-boosted. If not then it returns p->normal_prio.
1046 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001047static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001048{
1049 p->normal_prio = normal_prio(p);
1050 /*
1051 * If we are RT tasks or we were boosted to RT priority,
1052 * keep the priority unchanged. Otherwise, update priority
1053 * to the normal priority:
1054 */
1055 if (!rt_prio(p->prio))
1056 return p->normal_prio;
1057 return p->prio;
1058}
1059
1060/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001061 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001063static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064{
Ingo Molnardd41f592007-07-09 18:51:59 +02001065 if (p->state == TASK_UNINTERRUPTIBLE)
1066 rq->nr_uninterruptible--;
1067
Ingo Molnar8159f872007-08-09 11:16:49 +02001068 enqueue_task(rq, p, wakeup);
Ingo Molnare5fa2232007-08-09 11:16:49 +02001069 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070}
1071
1072/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 * deactivate_task - remove a task from the runqueue.
1074 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001075static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076{
Ingo Molnardd41f592007-07-09 18:51:59 +02001077 if (p->state == TASK_UNINTERRUPTIBLE)
1078 rq->nr_uninterruptible++;
1079
Ingo Molnar69be72c2007-08-09 11:16:49 +02001080 dequeue_task(rq, p, sleep);
Ingo Molnardb531812007-08-09 11:16:49 +02001081 dec_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082}
1083
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084/**
1085 * task_curr - is this task currently executing on a CPU?
1086 * @p: the task in question.
1087 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001088inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089{
1090 return cpu_curr(task_cpu(p)) == p;
1091}
1092
Peter Williams2dd73a42006-06-27 02:54:34 -07001093/* Used instead of source_load when we know the type == 0 */
1094unsigned long weighted_cpuload(const int cpu)
1095{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02001096 return cpu_rq(cpu)->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02001097}
1098
1099static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1100{
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001101 set_task_cfs_rq(p, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001102#ifdef CONFIG_SMP
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001103 /*
1104 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1105 * successfuly executed on another CPU. We must ensure that updates of
1106 * per-task data have been completed by this moment.
1107 */
1108 smp_wmb();
Ingo Molnardd41f592007-07-09 18:51:59 +02001109 task_thread_info(p)->cpu = cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02001110#endif
Peter Williams2dd73a42006-06-27 02:54:34 -07001111}
1112
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113#ifdef CONFIG_SMP
Ingo Molnarc65cc872007-07-09 18:51:58 +02001114
Ingo Molnarcc367732007-10-15 17:00:18 +02001115/*
1116 * Is this task likely cache-hot:
1117 */
1118static inline int
1119task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1120{
1121 s64 delta;
1122
1123 if (p->sched_class != &fair_sched_class)
1124 return 0;
1125
Ingo Molnar6bc16652007-10-15 17:00:18 +02001126 if (sysctl_sched_migration_cost == -1)
1127 return 1;
1128 if (sysctl_sched_migration_cost == 0)
1129 return 0;
1130
Ingo Molnarcc367732007-10-15 17:00:18 +02001131 delta = now - p->se.exec_start;
1132
1133 return delta < (s64)sysctl_sched_migration_cost;
1134}
1135
1136
Ingo Molnardd41f592007-07-09 18:51:59 +02001137void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02001138{
Ingo Molnardd41f592007-07-09 18:51:59 +02001139 int old_cpu = task_cpu(p);
1140 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001141 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1142 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnarbbdba7c2007-10-15 17:00:06 +02001143 u64 clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001144
1145 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001146
1147#ifdef CONFIG_SCHEDSTATS
1148 if (p->se.wait_start)
1149 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001150 if (p->se.sleep_start)
1151 p->se.sleep_start -= clock_offset;
1152 if (p->se.block_start)
1153 p->se.block_start -= clock_offset;
Ingo Molnarcc367732007-10-15 17:00:18 +02001154 if (old_cpu != new_cpu) {
1155 schedstat_inc(p, se.nr_migrations);
1156 if (task_hot(p, old_rq->clock, NULL))
1157 schedstat_inc(p, se.nr_forced2_migrations);
1158 }
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001159#endif
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001160 p->se.vruntime -= old_cfsrq->min_vruntime -
1161 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02001162
1163 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02001164}
1165
Ingo Molnar70b97a72006-07-03 00:25:42 -07001166struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168
Ingo Molnar36c8b582006-07-03 00:25:41 -07001169 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 int dest_cpu;
1171
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001173};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174
1175/*
1176 * The task's runqueue lock must be held.
1177 * Returns true if you have to wait for migration thread.
1178 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001179static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07001180migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001182 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183
1184 /*
1185 * If the task is not on a runqueue (and not running), then
1186 * it is sufficient to simply update the task's cpu field.
1187 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001188 if (!p->se.on_rq && !task_running(rq, p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 set_task_cpu(p, dest_cpu);
1190 return 0;
1191 }
1192
1193 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 req->task = p;
1195 req->dest_cpu = dest_cpu;
1196 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07001197
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 return 1;
1199}
1200
1201/*
1202 * wait_task_inactive - wait for a thread to unschedule.
1203 *
1204 * The caller must ensure that the task *will* unschedule sometime soon,
1205 * else this function might spin for a *long* time. This function can't
1206 * be called with interrupts off, or it may introduce deadlock with
1207 * smp_call_function() if an IPI is sent by the same process we are
1208 * waiting to become inactive.
1209 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001210void wait_task_inactive(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211{
1212 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02001213 int running, on_rq;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001214 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215
Andi Kleen3a5c3592007-10-15 17:00:14 +02001216 for (;;) {
1217 /*
1218 * We do the initial early heuristics without holding
1219 * any task-queue locks at all. We'll only try to get
1220 * the runqueue lock when things look like they will
1221 * work out!
1222 */
1223 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001224
Andi Kleen3a5c3592007-10-15 17:00:14 +02001225 /*
1226 * If the task is actively running on another CPU
1227 * still, just relax and busy-wait without holding
1228 * any locks.
1229 *
1230 * NOTE! Since we don't hold any locks, it's not
1231 * even sure that "rq" stays as the right runqueue!
1232 * But we don't care, since "task_running()" will
1233 * return false if the runqueue has changed and p
1234 * is actually now running somewhere else!
1235 */
1236 while (task_running(rq, p))
1237 cpu_relax();
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001238
Andi Kleen3a5c3592007-10-15 17:00:14 +02001239 /*
1240 * Ok, time to look more closely! We need the rq
1241 * lock now, to be *sure*. If we're wrong, we'll
1242 * just go back and repeat.
1243 */
1244 rq = task_rq_lock(p, &flags);
1245 running = task_running(rq, p);
1246 on_rq = p->se.on_rq;
1247 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001248
Andi Kleen3a5c3592007-10-15 17:00:14 +02001249 /*
1250 * Was it really running after all now that we
1251 * checked with the proper locks actually held?
1252 *
1253 * Oops. Go back and try again..
1254 */
1255 if (unlikely(running)) {
1256 cpu_relax();
1257 continue;
1258 }
1259
1260 /*
1261 * It's not enough that it's not actively running,
1262 * it must be off the runqueue _entirely_, and not
1263 * preempted!
1264 *
1265 * So if it wa still runnable (but just not actively
1266 * running right now), it's preempted, and we should
1267 * yield - it could be a while.
1268 */
1269 if (unlikely(on_rq)) {
1270 schedule_timeout_uninterruptible(1);
1271 continue;
1272 }
1273
1274 /*
1275 * Ahh, all good. It wasn't running, and it wasn't
1276 * runnable, which means that it will never become
1277 * running in the future either. We're all done!
1278 */
1279 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281}
1282
1283/***
1284 * kick_process - kick a running thread to enter/exit the kernel
1285 * @p: the to-be-kicked thread
1286 *
1287 * Cause a process which is running on another CPU to enter
1288 * kernel-mode, without any delay. (to get signals handled.)
1289 *
1290 * NOTE: this function doesnt have to take the runqueue lock,
1291 * because all it wants to ensure is that the remote task enters
1292 * the kernel. If the IPI races and the task has been migrated
1293 * to another CPU then no harm is done and the purpose has been
1294 * achieved as well.
1295 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001296void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297{
1298 int cpu;
1299
1300 preempt_disable();
1301 cpu = task_cpu(p);
1302 if ((cpu != smp_processor_id()) && task_curr(p))
1303 smp_send_reschedule(cpu);
1304 preempt_enable();
1305}
1306
1307/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001308 * Return a low guess at the load of a migration-source cpu weighted
1309 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 *
1311 * We want to under-estimate the load of migration sources, to
1312 * balance conservatively.
1313 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001314static unsigned long source_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08001315{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001316 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001317 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001318
Peter Williams2dd73a42006-06-27 02:54:34 -07001319 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02001320 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07001321
Ingo Molnardd41f592007-07-09 18:51:59 +02001322 return min(rq->cpu_load[type-1], total);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323}
1324
1325/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001326 * Return a high guess at the load of a migration-target cpu weighted
1327 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001329static unsigned long target_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08001330{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001331 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001332 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001333
Peter Williams2dd73a42006-06-27 02:54:34 -07001334 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02001335 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07001336
Ingo Molnardd41f592007-07-09 18:51:59 +02001337 return max(rq->cpu_load[type-1], total);
Peter Williams2dd73a42006-06-27 02:54:34 -07001338}
1339
1340/*
1341 * Return the average load per task on the cpu's run queue
1342 */
1343static inline unsigned long cpu_avg_load_per_task(int cpu)
1344{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001345 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001346 unsigned long total = weighted_cpuload(cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07001347 unsigned long n = rq->nr_running;
1348
Ingo Molnardd41f592007-07-09 18:51:59 +02001349 return n ? total / n : SCHED_LOAD_SCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350}
1351
Nick Piggin147cbb42005-06-25 14:57:19 -07001352/*
1353 * find_idlest_group finds and returns the least busy CPU group within the
1354 * domain.
1355 */
1356static struct sched_group *
1357find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1358{
1359 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1360 unsigned long min_load = ULONG_MAX, this_load = 0;
1361 int load_idx = sd->forkexec_idx;
1362 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1363
1364 do {
1365 unsigned long load, avg_load;
1366 int local_group;
1367 int i;
1368
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001369 /* Skip over this group if it has no CPUs allowed */
1370 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
Andi Kleen3a5c3592007-10-15 17:00:14 +02001371 continue;
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001372
Nick Piggin147cbb42005-06-25 14:57:19 -07001373 local_group = cpu_isset(this_cpu, group->cpumask);
Nick Piggin147cbb42005-06-25 14:57:19 -07001374
1375 /* Tally up the load of all CPUs in the group */
1376 avg_load = 0;
1377
1378 for_each_cpu_mask(i, group->cpumask) {
1379 /* Bias balancing toward cpus of our domain */
1380 if (local_group)
1381 load = source_load(i, load_idx);
1382 else
1383 load = target_load(i, load_idx);
1384
1385 avg_load += load;
1386 }
1387
1388 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07001389 avg_load = sg_div_cpu_power(group,
1390 avg_load * SCHED_LOAD_SCALE);
Nick Piggin147cbb42005-06-25 14:57:19 -07001391
1392 if (local_group) {
1393 this_load = avg_load;
1394 this = group;
1395 } else if (avg_load < min_load) {
1396 min_load = avg_load;
1397 idlest = group;
1398 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02001399 } while (group = group->next, group != sd->groups);
Nick Piggin147cbb42005-06-25 14:57:19 -07001400
1401 if (!idlest || 100*this_load < imbalance*min_load)
1402 return NULL;
1403 return idlest;
1404}
1405
1406/*
Satoru Takeuchi0feaece2006-10-03 01:14:10 -07001407 * find_idlest_cpu - find the idlest cpu among the cpus in group.
Nick Piggin147cbb42005-06-25 14:57:19 -07001408 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07001409static int
1410find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
Nick Piggin147cbb42005-06-25 14:57:19 -07001411{
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001412 cpumask_t tmp;
Nick Piggin147cbb42005-06-25 14:57:19 -07001413 unsigned long load, min_load = ULONG_MAX;
1414 int idlest = -1;
1415 int i;
1416
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001417 /* Traverse only the allowed CPUs */
1418 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1419
1420 for_each_cpu_mask(i, tmp) {
Peter Williams2dd73a42006-06-27 02:54:34 -07001421 load = weighted_cpuload(i);
Nick Piggin147cbb42005-06-25 14:57:19 -07001422
1423 if (load < min_load || (load == min_load && i == this_cpu)) {
1424 min_load = load;
1425 idlest = i;
1426 }
1427 }
1428
1429 return idlest;
1430}
1431
Nick Piggin476d1392005-06-25 14:57:29 -07001432/*
1433 * sched_balance_self: balance the current task (running on cpu) in domains
1434 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1435 * SD_BALANCE_EXEC.
1436 *
1437 * Balance, ie. select the least loaded group.
1438 *
1439 * Returns the target CPU number, or the same CPU if no balancing is needed.
1440 *
1441 * preempt must be disabled.
1442 */
1443static int sched_balance_self(int cpu, int flag)
1444{
1445 struct task_struct *t = current;
1446 struct sched_domain *tmp, *sd = NULL;
Nick Piggin147cbb42005-06-25 14:57:19 -07001447
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001448 for_each_domain(cpu, tmp) {
Ingo Molnar9761eea2007-07-09 18:52:00 +02001449 /*
1450 * If power savings logic is enabled for a domain, stop there.
1451 */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07001452 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1453 break;
Nick Piggin476d1392005-06-25 14:57:29 -07001454 if (tmp->flags & flag)
1455 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001456 }
Nick Piggin476d1392005-06-25 14:57:29 -07001457
1458 while (sd) {
1459 cpumask_t span;
1460 struct sched_group *group;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001461 int new_cpu, weight;
1462
1463 if (!(sd->flags & flag)) {
1464 sd = sd->child;
1465 continue;
1466 }
Nick Piggin476d1392005-06-25 14:57:29 -07001467
1468 span = sd->span;
1469 group = find_idlest_group(sd, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001470 if (!group) {
1471 sd = sd->child;
1472 continue;
1473 }
Nick Piggin476d1392005-06-25 14:57:29 -07001474
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001475 new_cpu = find_idlest_cpu(group, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001476 if (new_cpu == -1 || new_cpu == cpu) {
1477 /* Now try balancing at a lower domain level of cpu */
1478 sd = sd->child;
1479 continue;
1480 }
Nick Piggin476d1392005-06-25 14:57:29 -07001481
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001482 /* Now try balancing at a lower domain level of new_cpu */
Nick Piggin476d1392005-06-25 14:57:29 -07001483 cpu = new_cpu;
Nick Piggin476d1392005-06-25 14:57:29 -07001484 sd = NULL;
1485 weight = cpus_weight(span);
1486 for_each_domain(cpu, tmp) {
1487 if (weight <= cpus_weight(tmp->span))
1488 break;
1489 if (tmp->flags & flag)
1490 sd = tmp;
1491 }
1492 /* while loop will break here if sd == NULL */
1493 }
1494
1495 return cpu;
1496}
1497
1498#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499
1500/*
1501 * wake_idle() will wake a task on an idle cpu if task->cpu is
1502 * not idle and an idle cpu is available. The span of cpus to
1503 * search starts with cpus closest then further out as needed,
1504 * so we always favor a closer, idle cpu.
1505 *
1506 * Returns the CPU we should wake onto.
1507 */
1508#if defined(ARCH_HAS_SCHED_WAKE_IDLE)
Ingo Molnar36c8b582006-07-03 00:25:41 -07001509static int wake_idle(int cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510{
1511 cpumask_t tmp;
1512 struct sched_domain *sd;
1513 int i;
1514
Siddha, Suresh B49531982007-05-08 00:33:01 -07001515 /*
1516 * If it is idle, then it is the best cpu to run this task.
1517 *
1518 * This cpu is also the best, if it has more than one task already.
1519 * Siblings must be also busy(in most cases) as they didn't already
1520 * pickup the extra load from this cpu and hence we need not check
1521 * sibling runqueue info. This will avoid the checks and cache miss
1522 * penalities associated with that.
1523 */
1524 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 return cpu;
1526
1527 for_each_domain(cpu, sd) {
1528 if (sd->flags & SD_WAKE_IDLE) {
Nick Piggine0f364f2005-06-25 14:57:06 -07001529 cpus_and(tmp, sd->span, p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 for_each_cpu_mask(i, tmp) {
Ingo Molnarcc367732007-10-15 17:00:18 +02001531 if (idle_cpu(i)) {
1532 if (i != task_cpu(p)) {
1533 schedstat_inc(p,
1534 se.nr_wakeups_idle);
1535 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536 return i;
Ingo Molnarcc367732007-10-15 17:00:18 +02001537 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 }
Ingo Molnar9761eea2007-07-09 18:52:00 +02001539 } else {
Nick Piggine0f364f2005-06-25 14:57:06 -07001540 break;
Ingo Molnar9761eea2007-07-09 18:52:00 +02001541 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 }
1543 return cpu;
1544}
1545#else
Ingo Molnar36c8b582006-07-03 00:25:41 -07001546static inline int wake_idle(int cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547{
1548 return cpu;
1549}
1550#endif
1551
1552/***
1553 * try_to_wake_up - wake up a thread
1554 * @p: the to-be-woken-up thread
1555 * @state: the mask of task states that can be woken
1556 * @sync: do a synchronous wakeup?
1557 *
1558 * Put it on the run-queue if it's not already there. The "current"
1559 * thread is always on the run-queue (except when the actual
1560 * re-schedule is in progress), and as such you're allowed to do
1561 * the simpler "current->state = TASK_RUNNING" to mark yourself
1562 * runnable without the overhead of this.
1563 *
1564 * returns failure only if the task is already active.
1565 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001566static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567{
Ingo Molnarcc367732007-10-15 17:00:18 +02001568 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 unsigned long flags;
1570 long old_state;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001571 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572#ifdef CONFIG_SMP
Nick Piggin78979862005-06-25 14:57:13 -07001573 struct sched_domain *sd, *this_sd = NULL;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001574 unsigned long load, this_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575 int new_cpu;
1576#endif
1577
1578 rq = task_rq_lock(p, &flags);
1579 old_state = p->state;
1580 if (!(old_state & state))
1581 goto out;
1582
Ingo Molnardd41f592007-07-09 18:51:59 +02001583 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 goto out_running;
1585
1586 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02001587 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 this_cpu = smp_processor_id();
1589
1590#ifdef CONFIG_SMP
1591 if (unlikely(task_running(rq, p)))
1592 goto out_activate;
1593
Nick Piggin78979862005-06-25 14:57:13 -07001594 new_cpu = cpu;
1595
Ingo Molnar2d723762007-10-15 17:00:12 +02001596 schedstat_inc(rq, ttwu_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 if (cpu == this_cpu) {
1598 schedstat_inc(rq, ttwu_local);
Nick Piggin78979862005-06-25 14:57:13 -07001599 goto out_set_cpu;
1600 }
1601
1602 for_each_domain(this_cpu, sd) {
1603 if (cpu_isset(cpu, sd->span)) {
1604 schedstat_inc(sd, ttwu_wake_remote);
1605 this_sd = sd;
1606 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607 }
1608 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609
Nick Piggin78979862005-06-25 14:57:13 -07001610 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 goto out_set_cpu;
1612
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 /*
Nick Piggin78979862005-06-25 14:57:13 -07001614 * Check for affine wakeup and passive balancing possibilities.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615 */
Nick Piggin78979862005-06-25 14:57:13 -07001616 if (this_sd) {
1617 int idx = this_sd->wake_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 unsigned int imbalance;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619
Nick Piggina3f21bc2005-06-25 14:57:15 -07001620 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1621
Nick Piggin78979862005-06-25 14:57:13 -07001622 load = source_load(cpu, idx);
1623 this_load = target_load(this_cpu, idx);
1624
Nick Piggin78979862005-06-25 14:57:13 -07001625 new_cpu = this_cpu; /* Wake to this CPU if we can */
1626
Nick Piggina3f21bc2005-06-25 14:57:15 -07001627 if (this_sd->flags & SD_WAKE_AFFINE) {
1628 unsigned long tl = this_load;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08001629 unsigned long tl_per_task;
1630
Ingo Molnar71e20f12007-10-15 17:00:19 +02001631 /*
1632 * Attract cache-cold tasks on sync wakeups:
1633 */
1634 if (sync && !task_hot(p, rq->clock, this_sd))
1635 goto out_set_cpu;
1636
Ingo Molnarcc367732007-10-15 17:00:18 +02001637 schedstat_inc(p, se.nr_wakeups_affine_attempts);
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08001638 tl_per_task = cpu_avg_load_per_task(this_cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07001639
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 /*
Nick Piggina3f21bc2005-06-25 14:57:15 -07001641 * If sync wakeup then subtract the (maximum possible)
1642 * effect of the currently running task from the load
1643 * of the current CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 */
Nick Piggina3f21bc2005-06-25 14:57:15 -07001645 if (sync)
Ingo Molnardd41f592007-07-09 18:51:59 +02001646 tl -= current->se.load.weight;
Nick Piggina3f21bc2005-06-25 14:57:15 -07001647
1648 if ((tl <= load &&
Peter Williams2dd73a42006-06-27 02:54:34 -07001649 tl + target_load(cpu, idx) <= tl_per_task) ||
Ingo Molnardd41f592007-07-09 18:51:59 +02001650 100*(tl + p->se.load.weight) <= imbalance*load) {
Nick Piggina3f21bc2005-06-25 14:57:15 -07001651 /*
1652 * This domain has SD_WAKE_AFFINE and
1653 * p is cache cold in this domain, and
1654 * there is no bad imbalance.
1655 */
1656 schedstat_inc(this_sd, ttwu_move_affine);
Ingo Molnarcc367732007-10-15 17:00:18 +02001657 schedstat_inc(p, se.nr_wakeups_affine);
Nick Piggina3f21bc2005-06-25 14:57:15 -07001658 goto out_set_cpu;
1659 }
1660 }
1661
1662 /*
1663 * Start passive balancing when half the imbalance_pct
1664 * limit is reached.
1665 */
1666 if (this_sd->flags & SD_WAKE_BALANCE) {
1667 if (imbalance*this_load <= 100*load) {
1668 schedstat_inc(this_sd, ttwu_move_balance);
Ingo Molnarcc367732007-10-15 17:00:18 +02001669 schedstat_inc(p, se.nr_wakeups_passive);
Nick Piggina3f21bc2005-06-25 14:57:15 -07001670 goto out_set_cpu;
1671 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 }
1673 }
1674
1675 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1676out_set_cpu:
1677 new_cpu = wake_idle(new_cpu, p);
1678 if (new_cpu != cpu) {
1679 set_task_cpu(p, new_cpu);
1680 task_rq_unlock(rq, &flags);
1681 /* might preempt at this point */
1682 rq = task_rq_lock(p, &flags);
1683 old_state = p->state;
1684 if (!(old_state & state))
1685 goto out;
Ingo Molnardd41f592007-07-09 18:51:59 +02001686 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 goto out_running;
1688
1689 this_cpu = smp_processor_id();
1690 cpu = task_cpu(p);
1691 }
1692
1693out_activate:
1694#endif /* CONFIG_SMP */
Ingo Molnarcc367732007-10-15 17:00:18 +02001695 schedstat_inc(p, se.nr_wakeups);
1696 if (sync)
1697 schedstat_inc(p, se.nr_wakeups_sync);
1698 if (orig_cpu != cpu)
1699 schedstat_inc(p, se.nr_wakeups_migrate);
1700 if (cpu == this_cpu)
1701 schedstat_inc(p, se.nr_wakeups_local);
1702 else
1703 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnar2daa3572007-08-09 11:16:51 +02001704 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02001705 activate_task(rq, p, 1);
Ingo Molnar9c63d9c2007-10-15 17:00:20 +02001706 check_preempt_curr(rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 success = 1;
1708
1709out_running:
1710 p->state = TASK_RUNNING;
1711out:
1712 task_rq_unlock(rq, &flags);
1713
1714 return success;
1715}
1716
Ingo Molnar36c8b582006-07-03 00:25:41 -07001717int fastcall wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718{
1719 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1720 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1721}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722EXPORT_SYMBOL(wake_up_process);
1723
Ingo Molnar36c8b582006-07-03 00:25:41 -07001724int fastcall wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725{
1726 return try_to_wake_up(p, state, 0);
1727}
1728
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729/*
1730 * Perform scheduler related setup for a newly forked process p.
1731 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02001732 *
1733 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001735static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736{
Ingo Molnardd41f592007-07-09 18:51:59 +02001737 p->se.exec_start = 0;
1738 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02001739 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001740
1741#ifdef CONFIG_SCHEDSTATS
1742 p->se.wait_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02001743 p->se.sum_sleep_runtime = 0;
1744 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02001745 p->se.block_start = 0;
1746 p->se.sleep_max = 0;
1747 p->se.block_max = 0;
1748 p->se.exec_max = 0;
Ingo Molnareba1ed42007-10-15 17:00:02 +02001749 p->se.slice_max = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02001750 p->se.wait_max = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001751#endif
Nick Piggin476d1392005-06-25 14:57:29 -07001752
Ingo Molnardd41f592007-07-09 18:51:59 +02001753 INIT_LIST_HEAD(&p->run_list);
1754 p->se.on_rq = 0;
Nick Piggin476d1392005-06-25 14:57:29 -07001755
Avi Kivitye107be32007-07-26 13:40:43 +02001756#ifdef CONFIG_PREEMPT_NOTIFIERS
1757 INIT_HLIST_HEAD(&p->preempt_notifiers);
1758#endif
1759
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 /*
1761 * We mark the process as running here, but have not actually
1762 * inserted it onto the runqueue yet. This guarantees that
1763 * nobody will actually run it, and a signal or other external
1764 * event cannot wake it up and insert it on the runqueue either.
