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> | |
6cd8a4bb | 73 | #include <linux/ftrace.h> |
1da177e4 | 74 | |
5517d86b | 75 | #include <asm/tlb.h> |
838225b4 | 76 | #include <asm/irq_regs.h> |
1da177e4 LT |
77 | |
78 | /* | |
79 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
80 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
81 | * and back. | |
82 | */ | |
83 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
84 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
85 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
86 | ||
87 | /* | |
88 | * 'User priority' is the nice value converted to something we | |
89 | * can work with better when scaling various scheduler parameters, | |
90 | * it's a [ 0 ... 39 ] range. | |
91 | */ | |
92 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
93 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
94 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
95 | ||
96 | /* | |
d7876a08 | 97 | * Helpers for converting nanosecond timing to jiffy resolution |
1da177e4 | 98 | */ |
d6322faf | 99 | #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) |
1da177e4 | 100 | |
6aa645ea IM |
101 | #define NICE_0_LOAD SCHED_LOAD_SCALE |
102 | #define NICE_0_SHIFT SCHED_LOAD_SHIFT | |
103 | ||
1da177e4 LT |
104 | /* |
105 | * These are the 'tuning knobs' of the scheduler: | |
106 | * | |
a4ec24b4 | 107 | * default timeslice is 100 msecs (used only for SCHED_RR tasks). |
1da177e4 LT |
108 | * Timeslices get refilled after they expire. |
109 | */ | |
1da177e4 | 110 | #define DEF_TIMESLICE (100 * HZ / 1000) |
2dd73a4f | 111 | |
d0b27fa7 PZ |
112 | /* |
113 | * single value that denotes runtime == period, ie unlimited time. | |
114 | */ | |
115 | #define RUNTIME_INF ((u64)~0ULL) | |
116 | ||
5517d86b ED |
117 | #ifdef CONFIG_SMP |
118 | /* | |
119 | * Divide a load by a sched group cpu_power : (load / sg->__cpu_power) | |
120 | * Since cpu_power is a 'constant', we can use a reciprocal divide. | |
121 | */ | |
122 | static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load) | |
123 | { | |
124 | return reciprocal_divide(load, sg->reciprocal_cpu_power); | |
125 | } | |
126 | ||
127 | /* | |
128 | * Each time a sched group cpu_power is changed, | |
129 | * we must compute its reciprocal value | |
130 | */ | |
131 | static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val) | |
132 | { | |
133 | sg->__cpu_power += val; | |
134 | sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power); | |
135 | } | |
136 | #endif | |
137 | ||
e05606d3 IM |
138 | static inline int rt_policy(int policy) |
139 | { | |
140 | if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR)) | |
141 | return 1; | |
142 | return 0; | |
143 | } | |
144 | ||
145 | static inline int task_has_rt_policy(struct task_struct *p) | |
146 | { | |
147 | return rt_policy(p->policy); | |
148 | } | |
149 | ||
1da177e4 | 150 | /* |
6aa645ea | 151 | * This is the priority-queue data structure of the RT scheduling class: |
1da177e4 | 152 | */ |
6aa645ea IM |
153 | struct rt_prio_array { |
154 | DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ | |
155 | struct list_head queue[MAX_RT_PRIO]; | |
156 | }; | |
157 | ||
d0b27fa7 | 158 | struct rt_bandwidth { |
ea736ed5 IM |
159 | /* nests inside the rq lock: */ |
160 | spinlock_t rt_runtime_lock; | |
161 | ktime_t rt_period; | |
162 | u64 rt_runtime; | |
163 | struct hrtimer rt_period_timer; | |
d0b27fa7 PZ |
164 | }; |
165 | ||
166 | static struct rt_bandwidth def_rt_bandwidth; | |
167 | ||
168 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); | |
169 | ||
170 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) | |
171 | { | |
172 | struct rt_bandwidth *rt_b = | |
173 | container_of(timer, struct rt_bandwidth, rt_period_timer); | |
174 | ktime_t now; | |
175 | int overrun; | |
176 | int idle = 0; | |
177 | ||
178 | for (;;) { | |
179 | now = hrtimer_cb_get_time(timer); | |
180 | overrun = hrtimer_forward(timer, now, rt_b->rt_period); | |
181 | ||
182 | if (!overrun) | |
183 | break; | |
184 | ||
185 | idle = do_sched_rt_period_timer(rt_b, overrun); | |
186 | } | |
187 | ||
188 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
189 | } | |
190 | ||
191 | static | |
192 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) | |
193 | { | |
194 | rt_b->rt_period = ns_to_ktime(period); | |
195 | rt_b->rt_runtime = runtime; | |
196 | ||
ac086bc2 PZ |
197 | spin_lock_init(&rt_b->rt_runtime_lock); |
198 | ||
d0b27fa7 PZ |
199 | hrtimer_init(&rt_b->rt_period_timer, |
200 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
201 | rt_b->rt_period_timer.function = sched_rt_period_timer; | |
202 | rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
203 | } | |
204 | ||
205 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) | |
206 | { | |
207 | ktime_t now; | |
208 | ||
209 | if (rt_b->rt_runtime == RUNTIME_INF) | |
210 | return; | |
211 | ||
212 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
213 | return; | |
214 | ||
215 | spin_lock(&rt_b->rt_runtime_lock); | |
216 | for (;;) { | |
217 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
218 | break; | |
219 | ||
220 | now = hrtimer_cb_get_time(&rt_b->rt_period_timer); | |
221 | hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period); | |
222 | hrtimer_start(&rt_b->rt_period_timer, | |
223 | rt_b->rt_period_timer.expires, | |
224 | HRTIMER_MODE_ABS); | |
225 | } | |
226 | spin_unlock(&rt_b->rt_runtime_lock); | |
227 | } | |
228 | ||
229 | #ifdef CONFIG_RT_GROUP_SCHED | |
230 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) | |
231 | { | |
232 | hrtimer_cancel(&rt_b->rt_period_timer); | |
233 | } | |
234 | #endif | |
235 | ||
712555ee HC |
236 | /* |
237 | * sched_domains_mutex serializes calls to arch_init_sched_domains, | |
238 | * detach_destroy_domains and partition_sched_domains. | |
239 | */ | |
240 | static DEFINE_MUTEX(sched_domains_mutex); | |
241 | ||
052f1dc7 | 242 | #ifdef CONFIG_GROUP_SCHED |
29f59db3 | 243 | |
68318b8e SV |
244 | #include <linux/cgroup.h> |
245 | ||
29f59db3 SV |
246 | struct cfs_rq; |
247 | ||
6f505b16 PZ |
248 | static LIST_HEAD(task_groups); |
249 | ||
29f59db3 | 250 | /* task group related information */ |
4cf86d77 | 251 | struct task_group { |
052f1dc7 | 252 | #ifdef CONFIG_CGROUP_SCHED |
68318b8e SV |
253 | struct cgroup_subsys_state css; |
254 | #endif | |
052f1dc7 PZ |
255 | |
256 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
29f59db3 SV |
257 | /* schedulable entities of this group on each cpu */ |
258 | struct sched_entity **se; | |
259 | /* runqueue "owned" by this group on each cpu */ | |
260 | struct cfs_rq **cfs_rq; | |
261 | unsigned long shares; | |
052f1dc7 PZ |
262 | #endif |
263 | ||
264 | #ifdef CONFIG_RT_GROUP_SCHED | |
265 | struct sched_rt_entity **rt_se; | |
266 | struct rt_rq **rt_rq; | |
267 | ||
d0b27fa7 | 268 | struct rt_bandwidth rt_bandwidth; |
052f1dc7 | 269 | #endif |
6b2d7700 | 270 | |
ae8393e5 | 271 | struct rcu_head rcu; |
6f505b16 | 272 | struct list_head list; |
f473aa5e PZ |
273 | |
274 | struct task_group *parent; | |
275 | struct list_head siblings; | |
276 | struct list_head children; | |
29f59db3 SV |
277 | }; |
278 | ||
354d60c2 | 279 | #ifdef CONFIG_USER_SCHED |
eff766a6 PZ |
280 | |
281 | /* | |
282 | * Root task group. | |
283 | * Every UID task group (including init_task_group aka UID-0) will | |
284 | * be a child to this group. | |
285 | */ | |
286 | struct task_group root_task_group; | |
287 | ||
052f1dc7 | 288 | #ifdef CONFIG_FAIR_GROUP_SCHED |
29f59db3 SV |
289 | /* Default task group's sched entity on each cpu */ |
290 | static DEFINE_PER_CPU(struct sched_entity, init_sched_entity); | |
291 | /* Default task group's cfs_rq on each cpu */ | |
292 | static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp; | |
052f1dc7 PZ |
293 | #endif |
294 | ||
295 | #ifdef CONFIG_RT_GROUP_SCHED | |
296 | static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity); | |
297 | static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp; | |
052f1dc7 | 298 | #endif |
eff766a6 PZ |
299 | #else |
300 | #define root_task_group init_task_group | |
354d60c2 | 301 | #endif |
6f505b16 | 302 | |
8ed36996 | 303 | /* task_group_lock serializes add/remove of task groups and also changes to |
ec2c507f SV |
304 | * a task group's cpu shares. |
305 | */ | |
8ed36996 | 306 | static DEFINE_SPINLOCK(task_group_lock); |
ec2c507f | 307 | |
052f1dc7 | 308 | #ifdef CONFIG_FAIR_GROUP_SCHED |
052f1dc7 PZ |
309 | #ifdef CONFIG_USER_SCHED |
310 | # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD) | |
311 | #else | |
312 | # define INIT_TASK_GROUP_LOAD NICE_0_LOAD | |
313 | #endif | |
314 | ||
cb4ad1ff MX |
315 | /* |
316 | * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems. | |
317 | * (The default weight is 1024 - so there's no practical | |
318 | * limitation from this.) | |
319 | */ | |
18d95a28 | 320 | #define MIN_SHARES 2 |
cb4ad1ff | 321 | #define MAX_SHARES (ULONG_MAX - 1) |
18d95a28 | 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 | ||
361 | #else | |
362 | ||
6f505b16 | 363 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } |
29f59db3 | 364 | |
052f1dc7 | 365 | #endif /* CONFIG_GROUP_SCHED */ |
29f59db3 | 366 | |
6aa645ea IM |
367 | /* CFS-related fields in a runqueue */ |
368 | struct cfs_rq { | |
369 | struct load_weight load; | |
370 | unsigned long nr_running; | |
371 | ||
