Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
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 | |
c31f2e8a IM |
19 | * 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 | |
b9131769 IM |
25 | * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, |
26 | * Thomas Gleixner, Mike Kravetz | |
1da177e4 LT |
27 | */ |
28 | ||
29 | #include <linux/mm.h> | |
30 | #include <linux/module.h> | |
31 | #include <linux/nmi.h> | |
32 | #include <linux/init.h> | |
dff06c15 | 33 | #include <linux/uaccess.h> |
1da177e4 LT |
34 | #include <linux/highmem.h> |
35 | #include <linux/smp_lock.h> | |
36 | #include <asm/mmu_context.h> | |
37 | #include <linux/interrupt.h> | |
c59ede7b | 38 | #include <linux/capability.h> |
1da177e4 LT |
39 | #include <linux/completion.h> |
40 | #include <linux/kernel_stat.h> | |
9a11b49a | 41 | #include <linux/debug_locks.h> |
1da177e4 LT |
42 | #include <linux/security.h> |
43 | #include <linux/notifier.h> | |
44 | #include <linux/profile.h> | |
7dfb7103 | 45 | #include <linux/freezer.h> |
198e2f18 | 46 | #include <linux/vmalloc.h> |
1da177e4 LT |
47 | #include <linux/blkdev.h> |
48 | #include <linux/delay.h> | |
b488893a | 49 | #include <linux/pid_namespace.h> |
1da177e4 LT |
50 | #include <linux/smp.h> |
51 | #include <linux/threads.h> | |
52 | #include <linux/timer.h> | |
53 | #include <linux/rcupdate.h> | |
54 | #include <linux/cpu.h> | |
55 | #include <linux/cpuset.h> | |
56 | #include <linux/percpu.h> | |
57 | #include <linux/kthread.h> | |
58 | #include <linux/seq_file.h> | |
e692ab53 | 59 | #include <linux/sysctl.h> |
1da177e4 LT |
60 | #include <linux/syscalls.h> |
61 | #include <linux/times.h> | |
8f0ab514 | 62 | #include <linux/tsacct_kern.h> |
c6fd91f0 | 63 | #include <linux/kprobes.h> |
0ff92245 | 64 | #include <linux/delayacct.h> |
5517d86b | 65 | #include <linux/reciprocal_div.h> |
dff06c15 | 66 | #include <linux/unistd.h> |
f5ff8422 | 67 | #include <linux/pagemap.h> |
8f4d37ec | 68 | #include <linux/hrtimer.h> |
30914a58 | 69 | #include <linux/tick.h> |
434d53b0 | 70 | #include <linux/bootmem.h> |
f00b45c1 PZ |
71 | #include <linux/debugfs.h> |
72 | #include <linux/ctype.h> | |
1da177e4 | 73 | |
5517d86b | 74 | #include <asm/tlb.h> |
838225b4 | 75 | #include <asm/irq_regs.h> |
1da177e4 | 76 | |
b035b6de AD |
77 | /* |
78 | * Scheduler clock - returns current time in nanosec units. | |
79 | * This is default implementation. | |
80 | * Architectures and sub-architectures can override this. | |
81 | */ | |
82 | unsigned long long __attribute__((weak)) sched_clock(void) | |
83 | { | |
d6322faf | 84 | return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ); |
b035b6de AD |
85 | } |
86 | ||
1da177e4 LT |
87 | /* |
88 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
89 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
90 | * and back. | |
91 | */ | |
92 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
93 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
94 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
95 | ||
96 | /* | |
97 | * 'User priority' is the nice value converted to something we | |
98 | * can work with better when scaling various scheduler parameters, | |
99 | * it's a [ 0 ... 39 ] range. | |
100 | */ | |
101 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
102 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
103 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
104 | ||
105 | /* | |
d7876a08 | 106 | * Helpers for converting nanosecond timing to jiffy resolution |
1da177e4 | 107 | */ |
d6322faf | 108 | #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) |
1da177e4 | 109 | |
6aa645ea IM |
110 | #define NICE_0_LOAD SCHED_LOAD_SCALE |
111 | #define NICE_0_SHIFT SCHED_LOAD_SHIFT | |
112 | ||
1da177e4 LT |
113 | /* |
114 | * These are the 'tuning knobs' of the scheduler: | |
115 | * | |
a4ec24b4 | 116 | * default timeslice is 100 msecs (used only for SCHED_RR tasks). |
1da177e4 LT |
117 | * Timeslices get refilled after they expire. |
118 | */ | |
1da177e4 | 119 | #define DEF_TIMESLICE (100 * HZ / 1000) |
2dd73a4f | 120 | |
d0b27fa7 PZ |
121 | /* |
122 | * single value that denotes runtime == period, ie unlimited time. | |
123 | */ | |
124 | #define RUNTIME_INF ((u64)~0ULL) | |
125 | ||
5517d86b ED |
126 | #ifdef CONFIG_SMP |
127 | /* | |
128 | * Divide a load by a sched group cpu_power : (load / sg->__cpu_power) | |
129 | * Since cpu_power is a 'constant', we can use a reciprocal divide. | |
130 | */ | |
131 | static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load) | |
132 | { | |
133 | return reciprocal_divide(load, sg->reciprocal_cpu_power); | |
134 | } | |
135 | ||
136 | /* | |
137 | * Each time a sched group cpu_power is changed, | |
138 | * we must compute its reciprocal value | |
139 | */ | |
140 | static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val) | |
141 | { | |
142 | sg->__cpu_power += val; | |
143 | sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power); | |
144 | } | |
145 | #endif | |
146 | ||
e05606d3 IM |
147 | static inline int rt_policy(int policy) |
148 | { | |
149 | if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR)) | |
150 | return 1; | |
151 | return 0; | |
152 | } | |
153 | ||
154 | static inline int task_has_rt_policy(struct task_struct *p) | |
155 | { | |
156 | return rt_policy(p->policy); | |
157 | } | |
158 | ||
1da177e4 | 159 | /* |
6aa645ea | 160 | * This is the priority-queue data structure of the RT scheduling class: |
1da177e4 | 161 | */ |
6aa645ea IM |
162 | struct rt_prio_array { |
163 | DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ | |
164 | struct list_head queue[MAX_RT_PRIO]; | |
165 | }; | |
166 | ||
d0b27fa7 | 167 | struct rt_bandwidth { |
ea736ed5 IM |
168 | /* nests inside the rq lock: */ |
169 | spinlock_t rt_runtime_lock; | |
170 | ktime_t rt_period; | |
171 | u64 rt_runtime; | |
172 | struct hrtimer rt_period_timer; | |
d0b27fa7 PZ |
173 | }; |
174 | ||
175 | static struct rt_bandwidth def_rt_bandwidth; | |
176 | ||
177 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); | |
178 | ||
179 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) | |
180 | { | |
181 | struct rt_bandwidth *rt_b = | |
182 | container_of(timer, struct rt_bandwidth, rt_period_timer); | |
183 | ktime_t now; | |
184 | int overrun; | |
185 | int idle = 0; | |
186 | ||
187 | for (;;) { | |
188 | now = hrtimer_cb_get_time(timer); | |
189 | overrun = hrtimer_forward(timer, now, rt_b->rt_period); | |
190 | ||
191 | if (!overrun) | |
192 | break; | |
193 | ||
194 | idle = do_sched_rt_period_timer(rt_b, overrun); | |
195 | } | |
196 | ||
197 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
198 | } | |
199 | ||
200 | static | |
201 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) | |
202 | { | |
203 | rt_b->rt_period = ns_to_ktime(period); | |
204 | rt_b->rt_runtime = runtime; | |
205 | ||
ac086bc2 PZ |
206 | spin_lock_init(&rt_b->rt_runtime_lock); |
207 | ||
d0b27fa7 PZ |
208 | hrtimer_init(&rt_b->rt_period_timer, |
209 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
210 | rt_b->rt_period_timer.function = sched_rt_period_timer; | |
211 | rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
212 | } | |
213 | ||
214 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) | |
215 | { | |
216 | ktime_t now; | |
217 | ||
218 | if (rt_b->rt_runtime == RUNTIME_INF) | |
219 | return; | |
220 | ||
221 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
222 | return; | |
223 | ||
224 | spin_lock(&rt_b->rt_runtime_lock); | |
225 | for (;;) { | |
226 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
227 | break; | |
228 | ||
229 | now = hrtimer_cb_get_time(&rt_b->rt_period_timer); | |
230 | hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period); | |
231 | hrtimer_start(&rt_b->rt_period_timer, | |
232 | rt_b->rt_period_timer.expires, | |
233 | HRTIMER_MODE_ABS); | |
234 | } | |
235 | spin_unlock(&rt_b->rt_runtime_lock); | |
236 | } | |
237 | ||
238 | #ifdef CONFIG_RT_GROUP_SCHED | |
239 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) | |
240 | { | |
241 | hrtimer_cancel(&rt_b->rt_period_timer); | |
242 | } | |
243 | #endif | |
244 | ||
052f1dc7 | 245 | #ifdef CONFIG_GROUP_SCHED |
29f59db3 | 246 | |
68318b8e SV |
247 | #include <linux/cgroup.h> |
248 | ||
29f59db3 SV |
249 | struct cfs_rq; |
250 | ||
6f505b16 PZ |
251 | static LIST_HEAD(task_groups); |
252 | ||
29f59db3 | 253 | /* task group related information */ |
4cf86d77 | 254 | struct task_group { |
052f1dc7 | 255 | #ifdef CONFIG_CGROUP_SCHED |
68318b8e SV |
256 | struct cgroup_subsys_state css; |
257 | #endif | |
052f1dc7 PZ |
258 | |
259 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
29f59db3 SV |
260 | /* schedulable entities of this group on each cpu */ |
261 | struct sched_entity **se; | |
262 | /* runqueue "owned" by this group on each cpu */ | |
263 | struct cfs_rq **cfs_rq; | |
264 | unsigned long shares; | |
052f1dc7 PZ |
265 | #endif |
266 | ||
267 | #ifdef CONFIG_RT_GROUP_SCHED | |
268 | struct sched_rt_entity **rt_se; | |
269 | struct rt_rq **rt_rq; | |
270 | ||
d0b27fa7 | 271 | struct rt_bandwidth rt_bandwidth; |
052f1dc7 | 272 | #endif |
6b2d7700 | 273 | |
ae8393e5 | 274 | struct rcu_head rcu; |
6f505b16 | 275 | struct list_head list; |
f473aa5e PZ |
276 | |
277 | struct task_group *parent; | |
278 | struct list_head siblings; | |
279 | struct list_head children; | |
29f59db3 SV |
280 | }; |
281 | ||
354d60c2 | 282 | #ifdef CONFIG_USER_SCHED |
eff766a6 PZ |
283 | |
284 | /* | |
285 | * Root task group. | |
286 | * Every UID task group (including init_task_group aka UID-0) will | |
287 | * be a child to this group. | |
288 | */ | |
289 | struct task_group root_task_group; | |
290 | ||
052f1dc7 | 291 | #ifdef CONFIG_FAIR_GROUP_SCHED |
29f59db3 SV |
292 | /* Default task group's sched entity on each cpu */ |
293 | static DEFINE_PER_CPU(struct sched_entity, init_sched_entity); | |
294 | /* Default task group's cfs_rq on each cpu */ | |
295 | static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp; | |
052f1dc7 PZ |
296 | #endif |
297 | ||
298 | #ifdef CONFIG_RT_GROUP_SCHED | |
299 | static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity); | |
300 | static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp; | |
052f1dc7 | 301 | #endif |
eff766a6 PZ |
302 | #else |
303 | #define root_task_group init_task_group | |
354d60c2 | 304 | #endif |
6f505b16 | 305 | |
8ed36996 | 306 | /* task_group_lock serializes add/remove of task groups and also changes to |
ec2c507f SV |
307 | * a task group's cpu shares. |
308 | */ | |
8ed36996 | 309 | static DEFINE_SPINLOCK(task_group_lock); |
ec2c507f | 310 | |
a1835615 SV |
311 | /* doms_cur_mutex serializes access to doms_cur[] array */ |
312 | static DEFINE_MUTEX(doms_cur_mutex); | |
313 | ||
052f1dc7 | 314 | #ifdef CONFIG_FAIR_GROUP_SCHED |
052f1dc7 PZ |
315 | #ifdef CONFIG_USER_SCHED |
316 | # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD) | |
317 | #else | |
318 | # define INIT_TASK_GROUP_LOAD NICE_0_LOAD | |
319 | #endif | |
320 | ||
18d95a28 PZ |
321 | #define MIN_SHARES 2 |
322 | ||
052f1dc7 PZ |
323 | static int init_task_group_load = INIT_TASK_GROUP_LOAD; |
324 | #endif | |
325 | ||
29f59db3 | 326 | /* Default task group. |
3a252015 | 327 | * Every task in system belong to this group at bootup. |
29f59db3 | 328 | */ |
434d53b0 | 329 | struct task_group init_task_group; |
29f59db3 SV |
330 | |
331 | /* return group to which a task belongs */ | |
4cf86d77 | 332 | static inline struct task_group *task_group(struct task_struct *p) |
29f59db3 | 333 | { |
4cf86d77 | 334 | struct task_group *tg; |
9b5b7751 | 335 | |
052f1dc7 | 336 | #ifdef CONFIG_USER_SCHED |
24e377a8 | 337 | tg = p->user->tg; |
052f1dc7 | 338 | #elif defined(CONFIG_CGROUP_SCHED) |
68318b8e SV |
339 | tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id), |
340 | struct task_group, css); | |
24e377a8 | 341 | #else |
41a2d6cf | 342 | tg = &init_task_group; |
24e377a8 | 343 | #endif |
9b5b7751 | 344 | return tg; |
29f59db3 SV |
345 | } |
346 | ||
347 | /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ | |
6f505b16 | 348 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) |
29f59db3 | 349 | { |
052f1dc7 | 350 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ce96b5ac DA |
351 | p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; |
352 | p->se.parent = task_group(p)->se[cpu]; | |
052f1dc7 | 353 | #endif |
6f505b16 | 354 | |
052f1dc7 | 355 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
356 | p->rt.rt_rq = task_group(p)->rt_rq[cpu]; |
357 | p->rt.parent = task_group(p)->rt_se[cpu]; | |
052f1dc7 | 358 | #endif |
29f59db3 SV |
359 | } |
360 | ||
a1835615 SV |
361 | static inline void lock_doms_cur(void) |
362 | { | |
363 | mutex_lock(&doms_cur_mutex); | |
364 | } | |
365 | ||
366 | static inline void unlock_doms_cur(void) | |
367 | { | |
368 | mutex_unlock(&doms_cur_mutex); | |
369 | } | |
370 | ||
29f59db3 SV |
371 | #else |
372 | ||
6f505b16 | 373 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } |
a1835615 SV |
374 | static inline void lock_doms_cur(void) { } |
375 | static inline void unlock_doms_cur(void) { } | |
29f59db3 | 376 | |
052f1dc7 | 377 | #endif /* CONFIG_GROUP_SCHED */ |
29f59db3 | 378 | |
6aa645ea IM |
379 | /* CFS-related fields in a runqueue */ |
380 | struct cfs_rq { | |
381 | struct load_weight load; | |
382 | unsigned long nr_running; | |
383 | ||
6aa645ea | 384 | u64 exec_clock; |
e9acbff6 | 385 | u64 min_vruntime; |
6aa645ea IM |
386 | |
387 | struct rb_root tasks_timeline; | |
388 | struct rb_node *rb_leftmost; | |
4a55bd5e PZ |
389 | |
390 | struct list_head tasks; | |
391 | struct list_head *balance_iterator; | |
392 | ||
393 | /* | |
394 | * 'curr' points to currently running entity on this cfs_rq. | |
6aa645ea IM |
395 | * It is set to NULL otherwise (i.e when none are currently running). |
396 | */ | |
aa2ac252 | 397 | struct sched_entity *curr, *next; |
ddc97297 PZ |
398 | |
399 | unsigned long nr_spread_over; | |
400 | ||
62160e3f | 401 | #ifdef CONFIG_FAIR_GROUP_SCHED |
6aa645ea IM |
402 | struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ |
403 | ||
41a2d6cf IM |
404 | /* |
405 | * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in | |
6aa645ea IM |
406 | * a hierarchy). Non-leaf lrqs hold other higher schedulable entities |
407 | * (like users, containers etc.) | |
408 | * | |
409 | * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This | |
410 | * list is used during load balance. | |
411 | */ | |
41a2d6cf IM |
412 | struct list_head leaf_cfs_rq_list; |
413 | struct task_group *tg; /* group that "owns" this runqueue */ | |
18d95a28 PZ |
414 | |
415 | #ifdef CONFIG_SMP | |
416 | unsigned long task_weight; | |
417 | unsigned long shares; | |
418 | /* | |
419 | * We need space to build a sched_domain wide view of the full task | |
420 | * group tree, in order to avoid depending on dynamic memory allocation | |
421 | * during the load balancing we place this in the per cpu task group | |
422 | * hierarchy. This limits the load balancing to one instance per cpu, | |
423 | * but more should not be needed anyway. | |
424 | */ | |
425 | struct aggregate_struct { | |
426 | /* | |
427 | * load = weight(cpus) * f(tg) | |
428 | * | |
429 | * Where f(tg) is the recursive weight fraction assigned to | |
430 | * this group. | |
431 | */ | |
432 | unsigned long load; | |
433 | ||
434 | /* | |
435 | * part of the group weight distributed to this span. | |
436 | */ | |
437 | unsigned long shares; | |
438 | ||
439 | /* | |
440 | * The sum of all runqueue weights within this span. | |
441 | */ | |
442 | unsigned long rq_weight; | |
443 | ||
444 | /* | |
445 | * Weight contributed by tasks; this is the part we can | |
446 | * influence by moving tasks around. | |
447 | */ | |
448 | unsigned long task_weight; | |
449 | } aggregate; | |
450 | #endif | |
6aa645ea IM |
451 | #endif |
452 | }; | |
1da177e4 | 453 | |
6aa645ea IM |
454 | /* Real-Time classes' related field in a runqueue: */ |
455 | struct rt_rq { | |
456 | struct rt_prio_array active; | |
63489e45 | 457 | unsigned long rt_nr_running; |
052f1dc7 | 458 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
459 | int highest_prio; /* highest queued rt task prio */ |
460 | #endif | |
fa85ae24 | 461 | #ifdef CONFIG_SMP |
73fe6aae | 462 | unsigned long rt_nr_migratory; |
a22d7fc1 | 463 | int overloaded; |
fa85ae24 | 464 | #endif |
6f505b16 | 465 | int rt_throttled; |
fa85ae24 | 466 | u64 rt_time; |
ac086bc2 | 467 | u64 rt_runtime; |
ea736ed5 | 468 | /* Nests inside the rq lock: */ |
ac086bc2 | 469 | spinlock_t rt_runtime_lock; |
6f505b16 | 470 | |
052f1dc7 | 471 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc PZ |
472 | unsigned long rt_nr_boosted; |
473 | ||
6f505b16 PZ |
474 | struct rq *rq; |
475 | struct list_head leaf_rt_rq_list; | |
476 | struct task_group *tg; | |
477 | struct sched_rt_entity *rt_se; | |
478 | #endif | |
6aa645ea IM |
479 | }; |
480 | ||
57d885fe GH |
481 | #ifdef CONFIG_SMP |
482 | ||
483 | /* | |
484 | * We add the notion of a root-domain which will be used to define per-domain | |
0eab9146 IM |
485 | * variables. Each exclusive cpuset essentially defines an island domain by |
486 | * fully partitioning the member cpus from any other cpuset. Whenever a new | |
57d885fe GH |
487 | * exclusive cpuset is created, we also create and attach a new root-domain |
488 | * object. | |
489 | * | |
57d885fe GH |
490 | */ |
491 | struct root_domain { | |
492 | atomic_t refcount; | |
493 | cpumask_t span; | |
494 | cpumask_t online; | |
637f5085 | 495 | |
0eab9146 | 496 | /* |
637f5085 GH |
497 | * The "RT overload" flag: it gets set if a CPU has more than |
498 | * one runnable RT task. | |
499 | */ | |
500 | cpumask_t rto_mask; | |
0eab9146 | 501 | atomic_t rto_count; |
57d885fe GH |
502 | }; |
503 | ||
dc938520 GH |
504 | /* |
505 | * By default the system creates a single root-domain with all cpus as | |
506 | * members (mimicking the global state we have today). | |
507 | */ | |
57d885fe GH |
508 | static struct root_domain def_root_domain; |
509 | ||
510 | #endif | |
511 | ||
1da177e4 LT |
512 | /* |
513 | * This is the main, per-CPU runqueue data structure. | |
514 | * | |
515 | * Locking rule: those places that want to lock multiple runqueues | |
516 | * (such as the load balancing or the thread migration code), lock | |
517 | * acquire operations must be ordered by ascending &runqueue. | |
518 | */ | |
70b97a7f | 519 | struct rq { |
d8016491 IM |
520 | /* runqueue lock: */ |
521 | spinlock_t lock; | |
1da177e4 LT |
522 | |
523 | /* | |
524 | * nr_running and cpu_load should be in the same cacheline because | |
525 | * remote CPUs use both these fields when doing load calculation. | |
526 | */ | |
527 | unsigned long nr_running; | |
6aa645ea IM |
528 | #define CPU_LOAD_IDX_MAX 5 |
529 | unsigned long cpu_load[CPU_LOAD_IDX_MAX]; | |
bdecea3a | 530 | unsigned char idle_at_tick; |
46cb4b7c | 531 | #ifdef CONFIG_NO_HZ |
15934a37 | 532 | unsigned long last_tick_seen; |
46cb4b7c SS |
533 | unsigned char in_nohz_recently; |
534 | #endif | |
d8016491 IM |
535 | /* capture load from *all* tasks on this cpu: */ |
536 | struct load_weight load; | |
6aa645ea IM |
537 | unsigned long nr_load_updates; |
538 | u64 nr_switches; | |
539 | ||
540 | struct cfs_rq cfs; | |
6f505b16 | 541 | struct rt_rq rt; |
6f505b16 | 542 | |
6aa645ea | 543 | #ifdef CONFIG_FAIR_GROUP_SCHED |
d8016491 IM |
544 | /* list of leaf cfs_rq on this cpu: */ |
545 | struct list_head leaf_cfs_rq_list; | |
052f1dc7 PZ |
546 | #endif |
547 | #ifdef CONFIG_RT_GROUP_SCHED | |
6f505b16 | 548 | struct list_head leaf_rt_rq_list; |
1da177e4 | 549 | #endif |
1da177e4 LT |
550 | |
551 | /* | |
552 | * This is part of a global counter where only the total sum | |
553 | * over all CPUs matters. A task can increase this counter on | |
554 | * one CPU and if it got migrated afterwards it may decrease | |
555 | * it on another CPU. Always updated under the runqueue lock: | |
556 | */ | |
557 | unsigned long nr_uninterruptible; | |
558 | ||
36c8b586 | 559 | struct task_struct *curr, *idle; |
c9819f45 | 560 | unsigned long next_balance; |
1da177e4 | 561 | struct mm_struct *prev_mm; |
6aa645ea | 562 | |
6aa645ea IM |
563 | u64 clock, prev_clock_raw; |
564 | s64 clock_max_delta; | |
565 | ||
cc203d24 | 566 | unsigned int clock_warps, clock_overflows, clock_underflows; |
2aa44d05 IM |
567 | u64 idle_clock; |
568 | unsigned int clock_deep_idle_events; | |
529c7726 | 569 | u64 tick_timestamp; |
6aa645ea | 570 | |
1da177e4 LT |
571 | atomic_t nr_iowait; |
572 | ||
573 | #ifdef CONFIG_SMP | |
0eab9146 | 574 | struct root_domain *rd; |
1da177e4 LT |
575 | struct sched_domain *sd; |
576 | ||
577 | /* For active balancing */ | |
578 | int active_balance; | |
579 | int push_cpu; | |
d8016491 IM |
580 | /* cpu of this runqueue: */ |
581 | int cpu; | |
1da177e4 | 582 | |
36c8b586 | 583 | struct task_struct *migration_thread; |
1da177e4 LT |
584 | struct list_head migration_queue; |
585 | #endif | |
586 | ||
8f4d37ec PZ |
587 | #ifdef CONFIG_SCHED_HRTICK |
588 | unsigned long hrtick_flags; | |
589 | ktime_t hrtick_expire; | |
590 | struct hrtimer hrtick_timer; | |
591 | #endif | |
592 | ||
1da177e4 LT |
593 | #ifdef CONFIG_SCHEDSTATS |
594 | /* latency stats */ | |
595 | struct sched_info rq_sched_info; | |
596 | ||
597 | /* sys_sched_yield() stats */ | |
480b9434 KC |
598 | unsigned int yld_exp_empty; |
599 | unsigned int yld_act_empty; | |
600 | unsigned int yld_both_empty; | |
601 | unsigned int yld_count; | |
1da177e4 LT |
602 | |
603 | /* schedule() stats */ | |
480b9434 KC |
604 | unsigned int sched_switch; |
605 | unsigned int sched_count; | |
606 | unsigned int sched_goidle; | |
1da177e4 LT |
607 | |
608 | /* try_to_wake_up() stats */ | |
480b9434 KC |
609 | unsigned int ttwu_count; |
610 | unsigned int ttwu_local; | |
b8efb561 IM |
611 | |
612 | /* BKL stats */ | |
480b9434 | 613 | unsigned int bkl_count; |
1da177e4 | 614 | #endif |
fcb99371 | 615 | struct lock_class_key rq_lock_key; |
1da177e4 LT |
616 | }; |
617 | ||
f34e3b61 | 618 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); |
1da177e4 | 619 | |
dd41f596 IM |
620 | static inline void check_preempt_curr(struct rq *rq, struct task_struct *p) |
621 | { | |
622 | rq->curr->sched_class->check_preempt_curr(rq, p); | |
623 | } | |
624 | ||
0a2966b4 CL |
625 | static inline int cpu_of(struct rq *rq) |
626 | { | |
627 | #ifdef CONFIG_SMP | |
628 | return rq->cpu; | |
629 | #else | |
630 | return 0; | |
631 | #endif | |
632 | } | |
633 | ||
15934a37 GC |
634 | #ifdef CONFIG_NO_HZ |
635 | static inline bool nohz_on(int cpu) | |
636 | { | |
637 | return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE; | |
638 | } | |
639 | ||
640 | static inline u64 max_skipped_ticks(struct rq *rq) | |
641 | { | |
642 | return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1; | |
643 | } | |
644 | ||
645 | static inline void update_last_tick_seen(struct rq *rq) | |
646 | { | |
647 | rq->last_tick_seen = jiffies; | |
648 | } | |
649 | #else | |
650 | static inline u64 max_skipped_ticks(struct rq *rq) | |
651 | { | |
652 | return 1; | |
653 | } | |
654 | ||
655 | static inline void update_last_tick_seen(struct rq *rq) | |
656 | { | |
657 | } | |
658 | #endif | |
659 | ||
20d315d4 | 660 | /* |
b04a0f4c IM |
661 | * Update the per-runqueue clock, as finegrained as the platform can give |
662 | * us, but without assuming monotonicity, etc.: | |
20d315d4 | 663 | */ |
b04a0f4c | 664 | static void __update_rq_clock(struct rq *rq) |
20d315d4 IM |
665 | { |
666 | u64 prev_raw = rq->prev_clock_raw; | |
667 | u64 now = sched_clock(); | |
668 | s64 delta = now - prev_raw; | |
669 | u64 clock = rq->clock; | |
670 | ||
b04a0f4c IM |
671 | #ifdef CONFIG_SCHED_DEBUG |
672 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
673 | #endif | |
20d315d4 IM |
674 | /* |
675 | * Protect against sched_clock() occasionally going backwards: | |
676 | */ | |
677 | if (unlikely(delta < 0)) { | |
678 | clock++; | |
679 | rq->clock_warps++; | |
680 | } else { | |
681 | /* | |
682 | * Catch too large forward jumps too: | |
683 | */ | |
15934a37 GC |
684 | u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC; |
685 | u64 max_time = rq->tick_timestamp + max_jump; | |
686 | ||
687 | if (unlikely(clock + delta > max_time)) { | |
688 | if (clock < max_time) | |
689 | clock = max_time; | |
529c7726 IM |
690 | else |
691 | clock++; | |
20d315d4 IM |
692 | rq->clock_overflows++; |
693 | } else { | |
694 | if (unlikely(delta > rq->clock_max_delta)) | |
695 | rq->clock_max_delta = delta; | |
696 | clock += delta; | |
697 | } | |
698 | } | |
699 | ||
700 | rq->prev_clock_raw = now; | |
701 | rq->clock = clock; | |
b04a0f4c | 702 | } |
20d315d4 | 703 | |
b04a0f4c IM |
704 | static void update_rq_clock(struct rq *rq) |
705 | { | |
706 | if (likely(smp_processor_id() == cpu_of(rq))) | |
707 | __update_rq_clock(rq); | |
20d315d4 IM |
708 | } |
709 | ||
674311d5 NP |
710 | /* |
711 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 712 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
713 | * |
714 | * The domain tree of any CPU may only be accessed from within | |
715 | * preempt-disabled sections. | |
716 | */ | |
48f24c4d IM |
717 | #define for_each_domain(cpu, __sd) \ |
718 | for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) | |
1da177e4 LT |
719 | |
720 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
721 | #define this_rq() (&__get_cpu_var(runqueues)) | |
722 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
723 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
724 | ||
bf5c91ba IM |
725 | /* |
726 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: | |
727 | */ | |
728 | #ifdef CONFIG_SCHED_DEBUG | |
729 | # define const_debug __read_mostly | |
730 | #else | |
731 | # define const_debug static const | |
732 | #endif | |
733 | ||
734 | /* | |
735 | * Debugging: various feature bits | |
736 | */ | |
f00b45c1 PZ |
737 | |
738 | #define SCHED_FEAT(name, enabled) \ | |
739 | __SCHED_FEAT_##name , | |
740 | ||
bf5c91ba | 741 | enum { |
f00b45c1 | 742 | #include "sched_features.h" |
bf5c91ba IM |
743 | }; |
744 | ||
f00b45c1 PZ |
745 | #undef SCHED_FEAT |
746 | ||
747 | #define SCHED_FEAT(name, enabled) \ | |
748 | (1UL << __SCHED_FEAT_##name) * enabled | | |
749 | ||
bf5c91ba | 750 | const_debug unsigned int sysctl_sched_features = |
f00b45c1 PZ |
751 | #include "sched_features.h" |
752 | 0; | |
753 | ||
754 | #undef SCHED_FEAT | |
755 | ||
756 | #ifdef CONFIG_SCHED_DEBUG | |
757 | #define SCHED_FEAT(name, enabled) \ | |
758 | #name , | |
759 | ||
983ed7a6 | 760 | static __read_mostly char *sched_feat_names[] = { |
f00b45c1 PZ |
761 | #include "sched_features.h" |
762 | NULL | |
763 | }; | |
764 | ||
765 | #undef SCHED_FEAT | |
766 | ||
983ed7a6 | 767 | static int sched_feat_open(struct inode *inode, struct file *filp) |
f00b45c1 PZ |
768 | { |
769 | filp->private_data = inode->i_private; | |
770 | return 0; | |
771 | } | |
772 | ||
773 | static ssize_t | |
774 | sched_feat_read(struct file *filp, char __user *ubuf, | |
775 | size_t cnt, loff_t *ppos) | |
776 | { | |
777 | char *buf; | |
778 | int r = 0; | |
779 | int len = 0; | |
780 | int i; | |
781 | ||
782 | for (i = 0; sched_feat_names[i]; i++) { | |
783 | len += strlen(sched_feat_names[i]); | |
784 | len += 4; | |
785 | } | |
786 | ||
787 | buf = kmalloc(len + 2, GFP_KERNEL); | |
788 | if (!buf) | |
789 | return -ENOMEM; | |
790 | ||
791 | for (i = 0; sched_feat_names[i]; i++) { | |
792 | if (sysctl_sched_features & (1UL << i)) | |
793 | r += sprintf(buf + r, "%s ", sched_feat_names[i]); | |
794 | else | |
c24b7c52 | 795 | r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]); |
f00b45c1 PZ |
796 | } |
797 | ||
798 | r += sprintf(buf + r, "\n"); | |
799 | WARN_ON(r >= len + 2); | |
800 | ||
801 | r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r); | |
802 | ||
803 | kfree(buf); | |
804 | ||
805 | return r; | |
806 | } | |
807 | ||
808 | static ssize_t | |
809 | sched_feat_write(struct file *filp, const char __user *ubuf, | |
810 | size_t cnt, loff_t *ppos) | |
811 | { | |
812 | char buf[64]; | |
813 | char *cmp = buf; | |
814 | int neg = 0; | |
815 | int i; | |
816 | ||
817 | if (cnt > 63) | |
818 | cnt = 63; | |
819 | ||
820 | if (copy_from_user(&buf, ubuf, cnt)) | |
821 | return -EFAULT; | |
822 | ||
823 | buf[cnt] = 0; | |
824 | ||
c24b7c52 | 825 | if (strncmp(buf, "NO_", 3) == 0) { |
f00b45c1 PZ |
826 | neg = 1; |
827 | cmp += 3; | |
828 | } | |
829 | ||
830 | for (i = 0; sched_feat_names[i]; i++) { | |
831 | int len = strlen(sched_feat_names[i]); | |
832 | ||
833 | if (strncmp(cmp, sched_feat_names[i], len) == 0) { | |
834 | if (neg) | |
835 | sysctl_sched_features &= ~(1UL << i); | |
836 | else | |
837 | sysctl_sched_features |= (1UL << i); | |
838 | break; | |
839 | } | |
840 | } | |
841 | ||
842 | if (!sched_feat_names[i]) | |
843 | return -EINVAL; | |
844 | ||
845 | filp->f_pos += cnt; | |
846 | ||
847 | return cnt; | |
848 | } | |
849 | ||
850 | static struct file_operations sched_feat_fops = { | |
851 | .open = sched_feat_open, | |
852 | .read = sched_feat_read, | |
853 | .write = sched_feat_write, | |
854 | }; | |
855 | ||
856 | static __init int sched_init_debug(void) | |
857 | { | |
f00b45c1 PZ |
858 | debugfs_create_file("sched_features", 0644, NULL, NULL, |
859 | &sched_feat_fops); | |
860 | ||
861 | return 0; | |
862 | } | |
863 | late_initcall(sched_init_debug); | |
864 | ||
865 | #endif | |
866 | ||
867 | #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) | |
bf5c91ba | 868 | |
b82d9fdd PZ |
869 | /* |
870 | * Number of tasks to iterate in a single balance run. | |
871 | * Limited because this is done with IRQs disabled. | |
872 | */ | |
873 | const_debug unsigned int sysctl_sched_nr_migrate = 32; | |
874 | ||
fa85ae24 | 875 | /* |
9f0c1e56 | 876 | * period over which we measure -rt task cpu usage in us. |
fa85ae24 PZ |
877 | * default: 1s |
878 | */ | |
9f0c1e56 | 879 | unsigned int sysctl_sched_rt_period = 1000000; |
fa85ae24 | 880 | |
6892b75e IM |
881 | static __read_mostly int scheduler_running; |
882 | ||
9f0c1e56 PZ |
883 | /* |
884 | * part of the period that we allow rt tasks to run in us. | |
885 | * default: 0.95s | |
886 | */ | |
887 | int sysctl_sched_rt_runtime = 950000; | |
fa85ae24 | 888 | |
d0b27fa7 PZ |
889 | static inline u64 global_rt_period(void) |
890 | { | |
891 | return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; | |
892 | } | |
893 | ||
894 | static inline u64 global_rt_runtime(void) | |
895 | { | |
896 | if (sysctl_sched_rt_period < 0) | |
897 | return RUNTIME_INF; | |
898 | ||
899 | return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; | |
900 | } | |
fa85ae24 | 901 | |
27ec4407 IM |
902 | static const unsigned long long time_sync_thresh = 100000; |
903 | ||
904 | static DEFINE_PER_CPU(unsigned long long, time_offset); | |
905 | static DEFINE_PER_CPU(unsigned long long, prev_cpu_time); | |
906 | ||
e436d800 | 907 | /* |
27ec4407 IM |
908 | * Global lock which we take every now and then to synchronize |
909 | * the CPUs time. This method is not warp-safe, but it's good | |
910 | * enough to synchronize slowly diverging time sources and thus | |
911 | * it's good enough for tracing: | |
e436d800 | 912 | */ |
27ec4407 IM |
913 | static DEFINE_SPINLOCK(time_sync_lock); |
914 | static unsigned long long prev_global_time; | |
915 | ||
916 | static unsigned long long __sync_cpu_clock(cycles_t time, int cpu) | |
917 | { | |
918 | unsigned long flags; | |
919 | ||
920 | spin_lock_irqsave(&time_sync_lock, flags); | |
921 | ||
922 | if (time < prev_global_time) { | |
923 | per_cpu(time_offset, cpu) += prev_global_time - time; | |
924 | time = prev_global_time; | |
925 | } else { | |
926 | prev_global_time = time; | |
927 | } | |
928 | ||
929 | spin_unlock_irqrestore(&time_sync_lock, flags); | |
930 | ||
931 | return time; | |
932 | } | |
933 | ||
934 | static unsigned long long __cpu_clock(int cpu) | |
e436d800 | 935 | { |
e436d800 IM |
936 | unsigned long long now; |
937 | unsigned long flags; | |
b04a0f4c | 938 | struct rq *rq; |
e436d800 | 939 | |
8ced5f69 IM |
940 | /* |
941 | * Only call sched_clock() if the scheduler has already been | |
942 | * initialized (some code might call cpu_clock() very early): | |
943 | */ | |
6892b75e IM |
944 | if (unlikely(!scheduler_running)) |
945 | return 0; | |
946 | ||
947 | local_irq_save(flags); | |
948 | rq = cpu_rq(cpu); | |
949 | update_rq_clock(rq); | |
b04a0f4c | 950 | now = rq->clock; |
2cd4d0ea | 951 | local_irq_restore(flags); |
e436d800 IM |
952 | |
953 | return now; | |
954 | } | |
27ec4407 IM |
955 | |
956 | /* | |
957 | * For kernel-internal use: high-speed (but slightly incorrect) per-cpu | |
958 | * clock constructed from sched_clock(): | |
959 | */ | |
960 | unsigned long long cpu_clock(int cpu) | |
961 | { | |
962 | unsigned long long prev_cpu_time, time, delta_time; | |
963 | ||
964 | prev_cpu_time = per_cpu(prev_cpu_time, cpu); | |
965 | time = __cpu_clock(cpu) + per_cpu(time_offset, cpu); | |
966 | delta_time = time-prev_cpu_time; | |
967 | ||
968 | if (unlikely(delta_time > time_sync_thresh)) | |
969 | time = __sync_cpu_clock(time, cpu); | |
970 | ||
971 | return time; | |
972 | } | |
a58f6f25 | 973 | EXPORT_SYMBOL_GPL(cpu_clock); |
e436d800 | 974 | |
1da177e4 | 975 | #ifndef prepare_arch_switch |
4866cde0 NP |
976 | # define prepare_arch_switch(next) do { } while (0) |
977 | #endif | |
978 | #ifndef finish_arch_switch | |
979 | # define finish_arch_switch(prev) do { } while (0) | |
980 | #endif | |
981 | ||
051a1d1a DA |
982 | static inline int task_current(struct rq *rq, struct task_struct *p) |
983 | { | |
984 | return rq->curr == p; | |
985 | } | |
986 | ||
4866cde0 | 987 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
70b97a7f | 988 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 | 989 | { |
051a1d1a | 990 | return task_current(rq, p); |
4866cde0 NP |
991 | } |
992 | ||
70b97a7f | 993 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
994 | { |
995 | } | |
996 | ||
70b97a7f | 997 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 | 998 | { |
da04c035 IM |
999 | #ifdef CONFIG_DEBUG_SPINLOCK |
1000 | /* this is a valid case when another task releases the spinlock */ | |
1001 | rq->lock.owner = current; | |
1002 | #endif | |
8a25d5de IM |
1003 | /* |
1004 | * If we are tracking spinlock dependencies then we have to | |
1005 | * fix up the runqueue lock - which gets 'carried over' from | |
1006 | * prev into current: | |
1007 | */ | |
1008 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); | |
1009 | ||
4866cde0 NP |
1010 | spin_unlock_irq(&rq->lock); |
1011 | } | |
1012 | ||
1013 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
70b97a7f | 1014 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 NP |
1015 | { |
1016 | #ifdef CONFIG_SMP | |
1017 | return p->oncpu; | |
1018 | #else | |
051a1d1a | 1019 | return task_current(rq, p); |
4866cde0 NP |
1020 | #endif |
1021 | } | |
1022 | ||
70b97a7f | 1023 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
1024 | { |
1025 | #ifdef CONFIG_SMP | |
1026 | /* | |
1027 | * We can optimise this out completely for !SMP, because the | |
1028 | * SMP rebalancing from interrupt is the only thing that cares | |
1029 | * here. | |
1030 | */ | |
1031 | next->oncpu = 1; | |
1032 | #endif | |
1033 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
1034 | spin_unlock_irq(&rq->lock); | |
1035 | #else | |
1036 | spin_unlock(&rq->lock); | |
1037 | #endif | |
1038 | } | |
1039 | ||
70b97a7f | 1040 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 NP |
1041 | { |
1042 | #ifdef CONFIG_SMP | |
1043 | /* | |
1044 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
1045 | * We must ensure this doesn't happen until the switch is completely | |
1046 | * finished. | |
1047 | */ | |
1048 | smp_wmb(); | |
1049 | prev->oncpu = 0; | |
1050 | #endif | |
1051 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
1052 | local_irq_enable(); | |
1da177e4 | 1053 | #endif |
4866cde0 NP |
1054 | } |
1055 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 | 1056 | |
b29739f9 IM |
1057 | /* |
1058 | * __task_rq_lock - lock the runqueue a given task resides on. | |
1059 | * Must be called interrupts disabled. | |
1060 | */ | |
70b97a7f | 1061 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
b29739f9 IM |
1062 | __acquires(rq->lock) |
1063 | { | |
3a5c359a AK |
1064 | for (;;) { |
1065 | struct rq *rq = task_rq(p); | |
1066 | spin_lock(&rq->lock); | |
1067 | if (likely(rq == task_rq(p))) | |
1068 | return rq; | |
b29739f9 | 1069 | spin_unlock(&rq->lock); |
b29739f9 | 1070 | } |
b29739f9 IM |
1071 | } |
1072 | ||
1da177e4 LT |
1073 | /* |
1074 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
41a2d6cf | 1075 | * interrupts. Note the ordering: we can safely lookup the task_rq without |
1da177e4 LT |
1076 | * explicitly disabling preemption. |
1077 | */ | |
70b97a7f | 1078 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
1da177e4 LT |
1079 | __acquires(rq->lock) |
1080 | { | |
70b97a7f | 1081 | struct rq *rq; |
1da177e4 | 1082 | |
3a5c359a AK |
1083 | for (;;) { |
1084 | local_irq_save(*flags); | |
1085 | rq = task_rq(p); | |
1086 | spin_lock(&rq->lock); | |
1087 | if (likely(rq == task_rq(p))) | |
1088 | return rq; | |
1da177e4 | 1089 | spin_unlock_irqrestore(&rq->lock, *flags); |
1da177e4 | 1090 | } |
1da177e4 LT |
1091 | } |
1092 | ||
a9957449 | 1093 | static void __task_rq_unlock(struct rq *rq) |
b29739f9 IM |
1094 | __releases(rq->lock) |
1095 | { | |
1096 | spin_unlock(&rq->lock); | |
1097 | } | |
1098 | ||
70b97a7f | 1099 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) |
1da177e4 LT |
1100 | __releases(rq->lock) |
1101 | { | |
1102 | spin_unlock_irqrestore(&rq->lock, *flags); | |
1103 | } | |
1104 | ||
1da177e4 | 1105 | /* |
cc2a73b5 | 1106 | * this_rq_lock - lock this runqueue and disable interrupts. |
1da177e4 | 1107 | */ |
a9957449 | 1108 | static struct rq *this_rq_lock(void) |
1da177e4 LT |
1109 | __acquires(rq->lock) |
1110 | { | |
70b97a7f | 1111 | struct rq *rq; |
1da177e4 LT |
1112 | |
1113 | local_irq_disable(); | |
1114 | rq = this_rq(); | |
1115 | spin_lock(&rq->lock); | |
1116 | ||
1117 | return rq; | |
1118 | } | |
1119 | ||
1b9f19c2 | 1120 | /* |
2aa44d05 | 1121 | * We are going deep-idle (irqs are disabled): |
1b9f19c2 | 1122 | */ |
2aa44d05 | 1123 | void sched_clock_idle_sleep_event(void) |
1b9f19c2 | 1124 | { |
2aa44d05 IM |
1125 | struct rq *rq = cpu_rq(smp_processor_id()); |
1126 | ||
d478c2cf | 1127 | WARN_ON(!irqs_disabled()); |
2aa44d05 IM |
1128 | spin_lock(&rq->lock); |
1129 | __update_rq_clock(rq); | |
1130 | spin_unlock(&rq->lock); | |
1131 | rq->clock_deep_idle_events++; | |
1132 | } | |
1133 | EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event); | |
1134 | ||
1135 | /* | |
1136 | * We just idled delta nanoseconds (called with irqs disabled): | |
1137 | */ | |
1138 | void sched_clock_idle_wakeup_event(u64 delta_ns) | |
1139 | { | |
1140 | struct rq *rq = cpu_rq(smp_processor_id()); | |
1141 | u64 now = sched_clock(); | |
1b9f19c2 | 1142 | |
d478c2cf | 1143 | WARN_ON(!irqs_disabled()); |
2aa44d05 IM |
1144 | rq->idle_clock += delta_ns; |
1145 | /* | |
1146 | * Override the previous timestamp and ignore all | |
1147 | * sched_clock() deltas that occured while we idled, | |
1148 | * and use the PM-provided delta_ns to advance the | |
1149 | * rq clock: | |
1150 | */ | |
1151 | spin_lock(&rq->lock); | |
1152 | rq->prev_clock_raw = now; | |
1153 | rq->clock += delta_ns; | |
1154 | spin_unlock(&rq->lock); | |
782daeee | 1155 | touch_softlockup_watchdog(); |
1b9f19c2 | 1156 | } |
2aa44d05 | 1157 | EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); |
1b9f19c2 | 1158 | |
8f4d37ec PZ |
1159 | static void __resched_task(struct task_struct *p, int tif_bit); |
1160 | ||
1161 | static inline void resched_task(struct task_struct *p) | |
1162 | { | |
1163 | __resched_task(p, TIF_NEED_RESCHED); | |
1164 | } | |
1165 | ||
1166 | #ifdef CONFIG_SCHED_HRTICK | |
1167 | /* | |
1168 | * Use HR-timers to deliver accurate preemption points. | |
1169 | * | |
1170 | * Its all a bit involved since we cannot program an hrt while holding the | |
1171 | * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a | |
1172 | * reschedule event. | |
1173 | * | |
1174 | * When we get rescheduled we reprogram the hrtick_timer outside of the | |
1175 | * rq->lock. | |
1176 | */ | |
1177 | static inline void resched_hrt(struct task_struct *p) | |
1178 | { | |
1179 | __resched_task(p, TIF_HRTICK_RESCHED); | |
1180 | } | |
1181 | ||
1182 | static inline void resched_rq(struct rq *rq) | |
1183 | { | |
1184 | unsigned long flags; | |
1185 | ||
1186 | spin_lock_irqsave(&rq->lock, flags); | |
1187 | resched_task(rq->curr); | |
1188 | spin_unlock_irqrestore(&rq->lock, flags); | |
1189 | } | |
1190 | ||
1191 | enum { | |
1192 | HRTICK_SET, /* re-programm hrtick_timer */ | |
1193 | HRTICK_RESET, /* not a new slice */ | |
1194 | }; | |
1195 | ||
1196 | /* | |
1197 | * Use hrtick when: | |
1198 | * - enabled by features | |
1199 | * - hrtimer is actually high res | |
1200 | */ | |
1201 | static inline int hrtick_enabled(struct rq *rq) | |
1202 | { | |
1203 | if (!sched_feat(HRTICK)) | |
1204 | return 0; | |
1205 | return hrtimer_is_hres_active(&rq->hrtick_timer); | |
1206 | } | |
1207 | ||
1208 | /* | |
1209 | * Called to set the hrtick timer state. | |
1210 | * | |
1211 | * called with rq->lock held and irqs disabled | |
1212 | */ | |
1213 | static void hrtick_start(struct rq *rq, u64 delay, int reset) | |
1214 | { | |
1215 | assert_spin_locked(&rq->lock); | |
1216 | ||
1217 | /* | |
1218 | * preempt at: now + delay | |
1219 | */ | |
1220 | rq->hrtick_expire = | |
1221 | ktime_add_ns(rq->hrtick_timer.base->get_time(), delay); | |
1222 | /* | |
1223 | * indicate we need to program the timer | |
1224 | */ | |
1225 | __set_bit(HRTICK_SET, &rq->hrtick_flags); | |
1226 | if (reset) | |
1227 | __set_bit(HRTICK_RESET, &rq->hrtick_flags); | |
1228 | ||
1229 | /* | |
1230 | * New slices are called from the schedule path and don't need a | |
1231 | * forced reschedule. | |
1232 | */ | |
1233 | if (reset) | |
1234 | resched_hrt(rq->curr); | |
1235 | } | |
1236 | ||
1237 | static void hrtick_clear(struct rq *rq) | |
1238 | { | |
1239 | if (hrtimer_active(&rq->hrtick_timer)) | |
1240 | hrtimer_cancel(&rq->hrtick_timer); | |
1241 | } | |
1242 | ||
1243 | /* | |
1244 | * Update the timer from the possible pending state. | |
1245 | */ | |
1246 | static void hrtick_set(struct rq *rq) | |
1247 | { | |
1248 | ktime_t time; | |
1249 | int set, reset; | |
1250 | unsigned long flags; | |
1251 | ||
1252 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1253 | ||
1254 | spin_lock_irqsave(&rq->lock, flags); | |
1255 | set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags); | |
1256 | reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags); | |
