blob: c02659f1bd095b5145575bb7cc33c77140906d01 [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>
47#include <linux/smp.h>
48#include <linux/threads.h>
49#include <linux/timer.h>
50#include <linux/rcupdate.h>
51#include <linux/cpu.h>
52#include <linux/cpuset.h>
53#include <linux/percpu.h>
54#include <linux/kthread.h>
55#include <linux/seq_file.h>
Nick Piggine692ab52007-07-26 13:40:43 +020056#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070057#include <linux/syscalls.h>
58#include <linux/times.h>
Jay Lan8f0ab512006-09-30 23:28:59 -070059#include <linux/tsacct_kern.h>
bibo maoc6fd91f2006-03-26 01:38:20 -080060#include <linux/kprobes.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070061#include <linux/delayacct.h>
Eric Dumazet5517d862007-05-08 00:32:57 -070062#include <linux/reciprocal_div.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020063#include <linux/unistd.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
Eric Dumazet5517d862007-05-08 00:32:57 -070065#include <asm/tlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070066
67/*
Alexey Dobriyanb035b6d2007-02-10 01:45:10 -080068 * Scheduler clock - returns current time in nanosec units.
69 * This is default implementation.
70 * Architectures and sub-architectures can override this.
71 */
72unsigned long long __attribute__((weak)) sched_clock(void)
73{
74 return (unsigned long long)jiffies * (1000000000 / HZ);
75}
76
77/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 * Convert user-nice values [ -20 ... 0 ... 19 ]
79 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
80 * and back.
81 */
82#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
83#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
84#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
85
86/*
87 * 'User priority' is the nice value converted to something we
88 * can work with better when scaling various scheduler parameters,
89 * it's a [ 0 ... 39 ] range.
90 */
91#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
92#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
93#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
94
95/*
96 * Some helpers for converting nanosecond timing to jiffy resolution
97 */
98#define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
99#define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
100
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200101#define NICE_0_LOAD SCHED_LOAD_SCALE
102#define NICE_0_SHIFT SCHED_LOAD_SHIFT
103
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104/*
105 * These are the 'tuning knobs' of the scheduler:
106 *
107 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
108 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
109 * Timeslices get refilled after they expire.
110 */
111#define MIN_TIMESLICE max(5 * HZ / 1000, 1)
112#define DEF_TIMESLICE (100 * HZ / 1000)
Peter Williams2dd73a42006-06-27 02:54:34 -0700113
Eric Dumazet5517d862007-05-08 00:32:57 -0700114#ifdef CONFIG_SMP
115/*
116 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
117 * Since cpu_power is a 'constant', we can use a reciprocal divide.
118 */
119static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
120{
121 return reciprocal_divide(load, sg->reciprocal_cpu_power);
122}
123
124/*
125 * Each time a sched group cpu_power is changed,
126 * we must compute its reciprocal value
127 */
128static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
129{
130 sg->__cpu_power += val;
131 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
132}
133#endif
134
Ingo Molnar634fa8c2007-07-09 18:52:00 +0200135#define SCALE_PRIO(x, prio) \
136 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
Borislav Petkov91fcdd42006-10-19 23:28:29 -0700137
Ingo Molnar634fa8c2007-07-09 18:52:00 +0200138/*
139 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
140 * to time slice values: [800ms ... 100ms ... 5ms]
141 */
142static unsigned int static_prio_timeslice(int static_prio)
Peter Williams2dd73a42006-06-27 02:54:34 -0700143{
Ingo Molnar634fa8c2007-07-09 18:52:00 +0200144 if (static_prio == NICE_TO_PRIO(19))
145 return 1;
146
147 if (static_prio < NICE_TO_PRIO(0))
148 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
149 else
150 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
Peter Williams2dd73a42006-06-27 02:54:34 -0700151}
152
Ingo Molnare05606d2007-07-09 18:51:59 +0200153static inline int rt_policy(int policy)
154{
155 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
156 return 1;
157 return 0;
158}
159
160static inline int task_has_rt_policy(struct task_struct *p)
161{
162 return rt_policy(p->policy);
163}
164
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165/*
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200166 * This is the priority-queue data structure of the RT scheduling class:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200168struct rt_prio_array {
169 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
170 struct list_head queue[MAX_RT_PRIO];
171};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200173struct load_stat {
174 struct load_weight load;
175 u64 load_update_start, load_update_last;
176 unsigned long delta_fair, delta_exec, delta_stat;
177};
178
179/* CFS-related fields in a runqueue */
180struct cfs_rq {
181 struct load_weight load;
182 unsigned long nr_running;
183
184 s64 fair_clock;
185 u64 exec_clock;
186 s64 wait_runtime;
187 u64 sleeper_bonus;
188 unsigned long wait_runtime_overruns, wait_runtime_underruns;
189
190 struct rb_root tasks_timeline;
191 struct rb_node *rb_leftmost;
192 struct rb_node *rb_load_balance_curr;
193#ifdef CONFIG_FAIR_GROUP_SCHED
194 /* 'curr' points to currently running entity on this cfs_rq.
195 * It is set to NULL otherwise (i.e when none are currently running).
196 */
197 struct sched_entity *curr;
198 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
199
200 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
201 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
202 * (like users, containers etc.)
203 *
204 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
205 * list is used during load balance.
206 */
207 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
208#endif
209};
210
211/* Real-Time classes' related field in a runqueue: */
212struct rt_rq {
213 struct rt_prio_array active;
214 int rt_load_balance_idx;
215 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
216};
217
218/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 * This is the main, per-CPU runqueue data structure.
220 *
221 * Locking rule: those places that want to lock multiple runqueues
222 * (such as the load balancing or the thread migration code), lock
223 * acquire operations must be ordered by ascending &runqueue.
224 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700225struct rq {
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200226 spinlock_t lock; /* runqueue lock */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227
228 /*
229 * nr_running and cpu_load should be in the same cacheline because
230 * remote CPUs use both these fields when doing load calculation.
231 */
232 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200233 #define CPU_LOAD_IDX_MAX 5
234 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh Bbdecea32007-05-08 00:32:48 -0700235 unsigned char idle_at_tick;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700236#ifdef CONFIG_NO_HZ
237 unsigned char in_nohz_recently;
238#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200239 struct load_stat ls; /* capture load from *all* tasks on this cpu */
240 unsigned long nr_load_updates;
241 u64 nr_switches;
242
243 struct cfs_rq cfs;
244#ifdef CONFIG_FAIR_GROUP_SCHED
245 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200247 struct rt_rq rt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248
249 /*
250 * This is part of a global counter where only the total sum
251 * over all CPUs matters. A task can increase this counter on
252 * one CPU and if it got migrated afterwards it may decrease
253 * it on another CPU. Always updated under the runqueue lock:
254 */
255 unsigned long nr_uninterruptible;
256
Ingo Molnar36c8b582006-07-03 00:25:41 -0700257 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800258 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200260
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200261 u64 clock, prev_clock_raw;
262 s64 clock_max_delta;
263
264 unsigned int clock_warps, clock_overflows;
265 unsigned int clock_unstable_events;
Ingo Molnar529c7722007-08-10 23:05:11 +0200266 u64 tick_timestamp;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200267
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 atomic_t nr_iowait;
269
270#ifdef CONFIG_SMP
271 struct sched_domain *sd;
272
273 /* For active balancing */
274 int active_balance;
275 int push_cpu;
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700276 int cpu; /* cpu of this runqueue */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277
Ingo Molnar36c8b582006-07-03 00:25:41 -0700278 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 struct list_head migration_queue;
280#endif
281
282#ifdef CONFIG_SCHEDSTATS
283 /* latency stats */
284 struct sched_info rq_sched_info;
285
286 /* sys_sched_yield() stats */
287 unsigned long yld_exp_empty;
288 unsigned long yld_act_empty;
289 unsigned long yld_both_empty;
290 unsigned long yld_cnt;
291
292 /* schedule() stats */
293 unsigned long sched_switch;
294 unsigned long sched_cnt;
295 unsigned long sched_goidle;
296
297 /* try_to_wake_up() stats */
298 unsigned long ttwu_cnt;
299 unsigned long ttwu_local;
300#endif
Ingo Molnarfcb99372006-07-03 00:25:10 -0700301 struct lock_class_key rq_lock_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302};
303
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700304static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Gautham R Shenoy5be93612007-05-09 02:34:04 -0700305static DEFINE_MUTEX(sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306
Ingo Molnardd41f592007-07-09 18:51:59 +0200307static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
308{
309 rq->curr->sched_class->check_preempt_curr(rq, p);
310}
311
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700312static inline int cpu_of(struct rq *rq)
313{
314#ifdef CONFIG_SMP
315 return rq->cpu;
316#else
317 return 0;
318#endif
319}
320
Nick Piggin674311d2005-06-25 14:57:27 -0700321/*
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200322 * Update the per-runqueue clock, as finegrained as the platform can give
323 * us, but without assuming monotonicity, etc.:
Ingo Molnar20d315d2007-07-09 18:51:58 +0200324 */
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200325static void __update_rq_clock(struct rq *rq)
Ingo Molnar20d315d2007-07-09 18:51:58 +0200326{
327 u64 prev_raw = rq->prev_clock_raw;
328 u64 now = sched_clock();
329 s64 delta = now - prev_raw;
330 u64 clock = rq->clock;
331
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200332#ifdef CONFIG_SCHED_DEBUG
333 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
334#endif
Ingo Molnar20d315d2007-07-09 18:51:58 +0200335 /*
336 * Protect against sched_clock() occasionally going backwards:
337 */
338 if (unlikely(delta < 0)) {
339 clock++;
340 rq->clock_warps++;
341 } else {
342 /*
343 * Catch too large forward jumps too:
344 */
Ingo Molnar529c7722007-08-10 23:05:11 +0200345 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
346 if (clock < rq->tick_timestamp + TICK_NSEC)
347 clock = rq->tick_timestamp + TICK_NSEC;
348 else
349 clock++;
Ingo Molnar20d315d2007-07-09 18:51:58 +0200350 rq->clock_overflows++;
351 } else {
352 if (unlikely(delta > rq->clock_max_delta))
353 rq->clock_max_delta = delta;
354 clock += delta;
355 }
356 }
357
358 rq->prev_clock_raw = now;
359 rq->clock = clock;
Ingo Molnar20d315d2007-07-09 18:51:58 +0200360}
361
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200362static void update_rq_clock(struct rq *rq)
Ingo Molnar20d315d2007-07-09 18:51:58 +0200363{
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200364 if (likely(smp_processor_id() == cpu_of(rq)))
365 __update_rq_clock(rq);
366}
Ingo Molnar20d315d2007-07-09 18:51:58 +0200367
Ingo Molnar20d315d2007-07-09 18:51:58 +0200368/*
Nick Piggin674311d2005-06-25 14:57:27 -0700369 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700370 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700371 *
372 * The domain tree of any CPU may only be accessed from within
373 * preempt-disabled sections.
374 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700375#define for_each_domain(cpu, __sd) \
376 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
378#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
379#define this_rq() (&__get_cpu_var(runqueues))
380#define task_rq(p) cpu_rq(task_cpu(p))
381#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
382
Ingo Molnare436d802007-07-19 21:28:35 +0200383/*
384 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
385 * clock constructed from sched_clock():
386 */
387unsigned long long cpu_clock(int cpu)
388{
Ingo Molnare436d802007-07-19 21:28:35 +0200389 unsigned long long now;
390 unsigned long flags;
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200391 struct rq *rq;
Ingo Molnare436d802007-07-19 21:28:35 +0200392
Ingo Molnar2cd4d0e2007-07-26 13:40:43 +0200393 local_irq_save(flags);
Ingo Molnarb04a0f42007-08-09 11:16:46 +0200394 rq = cpu_rq(cpu);
395 update_rq_clock(rq);
396 now = rq->clock;
Ingo Molnar2cd4d0e2007-07-26 13:40:43 +0200397 local_irq_restore(flags);
Ingo Molnare436d802007-07-19 21:28:35 +0200398
399 return now;
400}
401
Ingo Molnar138a8ae2007-07-09 18:51:58 +0200402#ifdef CONFIG_FAIR_GROUP_SCHED
403/* Change a task's ->cfs_rq if it moves across CPUs */
404static inline void set_task_cfs_rq(struct task_struct *p)
405{
406 p->se.cfs_rq = &task_rq(p)->cfs;
407}
408#else
409static inline void set_task_cfs_rq(struct task_struct *p)
410{
411}
412#endif
413
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700415# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700417#ifndef finish_arch_switch
418# define finish_arch_switch(prev) do { } while (0)
419#endif
420
421#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700422static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700423{
424 return rq->curr == p;
425}
426
Ingo Molnar70b97a72006-07-03 00:25:42 -0700427static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700428{
429}
430
Ingo Molnar70b97a72006-07-03 00:25:42 -0700431static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700432{
Ingo Molnarda04c032005-09-13 11:17:59 +0200433#ifdef CONFIG_DEBUG_SPINLOCK
434 /* this is a valid case when another task releases the spinlock */
435 rq->lock.owner = current;
436#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700437 /*
438 * If we are tracking spinlock dependencies then we have to
439 * fix up the runqueue lock - which gets 'carried over' from
440 * prev into current:
441 */
442 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
443
Nick Piggin4866cde2005-06-25 14:57:23 -0700444 spin_unlock_irq(&rq->lock);
445}
446
447#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700448static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700449{
450#ifdef CONFIG_SMP
451 return p->oncpu;
452#else
453 return rq->curr == p;
454#endif
455}
456
Ingo Molnar70b97a72006-07-03 00:25:42 -0700457static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700458{
459#ifdef CONFIG_SMP
460 /*
461 * We can optimise this out completely for !SMP, because the
462 * SMP rebalancing from interrupt is the only thing that cares
463 * here.
464 */
465 next->oncpu = 1;
466#endif
467#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
468 spin_unlock_irq(&rq->lock);
469#else
470 spin_unlock(&rq->lock);
471#endif
472}
473
Ingo Molnar70b97a72006-07-03 00:25:42 -0700474static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700475{
476#ifdef CONFIG_SMP
477 /*
478 * After ->oncpu is cleared, the task can be moved to a different CPU.
479 * We must ensure this doesn't happen until the switch is completely
480 * finished.
481 */
482 smp_wmb();
483 prev->oncpu = 0;
484#endif
485#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
486 local_irq_enable();
487#endif
488}
489#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490
491/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700492 * __task_rq_lock - lock the runqueue a given task resides on.
493 * Must be called interrupts disabled.
494 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700495static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700496 __acquires(rq->lock)
497{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700498 struct rq *rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700499
500repeat_lock_task:
501 rq = task_rq(p);
502 spin_lock(&rq->lock);
503 if (unlikely(rq != task_rq(p))) {
504 spin_unlock(&rq->lock);
505 goto repeat_lock_task;
506 }
507 return rq;
508}
509
510/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 * task_rq_lock - lock the runqueue a given task resides on and disable
512 * interrupts. Note the ordering: we can safely lookup the task_rq without
513 * explicitly disabling preemption.
514 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700515static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 __acquires(rq->lock)
517{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700518 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519
520repeat_lock_task:
521 local_irq_save(*flags);
522 rq = task_rq(p);
523 spin_lock(&rq->lock);
524 if (unlikely(rq != task_rq(p))) {
525 spin_unlock_irqrestore(&rq->lock, *flags);
526 goto repeat_lock_task;
527 }
528 return rq;
529}
530
Ingo Molnar70b97a72006-07-03 00:25:42 -0700531static inline void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700532 __releases(rq->lock)
533{
534 spin_unlock(&rq->lock);
535}
536
Ingo Molnar70b97a72006-07-03 00:25:42 -0700537static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 __releases(rq->lock)
539{
540 spin_unlock_irqrestore(&rq->lock, *flags);
541}
542
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -0800544 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700546static inline struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 __acquires(rq->lock)
548{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700549 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550
551 local_irq_disable();
552 rq = this_rq();
553 spin_lock(&rq->lock);
554
555 return rq;
556}
557
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200558/*
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200559 * CPU frequency is/was unstable - start new by setting prev_clock_raw:
560 */
561void sched_clock_unstable_event(void)
562{
563 unsigned long flags;
564 struct rq *rq;
565
566 rq = task_rq_lock(current, &flags);
567 rq->prev_clock_raw = sched_clock();
568 rq->clock_unstable_events++;
569 task_rq_unlock(rq, &flags);
570}
571
572/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200573 * resched_task - mark a task 'to be rescheduled now'.
574 *
575 * On UP this means the setting of the need_resched flag, on SMP it
576 * might also involve a cross-CPU call to trigger the scheduler on
577 * the target CPU.
578 */
579#ifdef CONFIG_SMP
580
581#ifndef tsk_is_polling
582#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
583#endif
584
585static void resched_task(struct task_struct *p)
586{
587 int cpu;
588
589 assert_spin_locked(&task_rq(p)->lock);
590
591 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
592 return;
593
594 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
595
596 cpu = task_cpu(p);
597 if (cpu == smp_processor_id())
598 return;
599
600 /* NEED_RESCHED must be visible before we test polling */
601 smp_mb();
602 if (!tsk_is_polling(p))
603 smp_send_reschedule(cpu);
604}
605
606static void resched_cpu(int cpu)
607{
608 struct rq *rq = cpu_rq(cpu);
609 unsigned long flags;
610
611 if (!spin_trylock_irqsave(&rq->lock, flags))
612 return;
613 resched_task(cpu_curr(cpu));
614 spin_unlock_irqrestore(&rq->lock, flags);
615}
616#else
617static inline void resched_task(struct task_struct *p)
618{
619 assert_spin_locked(&task_rq(p)->lock);
620 set_tsk_need_resched(p);
621}
622#endif
623
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200624static u64 div64_likely32(u64 divident, unsigned long divisor)
625{
626#if BITS_PER_LONG == 32
627 if (likely(divident <= 0xffffffffULL))
628 return (u32)divident / divisor;
629 do_div(divident, divisor);
630
631 return divident;
632#else
633 return divident / divisor;
634#endif
635}
636
637#if BITS_PER_LONG == 32
638# define WMULT_CONST (~0UL)
639#else
640# define WMULT_CONST (1UL << 32)
641#endif
642
643#define WMULT_SHIFT 32
644
Ingo Molnar194081e2007-08-09 11:16:51 +0200645/*
646 * Shift right and round:
647 */
648#define RSR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
649
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +0200650static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200651calc_delta_mine(unsigned long delta_exec, unsigned long weight,
652 struct load_weight *lw)
653{
654 u64 tmp;
655
656 if (unlikely(!lw->inv_weight))
Ingo Molnar194081e2007-08-09 11:16:51 +0200657 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200658
659 tmp = (u64)delta_exec * weight;
660 /*
661 * Check whether we'd overflow the 64-bit multiplication:
662 */
Ingo Molnar194081e2007-08-09 11:16:51 +0200663 if (unlikely(tmp > WMULT_CONST))
664 tmp = RSR(RSR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
665 WMULT_SHIFT/2);
666 else
667 tmp = RSR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200668
Ingo Molnarecf691d2007-08-02 17:41:40 +0200669 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200670}
671
672static inline unsigned long
673calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
674{
675 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
676}
677
678static void update_load_add(struct load_weight *lw, unsigned long inc)
679{
680 lw->weight += inc;
681 lw->inv_weight = 0;
682}
683
684static void update_load_sub(struct load_weight *lw, unsigned long dec)
685{
686 lw->weight -= dec;
687 lw->inv_weight = 0;
688}
689
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690/*
Peter Williams2dd73a42006-06-27 02:54:34 -0700691 * To aid in avoiding the subversion of "niceness" due to uneven distribution
692 * of tasks with abnormal "nice" values across CPUs the contribution that
693 * each task makes to its run queue's load is weighted according to its
694 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
695 * scaled version of the new time slice allocation that they receive on time
696 * slice expiry etc.
697 */
698
Ingo Molnardd41f592007-07-09 18:51:59 +0200699#define WEIGHT_IDLEPRIO 2
700#define WMULT_IDLEPRIO (1 << 31)
701
702/*
703 * Nice levels are multiplicative, with a gentle 10% change for every
704 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
705 * nice 1, it will get ~10% less CPU time than another CPU-bound task
706 * that remained on nice 0.
707 *
708 * The "10% effect" is relative and cumulative: from _any_ nice level,
709 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +0200710 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
711 * If a task goes up by ~10% and another task goes down by ~10% then
712 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +0200713 */
714static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +0200715 /* -20 */ 88761, 71755, 56483, 46273, 36291,
716 /* -15 */ 29154, 23254, 18705, 14949, 11916,
717 /* -10 */ 9548, 7620, 6100, 4904, 3906,
718 /* -5 */ 3121, 2501, 1991, 1586, 1277,
719 /* 0 */ 1024, 820, 655, 526, 423,
720 /* 5 */ 335, 272, 215, 172, 137,
721 /* 10 */ 110, 87, 70, 56, 45,
722 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +0200723};
724
Ingo Molnar5714d2d2007-07-16 09:46:31 +0200725/*
726 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
727 *
728 * In cases where the weight does not change often, we can use the
729 * precalculated inverse to speed up arithmetics by turning divisions
730 * into multiplications:
731 */
Ingo Molnardd41f592007-07-09 18:51:59 +0200732static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +0200733 /* -20 */ 48388, 59856, 76040, 92818, 118348,
734 /* -15 */ 147320, 184698, 229616, 287308, 360437,
735 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
736 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
737 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
738 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
739 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
740 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +0200741};
Peter Williams2dd73a42006-06-27 02:54:34 -0700742
Ingo Molnardd41f592007-07-09 18:51:59 +0200743static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
744
745/*
746 * runqueue iterator, to support SMP load-balancing between different
747 * scheduling classes, without having to expose their internal data
748 * structures to the load-balancing proper:
749 */
750struct rq_iterator {
751 void *arg;
752 struct task_struct *(*start)(void *);
753 struct task_struct *(*next)(void *);
754};
755
756static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
757 unsigned long max_nr_move, unsigned long max_load_move,
758 struct sched_domain *sd, enum cpu_idle_type idle,
759 int *all_pinned, unsigned long *load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +0200760 int *this_best_prio, struct rq_iterator *iterator);
Ingo Molnardd41f592007-07-09 18:51:59 +0200761
762#include "sched_stats.h"
763#include "sched_rt.c"
764#include "sched_fair.c"
765#include "sched_idletask.c"
766#ifdef CONFIG_SCHED_DEBUG
767# include "sched_debug.c"
768#endif
769
770#define sched_class_highest (&rt_sched_class)
771
Ingo Molnar9c217242007-08-02 17:41:40 +0200772static void __update_curr_load(struct rq *rq, struct load_stat *ls)
773{
774 if (rq->curr != rq->idle && ls->load.weight) {
775 ls->delta_exec += ls->delta_stat;
776 ls->delta_fair += calc_delta_fair(ls->delta_stat, &ls->load);
777 ls->delta_stat = 0;
778 }
779}
780
781/*
782 * Update delta_exec, delta_fair fields for rq.
783 *
784 * delta_fair clock advances at a rate inversely proportional to
785 * total load (rq->ls.load.weight) on the runqueue, while
786 * delta_exec advances at the same rate as wall-clock (provided
787 * cpu is not idle).
788 *
789 * delta_exec / delta_fair is a measure of the (smoothened) load on this
790 * runqueue over any given interval. This (smoothened) load is used
791 * during load balance.
792 *
793 * This function is called /before/ updating rq->ls.load
794 * and when switching tasks.
795 */
Ingo Molnar84a1d7a2007-08-09 11:16:49 +0200796static void update_curr_load(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +0200797{
798 struct load_stat *ls = &rq->ls;
799 u64 start;
800
801 start = ls->load_update_start;
Ingo Molnard2819182007-08-09 11:16:47 +0200802 ls->load_update_start = rq->clock;
803 ls->delta_stat += rq->clock - start;
Ingo Molnar9c217242007-08-02 17:41:40 +0200804 /*
805 * Stagger updates to ls->delta_fair. Very frequent updates
806 * can be expensive.
807 */
808 if (ls->delta_stat >= sysctl_sched_stat_granularity)
809 __update_curr_load(rq, ls);
810}
811
Ingo Molnar29b4b622007-08-09 11:16:49 +0200812static inline void inc_load(struct rq *rq, const struct task_struct *p)
Ingo Molnar9c217242007-08-02 17:41:40 +0200813{
Ingo Molnar84a1d7a2007-08-09 11:16:49 +0200814 update_curr_load(rq);
Ingo Molnar9c217242007-08-02 17:41:40 +0200815 update_load_add(&rq->ls.load, p->se.load.weight);
816}
817
Ingo Molnar79b5ddd2007-08-09 11:16:49 +0200818static inline void dec_load(struct rq *rq, const struct task_struct *p)
Ingo Molnar9c217242007-08-02 17:41:40 +0200819{
Ingo Molnar84a1d7a2007-08-09 11:16:49 +0200820 update_curr_load(rq);
Ingo Molnar9c217242007-08-02 17:41:40 +0200821 update_load_sub(&rq->ls.load, p->se.load.weight);
822}
823
Ingo Molnare5fa2232007-08-09 11:16:49 +0200824static void inc_nr_running(struct task_struct *p, struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +0200825{
826 rq->nr_running++;
Ingo Molnar29b4b622007-08-09 11:16:49 +0200827 inc_load(rq, p);
Ingo Molnar9c217242007-08-02 17:41:40 +0200828}
829
Ingo Molnardb531812007-08-09 11:16:49 +0200830static void dec_nr_running(struct task_struct *p, struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +0200831{
832 rq->nr_running--;
Ingo Molnar79b5ddd2007-08-09 11:16:49 +0200833 dec_load(rq, p);
Ingo Molnar9c217242007-08-02 17:41:40 +0200834}
835
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200836static void set_load_weight(struct task_struct *p)
837{
Ingo Molnardd41f592007-07-09 18:51:59 +0200838 task_rq(p)->cfs.wait_runtime -= p->se.wait_runtime;
839 p->se.wait_runtime = 0;
840
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200841 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +0200842 p->se.load.weight = prio_to_weight[0] * 2;
843 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
844 return;
845 }
846
847 /*
848 * SCHED_IDLE tasks get minimal weight:
849 */
850 if (p->policy == SCHED_IDLE) {
851 p->se.load.weight = WEIGHT_IDLEPRIO;
852 p->se.load.inv_weight = WMULT_IDLEPRIO;
853 return;
854 }
855
856 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
857 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200858}
859
Ingo Molnar8159f872007-08-09 11:16:49 +0200860static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +0200861{
862 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +0200863 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +0200864 p->se.on_rq = 1;
865}
866
Ingo Molnar69be72c2007-08-09 11:16:49 +0200867static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +0200868{
Ingo Molnarf02231e2007-08-09 11:16:48 +0200869 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +0200870 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +0200871}
872
873/*
Ingo Molnardd41f592007-07-09 18:51:59 +0200874 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +0200875 */
Ingo Molnar14531182007-07-09 18:51:59 +0200876static inline int __normal_prio(struct task_struct *p)
877{
Ingo Molnardd41f592007-07-09 18:51:59 +0200878 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +0200879}
880
881/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700882 * Calculate the expected normal priority: i.e. priority
883 * without taking RT-inheritance into account. Might be
884 * boosted by interactivity modifiers. Changes upon fork,
885 * setprio syscalls, and whenever the interactivity
886 * estimator recalculates.