6aa645ea | 372 | u64 exec_clock; |
e9acbff6 | 373 | u64 min_vruntime; |
6aa645ea IM |
374 | |
375 | struct rb_root tasks_timeline; | |
376 | struct rb_node *rb_leftmost; | |
4a55bd5e PZ |
377 | |
378 | struct list_head tasks; | |
379 | struct list_head *balance_iterator; | |
380 | ||
381 | /* | |
382 | * 'curr' points to currently running entity on this cfs_rq. | |
6aa645ea IM |
383 | * It is set to NULL otherwise (i.e when none are currently running). |
384 | */ | |
aa2ac252 | 385 | struct sched_entity *curr, *next; |
ddc97297 PZ |
386 | |
387 | unsigned long nr_spread_over; | |
388 | ||
62160e3f | 389 | #ifdef CONFIG_FAIR_GROUP_SCHED |
6aa645ea IM |
390 | struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ |
391 | ||
41a2d6cf IM |
392 | /* |
393 | * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in | |
6aa645ea IM |
394 | * a hierarchy). Non-leaf lrqs hold other higher schedulable entities |
395 | * (like users, containers etc.) | |
396 | * | |
397 | * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This | |
398 | * list is used during load balance. | |
399 | */ | |
41a2d6cf IM |
400 | struct list_head leaf_cfs_rq_list; |
401 | struct task_group *tg; /* group that "owns" this runqueue */ | |
18d95a28 PZ |
402 | |
403 | #ifdef CONFIG_SMP | |
404 | unsigned long task_weight; | |
405 | unsigned long shares; | |
406 | /* | |
407 | * We need space to build a sched_domain wide view of the full task | |
408 | * group tree, in order to avoid depending on dynamic memory allocation | |
409 | * during the load balancing we place this in the per cpu task group | |
410 | * hierarchy. This limits the load balancing to one instance per cpu, | |
411 | * but more should not be needed anyway. | |
412 | */ | |
413 | struct aggregate_struct { | |
414 | /* | |
415 | * load = weight(cpus) * f(tg) | |
416 | * | |
417 | * Where f(tg) is the recursive weight fraction assigned to | |
418 | * this group. | |
419 | */ | |
420 | unsigned long load; | |
421 | ||
422 | /* | |
423 | * part of the group weight distributed to this span. | |
424 | */ | |
425 | unsigned long shares; | |
426 | ||
427 | /* | |
428 | * The sum of all runqueue weights within this span. | |
429 | */ | |
430 | unsigned long rq_weight; | |
431 | ||
432 | /* | |
433 | * Weight contributed by tasks; this is the part we can | |
434 | * influence by moving tasks around. | |
435 | */ | |
436 | unsigned long task_weight; | |
437 | } aggregate; | |
438 | #endif | |
6aa645ea IM |
439 | #endif |
440 | }; | |
1da177e4 | 441 | |
6aa645ea IM |
442 | /* Real-Time classes' related field in a runqueue: */ |
443 | struct rt_rq { | |
444 | struct rt_prio_array active; | |
63489e45 | 445 | unsigned long rt_nr_running; |
052f1dc7 | 446 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
447 | int highest_prio; /* highest queued rt task prio */ |
448 | #endif | |
fa85ae24 | 449 | #ifdef CONFIG_SMP |
73fe6aae | 450 | unsigned long rt_nr_migratory; |
a22d7fc1 | 451 | int overloaded; |
fa85ae24 | 452 | #endif |
6f505b16 | 453 | int rt_throttled; |
fa85ae24 | 454 | u64 rt_time; |
ac086bc2 | 455 | u64 rt_runtime; |
ea736ed5 | 456 | /* Nests inside the rq lock: */ |
ac086bc2 | 457 | spinlock_t rt_runtime_lock; |
6f505b16 | 458 | |
052f1dc7 | 459 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc PZ |
460 | unsigned long rt_nr_boosted; |
461 | ||
6f505b16 PZ |
462 | struct rq *rq; |
463 | struct list_head leaf_rt_rq_list; | |
464 | struct task_group *tg; | |
465 | struct sched_rt_entity *rt_se; | |
466 | #endif | |
6aa645ea IM |
467 | }; |
468 | ||
57d885fe GH |
469 | #ifdef CONFIG_SMP |
470 | ||
471 | /* | |
472 | * We add the notion of a root-domain which will be used to define per-domain | |
0eab9146 IM |
473 | * variables. Each exclusive cpuset essentially defines an island domain by |
474 | * fully partitioning the member cpus from any other cpuset. Whenever a new | |
57d885fe GH |
475 | * exclusive cpuset is created, we also create and attach a new root-domain |
476 | * object. | |
477 | * | |
57d885fe GH |
478 | */ |
479 | struct root_domain { | |
480 | atomic_t refcount; | |
481 | cpumask_t span; | |
482 | cpumask_t online; | |
637f5085 | 483 | |
0eab9146 | 484 | /* |
637f5085 GH |
485 | * The "RT overload" flag: it gets set if a CPU has more than |
486 | * one runnable RT task. | |
487 | */ | |
488 | cpumask_t rto_mask; | |
0eab9146 | 489 | atomic_t rto_count; |
57d885fe GH |
490 | }; |
491 | ||
dc938520 GH |
492 | /* |
493 | * By default the system creates a single root-domain with all cpus as | |
494 | * members (mimicking the global state we have today). | |
495 | */ | |
57d885fe GH |
496 | static struct root_domain def_root_domain; |
497 | ||
498 | #endif | |
499 | ||
1da177e4 LT |
500 | /* |
501 | * This is the main, per-CPU runqueue data structure. | |
502 | * | |
503 | * Locking rule: those places that want to lock multiple runqueues | |
504 | * (such as the load balancing or the thread migration code), lock | |
505 | * acquire operations must be ordered by ascending &runqueue. | |
506 | */ | |
70b97a7f | 507 | struct rq { |
d8016491 IM |
508 | /* runqueue lock: */ |
509 | spinlock_t lock; | |
1da177e4 LT |
510 | |
511 | /* | |
512 | * nr_running and cpu_load should be in the same cacheline because | |
513 | * remote CPUs use both these fields when doing load calculation. | |
514 | */ | |
515 | unsigned long nr_running; | |
6aa645ea IM |
516 | #define CPU_LOAD_IDX_MAX 5 |
517 | unsigned long cpu_load[CPU_LOAD_IDX_MAX]; | |
bdecea3a | 518 | unsigned char idle_at_tick; |
46cb4b7c | 519 | #ifdef CONFIG_NO_HZ |
15934a37 | 520 | unsigned long last_tick_seen; |
46cb4b7c SS |
521 | unsigned char in_nohz_recently; |
522 | #endif | |
d8016491 IM |
523 | /* capture load from *all* tasks on this cpu: */ |
524 | struct load_weight load; | |
6aa645ea IM |
525 | unsigned long nr_load_updates; |
526 | u64 nr_switches; | |
527 | ||
528 | struct cfs_rq cfs; | |
6f505b16 | 529 | struct rt_rq rt; |
6f505b16 | 530 | |
6aa645ea | 531 | #ifdef CONFIG_FAIR_GROUP_SCHED |
d8016491 IM |
532 | /* list of leaf cfs_rq on this cpu: */ |
533 | struct list_head leaf_cfs_rq_list; | |
052f1dc7 PZ |
534 | #endif |
535 | #ifdef CONFIG_RT_GROUP_SCHED | |
6f505b16 | 536 | struct list_head leaf_rt_rq_list; |
1da177e4 | 537 | #endif |
1da177e4 LT |
538 | |
539 | /* | |
540 | * This is part of a global counter where only the total sum | |
541 | * over all CPUs matters. A task can increase this counter on | |
542 | * one CPU and if it got migrated afterwards it may decrease | |
543 | * it on another CPU. Always updated under the runqueue lock: | |
544 | */ | |
545 | unsigned long nr_uninterruptible; | |
546 | ||
36c8b586 | 547 | struct task_struct *curr, *idle; |
c9819f45 | 548 | unsigned long next_balance; |
1da177e4 | 549 | struct mm_struct *prev_mm; |
6aa645ea | 550 | |
3e51f33f | 551 | u64 clock; |
6aa645ea | 552 | |
1da177e4 LT |
553 | atomic_t nr_iowait; |
554 | ||
555 | #ifdef CONFIG_SMP | |
0eab9146 | 556 | struct root_domain *rd; |
1da177e4 LT |
557 | struct sched_domain *sd; |
558 | ||
559 | /* For active balancing */ | |
560 | int active_balance; | |
561 | int push_cpu; | |
d8016491 IM |
562 | /* cpu of this runqueue: */ |
563 | int cpu; | |
1da177e4 | 564 | |
36c8b586 | 565 | struct task_struct *migration_thread; |
1da177e4 LT |
566 | struct list_head migration_queue; |
567 | #endif | |
568 | ||
8f4d37ec PZ |
569 | #ifdef CONFIG_SCHED_HRTICK |
570 | unsigned long hrtick_flags; | |
571 | ktime_t hrtick_expire; | |
572 | struct hrtimer hrtick_timer; | |
573 | #endif | |
574 | ||
1da177e4 LT |
575 | #ifdef CONFIG_SCHEDSTATS |
576 | /* latency stats */ | |
577 | struct sched_info rq_sched_info; | |
578 | ||
579 | /* sys_sched_yield() stats */ | |
480b9434 KC |
580 | unsigned int yld_exp_empty; |
581 | unsigned int yld_act_empty; | |
582 | unsigned int yld_both_empty; | |
583 | unsigned int yld_count; | |
1da177e4 LT |
584 | |
585 | /* schedule() stats */ | |
480b9434 KC |
586 | unsigned int sched_switch; |
587 | unsigned int sched_count; | |
588 | unsigned int sched_goidle; | |
1da177e4 LT |
589 | |
590 | /* try_to_wake_up() stats */ | |
480b9434 KC |
591 | unsigned int ttwu_count; |
592 | unsigned int ttwu_local; | |
b8efb561 IM |
593 | |
594 | /* BKL stats */ | |
480b9434 | 595 | unsigned int bkl_count; |
1da177e4 | 596 | #endif |
fcb99371 | 597 | struct lock_class_key rq_lock_key; |
1da177e4 LT |
598 | }; |
599 | ||
f34e3b61 | 600 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); |
1da177e4 | 601 | |
dd41f596 IM |
602 | static inline void check_preempt_curr(struct rq *rq, struct task_struct *p) |
603 | { | |
604 | rq->curr->sched_class->check_preempt_curr(rq, p); | |
605 | } | |
606 | ||
0a2966b4 CL |
607 | static inline int cpu_of(struct rq *rq) |
608 | { | |
609 | #ifdef CONFIG_SMP | |
610 | return rq->cpu; | |
611 | #else | |
612 | return 0; | |
613 | #endif | |
614 | } | |
615 | ||
674311d5 NP |
616 | /* |
617 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 618 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
619 | * |
620 | * The domain tree of any CPU may only be accessed from within | |
621 | * preempt-disabled sections. | |
622 | */ | |
48f24c4d IM |
623 | #define for_each_domain(cpu, __sd) \ |
624 | for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) | |
1da177e4 LT |
625 | |
626 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
627 | #define this_rq() (&__get_cpu_var(runqueues)) | |