1257 | time = rq->hrtick_expire; | |
1258 | clear_thread_flag(TIF_HRTICK_RESCHED); | |
1259 | spin_unlock_irqrestore(&rq->lock, flags); | |
1260 | ||
1261 | if (set) { | |
1262 | hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS); | |
1263 | if (reset && !hrtimer_active(&rq->hrtick_timer)) | |
1264 | resched_rq(rq); | |
1265 | } else | |
1266 | hrtick_clear(rq); | |
1267 | } | |
1268 | ||
1269 | /* | |
1270 | * High-resolution timer tick. | |
1271 | * Runs from hardirq context with interrupts disabled. | |
1272 | */ | |
1273 | static enum hrtimer_restart hrtick(struct hrtimer *timer) | |
1274 | { | |
1275 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); | |
1276 | ||
1277 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1278 | ||
1279 | spin_lock(&rq->lock); | |
1280 | __update_rq_clock(rq); | |
1281 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); | |
1282 | spin_unlock(&rq->lock); | |
1283 | ||
1284 | return HRTIMER_NORESTART; | |
1285 | } | |
1286 | ||
1287 | static inline void init_rq_hrtick(struct rq *rq) | |
1288 | { | |
1289 | rq->hrtick_flags = 0; | |
1290 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1291 | rq->hrtick_timer.function = hrtick; | |
1292 | rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
1293 | } | |
1294 | ||
1295 | void hrtick_resched(void) | |
1296 | { | |
1297 | struct rq *rq; | |
1298 | unsigned long flags; | |
1299 | ||
1300 | if (!test_thread_flag(TIF_HRTICK_RESCHED)) | |
1301 | return; | |
1302 | ||
1303 | local_irq_save(flags); | |
1304 | rq = cpu_rq(smp_processor_id()); | |
1305 | hrtick_set(rq); | |
1306 | local_irq_restore(flags); | |
1307 | } | |
1308 | #else | |
1309 | static inline void hrtick_clear(struct rq *rq) | |
1310 | { | |
1311 | } | |
1312 | ||
1313 | static inline void hrtick_set(struct rq *rq) | |
1314 | { | |
1315 | } | |
1316 | ||
1317 | static inline void init_rq_hrtick(struct rq *rq) | |
1318 | { | |
1319 | } | |
1320 | ||
1321 | void hrtick_resched(void) | |
1322 | { | |
1323 | } | |
1324 | #endif | |
1325 | ||
c24d20db IM |
1326 | /* |
1327 | * resched_task - mark a task 'to be rescheduled now'. | |
1328 | * | |
1329 | * On UP this means the setting of the need_resched flag, on SMP it | |
1330 | * might also involve a cross-CPU call to trigger the scheduler on | |
1331 | * the target CPU. | |
1332 | */ | |
1333 | #ifdef CONFIG_SMP | |
1334 | ||
1335 | #ifndef tsk_is_polling | |
1336 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
1337 | #endif | |
1338 | ||
8f4d37ec | 1339 | static void __resched_task(struct task_struct *p, int tif_bit) |
c24d20db IM |
1340 | { |
1341 | int cpu; | |
1342 | ||
1343 | assert_spin_locked(&task_rq(p)->lock); | |
1344 | ||
8f4d37ec | 1345 | if (unlikely(test_tsk_thread_flag(p, tif_bit))) |
c24d20db IM |
1346 | return; |
1347 | ||
8f4d37ec | 1348 | set_tsk_thread_flag(p, tif_bit); |
c24d20db IM |
1349 | |
1350 | cpu = task_cpu(p); | |
1351 | if (cpu == smp_processor_id()) | |
1352 | return; | |
1353 | ||
1354 | /* NEED_RESCHED must be visible before we test polling */ | |
1355 | smp_mb(); | |
1356 | if (!tsk_is_polling(p)) | |
1357 | smp_send_reschedule(cpu); | |
1358 | } | |
1359 | ||
1360 | static void resched_cpu(int cpu) | |
1361 | { | |
1362 | struct rq *rq = cpu_rq(cpu); | |
1363 | unsigned long flags; | |
1364 | ||
1365 | if (!spin_trylock_irqsave(&rq->lock, flags)) | |
1366 | return; | |
1367 | resched_task(cpu_curr(cpu)); | |
1368 | spin_unlock_irqrestore(&rq->lock, flags); | |
1369 | } | |
06d8308c TG |
1370 | |
1371 | #ifdef CONFIG_NO_HZ | |
1372 | /* | |
1373 | * When add_timer_on() enqueues a timer into the timer wheel of an | |
1374 | * idle CPU then this timer might expire before the next timer event | |
1375 | * which is scheduled to wake up that CPU. In case of a completely | |
1376 | * idle system the next event might even be infinite time into the | |
1377 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and | |
1378 | * leaves the inner idle loop so the newly added timer is taken into | |
1379 | * account when the CPU goes back to idle and evaluates the timer | |
1380 | * wheel for the next timer event. | |
1381 | */ | |
1382 | void wake_up_idle_cpu(int cpu) | |
1383 | { | |
1384 | struct rq *rq = cpu_rq(cpu); | |
1385 | ||
1386 | if (cpu == smp_processor_id()) | |
1387 | return; | |
1388 | ||
1389 | /* | |
1390 | * This is safe, as this function is called with the timer | |
1391 | * wheel base lock of (cpu) held. When the CPU is on the way | |
1392 | * to idle and has not yet set rq->curr to idle then it will | |
1393 | * be serialized on the timer wheel base lock and take the new | |
1394 | * timer into account automatically. | |
1395 | */ | |
1396 | if (rq->curr != rq->idle) | |
1397 | return; | |
1398 | ||
1399 | /* | |
1400 | * We can set TIF_RESCHED on the idle task of the other CPU | |
1401 | * lockless. The worst case is that the other CPU runs the | |
1402 | * idle task through an additional NOOP schedule() | |
1403 | */ | |
1404 | set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED); | |
1405 | ||
1406 | /* NEED_RESCHED must be visible before we test polling */ | |
1407 | smp_mb(); | |
1408 | if (!tsk_is_polling(rq->idle)) | |
1409 | smp_send_reschedule(cpu); | |
1410 | } | |
1411 | #endif | |
1412 | ||
c24d20db | 1413 | #else |
8f4d37ec | 1414 | static void __resched_task(struct task_struct *p, int tif_bit) |
c24d20db IM |
1415 | { |
1416 | assert_spin_locked(&task_rq(p)->lock); | |
8f4d37ec | 1417 | set_tsk_thread_flag(p, tif_bit); |
c24d20db IM |
1418 | } |
1419 | #endif | |
1420 | ||
45bf76df IM |
1421 | #if BITS_PER_LONG == 32 |
1422 | # define WMULT_CONST (~0UL) | |
1423 | #else | |
1424 | # define WMULT_CONST (1UL << 32) | |
1425 | #endif | |
1426 | ||
1427 | #define WMULT_SHIFT 32 | |
1428 | ||
194081eb IM |
1429 | /* |
1430 | * Shift right and round: | |
1431 | */ | |
cf2ab469 | 1432 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) |
194081eb | 1433 | |
8f1bc385 PZ |
1434 | /* |
1435 | * delta *= weight / lw | |
1436 | */ | |
cb1c4fc9 | 1437 | static unsigned long |
45bf76df IM |
1438 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, |
1439 | struct load_weight *lw) | |
1440 | { | |
1441 | u64 tmp; | |
1442 | ||
e05510d0 PZ |
1443 | if (!lw->inv_weight) |
1444 | lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1); | |
45bf76df IM |
1445 | |
1446 | tmp = (u64)delta_exec * weight; | |
1447 | /* | |
1448 | * Check whether we'd overflow the 64-bit multiplication: | |
1449 | */ | |
194081eb | 1450 | if (unlikely(tmp > WMULT_CONST)) |
cf2ab469 | 1451 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, |
194081eb IM |
1452 | WMULT_SHIFT/2); |
1453 | else | |
cf2ab469 | 1454 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); |
45bf76df | 1455 | |
ecf691da | 1456 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); |
45bf76df IM |
1457 | } |
1458 | ||
1091985b | 1459 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
45bf76df IM |
1460 | { |
1461 | lw->weight += inc; | |
e89996ae | 1462 | lw->inv_weight = 0; |
45bf76df IM |
1463 | } |
1464 | ||
1091985b | 1465 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) |
45bf76df IM |
1466 | { |
1467 | lw->weight -= dec; | |
e89996ae | 1468 | lw->inv_weight = 0; |
45bf76df IM |
1469 | } |
1470 | ||
2dd73a4f PW |
1471 | /* |
1472 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
1473 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
1474 | * each task makes to its run queue's load is weighted according to its | |
41a2d6cf | 1475 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a |
2dd73a4f PW |
1476 | * scaled version of the new time slice allocation that they receive on time |
1477 | * slice expiry etc. | |
1478 | */ | |
1479 | ||
dd41f596 IM |
1480 | #define WEIGHT_IDLEPRIO 2 |
1481 | #define WMULT_IDLEPRIO (1 << 31) | |
1482 | ||
1483 | /* | |
1484 | * Nice levels are multiplicative, with a gentle 10% change for every | |
1485 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to | |
1486 | * nice 1, it will get ~10% less CPU time than another CPU-bound task | |
1487 | * that remained on nice 0. | |
1488 | * | |
1489 | * The "10% effect" is relative and cumulative: from _any_ nice level, | |
1490 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level | |
f9153ee6 IM |
1491 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. |
1492 | * If a task goes up by ~10% and another task goes down by ~10% then | |
1493 | * the relative distance between them is ~25%.) | |
dd41f596 IM |
1494 | */ |
1495 | static const int prio_to_weight[40] = { | |
254753dc IM |
1496 | /* -20 */ 88761, 71755, 56483, 46273, 36291, |
1497 | /* -15 */ 29154, 23254, 18705, 14949, 11916, | |
1498 | /* -10 */ 9548, 7620, 6100, 4904, 3906, | |
1499 | /* -5 */ 3121, 2501, 1991, 1586, 1277, | |
1500 | /* 0 */ 1024, 820, 655, 526, 423, | |
1501 | /* 5 */ 335, 272, 215, 172, 137, | |
1502 | /* 10 */ 110, 87, 70, 56, 45, | |
1503 | /* 15 */ 36, 29, 23, 18, 15, | |
dd41f596 IM |
1504 | }; |
1505 | ||
5714d2de IM |
1506 | /* |
1507 | * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. | |
1508 | * | |
1509 | * In cases where the weight does not change often, we can use the | |
1510 | * precalculated inverse to speed up arithmetics by turning divisions | |
1511 | * into multiplications: | |
1512 | */ | |
dd41f596 | 1513 | static const u32 prio_to_wmult[40] = { |
254753dc IM |
1514 | /* -20 */ 48388, 59856, 76040, 92818, 118348, |
1515 | /* -15 */ 147320, 184698, 229616, 287308, 360437, | |
1516 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, | |
1517 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, | |
1518 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, | |
1519 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, | |
1520 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, | |
1521 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, | |
dd41f596 | 1522 | }; |
2dd73a4f | 1523 | |
dd41f596 IM |
1524 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup); |
1525 | ||
1526 | /* | |
1527 | * runqueue iterator, to support SMP load-balancing between different | |
1528 | * scheduling classes, without having to expose their internal data | |
1529 | * structures to the load-balancing proper: | |
1530 | */ | |
1531 | struct rq_iterator { | |
1532 | void *arg; | |
1533 | struct task_struct *(*start)(void *); | |
1534 | struct task_struct *(*next)(void *); | |
1535 | }; | |
1536 | ||
e1d1484f PW |
1537 | #ifdef CONFIG_SMP |
1538 | static unsigned long | |
1539 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1540 | unsigned long max_load_move, struct sched_domain *sd, | |
1541 | enum cpu_idle_type idle, int *all_pinned, | |
1542 | int *this_best_prio, struct rq_iterator *iterator); | |
1543 | ||
1544 | static int | |
1545 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1546 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1547 | struct rq_iterator *iterator); | |
e1d1484f | 1548 | #endif |
dd41f596 | 1549 | |
d842de87 SV |
1550 | #ifdef CONFIG_CGROUP_CPUACCT |
1551 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime); | |
1552 | #else | |
1553 | static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} | |
1554 | #endif | |
1555 | ||
18d95a28 PZ |
1556 | static inline void inc_cpu_load(struct rq *rq, unsigned long load) |
1557 | { | |
1558 | update_load_add(&rq->load, load); | |
1559 | } | |
1560 | ||
1561 | static inline void dec_cpu_load(struct rq *rq, unsigned long load) | |
1562 | { | |
1563 | update_load_sub(&rq->load, load); | |
1564 | } | |
1565 | ||
e7693a36 GH |
1566 | #ifdef CONFIG_SMP |
1567 | static unsigned long source_load(int cpu, int type); | |
1568 | static unsigned long target_load(int cpu, int type); | |
1569 | static unsigned long cpu_avg_load_per_task(int cpu); | |
1570 | static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd); | |
18d95a28 PZ |
1571 | |
1572 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1573 | ||
1574 | /* | |
1575 | * Group load balancing. | |
1576 | * | |
1577 | * We calculate a few balance domain wide aggregate numbers; load and weight. | |
1578 | * Given the pictures below, and assuming each item has equal weight: | |
1579 | * | |
1580 | * root 1 - thread | |
1581 | * / | \ A - group | |
1582 | * A 1 B | |
1583 | * /|\ / \ | |
1584 | * C 2 D 3 4 | |
1585 | * | | | |
1586 | * 5 6 | |
1587 | * | |
1588 | * load: | |
1589 | * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd, | |
1590 | * which equals 1/9-th of the total load. | |
1591 | * | |
1592 | * shares: | |
1593 | * The weight of this group on the selected cpus. | |
1594 | * | |
1595 | * rq_weight: | |
1596 | * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while | |
1597 | * B would get 2. | |
1598 | * | |
1599 | * task_weight: | |
1600 | * Part of the rq_weight contributed by tasks; all groups except B would | |
1601 | * get 1, B gets 2. | |
1602 | */ | |
1603 | ||
1604 | static inline struct aggregate_struct * | |
1605 | aggregate(struct task_group *tg, struct sched_domain *sd) | |
1606 | { | |
1607 | return &tg->cfs_rq[sd->first_cpu]->aggregate; | |
1608 | } | |
1609 | ||
1610 | typedef void (*aggregate_func)(struct task_group *, struct sched_domain *); | |
1611 | ||
1612 | /* | |
1613 | * Iterate the full tree, calling @down when first entering a node and @up when | |
1614 | * leaving it for the final time. | |
1615 | */ | |
1616 | static | |
1617 | void aggregate_walk_tree(aggregate_func down, aggregate_func up, | |
1618 | struct sched_domain *sd) | |
1619 | { | |
1620 | struct task_group *parent, *child; | |
1621 | ||
1622 | rcu_read_lock(); | |
1623 | parent = &root_task_group; | |
1624 | down: | |
1625 | (*down)(parent, sd); | |
1626 | list_for_each_entry_rcu(child, &parent->children, siblings) { | |
1627 | parent = child; | |
1628 | goto down; | |
1629 | ||
1630 | up: | |
1631 | continue; | |
1632 | } | |
1633 | (*up)(parent, sd); | |
1634 | ||
1635 | child = parent; | |
1636 | parent = parent->parent; | |
1637 | if (parent) | |
1638 | goto up; | |
1639 | rcu_read_unlock(); | |
1640 | } | |
1641 | ||
1642 | /* | |
1643 | * Calculate the aggregate runqueue weight. | |
1644 | */ | |
1645 | static | |
1646 | void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd) | |
1647 | { | |
1648 | unsigned long rq_weight = 0; | |
1649 | unsigned long task_weight = 0; | |
1650 | int i; | |
1651 | ||
1652 | for_each_cpu_mask(i, sd->span) { | |
1653 | rq_weight += tg->cfs_rq[i]->load.weight; | |
1654 | task_weight += tg->cfs_rq[i]->task_weight; | |
1655 | } | |
1656 | ||
1657 | aggregate(tg, sd)->rq_weight = rq_weight; | |
1658 | aggregate(tg, sd)->task_weight = task_weight; | |
1659 | } | |
1660 | ||
18d95a28 PZ |
1661 | /* |
1662 | * Compute the weight of this group on the given cpus. | |
1663 | */ | |
1664 | static | |
1665 | void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd) | |
1666 | { | |
1667 | unsigned long shares = 0; | |
1668 | int i; | |
1669 | ||
18d95a28 PZ |
1670 | for_each_cpu_mask(i, sd->span) |
1671 | shares += tg->cfs_rq[i]->shares; | |
1672 | ||
3f5087a2 PZ |
1673 | if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares) |
1674 | shares = tg->shares; | |
18d95a28 PZ |
1675 | |
1676 | aggregate(tg, sd)->shares = shares; | |
1677 | } | |
1678 | ||
1679 | /* | |
1680 | * Compute the load fraction assigned to this group, relies on the aggregate | |
1681 | * weight and this group's parent's load, i.e. top-down. | |
1682 | */ | |
1683 | static | |
1684 | void aggregate_group_load(struct task_group *tg, struct sched_domain *sd) | |
1685 | { | |
1686 | unsigned long load; | |
1687 | ||
1688 | if (!tg->parent) { | |
1689 | int i; | |
1690 | ||
1691 | load = 0; | |
1692 | for_each_cpu_mask(i, sd->span) | |
1693 | load += cpu_rq(i)->load.weight; | |
1694 | ||
1695 | } else { | |
1696 | load = aggregate(tg->parent, sd)->load; | |
1697 | ||
1698 | /* | |
1699 | * shares is our weight in the parent's rq so | |
1700 | * shares/parent->rq_weight gives our fraction of the load | |
1701 | */ | |
1702 | load *= aggregate(tg, sd)->shares; | |
1703 | load /= aggregate(tg->parent, sd)->rq_weight + 1; | |
1704 | } | |
1705 | ||
1706 | aggregate(tg, sd)->load = load; | |
1707 | } | |
1708 | ||
1709 | static void __set_se_shares(struct sched_entity *se, unsigned long shares); | |
1710 | ||
1711 | /* | |
1712 | * Calculate and set the cpu's group shares. | |
1713 | */ | |
1714 | static void | |
1715 | __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd, | |
1716 | int tcpu) | |
1717 | { | |
1718 | int boost = 0; | |
1719 | unsigned long shares; | |
1720 | unsigned long rq_weight; | |
1721 | ||
1722 | if (!tg->se[tcpu]) | |
1723 | return; | |
1724 | ||
1725 | rq_weight = tg->cfs_rq[tcpu]->load.weight; | |
1726 | ||
1727 | /* | |
1728 | * If there are currently no tasks on the cpu pretend there is one of | |
1729 | * average load so that when a new task gets to run here it will not | |
1730 | * get delayed by group starvation. | |
1731 | */ | |
1732 | if (!rq_weight) { | |
1733 | boost = 1; | |
1734 | rq_weight = NICE_0_LOAD; | |
1735 | } | |
1736 | ||
1737 | /* | |
1738 | * \Sum shares * rq_weight | |
1739 | * shares = ----------------------- | |
1740 | * \Sum rq_weight | |
1741 | * | |
1742 | */ | |
1743 | shares = aggregate(tg, sd)->shares * rq_weight; | |
1744 | shares /= aggregate(tg, sd)->rq_weight + 1; | |
1745 | ||
1746 | /* | |
1747 | * record the actual number of shares, not the boosted amount. | |
1748 | */ | |
1749 | tg->cfs_rq[tcpu]->shares = boost ? 0 : shares; | |
1750 | ||
1751 | if (shares < MIN_SHARES) | |
1752 | shares = MIN_SHARES; | |
1753 | ||
1754 | __set_se_shares(tg->se[tcpu], shares); | |
1755 | } | |
1756 | ||
1757 | /* | |
1758 | * Re-adjust the weights on the cpu the task came from and on the cpu the | |
1759 | * task went to. | |
1760 | */ | |
1761 | static void | |
1762 | __move_group_shares(struct task_group *tg, struct sched_domain *sd, | |
1763 | int scpu, int dcpu) | |
1764 | { | |
1765 | unsigned long shares; | |
1766 | ||
1767 | shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares; | |
1768 | ||
1769 | __update_group_shares_cpu(tg, sd, scpu); | |
1770 | __update_group_shares_cpu(tg, sd, dcpu); | |
1771 | ||
1772 | /* | |
1773 | * ensure we never loose shares due to rounding errors in the | |
1774 | * above redistribution. | |
1775 | */ | |
1776 | shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares; | |
1777 | if (shares) | |
1778 | tg->cfs_rq[dcpu]->shares += shares; | |
1779 | } | |
1780 | ||
1781 | /* | |
1782 | * Because changing a group's shares changes the weight of the super-group | |
1783 | * we need to walk up the tree and change all shares until we hit the root. | |
1784 | */ | |
1785 | static void | |
1786 | move_group_shares(struct task_group *tg, struct sched_domain *sd, | |
1787 | int scpu, int dcpu) | |
1788 | { | |
1789 | while (tg) { | |
1790 | __move_group_shares(tg, sd, scpu, dcpu); | |
1791 | tg = tg->parent; | |
1792 | } | |
1793 | } | |
1794 | ||
1795 | static | |
1796 | void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd) | |
1797 | { | |
1798 | unsigned long shares = aggregate(tg, sd)->shares; | |
1799 | int i; | |
1800 | ||
1801 | for_each_cpu_mask(i, sd->span) { | |
1802 | struct rq *rq = cpu_rq(i); | |
1803 | unsigned long flags; | |
1804 | ||
1805 | spin_lock_irqsave(&rq->lock, flags); | |
1806 | __update_group_shares_cpu(tg, sd, i); | |
1807 | spin_unlock_irqrestore(&rq->lock, flags); | |
1808 | } | |
1809 | ||
1810 | aggregate_group_shares(tg, sd); | |
1811 | ||
1812 | /* | |
1813 | * ensure we never loose shares due to rounding errors in the | |
1814 | * above redistribution. | |
1815 | */ | |
1816 | shares -= aggregate(tg, sd)->shares; | |
1817 | if (shares) { | |
1818 | tg->cfs_rq[sd->first_cpu]->shares += shares; | |
1819 | aggregate(tg, sd)->shares += shares; | |
1820 | } | |
1821 | } | |
1822 | ||
1823 | /* | |
1824 | * Calculate the accumulative weight and recursive load of each task group | |
1825 | * while walking down the tree. | |
1826 | */ | |
1827 | static | |
1828 | void aggregate_get_down(struct task_group *tg, struct sched_domain *sd) | |
1829 | { | |
1830 | aggregate_group_weight(tg, sd); | |
1831 | aggregate_group_shares(tg, sd); | |
1832 | aggregate_group_load(tg, sd); | |
1833 | } | |
1834 | ||
1835 | /* | |
1836 | * Rebalance the cpu shares while walking back up the tree. | |
1837 | */ | |
1838 | static | |
1839 | void aggregate_get_up(struct task_group *tg, struct sched_domain *sd) | |
1840 | { | |
1841 | aggregate_group_set_shares(tg, sd); | |
1842 | } | |
1843 | ||
1844 | static DEFINE_PER_CPU(spinlock_t, aggregate_lock); | |
1845 | ||
1846 | static void __init init_aggregate(void) | |
1847 | { | |
1848 | int i; | |
1849 | ||
1850 | for_each_possible_cpu(i) | |
1851 | spin_lock_init(&per_cpu(aggregate_lock, i)); | |
1852 | } | |
1853 | ||
1854 | static int get_aggregate(struct sched_domain *sd) | |
1855 | { | |
1856 | if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu))) | |
1857 | return 0; | |
1858 | ||
1859 | aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd); | |
1860 | return 1; | |
1861 | } | |
1862 | ||
1863 | static void put_aggregate(struct sched_domain *sd) | |
1864 | { | |
1865 | spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu)); | |
1866 | } | |
1867 | ||
1868 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1869 | { | |
1870 | cfs_rq->shares = shares; | |
1871 | } | |
1872 | ||
1873 | #else | |
1874 | ||
1875 | static inline void init_aggregate(void) | |
1876 | { | |
1877 | } | |
1878 | ||
1879 | static inline int get_aggregate(struct sched_domain *sd) | |
1880 | { | |
1881 | return 0; | |
1882 | } | |
1883 | ||
1884 | static inline void put_aggregate(struct sched_domain *sd) | |
1885 | { | |
1886 | } | |
1887 | #endif | |
1888 | ||
1889 | #else /* CONFIG_SMP */ | |
1890 | ||
1891 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1892 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1893 | { | |
1894 | } | |
1895 | #endif | |
1896 | ||
e7693a36 GH |
1897 | #endif /* CONFIG_SMP */ |
1898 | ||
dd41f596 | 1899 | #include "sched_stats.h" |
dd41f596 | 1900 | #include "sched_idletask.c" |
5522d5d5 IM |
1901 | #include "sched_fair.c" |
1902 | #include "sched_rt.c" | |
dd41f596 IM |
1903 | #ifdef CONFIG_SCHED_DEBUG |
1904 | # include "sched_debug.c" | |
1905 | #endif | |
1906 | ||
1907 | #define sched_class_highest (&rt_sched_class) | |
1908 | ||
18d95a28 | 1909 | static void inc_nr_running(struct rq *rq) |
9c217245 IM |
1910 | { |
1911 | rq->nr_running++; | |
9c217245 IM |
1912 | } |
1913 | ||
18d95a28 | 1914 | static void dec_nr_running(struct rq *rq) |
9c217245 IM |
1915 | { |
1916 | rq->nr_running--; | |
9c217245 IM |
1917 | } |
1918 | ||
45bf76df IM |
1919 | static void set_load_weight(struct task_struct *p) |
1920 | { | |
1921 | if (task_has_rt_policy(p)) { | |
dd41f596 IM |
1922 | p->se.load.weight = prio_to_weight[0] * 2; |
1923 | p->se.load.inv_weight = prio_to_wmult[0] >> 1; | |
1924 | return; | |
1925 | } | |
45bf76df | 1926 | |
dd41f596 IM |
1927 | /* |
1928 | * SCHED_IDLE tasks get minimal weight: | |
1929 | */ | |
1930 | if (p->policy == SCHED_IDLE) { | |
1931 | p->se.load.weight = WEIGHT_IDLEPRIO; | |
1932 | p->se.load.inv_weight = WMULT_IDLEPRIO; | |
1933 | return; | |
1934 | } | |
71f8bd46 | 1935 | |
dd41f596 IM |
1936 | p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO]; |
1937 | p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO]; | |
71f8bd46 IM |
1938 | } |
1939 | ||
8159f87e | 1940 | static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup) |
71f8bd46 | 1941 | { |
dd41f596 | 1942 | sched_info_queued(p); |
fd390f6a | 1943 | p->sched_class->enqueue_task(rq, p, wakeup); |
dd41f596 | 1944 | p->se.on_rq = 1; |
71f8bd46 IM |
1945 | } |
1946 | ||
69be72c1 | 1947 | static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep) |
71f8bd46 | 1948 | { |
f02231e5 | 1949 | p->sched_class->dequeue_task(rq, p, sleep); |
dd41f596 | 1950 | p->se.on_rq = 0; |
71f8bd46 IM |
1951 | } |
1952 | ||
14531189 | 1953 | /* |
dd41f596 | 1954 | * __normal_prio - return the priority that is based on the static prio |
14531189 | 1955 | */ |
14531189 IM |
1956 | static inline int __normal_prio(struct task_struct *p) |
1957 | { | |
dd41f596 | 1958 | return p->static_prio; |
14531189 IM |
1959 | } |
1960 | ||
b29739f9 IM |
1961 | /* |
1962 | * Calculate the expected normal priority: i.e. priority | |
1963 | * without taking RT-inheritance into account. Might be | |
1964 | * boosted by interactivity modifiers. Changes upon fork, | |
1965 | * setprio syscalls, and whenever the interactivity | |
1966 | * estimator recalculates. | |
1967 | */ | |
36c8b586 | 1968 | static inline int normal_prio(struct task_struct *p) |
b29739f9 IM |
1969 | { |
1970 | int prio; | |
1971 | ||
e05606d3 | 1972 | if (task_has_rt_policy(p)) |
b29739f9 IM |
1973 | prio = MAX_RT_PRIO-1 - p->rt_priority; |
1974 | else | |
1975 | prio = __normal_prio(p); | |
1976 | return prio; | |
1977 | } | |
1978 | ||
1979 | /* | |
1980 | * Calculate the current priority, i.e. the priority | |
1981 | * taken into account by the scheduler. This value might | |
1982 | * be boosted by RT tasks, or might be boosted by | |
1983 | * interactivity modifiers. Will be RT if the task got | |
1984 | * RT-boosted. If not then it returns p->normal_prio. | |
1985 | */ | |
36c8b586 | 1986 | static int effective_prio(struct task_struct *p) |
b29739f9 IM |
1987 | { |
1988 | p->normal_prio = normal_prio(p); | |
1989 | /* | |
1990 | * If we are RT tasks or we were boosted to RT priority, | |
1991 | * keep the priority unchanged. Otherwise, update priority | |
1992 | * to the normal priority: | |
1993 | */ | |
1994 | if (!rt_prio(p->prio)) | |
1995 | return p->normal_prio; | |
1996 | return p->prio; | |
1997 | } | |
1998 | ||
1da177e4 | 1999 | /* |
dd41f596 | 2000 | * activate_task - move a task to the runqueue. |
1da177e4 | 2001 | */ |
dd41f596 | 2002 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup) |
1da177e4 | 2003 | { |
d9514f6c | 2004 | if (task_contributes_to_load(p)) |
dd41f596 | 2005 | rq->nr_uninterruptible--; |
1da177e4 | 2006 | |
8159f87e | 2007 | enqueue_task(rq, p, wakeup); |
18d95a28 | 2008 | inc_nr_running(rq); |
1da177e4 LT |
2009 | } |
2010 | ||
1da177e4 LT |
2011 | /* |
2012 | * deactivate_task - remove a task from the runqueue. | |
2013 | */ | |
2e1cb74a | 2014 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep) |
1da177e4 | 2015 | { |
d9514f6c | 2016 | if (task_contributes_to_load(p)) |
dd41f596 IM |
2017 | rq->nr_uninterruptible++; |
2018 | ||
69be72c1 | 2019 | dequeue_task(rq, p, sleep); |
18d95a28 | 2020 | dec_nr_running(rq); |
1da177e4 LT |
2021 | } |
2022 | ||
1da177e4 LT |
2023 | /** |
2024 | * task_curr - is this task currently executing on a CPU? | |
2025 | * @p: the task in question. | |
2026 | */ | |
36c8b586 | 2027 | inline int task_curr(const struct task_struct *p) |
1da177e4 LT |
2028 | { |
2029 | return cpu_curr(task_cpu(p)) == p; | |
2030 | } | |
2031 | ||
2dd73a4f PW |
2032 | /* Used instead of source_load when we know the type == 0 */ |
2033 | unsigned long weighted_cpuload(const int cpu) | |
2034 | { | |
495eca49 | 2035 | return cpu_rq(cpu)->load.weight; |
dd41f596 IM |
2036 | } |
2037 | ||
2038 | static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) | |
2039 | { | |
6f505b16 | 2040 | set_task_rq(p, cpu); |
dd41f596 | 2041 | #ifdef CONFIG_SMP |
ce96b5ac DA |
2042 | /* |
2043 | * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be | |
2044 | * successfuly executed on another CPU. We must ensure that updates of | |
2045 | * per-task data have been completed by this moment. | |
2046 | */ | |
2047 | smp_wmb(); | |
dd41f596 | 2048 | task_thread_info(p)->cpu = cpu; |
dd41f596 | 2049 | #endif |
2dd73a4f PW |
2050 | } |
2051 | ||
cb469845 SR |
2052 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, |
2053 | const struct sched_class *prev_class, | |
2054 | int oldprio, int running) | |
2055 | { | |
2056 | if (prev_class != p->sched_class) { | |
2057 | if (prev_class->switched_from) | |
2058 | prev_class->switched_from(rq, p, running); | |
2059 | p->sched_class->switched_to(rq, p, running); | |
2060 | } else | |
2061 | p->sched_class->prio_changed(rq, p, oldprio, running); | |
2062 | } | |
2063 | ||
1da177e4 | 2064 | #ifdef CONFIG_SMP |
c65cc870 | 2065 | |
cc367732 IM |
2066 | /* |
2067 | * Is this task likely cache-hot: | |
2068 | */ | |
e7693a36 | 2069 | static int |
cc367732 IM |
2070 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) |
2071 | { | |
2072 | s64 delta; | |
2073 | ||
f540a608 IM |
2074 | /* |
2075 | * Buddy candidates are cache hot: | |
2076 | */ | |
d25ce4cd | 2077 | if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next)) |
f540a608 IM |
2078 | return 1; |
2079 | ||
cc367732 IM |
2080 | if (p->sched_class != &fair_sched_class) |
2081 | return 0; | |
2082 | ||
6bc1665b IM |
2083 | if (sysctl_sched_migration_cost == -1) |
2084 | return 1; | |
2085 | if (sysctl_sched_migration_cost == 0) | |
2086 | return 0; | |
2087 | ||
cc367732 IM |
2088 | delta = now - p->se.exec_start; |
2089 | ||
2090 | return delta < (s64)sysctl_sched_migration_cost; | |
2091 | } | |
2092 | ||
2093 | ||
dd41f596 | 2094 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
c65cc870 | 2095 | { |
dd41f596 IM |
2096 | int old_cpu = task_cpu(p); |
2097 | struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu); | |
2830cf8c SV |
2098 | struct cfs_rq *old_cfsrq = task_cfs_rq(p), |
2099 | *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu); | |
bbdba7c0 | 2100 | u64 clock_offset; |
dd41f596 IM |
2101 | |
2102 | clock_offset = old_rq->clock - new_rq->clock; | |
6cfb0d5d IM |
2103 | |
2104 | #ifdef CONFIG_SCHEDSTATS | |
2105 | if (p->se.wait_start) | |
2106 | p->se.wait_start -= clock_offset; | |
dd41f596 IM |
2107 | if (p->se.sleep_start) |
2108 | p->se.sleep_start -= clock_offset; | |
2109 | if (p->se.block_start) | |
2110 | p->se.block_start -= clock_offset; | |
cc367732 IM |
2111 | if (old_cpu != new_cpu) { |
2112 | schedstat_inc(p, se.nr_migrations); | |
2113 | if (task_hot(p, old_rq->clock, NULL)) | |
2114 | schedstat_inc(p, se.nr_forced2_migrations); | |
2115 | } | |
6cfb0d5d | 2116 | #endif |
2830cf8c SV |
2117 | p->se.vruntime -= old_cfsrq->min_vruntime - |
2118 | new_cfsrq->min_vruntime; | |
dd41f596 IM |
2119 | |
2120 | __set_task_cpu(p, new_cpu); | |
c65cc870 IM |
2121 | } |
2122 | ||
70b97a7f | 2123 | struct migration_req { |
1da177e4 | 2124 | struct list_head list; |
1da177e4 | 2125 | |
36c8b586 | 2126 | struct task_struct *task; |
1da177e4 LT |
2127 | int dest_cpu; |
2128 | ||
1da177e4 | 2129 | struct completion done; |
70b97a7f | 2130 | }; |
1da177e4 LT |
2131 | |
2132 | /* | |
2133 | * The task's runqueue lock must be held. | |
2134 | * Returns true if you have to wait for migration thread. | |
2135 | */ | |
36c8b586 | 2136 | static int |
70b97a7f | 2137 | migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req) |
1da177e4 | 2138 | { |
70b97a7f | 2139 | struct rq *rq = task_rq(p); |
1da177e4 LT |
2140 | |
2141 | /* | |
2142 | * If the task is not on a runqueue (and not running), then | |
2143 | * it is sufficient to simply update the task's cpu field. | |
2144 | */ | |
dd41f596 | 2145 | if (!p->se.on_rq && !task_running(rq, p)) { |
1da177e4 LT |
2146 | set_task_cpu(p, dest_cpu); |
2147 | return 0; | |
2148 | } | |
2149 | ||
2150 | init_completion(&req->done); | |
1da177e4 LT |
2151 | req->task = p; |
2152 | req->dest_cpu = dest_cpu; | |
2153 | list_add(&req->list, &rq->migration_queue); | |
48f24c4d | 2154 | |
1da177e4 LT |
2155 | return 1; |
2156 | } | |
2157 | ||
2158 | /* | |
2159 | * wait_task_inactive - wait for a thread to unschedule. | |
2160 | * | |
2161 | * The caller must ensure that the task *will* unschedule sometime soon, | |
2162 | * else this function might spin for a *long* time. This function can't | |
2163 | * be called with interrupts off, or it may introduce deadlock with | |
2164 | * smp_call_function() if an IPI is sent by the same process we are | |
2165 | * waiting to become inactive. | |
2166 | */ | |
36c8b586 | 2167 | void wait_task_inactive(struct task_struct *p) |
1da177e4 LT |
2168 | { |
2169 | unsigned long flags; | |
dd41f596 | 2170 | int running, on_rq; |
70b97a7f | 2171 | struct rq *rq; |
1da177e4 | 2172 | |
3a5c359a AK |
2173 | for (;;) { |
2174 | /* | |
2175 | * We do the initial early heuristics without holding | |
2176 | * any task-queue locks at all. We'll only try to get | |
2177 | * the runqueue lock when things look like they will | |
2178 | * work out! | |
2179 | */ | |
2180 | rq = task_rq(p); | |
fa490cfd | 2181 | |
3a5c359a AK |
2182 | /* |
2183 | * If the task is actively running on another CPU | |
2184 | * still, just relax and busy-wait without holding | |
2185 | * any locks. | |
2186 | * | |
2187 | * NOTE! Since we don't hold any locks, it's not | |
2188 | * even sure that "rq" stays as the right runqueue! | |
2189 | * But we don't care, since "task_running()" will | |
2190 | * return false if the runqueue has changed and p | |
2191 | * is actually now running somewhere else! | |
2192 | */ | |
2193 | while (task_running(rq, p)) | |
2194 | cpu_relax(); | |
fa490cfd | 2195 | |
3a5c359a AK |
2196 | /* |
2197 | * Ok, time to look more closely! We need the rq | |
2198 | * lock now, to be *sure*. If we're wrong, we'll | |
2199 | * just go back and repeat. | |
2200 | */ | |
2201 | rq = task_rq_lock(p, &flags); | |
2202 | running = task_running(rq, p); | |
2203 | on_rq = p->se.on_rq; | |
2204 | task_rq_unlock(rq, &flags); | |
fa490cfd | 2205 | |
3a5c359a AK |
2206 | /* |
2207 | * Was it really running after all now that we | |
2208 | * checked with the proper locks actually held? | |
2209 | * | |
2210 | * Oops. Go back and try again.. | |
2211 | */ | |
2212 | if (unlikely(running)) { | |
2213 | cpu_relax(); | |
2214 | continue; | |
2215 | } | |
fa490cfd | 2216 | |
3a5c359a AK |
2217 | /* |
2218 | * It's not enough that it's not actively running, | |
2219 | * it must be off the runqueue _entirely_, and not | |
2220 | * preempted! | |
2221 | * | |
2222 | * So if it wa still runnable (but just not actively | |
2223 | * running right now), it's preempted, and we should | |
2224 | * yield - it could be a while. | |
2225 | */ | |
2226 | if (unlikely(on_rq)) { | |
2227 | schedule_timeout_uninterruptible(1); | |
2228 | continue; | |
2229 | } | |
fa490cfd | 2230 | |
3a5c359a AK |
2231 | /* |
2232 | * Ahh, all good. It wasn't running, and it wasn't | |
2233 | * runnable, which means that it will never become | |
2234 | * running in the future either. We're all done! | |
2235 | */ | |
2236 | break; | |
2237 | } | |
1da177e4 LT |
2238 | } |
2239 | ||
2240 | /*** | |
2241 | * kick_process - kick a running thread to enter/exit the kernel | |
2242 | * @p: the to-be-kicked thread | |
2243 | * | |
2244 | * Cause a process which is running on another CPU to enter | |
2245 | * kernel-mode, without any delay. (to get signals handled.) | |
2246 | * | |
2247 | * NOTE: this function doesnt have to take the runqueue lock, | |
2248 | * because all it wants to ensure is that the remote task enters | |
2249 | * the kernel. If the IPI races and the task has been migrated | |
2250 | * to another CPU then no harm is done and the purpose has been | |
2251 | * achieved as well. | |
2252 | */ | |
36c8b586 | 2253 | void kick_process(struct task_struct *p) |
1da177e4 LT |
2254 | { |
2255 | int cpu; | |
2256 | ||
2257 | preempt_disable(); | |
2258 | cpu = task_cpu(p); | |
2259 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
2260 | smp_send_reschedule(cpu); | |
2261 | preempt_enable(); | |
2262 | } | |
2263 | ||
2264 | /* | |
2dd73a4f PW |
2265 | * Return a low guess at the load of a migration-source cpu weighted |
2266 | * according to the scheduling class and "nice" value. | |
1da177e4 LT |
2267 | * |
2268 | * We want to under-estimate the load of migration sources, to | |
2269 | * balance conservatively. | |
2270 | */ | |
a9957449 | 2271 | static unsigned long source_load(int cpu, int type) |
1da177e4 | 2272 | { |
70b97a7f | 2273 | struct rq *rq = cpu_rq(cpu); |
dd41f596 | 2274 | unsigned long total = weighted_cpuload(cpu); |
2dd73a4f | 2275 | |
3b0bd9bc | 2276 | if (type == 0) |
dd41f596 | 2277 | return total; |
b910472d | 2278 | |
dd41f596 | 2279 | return min(rq->cpu_load[type-1], total); |
1da177e4 LT |
2280 | } |
2281 | ||
2282 | /* | |
2dd73a4f PW |
2283 | * Return a high guess at the load of a migration-target cpu weighted |
2284 | * according to the scheduling class and "nice" value. | |
1da177e4 | 2285 | */ |
a9957449 | 2286 | static unsigned long target_load(int cpu, int type) |
1da177e4 | 2287 | { |
70b97a7f | 2288 | struct rq *rq = cpu_rq(cpu); |
dd41f596 | 2289 | unsigned long total = weighted_cpuload(cpu); |
2dd73a4f | 2290 | |
7897986b | 2291 | if (type == 0) |
dd41f596 | 2292 | return total; |
3b0bd9bc | 2293 | |
dd41f596 | 2294 | return max(rq->cpu_load[type-1], total); |
2dd73a4f PW |
2295 | } |
2296 | ||
2297 | /* | |
2298 | * Return the average load per task on the cpu's run queue | |
2299 | */ | |
e7693a36 | 2300 | static unsigned long cpu_avg_load_per_task(int cpu) |
2dd73a4f | 2301 | { |
70b97a7f | 2302 | struct rq *rq = cpu_rq(cpu); |
dd41f596 | 2303 | unsigned long total = weighted_cpuload(cpu); |
2dd73a4f PW |
2304 | unsigned long n = rq->nr_running; |
2305 | ||
dd41f596 | 2306 | return n ? total / n : SCHED_LOAD_SCALE; |
1da177e4 LT |
2307 | } |
2308 | ||
147cbb4b NP |
2309 | /* |
2310 | * find_idlest_group finds and returns the least busy CPU group within the | |
2311 | * domain. | |
2312 | */ | |
2313 | static struct sched_group * | |
2314 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
2315 | { | |
2316 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
2317 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
2318 | int load_idx = sd->forkexec_idx; | |
2319 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
2320 | ||
2321 | do { | |
2322 | unsigned long load, avg_load; | |
2323 | int local_group; | |
2324 | int i; | |
2325 | ||
da5a5522 BD |
2326 | /* Skip over this group if it has no CPUs allowed */ |
2327 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
3a5c359a | 2328 | continue; |
da5a5522 | 2329 | |
147cbb4b | 2330 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
2331 | |
2332 | /* Tally up the load of all CPUs in the group */ | |
2333 | avg_load = 0; | |
2334 | ||
2335 | for_each_cpu_mask(i, group->cpumask) { | |
2336 | /* Bias balancing toward cpus of our domain */ | |
2337 | if (local_group) | |
2338 | load = source_load(i, load_idx); | |
2339 | else | |
2340 | load = target_load(i, load_idx); | |
2341 | ||
2342 | avg_load += load; | |
2343 | } | |
2344 | ||
2345 | /* Adjust by relative CPU power of the group */ | |
5517d86b ED |
2346 | avg_load = sg_div_cpu_power(group, |
2347 | avg_load * SCHED_LOAD_SCALE); | |
147cbb4b NP |
2348 | |
2349 | if (local_group) { | |
2350 | this_load = avg_load; | |
2351 | this = group; | |
2352 | } else if (avg_load < min_load) { | |
2353 | min_load = avg_load; | |
2354 | idlest = group; | |
2355 | } | |
3a5c359a | 2356 | } while (group = group->next, group != sd->groups); |
147cbb4b NP |
2357 | |
2358 | if (!idlest || 100*this_load < imbalance*min_load) | |
2359 | return NULL; | |
2360 | return idlest; | |
2361 | } | |
2362 | ||
2363 | /* | |
0feaece9 | 2364 | * find_idlest_cpu - find the idlest cpu among the cpus in group. |
147cbb4b | 2365 | */ |
95cdf3b7 | 2366 | static int |
7c16ec58 MT |
2367 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu, |
2368 | cpumask_t *tmp) | |
147cbb4b NP |
2369 | { |
2370 | unsigned long load, min_load = ULONG_MAX; | |
2371 | int idlest = -1; | |
2372 | int i; | |
2373 | ||
da5a5522 | 2374 | /* Traverse only the allowed CPUs */ |
7c16ec58 | 2375 | cpus_and(*tmp, group->cpumask, p->cpus_allowed); |
da5a5522 | 2376 | |
7c16ec58 | 2377 | for_each_cpu_mask(i, *tmp) { |
2dd73a4f | 2378 | load = weighted_cpuload(i); |
147cbb4b NP |
2379 | |
2380 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
2381 | min_load = load; | |
2382 | idlest = i; | |
2383 | } | |
2384 | } | |
2385 | ||
2386 | return idlest; | |
2387 | } | |
2388 | ||
476d139c NP |
2389 | /* |
2390 | * sched_balance_self: balance the current task (running on cpu) in domains | |
2391 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
2392 | * SD_BALANCE_EXEC. | |
2393 | * | |
2394 | * Balance, ie. select the least loaded group. | |
2395 | * | |
2396 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
2397 | * | |
2398 | * preempt must be disabled. | |
2399 | */ | |
2400 | static int sched_balance_self(int cpu, int flag) | |
2401 | { | |
2402 | struct task_struct *t = current; | |
2403 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 2404 | |
c96d145e | 2405 | for_each_domain(cpu, tmp) { |
9761eea8 IM |
2406 | /* |
2407 | * If power savings logic is enabled for a domain, stop there. | |
2408 | */ | |
5c45bf27 SS |
2409 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
2410 | break; | |
476d139c NP |
2411 | if (tmp->flags & flag) |
2412 | sd = tmp; | |
c96d145e | 2413 | } |
476d139c NP |
2414 | |
2415 | while (sd) { | |
7c16ec58 | 2416 | cpumask_t span, tmpmask; |
476d139c | 2417 | struct sched_group *group; |
1a848870 SS |
2418 | int new_cpu, weight; |
2419 | ||
2420 | if (!(sd->flags & flag)) { | |
2421 | sd = sd->child; | |
2422 | continue; | |
2423 | } | |
476d139c NP |
2424 | |
2425 | span = sd->span; | |
2426 | group = find_idlest_group(sd, t, cpu); | |
1a848870 SS |
2427 | if (!group) { |
2428 | sd = sd->child; | |
2429 | continue; | |
2430 | } | |
476d139c | 2431 | |
7c16ec58 | 2432 | new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask); |
1a848870 SS |
2433 | if (new_cpu == -1 || new_cpu == cpu) { |
2434 | /* Now try balancing at a lower domain level of cpu */ | |
2435 | sd = sd->child; | |
2436 | continue; | |
2437 | } | |
476d139c | 2438 | |
1a848870 | 2439 | /* Now try balancing at a lower domain level of new_cpu */ |
476d139c | 2440 | cpu = new_cpu; |
476d139c NP |
2441 | sd = NULL; |
2442 | weight = cpus_weight(span); | |
2443 | for_each_domain(cpu, tmp) { | |
2444 | if (weight <= cpus_weight(tmp->span)) | |
2445 | break; | |
2446 | if (tmp->flags & flag) | |
2447 | sd = tmp; | |
2448 | } | |
2449 | /* while loop will break here if sd == NULL */ | |
2450 | } | |
2451 | ||
2452 | return cpu; | |
2453 | } | |
2454 | ||
2455 | #endif /* CONFIG_SMP */ | |
1da177e4 | 2456 | |
1da177e4 LT |
2457 | /*** |
2458 | * try_to_wake_up - wake up a thread | |
2459 | * @p: the to-be-woken-up thread | |
2460 | * @state: the mask of task states that can be woken | |
2461 | * @sync: do a synchronous wakeup? | |
2462 | * | |
2463 | * Put it on the run-queue if it's not already there. The "current" | |
2464 | * thread is always on the run-queue (except when the actual | |
2465 | * re-schedule is in progress), and as such you're allowed to do | |
2466 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