887 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700888static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700889{
890 int prio;
891
Ingo Molnare05606d2007-07-09 18:51:59 +0200892 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -0700893 prio = MAX_RT_PRIO-1 - p->rt_priority;
894 else
895 prio = __normal_prio(p);
896 return prio;
897}
898
899/*
900 * Calculate the current priority, i.e. the priority
901 * taken into account by the scheduler. This value might
902 * be boosted by RT tasks, or might be boosted by
903 * interactivity modifiers. Will be RT if the task got
904 * RT-boosted. If not then it returns p->normal_prio.
905 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700906static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700907{
908 p->normal_prio = normal_prio(p);
909 /*
910 * If we are RT tasks or we were boosted to RT priority,
911 * keep the priority unchanged. Otherwise, update priority
912 * to the normal priority:
913 */
914 if (!rt_prio(p->prio))
915 return p->normal_prio;
916 return p->prio;
917}
918
919/*
Ingo Molnardd41f592007-07-09 18:51:59 +0200920 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 */
Ingo Molnardd41f592007-07-09 18:51:59 +0200922static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923{
Ingo Molnardd41f592007-07-09 18:51:59 +0200924 if (p->state == TASK_UNINTERRUPTIBLE)
925 rq->nr_uninterruptible--;
926
Ingo Molnar8159f872007-08-09 11:16:49 +0200927 enqueue_task(rq, p, wakeup);
Ingo Molnare5fa2232007-08-09 11:16:49 +0200928 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929}
930
931/*
Ingo Molnardd41f592007-07-09 18:51:59 +0200932 * activate_idle_task - move idle task to the _front_ of runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 */
Ingo Molnardd41f592007-07-09 18:51:59 +0200934static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935{
Ingo Molnara8e504d2007-08-09 11:16:47 +0200936 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937
Ingo Molnardd41f592007-07-09 18:51:59 +0200938 if (p->state == TASK_UNINTERRUPTIBLE)
939 rq->nr_uninterruptible--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940
Ingo Molnar8159f872007-08-09 11:16:49 +0200941 enqueue_task(rq, p, 0);
Ingo Molnare5fa2232007-08-09 11:16:49 +0200942 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943}
944
945/*
946 * deactivate_task - remove a task from the runqueue.
947 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +0200948static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949{
Ingo Molnardd41f592007-07-09 18:51:59 +0200950 if (p->state == TASK_UNINTERRUPTIBLE)
951 rq->nr_uninterruptible++;
952
Ingo Molnar69be72c2007-08-09 11:16:49 +0200953 dequeue_task(rq, p, sleep);
Ingo Molnardb531812007-08-09 11:16:49 +0200954 dec_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955}
956
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957/**
958 * task_curr - is this task currently executing on a CPU?
959 * @p: the task in question.
960 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700961inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962{
963 return cpu_curr(task_cpu(p)) == p;
964}
965
Peter Williams2dd73a42006-06-27 02:54:34 -0700966/* Used instead of source_load when we know the type == 0 */
967unsigned long weighted_cpuload(const int cpu)
968{
Ingo Molnardd41f592007-07-09 18:51:59 +0200969 return cpu_rq(cpu)->ls.load.weight;
970}
971
972static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
973{
974#ifdef CONFIG_SMP
975 task_thread_info(p)->cpu = cpu;
976 set_task_cfs_rq(p);
977#endif
Peter Williams2dd73a42006-06-27 02:54:34 -0700978}
979
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980#ifdef CONFIG_SMP
Ingo Molnarc65cc872007-07-09 18:51:58 +0200981
Ingo Molnardd41f592007-07-09 18:51:59 +0200982void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +0200983{
Ingo Molnardd41f592007-07-09 18:51:59 +0200984 int old_cpu = task_cpu(p);
985 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
986 u64 clock_offset, fair_clock_offset;
987
988 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +0200989 fair_clock_offset = old_rq->cfs.fair_clock - new_rq->cfs.fair_clock;
990
Ingo Molnardd41f592007-07-09 18:51:59 +0200991 if (p->se.wait_start_fair)
992 p->se.wait_start_fair -= fair_clock_offset;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +0200993 if (p->se.sleep_start_fair)
994 p->se.sleep_start_fair -= fair_clock_offset;
995
996#ifdef CONFIG_SCHEDSTATS
997 if (p->se.wait_start)
998 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +0200999 if (p->se.sleep_start)
1000 p->se.sleep_start -= clock_offset;
1001 if (p->se.block_start)
1002 p->se.block_start -= clock_offset;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001003#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001004
1005 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02001006}
1007
Ingo Molnar70b97a72006-07-03 00:25:42 -07001008struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010
Ingo Molnar36c8b582006-07-03 00:25:41 -07001011 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 int dest_cpu;
1013
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001015};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016
1017/*
1018 * The task's runqueue lock must be held.
1019 * Returns true if you have to wait for migration thread.
1020 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001021static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07001022migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001024 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025
1026 /*
1027 * If the task is not on a runqueue (and not running), then
1028 * it is sufficient to simply update the task's cpu field.
1029 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001030 if (!p->se.on_rq && !task_running(rq, p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 set_task_cpu(p, dest_cpu);
1032 return 0;
1033 }
1034
1035 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 req->task = p;
1037 req->dest_cpu = dest_cpu;
1038 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07001039
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 return 1;
1041}
1042
1043/*
1044 * wait_task_inactive - wait for a thread to unschedule.
1045 *
1046 * The caller must ensure that the task *will* unschedule sometime soon,
1047 * else this function might spin for a *long* time. This function can't
1048 * be called with interrupts off, or it may introduce deadlock with
1049 * smp_call_function() if an IPI is sent by the same process we are
1050 * waiting to become inactive.
1051 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001052void wait_task_inactive(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053{
1054 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02001055 int running, on_rq;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001056 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057
1058repeat:
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001059 /*
1060 * We do the initial early heuristics without holding
1061 * any task-queue locks at all. We'll only try to get
1062 * the runqueue lock when things look like they will
1063 * work out!
1064 */
1065 rq = task_rq(p);
1066
1067 /*
1068 * If the task is actively running on another CPU
1069 * still, just relax and busy-wait without holding
1070 * any locks.
1071 *
1072 * NOTE! Since we don't hold any locks, it's not
1073 * even sure that "rq" stays as the right runqueue!
1074 * But we don't care, since "task_running()" will
1075 * return false if the runqueue has changed and p
1076 * is actually now running somewhere else!
1077 */
1078 while (task_running(rq, p))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079 cpu_relax();
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001080
1081 /*
1082 * Ok, time to look more closely! We need the rq
1083 * lock now, to be *sure*. If we're wrong, we'll
1084 * just go back and repeat.
1085 */
1086 rq = task_rq_lock(p, &flags);
1087 running = task_running(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02001088 on_rq = p->se.on_rq;
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001089 task_rq_unlock(rq, &flags);
1090
1091 /*
1092 * Was it really running after all now that we
1093 * checked with the proper locks actually held?
1094 *
1095 * Oops. Go back and try again..
1096 */
1097 if (unlikely(running)) {
1098 cpu_relax();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 goto repeat;
1100 }
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001101
1102 /*
1103 * It's not enough that it's not actively running,
1104 * it must be off the runqueue _entirely_, and not
1105 * preempted!
1106 *
1107 * So if it wa still runnable (but just not actively
1108 * running right now), it's preempted, and we should
1109 * yield - it could be a while.
1110 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001111 if (unlikely(on_rq)) {
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001112 yield();
1113 goto repeat;
1114 }
1115
1116 /*
1117 * Ahh, all good. It wasn't running, and it wasn't
1118 * runnable, which means that it will never become
1119 * running in the future either. We're all done!
1120 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121}
1122
1123/***
1124 * kick_process - kick a running thread to enter/exit the kernel
1125 * @p: the to-be-kicked thread
1126 *
1127 * Cause a process which is running on another CPU to enter
1128 * kernel-mode, without any delay. (to get signals handled.)
1129 *
1130 * NOTE: this function doesnt have to take the runqueue lock,
1131 * because all it wants to ensure is that the remote task enters
1132 * the kernel. If the IPI races and the task has been migrated
1133 * to another CPU then no harm is done and the purpose has been
1134 * achieved as well.
1135 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001136void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137{
1138 int cpu;
1139
1140 preempt_disable();
1141 cpu = task_cpu(p);
1142 if ((cpu != smp_processor_id()) && task_curr(p))
1143 smp_send_reschedule(cpu);
1144 preempt_enable();
1145}
1146
1147/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001148 * Return a low guess at the load of a migration-source cpu weighted
1149 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 *
1151 * We want to under-estimate the load of migration sources, to
1152 * balance conservatively.
1153 */
Con Kolivasb9104722005-11-08 21:38:55 -08001154static inline unsigned long source_load(int cpu, int type)
1155{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001156 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001157 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001158
Peter Williams2dd73a42006-06-27 02:54:34 -07001159 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02001160 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07001161
Ingo Molnardd41f592007-07-09 18:51:59 +02001162 return min(rq->cpu_load[type-1], total);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163}
1164
1165/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001166 * Return a high guess at the load of a migration-target cpu weighted
1167 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 */
Con Kolivasb9104722005-11-08 21:38:55 -08001169static inline unsigned long target_load(int cpu, int type)
1170{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001171 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001172 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08001173
Peter Williams2dd73a42006-06-27 02:54:34 -07001174 if (type == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02001175 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07001176
Ingo Molnardd41f592007-07-09 18:51:59 +02001177 return max(rq->cpu_load[type-1], total);
Peter Williams2dd73a42006-06-27 02:54:34 -07001178}
1179
1180/*
1181 * Return the average load per task on the cpu's run queue
1182 */
1183static inline unsigned long cpu_avg_load_per_task(int cpu)
1184{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001185 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001186 unsigned long total = weighted_cpuload(cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07001187 unsigned long n = rq->nr_running;
1188
Ingo Molnardd41f592007-07-09 18:51:59 +02001189 return n ? total / n : SCHED_LOAD_SCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190}
1191
Nick Piggin147cbb42005-06-25 14:57:19 -07001192/*
1193 * find_idlest_group finds and returns the least busy CPU group within the
1194 * domain.
1195 */
1196static struct sched_group *
1197find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1198{
1199 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1200 unsigned long min_load = ULONG_MAX, this_load = 0;
1201 int load_idx = sd->forkexec_idx;
1202 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1203
1204 do {
1205 unsigned long load, avg_load;
1206 int local_group;
1207 int i;
1208
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001209 /* Skip over this group if it has no CPUs allowed */
1210 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1211 goto nextgroup;
1212
Nick Piggin147cbb42005-06-25 14:57:19 -07001213 local_group = cpu_isset(this_cpu, group->cpumask);
Nick Piggin147cbb42005-06-25 14:57:19 -07001214
1215 /* Tally up the load of all CPUs in the group */
1216 avg_load = 0;
1217
1218 for_each_cpu_mask(i, group->cpumask) {
1219 /* Bias balancing toward cpus of our domain */
1220 if (local_group)
1221 load = source_load(i, load_idx);
1222 else
1223 load = target_load(i, load_idx);
1224
1225 avg_load += load;
1226 }
1227
1228 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07001229 avg_load = sg_div_cpu_power(group,
1230 avg_load * SCHED_LOAD_SCALE);
Nick Piggin147cbb42005-06-25 14:57:19 -07001231
1232 if (local_group) {
1233 this_load = avg_load;
1234 this = group;
1235 } else if (avg_load < min_load) {
1236 min_load = avg_load;
1237 idlest = group;
1238 }
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001239nextgroup:
Nick Piggin147cbb42005-06-25 14:57:19 -07001240 group = group->next;
1241 } while (group != sd->groups);
1242
1243 if (!idlest || 100*this_load < imbalance*min_load)
1244 return NULL;
1245 return idlest;
1246}
1247
1248/*
Satoru Takeuchi0feaece2006-10-03 01:14:10 -07001249 * find_idlest_cpu - find the idlest cpu among the cpus in group.
Nick Piggin147cbb42005-06-25 14:57:19 -07001250 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07001251static int
1252find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
Nick Piggin147cbb42005-06-25 14:57:19 -07001253{
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001254 cpumask_t tmp;
Nick Piggin147cbb42005-06-25 14:57:19 -07001255 unsigned long load, min_load = ULONG_MAX;
1256 int idlest = -1;
1257 int i;
1258
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001259 /* Traverse only the allowed CPUs */
1260 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1261
1262 for_each_cpu_mask(i, tmp) {
Peter Williams2dd73a42006-06-27 02:54:34 -07001263 load = weighted_cpuload(i);
Nick Piggin147cbb42005-06-25 14:57:19 -07001264
1265 if (load < min_load || (load == min_load && i == this_cpu)) {
1266 min_load = load;
1267 idlest = i;
1268 }
1269 }
1270
1271 return idlest;
1272}
1273
Nick Piggin476d1392005-06-25 14:57:29 -07001274/*
1275 * sched_balance_self: balance the current task (running on cpu) in domains
1276 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1277 * SD_BALANCE_EXEC.
1278 *
1279 * Balance, ie. select the least loaded group.
1280 *
1281 * Returns the target CPU number, or the same CPU if no balancing is needed.
1282 *
1283 * preempt must be disabled.
1284 */
1285static int sched_balance_self(int cpu, int flag)
1286{
1287 struct task_struct *t = current;
1288 struct sched_domain *tmp, *sd = NULL;
Nick Piggin147cbb42005-06-25 14:57:19 -07001289
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001290 for_each_domain(cpu, tmp) {
Ingo Molnar9761eea2007-07-09 18:52:00 +02001291 /*
1292 * If power savings logic is enabled for a domain, stop there.
1293 */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07001294 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1295 break;
Nick Piggin476d1392005-06-25 14:57:29 -07001296 if (tmp->flags & flag)
1297 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07001298 }
Nick Piggin476d1392005-06-25 14:57:29 -07001299
1300 while (sd) {
1301 cpumask_t span;
1302 struct sched_group *group;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001303 int new_cpu, weight;
1304
1305 if (!(sd->flags & flag)) {
1306 sd = sd->child;
1307 continue;
1308 }
Nick Piggin476d1392005-06-25 14:57:29 -07001309
1310 span = sd->span;
1311 group = find_idlest_group(sd, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001312 if (!group) {
1313 sd = sd->child;
1314 continue;
1315 }
Nick Piggin476d1392005-06-25 14:57:29 -07001316
M.Baris Demirayda5a5522005-09-10 00:26:09 -07001317 new_cpu = find_idlest_cpu(group, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001318 if (new_cpu == -1 || new_cpu == cpu) {
1319 /* Now try balancing at a lower domain level of cpu */
1320 sd = sd->child;
1321 continue;
1322 }
Nick Piggin476d1392005-06-25 14:57:29 -07001323
Siddha, Suresh B1a848872006-10-03 01:14:08 -07001324 /* Now try balancing at a lower domain level of new_cpu */
Nick Piggin476d1392005-06-25 14:57:29 -07001325 cpu = new_cpu;
Nick Piggin476d1392005-06-25 14:57:29 -07001326 sd = NULL;
1327 weight = cpus_weight(span);
1328 for_each_domain(cpu, tmp) {
1329 if (weight <= cpus_weight(tmp->span))
1330 break;
1331 if (tmp->flags & flag)
1332 sd = tmp;
1333 }
1334 /* while loop will break here if sd == NULL */
1335 }
1336
1337 return cpu;
1338}
1339
1340#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341
1342/*
1343 * wake_idle() will wake a task on an idle cpu if task->cpu is
1344 * not idle and an idle cpu is available. The span of cpus to
1345 * search starts with cpus closest then further out as needed,
1346 * so we always favor a closer, idle cpu.
1347 *
1348 * Returns the CPU we should wake onto.
1349 */
1350#if defined(ARCH_HAS_SCHED_WAKE_IDLE)
Ingo Molnar36c8b582006-07-03 00:25:41 -07001351static int wake_idle(int cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352{
1353 cpumask_t tmp;
1354 struct sched_domain *sd;
1355 int i;
1356
Siddha, Suresh B49531982007-05-08 00:33:01 -07001357 /*
1358 * If it is idle, then it is the best cpu to run this task.
1359 *
1360 * This cpu is also the best, if it has more than one task already.
1361 * Siblings must be also busy(in most cases) as they didn't already
1362 * pickup the extra load from this cpu and hence we need not check
1363 * sibling runqueue info. This will avoid the checks and cache miss
1364 * penalities associated with that.
1365 */
1366 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 return cpu;
1368
1369 for_each_domain(cpu, sd) {
1370 if (sd->flags & SD_WAKE_IDLE) {
Nick Piggine0f364f2005-06-25 14:57:06 -07001371 cpus_and(tmp, sd->span, p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 for_each_cpu_mask(i, tmp) {
1373 if (idle_cpu(i))
1374 return i;
1375 }
Ingo Molnar9761eea2007-07-09 18:52:00 +02001376 } else {
Nick Piggine0f364f2005-06-25 14:57:06 -07001377 break;
Ingo Molnar9761eea2007-07-09 18:52:00 +02001378 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 }
1380 return cpu;
1381}
1382#else
Ingo Molnar36c8b582006-07-03 00:25:41 -07001383static inline int wake_idle(int cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384{
1385 return cpu;
1386}
1387#endif
1388
1389/***
1390 * try_to_wake_up - wake up a thread
1391 * @p: the to-be-woken-up thread
1392 * @state: the mask of task states that can be woken
1393 * @sync: do a synchronous wakeup?
1394 *
1395 * Put it on the run-queue if it's not already there. The "current"
1396 * thread is always on the run-queue (except when the actual
1397 * re-schedule is in progress), and as such you're allowed to do
1398 * the simpler "current->state = TASK_RUNNING" to mark yourself
1399 * runnable without the overhead of this.
1400 *
1401 * returns failure only if the task is already active.
1402 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001403static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404{
1405 int cpu, this_cpu, success = 0;
1406 unsigned long flags;
1407 long old_state;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001408 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409#ifdef CONFIG_SMP
Nick Piggin78979862005-06-25 14:57:13 -07001410 struct sched_domain *sd, *this_sd = NULL;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001411 unsigned long load, this_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 int new_cpu;
1413#endif
1414
1415 rq = task_rq_lock(p, &flags);
1416 old_state = p->state;
1417 if (!(old_state & state))
1418 goto out;
1419
Ingo Molnardd41f592007-07-09 18:51:59 +02001420 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 goto out_running;
1422
1423 cpu = task_cpu(p);
1424 this_cpu = smp_processor_id();
1425
1426#ifdef CONFIG_SMP
1427 if (unlikely(task_running(rq, p)))
1428 goto out_activate;
1429
Nick Piggin78979862005-06-25 14:57:13 -07001430 new_cpu = cpu;
1431
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 schedstat_inc(rq, ttwu_cnt);
1433 if (cpu == this_cpu) {
1434 schedstat_inc(rq, ttwu_local);
Nick Piggin78979862005-06-25 14:57:13 -07001435 goto out_set_cpu;
1436 }
1437
1438 for_each_domain(this_cpu, sd) {
1439 if (cpu_isset(cpu, sd->span)) {
1440 schedstat_inc(sd, ttwu_wake_remote);
1441 this_sd = sd;
1442 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443 }
1444 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445
Nick Piggin78979862005-06-25 14:57:13 -07001446 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 goto out_set_cpu;
1448
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449 /*
Nick Piggin78979862005-06-25 14:57:13 -07001450 * Check for affine wakeup and passive balancing possibilities.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451 */
Nick Piggin78979862005-06-25 14:57:13 -07001452 if (this_sd) {
1453 int idx = this_sd->wake_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 unsigned int imbalance;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455
Nick Piggina3f21bc2005-06-25 14:57:15 -07001456 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1457
Nick Piggin78979862005-06-25 14:57:13 -07001458 load = source_load(cpu, idx);
1459 this_load = target_load(this_cpu, idx);
1460
Nick Piggin78979862005-06-25 14:57:13 -07001461 new_cpu = this_cpu; /* Wake to this CPU if we can */
1462
Nick Piggina3f21bc2005-06-25 14:57:15 -07001463 if (this_sd->flags & SD_WAKE_AFFINE) {
1464 unsigned long tl = this_load;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08001465 unsigned long tl_per_task;
1466
1467 tl_per_task = cpu_avg_load_per_task(this_cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07001468
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 /*
Nick Piggina3f21bc2005-06-25 14:57:15 -07001470 * If sync wakeup then subtract the (maximum possible)
1471 * effect of the currently running task from the load
1472 * of the current CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 */
Nick Piggina3f21bc2005-06-25 14:57:15 -07001474 if (sync)
Ingo Molnardd41f592007-07-09 18:51:59 +02001475 tl -= current->se.load.weight;
Nick Piggina3f21bc2005-06-25 14:57:15 -07001476
1477 if ((tl <= load &&
Peter Williams2dd73a42006-06-27 02:54:34 -07001478 tl + target_load(cpu, idx) <= tl_per_task) ||
Ingo Molnardd41f592007-07-09 18:51:59 +02001479 100*(tl + p->se.load.weight) <= imbalance*load) {
Nick Piggina3f21bc2005-06-25 14:57:15 -07001480 /*
1481 * This domain has SD_WAKE_AFFINE and
1482 * p is cache cold in this domain, and
1483 * there is no bad imbalance.
1484 */
1485 schedstat_inc(this_sd, ttwu_move_affine);
1486 goto out_set_cpu;
1487 }
1488 }
1489
1490 /*
1491 * Start passive balancing when half the imbalance_pct
1492 * limit is reached.
1493 */
1494 if (this_sd->flags & SD_WAKE_BALANCE) {
1495 if (imbalance*this_load <= 100*load) {
1496 schedstat_inc(this_sd, ttwu_move_balance);
1497 goto out_set_cpu;
1498 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499 }
1500 }
1501
1502 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1503out_set_cpu:
1504 new_cpu = wake_idle(new_cpu, p);
1505 if (new_cpu != cpu) {
1506 set_task_cpu(p, new_cpu);
1507 task_rq_unlock(rq, &flags);
1508 /* might preempt at this point */
1509 rq = task_rq_lock(p, &flags);
1510 old_state = p->state;
1511 if (!(old_state & state))
1512 goto out;
Ingo Molnardd41f592007-07-09 18:51:59 +02001513 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514 goto out_running;
1515
1516 this_cpu = smp_processor_id();
1517 cpu = task_cpu(p);
1518 }
1519
1520out_activate:
1521#endif /* CONFIG_SMP */
Ingo Molnar2daa3572007-08-09 11:16:51 +02001522 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02001523 activate_task(rq, p, 1);
Ingo Molnard79fc0f2005-09-10 00:26:12 -07001524 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 * Sync wakeups (i.e. those types of wakeups where the waker
1526 * has indicated that it will leave the CPU in short order)
1527 * don't trigger a preemption, if the woken up task will run on
1528 * this cpu. (in this case the 'I will reschedule' promise of
1529 * the waker guarantees that the freshly woken up task is going
1530 * to be considered on this CPU.)
1531 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001532 if (!sync || cpu != this_cpu)
1533 check_preempt_curr(rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 success = 1;
1535
1536out_running:
1537 p->state = TASK_RUNNING;
1538out:
1539 task_rq_unlock(rq, &flags);
1540
1541 return success;
1542}
1543
Ingo Molnar36c8b582006-07-03 00:25:41 -07001544int fastcall wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545{
1546 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1547 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1548}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549EXPORT_SYMBOL(wake_up_process);
1550
Ingo Molnar36c8b582006-07-03 00:25:41 -07001551int fastcall wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552{
1553 return try_to_wake_up(p, state, 0);
1554}
1555
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556/*
1557 * Perform scheduler related setup for a newly forked process p.
1558 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02001559 *
1560 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001562static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563{
Ingo Molnardd41f592007-07-09 18:51:59 +02001564 p->se.wait_start_fair = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02001565 p->se.exec_start = 0;
1566 p->se.sum_exec_runtime = 0;
1567 p->se.delta_exec = 0;
1568 p->se.delta_fair_run = 0;
1569 p->se.delta_fair_sleep = 0;
1570 p->se.wait_runtime = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001571 p->se.sleep_start_fair = 0;
1572
1573#ifdef CONFIG_SCHEDSTATS
1574 p->se.wait_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02001575 p->se.sum_wait_runtime = 0;
1576 p->se.sum_sleep_runtime = 0;
1577 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02001578 p->se.block_start = 0;
1579 p->se.sleep_max = 0;
1580 p->se.block_max = 0;
1581 p->se.exec_max = 0;
1582 p->se.wait_max = 0;
1583 p->se.wait_runtime_overruns = 0;
1584 p->se.wait_runtime_underruns = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001585#endif
Nick Piggin476d1392005-06-25 14:57:29 -07001586
Ingo Molnardd41f592007-07-09 18:51:59 +02001587 INIT_LIST_HEAD(&p->run_list);
1588 p->se.on_rq = 0;
Nick Piggin476d1392005-06-25 14:57:29 -07001589
Avi Kivitye107be32007-07-26 13:40:43 +02001590#ifdef CONFIG_PREEMPT_NOTIFIERS
1591 INIT_HLIST_HEAD(&p->preempt_notifiers);
1592#endif
1593
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 /*
1595 * We mark the process as running here, but have not actually
1596 * inserted it onto the runqueue yet. This guarantees that
1597 * nobody will actually run it, and a signal or other external
1598 * event cannot wake it up and insert it on the runqueue either.
1599 */
1600 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02001601}
1602
1603/*
1604 * fork()/clone()-time setup:
1605 */
1606void sched_fork(struct task_struct *p, int clone_flags)
1607{
1608 int cpu = get_cpu();
1609
1610 __sched_fork(p);
1611
1612#ifdef CONFIG_SMP
1613 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1614#endif
1615 __set_task_cpu(p, cpu);
Ingo Molnarb29739f2006-06-27 02:54:51 -07001616
1617 /*
1618 * Make sure we do not leak PI boosting priority to the child:
1619 */
1620 p->prio = current->normal_prio;
1621
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07001622#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02001623 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07001624 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08001626#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07001627 p->oncpu = 0;
1628#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07001630 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08001631 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632#endif
Nick Piggin476d1392005-06-25 14:57:29 -07001633 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634}
1635
1636/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001637 * After fork, child runs first. (default) If set to 0 then
1638 * parent will (try to) run first.
1639 */
1640unsigned int __read_mostly sysctl_sched_child_runs_first = 1;
1641
1642/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 * wake_up_new_task - wake up a newly created task for the first time.
1644 *
1645 * This function will do some initial scheduler statistics housekeeping
1646 * that must be done for every newly created context, then puts the task
1647 * on the runqueue and wakes it.