628 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
629 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
630 | ||
3e51f33f PZ |
631 | static inline void update_rq_clock(struct rq *rq) |
632 | { | |
633 | rq->clock = sched_clock_cpu(cpu_of(rq)); | |
634 | } | |
635 | ||
bf5c91ba IM |
636 | /* |
637 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: | |
638 | */ | |
639 | #ifdef CONFIG_SCHED_DEBUG | |
640 | # define const_debug __read_mostly | |
641 | #else | |
642 | # define const_debug static const | |
643 | #endif | |
644 | ||
645 | /* | |
646 | * Debugging: various feature bits | |
647 | */ | |
f00b45c1 PZ |
648 | |
649 | #define SCHED_FEAT(name, enabled) \ | |
650 | __SCHED_FEAT_##name , | |
651 | ||
bf5c91ba | 652 | enum { |
f00b45c1 | 653 | #include "sched_features.h" |
bf5c91ba IM |
654 | }; |
655 | ||
f00b45c1 PZ |
656 | #undef SCHED_FEAT |
657 | ||
658 | #define SCHED_FEAT(name, enabled) \ | |
659 | (1UL << __SCHED_FEAT_##name) * enabled | | |
660 | ||
bf5c91ba | 661 | const_debug unsigned int sysctl_sched_features = |
f00b45c1 PZ |
662 | #include "sched_features.h" |
663 | 0; | |
664 | ||
665 | #undef SCHED_FEAT | |
666 | ||
667 | #ifdef CONFIG_SCHED_DEBUG | |
668 | #define SCHED_FEAT(name, enabled) \ | |
669 | #name , | |
670 | ||
983ed7a6 | 671 | static __read_mostly char *sched_feat_names[] = { |
f00b45c1 PZ |
672 | #include "sched_features.h" |
673 | NULL | |
674 | }; | |
675 | ||
676 | #undef SCHED_FEAT | |
677 | ||
983ed7a6 | 678 | static int sched_feat_open(struct inode *inode, struct file *filp) |
f00b45c1 PZ |
679 | { |
680 | filp->private_data = inode->i_private; | |
681 | return 0; | |
682 | } | |
683 | ||
684 | static ssize_t | |
685 | sched_feat_read(struct file *filp, char __user *ubuf, | |
686 | size_t cnt, loff_t *ppos) | |
687 | { | |
688 | char *buf; | |
689 | int r = 0; | |
690 | int len = 0; | |
691 | int i; | |
692 | ||
693 | for (i = 0; sched_feat_names[i]; i++) { | |
694 | len += strlen(sched_feat_names[i]); | |
695 | len += 4; | |
696 | } | |
697 | ||
698 | buf = kmalloc(len + 2, GFP_KERNEL); | |
699 | if (!buf) | |
700 | return -ENOMEM; | |
701 | ||
702 | for (i = 0; sched_feat_names[i]; i++) { | |
703 | if (sysctl_sched_features & (1UL << i)) | |
704 | r += sprintf(buf + r, "%s ", sched_feat_names[i]); | |
705 | else | |
c24b7c52 | 706 | r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]); |
f00b45c1 PZ |
707 | } |
708 | ||
709 | r += sprintf(buf + r, "\n"); | |
710 | WARN_ON(r >= len + 2); | |
711 | ||
712 | r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r); | |
713 | ||
714 | kfree(buf); | |
715 | ||
716 | return r; | |
717 | } | |
718 | ||
719 | static ssize_t | |
720 | sched_feat_write(struct file *filp, const char __user *ubuf, | |
721 | size_t cnt, loff_t *ppos) | |
722 | { | |
723 | char buf[64]; | |
724 | char *cmp = buf; | |
725 | int neg = 0; | |
726 | int i; | |
727 | ||
728 | if (cnt > 63) | |
729 | cnt = 63; | |
730 | ||
731 | if (copy_from_user(&buf, ubuf, cnt)) | |
732 | return -EFAULT; | |
733 | ||
734 | buf[cnt] = 0; | |
735 | ||
c24b7c52 | 736 | if (strncmp(buf, "NO_", 3) == 0) { |
f00b45c1 PZ |
737 | neg = 1; |
738 | cmp += 3; | |
739 | } | |
740 | ||
741 | for (i = 0; sched_feat_names[i]; i++) { | |
742 | int len = strlen(sched_feat_names[i]); | |
743 | ||
744 | if (strncmp(cmp, sched_feat_names[i], len) == 0) { | |
745 | if (neg) | |
746 | sysctl_sched_features &= ~(1UL << i); | |
747 | else | |
748 | sysctl_sched_features |= (1UL << i); | |
749 | break; | |
750 | } | |
751 | } | |
752 | ||
753 | if (!sched_feat_names[i]) | |
754 | return -EINVAL; | |
755 | ||
756 | filp->f_pos += cnt; | |
757 | ||
758 | return cnt; | |
759 | } | |
760 | ||
761 | static struct file_operations sched_feat_fops = { | |
762 | .open = sched_feat_open, | |
763 | .read = sched_feat_read, | |
764 | .write = sched_feat_write, | |
765 | }; | |
766 | ||
767 | static __init int sched_init_debug(void) | |
768 | { | |
f00b45c1 PZ |
769 | debugfs_create_file("sched_features", 0644, NULL, NULL, |
770 | &sched_feat_fops); | |
771 | ||
772 | return 0; | |
773 | } | |
774 | late_initcall(sched_init_debug); | |
775 | ||
776 | #endif | |
777 | ||
778 | #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) | |
bf5c91ba | 779 | |
b82d9fdd PZ |
780 | /* |
781 | * Number of tasks to iterate in a single balance run. | |
782 | * Limited because this is done with IRQs disabled. | |
783 | */ | |
784 | const_debug unsigned int sysctl_sched_nr_migrate = 32; | |
785 | ||
fa85ae24 | 786 | /* |
9f0c1e56 | 787 | * period over which we measure -rt task cpu usage in us. |
fa85ae24 PZ |
788 | * default: 1s |
789 | */ | |
9f0c1e56 | 790 | unsigned int sysctl_sched_rt_period = 1000000; |
fa85ae24 | 791 | |
6892b75e IM |
792 | static __read_mostly int scheduler_running; |
793 | ||
9f0c1e56 PZ |
794 | /* |
795 | * part of the period that we allow rt tasks to run in us. | |
796 | * default: 0.95s | |
797 | */ | |
798 | int sysctl_sched_rt_runtime = 950000; | |
fa85ae24 | 799 | |
d0b27fa7 PZ |
800 | static inline u64 global_rt_period(void) |
801 | { | |
802 | return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; | |
803 | } | |
804 | ||
805 | static inline u64 global_rt_runtime(void) | |
806 | { | |
807 | if (sysctl_sched_rt_period < 0) | |
808 | return RUNTIME_INF; | |
809 | ||
810 | return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; | |
811 | } | |
fa85ae24 | 812 | |
690229a0 | 813 | unsigned long long time_sync_thresh = 100000; |
27ec4407 IM |
814 | |
815 | static DEFINE_PER_CPU(unsigned long long, time_offset); | |
816 | static DEFINE_PER_CPU(unsigned long long, prev_cpu_time); | |
817 | ||
e436d800 | 818 | /* |
27ec4407 IM |
819 | * Global lock which we take every now and then to synchronize |
820 | * the CPUs time. This method is not warp-safe, but it's good | |
821 | * enough to synchronize slowly diverging time sources and thus | |
822 | * it's good enough for tracing: | |
e436d800 | 823 | */ |
27ec4407 IM |
824 | static DEFINE_SPINLOCK(time_sync_lock); |
825 | static unsigned long long prev_global_time; | |
826 | ||
dfbf4a1b | 827 | static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu) |
27ec4407 | 828 | { |
dfbf4a1b IM |
829 | /* |
830 | * We want this inlined, to not get tracer function calls | |
831 | * in this critical section: | |
832 | */ | |
833 | spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_); | |
834 | __raw_spin_lock(&time_sync_lock.raw_lock); | |
27ec4407 IM |
835 | |
836 | if (time < prev_global_time) { | |
837 | per_cpu(time_offset, cpu) += prev_global_time - time; | |
838 | time = prev_global_time; | |
839 | } else { | |
840 | prev_global_time = time; | |
841 | } | |
842 | ||
dfbf4a1b IM |
843 | __raw_spin_unlock(&time_sync_lock.raw_lock); |
844 | spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_); | |
27ec4407 IM |
845 | |
846 | return time; | |
847 | } | |
848 | ||
849 | static unsigned long long __cpu_clock(int cpu) | |
e436d800 | 850 | { |
e436d800 | 851 | unsigned long long now; |
e436d800 | 852 | |
8ced5f69 IM |
853 | /* |
854 | * Only call sched_clock() if the scheduler has already been | |
855 | * initialized (some code might call cpu_clock() very early): | |
856 | */ | |
6892b75e IM |
857 | if (unlikely(!scheduler_running)) |
858 | return 0; | |
859 | ||
3e51f33f | 860 | now = sched_clock_cpu(cpu); |
e436d800 IM |
861 | |
862 | return now; | |
863 | } | |
27ec4407 IM |
864 | |
865 | /* | |
866 | * For kernel-internal use: high-speed (but slightly incorrect) per-cpu | |
867 | * clock constructed from sched_clock(): | |
868 | */ | |
869 | unsigned long long cpu_clock(int cpu) | |
870 | { | |
871 | unsigned long long prev_cpu_time, time, delta_time; | |
dfbf4a1b | 872 | unsigned long flags; |
27ec4407 | 873 | |
dfbf4a1b | 874 | local_irq_save(flags); |
27ec4407 IM |
875 | prev_cpu_time = per_cpu(prev_cpu_time, cpu); |
876 | time = __cpu_clock(cpu) + per_cpu(time_offset, cpu); | |
877 | delta_time = time-prev_cpu_time; | |
878 | ||
dfbf4a1b | 879 | if (unlikely(delta_time > time_sync_thresh)) { |
27ec4407 | 880 | time = __sync_cpu_clock(time, cpu); |
dfbf4a1b IM |
881 | per_cpu(prev_cpu_time, cpu) = time; |
882 | } | |
883 | local_irq_restore(flags); | |
27ec4407 IM |
884 | |
885 | return time; | |
886 | } | |
a58f6f25 | 887 | EXPORT_SYMBOL_GPL(cpu_clock); |
e436d800 | 888 | |
1da177e4 | 889 | #ifndef prepare_arch_switch |
4866cde0 NP |
890 | # define prepare_arch_switch(next) do { } while (0) |
891 | #endif | |
892 | #ifndef finish_arch_switch | |
893 | # define finish_arch_switch(prev) do { } while (0) | |
894 | #endif | |
895 | ||
051a1d1a DA |
896 | static inline int task_current(struct rq *rq, struct task_struct *p) |
897 | { | |
898 | return rq->curr == p; | |
899 | } | |
900 | ||
4866cde0 | 901 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
70b97a7f | 902 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 | 903 | { |
051a1d1a | 904 | return task_current(rq, p); |
4866cde0 NP |
905 | } |
906 | ||
70b97a7f | 907 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
908 | { |
909 | } | |
910 | ||
70b97a7f | 911 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 | 912 | { |
da04c035 IM |
913 | #ifdef CONFIG_DEBUG_SPINLOCK |
914 | /* this is a valid case when another task releases the spinlock */ | |
915 | rq->lock.owner = current; | |
916 | #endif | |
8a25d5de IM |
917 | /* |