2467 | * runnable without the overhead of this. | |
2468 | * | |
2469 | * returns failure only if the task is already active. | |
2470 | */ | |
36c8b586 | 2471 | static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync) |
1da177e4 | 2472 | { |
cc367732 | 2473 | int cpu, orig_cpu, this_cpu, success = 0; |
1da177e4 LT |
2474 | unsigned long flags; |
2475 | long old_state; | |
70b97a7f | 2476 | struct rq *rq; |
1da177e4 | 2477 | |
b85d0667 IM |
2478 | if (!sched_feat(SYNC_WAKEUPS)) |
2479 | sync = 0; | |
2480 | ||
04e2f174 | 2481 | smp_wmb(); |
1da177e4 LT |
2482 | rq = task_rq_lock(p, &flags); |
2483 | old_state = p->state; | |
2484 | if (!(old_state & state)) | |
2485 | goto out; | |
2486 | ||
dd41f596 | 2487 | if (p->se.on_rq) |
1da177e4 LT |
2488 | goto out_running; |
2489 | ||
2490 | cpu = task_cpu(p); | |
cc367732 | 2491 | orig_cpu = cpu; |
1da177e4 LT |
2492 | this_cpu = smp_processor_id(); |
2493 | ||
2494 | #ifdef CONFIG_SMP | |
2495 | if (unlikely(task_running(rq, p))) | |
2496 | goto out_activate; | |
2497 | ||
5d2f5a61 DA |
2498 | cpu = p->sched_class->select_task_rq(p, sync); |
2499 | if (cpu != orig_cpu) { | |
2500 | set_task_cpu(p, cpu); | |
1da177e4 LT |
2501 | task_rq_unlock(rq, &flags); |
2502 | /* might preempt at this point */ | |
2503 | rq = task_rq_lock(p, &flags); | |
2504 | old_state = p->state; | |
2505 | if (!(old_state & state)) | |
2506 | goto out; | |
dd41f596 | 2507 | if (p->se.on_rq) |
1da177e4 LT |
2508 | goto out_running; |
2509 | ||
2510 | this_cpu = smp_processor_id(); | |
2511 | cpu = task_cpu(p); | |
2512 | } | |
2513 | ||
e7693a36 GH |
2514 | #ifdef CONFIG_SCHEDSTATS |
2515 | schedstat_inc(rq, ttwu_count); | |
2516 | if (cpu == this_cpu) | |
2517 | schedstat_inc(rq, ttwu_local); | |
2518 | else { | |
2519 | struct sched_domain *sd; | |
2520 | for_each_domain(this_cpu, sd) { | |
2521 | if (cpu_isset(cpu, sd->span)) { | |
2522 | schedstat_inc(sd, ttwu_wake_remote); | |
2523 | break; | |
2524 | } | |
2525 | } | |
2526 | } | |
e7693a36 GH |
2527 | #endif |
2528 | ||
1da177e4 LT |
2529 | out_activate: |
2530 | #endif /* CONFIG_SMP */ | |
cc367732 IM |
2531 | schedstat_inc(p, se.nr_wakeups); |
2532 | if (sync) | |
2533 | schedstat_inc(p, se.nr_wakeups_sync); | |
2534 | if (orig_cpu != cpu) | |
2535 | schedstat_inc(p, se.nr_wakeups_migrate); | |
2536 | if (cpu == this_cpu) | |
2537 | schedstat_inc(p, se.nr_wakeups_local); | |
2538 | else | |
2539 | schedstat_inc(p, se.nr_wakeups_remote); | |
2daa3577 | 2540 | update_rq_clock(rq); |
dd41f596 | 2541 | activate_task(rq, p, 1); |
1da177e4 LT |
2542 | success = 1; |
2543 | ||
2544 | out_running: | |
4ae7d5ce IM |
2545 | check_preempt_curr(rq, p); |
2546 | ||
1da177e4 | 2547 | p->state = TASK_RUNNING; |
9a897c5a SR |
2548 | #ifdef CONFIG_SMP |
2549 | if (p->sched_class->task_wake_up) | |
2550 | p->sched_class->task_wake_up(rq, p); | |
2551 | #endif | |
1da177e4 LT |
2552 | out: |
2553 | task_rq_unlock(rq, &flags); | |
2554 | ||
2555 | return success; | |
2556 | } | |
2557 | ||
7ad5b3a5 | 2558 | int wake_up_process(struct task_struct *p) |
1da177e4 | 2559 | { |
d9514f6c | 2560 | return try_to_wake_up(p, TASK_ALL, 0); |
1da177e4 | 2561 | } |
1da177e4 LT |
2562 | EXPORT_SYMBOL(wake_up_process); |
2563 | ||
7ad5b3a5 | 2564 | int wake_up_state(struct task_struct *p, unsigned int state) |
1da177e4 LT |
2565 | { |
2566 | return try_to_wake_up(p, state, 0); | |
2567 | } | |
2568 | ||
1da177e4 LT |
2569 | /* |
2570 | * Perform scheduler related setup for a newly forked process p. | |
2571 | * p is forked by current. | |
dd41f596 IM |
2572 | * |
2573 | * __sched_fork() is basic setup used by init_idle() too: | |
2574 | */ | |
2575 | static void __sched_fork(struct task_struct *p) | |
2576 | { | |
dd41f596 IM |
2577 | p->se.exec_start = 0; |
2578 | p->se.sum_exec_runtime = 0; | |
f6cf891c | 2579 | p->se.prev_sum_exec_runtime = 0; |
4ae7d5ce IM |
2580 | p->se.last_wakeup = 0; |
2581 | p->se.avg_overlap = 0; | |
6cfb0d5d IM |
2582 | |
2583 | #ifdef CONFIG_SCHEDSTATS | |
2584 | p->se.wait_start = 0; | |
dd41f596 IM |
2585 | p->se.sum_sleep_runtime = 0; |
2586 | p->se.sleep_start = 0; | |
dd41f596 IM |
2587 | p->se.block_start = 0; |
2588 | p->se.sleep_max = 0; | |
2589 | p->se.block_max = 0; | |
2590 | p->se.exec_max = 0; | |
eba1ed4b | 2591 | p->se.slice_max = 0; |
dd41f596 | 2592 | p->se.wait_max = 0; |
6cfb0d5d | 2593 | #endif |
476d139c | 2594 | |
fa717060 | 2595 | INIT_LIST_HEAD(&p->rt.run_list); |
dd41f596 | 2596 | p->se.on_rq = 0; |
4a55bd5e | 2597 | INIT_LIST_HEAD(&p->se.group_node); |
476d139c | 2598 | |
e107be36 AK |
2599 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2600 | INIT_HLIST_HEAD(&p->preempt_notifiers); | |
2601 | #endif | |
2602 | ||
1da177e4 LT |
2603 | /* |
2604 | * We mark the process as running here, but have not actually | |
2605 | * inserted it onto the runqueue yet. This guarantees that | |
2606 | * nobody will actually run it, and a signal or other external | |
2607 | * event cannot wake it up and insert it on the runqueue either. | |
2608 | */ | |
2609 | p->state = TASK_RUNNING; | |
dd41f596 IM |
2610 | } |
2611 | ||
2612 | /* | |
2613 | * fork()/clone()-time setup: | |
2614 | */ | |
2615 | void sched_fork(struct task_struct *p, int clone_flags) | |
2616 | { | |
2617 | int cpu = get_cpu(); | |
2618 | ||
2619 | __sched_fork(p); | |
2620 | ||
2621 | #ifdef CONFIG_SMP | |
2622 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
2623 | #endif | |
02e4bac2 | 2624 | set_task_cpu(p, cpu); |
b29739f9 IM |
2625 | |
2626 | /* | |
2627 | * Make sure we do not leak PI boosting priority to the child: | |
2628 | */ | |
2629 | p->prio = current->normal_prio; | |
2ddbf952 HS |
2630 | if (!rt_prio(p->prio)) |
2631 | p->sched_class = &fair_sched_class; | |
b29739f9 | 2632 | |
52f17b6c | 2633 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
dd41f596 | 2634 | if (likely(sched_info_on())) |
52f17b6c | 2635 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
1da177e4 | 2636 | #endif |
d6077cb8 | 2637 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4866cde0 NP |
2638 | p->oncpu = 0; |
2639 | #endif | |
1da177e4 | 2640 | #ifdef CONFIG_PREEMPT |
4866cde0 | 2641 | /* Want to start with kernel preemption disabled. */ |
a1261f54 | 2642 | task_thread_info(p)->preempt_count = 1; |
1da177e4 | 2643 | #endif |
476d139c | 2644 | put_cpu(); |
1da177e4 LT |
2645 | } |
2646 | ||
2647 | /* | |
2648 | * wake_up_new_task - wake up a newly created task for the first time. | |
2649 | * | |
2650 | * This function will do some initial scheduler statistics housekeeping | |
2651 | * that must be done for every newly created context, then puts the task | |
2652 | * on the runqueue and wakes it. | |
2653 | */ | |
7ad5b3a5 | 2654 | void wake_up_new_task(struct task_struct *p, unsigned long clone_flags) |
1da177e4 LT |
2655 | { |
2656 | unsigned long flags; | |
dd41f596 | 2657 | struct rq *rq; |
1da177e4 LT |
2658 | |
2659 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 2660 | BUG_ON(p->state != TASK_RUNNING); |
a8e504d2 | 2661 | update_rq_clock(rq); |
1da177e4 LT |
2662 | |
2663 | p->prio = effective_prio(p); | |
2664 | ||
b9dca1e0 | 2665 | if (!p->sched_class->task_new || !current->se.on_rq) { |
dd41f596 | 2666 | activate_task(rq, p, 0); |
1da177e4 | 2667 | } else { |
1da177e4 | 2668 | /* |
dd41f596 IM |
2669 | * Let the scheduling class do new task startup |
2670 | * management (if any): | |
1da177e4 | 2671 | */ |
ee0827d8 | 2672 | p->sched_class->task_new(rq, p); |
18d95a28 | 2673 | inc_nr_running(rq); |
1da177e4 | 2674 | } |
dd41f596 | 2675 | check_preempt_curr(rq, p); |
9a897c5a SR |
2676 | #ifdef CONFIG_SMP |
2677 | if (p->sched_class->task_wake_up) | |
2678 | p->sched_class->task_wake_up(rq, p); | |
2679 | #endif | |
dd41f596 | 2680 | task_rq_unlock(rq, &flags); |
1da177e4 LT |
2681 | } |
2682 | ||
e107be36 AK |
2683 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2684 | ||
2685 | /** | |
421cee29 RD |
2686 | * preempt_notifier_register - tell me when current is being being preempted & rescheduled |
2687 | * @notifier: notifier struct to register | |
e107be36 AK |
2688 | */ |
2689 | void preempt_notifier_register(struct preempt_notifier *notifier) | |
2690 | { | |
2691 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); | |
2692 | } | |
2693 | EXPORT_SYMBOL_GPL(preempt_notifier_register); | |
2694 | ||
2695 | /** | |
2696 | * preempt_notifier_unregister - no longer interested in preemption notifications | |
421cee29 | 2697 | * @notifier: notifier struct to unregister |
e107be36 AK |
2698 | * |
2699 | * This is safe to call from within a preemption notifier. | |
2700 | */ | |
2701 | void preempt_notifier_unregister(struct preempt_notifier *notifier) | |
2702 | { | |
2703 | hlist_del(¬ifier->link); | |
2704 | } | |
2705 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); | |
2706 | ||
2707 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2708 | { | |
2709 | struct preempt_notifier *notifier; | |
2710 | struct hlist_node *node; | |
2711 | ||
2712 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2713 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); | |
2714 | } | |
2715 | ||
2716 | static void | |
2717 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2718 | struct task_struct *next) | |
2719 | { | |
2720 | struct preempt_notifier *notifier; | |
2721 | struct hlist_node *node; | |
2722 | ||
2723 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2724 | notifier->ops->sched_out(notifier, next); | |
2725 | } | |
2726 | ||
2727 | #else | |
2728 | ||
2729 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2730 | { | |
2731 | } | |
2732 | ||
2733 | static void | |
2734 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2735 | struct task_struct *next) | |
2736 | { | |
2737 | } | |
2738 | ||
2739 | #endif | |
2740 | ||
4866cde0 NP |
2741 | /** |
2742 | * prepare_task_switch - prepare to switch tasks | |
2743 | * @rq: the runqueue preparing to switch | |
421cee29 | 2744 | * @prev: the current task that is being switched out |
4866cde0 NP |
2745 | * @next: the task we are going to switch to. |
2746 | * | |
2747 | * This is called with the rq lock held and interrupts off. It must | |
2748 | * be paired with a subsequent finish_task_switch after the context | |
2749 | * switch. | |
2750 | * | |
2751 | * prepare_task_switch sets up locking and calls architecture specific | |
2752 | * hooks. | |
2753 | */ | |
e107be36 AK |
2754 | static inline void |
2755 | prepare_task_switch(struct rq *rq, struct task_struct *prev, | |
2756 | struct task_struct *next) | |
4866cde0 | 2757 | { |
e107be36 | 2758 | fire_sched_out_preempt_notifiers(prev, next); |
4866cde0 NP |
2759 | prepare_lock_switch(rq, next); |
2760 | prepare_arch_switch(next); | |
2761 | } | |
2762 | ||
1da177e4 LT |
2763 | /** |
2764 | * finish_task_switch - clean up after a task-switch | |
344babaa | 2765 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
2766 | * @prev: the thread we just switched away from. |
2767 | * | |
4866cde0 NP |
2768 | * finish_task_switch must be called after the context switch, paired |
2769 | * with a prepare_task_switch call before the context switch. | |
2770 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
2771 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
2772 | * |
2773 | * Note that we may have delayed dropping an mm in context_switch(). If | |
41a2d6cf | 2774 | * so, we finish that here outside of the runqueue lock. (Doing it |
1da177e4 LT |
2775 | * with the lock held can cause deadlocks; see schedule() for |
2776 | * details.) | |
2777 | */ | |
a9957449 | 2778 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) |
1da177e4 LT |
2779 | __releases(rq->lock) |
2780 | { | |
1da177e4 | 2781 | struct mm_struct *mm = rq->prev_mm; |
55a101f8 | 2782 | long prev_state; |
1da177e4 LT |
2783 | |
2784 | rq->prev_mm = NULL; | |
2785 | ||
2786 | /* | |
2787 | * A task struct has one reference for the use as "current". | |
c394cc9f | 2788 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
55a101f8 ON |
2789 | * schedule one last time. The schedule call will never return, and |
2790 | * the scheduled task must drop that reference. | |
c394cc9f | 2791 | * The test for TASK_DEAD must occur while the runqueue locks are |
1da177e4 LT |
2792 | * still held, otherwise prev could be scheduled on another cpu, die |
2793 | * there before we look at prev->state, and then the reference would | |
2794 | * be dropped twice. | |
2795 | * Manfred Spraul <manfred@colorfullife.com> | |
2796 | */ | |
55a101f8 | 2797 | prev_state = prev->state; |
4866cde0 NP |
2798 | finish_arch_switch(prev); |
2799 | finish_lock_switch(rq, prev); | |
9a897c5a SR |
2800 | #ifdef CONFIG_SMP |
2801 | if (current->sched_class->post_schedule) | |
2802 | current->sched_class->post_schedule(rq); | |
2803 | #endif | |
e8fa1362 | 2804 | |
e107be36 | 2805 | fire_sched_in_preempt_notifiers(current); |
1da177e4 LT |
2806 | if (mm) |
2807 | mmdrop(mm); | |
c394cc9f | 2808 | if (unlikely(prev_state == TASK_DEAD)) { |
c6fd91f0 | 2809 | /* |
2810 | * Remove function-return probe instances associated with this | |
2811 | * task and put them back on the free list. | |
9761eea8 | 2812 | */ |
c6fd91f0 | 2813 | kprobe_flush_task(prev); |
1da177e4 | 2814 | put_task_struct(prev); |
c6fd91f0 | 2815 | } |
1da177e4 LT |
2816 | } |
2817 | ||
2818 | /** | |
2819 | * schedule_tail - first thing a freshly forked thread must call. | |
2820 | * @prev: the thread we just switched away from. | |
2821 | */ | |
36c8b586 | 2822 | asmlinkage void schedule_tail(struct task_struct *prev) |
1da177e4 LT |
2823 | __releases(rq->lock) |
2824 | { | |
70b97a7f IM |
2825 | struct rq *rq = this_rq(); |
2826 | ||
4866cde0 NP |
2827 | finish_task_switch(rq, prev); |
2828 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
2829 | /* In this case, finish_task_switch does not reenable preemption */ | |
2830 | preempt_enable(); | |
2831 | #endif | |
1da177e4 | 2832 | if (current->set_child_tid) |
b488893a | 2833 | put_user(task_pid_vnr(current), current->set_child_tid); |
1da177e4 LT |
2834 | } |
2835 | ||
2836 | /* | |
2837 | * context_switch - switch to the new MM and the new | |
2838 | * thread's register state. | |
2839 | */ | |
dd41f596 | 2840 | static inline void |
70b97a7f | 2841 | context_switch(struct rq *rq, struct task_struct *prev, |
36c8b586 | 2842 | struct task_struct *next) |
1da177e4 | 2843 | { |
dd41f596 | 2844 | struct mm_struct *mm, *oldmm; |
1da177e4 | 2845 | |
e107be36 | 2846 | prepare_task_switch(rq, prev, next); |
dd41f596 IM |
2847 | mm = next->mm; |
2848 | oldmm = prev->active_mm; | |
9226d125 ZA |
2849 | /* |
2850 | * For paravirt, this is coupled with an exit in switch_to to | |
2851 | * combine the page table reload and the switch backend into | |
2852 | * one hypercall. | |
2853 | */ | |
2854 | arch_enter_lazy_cpu_mode(); | |
2855 | ||
dd41f596 | 2856 | if (unlikely(!mm)) { |
1da177e4 LT |
2857 | next->active_mm = oldmm; |
2858 | atomic_inc(&oldmm->mm_count); | |
2859 | enter_lazy_tlb(oldmm, next); | |
2860 | } else | |
2861 | switch_mm(oldmm, mm, next); | |
2862 | ||
dd41f596 | 2863 | if (unlikely(!prev->mm)) { |
1da177e4 | 2864 | prev->active_mm = NULL; |
1da177e4 LT |
2865 | rq->prev_mm = oldmm; |
2866 | } | |
3a5f5e48 IM |
2867 | /* |
2868 | * Since the runqueue lock will be released by the next | |
2869 | * task (which is an invalid locking op but in the case | |
2870 | * of the scheduler it's an obvious special-case), so we | |
2871 | * do an early lockdep release here: | |
2872 | */ | |
2873 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
8a25d5de | 2874 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
3a5f5e48 | 2875 | #endif |
1da177e4 LT |
2876 | |
2877 | /* Here we just switch the register state and the stack. */ | |
2878 | switch_to(prev, next, prev); | |
2879 | ||
dd41f596 IM |
2880 | barrier(); |
2881 | /* | |
2882 | * this_rq must be evaluated again because prev may have moved | |
2883 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
2884 | * frame will be invalid. | |
2885 | */ | |
2886 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
2887 | } |
2888 | ||
2889 | /* | |
2890 | * nr_running, nr_uninterruptible and nr_context_switches: | |
2891 | * | |
2892 | * externally visible scheduler statistics: current number of runnable | |
2893 | * threads, current number of uninterruptible-sleeping threads, total | |
2894 | * number of context switches performed since bootup. | |
2895 | */ | |
2896 | unsigned long nr_running(void) | |
2897 | { | |
2898 | unsigned long i, sum = 0; | |
2899 | ||
2900 | for_each_online_cpu(i) | |
2901 | sum += cpu_rq(i)->nr_running; | |
2902 | ||
2903 | return sum; | |
2904 | } | |
2905 | ||
2906 | unsigned long nr_uninterruptible(void) | |
2907 | { | |
2908 | unsigned long i, sum = 0; | |
2909 | ||
0a945022 | 2910 | for_each_possible_cpu(i) |
1da177e4 LT |
2911 | sum += cpu_rq(i)->nr_uninterruptible; |
2912 | ||
2913 | /* | |
2914 | * Since we read the counters lockless, it might be slightly | |
2915 | * inaccurate. Do not allow it to go below zero though: | |
2916 | */ | |
2917 | if (unlikely((long)sum < 0)) | |
2918 | sum = 0; | |
2919 | ||
2920 | return sum; | |
2921 | } | |
2922 | ||
2923 | unsigned long long nr_context_switches(void) | |
2924 | { | |
cc94abfc SR |
2925 | int i; |
2926 | unsigned long long sum = 0; | |
1da177e4 | 2927 | |
0a945022 | 2928 | for_each_possible_cpu(i) |
1da177e4 LT |
2929 | sum += cpu_rq(i)->nr_switches; |
2930 | ||
2931 | return sum; | |
2932 | } | |
2933 | ||
2934 | unsigned long nr_iowait(void) | |
2935 | { | |
2936 | unsigned long i, sum = 0; | |
2937 | ||
0a945022 | 2938 | for_each_possible_cpu(i) |
1da177e4 LT |
2939 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
2940 | ||
2941 | return sum; | |
2942 | } | |
2943 | ||
db1b1fef JS |
2944 | unsigned long nr_active(void) |
2945 | { | |
2946 | unsigned long i, running = 0, uninterruptible = 0; | |
2947 | ||
2948 | for_each_online_cpu(i) { | |
2949 | running += cpu_rq(i)->nr_running; | |
2950 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
2951 | } | |
2952 | ||
2953 | if (unlikely((long)uninterruptible < 0)) | |
2954 | uninterruptible = 0; | |
2955 | ||
2956 | return running + uninterruptible; | |
2957 | } | |
2958 | ||
48f24c4d | 2959 | /* |
dd41f596 IM |
2960 | * Update rq->cpu_load[] statistics. This function is usually called every |
2961 | * scheduler tick (TICK_NSEC). | |
48f24c4d | 2962 | */ |
dd41f596 | 2963 | static void update_cpu_load(struct rq *this_rq) |
48f24c4d | 2964 | { |
495eca49 | 2965 | unsigned long this_load = this_rq->load.weight; |
dd41f596 IM |
2966 | int i, scale; |
2967 | ||
2968 | this_rq->nr_load_updates++; | |
dd41f596 IM |
2969 | |
2970 | /* Update our load: */ | |
2971 | for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
2972 | unsigned long old_load, new_load; | |
2973 | ||
2974 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
2975 | ||
2976 | old_load = this_rq->cpu_load[i]; | |
2977 | new_load = this_load; | |
a25707f3 IM |
2978 | /* |
2979 | * Round up the averaging division if load is increasing. This | |
2980 | * prevents us from getting stuck on 9 if the load is 10, for | |
2981 | * example. | |
2982 | */ | |
2983 | if (new_load > old_load) | |
2984 | new_load += scale-1; | |
dd41f596 IM |
2985 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i; |
2986 | } | |
48f24c4d IM |
2987 | } |
2988 | ||
dd41f596 IM |
2989 | #ifdef CONFIG_SMP |
2990 | ||
1da177e4 LT |
2991 | /* |
2992 | * double_rq_lock - safely lock two runqueues | |
2993 | * | |
2994 | * Note this does not disable interrupts like task_rq_lock, | |
2995 | * you need to do so manually before calling. | |
2996 | */ | |
70b97a7f | 2997 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
2998 | __acquires(rq1->lock) |
2999 | __acquires(rq2->lock) | |
3000 | { | |
054b9108 | 3001 | BUG_ON(!irqs_disabled()); |
1da177e4 LT |
3002 | if (rq1 == rq2) { |
3003 | spin_lock(&rq1->lock); | |
3004 | __acquire(rq2->lock); /* Fake it out ;) */ | |
3005 | } else { | |
c96d145e | 3006 | if (rq1 < rq2) { |
1da177e4 LT |
3007 | spin_lock(&rq1->lock); |
3008 | spin_lock(&rq2->lock); | |
3009 | } else { | |
3010 | spin_lock(&rq2->lock); | |
3011 | spin_lock(&rq1->lock); | |
3012 | } | |
3013 | } | |
6e82a3be IM |
3014 | update_rq_clock(rq1); |
3015 | update_rq_clock(rq2); | |
1da177e4 LT |
3016 | } |
3017 | ||
3018 | /* | |
3019 | * double_rq_unlock - safely unlock two runqueues | |
3020 | * | |
3021 | * Note this does not restore interrupts like task_rq_unlock, | |
3022 | * you need to do so manually after calling. | |
3023 | */ | |
70b97a7f | 3024 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
3025 | __releases(rq1->lock) |
3026 | __releases(rq2->lock) | |
3027 | { | |
3028 | spin_unlock(&rq1->lock); | |
3029 | if (rq1 != rq2) | |
3030 | spin_unlock(&rq2->lock); | |
3031 | else | |
3032 | __release(rq2->lock); | |
3033 | } | |
3034 | ||
3035 | /* | |
3036 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
3037 | */ | |
e8fa1362 | 3038 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest) |
1da177e4 LT |
3039 | __releases(this_rq->lock) |
3040 | __acquires(busiest->lock) | |
3041 | __acquires(this_rq->lock) | |
3042 | { | |
e8fa1362 SR |
3043 | int ret = 0; |
3044 | ||
054b9108 KK |
3045 | if (unlikely(!irqs_disabled())) { |
3046 | /* printk() doesn't work good under rq->lock */ | |
3047 | spin_unlock(&this_rq->lock); | |
3048 | BUG_ON(1); | |
3049 | } | |
1da177e4 | 3050 | if (unlikely(!spin_trylock(&busiest->lock))) { |
c96d145e | 3051 | if (busiest < this_rq) { |
1da177e4 LT |
3052 | spin_unlock(&this_rq->lock); |
3053 | spin_lock(&busiest->lock); | |
3054 | spin_lock(&this_rq->lock); | |
e8fa1362 | 3055 | ret = 1; |
1da177e4 LT |
3056 | } else |
3057 | spin_lock(&busiest->lock); | |
3058 | } | |
e8fa1362 | 3059 | return ret; |
1da177e4 LT |
3060 | } |
3061 | ||
1da177e4 LT |
3062 | /* |
3063 | * If dest_cpu is allowed for this process, migrate the task to it. | |
3064 | * This is accomplished by forcing the cpu_allowed mask to only | |
41a2d6cf | 3065 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then |
1da177e4 LT |
3066 | * the cpu_allowed mask is restored. |
3067 | */ | |
36c8b586 | 3068 | static void sched_migrate_task(struct task_struct *p, int dest_cpu) |
1da177e4 | 3069 | { |
70b97a7f | 3070 | struct migration_req req; |
1da177e4 | 3071 | unsigned long flags; |
70b97a7f | 3072 | struct rq *rq; |
1da177e4 LT |
3073 | |
3074 | rq = task_rq_lock(p, &flags); | |
3075 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
3076 | || unlikely(cpu_is_offline(dest_cpu))) | |
3077 | goto out; | |
3078 | ||
3079 | /* force the process onto the specified CPU */ | |
3080 | if (migrate_task(p, dest_cpu, &req)) { | |
3081 | /* Need to wait for migration thread (might exit: take ref). */ | |
3082 | struct task_struct *mt = rq->migration_thread; | |
36c8b586 | 3083 | |
1da177e4 LT |
3084 | get_task_struct(mt); |
3085 | task_rq_unlock(rq, &flags); | |
3086 | wake_up_process(mt); | |
3087 | put_task_struct(mt); | |
3088 | wait_for_completion(&req.done); | |
36c8b586 | 3089 | |
1da177e4 LT |
3090 | return; |
3091 | } | |
3092 | out: | |
3093 | task_rq_unlock(rq, &flags); | |
3094 | } | |
3095 | ||
3096 | /* | |
476d139c NP |
3097 | * sched_exec - execve() is a valuable balancing opportunity, because at |
3098 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
3099 | */ |
3100 | void sched_exec(void) | |
3101 | { | |
1da177e4 | 3102 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 3103 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 3104 | put_cpu(); |
476d139c NP |
3105 | if (new_cpu != this_cpu) |
3106 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
3107 | } |
3108 | ||
3109 | /* | |
3110 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
3111 | * Both runqueues must be locked. | |
3112 | */ | |
dd41f596 IM |
3113 | static void pull_task(struct rq *src_rq, struct task_struct *p, |
3114 | struct rq *this_rq, int this_cpu) | |
1da177e4 | 3115 | { |
2e1cb74a | 3116 | deactivate_task(src_rq, p, 0); |
1da177e4 | 3117 | set_task_cpu(p, this_cpu); |
dd41f596 | 3118 | activate_task(this_rq, p, 0); |
1da177e4 LT |
3119 | /* |
3120 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
3121 | * to be always true for them. | |
3122 | */ | |
dd41f596 | 3123 | check_preempt_curr(this_rq, p); |
1da177e4 LT |
3124 | } |
3125 | ||
3126 | /* | |
3127 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3128 | */ | |
858119e1 | 3129 | static |
70b97a7f | 3130 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, |
d15bcfdb | 3131 | struct sched_domain *sd, enum cpu_idle_type idle, |
95cdf3b7 | 3132 | int *all_pinned) |
1da177e4 LT |
3133 | { |
3134 | /* | |
3135 | * We do not migrate tasks that are: | |
3136 | * 1) running (obviously), or | |
3137 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
3138 | * 3) are cache-hot on their current CPU. | |
3139 | */ | |
cc367732 IM |
3140 | if (!cpu_isset(this_cpu, p->cpus_allowed)) { |
3141 | schedstat_inc(p, se.nr_failed_migrations_affine); | |
1da177e4 | 3142 | return 0; |
cc367732 | 3143 | } |
81026794 NP |
3144 | *all_pinned = 0; |
3145 | ||
cc367732 IM |
3146 | if (task_running(rq, p)) { |
3147 | schedstat_inc(p, se.nr_failed_migrations_running); | |
81026794 | 3148 | return 0; |
cc367732 | 3149 | } |
1da177e4 | 3150 | |
da84d961 IM |
3151 | /* |
3152 | * Aggressive migration if: | |
3153 | * 1) task is cache cold, or | |
3154 | * 2) too many balance attempts have failed. | |
3155 | */ | |
3156 | ||
6bc1665b IM |
3157 | if (!task_hot(p, rq->clock, sd) || |
3158 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
da84d961 | 3159 | #ifdef CONFIG_SCHEDSTATS |
cc367732 | 3160 | if (task_hot(p, rq->clock, sd)) { |
da84d961 | 3161 | schedstat_inc(sd, lb_hot_gained[idle]); |
cc367732 IM |
3162 | schedstat_inc(p, se.nr_forced_migrations); |
3163 | } | |
da84d961 IM |
3164 | #endif |
3165 | return 1; | |
3166 | } | |
3167 | ||
cc367732 IM |
3168 | if (task_hot(p, rq->clock, sd)) { |
3169 | schedstat_inc(p, se.nr_failed_migrations_hot); | |
da84d961 | 3170 | return 0; |
cc367732 | 3171 | } |
1da177e4 LT |
3172 | return 1; |
3173 | } | |
3174 | ||
e1d1484f PW |
3175 | static unsigned long |
3176 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3177 | unsigned long max_load_move, struct sched_domain *sd, | |
3178 | enum cpu_idle_type idle, int *all_pinned, | |
3179 | int *this_best_prio, struct rq_iterator *iterator) | |
1da177e4 | 3180 | { |
b82d9fdd | 3181 | int loops = 0, pulled = 0, pinned = 0, skip_for_load; |
dd41f596 IM |
3182 | struct task_struct *p; |
3183 | long rem_load_move = max_load_move; | |
1da177e4 | 3184 | |
e1d1484f | 3185 | if (max_load_move == 0) |
1da177e4 LT |
3186 | goto out; |
3187 | ||
81026794 NP |
3188 | pinned = 1; |
3189 | ||
1da177e4 | 3190 | /* |
dd41f596 | 3191 | * Start the load-balancing iterator: |
1da177e4 | 3192 | */ |
dd41f596 IM |
3193 | p = iterator->start(iterator->arg); |
3194 | next: | |
b82d9fdd | 3195 | if (!p || loops++ > sysctl_sched_nr_migrate) |
1da177e4 | 3196 | goto out; |
50ddd969 | 3197 | /* |
b82d9fdd | 3198 | * To help distribute high priority tasks across CPUs we don't |
50ddd969 PW |
3199 | * skip a task if it will be the highest priority task (i.e. smallest |
3200 | * prio value) on its new queue regardless of its load weight | |
3201 | */ | |
dd41f596 IM |
3202 | skip_for_load = (p->se.load.weight >> 1) > rem_load_move + |
3203 | SCHED_LOAD_SCALE_FUZZ; | |
a4ac01c3 | 3204 | if ((skip_for_load && p->prio >= *this_best_prio) || |
dd41f596 | 3205 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { |
dd41f596 IM |
3206 | p = iterator->next(iterator->arg); |
3207 | goto next; | |
1da177e4 LT |
3208 | } |
3209 | ||
dd41f596 | 3210 | pull_task(busiest, p, this_rq, this_cpu); |
1da177e4 | 3211 | pulled++; |
dd41f596 | 3212 | rem_load_move -= p->se.load.weight; |
1da177e4 | 3213 | |
2dd73a4f | 3214 | /* |
b82d9fdd | 3215 | * We only want to steal up to the prescribed amount of weighted load. |
2dd73a4f | 3216 | */ |
e1d1484f | 3217 | if (rem_load_move > 0) { |
a4ac01c3 PW |
3218 | if (p->prio < *this_best_prio) |
3219 | *this_best_prio = p->prio; | |
dd41f596 IM |
3220 | p = iterator->next(iterator->arg); |
3221 | goto next; | |
1da177e4 LT |
3222 | } |
3223 | out: | |
3224 | /* | |
e1d1484f | 3225 | * Right now, this is one of only two places pull_task() is called, |
1da177e4 LT |
3226 | * so we can safely collect pull_task() stats here rather than |
3227 | * inside pull_task(). | |
3228 | */ | |
3229 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
3230 | |
3231 | if (all_pinned) | |
3232 | *all_pinned = pinned; | |
e1d1484f PW |
3233 | |
3234 | return max_load_move - rem_load_move; | |
1da177e4 LT |
3235 | } |
3236 | ||
dd41f596 | 3237 | /* |
43010659 PW |
3238 | * move_tasks tries to move up to max_load_move weighted load from busiest to |
3239 | * this_rq, as part of a balancing operation within domain "sd". | |
3240 | * Returns 1 if successful and 0 otherwise. | |
dd41f596 IM |
3241 | * |
3242 | * Called with both runqueues locked. | |
3243 | */ | |
3244 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
43010659 | 3245 | unsigned long max_load_move, |
dd41f596 IM |
3246 | struct sched_domain *sd, enum cpu_idle_type idle, |
3247 | int *all_pinned) | |
3248 | { | |
5522d5d5 | 3249 | const struct sched_class *class = sched_class_highest; |
43010659 | 3250 | unsigned long total_load_moved = 0; |
a4ac01c3 | 3251 | int this_best_prio = this_rq->curr->prio; |
dd41f596 IM |
3252 | |
3253 | do { | |
43010659 PW |
3254 | total_load_moved += |
3255 | class->load_balance(this_rq, this_cpu, busiest, | |
e1d1484f | 3256 | max_load_move - total_load_moved, |
a4ac01c3 | 3257 | sd, idle, all_pinned, &this_best_prio); |
dd41f596 | 3258 | class = class->next; |
43010659 | 3259 | } while (class && max_load_move > total_load_moved); |
dd41f596 | 3260 | |
43010659 PW |
3261 | return total_load_moved > 0; |
3262 | } | |
3263 | ||
e1d1484f PW |
3264 | static int |
3265 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3266 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3267 | struct rq_iterator *iterator) | |
3268 | { | |
3269 | struct task_struct *p = iterator->start(iterator->arg); | |
3270 | int pinned = 0; | |
3271 | ||
3272 | while (p) { | |
3273 | if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { | |
3274 | pull_task(busiest, p, this_rq, this_cpu); | |
3275 | /* | |
3276 | * Right now, this is only the second place pull_task() | |
3277 | * is called, so we can safely collect pull_task() | |
3278 | * stats here rather than inside pull_task(). | |
3279 | */ | |
3280 | schedstat_inc(sd, lb_gained[idle]); | |
3281 | ||
3282 | return 1; | |
3283 | } | |
3284 | p = iterator->next(iterator->arg); | |
3285 | } | |
3286 | ||
3287 | return 0; | |
3288 | } | |
3289 | ||
43010659 PW |
3290 | /* |
3291 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3292 | * part of active balancing operations within "domain". | |
3293 | * Returns 1 if successful and 0 otherwise. | |
3294 | * | |
3295 | * Called with both runqueues locked. | |
3296 | */ | |
3297 | static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3298 | struct sched_domain *sd, enum cpu_idle_type idle) | |
3299 | { | |
5522d5d5 | 3300 | const struct sched_class *class; |
43010659 PW |
3301 | |
3302 | for (class = sched_class_highest; class; class = class->next) | |
e1d1484f | 3303 | if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle)) |
43010659 PW |
3304 | return 1; |
3305 | ||
3306 | return 0; | |
dd41f596 IM |
3307 | } |
3308 | ||
1da177e4 LT |
3309 | /* |
3310 | * find_busiest_group finds and returns the busiest CPU group within the | |
48f24c4d IM |
3311 | * domain. It calculates and returns the amount of weighted load which |
3312 | * should be moved to restore balance via the imbalance parameter. | |
1da177e4 LT |
3313 | */ |
3314 | static struct sched_group * | |
3315 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
dd41f596 | 3316 | unsigned long *imbalance, enum cpu_idle_type idle, |
7c16ec58 | 3317 | int *sd_idle, const cpumask_t *cpus, int *balance) |
1da177e4 LT |
3318 | { |
3319 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
3320 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 3321 | unsigned long max_pull; |
2dd73a4f PW |
3322 | unsigned long busiest_load_per_task, busiest_nr_running; |
3323 | unsigned long this_load_per_task, this_nr_running; | |
908a7c1b | 3324 | int load_idx, group_imb = 0; |
5c45bf27 SS |
3325 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
3326 | int power_savings_balance = 1; | |
3327 | unsigned long leader_nr_running = 0, min_load_per_task = 0; | |
3328 | unsigned long min_nr_running = ULONG_MAX; | |
3329 | struct sched_group *group_min = NULL, *group_leader = NULL; | |
3330 | #endif | |
1da177e4 LT |
3331 | |
3332 | max_load = this_load = total_load = total_pwr = 0; | |
2dd73a4f PW |
3333 | busiest_load_per_task = busiest_nr_running = 0; |
3334 | this_load_per_task = this_nr_running = 0; | |
d15bcfdb | 3335 | if (idle == CPU_NOT_IDLE) |
7897986b | 3336 | load_idx = sd->busy_idx; |
d15bcfdb | 3337 | else if (idle == CPU_NEWLY_IDLE) |
7897986b NP |
3338 | load_idx = sd->newidle_idx; |
3339 | else | |
3340 | load_idx = sd->idle_idx; | |
1da177e4 LT |
3341 | |
3342 | do { | |
908a7c1b | 3343 | unsigned long load, group_capacity, max_cpu_load, min_cpu_load; |
1da177e4 LT |
3344 | int local_group; |
3345 | int i; | |
908a7c1b | 3346 | int __group_imb = 0; |
783609c6 | 3347 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
2dd73a4f | 3348 | unsigned long sum_nr_running, sum_weighted_load; |
1da177e4 LT |
3349 | |
3350 | local_group = cpu_isset(this_cpu, group->cpumask); | |
3351 | ||
783609c6 SS |
3352 | if (local_group) |
3353 | balance_cpu = first_cpu(group->cpumask); | |
3354 | ||
1da177e4 | 3355 | /* Tally up the load of all CPUs in the group */ |
2dd73a4f | 3356 | sum_weighted_load = sum_nr_running = avg_load = 0; |
908a7c1b KC |
3357 | max_cpu_load = 0; |
3358 | min_cpu_load = ~0UL; | |
1da177e4 LT |
3359 | |
3360 | for_each_cpu_mask(i, group->cpumask) { | |
0a2966b4 CL |
3361 | struct rq *rq; |
3362 | ||
3363 | if (!cpu_isset(i, *cpus)) | |
3364 | continue; | |
3365 | ||
3366 | rq = cpu_rq(i); | |
2dd73a4f | 3367 | |
9439aab8 | 3368 | if (*sd_idle && rq->nr_running) |
5969fe06 NP |
3369 | *sd_idle = 0; |
3370 | ||
1da177e4 | 3371 | /* Bias balancing toward cpus of our domain */ |
783609c6 SS |
3372 | if (local_group) { |
3373 | if (idle_cpu(i) && !first_idle_cpu) { | |
3374 | first_idle_cpu = 1; | |
3375 | balance_cpu = i; | |
3376 | } | |
3377 | ||
a2000572 | 3378 | load = target_load(i, load_idx); |
908a7c1b | 3379 | } else { |
a2000572 | 3380 | load = source_load(i, load_idx); |
908a7c1b KC |
3381 | if (load > max_cpu_load) |
3382 | max_cpu_load = load; | |
3383 | if (min_cpu_load > load) | |
3384 | min_cpu_load = load; | |
3385 | } | |
1da177e4 LT |
3386 | |
3387 | avg_load += load; | |
2dd73a4f | 3388 | sum_nr_running += rq->nr_running; |
dd41f596 | 3389 | sum_weighted_load += weighted_cpuload(i); |
1da177e4 LT |
3390 | } |
3391 | ||
783609c6 SS |
3392 | /* |
3393 | * First idle cpu or the first cpu(busiest) in this sched group | |
3394 | * is eligible for doing load balancing at this and above | |
9439aab8 SS |
3395 | * domains. In the newly idle case, we will allow all the cpu's |
3396 | * to do the newly idle load balance. | |
783609c6 | 3397 | */ |
9439aab8 SS |
3398 | if (idle != CPU_NEWLY_IDLE && local_group && |
3399 | balance_cpu != this_cpu && balance) { | |
783609c6 SS |
3400 | *balance = 0; |
3401 | goto ret; | |
3402 | } | |
3403 | ||
1da177e4 | 3404 | total_load += avg_load; |
5517d86b | 3405 | total_pwr += group->__cpu_power; |
1da177e4 LT |
3406 | |
3407 | /* Adjust by relative CPU power of the group */ | |
5517d86b ED |
3408 | avg_load = sg_div_cpu_power(group, |
3409 | avg_load * SCHED_LOAD_SCALE); | |
1da177e4 | 3410 | |
908a7c1b KC |
3411 | if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE) |
3412 | __group_imb = 1; | |
3413 | ||
5517d86b | 3414 | group_capacity = group->__cpu_power / SCHED_LOAD_SCALE; |
5c45bf27 | 3415 | |
1da177e4 LT |
3416 | if (local_group) { |
3417 | this_load = avg_load; | |
3418 | this = group; | |
2dd73a4f PW |
3419 | this_nr_running = sum_nr_running; |
3420 | this_load_per_task = sum_weighted_load; | |
3421 | } else if (avg_load > max_load && | |
908a7c1b | 3422 | (sum_nr_running > group_capacity || __group_imb)) { |
1da177e4 LT |
3423 | max_load = avg_load; |
3424 | busiest = group; | |
2dd73a4f PW |
3425 | busiest_nr_running = sum_nr_running; |
3426 | busiest_load_per_task = sum_weighted_load; | |
908a7c1b | 3427 | group_imb = __group_imb; |
1da177e4 | 3428 | } |
5c45bf27 SS |
3429 | |
3430 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3431 | /* | |
3432 | * Busy processors will not participate in power savings | |
3433 | * balance. | |
3434 | */ | |
dd41f596 IM |
3435 | if (idle == CPU_NOT_IDLE || |
3436 | !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
3437 | goto group_next; | |
5c45bf27 SS |
3438 | |
3439 | /* | |
3440 | * If the local group is idle or completely loaded | |
3441 | * no need to do power savings balance at this domain | |
3442 | */ | |
3443 | if (local_group && (this_nr_running >= group_capacity || | |
3444 | !this_nr_running)) | |
3445 | power_savings_balance = 0; | |
3446 | ||
dd41f596 | 3447 | /* |
5c45bf27 SS |
3448 | * If a group is already running at full capacity or idle, |
3449 | * don't include that group in power savings calculations | |
dd41f596 IM |
3450 | */ |
3451 | if (!power_savings_balance || sum_nr_running >= group_capacity | |
5c45bf27 | 3452 | || !sum_nr_running) |
dd41f596 | 3453 | goto group_next; |
5c45bf27 | 3454 | |
dd41f596 | 3455 | /* |
5c45bf27 | 3456 | * Calculate the group which has the least non-idle load. |
dd41f596 IM |
3457 | * This is the group from where we need to pick up the load |
3458 | * for saving power | |
3459 | */ | |
3460 | if ((sum_nr_running < min_nr_running) || | |
3461 | (sum_nr_running == min_nr_running && | |
5c45bf27 SS |
3462 | first_cpu(group->cpumask) < |
3463 | first_cpu(group_min->cpumask))) { | |
dd41f596 IM |
3464 | group_min = group; |
3465 | min_nr_running = sum_nr_running; | |
5c45bf27 SS |
3466 | min_load_per_task = sum_weighted_load / |
3467 | sum_nr_running; | |
dd41f596 | 3468 | } |
5c45bf27 | 3469 | |
dd41f596 | 3470 | /* |
5c45bf27 | 3471 | * Calculate the group which is almost near its |
dd41f596 IM |
3472 | * capacity but still has some space to pick up some load |
3473 | * from other group and save more power | |
3474 | */ | |
3475 | if (sum_nr_running <= group_capacity - 1) { | |
3476 | if (sum_nr_running > leader_nr_running || | |
3477 | (sum_nr_running == leader_nr_running && | |
3478 | first_cpu(group->cpumask) > | |
3479 | first_cpu(group_leader->cpumask))) { | |
3480 | group_leader = group; | |
3481 | leader_nr_running = sum_nr_running; | |
3482 | } | |
48f24c4d | 3483 | } |
5c45bf27 SS |
3484 | group_next: |
3485 | #endif | |
1da177e4 LT |
3486 | group = group->next; |
3487 | } while (group != sd->groups); | |
3488 | ||
2dd73a4f | 3489 | if (!busiest || this_load >= max_load || busiest_nr_running == 0) |
1da177e4 LT |
3490 | goto out_balanced; |
3491 | ||
3492 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
3493 | ||
3494 | if (this_load >= avg_load || | |
3495 | 100*max_load <= sd->imbalance_pct*this_load) | |
3496 | goto out_balanced; | |
3497 | ||
2dd73a4f | 3498 | busiest_load_per_task /= busiest_nr_running; |
908a7c1b KC |
3499 | if (group_imb) |
3500 | busiest_load_per_task = min(busiest_load_per_task, avg_load); | |
3501 | ||
1da177e4 LT |
3502 | /* |
3503 | * We're trying to get all the cpus to the average_load, so we don't | |
3504 | * want to push ourselves above the average load, nor do we wish to | |
3505 | * reduce the max loaded cpu below the average load, as either of these | |
3506 | * actions would just result in more rebalancing later, and ping-pong | |
3507 | * tasks around. Thus we look for the minimum possible imbalance. | |
3508 | * Negative imbalances (*we* are more loaded than anyone else) will | |
3509 | * be counted as no imbalance for these purposes -- we can't fix that | |
41a2d6cf | 3510 | * by pulling tasks to us. Be careful of negative numbers as they'll |
1da177e4 LT |
3511 | * appear as very large values with unsigned longs. |
3512 | */ | |
2dd73a4f PW |
3513 | if (max_load <= busiest_load_per_task) |
3514 | goto out_balanced; | |
3515 | ||
3516 | /* | |
3517 | * In the presence of smp nice balancing, certain scenarios can have | |
3518 | * max load less than avg load(as we skip the groups at or below | |
3519 | * its cpu_power, while calculating max_load..) | |
3520 | */ | |
3521 | if (max_load < avg_load) { | |
3522 | *imbalance = 0; | |
3523 | goto small_imbalance; | |
3524 | } | |
0c117f1b SS |
3525 | |
3526 | /* Don't want to pull so many tasks that a group would go idle */ | |
2dd73a4f | 3527 | max_pull = min(max_load - avg_load, max_load - busiest_load_per_task); |
0c117f1b | 3528 | |
1da177e4 | 3529 | /* How much load to actually move to equalise the imbalance */ |
5517d86b ED |
3530 | *imbalance = min(max_pull * busiest->__cpu_power, |
3531 | (avg_load - this_load) * this->__cpu_power) | |
1da177e4 LT |
3532 | / SCHED_LOAD_SCALE; |
3533 | ||
2dd73a4f PW |
3534 | /* |
3535 | * if *imbalance is less than the average load per runnable task | |
3536 | * there is no gaurantee that any tasks will be moved so we'll have | |
3537 | * a think about bumping its value to force at least one task to be | |
3538 | * moved | |
3539 | */ | |
7fd0d2dd | 3540 | if (*imbalance < busiest_load_per_task) { |
48f24c4d | 3541 | unsigned long tmp, pwr_now, pwr_move; |
2dd73a4f PW |
3542 | unsigned int imbn; |
3543 | ||
3544 | small_imbalance: | |
3545 | pwr_move = pwr_now = 0; | |
3546 | imbn = 2; | |
3547 | if (this_nr_running) { | |
3548 | this_load_per_task /= this_nr_running; | |
3549 | if (busiest_load_per_task > this_load_per_task) | |
3550 | imbn = 1; | |
3551 | } else | |
3552 | this_load_per_task = SCHED_LOAD_SCALE; | |
1da177e4 | 3553 | |
dd41f596 IM |
3554 | if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >= |
3555 | busiest_load_per_task * imbn) { | |
2dd73a4f | 3556 | *imbalance = busiest_load_per_task; |
1da177e4 LT |
3557 | return busiest; |
3558 | } | |
3559 | ||
3560 | /* | |
3561 | * OK, we don't have enough imbalance to justify moving tasks, | |
3562 | * however we may be able to increase total CPU power used by | |
3563 | * moving them. | |
3564 | */ | |
3565 | ||
5517d86b ED |
3566 | pwr_now += busiest->__cpu_power * |
3567 | min(busiest_load_per_task, max_load); | |
3568 | pwr_now += this->__cpu_power * | |
3569 | min(this_load_per_task, this_load); | |
1da177e4 LT |
3570 | pwr_now /= SCHED_LOAD_SCALE; |
3571 | ||
3572 | /* Amount of load we'd subtract */ | |
5517d86b ED |