1648 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001649void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650{
1651 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02001652 struct rq *rq;
1653 int this_cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654
1655 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnardd41f592007-07-09 18:51:59 +02001657 this_cpu = smp_processor_id(); /* parent's CPU */
Ingo Molnara8e504d2007-08-09 11:16:47 +02001658 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659
1660 p->prio = effective_prio(p);
1661
Ingo Molnarcad60d92007-08-02 17:41:40 +02001662 if (!p->sched_class->task_new || !sysctl_sched_child_runs_first ||
1663 (clone_flags & CLONE_VM) || task_cpu(p) != this_cpu ||
1664 !current->se.on_rq) {
1665
Ingo Molnardd41f592007-07-09 18:51:59 +02001666 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02001669 * Let the scheduling class do new task startup
1670 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02001672 p->sched_class->task_new(rq, p);
Ingo Molnare5fa2232007-08-09 11:16:49 +02001673 inc_nr_running(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 }
Ingo Molnardd41f592007-07-09 18:51:59 +02001675 check_preempt_curr(rq, p);
1676 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677}
1678
Avi Kivitye107be32007-07-26 13:40:43 +02001679#ifdef CONFIG_PREEMPT_NOTIFIERS
1680
1681/**
Randy Dunlap421cee22007-07-31 00:37:50 -07001682 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1683 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02001684 */
1685void preempt_notifier_register(struct preempt_notifier *notifier)
1686{
1687 hlist_add_head(&notifier->link, &current->preempt_notifiers);
1688}
1689EXPORT_SYMBOL_GPL(preempt_notifier_register);
1690
1691/**
1692 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07001693 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02001694 *
1695 * This is safe to call from within a preemption notifier.
1696 */
1697void preempt_notifier_unregister(struct preempt_notifier *notifier)
1698{
1699 hlist_del(&notifier->link);
1700}
1701EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1702
1703static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1704{
1705 struct preempt_notifier *notifier;
1706 struct hlist_node *node;
1707
1708 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1709 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1710}
1711
1712static void
1713fire_sched_out_preempt_notifiers(struct task_struct *curr,
1714 struct task_struct *next)
1715{
1716 struct preempt_notifier *notifier;
1717 struct hlist_node *node;
1718
1719 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1720 notifier->ops->sched_out(notifier, next);
1721}
1722
1723#else
1724
1725static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1726{
1727}
1728
1729static void
1730fire_sched_out_preempt_notifiers(struct task_struct *curr,
1731 struct task_struct *next)
1732{
1733}
1734
1735#endif
1736
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737/**
Nick Piggin4866cde2005-06-25 14:57:23 -07001738 * prepare_task_switch - prepare to switch tasks
1739 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07001740 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07001741 * @next: the task we are going to switch to.
1742 *
1743 * This is called with the rq lock held and interrupts off. It must
1744 * be paired with a subsequent finish_task_switch after the context
1745 * switch.
1746 *
1747 * prepare_task_switch sets up locking and calls architecture specific
1748 * hooks.
1749 */
Avi Kivitye107be32007-07-26 13:40:43 +02001750static inline void
1751prepare_task_switch(struct rq *rq, struct task_struct *prev,
1752 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07001753{
Avi Kivitye107be32007-07-26 13:40:43 +02001754 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07001755 prepare_lock_switch(rq, next);
1756 prepare_arch_switch(next);
1757}
1758
1759/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04001761 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762 * @prev: the thread we just switched away from.
1763 *
Nick Piggin4866cde2005-06-25 14:57:23 -07001764 * finish_task_switch must be called after the context switch, paired
1765 * with a prepare_task_switch call before the context switch.
1766 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1767 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 *
1769 * Note that we may have delayed dropping an mm in context_switch(). If
1770 * so, we finish that here outside of the runqueue lock. (Doing it
1771 * with the lock held can cause deadlocks; see schedule() for
1772 * details.)
1773 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07001774static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775 __releases(rq->lock)
1776{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07001778 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779
1780 rq->prev_mm = NULL;
1781
1782 /*
1783 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07001784 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07001785 * schedule one last time. The schedule call will never return, and
1786 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07001787 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 * still held, otherwise prev could be scheduled on another cpu, die
1789 * there before we look at prev->state, and then the reference would
1790 * be dropped twice.
1791 * Manfred Spraul <manfred@colorfullife.com>
1792 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07001793 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07001794 finish_arch_switch(prev);
1795 finish_lock_switch(rq, prev);
Avi Kivitye107be32007-07-26 13:40:43 +02001796 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 if (mm)
1798 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07001799 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08001800 /*
1801 * Remove function-return probe instances associated with this
1802 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02001803 */
bibo maoc6fd91f2006-03-26 01:38:20 -08001804 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08001806 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807}
1808
1809/**
1810 * schedule_tail - first thing a freshly forked thread must call.
1811 * @prev: the thread we just switched away from.
1812 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001813asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814 __releases(rq->lock)
1815{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001816 struct rq *rq = this_rq();
1817
Nick Piggin4866cde2005-06-25 14:57:23 -07001818 finish_task_switch(rq, prev);
1819#ifdef __ARCH_WANT_UNLOCKED_CTXSW
1820 /* In this case, finish_task_switch does not reenable preemption */
1821 preempt_enable();
1822#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 if (current->set_child_tid)
1824 put_user(current->pid, current->set_child_tid);
1825}
1826
1827/*
1828 * context_switch - switch to the new MM and the new
1829 * thread's register state.
1830 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001831static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07001832context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07001833 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834{
Ingo Molnardd41f592007-07-09 18:51:59 +02001835 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836
Avi Kivitye107be32007-07-26 13:40:43 +02001837 prepare_task_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02001838 mm = next->mm;
1839 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01001840 /*
1841 * For paravirt, this is coupled with an exit in switch_to to
1842 * combine the page table reload and the switch backend into
1843 * one hypercall.
1844 */
1845 arch_enter_lazy_cpu_mode();
1846
Ingo Molnardd41f592007-07-09 18:51:59 +02001847 if (unlikely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848 next->active_mm = oldmm;
1849 atomic_inc(&oldmm->mm_count);
1850 enter_lazy_tlb(oldmm, next);
1851 } else
1852 switch_mm(oldmm, mm, next);
1853
Ingo Molnardd41f592007-07-09 18:51:59 +02001854 if (unlikely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 rq->prev_mm = oldmm;
1857 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07001858 /*
1859 * Since the runqueue lock will be released by the next
1860 * task (which is an invalid locking op but in the case
1861 * of the scheduler it's an obvious special-case), so we
1862 * do an early lockdep release here:
1863 */
1864#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07001865 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07001866#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867
1868 /* Here we just switch the register state and the stack. */
1869 switch_to(prev, next, prev);
1870
Ingo Molnardd41f592007-07-09 18:51:59 +02001871 barrier();
1872 /*
1873 * this_rq must be evaluated again because prev may have moved
1874 * CPUs since it called schedule(), thus the 'rq' on its stack
1875 * frame will be invalid.
1876 */
1877 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878}
1879
1880/*
1881 * nr_running, nr_uninterruptible and nr_context_switches:
1882 *
1883 * externally visible scheduler statistics: current number of runnable
1884 * threads, current number of uninterruptible-sleeping threads, total
1885 * number of context switches performed since bootup.
1886 */
1887unsigned long nr_running(void)
1888{
1889 unsigned long i, sum = 0;
1890
1891 for_each_online_cpu(i)
1892 sum += cpu_rq(i)->nr_running;
1893
1894 return sum;
1895}
1896
1897unsigned long nr_uninterruptible(void)
1898{
1899 unsigned long i, sum = 0;
1900
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08001901 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 sum += cpu_rq(i)->nr_uninterruptible;
1903
1904 /*
1905 * Since we read the counters lockless, it might be slightly
1906 * inaccurate. Do not allow it to go below zero though:
1907 */
1908 if (unlikely((long)sum < 0))
1909 sum = 0;
1910
1911 return sum;
1912}
1913
1914unsigned long long nr_context_switches(void)
1915{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07001916 int i;
1917 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08001919 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 sum += cpu_rq(i)->nr_switches;
1921
1922 return sum;
1923}
1924
1925unsigned long nr_iowait(void)
1926{
1927 unsigned long i, sum = 0;
1928
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08001929 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1931
1932 return sum;
1933}
1934
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08001935unsigned long nr_active(void)
1936{
1937 unsigned long i, running = 0, uninterruptible = 0;
1938
1939 for_each_online_cpu(i) {
1940 running += cpu_rq(i)->nr_running;
1941 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1942 }
1943
1944 if (unlikely((long)uninterruptible < 0))
1945 uninterruptible = 0;
1946
1947 return running + uninterruptible;
1948}
1949
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001951 * Update rq->cpu_load[] statistics. This function is usually called every
1952 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07001953 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001954static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07001955{
Ingo Molnardd41f592007-07-09 18:51:59 +02001956 u64 fair_delta64, exec_delta64, idle_delta64, sample_interval64, tmp64;
1957 unsigned long total_load = this_rq->ls.load.weight;
1958 unsigned long this_load = total_load;
1959 struct load_stat *ls = &this_rq->ls;
Ingo Molnardd41f592007-07-09 18:51:59 +02001960 int i, scale;
1961
1962 this_rq->nr_load_updates++;
1963 if (unlikely(!(sysctl_sched_features & SCHED_FEAT_PRECISE_CPU_LOAD)))
1964 goto do_avg;
1965
1966 /* Update delta_fair/delta_exec fields first */
Ingo Molnar84a1d7a2007-08-09 11:16:49 +02001967 update_curr_load(this_rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02001968
1969 fair_delta64 = ls->delta_fair + 1;
1970 ls->delta_fair = 0;
1971
1972 exec_delta64 = ls->delta_exec + 1;
1973 ls->delta_exec = 0;
1974
Ingo Molnard2819182007-08-09 11:16:47 +02001975 sample_interval64 = this_rq->clock - ls->load_update_last;
1976 ls->load_update_last = this_rq->clock;
Ingo Molnardd41f592007-07-09 18:51:59 +02001977
1978 if ((s64)sample_interval64 < (s64)TICK_NSEC)
1979 sample_interval64 = TICK_NSEC;
1980
1981 if (exec_delta64 > sample_interval64)
1982 exec_delta64 = sample_interval64;
1983
1984 idle_delta64 = sample_interval64 - exec_delta64;
1985
1986 tmp64 = div64_64(SCHED_LOAD_SCALE * exec_delta64, fair_delta64);
1987 tmp64 = div64_64(tmp64 * exec_delta64, sample_interval64);
1988
1989 this_load = (unsigned long)tmp64;
1990
1991do_avg:
1992
1993 /* Update our load: */
1994 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
1995 unsigned long old_load, new_load;
1996
1997 /* scale is effectively 1 << i now, and >> i divides by scale */
1998
1999 old_load = this_rq->cpu_load[i];
2000 new_load = this_load;
2001
2002 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2003 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002004}
2005
Ingo Molnardd41f592007-07-09 18:51:59 +02002006#ifdef CONFIG_SMP
2007
Ingo Molnar48f24c42006-07-03 00:25:40 -07002008/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009 * double_rq_lock - safely lock two runqueues
2010 *
2011 * Note this does not disable interrupts like task_rq_lock,
2012 * you need to do so manually before calling.
2013 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002014static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 __acquires(rq1->lock)
2016 __acquires(rq2->lock)
2017{
Kirill Korotaev054b9102006-12-10 02:20:11 -08002018 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 if (rq1 == rq2) {
2020 spin_lock(&rq1->lock);
2021 __acquire(rq2->lock); /* Fake it out ;) */
2022 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002023 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024 spin_lock(&rq1->lock);
2025 spin_lock(&rq2->lock);
2026 } else {
2027 spin_lock(&rq2->lock);
2028 spin_lock(&rq1->lock);
2029 }
2030 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002031 update_rq_clock(rq1);
2032 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033}
2034
2035/*
2036 * double_rq_unlock - safely unlock two runqueues
2037 *
2038 * Note this does not restore interrupts like task_rq_unlock,
2039 * you need to do so manually after calling.
2040 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002041static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 __releases(rq1->lock)
2043 __releases(rq2->lock)
2044{
2045 spin_unlock(&rq1->lock);
2046 if (rq1 != rq2)
2047 spin_unlock(&rq2->lock);
2048 else
2049 __release(rq2->lock);
2050}
2051
2052/*
2053 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2054 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002055static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056 __releases(this_rq->lock)
2057 __acquires(busiest->lock)
2058 __acquires(this_rq->lock)
2059{
Kirill Korotaev054b9102006-12-10 02:20:11 -08002060 if (unlikely(!irqs_disabled())) {
2061 /* printk() doesn't work good under rq->lock */
2062 spin_unlock(&this_rq->lock);
2063 BUG_ON(1);
2064 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 if (unlikely(!spin_trylock(&busiest->lock))) {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002066 if (busiest < this_rq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 spin_unlock(&this_rq->lock);
2068 spin_lock(&busiest->lock);
2069 spin_lock(&this_rq->lock);
2070 } else
2071 spin_lock(&busiest->lock);
2072 }
2073}
2074
2075/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 * If dest_cpu is allowed for this process, migrate the task to it.
2077 * This is accomplished by forcing the cpu_allowed mask to only
2078 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2079 * the cpu_allowed mask is restored.
2080 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002081static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002083 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002085 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086
2087 rq = task_rq_lock(p, &flags);
2088 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2089 || unlikely(cpu_is_offline(dest_cpu)))
2090 goto out;
2091
2092 /* force the process onto the specified CPU */
2093 if (migrate_task(p, dest_cpu, &req)) {
2094 /* Need to wait for migration thread (might exit: take ref). */
2095 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07002096
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 get_task_struct(mt);
2098 task_rq_unlock(rq, &flags);
2099 wake_up_process(mt);
2100 put_task_struct(mt);
2101 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07002102
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 return;
2104 }
2105out:
2106 task_rq_unlock(rq, &flags);
2107}
2108
2109/*
Nick Piggin476d1392005-06-25 14:57:29 -07002110 * sched_exec - execve() is a valuable balancing opportunity, because at
2111 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 */
2113void sched_exec(void)
2114{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 int new_cpu, this_cpu = get_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002116 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002118 if (new_cpu != this_cpu)
2119 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120}
2121
2122/*
2123 * pull_task - move a task from a remote runqueue to the local runqueue.
2124 * Both runqueues must be locked.
2125 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002126static void pull_task(struct rq *src_rq, struct task_struct *p,
2127 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02002129 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002131 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 /*
2133 * Note that idle threads have a prio of MAX_PRIO, for this test
2134 * to be always true for them.
2135 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002136 check_preempt_curr(this_rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137}
2138
2139/*
2140 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2141 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08002142static
Ingo Molnar70b97a72006-07-03 00:25:42 -07002143int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002144 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002145 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146{
2147 /*
2148 * We do not migrate tasks that are:
2149 * 1) running (obviously), or
2150 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2151 * 3) are cache-hot on their current CPU.
2152 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153 if (!cpu_isset(this_cpu, p->cpus_allowed))
2154 return 0;
Nick Piggin81026792005-06-25 14:57:07 -07002155 *all_pinned = 0;
2156
2157 if (task_running(rq, p))
2158 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159
2160 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02002161 * Aggressive migration if too many balance attempts have failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002163 if (sd->nr_balance_failed > sd->cache_nice_tries)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 return 1;
2165
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 return 1;
2167}
2168
Ingo Molnardd41f592007-07-09 18:51:59 +02002169static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2170 unsigned long max_nr_move, unsigned long max_load_move,
2171 struct sched_domain *sd, enum cpu_idle_type idle,
2172 int *all_pinned, unsigned long *load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002173 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02002174{
2175 int pulled = 0, pinned = 0, skip_for_load;
2176 struct task_struct *p;
2177 long rem_load_move = max_load_move;
2178
2179 if (max_nr_move == 0 || max_load_move == 0)
2180 goto out;
2181
2182 pinned = 1;
2183
2184 /*
2185 * Start the load-balancing iterator:
2186 */
2187 p = iterator->start(iterator->arg);
2188next:
2189 if (!p)
2190 goto out;
2191 /*
2192 * To help distribute high priority tasks accross CPUs we don't
2193 * skip a task if it will be the highest priority task (i.e. smallest
2194 * prio value) on its new queue regardless of its load weight
2195 */
2196 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2197 SCHED_LOAD_SCALE_FUZZ;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002198 if ((skip_for_load && p->prio >= *this_best_prio) ||
Ingo Molnardd41f592007-07-09 18:51:59 +02002199 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002200 p = iterator->next(iterator->arg);
2201 goto next;
2202 }
2203
2204 pull_task(busiest, p, this_rq, this_cpu);
2205 pulled++;
2206 rem_load_move -= p->se.load.weight;
2207
2208 /*
2209 * We only want to steal up to the prescribed number of tasks
2210 * and the prescribed amount of weighted load.
2211 */
2212 if (pulled < max_nr_move && rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002213 if (p->prio < *this_best_prio)
2214 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02002215 p = iterator->next(iterator->arg);
2216 goto next;
2217 }
2218out:
2219 /*
2220 * Right now, this is the only place pull_task() is called,
2221 * so we can safely collect pull_task() stats here rather than
2222 * inside pull_task().
2223 */
2224 schedstat_add(sd, lb_gained[idle], pulled);
2225
2226 if (all_pinned)
2227 *all_pinned = pinned;
2228 *load_moved = max_load_move - rem_load_move;
2229 return pulled;
2230}
Ingo Molnar48f24c42006-07-03 00:25:40 -07002231
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232/*
Peter Williams43010652007-08-09 11:16:46 +02002233 * move_tasks tries to move up to max_load_move weighted load from busiest to
2234 * this_rq, as part of a balancing operation within domain "sd".
2235 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236 *
2237 * Called with both runqueues locked.
2238 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002239static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02002240 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002241 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07002242 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243{
Ingo Molnardd41f592007-07-09 18:51:59 +02002244 struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02002245 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002246 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247
Ingo Molnardd41f592007-07-09 18:51:59 +02002248 do {
Peter Williams43010652007-08-09 11:16:46 +02002249 total_load_moved +=
2250 class->load_balance(this_rq, this_cpu, busiest,
2251 ULONG_MAX, max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002252 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02002253 class = class->next;
Peter Williams43010652007-08-09 11:16:46 +02002254 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255
Peter Williams43010652007-08-09 11:16:46 +02002256 return total_load_moved > 0;
2257}
2258
2259/*
2260 * move_one_task tries to move exactly one task from busiest to this_rq, as
2261 * part of active balancing operations within "domain".
2262 * Returns 1 if successful and 0 otherwise.
2263 *
2264 * Called with both runqueues locked.
2265 */
2266static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2267 struct sched_domain *sd, enum cpu_idle_type idle)
2268{
2269 struct sched_class *class;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002270 int this_best_prio = MAX_PRIO;
Peter Williams43010652007-08-09 11:16:46 +02002271
2272 for (class = sched_class_highest; class; class = class->next)
2273 if (class->load_balance(this_rq, this_cpu, busiest,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002274 1, ULONG_MAX, sd, idle, NULL,
2275 &this_best_prio))
Peter Williams43010652007-08-09 11:16:46 +02002276 return 1;
2277
2278 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279}
2280
2281/*
2282 * find_busiest_group finds and returns the busiest CPU group within the
Ingo Molnar48f24c42006-07-03 00:25:40 -07002283 * domain. It calculates and returns the amount of weighted load which
2284 * should be moved to restore balance via the imbalance parameter.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 */
2286static struct sched_group *
2287find_busiest_group(struct sched_domain *sd, int this_cpu,
Ingo Molnardd41f592007-07-09 18:51:59 +02002288 unsigned long *imbalance, enum cpu_idle_type idle,
2289 int *sd_idle, cpumask_t *cpus, int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290{
2291 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2292 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002293 unsigned long max_pull;
Peter Williams2dd73a42006-06-27 02:54:34 -07002294 unsigned long busiest_load_per_task, busiest_nr_running;
2295 unsigned long this_load_per_task, this_nr_running;
Nick Piggin78979862005-06-25 14:57:13 -07002296 int load_idx;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002297#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2298 int power_savings_balance = 1;
2299 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2300 unsigned long min_nr_running = ULONG_MAX;
2301 struct sched_group *group_min = NULL, *group_leader = NULL;
2302#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303
2304 max_load = this_load = total_load = total_pwr = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07002305 busiest_load_per_task = busiest_nr_running = 0;
2306 this_load_per_task = this_nr_running = 0;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002307 if (idle == CPU_NOT_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07002308 load_idx = sd->busy_idx;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002309 else if (idle == CPU_NEWLY_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07002310 load_idx = sd->newidle_idx;
2311 else
2312 load_idx = sd->idle_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313
2314 do {
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002315 unsigned long load, group_capacity;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 int local_group;
2317 int i;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002318 unsigned int balance_cpu = -1, first_idle_cpu = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07002319 unsigned long sum_nr_running, sum_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320
2321 local_group = cpu_isset(this_cpu, group->cpumask);
2322
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002323 if (local_group)
2324 balance_cpu = first_cpu(group->cpumask);
2325
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 /* Tally up the load of all CPUs in the group */
Peter Williams2dd73a42006-06-27 02:54:34 -07002327 sum_weighted_load = sum_nr_running = avg_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328
2329 for_each_cpu_mask(i, group->cpumask) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002330 struct rq *rq;
2331
2332 if (!cpu_isset(i, *cpus))
2333 continue;
2334
2335 rq = cpu_rq(i);
Peter Williams2dd73a42006-06-27 02:54:34 -07002336
Suresh Siddha9439aab2007-07-19 21:28:35 +02002337 if (*sd_idle && rq->nr_running)
Nick Piggin5969fe02005-09-10 00:26:19 -07002338 *sd_idle = 0;
2339
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 /* Bias balancing toward cpus of our domain */
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002341 if (local_group) {
2342 if (idle_cpu(i) && !first_idle_cpu) {
2343 first_idle_cpu = 1;
2344 balance_cpu = i;
2345 }
2346
Nick Piggina2000572006-02-10 01:51:02 -08002347 load = target_load(i, load_idx);
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002348 } else
Nick Piggina2000572006-02-10 01:51:02 -08002349 load = source_load(i, load_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350
2351 avg_load += load;
Peter Williams2dd73a42006-06-27 02:54:34 -07002352 sum_nr_running += rq->nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02002353 sum_weighted_load += weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 }
2355
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002356 /*
2357 * First idle cpu or the first cpu(busiest) in this sched group
2358 * is eligible for doing load balancing at this and above
Suresh Siddha9439aab2007-07-19 21:28:35 +02002359 * domains. In the newly idle case, we will allow all the cpu's
2360 * to do the newly idle load balance.
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002361 */
Suresh Siddha9439aab2007-07-19 21:28:35 +02002362 if (idle != CPU_NEWLY_IDLE && local_group &&
2363 balance_cpu != this_cpu && balance) {
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002364 *balance = 0;
2365 goto ret;
2366 }
2367
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 total_load += avg_load;
Eric Dumazet5517d862007-05-08 00:32:57 -07002369 total_pwr += group->__cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370
2371 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07002372 avg_load = sg_div_cpu_power(group,
2373 avg_load * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374
Eric Dumazet5517d862007-05-08 00:32:57 -07002375 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002376
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377 if (local_group) {
2378 this_load = avg_load;
2379 this = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07002380 this_nr_running = sum_nr_running;
2381 this_load_per_task = sum_weighted_load;
2382 } else if (avg_load > max_load &&
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002383 sum_nr_running > group_capacity) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 max_load = avg_load;
2385 busiest = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07002386 busiest_nr_running = sum_nr_running;
2387 busiest_load_per_task = sum_weighted_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002389
2390#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2391 /*
2392 * Busy processors will not participate in power savings
2393 * balance.
2394 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002395 if (idle == CPU_NOT_IDLE ||
2396 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2397 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002398
2399 /*
2400 * If the local group is idle or completely loaded
2401 * no need to do power savings balance at this domain
2402 */
2403 if (local_group && (this_nr_running >= group_capacity ||
2404 !this_nr_running))
2405 power_savings_balance = 0;
2406
Ingo Molnardd41f592007-07-09 18:51:59 +02002407 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002408 * If a group is already running at full capacity or idle,
2409 * don't include that group in power savings calculations
Ingo Molnardd41f592007-07-09 18:51:59 +02002410 */
2411 if (!power_savings_balance || sum_nr_running >= group_capacity
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002412 || !sum_nr_running)
Ingo Molnardd41f592007-07-09 18:51:59 +02002413 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002414
Ingo Molnardd41f592007-07-09 18:51:59 +02002415 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002416 * Calculate the group which has the least non-idle load.
Ingo Molnardd41f592007-07-09 18:51:59 +02002417 * This is the group from where we need to pick up the load
2418 * for saving power
2419 */
2420 if ((sum_nr_running < min_nr_running) ||
2421 (sum_nr_running == min_nr_running &&
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002422 first_cpu(group->cpumask) <
2423 first_cpu(group_min->cpumask))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002424 group_min = group;
2425 min_nr_running = sum_nr_running;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002426 min_load_per_task = sum_weighted_load /
2427 sum_nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02002428 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002429
Ingo Molnardd41f592007-07-09 18:51:59 +02002430 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002431 * Calculate the group which is almost near its
Ingo Molnardd41f592007-07-09 18:51:59 +02002432 * capacity but still has some space to pick up some load
2433 * from other group and save more power
2434 */
2435 if (sum_nr_running <= group_capacity - 1) {
2436 if (sum_nr_running > leader_nr_running ||
2437 (sum_nr_running == leader_nr_running &&
2438 first_cpu(group->cpumask) >
2439 first_cpu(group_leader->cpumask))) {
2440 group_leader = group;
2441 leader_nr_running = sum_nr_running;
2442 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002443 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002444group_next:
2445#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 group = group->next;
2447 } while (group != sd->groups);
2448
Peter Williams2dd73a42006-06-27 02:54:34 -07002449 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 goto out_balanced;
2451
2452 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2453
2454 if (this_load >= avg_load ||
2455 100*max_load <= sd->imbalance_pct*this_load)
2456 goto out_balanced;
2457
Peter Williams2dd73a42006-06-27 02:54:34 -07002458 busiest_load_per_task /= busiest_nr_running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 /*
2460 * We're trying to get all the cpus to the average_load, so we don't
2461 * want to push ourselves above the average load, nor do we wish to
2462 * reduce the max loaded cpu below the average load, as either of these
2463 * actions would just result in more rebalancing later, and ping-pong
2464 * tasks around. Thus we look for the minimum possible imbalance.
2465 * Negative imbalances (*we* are more loaded than anyone else) will
2466 * be counted as no imbalance for these purposes -- we can't fix that
2467 * by pulling tasks to us. Be careful of negative numbers as they'll
2468 * appear as very large values with unsigned longs.
2469 */
Peter Williams2dd73a42006-06-27 02:54:34 -07002470 if (max_load <= busiest_load_per_task)
2471 goto out_balanced;
2472
2473 /*
2474 * In the presence of smp nice balancing, certain scenarios can have
2475 * max load less than avg load(as we skip the groups at or below
2476 * its cpu_power, while calculating max_load..)