1765 */
1766 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02001767}
1768
1769/*
1770 * fork()/clone()-time setup:
1771 */
1772void sched_fork(struct task_struct *p, int clone_flags)
1773{
1774 int cpu = get_cpu();
1775
1776 __sched_fork(p);
1777
1778#ifdef CONFIG_SMP
1779 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1780#endif
Ingo Molnar02e4bac2007-10-15 17:00:11 +02001781 set_task_cpu(p, cpu);
Ingo Molnarb29739f2006-06-27 02:54:51 -07001782
1783 /*
1784 * Make sure we do not leak PI boosting priority to the child:
1785 */
1786 p->prio = current->normal_prio;
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02001787 if (!rt_prio(p->prio))
1788 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07001789
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07001790#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02001791 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07001792 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08001794#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07001795 p->oncpu = 0;
1796#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07001798 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08001799 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800#endif
Nick Piggin476d1392005-06-25 14:57:29 -07001801 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802}
1803
1804/*
1805 * wake_up_new_task - wake up a newly created task for the first time.
1806 *
1807 * This function will do some initial scheduler statistics housekeeping
1808 * that must be done for every newly created context, then puts the task
1809 * on the runqueue and wakes it.
1810 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001811void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812{
1813 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02001814 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815
1816 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02001818 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819
1820 p->prio = effective_prio(p);
1821
Srivatsa Vaddagirib9dca1e2007-10-17 16:55:11 +02001822 if (!p->sched_class->task_new || !current->se.on_rq) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001823 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02001826 * Let the scheduling class do new task startup
1827 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02001829 p->sched_class->task_new(rq, p);
Ingo Molnare5fa2232007-08-09 11:16:49 +02001830 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 }
Ingo Molnardd41f592007-07-09 18:51:59 +02001832 check_preempt_curr(rq, p);
1833 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834}
1835
Avi Kivitye107be32007-07-26 13:40:43 +02001836#ifdef CONFIG_PREEMPT_NOTIFIERS
1837
1838/**
Randy Dunlap421cee22007-07-31 00:37:50 -07001839 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1840 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02001841 */
1842void preempt_notifier_register(struct preempt_notifier *notifier)
1843{
1844 hlist_add_head(&notifier->link, &current->preempt_notifiers);
1845}
1846EXPORT_SYMBOL_GPL(preempt_notifier_register);
1847
1848/**
1849 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07001850 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02001851 *
1852 * This is safe to call from within a preemption notifier.
1853 */
1854void preempt_notifier_unregister(struct preempt_notifier *notifier)
1855{
1856 hlist_del(&notifier->link);
1857}
1858EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1859
1860static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1861{
1862 struct preempt_notifier *notifier;
1863 struct hlist_node *node;
1864
1865 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1866 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1867}
1868
1869static void
1870fire_sched_out_preempt_notifiers(struct task_struct *curr,
1871 struct task_struct *next)
1872{
1873 struct preempt_notifier *notifier;
1874 struct hlist_node *node;
1875
1876 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1877 notifier->ops->sched_out(notifier, next);
1878}
1879
1880#else
1881
1882static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1883{
1884}
1885
1886static void
1887fire_sched_out_preempt_notifiers(struct task_struct *curr,
1888 struct task_struct *next)
1889{
1890}
1891
1892#endif
1893
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894/**
Nick Piggin4866cde2005-06-25 14:57:23 -07001895 * prepare_task_switch - prepare to switch tasks
1896 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07001897 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07001898 * @next: the task we are going to switch to.
1899 *
1900 * This is called with the rq lock held and interrupts off. It must
1901 * be paired with a subsequent finish_task_switch after the context
1902 * switch.
1903 *
1904 * prepare_task_switch sets up locking and calls architecture specific
1905 * hooks.
1906 */
Avi Kivitye107be32007-07-26 13:40:43 +02001907static inline void
1908prepare_task_switch(struct rq *rq, struct task_struct *prev,
1909 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07001910{
Avi Kivitye107be32007-07-26 13:40:43 +02001911 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07001912 prepare_lock_switch(rq, next);
1913 prepare_arch_switch(next);
1914}
1915
1916/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04001918 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 * @prev: the thread we just switched away from.
1920 *
Nick Piggin4866cde2005-06-25 14:57:23 -07001921 * finish_task_switch must be called after the context switch, paired
1922 * with a prepare_task_switch call before the context switch.
1923 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1924 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 *
1926 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001927 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928 * with the lock held can cause deadlocks; see schedule() for
1929 * details.)
1930 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001931static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932 __releases(rq->lock)
1933{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07001935 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936
1937 rq->prev_mm = NULL;
1938
1939 /*
1940 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07001941 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07001942 * schedule one last time. The schedule call will never return, and
1943 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07001944 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945 * still held, otherwise prev could be scheduled on another cpu, die
1946 * there before we look at prev->state, and then the reference would
1947 * be dropped twice.
1948 * Manfred Spraul <manfred@colorfullife.com>
1949 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07001950 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07001951 finish_arch_switch(prev);
1952 finish_lock_switch(rq, prev);
Avi Kivitye107be32007-07-26 13:40:43 +02001953 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 if (mm)
1955 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07001956 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08001957 /*
1958 * Remove function-return probe instances associated with this
1959 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02001960 */
bibo maoc6fd91f2006-03-26 01:38:20 -08001961 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08001963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964}
1965
1966/**
1967 * schedule_tail - first thing a freshly forked thread must call.
1968 * @prev: the thread we just switched away from.
1969 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001970asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 __releases(rq->lock)
1972{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001973 struct rq *rq = this_rq();
1974
Nick Piggin4866cde2005-06-25 14:57:23 -07001975 finish_task_switch(rq, prev);
1976#ifdef __ARCH_WANT_UNLOCKED_CTXSW
1977 /* In this case, finish_task_switch does not reenable preemption */
1978 preempt_enable();
1979#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07001981 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982}
1983
1984/*
1985 * context_switch - switch to the new MM and the new
1986 * thread's register state.
1987 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001988static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07001989context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07001990 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991{
Ingo Molnardd41f592007-07-09 18:51:59 +02001992 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993
Avi Kivitye107be32007-07-26 13:40:43 +02001994 prepare_task_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02001995 mm = next->mm;
1996 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01001997 /*
1998 * For paravirt, this is coupled with an exit in switch_to to
1999 * combine the page table reload and the switch backend into
2000 * one hypercall.
2001 */
2002 arch_enter_lazy_cpu_mode();
2003
Ingo Molnardd41f592007-07-09 18:51:59 +02002004 if (unlikely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 next->active_mm = oldmm;
2006 atomic_inc(&oldmm->mm_count);
2007 enter_lazy_tlb(oldmm, next);
2008 } else
2009 switch_mm(oldmm, mm, next);
2010
Ingo Molnardd41f592007-07-09 18:51:59 +02002011 if (unlikely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 rq->prev_mm = oldmm;
2014 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002015 /*
2016 * Since the runqueue lock will be released by the next
2017 * task (which is an invalid locking op but in the case
2018 * of the scheduler it's an obvious special-case), so we
2019 * do an early lockdep release here:
2020 */
2021#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002022 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002023#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024
2025 /* Here we just switch the register state and the stack. */
2026 switch_to(prev, next, prev);
2027
Ingo Molnardd41f592007-07-09 18:51:59 +02002028 barrier();
2029 /*
2030 * this_rq must be evaluated again because prev may have moved
2031 * CPUs since it called schedule(), thus the 'rq' on its stack
2032 * frame will be invalid.
2033 */
2034 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035}
2036
2037/*
2038 * nr_running, nr_uninterruptible and nr_context_switches:
2039 *
2040 * externally visible scheduler statistics: current number of runnable
2041 * threads, current number of uninterruptible-sleeping threads, total
2042 * number of context switches performed since bootup.
2043 */
2044unsigned long nr_running(void)
2045{
2046 unsigned long i, sum = 0;
2047
2048 for_each_online_cpu(i)
2049 sum += cpu_rq(i)->nr_running;
2050
2051 return sum;
2052}
2053
2054unsigned long nr_uninterruptible(void)
2055{
2056 unsigned long i, sum = 0;
2057
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002058 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 sum += cpu_rq(i)->nr_uninterruptible;
2060
2061 /*
2062 * Since we read the counters lockless, it might be slightly
2063 * inaccurate. Do not allow it to go below zero though:
2064 */
2065 if (unlikely((long)sum < 0))
2066 sum = 0;
2067
2068 return sum;
2069}
2070
2071unsigned long long nr_context_switches(void)
2072{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002073 int i;
2074 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002076 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 sum += cpu_rq(i)->nr_switches;
2078
2079 return sum;
2080}
2081
2082unsigned long nr_iowait(void)
2083{
2084 unsigned long i, sum = 0;
2085
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002086 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2088
2089 return sum;
2090}
2091
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002092unsigned long nr_active(void)
2093{
2094 unsigned long i, running = 0, uninterruptible = 0;
2095
2096 for_each_online_cpu(i) {
2097 running += cpu_rq(i)->nr_running;
2098 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2099 }
2100
2101 if (unlikely((long)uninterruptible < 0))
2102 uninterruptible = 0;
2103
2104 return running + uninterruptible;
2105}
2106
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107/*
Ingo Molnardd41f592007-07-09 18:51:59 +02002108 * Update rq->cpu_load[] statistics. This function is usually called every
2109 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07002110 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002111static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002112{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02002113 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02002114 int i, scale;
2115
2116 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02002117
2118 /* Update our load: */
2119 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2120 unsigned long old_load, new_load;
2121
2122 /* scale is effectively 1 << i now, and >> i divides by scale */
2123
2124 old_load = this_rq->cpu_load[i];
2125 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02002126 /*
2127 * Round up the averaging division if load is increasing. This
2128 * prevents us from getting stuck on 9 if the load is 10, for
2129 * example.
2130 */
2131 if (new_load > old_load)
2132 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02002133 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2134 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002135}
2136
Ingo Molnardd41f592007-07-09 18:51:59 +02002137#ifdef CONFIG_SMP
2138
Ingo Molnar48f24c42006-07-03 00:25:40 -07002139/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140 * double_rq_lock - safely lock two runqueues
2141 *
2142 * Note this does not disable interrupts like task_rq_lock,
2143 * you need to do so manually before calling.
2144 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002145static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 __acquires(rq1->lock)
2147 __acquires(rq2->lock)
2148{
Kirill Korotaev054b9102006-12-10 02:20:11 -08002149 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 if (rq1 == rq2) {
2151 spin_lock(&rq1->lock);
2152 __acquire(rq2->lock); /* Fake it out ;) */
2153 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002154 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155 spin_lock(&rq1->lock);
2156 spin_lock(&rq2->lock);
2157 } else {
2158 spin_lock(&rq2->lock);
2159 spin_lock(&rq1->lock);
2160 }
2161 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002162 update_rq_clock(rq1);
2163 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164}
2165
2166/*
2167 * double_rq_unlock - safely unlock two runqueues
2168 *
2169 * Note this does not restore interrupts like task_rq_unlock,
2170 * you need to do so manually after calling.
2171 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002172static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 __releases(rq1->lock)
2174 __releases(rq2->lock)
2175{
2176 spin_unlock(&rq1->lock);
2177 if (rq1 != rq2)
2178 spin_unlock(&rq2->lock);
2179 else
2180 __release(rq2->lock);
2181}
2182
2183/*
2184 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2185 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002186static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 __releases(this_rq->lock)
2188 __acquires(busiest->lock)
2189 __acquires(this_rq->lock)
2190{
Kirill Korotaev054b9102006-12-10 02:20:11 -08002191 if (unlikely(!irqs_disabled())) {
2192 /* printk() doesn't work good under rq->lock */
2193 spin_unlock(&this_rq->lock);
2194 BUG_ON(1);
2195 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 if (unlikely(!spin_trylock(&busiest->lock))) {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002197 if (busiest < this_rq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 spin_unlock(&this_rq->lock);
2199 spin_lock(&busiest->lock);
2200 spin_lock(&this_rq->lock);
2201 } else
2202 spin_lock(&busiest->lock);
2203 }
2204}
2205
2206/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 * If dest_cpu is allowed for this process, migrate the task to it.
2208 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002209 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 * the cpu_allowed mask is restored.
2211 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002212static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002214 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002216 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217
2218 rq = task_rq_lock(p, &flags);
2219 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2220 || unlikely(cpu_is_offline(dest_cpu)))
2221 goto out;
2222
2223 /* force the process onto the specified CPU */
2224 if (migrate_task(p, dest_cpu, &req)) {
2225 /* Need to wait for migration thread (might exit: take ref). */
2226 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07002227
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228 get_task_struct(mt);
2229 task_rq_unlock(rq, &flags);
2230 wake_up_process(mt);
2231 put_task_struct(mt);
2232 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07002233
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 return;
2235 }
2236out:
2237 task_rq_unlock(rq, &flags);
2238}
2239
2240/*
Nick Piggin476d1392005-06-25 14:57:29 -07002241 * sched_exec - execve() is a valuable balancing opportunity, because at
2242 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 */
2244void sched_exec(void)
2245{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 int new_cpu, this_cpu = get_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002247 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002249 if (new_cpu != this_cpu)
2250 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251}
2252
2253/*
2254 * pull_task - move a task from a remote runqueue to the local runqueue.
2255 * Both runqueues must be locked.
2256 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002257static void pull_task(struct rq *src_rq, struct task_struct *p,
2258 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02002260 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002262 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 /*
2264 * Note that idle threads have a prio of MAX_PRIO, for this test
2265 * to be always true for them.
2266 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002267 check_preempt_curr(this_rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268}
2269
2270/*
2271 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2272 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08002273static
Ingo Molnar70b97a72006-07-03 00:25:42 -07002274int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002275 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002276 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277{
2278 /*
2279 * We do not migrate tasks that are:
2280 * 1) running (obviously), or
2281 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2282 * 3) are cache-hot on their current CPU.
2283 */
Ingo Molnarcc367732007-10-15 17:00:18 +02002284 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2285 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002287 }
Nick Piggin81026792005-06-25 14:57:07 -07002288 *all_pinned = 0;
2289
Ingo Molnarcc367732007-10-15 17:00:18 +02002290 if (task_running(rq, p)) {
2291 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07002292 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002293 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294
Ingo Molnarda84d962007-10-15 17:00:18 +02002295 /*
2296 * Aggressive migration if:
2297 * 1) task is cache cold, or
2298 * 2) too many balance attempts have failed.
2299 */
2300
Ingo Molnar6bc16652007-10-15 17:00:18 +02002301 if (!task_hot(p, rq->clock, sd) ||
2302 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002303#ifdef CONFIG_SCHEDSTATS
Ingo Molnarcc367732007-10-15 17:00:18 +02002304 if (task_hot(p, rq->clock, sd)) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002305 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02002306 schedstat_inc(p, se.nr_forced_migrations);
2307 }
Ingo Molnarda84d962007-10-15 17:00:18 +02002308#endif
2309 return 1;
2310 }
2311
Ingo Molnarcc367732007-10-15 17:00:18 +02002312 if (task_hot(p, rq->clock, sd)) {
2313 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02002314 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002315 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 return 1;
2317}
2318
Peter Williamse1d14842007-10-24 18:23:51 +02002319static unsigned long
2320balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2321 unsigned long max_load_move, struct sched_domain *sd,
2322 enum cpu_idle_type idle, int *all_pinned,
2323 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02002324{
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002325 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
Ingo Molnardd41f592007-07-09 18:51:59 +02002326 struct task_struct *p;
2327 long rem_load_move = max_load_move;
2328
Peter Williamse1d14842007-10-24 18:23:51 +02002329 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02002330 goto out;
2331
2332 pinned = 1;
2333
2334 /*
2335 * Start the load-balancing iterator:
2336 */
2337 p = iterator->start(iterator->arg);
2338next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002339 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02002340 goto out;
2341 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002342 * To help distribute high priority tasks across CPUs we don't
Ingo Molnardd41f592007-07-09 18:51:59 +02002343 * skip a task if it will be the highest priority task (i.e. smallest
2344 * prio value) on its new queue regardless of its load weight
2345 */
2346 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2347 SCHED_LOAD_SCALE_FUZZ;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002348 if ((skip_for_load && p->prio >= *this_best_prio) ||
Ingo Molnardd41f592007-07-09 18:51:59 +02002349 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002350 p = iterator->next(iterator->arg);
2351 goto next;
2352 }
2353
2354 pull_task(busiest, p, this_rq, this_cpu);
2355 pulled++;
2356 rem_load_move -= p->se.load.weight;
2357
2358 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002359 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02002360 */
Peter Williamse1d14842007-10-24 18:23:51 +02002361 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002362 if (p->prio < *this_best_prio)
2363 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02002364 p = iterator->next(iterator->arg);
2365 goto next;
2366 }
2367out:
2368 /*
Peter Williamse1d14842007-10-24 18:23:51 +02002369 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02002370 * so we can safely collect pull_task() stats here rather than
2371 * inside pull_task().
2372 */
2373 schedstat_add(sd, lb_gained[idle], pulled);
2374
2375 if (all_pinned)
2376 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02002377
2378 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02002379}
Ingo Molnar48f24c42006-07-03 00:25:40 -07002380
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381/*
Peter Williams43010652007-08-09 11:16:46 +02002382 * move_tasks tries to move up to max_load_move weighted load from busiest to
2383 * this_rq, as part of a balancing operation within domain "sd".
2384 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385 *
2386 * Called with both runqueues locked.
2387 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002388static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02002389 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002390 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07002391 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02002393 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02002394 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002395 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396
Ingo Molnardd41f592007-07-09 18:51:59 +02002397 do {
Peter Williams43010652007-08-09 11:16:46 +02002398 total_load_moved +=
2399 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02002400 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002401 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02002402 class = class->next;
Peter Williams43010652007-08-09 11:16:46 +02002403 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404
Peter Williams43010652007-08-09 11:16:46 +02002405 return total_load_moved > 0;
2406}
2407
Peter Williamse1d14842007-10-24 18:23:51 +02002408static int
2409iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2410 struct sched_domain *sd, enum cpu_idle_type idle,
2411 struct rq_iterator *iterator)
2412{
2413 struct task_struct *p = iterator->start(iterator->arg);
2414 int pinned = 0;
2415
2416 while (p) {
2417 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2418 pull_task(busiest, p, this_rq, this_cpu);
2419 /*
2420 * Right now, this is only the second place pull_task()
2421 * is called, so we can safely collect pull_task()
2422 * stats here rather than inside pull_task().
2423 */
2424 schedstat_inc(sd, lb_gained[idle]);
2425
2426 return 1;
2427 }
2428 p = iterator->next(iterator->arg);
2429 }
2430
2431 return 0;
2432}
2433
Peter Williams43010652007-08-09 11:16:46 +02002434/*
2435 * move_one_task tries to move exactly one task from busiest to this_rq, as
2436 * part of active balancing operations within "domain".
2437 * Returns 1 if successful and 0 otherwise.
2438 *
2439 * Called with both runqueues locked.
2440 */
2441static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2442 struct sched_domain *sd, enum cpu_idle_type idle)
2443{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02002444 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02002445
2446 for (class = sched_class_highest; class; class = class->next)
Peter Williamse1d14842007-10-24 18:23:51 +02002447 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02002448 return 1;
2449
2450 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451}
2452
2453/*
2454 * find_busiest_group finds and returns the busiest CPU group within the
Ingo Molnar48f24c42006-07-03 00:25:40 -07002455 * domain. It calculates and returns the amount of weighted load which
2456 * should be moved to restore balance via the imbalance parameter.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 */
2458static struct sched_group *
2459find_busiest_group(struct sched_domain *sd, int this_cpu,
Ingo Molnardd41f592007-07-09 18:51:59 +02002460 unsigned long *imbalance, enum cpu_idle_type idle,
2461 int *sd_idle, cpumask_t *cpus, int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462{
2463 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2464 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002465 unsigned long max_pull;
Peter Williams2dd73a42006-06-27 02:54:34 -07002466 unsigned long busiest_load_per_task, busiest_nr_running;
2467 unsigned long this_load_per_task, this_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02002468 int load_idx, group_imb = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002469#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2470 int power_savings_balance = 1;
2471 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2472 unsigned long min_nr_running = ULONG_MAX;
2473 struct sched_group *group_min = NULL, *group_leader = NULL;
2474#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475
2476 max_load = this_load = total_load = total_pwr = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07002477 busiest_load_per_task = busiest_nr_running = 0;
2478 this_load_per_task = this_nr_running = 0;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002479 if (idle == CPU_NOT_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07002480 load_idx = sd->busy_idx;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002481 else if (idle == CPU_NEWLY_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07002482 load_idx = sd->newidle_idx;
2483 else
2484 load_idx = sd->idle_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485
2486 do {
Ken Chen908a7c12007-10-17 16:55:11 +02002487 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 int local_group;
2489 int i;
Ken Chen908a7c12007-10-17 16:55:11 +02002490 int __group_imb = 0;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002491 unsigned int balance_cpu = -1, first_idle_cpu = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07002492 unsigned long sum_nr_running, sum_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493
2494 local_group = cpu_isset(this_cpu, group->cpumask);
2495
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002496 if (local_group)
2497 balance_cpu = first_cpu(group->cpumask);
2498
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 /* Tally up the load of all CPUs in the group */
Peter Williams2dd73a42006-06-27 02:54:34 -07002500 sum_weighted_load = sum_nr_running = avg_load = 0;
Ken Chen908a7c12007-10-17 16:55:11 +02002501 max_cpu_load = 0;
2502 min_cpu_load = ~0UL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503
2504 for_each_cpu_mask(i, group->cpumask) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002505 struct rq *rq;
2506
2507 if (!cpu_isset(i, *cpus))
2508 continue;
2509
2510 rq = cpu_rq(i);
Peter Williams2dd73a42006-06-27 02:54:34 -07002511
Suresh Siddha9439aab2007-07-19 21:28:35 +02002512 if (*sd_idle && rq->nr_running)
Nick Piggin5969fe02005-09-10 00:26:19 -07002513 *sd_idle = 0;
2514
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 /* Bias balancing toward cpus of our domain */
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002516 if (local_group) {
2517 if (idle_cpu(i) && !first_idle_cpu) {
2518 first_idle_cpu = 1;
2519 balance_cpu = i;
2520 }
2521
Nick Piggina2000572006-02-10 01:51:02 -08002522 load = target_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02002523 } else {
Nick Piggina2000572006-02-10 01:51:02 -08002524 load = source_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02002525 if (load > max_cpu_load)
2526 max_cpu_load = load;
2527 if (min_cpu_load > load)
2528 min_cpu_load = load;
2529 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530
2531 avg_load += load;
Peter Williams2dd73a42006-06-27 02:54:34 -07002532 sum_nr_running += rq->nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02002533 sum_weighted_load += weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 }
2535
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002536 /*
2537 * First idle cpu or the first cpu(busiest) in this sched group
2538 * is eligible for doing load balancing at this and above
Suresh Siddha9439aab2007-07-19 21:28:35 +02002539 * domains. In the newly idle case, we will allow all the cpu's
2540 * to do the newly idle load balance.
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002541 */
Suresh Siddha9439aab2007-07-19 21:28:35 +02002542 if (idle != CPU_NEWLY_IDLE && local_group &&
2543 balance_cpu != this_cpu && balance) {
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002544 *balance = 0;
2545 goto ret;
2546 }
2547
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 total_load += avg_load;
Eric Dumazet5517d862007-05-08 00:32:57 -07002549 total_pwr += group->__cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550
2551 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07002552 avg_load = sg_div_cpu_power(group,
2553 avg_load * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554
Ken Chen908a7c12007-10-17 16:55:11 +02002555 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2556 __group_imb = 1;
2557
Eric Dumazet5517d862007-05-08 00:32:57 -07002558 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002559
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 if (local_group) {
2561 this_load = avg_load;
2562 this = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07002563 this_nr_running = sum_nr_running;
2564 this_load_per_task = sum_weighted_load;
2565 } else if (avg_load > max_load &&
Ken Chen908a7c12007-10-17 16:55:11 +02002566 (sum_nr_running > group_capacity || __group_imb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 max_load = avg_load;
2568 busiest = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07002569 busiest_nr_running = sum_nr_running;
2570 busiest_load_per_task = sum_weighted_load;
Ken Chen908a7c12007-10-17 16:55:11 +02002571 group_imb = __group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002573
2574#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2575 /*
2576 * Busy processors will not participate in power savings
2577 * balance.
2578 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002579 if (idle == CPU_NOT_IDLE ||
2580 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2581 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002582
2583 /*
2584 * If the local group is idle or completely loaded
2585 * no need to do power savings balance at this domain
2586 */
2587 if (local_group && (this_nr_running >= group_capacity ||
2588 !this_nr_running))
2589 power_savings_balance = 0;
2590
Ingo Molnardd41f592007-07-09 18:51:59 +02002591 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002592 * If a group is already running at full capacity or idle,
2593 * don't include that group in power savings calculations
Ingo Molnardd41f592007-07-09 18:51:59 +02002594 */
2595 if (!power_savings_balance || sum_nr_running >= group_capacity
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002596 || !sum_nr_running)
Ingo Molnardd41f592007-07-09 18:51:59 +02002597 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002598
Ingo Molnardd41f592007-07-09 18:51:59 +02002599 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002600 * Calculate the group which has the least non-idle load.
Ingo Molnardd41f592007-07-09 18:51:59 +02002601 * This is the group from where we need to pick up the load
2602 * for saving power
2603 */
2604 if ((sum_nr_running < min_nr_running) ||
2605 (sum_nr_running == min_nr_running &&
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002606 first_cpu(group->cpumask) <
2607 first_cpu(group_min->cpumask))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002608 group_min = group;
2609 min_nr_running = sum_nr_running;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002610 min_load_per_task = sum_weighted_load /
2611 sum_nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02002612 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002613
Ingo Molnardd41f592007-07-09 18:51:59 +02002614 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002615 * Calculate the group which is almost near its
Ingo Molnardd41f592007-07-09 18:51:59 +02002616 * capacity but still has some space to pick up some load
2617 * from other group and save more power
2618 */
2619 if (sum_nr_running <= group_capacity - 1) {
2620 if (sum_nr_running > leader_nr_running ||
2621 (sum_nr_running == leader_nr_running &&
2622 first_cpu(group->cpumask) >
2623 first_cpu(group_leader->cpumask))) {
2624 group_leader = group;
2625 leader_nr_running = sum_nr_running;
2626 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002627 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002628group_next:
2629#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630 group = group->next;
2631 } while (group != sd->groups);
2632
Peter Williams2dd73a42006-06-27 02:54:34 -07002633 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634 goto out_balanced;
2635
2636 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2637
2638 if (this_load >= avg_load ||
2639 100*max_load <= sd->imbalance_pct*this_load)
2640 goto out_balanced;
2641
Peter Williams2dd73a42006-06-27 02:54:34 -07002642 busiest_load_per_task /= busiest_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02002643 if (group_imb)
2644 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2645
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646 /*
2647 * We're trying to get all the cpus to the average_load, so we don't
2648 * want to push ourselves above the average load, nor do we wish to
2649 * reduce the max loaded cpu below the average load, as either of these
2650 * actions would just result in more rebalancing later, and ping-pong
2651 * tasks around. Thus we look for the minimum possible imbalance.