3573 | tmp = sg_div_cpu_power(busiest, |
3574 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
1da177e4 | 3575 | if (max_load > tmp) |
5517d86b | 3576 | pwr_move += busiest->__cpu_power * |
2dd73a4f | 3577 | min(busiest_load_per_task, max_load - tmp); |
1da177e4 LT |
3578 | |
3579 | /* Amount of load we'd add */ | |
5517d86b | 3580 | if (max_load * busiest->__cpu_power < |
33859f7f | 3581 | busiest_load_per_task * SCHED_LOAD_SCALE) |
5517d86b ED |
3582 | tmp = sg_div_cpu_power(this, |
3583 | max_load * busiest->__cpu_power); | |
1da177e4 | 3584 | else |
5517d86b ED |
3585 | tmp = sg_div_cpu_power(this, |
3586 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
3587 | pwr_move += this->__cpu_power * | |
3588 | min(this_load_per_task, this_load + tmp); | |
1da177e4 LT |
3589 | pwr_move /= SCHED_LOAD_SCALE; |
3590 | ||
3591 | /* Move if we gain throughput */ | |
7fd0d2dd SS |
3592 | if (pwr_move > pwr_now) |
3593 | *imbalance = busiest_load_per_task; | |
1da177e4 LT |
3594 | } |
3595 | ||
1da177e4 LT |
3596 | return busiest; |
3597 | ||
3598 | out_balanced: | |
5c45bf27 | 3599 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
d15bcfdb | 3600 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) |
5c45bf27 | 3601 | goto ret; |
1da177e4 | 3602 | |
5c45bf27 SS |
3603 | if (this == group_leader && group_leader != group_min) { |
3604 | *imbalance = min_load_per_task; | |
3605 | return group_min; | |
3606 | } | |
5c45bf27 | 3607 | #endif |
783609c6 | 3608 | ret: |
1da177e4 LT |
3609 | *imbalance = 0; |
3610 | return NULL; | |
3611 | } | |
3612 | ||
3613 | /* | |
3614 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3615 | */ | |
70b97a7f | 3616 | static struct rq * |
d15bcfdb | 3617 | find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle, |
7c16ec58 | 3618 | unsigned long imbalance, const cpumask_t *cpus) |
1da177e4 | 3619 | { |
70b97a7f | 3620 | struct rq *busiest = NULL, *rq; |
2dd73a4f | 3621 | unsigned long max_load = 0; |
1da177e4 LT |
3622 | int i; |
3623 | ||
3624 | for_each_cpu_mask(i, group->cpumask) { | |
dd41f596 | 3625 | unsigned long wl; |
0a2966b4 CL |
3626 | |
3627 | if (!cpu_isset(i, *cpus)) | |
3628 | continue; | |
3629 | ||
48f24c4d | 3630 | rq = cpu_rq(i); |
dd41f596 | 3631 | wl = weighted_cpuload(i); |
2dd73a4f | 3632 | |
dd41f596 | 3633 | if (rq->nr_running == 1 && wl > imbalance) |
2dd73a4f | 3634 | continue; |
1da177e4 | 3635 | |
dd41f596 IM |
3636 | if (wl > max_load) { |
3637 | max_load = wl; | |
48f24c4d | 3638 | busiest = rq; |
1da177e4 LT |
3639 | } |
3640 | } | |
3641 | ||
3642 | return busiest; | |
3643 | } | |
3644 | ||
77391d71 NP |
3645 | /* |
3646 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3647 | * so long as it is large enough. | |
3648 | */ | |
3649 | #define MAX_PINNED_INTERVAL 512 | |
3650 | ||
1da177e4 LT |
3651 | /* |
3652 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3653 | * tasks if there is an imbalance. | |
1da177e4 | 3654 | */ |
70b97a7f | 3655 | static int load_balance(int this_cpu, struct rq *this_rq, |
d15bcfdb | 3656 | struct sched_domain *sd, enum cpu_idle_type idle, |
7c16ec58 | 3657 | int *balance, cpumask_t *cpus) |
1da177e4 | 3658 | { |
43010659 | 3659 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; |
1da177e4 | 3660 | struct sched_group *group; |
1da177e4 | 3661 | unsigned long imbalance; |
70b97a7f | 3662 | struct rq *busiest; |
fe2eea3f | 3663 | unsigned long flags; |
18d95a28 | 3664 | int unlock_aggregate; |
5969fe06 | 3665 | |
7c16ec58 MT |
3666 | cpus_setall(*cpus); |
3667 | ||
18d95a28 PZ |
3668 | unlock_aggregate = get_aggregate(sd); |
3669 | ||
89c4710e SS |
3670 | /* |
3671 | * When power savings policy is enabled for the parent domain, idle | |
3672 | * sibling can pick up load irrespective of busy siblings. In this case, | |
dd41f596 | 3673 | * let the state of idle sibling percolate up as CPU_IDLE, instead of |
d15bcfdb | 3674 | * portraying it as CPU_NOT_IDLE. |
89c4710e | 3675 | */ |
d15bcfdb | 3676 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3677 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 3678 | sd_idle = 1; |
1da177e4 | 3679 | |
2d72376b | 3680 | schedstat_inc(sd, lb_count[idle]); |
1da177e4 | 3681 | |
0a2966b4 CL |
3682 | redo: |
3683 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
7c16ec58 | 3684 | cpus, balance); |
783609c6 | 3685 | |
06066714 | 3686 | if (*balance == 0) |
783609c6 | 3687 | goto out_balanced; |
783609c6 | 3688 | |
1da177e4 LT |
3689 | if (!group) { |
3690 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3691 | goto out_balanced; | |
3692 | } | |
3693 | ||
7c16ec58 | 3694 | busiest = find_busiest_queue(group, idle, imbalance, cpus); |
1da177e4 LT |
3695 | if (!busiest) { |
3696 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3697 | goto out_balanced; | |
3698 | } | |
3699 | ||
db935dbd | 3700 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
3701 | |
3702 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3703 | ||
43010659 | 3704 | ld_moved = 0; |
1da177e4 LT |
3705 | if (busiest->nr_running > 1) { |
3706 | /* | |
3707 | * Attempt to move tasks. If find_busiest_group has found | |
3708 | * an imbalance but busiest->nr_running <= 1, the group is | |
43010659 | 3709 | * still unbalanced. ld_moved simply stays zero, so it is |
1da177e4 LT |
3710 | * correctly treated as an imbalance. |
3711 | */ | |
fe2eea3f | 3712 | local_irq_save(flags); |
e17224bf | 3713 | double_rq_lock(this_rq, busiest); |
43010659 | 3714 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
48f24c4d | 3715 | imbalance, sd, idle, &all_pinned); |
e17224bf | 3716 | double_rq_unlock(this_rq, busiest); |
fe2eea3f | 3717 | local_irq_restore(flags); |
81026794 | 3718 | |
46cb4b7c SS |
3719 | /* |
3720 | * some other cpu did the load balance for us. | |
3721 | */ | |
43010659 | 3722 | if (ld_moved && this_cpu != smp_processor_id()) |
46cb4b7c SS |
3723 | resched_cpu(this_cpu); |
3724 | ||
81026794 | 3725 | /* All tasks on this runqueue were pinned by CPU affinity */ |
0a2966b4 | 3726 | if (unlikely(all_pinned)) { |
7c16ec58 MT |
3727 | cpu_clear(cpu_of(busiest), *cpus); |
3728 | if (!cpus_empty(*cpus)) | |
0a2966b4 | 3729 | goto redo; |
81026794 | 3730 | goto out_balanced; |
0a2966b4 | 3731 | } |
1da177e4 | 3732 | } |
81026794 | 3733 | |
43010659 | 3734 | if (!ld_moved) { |
1da177e4 LT |
3735 | schedstat_inc(sd, lb_failed[idle]); |
3736 | sd->nr_balance_failed++; | |
3737 | ||
3738 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 | 3739 | |
fe2eea3f | 3740 | spin_lock_irqsave(&busiest->lock, flags); |
fa3b6ddc SS |
3741 | |
3742 | /* don't kick the migration_thread, if the curr | |
3743 | * task on busiest cpu can't be moved to this_cpu | |
3744 | */ | |
3745 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
fe2eea3f | 3746 | spin_unlock_irqrestore(&busiest->lock, flags); |
fa3b6ddc SS |
3747 | all_pinned = 1; |
3748 | goto out_one_pinned; | |
3749 | } | |
3750 | ||
1da177e4 LT |
3751 | if (!busiest->active_balance) { |
3752 | busiest->active_balance = 1; | |
3753 | busiest->push_cpu = this_cpu; | |
81026794 | 3754 | active_balance = 1; |
1da177e4 | 3755 | } |
fe2eea3f | 3756 | spin_unlock_irqrestore(&busiest->lock, flags); |
81026794 | 3757 | if (active_balance) |
1da177e4 LT |
3758 | wake_up_process(busiest->migration_thread); |
3759 | ||
3760 | /* | |
3761 | * We've kicked active balancing, reset the failure | |
3762 | * counter. | |
3763 | */ | |
39507451 | 3764 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 3765 | } |
81026794 | 3766 | } else |
1da177e4 LT |
3767 | sd->nr_balance_failed = 0; |
3768 | ||
81026794 | 3769 | if (likely(!active_balance)) { |
1da177e4 LT |
3770 | /* We were unbalanced, so reset the balancing interval */ |
3771 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
3772 | } else { |
3773 | /* | |
3774 | * If we've begun active balancing, start to back off. This | |
3775 | * case may not be covered by the all_pinned logic if there | |
3776 | * is only 1 task on the busy runqueue (because we don't call | |
3777 | * move_tasks). | |
3778 | */ | |
3779 | if (sd->balance_interval < sd->max_interval) | |
3780 | sd->balance_interval *= 2; | |
1da177e4 LT |
3781 | } |
3782 | ||
43010659 | 3783 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3784 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
18d95a28 PZ |
3785 | ld_moved = -1; |
3786 | ||
3787 | goto out; | |
1da177e4 LT |
3788 | |
3789 | out_balanced: | |
1da177e4 LT |
3790 | schedstat_inc(sd, lb_balanced[idle]); |
3791 | ||
16cfb1c0 | 3792 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
3793 | |
3794 | out_one_pinned: | |
1da177e4 | 3795 | /* tune up the balancing interval */ |
77391d71 NP |
3796 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
3797 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
3798 | sd->balance_interval *= 2; |
3799 | ||
48f24c4d | 3800 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3801 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
18d95a28 PZ |
3802 | ld_moved = -1; |
3803 | else | |
3804 | ld_moved = 0; | |
3805 | out: | |
3806 | if (unlock_aggregate) | |
3807 | put_aggregate(sd); | |
3808 | return ld_moved; | |
1da177e4 LT |
3809 | } |
3810 | ||
3811 | /* | |
3812 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3813 | * tasks if there is an imbalance. | |
3814 | * | |
d15bcfdb | 3815 | * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE). |
1da177e4 LT |
3816 | * this_rq is locked. |
3817 | */ | |
48f24c4d | 3818 | static int |
7c16ec58 MT |
3819 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd, |
3820 | cpumask_t *cpus) | |
1da177e4 LT |
3821 | { |
3822 | struct sched_group *group; | |
70b97a7f | 3823 | struct rq *busiest = NULL; |
1da177e4 | 3824 | unsigned long imbalance; |
43010659 | 3825 | int ld_moved = 0; |
5969fe06 | 3826 | int sd_idle = 0; |
969bb4e4 | 3827 | int all_pinned = 0; |
7c16ec58 MT |
3828 | |
3829 | cpus_setall(*cpus); | |
5969fe06 | 3830 | |
89c4710e SS |
3831 | /* |
3832 | * When power savings policy is enabled for the parent domain, idle | |
3833 | * sibling can pick up load irrespective of busy siblings. In this case, | |
3834 | * let the state of idle sibling percolate up as IDLE, instead of | |
d15bcfdb | 3835 | * portraying it as CPU_NOT_IDLE. |
89c4710e SS |
3836 | */ |
3837 | if (sd->flags & SD_SHARE_CPUPOWER && | |
3838 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 | 3839 | sd_idle = 1; |
1da177e4 | 3840 | |
2d72376b | 3841 | schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]); |
0a2966b4 | 3842 | redo: |
d15bcfdb | 3843 | group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE, |
7c16ec58 | 3844 | &sd_idle, cpus, NULL); |
1da177e4 | 3845 | if (!group) { |
d15bcfdb | 3846 | schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]); |
16cfb1c0 | 3847 | goto out_balanced; |
1da177e4 LT |
3848 | } |
3849 | ||
7c16ec58 | 3850 | busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus); |
db935dbd | 3851 | if (!busiest) { |
d15bcfdb | 3852 | schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]); |
16cfb1c0 | 3853 | goto out_balanced; |
1da177e4 LT |
3854 | } |
3855 | ||
db935dbd NP |
3856 | BUG_ON(busiest == this_rq); |
3857 | ||
d15bcfdb | 3858 | schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance); |
d6d5cfaf | 3859 | |
43010659 | 3860 | ld_moved = 0; |
d6d5cfaf NP |
3861 | if (busiest->nr_running > 1) { |
3862 | /* Attempt to move tasks */ | |
3863 | double_lock_balance(this_rq, busiest); | |
6e82a3be IM |
3864 | /* this_rq->clock is already updated */ |
3865 | update_rq_clock(busiest); | |
43010659 | 3866 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
969bb4e4 SS |
3867 | imbalance, sd, CPU_NEWLY_IDLE, |
3868 | &all_pinned); | |
d6d5cfaf | 3869 | spin_unlock(&busiest->lock); |
0a2966b4 | 3870 | |
969bb4e4 | 3871 | if (unlikely(all_pinned)) { |
7c16ec58 MT |
3872 | cpu_clear(cpu_of(busiest), *cpus); |
3873 | if (!cpus_empty(*cpus)) | |
0a2966b4 CL |
3874 | goto redo; |
3875 | } | |
d6d5cfaf NP |
3876 | } |
3877 | ||
43010659 | 3878 | if (!ld_moved) { |
d15bcfdb | 3879 | schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]); |
89c4710e SS |
3880 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
3881 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 NP |
3882 | return -1; |
3883 | } else | |
16cfb1c0 | 3884 | sd->nr_balance_failed = 0; |
1da177e4 | 3885 | |
43010659 | 3886 | return ld_moved; |
16cfb1c0 NP |
3887 | |
3888 | out_balanced: | |
d15bcfdb | 3889 | schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]); |
48f24c4d | 3890 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3891 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 3892 | return -1; |
16cfb1c0 | 3893 | sd->nr_balance_failed = 0; |
48f24c4d | 3894 | |
16cfb1c0 | 3895 | return 0; |
1da177e4 LT |
3896 | } |
3897 | ||
3898 | /* | |
3899 | * idle_balance is called by schedule() if this_cpu is about to become | |
3900 | * idle. Attempts to pull tasks from other CPUs. | |
3901 | */ | |
70b97a7f | 3902 | static void idle_balance(int this_cpu, struct rq *this_rq) |
1da177e4 LT |
3903 | { |
3904 | struct sched_domain *sd; | |
dd41f596 IM |
3905 | int pulled_task = -1; |
3906 | unsigned long next_balance = jiffies + HZ; | |
7c16ec58 | 3907 | cpumask_t tmpmask; |
1da177e4 LT |
3908 | |
3909 | for_each_domain(this_cpu, sd) { | |
92c4ca5c CL |
3910 | unsigned long interval; |
3911 | ||
3912 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3913 | continue; | |
3914 | ||
3915 | if (sd->flags & SD_BALANCE_NEWIDLE) | |
48f24c4d | 3916 | /* If we've pulled tasks over stop searching: */ |
7c16ec58 MT |
3917 | pulled_task = load_balance_newidle(this_cpu, this_rq, |
3918 | sd, &tmpmask); | |
92c4ca5c CL |
3919 | |
3920 | interval = msecs_to_jiffies(sd->balance_interval); | |
3921 | if (time_after(next_balance, sd->last_balance + interval)) | |
3922 | next_balance = sd->last_balance + interval; | |
3923 | if (pulled_task) | |
3924 | break; | |
1da177e4 | 3925 | } |
dd41f596 | 3926 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
1bd77f2d CL |
3927 | /* |
3928 | * We are going idle. next_balance may be set based on | |
3929 | * a busy processor. So reset next_balance. | |
3930 | */ | |
3931 | this_rq->next_balance = next_balance; | |
dd41f596 | 3932 | } |
1da177e4 LT |
3933 | } |
3934 | ||
3935 | /* | |
3936 | * active_load_balance is run by migration threads. It pushes running tasks | |
3937 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
3938 | * running on each physical CPU where possible, and avoids physical / | |
3939 | * logical imbalances. | |
3940 | * | |
3941 | * Called with busiest_rq locked. | |
3942 | */ | |
70b97a7f | 3943 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) |
1da177e4 | 3944 | { |
39507451 | 3945 | int target_cpu = busiest_rq->push_cpu; |
70b97a7f IM |
3946 | struct sched_domain *sd; |
3947 | struct rq *target_rq; | |
39507451 | 3948 | |
48f24c4d | 3949 | /* Is there any task to move? */ |
39507451 | 3950 | if (busiest_rq->nr_running <= 1) |
39507451 NP |
3951 | return; |
3952 | ||
3953 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
3954 | |
3955 | /* | |
39507451 | 3956 | * This condition is "impossible", if it occurs |
41a2d6cf | 3957 | * we need to fix it. Originally reported by |
39507451 | 3958 | * Bjorn Helgaas on a 128-cpu setup. |
1da177e4 | 3959 | */ |
39507451 | 3960 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 3961 | |
39507451 NP |
3962 | /* move a task from busiest_rq to target_rq */ |
3963 | double_lock_balance(busiest_rq, target_rq); | |
6e82a3be IM |
3964 | update_rq_clock(busiest_rq); |
3965 | update_rq_clock(target_rq); | |
39507451 NP |
3966 | |
3967 | /* Search for an sd spanning us and the target CPU. */ | |
c96d145e | 3968 | for_each_domain(target_cpu, sd) { |
39507451 | 3969 | if ((sd->flags & SD_LOAD_BALANCE) && |
48f24c4d | 3970 | cpu_isset(busiest_cpu, sd->span)) |
39507451 | 3971 | break; |
c96d145e | 3972 | } |
39507451 | 3973 | |
48f24c4d | 3974 | if (likely(sd)) { |
2d72376b | 3975 | schedstat_inc(sd, alb_count); |
39507451 | 3976 | |
43010659 PW |
3977 | if (move_one_task(target_rq, target_cpu, busiest_rq, |
3978 | sd, CPU_IDLE)) | |
48f24c4d IM |
3979 | schedstat_inc(sd, alb_pushed); |
3980 | else | |
3981 | schedstat_inc(sd, alb_failed); | |
3982 | } | |
39507451 | 3983 | spin_unlock(&target_rq->lock); |
1da177e4 LT |
3984 | } |
3985 | ||
46cb4b7c SS |
3986 | #ifdef CONFIG_NO_HZ |
3987 | static struct { | |
3988 | atomic_t load_balancer; | |
41a2d6cf | 3989 | cpumask_t cpu_mask; |
46cb4b7c SS |
3990 | } nohz ____cacheline_aligned = { |
3991 | .load_balancer = ATOMIC_INIT(-1), | |
3992 | .cpu_mask = CPU_MASK_NONE, | |
3993 | }; | |
3994 | ||
7835b98b | 3995 | /* |
46cb4b7c SS |
3996 | * This routine will try to nominate the ilb (idle load balancing) |
3997 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
3998 | * load balancing on behalf of all those cpus. If all the cpus in the system | |
3999 | * go into this tickless mode, then there will be no ilb owner (as there is | |
4000 | * no need for one) and all the cpus will sleep till the next wakeup event | |
4001 | * arrives... | |
4002 | * | |
4003 | * For the ilb owner, tick is not stopped. And this tick will be used | |
4004 | * for idle load balancing. ilb owner will still be part of | |
4005 | * nohz.cpu_mask.. | |
7835b98b | 4006 | * |
46cb4b7c SS |
4007 | * While stopping the tick, this cpu will become the ilb owner if there |
4008 | * is no other owner. And will be the owner till that cpu becomes busy | |
4009 | * or if all cpus in the system stop their ticks at which point | |
4010 | * there is no need for ilb owner. | |
4011 | * | |
4012 | * When the ilb owner becomes busy, it nominates another owner, during the | |
4013 | * next busy scheduler_tick() | |
4014 | */ | |
4015 | int select_nohz_load_balancer(int stop_tick) | |
4016 | { | |
4017 | int cpu = smp_processor_id(); | |
4018 | ||
4019 | if (stop_tick) { | |
4020 | cpu_set(cpu, nohz.cpu_mask); | |
4021 | cpu_rq(cpu)->in_nohz_recently = 1; | |
4022 | ||
4023 | /* | |
4024 | * If we are going offline and still the leader, give up! | |
4025 | */ | |
4026 | if (cpu_is_offline(cpu) && | |
4027 | atomic_read(&nohz.load_balancer) == cpu) { | |
4028 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
4029 | BUG(); | |
4030 | return 0; | |
4031 | } | |
4032 | ||
4033 | /* time for ilb owner also to sleep */ | |
4034 | if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
4035 | if (atomic_read(&nohz.load_balancer) == cpu) | |
4036 | atomic_set(&nohz.load_balancer, -1); | |
4037 | return 0; | |
4038 | } | |
4039 | ||
4040 | if (atomic_read(&nohz.load_balancer) == -1) { | |
4041 | /* make me the ilb owner */ | |
4042 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) | |
4043 | return 1; | |
4044 | } else if (atomic_read(&nohz.load_balancer) == cpu) | |
4045 | return 1; | |
4046 | } else { | |
4047 | if (!cpu_isset(cpu, nohz.cpu_mask)) | |
4048 | return 0; | |
4049 | ||
4050 | cpu_clear(cpu, nohz.cpu_mask); | |
4051 | ||
4052 | if (atomic_read(&nohz.load_balancer) == cpu) | |
4053 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
4054 | BUG(); | |
4055 | } | |
4056 | return 0; | |
4057 | } | |
4058 | #endif | |
4059 | ||
4060 | static DEFINE_SPINLOCK(balancing); | |
4061 | ||
4062 | /* | |
7835b98b CL |
4063 | * It checks each scheduling domain to see if it is due to be balanced, |
4064 | * and initiates a balancing operation if so. | |
4065 | * | |
4066 | * Balancing parameters are set up in arch_init_sched_domains. | |
4067 | */ | |
a9957449 | 4068 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) |
7835b98b | 4069 | { |
46cb4b7c SS |
4070 | int balance = 1; |
4071 | struct rq *rq = cpu_rq(cpu); | |
7835b98b CL |
4072 | unsigned long interval; |
4073 | struct sched_domain *sd; | |
46cb4b7c | 4074 | /* Earliest time when we have to do rebalance again */ |
c9819f45 | 4075 | unsigned long next_balance = jiffies + 60*HZ; |
f549da84 | 4076 | int update_next_balance = 0; |
7c16ec58 | 4077 | cpumask_t tmp; |
1da177e4 | 4078 | |
46cb4b7c | 4079 | for_each_domain(cpu, sd) { |
1da177e4 LT |
4080 | if (!(sd->flags & SD_LOAD_BALANCE)) |
4081 | continue; | |
4082 | ||
4083 | interval = sd->balance_interval; | |
d15bcfdb | 4084 | if (idle != CPU_IDLE) |
1da177e4 LT |
4085 | interval *= sd->busy_factor; |
4086 | ||
4087 | /* scale ms to jiffies */ | |
4088 | interval = msecs_to_jiffies(interval); | |
4089 | if (unlikely(!interval)) | |
4090 | interval = 1; | |
dd41f596 IM |
4091 | if (interval > HZ*NR_CPUS/10) |
4092 | interval = HZ*NR_CPUS/10; | |
4093 | ||
1da177e4 | 4094 | |
08c183f3 CL |
4095 | if (sd->flags & SD_SERIALIZE) { |
4096 | if (!spin_trylock(&balancing)) | |
4097 | goto out; | |
4098 | } | |
4099 | ||
c9819f45 | 4100 | if (time_after_eq(jiffies, sd->last_balance + interval)) { |
7c16ec58 | 4101 | if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) { |
fa3b6ddc SS |
4102 | /* |
4103 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
4104 | * longer idle, or one of our SMT siblings is |
4105 | * not idle. | |
4106 | */ | |
d15bcfdb | 4107 | idle = CPU_NOT_IDLE; |
1da177e4 | 4108 | } |
1bd77f2d | 4109 | sd->last_balance = jiffies; |
1da177e4 | 4110 | } |
08c183f3 CL |
4111 | if (sd->flags & SD_SERIALIZE) |
4112 | spin_unlock(&balancing); | |
4113 | out: | |
f549da84 | 4114 | if (time_after(next_balance, sd->last_balance + interval)) { |
c9819f45 | 4115 | next_balance = sd->last_balance + interval; |
f549da84 SS |
4116 | update_next_balance = 1; |
4117 | } | |
783609c6 SS |
4118 | |
4119 | /* | |
4120 | * Stop the load balance at this level. There is another | |
4121 | * CPU in our sched group which is doing load balancing more | |
4122 | * actively. | |
4123 | */ | |
4124 | if (!balance) | |
4125 | break; | |
1da177e4 | 4126 | } |
f549da84 SS |
4127 | |
4128 | /* | |
4129 | * next_balance will be updated only when there is a need. | |
4130 | * When the cpu is attached to null domain for ex, it will not be | |
4131 | * updated. | |
4132 | */ | |
4133 | if (likely(update_next_balance)) | |
4134 | rq->next_balance = next_balance; | |
46cb4b7c SS |
4135 | } |
4136 | ||
4137 | /* | |
4138 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
4139 | * In CONFIG_NO_HZ case, the idle load balance owner will do the | |
4140 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
4141 | */ | |
4142 | static void run_rebalance_domains(struct softirq_action *h) | |
4143 | { | |
dd41f596 IM |
4144 | int this_cpu = smp_processor_id(); |
4145 | struct rq *this_rq = cpu_rq(this_cpu); | |
4146 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
4147 | CPU_IDLE : CPU_NOT_IDLE; | |
46cb4b7c | 4148 | |
dd41f596 | 4149 | rebalance_domains(this_cpu, idle); |
46cb4b7c SS |
4150 | |
4151 | #ifdef CONFIG_NO_HZ | |
4152 | /* | |
4153 | * If this cpu is the owner for idle load balancing, then do the | |
4154 | * balancing on behalf of the other idle cpus whose ticks are | |
4155 | * stopped. | |
4156 | */ | |
dd41f596 IM |
4157 | if (this_rq->idle_at_tick && |
4158 | atomic_read(&nohz.load_balancer) == this_cpu) { | |
46cb4b7c SS |
4159 | cpumask_t cpus = nohz.cpu_mask; |
4160 | struct rq *rq; | |
4161 | int balance_cpu; | |
4162 | ||
dd41f596 | 4163 | cpu_clear(this_cpu, cpus); |
46cb4b7c SS |
4164 | for_each_cpu_mask(balance_cpu, cpus) { |
4165 | /* | |
4166 | * If this cpu gets work to do, stop the load balancing | |
4167 | * work being done for other cpus. Next load | |
4168 | * balancing owner will pick it up. | |
4169 | */ | |
4170 | if (need_resched()) | |
4171 | break; | |
4172 | ||
de0cf899 | 4173 | rebalance_domains(balance_cpu, CPU_IDLE); |
46cb4b7c SS |
4174 | |
4175 | rq = cpu_rq(balance_cpu); | |
dd41f596 IM |
4176 | if (time_after(this_rq->next_balance, rq->next_balance)) |
4177 | this_rq->next_balance = rq->next_balance; | |
46cb4b7c SS |
4178 | } |
4179 | } | |
4180 | #endif | |
4181 | } | |
4182 | ||
4183 | /* | |
4184 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
4185 | * | |
4186 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new | |
4187 | * idle load balancing owner or decide to stop the periodic load balancing, | |
4188 | * if the whole system is idle. | |
4189 | */ | |
dd41f596 | 4190 | static inline void trigger_load_balance(struct rq *rq, int cpu) |
46cb4b7c | 4191 | { |
46cb4b7c SS |
4192 | #ifdef CONFIG_NO_HZ |
4193 | /* | |
4194 | * If we were in the nohz mode recently and busy at the current | |
4195 | * scheduler tick, then check if we need to nominate new idle | |
4196 | * load balancer. | |
4197 | */ | |
4198 | if (rq->in_nohz_recently && !rq->idle_at_tick) { | |
4199 | rq->in_nohz_recently = 0; | |
4200 | ||
4201 | if (atomic_read(&nohz.load_balancer) == cpu) { | |
4202 | cpu_clear(cpu, nohz.cpu_mask); | |
4203 | atomic_set(&nohz.load_balancer, -1); | |
4204 | } | |
4205 | ||
4206 | if (atomic_read(&nohz.load_balancer) == -1) { | |
4207 | /* | |
4208 | * simple selection for now: Nominate the | |
4209 | * first cpu in the nohz list to be the next | |
4210 | * ilb owner. | |
4211 | * | |
4212 | * TBD: Traverse the sched domains and nominate | |
4213 | * the nearest cpu in the nohz.cpu_mask. | |
4214 | */ | |
4215 | int ilb = first_cpu(nohz.cpu_mask); | |
4216 | ||
434d53b0 | 4217 | if (ilb < nr_cpu_ids) |
46cb4b7c SS |
4218 | resched_cpu(ilb); |
4219 | } | |
4220 | } | |
4221 | ||
4222 | /* | |
4223 | * If this cpu is idle and doing idle load balancing for all the | |
4224 | * cpus with ticks stopped, is it time for that to stop? | |
4225 | */ | |
4226 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && | |
4227 | cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
4228 | resched_cpu(cpu); | |
4229 | return; | |
4230 | } | |
4231 | ||
4232 | /* | |
4233 | * If this cpu is idle and the idle load balancing is done by | |
4234 | * someone else, then no need raise the SCHED_SOFTIRQ | |
4235 | */ | |
4236 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && | |
4237 | cpu_isset(cpu, nohz.cpu_mask)) | |
4238 | return; | |
4239 | #endif | |
4240 | if (time_after_eq(jiffies, rq->next_balance)) | |
4241 | raise_softirq(SCHED_SOFTIRQ); | |
1da177e4 | 4242 | } |
dd41f596 IM |
4243 | |
4244 | #else /* CONFIG_SMP */ | |
4245 | ||
1da177e4 LT |
4246 | /* |
4247 | * on UP we do not need to balance between CPUs: | |
4248 | */ | |
70b97a7f | 4249 | static inline void idle_balance(int cpu, struct rq *rq) |
1da177e4 LT |
4250 | { |
4251 | } | |
dd41f596 | 4252 | |
1da177e4 LT |
4253 | #endif |
4254 | ||
1da177e4 LT |
4255 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
4256 | ||
4257 | EXPORT_PER_CPU_SYMBOL(kstat); | |
4258 | ||
4259 | /* | |
41b86e9c IM |
4260 | * Return p->sum_exec_runtime plus any more ns on the sched_clock |
4261 | * that have not yet been banked in case the task is currently running. | |
1da177e4 | 4262 | */ |
41b86e9c | 4263 | unsigned long long task_sched_runtime(struct task_struct *p) |
1da177e4 | 4264 | { |
1da177e4 | 4265 | unsigned long flags; |
41b86e9c IM |
4266 | u64 ns, delta_exec; |
4267 | struct rq *rq; | |
48f24c4d | 4268 | |
41b86e9c IM |
4269 | rq = task_rq_lock(p, &flags); |
4270 | ns = p->se.sum_exec_runtime; | |
051a1d1a | 4271 | if (task_current(rq, p)) { |
a8e504d2 IM |
4272 | update_rq_clock(rq); |
4273 | delta_exec = rq->clock - p->se.exec_start; | |
41b86e9c IM |
4274 | if ((s64)delta_exec > 0) |
4275 | ns += delta_exec; | |
4276 | } | |
4277 | task_rq_unlock(rq, &flags); | |
48f24c4d | 4278 | |
1da177e4 LT |
4279 | return ns; |
4280 | } | |
4281 | ||
1da177e4 LT |
4282 | /* |
4283 | * Account user cpu time to a process. | |
4284 | * @p: the process that the cpu time gets accounted to | |
1da177e4 LT |
4285 | * @cputime: the cpu time spent in user space since the last update |
4286 | */ | |
4287 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
4288 | { | |
4289 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4290 | cputime64_t tmp; | |
4291 | ||
4292 | p->utime = cputime_add(p->utime, cputime); | |
4293 | ||
4294 | /* Add user time to cpustat. */ | |
4295 | tmp = cputime_to_cputime64(cputime); | |
4296 | if (TASK_NICE(p) > 0) | |
4297 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
4298 | else | |
4299 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
4300 | } | |
4301 | ||
94886b84 LV |
4302 | /* |
4303 | * Account guest cpu time to a process. | |
4304 | * @p: the process that the cpu time gets accounted to | |
4305 | * @cputime: the cpu time spent in virtual machine since the last update | |
4306 | */ | |
f7402e03 | 4307 | static void account_guest_time(struct task_struct *p, cputime_t cputime) |
94886b84 LV |
4308 | { |
4309 | cputime64_t tmp; | |
4310 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4311 | ||
4312 | tmp = cputime_to_cputime64(cputime); | |
4313 | ||
4314 | p->utime = cputime_add(p->utime, cputime); | |
4315 | p->gtime = cputime_add(p->gtime, cputime); | |
4316 | ||
4317 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
4318 | cpustat->guest = cputime64_add(cpustat->guest, tmp); | |
4319 | } | |
4320 | ||
c66f08be MN |
4321 | /* |
4322 | * Account scaled user cpu time to a process. | |
4323 | * @p: the process that the cpu time gets accounted to | |
4324 | * @cputime: the cpu time spent in user space since the last update | |
4325 | */ | |
4326 | void account_user_time_scaled(struct task_struct *p, cputime_t cputime) | |
4327 | { | |
4328 | p->utimescaled = cputime_add(p->utimescaled, cputime); | |
4329 | } | |
4330 | ||
1da177e4 LT |
4331 | /* |
4332 | * Account system cpu time to a process. | |
4333 | * @p: the process that the cpu time gets accounted to | |
4334 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
4335 | * @cputime: the cpu time spent in kernel space since the last update | |
4336 | */ | |
4337 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
4338 | cputime_t cputime) | |
4339 | { | |
4340 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
70b97a7f | 4341 | struct rq *rq = this_rq(); |
1da177e4 LT |
4342 | cputime64_t tmp; |
4343 | ||
983ed7a6 HH |
4344 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { |
4345 | account_guest_time(p, cputime); | |
4346 | return; | |
4347 | } | |
94886b84 | 4348 | |
1da177e4 LT |
4349 | p->stime = cputime_add(p->stime, cputime); |
4350 | ||
4351 | /* Add system time to cpustat. */ | |
4352 | tmp = cputime_to_cputime64(cputime); | |
4353 | if (hardirq_count() - hardirq_offset) | |
4354 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
4355 | else if (softirq_count()) | |
4356 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
cfb52856 | 4357 | else if (p != rq->idle) |
1da177e4 | 4358 | cpustat->system = cputime64_add(cpustat->system, tmp); |
cfb52856 | 4359 | else if (atomic_read(&rq->nr_iowait) > 0) |
1da177e4 LT |
4360 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
4361 | else | |
4362 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
4363 | /* Account for system time used */ | |
4364 | acct_update_integrals(p); | |
1da177e4 LT |
4365 | } |
4366 | ||
c66f08be MN |
4367 | /* |
4368 | * Account scaled system cpu time to a process. | |
4369 | * @p: the process that the cpu time gets accounted to | |
4370 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
4371 | * @cputime: the cpu time spent in kernel space since the last update | |
4372 | */ | |
4373 | void account_system_time_scaled(struct task_struct *p, cputime_t cputime) | |
4374 | { | |
4375 | p->stimescaled = cputime_add(p->stimescaled, cputime); | |
4376 | } | |
4377 | ||
1da177e4 LT |
4378 | /* |
4379 | * Account for involuntary wait time. | |
4380 | * @p: the process from which the cpu time has been stolen | |
4381 | * @steal: the cpu time spent in involuntary wait | |
4382 | */ | |
4383 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
4384 | { | |
4385 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4386 | cputime64_t tmp = cputime_to_cputime64(steal); | |
70b97a7f | 4387 | struct rq *rq = this_rq(); |
1da177e4 LT |
4388 | |
4389 | if (p == rq->idle) { | |
4390 | p->stime = cputime_add(p->stime, steal); | |
4391 | if (atomic_read(&rq->nr_iowait) > 0) | |
4392 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
4393 | else | |
4394 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
cfb52856 | 4395 | } else |
1da177e4 LT |
4396 | cpustat->steal = cputime64_add(cpustat->steal, tmp); |
4397 | } | |
4398 | ||
7835b98b CL |
4399 | /* |
4400 | * This function gets called by the timer code, with HZ frequency. | |
4401 | * We call it with interrupts disabled. | |
4402 | * | |
4403 | * It also gets called by the fork code, when changing the parent's | |
4404 | * timeslices. | |
4405 | */ | |
4406 | void scheduler_tick(void) | |
4407 | { | |
7835b98b CL |
4408 | int cpu = smp_processor_id(); |
4409 | struct rq *rq = cpu_rq(cpu); | |
dd41f596 | 4410 | struct task_struct *curr = rq->curr; |
529c7726 | 4411 | u64 next_tick = rq->tick_timestamp + TICK_NSEC; |
dd41f596 IM |
4412 | |
4413 | spin_lock(&rq->lock); | |
546fe3c9 | 4414 | __update_rq_clock(rq); |
529c7726 IM |
4415 | /* |
4416 | * Let rq->clock advance by at least TICK_NSEC: | |
4417 | */ | |
cc203d24 | 4418 | if (unlikely(rq->clock < next_tick)) { |
529c7726 | 4419 | rq->clock = next_tick; |
cc203d24 GC |
4420 | rq->clock_underflows++; |
4421 | } | |
529c7726 | 4422 | rq->tick_timestamp = rq->clock; |
15934a37 | 4423 | update_last_tick_seen(rq); |
f1a438d8 | 4424 | update_cpu_load(rq); |
fa85ae24 | 4425 | curr->sched_class->task_tick(rq, curr, 0); |
dd41f596 | 4426 | spin_unlock(&rq->lock); |
7835b98b | 4427 | |
e418e1c2 | 4428 | #ifdef CONFIG_SMP |
dd41f596 IM |
4429 | rq->idle_at_tick = idle_cpu(cpu); |
4430 | trigger_load_balance(rq, cpu); | |
e418e1c2 | 4431 | #endif |
1da177e4 LT |
4432 | } |
4433 | ||
1da177e4 LT |
4434 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) |
4435 | ||
43627582 | 4436 | void __kprobes add_preempt_count(int val) |
1da177e4 LT |
4437 | { |
4438 | /* | |
4439 | * Underflow? | |
4440 | */ | |
9a11b49a IM |
4441 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
4442 | return; | |
1da177e4 LT |
4443 | preempt_count() += val; |
4444 | /* | |
4445 | * Spinlock count overflowing soon? | |
4446 | */ | |
33859f7f MOS |
4447 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
4448 | PREEMPT_MASK - 10); | |
1da177e4 LT |
4449 | } |
4450 | EXPORT_SYMBOL(add_preempt_count); | |
4451 | ||
43627582 | 4452 | void __kprobes sub_preempt_count(int val) |
1da177e4 LT |
4453 | { |
4454 | /* | |
4455 | * Underflow? | |
4456 | */ | |
9a11b49a IM |
4457 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
4458 | return; | |
1da177e4 LT |
4459 | /* |
4460 | * Is the spinlock portion underflowing? | |
4461 | */ | |
9a11b49a IM |
4462 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
4463 | !(preempt_count() & PREEMPT_MASK))) | |
4464 | return; | |
4465 | ||
1da177e4 LT |
4466 | preempt_count() -= val; |
4467 | } | |
4468 | EXPORT_SYMBOL(sub_preempt_count); | |
4469 | ||
4470 | #endif | |
4471 | ||
4472 | /* | |
dd41f596 | 4473 | * Print scheduling while atomic bug: |
1da177e4 | 4474 | */ |
dd41f596 | 4475 | static noinline void __schedule_bug(struct task_struct *prev) |
1da177e4 | 4476 | { |
838225b4 SS |
4477 | struct pt_regs *regs = get_irq_regs(); |
4478 | ||
4479 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", | |
4480 | prev->comm, prev->pid, preempt_count()); | |
4481 | ||
dd41f596 IM |
4482 | debug_show_held_locks(prev); |
4483 | if (irqs_disabled()) | |
4484 | print_irqtrace_events(prev); | |
838225b4 SS |
4485 | |
4486 | if (regs) | |
4487 | show_regs(regs); | |
4488 | else | |
4489 | dump_stack(); | |
dd41f596 | 4490 | } |
1da177e4 | 4491 | |
dd41f596 IM |
4492 | /* |
4493 | * Various schedule()-time debugging checks and statistics: | |
4494 | */ | |
4495 | static inline void schedule_debug(struct task_struct *prev) | |
4496 | { | |
1da177e4 | 4497 | /* |
41a2d6cf | 4498 | * Test if we are atomic. Since do_exit() needs to call into |
1da177e4 LT |
4499 | * schedule() atomically, we ignore that path for now. |
4500 | * Otherwise, whine if we are scheduling when we should not be. | |
4501 | */ | |
dd41f596 IM |
4502 | if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state)) |
4503 | __schedule_bug(prev); | |
4504 | ||
1da177e4 LT |
4505 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
4506 | ||
2d72376b | 4507 | schedstat_inc(this_rq(), sched_count); |
b8efb561 IM |
4508 | #ifdef CONFIG_SCHEDSTATS |
4509 | if (unlikely(prev->lock_depth >= 0)) { | |
2d72376b IM |
4510 | schedstat_inc(this_rq(), bkl_count); |
4511 | schedstat_inc(prev, sched_info.bkl_count); | |
b8efb561 IM |
4512 | } |
4513 | #endif | |
dd41f596 IM |
4514 | } |
4515 | ||
4516 | /* | |
4517 | * Pick up the highest-prio task: | |
4518 | */ | |
4519 | static inline struct task_struct * | |
ff95f3df | 4520 | pick_next_task(struct rq *rq, struct task_struct *prev) |
dd41f596 | 4521 | { |
5522d5d5 | 4522 | const struct sched_class *class; |
dd41f596 | 4523 | struct task_struct *p; |
1da177e4 LT |
4524 | |
4525 | /* | |
dd41f596 IM |
4526 | * Optimization: we know that if all tasks are in |
4527 | * the fair class we can call that function directly: | |
1da177e4 | 4528 | */ |
dd41f596 | 4529 | if (likely(rq->nr_running == rq->cfs.nr_running)) { |
fb8d4724 | 4530 | p = fair_sched_class.pick_next_task(rq); |
dd41f596 IM |
4531 | if (likely(p)) |
4532 | return p; | |
1da177e4 LT |
4533 | } |
4534 | ||
dd41f596 IM |
4535 | class = sched_class_highest; |
4536 | for ( ; ; ) { | |
fb8d4724 | 4537 | p = class->pick_next_task(rq); |
dd41f596 IM |
4538 | if (p) |
4539 | return p; | |
4540 | /* | |
4541 | * Will never be NULL as the idle class always | |
4542 | * returns a non-NULL p: | |
4543 | */ | |
4544 | class = class->next; | |
4545 | } | |
4546 | } | |
1da177e4 | 4547 | |
dd41f596 IM |
4548 | /* |
4549 | * schedule() is the main scheduler function. | |
4550 | */ | |
4551 | asmlinkage void __sched schedule(void) | |
4552 | { | |
4553 | struct task_struct *prev, *next; | |
67ca7bde | 4554 | unsigned long *switch_count; |
dd41f596 | 4555 | struct rq *rq; |
dd41f596 IM |
4556 | int cpu; |
4557 | ||
4558 | need_resched: | |
4559 | preempt_disable(); | |
4560 | cpu = smp_processor_id(); | |
4561 | rq = cpu_rq(cpu); | |
4562 | rcu_qsctr_inc(cpu); | |
4563 | prev = rq->curr; | |
4564 | switch_count = &prev->nivcsw; | |
4565 | ||
4566 | release_kernel_lock(prev); | |
4567 | need_resched_nonpreemptible: | |
4568 | ||
4569 | schedule_debug(prev); | |
1da177e4 | 4570 | |
8f4d37ec PZ |
4571 | hrtick_clear(rq); |
4572 | ||
1e819950 IM |
4573 | /* |
4574 | * Do the rq-clock update outside the rq lock: | |
4575 | */ | |
4576 | local_irq_disable(); | |
c1b3da3e | 4577 | __update_rq_clock(rq); |
1e819950 IM |
4578 | spin_lock(&rq->lock); |
4579 | clear_tsk_need_resched(prev); | |
1da177e4 | 4580 | |
1da177e4 | 4581 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
1da177e4 | 4582 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && |
23e3c3cd | 4583 | signal_pending(prev))) { |
1da177e4 | 4584 | prev->state = TASK_RUNNING; |
dd41f596 | 4585 | } else { |
2e1cb74a | 4586 | deactivate_task(rq, prev, 1); |
1da177e4 | 4587 | } |
dd41f596 | 4588 | switch_count = &prev->nvcsw; |
1da177e4 LT |
4589 | } |
4590 | ||
9a897c5a SR |
4591 | #ifdef CONFIG_SMP |
4592 | if (prev->sched_class->pre_schedule) | |
4593 | prev->sched_class->pre_schedule(rq, prev); | |
4594 | #endif | |
f65eda4f | 4595 | |
dd41f596 | 4596 | if (unlikely(!rq->nr_running)) |
1da177e4 | 4597 | idle_balance(cpu, rq); |
1da177e4 | 4598 | |
31ee529c | 4599 | prev->sched_class->put_prev_task(rq, prev); |
ff95f3df | 4600 | next = pick_next_task(rq, prev); |
1da177e4 LT |
4601 | |
4602 | sched_info_switch(prev, next); | |
dd41f596 | 4603 | |
1da177e4 | 4604 | if (likely(prev != next)) { |
1da177e4 LT |
4605 | rq->nr_switches++; |
4606 | rq->curr = next; | |
4607 | ++*switch_count; | |
4608 | ||
dd41f596 | 4609 | context_switch(rq, prev, next); /* unlocks the rq */ |
8f4d37ec PZ |
4610 | /* |
4611 | * the context switch might have flipped the stack from under | |
4612 | * us, hence refresh the local variables. | |
4613 | */ | |
4614 | cpu = smp_processor_id(); | |
4615 | rq = cpu_rq(cpu); | |
1da177e4 LT |
4616 | } else |
4617 | spin_unlock_irq(&rq->lock); | |
4618 | ||
8f4d37ec PZ |
4619 | hrtick_set(rq); |
4620 | ||
4621 | if (unlikely(reacquire_kernel_lock(current) < 0)) | |
1da177e4 | 4622 | goto need_resched_nonpreemptible; |
8f4d37ec | 4623 | |
1da177e4 LT |
4624 | preempt_enable_no_resched(); |
4625 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
4626 | goto need_resched; | |
4627 | } | |
1da177e4 LT |
4628 | EXPORT_SYMBOL(schedule); |
4629 | ||
4630 | #ifdef CONFIG_PREEMPT | |
4631 | /* | |
2ed6e34f | 4632 | * this is the entry point to schedule() from in-kernel preemption |
41a2d6cf | 4633 | * off of preempt_enable. Kernel preemptions off return from interrupt |
1da177e4 LT |
4634 | * occur there and call schedule directly. |
4635 | */ | |
4636 | asmlinkage void __sched preempt_schedule(void) | |
4637 | { | |
4638 | struct thread_info *ti = current_thread_info(); | |
1da177e4 LT |
4639 | struct task_struct *task = current; |
4640 | int saved_lock_depth; | |
6478d880 | 4641 | |
1da177e4 LT |
4642 | /* |
4643 | * If there is a non-zero preempt_count or interrupts are disabled, | |
41a2d6cf | 4644 | * we do not want to preempt the current task. Just return.. |
1da177e4 | 4645 | */ |
beed33a8 | 4646 | if (likely(ti->preempt_count || irqs_disabled())) |
1da177e4 LT |
4647 | return; |
4648 | ||
3a5c359a AK |
4649 | do { |
4650 | add_preempt_count(PREEMPT_ACTIVE); | |
4651 | ||
4652 | /* | |
4653 | * We keep the big kernel semaphore locked, but we | |
4654 | * clear ->lock_depth so that schedule() doesnt | |
4655 | * auto-release the semaphore: | |
4656 | */ | |
3a5c359a AK |
4657 | saved_lock_depth = task->lock_depth; |
4658 | task->lock_depth = -1; | |
3a5c359a | 4659 | schedule(); |
3a5c359a | 4660 | task->lock_depth = saved_lock_depth; |
3a5c359a | 4661 | sub_preempt_count(PREEMPT_ACTIVE); |
1da177e4 | 4662 | |
3a5c359a AK |
4663 | /* |
4664 | * Check again in case we missed a preemption opportunity | |
4665 | * between schedule and now. | |
4666 | */ | |
4667 | barrier(); | |
4668 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); | |
1da177e4 | 4669 | } |
1da177e4 LT |
4670 | EXPORT_SYMBOL(preempt_schedule); |
4671 | ||
4672 | /* | |
2ed6e34f | 4673 | * this is the entry point to schedule() from kernel preemption |
1da177e4 LT |
4674 | * off of irq context. |
4675 | * Note, that this is called and return with irqs disabled. This will | |
4676 | * protect us against recursive calling from irq. | |
4677 | */ | |
4678 | asmlinkage void __sched preempt_schedule_irq(void) | |
4679 | { | |
4680 | struct thread_info *ti = current_thread_info(); | |
1da177e4 LT |
4681 | struct task_struct *task = current; |
4682 | int saved_lock_depth; | |
6478d880 | 4683 | |
2ed6e34f | 4684 | /* Catch callers which need to be fixed */ |