918 | * If we are tracking spinlock dependencies then we have to | |
919 | * fix up the runqueue lock - which gets 'carried over' from | |
920 | * prev into current: | |
921 | */ | |
922 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); | |
923 | ||
4866cde0 NP |
924 | spin_unlock_irq(&rq->lock); |
925 | } | |
926 | ||
927 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
70b97a7f | 928 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 NP |
929 | { |
930 | #ifdef CONFIG_SMP | |
931 | return p->oncpu; | |
932 | #else | |
051a1d1a | 933 | return task_current(rq, p); |
4866cde0 NP |
934 | #endif |
935 | } | |
936 | ||
70b97a7f | 937 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
938 | { |
939 | #ifdef CONFIG_SMP | |
940 | /* | |
941 | * We can optimise this out completely for !SMP, because the | |
942 | * SMP rebalancing from interrupt is the only thing that cares | |
943 | * here. | |
944 | */ | |
945 | next->oncpu = 1; | |
946 | #endif | |
947 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
948 | spin_unlock_irq(&rq->lock); | |
949 | #else | |
950 | spin_unlock(&rq->lock); | |
951 | #endif | |
952 | } | |
953 | ||
70b97a7f | 954 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 NP |
955 | { |
956 | #ifdef CONFIG_SMP | |
957 | /* | |
958 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
959 | * We must ensure this doesn't happen until the switch is completely | |
960 | * finished. | |
961 | */ | |
962 | smp_wmb(); | |
963 | prev->oncpu = 0; | |
964 | #endif | |
965 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
966 | local_irq_enable(); | |
1da177e4 | 967 | #endif |
4866cde0 NP |
968 | } |
969 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 | 970 | |
b29739f9 IM |
971 | /* |
972 | * __task_rq_lock - lock the runqueue a given task resides on. | |
973 | * Must be called interrupts disabled. | |
974 | */ | |
70b97a7f | 975 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
b29739f9 IM |
976 | __acquires(rq->lock) |
977 | { | |
3a5c359a AK |
978 | for (;;) { |
979 | struct rq *rq = task_rq(p); | |
980 | spin_lock(&rq->lock); | |
981 | if (likely(rq == task_rq(p))) | |
982 | return rq; | |
b29739f9 | 983 | spin_unlock(&rq->lock); |
b29739f9 | 984 | } |
b29739f9 IM |
985 | } |
986 | ||
1da177e4 LT |
987 | /* |
988 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
41a2d6cf | 989 | * interrupts. Note the ordering: we can safely lookup the task_rq without |
1da177e4 LT |
990 | * explicitly disabling preemption. |
991 | */ | |
70b97a7f | 992 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
1da177e4 LT |
993 | __acquires(rq->lock) |
994 | { | |
70b97a7f | 995 | struct rq *rq; |
1da177e4 | 996 | |
3a5c359a AK |
997 | for (;;) { |
998 | local_irq_save(*flags); | |
999 | rq = task_rq(p); | |
1000 | spin_lock(&rq->lock); | |
1001 | if (likely(rq == task_rq(p))) | |
1002 | return rq; | |
1da177e4 | 1003 | spin_unlock_irqrestore(&rq->lock, *flags); |
1da177e4 | 1004 | } |
1da177e4 LT |
1005 | } |
1006 | ||
a9957449 | 1007 | static void __task_rq_unlock(struct rq *rq) |
b29739f9 IM |
1008 | __releases(rq->lock) |
1009 | { | |
1010 | spin_unlock(&rq->lock); | |
1011 | } | |
1012 | ||
70b97a7f | 1013 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) |
1da177e4 LT |
1014 | __releases(rq->lock) |
1015 | { | |
1016 | spin_unlock_irqrestore(&rq->lock, *flags); | |
1017 | } | |
1018 | ||
1da177e4 | 1019 | /* |
cc2a73b5 | 1020 | * this_rq_lock - lock this runqueue and disable interrupts. |
1da177e4 | 1021 | */ |
a9957449 | 1022 | static struct rq *this_rq_lock(void) |
1da177e4 LT |
1023 | __acquires(rq->lock) |
1024 | { | |
70b97a7f | 1025 | struct rq *rq; |
1da177e4 LT |
1026 | |
1027 | local_irq_disable(); | |
1028 | rq = this_rq(); | |
1029 | spin_lock(&rq->lock); | |
1030 | ||
1031 | return rq; | |
1032 | } | |
1033 | ||
8f4d37ec PZ |
1034 | static void __resched_task(struct task_struct *p, int tif_bit); |
1035 | ||
1036 | static inline void resched_task(struct task_struct *p) | |
1037 | { | |
1038 | __resched_task(p, TIF_NEED_RESCHED); | |
1039 | } | |
1040 | ||
1041 | #ifdef CONFIG_SCHED_HRTICK | |
1042 | /* | |
1043 | * Use HR-timers to deliver accurate preemption points. | |
1044 | * | |
1045 | * Its all a bit involved since we cannot program an hrt while holding the | |
1046 | * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a | |
1047 | * reschedule event. | |
1048 | * | |
1049 | * When we get rescheduled we reprogram the hrtick_timer outside of the | |
1050 | * rq->lock. | |
1051 | */ | |
1052 | static inline void resched_hrt(struct task_struct *p) | |
1053 | { | |
1054 | __resched_task(p, TIF_HRTICK_RESCHED); | |
1055 | } | |
1056 | ||
1057 | static inline void resched_rq(struct rq *rq) | |
1058 | { | |
1059 | unsigned long flags; | |
1060 | ||
1061 | spin_lock_irqsave(&rq->lock, flags); | |
1062 | resched_task(rq->curr); | |
1063 | spin_unlock_irqrestore(&rq->lock, flags); | |
1064 | } | |
1065 | ||
1066 | enum { | |
1067 | HRTICK_SET, /* re-programm hrtick_timer */ | |
1068 | HRTICK_RESET, /* not a new slice */ | |
b328ca18 | 1069 | HRTICK_BLOCK, /* stop hrtick operations */ |
8f4d37ec PZ |
1070 | }; |
1071 | ||
1072 | /* | |
1073 | * Use hrtick when: | |
1074 | * - enabled by features | |
1075 | * - hrtimer is actually high res | |
1076 | */ | |
1077 | static inline int hrtick_enabled(struct rq *rq) | |
1078 | { | |
1079 | if (!sched_feat(HRTICK)) | |
1080 | return 0; | |
b328ca18 PZ |
1081 | if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags))) |
1082 | return 0; | |
8f4d37ec PZ |
1083 | return hrtimer_is_hres_active(&rq->hrtick_timer); |
1084 | } | |
1085 | ||
1086 | /* | |
1087 | * Called to set the hrtick timer state. | |
1088 | * | |
1089 | * called with rq->lock held and irqs disabled | |
1090 | */ | |
1091 | static void hrtick_start(struct rq *rq, u64 delay, int reset) | |
1092 | { | |
1093 | assert_spin_locked(&rq->lock); | |
1094 | ||
1095 | /* | |
1096 | * preempt at: now + delay | |
1097 | */ | |
1098 | rq->hrtick_expire = | |
1099 | ktime_add_ns(rq->hrtick_timer.base->get_time(), delay); | |
1100 | /* | |
1101 | * indicate we need to program the timer | |
1102 | */ | |
1103 | __set_bit(HRTICK_SET, &rq->hrtick_flags); | |
1104 | if (reset) | |
1105 | __set_bit(HRTICK_RESET, &rq->hrtick_flags); | |
1106 | ||
1107 | /* | |
1108 | * New slices are called from the schedule path and don't need a | |
1109 | * forced reschedule. | |
1110 | */ | |
1111 | if (reset) | |
1112 | resched_hrt(rq->curr); | |
1113 | } | |
1114 | ||
1115 | static void hrtick_clear(struct rq *rq) | |
1116 | { | |
1117 | if (hrtimer_active(&rq->hrtick_timer)) | |
1118 | hrtimer_cancel(&rq->hrtick_timer); | |
1119 | } | |
1120 | ||
1121 | /* | |
1122 | * Update the timer from the possible pending state. | |
1123 | */ | |
1124 | static void hrtick_set(struct rq *rq) | |
1125 | { | |
1126 | ktime_t time; | |
1127 | int set, reset; | |
1128 | unsigned long flags; | |
1129 | ||
1130 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1131 | ||
1132 | spin_lock_irqsave(&rq->lock, flags); | |
1133 | set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags); | |
1134 | reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags); | |
1135 | time = rq->hrtick_expire; | |
1136 | clear_thread_flag(TIF_HRTICK_RESCHED); | |
1137 | spin_unlock_irqrestore(&rq->lock, flags); | |
1138 | ||
1139 | if (set) { | |
1140 | hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS); | |
1141 | if (reset && !hrtimer_active(&rq->hrtick_timer)) | |
1142 | resched_rq(rq); | |
1143 | } else | |
1144 | hrtick_clear(rq); | |
1145 | } | |
1146 | ||
1147 | /* | |
1148 | * High-resolution timer tick. | |
1149 | * Runs from hardirq context with interrupts disabled. | |
1150 | */ | |
1151 | static enum hrtimer_restart hrtick(struct hrtimer *timer) | |
1152 | { | |
1153 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); | |
1154 | ||
1155 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1156 | ||
1157 | spin_lock(&rq->lock); | |
3e51f33f | 1158 | update_rq_clock(rq); |
8f4d37ec PZ |
1159 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); |
1160 | spin_unlock(&rq->lock); | |
1161 | ||
1162 | return HRTIMER_NORESTART; | |
1163 | } | |
1164 | ||
b328ca18 PZ |
1165 | static void hotplug_hrtick_disable(int cpu) |
1166 | { | |
1167 | struct rq *rq = cpu_rq(cpu); | |
1168 | unsigned long flags; | |
1169 | ||
1170 | spin_lock_irqsave(&rq->lock, flags); | |
1171 | rq->hrtick_flags = 0; | |
1172 | __set_bit(HRTICK_BLOCK, &rq->hrtick_flags); | |
1173 | spin_unlock_irqrestore(&rq->lock, flags); | |
1174 | ||
1175 | hrtick_clear(rq); | |
1176 | } | |
1177 | ||
1178 | static void hotplug_hrtick_enable(int cpu) | |
1179 | { | |
1180 | struct rq *rq = cpu_rq(cpu); | |
1181 | unsigned long flags; | |
1182 | ||
1183 | spin_lock_irqsave(&rq->lock, flags); | |
1184 | __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags); | |
1185 | spin_unlock_irqrestore(&rq->lock, flags); | |
1186 | } | |
1187 | ||
1188 | static int | |
1189 | hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
1190 | { | |
1191 | int cpu = (int)(long)hcpu; | |
1192 | ||
1193 | switch (action) { | |
1194 | case CPU_UP_CANCELED: | |
1195 | case CPU_UP_CANCELED_FROZEN: | |
1196 | case CPU_DOWN_PREPARE: | |
1197 | case CPU_DOWN_PREPARE_FROZEN: | |
1198 | case CPU_DEAD: | |
1199 | case CPU_DEAD_FROZEN: | |
1200 | hotplug_hrtick_disable(cpu); | |