2652 * Negative imbalances (*we* are more loaded than anyone else) will
2653 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002654 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 * appear as very large values with unsigned longs.
2656 */
Peter Williams2dd73a42006-06-27 02:54:34 -07002657 if (max_load <= busiest_load_per_task)
2658 goto out_balanced;
2659
2660 /*
2661 * In the presence of smp nice balancing, certain scenarios can have
2662 * max load less than avg load(as we skip the groups at or below
2663 * its cpu_power, while calculating max_load..)
2664 */
2665 if (max_load < avg_load) {
2666 *imbalance = 0;
2667 goto small_imbalance;
2668 }
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002669
2670 /* Don't want to pull so many tasks that a group would go idle */
Peter Williams2dd73a42006-06-27 02:54:34 -07002671 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002672
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 /* How much load to actually move to equalise the imbalance */
Eric Dumazet5517d862007-05-08 00:32:57 -07002674 *imbalance = min(max_pull * busiest->__cpu_power,
2675 (avg_load - this_load) * this->__cpu_power)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 / SCHED_LOAD_SCALE;
2677
Peter Williams2dd73a42006-06-27 02:54:34 -07002678 /*
2679 * if *imbalance is less than the average load per runnable task
2680 * there is no gaurantee that any tasks will be moved so we'll have
2681 * a think about bumping its value to force at least one task to be
2682 * moved
2683 */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02002684 if (*imbalance < busiest_load_per_task) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07002685 unsigned long tmp, pwr_now, pwr_move;
Peter Williams2dd73a42006-06-27 02:54:34 -07002686 unsigned int imbn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687
Peter Williams2dd73a42006-06-27 02:54:34 -07002688small_imbalance:
2689 pwr_move = pwr_now = 0;
2690 imbn = 2;
2691 if (this_nr_running) {
2692 this_load_per_task /= this_nr_running;
2693 if (busiest_load_per_task > this_load_per_task)
2694 imbn = 1;
2695 } else
2696 this_load_per_task = SCHED_LOAD_SCALE;
2697
Ingo Molnardd41f592007-07-09 18:51:59 +02002698 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2699 busiest_load_per_task * imbn) {
Peter Williams2dd73a42006-06-27 02:54:34 -07002700 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701 return busiest;
2702 }
2703
2704 /*
2705 * OK, we don't have enough imbalance to justify moving tasks,
2706 * however we may be able to increase total CPU power used by
2707 * moving them.
2708 */
2709
Eric Dumazet5517d862007-05-08 00:32:57 -07002710 pwr_now += busiest->__cpu_power *
2711 min(busiest_load_per_task, max_load);
2712 pwr_now += this->__cpu_power *
2713 min(this_load_per_task, this_load);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 pwr_now /= SCHED_LOAD_SCALE;
2715
2716 /* Amount of load we'd subtract */
Eric Dumazet5517d862007-05-08 00:32:57 -07002717 tmp = sg_div_cpu_power(busiest,
2718 busiest_load_per_task * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719 if (max_load > tmp)
Eric Dumazet5517d862007-05-08 00:32:57 -07002720 pwr_move += busiest->__cpu_power *
Peter Williams2dd73a42006-06-27 02:54:34 -07002721 min(busiest_load_per_task, max_load - tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722
2723 /* Amount of load we'd add */
Eric Dumazet5517d862007-05-08 00:32:57 -07002724 if (max_load * busiest->__cpu_power <
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08002725 busiest_load_per_task * SCHED_LOAD_SCALE)
Eric Dumazet5517d862007-05-08 00:32:57 -07002726 tmp = sg_div_cpu_power(this,
2727 max_load * busiest->__cpu_power);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 else
Eric Dumazet5517d862007-05-08 00:32:57 -07002729 tmp = sg_div_cpu_power(this,
2730 busiest_load_per_task * SCHED_LOAD_SCALE);
2731 pwr_move += this->__cpu_power *
2732 min(this_load_per_task, this_load + tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733 pwr_move /= SCHED_LOAD_SCALE;
2734
2735 /* Move if we gain throughput */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02002736 if (pwr_move > pwr_now)
2737 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738 }
2739
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740 return busiest;
2741
2742out_balanced:
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002743#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002744 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002745 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002747 if (this == group_leader && group_leader != group_min) {
2748 *imbalance = min_load_per_task;
2749 return group_min;
2750 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002751#endif
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002752ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753 *imbalance = 0;
2754 return NULL;
2755}
2756
2757/*
2758 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2759 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002760static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002761find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002762 unsigned long imbalance, cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002764 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07002765 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 int i;
2767
2768 for_each_cpu_mask(i, group->cpumask) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002769 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002770
2771 if (!cpu_isset(i, *cpus))
2772 continue;
2773
Ingo Molnar48f24c42006-07-03 00:25:40 -07002774 rq = cpu_rq(i);
Ingo Molnardd41f592007-07-09 18:51:59 +02002775 wl = weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776
Ingo Molnardd41f592007-07-09 18:51:59 +02002777 if (rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07002778 continue;
2779
Ingo Molnardd41f592007-07-09 18:51:59 +02002780 if (wl > max_load) {
2781 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002782 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 }
2784 }
2785
2786 return busiest;
2787}
2788
2789/*
Nick Piggin77391d72005-06-25 14:57:30 -07002790 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2791 * so long as it is large enough.
2792 */
2793#define MAX_PINNED_INTERVAL 512
2794
2795/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2797 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002799static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002800 struct sched_domain *sd, enum cpu_idle_type idle,
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002801 int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802{
Peter Williams43010652007-08-09 11:16:46 +02002803 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002806 struct rq *busiest;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002807 cpumask_t cpus = CPU_MASK_ALL;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002808 unsigned long flags;
Nick Piggin5969fe02005-09-10 00:26:19 -07002809
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002810 /*
2811 * When power savings policy is enabled for the parent domain, idle
2812 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02002813 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002814 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002815 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002816 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002817 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002818 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819
Ingo Molnar2d723762007-10-15 17:00:12 +02002820 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002822redo:
2823 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002824 &cpus, balance);
2825
Chen, Kenneth W06066712006-12-10 02:20:35 -08002826 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002827 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002828
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 if (!group) {
2830 schedstat_inc(sd, lb_nobusyg[idle]);
2831 goto out_balanced;
2832 }
2833
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002834 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 if (!busiest) {
2836 schedstat_inc(sd, lb_nobusyq[idle]);
2837 goto out_balanced;
2838 }
2839
Nick Piggindb935db2005-06-25 14:57:11 -07002840 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841
2842 schedstat_add(sd, lb_imbalance[idle], imbalance);
2843
Peter Williams43010652007-08-09 11:16:46 +02002844 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 if (busiest->nr_running > 1) {
2846 /*
2847 * Attempt to move tasks. If find_busiest_group has found
2848 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02002849 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850 * correctly treated as an imbalance.
2851 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002852 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07002853 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02002854 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07002855 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07002856 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002857 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07002858
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07002859 /*
2860 * some other cpu did the load balance for us.
2861 */
Peter Williams43010652007-08-09 11:16:46 +02002862 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07002863 resched_cpu(this_cpu);
2864
Nick Piggin81026792005-06-25 14:57:07 -07002865 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002866 if (unlikely(all_pinned)) {
2867 cpu_clear(cpu_of(busiest), cpus);
2868 if (!cpus_empty(cpus))
2869 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07002870 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002871 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 }
Nick Piggin81026792005-06-25 14:57:07 -07002873
Peter Williams43010652007-08-09 11:16:46 +02002874 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 schedstat_inc(sd, lb_failed[idle]);
2876 sd->nr_balance_failed++;
2877
2878 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002880 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07002881
2882 /* don't kick the migration_thread, if the curr
2883 * task on busiest cpu can't be moved to this_cpu
2884 */
2885 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002886 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07002887 all_pinned = 1;
2888 goto out_one_pinned;
2889 }
2890
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 if (!busiest->active_balance) {
2892 busiest->active_balance = 1;
2893 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07002894 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002896 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07002897 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 wake_up_process(busiest->migration_thread);
2899
2900 /*
2901 * We've kicked active balancing, reset the failure
2902 * counter.
2903 */
Nick Piggin39507452005-06-25 14:57:09 -07002904 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 }
Nick Piggin81026792005-06-25 14:57:07 -07002906 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907 sd->nr_balance_failed = 0;
2908
Nick Piggin81026792005-06-25 14:57:07 -07002909 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 /* We were unbalanced, so reset the balancing interval */
2911 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07002912 } else {
2913 /*
2914 * If we've begun active balancing, start to back off. This
2915 * case may not be covered by the all_pinned logic if there
2916 * is only 1 task on the busy runqueue (because we don't call
2917 * move_tasks).
2918 */
2919 if (sd->balance_interval < sd->max_interval)
2920 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 }
2922
Peter Williams43010652007-08-09 11:16:46 +02002923 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002924 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002925 return -1;
Peter Williams43010652007-08-09 11:16:46 +02002926 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927
2928out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 schedstat_inc(sd, lb_balanced[idle]);
2930
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002931 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07002932
2933out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07002935 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2936 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 sd->balance_interval *= 2;
2938
Ingo Molnar48f24c42006-07-03 00:25:40 -07002939 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002940 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002941 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 return 0;
2943}
2944
2945/*
2946 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2947 * tasks if there is an imbalance.
2948 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002949 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 * this_rq is locked.
2951 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07002952static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07002953load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954{
2955 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002956 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02002958 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07002959 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02002960 int all_pinned = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002961 cpumask_t cpus = CPU_MASK_ALL;
Nick Piggin5969fe02005-09-10 00:26:19 -07002962
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002963 /*
2964 * When power savings policy is enabled for the parent domain, idle
2965 * sibling can pick up load irrespective of busy siblings. In this case,
2966 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002967 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002968 */
2969 if (sd->flags & SD_SHARE_CPUPOWER &&
2970 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002971 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972
Ingo Molnar2d723762007-10-15 17:00:12 +02002973 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002974redo:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002975 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002976 &sd_idle, &cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002978 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002979 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980 }
2981
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002982 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002983 &cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07002984 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002985 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002986 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 }
2988
Nick Piggindb935db2005-06-25 14:57:11 -07002989 BUG_ON(busiest == this_rq);
2990
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002991 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07002992
Peter Williams43010652007-08-09 11:16:46 +02002993 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07002994 if (busiest->nr_running > 1) {
2995 /* Attempt to move tasks */
2996 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002997 /* this_rq->clock is already updated */
2998 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02002999 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003000 imbalance, sd, CPU_NEWLY_IDLE,
3001 &all_pinned);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003002 spin_unlock(&busiest->lock);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003003
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003004 if (unlikely(all_pinned)) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003005 cpu_clear(cpu_of(busiest), cpus);
3006 if (!cpus_empty(cpus))
3007 goto redo;
3008 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003009 }
3010
Peter Williams43010652007-08-09 11:16:46 +02003011 if (!ld_moved) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003012 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003013 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3014 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003015 return -1;
3016 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003017 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018
Peter Williams43010652007-08-09 11:16:46 +02003019 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003020
3021out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003022 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003023 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003024 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003025 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003026 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003027
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003028 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029}
3030
3031/*
3032 * idle_balance is called by schedule() if this_cpu is about to become
3033 * idle. Attempts to pull tasks from other CPUs.
3034 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003035static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036{
3037 struct sched_domain *sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02003038 int pulled_task = -1;
3039 unsigned long next_balance = jiffies + HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040
3041 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003042 unsigned long interval;
3043
3044 if (!(sd->flags & SD_LOAD_BALANCE))
3045 continue;
3046
3047 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07003048 /* If we've pulled tasks over stop searching: */
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003049 pulled_task = load_balance_newidle(this_cpu,
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003050 this_rq, sd);
3051
3052 interval = msecs_to_jiffies(sd->balance_interval);
3053 if (time_after(next_balance, sd->last_balance + interval))
3054 next_balance = sd->last_balance + interval;
3055 if (pulled_task)
3056 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003058 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003059 /*
3060 * We are going idle. next_balance may be set based on
3061 * a busy processor. So reset next_balance.
3062 */
3063 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02003064 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065}
3066
3067/*
3068 * active_load_balance is run by migration threads. It pushes running tasks
3069 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3070 * running on each physical CPU where possible, and avoids physical /
3071 * logical imbalances.
3072 *
3073 * Called with busiest_rq locked.
3074 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003075static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076{
Nick Piggin39507452005-06-25 14:57:09 -07003077 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003078 struct sched_domain *sd;
3079 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07003080
Ingo Molnar48f24c42006-07-03 00:25:40 -07003081 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07003082 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07003083 return;
3084
3085 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086
3087 /*
Nick Piggin39507452005-06-25 14:57:09 -07003088 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003089 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07003090 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 */
Nick Piggin39507452005-06-25 14:57:09 -07003092 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093
Nick Piggin39507452005-06-25 14:57:09 -07003094 /* move a task from busiest_rq to target_rq */
3095 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003096 update_rq_clock(busiest_rq);
3097 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098
Nick Piggin39507452005-06-25 14:57:09 -07003099 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003100 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07003101 if ((sd->flags & SD_LOAD_BALANCE) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07003102 cpu_isset(busiest_cpu, sd->span))
Nick Piggin39507452005-06-25 14:57:09 -07003103 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003104 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105
Ingo Molnar48f24c42006-07-03 00:25:40 -07003106 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02003107 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108
Peter Williams43010652007-08-09 11:16:46 +02003109 if (move_one_task(target_rq, target_cpu, busiest_rq,
3110 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07003111 schedstat_inc(sd, alb_pushed);
3112 else
3113 schedstat_inc(sd, alb_failed);
3114 }
Nick Piggin39507452005-06-25 14:57:09 -07003115 spin_unlock(&target_rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116}
3117
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003118#ifdef CONFIG_NO_HZ
3119static struct {
3120 atomic_t load_balancer;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003121 cpumask_t cpu_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003122} nohz ____cacheline_aligned = {
3123 .load_balancer = ATOMIC_INIT(-1),
3124 .cpu_mask = CPU_MASK_NONE,
3125};
3126
Christoph Lameter7835b982006-12-10 02:20:22 -08003127/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003128 * This routine will try to nominate the ilb (idle load balancing)
3129 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3130 * load balancing on behalf of all those cpus. If all the cpus in the system
3131 * go into this tickless mode, then there will be no ilb owner (as there is
3132 * no need for one) and all the cpus will sleep till the next wakeup event
3133 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08003134 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003135 * For the ilb owner, tick is not stopped. And this tick will be used
3136 * for idle load balancing. ilb owner will still be part of
3137 * nohz.cpu_mask..
3138 *
3139 * While stopping the tick, this cpu will become the ilb owner if there
3140 * is no other owner. And will be the owner till that cpu becomes busy
3141 * or if all cpus in the system stop their ticks at which point
3142 * there is no need for ilb owner.
3143 *
3144 * When the ilb owner becomes busy, it nominates another owner, during the
3145 * next busy scheduler_tick()
3146 */
3147int select_nohz_load_balancer(int stop_tick)
3148{
3149 int cpu = smp_processor_id();
3150
3151 if (stop_tick) {
3152 cpu_set(cpu, nohz.cpu_mask);
3153 cpu_rq(cpu)->in_nohz_recently = 1;
3154
3155 /*
3156 * If we are going offline and still the leader, give up!
3157 */
3158 if (cpu_is_offline(cpu) &&
3159 atomic_read(&nohz.load_balancer) == cpu) {
3160 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3161 BUG();
3162 return 0;
3163 }
3164
3165 /* time for ilb owner also to sleep */
3166 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3167 if (atomic_read(&nohz.load_balancer) == cpu)
3168 atomic_set(&nohz.load_balancer, -1);
3169 return 0;
3170 }
3171
3172 if (atomic_read(&nohz.load_balancer) == -1) {
3173 /* make me the ilb owner */
3174 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3175 return 1;
3176 } else if (atomic_read(&nohz.load_balancer) == cpu)
3177 return 1;
3178 } else {
3179 if (!cpu_isset(cpu, nohz.cpu_mask))
3180 return 0;
3181
3182 cpu_clear(cpu, nohz.cpu_mask);
3183
3184 if (atomic_read(&nohz.load_balancer) == cpu)
3185 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3186 BUG();
3187 }
3188 return 0;
3189}
3190#endif
3191
3192static DEFINE_SPINLOCK(balancing);
3193
3194/*
Christoph Lameter7835b982006-12-10 02:20:22 -08003195 * It checks each scheduling domain to see if it is due to be balanced,
3196 * and initiates a balancing operation if so.
3197 *
3198 * Balancing parameters are set up in arch_init_sched_domains.
3199 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02003200static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08003201{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003202 int balance = 1;
3203 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08003204 unsigned long interval;
3205 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003206 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08003207 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003208 int update_next_balance = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003210 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211 if (!(sd->flags & SD_LOAD_BALANCE))
3212 continue;
3213
3214 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003215 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216 interval *= sd->busy_factor;
3217
3218 /* scale ms to jiffies */
3219 interval = msecs_to_jiffies(interval);
3220 if (unlikely(!interval))
3221 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003222 if (interval > HZ*NR_CPUS/10)
3223 interval = HZ*NR_CPUS/10;
3224
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225
Christoph Lameter08c183f2006-12-10 02:20:29 -08003226 if (sd->flags & SD_SERIALIZE) {
3227 if (!spin_trylock(&balancing))
3228 goto out;
3229 }
3230
Christoph Lameterc9819f42006-12-10 02:20:25 -08003231 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003232 if (load_balance(cpu, rq, sd, idle, &balance)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003233 /*
3234 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07003235 * longer idle, or one of our SMT siblings is
3236 * not idle.
3237 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003238 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003240 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241 }
Christoph Lameter08c183f2006-12-10 02:20:29 -08003242 if (sd->flags & SD_SERIALIZE)
3243 spin_unlock(&balancing);
3244out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02003245 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08003246 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003247 update_next_balance = 1;
3248 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003249
3250 /*
3251 * Stop the load balance at this level. There is another
3252 * CPU in our sched group which is doing load balancing more
3253 * actively.
3254 */
3255 if (!balance)
3256 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02003258
3259 /*
3260 * next_balance will be updated only when there is a need.
3261 * When the cpu is attached to null domain for ex, it will not be
3262 * updated.
3263 */
3264 if (likely(update_next_balance))
3265 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003266}
3267
3268/*
3269 * run_rebalance_domains is triggered when needed from the scheduler tick.
3270 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3271 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3272 */
3273static void run_rebalance_domains(struct softirq_action *h)
3274{
Ingo Molnardd41f592007-07-09 18:51:59 +02003275 int this_cpu = smp_processor_id();
3276 struct rq *this_rq = cpu_rq(this_cpu);
3277 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3278 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003279
Ingo Molnardd41f592007-07-09 18:51:59 +02003280 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003281
3282#ifdef CONFIG_NO_HZ
3283 /*
3284 * If this cpu is the owner for idle load balancing, then do the
3285 * balancing on behalf of the other idle cpus whose ticks are
3286 * stopped.
3287 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003288 if (this_rq->idle_at_tick &&
3289 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003290 cpumask_t cpus = nohz.cpu_mask;
3291 struct rq *rq;
3292 int balance_cpu;
3293
Ingo Molnardd41f592007-07-09 18:51:59 +02003294 cpu_clear(this_cpu, cpus);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003295 for_each_cpu_mask(balance_cpu, cpus) {
3296 /*
3297 * If this cpu gets work to do, stop the load balancing
3298 * work being done for other cpus. Next load
3299 * balancing owner will pick it up.
3300 */
3301 if (need_resched())
3302 break;
3303
Oleg Nesterovde0cf892007-08-12 18:08:19 +02003304 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003305
3306 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003307 if (time_after(this_rq->next_balance, rq->next_balance))
3308 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003309 }
3310 }
3311#endif
3312}
3313
3314/*
3315 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3316 *
3317 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3318 * idle load balancing owner or decide to stop the periodic load balancing,
3319 * if the whole system is idle.
3320 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003321static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003322{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003323#ifdef CONFIG_NO_HZ
3324 /*
3325 * If we were in the nohz mode recently and busy at the current
3326 * scheduler tick, then check if we need to nominate new idle
3327 * load balancer.
3328 */
3329 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3330 rq->in_nohz_recently = 0;
3331
3332 if (atomic_read(&nohz.load_balancer) == cpu) {
3333 cpu_clear(cpu, nohz.cpu_mask);
3334 atomic_set(&nohz.load_balancer, -1);
3335 }
3336
3337 if (atomic_read(&nohz.load_balancer) == -1) {
3338 /*
3339 * simple selection for now: Nominate the
3340 * first cpu in the nohz list to be the next
3341 * ilb owner.
3342 *
3343 * TBD: Traverse the sched domains and nominate
3344 * the nearest cpu in the nohz.cpu_mask.
3345 */
3346 int ilb = first_cpu(nohz.cpu_mask);
3347
3348 if (ilb != NR_CPUS)
3349 resched_cpu(ilb);
3350 }
3351 }
3352
3353 /*
3354 * If this cpu is idle and doing idle load balancing for all the
3355 * cpus with ticks stopped, is it time for that to stop?
3356 */
3357 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3358 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3359 resched_cpu(cpu);
3360 return;
3361 }
3362
3363 /*
3364 * If this cpu is idle and the idle load balancing is done by
3365 * someone else, then no need raise the SCHED_SOFTIRQ
3366 */
3367 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3368 cpu_isset(cpu, nohz.cpu_mask))
3369 return;
3370#endif
3371 if (time_after_eq(jiffies, rq->next_balance))
3372 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373}
Ingo Molnardd41f592007-07-09 18:51:59 +02003374
3375#else /* CONFIG_SMP */
3376
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377/*
3378 * on UP we do not need to balance between CPUs:
3379 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003380static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381{
3382}
Ingo Molnardd41f592007-07-09 18:51:59 +02003383
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384#endif
3385
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386DEFINE_PER_CPU(struct kernel_stat, kstat);
3387
3388EXPORT_PER_CPU_SYMBOL(kstat);
3389
3390/*
Ingo Molnar41b86e92007-07-09 18:51:58 +02003391 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3392 * that have not yet been banked in case the task is currently running.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393 */
Ingo Molnar41b86e92007-07-09 18:51:58 +02003394unsigned long long task_sched_runtime(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003395{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003397 u64 ns, delta_exec;
3398 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003399
Ingo Molnar41b86e92007-07-09 18:51:58 +02003400 rq = task_rq_lock(p, &flags);
3401 ns = p->se.sum_exec_runtime;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01003402 if (task_current(rq, p)) {
Ingo Molnara8e504d2007-08-09 11:16:47 +02003403 update_rq_clock(rq);
3404 delta_exec = rq->clock - p->se.exec_start;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003405 if ((s64)delta_exec > 0)
3406 ns += delta_exec;
3407 }
3408 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003409
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 return ns;
3411}
3412
3413/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414 * Account user cpu time to a process.
3415 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416 * @cputime: the cpu time spent in user space since the last update
3417 */
3418void account_user_time(struct task_struct *p, cputime_t cputime)
3419{
3420 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3421 cputime64_t tmp;
3422
3423 p->utime = cputime_add(p->utime, cputime);
3424
3425 /* Add user time to cpustat. */
3426 tmp = cputime_to_cputime64(cputime);
3427 if (TASK_NICE(p) > 0)
3428 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3429 else
3430 cpustat->user = cputime64_add(cpustat->user, tmp);
3431}
3432
3433/*
Laurent Vivier94886b82007-10-15 17:00:19 +02003434 * Account guest cpu time to a process.
3435 * @p: the process that the cpu time gets accounted to
3436 * @cputime: the cpu time spent in virtual machine since the last update
3437 */
Adrian Bunkf7402e02007-10-29 21:18:10 +01003438static void account_guest_time(struct task_struct *p, cputime_t cputime)
Laurent Vivier94886b82007-10-15 17:00:19 +02003439{
3440 cputime64_t tmp;
3441 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3442
3443 tmp = cputime_to_cputime64(cputime);
3444
3445 p->utime = cputime_add(p->utime, cputime);
3446 p->gtime = cputime_add(p->gtime, cputime);
3447
3448 cpustat->user = cputime64_add(cpustat->user, tmp);
3449 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3450}
3451
3452/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07003453 * Account scaled user cpu time to a process.
3454 * @p: the process that the cpu time gets accounted to
3455 * @cputime: the cpu time spent in user space since the last update
3456 */
3457void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3458{
3459 p->utimescaled = cputime_add(p->utimescaled, cputime);
3460}
3461
3462/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463 * Account system cpu time to a process.
3464 * @p: the process that the cpu time gets accounted to
3465 * @hardirq_offset: the offset to subtract from hardirq_count()
3466 * @cputime: the cpu time spent in kernel space since the last update
3467 */
3468void account_system_time(struct task_struct *p, int hardirq_offset,
3469 cputime_t cputime)
3470{
3471 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003472 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473 cputime64_t tmp;
3474
Christian Borntraeger97783852007-11-15 20:57:39 +01003475 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3476 return account_guest_time(p, cputime);
Laurent Vivier94886b82007-10-15 17:00:19 +02003477
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478 p->stime = cputime_add(p->stime, cputime);
3479
3480 /* Add system time to cpustat. */
3481 tmp = cputime_to_cputime64(cputime);
3482 if (hardirq_count() - hardirq_offset)
3483 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3484 else if (softirq_count())
3485 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08003486 else if (p != rq->idle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487 cpustat->system = cputime64_add(cpustat->system, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08003488 else if (atomic_read(&rq->nr_iowait) > 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3490 else
3491 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3492 /* Account for system time used */
3493 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494}
3495
3496/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07003497 * Account scaled system cpu time to a process.