2477 */
2478 if (max_load < avg_load) {
2479 *imbalance = 0;
2480 goto small_imbalance;
2481 }
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002482
2483 /* Don't want to pull so many tasks that a group would go idle */
Peter Williams2dd73a42006-06-27 02:54:34 -07002484 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07002485
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486 /* How much load to actually move to equalise the imbalance */
Eric Dumazet5517d862007-05-08 00:32:57 -07002487 *imbalance = min(max_pull * busiest->__cpu_power,
2488 (avg_load - this_load) * this->__cpu_power)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 / SCHED_LOAD_SCALE;
2490
Peter Williams2dd73a42006-06-27 02:54:34 -07002491 /*
2492 * if *imbalance is less than the average load per runnable task
2493 * there is no gaurantee that any tasks will be moved so we'll have
2494 * a think about bumping its value to force at least one task to be
2495 * moved
2496 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002497 if (*imbalance + SCHED_LOAD_SCALE_FUZZ < busiest_load_per_task/2) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07002498 unsigned long tmp, pwr_now, pwr_move;
Peter Williams2dd73a42006-06-27 02:54:34 -07002499 unsigned int imbn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500
Peter Williams2dd73a42006-06-27 02:54:34 -07002501small_imbalance:
2502 pwr_move = pwr_now = 0;
2503 imbn = 2;
2504 if (this_nr_running) {
2505 this_load_per_task /= this_nr_running;
2506 if (busiest_load_per_task > this_load_per_task)
2507 imbn = 1;
2508 } else
2509 this_load_per_task = SCHED_LOAD_SCALE;
2510
Ingo Molnardd41f592007-07-09 18:51:59 +02002511 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2512 busiest_load_per_task * imbn) {
Peter Williams2dd73a42006-06-27 02:54:34 -07002513 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 return busiest;
2515 }
2516
2517 /*
2518 * OK, we don't have enough imbalance to justify moving tasks,
2519 * however we may be able to increase total CPU power used by
2520 * moving them.
2521 */
2522
Eric Dumazet5517d862007-05-08 00:32:57 -07002523 pwr_now += busiest->__cpu_power *
2524 min(busiest_load_per_task, max_load);
2525 pwr_now += this->__cpu_power *
2526 min(this_load_per_task, this_load);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 pwr_now /= SCHED_LOAD_SCALE;
2528
2529 /* Amount of load we'd subtract */
Eric Dumazet5517d862007-05-08 00:32:57 -07002530 tmp = sg_div_cpu_power(busiest,
2531 busiest_load_per_task * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 if (max_load > tmp)
Eric Dumazet5517d862007-05-08 00:32:57 -07002533 pwr_move += busiest->__cpu_power *
Peter Williams2dd73a42006-06-27 02:54:34 -07002534 min(busiest_load_per_task, max_load - tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535
2536 /* Amount of load we'd add */
Eric Dumazet5517d862007-05-08 00:32:57 -07002537 if (max_load * busiest->__cpu_power <
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08002538 busiest_load_per_task * SCHED_LOAD_SCALE)
Eric Dumazet5517d862007-05-08 00:32:57 -07002539 tmp = sg_div_cpu_power(this,
2540 max_load * busiest->__cpu_power);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 else
Eric Dumazet5517d862007-05-08 00:32:57 -07002542 tmp = sg_div_cpu_power(this,
2543 busiest_load_per_task * SCHED_LOAD_SCALE);
2544 pwr_move += this->__cpu_power *
2545 min(this_load_per_task, this_load + tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 pwr_move /= SCHED_LOAD_SCALE;
2547
2548 /* Move if we gain throughput */
2549 if (pwr_move <= pwr_now)
2550 goto out_balanced;
2551
Peter Williams2dd73a42006-06-27 02:54:34 -07002552 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553 }
2554
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 return busiest;
2556
2557out_balanced:
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002558#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002559 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002560 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002562 if (this == group_leader && group_leader != group_min) {
2563 *imbalance = min_load_per_task;
2564 return group_min;
2565 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002566#endif
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002567ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568 *imbalance = 0;
2569 return NULL;
2570}
2571
2572/*
2573 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2574 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002575static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002576find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002577 unsigned long imbalance, cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002579 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07002580 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581 int i;
2582
2583 for_each_cpu_mask(i, group->cpumask) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002584 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002585
2586 if (!cpu_isset(i, *cpus))
2587 continue;
2588
Ingo Molnar48f24c42006-07-03 00:25:40 -07002589 rq = cpu_rq(i);
Ingo Molnardd41f592007-07-09 18:51:59 +02002590 wl = weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591
Ingo Molnardd41f592007-07-09 18:51:59 +02002592 if (rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07002593 continue;
2594
Ingo Molnardd41f592007-07-09 18:51:59 +02002595 if (wl > max_load) {
2596 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002597 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598 }
2599 }
2600
2601 return busiest;
2602}
2603
2604/*
Nick Piggin77391d72005-06-25 14:57:30 -07002605 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2606 * so long as it is large enough.
2607 */
2608#define MAX_PINNED_INTERVAL 512
2609
2610/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2612 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002614static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002615 struct sched_domain *sd, enum cpu_idle_type idle,
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002616 int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617{
Peter Williams43010652007-08-09 11:16:46 +02002618 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002621 struct rq *busiest;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002622 cpumask_t cpus = CPU_MASK_ALL;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002623 unsigned long flags;
Nick Piggin5969fe02005-09-10 00:26:19 -07002624
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002625 /*
2626 * When power savings policy is enabled for the parent domain, idle
2627 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02002628 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002629 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002630 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002631 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002632 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002633 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 schedstat_inc(sd, lb_cnt[idle]);
2636
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002637redo:
2638 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002639 &cpus, balance);
2640
Chen, Kenneth W06066712006-12-10 02:20:35 -08002641 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002642 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002643
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 if (!group) {
2645 schedstat_inc(sd, lb_nobusyg[idle]);
2646 goto out_balanced;
2647 }
2648
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002649 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650 if (!busiest) {
2651 schedstat_inc(sd, lb_nobusyq[idle]);
2652 goto out_balanced;
2653 }
2654
Nick Piggindb935db2005-06-25 14:57:11 -07002655 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656
2657 schedstat_add(sd, lb_imbalance[idle], imbalance);
2658
Peter Williams43010652007-08-09 11:16:46 +02002659 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 if (busiest->nr_running > 1) {
2661 /*
2662 * Attempt to move tasks. If find_busiest_group has found
2663 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02002664 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 * correctly treated as an imbalance.
2666 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002667 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07002668 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02002669 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07002670 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07002671 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002672 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07002673
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07002674 /*
2675 * some other cpu did the load balance for us.
2676 */
Peter Williams43010652007-08-09 11:16:46 +02002677 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07002678 resched_cpu(this_cpu);
2679
Nick Piggin81026792005-06-25 14:57:07 -07002680 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002681 if (unlikely(all_pinned)) {
2682 cpu_clear(cpu_of(busiest), cpus);
2683 if (!cpus_empty(cpus))
2684 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07002685 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002686 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687 }
Nick Piggin81026792005-06-25 14:57:07 -07002688
Peter Williams43010652007-08-09 11:16:46 +02002689 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 schedstat_inc(sd, lb_failed[idle]);
2691 sd->nr_balance_failed++;
2692
2693 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002695 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07002696
2697 /* don't kick the migration_thread, if the curr
2698 * task on busiest cpu can't be moved to this_cpu
2699 */
2700 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002701 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07002702 all_pinned = 1;
2703 goto out_one_pinned;
2704 }
2705
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 if (!busiest->active_balance) {
2707 busiest->active_balance = 1;
2708 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07002709 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08002711 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07002712 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 wake_up_process(busiest->migration_thread);
2714
2715 /*
2716 * We've kicked active balancing, reset the failure
2717 * counter.
2718 */
Nick Piggin39507452005-06-25 14:57:09 -07002719 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 }
Nick Piggin81026792005-06-25 14:57:07 -07002721 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 sd->nr_balance_failed = 0;
2723
Nick Piggin81026792005-06-25 14:57:07 -07002724 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 /* We were unbalanced, so reset the balancing interval */
2726 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07002727 } else {
2728 /*
2729 * If we've begun active balancing, start to back off. This
2730 * case may not be covered by the all_pinned logic if there
2731 * is only 1 task on the busy runqueue (because we don't call
2732 * move_tasks).
2733 */
2734 if (sd->balance_interval < sd->max_interval)
2735 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736 }
2737
Peter Williams43010652007-08-09 11:16:46 +02002738 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002739 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002740 return -1;
Peter Williams43010652007-08-09 11:16:46 +02002741 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742
2743out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 schedstat_inc(sd, lb_balanced[idle]);
2745
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002746 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07002747
2748out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07002750 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2751 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 sd->balance_interval *= 2;
2753
Ingo Molnar48f24c42006-07-03 00:25:40 -07002754 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002755 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002756 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757 return 0;
2758}
2759
2760/*
2761 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2762 * tasks if there is an imbalance.
2763 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002764 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 * this_rq is locked.
2766 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07002767static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07002768load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769{
2770 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002771 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02002773 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07002774 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02002775 int all_pinned = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002776 cpumask_t cpus = CPU_MASK_ALL;
Nick Piggin5969fe02005-09-10 00:26:19 -07002777
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002778 /*
2779 * When power savings policy is enabled for the parent domain, idle
2780 * sibling can pick up load irrespective of busy siblings. In this case,
2781 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002782 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002783 */
2784 if (sd->flags & SD_SHARE_CPUPOWER &&
2785 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002786 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002788 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002789redo:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002790 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Siddha, Suresh B783609c2006-12-10 02:20:33 -08002791 &sd_idle, &cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002793 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002794 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795 }
2796
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002797 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002798 &cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07002799 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002800 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002801 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 }
2803
Nick Piggindb935db2005-06-25 14:57:11 -07002804 BUG_ON(busiest == this_rq);
2805
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002806 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07002807
Peter Williams43010652007-08-09 11:16:46 +02002808 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07002809 if (busiest->nr_running > 1) {
2810 /* Attempt to move tasks */
2811 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002812 /* this_rq->clock is already updated */
2813 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02002814 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02002815 imbalance, sd, CPU_NEWLY_IDLE,
2816 &all_pinned);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07002817 spin_unlock(&busiest->lock);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002818
Suresh Siddha969bb4e2007-07-19 21:28:35 +02002819 if (unlikely(all_pinned)) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07002820 cpu_clear(cpu_of(busiest), cpus);
2821 if (!cpus_empty(cpus))
2822 goto redo;
2823 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07002824 }
2825
Peter Williams43010652007-08-09 11:16:46 +02002826 if (!ld_moved) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002827 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002828 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2829 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002830 return -1;
2831 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002832 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833
Peter Williams43010652007-08-09 11:16:46 +02002834 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002835
2836out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002837 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002838 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07002839 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07002840 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002841 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002842
Nick Piggin16cfb1c2005-06-25 14:57:08 -07002843 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844}
2845
2846/*
2847 * idle_balance is called by schedule() if this_cpu is about to become
2848 * idle. Attempts to pull tasks from other CPUs.
2849 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002850static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851{
2852 struct sched_domain *sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02002853 int pulled_task = -1;
2854 unsigned long next_balance = jiffies + HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855
2856 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07002857 unsigned long interval;
2858
2859 if (!(sd->flags & SD_LOAD_BALANCE))
2860 continue;
2861
2862 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002863 /* If we've pulled tasks over stop searching: */
Christoph Lameter1bd77f22006-12-10 02:20:27 -08002864 pulled_task = load_balance_newidle(this_cpu,
Christoph Lameter92c4ca52007-06-23 17:16:33 -07002865 this_rq, sd);
2866
2867 interval = msecs_to_jiffies(sd->balance_interval);
2868 if (time_after(next_balance, sd->last_balance + interval))
2869 next_balance = sd->last_balance + interval;
2870 if (pulled_task)
2871 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 }
Ingo Molnardd41f592007-07-09 18:51:59 +02002873 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08002874 /*
2875 * We are going idle. next_balance may be set based on
2876 * a busy processor. So reset next_balance.
2877 */
2878 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02002879 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880}
2881
2882/*
2883 * active_load_balance is run by migration threads. It pushes running tasks
2884 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2885 * running on each physical CPU where possible, and avoids physical /
2886 * logical imbalances.
2887 *
2888 * Called with busiest_rq locked.
2889 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002890static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891{
Nick Piggin39507452005-06-25 14:57:09 -07002892 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002893 struct sched_domain *sd;
2894 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07002895
Ingo Molnar48f24c42006-07-03 00:25:40 -07002896 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07002897 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07002898 return;
2899
2900 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901
2902 /*
Nick Piggin39507452005-06-25 14:57:09 -07002903 * This condition is "impossible", if it occurs
2904 * we need to fix it. Originally reported by
2905 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906 */
Nick Piggin39507452005-06-25 14:57:09 -07002907 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908
Nick Piggin39507452005-06-25 14:57:09 -07002909 /* move a task from busiest_rq to target_rq */
2910 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002911 update_rq_clock(busiest_rq);
2912 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913
Nick Piggin39507452005-06-25 14:57:09 -07002914 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002915 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07002916 if ((sd->flags & SD_LOAD_BALANCE) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07002917 cpu_isset(busiest_cpu, sd->span))
Nick Piggin39507452005-06-25 14:57:09 -07002918 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002919 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920
Ingo Molnar48f24c42006-07-03 00:25:40 -07002921 if (likely(sd)) {
2922 schedstat_inc(sd, alb_cnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923
Peter Williams43010652007-08-09 11:16:46 +02002924 if (move_one_task(target_rq, target_cpu, busiest_rq,
2925 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07002926 schedstat_inc(sd, alb_pushed);
2927 else
2928 schedstat_inc(sd, alb_failed);
2929 }
Nick Piggin39507452005-06-25 14:57:09 -07002930 spin_unlock(&target_rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931}
2932
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07002933#ifdef CONFIG_NO_HZ
2934static struct {
2935 atomic_t load_balancer;
2936 cpumask_t cpu_mask;
2937} nohz ____cacheline_aligned = {
2938 .load_balancer = ATOMIC_INIT(-1),
2939 .cpu_mask = CPU_MASK_NONE,
2940};
2941
Christoph Lameter7835b982006-12-10 02:20:22 -08002942/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07002943 * This routine will try to nominate the ilb (idle load balancing)
2944 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2945 * load balancing on behalf of all those cpus. If all the cpus in the system
2946 * go into this tickless mode, then there will be no ilb owner (as there is
2947 * no need for one) and all the cpus will sleep till the next wakeup event
2948 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08002949 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07002950 * For the ilb owner, tick is not stopped. And this tick will be used
2951 * for idle load balancing. ilb owner will still be part of
2952 * nohz.cpu_mask..
2953 *
2954 * While stopping the tick, this cpu will become the ilb owner if there
2955 * is no other owner. And will be the owner till that cpu becomes busy
2956 * or if all cpus in the system stop their ticks at which point
2957 * there is no need for ilb owner.
2958 *
2959 * When the ilb owner becomes busy, it nominates another owner, during the
2960 * next busy scheduler_tick()
2961 */
2962int select_nohz_load_balancer(int stop_tick)
2963{
2964 int cpu = smp_processor_id();
2965
2966 if (stop_tick) {
2967 cpu_set(cpu, nohz.cpu_mask);
2968 cpu_rq(cpu)->in_nohz_recently = 1;
2969
2970 /*
2971 * If we are going offline and still the leader, give up!
2972 */
2973 if (cpu_is_offline(cpu) &&
2974 atomic_read(&nohz.load_balancer) == cpu) {
2975 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2976 BUG();
2977 return 0;
2978 }
2979
2980 /* time for ilb owner also to sleep */
2981 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2982 if (atomic_read(&nohz.load_balancer) == cpu)
2983 atomic_set(&nohz.load_balancer, -1);
2984 return 0;
2985 }
2986
2987 if (atomic_read(&nohz.load_balancer) == -1) {
2988 /* make me the ilb owner */
2989 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2990 return 1;
2991 } else if (atomic_read(&nohz.load_balancer) == cpu)
2992 return 1;
2993 } else {
2994 if (!cpu_isset(cpu, nohz.cpu_mask))
2995 return 0;
2996
2997 cpu_clear(cpu, nohz.cpu_mask);
2998
2999 if (atomic_read(&nohz.load_balancer) == cpu)
3000 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3001 BUG();
3002 }
3003 return 0;
3004}
3005#endif
3006
3007static DEFINE_SPINLOCK(balancing);
3008
3009/*
Christoph Lameter7835b982006-12-10 02:20:22 -08003010 * It checks each scheduling domain to see if it is due to be balanced,
3011 * and initiates a balancing operation if so.
3012 *
3013 * Balancing parameters are set up in arch_init_sched_domains.
3014 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003015static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08003016{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003017 int balance = 1;
3018 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08003019 unsigned long interval;
3020 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003021 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08003022 unsigned long next_balance = jiffies + 60*HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003024 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 if (!(sd->flags & SD_LOAD_BALANCE))
3026 continue;
3027
3028 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003029 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 interval *= sd->busy_factor;
3031
3032 /* scale ms to jiffies */
3033 interval = msecs_to_jiffies(interval);
3034 if (unlikely(!interval))
3035 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003036 if (interval > HZ*NR_CPUS/10)
3037 interval = HZ*NR_CPUS/10;
3038
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039
Christoph Lameter08c183f2006-12-10 02:20:29 -08003040 if (sd->flags & SD_SERIALIZE) {
3041 if (!spin_trylock(&balancing))
3042 goto out;
3043 }
3044
Christoph Lameterc9819f42006-12-10 02:20:25 -08003045 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003046 if (load_balance(cpu, rq, sd, idle, &balance)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003047 /*
3048 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07003049 * longer idle, or one of our SMT siblings is
3050 * not idle.
3051 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003052 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003054 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055 }
Christoph Lameter08c183f2006-12-10 02:20:29 -08003056 if (sd->flags & SD_SERIALIZE)
3057 spin_unlock(&balancing);
3058out:
Christoph Lameterc9819f42006-12-10 02:20:25 -08003059 if (time_after(next_balance, sd->last_balance + interval))
3060 next_balance = sd->last_balance + interval;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003061
3062 /*
3063 * Stop the load balance at this level. There is another
3064 * CPU in our sched group which is doing load balancing more
3065 * actively.
3066 */
3067 if (!balance)
3068 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003069 }
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003070 rq->next_balance = next_balance;
3071}
3072
3073/*
3074 * run_rebalance_domains is triggered when needed from the scheduler tick.
3075 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3076 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3077 */
3078static void run_rebalance_domains(struct softirq_action *h)
3079{
Ingo Molnardd41f592007-07-09 18:51:59 +02003080 int this_cpu = smp_processor_id();
3081 struct rq *this_rq = cpu_rq(this_cpu);
3082 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3083 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003084
Ingo Molnardd41f592007-07-09 18:51:59 +02003085 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003086
3087#ifdef CONFIG_NO_HZ
3088 /*
3089 * If this cpu is the owner for idle load balancing, then do the
3090 * balancing on behalf of the other idle cpus whose ticks are
3091 * stopped.
3092 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003093 if (this_rq->idle_at_tick &&
3094 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003095 cpumask_t cpus = nohz.cpu_mask;
3096 struct rq *rq;
3097 int balance_cpu;
3098
Ingo Molnardd41f592007-07-09 18:51:59 +02003099 cpu_clear(this_cpu, cpus);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003100 for_each_cpu_mask(balance_cpu, cpus) {
3101 /*
3102 * If this cpu gets work to do, stop the load balancing
3103 * work being done for other cpus. Next load
3104 * balancing owner will pick it up.
3105 */
3106 if (need_resched())
3107 break;
3108
Ingo Molnardd41f592007-07-09 18:51:59 +02003109 rebalance_domains(balance_cpu, SCHED_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003110
3111 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003112 if (time_after(this_rq->next_balance, rq->next_balance))
3113 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003114 }
3115 }
3116#endif
3117}
3118
3119/*
3120 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3121 *
3122 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3123 * idle load balancing owner or decide to stop the periodic load balancing,
3124 * if the whole system is idle.
3125 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003126static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003127{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003128#ifdef CONFIG_NO_HZ
3129 /*
3130 * If we were in the nohz mode recently and busy at the current
3131 * scheduler tick, then check if we need to nominate new idle
3132 * load balancer.
3133 */
3134 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3135 rq->in_nohz_recently = 0;
3136
3137 if (atomic_read(&nohz.load_balancer) == cpu) {
3138 cpu_clear(cpu, nohz.cpu_mask);
3139 atomic_set(&nohz.load_balancer, -1);
3140 }
3141
3142 if (atomic_read(&nohz.load_balancer) == -1) {
3143 /*
3144 * simple selection for now: Nominate the
3145 * first cpu in the nohz list to be the next
3146 * ilb owner.
3147 *
3148 * TBD: Traverse the sched domains and nominate
3149 * the nearest cpu in the nohz.cpu_mask.
3150 */
3151 int ilb = first_cpu(nohz.cpu_mask);
3152
3153 if (ilb != NR_CPUS)
3154 resched_cpu(ilb);
3155 }
3156 }
3157
3158 /*
3159 * If this cpu is idle and doing idle load balancing for all the
3160 * cpus with ticks stopped, is it time for that to stop?
3161 */
3162 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3163 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3164 resched_cpu(cpu);
3165 return;
3166 }
3167
3168 /*
3169 * If this cpu is idle and the idle load balancing is done by
3170 * someone else, then no need raise the SCHED_SOFTIRQ
3171 */
3172 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3173 cpu_isset(cpu, nohz.cpu_mask))
3174 return;
3175#endif
3176 if (time_after_eq(jiffies, rq->next_balance))
3177 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178}
Ingo Molnardd41f592007-07-09 18:51:59 +02003179
3180#else /* CONFIG_SMP */
3181
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182/*
3183 * on UP we do not need to balance between CPUs:
3184 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003185static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186{
3187}
Ingo Molnardd41f592007-07-09 18:51:59 +02003188
3189/* Avoid "used but not defined" warning on UP */
3190static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3191 unsigned long max_nr_move, unsigned long max_load_move,
3192 struct sched_domain *sd, enum cpu_idle_type idle,
3193 int *all_pinned, unsigned long *load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003194 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02003195{
3196 *load_moved = 0;
3197
3198 return 0;
3199}
3200
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201#endif
3202
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203DEFINE_PER_CPU(struct kernel_stat, kstat);
3204
3205EXPORT_PER_CPU_SYMBOL(kstat);
3206
3207/*
Ingo Molnar41b86e92007-07-09 18:51:58 +02003208 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3209 * that have not yet been banked in case the task is currently running.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210 */
Ingo Molnar41b86e92007-07-09 18:51:58 +02003211unsigned long long task_sched_runtime(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003213 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003214 u64 ns, delta_exec;
3215 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003216
Ingo Molnar41b86e92007-07-09 18:51:58 +02003217 rq = task_rq_lock(p, &flags);
3218 ns = p->se.sum_exec_runtime;
3219 if (rq->curr == p) {
Ingo Molnara8e504d2007-08-09 11:16:47 +02003220 update_rq_clock(rq);
3221 delta_exec = rq->clock - p->se.exec_start;
Ingo Molnar41b86e92007-07-09 18:51:58 +02003222 if ((s64)delta_exec > 0)
3223 ns += delta_exec;
3224 }
3225 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003226
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227 return ns;
3228}
3229
3230/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231 * Account user cpu time to a process.
3232 * @p: the process that the cpu time gets accounted to
3233 * @hardirq_offset: the offset to subtract from hardirq_count()
3234 * @cputime: the cpu time spent in user space since the last update
3235 */
3236void account_user_time(struct task_struct *p, cputime_t cputime)
3237{
3238 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3239 cputime64_t tmp;
3240
3241 p->utime = cputime_add(p->utime, cputime);
3242
3243 /* Add user time to cpustat. */
3244 tmp = cputime_to_cputime64(cputime);
3245 if (TASK_NICE(p) > 0)
3246 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3247 else
3248 cpustat->user = cputime64_add(cpustat->user, tmp);
3249}
3250
3251/*
3252 * Account system cpu time to a process.
3253 * @p: the process that the cpu time gets accounted to
3254 * @hardirq_offset: the offset to subtract from hardirq_count()
3255 * @cputime: the cpu time spent in kernel space since the last update
3256 */
3257void account_system_time(struct task_struct *p, int hardirq_offset,
3258 cputime_t cputime)
3259{
3260 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003261 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003262 cputime64_t tmp;
3263
3264 p->stime = cputime_add(p->stime, cputime);
3265
3266 /* Add system time to cpustat. */
3267 tmp = cputime_to_cputime64(cputime);
3268 if (hardirq_count() - hardirq_offset)
3269 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3270 else if (softirq_count())
3271 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3272 else if (p != rq->idle)
3273 cpustat->system = cputime64_add(cpustat->system, tmp);
3274 else if (atomic_read(&rq->nr_iowait) > 0)
3275 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3276 else
3277 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3278 /* Account for system time used */
3279 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280}
3281
3282/*
3283 * Account for involuntary wait time.
3284 * @p: the process from which the cpu time has been stolen
3285 * @steal: the cpu time spent in involuntary wait
3286 */
3287void account_steal_time(struct task_struct *p, cputime_t steal)
3288{
3289 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3290 cputime64_t tmp = cputime_to_cputime64(steal);
Ingo Molnar70b97a72006-07-03 00:25:42 -07003291 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292
3293 if (p == rq->idle) {
3294 p->stime = cputime_add(p->stime, steal);
3295 if (atomic_read(&rq->nr_iowait) > 0)
3296 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3297 else
3298 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3299 } else
3300 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3301}
3302
Christoph Lameter7835b982006-12-10 02:20:22 -08003303/*
3304 * This function gets called by the timer code, with HZ frequency.
3305 * We call it with interrupts disabled.
3306 *
3307 * It also gets called by the fork code, when changing the parent's
3308 * timeslices.
3309 */
3310void scheduler_tick(void)
3311{
Christoph Lameter7835b982006-12-10 02:20:22 -08003312 int cpu = smp_processor_id();
3313 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003314 struct task_struct *curr = rq->curr;
Ingo Molnar529c7722007-08-10 23:05:11 +02003315 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
Christoph Lameter7835b982006-12-10 02:20:22 -08003316
Ingo Molnardd41f592007-07-09 18:51:59 +02003317 spin_lock(&rq->lock);
Ingo Molnar546fe3c2007-08-09 11:16:51 +02003318 __update_rq_clock(rq);
Ingo Molnar529c7722007-08-10 23:05:11 +02003319 /*
3320 * Let rq->clock advance by at least TICK_NSEC:
3321 */
3322 if (unlikely(rq->clock < next_tick))
3323 rq->clock = next_tick;
3324 rq->tick_timestamp = rq->clock;
Ingo Molnarf1a438d2007-08-09 11:16:45 +02003325 update_cpu_load(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02003326 if (curr != rq->idle) /* FIXME: needed? */
3327 curr->sched_class->task_tick(rq, curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02003328 spin_unlock(&rq->lock);
3329
Christoph Lametere418e1c2006-12-10 02:20:23 -08003330#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02003331 rq->idle_at_tick = idle_cpu(cpu);
3332 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08003333#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334}
3335
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3337
3338void fastcall add_preempt_count(int val)
3339{
3340 /*
3341 * Underflow?
3342 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003343 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3344 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345 preempt_count() += val;
3346 /*
3347 * Spinlock count overflowing soon?