1201 | return NOTIFY_OK; | |
1202 | ||
1203 | case CPU_UP_PREPARE: | |
1204 | case CPU_UP_PREPARE_FROZEN: | |
1205 | case CPU_DOWN_FAILED: | |
1206 | case CPU_DOWN_FAILED_FROZEN: | |
1207 | case CPU_ONLINE: | |
1208 | case CPU_ONLINE_FROZEN: | |
1209 | hotplug_hrtick_enable(cpu); | |
1210 | return NOTIFY_OK; | |
1211 | } | |
1212 | ||
1213 | return NOTIFY_DONE; | |
1214 | } | |
1215 | ||
1216 | static void init_hrtick(void) | |
1217 | { | |
1218 | hotcpu_notifier(hotplug_hrtick, 0); | |
1219 | } | |
1220 | ||
1221 | static void init_rq_hrtick(struct rq *rq) | |
8f4d37ec PZ |
1222 | { |
1223 | rq->hrtick_flags = 0; | |
1224 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1225 | rq->hrtick_timer.function = hrtick; | |
1226 | rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
1227 | } | |
1228 | ||
1229 | void hrtick_resched(void) | |
1230 | { | |
1231 | struct rq *rq; | |
1232 | unsigned long flags; | |
1233 | ||
1234 | if (!test_thread_flag(TIF_HRTICK_RESCHED)) | |
1235 | return; | |
1236 | ||
1237 | local_irq_save(flags); | |
1238 | rq = cpu_rq(smp_processor_id()); | |
1239 | hrtick_set(rq); | |
1240 | local_irq_restore(flags); | |
1241 | } | |
1242 | #else | |
1243 | static inline void hrtick_clear(struct rq *rq) | |
1244 | { | |
1245 | } | |
1246 | ||
1247 | static inline void hrtick_set(struct rq *rq) | |
1248 | { | |
1249 | } | |
1250 | ||
1251 | static inline void init_rq_hrtick(struct rq *rq) | |
1252 | { | |
1253 | } | |
1254 | ||
1255 | void hrtick_resched(void) | |
1256 | { | |
1257 | } | |
b328ca18 PZ |
1258 | |
1259 | static inline void init_hrtick(void) | |
1260 | { | |
1261 | } | |
8f4d37ec PZ |
1262 | #endif |
1263 | ||
c24d20db IM |
1264 | /* |
1265 | * resched_task - mark a task 'to be rescheduled now'. | |
1266 | * | |
1267 | * On UP this means the setting of the need_resched flag, on SMP it | |
1268 | * might also involve a cross-CPU call to trigger the scheduler on | |
1269 | * the target CPU. | |
1270 | */ | |
1271 | #ifdef CONFIG_SMP | |
1272 | ||
1273 | #ifndef tsk_is_polling | |
1274 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
1275 | #endif | |
1276 | ||
8f4d37ec | 1277 | static void __resched_task(struct task_struct *p, int tif_bit) |
c24d20db IM |
1278 | { |
1279 | int cpu; | |
1280 | ||
1281 | assert_spin_locked(&task_rq(p)->lock); | |
1282 | ||
8f4d37ec | 1283 | if (unlikely(test_tsk_thread_flag(p, tif_bit))) |
c24d20db IM |
1284 | return; |
1285 | ||
8f4d37ec | 1286 | set_tsk_thread_flag(p, tif_bit); |
c24d20db IM |
1287 | |
1288 | cpu = task_cpu(p); | |
1289 | if (cpu == smp_processor_id()) | |
1290 | return; | |
1291 | ||
1292 | /* NEED_RESCHED must be visible before we test polling */ | |
1293 | smp_mb(); | |
1294 | if (!tsk_is_polling(p)) | |
1295 | smp_send_reschedule(cpu); | |
1296 | } | |
1297 | ||
1298 | static void resched_cpu(int cpu) | |
1299 | { | |
1300 | struct rq *rq = cpu_rq(cpu); | |
1301 | unsigned long flags; | |
1302 | ||
1303 | if (!spin_trylock_irqsave(&rq->lock, flags)) | |
1304 | return; | |
1305 | resched_task(cpu_curr(cpu)); | |
1306 | spin_unlock_irqrestore(&rq->lock, flags); | |
1307 | } | |
06d8308c TG |
1308 | |
1309 | #ifdef CONFIG_NO_HZ | |
1310 | /* | |
1311 | * When add_timer_on() enqueues a timer into the timer wheel of an | |
1312 | * idle CPU then this timer might expire before the next timer event | |
1313 | * which is scheduled to wake up that CPU. In case of a completely | |
1314 | * idle system the next event might even be infinite time into the | |
1315 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and | |
1316 | * leaves the inner idle loop so the newly added timer is taken into | |
1317 | * account when the CPU goes back to idle and evaluates the timer | |
1318 | * wheel for the next timer event. | |
1319 | */ | |
1320 | void wake_up_idle_cpu(int cpu) | |
1321 | { | |
1322 | struct rq *rq = cpu_rq(cpu); | |
1323 | ||
1324 | if (cpu == smp_processor_id()) | |
1325 | return; | |
1326 | ||
1327 | /* | |
1328 | * This is safe, as this function is called with the timer | |
1329 | * wheel base lock of (cpu) held. When the CPU is on the way | |
1330 | * to idle and has not yet set rq->curr to idle then it will | |
1331 | * be serialized on the timer wheel base lock and take the new | |
1332 | * timer into account automatically. | |
1333 | */ | |
1334 | if (rq->curr != rq->idle) | |
1335 | return; | |
1336 | ||
1337 | /* | |
1338 | * We can set TIF_RESCHED on the idle task of the other CPU | |
1339 | * lockless. The worst case is that the other CPU runs the | |
1340 | * idle task through an additional NOOP schedule() | |
1341 | */ | |
1342 | set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED); | |
1343 | ||
1344 | /* NEED_RESCHED must be visible before we test polling */ | |
1345 | smp_mb(); | |
1346 | if (!tsk_is_polling(rq->idle)) | |
1347 | smp_send_reschedule(cpu); | |
1348 | } | |
1349 | #endif | |
1350 | ||
c24d20db | 1351 | #else |
8f4d37ec | 1352 | static void __resched_task(struct task_struct *p, int tif_bit) |
c24d20db IM |
1353 | { |
1354 | assert_spin_locked(&task_rq(p)->lock); | |
8f4d37ec | 1355 | set_tsk_thread_flag(p, tif_bit); |
c24d20db IM |
1356 | } |
1357 | #endif | |
1358 | ||
45bf76df IM |
1359 | #if BITS_PER_LONG == 32 |
1360 | # define WMULT_CONST (~0UL) | |
1361 | #else | |
1362 | # define WMULT_CONST (1UL << 32) | |
1363 | #endif | |
1364 | ||
1365 | #define WMULT_SHIFT 32 | |
1366 | ||
194081eb IM |
1367 | /* |
1368 | * Shift right and round: | |
1369 | */ | |
cf2ab469 | 1370 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) |
194081eb | 1371 | |
8f1bc385 PZ |
1372 | /* |
1373 | * delta *= weight / lw | |
1374 | */ | |
cb1c4fc9 | 1375 | static unsigned long |
45bf76df IM |
1376 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, |
1377 | struct load_weight *lw) | |
1378 | { | |
1379 | u64 tmp; | |
1380 | ||
e05510d0 PZ |
1381 | if (!lw->inv_weight) |
1382 | lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1); | |
45bf76df IM |
1383 | |
1384 | tmp = (u64)delta_exec * weight; | |
1385 | /* | |
1386 | * Check whether we'd overflow the 64-bit multiplication: | |
1387 | */ | |
194081eb | 1388 | if (unlikely(tmp > WMULT_CONST)) |
cf2ab469 | 1389 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, |
194081eb IM |
1390 | WMULT_SHIFT/2); |
1391 | else | |
cf2ab469 | 1392 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); |
45bf76df | 1393 | |
ecf691da | 1394 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); |
45bf76df IM |
1395 | } |
1396 | ||
1091985b | 1397 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
45bf76df IM |
1398 | { |
1399 | lw->weight += inc; | |
e89996ae | 1400 | lw->inv_weight = 0; |
45bf76df IM |
1401 | } |
1402 | ||
1091985b | 1403 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) |
45bf76df IM |
1404 | { |
1405 | lw->weight -= dec; | |
e89996ae | 1406 | lw->inv_weight = 0; |
45bf76df IM |
1407 | } |
1408 | ||
2dd73a4f PW |
1409 | /* |
1410 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
1411 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
1412 | * each task makes to its run queue's load is weighted according to its | |
41a2d6cf | 1413 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a |
2dd73a4f PW |
1414 | * scaled version of the new time slice allocation that they receive on time |
1415 | * slice expiry etc. | |
1416 | */ | |
1417 | ||
dd41f596 IM |
1418 | #define WEIGHT_IDLEPRIO 2 |
1419 | #define WMULT_IDLEPRIO (1 << 31) | |
1420 | ||
1421 | /* | |
1422 | * Nice levels are multiplicative, with a gentle 10% change for every | |
1423 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to | |
1424 | * nice 1, it will get ~10% less CPU time than another CPU-bound task | |
1425 | * that remained on nice 0. | |
1426 | * | |
1427 | * The "10% effect" is relative and cumulative: from _any_ nice level, | |
1428 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level | |
f9153ee6 IM |
1429 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. |
1430 | * If a task goes up by ~10% and another task goes down by ~10% then | |
1431 | * the relative distance between them is ~25%.) | |
dd41f596 IM |
1432 | */ |
1433 | static const int prio_to_weight[40] = { | |
254753dc IM |
1434 | /* -20 */ 88761, 71755, 56483, 46273, 36291, |
1435 | /* -15 */ 29154, 23254, 18705, 14949, 11916, | |
1436 | /* -10 */ 9548, 7620, 6100, 4904, 3906, | |
1437 | /* -5 */ 3121, 2501, 1991, 1586, 1277, | |
1438 | /* 0 */ 1024, 820, 655, 526, 423, | |
1439 | /* 5 */ 335, 272, 215, 172, 137, | |
1440 | /* 10 */ 110, 87, 70, 56, 45, | |
1441 | /* 15 */ 36, 29, 23, 18, 15, | |
dd41f596 IM |
1442 | }; |
1443 | ||
5714d2de IM |
1444 | /* |
1445 | * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. | |
1446 | * | |
1447 | * In cases where the weight does not change often, we can use the | |
1448 | * precalculated inverse to speed up arithmetics by turning divisions | |
1449 | * into multiplications: | |
1450 | */ | |
dd41f596 | 1451 | static const u32 prio_to_wmult[40] = { |
254753dc IM |
1452 | /* -20 */ 48388, 59856, 76040, 92818, 118348, |
1453 | /* -15 */ 147320, 184698, 229616, 287308, 360437, | |
1454 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, | |
1455 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, | |
1456 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, | |
1457 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, | |