3498 * @p: the process that the cpu time gets accounted to
3499 * @hardirq_offset: the offset to subtract from hardirq_count()
3500 * @cputime: the cpu time spent in kernel space since the last update
3501 */
3502void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3503{
3504 p->stimescaled = cputime_add(p->stimescaled, cputime);
3505}
3506
3507/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 * Account for involuntary wait time.
3509 * @p: the process from which the cpu time has been stolen
3510 * @steal: the cpu time spent in involuntary wait
3511 */
3512void account_steal_time(struct task_struct *p, cputime_t steal)
3513{
3514 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3515 cputime64_t tmp = cputime_to_cputime64(steal);
Ingo Molnar70b97a72006-07-03 00:25:42 -07003516 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517
3518 if (p == rq->idle) {
3519 p->stime = cputime_add(p->stime, steal);
3520 if (atomic_read(&rq->nr_iowait) > 0)
3521 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3522 else
3523 cpustat->idle = cputime64_add(cpustat->idle, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08003524 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3526}
3527
Christoph Lameter7835b982006-12-10 02:20:22 -08003528/*
3529 * This function gets called by the timer code, with HZ frequency.
3530 * We call it with interrupts disabled.
3531 *
3532 * It also gets called by the fork code, when changing the parent's
3533 * timeslices.
3534 */
3535void scheduler_tick(void)
3536{
Christoph Lameter7835b982006-12-10 02:20:22 -08003537 int cpu = smp_processor_id();
3538 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003539 struct task_struct *curr = rq->curr;
Ingo Molnar529c7722007-08-10 23:05:11 +02003540 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
Christoph Lameter7835b982006-12-10 02:20:22 -08003541
Ingo Molnardd41f592007-07-09 18:51:59 +02003542 spin_lock(&rq->lock);
Ingo Molnar546fe3c2007-08-09 11:16:51 +02003543 __update_rq_clock(rq);
Ingo Molnar529c7722007-08-10 23:05:11 +02003544 /*
3545 * Let rq->clock advance by at least TICK_NSEC:
3546 */
3547 if (unlikely(rq->clock < next_tick))
3548 rq->clock = next_tick;
3549 rq->tick_timestamp = rq->clock;
Ingo Molnarf1a438d2007-08-09 11:16:45 +02003550 update_cpu_load(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02003551 if (curr != rq->idle) /* FIXME: needed? */
3552 curr->sched_class->task_tick(rq, curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02003553 spin_unlock(&rq->lock);
3554
Christoph Lametere418e1c2006-12-10 02:20:23 -08003555#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02003556 rq->idle_at_tick = idle_cpu(cpu);
3557 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08003558#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559}
3560
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3562
3563void fastcall add_preempt_count(int val)
3564{
3565 /*
3566 * Underflow?
3567 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003568 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3569 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 preempt_count() += val;
3571 /*
3572 * Spinlock count overflowing soon?
3573 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08003574 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3575 PREEMPT_MASK - 10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576}
3577EXPORT_SYMBOL(add_preempt_count);
3578
3579void fastcall sub_preempt_count(int val)
3580{
3581 /*
3582 * Underflow?
3583 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003584 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3585 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 /*
3587 * Is the spinlock portion underflowing?
3588 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003589 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3590 !(preempt_count() & PREEMPT_MASK)))
3591 return;
3592
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 preempt_count() -= val;
3594}
3595EXPORT_SYMBOL(sub_preempt_count);
3596
3597#endif
3598
3599/*
Ingo Molnardd41f592007-07-09 18:51:59 +02003600 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003602static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603{
Satyam Sharma838225b2007-10-24 18:23:50 +02003604 struct pt_regs *regs = get_irq_regs();
3605
3606 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3607 prev->comm, prev->pid, preempt_count());
3608
Ingo Molnardd41f592007-07-09 18:51:59 +02003609 debug_show_held_locks(prev);
3610 if (irqs_disabled())
3611 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02003612
3613 if (regs)
3614 show_regs(regs);
3615 else
3616 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02003617}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618
Ingo Molnardd41f592007-07-09 18:51:59 +02003619/*
3620 * Various schedule()-time debugging checks and statistics:
3621 */
3622static inline void schedule_debug(struct task_struct *prev)
3623{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003625 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626 * schedule() atomically, we ignore that path for now.
3627 * Otherwise, whine if we are scheduling when we should not be.
3628 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003629 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3630 __schedule_bug(prev);
3631
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3633
Ingo Molnar2d723762007-10-15 17:00:12 +02003634 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02003635#ifdef CONFIG_SCHEDSTATS
3636 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02003637 schedstat_inc(this_rq(), bkl_count);
3638 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02003639 }
3640#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02003641}
3642
3643/*
3644 * Pick up the highest-prio task:
3645 */
3646static inline struct task_struct *
Ingo Molnarff95f3d2007-08-09 11:16:49 +02003647pick_next_task(struct rq *rq, struct task_struct *prev)
Ingo Molnardd41f592007-07-09 18:51:59 +02003648{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003649 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02003650 struct task_struct *p;
3651
3652 /*
3653 * Optimization: we know that if all tasks are in
3654 * the fair class we can call that function directly:
3655 */
3656 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02003657 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02003658 if (likely(p))
3659 return p;
3660 }
3661
3662 class = sched_class_highest;
3663 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02003664 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02003665 if (p)
3666 return p;
3667 /*
3668 * Will never be NULL as the idle class always
3669 * returns a non-NULL p:
3670 */
3671 class = class->next;
3672 }
3673}
3674
3675/*
3676 * schedule() is the main scheduler function.
3677 */
3678asmlinkage void __sched schedule(void)
3679{
3680 struct task_struct *prev, *next;
3681 long *switch_count;
3682 struct rq *rq;
Ingo Molnardd41f592007-07-09 18:51:59 +02003683 int cpu;
3684
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685need_resched:
3686 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02003687 cpu = smp_processor_id();
3688 rq = cpu_rq(cpu);
3689 rcu_qsctr_inc(cpu);
3690 prev = rq->curr;
3691 switch_count = &prev->nivcsw;
3692
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693 release_kernel_lock(prev);
3694need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695
Ingo Molnardd41f592007-07-09 18:51:59 +02003696 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697
Ingo Molnar1e819952007-10-15 17:00:13 +02003698 /*
3699 * Do the rq-clock update outside the rq lock:
3700 */
3701 local_irq_disable();
Ingo Molnarc1b3da32007-08-09 11:16:47 +02003702 __update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02003703 spin_lock(&rq->lock);
3704 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705
Ingo Molnardd41f592007-07-09 18:51:59 +02003706 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3707 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3708 unlikely(signal_pending(prev)))) {
3709 prev->state = TASK_RUNNING;
3710 } else {
Ingo Molnar2e1cb742007-08-09 11:16:49 +02003711 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02003712 }
3713 switch_count = &prev->nvcsw;
3714 }
3715
3716 if (unlikely(!rq->nr_running))
3717 idle_balance(cpu, rq);
3718
Ingo Molnar31ee5292007-08-09 11:16:49 +02003719 prev->sched_class->put_prev_task(rq, prev);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02003720 next = pick_next_task(rq, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721
3722 sched_info_switch(prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02003723
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724 if (likely(prev != next)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725 rq->nr_switches++;
3726 rq->curr = next;
3727 ++*switch_count;
3728
Ingo Molnardd41f592007-07-09 18:51:59 +02003729 context_switch(rq, prev, next); /* unlocks the rq */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730 } else
3731 spin_unlock_irq(&rq->lock);
3732
Ingo Molnardd41f592007-07-09 18:51:59 +02003733 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3734 cpu = smp_processor_id();
3735 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736 goto need_resched_nonpreemptible;
Ingo Molnardd41f592007-07-09 18:51:59 +02003737 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738 preempt_enable_no_resched();
3739 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3740 goto need_resched;
3741}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003742EXPORT_SYMBOL(schedule);
3743
3744#ifdef CONFIG_PREEMPT
3745/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003746 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003747 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748 * occur there and call schedule directly.
3749 */
3750asmlinkage void __sched preempt_schedule(void)
3751{
3752 struct thread_info *ti = current_thread_info();
3753#ifdef CONFIG_PREEMPT_BKL
3754 struct task_struct *task = current;
3755 int saved_lock_depth;
3756#endif
3757 /*
3758 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003759 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07003761 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762 return;
3763
Andi Kleen3a5c3592007-10-15 17:00:14 +02003764 do {
3765 add_preempt_count(PREEMPT_ACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766
Andi Kleen3a5c3592007-10-15 17:00:14 +02003767 /*
3768 * We keep the big kernel semaphore locked, but we
3769 * clear ->lock_depth so that schedule() doesnt
3770 * auto-release the semaphore:
3771 */
3772#ifdef CONFIG_PREEMPT_BKL
3773 saved_lock_depth = task->lock_depth;
3774 task->lock_depth = -1;
3775#endif
3776 schedule();
3777#ifdef CONFIG_PREEMPT_BKL
3778 task->lock_depth = saved_lock_depth;
3779#endif
3780 sub_preempt_count(PREEMPT_ACTIVE);
3781
3782 /*
3783 * Check again in case we missed a preemption opportunity
3784 * between schedule and now.
3785 */
3786 barrier();
3787 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789EXPORT_SYMBOL(preempt_schedule);
3790
3791/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003792 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07003793 * off of irq context.
3794 * Note, that this is called and return with irqs disabled. This will
3795 * protect us against recursive calling from irq.
3796 */
3797asmlinkage void __sched preempt_schedule_irq(void)
3798{
3799 struct thread_info *ti = current_thread_info();
3800#ifdef CONFIG_PREEMPT_BKL
3801 struct task_struct *task = current;
3802 int saved_lock_depth;
3803#endif
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003804 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805 BUG_ON(ti->preempt_count || !irqs_disabled());
3806
Andi Kleen3a5c3592007-10-15 17:00:14 +02003807 do {
3808 add_preempt_count(PREEMPT_ACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809
Andi Kleen3a5c3592007-10-15 17:00:14 +02003810 /*
3811 * We keep the big kernel semaphore locked, but we
3812 * clear ->lock_depth so that schedule() doesnt
3813 * auto-release the semaphore:
3814 */
3815#ifdef CONFIG_PREEMPT_BKL
3816 saved_lock_depth = task->lock_depth;
3817 task->lock_depth = -1;
3818#endif
3819 local_irq_enable();
3820 schedule();
3821 local_irq_disable();
3822#ifdef CONFIG_PREEMPT_BKL
3823 task->lock_depth = saved_lock_depth;
3824#endif
3825 sub_preempt_count(PREEMPT_ACTIVE);
3826
3827 /*
3828 * Check again in case we missed a preemption opportunity
3829 * between schedule and now.
3830 */
3831 barrier();
3832 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833}
3834
3835#endif /* CONFIG_PREEMPT */
3836
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003837int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3838 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839{
Ingo Molnar48f24c42006-07-03 00:25:40 -07003840 return try_to_wake_up(curr->private, mode, sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842EXPORT_SYMBOL(default_wake_function);
3843
3844/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003845 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3846 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847 * number) then we wake all the non-exclusive tasks and one exclusive task.
3848 *
3849 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003850 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3852 */
3853static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3854 int nr_exclusive, int sync, void *key)
3855{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02003856 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003857
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02003858 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07003859 unsigned flags = curr->flags;
3860
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 if (curr->func(curr, mode, sync, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07003862 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863 break;
3864 }
3865}
3866
3867/**
3868 * __wake_up - wake up threads blocked on a waitqueue.
3869 * @q: the waitqueue
3870 * @mode: which threads
3871 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07003872 * @key: is directly passed to the wakeup function
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 */
3874void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003875 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876{
3877 unsigned long flags;
3878
3879 spin_lock_irqsave(&q->lock, flags);
3880 __wake_up_common(q, mode, nr_exclusive, 0, key);
3881 spin_unlock_irqrestore(&q->lock, flags);
3882}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883EXPORT_SYMBOL(__wake_up);
3884
3885/*
3886 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3887 */
3888void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3889{
3890 __wake_up_common(q, mode, 1, 0, NULL);
3891}
3892
3893/**
Martin Waitz67be2dd2005-05-01 08:59:26 -07003894 * __wake_up_sync - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003895 * @q: the waitqueue
3896 * @mode: which threads
3897 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3898 *
3899 * The sync wakeup differs that the waker knows that it will schedule
3900 * away soon, so while the target thread will be woken up, it will not
3901 * be migrated to another CPU - ie. the two threads are 'synchronized'
3902 * with each other. This can prevent needless bouncing between CPUs.
3903 *
3904 * On UP it can prevent extra preemption.
3905 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003906void fastcall
3907__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908{
3909 unsigned long flags;
3910 int sync = 1;
3911
3912 if (unlikely(!q))
3913 return;
3914
3915 if (unlikely(!nr_exclusive))
3916 sync = 0;
3917
3918 spin_lock_irqsave(&q->lock, flags);
3919 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3920 spin_unlock_irqrestore(&q->lock, flags);
3921}
3922EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3923
Ingo Molnarb15136e2007-10-24 18:23:48 +02003924void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925{
3926 unsigned long flags;
3927
3928 spin_lock_irqsave(&x->wait.lock, flags);
3929 x->done++;
3930 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3931 1, 0, NULL);
3932 spin_unlock_irqrestore(&x->wait.lock, flags);
3933}
3934EXPORT_SYMBOL(complete);
3935
Ingo Molnarb15136e2007-10-24 18:23:48 +02003936void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937{
3938 unsigned long flags;
3939
3940 spin_lock_irqsave(&x->wait.lock, flags);
3941 x->done += UINT_MAX/2;
3942 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3943 0, 0, NULL);
3944 spin_unlock_irqrestore(&x->wait.lock, flags);
3945}
3946EXPORT_SYMBOL(complete_all);
3947
Andi Kleen8cbbe862007-10-15 17:00:14 +02003948static inline long __sched
3949do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 if (!x->done) {
3952 DECLARE_WAITQUEUE(wait, current);
3953
3954 wait.flags |= WQ_FLAG_EXCLUSIVE;
3955 __add_wait_queue_tail(&x->wait, &wait);
3956 do {
Andi Kleen8cbbe862007-10-15 17:00:14 +02003957 if (state == TASK_INTERRUPTIBLE &&
3958 signal_pending(current)) {
3959 __remove_wait_queue(&x->wait, &wait);
3960 return -ERESTARTSYS;
3961 }
3962 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003963 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02003964 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 spin_lock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02003966 if (!timeout) {
3967 __remove_wait_queue(&x->wait, &wait);
3968 return timeout;
3969 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970 } while (!x->done);
3971 __remove_wait_queue(&x->wait, &wait);
3972 }
3973 x->done--;
Andi Kleen8cbbe862007-10-15 17:00:14 +02003974 return timeout;
3975}
3976
3977static long __sched
3978wait_for_common(struct completion *x, long timeout, int state)
3979{
3980 might_sleep();
3981
3982 spin_lock_irq(&x->wait.lock);
3983 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02003985 return timeout;
3986}
3987
Ingo Molnarb15136e2007-10-24 18:23:48 +02003988void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02003989{
3990 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991}
3992EXPORT_SYMBOL(wait_for_completion);
3993
Ingo Molnarb15136e2007-10-24 18:23:48 +02003994unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07003995wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3996{
Andi Kleen8cbbe862007-10-15 17:00:14 +02003997 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998}
3999EXPORT_SYMBOL(wait_for_completion_timeout);
4000
Andi Kleen8cbbe862007-10-15 17:00:14 +02004001int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002{
Andi Kleen51e97992007-10-18 21:32:55 +02004003 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4004 if (t == -ERESTARTSYS)
4005 return t;
4006 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007}
4008EXPORT_SYMBOL(wait_for_completion_interruptible);
4009
Ingo Molnarb15136e2007-10-24 18:23:48 +02004010unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011wait_for_completion_interruptible_timeout(struct completion *x,
4012 unsigned long timeout)
4013{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004014 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015}
4016EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4017
Andi Kleen8cbbe862007-10-15 17:00:14 +02004018static long __sched
4019sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02004020{
4021 unsigned long flags;
4022 wait_queue_t wait;
4023
4024 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025
Andi Kleen8cbbe862007-10-15 17:00:14 +02004026 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027
Andi Kleen8cbbe862007-10-15 17:00:14 +02004028 spin_lock_irqsave(&q->lock, flags);
4029 __add_wait_queue(q, &wait);
4030 spin_unlock(&q->lock);
4031 timeout = schedule_timeout(timeout);
4032 spin_lock_irq(&q->lock);
4033 __remove_wait_queue(q, &wait);
4034 spin_unlock_irqrestore(&q->lock, flags);
4035
4036 return timeout;
4037}
4038
4039void __sched interruptible_sleep_on(wait_queue_head_t *q)
4040{
4041 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043EXPORT_SYMBOL(interruptible_sleep_on);
4044
Ingo Molnar0fec1712007-07-09 18:52:01 +02004045long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004046interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004048 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4051
Ingo Molnar0fec1712007-07-09 18:52:01 +02004052void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004054 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056EXPORT_SYMBOL(sleep_on);
4057
Ingo Molnar0fec1712007-07-09 18:52:01 +02004058long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004060 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004061}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062EXPORT_SYMBOL(sleep_on_timeout);
4063
Ingo Molnarb29739f2006-06-27 02:54:51 -07004064#ifdef CONFIG_RT_MUTEXES
4065
4066/*
4067 * rt_mutex_setprio - set the current priority of a task
4068 * @p: task
4069 * @prio: prio value (kernel-internal form)
4070 *
4071 * This function changes the 'effective' priority of a task. It does
4072 * not touch ->normal_prio like __setscheduler().
4073 *
4074 * Used by the rt_mutex code to implement priority inheritance logic.
4075 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004076void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07004077{
4078 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004079 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004080 struct rq *rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004081
4082 BUG_ON(prio < 0 || prio > MAX_PRIO);
4083
4084 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004085 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004086
Andrew Mortond5f9f942007-05-08 20:27:06 -07004087 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004088 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004089 running = task_current(rq, p);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004090 if (on_rq) {
Ingo Molnar69be72c2007-08-09 11:16:49 +02004091 dequeue_task(rq, p, 0);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004092 if (running)
4093 p->sched_class->put_prev_task(rq, p);
4094 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004095
4096 if (rt_prio(prio))
4097 p->sched_class = &rt_sched_class;
4098 else
4099 p->sched_class = &fair_sched_class;
4100
Ingo Molnarb29739f2006-06-27 02:54:51 -07004101 p->prio = prio;
4102
Ingo Molnardd41f592007-07-09 18:51:59 +02004103 if (on_rq) {
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004104 if (running)
4105 p->sched_class->set_curr_task(rq);
Ingo Molnar8159f872007-08-09 11:16:49 +02004106 enqueue_task(rq, p, 0);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004107 /*
4108 * Reschedule if we are currently running on this runqueue and
Andrew Mortond5f9f942007-05-08 20:27:06 -07004109 * our priority decreased, or if we are not currently running on
4110 * this runqueue and our priority is higher than the current's
Ingo Molnarb29739f2006-06-27 02:54:51 -07004111 */
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004112 if (running) {
Andrew Mortond5f9f942007-05-08 20:27:06 -07004113 if (p->prio > oldprio)
4114 resched_task(rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02004115 } else {
4116 check_preempt_curr(rq, p);
4117 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07004118 }
4119 task_rq_unlock(rq, &flags);
4120}
4121
4122#endif
4123
Ingo Molnar36c8b582006-07-03 00:25:41 -07004124void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125{
Ingo Molnardd41f592007-07-09 18:51:59 +02004126 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004128 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004129
4130 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4131 return;
4132 /*
4133 * We have to be careful, if called from sys_setpriority(),
4134 * the task might be in the middle of scheduling on another CPU.
4135 */
4136 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004137 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138 /*
4139 * The RT priorities are set via sched_setscheduler(), but we still
4140 * allow the 'normal' nice value to be set - but as expected
4141 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02004142 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143 */
Ingo Molnare05606d2007-07-09 18:51:59 +02004144 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145 p->static_prio = NICE_TO_PRIO(nice);
4146 goto out_unlock;
4147 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004148 on_rq = p->se.on_rq;
Srivatsa Vaddagiri58e2d4c2008-01-25 21:08:00 +01004149 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02004150 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004151
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07004153 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004154 old_prio = p->prio;
4155 p->prio = effective_prio(p);
4156 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004157
Ingo Molnardd41f592007-07-09 18:51:59 +02004158 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004159 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07004161 * If the task increased its priority or is running and
4162 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07004164 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004165 resched_task(rq->curr);
4166 }
4167out_unlock:
4168 task_rq_unlock(rq, &flags);
4169}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170EXPORT_SYMBOL(set_user_nice);
4171
Matt Mackalle43379f2005-05-01 08:59:00 -07004172/*
4173 * can_nice - check if a task can reduce its nice value
4174 * @p: task
4175 * @nice: nice value
4176 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004177int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07004178{
Matt Mackall024f4742005-08-18 11:24:19 -07004179 /* convert nice value [19,-20] to rlimit style value [1,40] */
4180 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004181
Matt Mackalle43379f2005-05-01 08:59:00 -07004182 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4183 capable(CAP_SYS_NICE));
4184}
4185
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186#ifdef __ARCH_WANT_SYS_NICE
4187
4188/*
4189 * sys_nice - change the priority of the current process.
4190 * @increment: priority increment
4191 *
4192 * sys_setpriority is a more generic, but much slower function that
4193 * does similar things.
4194 */
4195asmlinkage long sys_nice(int increment)
4196{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004197 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198
4199 /*
4200 * Setpriority might change our priority at the same moment.
4201 * We don't have to worry. Conceptually one call occurs first
4202 * and we have a single winner.
4203 */
Matt Mackalle43379f2005-05-01 08:59:00 -07004204 if (increment < -40)
4205 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004206 if (increment > 40)
4207 increment = 40;
4208
4209 nice = PRIO_TO_NICE(current->static_prio) + increment;
4210 if (nice < -20)
4211 nice = -20;
4212 if (nice > 19)
4213 nice = 19;
4214
Matt Mackalle43379f2005-05-01 08:59:00 -07004215 if (increment < 0 && !can_nice(current, nice))
4216 return -EPERM;
4217
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 retval = security_task_setnice(current, nice);
4219 if (retval)
4220 return retval;
4221
4222 set_user_nice(current, nice);
4223 return 0;
4224}
4225
4226#endif
4227
4228/**
4229 * task_prio - return the priority value of a given task.
4230 * @p: the task in question.
4231 *
4232 * This is the priority value as seen by users in /proc.
4233 * RT tasks are offset by -200. Normal tasks are centered
4234 * around 0, value goes from -16 to +15.
4235 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004236int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004237{
4238 return p->prio - MAX_RT_PRIO;
4239}
4240
4241/**
4242 * task_nice - return the nice value of a given task.
4243 * @p: the task in question.
4244 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004245int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004246{
4247 return TASK_NICE(p);
4248}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004249EXPORT_SYMBOL_GPL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250
4251/**
4252 * idle_cpu - is a given cpu idle currently?
4253 * @cpu: the processor in question.
4254 */
4255int idle_cpu(int cpu)
4256{
4257 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4258}
4259
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260/**
4261 * idle_task - return the idle task for a given cpu.
4262 * @cpu: the processor in question.
4263 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004264struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004265{
4266 return cpu_rq(cpu)->idle;
4267}
4268
4269/**
4270 * find_process_by_pid - find a process with a matching PID value.
4271 * @pid: the pid in question.
4272 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02004273static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004274{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07004275 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004276}
4277
4278/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02004279static void
4280__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281{
Ingo Molnardd41f592007-07-09 18:51:59 +02004282 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004283
Linus Torvalds1da177e2005-04-16 15:20:36 -07004284 p->policy = policy;
Ingo Molnardd41f592007-07-09 18:51:59 +02004285 switch (p->policy) {
4286 case SCHED_NORMAL:
4287 case SCHED_BATCH:
4288 case SCHED_IDLE:
4289 p->sched_class = &fair_sched_class;
4290 break;
4291 case SCHED_FIFO:
4292 case SCHED_RR:
4293 p->sched_class = &rt_sched_class;
4294 break;
4295 }
4296
Linus Torvalds1da177e2005-04-16 15:20:36 -07004297 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004298 p->normal_prio = normal_prio(p);
4299 /* we are holding p->pi_lock already */
4300 p->prio = rt_mutex_getprio(p);
Peter Williams2dd73a42006-06-27 02:54:34 -07004301 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302}
4303
4304/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004305 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004306 * @p: the task in question.
4307 * @policy: new policy.
4308 * @param: structure containing the new RT priority.
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004309 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004310 * NOTE that the task may be already dead.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004311 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004312int sched_setscheduler(struct task_struct *p, int policy,
4313 struct sched_param *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004314{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004315 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004317 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004318
Steven Rostedt66e53932006-06-27 02:54:44 -07004319 /* may grab non-irq protected spin_locks */
4320 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07004321recheck:
4322 /* double check policy once rq lock held */
4323 if (policy < 0)
4324 policy = oldpolicy = p->policy;
4325 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
Ingo Molnardd41f592007-07-09 18:51:59 +02004326 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4327 policy != SCHED_IDLE)
Ingo Molnarb0a94992006-01-14 13:20:41 -08004328 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004329 /*
4330 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02004331 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4332 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004333 */
4334 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004335 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04004336 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004337 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02004338 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004339 return -EINVAL;
4340
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004341 /*
4342 * Allow unprivileged RT tasks to decrease priority:
4343 */
4344 if (!capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02004345 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004346 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004347
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004348 if (!lock_task_sighand(p, &flags))
4349 return -ESRCH;
4350 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4351 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004352
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004353 /* can't set/change the rt policy */
4354 if (policy != p->policy && !rlim_rtprio)
4355 return -EPERM;
4356
4357 /* can't increase priority */
4358 if (param->sched_priority > p->rt_priority &&
4359 param->sched_priority > rlim_rtprio)
4360 return -EPERM;
4361 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004362 /*
4363 * Like positive nice levels, dont allow tasks to
4364 * move out of SCHED_IDLE either:
4365 */
4366 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4367 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004368
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004369 /* can't change other user's priorities */
4370 if ((current->euid != p->euid) &&
4371 (current->euid != p->uid))
4372 return -EPERM;
4373 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004374
4375 retval = security_task_setscheduler(p, policy, param);
4376 if (retval)
4377 return retval;
4378 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07004379 * make sure no PI-waiters arrive (or leave) while we are
4380 * changing the priority of the task:
4381 */
4382 spin_lock_irqsave(&p->pi_lock, flags);
4383 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004384 * To be able to change p->policy safely, the apropriate
4385 * runqueue lock must be held.