3348 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08003349 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3350 PREEMPT_MASK - 10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351}
3352EXPORT_SYMBOL(add_preempt_count);
3353
3354void fastcall sub_preempt_count(int val)
3355{
3356 /*
3357 * Underflow?
3358 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003359 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3360 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 /*
3362 * Is the spinlock portion underflowing?
3363 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07003364 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3365 !(preempt_count() & PREEMPT_MASK)))
3366 return;
3367
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368 preempt_count() -= val;
3369}
3370EXPORT_SYMBOL(sub_preempt_count);
3371
3372#endif
3373
3374/*
Ingo Molnardd41f592007-07-09 18:51:59 +02003375 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003377static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378{
Ingo Molnardd41f592007-07-09 18:51:59 +02003379 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3380 prev->comm, preempt_count(), prev->pid);
3381 debug_show_held_locks(prev);
3382 if (irqs_disabled())
3383 print_irqtrace_events(prev);
3384 dump_stack();
3385}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386
Ingo Molnardd41f592007-07-09 18:51:59 +02003387/*
3388 * Various schedule()-time debugging checks and statistics:
3389 */
3390static inline void schedule_debug(struct task_struct *prev)
3391{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392 /*
3393 * Test if we are atomic. Since do_exit() needs to call into
3394 * schedule() atomically, we ignore that path for now.
3395 * Otherwise, whine if we are scheduling when we should not be.
3396 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003397 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3398 __schedule_bug(prev);
3399
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3401
Ingo Molnardd41f592007-07-09 18:51:59 +02003402 schedstat_inc(this_rq(), sched_cnt);
3403}
3404
3405/*
3406 * Pick up the highest-prio task:
3407 */
3408static inline struct task_struct *
Ingo Molnarff95f3d2007-08-09 11:16:49 +02003409pick_next_task(struct rq *rq, struct task_struct *prev)
Ingo Molnardd41f592007-07-09 18:51:59 +02003410{
3411 struct sched_class *class;
3412 struct task_struct *p;
3413
3414 /*
3415 * Optimization: we know that if all tasks are in
3416 * the fair class we can call that function directly:
3417 */
3418 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02003419 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02003420 if (likely(p))
3421 return p;
3422 }
3423
3424 class = sched_class_highest;
3425 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02003426 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02003427 if (p)
3428 return p;
3429 /*
3430 * Will never be NULL as the idle class always
3431 * returns a non-NULL p:
3432 */
3433 class = class->next;
3434 }
3435}
3436
3437/*
3438 * schedule() is the main scheduler function.
3439 */
3440asmlinkage void __sched schedule(void)
3441{
3442 struct task_struct *prev, *next;
3443 long *switch_count;
3444 struct rq *rq;
Ingo Molnardd41f592007-07-09 18:51:59 +02003445 int cpu;
3446
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447need_resched:
3448 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02003449 cpu = smp_processor_id();
3450 rq = cpu_rq(cpu);
3451 rcu_qsctr_inc(cpu);
3452 prev = rq->curr;
3453 switch_count = &prev->nivcsw;
3454
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455 release_kernel_lock(prev);
3456need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457
Ingo Molnardd41f592007-07-09 18:51:59 +02003458 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459
3460 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461 clear_tsk_need_resched(prev);
Ingo Molnarc1b3da32007-08-09 11:16:47 +02003462 __update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463
Ingo Molnardd41f592007-07-09 18:51:59 +02003464 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3465 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3466 unlikely(signal_pending(prev)))) {
3467 prev->state = TASK_RUNNING;
3468 } else {
Ingo Molnar2e1cb742007-08-09 11:16:49 +02003469 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02003470 }
3471 switch_count = &prev->nvcsw;
3472 }
3473
3474 if (unlikely(!rq->nr_running))
3475 idle_balance(cpu, rq);
3476
Ingo Molnar31ee5292007-08-09 11:16:49 +02003477 prev->sched_class->put_prev_task(rq, prev);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02003478 next = pick_next_task(rq, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479
3480 sched_info_switch(prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02003481
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 if (likely(prev != next)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483 rq->nr_switches++;
3484 rq->curr = next;
3485 ++*switch_count;
3486
Ingo Molnardd41f592007-07-09 18:51:59 +02003487 context_switch(rq, prev, next); /* unlocks the rq */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488 } else
3489 spin_unlock_irq(&rq->lock);
3490
Ingo Molnardd41f592007-07-09 18:51:59 +02003491 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3492 cpu = smp_processor_id();
3493 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 goto need_resched_nonpreemptible;
Ingo Molnardd41f592007-07-09 18:51:59 +02003495 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496 preempt_enable_no_resched();
3497 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3498 goto need_resched;
3499}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500EXPORT_SYMBOL(schedule);
3501
3502#ifdef CONFIG_PREEMPT
3503/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003504 * this is the entry point to schedule() from in-kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 * off of preempt_enable. Kernel preemptions off return from interrupt
3506 * occur there and call schedule directly.
3507 */
3508asmlinkage void __sched preempt_schedule(void)
3509{
3510 struct thread_info *ti = current_thread_info();
3511#ifdef CONFIG_PREEMPT_BKL
3512 struct task_struct *task = current;
3513 int saved_lock_depth;
3514#endif
3515 /*
3516 * If there is a non-zero preempt_count or interrupts are disabled,
3517 * we do not want to preempt the current task. Just return..
3518 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07003519 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 return;
3521
3522need_resched:
3523 add_preempt_count(PREEMPT_ACTIVE);
3524 /*
3525 * We keep the big kernel semaphore locked, but we
3526 * clear ->lock_depth so that schedule() doesnt
3527 * auto-release the semaphore:
3528 */
3529#ifdef CONFIG_PREEMPT_BKL
3530 saved_lock_depth = task->lock_depth;
3531 task->lock_depth = -1;
3532#endif
3533 schedule();
3534#ifdef CONFIG_PREEMPT_BKL
3535 task->lock_depth = saved_lock_depth;
3536#endif
3537 sub_preempt_count(PREEMPT_ACTIVE);
3538
3539 /* we could miss a preemption opportunity between schedule and now */
3540 barrier();
3541 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3542 goto need_resched;
3543}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544EXPORT_SYMBOL(preempt_schedule);
3545
3546/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003547 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 * off of irq context.
3549 * Note, that this is called and return with irqs disabled. This will
3550 * protect us against recursive calling from irq.
3551 */
3552asmlinkage void __sched preempt_schedule_irq(void)
3553{
3554 struct thread_info *ti = current_thread_info();
3555#ifdef CONFIG_PREEMPT_BKL
3556 struct task_struct *task = current;
3557 int saved_lock_depth;
3558#endif
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003559 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 BUG_ON(ti->preempt_count || !irqs_disabled());
3561
3562need_resched:
3563 add_preempt_count(PREEMPT_ACTIVE);
3564 /*
3565 * We keep the big kernel semaphore locked, but we
3566 * clear ->lock_depth so that schedule() doesnt
3567 * auto-release the semaphore:
3568 */
3569#ifdef CONFIG_PREEMPT_BKL
3570 saved_lock_depth = task->lock_depth;
3571 task->lock_depth = -1;
3572#endif
3573 local_irq_enable();
3574 schedule();
3575 local_irq_disable();
3576#ifdef CONFIG_PREEMPT_BKL
3577 task->lock_depth = saved_lock_depth;
3578#endif
3579 sub_preempt_count(PREEMPT_ACTIVE);
3580
3581 /* we could miss a preemption opportunity between schedule and now */
3582 barrier();
3583 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3584 goto need_resched;
3585}
3586
3587#endif /* CONFIG_PREEMPT */
3588
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003589int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3590 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591{
Ingo Molnar48f24c42006-07-03 00:25:40 -07003592 return try_to_wake_up(curr->private, mode, sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594EXPORT_SYMBOL(default_wake_function);
3595
3596/*
3597 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3598 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3599 * number) then we wake all the non-exclusive tasks and one exclusive task.
3600 *
3601 * There are circumstances in which we can try to wake a task which has already
3602 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3603 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3604 */
3605static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3606 int nr_exclusive, int sync, void *key)
3607{
3608 struct list_head *tmp, *next;
3609
3610 list_for_each_safe(tmp, next, &q->task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07003611 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3612 unsigned flags = curr->flags;
3613
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614 if (curr->func(curr, mode, sync, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07003615 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616 break;
3617 }
3618}
3619
3620/**
3621 * __wake_up - wake up threads blocked on a waitqueue.
3622 * @q: the waitqueue
3623 * @mode: which threads
3624 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07003625 * @key: is directly passed to the wakeup function
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626 */
3627void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003628 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629{
3630 unsigned long flags;
3631
3632 spin_lock_irqsave(&q->lock, flags);
3633 __wake_up_common(q, mode, nr_exclusive, 0, key);
3634 spin_unlock_irqrestore(&q->lock, flags);
3635}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636EXPORT_SYMBOL(__wake_up);
3637
3638/*
3639 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3640 */
3641void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3642{
3643 __wake_up_common(q, mode, 1, 0, NULL);
3644}
3645
3646/**
Martin Waitz67be2dd2005-05-01 08:59:26 -07003647 * __wake_up_sync - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648 * @q: the waitqueue
3649 * @mode: which threads
3650 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3651 *
3652 * The sync wakeup differs that the waker knows that it will schedule
3653 * away soon, so while the target thread will be woken up, it will not
3654 * be migrated to another CPU - ie. the two threads are 'synchronized'
3655 * with each other. This can prevent needless bouncing between CPUs.
3656 *
3657 * On UP it can prevent extra preemption.
3658 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003659void fastcall
3660__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003661{
3662 unsigned long flags;
3663 int sync = 1;
3664
3665 if (unlikely(!q))
3666 return;
3667
3668 if (unlikely(!nr_exclusive))
3669 sync = 0;
3670
3671 spin_lock_irqsave(&q->lock, flags);
3672 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3673 spin_unlock_irqrestore(&q->lock, flags);
3674}
3675EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3676
3677void fastcall complete(struct completion *x)
3678{
3679 unsigned long flags;
3680
3681 spin_lock_irqsave(&x->wait.lock, flags);
3682 x->done++;
3683 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3684 1, 0, NULL);
3685 spin_unlock_irqrestore(&x->wait.lock, flags);
3686}
3687EXPORT_SYMBOL(complete);
3688
3689void fastcall complete_all(struct completion *x)
3690{
3691 unsigned long flags;
3692
3693 spin_lock_irqsave(&x->wait.lock, flags);
3694 x->done += UINT_MAX/2;
3695 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3696 0, 0, NULL);
3697 spin_unlock_irqrestore(&x->wait.lock, flags);
3698}
3699EXPORT_SYMBOL(complete_all);
3700
3701void fastcall __sched wait_for_completion(struct completion *x)
3702{
3703 might_sleep();
Ingo Molnar48f24c42006-07-03 00:25:40 -07003704
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 spin_lock_irq(&x->wait.lock);
3706 if (!x->done) {
3707 DECLARE_WAITQUEUE(wait, current);
3708
3709 wait.flags |= WQ_FLAG_EXCLUSIVE;
3710 __add_wait_queue_tail(&x->wait, &wait);
3711 do {
3712 __set_current_state(TASK_UNINTERRUPTIBLE);
3713 spin_unlock_irq(&x->wait.lock);
3714 schedule();
3715 spin_lock_irq(&x->wait.lock);
3716 } while (!x->done);
3717 __remove_wait_queue(&x->wait, &wait);
3718 }
3719 x->done--;
3720 spin_unlock_irq(&x->wait.lock);
3721}
3722EXPORT_SYMBOL(wait_for_completion);
3723
3724unsigned long fastcall __sched
3725wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3726{
3727 might_sleep();
3728
3729 spin_lock_irq(&x->wait.lock);
3730 if (!x->done) {
3731 DECLARE_WAITQUEUE(wait, current);
3732
3733 wait.flags |= WQ_FLAG_EXCLUSIVE;
3734 __add_wait_queue_tail(&x->wait, &wait);
3735 do {
3736 __set_current_state(TASK_UNINTERRUPTIBLE);
3737 spin_unlock_irq(&x->wait.lock);
3738 timeout = schedule_timeout(timeout);
3739 spin_lock_irq(&x->wait.lock);
3740 if (!timeout) {
3741 __remove_wait_queue(&x->wait, &wait);
3742 goto out;
3743 }
3744 } while (!x->done);
3745 __remove_wait_queue(&x->wait, &wait);
3746 }
3747 x->done--;
3748out:
3749 spin_unlock_irq(&x->wait.lock);
3750 return timeout;
3751}
3752EXPORT_SYMBOL(wait_for_completion_timeout);
3753
3754int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3755{
3756 int ret = 0;
3757
3758 might_sleep();
3759
3760 spin_lock_irq(&x->wait.lock);
3761 if (!x->done) {
3762 DECLARE_WAITQUEUE(wait, current);
3763
3764 wait.flags |= WQ_FLAG_EXCLUSIVE;
3765 __add_wait_queue_tail(&x->wait, &wait);
3766 do {
3767 if (signal_pending(current)) {
3768 ret = -ERESTARTSYS;
3769 __remove_wait_queue(&x->wait, &wait);
3770 goto out;
3771 }
3772 __set_current_state(TASK_INTERRUPTIBLE);
3773 spin_unlock_irq(&x->wait.lock);
3774 schedule();
3775 spin_lock_irq(&x->wait.lock);
3776 } while (!x->done);
3777 __remove_wait_queue(&x->wait, &wait);
3778 }
3779 x->done--;
3780out:
3781 spin_unlock_irq(&x->wait.lock);
3782
3783 return ret;
3784}
3785EXPORT_SYMBOL(wait_for_completion_interruptible);
3786
3787unsigned long fastcall __sched
3788wait_for_completion_interruptible_timeout(struct completion *x,
3789 unsigned long timeout)
3790{
3791 might_sleep();
3792
3793 spin_lock_irq(&x->wait.lock);
3794 if (!x->done) {
3795 DECLARE_WAITQUEUE(wait, current);
3796
3797 wait.flags |= WQ_FLAG_EXCLUSIVE;
3798 __add_wait_queue_tail(&x->wait, &wait);
3799 do {
3800 if (signal_pending(current)) {
3801 timeout = -ERESTARTSYS;
3802 __remove_wait_queue(&x->wait, &wait);
3803 goto out;
3804 }
3805 __set_current_state(TASK_INTERRUPTIBLE);
3806 spin_unlock_irq(&x->wait.lock);
3807 timeout = schedule_timeout(timeout);
3808 spin_lock_irq(&x->wait.lock);
3809 if (!timeout) {
3810 __remove_wait_queue(&x->wait, &wait);
3811 goto out;
3812 }
3813 } while (!x->done);
3814 __remove_wait_queue(&x->wait, &wait);
3815 }
3816 x->done--;
3817out:
3818 spin_unlock_irq(&x->wait.lock);
3819 return timeout;
3820}
3821EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3822
Ingo Molnar0fec1712007-07-09 18:52:01 +02003823static inline void
3824sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825{
Ingo Molnar0fec1712007-07-09 18:52:01 +02003826 spin_lock_irqsave(&q->lock, *flags);
3827 __add_wait_queue(q, wait);
3828 spin_unlock(&q->lock);
3829}
3830
3831static inline void
3832sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3833{
3834 spin_lock_irq(&q->lock);
3835 __remove_wait_queue(q, wait);
3836 spin_unlock_irqrestore(&q->lock, *flags);
3837}
3838
3839void __sched interruptible_sleep_on(wait_queue_head_t *q)
3840{
3841 unsigned long flags;
3842 wait_queue_t wait;
3843
3844 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845
3846 current->state = TASK_INTERRUPTIBLE;
3847
Ingo Molnar0fec1712007-07-09 18:52:01 +02003848 sleep_on_head(q, &wait, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003849 schedule();
Ingo Molnar0fec1712007-07-09 18:52:01 +02003850 sleep_on_tail(q, &wait, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852EXPORT_SYMBOL(interruptible_sleep_on);
3853
Ingo Molnar0fec1712007-07-09 18:52:01 +02003854long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003855interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856{
Ingo Molnar0fec1712007-07-09 18:52:01 +02003857 unsigned long flags;
3858 wait_queue_t wait;
3859
3860 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861
3862 current->state = TASK_INTERRUPTIBLE;
3863
Ingo Molnar0fec1712007-07-09 18:52:01 +02003864 sleep_on_head(q, &wait, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003865 timeout = schedule_timeout(timeout);
Ingo Molnar0fec1712007-07-09 18:52:01 +02003866 sleep_on_tail(q, &wait, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867
3868 return timeout;
3869}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3871
Ingo Molnar0fec1712007-07-09 18:52:01 +02003872void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873{
Ingo Molnar0fec1712007-07-09 18:52:01 +02003874 unsigned long flags;
3875 wait_queue_t wait;
3876
3877 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878
3879 current->state = TASK_UNINTERRUPTIBLE;
3880
Ingo Molnar0fec1712007-07-09 18:52:01 +02003881 sleep_on_head(q, &wait, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003882 schedule();
Ingo Molnar0fec1712007-07-09 18:52:01 +02003883 sleep_on_tail(q, &wait, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885EXPORT_SYMBOL(sleep_on);
3886
Ingo Molnar0fec1712007-07-09 18:52:01 +02003887long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888{
Ingo Molnar0fec1712007-07-09 18:52:01 +02003889 unsigned long flags;
3890 wait_queue_t wait;
3891
3892 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893
3894 current->state = TASK_UNINTERRUPTIBLE;
3895
Ingo Molnar0fec1712007-07-09 18:52:01 +02003896 sleep_on_head(q, &wait, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897 timeout = schedule_timeout(timeout);
Ingo Molnar0fec1712007-07-09 18:52:01 +02003898 sleep_on_tail(q, &wait, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899
3900 return timeout;
3901}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902EXPORT_SYMBOL(sleep_on_timeout);
3903
Ingo Molnarb29739f2006-06-27 02:54:51 -07003904#ifdef CONFIG_RT_MUTEXES
3905
3906/*
3907 * rt_mutex_setprio - set the current priority of a task
3908 * @p: task
3909 * @prio: prio value (kernel-internal form)
3910 *
3911 * This function changes the 'effective' priority of a task. It does
3912 * not touch ->normal_prio like __setscheduler().
3913 *
3914 * Used by the rt_mutex code to implement priority inheritance logic.
3915 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003916void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07003917{
3918 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02003919 int oldprio, on_rq;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003920 struct rq *rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -07003921
3922 BUG_ON(prio < 0 || prio > MAX_PRIO);
3923
3924 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02003925 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07003926
Andrew Mortond5f9f942007-05-08 20:27:06 -07003927 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02003928 on_rq = p->se.on_rq;
3929 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02003930 dequeue_task(rq, p, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02003931
3932 if (rt_prio(prio))
3933 p->sched_class = &rt_sched_class;
3934 else
3935 p->sched_class = &fair_sched_class;
3936
Ingo Molnarb29739f2006-06-27 02:54:51 -07003937 p->prio = prio;
3938
Ingo Molnardd41f592007-07-09 18:51:59 +02003939 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02003940 enqueue_task(rq, p, 0);
Ingo Molnarb29739f2006-06-27 02:54:51 -07003941 /*
3942 * Reschedule if we are currently running on this runqueue and
Andrew Mortond5f9f942007-05-08 20:27:06 -07003943 * our priority decreased, or if we are not currently running on
3944 * this runqueue and our priority is higher than the current's
Ingo Molnarb29739f2006-06-27 02:54:51 -07003945 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07003946 if (task_running(rq, p)) {
3947 if (p->prio > oldprio)
3948 resched_task(rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02003949 } else {
3950 check_preempt_curr(rq, p);
3951 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07003952 }
3953 task_rq_unlock(rq, &flags);
3954}
3955
3956#endif
3957
Ingo Molnar36c8b582006-07-03 00:25:41 -07003958void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959{
Ingo Molnardd41f592007-07-09 18:51:59 +02003960 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003962 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003963
3964 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3965 return;
3966 /*
3967 * We have to be careful, if called from sys_setpriority(),
3968 * the task might be in the middle of scheduling on another CPU.
3969 */
3970 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02003971 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 /*
3973 * The RT priorities are set via sched_setscheduler(), but we still
3974 * allow the 'normal' nice value to be set - but as expected
3975 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02003976 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977 */
Ingo Molnare05606d2007-07-09 18:51:59 +02003978 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979 p->static_prio = NICE_TO_PRIO(nice);
3980 goto out_unlock;
3981 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003982 on_rq = p->se.on_rq;
3983 if (on_rq) {
Ingo Molnar69be72c2007-08-09 11:16:49 +02003984 dequeue_task(rq, p, 0);
Ingo Molnar79b5ddd2007-08-09 11:16:49 +02003985 dec_load(rq, p);
Peter Williams2dd73a42006-06-27 02:54:34 -07003986 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07003989 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07003990 old_prio = p->prio;
3991 p->prio = effective_prio(p);
3992 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993
Ingo Molnardd41f592007-07-09 18:51:59 +02003994 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02003995 enqueue_task(rq, p, 0);
Ingo Molnar29b4b622007-08-09 11:16:49 +02003996 inc_load(rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07003998 * If the task increased its priority or is running and
3999 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07004001 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002 resched_task(rq->curr);
4003 }
4004out_unlock:
4005 task_rq_unlock(rq, &flags);
4006}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007EXPORT_SYMBOL(set_user_nice);
4008
Matt Mackalle43379f2005-05-01 08:59:00 -07004009/*
4010 * can_nice - check if a task can reduce its nice value
4011 * @p: task
4012 * @nice: nice value
4013 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004014int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07004015{
Matt Mackall024f4742005-08-18 11:24:19 -07004016 /* convert nice value [19,-20] to rlimit style value [1,40] */
4017 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004018
Matt Mackalle43379f2005-05-01 08:59:00 -07004019 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4020 capable(CAP_SYS_NICE));
4021}
4022
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023#ifdef __ARCH_WANT_SYS_NICE
4024
4025/*
4026 * sys_nice - change the priority of the current process.
4027 * @increment: priority increment
4028 *
4029 * sys_setpriority is a more generic, but much slower function that
4030 * does similar things.
4031 */
4032asmlinkage long sys_nice(int increment)
4033{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004034 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035
4036 /*
4037 * Setpriority might change our priority at the same moment.
4038 * We don't have to worry. Conceptually one call occurs first
4039 * and we have a single winner.
4040 */
Matt Mackalle43379f2005-05-01 08:59:00 -07004041 if (increment < -40)
4042 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043 if (increment > 40)
4044 increment = 40;
4045
4046 nice = PRIO_TO_NICE(current->static_prio) + increment;
4047 if (nice < -20)
4048 nice = -20;
4049 if (nice > 19)
4050 nice = 19;
4051
Matt Mackalle43379f2005-05-01 08:59:00 -07004052 if (increment < 0 && !can_nice(current, nice))
4053 return -EPERM;
4054
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 retval = security_task_setnice(current, nice);
4056 if (retval)
4057 return retval;
4058
4059 set_user_nice(current, nice);
4060 return 0;
4061}
4062
4063#endif
4064
4065/**
4066 * task_prio - return the priority value of a given task.
4067 * @p: the task in question.
4068 *
4069 * This is the priority value as seen by users in /proc.
4070 * RT tasks are offset by -200. Normal tasks are centered
4071 * around 0, value goes from -16 to +15.
4072 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004073int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074{
4075 return p->prio - MAX_RT_PRIO;
4076}
4077
4078/**
4079 * task_nice - return the nice value of a given task.
4080 * @p: the task in question.
4081 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004082int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083{
4084 return TASK_NICE(p);
4085}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086EXPORT_SYMBOL_GPL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087
4088/**
4089 * idle_cpu - is a given cpu idle currently?
4090 * @cpu: the processor in question.
4091 */
4092int idle_cpu(int cpu)
4093{
4094 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4095}
4096
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097/**
4098 * idle_task - return the idle task for a given cpu.
4099 * @cpu: the processor in question.
4100 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004101struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102{
4103 return cpu_rq(cpu)->idle;
4104}
4105
4106/**
4107 * find_process_by_pid - find a process with a matching PID value.
4108 * @pid: the pid in question.
4109 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004110static inline struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111{
4112 return pid ? find_task_by_pid(pid) : current;
4113}
4114
4115/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02004116static void
4117__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004118{
Ingo Molnardd41f592007-07-09 18:51:59 +02004119 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004120
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121 p->policy = policy;
Ingo Molnardd41f592007-07-09 18:51:59 +02004122 switch (p->policy) {
4123 case SCHED_NORMAL:
4124 case SCHED_BATCH:
4125 case SCHED_IDLE:
4126 p->sched_class = &fair_sched_class;
4127 break;
4128 case SCHED_FIFO:
4129 case SCHED_RR:
4130 p->sched_class = &rt_sched_class;
4131 break;
4132 }
4133
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004135 p->normal_prio = normal_prio(p);
4136 /* we are holding p->pi_lock already */
4137 p->prio = rt_mutex_getprio(p);
Peter Williams2dd73a42006-06-27 02:54:34 -07004138 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004139}
4140
4141/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004142 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143 * @p: the task in question.
4144 * @policy: new policy.
4145 * @param: structure containing the new RT priority.
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004146 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004147 * NOTE that the task may be already dead.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004149int sched_setscheduler(struct task_struct *p, int policy,
4150 struct sched_param *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004151{
Ingo Molnardd41f592007-07-09 18:51:59 +02004152 int retval, oldprio, oldpolicy = -1, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004154 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155
Steven Rostedt66e53932006-06-27 02:54:44 -07004156 /* may grab non-irq protected spin_locks */
4157 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158recheck:
4159 /* double check policy once rq lock held */
4160 if (policy < 0)
4161 policy = oldpolicy = p->policy;
4162 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
Ingo Molnardd41f592007-07-09 18:51:59 +02004163 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4164 policy != SCHED_IDLE)
Ingo Molnarb0a94992006-01-14 13:20:41 -08004165 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004166 /*
4167 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02004168 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4169 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170 */
4171 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004172 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04004173 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004174 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02004175 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004176 return -EINVAL;
4177
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004178 /*
4179 * Allow unprivileged RT tasks to decrease priority:
4180 */
4181 if (!capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02004182 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004183 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004184
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004185 if (!lock_task_sighand(p, &flags))
4186 return -ESRCH;
4187 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4188 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004189
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004190 /* can't set/change the rt policy */
4191 if (policy != p->policy && !rlim_rtprio)
4192 return -EPERM;
4193
4194 /* can't increase priority */
4195 if (param->sched_priority > p->rt_priority &&
4196 param->sched_priority > rlim_rtprio)
4197 return -EPERM;
4198 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004199 /*
4200 * Like positive nice levels, dont allow tasks to
4201 * move out of SCHED_IDLE either:
4202 */
4203 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4204 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07004205
Olivier Croquette37e4ab32005-06-25 14:57:32 -07004206 /* can't change other user's priorities */
4207 if ((current->euid != p->euid) &&
4208 (current->euid != p->uid))
4209 return -EPERM;
4210 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211
4212 retval = security_task_setscheduler(p, policy, param);
4213 if (retval)
4214 return retval;
4215 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07004216 * make sure no PI-waiters arrive (or leave) while we are
4217 * changing the priority of the task:
4218 */
4219 spin_lock_irqsave(&p->pi_lock, flags);
4220 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221 * To be able to change p->policy safely, the apropriate
4222 * runqueue lock must be held.