1458 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, | |
1459 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, | |
dd41f596 | 1460 | }; |
2dd73a4f | 1461 | |
dd41f596 IM |
1462 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup); |
1463 | ||
1464 | /* | |
1465 | * runqueue iterator, to support SMP load-balancing between different | |
1466 | * scheduling classes, without having to expose their internal data | |
1467 | * structures to the load-balancing proper: | |
1468 | */ | |
1469 | struct rq_iterator { | |
1470 | void *arg; | |
1471 | struct task_struct *(*start)(void *); | |
1472 | struct task_struct *(*next)(void *); | |
1473 | }; | |
1474 | ||
e1d1484f PW |
1475 | #ifdef CONFIG_SMP |
1476 | static unsigned long | |
1477 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1478 | unsigned long max_load_move, struct sched_domain *sd, | |
1479 | enum cpu_idle_type idle, int *all_pinned, | |
1480 | int *this_best_prio, struct rq_iterator *iterator); | |
1481 | ||
1482 | static int | |
1483 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1484 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1485 | struct rq_iterator *iterator); | |
e1d1484f | 1486 | #endif |
dd41f596 | 1487 | |
d842de87 SV |
1488 | #ifdef CONFIG_CGROUP_CPUACCT |
1489 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime); | |
1490 | #else | |
1491 | static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} | |
1492 | #endif | |
1493 | ||
18d95a28 PZ |
1494 | static inline void inc_cpu_load(struct rq *rq, unsigned long load) |
1495 | { | |
1496 | update_load_add(&rq->load, load); | |
1497 | } | |
1498 | ||
1499 | static inline void dec_cpu_load(struct rq *rq, unsigned long load) | |
1500 | { | |
1501 | update_load_sub(&rq->load, load); | |
1502 | } | |
1503 | ||
e7693a36 GH |
1504 | #ifdef CONFIG_SMP |
1505 | static unsigned long source_load(int cpu, int type); | |
1506 | static unsigned long target_load(int cpu, int type); | |
1507 | static unsigned long cpu_avg_load_per_task(int cpu); | |
1508 | static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd); | |
18d95a28 PZ |
1509 | |
1510 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1511 | ||
1512 | /* | |
1513 | * Group load balancing. | |
1514 | * | |
1515 | * We calculate a few balance domain wide aggregate numbers; load and weight. | |
1516 | * Given the pictures below, and assuming each item has equal weight: | |
1517 | * | |
1518 | * root 1 - thread | |
1519 | * / | \ A - group | |
1520 | * A 1 B | |
1521 | * /|\ / \ | |
1522 | * C 2 D 3 4 | |
1523 | * | | | |
1524 | * 5 6 | |
1525 | * | |
1526 | * load: | |
1527 | * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd, | |
1528 | * which equals 1/9-th of the total load. | |
1529 | * | |
1530 | * shares: | |
1531 | * The weight of this group on the selected cpus. | |
1532 | * | |
1533 | * rq_weight: | |
1534 | * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while | |
1535 | * B would get 2. | |
1536 | * | |
1537 | * task_weight: | |
1538 | * Part of the rq_weight contributed by tasks; all groups except B would | |
1539 | * get 1, B gets 2. | |
1540 | */ | |
1541 | ||
1542 | static inline struct aggregate_struct * | |
1543 | aggregate(struct task_group *tg, struct sched_domain *sd) | |
1544 | { | |
1545 | return &tg->cfs_rq[sd->first_cpu]->aggregate; | |
1546 | } | |
1547 | ||
1548 | typedef void (*aggregate_func)(struct task_group *, struct sched_domain *); | |
1549 | ||
1550 | /* | |
1551 | * Iterate the full tree, calling @down when first entering a node and @up when | |
1552 | * leaving it for the final time. | |
1553 | */ | |
1554 | static | |
1555 | void aggregate_walk_tree(aggregate_func down, aggregate_func up, | |
1556 | struct sched_domain *sd) | |
1557 | { | |
1558 | struct task_group *parent, *child; | |
1559 | ||
1560 | rcu_read_lock(); | |
1561 | parent = &root_task_group; | |
1562 | down: | |
1563 | (*down)(parent, sd); | |
1564 | list_for_each_entry_rcu(child, &parent->children, siblings) { | |
1565 | parent = child; | |
1566 | goto down; | |
1567 | ||
1568 | up: | |
1569 | continue; | |
1570 | } | |
1571 | (*up)(parent, sd); | |
1572 | ||
1573 | child = parent; | |
1574 | parent = parent->parent; | |
1575 | if (parent) | |
1576 | goto up; | |
1577 | rcu_read_unlock(); | |
1578 | } | |
1579 | ||
1580 | /* | |
1581 | * Calculate the aggregate runqueue weight. | |
1582 | */ | |
1583 | static | |
1584 | void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd) | |
1585 | { | |
1586 | unsigned long rq_weight = 0; | |
1587 | unsigned long task_weight = 0; | |
1588 | int i; | |
1589 | ||
1590 | for_each_cpu_mask(i, sd->span) { | |
1591 | rq_weight += tg->cfs_rq[i]->load.weight; | |
1592 | task_weight += tg->cfs_rq[i]->task_weight; | |
1593 | } | |
1594 | ||
1595 | aggregate(tg, sd)->rq_weight = rq_weight; | |
1596 | aggregate(tg, sd)->task_weight = task_weight; | |
1597 | } | |
1598 | ||
18d95a28 PZ |
1599 | /* |
1600 | * Compute the weight of this group on the given cpus. | |
1601 | */ | |
1602 | static | |
1603 | void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd) | |
1604 | { | |
1605 | unsigned long shares = 0; | |
1606 | int i; | |
1607 | ||
18d95a28 PZ |
1608 | for_each_cpu_mask(i, sd->span) |
1609 | shares += tg->cfs_rq[i]->shares; | |
1610 | ||
3f5087a2 PZ |
1611 | if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares) |
1612 | shares = tg->shares; | |
18d95a28 PZ |
1613 | |
1614 | aggregate(tg, sd)->shares = shares; | |
1615 | } | |
1616 | ||
1617 | /* | |
1618 | * Compute the load fraction assigned to this group, relies on the aggregate | |
1619 | * weight and this group's parent's load, i.e. top-down. | |
1620 | */ | |
1621 | static | |
1622 | void aggregate_group_load(struct task_group *tg, struct sched_domain *sd) | |
1623 | { | |
1624 | unsigned long load; | |
1625 | ||
1626 | if (!tg->parent) { | |
1627 | int i; | |
1628 | ||
1629 | load = 0; | |
1630 | for_each_cpu_mask(i, sd->span) | |
1631 | load += cpu_rq(i)->load.weight; | |
1632 | ||
1633 | } else { | |
1634 | load = aggregate(tg->parent, sd)->load; | |
1635 | ||
1636 | /* | |
1637 | * shares is our weight in the parent's rq so | |
1638 | * shares/parent->rq_weight gives our fraction of the load | |
1639 | */ | |
1640 | load *= aggregate(tg, sd)->shares; | |
1641 | load /= aggregate(tg->parent, sd)->rq_weight + 1; | |
1642 | } | |
1643 | ||
1644 | aggregate(tg, sd)->load = load; | |
1645 | } | |
1646 | ||
1647 | static void __set_se_shares(struct sched_entity *se, unsigned long shares); | |
1648 | ||
1649 | /* | |
1650 | * Calculate and set the cpu's group shares. | |
1651 | */ | |
1652 | static void | |
1653 | __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd, | |
1654 | int tcpu) | |
1655 | { | |
1656 | int boost = 0; | |
1657 | unsigned long shares; | |
1658 | unsigned long rq_weight; | |
1659 | ||
1660 | if (!tg->se[tcpu]) | |
1661 | return; | |
1662 | ||
1663 | rq_weight = tg->cfs_rq[tcpu]->load.weight; | |
1664 | ||
1665 | /* | |
1666 | * If there are currently no tasks on the cpu pretend there is one of | |
1667 | * average load so that when a new task gets to run here it will not | |
1668 | * get delayed by group starvation. | |
1669 | */ | |
1670 | if (!rq_weight) { | |
1671 | boost = 1; | |
1672 | rq_weight = NICE_0_LOAD; | |
1673 | } | |
1674 | ||
1675 | /* | |
1676 | * \Sum shares * rq_weight | |
1677 | * shares = ----------------------- | |
1678 | * \Sum rq_weight | |
1679 | * | |
1680 | */ | |
1681 | shares = aggregate(tg, sd)->shares * rq_weight; | |
1682 | shares /= aggregate(tg, sd)->rq_weight + 1; | |
1683 | ||
1684 | /* | |
1685 | * record the actual number of shares, not the boosted amount. | |
1686 | */ | |
1687 | tg->cfs_rq[tcpu]->shares = boost ? 0 : shares; | |
1688 | ||
1689 | if (shares < MIN_SHARES) | |
1690 | shares = MIN_SHARES; | |
cb4ad1ff MX |
1691 | else if (shares > MAX_SHARES) |
1692 | shares = MAX_SHARES; | |
18d95a28 PZ |
1693 | |
1694 | __set_se_shares(tg->se[tcpu], shares); | |
1695 | } | |
1696 | ||
1697 | /* | |
1698 | * Re-adjust the weights on the cpu the task came from and on the cpu the | |
1699 | * task went to. | |
1700 | */ | |
1701 | static void | |
1702 | __move_group_shares(struct task_group *tg, struct sched_domain *sd, | |
1703 | int scpu, int dcpu) | |
1704 | { | |
1705 | unsigned long shares; | |
1706 | ||
1707 | shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares; | |
1708 | ||
1709 | __update_group_shares_cpu(tg, sd, scpu); | |
1710 | __update_group_shares_cpu(tg, sd, dcpu); | |
1711 | ||
1712 | /* | |
1713 | * ensure we never loose shares due to rounding errors in the | |
1714 | * above redistribution. | |
1715 | */ | |
1716 | shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares; | |
1717 | if (shares) | |
1718 | tg->cfs_rq[dcpu]->shares += shares; | |
1719 | } | |
1720 | ||
1721 | /* | |
1722 | * Because changing a group's shares changes the weight of the super-group | |
1723 | * we need to walk up the tree and change all shares until we hit the root. | |
1724 | */ | |
1725 | static void | |
1726 | move_group_shares(struct task_group *tg, struct sched_domain *sd, | |
1727 | int scpu, int dcpu) | |
1728 | { | |
1729 | while (tg) { | |
1730 | __move_group_shares(tg, sd, scpu, dcpu); | |
1731 | tg = tg->parent; | |
1732 | } | |
1733 | } | |
1734 | ||
1735 | static | |
1736 | void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd) | |
1737 | { | |