4386 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07004387 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004388 /* recheck policy now with rq lock held */
4389 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4390 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004391 __task_rq_unlock(rq);
4392 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004393 goto recheck;
4394 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02004395 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004396 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004397 running = task_current(rq, p);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004398 if (on_rq) {
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004399 deactivate_task(rq, p, 0);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004400 if (running)
4401 p->sched_class->put_prev_task(rq, p);
4402 }
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02004403
Linus Torvalds1da177e2005-04-16 15:20:36 -07004404 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004405 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02004406
Ingo Molnardd41f592007-07-09 18:51:59 +02004407 if (on_rq) {
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004408 if (running)
4409 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004410 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004411 /*
4412 * Reschedule if we are currently running on this runqueue and
Andrew Mortond5f9f942007-05-08 20:27:06 -07004413 * our priority decreased, or if we are not currently running on
4414 * this runqueue and our priority is higher than the current's
Linus Torvalds1da177e2005-04-16 15:20:36 -07004415 */
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004416 if (running) {
Andrew Mortond5f9f942007-05-08 20:27:06 -07004417 if (p->prio > oldprio)
4418 resched_task(rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02004419 } else {
4420 check_preempt_curr(rq, p);
4421 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004422 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07004423 __task_rq_unlock(rq);
4424 spin_unlock_irqrestore(&p->pi_lock, flags);
4425
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07004426 rt_mutex_adjust_pi(p);
4427
Linus Torvalds1da177e2005-04-16 15:20:36 -07004428 return 0;
4429}
4430EXPORT_SYMBOL_GPL(sched_setscheduler);
4431
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004432static int
4433do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004434{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004435 struct sched_param lparam;
4436 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004437 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004438
4439 if (!param || pid < 0)
4440 return -EINVAL;
4441 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4442 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004443
4444 rcu_read_lock();
4445 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004446 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004447 if (p != NULL)
4448 retval = sched_setscheduler(p, policy, &lparam);
4449 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07004450
Linus Torvalds1da177e2005-04-16 15:20:36 -07004451 return retval;
4452}
4453
4454/**
4455 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4456 * @pid: the pid in question.
4457 * @policy: new policy.
4458 * @param: structure containing the new RT priority.
4459 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004460asmlinkage long
4461sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004462{
Jason Baronc21761f2006-01-18 17:43:03 -08004463 /* negative values for policy are not valid */
4464 if (policy < 0)
4465 return -EINVAL;
4466
Linus Torvalds1da177e2005-04-16 15:20:36 -07004467 return do_sched_setscheduler(pid, policy, param);
4468}
4469
4470/**
4471 * sys_sched_setparam - set/change the RT priority of a thread
4472 * @pid: the pid in question.
4473 * @param: structure containing the new RT priority.
4474 */
4475asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4476{
4477 return do_sched_setscheduler(pid, -1, param);
4478}
4479
4480/**
4481 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4482 * @pid: the pid in question.
4483 */
4484asmlinkage long sys_sched_getscheduler(pid_t pid)
4485{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004486 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004487 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004488
4489 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004490 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004491
4492 retval = -ESRCH;
4493 read_lock(&tasklist_lock);
4494 p = find_process_by_pid(pid);
4495 if (p) {
4496 retval = security_task_getscheduler(p);
4497 if (!retval)
4498 retval = p->policy;
4499 }
4500 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004501 return retval;
4502}
4503
4504/**
4505 * sys_sched_getscheduler - get the RT priority of a thread
4506 * @pid: the pid in question.
4507 * @param: structure containing the RT priority.
4508 */
4509asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4510{
4511 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004512 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004513 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004514
4515 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004516 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004517
4518 read_lock(&tasklist_lock);
4519 p = find_process_by_pid(pid);
4520 retval = -ESRCH;
4521 if (!p)
4522 goto out_unlock;
4523
4524 retval = security_task_getscheduler(p);
4525 if (retval)
4526 goto out_unlock;
4527
4528 lp.sched_priority = p->rt_priority;
4529 read_unlock(&tasklist_lock);
4530
4531 /*
4532 * This one might sleep, we cannot do it with a spinlock held ...
4533 */
4534 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4535
Linus Torvalds1da177e2005-04-16 15:20:36 -07004536 return retval;
4537
4538out_unlock:
4539 read_unlock(&tasklist_lock);
4540 return retval;
4541}
4542
4543long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4544{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004545 cpumask_t cpus_allowed;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004546 struct task_struct *p;
4547 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004548
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004549 mutex_lock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004550 read_lock(&tasklist_lock);
4551
4552 p = find_process_by_pid(pid);
4553 if (!p) {
4554 read_unlock(&tasklist_lock);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004555 mutex_unlock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004556 return -ESRCH;
4557 }
4558
4559 /*
4560 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004561 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07004562 * usage count and then drop tasklist_lock.
4563 */
4564 get_task_struct(p);
4565 read_unlock(&tasklist_lock);
4566
4567 retval = -EPERM;
4568 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4569 !capable(CAP_SYS_NICE))
4570 goto out_unlock;
4571
David Quigleye7834f82006-06-23 02:03:59 -07004572 retval = security_task_setscheduler(p, 0, NULL);
4573 if (retval)
4574 goto out_unlock;
4575
Linus Torvalds1da177e2005-04-16 15:20:36 -07004576 cpus_allowed = cpuset_cpus_allowed(p);
4577 cpus_and(new_mask, new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07004578 again:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004579 retval = set_cpus_allowed(p, new_mask);
4580
Paul Menage8707d8b2007-10-18 23:40:22 -07004581 if (!retval) {
4582 cpus_allowed = cpuset_cpus_allowed(p);
4583 if (!cpus_subset(new_mask, cpus_allowed)) {
4584 /*
4585 * We must have raced with a concurrent cpuset
4586 * update. Just reset the cpus_allowed to the
4587 * cpuset's cpus_allowed
4588 */
4589 new_mask = cpus_allowed;
4590 goto again;
4591 }
4592 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004593out_unlock:
4594 put_task_struct(p);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004595 mutex_unlock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004596 return retval;
4597}
4598
4599static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4600 cpumask_t *new_mask)
4601{
4602 if (len < sizeof(cpumask_t)) {
4603 memset(new_mask, 0, sizeof(cpumask_t));
4604 } else if (len > sizeof(cpumask_t)) {
4605 len = sizeof(cpumask_t);
4606 }
4607 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4608}
4609
4610/**
4611 * sys_sched_setaffinity - set the cpu affinity of a process
4612 * @pid: pid of the process
4613 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4614 * @user_mask_ptr: user-space pointer to the new cpu mask
4615 */
4616asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4617 unsigned long __user *user_mask_ptr)
4618{
4619 cpumask_t new_mask;
4620 int retval;
4621
4622 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4623 if (retval)
4624 return retval;
4625
4626 return sched_setaffinity(pid, new_mask);
4627}
4628
4629/*
4630 * Represents all cpu's present in the system
4631 * In systems capable of hotplug, this map could dynamically grow
4632 * as new cpu's are detected in the system via any platform specific
4633 * method, such as ACPI for e.g.
4634 */
4635
Andi Kleen4cef0c62006-01-11 22:44:57 +01004636cpumask_t cpu_present_map __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004637EXPORT_SYMBOL(cpu_present_map);
4638
4639#ifndef CONFIG_SMP
Andi Kleen4cef0c62006-01-11 22:44:57 +01004640cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07004641EXPORT_SYMBOL(cpu_online_map);
4642
Andi Kleen4cef0c62006-01-11 22:44:57 +01004643cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07004644EXPORT_SYMBOL(cpu_possible_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004645#endif
4646
4647long sched_getaffinity(pid_t pid, cpumask_t *mask)
4648{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004649 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004650 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004651
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004652 mutex_lock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004653 read_lock(&tasklist_lock);
4654
4655 retval = -ESRCH;
4656 p = find_process_by_pid(pid);
4657 if (!p)
4658 goto out_unlock;
4659
David Quigleye7834f82006-06-23 02:03:59 -07004660 retval = security_task_getscheduler(p);
4661 if (retval)
4662 goto out_unlock;
4663
Jack Steiner2f7016d2006-02-01 03:05:18 -08004664 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004665
4666out_unlock:
4667 read_unlock(&tasklist_lock);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004668 mutex_unlock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004669
Ulrich Drepper9531b622007-08-09 11:16:46 +02004670 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004671}
4672
4673/**
4674 * sys_sched_getaffinity - get the cpu affinity of a process
4675 * @pid: pid of the process
4676 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4677 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4678 */
4679asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4680 unsigned long __user *user_mask_ptr)
4681{
4682 int ret;
4683 cpumask_t mask;
4684
4685 if (len < sizeof(cpumask_t))
4686 return -EINVAL;
4687
4688 ret = sched_getaffinity(pid, &mask);
4689 if (ret < 0)
4690 return ret;
4691
4692 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4693 return -EFAULT;
4694
4695 return sizeof(cpumask_t);
4696}
4697
4698/**
4699 * sys_sched_yield - yield the current processor to other threads.
4700 *
Ingo Molnardd41f592007-07-09 18:51:59 +02004701 * This function yields the current CPU to other tasks. If there are no
4702 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004703 */
4704asmlinkage long sys_sched_yield(void)
4705{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004706 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004707
Ingo Molnar2d723762007-10-15 17:00:12 +02004708 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02004709 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004710
4711 /*
4712 * Since we are going to call schedule() anyway, there's
4713 * no need to preempt or enable interrupts:
4714 */
4715 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07004716 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004717 _raw_spin_unlock(&rq->lock);
4718 preempt_enable_no_resched();
4719
4720 schedule();
4721
4722 return 0;
4723}
4724
Andrew Mortone7b38402006-06-30 01:56:00 -07004725static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004726{
Ingo Molnar8e0a43d2006-06-23 02:05:23 -07004727#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4728 __might_sleep(__FILE__, __LINE__);
4729#endif
Ingo Molnar5bbcfd92005-07-07 17:57:04 -07004730 /*
4731 * The BKS might be reacquired before we have dropped
4732 * PREEMPT_ACTIVE, which could trigger a second
4733 * cond_resched() call.
4734 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004735 do {
4736 add_preempt_count(PREEMPT_ACTIVE);
4737 schedule();
4738 sub_preempt_count(PREEMPT_ACTIVE);
4739 } while (need_resched());
4740}
4741
4742int __sched cond_resched(void)
4743{
Ingo Molnar94142322006-12-29 16:48:13 -08004744 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4745 system_state == SYSTEM_RUNNING) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004746 __cond_resched();
4747 return 1;
4748 }
4749 return 0;
4750}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004751EXPORT_SYMBOL(cond_resched);
4752
4753/*
4754 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4755 * call schedule, and on return reacquire the lock.
4756 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004757 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07004758 * operations here to prevent schedule() from being called twice (once via
4759 * spin_unlock(), once by hand).
4760 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004761int cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004762{
Jan Kara6df3cec2005-06-13 15:52:32 -07004763 int ret = 0;
4764
Linus Torvalds1da177e2005-04-16 15:20:36 -07004765 if (need_lockbreak(lock)) {
4766 spin_unlock(lock);
4767 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07004768 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004769 spin_lock(lock);
4770 }
Ingo Molnar94142322006-12-29 16:48:13 -08004771 if (need_resched() && system_state == SYSTEM_RUNNING) {
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07004772 spin_release(&lock->dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004773 _raw_spin_unlock(lock);
4774 preempt_enable_no_resched();
4775 __cond_resched();
Jan Kara6df3cec2005-06-13 15:52:32 -07004776 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004777 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004778 }
Jan Kara6df3cec2005-06-13 15:52:32 -07004779 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004780}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004781EXPORT_SYMBOL(cond_resched_lock);
4782
4783int __sched cond_resched_softirq(void)
4784{
4785 BUG_ON(!in_softirq());
4786
Ingo Molnar94142322006-12-29 16:48:13 -08004787 if (need_resched() && system_state == SYSTEM_RUNNING) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07004788 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004789 __cond_resched();
4790 local_bh_disable();
4791 return 1;
4792 }
4793 return 0;
4794}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004795EXPORT_SYMBOL(cond_resched_softirq);
4796
Linus Torvalds1da177e2005-04-16 15:20:36 -07004797/**
4798 * yield - yield the current processor to other threads.
4799 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004800 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07004801 * thread runnable and calls sys_sched_yield().
4802 */
4803void __sched yield(void)
4804{
4805 set_current_state(TASK_RUNNING);
4806 sys_sched_yield();
4807}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004808EXPORT_SYMBOL(yield);
4809
4810/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004811 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07004812 * that process accounting knows that this is a task in IO wait state.
4813 *
4814 * But don't do that if it is a deliberate, throttling IO wait (this task
4815 * has set its backing_dev_info: the queue against which it should throttle)
4816 */
4817void __sched io_schedule(void)
4818{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004819 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004820
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004821 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004822 atomic_inc(&rq->nr_iowait);
4823 schedule();
4824 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004825 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004826}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004827EXPORT_SYMBOL(io_schedule);
4828
4829long __sched io_schedule_timeout(long timeout)
4830{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004831 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004832 long ret;
4833
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004834 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004835 atomic_inc(&rq->nr_iowait);
4836 ret = schedule_timeout(timeout);
4837 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004838 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004839 return ret;
4840}
4841
4842/**
4843 * sys_sched_get_priority_max - return maximum RT priority.
4844 * @policy: scheduling class.
4845 *
4846 * this syscall returns the maximum rt_priority that can be used
4847 * by a given scheduling class.
4848 */
4849asmlinkage long sys_sched_get_priority_max(int policy)
4850{
4851 int ret = -EINVAL;
4852
4853 switch (policy) {
4854 case SCHED_FIFO:
4855 case SCHED_RR:
4856 ret = MAX_USER_RT_PRIO-1;
4857 break;
4858 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08004859 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02004860 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004861 ret = 0;
4862 break;
4863 }
4864 return ret;
4865}
4866
4867/**
4868 * sys_sched_get_priority_min - return minimum RT priority.
4869 * @policy: scheduling class.
4870 *
4871 * this syscall returns the minimum rt_priority that can be used
4872 * by a given scheduling class.
4873 */
4874asmlinkage long sys_sched_get_priority_min(int policy)
4875{
4876 int ret = -EINVAL;
4877
4878 switch (policy) {
4879 case SCHED_FIFO:
4880 case SCHED_RR:
4881 ret = 1;
4882 break;
4883 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08004884 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02004885 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004886 ret = 0;
4887 }
4888 return ret;
4889}
4890
4891/**
4892 * sys_sched_rr_get_interval - return the default timeslice of a process.
4893 * @pid: pid of the process.
4894 * @interval: userspace pointer to the timeslice value.
4895 *
4896 * this syscall writes the default timeslice value of a given process
4897 * into the user-space timespec buffer. A value of '0' means infinity.
4898 */
4899asmlinkage
4900long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4901{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004902 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02004903 unsigned int time_slice;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004904 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004905 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004906
4907 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004908 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004909
4910 retval = -ESRCH;
4911 read_lock(&tasklist_lock);
4912 p = find_process_by_pid(pid);
4913 if (!p)
4914 goto out_unlock;
4915
4916 retval = security_task_getscheduler(p);
4917 if (retval)
4918 goto out_unlock;
4919
Ingo Molnar77034932007-12-04 17:04:39 +01004920 /*
4921 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4922 * tasks that are on an otherwise idle runqueue:
4923 */
4924 time_slice = 0;
4925 if (p->policy == SCHED_RR) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02004926 time_slice = DEF_TIMESLICE;
Ingo Molnar77034932007-12-04 17:04:39 +01004927 } else {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02004928 struct sched_entity *se = &p->se;
4929 unsigned long flags;
4930 struct rq *rq;
4931
4932 rq = task_rq_lock(p, &flags);
Ingo Molnar77034932007-12-04 17:04:39 +01004933 if (rq->cfs.load.weight)
4934 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02004935 task_rq_unlock(rq, &flags);
4936 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004937 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02004938 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004939 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004940 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004941
Linus Torvalds1da177e2005-04-16 15:20:36 -07004942out_unlock:
4943 read_unlock(&tasklist_lock);
4944 return retval;
4945}
4946
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004947static const char stat_nam[] = "RSDTtZX";
Ingo Molnar36c8b582006-07-03 00:25:41 -07004948
4949static void show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004950{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004951 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004952 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004953
Linus Torvalds1da177e2005-04-16 15:20:36 -07004954 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea792007-10-18 21:32:56 +02004955 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004956 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02004957#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07004958 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02004959 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004960 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02004961 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004962#else
4963 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02004964 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004965 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02004966 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004967#endif
4968#ifdef CONFIG_DEBUG_STACK_USAGE
4969 {
Al Viro10ebffd2005-11-13 16:06:56 -08004970 unsigned long *n = end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004971 while (!*n)
4972 n++;
Al Viro10ebffd2005-11-13 16:06:56 -08004973 free = (unsigned long)n - (unsigned long)end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004974 }
4975#endif
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -07004976 printk(KERN_CONT "%5lu %5d %6d\n", free,
Roland McGrathfcfd50a2008-01-09 00:03:23 -08004977 task_pid_nr(p), task_pid_nr(p->real_parent));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004978
4979 if (state != TASK_RUNNING)
4980 show_stack(p, NULL);
4981}
4982
Ingo Molnare59e2ae2006-12-06 20:35:59 -08004983void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004984{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004985 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004986
Ingo Molnar4bd77322007-07-11 21:21:47 +02004987#if BITS_PER_LONG == 32
4988 printk(KERN_INFO
4989 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004990#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02004991 printk(KERN_INFO
4992 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004993#endif
4994 read_lock(&tasklist_lock);
4995 do_each_thread(g, p) {
4996 /*
4997 * reset the NMI-timeout, listing all files on a slow
4998 * console might take alot of time:
4999 */
5000 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07005001 if (!state_filter || (p->state & state_filter))
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005002 show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005003 } while_each_thread(g, p);
5004
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07005005 touch_all_softlockup_watchdogs();
5006
Ingo Molnardd41f592007-07-09 18:51:59 +02005007#ifdef CONFIG_SCHED_DEBUG
5008 sysrq_sched_debug_show();
5009#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005010 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005011 /*
5012 * Only show locks if all tasks are dumped:
5013 */
5014 if (state_filter == -1)
5015 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005016}
5017
Ingo Molnar1df21052007-07-09 18:51:58 +02005018void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5019{
Ingo Molnardd41f592007-07-09 18:51:59 +02005020 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02005021}
5022
Ingo Molnarf340c0d2005-06-28 16:40:42 +02005023/**
5024 * init_idle - set up an idle thread for a given CPU
5025 * @idle: task in question
5026 * @cpu: cpu the idle task belongs to
5027 *
5028 * NOTE: this function does not set the idle thread's NEED_RESCHED
5029 * flag, to make booting more robust.
5030 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07005031void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005032{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005033 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005034 unsigned long flags;
5035
Ingo Molnardd41f592007-07-09 18:51:59 +02005036 __sched_fork(idle);
5037 idle->se.exec_start = sched_clock();
5038
Ingo Molnarb29739f2006-06-27 02:54:51 -07005039 idle->prio = idle->normal_prio = MAX_PRIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005040 idle->cpus_allowed = cpumask_of_cpu(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005041 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005042
5043 spin_lock_irqsave(&rq->lock, flags);
5044 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07005045#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5046 idle->oncpu = 1;
5047#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005048 spin_unlock_irqrestore(&rq->lock, flags);
5049
5050 /* Set the preempt count _outside_ the spinlocks! */
5051#if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
Al Viroa1261f52005-11-13 16:06:55 -08005052 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005053#else
Al Viroa1261f52005-11-13 16:06:55 -08005054 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005055#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02005056 /*
5057 * The idle tasks have their own, simple scheduling class:
5058 */
5059 idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005060}
5061
5062/*
5063 * In a system that switches off the HZ timer nohz_cpu_mask
5064 * indicates which cpus entered this state. This is used
5065 * in the rcu update to wait only for active cpus. For system
5066 * which do not switch off the HZ timer nohz_cpu_mask should
5067 * always be CPU_MASK_NONE.
5068 */
5069cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5070
Ingo Molnar19978ca2007-11-09 22:39:38 +01005071/*
5072 * Increase the granularity value when there are more CPUs,
5073 * because with more CPUs the 'effective latency' as visible
5074 * to users decreases. But the relationship is not linear,
5075 * so pick a second-best guess by going with the log2 of the
5076 * number of CPUs.
5077 *
5078 * This idea comes from the SD scheduler of Con Kolivas:
5079 */
5080static inline void sched_init_granularity(void)
5081{
5082 unsigned int factor = 1 + ilog2(num_online_cpus());
5083 const unsigned long limit = 200000000;
5084
5085 sysctl_sched_min_granularity *= factor;
5086 if (sysctl_sched_min_granularity > limit)
5087 sysctl_sched_min_granularity = limit;
5088
5089 sysctl_sched_latency *= factor;
5090 if (sysctl_sched_latency > limit)
5091 sysctl_sched_latency = limit;
5092
5093 sysctl_sched_wakeup_granularity *= factor;
5094 sysctl_sched_batch_wakeup_granularity *= factor;
5095}
5096
Linus Torvalds1da177e2005-04-16 15:20:36 -07005097#ifdef CONFIG_SMP
5098/*
5099 * This is how migration works:
5100 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07005101 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005102 * runqueue and wake up that CPU's migration thread.
5103 * 2) we down() the locked semaphore => thread blocks.
5104 * 3) migration thread wakes up (implicitly it forces the migrated
5105 * thread off the CPU)
5106 * 4) it gets the migration request and checks whether the migrated
5107 * task is still in the wrong runqueue.
5108 * 5) if it's in the wrong runqueue then the migration thread removes
5109 * it and puts it into the right queue.
5110 * 6) migration thread up()s the semaphore.
5111 * 7) we wake up and the migration is done.
5112 */
5113
5114/*
5115 * Change a given task's CPU affinity. Migrate the thread to a
5116 * proper CPU and schedule it away if the CPU it's executing on
5117 * is removed from the allowed bitmask.
5118 *
5119 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005120 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07005121 * call is not atomic; no spinlocks may be held.
5122 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005123int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005124{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005125 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005126 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005127 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005128 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005129
5130 rq = task_rq_lock(p, &flags);
5131 if (!cpus_intersects(new_mask, cpu_online_map)) {
5132 ret = -EINVAL;
5133 goto out;
5134 }
5135
5136 p->cpus_allowed = new_mask;
5137 /* Can the task run on the task's current CPU? If so, we're done */
5138 if (cpu_isset(task_cpu(p), new_mask))
5139 goto out;
5140
5141 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5142 /* Need help from migration thread: drop lock and wait. */
5143 task_rq_unlock(rq, &flags);
5144 wake_up_process(rq->migration_thread);
5145 wait_for_completion(&req.done);
5146 tlb_migrate_finish(p->mm);
5147 return 0;
5148 }
5149out:
5150 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005151
Linus Torvalds1da177e2005-04-16 15:20:36 -07005152 return ret;
5153}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005154EXPORT_SYMBOL_GPL(set_cpus_allowed);
5155
5156/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005157 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07005158 * this because either it can't run here any more (set_cpus_allowed()
5159 * away from this CPU, or CPU going down), or because we're
5160 * attempting to rebalance this task on exec (sched_exec).
5161 *
5162 * So we race with normal scheduler movements, but that's OK, as long
5163 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07005164 *
5165 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005166 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07005167static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005168{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005169 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02005170 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005171
5172 if (unlikely(cpu_is_offline(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07005173 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005174
5175 rq_src = cpu_rq(src_cpu);
5176 rq_dest = cpu_rq(dest_cpu);
5177
5178 double_rq_lock(rq_src, rq_dest);
5179 /* Already moved. */
5180 if (task_cpu(p) != src_cpu)
5181 goto out;
5182 /* Affinity changed (again). */
5183 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5184 goto out;
5185
Ingo Molnardd41f592007-07-09 18:51:59 +02005186 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02005187 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005188 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02005189
Linus Torvalds1da177e2005-04-16 15:20:36 -07005190 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005191 if (on_rq) {
5192 activate_task(rq_dest, p, 0);
5193 check_preempt_curr(rq_dest, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005194 }
Kirill Korotaevefc30812006-06-27 02:54:32 -07005195 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005196out:
5197 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07005198 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005199}
5200
5201/*
5202 * migration_thread - this is a highprio system thread that performs
5203 * thread migration by bumping thread off CPU then 'pushing' onto
5204 * another runqueue.