1da177e4 LT |
4685 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
4686 | ||
3a5c359a AK |
4687 | do { |
4688 | add_preempt_count(PREEMPT_ACTIVE); | |
4689 | ||
4690 | /* | |
4691 | * We keep the big kernel semaphore locked, but we | |
4692 | * clear ->lock_depth so that schedule() doesnt | |
4693 | * auto-release the semaphore: | |
4694 | */ | |
3a5c359a AK |
4695 | saved_lock_depth = task->lock_depth; |
4696 | task->lock_depth = -1; | |
3a5c359a AK |
4697 | local_irq_enable(); |
4698 | schedule(); | |
4699 | local_irq_disable(); | |
3a5c359a | 4700 | task->lock_depth = saved_lock_depth; |
3a5c359a | 4701 | sub_preempt_count(PREEMPT_ACTIVE); |
1da177e4 | 4702 | |
3a5c359a AK |
4703 | /* |
4704 | * Check again in case we missed a preemption opportunity | |
4705 | * between schedule and now. | |
4706 | */ | |
4707 | barrier(); | |
4708 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); | |
1da177e4 LT |
4709 | } |
4710 | ||
4711 | #endif /* CONFIG_PREEMPT */ | |
4712 | ||
95cdf3b7 IM |
4713 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
4714 | void *key) | |
1da177e4 | 4715 | { |
48f24c4d | 4716 | return try_to_wake_up(curr->private, mode, sync); |
1da177e4 | 4717 | } |
1da177e4 LT |
4718 | EXPORT_SYMBOL(default_wake_function); |
4719 | ||
4720 | /* | |
41a2d6cf IM |
4721 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
4722 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
1da177e4 LT |
4723 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
4724 | * | |
4725 | * There are circumstances in which we can try to wake a task which has already | |
41a2d6cf | 4726 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
1da177e4 LT |
4727 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
4728 | */ | |
4729 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
4730 | int nr_exclusive, int sync, void *key) | |
4731 | { | |
2e45874c | 4732 | wait_queue_t *curr, *next; |
1da177e4 | 4733 | |
2e45874c | 4734 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { |
48f24c4d IM |
4735 | unsigned flags = curr->flags; |
4736 | ||
1da177e4 | 4737 | if (curr->func(curr, mode, sync, key) && |
48f24c4d | 4738 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) |
1da177e4 LT |
4739 | break; |
4740 | } | |
4741 | } | |
4742 | ||
4743 | /** | |
4744 | * __wake_up - wake up threads blocked on a waitqueue. | |
4745 | * @q: the waitqueue | |
4746 | * @mode: which threads | |
4747 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 4748 | * @key: is directly passed to the wakeup function |
1da177e4 | 4749 | */ |
7ad5b3a5 | 4750 | void __wake_up(wait_queue_head_t *q, unsigned int mode, |
95cdf3b7 | 4751 | int nr_exclusive, void *key) |
1da177e4 LT |
4752 | { |
4753 | unsigned long flags; | |
4754 | ||
4755 | spin_lock_irqsave(&q->lock, flags); | |
4756 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
4757 | spin_unlock_irqrestore(&q->lock, flags); | |
4758 | } | |
1da177e4 LT |
4759 | EXPORT_SYMBOL(__wake_up); |
4760 | ||
4761 | /* | |
4762 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
4763 | */ | |
7ad5b3a5 | 4764 | void __wake_up_locked(wait_queue_head_t *q, unsigned int mode) |
1da177e4 LT |
4765 | { |
4766 | __wake_up_common(q, mode, 1, 0, NULL); | |
4767 | } | |
4768 | ||
4769 | /** | |
67be2dd1 | 4770 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
4771 | * @q: the waitqueue |
4772 | * @mode: which threads | |
4773 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
4774 | * | |
4775 | * The sync wakeup differs that the waker knows that it will schedule | |
4776 | * away soon, so while the target thread will be woken up, it will not | |
4777 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
4778 | * with each other. This can prevent needless bouncing between CPUs. | |
4779 | * | |
4780 | * On UP it can prevent extra preemption. | |
4781 | */ | |
7ad5b3a5 | 4782 | void |
95cdf3b7 | 4783 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
1da177e4 LT |
4784 | { |
4785 | unsigned long flags; | |
4786 | int sync = 1; | |
4787 | ||
4788 | if (unlikely(!q)) | |
4789 | return; | |
4790 | ||
4791 | if (unlikely(!nr_exclusive)) | |
4792 | sync = 0; | |
4793 | ||
4794 | spin_lock_irqsave(&q->lock, flags); | |
4795 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
4796 | spin_unlock_irqrestore(&q->lock, flags); | |
4797 | } | |
4798 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
4799 | ||
b15136e9 | 4800 | void complete(struct completion *x) |
1da177e4 LT |
4801 | { |
4802 | unsigned long flags; | |
4803 | ||
4804 | spin_lock_irqsave(&x->wait.lock, flags); | |
4805 | x->done++; | |
d9514f6c | 4806 | __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); |
1da177e4 LT |
4807 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4808 | } | |
4809 | EXPORT_SYMBOL(complete); | |
4810 | ||
b15136e9 | 4811 | void complete_all(struct completion *x) |
1da177e4 LT |
4812 | { |
4813 | unsigned long flags; | |
4814 | ||
4815 | spin_lock_irqsave(&x->wait.lock, flags); | |
4816 | x->done += UINT_MAX/2; | |
d9514f6c | 4817 | __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); |
1da177e4 LT |
4818 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4819 | } | |
4820 | EXPORT_SYMBOL(complete_all); | |
4821 | ||
8cbbe86d AK |
4822 | static inline long __sched |
4823 | do_wait_for_common(struct completion *x, long timeout, int state) | |
1da177e4 | 4824 | { |
1da177e4 LT |
4825 | if (!x->done) { |
4826 | DECLARE_WAITQUEUE(wait, current); | |
4827 | ||
4828 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
4829 | __add_wait_queue_tail(&x->wait, &wait); | |
4830 | do { | |
009e577e MW |
4831 | if ((state == TASK_INTERRUPTIBLE && |
4832 | signal_pending(current)) || | |
4833 | (state == TASK_KILLABLE && | |
4834 | fatal_signal_pending(current))) { | |
8cbbe86d AK |
4835 | __remove_wait_queue(&x->wait, &wait); |
4836 | return -ERESTARTSYS; | |
4837 | } | |
4838 | __set_current_state(state); | |
1da177e4 LT |
4839 | spin_unlock_irq(&x->wait.lock); |
4840 | timeout = schedule_timeout(timeout); | |
4841 | spin_lock_irq(&x->wait.lock); | |
4842 | if (!timeout) { | |
4843 | __remove_wait_queue(&x->wait, &wait); | |
8cbbe86d | 4844 | return timeout; |
1da177e4 LT |
4845 | } |
4846 | } while (!x->done); | |
4847 | __remove_wait_queue(&x->wait, &wait); | |
4848 | } | |
4849 | x->done--; | |
1da177e4 LT |
4850 | return timeout; |
4851 | } | |
1da177e4 | 4852 | |
8cbbe86d AK |
4853 | static long __sched |
4854 | wait_for_common(struct completion *x, long timeout, int state) | |
1da177e4 | 4855 | { |
1da177e4 LT |
4856 | might_sleep(); |
4857 | ||
4858 | spin_lock_irq(&x->wait.lock); | |
8cbbe86d | 4859 | timeout = do_wait_for_common(x, timeout, state); |
1da177e4 | 4860 | spin_unlock_irq(&x->wait.lock); |
8cbbe86d AK |
4861 | return timeout; |
4862 | } | |
1da177e4 | 4863 | |
b15136e9 | 4864 | void __sched wait_for_completion(struct completion *x) |
8cbbe86d AK |
4865 | { |
4866 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); | |
1da177e4 | 4867 | } |
8cbbe86d | 4868 | EXPORT_SYMBOL(wait_for_completion); |
1da177e4 | 4869 | |
b15136e9 | 4870 | unsigned long __sched |
8cbbe86d | 4871 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
1da177e4 | 4872 | { |
8cbbe86d | 4873 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); |
1da177e4 | 4874 | } |
8cbbe86d | 4875 | EXPORT_SYMBOL(wait_for_completion_timeout); |
1da177e4 | 4876 | |
8cbbe86d | 4877 | int __sched wait_for_completion_interruptible(struct completion *x) |
0fec171c | 4878 | { |
51e97990 AK |
4879 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); |
4880 | if (t == -ERESTARTSYS) | |
4881 | return t; | |
4882 | return 0; | |
0fec171c | 4883 | } |
8cbbe86d | 4884 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
1da177e4 | 4885 | |
b15136e9 | 4886 | unsigned long __sched |
8cbbe86d AK |
4887 | wait_for_completion_interruptible_timeout(struct completion *x, |
4888 | unsigned long timeout) | |
0fec171c | 4889 | { |
8cbbe86d | 4890 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); |
0fec171c | 4891 | } |
8cbbe86d | 4892 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
1da177e4 | 4893 | |
009e577e MW |
4894 | int __sched wait_for_completion_killable(struct completion *x) |
4895 | { | |
4896 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); | |
4897 | if (t == -ERESTARTSYS) | |
4898 | return t; | |
4899 | return 0; | |
4900 | } | |
4901 | EXPORT_SYMBOL(wait_for_completion_killable); | |
4902 | ||
8cbbe86d AK |
4903 | static long __sched |
4904 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) | |
1da177e4 | 4905 | { |
0fec171c IM |
4906 | unsigned long flags; |
4907 | wait_queue_t wait; | |
4908 | ||
4909 | init_waitqueue_entry(&wait, current); | |
1da177e4 | 4910 | |
8cbbe86d | 4911 | __set_current_state(state); |
1da177e4 | 4912 | |
8cbbe86d AK |
4913 | spin_lock_irqsave(&q->lock, flags); |
4914 | __add_wait_queue(q, &wait); | |
4915 | spin_unlock(&q->lock); | |
4916 | timeout = schedule_timeout(timeout); | |
4917 | spin_lock_irq(&q->lock); | |
4918 | __remove_wait_queue(q, &wait); | |
4919 | spin_unlock_irqrestore(&q->lock, flags); | |
4920 | ||
4921 | return timeout; | |
4922 | } | |
4923 | ||
4924 | void __sched interruptible_sleep_on(wait_queue_head_t *q) | |
4925 | { | |
4926 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | |
1da177e4 | 4927 | } |
1da177e4 LT |
4928 | EXPORT_SYMBOL(interruptible_sleep_on); |
4929 | ||
0fec171c | 4930 | long __sched |
95cdf3b7 | 4931 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
1da177e4 | 4932 | { |
8cbbe86d | 4933 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); |
1da177e4 | 4934 | } |
1da177e4 LT |
4935 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
4936 | ||
0fec171c | 4937 | void __sched sleep_on(wait_queue_head_t *q) |
1da177e4 | 4938 | { |
8cbbe86d | 4939 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
1da177e4 | 4940 | } |
1da177e4 LT |
4941 | EXPORT_SYMBOL(sleep_on); |
4942 | ||
0fec171c | 4943 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
1da177e4 | 4944 | { |
8cbbe86d | 4945 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); |
1da177e4 | 4946 | } |
1da177e4 LT |
4947 | EXPORT_SYMBOL(sleep_on_timeout); |
4948 | ||
b29739f9 IM |
4949 | #ifdef CONFIG_RT_MUTEXES |
4950 | ||
4951 | /* | |
4952 | * rt_mutex_setprio - set the current priority of a task | |
4953 | * @p: task | |
4954 | * @prio: prio value (kernel-internal form) | |
4955 | * | |
4956 | * This function changes the 'effective' priority of a task. It does | |
4957 | * not touch ->normal_prio like __setscheduler(). | |
4958 | * | |
4959 | * Used by the rt_mutex code to implement priority inheritance logic. | |
4960 | */ | |
36c8b586 | 4961 | void rt_mutex_setprio(struct task_struct *p, int prio) |
b29739f9 IM |
4962 | { |
4963 | unsigned long flags; | |
83b699ed | 4964 | int oldprio, on_rq, running; |
70b97a7f | 4965 | struct rq *rq; |
cb469845 | 4966 | const struct sched_class *prev_class = p->sched_class; |
b29739f9 IM |
4967 | |
4968 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
4969 | ||
4970 | rq = task_rq_lock(p, &flags); | |
a8e504d2 | 4971 | update_rq_clock(rq); |
b29739f9 | 4972 | |
d5f9f942 | 4973 | oldprio = p->prio; |
dd41f596 | 4974 | on_rq = p->se.on_rq; |
051a1d1a | 4975 | running = task_current(rq, p); |
0e1f3483 | 4976 | if (on_rq) |
69be72c1 | 4977 | dequeue_task(rq, p, 0); |
0e1f3483 HS |
4978 | if (running) |
4979 | p->sched_class->put_prev_task(rq, p); | |
dd41f596 IM |
4980 | |
4981 | if (rt_prio(prio)) | |
4982 | p->sched_class = &rt_sched_class; | |
4983 | else | |
4984 | p->sched_class = &fair_sched_class; | |
4985 | ||
b29739f9 IM |
4986 | p->prio = prio; |
4987 | ||
0e1f3483 HS |
4988 | if (running) |
4989 | p->sched_class->set_curr_task(rq); | |
dd41f596 | 4990 | if (on_rq) { |
8159f87e | 4991 | enqueue_task(rq, p, 0); |
cb469845 SR |
4992 | |
4993 | check_class_changed(rq, p, prev_class, oldprio, running); | |
b29739f9 IM |
4994 | } |
4995 | task_rq_unlock(rq, &flags); | |
4996 | } | |
4997 | ||
4998 | #endif | |
4999 | ||
36c8b586 | 5000 | void set_user_nice(struct task_struct *p, long nice) |
1da177e4 | 5001 | { |
dd41f596 | 5002 | int old_prio, delta, on_rq; |
1da177e4 | 5003 | unsigned long flags; |
70b97a7f | 5004 | struct rq *rq; |
1da177e4 LT |
5005 | |
5006 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
5007 | return; | |
5008 | /* | |
5009 | * We have to be careful, if called from sys_setpriority(), | |
5010 | * the task might be in the middle of scheduling on another CPU. | |
5011 | */ | |
5012 | rq = task_rq_lock(p, &flags); | |
a8e504d2 | 5013 | update_rq_clock(rq); |
1da177e4 LT |
5014 | /* |
5015 | * The RT priorities are set via sched_setscheduler(), but we still | |
5016 | * allow the 'normal' nice value to be set - but as expected | |
5017 | * it wont have any effect on scheduling until the task is | |
dd41f596 | 5018 | * SCHED_FIFO/SCHED_RR: |
1da177e4 | 5019 | */ |
e05606d3 | 5020 | if (task_has_rt_policy(p)) { |
1da177e4 LT |
5021 | p->static_prio = NICE_TO_PRIO(nice); |
5022 | goto out_unlock; | |
5023 | } | |
dd41f596 | 5024 | on_rq = p->se.on_rq; |
18d95a28 | 5025 | if (on_rq) |
69be72c1 | 5026 | dequeue_task(rq, p, 0); |
1da177e4 | 5027 | |
1da177e4 | 5028 | p->static_prio = NICE_TO_PRIO(nice); |
2dd73a4f | 5029 | set_load_weight(p); |
b29739f9 IM |
5030 | old_prio = p->prio; |
5031 | p->prio = effective_prio(p); | |
5032 | delta = p->prio - old_prio; | |
1da177e4 | 5033 | |
dd41f596 | 5034 | if (on_rq) { |
8159f87e | 5035 | enqueue_task(rq, p, 0); |
1da177e4 | 5036 | /* |
d5f9f942 AM |
5037 | * If the task increased its priority or is running and |
5038 | * lowered its priority, then reschedule its CPU: | |
1da177e4 | 5039 | */ |
d5f9f942 | 5040 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
1da177e4 LT |
5041 | resched_task(rq->curr); |
5042 | } | |
5043 | out_unlock: | |
5044 | task_rq_unlock(rq, &flags); | |
5045 | } | |
1da177e4 LT |
5046 | EXPORT_SYMBOL(set_user_nice); |
5047 | ||
e43379f1 MM |
5048 | /* |
5049 | * can_nice - check if a task can reduce its nice value | |
5050 | * @p: task | |
5051 | * @nice: nice value | |
5052 | */ | |
36c8b586 | 5053 | int can_nice(const struct task_struct *p, const int nice) |
e43379f1 | 5054 | { |
024f4747 MM |
5055 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
5056 | int nice_rlim = 20 - nice; | |
48f24c4d | 5057 | |
e43379f1 MM |
5058 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
5059 | capable(CAP_SYS_NICE)); | |
5060 | } | |
5061 | ||
1da177e4 LT |
5062 | #ifdef __ARCH_WANT_SYS_NICE |
5063 | ||
5064 | /* | |
5065 | * sys_nice - change the priority of the current process. | |
5066 | * @increment: priority increment | |
5067 | * | |
5068 | * sys_setpriority is a more generic, but much slower function that | |
5069 | * does similar things. | |
5070 | */ | |
5071 | asmlinkage long sys_nice(int increment) | |
5072 | { | |
48f24c4d | 5073 | long nice, retval; |
1da177e4 LT |
5074 | |
5075 | /* | |
5076 | * Setpriority might change our priority at the same moment. | |
5077 | * We don't have to worry. Conceptually one call occurs first | |
5078 | * and we have a single winner. | |
5079 | */ | |
e43379f1 MM |
5080 | if (increment < -40) |
5081 | increment = -40; | |
1da177e4 LT |
5082 | if (increment > 40) |
5083 | increment = 40; | |
5084 | ||
5085 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
5086 | if (nice < -20) | |
5087 | nice = -20; | |
5088 | if (nice > 19) | |
5089 | nice = 19; | |
5090 | ||
e43379f1 MM |
5091 | if (increment < 0 && !can_nice(current, nice)) |
5092 | return -EPERM; | |
5093 | ||
1da177e4 LT |
5094 | retval = security_task_setnice(current, nice); |
5095 | if (retval) | |
5096 | return retval; | |
5097 | ||
5098 | set_user_nice(current, nice); | |
5099 | return 0; | |
5100 | } | |
5101 | ||
5102 | #endif | |
5103 | ||
5104 | /** | |
5105 | * task_prio - return the priority value of a given task. | |
5106 | * @p: the task in question. | |
5107 | * | |
5108 | * This is the priority value as seen by users in /proc. | |
5109 | * RT tasks are offset by -200. Normal tasks are centered | |
5110 | * around 0, value goes from -16 to +15. | |
5111 | */ | |
36c8b586 | 5112 | int task_prio(const struct task_struct *p) |
1da177e4 LT |
5113 | { |
5114 | return p->prio - MAX_RT_PRIO; | |
5115 | } | |
5116 | ||
5117 | /** | |
5118 | * task_nice - return the nice value of a given task. | |
5119 | * @p: the task in question. | |
5120 | */ | |
36c8b586 | 5121 | int task_nice(const struct task_struct *p) |
1da177e4 LT |
5122 | { |
5123 | return TASK_NICE(p); | |
5124 | } | |
150d8bed | 5125 | EXPORT_SYMBOL(task_nice); |
1da177e4 LT |
5126 | |
5127 | /** | |
5128 | * idle_cpu - is a given cpu idle currently? | |
5129 | * @cpu: the processor in question. | |
5130 | */ | |
5131 | int idle_cpu(int cpu) | |
5132 | { | |
5133 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
5134 | } | |
5135 | ||
1da177e4 LT |
5136 | /** |
5137 | * idle_task - return the idle task for a given cpu. | |
5138 | * @cpu: the processor in question. | |
5139 | */ | |
36c8b586 | 5140 | struct task_struct *idle_task(int cpu) |
1da177e4 LT |
5141 | { |
5142 | return cpu_rq(cpu)->idle; | |
5143 | } | |
5144 | ||
5145 | /** | |
5146 | * find_process_by_pid - find a process with a matching PID value. | |
5147 | * @pid: the pid in question. | |
5148 | */ | |
a9957449 | 5149 | static struct task_struct *find_process_by_pid(pid_t pid) |
1da177e4 | 5150 | { |
228ebcbe | 5151 | return pid ? find_task_by_vpid(pid) : current; |
1da177e4 LT |
5152 | } |
5153 | ||
5154 | /* Actually do priority change: must hold rq lock. */ | |
dd41f596 IM |
5155 | static void |
5156 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) | |
1da177e4 | 5157 | { |
dd41f596 | 5158 | BUG_ON(p->se.on_rq); |
48f24c4d | 5159 | |
1da177e4 | 5160 | p->policy = policy; |
dd41f596 IM |
5161 | switch (p->policy) { |
5162 | case SCHED_NORMAL: | |
5163 | case SCHED_BATCH: | |
5164 | case SCHED_IDLE: | |
5165 | p->sched_class = &fair_sched_class; | |
5166 | break; | |
5167 | case SCHED_FIFO: | |
5168 | case SCHED_RR: | |
5169 | p->sched_class = &rt_sched_class; | |
5170 | break; | |
5171 | } | |
5172 | ||
1da177e4 | 5173 | p->rt_priority = prio; |
b29739f9 IM |
5174 | p->normal_prio = normal_prio(p); |
5175 | /* we are holding p->pi_lock already */ | |
5176 | p->prio = rt_mutex_getprio(p); | |
2dd73a4f | 5177 | set_load_weight(p); |
1da177e4 LT |
5178 | } |
5179 | ||
5180 | /** | |
72fd4a35 | 5181 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. |
1da177e4 LT |
5182 | * @p: the task in question. |
5183 | * @policy: new policy. | |
5184 | * @param: structure containing the new RT priority. | |
5fe1d75f | 5185 | * |
72fd4a35 | 5186 | * NOTE that the task may be already dead. |
1da177e4 | 5187 | */ |
95cdf3b7 IM |
5188 | int sched_setscheduler(struct task_struct *p, int policy, |
5189 | struct sched_param *param) | |
1da177e4 | 5190 | { |
83b699ed | 5191 | int retval, oldprio, oldpolicy = -1, on_rq, running; |
1da177e4 | 5192 | unsigned long flags; |
cb469845 | 5193 | const struct sched_class *prev_class = p->sched_class; |
70b97a7f | 5194 | struct rq *rq; |
1da177e4 | 5195 | |
66e5393a SR |
5196 | /* may grab non-irq protected spin_locks */ |
5197 | BUG_ON(in_interrupt()); | |
1da177e4 LT |
5198 | recheck: |
5199 | /* double check policy once rq lock held */ | |
5200 | if (policy < 0) | |
5201 | policy = oldpolicy = p->policy; | |
5202 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
dd41f596 IM |
5203 | policy != SCHED_NORMAL && policy != SCHED_BATCH && |
5204 | policy != SCHED_IDLE) | |
b0a9499c | 5205 | return -EINVAL; |
1da177e4 LT |
5206 | /* |
5207 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
dd41f596 IM |
5208 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, |
5209 | * SCHED_BATCH and SCHED_IDLE is 0. | |
1da177e4 LT |
5210 | */ |
5211 | if (param->sched_priority < 0 || | |
95cdf3b7 | 5212 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 5213 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 5214 | return -EINVAL; |
e05606d3 | 5215 | if (rt_policy(policy) != (param->sched_priority != 0)) |
1da177e4 LT |
5216 | return -EINVAL; |
5217 | ||
37e4ab3f OC |
5218 | /* |
5219 | * Allow unprivileged RT tasks to decrease priority: | |
5220 | */ | |
5221 | if (!capable(CAP_SYS_NICE)) { | |
e05606d3 | 5222 | if (rt_policy(policy)) { |
8dc3e909 | 5223 | unsigned long rlim_rtprio; |
8dc3e909 ON |
5224 | |
5225 | if (!lock_task_sighand(p, &flags)) | |
5226 | return -ESRCH; | |
5227 | rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; | |
5228 | unlock_task_sighand(p, &flags); | |
5229 | ||
5230 | /* can't set/change the rt policy */ | |
5231 | if (policy != p->policy && !rlim_rtprio) | |
5232 | return -EPERM; | |
5233 | ||
5234 | /* can't increase priority */ | |
5235 | if (param->sched_priority > p->rt_priority && | |
5236 | param->sched_priority > rlim_rtprio) | |
5237 | return -EPERM; | |
5238 | } | |
dd41f596 IM |
5239 | /* |
5240 | * Like positive nice levels, dont allow tasks to | |
5241 | * move out of SCHED_IDLE either: | |
5242 | */ | |
5243 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) | |
5244 | return -EPERM; | |
5fe1d75f | 5245 | |
37e4ab3f OC |
5246 | /* can't change other user's priorities */ |
5247 | if ((current->euid != p->euid) && | |
5248 | (current->euid != p->uid)) | |
5249 | return -EPERM; | |
5250 | } | |
1da177e4 | 5251 | |
b68aa230 PZ |
5252 | #ifdef CONFIG_RT_GROUP_SCHED |
5253 | /* | |
5254 | * Do not allow realtime tasks into groups that have no runtime | |
5255 | * assigned. | |
5256 | */ | |
d0b27fa7 | 5257 | if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0) |
b68aa230 PZ |
5258 | return -EPERM; |
5259 | #endif | |
5260 | ||
1da177e4 LT |
5261 | retval = security_task_setscheduler(p, policy, param); |
5262 | if (retval) | |
5263 | return retval; | |
b29739f9 IM |
5264 | /* |
5265 | * make sure no PI-waiters arrive (or leave) while we are | |
5266 | * changing the priority of the task: | |
5267 | */ | |
5268 | spin_lock_irqsave(&p->pi_lock, flags); | |
1da177e4 LT |
5269 | /* |
5270 | * To be able to change p->policy safely, the apropriate | |
5271 | * runqueue lock must be held. | |
5272 | */ | |
b29739f9 | 5273 | rq = __task_rq_lock(p); |
1da177e4 LT |
5274 | /* recheck policy now with rq lock held */ |
5275 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
5276 | policy = oldpolicy = -1; | |
b29739f9 IM |
5277 | __task_rq_unlock(rq); |
5278 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
5279 | goto recheck; |
5280 | } | |
2daa3577 | 5281 | update_rq_clock(rq); |
dd41f596 | 5282 | on_rq = p->se.on_rq; |
051a1d1a | 5283 | running = task_current(rq, p); |
0e1f3483 | 5284 | if (on_rq) |
2e1cb74a | 5285 | deactivate_task(rq, p, 0); |
0e1f3483 HS |
5286 | if (running) |
5287 | p->sched_class->put_prev_task(rq, p); | |
f6b53205 | 5288 | |
1da177e4 | 5289 | oldprio = p->prio; |
dd41f596 | 5290 | __setscheduler(rq, p, policy, param->sched_priority); |
f6b53205 | 5291 | |
0e1f3483 HS |
5292 | if (running) |
5293 | p->sched_class->set_curr_task(rq); | |
dd41f596 IM |
5294 | if (on_rq) { |
5295 | activate_task(rq, p, 0); | |
cb469845 SR |
5296 | |
5297 | check_class_changed(rq, p, prev_class, oldprio, running); | |
1da177e4 | 5298 | } |
b29739f9 IM |
5299 | __task_rq_unlock(rq); |
5300 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
5301 | ||
95e02ca9 TG |
5302 | rt_mutex_adjust_pi(p); |
5303 | ||
1da177e4 LT |
5304 | return 0; |
5305 | } | |
5306 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
5307 | ||
95cdf3b7 IM |
5308 | static int |
5309 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 | 5310 | { |
1da177e4 LT |
5311 | struct sched_param lparam; |
5312 | struct task_struct *p; | |
36c8b586 | 5313 | int retval; |
1da177e4 LT |
5314 | |
5315 | if (!param || pid < 0) | |
5316 | return -EINVAL; | |
5317 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
5318 | return -EFAULT; | |
5fe1d75f ON |
5319 | |
5320 | rcu_read_lock(); | |
5321 | retval = -ESRCH; | |
1da177e4 | 5322 | p = find_process_by_pid(pid); |
5fe1d75f ON |
5323 | if (p != NULL) |
5324 | retval = sched_setscheduler(p, policy, &lparam); | |
5325 | rcu_read_unlock(); | |
36c8b586 | 5326 | |
1da177e4 LT |
5327 | return retval; |
5328 | } | |
5329 | ||
5330 | /** | |
5331 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
5332 | * @pid: the pid in question. | |
5333 | * @policy: new policy. | |
5334 | * @param: structure containing the new RT priority. | |
5335 | */ | |
41a2d6cf IM |
5336 | asmlinkage long |
5337 | sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 | 5338 | { |
c21761f1 JB |
5339 | /* negative values for policy are not valid */ |
5340 | if (policy < 0) | |
5341 | return -EINVAL; | |
5342 | ||
1da177e4 LT |
5343 | return do_sched_setscheduler(pid, policy, param); |
5344 | } | |
5345 | ||
5346 | /** | |
5347 | * sys_sched_setparam - set/change the RT priority of a thread | |
5348 | * @pid: the pid in question. | |
5349 | * @param: structure containing the new RT priority. | |
5350 | */ | |
5351 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
5352 | { | |
5353 | return do_sched_setscheduler(pid, -1, param); | |
5354 | } | |
5355 | ||
5356 | /** | |
5357 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
5358 | * @pid: the pid in question. | |
5359 | */ | |
5360 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
5361 | { | |
36c8b586 | 5362 | struct task_struct *p; |
3a5c359a | 5363 | int retval; |
1da177e4 LT |
5364 | |
5365 | if (pid < 0) | |
3a5c359a | 5366 | return -EINVAL; |
1da177e4 LT |
5367 | |
5368 | retval = -ESRCH; | |
5369 | read_lock(&tasklist_lock); | |
5370 | p = find_process_by_pid(pid); | |
5371 | if (p) { | |
5372 | retval = security_task_getscheduler(p); | |
5373 | if (!retval) | |
5374 | retval = p->policy; | |
5375 | } | |
5376 | read_unlock(&tasklist_lock); | |
1da177e4 LT |
5377 | return retval; |
5378 | } | |
5379 | ||
5380 | /** | |
5381 | * sys_sched_getscheduler - get the RT priority of a thread | |
5382 | * @pid: the pid in question. | |
5383 | * @param: structure containing the RT priority. | |
5384 | */ | |
5385 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
5386 | { | |
5387 | struct sched_param lp; | |
36c8b586 | 5388 | struct task_struct *p; |
3a5c359a | 5389 | int retval; |
1da177e4 LT |
5390 | |
5391 | if (!param || pid < 0) | |
3a5c359a | 5392 | return -EINVAL; |
1da177e4 LT |
5393 | |
5394 | read_lock(&tasklist_lock); | |
5395 | p = find_process_by_pid(pid); | |
5396 | retval = -ESRCH; | |
5397 | if (!p) | |
5398 | goto out_unlock; | |
5399 | ||
5400 | retval = security_task_getscheduler(p); | |
5401 | if (retval) | |
5402 | goto out_unlock; | |
5403 | ||
5404 | lp.sched_priority = p->rt_priority; | |
5405 | read_unlock(&tasklist_lock); | |
5406 | ||
5407 | /* | |
5408 | * This one might sleep, we cannot do it with a spinlock held ... | |
5409 | */ | |
5410 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
5411 | ||
1da177e4 LT |
5412 | return retval; |
5413 | ||
5414 | out_unlock: | |
5415 | read_unlock(&tasklist_lock); | |
5416 | return retval; | |
5417 | } | |
5418 | ||
b53e921b | 5419 | long sched_setaffinity(pid_t pid, const cpumask_t *in_mask) |
1da177e4 | 5420 | { |
1da177e4 | 5421 | cpumask_t cpus_allowed; |
b53e921b | 5422 | cpumask_t new_mask = *in_mask; |
36c8b586 IM |
5423 | struct task_struct *p; |
5424 | int retval; | |
1da177e4 | 5425 | |
95402b38 | 5426 | get_online_cpus(); |
1da177e4 LT |
5427 | read_lock(&tasklist_lock); |
5428 | ||
5429 | p = find_process_by_pid(pid); | |
5430 | if (!p) { | |
5431 | read_unlock(&tasklist_lock); | |
95402b38 | 5432 | put_online_cpus(); |
1da177e4 LT |
5433 | return -ESRCH; |
5434 | } | |
5435 | ||
5436 | /* | |
5437 | * It is not safe to call set_cpus_allowed with the | |
41a2d6cf | 5438 | * tasklist_lock held. We will bump the task_struct's |
1da177e4 LT |
5439 | * usage count and then drop tasklist_lock. |
5440 | */ | |
5441 | get_task_struct(p); | |
5442 | read_unlock(&tasklist_lock); | |
5443 | ||
5444 | retval = -EPERM; | |
5445 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
5446 | !capable(CAP_SYS_NICE)) | |
5447 | goto out_unlock; | |
5448 | ||
e7834f8f DQ |
5449 | retval = security_task_setscheduler(p, 0, NULL); |
5450 | if (retval) | |
5451 | goto out_unlock; | |
5452 | ||
f9a86fcb | 5453 | cpuset_cpus_allowed(p, &cpus_allowed); |
1da177e4 | 5454 | cpus_and(new_mask, new_mask, cpus_allowed); |
8707d8b8 | 5455 | again: |
7c16ec58 | 5456 | retval = set_cpus_allowed_ptr(p, &new_mask); |
1da177e4 | 5457 | |
8707d8b8 | 5458 | if (!retval) { |
f9a86fcb | 5459 | cpuset_cpus_allowed(p, &cpus_allowed); |
8707d8b8 PM |
5460 | if (!cpus_subset(new_mask, cpus_allowed)) { |
5461 | /* | |
5462 | * We must have raced with a concurrent cpuset | |
5463 | * update. Just reset the cpus_allowed to the | |
5464 | * cpuset's cpus_allowed | |
5465 | */ | |
5466 | new_mask = cpus_allowed; | |
5467 | goto again; | |
5468 | } | |
5469 | } | |
1da177e4 LT |
5470 | out_unlock: |
5471 | put_task_struct(p); | |
95402b38 | 5472 | put_online_cpus(); |
1da177e4 LT |
5473 | return retval; |
5474 | } | |
5475 | ||
5476 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
5477 | cpumask_t *new_mask) | |
5478 | { | |
5479 | if (len < sizeof(cpumask_t)) { | |
5480 | memset(new_mask, 0, sizeof(cpumask_t)); | |
5481 | } else if (len > sizeof(cpumask_t)) { | |
5482 | len = sizeof(cpumask_t); | |
5483 | } | |
5484 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
5485 | } | |
5486 | ||
5487 | /** | |
5488 | * sys_sched_setaffinity - set the cpu affinity of a process | |
5489 | * @pid: pid of the process | |
5490 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5491 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
5492 | */ | |
5493 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
5494 | unsigned long __user *user_mask_ptr) | |
5495 | { | |
5496 | cpumask_t new_mask; | |
5497 | int retval; | |
5498 | ||
5499 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
5500 | if (retval) | |
5501 | return retval; | |
5502 | ||
b53e921b | 5503 | return sched_setaffinity(pid, &new_mask); |
1da177e4 LT |
5504 | } |
5505 | ||
5506 | /* | |
5507 | * Represents all cpu's present in the system | |
5508 | * In systems capable of hotplug, this map could dynamically grow | |
5509 | * as new cpu's are detected in the system via any platform specific | |
5510 | * method, such as ACPI for e.g. | |
5511 | */ | |
5512 | ||
4cef0c61 | 5513 | cpumask_t cpu_present_map __read_mostly; |
1da177e4 LT |
5514 | EXPORT_SYMBOL(cpu_present_map); |
5515 | ||
5516 | #ifndef CONFIG_SMP | |
4cef0c61 | 5517 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 GB |
5518 | EXPORT_SYMBOL(cpu_online_map); |
5519 | ||
4cef0c61 | 5520 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 | 5521 | EXPORT_SYMBOL(cpu_possible_map); |
1da177e4 LT |
5522 | #endif |
5523 | ||
5524 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
5525 | { | |
36c8b586 | 5526 | struct task_struct *p; |
1da177e4 | 5527 | int retval; |
1da177e4 | 5528 | |
95402b38 | 5529 | get_online_cpus(); |
1da177e4 LT |
5530 | read_lock(&tasklist_lock); |
5531 | ||
5532 | retval = -ESRCH; | |
5533 | p = find_process_by_pid(pid); | |
5534 | if (!p) | |
5535 | goto out_unlock; | |
5536 | ||
e7834f8f DQ |
5537 | retval = security_task_getscheduler(p); |
5538 | if (retval) | |
5539 | goto out_unlock; | |
5540 | ||
2f7016d9 | 5541 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
1da177e4 LT |
5542 | |
5543 | out_unlock: | |
5544 | read_unlock(&tasklist_lock); | |
95402b38 | 5545 | put_online_cpus(); |
1da177e4 | 5546 | |
9531b62f | 5547 | return retval; |
1da177e4 LT |
5548 | } |
5549 | ||
5550 | /** | |
5551 | * sys_sched_getaffinity - get the cpu affinity of a process | |
5552 | * @pid: pid of the process | |
5553 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5554 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
5555 | */ | |
5556 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
5557 | unsigned long __user *user_mask_ptr) | |
5558 | { | |
5559 | int ret; | |
5560 | cpumask_t mask; | |
5561 | ||
5562 | if (len < sizeof(cpumask_t)) | |
5563 | return -EINVAL; | |
5564 | ||
5565 | ret = sched_getaffinity(pid, &mask); | |
5566 | if (ret < 0) | |
5567 | return ret; | |
5568 | ||
5569 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
5570 | return -EFAULT; | |
5571 | ||
5572 | return sizeof(cpumask_t); | |
5573 | } | |
5574 | ||
5575 | /** | |
5576 | * sys_sched_yield - yield the current processor to other threads. | |
5577 | * | |
dd41f596 IM |
5578 | * This function yields the current CPU to other tasks. If there are no |
5579 | * other threads running on this CPU then this function will return. | |
1da177e4 LT |
5580 | */ |
5581 | asmlinkage long sys_sched_yield(void) | |
5582 | { | |
70b97a7f | 5583 | struct rq *rq = this_rq_lock(); |
1da177e4 | 5584 | |
2d72376b | 5585 | schedstat_inc(rq, yld_count); |
4530d7ab | 5586 | current->sched_class->yield_task(rq); |
1da177e4 LT |
5587 | |
5588 | /* | |
5589 | * Since we are going to call schedule() anyway, there's | |
5590 | * no need to preempt or enable interrupts: | |
5591 | */ | |
5592 | __release(rq->lock); | |
8a25d5de | 5593 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
1da177e4 LT |
5594 | _raw_spin_unlock(&rq->lock); |
5595 | preempt_enable_no_resched(); | |
5596 | ||
5597 | schedule(); | |
5598 | ||
5599 | return 0; | |
5600 | } | |
5601 | ||
e7b38404 | 5602 | static void __cond_resched(void) |
1da177e4 | 5603 | { |
8e0a43d8 IM |
5604 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
5605 | __might_sleep(__FILE__, __LINE__); | |
5606 | #endif | |
5bbcfd90 IM |
5607 | /* |
5608 | * The BKS might be reacquired before we have dropped | |
5609 | * PREEMPT_ACTIVE, which could trigger a second | |
5610 | * cond_resched() call. | |
5611 | */ | |
1da177e4 LT |
5612 | do { |
5613 | add_preempt_count(PREEMPT_ACTIVE); | |
5614 | schedule(); | |
5615 | sub_preempt_count(PREEMPT_ACTIVE); | |
5616 | } while (need_resched()); | |
5617 | } | |
5618 | ||
02b67cc3 HX |
5619 | #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY) |
5620 | int __sched _cond_resched(void) | |
1da177e4 | 5621 | { |
9414232f IM |
5622 | if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) && |
5623 | system_state == SYSTEM_RUNNING) { | |
1da177e4 LT |
5624 | __cond_resched(); |
5625 | return 1; | |
5626 | } | |
5627 | return 0; | |
5628 | } | |
02b67cc3 HX |
5629 | EXPORT_SYMBOL(_cond_resched); |
5630 | #endif | |
1da177e4 LT |
5631 | |
5632 | /* | |
5633 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
5634 | * call schedule, and on return reacquire the lock. | |
5635 | * | |
41a2d6cf | 5636 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
1da177e4 LT |
5637 | * operations here to prevent schedule() from being called twice (once via |
5638 | * spin_unlock(), once by hand). | |
5639 | */ | |
95cdf3b7 | 5640 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 5641 | { |
95c354fe | 5642 | int resched = need_resched() && system_state == SYSTEM_RUNNING; |
6df3cecb JK |
5643 | int ret = 0; |
5644 | ||
95c354fe | 5645 | if (spin_needbreak(lock) || resched) { |
1da177e4 | 5646 | spin_unlock(lock); |
95c354fe NP |
5647 | if (resched && need_resched()) |
5648 | __cond_resched(); | |
5649 | else | |
5650 | cpu_relax(); | |
6df3cecb | 5651 | ret = 1; |
1da177e4 | 5652 | spin_lock(lock); |
1da177e4 | 5653 | } |
6df3cecb | 5654 | return ret; |
1da177e4 | 5655 | } |
1da177e4 LT |
5656 | EXPORT_SYMBOL(cond_resched_lock); |
5657 | ||
5658 | int __sched cond_resched_softirq(void) | |
5659 | { | |
5660 | BUG_ON(!in_softirq()); | |
5661 | ||
9414232f | 5662 | if (need_resched() && system_state == SYSTEM_RUNNING) { |
98d82567 | 5663 | local_bh_enable(); |
1da177e4 LT |
5664 | __cond_resched(); |
5665 | local_bh_disable(); | |
5666 | return 1; | |
5667 | } | |
5668 | return 0; | |
5669 | } | |
1da177e4 LT |
5670 | EXPORT_SYMBOL(cond_resched_softirq); |
5671 | ||
1da177e4 LT |
5672 | /** |
5673 | * yield - yield the current processor to other threads. | |
5674 | * | |
72fd4a35 | 5675 | * This is a shortcut for kernel-space yielding - it marks the |
1da177e4 LT |
5676 | * thread runnable and calls sys_sched_yield(). |
5677 | */ | |
5678 | void __sched yield(void) | |
5679 | { | |
5680 | set_current_state(TASK_RUNNING); | |
5681 | sys_sched_yield(); | |
5682 | } | |
1da177e4 LT |
5683 | EXPORT_SYMBOL(yield); |
5684 | ||
5685 | /* | |
41a2d6cf | 5686 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
1da177e4 LT |
5687 | * that process accounting knows that this is a task in IO wait state. |
5688 | * | |
5689 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
5690 | * has set its backing_dev_info: the queue against which it should throttle) | |
5691 | */ | |
5692 | void __sched io_schedule(void) | |
5693 | { | |
70b97a7f | 5694 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 | 5695 | |
0ff92245 | 5696 | delayacct_blkio_start(); |
1da177e4 LT |
5697 | atomic_inc(&rq->nr_iowait); |
5698 | schedule(); | |
5699 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 5700 | delayacct_blkio_end(); |
1da177e4 | 5701 | } |
1da177e4 LT |
5702 | EXPORT_SYMBOL(io_schedule); |
5703 | ||
5704 | long __sched io_schedule_timeout(long timeout) | |
5705 | { | |
70b97a7f | 5706 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
5707 | long ret; |
5708 | ||
0ff92245 | 5709 | delayacct_blkio_start(); |
1da177e4 LT |
5710 | atomic_inc(&rq->nr_iowait); |
5711 | ret = schedule_timeout(timeout); | |
5712 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 5713 | delayacct_blkio_end(); |
1da177e4 LT |
5714 | return ret; |
5715 | } | |
5716 | ||
5717 | /** | |
5718 | * sys_sched_get_priority_max - return maximum RT priority. | |
5719 | * @policy: scheduling class. | |
5720 | * | |
5721 | * this syscall returns the maximum rt_priority that can be used | |
5722 | * by a given scheduling class. | |
5723 | */ | |
5724 | asmlinkage long sys_sched_get_priority_max(int policy) | |
5725 | { | |
5726 | int ret = -EINVAL; | |
5727 | ||
5728 | switch (policy) { | |
5729 | case SCHED_FIFO: | |
5730 | case SCHED_RR: | |
5731 | ret = MAX_USER_RT_PRIO-1; | |
5732 | break; | |
5733 | case SCHED_NORMAL: | |
b0a9499c | 5734 | case SCHED_BATCH: |
dd41f596 | 5735 | case SCHED_IDLE: |
1da177e4 LT |
5736 | ret = 0; |
5737 | break; | |
5738 | } | |
5739 | return ret; | |
5740 | } | |
5741 | ||
5742 | /** | |
5743 | * sys_sched_get_priority_min - return minimum RT priority. | |
5744 | * @policy: scheduling class. | |
5745 | * | |
5746 | * this syscall returns the minimum rt_priority that can be used | |
5747 | * by a given scheduling class. | |
5748 | */ | |
5749 | asmlinkage long sys_sched_get_priority_min(int policy) | |
5750 | { | |
5751 | int ret = -EINVAL; | |
5752 | ||
5753 | switch (policy) { | |
5754 | case SCHED_FIFO: | |
5755 | case SCHED_RR: | |
5756 | ret = 1; | |
5757 | break; | |
5758 | case SCHED_NORMAL: | |
b0a9499c | 5759 | case SCHED_BATCH: |
dd41f596 | 5760 | case SCHED_IDLE: |
1da177e4 LT |
5761 | ret = 0; |
5762 | } | |
5763 | return ret; | |
5764 | } | |
5765 | ||
5766 | /** | |
5767 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
5768 | * @pid: pid of the process. | |
5769 | * @interval: userspace pointer to the timeslice value. | |
5770 | * | |
5771 | * this syscall writes the default timeslice value of a given process | |
5772 | * into the user-space timespec buffer. A value of '0' means infinity. | |
5773 | */ | |
5774 | asmlinkage | |
5775 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
5776 | { | |
36c8b586 | 5777 | struct task_struct *p; |
a4ec24b4 | 5778 | unsigned int time_slice; |
3a5c359a | 5779 | int retval; |
1da177e4 | 5780 | struct timespec t; |
1da177e4 LT |
5781 | |
5782 | if (pid < 0) | |
3a5c359a | 5783 | return -EINVAL; |
1da177e4 LT |
5784 | |
5785 | retval = -ESRCH; | |
5786 | read_lock(&tasklist_lock); | |
5787 | p = find_process_by_pid(pid); | |
5788 | if (!p) | |
5789 | goto out_unlock; | |
5790 | ||
5791 | retval = security_task_getscheduler(p); | |
5792 | if (retval) | |
5793 | goto out_unlock; | |
5794 | ||
77034937 IM |
5795 | /* |
5796 | * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER | |
5797 | * tasks that are on an otherwise idle runqueue: | |
5798 | */ | |
5799 | time_slice = 0; | |
5800 | if (p->policy == SCHED_RR) { | |
a4ec24b4 | 5801 | time_slice = DEF_TIMESLICE; |
1868f958 | 5802 | } else if (p->policy != SCHED_FIFO) { |
a4ec24b4 DA |
5803 | struct sched_entity *se = &p->se; |
5804 | unsigned long flags; | |
5805 | struct rq *rq; | |
5806 | ||
5807 | rq = task_rq_lock(p, &flags); | |
77034937 IM |
5808 | if (rq->cfs.load.weight) |
5809 | time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
a4ec24b4 DA |
5810 | task_rq_unlock(rq, &flags); |
5811 | } | |
1da177e4 | 5812 | read_unlock(&tasklist_lock); |
a4ec24b4 | 5813 | jiffies_to_timespec(time_slice, &t); |
1da177e4 | 5814 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
1da177e4 | 5815 | return retval; |
3a5c359a | 5816 | |
1da177e4 LT |
5817 | out_unlock: |
5818 | read_unlock(&tasklist_lock); | |
5819 | return retval; | |
5820 | } | |
5821 | ||
2ed6e34f | 5822 | static const char stat_nam[] = "RSDTtZX"; |
36c8b586 | 5823 | |
82a1fcb9 | 5824 | void sched_show_task(struct task_struct *p) |
1da177e4 | 5825 | { |
1da177e4 | 5826 | unsigned long free = 0; |
36c8b586 | 5827 | unsigned state; |
1da177e4 | 5828 | |
1da177e4 | 5829 | state = p->state ? __ffs(p->state) + 1 : 0; |
cc4ea795 | 5830 | printk(KERN_INFO "%-13.13s %c", p->comm, |
2ed6e34f | 5831 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); |
4bd77321 | 5832 | #if BITS_PER_LONG == 32 |
1da177e4 | 5833 | if (state == TASK_RUNNING) |
cc4ea795 | 5834 | printk(KERN_CONT " running "); |
1da177e4 | 5835 | else |
cc4ea795 | 5836 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); |
1da177e4 LT |
5837 | #else |
5838 | if (state == TASK_RUNNING) | |
cc4ea795 | 5839 | printk(KERN_CONT " running task "); |
1da177e4 | 5840 | else |
cc4ea795 | 5841 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); |
1da177e4 LT |
5842 | #endif |
5843 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
5844 | { | |
10ebffde | 5845 | unsigned long *n = end_of_stack(p); |
1da177e4 LT |
5846 | while (!*n) |
5847 | n++; | |
10ebffde | 5848 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
1da177e4 LT |
5849 | } |
5850 | #endif | |
ba25f9dc | 5851 | printk(KERN_CONT "%5lu %5d %6d\n", free, |
fcfd50af | 5852 | task_pid_nr(p), task_pid_nr(p->real_parent)); |
1da177e4 | 5853 | |
5fb5e6de | 5854 | show_stack(p, NULL); |
1da177e4 LT |
5855 | } |
5856 | ||
e59e2ae2 | 5857 | void show_state_filter(unsigned long state_filter) |
1da177e4 | 5858 | { |
36c8b586 | 5859 | struct task_struct *g, *p; |
1da177e4 | 5860 | |
4bd77321 IM |
5861 | #if BITS_PER_LONG == 32 |
5862 | printk(KERN_INFO | |
5863 | " task PC stack pid father\n"); | |