4223 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07004224 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225 /* recheck policy now with rq lock held */
4226 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4227 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004228 __task_rq_unlock(rq);
4229 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230 goto recheck;
4231 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02004232 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004233 on_rq = p->se.on_rq;
Ingo Molnar2daa3572007-08-09 11:16:51 +02004234 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004235 deactivate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004237 __setscheduler(rq, p, policy, param->sched_priority);
4238 if (on_rq) {
4239 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240 /*
4241 * Reschedule if we are currently running on this runqueue and
Andrew Mortond5f9f942007-05-08 20:27:06 -07004242 * our priority decreased, or if we are not currently running on
4243 * this runqueue and our priority is higher than the current's
Linus Torvalds1da177e2005-04-16 15:20:36 -07004244 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07004245 if (task_running(rq, p)) {
4246 if (p->prio > oldprio)
4247 resched_task(rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02004248 } else {
4249 check_preempt_curr(rq, p);
4250 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07004252 __task_rq_unlock(rq);
4253 spin_unlock_irqrestore(&p->pi_lock, flags);
4254
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07004255 rt_mutex_adjust_pi(p);
4256
Linus Torvalds1da177e2005-04-16 15:20:36 -07004257 return 0;
4258}
4259EXPORT_SYMBOL_GPL(sched_setscheduler);
4260
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004261static int
4262do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004263{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264 struct sched_param lparam;
4265 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004266 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267
4268 if (!param || pid < 0)
4269 return -EINVAL;
4270 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4271 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004272
4273 rcu_read_lock();
4274 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004275 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004276 if (p != NULL)
4277 retval = sched_setscheduler(p, policy, &lparam);
4278 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07004279
Linus Torvalds1da177e2005-04-16 15:20:36 -07004280 return retval;
4281}
4282
4283/**
4284 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4285 * @pid: the pid in question.
4286 * @policy: new policy.
4287 * @param: structure containing the new RT priority.
4288 */
4289asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4290 struct sched_param __user *param)
4291{
Jason Baronc21761f2006-01-18 17:43:03 -08004292 /* negative values for policy are not valid */
4293 if (policy < 0)
4294 return -EINVAL;
4295
Linus Torvalds1da177e2005-04-16 15:20:36 -07004296 return do_sched_setscheduler(pid, policy, param);
4297}
4298
4299/**
4300 * sys_sched_setparam - set/change the RT priority of a thread
4301 * @pid: the pid in question.
4302 * @param: structure containing the new RT priority.
4303 */
4304asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4305{
4306 return do_sched_setscheduler(pid, -1, param);
4307}
4308
4309/**
4310 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4311 * @pid: the pid in question.
4312 */
4313asmlinkage long sys_sched_getscheduler(pid_t pid)
4314{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004315 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316 int retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004317
4318 if (pid < 0)
4319 goto out_nounlock;
4320
4321 retval = -ESRCH;
4322 read_lock(&tasklist_lock);
4323 p = find_process_by_pid(pid);
4324 if (p) {
4325 retval = security_task_getscheduler(p);
4326 if (!retval)
4327 retval = p->policy;
4328 }
4329 read_unlock(&tasklist_lock);
4330
4331out_nounlock:
4332 return retval;
4333}
4334
4335/**
4336 * sys_sched_getscheduler - get the RT priority of a thread
4337 * @pid: the pid in question.
4338 * @param: structure containing the RT priority.
4339 */
4340asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4341{
4342 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004343 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004344 int retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004345
4346 if (!param || pid < 0)
4347 goto out_nounlock;
4348
4349 read_lock(&tasklist_lock);
4350 p = find_process_by_pid(pid);
4351 retval = -ESRCH;
4352 if (!p)
4353 goto out_unlock;
4354
4355 retval = security_task_getscheduler(p);
4356 if (retval)
4357 goto out_unlock;
4358
4359 lp.sched_priority = p->rt_priority;
4360 read_unlock(&tasklist_lock);
4361
4362 /*
4363 * This one might sleep, we cannot do it with a spinlock held ...
4364 */
4365 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4366
4367out_nounlock:
4368 return retval;
4369
4370out_unlock:
4371 read_unlock(&tasklist_lock);
4372 return retval;
4373}
4374
4375long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4376{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004377 cpumask_t cpus_allowed;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004378 struct task_struct *p;
4379 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004380
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004381 mutex_lock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004382 read_lock(&tasklist_lock);
4383
4384 p = find_process_by_pid(pid);
4385 if (!p) {
4386 read_unlock(&tasklist_lock);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004387 mutex_unlock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004388 return -ESRCH;
4389 }
4390
4391 /*
4392 * It is not safe to call set_cpus_allowed with the
4393 * tasklist_lock held. We will bump the task_struct's
4394 * usage count and then drop tasklist_lock.
4395 */
4396 get_task_struct(p);
4397 read_unlock(&tasklist_lock);
4398
4399 retval = -EPERM;
4400 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4401 !capable(CAP_SYS_NICE))
4402 goto out_unlock;
4403
David Quigleye7834f82006-06-23 02:03:59 -07004404 retval = security_task_setscheduler(p, 0, NULL);
4405 if (retval)
4406 goto out_unlock;
4407
Linus Torvalds1da177e2005-04-16 15:20:36 -07004408 cpus_allowed = cpuset_cpus_allowed(p);
4409 cpus_and(new_mask, new_mask, cpus_allowed);
4410 retval = set_cpus_allowed(p, new_mask);
4411
4412out_unlock:
4413 put_task_struct(p);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004414 mutex_unlock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004415 return retval;
4416}
4417
4418static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4419 cpumask_t *new_mask)
4420{
4421 if (len < sizeof(cpumask_t)) {
4422 memset(new_mask, 0, sizeof(cpumask_t));
4423 } else if (len > sizeof(cpumask_t)) {
4424 len = sizeof(cpumask_t);
4425 }
4426 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4427}
4428
4429/**
4430 * sys_sched_setaffinity - set the cpu affinity of a process
4431 * @pid: pid of the process
4432 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4433 * @user_mask_ptr: user-space pointer to the new cpu mask
4434 */
4435asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4436 unsigned long __user *user_mask_ptr)
4437{
4438 cpumask_t new_mask;
4439 int retval;
4440
4441 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4442 if (retval)
4443 return retval;
4444
4445 return sched_setaffinity(pid, new_mask);
4446}
4447
4448/*
4449 * Represents all cpu's present in the system
4450 * In systems capable of hotplug, this map could dynamically grow
4451 * as new cpu's are detected in the system via any platform specific
4452 * method, such as ACPI for e.g.
4453 */
4454
Andi Kleen4cef0c62006-01-11 22:44:57 +01004455cpumask_t cpu_present_map __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004456EXPORT_SYMBOL(cpu_present_map);
4457
4458#ifndef CONFIG_SMP
Andi Kleen4cef0c62006-01-11 22:44:57 +01004459cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07004460EXPORT_SYMBOL(cpu_online_map);
4461
Andi Kleen4cef0c62006-01-11 22:44:57 +01004462cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
Greg Bankse16b38f2006-10-02 02:17:40 -07004463EXPORT_SYMBOL(cpu_possible_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004464#endif
4465
4466long sched_getaffinity(pid_t pid, cpumask_t *mask)
4467{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004468 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004469 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004470
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004471 mutex_lock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004472 read_lock(&tasklist_lock);
4473
4474 retval = -ESRCH;
4475 p = find_process_by_pid(pid);
4476 if (!p)
4477 goto out_unlock;
4478
David Quigleye7834f82006-06-23 02:03:59 -07004479 retval = security_task_getscheduler(p);
4480 if (retval)
4481 goto out_unlock;
4482
Jack Steiner2f7016d2006-02-01 03:05:18 -08004483 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004484
4485out_unlock:
4486 read_unlock(&tasklist_lock);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07004487 mutex_unlock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004488
Ulrich Drepper9531b622007-08-09 11:16:46 +02004489 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004490}
4491
4492/**
4493 * sys_sched_getaffinity - get the cpu affinity of a process
4494 * @pid: pid of the process
4495 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4496 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4497 */
4498asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4499 unsigned long __user *user_mask_ptr)
4500{
4501 int ret;
4502 cpumask_t mask;
4503
4504 if (len < sizeof(cpumask_t))
4505 return -EINVAL;
4506
4507 ret = sched_getaffinity(pid, &mask);
4508 if (ret < 0)
4509 return ret;
4510
4511 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4512 return -EFAULT;
4513
4514 return sizeof(cpumask_t);
4515}
4516
4517/**
4518 * sys_sched_yield - yield the current processor to other threads.
4519 *
Ingo Molnardd41f592007-07-09 18:51:59 +02004520 * This function yields the current CPU to other tasks. If there are no
4521 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004522 */
4523asmlinkage long sys_sched_yield(void)
4524{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004525 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004526
4527 schedstat_inc(rq, yld_cnt);
Ingo Molnardd41f592007-07-09 18:51:59 +02004528 if (unlikely(rq->nr_running == 1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004529 schedstat_inc(rq, yld_act_empty);
Ingo Molnardd41f592007-07-09 18:51:59 +02004530 else
4531 current->sched_class->yield_task(rq, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004532
4533 /*
4534 * Since we are going to call schedule() anyway, there's
4535 * no need to preempt or enable interrupts:
4536 */
4537 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07004538 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004539 _raw_spin_unlock(&rq->lock);
4540 preempt_enable_no_resched();
4541
4542 schedule();
4543
4544 return 0;
4545}
4546
Andrew Mortone7b38402006-06-30 01:56:00 -07004547static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004548{
Ingo Molnar8e0a43d2006-06-23 02:05:23 -07004549#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4550 __might_sleep(__FILE__, __LINE__);
4551#endif
Ingo Molnar5bbcfd92005-07-07 17:57:04 -07004552 /*
4553 * The BKS might be reacquired before we have dropped
4554 * PREEMPT_ACTIVE, which could trigger a second
4555 * cond_resched() call.
4556 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004557 do {
4558 add_preempt_count(PREEMPT_ACTIVE);
4559 schedule();
4560 sub_preempt_count(PREEMPT_ACTIVE);
4561 } while (need_resched());
4562}
4563
4564int __sched cond_resched(void)
4565{
Ingo Molnar94142322006-12-29 16:48:13 -08004566 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4567 system_state == SYSTEM_RUNNING) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004568 __cond_resched();
4569 return 1;
4570 }
4571 return 0;
4572}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004573EXPORT_SYMBOL(cond_resched);
4574
4575/*
4576 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4577 * call schedule, and on return reacquire the lock.
4578 *
4579 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4580 * operations here to prevent schedule() from being called twice (once via
4581 * spin_unlock(), once by hand).
4582 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004583int cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004584{
Jan Kara6df3cec2005-06-13 15:52:32 -07004585 int ret = 0;
4586
Linus Torvalds1da177e2005-04-16 15:20:36 -07004587 if (need_lockbreak(lock)) {
4588 spin_unlock(lock);
4589 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07004590 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004591 spin_lock(lock);
4592 }
Ingo Molnar94142322006-12-29 16:48:13 -08004593 if (need_resched() && system_state == SYSTEM_RUNNING) {
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07004594 spin_release(&lock->dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004595 _raw_spin_unlock(lock);
4596 preempt_enable_no_resched();
4597 __cond_resched();
Jan Kara6df3cec2005-06-13 15:52:32 -07004598 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004599 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004600 }
Jan Kara6df3cec2005-06-13 15:52:32 -07004601 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004602}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004603EXPORT_SYMBOL(cond_resched_lock);
4604
4605int __sched cond_resched_softirq(void)
4606{
4607 BUG_ON(!in_softirq());
4608
Ingo Molnar94142322006-12-29 16:48:13 -08004609 if (need_resched() && system_state == SYSTEM_RUNNING) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07004610 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004611 __cond_resched();
4612 local_bh_disable();
4613 return 1;
4614 }
4615 return 0;
4616}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004617EXPORT_SYMBOL(cond_resched_softirq);
4618
Linus Torvalds1da177e2005-04-16 15:20:36 -07004619/**
4620 * yield - yield the current processor to other threads.
4621 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08004622 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07004623 * thread runnable and calls sys_sched_yield().
4624 */
4625void __sched yield(void)
4626{
4627 set_current_state(TASK_RUNNING);
4628 sys_sched_yield();
4629}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004630EXPORT_SYMBOL(yield);
4631
4632/*
4633 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4634 * that process accounting knows that this is a task in IO wait state.
4635 *
4636 * But don't do that if it is a deliberate, throttling IO wait (this task
4637 * has set its backing_dev_info: the queue against which it should throttle)
4638 */
4639void __sched io_schedule(void)
4640{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004641 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004642
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004643 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004644 atomic_inc(&rq->nr_iowait);
4645 schedule();
4646 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004647 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004648}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004649EXPORT_SYMBOL(io_schedule);
4650
4651long __sched io_schedule_timeout(long timeout)
4652{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004653 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004654 long ret;
4655
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004656 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004657 atomic_inc(&rq->nr_iowait);
4658 ret = schedule_timeout(timeout);
4659 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004660 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004661 return ret;
4662}
4663
4664/**
4665 * sys_sched_get_priority_max - return maximum RT priority.
4666 * @policy: scheduling class.
4667 *
4668 * this syscall returns the maximum rt_priority that can be used
4669 * by a given scheduling class.
4670 */
4671asmlinkage long sys_sched_get_priority_max(int policy)
4672{
4673 int ret = -EINVAL;
4674
4675 switch (policy) {
4676 case SCHED_FIFO:
4677 case SCHED_RR:
4678 ret = MAX_USER_RT_PRIO-1;
4679 break;
4680 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08004681 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02004682 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004683 ret = 0;
4684 break;
4685 }
4686 return ret;
4687}
4688
4689/**
4690 * sys_sched_get_priority_min - return minimum RT priority.
4691 * @policy: scheduling class.
4692 *
4693 * this syscall returns the minimum rt_priority that can be used
4694 * by a given scheduling class.
4695 */
4696asmlinkage long sys_sched_get_priority_min(int policy)
4697{
4698 int ret = -EINVAL;
4699
4700 switch (policy) {
4701 case SCHED_FIFO:
4702 case SCHED_RR:
4703 ret = 1;
4704 break;
4705 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08004706 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02004707 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004708 ret = 0;
4709 }
4710 return ret;
4711}
4712
4713/**
4714 * sys_sched_rr_get_interval - return the default timeslice of a process.
4715 * @pid: pid of the process.
4716 * @interval: userspace pointer to the timeslice value.
4717 *
4718 * this syscall writes the default timeslice value of a given process
4719 * into the user-space timespec buffer. A value of '0' means infinity.
4720 */
4721asmlinkage
4722long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4723{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004724 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004725 int retval = -EINVAL;
4726 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004727
4728 if (pid < 0)
4729 goto out_nounlock;
4730
4731 retval = -ESRCH;
4732 read_lock(&tasklist_lock);
4733 p = find_process_by_pid(pid);
4734 if (!p)
4735 goto out_unlock;
4736
4737 retval = security_task_getscheduler(p);
4738 if (retval)
4739 goto out_unlock;
4740
Peter Williamsb78709c2006-06-26 16:58:00 +10004741 jiffies_to_timespec(p->policy == SCHED_FIFO ?
Ingo Molnardd41f592007-07-09 18:51:59 +02004742 0 : static_prio_timeslice(p->static_prio), &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004743 read_unlock(&tasklist_lock);
4744 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4745out_nounlock:
4746 return retval;
4747out_unlock:
4748 read_unlock(&tasklist_lock);
4749 return retval;
4750}
4751
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004752static const char stat_nam[] = "RSDTtZX";
Ingo Molnar36c8b582006-07-03 00:25:41 -07004753
4754static void show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004755{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004756 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004757 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004758
Linus Torvalds1da177e2005-04-16 15:20:36 -07004759 state = p->state ? __ffs(p->state) + 1 : 0;
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004760 printk("%-13.13s %c", p->comm,
4761 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02004762#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07004763 if (state == TASK_RUNNING)
Ingo Molnar4bd77322007-07-11 21:21:47 +02004764 printk(" running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004765 else
Ingo Molnar4bd77322007-07-11 21:21:47 +02004766 printk(" %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004767#else
4768 if (state == TASK_RUNNING)
Ingo Molnar4bd77322007-07-11 21:21:47 +02004769 printk(" running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004770 else
4771 printk(" %016lx ", thread_saved_pc(p));
4772#endif
4773#ifdef CONFIG_DEBUG_STACK_USAGE
4774 {
Al Viro10ebffd2005-11-13 16:06:56 -08004775 unsigned long *n = end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004776 while (!*n)
4777 n++;
Al Viro10ebffd2005-11-13 16:06:56 -08004778 free = (unsigned long)n - (unsigned long)end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004779 }
4780#endif
Ingo Molnar4bd77322007-07-11 21:21:47 +02004781 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004782
4783 if (state != TASK_RUNNING)
4784 show_stack(p, NULL);
4785}
4786
Ingo Molnare59e2ae2006-12-06 20:35:59 -08004787void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004788{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004789 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004790
Ingo Molnar4bd77322007-07-11 21:21:47 +02004791#if BITS_PER_LONG == 32
4792 printk(KERN_INFO
4793 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004794#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02004795 printk(KERN_INFO
4796 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004797#endif
4798 read_lock(&tasklist_lock);
4799 do_each_thread(g, p) {
4800 /*
4801 * reset the NMI-timeout, listing all files on a slow
4802 * console might take alot of time:
4803 */
4804 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07004805 if (!state_filter || (p->state & state_filter))
Ingo Molnare59e2ae2006-12-06 20:35:59 -08004806 show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004807 } while_each_thread(g, p);
4808
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07004809 touch_all_softlockup_watchdogs();
4810
Ingo Molnardd41f592007-07-09 18:51:59 +02004811#ifdef CONFIG_SCHED_DEBUG
4812 sysrq_sched_debug_show();
4813#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004814 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08004815 /*
4816 * Only show locks if all tasks are dumped:
4817 */
4818 if (state_filter == -1)
4819 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004820}
4821
Ingo Molnar1df21052007-07-09 18:51:58 +02004822void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4823{
Ingo Molnardd41f592007-07-09 18:51:59 +02004824 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02004825}
4826
Ingo Molnarf340c0d2005-06-28 16:40:42 +02004827/**
4828 * init_idle - set up an idle thread for a given CPU
4829 * @idle: task in question
4830 * @cpu: cpu the idle task belongs to
4831 *
4832 * NOTE: this function does not set the idle thread's NEED_RESCHED
4833 * flag, to make booting more robust.
4834 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07004835void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004836{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004837 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004838 unsigned long flags;
4839
Ingo Molnardd41f592007-07-09 18:51:59 +02004840 __sched_fork(idle);
4841 idle->se.exec_start = sched_clock();
4842
Ingo Molnarb29739f2006-06-27 02:54:51 -07004843 idle->prio = idle->normal_prio = MAX_PRIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004844 idle->cpus_allowed = cpumask_of_cpu(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004845 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004846
4847 spin_lock_irqsave(&rq->lock, flags);
4848 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07004849#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4850 idle->oncpu = 1;
4851#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004852 spin_unlock_irqrestore(&rq->lock, flags);
4853
4854 /* Set the preempt count _outside_ the spinlocks! */
4855#if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
Al Viroa1261f52005-11-13 16:06:55 -08004856 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004857#else
Al Viroa1261f52005-11-13 16:06:55 -08004858 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004859#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02004860 /*
4861 * The idle tasks have their own, simple scheduling class:
4862 */
4863 idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004864}
4865
4866/*
4867 * In a system that switches off the HZ timer nohz_cpu_mask
4868 * indicates which cpus entered this state. This is used
4869 * in the rcu update to wait only for active cpus. For system
4870 * which do not switch off the HZ timer nohz_cpu_mask should
4871 * always be CPU_MASK_NONE.
4872 */
4873cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4874
Ingo Molnardd41f592007-07-09 18:51:59 +02004875/*
4876 * Increase the granularity value when there are more CPUs,
4877 * because with more CPUs the 'effective latency' as visible
4878 * to users decreases. But the relationship is not linear,
4879 * so pick a second-best guess by going with the log2 of the
4880 * number of CPUs.
4881 *
4882 * This idea comes from the SD scheduler of Con Kolivas:
4883 */
4884static inline void sched_init_granularity(void)
4885{
4886 unsigned int factor = 1 + ilog2(num_online_cpus());
Ingo Molnara5968df2007-07-11 21:21:47 +02004887 const unsigned long gran_limit = 100000000;
Ingo Molnardd41f592007-07-09 18:51:59 +02004888
4889 sysctl_sched_granularity *= factor;
4890 if (sysctl_sched_granularity > gran_limit)
4891 sysctl_sched_granularity = gran_limit;
4892
4893 sysctl_sched_runtime_limit = sysctl_sched_granularity * 4;
4894 sysctl_sched_wakeup_granularity = sysctl_sched_granularity / 2;
4895}
4896
Linus Torvalds1da177e2005-04-16 15:20:36 -07004897#ifdef CONFIG_SMP
4898/*
4899 * This is how migration works:
4900 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07004901 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07004902 * runqueue and wake up that CPU's migration thread.
4903 * 2) we down() the locked semaphore => thread blocks.
4904 * 3) migration thread wakes up (implicitly it forces the migrated
4905 * thread off the CPU)
4906 * 4) it gets the migration request and checks whether the migrated
4907 * task is still in the wrong runqueue.
4908 * 5) if it's in the wrong runqueue then the migration thread removes
4909 * it and puts it into the right queue.
4910 * 6) migration thread up()s the semaphore.
4911 * 7) we wake up and the migration is done.
4912 */
4913
4914/*
4915 * Change a given task's CPU affinity. Migrate the thread to a
4916 * proper CPU and schedule it away if the CPU it's executing on
4917 * is removed from the allowed bitmask.
4918 *
4919 * NOTE: the caller must have a valid reference to the task, the
4920 * task must not exit() & deallocate itself prematurely. The
4921 * call is not atomic; no spinlocks may be held.
4922 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004923int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004924{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004925 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004926 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004927 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004928 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004929
4930 rq = task_rq_lock(p, &flags);
4931 if (!cpus_intersects(new_mask, cpu_online_map)) {
4932 ret = -EINVAL;
4933 goto out;
4934 }
4935
4936 p->cpus_allowed = new_mask;
4937 /* Can the task run on the task's current CPU? If so, we're done */
4938 if (cpu_isset(task_cpu(p), new_mask))
4939 goto out;
4940
4941 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4942 /* Need help from migration thread: drop lock and wait. */
4943 task_rq_unlock(rq, &flags);
4944 wake_up_process(rq->migration_thread);
4945 wait_for_completion(&req.done);
4946 tlb_migrate_finish(p->mm);
4947 return 0;
4948 }
4949out:
4950 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004951
Linus Torvalds1da177e2005-04-16 15:20:36 -07004952 return ret;
4953}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004954EXPORT_SYMBOL_GPL(set_cpus_allowed);
4955
4956/*
4957 * Move (not current) task off this cpu, onto dest cpu. We're doing
4958 * this because either it can't run here any more (set_cpus_allowed()
4959 * away from this CPU, or CPU going down), or because we're
4960 * attempting to rebalance this task on exec (sched_exec).
4961 *
4962 * So we race with normal scheduler movements, but that's OK, as long
4963 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07004964 *
4965 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004966 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07004967static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004968{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004969 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02004970 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004971
4972 if (unlikely(cpu_is_offline(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07004973 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004974
4975 rq_src = cpu_rq(src_cpu);
4976 rq_dest = cpu_rq(dest_cpu);
4977
4978 double_rq_lock(rq_src, rq_dest);
4979 /* Already moved. */
4980 if (task_cpu(p) != src_cpu)
4981 goto out;
4982 /* Affinity changed (again). */
4983 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4984 goto out;
4985
Ingo Molnardd41f592007-07-09 18:51:59 +02004986 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004987 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004988 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004989
Linus Torvalds1da177e2005-04-16 15:20:36 -07004990 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004991 if (on_rq) {
4992 activate_task(rq_dest, p, 0);
4993 check_preempt_curr(rq_dest, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004994 }
Kirill Korotaevefc30812006-06-27 02:54:32 -07004995 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004996out:
4997 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07004998 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004999}
5000
5001/*
5002 * migration_thread - this is a highprio system thread that performs
5003 * thread migration by bumping thread off CPU then 'pushing' onto
5004 * another runqueue.
5005 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005006static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005007{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005008 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005009 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005010
5011 rq = cpu_rq(cpu);
5012 BUG_ON(rq->migration_thread != current);
5013
5014 set_current_state(TASK_INTERRUPTIBLE);
5015 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005016 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005017 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005018
Linus Torvalds1da177e2005-04-16 15:20:36 -07005019 spin_lock_irq(&rq->lock);
5020
5021 if (cpu_is_offline(cpu)) {
5022 spin_unlock_irq(&rq->lock);
5023 goto wait_to_die;
5024 }
5025
5026 if (rq->active_balance) {
5027 active_load_balance(rq, cpu);
5028 rq->active_balance = 0;
5029 }
5030
5031 head = &rq->migration_queue;
5032
5033 if (list_empty(head)) {
5034 spin_unlock_irq(&rq->lock);
5035 schedule();
5036 set_current_state(TASK_INTERRUPTIBLE);
5037 continue;
5038 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07005039 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005040 list_del_init(head->next);
5041
Nick Piggin674311d2005-06-25 14:57:27 -07005042 spin_unlock(&rq->lock);
5043 __migrate_task(req->task, cpu, req->dest_cpu);
5044 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005045
5046 complete(&req->done);
5047 }
5048 __set_current_state(TASK_RUNNING);
5049 return 0;
5050
5051wait_to_die:
5052 /* Wait for kthread_stop */
5053 set_current_state(TASK_INTERRUPTIBLE);
5054 while (!kthread_should_stop()) {
5055 schedule();
5056 set_current_state(TASK_INTERRUPTIBLE);
5057 }
5058 __set_current_state(TASK_RUNNING);
5059 return 0;
5060}
5061
5062#ifdef CONFIG_HOTPLUG_CPU
Kirill Korotaev054b9102006-12-10 02:20:11 -08005063/*
5064 * Figure out where task on dead CPU should go, use force if neccessary.
5065 * NOTE: interrupts should be disabled by the caller
5066 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005067static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005068{
Kirill Korotaevefc30812006-06-27 02:54:32 -07005069 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005070 cpumask_t mask;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005071 struct rq *rq;
5072 int dest_cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005073
Kirill Korotaevefc30812006-06-27 02:54:32 -07005074restart:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005075 /* On same node? */
5076 mask = node_to_cpumask(cpu_to_node(dead_cpu));
Ingo Molnar48f24c42006-07-03 00:25:40 -07005077 cpus_and(mask, mask, p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005078 dest_cpu = any_online_cpu(mask);
5079
5080 /* On any allowed CPU? */
5081 if (dest_cpu == NR_CPUS)
Ingo Molnar48f24c42006-07-03 00:25:40 -07005082 dest_cpu = any_online_cpu(p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005083
5084 /* No more Mr. Nice Guy. */
5085 if (dest_cpu == NR_CPUS) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07005086 rq = task_rq_lock(p, &flags);
5087 cpus_setall(p->cpus_allowed);
5088 dest_cpu = any_online_cpu(p->cpus_allowed);
Kirill Korotaevefc30812006-06-27 02:54:32 -07005089 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005090
5091 /*
5092 * Don't tell them about moving exiting tasks or
5093 * kernel threads (both mm NULL), since they never
5094 * leave kernel.