1738 | unsigned long shares = aggregate(tg, sd)->shares; | |
1739 | int i; | |
1740 | ||
1741 | for_each_cpu_mask(i, sd->span) { | |
1742 | struct rq *rq = cpu_rq(i); | |
1743 | unsigned long flags; | |
1744 | ||
1745 | spin_lock_irqsave(&rq->lock, flags); | |
1746 | __update_group_shares_cpu(tg, sd, i); | |
1747 | spin_unlock_irqrestore(&rq->lock, flags); | |
1748 | } | |
1749 | ||
1750 | aggregate_group_shares(tg, sd); | |
1751 | ||
1752 | /* | |
1753 | * ensure we never loose shares due to rounding errors in the | |
1754 | * above redistribution. | |
1755 | */ | |
1756 | shares -= aggregate(tg, sd)->shares; | |
1757 | if (shares) { | |
1758 | tg->cfs_rq[sd->first_cpu]->shares += shares; | |
1759 | aggregate(tg, sd)->shares += shares; | |
1760 | } | |
1761 | } | |
1762 | ||
1763 | /* | |
1764 | * Calculate the accumulative weight and recursive load of each task group | |
1765 | * while walking down the tree. | |
1766 | */ | |
1767 | static | |
1768 | void aggregate_get_down(struct task_group *tg, struct sched_domain *sd) | |
1769 | { | |
1770 | aggregate_group_weight(tg, sd); | |
1771 | aggregate_group_shares(tg, sd); | |
1772 | aggregate_group_load(tg, sd); | |
1773 | } | |
1774 | ||
1775 | /* | |
1776 | * Rebalance the cpu shares while walking back up the tree. | |
1777 | */ | |
1778 | static | |
1779 | void aggregate_get_up(struct task_group *tg, struct sched_domain *sd) | |
1780 | { | |
1781 | aggregate_group_set_shares(tg, sd); | |
1782 | } | |
1783 | ||
1784 | static DEFINE_PER_CPU(spinlock_t, aggregate_lock); | |
1785 | ||
1786 | static void __init init_aggregate(void) | |
1787 | { | |
1788 | int i; | |
1789 | ||
1790 | for_each_possible_cpu(i) | |
1791 | spin_lock_init(&per_cpu(aggregate_lock, i)); | |
1792 | } | |
1793 | ||
1794 | static int get_aggregate(struct sched_domain *sd) | |
1795 | { | |
1796 | if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu))) | |
1797 | return 0; | |
1798 | ||
1799 | aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd); | |
1800 | return 1; | |
1801 | } | |
1802 | ||
1803 | static void put_aggregate(struct sched_domain *sd) | |
1804 | { | |
1805 | spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu)); | |
1806 | } | |
1807 | ||
1808 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1809 | { | |
1810 | cfs_rq->shares = shares; | |
1811 | } | |
1812 | ||
1813 | #else | |
1814 | ||
1815 | static inline void init_aggregate(void) | |
1816 | { | |
1817 | } | |
1818 | ||
1819 | static inline int get_aggregate(struct sched_domain *sd) | |
1820 | { | |
1821 | return 0; | |
1822 | } | |
1823 | ||
1824 | static inline void put_aggregate(struct sched_domain *sd) | |
1825 | { | |
1826 | } | |
1827 | #endif | |
1828 | ||
1829 | #else /* CONFIG_SMP */ | |
1830 | ||
1831 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1832 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1833 | { | |
1834 | } | |
1835 | #endif | |
1836 | ||
e7693a36 GH |
1837 | #endif /* CONFIG_SMP */ |
1838 | ||
dd41f596 | 1839 | #include "sched_stats.h" |
dd41f596 | 1840 | #include "sched_idletask.c" |
5522d5d5 IM |
1841 | #include "sched_fair.c" |
1842 | #include "sched_rt.c" | |
dd41f596 IM |
1843 | #ifdef CONFIG_SCHED_DEBUG |
1844 | # include "sched_debug.c" | |
1845 | #endif | |
1846 | ||
1847 | #define sched_class_highest (&rt_sched_class) | |
1848 | ||
18d95a28 | 1849 | static void inc_nr_running(struct rq *rq) |
9c217245 IM |
1850 | { |
1851 | rq->nr_running++; | |
9c217245 IM |
1852 | } |
1853 | ||
18d95a28 | 1854 | static void dec_nr_running(struct rq *rq) |
9c217245 IM |
1855 | { |
1856 | rq->nr_running--; | |
9c217245 IM |
1857 | } |
1858 | ||
45bf76df IM |
1859 | static void set_load_weight(struct task_struct *p) |
1860 | { | |
1861 | if (task_has_rt_policy(p)) { | |
dd41f596 IM |
1862 | p->se.load.weight = prio_to_weight[0] * 2; |
1863 | p->se.load.inv_weight = prio_to_wmult[0] >> 1; | |
1864 | return; | |
1865 | } | |
45bf76df | 1866 | |
dd41f596 IM |
1867 | /* |
1868 | * SCHED_IDLE tasks get minimal weight: | |
1869 | */ | |
1870 | if (p->policy == SCHED_IDLE) { | |
1871 | p->se.load.weight = WEIGHT_IDLEPRIO; | |
1872 | p->se.load.inv_weight = WMULT_IDLEPRIO; | |
1873 | return; | |
1874 | } | |
71f8bd46 | 1875 | |
dd41f596 IM |
1876 | p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO]; |
1877 | p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO]; | |
71f8bd46 IM |
1878 | } |
1879 | ||
8159f87e | 1880 | static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup) |
71f8bd46 | 1881 | { |
dd41f596 | 1882 | sched_info_queued(p); |
fd390f6a | 1883 | p->sched_class->enqueue_task(rq, p, wakeup); |
dd41f596 | 1884 | p->se.on_rq = 1; |
71f8bd46 IM |
1885 | } |
1886 | ||
69be72c1 | 1887 | static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep) |
71f8bd46 | 1888 | { |
f02231e5 | 1889 | p->sched_class->dequeue_task(rq, p, sleep); |
dd41f596 | 1890 | p->se.on_rq = 0; |
71f8bd46 IM |
1891 | } |
1892 | ||
14531189 | 1893 | /* |
dd41f596 | 1894 | * __normal_prio - return the priority that is based on the static prio |
14531189 | 1895 | */ |
14531189 IM |
1896 | static inline int __normal_prio(struct task_struct *p) |
1897 | { | |
dd41f596 | 1898 | return p->static_prio; |
14531189 IM |
1899 | } |
1900 | ||
b29739f9 IM |
1901 | /* |
1902 | * Calculate the expected normal priority: i.e. priority | |
1903 | * without taking RT-inheritance into account. Might be | |
1904 | * boosted by interactivity modifiers. Changes upon fork, | |
1905 | * setprio syscalls, and whenever the interactivity | |
1906 | * estimator recalculates. | |
1907 | */ | |
36c8b586 | 1908 | static inline int normal_prio(struct task_struct *p) |
b29739f9 IM |
1909 | { |
1910 | int prio; | |
1911 | ||
e05606d3 | 1912 | if (task_has_rt_policy(p)) |
b29739f9 IM |
1913 | prio = MAX_RT_PRIO-1 - p->rt_priority; |
1914 | else | |
1915 | prio = __normal_prio(p); | |
1916 | return prio; | |
1917 | } | |
1918 | ||
1919 | /* | |
1920 | * Calculate the current priority, i.e. the priority | |
1921 | * taken into account by the scheduler. This value might | |
1922 | * be boosted by RT tasks, or might be boosted by | |
1923 | * interactivity modifiers. Will be RT if the task got | |
1924 | * RT-boosted. If not then it returns p->normal_prio. | |
1925 | */ | |
36c8b586 | 1926 | static int effective_prio(struct task_struct *p) |
b29739f9 IM |
1927 | { |
1928 | p->normal_prio = normal_prio(p); | |
1929 | /* | |
1930 | * If we are RT tasks or we were boosted to RT priority, | |
1931 | * keep the priority unchanged. Otherwise, update priority | |
1932 | * to the normal priority: | |
1933 | */ | |
1934 | if (!rt_prio(p->prio)) | |
1935 | return p->normal_prio; | |
1936 | return p->prio; | |
1937 | } | |
1938 | ||
1da177e4 | 1939 | /* |
dd41f596 | 1940 | * activate_task - move a task to the runqueue. |
1da177e4 | 1941 | */ |
dd41f596 | 1942 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup) |
1da177e4 | 1943 | { |
d9514f6c | 1944 | if (task_contributes_to_load(p)) |
dd41f596 | 1945 | rq->nr_uninterruptible--; |
1da177e4 | 1946 | |
8159f87e | 1947 | enqueue_task(rq, p, wakeup); |
18d95a28 | 1948 | inc_nr_running(rq); |
1da177e4 LT |
1949 | } |
1950 | ||
1da177e4 LT |
1951 | /* |
1952 | * deactivate_task - remove a task from the runqueue. | |
1953 | */ | |
2e1cb74a | 1954 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep) |
1da177e4 | 1955 | { |
d9514f6c | 1956 | if (task_contributes_to_load(p)) |
dd41f596 IM |
1957 | rq->nr_uninterruptible++; |
1958 | ||
69be72c1 | 1959 | dequeue_task(rq, p, sleep); |
18d95a28 | 1960 | dec_nr_running(rq); |
1da177e4 LT |
1961 | } |
1962 | ||
1da177e4 LT |
1963 | /** |
1964 | * task_curr - is this task currently executing on a CPU? | |
1965 | * @p: the task in question. | |
1966 | */ | |
36c8b586 | 1967 | inline int task_curr(const struct task_struct *p) |
1da177e4 LT |
1968 | { |
1969 | return cpu_curr(task_cpu(p)) == p; | |
1970 | } | |
1971 | ||
2dd73a4f PW |
1972 | /* Used instead of source_load when we know the type == 0 */ |
1973 | unsigned long weighted_cpuload(const int cpu) | |
1974 | { | |
495eca49 | 1975 | return cpu_rq(cpu)->load.weight; |
dd41f596 IM |
1976 | } |
1977 | ||
1978 | static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) | |
1979 | { | |
6f505b16 | 1980 | set_task_rq(p, cpu); |
dd41f596 | 1981 | #ifdef CONFIG_SMP |
ce96b5ac DA |
1982 | /* |
1983 | * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be | |
1984 | * successfuly executed on another CPU. We must ensure that updates of | |
1985 | * per-task data have been completed by this moment. | |
1986 | */ | |
1987 | smp_wmb(); | |
dd41f596 | 1988 | task_thread_info(p)->cpu = cpu; |
dd41f596 | 1989 | #endif |
2dd73a4f PW |
1990 | } |
1991 | ||
cb469845 SR |
1992 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, |
1993 | const struct sched_class *prev_class, | |
1994 | int oldprio, int running) | |
1995 | { | |
1996 | if (prev_class != p->sched_class) { | |
1997 | if (prev_class->switched_from) | |
1998 | prev_class->switched_from(rq, p, running); | |
1999 | p->sched_class->switched_to(rq, p, running); | |
2000 | } else | |
2001 | p->sched_class->prio_changed(rq, p, oldprio, running); | |
2002 | } | |
2003 | ||
1da177e4 | 2004 | #ifdef CONFIG_SMP |
c65cc870 | 2005 | |
cc367732 IM |
2006 | /* |
2007 | * Is this task likely cache-hot: | |
2008 | */ | |
e7693a36 | 2009 | static int |
cc367732 IM |