5205 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005206static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005207{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005208 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005209 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005210
5211 rq = cpu_rq(cpu);
5212 BUG_ON(rq->migration_thread != current);
5213
5214 set_current_state(TASK_INTERRUPTIBLE);
5215 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005216 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005217 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005218
Linus Torvalds1da177e2005-04-16 15:20:36 -07005219 spin_lock_irq(&rq->lock);
5220
5221 if (cpu_is_offline(cpu)) {
5222 spin_unlock_irq(&rq->lock);
5223 goto wait_to_die;
5224 }
5225
5226 if (rq->active_balance) {
5227 active_load_balance(rq, cpu);
5228 rq->active_balance = 0;
5229 }
5230
5231 head = &rq->migration_queue;
5232
5233 if (list_empty(head)) {
5234 spin_unlock_irq(&rq->lock);
5235 schedule();
5236 set_current_state(TASK_INTERRUPTIBLE);
5237 continue;
5238 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07005239 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005240 list_del_init(head->next);
5241
Nick Piggin674311d2005-06-25 14:57:27 -07005242 spin_unlock(&rq->lock);
5243 __migrate_task(req->task, cpu, req->dest_cpu);
5244 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005245
5246 complete(&req->done);
5247 }
5248 __set_current_state(TASK_RUNNING);
5249 return 0;
5250
5251wait_to_die:
5252 /* Wait for kthread_stop */
5253 set_current_state(TASK_INTERRUPTIBLE);
5254 while (!kthread_should_stop()) {
5255 schedule();
5256 set_current_state(TASK_INTERRUPTIBLE);
5257 }
5258 __set_current_state(TASK_RUNNING);
5259 return 0;
5260}
5261
5262#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005263
5264static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5265{
5266 int ret;
5267
5268 local_irq_disable();
5269 ret = __migrate_task(p, src_cpu, dest_cpu);
5270 local_irq_enable();
5271 return ret;
5272}
5273
Kirill Korotaev054b9102006-12-10 02:20:11 -08005274/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02005275 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08005276 * NOTE: interrupts should be disabled by the caller
5277 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005278static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005279{
Kirill Korotaevefc30812006-06-27 02:54:32 -07005280 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005281 cpumask_t mask;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005282 struct rq *rq;
5283 int dest_cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005284
Andi Kleen3a5c3592007-10-15 17:00:14 +02005285 do {
5286 /* On same node? */
5287 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5288 cpus_and(mask, mask, p->cpus_allowed);
5289 dest_cpu = any_online_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005290
Andi Kleen3a5c3592007-10-15 17:00:14 +02005291 /* On any allowed CPU? */
5292 if (dest_cpu == NR_CPUS)
5293 dest_cpu = any_online_cpu(p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005294
Andi Kleen3a5c3592007-10-15 17:00:14 +02005295 /* No more Mr. Nice Guy. */
5296 if (dest_cpu == NR_CPUS) {
Cliff Wickman470fd642007-10-18 23:40:46 -07005297 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5298 /*
5299 * Try to stay on the same cpuset, where the
5300 * current cpuset may be a subset of all cpus.
5301 * The cpuset_cpus_allowed_locked() variant of
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005302 * cpuset_cpus_allowed() will not block. It must be
Cliff Wickman470fd642007-10-18 23:40:46 -07005303 * called within calls to cpuset_lock/cpuset_unlock.
5304 */
Andi Kleen3a5c3592007-10-15 17:00:14 +02005305 rq = task_rq_lock(p, &flags);
Cliff Wickman470fd642007-10-18 23:40:46 -07005306 p->cpus_allowed = cpus_allowed;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005307 dest_cpu = any_online_cpu(p->cpus_allowed);
5308 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005309
Andi Kleen3a5c3592007-10-15 17:00:14 +02005310 /*
5311 * Don't tell them about moving exiting tasks or
5312 * kernel threads (both mm NULL), since they never
5313 * leave kernel.
5314 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005315 if (p->mm && printk_ratelimit()) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02005316 printk(KERN_INFO "process %d (%s) no "
5317 "longer affine to cpu%d\n",
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005318 task_pid_nr(p), p->comm, dead_cpu);
5319 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02005320 }
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005321 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005322}
5323
5324/*
5325 * While a dead CPU has no uninterruptible tasks queued at this point,
5326 * it might still have a nonzero ->nr_uninterruptible counter, because
5327 * for performance reasons the counter is not stricly tracking tasks to
5328 * their home CPUs. So we just add the counter to another CPU's counter,
5329 * to keep the global sum constant after CPU-down:
5330 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07005331static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005332{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005333 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005334 unsigned long flags;
5335
5336 local_irq_save(flags);
5337 double_rq_lock(rq_src, rq_dest);
5338 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5339 rq_src->nr_uninterruptible = 0;
5340 double_rq_unlock(rq_src, rq_dest);
5341 local_irq_restore(flags);
5342}
5343
5344/* Run through task list and migrate tasks from the dead cpu. */
5345static void migrate_live_tasks(int src_cpu)
5346{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005347 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005348
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005349 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005350
Ingo Molnar48f24c42006-07-03 00:25:40 -07005351 do_each_thread(t, p) {
5352 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005353 continue;
5354
Ingo Molnar48f24c42006-07-03 00:25:40 -07005355 if (task_cpu(p) == src_cpu)
5356 move_task_off_dead_cpu(src_cpu, p);
5357 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005358
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005359 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005360}
5361
Ingo Molnardd41f592007-07-09 18:51:59 +02005362/*
5363 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01005364 * It does so by boosting its priority to highest possible.
5365 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005366 */
5367void sched_idle_next(void)
5368{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005369 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07005370 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005371 struct task_struct *p = rq->idle;
5372 unsigned long flags;
5373
5374 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005375 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005376
Ingo Molnar48f24c42006-07-03 00:25:40 -07005377 /*
5378 * Strictly not necessary since rest of the CPUs are stopped by now
5379 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005380 */
5381 spin_lock_irqsave(&rq->lock, flags);
5382
Ingo Molnardd41f592007-07-09 18:51:59 +02005383 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005384
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01005385 update_rq_clock(rq);
5386 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005387
5388 spin_unlock_irqrestore(&rq->lock, flags);
5389}
5390
Ingo Molnar48f24c42006-07-03 00:25:40 -07005391/*
5392 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07005393 * offline.
5394 */
5395void idle_task_exit(void)
5396{
5397 struct mm_struct *mm = current->active_mm;
5398
5399 BUG_ON(cpu_online(smp_processor_id()));
5400
5401 if (mm != &init_mm)
5402 switch_mm(mm, &init_mm, current);
5403 mmdrop(mm);
5404}
5405
Kirill Korotaev054b9102006-12-10 02:20:11 -08005406/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005407static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005408{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005409 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005410
5411 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07005412 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005413
5414 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07005415 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005416
Ingo Molnar48f24c42006-07-03 00:25:40 -07005417 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005418
5419 /*
5420 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005421 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005422 * fine.
5423 */
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005424 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005425 move_task_off_dead_cpu(dead_cpu, p);
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07005426 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005427
Ingo Molnar48f24c42006-07-03 00:25:40 -07005428 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005429}
5430
5431/* release_task() removes task from tasklist, so we won't find dead tasks. */
5432static void migrate_dead_tasks(unsigned int dead_cpu)
5433{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005434 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005435 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005436
Ingo Molnardd41f592007-07-09 18:51:59 +02005437 for ( ; ; ) {
5438 if (!rq->nr_running)
5439 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02005440 update_rq_clock(rq);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02005441 next = pick_next_task(rq, rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02005442 if (!next)
5443 break;
5444 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02005445
Linus Torvalds1da177e2005-04-16 15:20:36 -07005446 }
5447}
5448#endif /* CONFIG_HOTPLUG_CPU */
5449
Nick Piggine692ab52007-07-26 13:40:43 +02005450#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5451
5452static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005453 {
5454 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005455 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005456 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01005457 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02005458};
5459
5460static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005461 {
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005462 .ctl_name = CTL_KERN,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005463 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005464 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005465 .child = sd_ctl_dir,
5466 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01005467 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02005468};
5469
5470static struct ctl_table *sd_alloc_ctl_entry(int n)
5471{
5472 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02005473 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02005474
Nick Piggine692ab52007-07-26 13:40:43 +02005475 return entry;
5476}
5477
Milton Miller6382bc92007-10-15 17:00:19 +02005478static void sd_free_ctl_entry(struct ctl_table **tablep)
5479{
Milton Millercd7900762007-10-17 16:55:11 +02005480 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02005481
Milton Millercd7900762007-10-17 16:55:11 +02005482 /*
5483 * In the intermediate directories, both the child directory and
5484 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005485 * will always be set. In the lowest directory the names are
Milton Millercd7900762007-10-17 16:55:11 +02005486 * static strings and all have proc handlers.
5487 */
5488 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02005489 if (entry->child)
5490 sd_free_ctl_entry(&entry->child);
Milton Millercd7900762007-10-17 16:55:11 +02005491 if (entry->proc_handler == NULL)
5492 kfree(entry->procname);
5493 }
Milton Miller6382bc92007-10-15 17:00:19 +02005494
5495 kfree(*tablep);
5496 *tablep = NULL;
5497}
5498
Nick Piggine692ab52007-07-26 13:40:43 +02005499static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02005500set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02005501 const char *procname, void *data, int maxlen,
5502 mode_t mode, proc_handler *proc_handler)
5503{
Nick Piggine692ab52007-07-26 13:40:43 +02005504 entry->procname = procname;
5505 entry->data = data;
5506 entry->maxlen = maxlen;
5507 entry->mode = mode;
5508 entry->proc_handler = proc_handler;
5509}
5510
5511static struct ctl_table *
5512sd_alloc_ctl_domain_table(struct sched_domain *sd)
5513{
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005514 struct ctl_table *table = sd_alloc_ctl_entry(12);
Nick Piggine692ab52007-07-26 13:40:43 +02005515
Milton Millerad1cdc12007-10-15 17:00:19 +02005516 if (table == NULL)
5517 return NULL;
5518
Alexey Dobriyane0361852007-08-09 11:16:46 +02005519 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02005520 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005521 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02005522 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005523 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005524 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005525 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005526 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005527 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005528 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005529 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005530 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005531 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005532 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005533 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02005534 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005535 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02005536 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005537 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02005538 &sd->cache_nice_tries,
5539 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005540 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02005541 sizeof(int), 0644, proc_dointvec_minmax);
Milton Miller6323469f2007-10-15 17:00:19 +02005542 /* &table[11] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02005543
5544 return table;
5545}
5546
Ingo Molnar9a4e7152007-11-28 15:52:56 +01005547static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02005548{
5549 struct ctl_table *entry, *table;
5550 struct sched_domain *sd;
5551 int domain_num = 0, i;
5552 char buf[32];
5553
5554 for_each_domain(cpu, sd)
5555 domain_num++;
5556 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02005557 if (table == NULL)
5558 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02005559
5560 i = 0;
5561 for_each_domain(cpu, sd) {
5562 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02005563 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005564 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02005565 entry->child = sd_alloc_ctl_domain_table(sd);
5566 entry++;
5567 i++;
5568 }
5569 return table;
5570}
5571
5572static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02005573static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02005574{
5575 int i, cpu_num = num_online_cpus();
5576 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5577 char buf[32];
5578
Milton Miller73785472007-10-24 18:23:48 +02005579 WARN_ON(sd_ctl_dir[0].child);
5580 sd_ctl_dir[0].child = entry;
5581
Milton Millerad1cdc12007-10-15 17:00:19 +02005582 if (entry == NULL)
5583 return;
5584
Milton Miller97b6ea72007-10-15 17:00:19 +02005585 for_each_online_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02005586 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02005587 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005588 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02005589 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02005590 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02005591 }
Milton Miller73785472007-10-24 18:23:48 +02005592
5593 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02005594 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5595}
Milton Miller6382bc92007-10-15 17:00:19 +02005596
Milton Miller73785472007-10-24 18:23:48 +02005597/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02005598static void unregister_sched_domain_sysctl(void)
5599{
Milton Miller73785472007-10-24 18:23:48 +02005600 if (sd_sysctl_header)
5601 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02005602 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02005603 if (sd_ctl_dir[0].child)
5604 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02005605}
Nick Piggine692ab52007-07-26 13:40:43 +02005606#else
Milton Miller6382bc92007-10-15 17:00:19 +02005607static void register_sched_domain_sysctl(void)
5608{
5609}
5610static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02005611{
5612}
5613#endif
5614
Linus Torvalds1da177e2005-04-16 15:20:36 -07005615/*
5616 * migration_call - callback that gets triggered when a CPU is added.
5617 * Here we can start up the necessary migration thread for the new CPU.
5618 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005619static int __cpuinit
5620migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005621{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005622 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005623 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005624 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005625 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005626
5627 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07005628 case CPU_LOCK_ACQUIRE:
5629 mutex_lock(&sched_hotcpu_mutex);
5630 break;
5631
Linus Torvalds1da177e2005-04-16 15:20:36 -07005632 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005633 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02005634 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005635 if (IS_ERR(p))
5636 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005637 kthread_bind(p, cpu);
5638 /* Must be high prio: stop_machine expects to yield to it. */
5639 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02005640 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005641 task_rq_unlock(rq, &flags);
5642 cpu_rq(cpu)->migration_thread = p;
5643 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005644
Linus Torvalds1da177e2005-04-16 15:20:36 -07005645 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005646 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02005647 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005648 wake_up_process(cpu_rq(cpu)->migration_thread);
5649 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005650
Linus Torvalds1da177e2005-04-16 15:20:36 -07005651#ifdef CONFIG_HOTPLUG_CPU
5652 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005653 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07005654 if (!cpu_rq(cpu)->migration_thread)
5655 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005656 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08005657 kthread_bind(cpu_rq(cpu)->migration_thread,
5658 any_online_cpu(cpu_online_map));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005659 kthread_stop(cpu_rq(cpu)->migration_thread);
5660 cpu_rq(cpu)->migration_thread = NULL;
5661 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005662
Linus Torvalds1da177e2005-04-16 15:20:36 -07005663 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005664 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07005665 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005666 migrate_live_tasks(cpu);
5667 rq = cpu_rq(cpu);
5668 kthread_stop(rq->migration_thread);
5669 rq->migration_thread = NULL;
5670 /* Idle task back to normal (off runqueue, low prio) */
Oleg Nesterovd2da2722007-10-16 23:30:56 -07005671 spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02005672 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005673 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005674 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02005675 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5676 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005677 migrate_dead_tasks(cpu);
Oleg Nesterovd2da2722007-10-16 23:30:56 -07005678 spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07005679 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005680 migrate_nr_uninterruptible(rq);
5681 BUG_ON(rq->nr_running != 0);
5682
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005683 /*
5684 * No need to migrate the tasks: it was best-effort if
5685 * they didn't take sched_hotcpu_mutex. Just wake up
5686 * the requestors.
5687 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005688 spin_lock_irq(&rq->lock);
5689 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005690 struct migration_req *req;
5691
Linus Torvalds1da177e2005-04-16 15:20:36 -07005692 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07005693 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005694 list_del_init(&req->list);
5695 complete(&req->done);
5696 }
5697 spin_unlock_irq(&rq->lock);
5698 break;
5699#endif
Gautham R Shenoy5be93612007-05-09 02:34:04 -07005700 case CPU_LOCK_RELEASE:
5701 mutex_unlock(&sched_hotcpu_mutex);
5702 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005703 }
5704 return NOTIFY_OK;
5705}
5706
5707/* Register at highest priority so that task migration (migrate_all_tasks)
5708 * happens before everything else.
5709 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07005710static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005711 .notifier_call = migration_call,
5712 .priority = 10
5713};
5714
Adrian Bunke6fe6642007-11-09 22:39:39 +01005715void __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005716{
5717 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07005718 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005719
5720 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07005721 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5722 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005723 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5724 register_cpu_notifier(&migration_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005725}
5726#endif
5727
5728#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07005729
5730/* Number of possible processor ids */
5731int nr_cpu_ids __read_mostly = NR_CPUS;
5732EXPORT_SYMBOL(nr_cpu_ids);
5733
Ingo Molnar3e9830d2007-10-15 17:00:13 +02005734#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005735
5736static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5737{
5738 struct sched_group *group = sd->groups;
5739 cpumask_t groupmask;
5740 char str[NR_CPUS];
5741
5742 cpumask_scnprintf(str, NR_CPUS, sd->span);
5743 cpus_clear(groupmask);
5744
5745 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5746
5747 if (!(sd->flags & SD_LOAD_BALANCE)) {
5748 printk("does not load-balance\n");
5749 if (sd->parent)
5750 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5751 " has parent");
5752 return -1;
5753 }
5754
5755 printk(KERN_CONT "span %s\n", str);
5756
5757 if (!cpu_isset(cpu, sd->span)) {
5758 printk(KERN_ERR "ERROR: domain->span does not contain "
5759 "CPU%d\n", cpu);
5760 }
5761 if (!cpu_isset(cpu, group->cpumask)) {
5762 printk(KERN_ERR "ERROR: domain->groups does not contain"
5763 " CPU%d\n", cpu);
5764 }
5765
5766 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5767 do {
5768 if (!group) {
5769 printk("\n");
5770 printk(KERN_ERR "ERROR: group is NULL\n");
5771 break;
5772 }
5773
5774 if (!group->__cpu_power) {
5775 printk(KERN_CONT "\n");
5776 printk(KERN_ERR "ERROR: domain->cpu_power not "
5777 "set\n");
5778 break;
5779 }
5780
5781 if (!cpus_weight(group->cpumask)) {
5782 printk(KERN_CONT "\n");
5783 printk(KERN_ERR "ERROR: empty group\n");
5784 break;
5785 }
5786
5787 if (cpus_intersects(groupmask, group->cpumask)) {
5788 printk(KERN_CONT "\n");
5789 printk(KERN_ERR "ERROR: repeated CPUs\n");
5790 break;
5791 }
5792
5793 cpus_or(groupmask, groupmask, group->cpumask);
5794
5795 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5796 printk(KERN_CONT " %s", str);
5797
5798 group = group->next;
5799 } while (group != sd->groups);
5800 printk(KERN_CONT "\n");
5801
5802 if (!cpus_equal(sd->span, groupmask))
5803 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5804
5805 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5806 printk(KERN_ERR "ERROR: parent span is not a superset "
5807 "of domain->span\n");
5808 return 0;
5809}
5810
Linus Torvalds1da177e2005-04-16 15:20:36 -07005811static void sched_domain_debug(struct sched_domain *sd, int cpu)
5812{
5813 int level = 0;
5814
Nick Piggin41c7ce92005-06-25 14:57:24 -07005815 if (!sd) {
5816 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5817 return;
5818 }
5819
Linus Torvalds1da177e2005-04-16 15:20:36 -07005820 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5821
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005822 for (;;) {
5823 if (sched_domain_debug_one(sd, cpu, level))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005824 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005825 level++;
5826 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005827 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005828 break;
5829 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005830}
5831#else
Ingo Molnar48f24c42006-07-03 00:25:40 -07005832# define sched_domain_debug(sd, cpu) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005833#endif
5834
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07005835static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07005836{
5837 if (cpus_weight(sd->span) == 1)
5838 return 1;
5839
5840 /* Following flags need at least 2 groups */
5841 if (sd->flags & (SD_LOAD_BALANCE |
5842 SD_BALANCE_NEWIDLE |
5843 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005844 SD_BALANCE_EXEC |
5845 SD_SHARE_CPUPOWER |
5846 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005847 if (sd->groups != sd->groups->next)
5848 return 0;
5849 }
5850
5851 /* Following flags don't use groups */
5852 if (sd->flags & (SD_WAKE_IDLE |
5853 SD_WAKE_AFFINE |
5854 SD_WAKE_BALANCE))
5855 return 0;
5856
5857 return 1;
5858}
5859
Ingo Molnar48f24c42006-07-03 00:25:40 -07005860static int
5861sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07005862{
5863 unsigned long cflags = sd->flags, pflags = parent->flags;
5864
5865 if (sd_degenerate(parent))
5866 return 1;
5867
5868 if (!cpus_equal(sd->span, parent->span))
5869 return 0;
5870
5871 /* Does parent contain flags not in child? */
5872 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5873 if (cflags & SD_WAKE_AFFINE)
5874 pflags &= ~SD_WAKE_BALANCE;
5875 /* Flags needing groups don't count if only 1 group in parent */
5876 if (parent->groups == parent->groups->next) {
5877 pflags &= ~(SD_LOAD_BALANCE |
5878 SD_BALANCE_NEWIDLE |
5879 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005880 SD_BALANCE_EXEC |
5881 SD_SHARE_CPUPOWER |
5882 SD_SHARE_PKG_RESOURCES);
Suresh Siddha245af2c2005-06-25 14:57:25 -07005883 }
5884 if (~cflags & pflags)
5885 return 0;
5886
5887 return 1;
5888}
5889
Linus Torvalds1da177e2005-04-16 15:20:36 -07005890/*
5891 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5892 * hold the hotplug lock.
5893 */
John Hawkes9c1cfda2005-09-06 15:18:14 -07005894static void cpu_attach_domain(struct sched_domain *sd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005895{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005896 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07005897 struct sched_domain *tmp;
5898
5899 /* Remove the sched domains which do not contribute to scheduling. */
5900 for (tmp = sd; tmp; tmp = tmp->parent) {
5901 struct sched_domain *parent = tmp->parent;
5902 if (!parent)
5903 break;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005904 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005905 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005906 if (parent->parent)
5907 parent->parent->child = tmp;
5908 }
Suresh Siddha245af2c2005-06-25 14:57:25 -07005909 }
5910
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005911 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005912 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005913 if (sd)
5914 sd->child = NULL;
5915 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005916
5917 sched_domain_debug(sd, cpu);
5918
Nick Piggin674311d2005-06-25 14:57:27 -07005919 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005920}
5921
5922/* cpus with isolated domains */
Tim Chen67af63a2006-12-22 01:07:50 -08005923static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005924
5925/* Setup the mask of cpus configured for isolated domains */
5926static int __init isolated_cpu_setup(char *str)
5927{
5928 int ints[NR_CPUS], i;
5929
5930 str = get_options(str, ARRAY_SIZE(ints), ints);
5931 cpus_clear(cpu_isolated_map);
5932 for (i = 1; i <= ints[0]; i++)
5933 if (ints[i] < NR_CPUS)
5934 cpu_set(ints[i], cpu_isolated_map);
5935 return 1;
5936}
5937
Ingo Molnar8927f492007-10-15 17:00:13 +02005938__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005939
5940/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005941 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5942 * to a function which identifies what group(along with sched group) a CPU
5943 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5944 * (due to the fact that we keep track of groups covered with a cpumask_t).
Linus Torvalds1da177e2005-04-16 15:20:36 -07005945 *
5946 * init_sched_build_groups will build a circular linked list of the groups
5947 * covered by the given span, and will set each group's ->cpumask correctly,
5948 * and ->cpu_power to 0.
5949 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005950static void
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005951init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5952 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5953 struct sched_group **sg))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005954{
5955 struct sched_group *first = NULL, *last = NULL;
5956 cpumask_t covered = CPU_MASK_NONE;
5957 int i;
5958
5959 for_each_cpu_mask(i, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005960 struct sched_group *sg;
5961 int group = group_fn(i, cpu_map, &sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005962 int j;
5963
5964 if (cpu_isset(i, covered))
5965 continue;
5966
5967 sg->cpumask = CPU_MASK_NONE;
Eric Dumazet5517d862007-05-08 00:32:57 -07005968 sg->__cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005969
5970 for_each_cpu_mask(j, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005971 if (group_fn(j, cpu_map, NULL) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005972 continue;
5973
5974 cpu_set(j, covered);
5975 cpu_set(j, sg->cpumask);
5976 }
5977 if (!first)
5978 first = sg;
5979 if (last)
5980 last->next = sg;
5981 last = sg;
5982 }
5983 last->next = first;
5984}
5985
John Hawkes9c1cfda2005-09-06 15:18:14 -07005986#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07005987
John Hawkes9c1cfda2005-09-06 15:18:14 -07005988#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08005989
John Hawkes9c1cfda2005-09-06 15:18:14 -07005990/**
5991 * find_next_best_node - find the next node to include in a sched_domain
5992 * @node: node whose sched_domain we're building
5993 * @used_nodes: nodes already in the sched_domain
5994 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005995 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07005996 * finds the closest node not already in the @used_nodes map.
5997 *
5998 * Should use nodemask_t.
5999 */
6000static int find_next_best_node(int node, unsigned long *used_nodes)
6001{
6002 int i, n, val, min_val, best_node = 0;
6003
6004 min_val = INT_MAX;
6005
6006 for (i = 0; i < MAX_NUMNODES; i++) {
6007 /* Start at @node */
6008 n = (node + i) % MAX_NUMNODES;
6009
6010 if (!nr_cpus_node(n))
6011 continue;
6012
6013 /* Skip already used nodes */
6014 if (test_bit(n, used_nodes))
6015 continue;
6016
6017 /* Simple min distance search */
6018 val = node_distance(node, n);
6019
6020 if (val < min_val) {
6021 min_val = val;
6022 best_node = n;
6023 }
6024 }
6025
6026 set_bit(best_node, used_nodes);
6027 return best_node;
6028}
6029
6030/**
6031 * sched_domain_node_span - get a cpumask for a node's sched_domain
6032 * @node: node whose cpumask we're constructing
6033 * @size: number of nodes to include in this span
6034 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006035 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07006036 * should be one that prevents unnecessary balancing, but also spreads tasks
6037 * out optimally.