1da177e4 | 5864 | #else |
4bd77321 IM |
5865 | printk(KERN_INFO |
5866 | " task PC stack pid father\n"); | |
1da177e4 LT |
5867 | #endif |
5868 | read_lock(&tasklist_lock); | |
5869 | do_each_thread(g, p) { | |
5870 | /* | |
5871 | * reset the NMI-timeout, listing all files on a slow | |
5872 | * console might take alot of time: | |
5873 | */ | |
5874 | touch_nmi_watchdog(); | |
39bc89fd | 5875 | if (!state_filter || (p->state & state_filter)) |
82a1fcb9 | 5876 | sched_show_task(p); |
1da177e4 LT |
5877 | } while_each_thread(g, p); |
5878 | ||
04c9167f JF |
5879 | touch_all_softlockup_watchdogs(); |
5880 | ||
dd41f596 IM |
5881 | #ifdef CONFIG_SCHED_DEBUG |
5882 | sysrq_sched_debug_show(); | |
5883 | #endif | |
1da177e4 | 5884 | read_unlock(&tasklist_lock); |
e59e2ae2 IM |
5885 | /* |
5886 | * Only show locks if all tasks are dumped: | |
5887 | */ | |
5888 | if (state_filter == -1) | |
5889 | debug_show_all_locks(); | |
1da177e4 LT |
5890 | } |
5891 | ||
1df21055 IM |
5892 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) |
5893 | { | |
dd41f596 | 5894 | idle->sched_class = &idle_sched_class; |
1df21055 IM |
5895 | } |
5896 | ||
f340c0d1 IM |
5897 | /** |
5898 | * init_idle - set up an idle thread for a given CPU | |
5899 | * @idle: task in question | |
5900 | * @cpu: cpu the idle task belongs to | |
5901 | * | |
5902 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
5903 | * flag, to make booting more robust. | |
5904 | */ | |
5c1e1767 | 5905 | void __cpuinit init_idle(struct task_struct *idle, int cpu) |
1da177e4 | 5906 | { |
70b97a7f | 5907 | struct rq *rq = cpu_rq(cpu); |
1da177e4 LT |
5908 | unsigned long flags; |
5909 | ||
dd41f596 IM |
5910 | __sched_fork(idle); |
5911 | idle->se.exec_start = sched_clock(); | |
5912 | ||
b29739f9 | 5913 | idle->prio = idle->normal_prio = MAX_PRIO; |
1da177e4 | 5914 | idle->cpus_allowed = cpumask_of_cpu(cpu); |
dd41f596 | 5915 | __set_task_cpu(idle, cpu); |
1da177e4 LT |
5916 | |
5917 | spin_lock_irqsave(&rq->lock, flags); | |
5918 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
5919 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
5920 | idle->oncpu = 1; | |
5921 | #endif | |
1da177e4 LT |
5922 | spin_unlock_irqrestore(&rq->lock, flags); |
5923 | ||
5924 | /* Set the preempt count _outside_ the spinlocks! */ | |
a1261f54 | 5925 | task_thread_info(idle)->preempt_count = 0; |
6478d880 | 5926 | |
dd41f596 IM |
5927 | /* |
5928 | * The idle tasks have their own, simple scheduling class: | |
5929 | */ | |
5930 | idle->sched_class = &idle_sched_class; | |
1da177e4 LT |
5931 | } |
5932 | ||
5933 | /* | |
5934 | * In a system that switches off the HZ timer nohz_cpu_mask | |
5935 | * indicates which cpus entered this state. This is used | |
5936 | * in the rcu update to wait only for active cpus. For system | |
5937 | * which do not switch off the HZ timer nohz_cpu_mask should | |
5938 | * always be CPU_MASK_NONE. | |
5939 | */ | |
5940 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
5941 | ||
19978ca6 IM |
5942 | /* |
5943 | * Increase the granularity value when there are more CPUs, | |
5944 | * because with more CPUs the 'effective latency' as visible | |
5945 | * to users decreases. But the relationship is not linear, | |
5946 | * so pick a second-best guess by going with the log2 of the | |
5947 | * number of CPUs. | |
5948 | * | |
5949 | * This idea comes from the SD scheduler of Con Kolivas: | |
5950 | */ | |
5951 | static inline void sched_init_granularity(void) | |
5952 | { | |
5953 | unsigned int factor = 1 + ilog2(num_online_cpus()); | |
5954 | const unsigned long limit = 200000000; | |
5955 | ||
5956 | sysctl_sched_min_granularity *= factor; | |
5957 | if (sysctl_sched_min_granularity > limit) | |
5958 | sysctl_sched_min_granularity = limit; | |
5959 | ||
5960 | sysctl_sched_latency *= factor; | |
5961 | if (sysctl_sched_latency > limit) | |
5962 | sysctl_sched_latency = limit; | |
5963 | ||
5964 | sysctl_sched_wakeup_granularity *= factor; | |
19978ca6 IM |
5965 | } |
5966 | ||
1da177e4 LT |
5967 | #ifdef CONFIG_SMP |
5968 | /* | |
5969 | * This is how migration works: | |
5970 | * | |
70b97a7f | 5971 | * 1) we queue a struct migration_req structure in the source CPU's |
1da177e4 LT |
5972 | * runqueue and wake up that CPU's migration thread. |
5973 | * 2) we down() the locked semaphore => thread blocks. | |
5974 | * 3) migration thread wakes up (implicitly it forces the migrated | |
5975 | * thread off the CPU) | |
5976 | * 4) it gets the migration request and checks whether the migrated | |
5977 | * task is still in the wrong runqueue. | |
5978 | * 5) if it's in the wrong runqueue then the migration thread removes | |
5979 | * it and puts it into the right queue. | |
5980 | * 6) migration thread up()s the semaphore. | |
5981 | * 7) we wake up and the migration is done. | |
5982 | */ | |
5983 | ||
5984 | /* | |
5985 | * Change a given task's CPU affinity. Migrate the thread to a | |
5986 | * proper CPU and schedule it away if the CPU it's executing on | |
5987 | * is removed from the allowed bitmask. | |
5988 | * | |
5989 | * NOTE: the caller must have a valid reference to the task, the | |
41a2d6cf | 5990 | * task must not exit() & deallocate itself prematurely. The |
1da177e4 LT |
5991 | * call is not atomic; no spinlocks may be held. |
5992 | */ | |
cd8ba7cd | 5993 | int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask) |
1da177e4 | 5994 | { |
70b97a7f | 5995 | struct migration_req req; |
1da177e4 | 5996 | unsigned long flags; |
70b97a7f | 5997 | struct rq *rq; |
48f24c4d | 5998 | int ret = 0; |
1da177e4 LT |
5999 | |
6000 | rq = task_rq_lock(p, &flags); | |
cd8ba7cd | 6001 | if (!cpus_intersects(*new_mask, cpu_online_map)) { |
1da177e4 LT |
6002 | ret = -EINVAL; |
6003 | goto out; | |
6004 | } | |
6005 | ||
73fe6aae | 6006 | if (p->sched_class->set_cpus_allowed) |
cd8ba7cd | 6007 | p->sched_class->set_cpus_allowed(p, new_mask); |
73fe6aae | 6008 | else { |
cd8ba7cd MT |
6009 | p->cpus_allowed = *new_mask; |
6010 | p->rt.nr_cpus_allowed = cpus_weight(*new_mask); | |
73fe6aae GH |
6011 | } |
6012 | ||
1da177e4 | 6013 | /* Can the task run on the task's current CPU? If so, we're done */ |
cd8ba7cd | 6014 | if (cpu_isset(task_cpu(p), *new_mask)) |
1da177e4 LT |
6015 | goto out; |
6016 | ||
cd8ba7cd | 6017 | if (migrate_task(p, any_online_cpu(*new_mask), &req)) { |
1da177e4 LT |
6018 | /* Need help from migration thread: drop lock and wait. */ |
6019 | task_rq_unlock(rq, &flags); | |
6020 | wake_up_process(rq->migration_thread); | |
6021 | wait_for_completion(&req.done); | |
6022 | tlb_migrate_finish(p->mm); | |
6023 | return 0; | |
6024 | } | |
6025 | out: | |
6026 | task_rq_unlock(rq, &flags); | |
48f24c4d | 6027 | |
1da177e4 LT |
6028 | return ret; |
6029 | } | |
cd8ba7cd | 6030 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); |
1da177e4 LT |
6031 | |
6032 | /* | |
41a2d6cf | 6033 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
1da177e4 LT |
6034 | * this because either it can't run here any more (set_cpus_allowed() |
6035 | * away from this CPU, or CPU going down), or because we're | |
6036 | * attempting to rebalance this task on exec (sched_exec). | |
6037 | * | |
6038 | * So we race with normal scheduler movements, but that's OK, as long | |
6039 | * as the task is no longer on this CPU. | |
efc30814 KK |
6040 | * |
6041 | * Returns non-zero if task was successfully migrated. | |
1da177e4 | 6042 | */ |
efc30814 | 6043 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
1da177e4 | 6044 | { |
70b97a7f | 6045 | struct rq *rq_dest, *rq_src; |
dd41f596 | 6046 | int ret = 0, on_rq; |
1da177e4 LT |
6047 | |
6048 | if (unlikely(cpu_is_offline(dest_cpu))) | |
efc30814 | 6049 | return ret; |
1da177e4 LT |
6050 | |
6051 | rq_src = cpu_rq(src_cpu); | |
6052 | rq_dest = cpu_rq(dest_cpu); | |
6053 | ||
6054 | double_rq_lock(rq_src, rq_dest); | |
6055 | /* Already moved. */ | |
6056 | if (task_cpu(p) != src_cpu) | |
6057 | goto out; | |
6058 | /* Affinity changed (again). */ | |
6059 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
6060 | goto out; | |
6061 | ||
dd41f596 | 6062 | on_rq = p->se.on_rq; |
6e82a3be | 6063 | if (on_rq) |
2e1cb74a | 6064 | deactivate_task(rq_src, p, 0); |
6e82a3be | 6065 | |
1da177e4 | 6066 | set_task_cpu(p, dest_cpu); |
dd41f596 IM |
6067 | if (on_rq) { |
6068 | activate_task(rq_dest, p, 0); | |
6069 | check_preempt_curr(rq_dest, p); | |
1da177e4 | 6070 | } |
efc30814 | 6071 | ret = 1; |
1da177e4 LT |
6072 | out: |
6073 | double_rq_unlock(rq_src, rq_dest); | |
efc30814 | 6074 | return ret; |
1da177e4 LT |
6075 | } |
6076 | ||
6077 | /* | |
6078 | * migration_thread - this is a highprio system thread that performs | |
6079 | * thread migration by bumping thread off CPU then 'pushing' onto | |
6080 | * another runqueue. | |
6081 | */ | |
95cdf3b7 | 6082 | static int migration_thread(void *data) |
1da177e4 | 6083 | { |
1da177e4 | 6084 | int cpu = (long)data; |
70b97a7f | 6085 | struct rq *rq; |
1da177e4 LT |
6086 | |
6087 | rq = cpu_rq(cpu); | |
6088 | BUG_ON(rq->migration_thread != current); | |
6089 | ||
6090 | set_current_state(TASK_INTERRUPTIBLE); | |
6091 | while (!kthread_should_stop()) { | |
70b97a7f | 6092 | struct migration_req *req; |
1da177e4 | 6093 | struct list_head *head; |
1da177e4 | 6094 | |
1da177e4 LT |
6095 | spin_lock_irq(&rq->lock); |
6096 | ||
6097 | if (cpu_is_offline(cpu)) { | |
6098 | spin_unlock_irq(&rq->lock); | |
6099 | goto wait_to_die; | |
6100 | } | |
6101 | ||
6102 | if (rq->active_balance) { | |
6103 | active_load_balance(rq, cpu); | |
6104 | rq->active_balance = 0; | |
6105 | } | |
6106 | ||
6107 | head = &rq->migration_queue; | |
6108 | ||
6109 | if (list_empty(head)) { | |
6110 | spin_unlock_irq(&rq->lock); | |
6111 | schedule(); | |
6112 | set_current_state(TASK_INTERRUPTIBLE); | |
6113 | continue; | |
6114 | } | |
70b97a7f | 6115 | req = list_entry(head->next, struct migration_req, list); |
1da177e4 LT |
6116 | list_del_init(head->next); |
6117 | ||
674311d5 NP |
6118 | spin_unlock(&rq->lock); |
6119 | __migrate_task(req->task, cpu, req->dest_cpu); | |
6120 | local_irq_enable(); | |
1da177e4 LT |
6121 | |
6122 | complete(&req->done); | |
6123 | } | |
6124 | __set_current_state(TASK_RUNNING); | |
6125 | return 0; | |
6126 | ||
6127 | wait_to_die: | |
6128 | /* Wait for kthread_stop */ | |
6129 | set_current_state(TASK_INTERRUPTIBLE); | |
6130 | while (!kthread_should_stop()) { | |
6131 | schedule(); | |
6132 | set_current_state(TASK_INTERRUPTIBLE); | |
6133 | } | |
6134 | __set_current_state(TASK_RUNNING); | |
6135 | return 0; | |
6136 | } | |
6137 | ||
6138 | #ifdef CONFIG_HOTPLUG_CPU | |
f7b4cddc ON |
6139 | |
6140 | static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu) | |
6141 | { | |
6142 | int ret; | |
6143 | ||
6144 | local_irq_disable(); | |
6145 | ret = __migrate_task(p, src_cpu, dest_cpu); | |
6146 | local_irq_enable(); | |
6147 | return ret; | |
6148 | } | |
6149 | ||
054b9108 | 6150 | /* |
3a4fa0a2 | 6151 | * Figure out where task on dead CPU should go, use force if necessary. |
054b9108 KK |
6152 | * NOTE: interrupts should be disabled by the caller |
6153 | */ | |
48f24c4d | 6154 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) |
1da177e4 | 6155 | { |
efc30814 | 6156 | unsigned long flags; |
1da177e4 | 6157 | cpumask_t mask; |
70b97a7f IM |
6158 | struct rq *rq; |
6159 | int dest_cpu; | |
1da177e4 | 6160 | |
3a5c359a AK |
6161 | do { |
6162 | /* On same node? */ | |
6163 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
6164 | cpus_and(mask, mask, p->cpus_allowed); | |
6165 | dest_cpu = any_online_cpu(mask); | |
6166 | ||
6167 | /* On any allowed CPU? */ | |
434d53b0 | 6168 | if (dest_cpu >= nr_cpu_ids) |
3a5c359a AK |
6169 | dest_cpu = any_online_cpu(p->cpus_allowed); |
6170 | ||
6171 | /* No more Mr. Nice Guy. */ | |
434d53b0 | 6172 | if (dest_cpu >= nr_cpu_ids) { |
f9a86fcb MT |
6173 | cpumask_t cpus_allowed; |
6174 | ||
6175 | cpuset_cpus_allowed_locked(p, &cpus_allowed); | |
470fd646 CW |
6176 | /* |
6177 | * Try to stay on the same cpuset, where the | |
6178 | * current cpuset may be a subset of all cpus. | |
6179 | * The cpuset_cpus_allowed_locked() variant of | |
41a2d6cf | 6180 | * cpuset_cpus_allowed() will not block. It must be |
470fd646 CW |
6181 | * called within calls to cpuset_lock/cpuset_unlock. |
6182 | */ | |
3a5c359a | 6183 | rq = task_rq_lock(p, &flags); |
470fd646 | 6184 | p->cpus_allowed = cpus_allowed; |
3a5c359a AK |
6185 | dest_cpu = any_online_cpu(p->cpus_allowed); |
6186 | task_rq_unlock(rq, &flags); | |
1da177e4 | 6187 | |
3a5c359a AK |
6188 | /* |
6189 | * Don't tell them about moving exiting tasks or | |
6190 | * kernel threads (both mm NULL), since they never | |
6191 | * leave kernel. | |
6192 | */ | |
41a2d6cf | 6193 | if (p->mm && printk_ratelimit()) { |
3a5c359a AK |
6194 | printk(KERN_INFO "process %d (%s) no " |
6195 | "longer affine to cpu%d\n", | |
41a2d6cf IM |
6196 | task_pid_nr(p), p->comm, dead_cpu); |
6197 | } | |
3a5c359a | 6198 | } |
f7b4cddc | 6199 | } while (!__migrate_task_irq(p, dead_cpu, dest_cpu)); |
1da177e4 LT |
6200 | } |
6201 | ||
6202 | /* | |
6203 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
6204 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
6205 | * for performance reasons the counter is not stricly tracking tasks to | |
6206 | * their home CPUs. So we just add the counter to another CPU's counter, | |
6207 | * to keep the global sum constant after CPU-down: | |
6208 | */ | |
70b97a7f | 6209 | static void migrate_nr_uninterruptible(struct rq *rq_src) |
1da177e4 | 6210 | { |
7c16ec58 | 6211 | struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR)); |
1da177e4 LT |
6212 | unsigned long flags; |
6213 | ||
6214 | local_irq_save(flags); | |
6215 | double_rq_lock(rq_src, rq_dest); | |
6216 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
6217 | rq_src->nr_uninterruptible = 0; | |
6218 | double_rq_unlock(rq_src, rq_dest); | |
6219 | local_irq_restore(flags); | |
6220 | } | |
6221 | ||
6222 | /* Run through task list and migrate tasks from the dead cpu. */ | |
6223 | static void migrate_live_tasks(int src_cpu) | |
6224 | { | |
48f24c4d | 6225 | struct task_struct *p, *t; |
1da177e4 | 6226 | |
f7b4cddc | 6227 | read_lock(&tasklist_lock); |
1da177e4 | 6228 | |
48f24c4d IM |
6229 | do_each_thread(t, p) { |
6230 | if (p == current) | |
1da177e4 LT |
6231 | continue; |
6232 | ||
48f24c4d IM |
6233 | if (task_cpu(p) == src_cpu) |
6234 | move_task_off_dead_cpu(src_cpu, p); | |
6235 | } while_each_thread(t, p); | |
1da177e4 | 6236 | |
f7b4cddc | 6237 | read_unlock(&tasklist_lock); |
1da177e4 LT |
6238 | } |
6239 | ||
dd41f596 IM |
6240 | /* |
6241 | * Schedules idle task to be the next runnable task on current CPU. | |
94bc9a7b DA |
6242 | * It does so by boosting its priority to highest possible. |
6243 | * Used by CPU offline code. | |
1da177e4 LT |
6244 | */ |
6245 | void sched_idle_next(void) | |
6246 | { | |
48f24c4d | 6247 | int this_cpu = smp_processor_id(); |
70b97a7f | 6248 | struct rq *rq = cpu_rq(this_cpu); |
1da177e4 LT |
6249 | struct task_struct *p = rq->idle; |
6250 | unsigned long flags; | |
6251 | ||
6252 | /* cpu has to be offline */ | |
48f24c4d | 6253 | BUG_ON(cpu_online(this_cpu)); |
1da177e4 | 6254 | |
48f24c4d IM |
6255 | /* |
6256 | * Strictly not necessary since rest of the CPUs are stopped by now | |
6257 | * and interrupts disabled on the current cpu. | |
1da177e4 LT |
6258 | */ |
6259 | spin_lock_irqsave(&rq->lock, flags); | |
6260 | ||
dd41f596 | 6261 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
48f24c4d | 6262 | |
94bc9a7b DA |
6263 | update_rq_clock(rq); |
6264 | activate_task(rq, p, 0); | |
1da177e4 LT |
6265 | |
6266 | spin_unlock_irqrestore(&rq->lock, flags); | |
6267 | } | |
6268 | ||
48f24c4d IM |
6269 | /* |
6270 | * Ensures that the idle task is using init_mm right before its cpu goes | |
1da177e4 LT |
6271 | * offline. |
6272 | */ | |
6273 | void idle_task_exit(void) | |
6274 | { | |
6275 | struct mm_struct *mm = current->active_mm; | |
6276 | ||
6277 | BUG_ON(cpu_online(smp_processor_id())); | |
6278 | ||
6279 | if (mm != &init_mm) | |
6280 | switch_mm(mm, &init_mm, current); | |
6281 | mmdrop(mm); | |
6282 | } | |
6283 | ||
054b9108 | 6284 | /* called under rq->lock with disabled interrupts */ |
36c8b586 | 6285 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) |
1da177e4 | 6286 | { |
70b97a7f | 6287 | struct rq *rq = cpu_rq(dead_cpu); |
1da177e4 LT |
6288 | |
6289 | /* Must be exiting, otherwise would be on tasklist. */ | |
270f722d | 6290 | BUG_ON(!p->exit_state); |
1da177e4 LT |
6291 | |
6292 | /* Cannot have done final schedule yet: would have vanished. */ | |
c394cc9f | 6293 | BUG_ON(p->state == TASK_DEAD); |
1da177e4 | 6294 | |
48f24c4d | 6295 | get_task_struct(p); |
1da177e4 LT |
6296 | |
6297 | /* | |
6298 | * Drop lock around migration; if someone else moves it, | |
41a2d6cf | 6299 | * that's OK. No task can be added to this CPU, so iteration is |
1da177e4 LT |
6300 | * fine. |
6301 | */ | |
f7b4cddc | 6302 | spin_unlock_irq(&rq->lock); |
48f24c4d | 6303 | move_task_off_dead_cpu(dead_cpu, p); |
f7b4cddc | 6304 | spin_lock_irq(&rq->lock); |
1da177e4 | 6305 | |
48f24c4d | 6306 | put_task_struct(p); |
1da177e4 LT |
6307 | } |
6308 | ||
6309 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
6310 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
6311 | { | |
70b97a7f | 6312 | struct rq *rq = cpu_rq(dead_cpu); |
dd41f596 | 6313 | struct task_struct *next; |
48f24c4d | 6314 | |
dd41f596 IM |
6315 | for ( ; ; ) { |
6316 | if (!rq->nr_running) | |
6317 | break; | |
a8e504d2 | 6318 | update_rq_clock(rq); |
ff95f3df | 6319 | next = pick_next_task(rq, rq->curr); |
dd41f596 IM |
6320 | if (!next) |
6321 | break; | |
6322 | migrate_dead(dead_cpu, next); | |
e692ab53 | 6323 | |
1da177e4 LT |
6324 | } |
6325 | } | |
6326 | #endif /* CONFIG_HOTPLUG_CPU */ | |
6327 | ||
e692ab53 NP |
6328 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) |
6329 | ||
6330 | static struct ctl_table sd_ctl_dir[] = { | |
e0361851 AD |
6331 | { |
6332 | .procname = "sched_domain", | |
c57baf1e | 6333 | .mode = 0555, |
e0361851 | 6334 | }, |
38605cae | 6335 | {0, }, |
e692ab53 NP |
6336 | }; |
6337 | ||
6338 | static struct ctl_table sd_ctl_root[] = { | |
e0361851 | 6339 | { |
c57baf1e | 6340 | .ctl_name = CTL_KERN, |
e0361851 | 6341 | .procname = "kernel", |
c57baf1e | 6342 | .mode = 0555, |
e0361851 AD |
6343 | .child = sd_ctl_dir, |
6344 | }, | |
38605cae | 6345 | {0, }, |
e692ab53 NP |
6346 | }; |
6347 | ||
6348 | static struct ctl_table *sd_alloc_ctl_entry(int n) | |
6349 | { | |
6350 | struct ctl_table *entry = | |
5cf9f062 | 6351 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); |
e692ab53 | 6352 | |
e692ab53 NP |
6353 | return entry; |
6354 | } | |
6355 | ||
6382bc90 MM |
6356 | static void sd_free_ctl_entry(struct ctl_table **tablep) |
6357 | { | |
cd790076 | 6358 | struct ctl_table *entry; |
6382bc90 | 6359 | |
cd790076 MM |
6360 | /* |
6361 | * In the intermediate directories, both the child directory and | |
6362 | * procname are dynamically allocated and could fail but the mode | |
41a2d6cf | 6363 | * will always be set. In the lowest directory the names are |
cd790076 MM |
6364 | * static strings and all have proc handlers. |
6365 | */ | |
6366 | for (entry = *tablep; entry->mode; entry++) { | |
6382bc90 MM |
6367 | if (entry->child) |
6368 | sd_free_ctl_entry(&entry->child); | |
cd790076 MM |
6369 | if (entry->proc_handler == NULL) |
6370 | kfree(entry->procname); | |
6371 | } | |
6382bc90 MM |
6372 | |
6373 | kfree(*tablep); | |
6374 | *tablep = NULL; | |
6375 | } | |
6376 | ||
e692ab53 | 6377 | static void |
e0361851 | 6378 | set_table_entry(struct ctl_table *entry, |
e692ab53 NP |
6379 | const char *procname, void *data, int maxlen, |
6380 | mode_t mode, proc_handler *proc_handler) | |
6381 | { | |
e692ab53 NP |
6382 | entry->procname = procname; |
6383 | entry->data = data; | |
6384 | entry->maxlen = maxlen; | |
6385 | entry->mode = mode; | |
6386 | entry->proc_handler = proc_handler; | |
6387 | } | |
6388 | ||
6389 | static struct ctl_table * | |
6390 | sd_alloc_ctl_domain_table(struct sched_domain *sd) | |
6391 | { | |
ace8b3d6 | 6392 | struct ctl_table *table = sd_alloc_ctl_entry(12); |
e692ab53 | 6393 | |
ad1cdc1d MM |
6394 | if (table == NULL) |
6395 | return NULL; | |
6396 | ||
e0361851 | 6397 | set_table_entry(&table[0], "min_interval", &sd->min_interval, |
e692ab53 | 6398 | sizeof(long), 0644, proc_doulongvec_minmax); |
e0361851 | 6399 | set_table_entry(&table[1], "max_interval", &sd->max_interval, |
e692ab53 | 6400 | sizeof(long), 0644, proc_doulongvec_minmax); |
e0361851 | 6401 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, |
e692ab53 | 6402 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6403 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, |
e692ab53 | 6404 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6405 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, |
e692ab53 | 6406 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6407 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, |
e692ab53 | 6408 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6409 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, |
e692ab53 | 6410 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6411 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, |
e692ab53 | 6412 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6413 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, |
e692ab53 | 6414 | sizeof(int), 0644, proc_dointvec_minmax); |
ace8b3d6 | 6415 | set_table_entry(&table[9], "cache_nice_tries", |
e692ab53 NP |
6416 | &sd->cache_nice_tries, |
6417 | sizeof(int), 0644, proc_dointvec_minmax); | |
ace8b3d6 | 6418 | set_table_entry(&table[10], "flags", &sd->flags, |
e692ab53 | 6419 | sizeof(int), 0644, proc_dointvec_minmax); |
6323469f | 6420 | /* &table[11] is terminator */ |
e692ab53 NP |
6421 | |
6422 | return table; | |
6423 | } | |
6424 | ||
9a4e7159 | 6425 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) |
e692ab53 NP |
6426 | { |
6427 | struct ctl_table *entry, *table; | |
6428 | struct sched_domain *sd; | |
6429 | int domain_num = 0, i; | |
6430 | char buf[32]; | |
6431 | ||
6432 | for_each_domain(cpu, sd) | |
6433 | domain_num++; | |
6434 | entry = table = sd_alloc_ctl_entry(domain_num + 1); | |
ad1cdc1d MM |
6435 | if (table == NULL) |
6436 | return NULL; | |
e692ab53 NP |
6437 | |
6438 | i = 0; | |
6439 | for_each_domain(cpu, sd) { | |
6440 | snprintf(buf, 32, "domain%d", i); | |
e692ab53 | 6441 | entry->procname = kstrdup(buf, GFP_KERNEL); |
c57baf1e | 6442 | entry->mode = 0555; |
e692ab53 NP |
6443 | entry->child = sd_alloc_ctl_domain_table(sd); |
6444 | entry++; | |
6445 | i++; | |
6446 | } | |
6447 | return table; | |
6448 | } | |
6449 | ||
6450 | static struct ctl_table_header *sd_sysctl_header; | |
6382bc90 | 6451 | static void register_sched_domain_sysctl(void) |
e692ab53 NP |
6452 | { |
6453 | int i, cpu_num = num_online_cpus(); | |
6454 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); | |
6455 | char buf[32]; | |
6456 | ||
7378547f MM |
6457 | WARN_ON(sd_ctl_dir[0].child); |
6458 | sd_ctl_dir[0].child = entry; | |
6459 | ||
ad1cdc1d MM |
6460 | if (entry == NULL) |
6461 | return; | |
6462 | ||
97b6ea7b | 6463 | for_each_online_cpu(i) { |
e692ab53 | 6464 | snprintf(buf, 32, "cpu%d", i); |
e692ab53 | 6465 | entry->procname = kstrdup(buf, GFP_KERNEL); |
c57baf1e | 6466 | entry->mode = 0555; |
e692ab53 | 6467 | entry->child = sd_alloc_ctl_cpu_table(i); |
97b6ea7b | 6468 | entry++; |
e692ab53 | 6469 | } |
7378547f MM |
6470 | |
6471 | WARN_ON(sd_sysctl_header); | |
e692ab53 NP |
6472 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); |
6473 | } | |
6382bc90 | 6474 | |
7378547f | 6475 | /* may be called multiple times per register */ |
6382bc90 MM |
6476 | static void unregister_sched_domain_sysctl(void) |
6477 | { | |
7378547f MM |
6478 | if (sd_sysctl_header) |
6479 | unregister_sysctl_table(sd_sysctl_header); | |
6382bc90 | 6480 | sd_sysctl_header = NULL; |
7378547f MM |
6481 | if (sd_ctl_dir[0].child) |
6482 | sd_free_ctl_entry(&sd_ctl_dir[0].child); | |
6382bc90 | 6483 | } |
e692ab53 | 6484 | #else |
6382bc90 MM |
6485 | static void register_sched_domain_sysctl(void) |
6486 | { | |
6487 | } | |
6488 | static void unregister_sched_domain_sysctl(void) | |
e692ab53 NP |
6489 | { |
6490 | } | |
6491 | #endif | |
6492 | ||
1da177e4 LT |
6493 | /* |
6494 | * migration_call - callback that gets triggered when a CPU is added. | |
6495 | * Here we can start up the necessary migration thread for the new CPU. | |
6496 | */ | |
48f24c4d IM |
6497 | static int __cpuinit |
6498 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
1da177e4 | 6499 | { |
1da177e4 | 6500 | struct task_struct *p; |
48f24c4d | 6501 | int cpu = (long)hcpu; |
1da177e4 | 6502 | unsigned long flags; |
70b97a7f | 6503 | struct rq *rq; |
1da177e4 LT |
6504 | |
6505 | switch (action) { | |
5be9361c | 6506 | |
1da177e4 | 6507 | case CPU_UP_PREPARE: |
8bb78442 | 6508 | case CPU_UP_PREPARE_FROZEN: |
dd41f596 | 6509 | p = kthread_create(migration_thread, hcpu, "migration/%d", cpu); |
1da177e4 LT |
6510 | if (IS_ERR(p)) |
6511 | return NOTIFY_BAD; | |
1da177e4 LT |
6512 | kthread_bind(p, cpu); |
6513 | /* Must be high prio: stop_machine expects to yield to it. */ | |
6514 | rq = task_rq_lock(p, &flags); | |
dd41f596 | 6515 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
1da177e4 LT |
6516 | task_rq_unlock(rq, &flags); |
6517 | cpu_rq(cpu)->migration_thread = p; | |
6518 | break; | |
48f24c4d | 6519 | |
1da177e4 | 6520 | case CPU_ONLINE: |
8bb78442 | 6521 | case CPU_ONLINE_FROZEN: |
3a4fa0a2 | 6522 | /* Strictly unnecessary, as first user will wake it. */ |
1da177e4 | 6523 | wake_up_process(cpu_rq(cpu)->migration_thread); |
1f94ef59 GH |
6524 | |
6525 | /* Update our root-domain */ | |
6526 | rq = cpu_rq(cpu); | |
6527 | spin_lock_irqsave(&rq->lock, flags); | |
6528 | if (rq->rd) { | |
6529 | BUG_ON(!cpu_isset(cpu, rq->rd->span)); | |
6530 | cpu_set(cpu, rq->rd->online); | |
6531 | } | |
6532 | spin_unlock_irqrestore(&rq->lock, flags); | |
1da177e4 | 6533 | break; |
48f24c4d | 6534 | |
1da177e4 LT |
6535 | #ifdef CONFIG_HOTPLUG_CPU |
6536 | case CPU_UP_CANCELED: | |
8bb78442 | 6537 | case CPU_UP_CANCELED_FROZEN: |
fc75cdfa HC |
6538 | if (!cpu_rq(cpu)->migration_thread) |
6539 | break; | |
41a2d6cf | 6540 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
a4c4af7c HC |
6541 | kthread_bind(cpu_rq(cpu)->migration_thread, |
6542 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
6543 | kthread_stop(cpu_rq(cpu)->migration_thread); |
6544 | cpu_rq(cpu)->migration_thread = NULL; | |
6545 | break; | |
48f24c4d | 6546 | |
1da177e4 | 6547 | case CPU_DEAD: |
8bb78442 | 6548 | case CPU_DEAD_FROZEN: |
470fd646 | 6549 | cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */ |
1da177e4 LT |
6550 | migrate_live_tasks(cpu); |
6551 | rq = cpu_rq(cpu); | |
6552 | kthread_stop(rq->migration_thread); | |
6553 | rq->migration_thread = NULL; | |
6554 | /* Idle task back to normal (off runqueue, low prio) */ | |
d2da272a | 6555 | spin_lock_irq(&rq->lock); |
a8e504d2 | 6556 | update_rq_clock(rq); |
2e1cb74a | 6557 | deactivate_task(rq, rq->idle, 0); |
1da177e4 | 6558 | rq->idle->static_prio = MAX_PRIO; |
dd41f596 IM |
6559 | __setscheduler(rq, rq->idle, SCHED_NORMAL, 0); |
6560 | rq->idle->sched_class = &idle_sched_class; | |
1da177e4 | 6561 | migrate_dead_tasks(cpu); |
d2da272a | 6562 | spin_unlock_irq(&rq->lock); |
470fd646 | 6563 | cpuset_unlock(); |
1da177e4 LT |
6564 | migrate_nr_uninterruptible(rq); |
6565 | BUG_ON(rq->nr_running != 0); | |
6566 | ||
41a2d6cf IM |
6567 | /* |
6568 | * No need to migrate the tasks: it was best-effort if | |
6569 | * they didn't take sched_hotcpu_mutex. Just wake up | |
6570 | * the requestors. | |
6571 | */ | |
1da177e4 LT |
6572 | spin_lock_irq(&rq->lock); |
6573 | while (!list_empty(&rq->migration_queue)) { | |
70b97a7f IM |
6574 | struct migration_req *req; |
6575 | ||
1da177e4 | 6576 | req = list_entry(rq->migration_queue.next, |
70b97a7f | 6577 | struct migration_req, list); |
1da177e4 LT |
6578 | list_del_init(&req->list); |
6579 | complete(&req->done); | |
6580 | } | |
6581 | spin_unlock_irq(&rq->lock); | |
6582 | break; | |
57d885fe | 6583 | |
08f503b0 GH |
6584 | case CPU_DYING: |
6585 | case CPU_DYING_FROZEN: | |
57d885fe GH |
6586 | /* Update our root-domain */ |
6587 | rq = cpu_rq(cpu); | |
6588 | spin_lock_irqsave(&rq->lock, flags); | |
6589 | if (rq->rd) { | |
6590 | BUG_ON(!cpu_isset(cpu, rq->rd->span)); | |
6591 | cpu_clear(cpu, rq->rd->online); | |
6592 | } | |
6593 | spin_unlock_irqrestore(&rq->lock, flags); | |
6594 | break; | |
1da177e4 LT |
6595 | #endif |
6596 | } | |
6597 | return NOTIFY_OK; | |
6598 | } | |
6599 | ||
6600 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
6601 | * happens before everything else. | |
6602 | */ | |
26c2143b | 6603 | static struct notifier_block __cpuinitdata migration_notifier = { |
1da177e4 LT |
6604 | .notifier_call = migration_call, |
6605 | .priority = 10 | |
6606 | }; | |
6607 | ||
e6fe6649 | 6608 | void __init migration_init(void) |
1da177e4 LT |
6609 | { |
6610 | void *cpu = (void *)(long)smp_processor_id(); | |
07dccf33 | 6611 | int err; |
48f24c4d IM |
6612 | |
6613 | /* Start one for the boot CPU: */ | |
07dccf33 AM |
6614 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
6615 | BUG_ON(err == NOTIFY_BAD); | |
1da177e4 LT |
6616 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
6617 | register_cpu_notifier(&migration_notifier); | |
1da177e4 LT |
6618 | } |
6619 | #endif | |
6620 | ||
6621 | #ifdef CONFIG_SMP | |
476f3534 | 6622 | |
3e9830dc | 6623 | #ifdef CONFIG_SCHED_DEBUG |
4dcf6aff | 6624 | |
7c16ec58 MT |
6625 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, |
6626 | cpumask_t *groupmask) | |
1da177e4 | 6627 | { |
4dcf6aff | 6628 | struct sched_group *group = sd->groups; |
434d53b0 | 6629 | char str[256]; |
1da177e4 | 6630 | |
434d53b0 | 6631 | cpulist_scnprintf(str, sizeof(str), sd->span); |
7c16ec58 | 6632 | cpus_clear(*groupmask); |
4dcf6aff IM |
6633 | |
6634 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); | |
6635 | ||
6636 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
6637 | printk("does not load-balance\n"); | |
6638 | if (sd->parent) | |
6639 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" | |
6640 | " has parent"); | |
6641 | return -1; | |
41c7ce9a NP |
6642 | } |
6643 | ||
4dcf6aff IM |
6644 | printk(KERN_CONT "span %s\n", str); |
6645 | ||
6646 | if (!cpu_isset(cpu, sd->span)) { | |
6647 | printk(KERN_ERR "ERROR: domain->span does not contain " | |
6648 | "CPU%d\n", cpu); | |
6649 | } | |
6650 | if (!cpu_isset(cpu, group->cpumask)) { | |
6651 | printk(KERN_ERR "ERROR: domain->groups does not contain" | |
6652 | " CPU%d\n", cpu); | |
6653 | } | |
1da177e4 | 6654 | |
4dcf6aff | 6655 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); |
1da177e4 | 6656 | do { |
4dcf6aff IM |
6657 | if (!group) { |
6658 | printk("\n"); | |
6659 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
1da177e4 LT |
6660 | break; |
6661 | } | |
6662 | ||
4dcf6aff IM |
6663 | if (!group->__cpu_power) { |
6664 | printk(KERN_CONT "\n"); | |
6665 | printk(KERN_ERR "ERROR: domain->cpu_power not " | |
6666 | "set\n"); | |
6667 | break; | |
6668 | } | |
1da177e4 | 6669 | |
4dcf6aff IM |
6670 | if (!cpus_weight(group->cpumask)) { |
6671 | printk(KERN_CONT "\n"); | |
6672 | printk(KERN_ERR "ERROR: empty group\n"); | |
6673 | break; | |
6674 | } | |
1da177e4 | 6675 | |
7c16ec58 | 6676 | if (cpus_intersects(*groupmask, group->cpumask)) { |
4dcf6aff IM |
6677 | printk(KERN_CONT "\n"); |
6678 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
6679 | break; | |
6680 | } | |
1da177e4 | 6681 | |
7c16ec58 | 6682 | cpus_or(*groupmask, *groupmask, group->cpumask); |
1da177e4 | 6683 | |
434d53b0 | 6684 | cpulist_scnprintf(str, sizeof(str), group->cpumask); |
4dcf6aff | 6685 | printk(KERN_CONT " %s", str); |
1da177e4 | 6686 | |
4dcf6aff IM |
6687 | group = group->next; |
6688 | } while (group != sd->groups); | |
6689 | printk(KERN_CONT "\n"); | |
1da177e4 | 6690 | |
7c16ec58 | 6691 | if (!cpus_equal(sd->span, *groupmask)) |
4dcf6aff | 6692 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
1da177e4 | 6693 | |
7c16ec58 | 6694 | if (sd->parent && !cpus_subset(*groupmask, sd->parent->span)) |
4dcf6aff IM |
6695 | printk(KERN_ERR "ERROR: parent span is not a superset " |
6696 | "of domain->span\n"); | |
6697 | return 0; | |
6698 | } | |
1da177e4 | 6699 | |
4dcf6aff IM |
6700 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
6701 | { | |
7c16ec58 | 6702 | cpumask_t *groupmask; |
4dcf6aff | 6703 | int level = 0; |
1da177e4 | 6704 | |
4dcf6aff IM |
6705 | if (!sd) { |
6706 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
6707 | return; | |
6708 | } | |
1da177e4 | 6709 | |
4dcf6aff IM |
6710 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
6711 | ||
7c16ec58 MT |
6712 | groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL); |
6713 | if (!groupmask) { | |
6714 | printk(KERN_DEBUG "Cannot load-balance (out of memory)\n"); | |
6715 | return; | |
6716 | } | |
6717 | ||
4dcf6aff | 6718 | for (;;) { |
7c16ec58 | 6719 | if (sched_domain_debug_one(sd, cpu, level, groupmask)) |
4dcf6aff | 6720 | break; |
1da177e4 LT |
6721 | level++; |
6722 | sd = sd->parent; | |
33859f7f | 6723 | if (!sd) |
4dcf6aff IM |
6724 | break; |
6725 | } | |
7c16ec58 | 6726 | kfree(groupmask); |
1da177e4 LT |
6727 | } |
6728 | #else | |
48f24c4d | 6729 | # define sched_domain_debug(sd, cpu) do { } while (0) |
1da177e4 LT |
6730 | #endif |
6731 | ||
1a20ff27 | 6732 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
6733 | { |
6734 | if (cpus_weight(sd->span) == 1) | |
6735 | return 1; | |
6736 | ||
6737 | /* Following flags need at least 2 groups */ | |
6738 | if (sd->flags & (SD_LOAD_BALANCE | | |
6739 | SD_BALANCE_NEWIDLE | | |
6740 | SD_BALANCE_FORK | | |
89c4710e SS |
6741 | SD_BALANCE_EXEC | |
6742 | SD_SHARE_CPUPOWER | | |
6743 | SD_SHARE_PKG_RESOURCES)) { | |
245af2c7 SS |
6744 | if (sd->groups != sd->groups->next) |
6745 | return 0; | |
6746 | } | |
6747 | ||
6748 | /* Following flags don't use groups */ | |
6749 | if (sd->flags & (SD_WAKE_IDLE | | |
6750 | SD_WAKE_AFFINE | | |
6751 | SD_WAKE_BALANCE)) | |
6752 | return 0; | |
6753 | ||
6754 | return 1; | |
6755 | } | |
6756 | ||
48f24c4d IM |
6757 | static int |
6758 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
245af2c7 SS |
6759 | { |
6760 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
6761 | ||
6762 | if (sd_degenerate(parent)) | |
6763 | return 1; | |
6764 | ||
6765 | if (!cpus_equal(sd->span, parent->span)) | |
6766 | return 0; | |
6767 | ||
6768 | /* Does parent contain flags not in child? */ | |
6769 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
6770 | if (cflags & SD_WAKE_AFFINE) | |
6771 | pflags &= ~SD_WAKE_BALANCE; | |
6772 | /* Flags needing groups don't count if only 1 group in parent */ | |
6773 | if (parent->groups == parent->groups->next) { | |
6774 | pflags &= ~(SD_LOAD_BALANCE | | |
6775 | SD_BALANCE_NEWIDLE | | |
6776 | SD_BALANCE_FORK | | |
89c4710e SS |
6777 | SD_BALANCE_EXEC | |
6778 | SD_SHARE_CPUPOWER | | |
6779 | SD_SHARE_PKG_RESOURCES); | |
245af2c7 SS |
6780 | } |
6781 | if (~cflags & pflags) | |
6782 | return 0; | |
6783 | ||
6784 | return 1; | |
6785 | } | |
6786 | ||
57d885fe GH |
6787 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) |
6788 | { | |
6789 | unsigned long flags; | |
6790 | const struct sched_class *class; | |
6791 | ||
6792 | spin_lock_irqsave(&rq->lock, flags); | |
6793 | ||
6794 | if (rq->rd) { | |
6795 | struct root_domain *old_rd = rq->rd; | |
6796 | ||
0eab9146 | 6797 | for (class = sched_class_highest; class; class = class->next) { |
57d885fe GH |
6798 | if (class->leave_domain) |
6799 | class->leave_domain(rq); | |
0eab9146 | 6800 | } |
57d885fe | 6801 | |
dc938520 GH |
6802 | cpu_clear(rq->cpu, old_rd->span); |
6803 | cpu_clear(rq->cpu, old_rd->online); | |
6804 | ||
57d885fe GH |
6805 | if (atomic_dec_and_test(&old_rd->refcount)) |
6806 | kfree(old_rd); | |
6807 | } | |
6808 | ||
6809 | atomic_inc(&rd->refcount); | |
6810 | rq->rd = rd; | |
6811 | ||
dc938520 | 6812 | cpu_set(rq->cpu, rd->span); |
1f94ef59 GH |
6813 | if (cpu_isset(rq->cpu, cpu_online_map)) |
6814 | cpu_set(rq->cpu, rd->online); | |
dc938520 | 6815 | |
0eab9146 | 6816 | for (class = sched_class_highest; class; class = class->next) { |
57d885fe GH |
6817 | if (class->join_domain) |
6818 | class->join_domain(rq); | |
0eab9146 | 6819 | } |
57d885fe GH |
6820 | |
6821 | spin_unlock_irqrestore(&rq->lock, flags); | |
6822 | } | |
6823 | ||
dc938520 | 6824 | static void init_rootdomain(struct root_domain *rd) |
57d885fe GH |
6825 | { |
6826 | memset(rd, 0, sizeof(*rd)); | |
6827 | ||
dc938520 GH |
6828 | cpus_clear(rd->span); |
6829 | cpus_clear(rd->online); | |
57d885fe GH |
6830 | } |
6831 | ||
6832 | static void init_defrootdomain(void) | |
6833 | { | |
dc938520 | 6834 | init_rootdomain(&def_root_domain); |
57d885fe GH |
6835 | atomic_set(&def_root_domain.refcount, 1); |
6836 | } | |
6837 | ||
dc938520 | 6838 | static struct root_domain *alloc_rootdomain(void) |
57d885fe GH |
6839 | { |
6840 | struct root_domain *rd; | |
6841 | ||
6842 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); | |
6843 | if (!rd) | |
6844 | return NULL; | |
6845 | ||
dc938520 | 6846 | init_rootdomain(rd); |
57d885fe GH |
6847 | |
6848 | return rd; | |
6849 | } | |
6850 | ||
1da177e4 | 6851 | /* |
0eab9146 | 6852 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
1da177e4 LT |
6853 | * hold the hotplug lock. |
6854 | */ | |
0eab9146 IM |
6855 | static void |
6856 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) | |
1da177e4 | 6857 | { |
70b97a7f | 6858 | struct rq *rq = cpu_rq(cpu); |
245af2c7 SS |
6859 | struct sched_domain *tmp; |
6860 | ||
6861 | /* Remove the sched domains which do not contribute to scheduling. */ | |
6862 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
6863 | struct sched_domain *parent = tmp->parent; | |
6864 | if (!parent) | |
6865 | break; | |
1a848870 | 6866 | if (sd_parent_degenerate(tmp, parent)) { |
245af2c7 | 6867 | tmp->parent = parent->parent; |
1a848870 SS |
6868 | if (parent->parent) |
6869 | parent->parent->child = tmp; | |
6870 | } | |
245af2c7 SS |
6871 | } |
6872 | ||
1a848870 | 6873 | if (sd && sd_degenerate(sd)) { |
245af2c7 | 6874 | sd = sd->parent; |
1a848870 SS |
6875 | if (sd) |
6876 | sd->child = NULL; | |
6877 | } | |
1da177e4 LT |
6878 | |
6879 | sched_domain_debug(sd, cpu); | |
6880 | ||
57d885fe | 6881 | rq_attach_root(rq, rd); |
674311d5 | 6882 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
6883 | } |
6884 | ||
6885 | /* cpus with isolated domains */ | |
67af63a6 | 6886 | static cpumask_t cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
6887 | |
6888 | /* Setup the mask of cpus configured for isolated domains */ | |
6889 | static int __init isolated_cpu_setup(char *str) | |
6890 | { | |
6891 | int ints[NR_CPUS], i; | |
6892 | ||
6893 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
6894 | cpus_clear(cpu_isolated_map); | |
6895 | for (i = 1; i <= ints[0]; i++) | |
6896 | if (ints[i] < NR_CPUS) | |
6897 | cpu_set(ints[i], cpu_isolated_map); | |
6898 | return 1; | |
6899 | } | |
6900 | ||
8927f494 | 6901 | __setup("isolcpus=", isolated_cpu_setup); |
1da177e4 LT |
6902 | |
6903 | /* | |
6711cab4 SS |
6904 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer |
6905 | * to a function which identifies what group(along with sched group) a CPU | |
6906 | * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS | |
6907 | * (due to the fact that we keep track of groups covered with a cpumask_t). | |
1da177e4 LT |
6908 | * |
6909 | * init_sched_build_groups will build a circular linked list of the groups | |
6910 | * covered by the given span, and will set each group's ->cpumask correctly, | |
6911 | * and ->cpu_power to 0. | |
6912 | */ | |
a616058b | 6913 | static void |
7c16ec58 | 6914 | init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map, |
6711cab4 | 6915 | int (*group_fn)(int cpu, const cpumask_t *cpu_map, |
7c16ec58 MT |
6916 | struct sched_group **sg, |
6917 | cpumask_t *tmpmask), | |
6918 | cpumask_t *covered, cpumask_t *tmpmask) | |
1da177e4 LT |
6919 | { |
6920 | struct sched_group *first = NULL, *last = NULL; | |
1da177e4 LT |
6921 | int i; |
6922 | ||
7c16ec58 MT |
6923 | cpus_clear(*covered); |
6924 | ||
6925 | for_each_cpu_mask(i, *span) { | |
6711cab4 | 6926 | struct sched_group *sg; |
7c16ec58 | 6927 | int group = group_fn(i, cpu_map, &sg, tmpmask); |
1da177e4 LT |
6928 | int j; |
6929 | ||
7c16ec58 | 6930 | if (cpu_isset(i, *covered)) |
1da177e4 LT |
6931 | continue; |
6932 | ||
7c16ec58 | 6933 | cpus_clear(sg->cpumask); |
5517d86b | 6934 | sg->__cpu_power = 0; |
1da177e4 | 6935 | |
7c16ec58 MT |
6936 | for_each_cpu_mask(j, *span) { |
6937 | if (group_fn(j, cpu_map, NULL, tmpmask) != group) | |
1da177e4 LT |
6938 | continue; |
6939 | ||
7c16ec58 | 6940 | cpu_set(j, *covered); |
1da177e4 LT |
6941 | cpu_set(j, sg->cpumask); |
6942 | } | |
6943 | if (!first) | |
6944 | first = sg; | |
6945 | if (last) | |
6946 | last->next = sg; | |
6947 | last = sg; | |
6948 | } | |
6949 | last->next = first; | |
6950 | } | |
6951 | ||
9c1cfda2 | 6952 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 6953 | |
9c1cfda2 | 6954 | #ifdef CONFIG_NUMA |
198e2f18 | 6955 | |
9c1cfda2 JH |
6956 | /** |
6957 | * find_next_best_node - find the next node to include in a sched_domain | |
6958 | * @node: node whose sched_domain we're building | |
6959 | * @used_nodes: nodes already in the sched_domain | |
6960 | * | |
41a2d6cf | 6961 | * Find the next node to include in a given scheduling domain. Simply |
9c1cfda2 JH |
6962 | * finds the closest node not already in the @used_nodes map. |
6963 | * | |
6964 | * Should use nodemask_t. | |
6965 | */ | |
c5f59f08 | 6966 | static int find_next_best_node(int node, nodemask_t *used_nodes) |
9c1cfda2 JH |
6967 | { |
6968 | int i, n, val, min_val, best_node = 0; | |
6969 | ||
6970 | min_val = INT_MAX; | |
6971 | ||
6972 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6973 | /* Start at @node */ | |
6974 | n = (node + i) % MAX_NUMNODES; | |
6975 | ||
6976 | if (!nr_cpus_node(n)) | |
6977 | continue; | |
6978 | ||
6979 | /* Skip already used nodes */ | |
c5f59f08 | 6980 | if (node_isset(n, *used_nodes)) |
9c1cfda2 JH |
6981 | continue; |
6982 | ||
6983 | /* Simple min distance search */ | |
6984 | val = node_distance(node, n); | |
6985 | ||
6986 | if (val < min_val) { | |
6987 | min_val = val; | |
6988 | best_node = n; | |
6989 | } | |
6990 | } | |
6991 | ||
c5f59f08 | 6992 | node_set(best_node, *used_nodes); |
9c1cfda2 JH |
6993 | return best_node; |
6994 | } | |
6995 | ||
6996 | /** | |
6997 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