5095 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005096 if (p->mm && printk_ratelimit())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005097 printk(KERN_INFO "process %d (%s) no "
5098 "longer affine to cpu%d\n",
Ingo Molnar48f24c42006-07-03 00:25:40 -07005099 p->pid, p->comm, dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005100 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07005101 if (!__migrate_task(p, dead_cpu, dest_cpu))
Kirill Korotaevefc30812006-06-27 02:54:32 -07005102 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005103}
5104
5105/*
5106 * While a dead CPU has no uninterruptible tasks queued at this point,
5107 * it might still have a nonzero ->nr_uninterruptible counter, because
5108 * for performance reasons the counter is not stricly tracking tasks to
5109 * their home CPUs. So we just add the counter to another CPU's counter,
5110 * to keep the global sum constant after CPU-down:
5111 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07005112static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005113{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005114 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005115 unsigned long flags;
5116
5117 local_irq_save(flags);
5118 double_rq_lock(rq_src, rq_dest);
5119 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5120 rq_src->nr_uninterruptible = 0;
5121 double_rq_unlock(rq_src, rq_dest);
5122 local_irq_restore(flags);
5123}
5124
5125/* Run through task list and migrate tasks from the dead cpu. */
5126static void migrate_live_tasks(int src_cpu)
5127{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005128 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005129
5130 write_lock_irq(&tasklist_lock);
5131
Ingo Molnar48f24c42006-07-03 00:25:40 -07005132 do_each_thread(t, p) {
5133 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005134 continue;
5135
Ingo Molnar48f24c42006-07-03 00:25:40 -07005136 if (task_cpu(p) == src_cpu)
5137 move_task_off_dead_cpu(src_cpu, p);
5138 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005139
5140 write_unlock_irq(&tasklist_lock);
5141}
5142
Ingo Molnardd41f592007-07-09 18:51:59 +02005143/*
5144 * Schedules idle task to be the next runnable task on current CPU.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005145 * It does so by boosting its priority to highest possible and adding it to
Ingo Molnar48f24c42006-07-03 00:25:40 -07005146 * the _front_ of the runqueue. Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005147 */
5148void sched_idle_next(void)
5149{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005150 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07005151 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005152 struct task_struct *p = rq->idle;
5153 unsigned long flags;
5154
5155 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005156 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005157
Ingo Molnar48f24c42006-07-03 00:25:40 -07005158 /*
5159 * Strictly not necessary since rest of the CPUs are stopped by now
5160 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005161 */
5162 spin_lock_irqsave(&rq->lock, flags);
5163
Ingo Molnardd41f592007-07-09 18:51:59 +02005164 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005165
5166 /* Add idle task to the _front_ of its priority queue: */
Ingo Molnardd41f592007-07-09 18:51:59 +02005167 activate_idle_task(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005168
5169 spin_unlock_irqrestore(&rq->lock, flags);
5170}
5171
Ingo Molnar48f24c42006-07-03 00:25:40 -07005172/*
5173 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07005174 * offline.
5175 */
5176void idle_task_exit(void)
5177{
5178 struct mm_struct *mm = current->active_mm;
5179
5180 BUG_ON(cpu_online(smp_processor_id()));
5181
5182 if (mm != &init_mm)
5183 switch_mm(mm, &init_mm, current);
5184 mmdrop(mm);
5185}
5186
Kirill Korotaev054b9102006-12-10 02:20:11 -08005187/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005188static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005189{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005190 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005191
5192 /* Must be exiting, otherwise would be on tasklist. */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005193 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005194
5195 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07005196 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005197
Ingo Molnar48f24c42006-07-03 00:25:40 -07005198 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005199
5200 /*
5201 * Drop lock around migration; if someone else moves it,
5202 * that's OK. No task can be added to this CPU, so iteration is
5203 * fine.
Kirill Korotaev054b9102006-12-10 02:20:11 -08005204 * NOTE: interrupts should be left disabled --dev@
Linus Torvalds1da177e2005-04-16 15:20:36 -07005205 */
Kirill Korotaev054b9102006-12-10 02:20:11 -08005206 spin_unlock(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005207 move_task_off_dead_cpu(dead_cpu, p);
Kirill Korotaev054b9102006-12-10 02:20:11 -08005208 spin_lock(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005209
Ingo Molnar48f24c42006-07-03 00:25:40 -07005210 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005211}
5212
5213/* release_task() removes task from tasklist, so we won't find dead tasks. */
5214static void migrate_dead_tasks(unsigned int dead_cpu)
5215{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005216 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005217 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005218
Ingo Molnardd41f592007-07-09 18:51:59 +02005219 for ( ; ; ) {
5220 if (!rq->nr_running)
5221 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02005222 update_rq_clock(rq);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02005223 next = pick_next_task(rq, rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02005224 if (!next)
5225 break;
5226 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02005227
Linus Torvalds1da177e2005-04-16 15:20:36 -07005228 }
5229}
5230#endif /* CONFIG_HOTPLUG_CPU */
5231
Nick Piggine692ab52007-07-26 13:40:43 +02005232#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5233
5234static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005235 {
5236 .procname = "sched_domain",
5237 .mode = 0755,
5238 },
Nick Piggine692ab52007-07-26 13:40:43 +02005239 {0,},
5240};
5241
5242static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005243 {
5244 .procname = "kernel",
5245 .mode = 0755,
5246 .child = sd_ctl_dir,
5247 },
Nick Piggine692ab52007-07-26 13:40:43 +02005248 {0,},
5249};
5250
5251static struct ctl_table *sd_alloc_ctl_entry(int n)
5252{
5253 struct ctl_table *entry =
5254 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5255
5256 BUG_ON(!entry);
5257 memset(entry, 0, n * sizeof(struct ctl_table));
5258
5259 return entry;
5260}
5261
5262static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02005263set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02005264 const char *procname, void *data, int maxlen,
5265 mode_t mode, proc_handler *proc_handler)
5266{
Nick Piggine692ab52007-07-26 13:40:43 +02005267 entry->procname = procname;
5268 entry->data = data;
5269 entry->maxlen = maxlen;
5270 entry->mode = mode;
5271 entry->proc_handler = proc_handler;
5272}
5273
5274static struct ctl_table *
5275sd_alloc_ctl_domain_table(struct sched_domain *sd)
5276{
5277 struct ctl_table *table = sd_alloc_ctl_entry(14);
5278
Alexey Dobriyane0361852007-08-09 11:16:46 +02005279 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02005280 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005281 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02005282 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005283 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005284 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005285 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005286 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005287 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005288 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005289 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005290 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005291 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02005292 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005293 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02005294 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005295 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02005296 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005297 set_table_entry(&table[10], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02005298 &sd->cache_nice_tries,
5299 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005300 set_table_entry(&table[12], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02005301 sizeof(int), 0644, proc_dointvec_minmax);
5302
5303 return table;
5304}
5305
5306static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5307{
5308 struct ctl_table *entry, *table;
5309 struct sched_domain *sd;
5310 int domain_num = 0, i;
5311 char buf[32];
5312
5313 for_each_domain(cpu, sd)
5314 domain_num++;
5315 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5316
5317 i = 0;
5318 for_each_domain(cpu, sd) {
5319 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02005320 entry->procname = kstrdup(buf, GFP_KERNEL);
5321 entry->mode = 0755;
5322 entry->child = sd_alloc_ctl_domain_table(sd);
5323 entry++;
5324 i++;
5325 }
5326 return table;
5327}
5328
5329static struct ctl_table_header *sd_sysctl_header;
5330static void init_sched_domain_sysctl(void)
5331{
5332 int i, cpu_num = num_online_cpus();
5333 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5334 char buf[32];
5335
5336 sd_ctl_dir[0].child = entry;
5337
5338 for (i = 0; i < cpu_num; i++, entry++) {
5339 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02005340 entry->procname = kstrdup(buf, GFP_KERNEL);
5341 entry->mode = 0755;
5342 entry->child = sd_alloc_ctl_cpu_table(i);
5343 }
5344 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5345}
5346#else
5347static void init_sched_domain_sysctl(void)
5348{
5349}
5350#endif
5351
Linus Torvalds1da177e2005-04-16 15:20:36 -07005352/*
5353 * migration_call - callback that gets triggered when a CPU is added.
5354 * Here we can start up the necessary migration thread for the new CPU.
5355 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005356static int __cpuinit
5357migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005358{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005359 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005360 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005361 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005362 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005363
5364 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07005365 case CPU_LOCK_ACQUIRE:
5366 mutex_lock(&sched_hotcpu_mutex);
5367 break;
5368
Linus Torvalds1da177e2005-04-16 15:20:36 -07005369 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005370 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02005371 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005372 if (IS_ERR(p))
5373 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005374 kthread_bind(p, cpu);
5375 /* Must be high prio: stop_machine expects to yield to it. */
5376 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02005377 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005378 task_rq_unlock(rq, &flags);
5379 cpu_rq(cpu)->migration_thread = p;
5380 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005381
Linus Torvalds1da177e2005-04-16 15:20:36 -07005382 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005383 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005384 /* Strictly unneccessary, as first user will wake it. */
5385 wake_up_process(cpu_rq(cpu)->migration_thread);
5386 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005387
Linus Torvalds1da177e2005-04-16 15:20:36 -07005388#ifdef CONFIG_HOTPLUG_CPU
5389 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005390 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07005391 if (!cpu_rq(cpu)->migration_thread)
5392 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005393 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08005394 kthread_bind(cpu_rq(cpu)->migration_thread,
5395 any_online_cpu(cpu_online_map));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005396 kthread_stop(cpu_rq(cpu)->migration_thread);
5397 cpu_rq(cpu)->migration_thread = NULL;
5398 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005399
Linus Torvalds1da177e2005-04-16 15:20:36 -07005400 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005401 case CPU_DEAD_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005402 migrate_live_tasks(cpu);
5403 rq = cpu_rq(cpu);
5404 kthread_stop(rq->migration_thread);
5405 rq->migration_thread = NULL;
5406 /* Idle task back to normal (off runqueue, low prio) */
5407 rq = task_rq_lock(rq->idle, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02005408 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005409 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005410 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02005411 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5412 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005413 migrate_dead_tasks(cpu);
5414 task_rq_unlock(rq, &flags);
5415 migrate_nr_uninterruptible(rq);
5416 BUG_ON(rq->nr_running != 0);
5417
5418 /* No need to migrate the tasks: it was best-effort if
Gautham R Shenoy5be93612007-05-09 02:34:04 -07005419 * they didn't take sched_hotcpu_mutex. Just wake up
Linus Torvalds1da177e2005-04-16 15:20:36 -07005420 * the requestors. */
5421 spin_lock_irq(&rq->lock);
5422 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07005423 struct migration_req *req;
5424
Linus Torvalds1da177e2005-04-16 15:20:36 -07005425 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07005426 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005427 list_del_init(&req->list);
5428 complete(&req->done);
5429 }
5430 spin_unlock_irq(&rq->lock);
5431 break;
5432#endif
Gautham R Shenoy5be93612007-05-09 02:34:04 -07005433 case CPU_LOCK_RELEASE:
5434 mutex_unlock(&sched_hotcpu_mutex);
5435 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005436 }
5437 return NOTIFY_OK;
5438}
5439
5440/* Register at highest priority so that task migration (migrate_all_tasks)
5441 * happens before everything else.
5442 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07005443static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005444 .notifier_call = migration_call,
5445 .priority = 10
5446};
5447
5448int __init migration_init(void)
5449{
5450 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07005451 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005452
5453 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07005454 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5455 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005456 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5457 register_cpu_notifier(&migration_notifier);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005458
Linus Torvalds1da177e2005-04-16 15:20:36 -07005459 return 0;
5460}
5461#endif
5462
5463#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07005464
5465/* Number of possible processor ids */
5466int nr_cpu_ids __read_mostly = NR_CPUS;
5467EXPORT_SYMBOL(nr_cpu_ids);
5468
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07005469#undef SCHED_DOMAIN_DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07005470#ifdef SCHED_DOMAIN_DEBUG
5471static void sched_domain_debug(struct sched_domain *sd, int cpu)
5472{
5473 int level = 0;
5474
Nick Piggin41c7ce92005-06-25 14:57:24 -07005475 if (!sd) {
5476 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5477 return;
5478 }
5479
Linus Torvalds1da177e2005-04-16 15:20:36 -07005480 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5481
5482 do {
5483 int i;
5484 char str[NR_CPUS];
5485 struct sched_group *group = sd->groups;
5486 cpumask_t groupmask;
5487
5488 cpumask_scnprintf(str, NR_CPUS, sd->span);
5489 cpus_clear(groupmask);
5490
5491 printk(KERN_DEBUG);
5492 for (i = 0; i < level + 1; i++)
5493 printk(" ");
5494 printk("domain %d: ", level);
5495
5496 if (!(sd->flags & SD_LOAD_BALANCE)) {
5497 printk("does not load-balance\n");
5498 if (sd->parent)
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005499 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5500 " has parent");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005501 break;
5502 }
5503
5504 printk("span %s\n", str);
5505
5506 if (!cpu_isset(cpu, sd->span))
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005507 printk(KERN_ERR "ERROR: domain->span does not contain "
5508 "CPU%d\n", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005509 if (!cpu_isset(cpu, group->cpumask))
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005510 printk(KERN_ERR "ERROR: domain->groups does not contain"
5511 " CPU%d\n", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005512
5513 printk(KERN_DEBUG);
5514 for (i = 0; i < level + 2; i++)
5515 printk(" ");
5516 printk("groups:");
5517 do {
5518 if (!group) {
5519 printk("\n");
5520 printk(KERN_ERR "ERROR: group is NULL\n");
5521 break;
5522 }
5523
Eric Dumazet5517d862007-05-08 00:32:57 -07005524 if (!group->__cpu_power) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005525 printk("\n");
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005526 printk(KERN_ERR "ERROR: domain->cpu_power not "
5527 "set\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005528 }
5529
5530 if (!cpus_weight(group->cpumask)) {
5531 printk("\n");
5532 printk(KERN_ERR "ERROR: empty group\n");
5533 }
5534
5535 if (cpus_intersects(groupmask, group->cpumask)) {
5536 printk("\n");
5537 printk(KERN_ERR "ERROR: repeated CPUs\n");
5538 }
5539
5540 cpus_or(groupmask, groupmask, group->cpumask);
5541
5542 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5543 printk(" %s", str);
5544
5545 group = group->next;
5546 } while (group != sd->groups);
5547 printk("\n");
5548
5549 if (!cpus_equal(sd->span, groupmask))
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005550 printk(KERN_ERR "ERROR: groups don't span "
5551 "domain->span\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005552
5553 level++;
5554 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005555 if (!sd)
5556 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005557
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005558 if (!cpus_subset(groupmask, sd->span))
5559 printk(KERN_ERR "ERROR: parent span is not a superset "
5560 "of domain->span\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005561
5562 } while (sd);
5563}
5564#else
Ingo Molnar48f24c42006-07-03 00:25:40 -07005565# define sched_domain_debug(sd, cpu) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005566#endif
5567
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07005568static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07005569{
5570 if (cpus_weight(sd->span) == 1)
5571 return 1;
5572
5573 /* Following flags need at least 2 groups */
5574 if (sd->flags & (SD_LOAD_BALANCE |
5575 SD_BALANCE_NEWIDLE |
5576 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005577 SD_BALANCE_EXEC |
5578 SD_SHARE_CPUPOWER |
5579 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005580 if (sd->groups != sd->groups->next)
5581 return 0;
5582 }
5583
5584 /* Following flags don't use groups */
5585 if (sd->flags & (SD_WAKE_IDLE |
5586 SD_WAKE_AFFINE |
5587 SD_WAKE_BALANCE))
5588 return 0;
5589
5590 return 1;
5591}
5592
Ingo Molnar48f24c42006-07-03 00:25:40 -07005593static int
5594sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07005595{
5596 unsigned long cflags = sd->flags, pflags = parent->flags;
5597
5598 if (sd_degenerate(parent))
5599 return 1;
5600
5601 if (!cpus_equal(sd->span, parent->span))
5602 return 0;
5603
5604 /* Does parent contain flags not in child? */
5605 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5606 if (cflags & SD_WAKE_AFFINE)
5607 pflags &= ~SD_WAKE_BALANCE;
5608 /* Flags needing groups don't count if only 1 group in parent */
5609 if (parent->groups == parent->groups->next) {
5610 pflags &= ~(SD_LOAD_BALANCE |
5611 SD_BALANCE_NEWIDLE |
5612 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005613 SD_BALANCE_EXEC |
5614 SD_SHARE_CPUPOWER |
5615 SD_SHARE_PKG_RESOURCES);
Suresh Siddha245af2c2005-06-25 14:57:25 -07005616 }
5617 if (~cflags & pflags)
5618 return 0;
5619
5620 return 1;
5621}
5622
Linus Torvalds1da177e2005-04-16 15:20:36 -07005623/*
5624 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5625 * hold the hotplug lock.
5626 */
John Hawkes9c1cfda2005-09-06 15:18:14 -07005627static void cpu_attach_domain(struct sched_domain *sd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005628{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005629 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07005630 struct sched_domain *tmp;
5631
5632 /* Remove the sched domains which do not contribute to scheduling. */
5633 for (tmp = sd; tmp; tmp = tmp->parent) {
5634 struct sched_domain *parent = tmp->parent;
5635 if (!parent)
5636 break;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005637 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005638 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005639 if (parent->parent)
5640 parent->parent->child = tmp;
5641 }
Suresh Siddha245af2c2005-06-25 14:57:25 -07005642 }
5643
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005644 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005645 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005646 if (sd)
5647 sd->child = NULL;
5648 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005649
5650 sched_domain_debug(sd, cpu);
5651
Nick Piggin674311d2005-06-25 14:57:27 -07005652 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005653}
5654
5655/* cpus with isolated domains */
Tim Chen67af63a2006-12-22 01:07:50 -08005656static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005657
5658/* Setup the mask of cpus configured for isolated domains */
5659static int __init isolated_cpu_setup(char *str)
5660{
5661 int ints[NR_CPUS], i;
5662
5663 str = get_options(str, ARRAY_SIZE(ints), ints);
5664 cpus_clear(cpu_isolated_map);
5665 for (i = 1; i <= ints[0]; i++)
5666 if (ints[i] < NR_CPUS)
5667 cpu_set(ints[i], cpu_isolated_map);
5668 return 1;
5669}
5670
5671__setup ("isolcpus=", isolated_cpu_setup);
5672
5673/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005674 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5675 * to a function which identifies what group(along with sched group) a CPU
5676 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5677 * (due to the fact that we keep track of groups covered with a cpumask_t).
Linus Torvalds1da177e2005-04-16 15:20:36 -07005678 *
5679 * init_sched_build_groups will build a circular linked list of the groups
5680 * covered by the given span, and will set each group's ->cpumask correctly,
5681 * and ->cpu_power to 0.
5682 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005683static void
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005684init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5685 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5686 struct sched_group **sg))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005687{
5688 struct sched_group *first = NULL, *last = NULL;
5689 cpumask_t covered = CPU_MASK_NONE;
5690 int i;
5691
5692 for_each_cpu_mask(i, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005693 struct sched_group *sg;
5694 int group = group_fn(i, cpu_map, &sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005695 int j;
5696
5697 if (cpu_isset(i, covered))
5698 continue;
5699
5700 sg->cpumask = CPU_MASK_NONE;
Eric Dumazet5517d862007-05-08 00:32:57 -07005701 sg->__cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005702
5703 for_each_cpu_mask(j, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005704 if (group_fn(j, cpu_map, NULL) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005705 continue;
5706
5707 cpu_set(j, covered);
5708 cpu_set(j, sg->cpumask);
5709 }
5710 if (!first)
5711 first = sg;
5712 if (last)
5713 last->next = sg;
5714 last = sg;
5715 }
5716 last->next = first;
5717}
5718
John Hawkes9c1cfda2005-09-06 15:18:14 -07005719#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07005720
John Hawkes9c1cfda2005-09-06 15:18:14 -07005721#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08005722
John Hawkes9c1cfda2005-09-06 15:18:14 -07005723/**
5724 * find_next_best_node - find the next node to include in a sched_domain
5725 * @node: node whose sched_domain we're building
5726 * @used_nodes: nodes already in the sched_domain
5727 *
5728 * Find the next node to include in a given scheduling domain. Simply
5729 * finds the closest node not already in the @used_nodes map.
5730 *
5731 * Should use nodemask_t.
5732 */
5733static int find_next_best_node(int node, unsigned long *used_nodes)
5734{
5735 int i, n, val, min_val, best_node = 0;
5736
5737 min_val = INT_MAX;
5738
5739 for (i = 0; i < MAX_NUMNODES; i++) {
5740 /* Start at @node */
5741 n = (node + i) % MAX_NUMNODES;
5742
5743 if (!nr_cpus_node(n))
5744 continue;
5745
5746 /* Skip already used nodes */
5747 if (test_bit(n, used_nodes))
5748 continue;
5749
5750 /* Simple min distance search */
5751 val = node_distance(node, n);
5752
5753 if (val < min_val) {
5754 min_val = val;
5755 best_node = n;
5756 }
5757 }
5758
5759 set_bit(best_node, used_nodes);
5760 return best_node;
5761}
5762
5763/**
5764 * sched_domain_node_span - get a cpumask for a node's sched_domain
5765 * @node: node whose cpumask we're constructing
5766 * @size: number of nodes to include in this span
5767 *
5768 * Given a node, construct a good cpumask for its sched_domain to span. It
5769 * should be one that prevents unnecessary balancing, but also spreads tasks
5770 * out optimally.
5771 */
5772static cpumask_t sched_domain_node_span(int node)
5773{
John Hawkes9c1cfda2005-09-06 15:18:14 -07005774 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005775 cpumask_t span, nodemask;
5776 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07005777
5778 cpus_clear(span);
5779 bitmap_zero(used_nodes, MAX_NUMNODES);
5780
5781 nodemask = node_to_cpumask(node);
5782 cpus_or(span, span, nodemask);
5783 set_bit(node, used_nodes);
5784
5785 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5786 int next_node = find_next_best_node(node, used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005787
John Hawkes9c1cfda2005-09-06 15:18:14 -07005788 nodemask = node_to_cpumask(next_node);
5789 cpus_or(span, span, nodemask);
5790 }
5791
5792 return span;
5793}
5794#endif
5795
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07005796int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005797
John Hawkes9c1cfda2005-09-06 15:18:14 -07005798/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07005799 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07005800 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005801#ifdef CONFIG_SCHED_SMT
5802static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005803static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005804
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005805static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5806 struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005807{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005808 if (sg)
5809 *sg = &per_cpu(sched_group_cpus, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005810 return cpu;
5811}
5812#endif
5813
Ingo Molnar48f24c42006-07-03 00:25:40 -07005814/*
5815 * multi-core sched-domains:
5816 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08005817#ifdef CONFIG_SCHED_MC
5818static DEFINE_PER_CPU(struct sched_domain, core_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005819static DEFINE_PER_CPU(struct sched_group, sched_group_core);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08005820#endif
5821
5822#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005823static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5824 struct sched_group **sg)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08005825{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005826 int group;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005827 cpumask_t mask = cpu_sibling_map[cpu];
5828 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005829 group = first_cpu(mask);
5830 if (sg)
5831 *sg = &per_cpu(sched_group_core, group);
5832 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08005833}
5834#elif defined(CONFIG_SCHED_MC)
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005835static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5836 struct sched_group **sg)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08005837{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005838 if (sg)
5839 *sg = &per_cpu(sched_group_core, cpu);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08005840 return cpu;
5841}
5842#endif
5843
Linus Torvalds1da177e2005-04-16 15:20:36 -07005844static DEFINE_PER_CPU(struct sched_domain, phys_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005845static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005846
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005847static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5848 struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005849{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005850 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005851#ifdef CONFIG_SCHED_MC
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08005852 cpumask_t mask = cpu_coregroup_map(cpu);
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005853 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005854 group = first_cpu(mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08005855#elif defined(CONFIG_SCHED_SMT)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005856 cpumask_t mask = cpu_sibling_map[cpu];
5857 cpus_and(mask, mask, *cpu_map);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005858 group = first_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005859#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005860 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005861#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005862 if (sg)
5863 *sg = &per_cpu(sched_group_phys, group);
5864 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005865}
5866
5867#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07005868/*
5869 * The init_sched_build_groups can't handle what we want to do with node
5870 * groups, so roll our own. Now each node has its own list of groups which
5871 * gets dynamically allocated.
5872 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005873static DEFINE_PER_CPU(struct sched_domain, node_domains);
John Hawkesd1b55132005-09-06 15:18:14 -07005874static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
John Hawkes9c1cfda2005-09-06 15:18:14 -07005875
5876static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005877static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07005878
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005879static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5880 struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005881{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005882 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5883 int group;
5884
5885 cpus_and(nodemask, nodemask, *cpu_map);
5886 group = first_cpu(nodemask);
5887
5888 if (sg)
5889 *sg = &per_cpu(sched_group_allnodes, group);
5890 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005891}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08005892
Siddha, Suresh B08069032006-03-27 01:15:23 -08005893static void init_numa_sched_groups_power(struct sched_group *group_head)
5894{
5895 struct sched_group *sg = group_head;
5896 int j;
5897
5898 if (!sg)
5899 return;
5900next_sg:
5901 for_each_cpu_mask(j, sg->cpumask) {
5902 struct sched_domain *sd;
5903
5904 sd = &per_cpu(phys_domains, j);
5905 if (j != first_cpu(sd->groups->cpumask)) {
5906 /*
5907 * Only add "power" once for each
5908 * physical package.
5909 */
5910 continue;
5911 }
5912
Eric Dumazet5517d862007-05-08 00:32:57 -07005913 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
Siddha, Suresh B08069032006-03-27 01:15:23 -08005914 }
5915 sg = sg->next;
5916 if (sg != group_head)
5917 goto next_sg;
5918}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005919#endif
5920
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005921#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07005922/* Free memory allocated for various sched_group structures */
5923static void free_sched_groups(const cpumask_t *cpu_map)
5924{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005925 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07005926
5927 for_each_cpu_mask(cpu, *cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07005928 struct sched_group **sched_group_nodes
5929 = sched_group_nodes_bycpu[cpu];
5930
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07005931 if (!sched_group_nodes)
5932 continue;
5933
5934 for (i = 0; i < MAX_NUMNODES; i++) {
5935 cpumask_t nodemask = node_to_cpumask(i);
5936 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5937
5938 cpus_and(nodemask, nodemask, *cpu_map);
5939 if (cpus_empty(nodemask))
5940 continue;
5941
5942 if (sg == NULL)
5943 continue;
5944 sg = sg->next;
5945next_sg:
5946 oldsg = sg;
5947 sg = sg->next;
5948 kfree(oldsg);
5949 if (oldsg != sched_group_nodes[i])
5950 goto next_sg;
5951 }
5952 kfree(sched_group_nodes);
5953 sched_group_nodes_bycpu[cpu] = NULL;
5954 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07005955}
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07005956#else
5957static void free_sched_groups(const cpumask_t *cpu_map)
5958{
5959}
5960#endif
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07005961
Linus Torvalds1da177e2005-04-16 15:20:36 -07005962/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005963 * Initialize sched groups cpu_power.