6038 */
6039static cpumask_t sched_domain_node_span(int node)
6040{
John Hawkes9c1cfda2005-09-06 15:18:14 -07006041 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006042 cpumask_t span, nodemask;
6043 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006044
6045 cpus_clear(span);
6046 bitmap_zero(used_nodes, MAX_NUMNODES);
6047
6048 nodemask = node_to_cpumask(node);
6049 cpus_or(span, span, nodemask);
6050 set_bit(node, used_nodes);
6051
6052 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6053 int next_node = find_next_best_node(node, used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006054
John Hawkes9c1cfda2005-09-06 15:18:14 -07006055 nodemask = node_to_cpumask(next_node);
6056 cpus_or(span, span, nodemask);
6057 }
6058
6059 return span;
6060}
6061#endif
6062
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006063int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006064
John Hawkes9c1cfda2005-09-06 15:18:14 -07006065/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07006066 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07006067 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006068#ifdef CONFIG_SCHED_SMT
6069static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006070static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006071
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006072static int
6073cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006074{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006075 if (sg)
6076 *sg = &per_cpu(sched_group_cpus, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006077 return cpu;
6078}
6079#endif
6080
Ingo Molnar48f24c42006-07-03 00:25:40 -07006081/*
6082 * multi-core sched-domains:
6083 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006084#ifdef CONFIG_SCHED_MC
6085static DEFINE_PER_CPU(struct sched_domain, core_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006086static DEFINE_PER_CPU(struct sched_group, sched_group_core);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006087#endif
6088
6089#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006090static int
6091cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006092{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006093 int group;
Mike Travisd5a74302007-10-16 01:24:05 -07006094 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006095 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006096 group = first_cpu(mask);
6097 if (sg)
6098 *sg = &per_cpu(sched_group_core, group);
6099 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006100}
6101#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006102static int
6103cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006104{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006105 if (sg)
6106 *sg = &per_cpu(sched_group_core, cpu);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006107 return cpu;
6108}
6109#endif
6110
Linus Torvalds1da177e2005-04-16 15:20:36 -07006111static DEFINE_PER_CPU(struct sched_domain, phys_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006112static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006113
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006114static int
6115cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006116{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006117 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006118#ifdef CONFIG_SCHED_MC
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006119 cpumask_t mask = cpu_coregroup_map(cpu);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006120 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006121 group = first_cpu(mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006122#elif defined(CONFIG_SCHED_SMT)
Mike Travisd5a74302007-10-16 01:24:05 -07006123 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006124 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006125 group = first_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006126#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006127 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006128#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006129 if (sg)
6130 *sg = &per_cpu(sched_group_phys, group);
6131 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006132}
6133
6134#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07006135/*
6136 * The init_sched_build_groups can't handle what we want to do with node
6137 * groups, so roll our own. Now each node has its own list of groups which
6138 * gets dynamically allocated.
6139 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006140static DEFINE_PER_CPU(struct sched_domain, node_domains);
John Hawkesd1b55132005-09-06 15:18:14 -07006141static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
John Hawkes9c1cfda2005-09-06 15:18:14 -07006142
6143static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006144static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006145
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006146static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6147 struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006148{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006149 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6150 int group;
6151
6152 cpus_and(nodemask, nodemask, *cpu_map);
6153 group = first_cpu(nodemask);
6154
6155 if (sg)
6156 *sg = &per_cpu(sched_group_allnodes, group);
6157 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006158}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006159
Siddha, Suresh B08069032006-03-27 01:15:23 -08006160static void init_numa_sched_groups_power(struct sched_group *group_head)
6161{
6162 struct sched_group *sg = group_head;
6163 int j;
6164
6165 if (!sg)
6166 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006167 do {
6168 for_each_cpu_mask(j, sg->cpumask) {
6169 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08006170
Andi Kleen3a5c3592007-10-15 17:00:14 +02006171 sd = &per_cpu(phys_domains, j);
6172 if (j != first_cpu(sd->groups->cpumask)) {
6173 /*
6174 * Only add "power" once for each
6175 * physical package.
6176 */
6177 continue;
6178 }
6179
6180 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
Siddha, Suresh B08069032006-03-27 01:15:23 -08006181 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02006182 sg = sg->next;
6183 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08006184}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006185#endif
6186
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006187#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006188/* Free memory allocated for various sched_group structures */
6189static void free_sched_groups(const cpumask_t *cpu_map)
6190{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006191 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006192
6193 for_each_cpu_mask(cpu, *cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006194 struct sched_group **sched_group_nodes
6195 = sched_group_nodes_bycpu[cpu];
6196
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006197 if (!sched_group_nodes)
6198 continue;
6199
6200 for (i = 0; i < MAX_NUMNODES; i++) {
6201 cpumask_t nodemask = node_to_cpumask(i);
6202 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6203
6204 cpus_and(nodemask, nodemask, *cpu_map);
6205 if (cpus_empty(nodemask))
6206 continue;
6207
6208 if (sg == NULL)
6209 continue;
6210 sg = sg->next;
6211next_sg:
6212 oldsg = sg;
6213 sg = sg->next;
6214 kfree(oldsg);
6215 if (oldsg != sched_group_nodes[i])
6216 goto next_sg;
6217 }
6218 kfree(sched_group_nodes);
6219 sched_group_nodes_bycpu[cpu] = NULL;
6220 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006221}
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006222#else
6223static void free_sched_groups(const cpumask_t *cpu_map)
6224{
6225}
6226#endif
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006227
Linus Torvalds1da177e2005-04-16 15:20:36 -07006228/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006229 * Initialize sched groups cpu_power.
6230 *
6231 * cpu_power indicates the capacity of sched group, which is used while
6232 * distributing the load between different sched groups in a sched domain.
6233 * Typically cpu_power for all the groups in a sched domain will be same unless
6234 * there are asymmetries in the topology. If there are asymmetries, group
6235 * having more cpu_power will pickup more load compared to the group having
6236 * less cpu_power.
6237 *
6238 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6239 * the maximum number of tasks a group can handle in the presence of other idle
6240 * or lightly loaded groups in the same sched domain.
6241 */
6242static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6243{
6244 struct sched_domain *child;
6245 struct sched_group *group;
6246
6247 WARN_ON(!sd || !sd->groups);
6248
6249 if (cpu != first_cpu(sd->groups->cpumask))
6250 return;
6251
6252 child = sd->child;
6253
Eric Dumazet5517d862007-05-08 00:32:57 -07006254 sd->groups->__cpu_power = 0;
6255
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006256 /*
6257 * For perf policy, if the groups in child domain share resources
6258 * (for example cores sharing some portions of the cache hierarchy
6259 * or SMT), then set this domain groups cpu_power such that each group
6260 * can handle only one task, when there are other idle groups in the
6261 * same sched domain.
6262 */
6263 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6264 (child->flags &
6265 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
Eric Dumazet5517d862007-05-08 00:32:57 -07006266 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006267 return;
6268 }
6269
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006270 /*
6271 * add cpu_power of each child group to this groups cpu_power
6272 */
6273 group = child->groups;
6274 do {
Eric Dumazet5517d862007-05-08 00:32:57 -07006275 sg_inc_cpu_power(sd->groups, group->__cpu_power);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006276 group = group->next;
6277 } while (group != child->groups);
6278}
6279
6280/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006281 * Build sched domains for a given set of cpus and attach the sched domains
6282 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07006283 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006284static int build_sched_domains(const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006285{
6286 int i;
John Hawkesd1b55132005-09-06 15:18:14 -07006287#ifdef CONFIG_NUMA
6288 struct sched_group **sched_group_nodes = NULL;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006289 int sd_allnodes = 0;
John Hawkesd1b55132005-09-06 15:18:14 -07006290
6291 /*
6292 * Allocate the per-node list of sched groups
6293 */
Milton Miller5cf9f062007-10-15 17:00:19 +02006294 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006295 GFP_KERNEL);
John Hawkesd1b55132005-09-06 15:18:14 -07006296 if (!sched_group_nodes) {
6297 printk(KERN_WARNING "Can not alloc sched group node list\n");
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006298 return -ENOMEM;
John Hawkesd1b55132005-09-06 15:18:14 -07006299 }
6300 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6301#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006302
6303 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006304 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006305 */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006306 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006307 struct sched_domain *sd = NULL, *p;
6308 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6309
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006310 cpus_and(nodemask, nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006311
6312#ifdef CONFIG_NUMA
Ingo Molnardd41f592007-07-09 18:51:59 +02006313 if (cpus_weight(*cpu_map) >
6314 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07006315 sd = &per_cpu(allnodes_domains, i);
6316 *sd = SD_ALLNODES_INIT;
6317 sd->span = *cpu_map;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006318 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006319 p = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006320 sd_allnodes = 1;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006321 } else
6322 p = NULL;
6323
Linus Torvalds1da177e2005-04-16 15:20:36 -07006324 sd = &per_cpu(node_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006325 *sd = SD_NODE_INIT;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006326 sd->span = sched_domain_node_span(cpu_to_node(i));
6327 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006328 if (p)
6329 p->child = sd;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006330 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006331#endif
6332
6333 p = sd;
6334 sd = &per_cpu(phys_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006335 *sd = SD_CPU_INIT;
6336 sd->span = nodemask;
6337 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006338 if (p)
6339 p->child = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006340 cpu_to_phys_group(i, cpu_map, &sd->groups);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006341
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006342#ifdef CONFIG_SCHED_MC
6343 p = sd;
6344 sd = &per_cpu(core_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006345 *sd = SD_MC_INIT;
6346 sd->span = cpu_coregroup_map(i);
6347 cpus_and(sd->span, sd->span, *cpu_map);
6348 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006349 p->child = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006350 cpu_to_core_group(i, cpu_map, &sd->groups);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006351#endif
6352
Linus Torvalds1da177e2005-04-16 15:20:36 -07006353#ifdef CONFIG_SCHED_SMT
6354 p = sd;
6355 sd = &per_cpu(cpu_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006356 *sd = SD_SIBLING_INIT;
Mike Travisd5a74302007-10-16 01:24:05 -07006357 sd->span = per_cpu(cpu_sibling_map, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006358 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006359 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006360 p->child = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006361 cpu_to_cpu_group(i, cpu_map, &sd->groups);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006362#endif
6363 }
6364
6365#ifdef CONFIG_SCHED_SMT
6366 /* Set up CPU (sibling) groups */
John Hawkes9c1cfda2005-09-06 15:18:14 -07006367 for_each_cpu_mask(i, *cpu_map) {
Mike Travisd5a74302007-10-16 01:24:05 -07006368 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006369 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006370 if (i != first_cpu(this_sibling_map))
6371 continue;
6372
Ingo Molnardd41f592007-07-09 18:51:59 +02006373 init_sched_build_groups(this_sibling_map, cpu_map,
6374 &cpu_to_cpu_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006375 }
6376#endif
6377
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006378#ifdef CONFIG_SCHED_MC
6379 /* Set up multi-core groups */
6380 for_each_cpu_mask(i, *cpu_map) {
6381 cpumask_t this_core_map = cpu_coregroup_map(i);
6382 cpus_and(this_core_map, this_core_map, *cpu_map);
6383 if (i != first_cpu(this_core_map))
6384 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02006385 init_sched_build_groups(this_core_map, cpu_map,
6386 &cpu_to_core_group);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006387 }
6388#endif
6389
Linus Torvalds1da177e2005-04-16 15:20:36 -07006390 /* Set up physical groups */
6391 for (i = 0; i < MAX_NUMNODES; i++) {
6392 cpumask_t nodemask = node_to_cpumask(i);
6393
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006394 cpus_and(nodemask, nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006395 if (cpus_empty(nodemask))
6396 continue;
6397
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006398 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006399 }
6400
6401#ifdef CONFIG_NUMA
6402 /* Set up node groups */
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006403 if (sd_allnodes)
Ingo Molnardd41f592007-07-09 18:51:59 +02006404 init_sched_build_groups(*cpu_map, cpu_map,
6405 &cpu_to_allnodes_group);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006406
6407 for (i = 0; i < MAX_NUMNODES; i++) {
6408 /* Set up node groups */
6409 struct sched_group *sg, *prev;
6410 cpumask_t nodemask = node_to_cpumask(i);
6411 cpumask_t domainspan;
6412 cpumask_t covered = CPU_MASK_NONE;
6413 int j;
6414
6415 cpus_and(nodemask, nodemask, *cpu_map);
John Hawkesd1b55132005-09-06 15:18:14 -07006416 if (cpus_empty(nodemask)) {
6417 sched_group_nodes[i] = NULL;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006418 continue;
John Hawkesd1b55132005-09-06 15:18:14 -07006419 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006420
6421 domainspan = sched_domain_node_span(i);
6422 cpus_and(domainspan, domainspan, *cpu_map);
6423
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07006424 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006425 if (!sg) {
6426 printk(KERN_WARNING "Can not alloc domain group for "
6427 "node %d\n", i);
6428 goto error;
6429 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006430 sched_group_nodes[i] = sg;
6431 for_each_cpu_mask(j, nodemask) {
6432 struct sched_domain *sd;
Ingo Molnar9761eea2007-07-09 18:52:00 +02006433
John Hawkes9c1cfda2005-09-06 15:18:14 -07006434 sd = &per_cpu(node_domains, j);
6435 sd->groups = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006436 }
Eric Dumazet5517d862007-05-08 00:32:57 -07006437 sg->__cpu_power = 0;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006438 sg->cpumask = nodemask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006439 sg->next = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006440 cpus_or(covered, covered, nodemask);
6441 prev = sg;
6442
6443 for (j = 0; j < MAX_NUMNODES; j++) {
6444 cpumask_t tmp, notcovered;
6445 int n = (i + j) % MAX_NUMNODES;
6446
6447 cpus_complement(notcovered, covered);
6448 cpus_and(tmp, notcovered, *cpu_map);
6449 cpus_and(tmp, tmp, domainspan);
6450 if (cpus_empty(tmp))
6451 break;
6452
6453 nodemask = node_to_cpumask(n);
6454 cpus_and(tmp, tmp, nodemask);
6455 if (cpus_empty(tmp))
6456 continue;
6457
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07006458 sg = kmalloc_node(sizeof(struct sched_group),
6459 GFP_KERNEL, i);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006460 if (!sg) {
6461 printk(KERN_WARNING
6462 "Can not alloc domain group for node %d\n", j);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006463 goto error;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006464 }
Eric Dumazet5517d862007-05-08 00:32:57 -07006465 sg->__cpu_power = 0;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006466 sg->cpumask = tmp;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006467 sg->next = prev->next;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006468 cpus_or(covered, covered, tmp);
6469 prev->next = sg;
6470 prev = sg;
6471 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006472 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006473#endif
6474
6475 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006476#ifdef CONFIG_SCHED_SMT
6477 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006478 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6479
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006480 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006481 }
6482#endif
6483#ifdef CONFIG_SCHED_MC
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006484 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006485 struct sched_domain *sd = &per_cpu(core_domains, i);
6486
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006487 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006488 }
6489#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006490
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006491 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006492 struct sched_domain *sd = &per_cpu(phys_domains, i);
6493
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006494 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006495 }
6496
John Hawkes9c1cfda2005-09-06 15:18:14 -07006497#ifdef CONFIG_NUMA
Siddha, Suresh B08069032006-03-27 01:15:23 -08006498 for (i = 0; i < MAX_NUMNODES; i++)
6499 init_numa_sched_groups_power(sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006500
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006501 if (sd_allnodes) {
6502 struct sched_group *sg;
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07006503
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006504 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07006505 init_numa_sched_groups_power(sg);
6506 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006507#endif
6508
Linus Torvalds1da177e2005-04-16 15:20:36 -07006509 /* Attach the domains */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006510 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006511 struct sched_domain *sd;
6512#ifdef CONFIG_SCHED_SMT
6513 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006514#elif defined(CONFIG_SCHED_MC)
6515 sd = &per_cpu(core_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006516#else
6517 sd = &per_cpu(phys_domains, i);
6518#endif
6519 cpu_attach_domain(sd, i);
6520 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006521
6522 return 0;
6523
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006524#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006525error:
6526 free_sched_groups(cpu_map);
6527 return -ENOMEM;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006528#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006529}
Paul Jackson029190c2007-10-18 23:40:20 -07006530
6531static cpumask_t *doms_cur; /* current sched domains */
6532static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6533
6534/*
6535 * Special case: If a kmalloc of a doms_cur partition (array of
6536 * cpumask_t) fails, then fallback to a single sched domain,
6537 * as determined by the single cpumask_t fallback_doms.
6538 */
6539static cpumask_t fallback_doms;
6540
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006541/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006542 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07006543 * For now this just excludes isolated cpus, but could be used to
6544 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006545 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006546static int arch_init_sched_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006547{
Milton Miller73785472007-10-24 18:23:48 +02006548 int err;
6549
Paul Jackson029190c2007-10-18 23:40:20 -07006550 ndoms_cur = 1;
6551 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6552 if (!doms_cur)
6553 doms_cur = &fallback_doms;
6554 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
Milton Miller73785472007-10-24 18:23:48 +02006555 err = build_sched_domains(doms_cur);
Milton Miller6382bc92007-10-15 17:00:19 +02006556 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02006557
6558 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006559}
6560
6561static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006562{
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006563 free_sched_groups(cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006564}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006565
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006566/*
6567 * Detach sched domains from a group of cpus specified in cpu_map
6568 * These cpus will now be attached to the NULL domain
6569 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08006570static void detach_destroy_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006571{
6572 int i;
6573
Milton Miller6382bc92007-10-15 17:00:19 +02006574 unregister_sched_domain_sysctl();
6575
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006576 for_each_cpu_mask(i, *cpu_map)
6577 cpu_attach_domain(NULL, i);
6578 synchronize_sched();
6579 arch_destroy_sched_domains(cpu_map);
6580}
6581
Paul Jackson029190c2007-10-18 23:40:20 -07006582/*
6583 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006584 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07006585 * doms_new[] to the current sched domain partitioning, doms_cur[].
6586 * It destroys each deleted domain and builds each new domain.
6587 *
6588 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006589 * The masks don't intersect (don't overlap.) We should setup one
6590 * sched domain for each mask. CPUs not in any of the cpumasks will
6591 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07006592 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6593 * it as it is.
6594 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006595 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6596 * ownership of it and will kfree it when done with it. If the caller
Paul Jackson029190c2007-10-18 23:40:20 -07006597 * failed the kmalloc call, then it can pass in doms_new == NULL,
6598 * and partition_sched_domains() will fallback to the single partition
6599 * 'fallback_doms'.
6600 *
6601 * Call with hotplug lock held
6602 */
6603void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6604{
6605 int i, j;
6606
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01006607 lock_doms_cur();
6608
Milton Miller73785472007-10-24 18:23:48 +02006609 /* always unregister in case we don't destroy any domains */
6610 unregister_sched_domain_sysctl();
6611
Paul Jackson029190c2007-10-18 23:40:20 -07006612 if (doms_new == NULL) {
6613 ndoms_new = 1;
6614 doms_new = &fallback_doms;
6615 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6616 }
6617
6618 /* Destroy deleted domains */
6619 for (i = 0; i < ndoms_cur; i++) {
6620 for (j = 0; j < ndoms_new; j++) {
6621 if (cpus_equal(doms_cur[i], doms_new[j]))
6622 goto match1;
6623 }
6624 /* no match - a current sched domain not in new doms_new[] */
6625 detach_destroy_domains(doms_cur + i);
6626match1:
6627 ;
6628 }
6629
6630 /* Build new domains */
6631 for (i = 0; i < ndoms_new; i++) {
6632 for (j = 0; j < ndoms_cur; j++) {
6633 if (cpus_equal(doms_new[i], doms_cur[j]))
6634 goto match2;
6635 }
6636 /* no match - add a new doms_new */
6637 build_sched_domains(doms_new + i);
6638match2:
6639 ;
6640 }
6641
6642 /* Remember the new sched domains */
6643 if (doms_cur != &fallback_doms)
6644 kfree(doms_cur);
6645 doms_cur = doms_new;
6646 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02006647
6648 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01006649
6650 unlock_doms_cur();
Paul Jackson029190c2007-10-18 23:40:20 -07006651}
6652
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006653#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Adrian Bunk6707de002007-08-12 18:08:19 +02006654static int arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006655{
6656 int err;
6657
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006658 mutex_lock(&sched_hotcpu_mutex);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006659 detach_destroy_domains(&cpu_online_map);
6660 err = arch_init_sched_domains(&cpu_online_map);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006661 mutex_unlock(&sched_hotcpu_mutex);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006662
6663 return err;
6664}
6665
6666static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6667{
6668 int ret;
6669
6670 if (buf[0] != '0' && buf[0] != '1')
6671 return -EINVAL;
6672
6673 if (smt)
6674 sched_smt_power_savings = (buf[0] == '1');
6675 else
6676 sched_mc_power_savings = (buf[0] == '1');
6677
6678 ret = arch_reinit_sched_domains();
6679
6680 return ret ? ret : count;
6681}
6682
Adrian Bunk6707de002007-08-12 18:08:19 +02006683#ifdef CONFIG_SCHED_MC
6684static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6685{
6686 return sprintf(page, "%u\n", sched_mc_power_savings);
6687}
6688static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6689 const char *buf, size_t count)
6690{
6691 return sched_power_savings_store(buf, count, 0);
6692}
6693static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6694 sched_mc_power_savings_store);
6695#endif
6696
6697#ifdef CONFIG_SCHED_SMT
6698static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6699{
6700 return sprintf(page, "%u\n", sched_smt_power_savings);
6701}
6702static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6703 const char *buf, size_t count)
6704{
6705 return sched_power_savings_store(buf, count, 1);
6706}
6707static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6708 sched_smt_power_savings_store);
6709#endif
6710
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006711int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6712{
6713 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006714
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006715#ifdef CONFIG_SCHED_SMT
6716 if (smt_capable())
6717 err = sysfs_create_file(&cls->kset.kobj,
6718 &attr_sched_smt_power_savings.attr);
6719#endif
6720#ifdef CONFIG_SCHED_MC
6721 if (!err && mc_capable())
6722 err = sysfs_create_file(&cls->kset.kobj,
6723 &attr_sched_mc_power_savings.attr);
6724#endif
6725 return err;
6726}
6727#endif
6728
Linus Torvalds1da177e2005-04-16 15:20:36 -07006729/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006730 * Force a reinitialization of the sched domains hierarchy. The domains
Linus Torvalds1da177e2005-04-16 15:20:36 -07006731 * and groups cannot be updated in place without racing with the balancing
Nick Piggin41c7ce92005-06-25 14:57:24 -07006732 * code, so we temporarily attach all running cpus to the NULL domain
Linus Torvalds1da177e2005-04-16 15:20:36 -07006733 * which will prevent rebalancing while the sched domains are recalculated.
6734 */
6735static int update_sched_domains(struct notifier_block *nfb,
6736 unsigned long action, void *hcpu)
6737{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006738 switch (action) {
6739 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006740 case CPU_UP_PREPARE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006741 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006742 case CPU_DOWN_PREPARE_FROZEN:
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006743 detach_destroy_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006744 return NOTIFY_OK;
6745
6746 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006747 case CPU_UP_CANCELED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006748 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006749 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006750 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006751 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006752 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006753 case CPU_DEAD_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006754 /*
6755 * Fall through and re-initialise the domains.
6756 */
6757 break;
6758 default:
6759 return NOTIFY_DONE;
6760 }
6761
6762 /* The hotplug lock is already held by cpu_up/cpu_down */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006763 arch_init_sched_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006764
6765 return NOTIFY_OK;
6766}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006767
6768void __init sched_init_smp(void)
6769{
Nick Piggin5c1e1762006-10-03 01:14:04 -07006770 cpumask_t non_isolated_cpus;
6771
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006772 mutex_lock(&sched_hotcpu_mutex);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006773 arch_init_sched_domains(&cpu_online_map);
Nathan Lynche5e56732007-01-10 23:15:28 -08006774 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
Nick Piggin5c1e1762006-10-03 01:14:04 -07006775 if (cpus_empty(non_isolated_cpus))
6776 cpu_set(smp_processor_id(), non_isolated_cpus);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006777 mutex_unlock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006778 /* XXX: Theoretical race here - CPU may be hotplugged now */
6779 hotcpu_notifier(update_sched_domains, 0);
Nick Piggin5c1e1762006-10-03 01:14:04 -07006780
6781 /* Move init over to a non-isolated CPU */
6782 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6783 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01006784 sched_init_granularity();
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01006785
6786#ifdef CONFIG_FAIR_GROUP_SCHED
6787 if (nr_cpu_ids == 1)
6788 return;
6789
6790 lb_monitor_task = kthread_create(load_balance_monitor, NULL,
6791 "group_balance");
6792 if (!IS_ERR(lb_monitor_task)) {
6793 lb_monitor_task->flags |= PF_NOFREEZE;
6794 wake_up_process(lb_monitor_task);
6795 } else {
6796 printk(KERN_ERR "Could not create load balance monitor thread"
6797 "(error = %ld) \n", PTR_ERR(lb_monitor_task));
6798 }
6799#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006800}
6801#else
6802void __init sched_init_smp(void)
6803{
Ingo Molnar19978ca2007-11-09 22:39:38 +01006804 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006805}
6806#endif /* CONFIG_SMP */
6807
6808int in_sched_functions(unsigned long addr)
6809{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006810 return in_lock_functions(addr) ||
6811 (addr >= (unsigned long)__sched_text_start
6812 && addr < (unsigned long)__sched_text_end);
6813}
6814
Alexey Dobriyana9957442007-10-15 17:00:13 +02006815static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02006816{
6817 cfs_rq->tasks_timeline = RB_ROOT;
Ingo Molnardd41f592007-07-09 18:51:59 +02006818#ifdef CONFIG_FAIR_GROUP_SCHED
6819 cfs_rq->rq = rq;
6820#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02006821 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02006822}
6823
Linus Torvalds1da177e2005-04-16 15:20:36 -07006824void __init sched_init(void)
6825{
Christoph Lameter476f3532007-05-06 14:48:58 -07006826 int highest_cpu = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006827 int i, j;
6828
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08006829 for_each_possible_cpu(i) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006830 struct rt_prio_array *array;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006831 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006832
6833 rq = cpu_rq(i);
6834 spin_lock_init(&rq->lock);
Ingo Molnarfcb99372006-07-03 00:25:10 -07006835 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
Nick Piggin78979862005-06-25 14:57:13 -07006836 rq->nr_running = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006837 rq->clock = 1;
6838 init_cfs_rq(&rq->cfs, rq);
6839#ifdef CONFIG_FAIR_GROUP_SCHED
6840 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Ingo Molnar3a252012007-10-15 17:00:12 +02006841 {
6842 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6843 struct sched_entity *se =
6844 &per_cpu(init_sched_entity, i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02006845
Ingo Molnar3a252012007-10-15 17:00:12 +02006846 init_cfs_rq_p[i] = cfs_rq;
6847 init_cfs_rq(cfs_rq, rq);
Ingo Molnar4cf86d72007-10-15 17:00:14 +02006848 cfs_rq->tg = &init_task_group;
Ingo Molnar3a252012007-10-15 17:00:12 +02006849 list_add(&cfs_rq->leaf_cfs_rq_list,
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02006850 &rq->leaf_cfs_rq_list);
6851
Ingo Molnar3a252012007-10-15 17:00:12 +02006852 init_sched_entity_p[i] = se;
6853 se->cfs_rq = &rq->cfs;
6854 se->my_q = cfs_rq;
Ingo Molnar4cf86d72007-10-15 17:00:14 +02006855 se->load.weight = init_task_group_load;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02006856 se->load.inv_weight =
Ingo Molnar4cf86d72007-10-15 17:00:14 +02006857 div64_64(1ULL<<32, init_task_group_load);
Ingo Molnar3a252012007-10-15 17:00:12 +02006858 se->parent = NULL;
6859 }
Ingo Molnar4cf86d72007-10-15 17:00:14 +02006860 init_task_group.shares = init_task_group_load;
Ingo Molnardd41f592007-07-09 18:51:59 +02006861#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006862
Ingo Molnardd41f592007-07-09 18:51:59 +02006863 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6864 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006865#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07006866 rq->sd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006867 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006868 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006869 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07006870 rq->cpu = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006871 rq->migration_thread = NULL;
6872 INIT_LIST_HEAD(&rq->migration_queue);
6873#endif
6874 atomic_set(&rq->nr_iowait, 0);
6875
Ingo Molnardd41f592007-07-09 18:51:59 +02006876 array = &rq->rt.active;
6877 for (j = 0; j < MAX_RT_PRIO; j++) {
6878 INIT_LIST_HEAD(array->queue + j);
6879 __clear_bit(j, array->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006880 }
Christoph Lameter476f3532007-05-06 14:48:58 -07006881 highest_cpu = i;
Ingo Molnardd41f592007-07-09 18:51:59 +02006882 /* delimiter for bitsearch: */
6883 __set_bit(MAX_RT_PRIO, array->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006884 }
6885
Peter Williams2dd73a42006-06-27 02:54:34 -07006886 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07006887
Avi Kivitye107be32007-07-26 13:40:43 +02006888#ifdef CONFIG_PREEMPT_NOTIFIERS
6889 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6890#endif
6891
Christoph Lameterc9819f42006-12-10 02:20:25 -08006892#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07006893 nr_cpu_ids = highest_cpu + 1;
Christoph Lameterc9819f42006-12-10 02:20:25 -08006894 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6895#endif
6896
Heiko Carstensb50f60c2006-07-30 03:03:52 -07006897#ifdef CONFIG_RT_MUTEXES
6898 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6899#endif
6900
Linus Torvalds1da177e2005-04-16 15:20:36 -07006901 /*
6902 * The boot idle thread does lazy MMU switching as well:
6903 */
6904 atomic_inc(&init_mm.mm_count);
6905 enter_lazy_tlb(&init_mm, current);
6906
6907 /*
6908 * Make us the idle thread. Technically, schedule() should not be
6909 * called from this thread, however somewhere below it might be,
6910 * but because we are the idle thread, we just pick up running again
6911 * when this runqueue becomes "idle".