6998 | * @node: node whose cpumask we're constructing | |
73486722 | 6999 | * @span: resulting cpumask |
9c1cfda2 | 7000 | * |
41a2d6cf | 7001 | * Given a node, construct a good cpumask for its sched_domain to span. It |
9c1cfda2 JH |
7002 | * should be one that prevents unnecessary balancing, but also spreads tasks |
7003 | * out optimally. | |
7004 | */ | |
4bdbaad3 | 7005 | static void sched_domain_node_span(int node, cpumask_t *span) |
9c1cfda2 | 7006 | { |
c5f59f08 | 7007 | nodemask_t used_nodes; |
c5f59f08 | 7008 | node_to_cpumask_ptr(nodemask, node); |
48f24c4d | 7009 | int i; |
9c1cfda2 | 7010 | |
4bdbaad3 | 7011 | cpus_clear(*span); |
c5f59f08 | 7012 | nodes_clear(used_nodes); |
9c1cfda2 | 7013 | |
4bdbaad3 | 7014 | cpus_or(*span, *span, *nodemask); |
c5f59f08 | 7015 | node_set(node, used_nodes); |
9c1cfda2 JH |
7016 | |
7017 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
c5f59f08 | 7018 | int next_node = find_next_best_node(node, &used_nodes); |
48f24c4d | 7019 | |
c5f59f08 | 7020 | node_to_cpumask_ptr_next(nodemask, next_node); |
4bdbaad3 | 7021 | cpus_or(*span, *span, *nodemask); |
9c1cfda2 | 7022 | } |
9c1cfda2 JH |
7023 | } |
7024 | #endif | |
7025 | ||
5c45bf27 | 7026 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
48f24c4d | 7027 | |
9c1cfda2 | 7028 | /* |
48f24c4d | 7029 | * SMT sched-domains: |
9c1cfda2 | 7030 | */ |
1da177e4 LT |
7031 | #ifdef CONFIG_SCHED_SMT |
7032 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
6711cab4 | 7033 | static DEFINE_PER_CPU(struct sched_group, sched_group_cpus); |
48f24c4d | 7034 | |
41a2d6cf | 7035 | static int |
7c16ec58 MT |
7036 | cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
7037 | cpumask_t *unused) | |
1da177e4 | 7038 | { |
6711cab4 SS |
7039 | if (sg) |
7040 | *sg = &per_cpu(sched_group_cpus, cpu); | |
1da177e4 LT |
7041 | return cpu; |
7042 | } | |
7043 | #endif | |
7044 | ||
48f24c4d IM |
7045 | /* |
7046 | * multi-core sched-domains: | |
7047 | */ | |
1e9f28fa SS |
7048 | #ifdef CONFIG_SCHED_MC |
7049 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
6711cab4 | 7050 | static DEFINE_PER_CPU(struct sched_group, sched_group_core); |
1e9f28fa SS |
7051 | #endif |
7052 | ||
7053 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
41a2d6cf | 7054 | static int |
7c16ec58 MT |
7055 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
7056 | cpumask_t *mask) | |
1e9f28fa | 7057 | { |
6711cab4 | 7058 | int group; |
7c16ec58 MT |
7059 | |
7060 | *mask = per_cpu(cpu_sibling_map, cpu); | |
7061 | cpus_and(*mask, *mask, *cpu_map); | |
7062 | group = first_cpu(*mask); | |
6711cab4 SS |
7063 | if (sg) |
7064 | *sg = &per_cpu(sched_group_core, group); | |
7065 | return group; | |
1e9f28fa SS |
7066 | } |
7067 | #elif defined(CONFIG_SCHED_MC) | |
41a2d6cf | 7068 | static int |
7c16ec58 MT |
7069 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
7070 | cpumask_t *unused) | |
1e9f28fa | 7071 | { |
6711cab4 SS |
7072 | if (sg) |
7073 | *sg = &per_cpu(sched_group_core, cpu); | |
1e9f28fa SS |
7074 | return cpu; |
7075 | } | |
7076 | #endif | |
7077 | ||
1da177e4 | 7078 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
6711cab4 | 7079 | static DEFINE_PER_CPU(struct sched_group, sched_group_phys); |
48f24c4d | 7080 | |
41a2d6cf | 7081 | static int |
7c16ec58 MT |
7082 | cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
7083 | cpumask_t *mask) | |
1da177e4 | 7084 | { |
6711cab4 | 7085 | int group; |
48f24c4d | 7086 | #ifdef CONFIG_SCHED_MC |
7c16ec58 MT |
7087 | *mask = cpu_coregroup_map(cpu); |
7088 | cpus_and(*mask, *mask, *cpu_map); | |
7089 | group = first_cpu(*mask); | |
1e9f28fa | 7090 | #elif defined(CONFIG_SCHED_SMT) |
7c16ec58 MT |
7091 | *mask = per_cpu(cpu_sibling_map, cpu); |
7092 | cpus_and(*mask, *mask, *cpu_map); | |
7093 | group = first_cpu(*mask); | |
1da177e4 | 7094 | #else |
6711cab4 | 7095 | group = cpu; |
1da177e4 | 7096 | #endif |
6711cab4 SS |
7097 | if (sg) |
7098 | *sg = &per_cpu(sched_group_phys, group); | |
7099 | return group; | |
1da177e4 LT |
7100 | } |
7101 | ||
7102 | #ifdef CONFIG_NUMA | |
1da177e4 | 7103 | /* |
9c1cfda2 JH |
7104 | * The init_sched_build_groups can't handle what we want to do with node |
7105 | * groups, so roll our own. Now each node has its own list of groups which | |
7106 | * gets dynamically allocated. | |
1da177e4 | 7107 | */ |
9c1cfda2 | 7108 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
434d53b0 | 7109 | static struct sched_group ***sched_group_nodes_bycpu; |
1da177e4 | 7110 | |
9c1cfda2 | 7111 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
6711cab4 | 7112 | static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes); |
9c1cfda2 | 7113 | |
6711cab4 | 7114 | static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map, |
7c16ec58 | 7115 | struct sched_group **sg, cpumask_t *nodemask) |
9c1cfda2 | 7116 | { |
6711cab4 SS |
7117 | int group; |
7118 | ||
7c16ec58 MT |
7119 | *nodemask = node_to_cpumask(cpu_to_node(cpu)); |
7120 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7121 | group = first_cpu(*nodemask); | |
6711cab4 SS |
7122 | |
7123 | if (sg) | |
7124 | *sg = &per_cpu(sched_group_allnodes, group); | |
7125 | return group; | |
1da177e4 | 7126 | } |
6711cab4 | 7127 | |
08069033 SS |
7128 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
7129 | { | |
7130 | struct sched_group *sg = group_head; | |
7131 | int j; | |
7132 | ||
7133 | if (!sg) | |
7134 | return; | |
3a5c359a AK |
7135 | do { |
7136 | for_each_cpu_mask(j, sg->cpumask) { | |
7137 | struct sched_domain *sd; | |
08069033 | 7138 | |
3a5c359a AK |
7139 | sd = &per_cpu(phys_domains, j); |
7140 | if (j != first_cpu(sd->groups->cpumask)) { | |
7141 | /* | |
7142 | * Only add "power" once for each | |
7143 | * physical package. | |
7144 | */ | |
7145 | continue; | |
7146 | } | |
08069033 | 7147 | |
3a5c359a AK |
7148 | sg_inc_cpu_power(sg, sd->groups->__cpu_power); |
7149 | } | |
7150 | sg = sg->next; | |
7151 | } while (sg != group_head); | |
08069033 | 7152 | } |
1da177e4 LT |
7153 | #endif |
7154 | ||
a616058b | 7155 | #ifdef CONFIG_NUMA |
51888ca2 | 7156 | /* Free memory allocated for various sched_group structures */ |
7c16ec58 | 7157 | static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask) |
51888ca2 | 7158 | { |
a616058b | 7159 | int cpu, i; |
51888ca2 SV |
7160 | |
7161 | for_each_cpu_mask(cpu, *cpu_map) { | |
51888ca2 SV |
7162 | struct sched_group **sched_group_nodes |
7163 | = sched_group_nodes_bycpu[cpu]; | |
7164 | ||
51888ca2 SV |
7165 | if (!sched_group_nodes) |
7166 | continue; | |
7167 | ||
7168 | for (i = 0; i < MAX_NUMNODES; i++) { | |
51888ca2 SV |
7169 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; |
7170 | ||
7c16ec58 MT |
7171 | *nodemask = node_to_cpumask(i); |
7172 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7173 | if (cpus_empty(*nodemask)) | |
51888ca2 SV |
7174 | continue; |
7175 | ||
7176 | if (sg == NULL) | |
7177 | continue; | |
7178 | sg = sg->next; | |
7179 | next_sg: | |
7180 | oldsg = sg; | |
7181 | sg = sg->next; | |
7182 | kfree(oldsg); | |
7183 | if (oldsg != sched_group_nodes[i]) | |
7184 | goto next_sg; | |
7185 | } | |
7186 | kfree(sched_group_nodes); | |
7187 | sched_group_nodes_bycpu[cpu] = NULL; | |
7188 | } | |
51888ca2 | 7189 | } |
a616058b | 7190 | #else |
7c16ec58 | 7191 | static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask) |
a616058b SS |
7192 | { |
7193 | } | |
7194 | #endif | |
51888ca2 | 7195 | |
89c4710e SS |
7196 | /* |
7197 | * Initialize sched groups cpu_power. | |
7198 | * | |
7199 | * cpu_power indicates the capacity of sched group, which is used while | |
7200 | * distributing the load between different sched groups in a sched domain. | |
7201 | * Typically cpu_power for all the groups in a sched domain will be same unless | |
7202 | * there are asymmetries in the topology. If there are asymmetries, group | |
7203 | * having more cpu_power will pickup more load compared to the group having | |
7204 | * less cpu_power. | |
7205 | * | |
7206 | * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents | |
7207 | * the maximum number of tasks a group can handle in the presence of other idle | |
7208 | * or lightly loaded groups in the same sched domain. | |
7209 | */ | |
7210 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) | |
7211 | { | |
7212 | struct sched_domain *child; | |
7213 | struct sched_group *group; | |
7214 | ||
7215 | WARN_ON(!sd || !sd->groups); | |
7216 | ||
7217 | if (cpu != first_cpu(sd->groups->cpumask)) | |
7218 | return; | |
7219 | ||
7220 | child = sd->child; | |
7221 | ||
5517d86b ED |
7222 | sd->groups->__cpu_power = 0; |
7223 | ||
89c4710e SS |
7224 | /* |
7225 | * For perf policy, if the groups in child domain share resources | |
7226 | * (for example cores sharing some portions of the cache hierarchy | |
7227 | * or SMT), then set this domain groups cpu_power such that each group | |
7228 | * can handle only one task, when there are other idle groups in the | |
7229 | * same sched domain. | |
7230 | */ | |
7231 | if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) && | |
7232 | (child->flags & | |
7233 | (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) { | |
5517d86b | 7234 | sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE); |
89c4710e SS |
7235 | return; |
7236 | } | |
7237 | ||
89c4710e SS |
7238 | /* |
7239 | * add cpu_power of each child group to this groups cpu_power | |
7240 | */ | |
7241 | group = child->groups; | |
7242 | do { | |
5517d86b | 7243 | sg_inc_cpu_power(sd->groups, group->__cpu_power); |
89c4710e SS |
7244 | group = group->next; |
7245 | } while (group != child->groups); | |
7246 | } | |
7247 | ||
7c16ec58 MT |
7248 | /* |
7249 | * Initializers for schedule domains | |
7250 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() | |
7251 | */ | |
7252 | ||
7253 | #define SD_INIT(sd, type) sd_init_##type(sd) | |
7254 | #define SD_INIT_FUNC(type) \ | |
7255 | static noinline void sd_init_##type(struct sched_domain *sd) \ | |
7256 | { \ | |
7257 | memset(sd, 0, sizeof(*sd)); \ | |
7258 | *sd = SD_##type##_INIT; \ | |
1d3504fc | 7259 | sd->level = SD_LV_##type; \ |
7c16ec58 MT |
7260 | } |
7261 | ||
7262 | SD_INIT_FUNC(CPU) | |
7263 | #ifdef CONFIG_NUMA | |
7264 | SD_INIT_FUNC(ALLNODES) | |
7265 | SD_INIT_FUNC(NODE) | |
7266 | #endif | |
7267 | #ifdef CONFIG_SCHED_SMT | |
7268 | SD_INIT_FUNC(SIBLING) | |
7269 | #endif | |
7270 | #ifdef CONFIG_SCHED_MC | |
7271 | SD_INIT_FUNC(MC) | |
7272 | #endif | |
7273 | ||
7274 | /* | |
7275 | * To minimize stack usage kmalloc room for cpumasks and share the | |
7276 | * space as the usage in build_sched_domains() dictates. Used only | |
7277 | * if the amount of space is significant. | |
7278 | */ | |
7279 | struct allmasks { | |
7280 | cpumask_t tmpmask; /* make this one first */ | |
7281 | union { | |
7282 | cpumask_t nodemask; | |
7283 | cpumask_t this_sibling_map; | |
7284 | cpumask_t this_core_map; | |
7285 | }; | |
7286 | cpumask_t send_covered; | |
7287 | ||
7288 | #ifdef CONFIG_NUMA | |
7289 | cpumask_t domainspan; | |
7290 | cpumask_t covered; | |
7291 | cpumask_t notcovered; | |
7292 | #endif | |
7293 | }; | |
7294 | ||
7295 | #if NR_CPUS > 128 | |
7296 | #define SCHED_CPUMASK_ALLOC 1 | |
7297 | #define SCHED_CPUMASK_FREE(v) kfree(v) | |
7298 | #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v | |
7299 | #else | |
7300 | #define SCHED_CPUMASK_ALLOC 0 | |
7301 | #define SCHED_CPUMASK_FREE(v) | |
7302 | #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v | |
7303 | #endif | |
7304 | ||
7305 | #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \ | |
7306 | ((unsigned long)(a) + offsetof(struct allmasks, v)) | |
7307 | ||
1d3504fc HS |
7308 | static int default_relax_domain_level = -1; |
7309 | ||
7310 | static int __init setup_relax_domain_level(char *str) | |
7311 | { | |
7312 | default_relax_domain_level = simple_strtoul(str, NULL, 0); | |
7313 | return 1; | |
7314 | } | |
7315 | __setup("relax_domain_level=", setup_relax_domain_level); | |
7316 | ||
7317 | static void set_domain_attribute(struct sched_domain *sd, | |
7318 | struct sched_domain_attr *attr) | |
7319 | { | |
7320 | int request; | |
7321 | ||
7322 | if (!attr || attr->relax_domain_level < 0) { | |
7323 | if (default_relax_domain_level < 0) | |
7324 | return; | |
7325 | else | |
7326 | request = default_relax_domain_level; | |
7327 | } else | |
7328 | request = attr->relax_domain_level; | |
7329 | if (request < sd->level) { | |
7330 | /* turn off idle balance on this domain */ | |
7331 | sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE); | |
7332 | } else { | |
7333 | /* turn on idle balance on this domain */ | |
7334 | sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE); | |
7335 | } | |
7336 | } | |
7337 | ||
1da177e4 | 7338 | /* |
1a20ff27 DG |
7339 | * Build sched domains for a given set of cpus and attach the sched domains |
7340 | * to the individual cpus | |
1da177e4 | 7341 | */ |
1d3504fc HS |
7342 | static int __build_sched_domains(const cpumask_t *cpu_map, |
7343 | struct sched_domain_attr *attr) | |
1da177e4 LT |
7344 | { |
7345 | int i; | |
57d885fe | 7346 | struct root_domain *rd; |
7c16ec58 MT |
7347 | SCHED_CPUMASK_DECLARE(allmasks); |
7348 | cpumask_t *tmpmask; | |
d1b55138 JH |
7349 | #ifdef CONFIG_NUMA |
7350 | struct sched_group **sched_group_nodes = NULL; | |
6711cab4 | 7351 | int sd_allnodes = 0; |
d1b55138 JH |
7352 | |
7353 | /* | |
7354 | * Allocate the per-node list of sched groups | |
7355 | */ | |
5cf9f062 | 7356 | sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *), |
41a2d6cf | 7357 | GFP_KERNEL); |
d1b55138 JH |
7358 | if (!sched_group_nodes) { |
7359 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
51888ca2 | 7360 | return -ENOMEM; |
d1b55138 | 7361 | } |
d1b55138 | 7362 | #endif |
1da177e4 | 7363 | |
dc938520 | 7364 | rd = alloc_rootdomain(); |
57d885fe GH |
7365 | if (!rd) { |
7366 | printk(KERN_WARNING "Cannot alloc root domain\n"); | |
7c16ec58 MT |
7367 | #ifdef CONFIG_NUMA |
7368 | kfree(sched_group_nodes); | |
7369 | #endif | |
57d885fe GH |
7370 | return -ENOMEM; |
7371 | } | |
7372 | ||
7c16ec58 MT |
7373 | #if SCHED_CPUMASK_ALLOC |
7374 | /* get space for all scratch cpumask variables */ | |
7375 | allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL); | |
7376 | if (!allmasks) { | |
7377 | printk(KERN_WARNING "Cannot alloc cpumask array\n"); | |
7378 | kfree(rd); | |
7379 | #ifdef CONFIG_NUMA | |
7380 | kfree(sched_group_nodes); | |
7381 | #endif | |
7382 | return -ENOMEM; | |
7383 | } | |
7384 | #endif | |
7385 | tmpmask = (cpumask_t *)allmasks; | |
7386 | ||
7387 | ||
7388 | #ifdef CONFIG_NUMA | |
7389 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
7390 | #endif | |
7391 | ||
1da177e4 | 7392 | /* |
1a20ff27 | 7393 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 7394 | */ |
1a20ff27 | 7395 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 7396 | struct sched_domain *sd = NULL, *p; |
7c16ec58 | 7397 | SCHED_CPUMASK_VAR(nodemask, allmasks); |
1da177e4 | 7398 | |
7c16ec58 MT |
7399 | *nodemask = node_to_cpumask(cpu_to_node(i)); |
7400 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
1da177e4 LT |
7401 | |
7402 | #ifdef CONFIG_NUMA | |
dd41f596 | 7403 | if (cpus_weight(*cpu_map) > |
7c16ec58 | 7404 | SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) { |
9c1cfda2 | 7405 | sd = &per_cpu(allnodes_domains, i); |
7c16ec58 | 7406 | SD_INIT(sd, ALLNODES); |
1d3504fc | 7407 | set_domain_attribute(sd, attr); |
9c1cfda2 | 7408 | sd->span = *cpu_map; |
18d95a28 | 7409 | sd->first_cpu = first_cpu(sd->span); |
7c16ec58 | 7410 | cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask); |
9c1cfda2 | 7411 | p = sd; |
6711cab4 | 7412 | sd_allnodes = 1; |
9c1cfda2 JH |
7413 | } else |
7414 | p = NULL; | |
7415 | ||
1da177e4 | 7416 | sd = &per_cpu(node_domains, i); |
7c16ec58 | 7417 | SD_INIT(sd, NODE); |
1d3504fc | 7418 | set_domain_attribute(sd, attr); |
4bdbaad3 | 7419 | sched_domain_node_span(cpu_to_node(i), &sd->span); |
18d95a28 | 7420 | sd->first_cpu = first_cpu(sd->span); |
9c1cfda2 | 7421 | sd->parent = p; |
1a848870 SS |
7422 | if (p) |
7423 | p->child = sd; | |
9c1cfda2 | 7424 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
7425 | #endif |
7426 | ||
7427 | p = sd; | |
7428 | sd = &per_cpu(phys_domains, i); | |
7c16ec58 | 7429 | SD_INIT(sd, CPU); |
1d3504fc | 7430 | set_domain_attribute(sd, attr); |
7c16ec58 | 7431 | sd->span = *nodemask; |
18d95a28 | 7432 | sd->first_cpu = first_cpu(sd->span); |
1da177e4 | 7433 | sd->parent = p; |
1a848870 SS |
7434 | if (p) |
7435 | p->child = sd; | |
7c16ec58 | 7436 | cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask); |
1da177e4 | 7437 | |
1e9f28fa SS |
7438 | #ifdef CONFIG_SCHED_MC |
7439 | p = sd; | |
7440 | sd = &per_cpu(core_domains, i); | |
7c16ec58 | 7441 | SD_INIT(sd, MC); |
1d3504fc | 7442 | set_domain_attribute(sd, attr); |
1e9f28fa | 7443 | sd->span = cpu_coregroup_map(i); |
18d95a28 | 7444 | sd->first_cpu = first_cpu(sd->span); |
1e9f28fa SS |
7445 | cpus_and(sd->span, sd->span, *cpu_map); |
7446 | sd->parent = p; | |
1a848870 | 7447 | p->child = sd; |
7c16ec58 | 7448 | cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask); |
1e9f28fa SS |
7449 | #endif |
7450 | ||
1da177e4 LT |
7451 | #ifdef CONFIG_SCHED_SMT |
7452 | p = sd; | |
7453 | sd = &per_cpu(cpu_domains, i); | |
7c16ec58 | 7454 | SD_INIT(sd, SIBLING); |
1d3504fc | 7455 | set_domain_attribute(sd, attr); |
d5a7430d | 7456 | sd->span = per_cpu(cpu_sibling_map, i); |
18d95a28 | 7457 | sd->first_cpu = first_cpu(sd->span); |
1a20ff27 | 7458 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 | 7459 | sd->parent = p; |
1a848870 | 7460 | p->child = sd; |
7c16ec58 | 7461 | cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask); |
1da177e4 LT |
7462 | #endif |
7463 | } | |
7464 | ||
7465 | #ifdef CONFIG_SCHED_SMT | |
7466 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 7467 | for_each_cpu_mask(i, *cpu_map) { |
7c16ec58 MT |
7468 | SCHED_CPUMASK_VAR(this_sibling_map, allmasks); |
7469 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7470 | ||
7471 | *this_sibling_map = per_cpu(cpu_sibling_map, i); | |
7472 | cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map); | |
7473 | if (i != first_cpu(*this_sibling_map)) | |
1da177e4 LT |
7474 | continue; |
7475 | ||
dd41f596 | 7476 | init_sched_build_groups(this_sibling_map, cpu_map, |
7c16ec58 MT |
7477 | &cpu_to_cpu_group, |
7478 | send_covered, tmpmask); | |
1da177e4 LT |
7479 | } |
7480 | #endif | |
7481 | ||
1e9f28fa SS |
7482 | #ifdef CONFIG_SCHED_MC |
7483 | /* Set up multi-core groups */ | |
7484 | for_each_cpu_mask(i, *cpu_map) { | |
7c16ec58 MT |
7485 | SCHED_CPUMASK_VAR(this_core_map, allmasks); |
7486 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7487 | ||
7488 | *this_core_map = cpu_coregroup_map(i); | |
7489 | cpus_and(*this_core_map, *this_core_map, *cpu_map); | |
7490 | if (i != first_cpu(*this_core_map)) | |
1e9f28fa | 7491 | continue; |
7c16ec58 | 7492 | |
dd41f596 | 7493 | init_sched_build_groups(this_core_map, cpu_map, |
7c16ec58 MT |
7494 | &cpu_to_core_group, |
7495 | send_covered, tmpmask); | |
1e9f28fa SS |
7496 | } |
7497 | #endif | |
7498 | ||
1da177e4 LT |
7499 | /* Set up physical groups */ |
7500 | for (i = 0; i < MAX_NUMNODES; i++) { | |
7c16ec58 MT |
7501 | SCHED_CPUMASK_VAR(nodemask, allmasks); |
7502 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
1da177e4 | 7503 | |
7c16ec58 MT |
7504 | *nodemask = node_to_cpumask(i); |
7505 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7506 | if (cpus_empty(*nodemask)) | |
1da177e4 LT |
7507 | continue; |
7508 | ||
7c16ec58 MT |
7509 | init_sched_build_groups(nodemask, cpu_map, |
7510 | &cpu_to_phys_group, | |
7511 | send_covered, tmpmask); | |
1da177e4 LT |
7512 | } |
7513 | ||
7514 | #ifdef CONFIG_NUMA | |
7515 | /* Set up node groups */ | |
7c16ec58 MT |
7516 | if (sd_allnodes) { |
7517 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7518 | ||
7519 | init_sched_build_groups(cpu_map, cpu_map, | |
7520 | &cpu_to_allnodes_group, | |
7521 | send_covered, tmpmask); | |
7522 | } | |
9c1cfda2 JH |
7523 | |
7524 | for (i = 0; i < MAX_NUMNODES; i++) { | |
7525 | /* Set up node groups */ | |
7526 | struct sched_group *sg, *prev; | |
7c16ec58 MT |
7527 | SCHED_CPUMASK_VAR(nodemask, allmasks); |
7528 | SCHED_CPUMASK_VAR(domainspan, allmasks); | |
7529 | SCHED_CPUMASK_VAR(covered, allmasks); | |
9c1cfda2 JH |
7530 | int j; |
7531 | ||
7c16ec58 MT |
7532 | *nodemask = node_to_cpumask(i); |
7533 | cpus_clear(*covered); | |
7534 | ||
7535 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7536 | if (cpus_empty(*nodemask)) { | |
d1b55138 | 7537 | sched_group_nodes[i] = NULL; |
9c1cfda2 | 7538 | continue; |
d1b55138 | 7539 | } |
9c1cfda2 | 7540 | |
4bdbaad3 | 7541 | sched_domain_node_span(i, domainspan); |
7c16ec58 | 7542 | cpus_and(*domainspan, *domainspan, *cpu_map); |
9c1cfda2 | 7543 | |
15f0b676 | 7544 | sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i); |
51888ca2 SV |
7545 | if (!sg) { |
7546 | printk(KERN_WARNING "Can not alloc domain group for " | |
7547 | "node %d\n", i); | |
7548 | goto error; | |
7549 | } | |
9c1cfda2 | 7550 | sched_group_nodes[i] = sg; |
7c16ec58 | 7551 | for_each_cpu_mask(j, *nodemask) { |
9c1cfda2 | 7552 | struct sched_domain *sd; |
9761eea8 | 7553 | |
9c1cfda2 JH |
7554 | sd = &per_cpu(node_domains, j); |
7555 | sd->groups = sg; | |
9c1cfda2 | 7556 | } |
5517d86b | 7557 | sg->__cpu_power = 0; |
7c16ec58 | 7558 | sg->cpumask = *nodemask; |
51888ca2 | 7559 | sg->next = sg; |
7c16ec58 | 7560 | cpus_or(*covered, *covered, *nodemask); |
9c1cfda2 JH |
7561 | prev = sg; |
7562 | ||
7563 | for (j = 0; j < MAX_NUMNODES; j++) { | |
7c16ec58 | 7564 | SCHED_CPUMASK_VAR(notcovered, allmasks); |
9c1cfda2 | 7565 | int n = (i + j) % MAX_NUMNODES; |
c5f59f08 | 7566 | node_to_cpumask_ptr(pnodemask, n); |
9c1cfda2 | 7567 | |
7c16ec58 MT |
7568 | cpus_complement(*notcovered, *covered); |
7569 | cpus_and(*tmpmask, *notcovered, *cpu_map); | |
7570 | cpus_and(*tmpmask, *tmpmask, *domainspan); | |
7571 | if (cpus_empty(*tmpmask)) | |
9c1cfda2 JH |
7572 | break; |
7573 | ||
7c16ec58 MT |
7574 | cpus_and(*tmpmask, *tmpmask, *pnodemask); |
7575 | if (cpus_empty(*tmpmask)) | |
9c1cfda2 JH |
7576 | continue; |
7577 | ||
15f0b676 SV |
7578 | sg = kmalloc_node(sizeof(struct sched_group), |
7579 | GFP_KERNEL, i); | |
9c1cfda2 JH |
7580 | if (!sg) { |
7581 | printk(KERN_WARNING | |
7582 | "Can not alloc domain group for node %d\n", j); | |
51888ca2 | 7583 | goto error; |
9c1cfda2 | 7584 | } |
5517d86b | 7585 | sg->__cpu_power = 0; |
7c16ec58 | 7586 | sg->cpumask = *tmpmask; |
51888ca2 | 7587 | sg->next = prev->next; |
7c16ec58 | 7588 | cpus_or(*covered, *covered, *tmpmask); |
9c1cfda2 JH |
7589 | prev->next = sg; |
7590 | prev = sg; | |
7591 | } | |
9c1cfda2 | 7592 | } |
1da177e4 LT |
7593 | #endif |
7594 | ||
7595 | /* Calculate CPU power for physical packages and nodes */ | |
5c45bf27 | 7596 | #ifdef CONFIG_SCHED_SMT |
1a20ff27 | 7597 | for_each_cpu_mask(i, *cpu_map) { |
dd41f596 IM |
7598 | struct sched_domain *sd = &per_cpu(cpu_domains, i); |
7599 | ||
89c4710e | 7600 | init_sched_groups_power(i, sd); |
5c45bf27 | 7601 | } |
1da177e4 | 7602 | #endif |
1e9f28fa | 7603 | #ifdef CONFIG_SCHED_MC |
5c45bf27 | 7604 | for_each_cpu_mask(i, *cpu_map) { |
dd41f596 IM |
7605 | struct sched_domain *sd = &per_cpu(core_domains, i); |
7606 | ||
89c4710e | 7607 | init_sched_groups_power(i, sd); |
5c45bf27 SS |
7608 | } |
7609 | #endif | |
1e9f28fa | 7610 | |
5c45bf27 | 7611 | for_each_cpu_mask(i, *cpu_map) { |
dd41f596 IM |
7612 | struct sched_domain *sd = &per_cpu(phys_domains, i); |
7613 | ||
89c4710e | 7614 | init_sched_groups_power(i, sd); |
1da177e4 LT |
7615 | } |
7616 | ||
9c1cfda2 | 7617 | #ifdef CONFIG_NUMA |
08069033 SS |
7618 | for (i = 0; i < MAX_NUMNODES; i++) |
7619 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
9c1cfda2 | 7620 | |
6711cab4 SS |
7621 | if (sd_allnodes) { |
7622 | struct sched_group *sg; | |
f712c0c7 | 7623 | |
7c16ec58 MT |
7624 | cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg, |
7625 | tmpmask); | |
f712c0c7 SS |
7626 | init_numa_sched_groups_power(sg); |
7627 | } | |
9c1cfda2 JH |
7628 | #endif |
7629 | ||
1da177e4 | 7630 | /* Attach the domains */ |
1a20ff27 | 7631 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
7632 | struct sched_domain *sd; |
7633 | #ifdef CONFIG_SCHED_SMT | |
7634 | sd = &per_cpu(cpu_domains, i); | |
1e9f28fa SS |
7635 | #elif defined(CONFIG_SCHED_MC) |
7636 | sd = &per_cpu(core_domains, i); | |
1da177e4 LT |
7637 | #else |
7638 | sd = &per_cpu(phys_domains, i); | |
7639 | #endif | |
57d885fe | 7640 | cpu_attach_domain(sd, rd, i); |
1da177e4 | 7641 | } |
51888ca2 | 7642 | |
7c16ec58 | 7643 | SCHED_CPUMASK_FREE((void *)allmasks); |
51888ca2 SV |
7644 | return 0; |
7645 | ||
a616058b | 7646 | #ifdef CONFIG_NUMA |
51888ca2 | 7647 | error: |
7c16ec58 MT |
7648 | free_sched_groups(cpu_map, tmpmask); |
7649 | SCHED_CPUMASK_FREE((void *)allmasks); | |
51888ca2 | 7650 | return -ENOMEM; |
a616058b | 7651 | #endif |
1da177e4 | 7652 | } |
029190c5 | 7653 | |
1d3504fc HS |
7654 | static int build_sched_domains(const cpumask_t *cpu_map) |
7655 | { | |
7656 | return __build_sched_domains(cpu_map, NULL); | |
7657 | } | |
7658 | ||
029190c5 PJ |
7659 | static cpumask_t *doms_cur; /* current sched domains */ |
7660 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ | |
1d3504fc HS |
7661 | static struct sched_domain_attr *dattr_cur; /* attribues of custom domains |
7662 | in 'doms_cur' */ | |
029190c5 PJ |
7663 | |
7664 | /* | |
7665 | * Special case: If a kmalloc of a doms_cur partition (array of | |
7666 | * cpumask_t) fails, then fallback to a single sched domain, | |
7667 | * as determined by the single cpumask_t fallback_doms. | |
7668 | */ | |
7669 | static cpumask_t fallback_doms; | |
7670 | ||
22e52b07 HC |
7671 | void __attribute__((weak)) arch_update_cpu_topology(void) |
7672 | { | |
7673 | } | |
7674 | ||
1a20ff27 | 7675 | /* |
41a2d6cf | 7676 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
029190c5 PJ |
7677 | * For now this just excludes isolated cpus, but could be used to |
7678 | * exclude other special cases in the future. | |
1a20ff27 | 7679 | */ |
51888ca2 | 7680 | static int arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 | 7681 | { |
7378547f MM |
7682 | int err; |
7683 | ||
22e52b07 | 7684 | arch_update_cpu_topology(); |
029190c5 PJ |
7685 | ndoms_cur = 1; |
7686 | doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL); | |
7687 | if (!doms_cur) | |
7688 | doms_cur = &fallback_doms; | |
7689 | cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map); | |
1d3504fc | 7690 | dattr_cur = NULL; |
7378547f | 7691 | err = build_sched_domains(doms_cur); |
6382bc90 | 7692 | register_sched_domain_sysctl(); |
7378547f MM |
7693 | |
7694 | return err; | |
1a20ff27 DG |
7695 | } |
7696 | ||
7c16ec58 MT |
7697 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map, |
7698 | cpumask_t *tmpmask) | |
1da177e4 | 7699 | { |
7c16ec58 | 7700 | free_sched_groups(cpu_map, tmpmask); |
9c1cfda2 | 7701 | } |
1da177e4 | 7702 | |
1a20ff27 DG |
7703 | /* |
7704 | * Detach sched domains from a group of cpus specified in cpu_map | |
7705 | * These cpus will now be attached to the NULL domain | |
7706 | */ | |
858119e1 | 7707 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
1a20ff27 | 7708 | { |
7c16ec58 | 7709 | cpumask_t tmpmask; |
1a20ff27 DG |
7710 | int i; |
7711 | ||
6382bc90 MM |
7712 | unregister_sched_domain_sysctl(); |
7713 | ||
1a20ff27 | 7714 | for_each_cpu_mask(i, *cpu_map) |
57d885fe | 7715 | cpu_attach_domain(NULL, &def_root_domain, i); |
1a20ff27 | 7716 | synchronize_sched(); |
7c16ec58 | 7717 | arch_destroy_sched_domains(cpu_map, &tmpmask); |
1a20ff27 DG |
7718 | } |
7719 | ||
1d3504fc HS |
7720 | /* handle null as "default" */ |
7721 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, | |
7722 | struct sched_domain_attr *new, int idx_new) | |
7723 | { | |
7724 | struct sched_domain_attr tmp; | |
7725 | ||
7726 | /* fast path */ | |
7727 | if (!new && !cur) | |
7728 | return 1; | |
7729 | ||
7730 | tmp = SD_ATTR_INIT; | |
7731 | return !memcmp(cur ? (cur + idx_cur) : &tmp, | |
7732 | new ? (new + idx_new) : &tmp, | |
7733 | sizeof(struct sched_domain_attr)); | |
7734 | } | |
7735 | ||
029190c5 PJ |
7736 | /* |
7737 | * Partition sched domains as specified by the 'ndoms_new' | |
41a2d6cf | 7738 | * cpumasks in the array doms_new[] of cpumasks. This compares |
029190c5 PJ |
7739 | * doms_new[] to the current sched domain partitioning, doms_cur[]. |
7740 | * It destroys each deleted domain and builds each new domain. | |
7741 | * | |
7742 | * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'. | |
41a2d6cf IM |
7743 | * The masks don't intersect (don't overlap.) We should setup one |
7744 | * sched domain for each mask. CPUs not in any of the cpumasks will | |
7745 | * not be load balanced. If the same cpumask appears both in the | |
029190c5 PJ |
7746 | * current 'doms_cur' domains and in the new 'doms_new', we can leave |
7747 | * it as it is. | |
7748 | * | |
41a2d6cf IM |
7749 | * The passed in 'doms_new' should be kmalloc'd. This routine takes |
7750 | * ownership of it and will kfree it when done with it. If the caller | |
029190c5 PJ |
7751 | * failed the kmalloc call, then it can pass in doms_new == NULL, |
7752 | * and partition_sched_domains() will fallback to the single partition | |
7753 | * 'fallback_doms'. | |
7754 | * | |
7755 | * Call with hotplug lock held | |
7756 | */ | |
1d3504fc HS |
7757 | void partition_sched_domains(int ndoms_new, cpumask_t *doms_new, |
7758 | struct sched_domain_attr *dattr_new) | |
029190c5 PJ |
7759 | { |
7760 | int i, j; | |
7761 | ||
a1835615 SV |
7762 | lock_doms_cur(); |
7763 | ||
7378547f MM |
7764 | /* always unregister in case we don't destroy any domains */ |
7765 | unregister_sched_domain_sysctl(); | |
7766 | ||
029190c5 PJ |
7767 | if (doms_new == NULL) { |
7768 | ndoms_new = 1; | |
7769 | doms_new = &fallback_doms; | |
7770 | cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map); | |
1d3504fc | 7771 | dattr_new = NULL; |
029190c5 PJ |
7772 | } |
7773 | ||
7774 | /* Destroy deleted domains */ | |
7775 | for (i = 0; i < ndoms_cur; i++) { | |
7776 | for (j = 0; j < ndoms_new; j++) { | |
1d3504fc HS |
7777 | if (cpus_equal(doms_cur[i], doms_new[j]) |
7778 | && dattrs_equal(dattr_cur, i, dattr_new, j)) | |
029190c5 PJ |
7779 | goto match1; |
7780 | } | |
7781 | /* no match - a current sched domain not in new doms_new[] */ | |
7782 | detach_destroy_domains(doms_cur + i); | |
7783 | match1: | |
7784 | ; | |
7785 | } | |
7786 | ||
7787 | /* Build new domains */ | |
7788 | for (i = 0; i < ndoms_new; i++) { | |
7789 | for (j = 0; j < ndoms_cur; j++) { | |
1d3504fc HS |
7790 | if (cpus_equal(doms_new[i], doms_cur[j]) |
7791 | && dattrs_equal(dattr_new, i, dattr_cur, j)) | |
029190c5 PJ |
7792 | goto match2; |
7793 | } | |
7794 | /* no match - add a new doms_new */ | |
1d3504fc HS |
7795 | __build_sched_domains(doms_new + i, |
7796 | dattr_new ? dattr_new + i : NULL); | |
029190c5 PJ |
7797 | match2: |
7798 | ; | |
7799 | } | |
7800 | ||
7801 | /* Remember the new sched domains */ | |
7802 | if (doms_cur != &fallback_doms) | |
7803 | kfree(doms_cur); | |
1d3504fc | 7804 | kfree(dattr_cur); /* kfree(NULL) is safe */ |
029190c5 | 7805 | doms_cur = doms_new; |
1d3504fc | 7806 | dattr_cur = dattr_new; |
029190c5 | 7807 | ndoms_cur = ndoms_new; |
7378547f MM |
7808 | |
7809 | register_sched_domain_sysctl(); | |
a1835615 SV |
7810 | |
7811 | unlock_doms_cur(); | |
029190c5 PJ |
7812 | } |
7813 | ||
5c45bf27 | 7814 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
9aefd0ab | 7815 | int arch_reinit_sched_domains(void) |
5c45bf27 SS |
7816 | { |
7817 | int err; | |
7818 | ||
95402b38 | 7819 | get_online_cpus(); |
5c45bf27 SS |
7820 | detach_destroy_domains(&cpu_online_map); |
7821 | err = arch_init_sched_domains(&cpu_online_map); | |
95402b38 | 7822 | put_online_cpus(); |
5c45bf27 SS |
7823 | |
7824 | return err; | |
7825 | } | |
7826 | ||
7827 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
7828 | { | |
7829 | int ret; | |
7830 | ||
7831 | if (buf[0] != '0' && buf[0] != '1') | |
7832 | return -EINVAL; | |
7833 | ||
7834 | if (smt) | |
7835 | sched_smt_power_savings = (buf[0] == '1'); | |
7836 | else | |
7837 | sched_mc_power_savings = (buf[0] == '1'); | |
7838 | ||
7839 | ret = arch_reinit_sched_domains(); | |
7840 | ||
7841 | return ret ? ret : count; | |
7842 | } | |
7843 | ||
5c45bf27 SS |
7844 | #ifdef CONFIG_SCHED_MC |
7845 | static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page) | |
7846 | { | |
7847 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
7848 | } | |
48f24c4d IM |
7849 | static ssize_t sched_mc_power_savings_store(struct sys_device *dev, |
7850 | const char *buf, size_t count) | |
5c45bf27 SS |
7851 | { |
7852 | return sched_power_savings_store(buf, count, 0); | |
7853 | } | |
6707de00 AB |
7854 | static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show, |
7855 | sched_mc_power_savings_store); | |
5c45bf27 SS |
7856 | #endif |
7857 | ||
7858 | #ifdef CONFIG_SCHED_SMT | |
7859 | static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page) | |
7860 | { | |
7861 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
7862 | } | |
48f24c4d IM |
7863 | static ssize_t sched_smt_power_savings_store(struct sys_device *dev, |
7864 | const char *buf, size_t count) | |
5c45bf27 SS |
7865 | { |
7866 | return sched_power_savings_store(buf, count, 1); | |
7867 | } | |
6707de00 AB |
7868 | static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show, |
7869 | sched_smt_power_savings_store); | |
7870 | #endif | |
7871 | ||
7872 | int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
7873 | { | |
7874 | int err = 0; | |
7875 | ||
7876 | #ifdef CONFIG_SCHED_SMT | |
7877 | if (smt_capable()) | |
7878 | err = sysfs_create_file(&cls->kset.kobj, | |
7879 | &attr_sched_smt_power_savings.attr); | |
7880 | #endif | |
7881 | #ifdef CONFIG_SCHED_MC | |
7882 | if (!err && mc_capable()) | |
7883 | err = sysfs_create_file(&cls->kset.kobj, | |
7884 | &attr_sched_mc_power_savings.attr); | |
7885 | #endif | |
7886 | return err; | |
7887 | } | |
5c45bf27 SS |
7888 | #endif |
7889 | ||
1da177e4 | 7890 | /* |
41a2d6cf | 7891 | * Force a reinitialization of the sched domains hierarchy. The domains |
1da177e4 | 7892 | * and groups cannot be updated in place without racing with the balancing |
41c7ce9a | 7893 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
7894 | * which will prevent rebalancing while the sched domains are recalculated. |
7895 | */ | |
7896 | static int update_sched_domains(struct notifier_block *nfb, | |
7897 | unsigned long action, void *hcpu) | |
7898 | { | |
1da177e4 LT |
7899 | switch (action) { |
7900 | case CPU_UP_PREPARE: | |
8bb78442 | 7901 | case CPU_UP_PREPARE_FROZEN: |
1da177e4 | 7902 | case CPU_DOWN_PREPARE: |
8bb78442 | 7903 | case CPU_DOWN_PREPARE_FROZEN: |
1a20ff27 | 7904 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
7905 | return NOTIFY_OK; |
7906 | ||
7907 | case CPU_UP_CANCELED: | |
8bb78442 | 7908 | case CPU_UP_CANCELED_FROZEN: |
1da177e4 | 7909 | case CPU_DOWN_FAILED: |
8bb78442 | 7910 | case CPU_DOWN_FAILED_FROZEN: |
1da177e4 | 7911 | case CPU_ONLINE: |
8bb78442 | 7912 | case CPU_ONLINE_FROZEN: |
1da177e4 | 7913 | case CPU_DEAD: |
8bb78442 | 7914 | case CPU_DEAD_FROZEN: |
1da177e4 LT |
7915 | /* |
7916 | * Fall through and re-initialise the domains. | |
7917 | */ | |
7918 | break; | |
7919 | default: | |
7920 | return NOTIFY_DONE; | |
7921 | } | |
7922 | ||
7923 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 7924 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
7925 | |
7926 | return NOTIFY_OK; | |
7927 | } | |
1da177e4 LT |
7928 | |
7929 | void __init sched_init_smp(void) | |
7930 | { | |
5c1e1767 NP |
7931 | cpumask_t non_isolated_cpus; |
7932 | ||
434d53b0 MT |
7933 | #if defined(CONFIG_NUMA) |
7934 | sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **), | |
7935 | GFP_KERNEL); | |
7936 | BUG_ON(sched_group_nodes_bycpu == NULL); | |
7937 | #endif | |
95402b38 | 7938 | get_online_cpus(); |
1a20ff27 | 7939 | arch_init_sched_domains(&cpu_online_map); |
e5e5673f | 7940 | cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map); |
5c1e1767 NP |
7941 | if (cpus_empty(non_isolated_cpus)) |
7942 | cpu_set(smp_processor_id(), non_isolated_cpus); | |
95402b38 | 7943 | put_online_cpus(); |
1da177e4 LT |
7944 | /* XXX: Theoretical race here - CPU may be hotplugged now */ |
7945 | hotcpu_notifier(update_sched_domains, 0); | |
5c1e1767 NP |
7946 | |
7947 | /* Move init over to a non-isolated CPU */ | |
7c16ec58 | 7948 | if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0) |
5c1e1767 | 7949 | BUG(); |
19978ca6 | 7950 | sched_init_granularity(); |
1da177e4 LT |
7951 | } |
7952 | #else | |
7953 | void __init sched_init_smp(void) | |
7954 | { | |
19978ca6 | 7955 | sched_init_granularity(); |
1da177e4 LT |
7956 | } |
7957 | #endif /* CONFIG_SMP */ | |
7958 | ||
7959 | int in_sched_functions(unsigned long addr) | |
7960 | { | |
1da177e4 LT |
7961 | return in_lock_functions(addr) || |
7962 | (addr >= (unsigned long)__sched_text_start | |
7963 | && addr < (unsigned long)__sched_text_end); | |
7964 | } | |
7965 | ||
a9957449 | 7966 | static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq) |
dd41f596 IM |
7967 | { |
7968 | cfs_rq->tasks_timeline = RB_ROOT; | |
4a55bd5e | 7969 | INIT_LIST_HEAD(&cfs_rq->tasks); |
dd41f596 IM |
7970 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7971 | cfs_rq->rq = rq; | |
7972 | #endif | |
67e9fb2a | 7973 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
dd41f596 IM |
7974 | } |
7975 | ||
fa85ae24 PZ |
7976 | static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) |
7977 | { | |
7978 | struct rt_prio_array *array; | |
7979 | int i; | |
7980 | ||
7981 | array = &rt_rq->active; | |
7982 | for (i = 0; i < MAX_RT_PRIO; i++) { | |
7983 | INIT_LIST_HEAD(array->queue + i); | |
7984 | __clear_bit(i, array->bitmap); | |
7985 | } | |
7986 | /* delimiter for bitsearch: */ | |
7987 | __set_bit(MAX_RT_PRIO, array->bitmap); | |
7988 | ||
052f1dc7 | 7989 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
48d5e258 PZ |
7990 | rt_rq->highest_prio = MAX_RT_PRIO; |
7991 | #endif | |
fa85ae24 PZ |
7992 | #ifdef CONFIG_SMP |
7993 | rt_rq->rt_nr_migratory = 0; | |
fa85ae24 PZ |
7994 | rt_rq->overloaded = 0; |
7995 | #endif | |
7996 | ||
7997 | rt_rq->rt_time = 0; | |
7998 | rt_rq->rt_throttled = 0; | |
ac086bc2 PZ |
7999 | rt_rq->rt_runtime = 0; |
8000 | spin_lock_init(&rt_rq->rt_runtime_lock); | |
6f505b16 | 8001 | |
052f1dc7 | 8002 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc | 8003 | rt_rq->rt_nr_boosted = 0; |
6f505b16 PZ |
8004 | rt_rq->rq = rq; |
8005 | #endif | |
fa85ae24 PZ |
8006 | } |
8007 | ||
6f505b16 | 8008 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ec7dc8ac DG |
8009 | static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, |
8010 | struct sched_entity *se, int cpu, int add, | |
8011 | struct sched_entity *parent) | |
6f505b16 | 8012 | { |
ec7dc8ac | 8013 | struct rq *rq = cpu_rq(cpu); |
6f505b16 PZ |
8014 | tg->cfs_rq[cpu] = cfs_rq; |
8015 | init_cfs_rq(cfs_rq, rq); | |
8016 | cfs_rq->tg = tg; | |
8017 | if (add) | |
8018 | list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list); | |
8019 | ||
8020 | tg->se[cpu] = se; | |
354d60c2 DG |
8021 | /* se could be NULL for init_task_group */ |
8022 | if (!se) | |
8023 | return; | |
8024 | ||
ec7dc8ac DG |
8025 | if (!parent) |
8026 | se->cfs_rq = &rq->cfs; | |
8027 | else | |
8028 | se->cfs_rq = parent->my_q; | |
8029 | ||
6f505b16 PZ |
8030 | se->my_q = cfs_rq; |
8031 | se->load.weight = tg->shares; | |
e05510d0 | 8032 | se->load.inv_weight = 0; |
ec7dc8ac | 8033 | se->parent = parent; |
6f505b16 | 8034 | } |
052f1dc7 | 8035 | #endif |
6f505b16 | 8036 | |
052f1dc7 | 8037 | #ifdef CONFIG_RT_GROUP_SCHED |
ec7dc8ac DG |
8038 | static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, |
8039 | struct sched_rt_entity *rt_se, int cpu, int add, | |
8040 | struct sched_rt_entity *parent) | |
6f505b16 | 8041 | { |
ec7dc8ac DG |
8042 | struct rq *rq = cpu_rq(cpu); |
8043 | ||
6f505b16 PZ |
8044 | tg->rt_rq[cpu] = rt_rq; |
8045 | init_rt_rq(rt_rq, rq); | |
8046 | rt_rq->tg = tg; | |
8047 | rt_rq->rt_se = rt_se; | |
ac086bc2 | 8048 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; |
6f505b16 PZ |
8049 | if (add) |
8050 | list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list); | |
8051 | ||
8052 | tg->rt_se[cpu] = rt_se; | |
354d60c2 DG |
8053 | if (!rt_se) |
8054 | return; | |
8055 | ||
ec7dc8ac DG |
8056 | if (!parent) |
8057 | rt_se->rt_rq = &rq->rt; | |
8058 | else | |
8059 | rt_se->rt_rq = parent->my_q; | |
8060 | ||
6f505b16 PZ |
8061 | rt_se->rt_rq = &rq->rt; |
8062 | rt_se->my_q = rt_rq; | |
ec7dc8ac | 8063 | rt_se->parent = parent; |
6f505b16 PZ |
8064 | INIT_LIST_HEAD(&rt_se->run_list); |
8065 | } | |
8066 | #endif | |
8067 | ||
1da177e4 LT |
8068 | void __init sched_init(void) |
8069 | { | |
dd41f596 | 8070 | int i, j; |
434d53b0 MT |
8071 | unsigned long alloc_size = 0, ptr; |
8072 | ||
8073 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8074 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
8075 | #endif | |
8076 | #ifdef CONFIG_RT_GROUP_SCHED | |
8077 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
eff766a6 PZ |
8078 | #endif |
8079 | #ifdef CONFIG_USER_SCHED | |
8080 | alloc_size *= 2; | |
434d53b0 MT |
8081 | #endif |
8082 | /* | |
8083 | * As sched_init() is called before page_alloc is setup, | |
8084 | * we use alloc_bootmem(). | |
8085 | */ | |
8086 | if (alloc_size) { | |
5a9d3225 | 8087 | ptr = (unsigned long)alloc_bootmem(alloc_size); |
434d53b0 MT |
8088 | |
8089 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8090 | init_task_group.se = (struct sched_entity **)ptr; | |
8091 | ptr += nr_cpu_ids * sizeof(void **); | |
8092 | ||
8093 | init_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
8094 | ptr += nr_cpu_ids * sizeof(void **); | |
eff766a6 PZ |
8095 | |
8096 | #ifdef CONFIG_USER_SCHED | |
8097 | root_task_group.se = (struct sched_entity **)ptr; | |
8098 | ptr += nr_cpu_ids * sizeof(void **); | |
8099 | ||
8100 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
8101 | ptr += nr_cpu_ids * sizeof(void **); | |
8102 | #endif | |
434d53b0 MT |
8103 | #endif |
8104 | #ifdef CONFIG_RT_GROUP_SCHED | |
8105 | init_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
8106 | ptr += nr_cpu_ids * sizeof(void **); | |
8107 | ||
8108 | init_task_group.rt_rq = (struct rt_rq **)ptr; | |
eff766a6 PZ |
8109 | ptr += nr_cpu_ids * sizeof(void **); |
8110 | ||
8111 | #ifdef CONFIG_USER_SCHED | |
8112 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
8113 | ptr += nr_cpu_ids * sizeof(void **); | |
8114 | ||