5964 *
5965 * cpu_power indicates the capacity of sched group, which is used while
5966 * distributing the load between different sched groups in a sched domain.
5967 * Typically cpu_power for all the groups in a sched domain will be same unless
5968 * there are asymmetries in the topology. If there are asymmetries, group
5969 * having more cpu_power will pickup more load compared to the group having
5970 * less cpu_power.
5971 *
5972 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5973 * the maximum number of tasks a group can handle in the presence of other idle
5974 * or lightly loaded groups in the same sched domain.
5975 */
5976static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5977{
5978 struct sched_domain *child;
5979 struct sched_group *group;
5980
5981 WARN_ON(!sd || !sd->groups);
5982
5983 if (cpu != first_cpu(sd->groups->cpumask))
5984 return;
5985
5986 child = sd->child;
5987
Eric Dumazet5517d862007-05-08 00:32:57 -07005988 sd->groups->__cpu_power = 0;
5989
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005990 /*
5991 * For perf policy, if the groups in child domain share resources
5992 * (for example cores sharing some portions of the cache hierarchy
5993 * or SMT), then set this domain groups cpu_power such that each group
5994 * can handle only one task, when there are other idle groups in the
5995 * same sched domain.
5996 */
5997 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5998 (child->flags &
5999 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
Eric Dumazet5517d862007-05-08 00:32:57 -07006000 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006001 return;
6002 }
6003
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006004 /*
6005 * add cpu_power of each child group to this groups cpu_power
6006 */
6007 group = child->groups;
6008 do {
Eric Dumazet5517d862007-05-08 00:32:57 -07006009 sg_inc_cpu_power(sd->groups, group->__cpu_power);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006010 group = group->next;
6011 } while (group != child->groups);
6012}
6013
6014/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006015 * Build sched domains for a given set of cpus and attach the sched domains
6016 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07006017 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006018static int build_sched_domains(const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006019{
6020 int i;
John Hawkesd1b55132005-09-06 15:18:14 -07006021#ifdef CONFIG_NUMA
6022 struct sched_group **sched_group_nodes = NULL;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006023 int sd_allnodes = 0;
John Hawkesd1b55132005-09-06 15:18:14 -07006024
6025 /*
6026 * Allocate the per-node list of sched groups
6027 */
Ingo Molnardd41f592007-07-09 18:51:59 +02006028 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
Srivatsa Vaddagirid3a5aa92006-06-27 02:54:39 -07006029 GFP_KERNEL);
John Hawkesd1b55132005-09-06 15:18:14 -07006030 if (!sched_group_nodes) {
6031 printk(KERN_WARNING "Can not alloc sched group node list\n");
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006032 return -ENOMEM;
John Hawkesd1b55132005-09-06 15:18:14 -07006033 }
6034 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6035#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006036
6037 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006038 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006039 */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006040 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006041 struct sched_domain *sd = NULL, *p;
6042 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6043
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006044 cpus_and(nodemask, nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006045
6046#ifdef CONFIG_NUMA
Ingo Molnardd41f592007-07-09 18:51:59 +02006047 if (cpus_weight(*cpu_map) >
6048 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07006049 sd = &per_cpu(allnodes_domains, i);
6050 *sd = SD_ALLNODES_INIT;
6051 sd->span = *cpu_map;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006052 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006053 p = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006054 sd_allnodes = 1;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006055 } else
6056 p = NULL;
6057
Linus Torvalds1da177e2005-04-16 15:20:36 -07006058 sd = &per_cpu(node_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006059 *sd = SD_NODE_INIT;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006060 sd->span = sched_domain_node_span(cpu_to_node(i));
6061 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006062 if (p)
6063 p->child = sd;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006064 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006065#endif
6066
6067 p = sd;
6068 sd = &per_cpu(phys_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006069 *sd = SD_CPU_INIT;
6070 sd->span = nodemask;
6071 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006072 if (p)
6073 p->child = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006074 cpu_to_phys_group(i, cpu_map, &sd->groups);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006075
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006076#ifdef CONFIG_SCHED_MC
6077 p = sd;
6078 sd = &per_cpu(core_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006079 *sd = SD_MC_INIT;
6080 sd->span = cpu_coregroup_map(i);
6081 cpus_and(sd->span, sd->span, *cpu_map);
6082 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006083 p->child = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006084 cpu_to_core_group(i, cpu_map, &sd->groups);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006085#endif
6086
Linus Torvalds1da177e2005-04-16 15:20:36 -07006087#ifdef CONFIG_SCHED_SMT
6088 p = sd;
6089 sd = &per_cpu(cpu_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006090 *sd = SD_SIBLING_INIT;
6091 sd->span = cpu_sibling_map[i];
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006092 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006093 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006094 p->child = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006095 cpu_to_cpu_group(i, cpu_map, &sd->groups);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006096#endif
6097 }
6098
6099#ifdef CONFIG_SCHED_SMT
6100 /* Set up CPU (sibling) groups */
John Hawkes9c1cfda2005-09-06 15:18:14 -07006101 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006102 cpumask_t this_sibling_map = cpu_sibling_map[i];
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006103 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006104 if (i != first_cpu(this_sibling_map))
6105 continue;
6106
Ingo Molnardd41f592007-07-09 18:51:59 +02006107 init_sched_build_groups(this_sibling_map, cpu_map,
6108 &cpu_to_cpu_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006109 }
6110#endif
6111
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006112#ifdef CONFIG_SCHED_MC
6113 /* Set up multi-core groups */
6114 for_each_cpu_mask(i, *cpu_map) {
6115 cpumask_t this_core_map = cpu_coregroup_map(i);
6116 cpus_and(this_core_map, this_core_map, *cpu_map);
6117 if (i != first_cpu(this_core_map))
6118 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02006119 init_sched_build_groups(this_core_map, cpu_map,
6120 &cpu_to_core_group);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006121 }
6122#endif
6123
Linus Torvalds1da177e2005-04-16 15:20:36 -07006124 /* Set up physical groups */
6125 for (i = 0; i < MAX_NUMNODES; i++) {
6126 cpumask_t nodemask = node_to_cpumask(i);
6127
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006128 cpus_and(nodemask, nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006129 if (cpus_empty(nodemask))
6130 continue;
6131
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006132 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006133 }
6134
6135#ifdef CONFIG_NUMA
6136 /* Set up node groups */
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006137 if (sd_allnodes)
Ingo Molnardd41f592007-07-09 18:51:59 +02006138 init_sched_build_groups(*cpu_map, cpu_map,
6139 &cpu_to_allnodes_group);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006140
6141 for (i = 0; i < MAX_NUMNODES; i++) {
6142 /* Set up node groups */
6143 struct sched_group *sg, *prev;
6144 cpumask_t nodemask = node_to_cpumask(i);
6145 cpumask_t domainspan;
6146 cpumask_t covered = CPU_MASK_NONE;
6147 int j;
6148
6149 cpus_and(nodemask, nodemask, *cpu_map);
John Hawkesd1b55132005-09-06 15:18:14 -07006150 if (cpus_empty(nodemask)) {
6151 sched_group_nodes[i] = NULL;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006152 continue;
John Hawkesd1b55132005-09-06 15:18:14 -07006153 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006154
6155 domainspan = sched_domain_node_span(i);
6156 cpus_and(domainspan, domainspan, *cpu_map);
6157
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07006158 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006159 if (!sg) {
6160 printk(KERN_WARNING "Can not alloc domain group for "
6161 "node %d\n", i);
6162 goto error;
6163 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006164 sched_group_nodes[i] = sg;
6165 for_each_cpu_mask(j, nodemask) {
6166 struct sched_domain *sd;
Ingo Molnar9761eea2007-07-09 18:52:00 +02006167
John Hawkes9c1cfda2005-09-06 15:18:14 -07006168 sd = &per_cpu(node_domains, j);
6169 sd->groups = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006170 }
Eric Dumazet5517d862007-05-08 00:32:57 -07006171 sg->__cpu_power = 0;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006172 sg->cpumask = nodemask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006173 sg->next = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006174 cpus_or(covered, covered, nodemask);
6175 prev = sg;
6176
6177 for (j = 0; j < MAX_NUMNODES; j++) {
6178 cpumask_t tmp, notcovered;
6179 int n = (i + j) % MAX_NUMNODES;
6180
6181 cpus_complement(notcovered, covered);
6182 cpus_and(tmp, notcovered, *cpu_map);
6183 cpus_and(tmp, tmp, domainspan);
6184 if (cpus_empty(tmp))
6185 break;
6186
6187 nodemask = node_to_cpumask(n);
6188 cpus_and(tmp, tmp, nodemask);
6189 if (cpus_empty(tmp))
6190 continue;
6191
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07006192 sg = kmalloc_node(sizeof(struct sched_group),
6193 GFP_KERNEL, i);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006194 if (!sg) {
6195 printk(KERN_WARNING
6196 "Can not alloc domain group for node %d\n", j);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006197 goto error;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006198 }
Eric Dumazet5517d862007-05-08 00:32:57 -07006199 sg->__cpu_power = 0;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006200 sg->cpumask = tmp;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006201 sg->next = prev->next;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006202 cpus_or(covered, covered, tmp);
6203 prev->next = sg;
6204 prev = sg;
6205 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006206 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006207#endif
6208
6209 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006210#ifdef CONFIG_SCHED_SMT
6211 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006212 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6213
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006214 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006215 }
6216#endif
6217#ifdef CONFIG_SCHED_MC
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006218 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006219 struct sched_domain *sd = &per_cpu(core_domains, i);
6220
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006221 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006222 }
6223#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006224
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006225 for_each_cpu_mask(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006226 struct sched_domain *sd = &per_cpu(phys_domains, i);
6227
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006228 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006229 }
6230
John Hawkes9c1cfda2005-09-06 15:18:14 -07006231#ifdef CONFIG_NUMA
Siddha, Suresh B08069032006-03-27 01:15:23 -08006232 for (i = 0; i < MAX_NUMNODES; i++)
6233 init_numa_sched_groups_power(sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006234
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006235 if (sd_allnodes) {
6236 struct sched_group *sg;
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07006237
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006238 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07006239 init_numa_sched_groups_power(sg);
6240 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006241#endif
6242
Linus Torvalds1da177e2005-04-16 15:20:36 -07006243 /* Attach the domains */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006244 for_each_cpu_mask(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006245 struct sched_domain *sd;
6246#ifdef CONFIG_SCHED_SMT
6247 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006248#elif defined(CONFIG_SCHED_MC)
6249 sd = &per_cpu(core_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006250#else
6251 sd = &per_cpu(phys_domains, i);
6252#endif
6253 cpu_attach_domain(sd, i);
6254 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006255
6256 return 0;
6257
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006258#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006259error:
6260 free_sched_groups(cpu_map);
6261 return -ENOMEM;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006262#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006263}
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006264/*
6265 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6266 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006267static int arch_init_sched_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006268{
6269 cpumask_t cpu_default_map;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006270 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006271
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006272 /*
6273 * Setup mask for cpus without special case scheduling requirements.
6274 * For now this just excludes isolated cpus, but could be used to
6275 * exclude other special cases in the future.
6276 */
6277 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6278
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006279 err = build_sched_domains(&cpu_default_map);
6280
6281 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006282}
6283
6284static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006285{
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006286 free_sched_groups(cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07006287}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006288
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006289/*
6290 * Detach sched domains from a group of cpus specified in cpu_map
6291 * These cpus will now be attached to the NULL domain
6292 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08006293static void detach_destroy_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006294{
6295 int i;
6296
6297 for_each_cpu_mask(i, *cpu_map)
6298 cpu_attach_domain(NULL, i);
6299 synchronize_sched();
6300 arch_destroy_sched_domains(cpu_map);
6301}
6302
6303/*
6304 * Partition sched domains as specified by the cpumasks below.
6305 * This attaches all cpus from the cpumasks to the NULL domain,
6306 * waits for a RCU quiescent period, recalculates sched
6307 * domain information and then attaches them back to the
6308 * correct sched domains
6309 * Call with hotplug lock held
6310 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006311int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006312{
6313 cpumask_t change_map;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006314 int err = 0;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006315
6316 cpus_and(*partition1, *partition1, cpu_online_map);
6317 cpus_and(*partition2, *partition2, cpu_online_map);
6318 cpus_or(change_map, *partition1, *partition2);
6319
6320 /* Detach sched domains from all of the affected cpus */
6321 detach_destroy_domains(&change_map);
6322 if (!cpus_empty(*partition1))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006323 err = build_sched_domains(partition1);
6324 if (!err && !cpus_empty(*partition2))
6325 err = build_sched_domains(partition2);
6326
6327 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006328}
6329
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006330#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Adrian Bunk6707de002007-08-12 18:08:19 +02006331static int arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006332{
6333 int err;
6334
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006335 mutex_lock(&sched_hotcpu_mutex);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006336 detach_destroy_domains(&cpu_online_map);
6337 err = arch_init_sched_domains(&cpu_online_map);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006338 mutex_unlock(&sched_hotcpu_mutex);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006339
6340 return err;
6341}
6342
6343static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6344{
6345 int ret;
6346
6347 if (buf[0] != '0' && buf[0] != '1')
6348 return -EINVAL;
6349
6350 if (smt)
6351 sched_smt_power_savings = (buf[0] == '1');
6352 else
6353 sched_mc_power_savings = (buf[0] == '1');
6354
6355 ret = arch_reinit_sched_domains();
6356
6357 return ret ? ret : count;
6358}
6359
Adrian Bunk6707de002007-08-12 18:08:19 +02006360#ifdef CONFIG_SCHED_MC
6361static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6362{
6363 return sprintf(page, "%u\n", sched_mc_power_savings);
6364}
6365static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6366 const char *buf, size_t count)
6367{
6368 return sched_power_savings_store(buf, count, 0);
6369}
6370static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6371 sched_mc_power_savings_store);
6372#endif
6373
6374#ifdef CONFIG_SCHED_SMT
6375static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6376{
6377 return sprintf(page, "%u\n", sched_smt_power_savings);
6378}
6379static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6380 const char *buf, size_t count)
6381{
6382 return sched_power_savings_store(buf, count, 1);
6383}
6384static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6385 sched_smt_power_savings_store);
6386#endif
6387
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006388int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6389{
6390 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006391
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07006392#ifdef CONFIG_SCHED_SMT
6393 if (smt_capable())
6394 err = sysfs_create_file(&cls->kset.kobj,
6395 &attr_sched_smt_power_savings.attr);
6396#endif
6397#ifdef CONFIG_SCHED_MC
6398 if (!err && mc_capable())
6399 err = sysfs_create_file(&cls->kset.kobj,
6400 &attr_sched_mc_power_savings.attr);
6401#endif
6402 return err;
6403}
6404#endif
6405
Linus Torvalds1da177e2005-04-16 15:20:36 -07006406/*
6407 * Force a reinitialization of the sched domains hierarchy. The domains
6408 * and groups cannot be updated in place without racing with the balancing
Nick Piggin41c7ce92005-06-25 14:57:24 -07006409 * code, so we temporarily attach all running cpus to the NULL domain
Linus Torvalds1da177e2005-04-16 15:20:36 -07006410 * which will prevent rebalancing while the sched domains are recalculated.
6411 */
6412static int update_sched_domains(struct notifier_block *nfb,
6413 unsigned long action, void *hcpu)
6414{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006415 switch (action) {
6416 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006417 case CPU_UP_PREPARE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006418 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006419 case CPU_DOWN_PREPARE_FROZEN:
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006420 detach_destroy_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006421 return NOTIFY_OK;
6422
6423 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006424 case CPU_UP_CANCELED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006425 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006426 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006427 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006428 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006429 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006430 case CPU_DEAD_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006431 /*
6432 * Fall through and re-initialise the domains.
6433 */
6434 break;
6435 default:
6436 return NOTIFY_DONE;
6437 }
6438
6439 /* The hotplug lock is already held by cpu_up/cpu_down */
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006440 arch_init_sched_domains(&cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006441
6442 return NOTIFY_OK;
6443}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006444
6445void __init sched_init_smp(void)
6446{
Nick Piggin5c1e1762006-10-03 01:14:04 -07006447 cpumask_t non_isolated_cpus;
6448
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006449 mutex_lock(&sched_hotcpu_mutex);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006450 arch_init_sched_domains(&cpu_online_map);
Nathan Lynche5e56732007-01-10 23:15:28 -08006451 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
Nick Piggin5c1e1762006-10-03 01:14:04 -07006452 if (cpus_empty(non_isolated_cpus))
6453 cpu_set(smp_processor_id(), non_isolated_cpus);
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006454 mutex_unlock(&sched_hotcpu_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006455 /* XXX: Theoretical race here - CPU may be hotplugged now */
6456 hotcpu_notifier(update_sched_domains, 0);
Nick Piggin5c1e1762006-10-03 01:14:04 -07006457
Nick Piggine692ab52007-07-26 13:40:43 +02006458 init_sched_domain_sysctl();
6459
Nick Piggin5c1e1762006-10-03 01:14:04 -07006460 /* Move init over to a non-isolated CPU */
6461 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6462 BUG();
Ingo Molnardd41f592007-07-09 18:51:59 +02006463 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006464}
6465#else
6466void __init sched_init_smp(void)
6467{
Ingo Molnardd41f592007-07-09 18:51:59 +02006468 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006469}
6470#endif /* CONFIG_SMP */
6471
6472int in_sched_functions(unsigned long addr)
6473{
6474 /* Linker adds these: start and end of __sched functions */
6475 extern char __sched_text_start[], __sched_text_end[];
Ingo Molnar48f24c42006-07-03 00:25:40 -07006476
Linus Torvalds1da177e2005-04-16 15:20:36 -07006477 return in_lock_functions(addr) ||
6478 (addr >= (unsigned long)__sched_text_start
6479 && addr < (unsigned long)__sched_text_end);
6480}
6481
Ingo Molnardd41f592007-07-09 18:51:59 +02006482static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6483{
6484 cfs_rq->tasks_timeline = RB_ROOT;
6485 cfs_rq->fair_clock = 1;
6486#ifdef CONFIG_FAIR_GROUP_SCHED
6487 cfs_rq->rq = rq;
6488#endif
6489}
6490
Linus Torvalds1da177e2005-04-16 15:20:36 -07006491void __init sched_init(void)
6492{
Ingo Molnardd41f592007-07-09 18:51:59 +02006493 u64 now = sched_clock();
Christoph Lameter476f3532007-05-06 14:48:58 -07006494 int highest_cpu = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006495 int i, j;
6496
6497 /*
6498 * Link up the scheduling class hierarchy:
6499 */
6500 rt_sched_class.next = &fair_sched_class;
6501 fair_sched_class.next = &idle_sched_class;
6502 idle_sched_class.next = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006503
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08006504 for_each_possible_cpu(i) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006505 struct rt_prio_array *array;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006506 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006507
6508 rq = cpu_rq(i);
6509 spin_lock_init(&rq->lock);
Ingo Molnarfcb99372006-07-03 00:25:10 -07006510 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
Nick Piggin78979862005-06-25 14:57:13 -07006511 rq->nr_running = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006512 rq->clock = 1;
6513 init_cfs_rq(&rq->cfs, rq);
6514#ifdef CONFIG_FAIR_GROUP_SCHED
6515 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6516 list_add(&rq->cfs.leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6517#endif
6518 rq->ls.load_update_last = now;
6519 rq->ls.load_update_start = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006520
Ingo Molnardd41f592007-07-09 18:51:59 +02006521 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6522 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006523#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07006524 rq->sd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006525 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006526 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006527 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07006528 rq->cpu = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006529 rq->migration_thread = NULL;
6530 INIT_LIST_HEAD(&rq->migration_queue);
6531#endif
6532 atomic_set(&rq->nr_iowait, 0);
6533
Ingo Molnardd41f592007-07-09 18:51:59 +02006534 array = &rq->rt.active;
6535 for (j = 0; j < MAX_RT_PRIO; j++) {
6536 INIT_LIST_HEAD(array->queue + j);
6537 __clear_bit(j, array->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006538 }
Christoph Lameter476f3532007-05-06 14:48:58 -07006539 highest_cpu = i;
Ingo Molnardd41f592007-07-09 18:51:59 +02006540 /* delimiter for bitsearch: */
6541 __set_bit(MAX_RT_PRIO, array->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006542 }
6543
Peter Williams2dd73a42006-06-27 02:54:34 -07006544 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07006545
Avi Kivitye107be32007-07-26 13:40:43 +02006546#ifdef CONFIG_PREEMPT_NOTIFIERS
6547 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6548#endif
6549
Christoph Lameterc9819f42006-12-10 02:20:25 -08006550#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07006551 nr_cpu_ids = highest_cpu + 1;
Christoph Lameterc9819f42006-12-10 02:20:25 -08006552 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6553#endif
6554
Heiko Carstensb50f60c2006-07-30 03:03:52 -07006555#ifdef CONFIG_RT_MUTEXES
6556 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6557#endif
6558
Linus Torvalds1da177e2005-04-16 15:20:36 -07006559 /*
6560 * The boot idle thread does lazy MMU switching as well:
6561 */
6562 atomic_inc(&init_mm.mm_count);
6563 enter_lazy_tlb(&init_mm, current);
6564
6565 /*
6566 * Make us the idle thread. Technically, schedule() should not be
6567 * called from this thread, however somewhere below it might be,
6568 * but because we are the idle thread, we just pick up running again
6569 * when this runqueue becomes "idle".
6570 */
6571 init_idle(current, smp_processor_id());
Ingo Molnardd41f592007-07-09 18:51:59 +02006572 /*
6573 * During early bootup we pretend to be a normal task:
6574 */
6575 current->sched_class = &fair_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006576}
6577
6578#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6579void __might_sleep(char *file, int line)
6580{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006581#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07006582 static unsigned long prev_jiffy; /* ratelimiting */
6583
6584 if ((in_atomic() || irqs_disabled()) &&
6585 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6586 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6587 return;
6588 prev_jiffy = jiffies;
Ingo Molnar91368d72006-03-23 03:00:54 -08006589 printk(KERN_ERR "BUG: sleeping function called from invalid"
Linus Torvalds1da177e2005-04-16 15:20:36 -07006590 " context at %s:%d\n", file, line);
6591 printk("in_atomic():%d, irqs_disabled():%d\n",
6592 in_atomic(), irqs_disabled());
Peter Zijlstraa4c410f2006-12-06 20:37:21 -08006593 debug_show_held_locks(current);
Ingo Molnar3117df02006-12-13 00:34:43 -08006594 if (irqs_disabled())
6595 print_irqtrace_events(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006596 dump_stack();
6597 }
6598#endif
6599}
6600EXPORT_SYMBOL(__might_sleep);
6601#endif
6602
6603#ifdef CONFIG_MAGIC_SYSRQ
6604void normalize_rt_tasks(void)
6605{
Ingo Molnara0f98a12007-06-17 18:37:45 +02006606 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006607 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006608 struct rq *rq;
Ingo Molnardd41f592007-07-09 18:51:59 +02006609 int on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006610
6611 read_lock_irq(&tasklist_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02006612 do_each_thread(g, p) {
Ingo Molnardd41f592007-07-09 18:51:59 +02006613 p->se.fair_key = 0;
6614 p->se.wait_runtime = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006615 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02006616 p->se.wait_start_fair = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006617 p->se.sleep_start_fair = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02006618#ifdef CONFIG_SCHEDSTATS
6619 p->se.wait_start = 0;
6620 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02006621 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02006622#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02006623 task_rq(p)->cfs.fair_clock = 0;
6624 task_rq(p)->clock = 0;
6625
6626 if (!rt_task(p)) {
6627 /*
6628 * Renice negative nice level userspace
6629 * tasks back to 0:
6630 */
6631 if (TASK_NICE(p) < 0 && p->mm)
6632 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006633 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02006634 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006635
Ingo Molnarb29739f2006-06-27 02:54:51 -07006636 spin_lock_irqsave(&p->pi_lock, flags);
6637 rq = __task_rq_lock(p);
Ingo Molnardd41f592007-07-09 18:51:59 +02006638#ifdef CONFIG_SMP
6639 /*
6640 * Do not touch the migration thread:
6641 */
6642 if (p == rq->migration_thread)
6643 goto out_unlock;
6644#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006645
Ingo Molnar2daa3572007-08-09 11:16:51 +02006646 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006647 on_rq = p->se.on_rq;
Ingo Molnar2daa3572007-08-09 11:16:51 +02006648 if (on_rq)
6649 deactivate_task(rq, p, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02006650 __setscheduler(rq, p, SCHED_NORMAL, 0);
6651 if (on_rq) {
Ingo Molnar2daa3572007-08-09 11:16:51 +02006652 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006653 resched_task(rq->curr);
6654 }
Ingo Molnardd41f592007-07-09 18:51:59 +02006655#ifdef CONFIG_SMP
6656 out_unlock:
6657#endif
Ingo Molnarb29739f2006-06-27 02:54:51 -07006658 __task_rq_unlock(rq);
6659 spin_unlock_irqrestore(&p->pi_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02006660 } while_each_thread(g, p);
6661
Linus Torvalds1da177e2005-04-16 15:20:36 -07006662 read_unlock_irq(&tasklist_lock);
6663}
6664
6665#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07006666
6667#ifdef CONFIG_IA64
6668/*
6669 * These functions are only useful for the IA64 MCA handling.
6670 *
6671 * They can only be called when the whole system has been
6672 * stopped - every CPU needs to be quiescent, and no scheduling
6673 * activity can take place. Using them for anything else would
6674 * be a serious bug, and as a result, they aren't even visible
6675 * under any other configuration.
6676 */
6677
6678/**
6679 * curr_task - return the current task for a given cpu.
6680 * @cpu: the processor in question.
6681 *
6682 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6683 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006684struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07006685{
6686 return cpu_curr(cpu);
6687}
6688
6689/**
6690 * set_curr_task - set the current task for a given cpu.
6691 * @cpu: the processor in question.
6692 * @p: the task pointer to set.
6693 *
6694 * Description: This function must only be used when non-maskable interrupts
6695 * are serviced on a separate stack. It allows the architecture to switch the
6696 * notion of the current task on a cpu in a non-blocking manner. This function
6697 * must be called with all CPU's synchronized, and interrupts disabled, the
6698 * and caller must save the original value of the current task (see
6699 * curr_task() above) and restore that value before reenabling interrupts and
6700 * re-starting the system.
6701 *
6702 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6703 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006704void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07006705{
6706 cpu_curr(cpu) = p;
6707}
6708
6709#endif