6912 */
6913 init_idle(current, smp_processor_id());
Ingo Molnardd41f592007-07-09 18:51:59 +02006914 /*
6915 * During early bootup we pretend to be a normal task:
6916 */
6917 current->sched_class = &fair_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006918}
6919
6920#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6921void __might_sleep(char *file, int line)
6922{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006923#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07006924 static unsigned long prev_jiffy; /* ratelimiting */
6925
6926 if ((in_atomic() || irqs_disabled()) &&
6927 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6928 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6929 return;
6930 prev_jiffy = jiffies;
Ingo Molnar91368d72006-03-23 03:00:54 -08006931 printk(KERN_ERR "BUG: sleeping function called from invalid"
Linus Torvalds1da177e2005-04-16 15:20:36 -07006932 " context at %s:%d\n", file, line);
6933 printk("in_atomic():%d, irqs_disabled():%d\n",
6934 in_atomic(), irqs_disabled());
Peter Zijlstraa4c410f2006-12-06 20:37:21 -08006935 debug_show_held_locks(current);
Ingo Molnar3117df02006-12-13 00:34:43 -08006936 if (irqs_disabled())
6937 print_irqtrace_events(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006938 dump_stack();
6939 }
6940#endif
6941}
6942EXPORT_SYMBOL(__might_sleep);
6943#endif
6944
6945#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02006946static void normalize_task(struct rq *rq, struct task_struct *p)
6947{
6948 int on_rq;
6949 update_rq_clock(rq);
6950 on_rq = p->se.on_rq;
6951 if (on_rq)
6952 deactivate_task(rq, p, 0);
6953 __setscheduler(rq, p, SCHED_NORMAL, 0);
6954 if (on_rq) {
6955 activate_task(rq, p, 0);
6956 resched_task(rq->curr);
6957 }
6958}
6959
Linus Torvalds1da177e2005-04-16 15:20:36 -07006960void normalize_rt_tasks(void)
6961{
Ingo Molnara0f98a12007-06-17 18:37:45 +02006962 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006963 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006964 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006965
6966 read_lock_irq(&tasklist_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02006967 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02006968 /*
6969 * Only normalize user tasks:
6970 */
6971 if (!p->mm)
6972 continue;
6973
Ingo Molnardd41f592007-07-09 18:51:59 +02006974 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02006975#ifdef CONFIG_SCHEDSTATS
6976 p->se.wait_start = 0;
6977 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006978 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02006979#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02006980 task_rq(p)->clock = 0;
6981
6982 if (!rt_task(p)) {
6983 /*
6984 * Renice negative nice level userspace
6985 * tasks back to 0:
6986 */
6987 if (TASK_NICE(p) < 0 && p->mm)
6988 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006989 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02006990 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006991
Ingo Molnarb29739f2006-06-27 02:54:51 -07006992 spin_lock_irqsave(&p->pi_lock, flags);
6993 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006994
Ingo Molnar178be792007-10-15 17:00:18 +02006995 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02006996
Ingo Molnarb29739f2006-06-27 02:54:51 -07006997 __task_rq_unlock(rq);
6998 spin_unlock_irqrestore(&p->pi_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02006999 } while_each_thread(g, p);
7000
Linus Torvalds1da177e2005-04-16 15:20:36 -07007001 read_unlock_irq(&tasklist_lock);
7002}
7003
7004#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07007005
7006#ifdef CONFIG_IA64
7007/*
7008 * These functions are only useful for the IA64 MCA handling.
7009 *
7010 * They can only be called when the whole system has been
7011 * stopped - every CPU needs to be quiescent, and no scheduling
7012 * activity can take place. Using them for anything else would
7013 * be a serious bug, and as a result, they aren't even visible
7014 * under any other configuration.
7015 */
7016
7017/**
7018 * curr_task - return the current task for a given cpu.
7019 * @cpu: the processor in question.
7020 *
7021 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7022 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007023struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07007024{
7025 return cpu_curr(cpu);
7026}
7027
7028/**
7029 * set_curr_task - set the current task for a given cpu.
7030 * @cpu: the processor in question.
7031 * @p: the task pointer to set.
7032 *
7033 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007034 * are serviced on a separate stack. It allows the architecture to switch the
7035 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07007036 * must be called with all CPU's synchronized, and interrupts disabled, the
7037 * and caller must save the original value of the current task (see
7038 * curr_task() above) and restore that value before reenabling interrupts and
7039 * re-starting the system.
7040 *
7041 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7042 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007043void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07007044{
7045 cpu_curr(cpu) = p;
7046}
7047
7048#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007049
7050#ifdef CONFIG_FAIR_GROUP_SCHED
7051
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007052#ifdef CONFIG_SMP
7053/*
7054 * distribute shares of all task groups among their schedulable entities,
7055 * to reflect load distrbution across cpus.
7056 */
7057static int rebalance_shares(struct sched_domain *sd, int this_cpu)
7058{
7059 struct cfs_rq *cfs_rq;
7060 struct rq *rq = cpu_rq(this_cpu);
7061 cpumask_t sdspan = sd->span;
7062 int balanced = 1;
7063
7064 /* Walk thr' all the task groups that we have */
7065 for_each_leaf_cfs_rq(rq, cfs_rq) {
7066 int i;
7067 unsigned long total_load = 0, total_shares;
7068 struct task_group *tg = cfs_rq->tg;
7069
7070 /* Gather total task load of this group across cpus */
7071 for_each_cpu_mask(i, sdspan)
7072 total_load += tg->cfs_rq[i]->load.weight;
7073
7074 /* Nothing to do if this group has no load */
7075 if (!total_load)
7076 continue;
7077
7078 /*
7079 * tg->shares represents the number of cpu shares the task group
7080 * is eligible to hold on a single cpu. On N cpus, it is
7081 * eligible to hold (N * tg->shares) number of cpu shares.
7082 */
7083 total_shares = tg->shares * cpus_weight(sdspan);
7084
7085 /*
7086 * redistribute total_shares across cpus as per the task load
7087 * distribution.
7088 */
7089 for_each_cpu_mask(i, sdspan) {
7090 unsigned long local_load, local_shares;
7091
7092 local_load = tg->cfs_rq[i]->load.weight;
7093 local_shares = (local_load * total_shares) / total_load;
7094 if (!local_shares)
7095 local_shares = MIN_GROUP_SHARES;
7096 if (local_shares == tg->se[i]->load.weight)
7097 continue;
7098
7099 spin_lock_irq(&cpu_rq(i)->lock);
7100 set_se_shares(tg->se[i], local_shares);
7101 spin_unlock_irq(&cpu_rq(i)->lock);
7102 balanced = 0;
7103 }
7104 }
7105
7106 return balanced;
7107}
7108
7109/*
7110 * How frequently should we rebalance_shares() across cpus?
7111 *
7112 * The more frequently we rebalance shares, the more accurate is the fairness
7113 * of cpu bandwidth distribution between task groups. However higher frequency
7114 * also implies increased scheduling overhead.
7115 *
7116 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7117 * consecutive calls to rebalance_shares() in the same sched domain.
7118 *
7119 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7120 * consecutive calls to rebalance_shares() in the same sched domain.
7121 *
7122 * These settings allows for the appropriate tradeoff between accuracy of
7123 * fairness and the associated overhead.
7124 *
7125 */
7126
7127/* default: 8ms, units: milliseconds */
7128const_debug unsigned int sysctl_sched_min_bal_int_shares = 8;
7129
7130/* default: 128ms, units: milliseconds */
7131const_debug unsigned int sysctl_sched_max_bal_int_shares = 128;
7132
7133/* kernel thread that runs rebalance_shares() periodically */
7134static int load_balance_monitor(void *unused)
7135{
7136 unsigned int timeout = sysctl_sched_min_bal_int_shares;
7137 struct sched_param schedparm;
7138 int ret;
7139
7140 /*
7141 * We don't want this thread's execution to be limited by the shares
7142 * assigned to default group (init_task_group). Hence make it run
7143 * as a SCHED_RR RT task at the lowest priority.
7144 */
7145 schedparm.sched_priority = 1;
7146 ret = sched_setscheduler(current, SCHED_RR, &schedparm);
7147 if (ret)
7148 printk(KERN_ERR "Couldn't set SCHED_RR policy for load balance"
7149 " monitor thread (error = %d) \n", ret);
7150
7151 while (!kthread_should_stop()) {
7152 int i, cpu, balanced = 1;
7153
7154 /* Prevent cpus going down or coming up */
7155 lock_cpu_hotplug();
7156 /* lockout changes to doms_cur[] array */
7157 lock_doms_cur();
7158 /*
7159 * Enter a rcu read-side critical section to safely walk rq->sd
7160 * chain on various cpus and to walk task group list
7161 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7162 */
7163 rcu_read_lock();
7164
7165 for (i = 0; i < ndoms_cur; i++) {
7166 cpumask_t cpumap = doms_cur[i];
7167 struct sched_domain *sd = NULL, *sd_prev = NULL;
7168
7169 cpu = first_cpu(cpumap);
7170
7171 /* Find the highest domain at which to balance shares */
7172 for_each_domain(cpu, sd) {
7173 if (!(sd->flags & SD_LOAD_BALANCE))
7174 continue;
7175 sd_prev = sd;
7176 }
7177
7178 sd = sd_prev;
7179 /* sd == NULL? No load balance reqd in this domain */
7180 if (!sd)
7181 continue;
7182
7183 balanced &= rebalance_shares(sd, cpu);
7184 }
7185
7186 rcu_read_unlock();
7187
7188 unlock_doms_cur();
7189 unlock_cpu_hotplug();
7190
7191 if (!balanced)
7192 timeout = sysctl_sched_min_bal_int_shares;
7193 else if (timeout < sysctl_sched_max_bal_int_shares)
7194 timeout *= 2;
7195
7196 msleep_interruptible(timeout);
7197 }
7198
7199 return 0;
7200}
7201#endif /* CONFIG_SMP */
7202
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007203/* allocate runqueue etc for a new task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02007204struct task_group *sched_create_group(void)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007205{
Ingo Molnar4cf86d72007-10-15 17:00:14 +02007206 struct task_group *tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007207 struct cfs_rq *cfs_rq;
7208 struct sched_entity *se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007209 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007210 int i;
7211
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007212 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7213 if (!tg)
7214 return ERR_PTR(-ENOMEM);
7215
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007216 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007217 if (!tg->cfs_rq)
7218 goto err;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007219 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007220 if (!tg->se)
7221 goto err;
7222
7223 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007224 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007225
7226 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7227 cpu_to_node(i));
7228 if (!cfs_rq)
7229 goto err;
7230
7231 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7232 cpu_to_node(i));
7233 if (!se)
7234 goto err;
7235
7236 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7237 memset(se, 0, sizeof(struct sched_entity));
7238
7239 tg->cfs_rq[i] = cfs_rq;
7240 init_cfs_rq(cfs_rq, rq);
7241 cfs_rq->tg = tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007242
7243 tg->se[i] = se;
7244 se->cfs_rq = &rq->cfs;
7245 se->my_q = cfs_rq;
7246 se->load.weight = NICE_0_LOAD;
7247 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7248 se->parent = NULL;
7249 }
7250
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007251 tg->shares = NICE_0_LOAD;
7252
7253 lock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007254 for_each_possible_cpu(i) {
7255 rq = cpu_rq(i);
7256 cfs_rq = tg->cfs_rq[i];
7257 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7258 }
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007259 unlock_task_group_list();
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007260
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007261 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007262
7263err:
7264 for_each_possible_cpu(i) {
Ingo Molnara65914b2007-10-15 17:00:13 +02007265 if (tg->cfs_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007266 kfree(tg->cfs_rq[i]);
Ingo Molnara65914b2007-10-15 17:00:13 +02007267 if (tg->se)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007268 kfree(tg->se[i]);
7269 }
Ingo Molnara65914b2007-10-15 17:00:13 +02007270 kfree(tg->cfs_rq);
7271 kfree(tg->se);
7272 kfree(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007273
7274 return ERR_PTR(-ENOMEM);
7275}
7276
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007277/* rcu callback to free various structures associated with a task group */
7278static void free_sched_group(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007279{
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +01007280 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7281 struct cfs_rq *cfs_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007282 struct sched_entity *se;
7283 int i;
7284
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007285 /* now it should be safe to free those cfs_rqs */
7286 for_each_possible_cpu(i) {
7287 cfs_rq = tg->cfs_rq[i];
7288 kfree(cfs_rq);
7289
7290 se = tg->se[i];
7291 kfree(se);
7292 }
7293
7294 kfree(tg->cfs_rq);
7295 kfree(tg->se);
7296 kfree(tg);
7297}
7298
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007299/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02007300void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007301{
James Bottomley7bae49d2007-10-29 21:18:11 +01007302 struct cfs_rq *cfs_rq = NULL;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007303 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007304
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007305 lock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007306 for_each_possible_cpu(i) {
7307 cfs_rq = tg->cfs_rq[i];
7308 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7309 }
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007310 unlock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007311
James Bottomley7bae49d2007-10-29 21:18:11 +01007312 BUG_ON(!cfs_rq);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007313
7314 /* wait for possible concurrent references to cfs_rqs complete */
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +01007315 call_rcu(&tg->rcu, free_sched_group);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007316}
7317
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007318/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +02007319 * The caller of this function should have put the task in its new group
7320 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7321 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007322 */
7323void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007324{
7325 int on_rq, running;
7326 unsigned long flags;
7327 struct rq *rq;
7328
7329 rq = task_rq_lock(tsk, &flags);
7330
Oleg Nesterovdae51f52007-11-15 20:57:40 +01007331 if (tsk->sched_class != &fair_sched_class) {
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01007332 set_task_cfs_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007333 goto done;
Oleg Nesterovdae51f52007-11-15 20:57:40 +01007334 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007335
7336 update_rq_clock(rq);
7337
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01007338 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007339 on_rq = tsk->se.on_rq;
7340
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02007341 if (on_rq) {
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007342 dequeue_task(rq, tsk, 0);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02007343 if (unlikely(running))
7344 tsk->sched_class->put_prev_task(rq, tsk);
7345 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007346
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01007347 set_task_cfs_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007348
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02007349 if (on_rq) {
7350 if (unlikely(running))
7351 tsk->sched_class->set_curr_task(rq);
Dmitry Adamushko7074bad2007-10-15 17:00:07 +02007352 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02007353 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007354
7355done:
7356 task_rq_unlock(rq, &flags);
7357}
7358
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007359/* rq->lock to be locked by caller */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007360static void set_se_shares(struct sched_entity *se, unsigned long shares)
7361{
7362 struct cfs_rq *cfs_rq = se->cfs_rq;
7363 struct rq *rq = cfs_rq->rq;
7364 int on_rq;
7365
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007366 if (!shares)
7367 shares = MIN_GROUP_SHARES;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007368
7369 on_rq = se->on_rq;
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007370 if (on_rq) {
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007371 dequeue_entity(cfs_rq, se, 0);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007372 dec_cpu_load(rq, se->load.weight);
7373 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007374
7375 se->load.weight = shares;
7376 se->load.inv_weight = div64_64((1ULL<<32), shares);
7377
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007378 if (on_rq) {
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007379 enqueue_entity(cfs_rq, se, 0);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007380 inc_cpu_load(rq, se->load.weight);
7381 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007382}
7383
Ingo Molnar4cf86d72007-10-15 17:00:14 +02007384int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007385{
7386 int i;
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007387 struct cfs_rq *cfs_rq;
7388 struct rq *rq;
Ingo Molnarc61935f2008-01-22 11:24:58 +01007389
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007390 lock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007391 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +02007392 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007393
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007394 if (shares < MIN_GROUP_SHARES)
7395 shares = MIN_GROUP_SHARES;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007396
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01007397 /*
7398 * Prevent any load balance activity (rebalance_shares,
7399 * load_balance_fair) from referring to this group first,
7400 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7401 */
7402 for_each_possible_cpu(i) {
7403 cfs_rq = tg->cfs_rq[i];
7404 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7405 }
7406
7407 /* wait for any ongoing reference to this group to finish */
7408 synchronize_sched();
7409
7410 /*
7411 * Now we are free to modify the group's share on each cpu
7412 * w/o tripping rebalance_share or load_balance_fair.
7413 */
7414 tg->shares = shares;
7415 for_each_possible_cpu(i) {
7416 spin_lock_irq(&cpu_rq(i)->lock);
7417 set_se_shares(tg->se[i], shares);
7418 spin_unlock_irq(&cpu_rq(i)->lock);
7419 }
7420
7421 /*
7422 * Enable load balance activity on this group, by inserting it back on
7423 * each cpu's rq->leaf_cfs_rq_list.
7424 */
7425 for_each_possible_cpu(i) {
7426 rq = cpu_rq(i);
7427 cfs_rq = tg->cfs_rq[i];
7428 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7429 }
Dhaval Giani5cb350b2007-10-15 17:00:14 +02007430done:
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +01007431 unlock_task_group_list();
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007432 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007433}
7434
Dhaval Giani5cb350b2007-10-15 17:00:14 +02007435unsigned long sched_group_shares(struct task_group *tg)
7436{
7437 return tg->shares;
7438}
7439
Ingo Molnar3a252012007-10-15 17:00:12 +02007440#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007441
7442#ifdef CONFIG_FAIR_CGROUP_SCHED
7443
7444/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02007445static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007446{
Paul Menage2b01dfe2007-10-24 18:23:50 +02007447 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7448 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007449}
7450
7451static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +02007452cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007453{
7454 struct task_group *tg;
7455
Paul Menage2b01dfe2007-10-24 18:23:50 +02007456 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007457 /* This is early initialization for the top cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02007458 init_task_group.css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007459 return &init_task_group.css;
7460 }
7461
7462 /* we support only 1-level deep hierarchical scheduler atm */
Paul Menage2b01dfe2007-10-24 18:23:50 +02007463 if (cgrp->parent->parent)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007464 return ERR_PTR(-EINVAL);
7465
7466 tg = sched_create_group();
7467 if (IS_ERR(tg))
7468 return ERR_PTR(-ENOMEM);
7469
7470 /* Bind the cgroup to task_group object we just created */
Paul Menage2b01dfe2007-10-24 18:23:50 +02007471 tg->css.cgroup = cgrp;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007472
7473 return &tg->css;
7474}
7475
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007476static void
7477cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007478{
Paul Menage2b01dfe2007-10-24 18:23:50 +02007479 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007480
7481 sched_destroy_group(tg);
7482}
7483
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007484static int
7485cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7486 struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007487{
7488 /* We don't support RT-tasks being in separate groups */
7489 if (tsk->sched_class != &fair_sched_class)
7490 return -EINVAL;
7491
7492 return 0;
7493}
7494
7495static void
Paul Menage2b01dfe2007-10-24 18:23:50 +02007496cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007497 struct cgroup *old_cont, struct task_struct *tsk)
7498{
7499 sched_move_task(tsk);
7500}
7501
Paul Menage2b01dfe2007-10-24 18:23:50 +02007502static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7503 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007504{
Paul Menage2b01dfe2007-10-24 18:23:50 +02007505 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007506}
7507
Paul Menage2b01dfe2007-10-24 18:23:50 +02007508static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007509{
Paul Menage2b01dfe2007-10-24 18:23:50 +02007510 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007511
7512 return (u64) tg->shares;
7513}
7514
Paul Menagefe5c7cc2007-10-29 21:18:11 +01007515static struct cftype cpu_files[] = {
7516 {
7517 .name = "shares",
7518 .read_uint = cpu_shares_read_uint,
7519 .write_uint = cpu_shares_write_uint,
7520 },
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007521};
7522
7523static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7524{
Paul Menagefe5c7cc2007-10-29 21:18:11 +01007525 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007526}
7527
7528struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +01007529 .name = "cpu",
7530 .create = cpu_cgroup_create,
7531 .destroy = cpu_cgroup_destroy,
7532 .can_attach = cpu_cgroup_can_attach,
7533 .attach = cpu_cgroup_attach,
7534 .populate = cpu_cgroup_populate,
7535 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07007536 .early_init = 1,
7537};
7538
7539#endif /* CONFIG_FAIR_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01007540
7541#ifdef CONFIG_CGROUP_CPUACCT
7542
7543/*
7544 * CPU accounting code for task groups.
7545 *
7546 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7547 * (balbir@in.ibm.com).
7548 */
7549
7550/* track cpu usage of a group of tasks */
7551struct cpuacct {
7552 struct cgroup_subsys_state css;
7553 /* cpuusage holds pointer to a u64-type object on every cpu */
7554 u64 *cpuusage;
7555};
7556
7557struct cgroup_subsys cpuacct_subsys;
7558
7559/* return cpu accounting group corresponding to this container */
7560static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7561{
7562 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7563 struct cpuacct, css);
7564}
7565
7566/* return cpu accounting group to which this task belongs */
7567static inline struct cpuacct *task_ca(struct task_struct *tsk)
7568{
7569 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7570 struct cpuacct, css);
7571}
7572
7573/* create a new cpu accounting group */
7574static struct cgroup_subsys_state *cpuacct_create(
7575 struct cgroup_subsys *ss, struct cgroup *cont)
7576{
7577 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7578
7579 if (!ca)
7580 return ERR_PTR(-ENOMEM);
7581
7582 ca->cpuusage = alloc_percpu(u64);
7583 if (!ca->cpuusage) {
7584 kfree(ca);
7585 return ERR_PTR(-ENOMEM);
7586 }
7587
7588 return &ca->css;
7589}
7590
7591/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007592static void
7593cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01007594{
7595 struct cpuacct *ca = cgroup_ca(cont);
7596
7597 free_percpu(ca->cpuusage);
7598 kfree(ca);
7599}
7600
7601/* return total cpu usage (in nanoseconds) of a group */
7602static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7603{
7604 struct cpuacct *ca = cgroup_ca(cont);
7605 u64 totalcpuusage = 0;
7606 int i;
7607
7608 for_each_possible_cpu(i) {
7609 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
7610
7611 /*
7612 * Take rq->lock to make 64-bit addition safe on 32-bit
7613 * platforms.
7614 */
7615 spin_lock_irq(&cpu_rq(i)->lock);
7616 totalcpuusage += *cpuusage;
7617 spin_unlock_irq(&cpu_rq(i)->lock);
7618 }
7619
7620 return totalcpuusage;
7621}
7622
7623static struct cftype files[] = {
7624 {
7625 .name = "usage",
7626 .read_uint = cpuusage_read,
7627 },
7628};
7629
7630static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7631{
7632 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
7633}
7634
7635/*
7636 * charge this task's execution time to its accounting group.
7637 *
7638 * called with rq->lock held.
7639 */
7640static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
7641{
7642 struct cpuacct *ca;
7643
7644 if (!cpuacct_subsys.active)
7645 return;
7646
7647 ca = task_ca(tsk);
7648 if (ca) {
7649 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
7650
7651 *cpuusage += cputime;
7652 }
7653}
7654
7655struct cgroup_subsys cpuacct_subsys = {
7656 .name = "cpuacct",
7657 .create = cpuacct_create,
7658 .destroy = cpuacct_destroy,
7659 .populate = cpuacct_populate,
7660 .subsys_id = cpuacct_subsys_id,
7661};
7662#endif /* CONFIG_CGROUP_CPUACCT */