8115 | root_task_group.rt_rq = (struct rt_rq **)ptr; | |
8116 | ptr += nr_cpu_ids * sizeof(void **); | |
8117 | #endif | |
434d53b0 MT |
8118 | #endif |
8119 | } | |
dd41f596 | 8120 | |
57d885fe | 8121 | #ifdef CONFIG_SMP |
18d95a28 | 8122 | init_aggregate(); |
57d885fe GH |
8123 | init_defrootdomain(); |
8124 | #endif | |
8125 | ||
d0b27fa7 PZ |
8126 | init_rt_bandwidth(&def_rt_bandwidth, |
8127 | global_rt_period(), global_rt_runtime()); | |
8128 | ||
8129 | #ifdef CONFIG_RT_GROUP_SCHED | |
8130 | init_rt_bandwidth(&init_task_group.rt_bandwidth, | |
8131 | global_rt_period(), global_rt_runtime()); | |
eff766a6 PZ |
8132 | #ifdef CONFIG_USER_SCHED |
8133 | init_rt_bandwidth(&root_task_group.rt_bandwidth, | |
8134 | global_rt_period(), RUNTIME_INF); | |
8135 | #endif | |
d0b27fa7 PZ |
8136 | #endif |
8137 | ||
052f1dc7 | 8138 | #ifdef CONFIG_GROUP_SCHED |
6f505b16 | 8139 | list_add(&init_task_group.list, &task_groups); |
f473aa5e PZ |
8140 | INIT_LIST_HEAD(&init_task_group.children); |
8141 | ||
8142 | #ifdef CONFIG_USER_SCHED | |
8143 | INIT_LIST_HEAD(&root_task_group.children); | |
8144 | init_task_group.parent = &root_task_group; | |
8145 | list_add(&init_task_group.siblings, &root_task_group.children); | |
8146 | #endif | |
6f505b16 PZ |
8147 | #endif |
8148 | ||
0a945022 | 8149 | for_each_possible_cpu(i) { |
70b97a7f | 8150 | struct rq *rq; |
1da177e4 LT |
8151 | |
8152 | rq = cpu_rq(i); | |
8153 | spin_lock_init(&rq->lock); | |
fcb99371 | 8154 | lockdep_set_class(&rq->lock, &rq->rq_lock_key); |
7897986b | 8155 | rq->nr_running = 0; |
dd41f596 | 8156 | rq->clock = 1; |
15934a37 | 8157 | update_last_tick_seen(rq); |
dd41f596 | 8158 | init_cfs_rq(&rq->cfs, rq); |
6f505b16 | 8159 | init_rt_rq(&rq->rt, rq); |
dd41f596 | 8160 | #ifdef CONFIG_FAIR_GROUP_SCHED |
4cf86d77 | 8161 | init_task_group.shares = init_task_group_load; |
6f505b16 | 8162 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
354d60c2 DG |
8163 | #ifdef CONFIG_CGROUP_SCHED |
8164 | /* | |
8165 | * How much cpu bandwidth does init_task_group get? | |
8166 | * | |
8167 | * In case of task-groups formed thr' the cgroup filesystem, it | |
8168 | * gets 100% of the cpu resources in the system. This overall | |
8169 | * system cpu resource is divided among the tasks of | |
8170 | * init_task_group and its child task-groups in a fair manner, | |
8171 | * based on each entity's (task or task-group's) weight | |
8172 | * (se->load.weight). | |
8173 | * | |
8174 | * In other words, if init_task_group has 10 tasks of weight | |
8175 | * 1024) and two child groups A0 and A1 (of weight 1024 each), | |
8176 | * then A0's share of the cpu resource is: | |
8177 | * | |
8178 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% | |
8179 | * | |
8180 | * We achieve this by letting init_task_group's tasks sit | |
8181 | * directly in rq->cfs (i.e init_task_group->se[] = NULL). | |
8182 | */ | |
ec7dc8ac | 8183 | init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL); |
354d60c2 | 8184 | #elif defined CONFIG_USER_SCHED |
eff766a6 PZ |
8185 | root_task_group.shares = NICE_0_LOAD; |
8186 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL); | |
354d60c2 DG |
8187 | /* |
8188 | * In case of task-groups formed thr' the user id of tasks, | |
8189 | * init_task_group represents tasks belonging to root user. | |
8190 | * Hence it forms a sibling of all subsequent groups formed. | |
8191 | * In this case, init_task_group gets only a fraction of overall | |
8192 | * system cpu resource, based on the weight assigned to root | |
8193 | * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished | |
8194 | * by letting tasks of init_task_group sit in a separate cfs_rq | |
8195 | * (init_cfs_rq) and having one entity represent this group of | |
8196 | * tasks in rq->cfs (i.e init_task_group->se[] != NULL). | |
8197 | */ | |
ec7dc8ac | 8198 | init_tg_cfs_entry(&init_task_group, |
6f505b16 | 8199 | &per_cpu(init_cfs_rq, i), |
eff766a6 PZ |
8200 | &per_cpu(init_sched_entity, i), i, 1, |
8201 | root_task_group.se[i]); | |
6f505b16 | 8202 | |
052f1dc7 | 8203 | #endif |
354d60c2 DG |
8204 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
8205 | ||
8206 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; | |
052f1dc7 | 8207 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 8208 | INIT_LIST_HEAD(&rq->leaf_rt_rq_list); |
354d60c2 | 8209 | #ifdef CONFIG_CGROUP_SCHED |
ec7dc8ac | 8210 | init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL); |
354d60c2 | 8211 | #elif defined CONFIG_USER_SCHED |
eff766a6 | 8212 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL); |
ec7dc8ac | 8213 | init_tg_rt_entry(&init_task_group, |
6f505b16 | 8214 | &per_cpu(init_rt_rq, i), |
eff766a6 PZ |
8215 | &per_cpu(init_sched_rt_entity, i), i, 1, |
8216 | root_task_group.rt_se[i]); | |
354d60c2 | 8217 | #endif |
dd41f596 | 8218 | #endif |
1da177e4 | 8219 | |
dd41f596 IM |
8220 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) |
8221 | rq->cpu_load[j] = 0; | |
1da177e4 | 8222 | #ifdef CONFIG_SMP |
41c7ce9a | 8223 | rq->sd = NULL; |
57d885fe | 8224 | rq->rd = NULL; |
1da177e4 | 8225 | rq->active_balance = 0; |
dd41f596 | 8226 | rq->next_balance = jiffies; |
1da177e4 | 8227 | rq->push_cpu = 0; |
0a2966b4 | 8228 | rq->cpu = i; |
1da177e4 LT |
8229 | rq->migration_thread = NULL; |
8230 | INIT_LIST_HEAD(&rq->migration_queue); | |
dc938520 | 8231 | rq_attach_root(rq, &def_root_domain); |
1da177e4 | 8232 | #endif |
8f4d37ec | 8233 | init_rq_hrtick(rq); |
1da177e4 | 8234 | atomic_set(&rq->nr_iowait, 0); |
1da177e4 LT |
8235 | } |
8236 | ||
2dd73a4f | 8237 | set_load_weight(&init_task); |
b50f60ce | 8238 | |
e107be36 AK |
8239 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
8240 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); | |
8241 | #endif | |
8242 | ||
c9819f45 CL |
8243 | #ifdef CONFIG_SMP |
8244 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL); | |
8245 | #endif | |
8246 | ||
b50f60ce HC |
8247 | #ifdef CONFIG_RT_MUTEXES |
8248 | plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); | |
8249 | #endif | |
8250 | ||
1da177e4 LT |
8251 | /* |
8252 | * The boot idle thread does lazy MMU switching as well: | |
8253 | */ | |
8254 | atomic_inc(&init_mm.mm_count); | |
8255 | enter_lazy_tlb(&init_mm, current); | |
8256 | ||
8257 | /* | |
8258 | * Make us the idle thread. Technically, schedule() should not be | |
8259 | * called from this thread, however somewhere below it might be, | |
8260 | * but because we are the idle thread, we just pick up running again | |
8261 | * when this runqueue becomes "idle". | |
8262 | */ | |
8263 | init_idle(current, smp_processor_id()); | |
dd41f596 IM |
8264 | /* |
8265 | * During early bootup we pretend to be a normal task: | |
8266 | */ | |
8267 | current->sched_class = &fair_sched_class; | |
6892b75e IM |
8268 | |
8269 | scheduler_running = 1; | |
1da177e4 LT |
8270 | } |
8271 | ||
8272 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
8273 | void __might_sleep(char *file, int line) | |
8274 | { | |
48f24c4d | 8275 | #ifdef in_atomic |
1da177e4 LT |
8276 | static unsigned long prev_jiffy; /* ratelimiting */ |
8277 | ||
8278 | if ((in_atomic() || irqs_disabled()) && | |
8279 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
8280 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
8281 | return; | |
8282 | prev_jiffy = jiffies; | |
91368d73 | 8283 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
1da177e4 LT |
8284 | " context at %s:%d\n", file, line); |
8285 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
8286 | in_atomic(), irqs_disabled()); | |
a4c410f0 | 8287 | debug_show_held_locks(current); |
3117df04 IM |
8288 | if (irqs_disabled()) |
8289 | print_irqtrace_events(current); | |
1da177e4 LT |
8290 | dump_stack(); |
8291 | } | |
8292 | #endif | |
8293 | } | |
8294 | EXPORT_SYMBOL(__might_sleep); | |
8295 | #endif | |
8296 | ||
8297 | #ifdef CONFIG_MAGIC_SYSRQ | |
3a5e4dc1 AK |
8298 | static void normalize_task(struct rq *rq, struct task_struct *p) |
8299 | { | |
8300 | int on_rq; | |
8301 | update_rq_clock(rq); | |
8302 | on_rq = p->se.on_rq; | |
8303 | if (on_rq) | |
8304 | deactivate_task(rq, p, 0); | |
8305 | __setscheduler(rq, p, SCHED_NORMAL, 0); | |
8306 | if (on_rq) { | |
8307 | activate_task(rq, p, 0); | |
8308 | resched_task(rq->curr); | |
8309 | } | |
8310 | } | |
8311 | ||
1da177e4 LT |
8312 | void normalize_rt_tasks(void) |
8313 | { | |
a0f98a1c | 8314 | struct task_struct *g, *p; |
1da177e4 | 8315 | unsigned long flags; |
70b97a7f | 8316 | struct rq *rq; |
1da177e4 | 8317 | |
4cf5d77a | 8318 | read_lock_irqsave(&tasklist_lock, flags); |
a0f98a1c | 8319 | do_each_thread(g, p) { |
178be793 IM |
8320 | /* |
8321 | * Only normalize user tasks: | |
8322 | */ | |
8323 | if (!p->mm) | |
8324 | continue; | |
8325 | ||
6cfb0d5d | 8326 | p->se.exec_start = 0; |
6cfb0d5d | 8327 | #ifdef CONFIG_SCHEDSTATS |
dd41f596 | 8328 | p->se.wait_start = 0; |
dd41f596 | 8329 | p->se.sleep_start = 0; |
dd41f596 | 8330 | p->se.block_start = 0; |
6cfb0d5d | 8331 | #endif |
dd41f596 IM |
8332 | task_rq(p)->clock = 0; |
8333 | ||
8334 | if (!rt_task(p)) { | |
8335 | /* | |
8336 | * Renice negative nice level userspace | |
8337 | * tasks back to 0: | |
8338 | */ | |
8339 | if (TASK_NICE(p) < 0 && p->mm) | |
8340 | set_user_nice(p, 0); | |
1da177e4 | 8341 | continue; |
dd41f596 | 8342 | } |
1da177e4 | 8343 | |
4cf5d77a | 8344 | spin_lock(&p->pi_lock); |
b29739f9 | 8345 | rq = __task_rq_lock(p); |
1da177e4 | 8346 | |
178be793 | 8347 | normalize_task(rq, p); |
3a5e4dc1 | 8348 | |
b29739f9 | 8349 | __task_rq_unlock(rq); |
4cf5d77a | 8350 | spin_unlock(&p->pi_lock); |
a0f98a1c IM |
8351 | } while_each_thread(g, p); |
8352 | ||
4cf5d77a | 8353 | read_unlock_irqrestore(&tasklist_lock, flags); |
1da177e4 LT |
8354 | } |
8355 | ||
8356 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
8357 | |
8358 | #ifdef CONFIG_IA64 | |
8359 | /* | |
8360 | * These functions are only useful for the IA64 MCA handling. | |
8361 | * | |
8362 | * They can only be called when the whole system has been | |
8363 | * stopped - every CPU needs to be quiescent, and no scheduling | |
8364 | * activity can take place. Using them for anything else would | |
8365 | * be a serious bug, and as a result, they aren't even visible | |
8366 | * under any other configuration. | |
8367 | */ | |
8368 | ||
8369 | /** | |
8370 | * curr_task - return the current task for a given cpu. | |
8371 | * @cpu: the processor in question. | |
8372 | * | |
8373 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8374 | */ | |
36c8b586 | 8375 | struct task_struct *curr_task(int cpu) |
1df5c10a LT |
8376 | { |
8377 | return cpu_curr(cpu); | |
8378 | } | |
8379 | ||
8380 | /** | |
8381 | * set_curr_task - set the current task for a given cpu. | |
8382 | * @cpu: the processor in question. | |
8383 | * @p: the task pointer to set. | |
8384 | * | |
8385 | * Description: This function must only be used when non-maskable interrupts | |
41a2d6cf IM |
8386 | * are serviced on a separate stack. It allows the architecture to switch the |
8387 | * notion of the current task on a cpu in a non-blocking manner. This function | |
1df5c10a LT |
8388 | * must be called with all CPU's synchronized, and interrupts disabled, the |
8389 | * and caller must save the original value of the current task (see | |
8390 | * curr_task() above) and restore that value before reenabling interrupts and | |
8391 | * re-starting the system. | |
8392 | * | |
8393 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8394 | */ | |
36c8b586 | 8395 | void set_curr_task(int cpu, struct task_struct *p) |
1df5c10a LT |
8396 | { |
8397 | cpu_curr(cpu) = p; | |
8398 | } | |
8399 | ||
8400 | #endif | |
29f59db3 | 8401 | |
bccbe08a PZ |
8402 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8403 | static void free_fair_sched_group(struct task_group *tg) | |
6f505b16 PZ |
8404 | { |
8405 | int i; | |
8406 | ||
8407 | for_each_possible_cpu(i) { | |
8408 | if (tg->cfs_rq) | |
8409 | kfree(tg->cfs_rq[i]); | |
8410 | if (tg->se) | |
8411 | kfree(tg->se[i]); | |
6f505b16 PZ |
8412 | } |
8413 | ||
8414 | kfree(tg->cfs_rq); | |
8415 | kfree(tg->se); | |
6f505b16 PZ |
8416 | } |
8417 | ||
ec7dc8ac DG |
8418 | static |
8419 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
29f59db3 | 8420 | { |
29f59db3 | 8421 | struct cfs_rq *cfs_rq; |
ec7dc8ac | 8422 | struct sched_entity *se, *parent_se; |
9b5b7751 | 8423 | struct rq *rq; |
29f59db3 SV |
8424 | int i; |
8425 | ||
434d53b0 | 8426 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); |
29f59db3 SV |
8427 | if (!tg->cfs_rq) |
8428 | goto err; | |
434d53b0 | 8429 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); |
29f59db3 SV |
8430 | if (!tg->se) |
8431 | goto err; | |
052f1dc7 PZ |
8432 | |
8433 | tg->shares = NICE_0_LOAD; | |
29f59db3 SV |
8434 | |
8435 | for_each_possible_cpu(i) { | |
9b5b7751 | 8436 | rq = cpu_rq(i); |
29f59db3 | 8437 | |
6f505b16 PZ |
8438 | cfs_rq = kmalloc_node(sizeof(struct cfs_rq), |
8439 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
29f59db3 SV |
8440 | if (!cfs_rq) |
8441 | goto err; | |
8442 | ||
6f505b16 PZ |
8443 | se = kmalloc_node(sizeof(struct sched_entity), |
8444 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
29f59db3 SV |
8445 | if (!se) |
8446 | goto err; | |
8447 | ||
ec7dc8ac DG |
8448 | parent_se = parent ? parent->se[i] : NULL; |
8449 | init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se); | |
bccbe08a PZ |
8450 | } |
8451 | ||
8452 | return 1; | |
8453 | ||
8454 | err: | |
8455 | return 0; | |
8456 | } | |
8457 | ||
8458 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8459 | { | |
8460 | list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list, | |
8461 | &cpu_rq(cpu)->leaf_cfs_rq_list); | |
8462 | } | |
8463 | ||
8464 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8465 | { | |
8466 | list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list); | |
8467 | } | |
8468 | #else | |
8469 | static inline void free_fair_sched_group(struct task_group *tg) | |
8470 | { | |
8471 | } | |
8472 | ||
ec7dc8ac DG |
8473 | static inline |
8474 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
bccbe08a PZ |
8475 | { |
8476 | return 1; | |
8477 | } | |
8478 | ||
8479 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8480 | { | |
8481 | } | |
8482 | ||
8483 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8484 | { | |
8485 | } | |
052f1dc7 PZ |
8486 | #endif |
8487 | ||
8488 | #ifdef CONFIG_RT_GROUP_SCHED | |
bccbe08a PZ |
8489 | static void free_rt_sched_group(struct task_group *tg) |
8490 | { | |
8491 | int i; | |
8492 | ||
d0b27fa7 PZ |
8493 | destroy_rt_bandwidth(&tg->rt_bandwidth); |
8494 | ||
bccbe08a PZ |
8495 | for_each_possible_cpu(i) { |
8496 | if (tg->rt_rq) | |
8497 | kfree(tg->rt_rq[i]); | |
8498 | if (tg->rt_se) | |
8499 | kfree(tg->rt_se[i]); | |
8500 | } | |
8501 | ||
8502 | kfree(tg->rt_rq); | |
8503 | kfree(tg->rt_se); | |
8504 | } | |
8505 | ||
ec7dc8ac DG |
8506 | static |
8507 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
bccbe08a PZ |
8508 | { |
8509 | struct rt_rq *rt_rq; | |
ec7dc8ac | 8510 | struct sched_rt_entity *rt_se, *parent_se; |
bccbe08a PZ |
8511 | struct rq *rq; |
8512 | int i; | |
8513 | ||
434d53b0 | 8514 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); |
bccbe08a PZ |
8515 | if (!tg->rt_rq) |
8516 | goto err; | |
434d53b0 | 8517 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); |
bccbe08a PZ |
8518 | if (!tg->rt_se) |
8519 | goto err; | |
8520 | ||
d0b27fa7 PZ |
8521 | init_rt_bandwidth(&tg->rt_bandwidth, |
8522 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); | |
bccbe08a PZ |
8523 | |
8524 | for_each_possible_cpu(i) { | |
8525 | rq = cpu_rq(i); | |
8526 | ||
6f505b16 PZ |
8527 | rt_rq = kmalloc_node(sizeof(struct rt_rq), |
8528 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8529 | if (!rt_rq) | |
8530 | goto err; | |
29f59db3 | 8531 | |
6f505b16 PZ |
8532 | rt_se = kmalloc_node(sizeof(struct sched_rt_entity), |
8533 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8534 | if (!rt_se) | |
8535 | goto err; | |
29f59db3 | 8536 | |
ec7dc8ac DG |
8537 | parent_se = parent ? parent->rt_se[i] : NULL; |
8538 | init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se); | |
29f59db3 SV |
8539 | } |
8540 | ||
bccbe08a PZ |
8541 | return 1; |
8542 | ||
8543 | err: | |
8544 | return 0; | |
8545 | } | |
8546 | ||
8547 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8548 | { | |
8549 | list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list, | |
8550 | &cpu_rq(cpu)->leaf_rt_rq_list); | |
8551 | } | |
8552 | ||
8553 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8554 | { | |
8555 | list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list); | |
8556 | } | |
8557 | #else | |
8558 | static inline void free_rt_sched_group(struct task_group *tg) | |
8559 | { | |
8560 | } | |
8561 | ||
ec7dc8ac DG |
8562 | static inline |
8563 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
bccbe08a PZ |
8564 | { |
8565 | return 1; | |
8566 | } | |
8567 | ||
8568 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8569 | { | |
8570 | } | |
8571 | ||
8572 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8573 | { | |
8574 | } | |
8575 | #endif | |
8576 | ||
d0b27fa7 | 8577 | #ifdef CONFIG_GROUP_SCHED |
bccbe08a PZ |
8578 | static void free_sched_group(struct task_group *tg) |
8579 | { | |
8580 | free_fair_sched_group(tg); | |
8581 | free_rt_sched_group(tg); | |
8582 | kfree(tg); | |
8583 | } | |
8584 | ||
8585 | /* allocate runqueue etc for a new task group */ | |
ec7dc8ac | 8586 | struct task_group *sched_create_group(struct task_group *parent) |
bccbe08a PZ |
8587 | { |
8588 | struct task_group *tg; | |
8589 | unsigned long flags; | |
8590 | int i; | |
8591 | ||
8592 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); | |
8593 | if (!tg) | |
8594 | return ERR_PTR(-ENOMEM); | |
8595 | ||
ec7dc8ac | 8596 | if (!alloc_fair_sched_group(tg, parent)) |
bccbe08a PZ |
8597 | goto err; |
8598 | ||
ec7dc8ac | 8599 | if (!alloc_rt_sched_group(tg, parent)) |
bccbe08a PZ |
8600 | goto err; |
8601 | ||
8ed36996 | 8602 | spin_lock_irqsave(&task_group_lock, flags); |
9b5b7751 | 8603 | for_each_possible_cpu(i) { |
bccbe08a PZ |
8604 | register_fair_sched_group(tg, i); |
8605 | register_rt_sched_group(tg, i); | |
9b5b7751 | 8606 | } |
6f505b16 | 8607 | list_add_rcu(&tg->list, &task_groups); |
f473aa5e PZ |
8608 | |
8609 | WARN_ON(!parent); /* root should already exist */ | |
8610 | ||
8611 | tg->parent = parent; | |
8612 | list_add_rcu(&tg->siblings, &parent->children); | |
8613 | INIT_LIST_HEAD(&tg->children); | |
8ed36996 | 8614 | spin_unlock_irqrestore(&task_group_lock, flags); |
29f59db3 | 8615 | |
9b5b7751 | 8616 | return tg; |
29f59db3 SV |
8617 | |
8618 | err: | |
6f505b16 | 8619 | free_sched_group(tg); |
29f59db3 SV |
8620 | return ERR_PTR(-ENOMEM); |
8621 | } | |
8622 | ||
9b5b7751 | 8623 | /* rcu callback to free various structures associated with a task group */ |
6f505b16 | 8624 | static void free_sched_group_rcu(struct rcu_head *rhp) |
29f59db3 | 8625 | { |
29f59db3 | 8626 | /* now it should be safe to free those cfs_rqs */ |
6f505b16 | 8627 | free_sched_group(container_of(rhp, struct task_group, rcu)); |
29f59db3 SV |
8628 | } |
8629 | ||
9b5b7751 | 8630 | /* Destroy runqueue etc associated with a task group */ |
4cf86d77 | 8631 | void sched_destroy_group(struct task_group *tg) |
29f59db3 | 8632 | { |
8ed36996 | 8633 | unsigned long flags; |
9b5b7751 | 8634 | int i; |
29f59db3 | 8635 | |
8ed36996 | 8636 | spin_lock_irqsave(&task_group_lock, flags); |
9b5b7751 | 8637 | for_each_possible_cpu(i) { |
bccbe08a PZ |
8638 | unregister_fair_sched_group(tg, i); |
8639 | unregister_rt_sched_group(tg, i); | |
9b5b7751 | 8640 | } |
6f505b16 | 8641 | list_del_rcu(&tg->list); |
f473aa5e | 8642 | list_del_rcu(&tg->siblings); |
8ed36996 | 8643 | spin_unlock_irqrestore(&task_group_lock, flags); |
9b5b7751 | 8644 | |
9b5b7751 | 8645 | /* wait for possible concurrent references to cfs_rqs complete */ |
6f505b16 | 8646 | call_rcu(&tg->rcu, free_sched_group_rcu); |
29f59db3 SV |
8647 | } |
8648 | ||
9b5b7751 | 8649 | /* change task's runqueue when it moves between groups. |
3a252015 IM |
8650 | * The caller of this function should have put the task in its new group |
8651 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to | |
8652 | * reflect its new group. | |
9b5b7751 SV |
8653 | */ |
8654 | void sched_move_task(struct task_struct *tsk) | |
29f59db3 SV |
8655 | { |
8656 | int on_rq, running; | |
8657 | unsigned long flags; | |
8658 | struct rq *rq; | |
8659 | ||
8660 | rq = task_rq_lock(tsk, &flags); | |
8661 | ||
29f59db3 SV |
8662 | update_rq_clock(rq); |
8663 | ||
051a1d1a | 8664 | running = task_current(rq, tsk); |
29f59db3 SV |
8665 | on_rq = tsk->se.on_rq; |
8666 | ||
0e1f3483 | 8667 | if (on_rq) |
29f59db3 | 8668 | dequeue_task(rq, tsk, 0); |
0e1f3483 HS |
8669 | if (unlikely(running)) |
8670 | tsk->sched_class->put_prev_task(rq, tsk); | |
29f59db3 | 8671 | |
6f505b16 | 8672 | set_task_rq(tsk, task_cpu(tsk)); |
29f59db3 | 8673 | |
810b3817 PZ |
8674 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8675 | if (tsk->sched_class->moved_group) | |
8676 | tsk->sched_class->moved_group(tsk); | |
8677 | #endif | |
8678 | ||
0e1f3483 HS |
8679 | if (unlikely(running)) |
8680 | tsk->sched_class->set_curr_task(rq); | |
8681 | if (on_rq) | |
7074badb | 8682 | enqueue_task(rq, tsk, 0); |
29f59db3 | 8683 | |
29f59db3 SV |
8684 | task_rq_unlock(rq, &flags); |
8685 | } | |
d0b27fa7 | 8686 | #endif |
29f59db3 | 8687 | |
052f1dc7 | 8688 | #ifdef CONFIG_FAIR_GROUP_SCHED |
18d95a28 | 8689 | static void __set_se_shares(struct sched_entity *se, unsigned long shares) |
29f59db3 SV |
8690 | { |
8691 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
29f59db3 SV |
8692 | int on_rq; |
8693 | ||
29f59db3 | 8694 | on_rq = se->on_rq; |
62fb1851 | 8695 | if (on_rq) |
29f59db3 SV |
8696 | dequeue_entity(cfs_rq, se, 0); |
8697 | ||
8698 | se->load.weight = shares; | |
e05510d0 | 8699 | se->load.inv_weight = 0; |
29f59db3 | 8700 | |
62fb1851 | 8701 | if (on_rq) |
29f59db3 | 8702 | enqueue_entity(cfs_rq, se, 0); |
18d95a28 | 8703 | } |
62fb1851 | 8704 | |
18d95a28 PZ |
8705 | static void set_se_shares(struct sched_entity *se, unsigned long shares) |
8706 | { | |
8707 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
8708 | struct rq *rq = cfs_rq->rq; | |
8709 | unsigned long flags; | |
8710 | ||
8711 | spin_lock_irqsave(&rq->lock, flags); | |
8712 | __set_se_shares(se, shares); | |
8713 | spin_unlock_irqrestore(&rq->lock, flags); | |
29f59db3 SV |
8714 | } |
8715 | ||
8ed36996 PZ |
8716 | static DEFINE_MUTEX(shares_mutex); |
8717 | ||
4cf86d77 | 8718 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) |
29f59db3 SV |
8719 | { |
8720 | int i; | |
8ed36996 | 8721 | unsigned long flags; |
c61935fd | 8722 | |
ec7dc8ac DG |
8723 | /* |
8724 | * We can't change the weight of the root cgroup. | |
8725 | */ | |
8726 | if (!tg->se[0]) | |
8727 | return -EINVAL; | |
8728 | ||
62fb1851 PZ |
8729 | /* |
8730 | * A weight of 0 or 1 can cause arithmetics problems. | |
8731 | * (The default weight is 1024 - so there's no practical | |
8732 | * limitation from this.) | |
8733 | */ | |
18d95a28 PZ |
8734 | if (shares < MIN_SHARES) |
8735 | shares = MIN_SHARES; | |
62fb1851 | 8736 | |
8ed36996 | 8737 | mutex_lock(&shares_mutex); |
9b5b7751 | 8738 | if (tg->shares == shares) |
5cb350ba | 8739 | goto done; |
29f59db3 | 8740 | |
8ed36996 | 8741 | spin_lock_irqsave(&task_group_lock, flags); |
bccbe08a PZ |
8742 | for_each_possible_cpu(i) |
8743 | unregister_fair_sched_group(tg, i); | |
f473aa5e | 8744 | list_del_rcu(&tg->siblings); |
8ed36996 | 8745 | spin_unlock_irqrestore(&task_group_lock, flags); |
6b2d7700 SV |
8746 | |
8747 | /* wait for any ongoing reference to this group to finish */ | |
8748 | synchronize_sched(); | |
8749 | ||
8750 | /* | |
8751 | * Now we are free to modify the group's share on each cpu | |
8752 | * w/o tripping rebalance_share or load_balance_fair. | |
8753 | */ | |
9b5b7751 | 8754 | tg->shares = shares; |
18d95a28 PZ |
8755 | for_each_possible_cpu(i) { |
8756 | /* | |
8757 | * force a rebalance | |
8758 | */ | |
8759 | cfs_rq_set_shares(tg->cfs_rq[i], 0); | |
8760 | set_se_shares(tg->se[i], shares/nr_cpu_ids); | |
8761 | } | |
29f59db3 | 8762 | |
6b2d7700 SV |
8763 | /* |
8764 | * Enable load balance activity on this group, by inserting it back on | |
8765 | * each cpu's rq->leaf_cfs_rq_list. | |
8766 | */ | |
8ed36996 | 8767 | spin_lock_irqsave(&task_group_lock, flags); |
bccbe08a PZ |
8768 | for_each_possible_cpu(i) |
8769 | register_fair_sched_group(tg, i); | |
f473aa5e | 8770 | list_add_rcu(&tg->siblings, &tg->parent->children); |
8ed36996 | 8771 | spin_unlock_irqrestore(&task_group_lock, flags); |
5cb350ba | 8772 | done: |
8ed36996 | 8773 | mutex_unlock(&shares_mutex); |
9b5b7751 | 8774 | return 0; |
29f59db3 SV |
8775 | } |
8776 | ||
5cb350ba DG |
8777 | unsigned long sched_group_shares(struct task_group *tg) |
8778 | { | |
8779 | return tg->shares; | |
8780 | } | |
052f1dc7 | 8781 | #endif |
5cb350ba | 8782 | |
052f1dc7 | 8783 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 8784 | /* |
9f0c1e56 | 8785 | * Ensure that the real time constraints are schedulable. |
6f505b16 | 8786 | */ |
9f0c1e56 PZ |
8787 | static DEFINE_MUTEX(rt_constraints_mutex); |
8788 | ||
8789 | static unsigned long to_ratio(u64 period, u64 runtime) | |
8790 | { | |
8791 | if (runtime == RUNTIME_INF) | |
8792 | return 1ULL << 16; | |
8793 | ||
6f6d6a1a | 8794 | return div64_u64(runtime << 16, period); |
9f0c1e56 PZ |
8795 | } |
8796 | ||
b40b2e8e PZ |
8797 | #ifdef CONFIG_CGROUP_SCHED |
8798 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
8799 | { | |
8800 | struct task_group *tgi, *parent = tg->parent; | |
8801 | unsigned long total = 0; | |
8802 | ||
8803 | if (!parent) { | |
8804 | if (global_rt_period() < period) | |
8805 | return 0; | |
8806 | ||
8807 | return to_ratio(period, runtime) < | |
8808 | to_ratio(global_rt_period(), global_rt_runtime()); | |
8809 | } | |
8810 | ||
8811 | if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period) | |
8812 | return 0; | |
8813 | ||
8814 | rcu_read_lock(); | |
8815 | list_for_each_entry_rcu(tgi, &parent->children, siblings) { | |
8816 | if (tgi == tg) | |
8817 | continue; | |
8818 | ||
8819 | total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period), | |
8820 | tgi->rt_bandwidth.rt_runtime); | |
8821 | } | |
8822 | rcu_read_unlock(); | |
8823 | ||
8824 | return total + to_ratio(period, runtime) < | |
8825 | to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period), | |
8826 | parent->rt_bandwidth.rt_runtime); | |
8827 | } | |
8828 | #elif defined CONFIG_USER_SCHED | |
9f0c1e56 | 8829 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) |
6f505b16 PZ |
8830 | { |
8831 | struct task_group *tgi; | |
8832 | unsigned long total = 0; | |
9f0c1e56 | 8833 | unsigned long global_ratio = |
d0b27fa7 | 8834 | to_ratio(global_rt_period(), global_rt_runtime()); |
6f505b16 PZ |
8835 | |
8836 | rcu_read_lock(); | |
9f0c1e56 PZ |
8837 | list_for_each_entry_rcu(tgi, &task_groups, list) { |
8838 | if (tgi == tg) | |
8839 | continue; | |
6f505b16 | 8840 | |
d0b27fa7 PZ |
8841 | total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period), |
8842 | tgi->rt_bandwidth.rt_runtime); | |
9f0c1e56 PZ |
8843 | } |
8844 | rcu_read_unlock(); | |
6f505b16 | 8845 | |
9f0c1e56 | 8846 | return total + to_ratio(period, runtime) < global_ratio; |
6f505b16 | 8847 | } |
b40b2e8e | 8848 | #endif |
6f505b16 | 8849 | |
521f1a24 DG |
8850 | /* Must be called with tasklist_lock held */ |
8851 | static inline int tg_has_rt_tasks(struct task_group *tg) | |
8852 | { | |
8853 | struct task_struct *g, *p; | |
8854 | do_each_thread(g, p) { | |
8855 | if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg) | |
8856 | return 1; | |
8857 | } while_each_thread(g, p); | |
8858 | return 0; | |
8859 | } | |
8860 | ||
d0b27fa7 PZ |
8861 | static int tg_set_bandwidth(struct task_group *tg, |
8862 | u64 rt_period, u64 rt_runtime) | |
6f505b16 | 8863 | { |
ac086bc2 | 8864 | int i, err = 0; |
9f0c1e56 | 8865 | |
9f0c1e56 | 8866 | mutex_lock(&rt_constraints_mutex); |
521f1a24 | 8867 | read_lock(&tasklist_lock); |
ac086bc2 | 8868 | if (rt_runtime == 0 && tg_has_rt_tasks(tg)) { |
521f1a24 DG |
8869 | err = -EBUSY; |
8870 | goto unlock; | |
8871 | } | |
9f0c1e56 PZ |
8872 | if (!__rt_schedulable(tg, rt_period, rt_runtime)) { |
8873 | err = -EINVAL; | |
8874 | goto unlock; | |
8875 | } | |
ac086bc2 PZ |
8876 | |
8877 | spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
d0b27fa7 PZ |
8878 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); |
8879 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
ac086bc2 PZ |
8880 | |
8881 | for_each_possible_cpu(i) { | |
8882 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
8883 | ||
8884 | spin_lock(&rt_rq->rt_runtime_lock); | |
8885 | rt_rq->rt_runtime = rt_runtime; | |
8886 | spin_unlock(&rt_rq->rt_runtime_lock); | |
8887 | } | |
8888 | spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
9f0c1e56 | 8889 | unlock: |
521f1a24 | 8890 | read_unlock(&tasklist_lock); |
9f0c1e56 PZ |
8891 | mutex_unlock(&rt_constraints_mutex); |
8892 | ||
8893 | return err; | |
6f505b16 PZ |
8894 | } |
8895 | ||
d0b27fa7 PZ |
8896 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) |
8897 | { | |
8898 | u64 rt_runtime, rt_period; | |
8899 | ||
8900 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8901 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
8902 | if (rt_runtime_us < 0) | |
8903 | rt_runtime = RUNTIME_INF; | |
8904 | ||
8905 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8906 | } | |
8907 | ||
9f0c1e56 PZ |
8908 | long sched_group_rt_runtime(struct task_group *tg) |
8909 | { | |
8910 | u64 rt_runtime_us; | |
8911 | ||
d0b27fa7 | 8912 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) |
9f0c1e56 PZ |
8913 | return -1; |
8914 | ||
d0b27fa7 | 8915 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; |
9f0c1e56 PZ |
8916 | do_div(rt_runtime_us, NSEC_PER_USEC); |
8917 | return rt_runtime_us; | |
8918 | } | |
d0b27fa7 PZ |
8919 | |
8920 | int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) | |
8921 | { | |
8922 | u64 rt_runtime, rt_period; | |
8923 | ||
8924 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; | |
8925 | rt_runtime = tg->rt_bandwidth.rt_runtime; | |
8926 | ||
8927 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8928 | } | |
8929 | ||
8930 | long sched_group_rt_period(struct task_group *tg) | |
8931 | { | |
8932 | u64 rt_period_us; | |
8933 | ||
8934 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8935 | do_div(rt_period_us, NSEC_PER_USEC); | |
8936 | return rt_period_us; | |
8937 | } | |
8938 | ||
8939 | static int sched_rt_global_constraints(void) | |
8940 | { | |
8941 | int ret = 0; | |
8942 | ||
8943 | mutex_lock(&rt_constraints_mutex); | |
8944 | if (!__rt_schedulable(NULL, 1, 0)) | |
8945 | ret = -EINVAL; | |
8946 | mutex_unlock(&rt_constraints_mutex); | |
8947 | ||
8948 | return ret; | |
8949 | } | |
8950 | #else | |
8951 | static int sched_rt_global_constraints(void) | |
8952 | { | |
ac086bc2 PZ |
8953 | unsigned long flags; |
8954 | int i; | |
8955 | ||
8956 | spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8957 | for_each_possible_cpu(i) { | |
8958 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
8959 | ||
8960 | spin_lock(&rt_rq->rt_runtime_lock); | |
8961 | rt_rq->rt_runtime = global_rt_runtime(); | |
8962 | spin_unlock(&rt_rq->rt_runtime_lock); | |
8963 | } | |
8964 | spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8965 | ||
d0b27fa7 PZ |
8966 | return 0; |
8967 | } | |
052f1dc7 | 8968 | #endif |
d0b27fa7 PZ |
8969 | |
8970 | int sched_rt_handler(struct ctl_table *table, int write, | |
8971 | struct file *filp, void __user *buffer, size_t *lenp, | |
8972 | loff_t *ppos) | |
8973 | { | |
8974 | int ret; | |
8975 | int old_period, old_runtime; | |
8976 | static DEFINE_MUTEX(mutex); | |
8977 | ||
8978 | mutex_lock(&mutex); | |
8979 | old_period = sysctl_sched_rt_period; | |
8980 | old_runtime = sysctl_sched_rt_runtime; | |
8981 | ||
8982 | ret = proc_dointvec(table, write, filp, buffer, lenp, ppos); | |
8983 | ||
8984 | if (!ret && write) { | |
8985 | ret = sched_rt_global_constraints(); | |
8986 | if (ret) { | |
8987 | sysctl_sched_rt_period = old_period; | |
8988 | sysctl_sched_rt_runtime = old_runtime; | |
8989 | } else { | |
8990 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
8991 | def_rt_bandwidth.rt_period = | |
8992 | ns_to_ktime(global_rt_period()); | |
8993 | } | |
8994 | } | |
8995 | mutex_unlock(&mutex); | |
8996 | ||
8997 | return ret; | |
8998 | } | |
68318b8e | 8999 | |
052f1dc7 | 9000 | #ifdef CONFIG_CGROUP_SCHED |
68318b8e SV |
9001 | |
9002 | /* return corresponding task_group object of a cgroup */ | |
2b01dfe3 | 9003 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) |
68318b8e | 9004 | { |
2b01dfe3 PM |
9005 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), |
9006 | struct task_group, css); | |
68318b8e SV |
9007 | } |
9008 | ||
9009 | static struct cgroup_subsys_state * | |
2b01dfe3 | 9010 | cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) |
68318b8e | 9011 | { |
ec7dc8ac | 9012 | struct task_group *tg, *parent; |
68318b8e | 9013 | |
2b01dfe3 | 9014 | if (!cgrp->parent) { |
68318b8e | 9015 | /* This is early initialization for the top cgroup */ |
2b01dfe3 | 9016 | init_task_group.css.cgroup = cgrp; |
68318b8e SV |
9017 | return &init_task_group.css; |
9018 | } | |
9019 | ||
ec7dc8ac DG |
9020 | parent = cgroup_tg(cgrp->parent); |
9021 | tg = sched_create_group(parent); | |
68318b8e SV |
9022 | if (IS_ERR(tg)) |
9023 | return ERR_PTR(-ENOMEM); | |
9024 | ||
9025 | /* Bind the cgroup to task_group object we just created */ | |
2b01dfe3 | 9026 | tg->css.cgroup = cgrp; |
68318b8e SV |
9027 | |
9028 | return &tg->css; | |
9029 | } | |
9030 | ||
41a2d6cf IM |
9031 | static void |
9032 | cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
68318b8e | 9033 | { |
2b01dfe3 | 9034 | struct task_group *tg = cgroup_tg(cgrp); |
68318b8e SV |
9035 | |
9036 | sched_destroy_group(tg); | |
9037 | } | |
9038 | ||
41a2d6cf IM |
9039 | static int |
9040 | cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, | |
9041 | struct task_struct *tsk) | |
68318b8e | 9042 | { |
b68aa230 PZ |
9043 | #ifdef CONFIG_RT_GROUP_SCHED |
9044 | /* Don't accept realtime tasks when there is no way for them to run */ | |
d0b27fa7 | 9045 | if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0) |
b68aa230 PZ |
9046 | return -EINVAL; |
9047 | #else | |
68318b8e SV |
9048 | /* We don't support RT-tasks being in separate groups */ |
9049 | if (tsk->sched_class != &fair_sched_class) | |
9050 | return -EINVAL; | |
b68aa230 | 9051 | #endif |
68318b8e SV |
9052 | |
9053 | return 0; | |
9054 | } | |
9055 | ||
9056 | static void | |
2b01dfe3 | 9057 | cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
68318b8e SV |
9058 | struct cgroup *old_cont, struct task_struct *tsk) |
9059 | { | |
9060 | sched_move_task(tsk); | |
9061 | } | |
9062 | ||
052f1dc7 | 9063 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f4c753b7 | 9064 | static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, |
2b01dfe3 | 9065 | u64 shareval) |
68318b8e | 9066 | { |
2b01dfe3 | 9067 | return sched_group_set_shares(cgroup_tg(cgrp), shareval); |
68318b8e SV |
9068 | } |
9069 | ||
f4c753b7 | 9070 | static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) |
68318b8e | 9071 | { |
2b01dfe3 | 9072 | struct task_group *tg = cgroup_tg(cgrp); |
68318b8e SV |
9073 | |
9074 | return (u64) tg->shares; | |
9075 | } | |
052f1dc7 | 9076 | #endif |
68318b8e | 9077 | |
052f1dc7 | 9078 | #ifdef CONFIG_RT_GROUP_SCHED |
ac086bc2 | 9079 | static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, |
06ecb27c | 9080 | s64 val) |
6f505b16 | 9081 | { |
06ecb27c | 9082 | return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); |
6f505b16 PZ |
9083 | } |
9084 | ||
06ecb27c | 9085 | static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) |
6f505b16 | 9086 | { |
06ecb27c | 9087 | return sched_group_rt_runtime(cgroup_tg(cgrp)); |
6f505b16 | 9088 | } |
d0b27fa7 PZ |
9089 | |
9090 | static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, | |
9091 | u64 rt_period_us) | |
9092 | { | |
9093 | return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); | |
9094 | } | |
9095 | ||
9096 | static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) | |
9097 | { | |
9098 | return sched_group_rt_period(cgroup_tg(cgrp)); | |
9099 | } | |
052f1dc7 | 9100 | #endif |
6f505b16 | 9101 | |
fe5c7cc2 | 9102 | static struct cftype cpu_files[] = { |
052f1dc7 | 9103 | #ifdef CONFIG_FAIR_GROUP_SCHED |
fe5c7cc2 PM |
9104 | { |
9105 | .name = "shares", | |
f4c753b7 PM |
9106 | .read_u64 = cpu_shares_read_u64, |
9107 | .write_u64 = cpu_shares_write_u64, | |
fe5c7cc2 | 9108 | }, |
052f1dc7 PZ |
9109 | #endif |
9110 | #ifdef CONFIG_RT_GROUP_SCHED | |
6f505b16 | 9111 | { |
9f0c1e56 | 9112 | .name = "rt_runtime_us", |
06ecb27c PM |
9113 | .read_s64 = cpu_rt_runtime_read, |
9114 | .write_s64 = cpu_rt_runtime_write, | |
6f505b16 | 9115 | }, |
d0b27fa7 PZ |
9116 | { |
9117 | .name = "rt_period_us", | |
f4c753b7 PM |
9118 | .read_u64 = cpu_rt_period_read_uint, |
9119 | .write_u64 = cpu_rt_period_write_uint, | |
d0b27fa7 | 9120 | }, |
052f1dc7 | 9121 | #endif |
68318b8e SV |
9122 | }; |
9123 | ||
9124 | static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) | |
9125 | { | |
fe5c7cc2 | 9126 | return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); |
68318b8e SV |
9127 | } |
9128 | ||
9129 | struct cgroup_subsys cpu_cgroup_subsys = { | |
38605cae IM |
9130 | .name = "cpu", |
9131 | .create = cpu_cgroup_create, | |
9132 | .destroy = cpu_cgroup_destroy, | |
9133 | .can_attach = cpu_cgroup_can_attach, | |
9134 | .attach = cpu_cgroup_attach, | |
9135 | .populate = cpu_cgroup_populate, | |
9136 | .subsys_id = cpu_cgroup_subsys_id, | |
68318b8e SV |
9137 | .early_init = 1, |
9138 | }; | |
9139 | ||
052f1dc7 | 9140 | #endif /* CONFIG_CGROUP_SCHED */ |
d842de87 SV |
9141 | |
9142 | #ifdef CONFIG_CGROUP_CPUACCT | |
9143 | ||
9144 | /* | |
9145 | * CPU accounting code for task groups. | |
9146 | * | |
9147 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh | |
9148 | * (balbir@in.ibm.com). | |
9149 | */ | |
9150 | ||
9151 | /* track cpu usage of a group of tasks */ | |
9152 | struct cpuacct { | |
9153 | struct cgroup_subsys_state css; | |
9154 | /* cpuusage holds pointer to a u64-type object on every cpu */ | |
9155 | u64 *cpuusage; | |
9156 | }; | |
9157 | ||
9158 | struct cgroup_subsys cpuacct_subsys; | |
9159 | ||
9160 | /* return cpu accounting group corresponding to this container */ | |
32cd756a | 9161 | static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) |
d842de87 | 9162 | { |
32cd756a | 9163 | return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), |
d842de87 SV |
9164 | struct cpuacct, css); |
9165 | } | |
9166 | ||
9167 | /* return cpu accounting group to which this task belongs */ | |
9168 | static inline struct cpuacct *task_ca(struct task_struct *tsk) | |
9169 | { | |
9170 | return container_of(task_subsys_state(tsk, cpuacct_subsys_id), | |
9171 | struct cpuacct, css); | |
9172 | } | |
9173 | ||
9174 | /* create a new cpu accounting group */ | |
9175 | static struct cgroup_subsys_state *cpuacct_create( | |
32cd756a | 9176 | struct cgroup_subsys *ss, struct cgroup *cgrp) |
d842de87 SV |
9177 | { |
9178 | struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); | |
9179 | ||
9180 | if (!ca) | |
9181 | return ERR_PTR(-ENOMEM); | |
9182 | ||
9183 | ca->cpuusage = alloc_percpu(u64); | |
9184 | if (!ca->cpuusage) { | |
9185 | kfree(ca); | |
9186 | return ERR_PTR(-ENOMEM); | |
9187 | } | |
9188 | ||
9189 | return &ca->css; | |
9190 | } | |
9191 | ||
9192 | /* destroy an existing cpu accounting group */ | |
41a2d6cf | 9193 | static void |
32cd756a | 9194 | cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) |
d842de87 | 9195 | { |
32cd756a | 9196 | struct cpuacct *ca = cgroup_ca(cgrp); |
d842de87 SV |
9197 | |
9198 | free_percpu(ca->cpuusage); | |
9199 | kfree(ca); | |
9200 | } | |
9201 | ||
9202 | /* return total cpu usage (in nanoseconds) of a group */ | |
32cd756a | 9203 | static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) |
d842de87 | 9204 | { |
32cd756a | 9205 | struct cpuacct *ca = cgroup_ca(cgrp); |
d842de87 SV |
9206 | u64 totalcpuusage = 0; |
9207 | int i; | |
9208 | ||
9209 | for_each_possible_cpu(i) { | |
9210 | u64 *cpuusage = percpu_ptr(ca->cpuusage, i); | |
9211 | ||
9212 | /* | |
9213 | * Take rq->lock to make 64-bit addition safe on 32-bit | |
9214 | * platforms. | |
9215 | */ | |
9216 | spin_lock_irq(&cpu_rq(i)->lock); | |
9217 | totalcpuusage += *cpuusage; | |
9218 | spin_unlock_irq(&cpu_rq(i)->lock); | |
9219 | } | |
9220 | ||
9221 | return totalcpuusage; | |
9222 | } | |
9223 | ||
0297b803 DG |
9224 | static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, |
9225 | u64 reset) | |
9226 | { | |
9227 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9228 | int err = 0; | |
9229 | int i; | |
9230 | ||
9231 | if (reset) { | |
9232 | err = -EINVAL; | |
9233 | goto out; | |
9234 | } | |
9235 | ||
9236 | for_each_possible_cpu(i) { | |
9237 | u64 *cpuusage = percpu_ptr(ca->cpuusage, i); | |
9238 | ||
9239 | spin_lock_irq(&cpu_rq(i)->lock); | |
9240 | *cpuusage = 0; | |
9241 | spin_unlock_irq(&cpu_rq(i)->lock); | |
9242 | } | |
9243 | out: | |
9244 | return err; | |
9245 | } | |
9246 | ||
d842de87 SV |
9247 | static struct cftype files[] = { |
9248 | { | |
9249 | .name = "usage", | |
f4c753b7 PM |
9250 | .read_u64 = cpuusage_read, |
9251 | .write_u64 = cpuusage_write, | |
d842de87 SV |
9252 | }, |
9253 | }; | |
9254 | ||
32cd756a | 9255 | static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) |
d842de87 | 9256 | { |
32cd756a | 9257 | return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); |
d842de87 SV |
9258 | } |
9259 | ||
9260 | /* | |
9261 | * charge this task's execution time to its accounting group. | |
9262 | * | |
9263 | * called with rq->lock held. | |
9264 | */ | |
9265 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime) | |
9266 | { | |
9267 | struct cpuacct *ca; | |
9268 | ||
9269 | if (!cpuacct_subsys.active) | |
9270 | return; | |
9271 | ||
9272 | ca = task_ca(tsk); | |
9273 | if (ca) { | |
9274 | u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk)); | |
9275 | ||
9276 | *cpuusage += cputime; | |
9277 | } | |
9278 | } | |
9279 | ||
9280 | struct cgroup_subsys cpuacct_subsys = { | |
9281 | .name = "cpuacct", | |
9282 | .create = cpuacct_create, | |
9283 | .destroy = cpuacct_destroy, | |
9284 | .populate = cpuacct_populate, | |
9285 | .subsys_id = cpuacct_subsys_id, | |
9286 | }; | |
9287 | #endif /* CONFIG_CGROUP_CPUACCT */ |