2010 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) |
2011 | { | |
2012 | s64 delta; | |
2013 | ||
f540a608 IM |
2014 | /* |
2015 | * Buddy candidates are cache hot: | |
2016 | */ | |
d25ce4cd | 2017 | if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next)) |
f540a608 IM |
2018 | return 1; |
2019 | ||
cc367732 IM |
2020 | if (p->sched_class != &fair_sched_class) |
2021 | return 0; | |
2022 | ||
6bc1665b IM |
2023 | if (sysctl_sched_migration_cost == -1) |
2024 | return 1; | |
2025 | if (sysctl_sched_migration_cost == 0) | |
2026 | return 0; | |
2027 | ||
cc367732 IM |
2028 | delta = now - p->se.exec_start; |
2029 | ||
2030 | return delta < (s64)sysctl_sched_migration_cost; | |
2031 | } | |
2032 | ||
2033 | ||
dd41f596 | 2034 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
c65cc870 | 2035 | { |
dd41f596 IM |
2036 | int old_cpu = task_cpu(p); |
2037 | struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu); | |
2830cf8c SV |
2038 | struct cfs_rq *old_cfsrq = task_cfs_rq(p), |
2039 | *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu); | |
bbdba7c0 | 2040 | u64 clock_offset; |
dd41f596 IM |
2041 | |
2042 | clock_offset = old_rq->clock - new_rq->clock; | |
6cfb0d5d IM |
2043 | |
2044 | #ifdef CONFIG_SCHEDSTATS | |
2045 | if (p->se.wait_start) | |
2046 | p->se.wait_start -= clock_offset; | |
dd41f596 IM |
2047 | if (p->se.sleep_start) |
2048 | p->se.sleep_start -= clock_offset; | |
2049 | if (p->se.block_start) | |
2050 | p->se.block_start -= clock_offset; | |
cc367732 IM |
2051 | if (old_cpu != new_cpu) { |
2052 | schedstat_inc(p, se.nr_migrations); | |
2053 | if (task_hot(p, old_rq->clock, NULL)) | |
2054 | schedstat_inc(p, se.nr_forced2_migrations); | |
2055 | } | |
6cfb0d5d | 2056 | #endif |
2830cf8c SV |
2057 | p->se.vruntime -= old_cfsrq->min_vruntime - |
2058 | new_cfsrq->min_vruntime; | |
dd41f596 IM |
2059 | |
2060 | __set_task_cpu(p, new_cpu); | |
c65cc870 IM |
2061 | } |
2062 | ||
70b97a7f | 2063 | struct migration_req { |
1da177e4 | 2064 | struct list_head list; |
1da177e4 | 2065 | |
36c8b586 | 2066 | struct task_struct *task; |
1da177e4 LT |
2067 | int dest_cpu; |
2068 | ||
1da177e4 | 2069 | struct completion done; |
70b97a7f | 2070 | }; |
1da177e4 LT |
2071 | |
2072 | /* | |
2073 | * The task's runqueue lock must be held. | |
2074 | * Returns true if you have to wait for migration thread. | |
2075 | */ | |
36c8b586 | 2076 | static int |
70b97a7f | 2077 | migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req) |
1da177e4 | 2078 | { |
70b97a7f | 2079 | struct rq *rq = task_rq(p); |
1da177e4 LT |
2080 | |
2081 | /* | |
2082 | * If the task is not on a runqueue (and not running), then | |
2083 | * it is sufficient to simply update the task's cpu field. | |
2084 | */ | |
dd41f596 | 2085 | if (!p->se.on_rq && !task_running(rq, p)) { |
1da177e4 LT |
2086 | set_task_cpu(p, dest_cpu); |
2087 | return 0; | |
2088 | } | |
2089 | ||
2090 | init_completion(&req->done); | |
1da177e4 LT |
2091 | req->task = p; |
2092 | req->dest_cpu = dest_cpu; | |
2093 | list_add(&req->list, &rq->migration_queue); | |
48f24c4d | 2094 | |
1da177e4 LT |
2095 | return 1; |
2096 | } | |
2097 | ||
2098 | /* | |
2099 | * wait_task_inactive - wait for a thread to unschedule. | |
2100 | * | |
2101 | * The caller must ensure that the task *will* unschedule sometime soon, | |
2102 | * else this function might spin for a *long* time. This function can't | |
2103 | * be called with interrupts off, or it may introduce deadlock with | |
2104 | * smp_call_function() if an IPI is sent by the same process we are | |
2105 | * waiting to become inactive. | |
2106 | */ | |
36c8b586 | 2107 | void wait_task_inactive(struct task_struct *p) |
1da177e4 LT |
2108 | { |
2109 | unsigned long flags; | |
dd41f596 | 2110 | int running, on_rq; |
70b97a7f | 2111 | struct rq *rq; |
1da177e4 | 2112 | |
3a5c359a AK |
2113 | for (;;) { |
2114 | /* | |
2115 | * We do the initial early heuristics without holding | |
2116 | * any task-queue locks at all. We'll only try to get | |
2117 | * the runqueue lock when things look like they will | |
2118 | * work out! | |
2119 | */ | |
2120 | rq = task_rq(p); | |
fa490cfd | 2121 | |
3a5c359a AK |
2122 | /* |
2123 | * If the task is actively running on another CPU | |
2124 | * still, just relax and busy-wait without holding | |
2125 | * any locks. | |
2126 | * | |
2127 | * NOTE! Since we don't hold any locks, it's not | |
2128 | * even sure that "rq" stays as the right runqueue! | |
2129 | * But we don't care, since "task_running()" will | |
2130 | * return false if the runqueue has changed and p | |
2131 | * is actually now running somewhere else! | |
2132 | */ | |
2133 | while (task_running(rq, p)) | |
2134 | cpu_relax(); | |
fa490cfd | 2135 | |
3a5c359a AK |
2136 | /* |
2137 | * Ok, time to look more closely! We need the rq | |
2138 | * lock now, to be *sure*. If we're wrong, we'll | |
2139 | * just go back and repeat. | |
2140 | */ | |
2141 | rq = task_rq_lock(p, &flags); | |
2142 | running = task_running(rq, p); | |
2143 | on_rq = p->se.on_rq; | |
2144 | task_rq_unlock(rq, &flags); | |
fa490cfd | 2145 | |
3a5c359a AK |
2146 | /* |
2147 | * Was it really running after all now that we | |
2148 | * checked with the proper locks actually held? | |
2149 | * | |
2150 | * Oops. Go back and try again.. | |
2151 | */ | |
2152 | if (unlikely(running)) { | |
2153 | cpu_relax(); | |
2154 | continue; | |
2155 | } | |
fa490cfd | 2156 | |
3a5c359a AK |
2157 | /* |
2158 | * It's not enough that it's not actively running, | |
2159 | * it must be off the runqueue _entirely_, and not | |
2160 | * preempted! | |
2161 | * | |
2162 | * So if it wa still runnable (but just not actively | |
2163 | * running right now), it's preempted, and we should | |
2164 | * yield - it could be a while. | |
2165 | */ | |
2166 | if (unlikely(on_rq)) { | |
2167 | schedule_timeout_uninterruptible(1); | |
2168 | continue; | |
2169 | } | |
fa490cfd | 2170 | |
3a5c359a AK |
2171 | /* |
2172 | * Ahh, all good. It wasn't running, and it wasn't | |
2173 | * runnable, which means that it will never become | |
2174 | * running in the future either. We're all done! | |
2175 | */ | |
2176 | break; | |
2177 | } | |
1da177e4 LT |
2178 | } |
2179 | ||
2180 | /*** | |
2181 | * kick_process - kick a running thread to enter/exit the kernel | |
2182 | * @p: the to-be-kicked thread | |
2183 | * | |
2184 | * Cause a process which is running on another CPU to enter | |
2185 | * kernel-mode, without any delay. (to get signals handled.) | |
2186 | * | |
2187 | * NOTE: this function doesnt have to take the runqueue lock, | |
2188 | * because all it wants to ensure is that the remote task enters | |
2189 | * the kernel. If the IPI races and the task has been migrated | |
2190 | * to another CPU then no harm is done and the purpose has been | |
2191 | * achieved as well. | |
2192 | */ | |
36c8b586 | 2193 | void kick_process(struct task_struct *p) |
1da177e4 LT |
2194 | { |
2195 | int cpu; | |
2196 | ||
2197 | preempt_disable(); | |
2198 | cpu = task_cpu(p); | |
2199 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
2200 | smp_send_reschedule(cpu); | |
2201 | preempt_enable(); | |
2202 | } | |
2203 | ||
2204 | /* | |
2dd73a4f PW |
2205 | * Return a low guess at the load of a migration-source cpu weighted |
2206 | * according to the scheduling class and "nice" value. | |
1da177e4 LT |
2207 | * |
2208 | * We want to under-estimate the load of migration sources, to | |
2209 | * balance conservatively. | |
2210 | */ | |
a9957449 | 2211 | static unsigned long source_load(int cpu, int type) |
1da177e4 | 2212 | { |
70b97a7f | 2213 | struct rq *rq = cpu_rq(cpu); |
dd41f596 | 2214 | unsigned long total = weighted_cpuload(cpu); |
2dd73a4f | 2215 | |
3b0bd9bc | 2216 | if (type == 0) |
dd41f596 | 2217 | return total; |
b910472d | 2218 | |
dd41f596 | 2219 | return min(rq->cpu_load[type-1], total); |
1da177e4 LT |
2220 | } |
2221 | ||
2222 | /* | |
2dd73a4f PW |
2223 | * Return a high guess at the load of a migration-target cpu weighted |
2224 | * according to the scheduling class and "nice" value. | |
1da177e4 | 2225 | */ |
a9957449 | 2226 | static unsigned long target_load(int cpu, int type) |
1da177e4 | 2227 | { |
70b97a7f | 2228 | struct rq *rq = cpu_rq(cpu); |
dd41f596 | 2229 | unsigned long total = weighted_cpuload(cpu); |
2dd73a4f | 2230 | |
7897986b | 2231 | if (type == 0) |
dd41f596 | 2232 | return total; |
3b0bd9bc | 2233 | |
dd41f596 | 2234 | return max(rq->cpu_load[type-1], total); |
2dd73a4f PW |
2235 | } |
2236 | ||
2237 | /* | |
2238 | * Return the average load per task on the cpu's run queue | |
2239 | */ | |
e7693a36 | 2240 | static unsigned long cpu_avg_load_per_task(int cpu) |
2dd73a4f | 2241 | { |
70b97a7f | 2242 | struct rq *rq = cpu_rq(cpu); |
dd41f596 | 2243 | unsigned long total = weighted_cpuload(cpu); |
2dd73a4f PW |
2244 | unsigned long n = rq->nr_running; |
2245 | ||
dd41f596 | 2246 | return n ? total / n : SCHED_LOAD_SCALE; |
1da177e4 LT |
2247 | } |
2248 | ||
147cbb4b NP |
2249 | /* |
2250 | * find_idlest_group finds and returns the least busy CPU group within the | |
2251 | * domain. | |
2252 | */ | |
2253 | static struct sched_group * | |
2254 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
2255 | { | |
2256 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
2257 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
2258 | int load_idx = sd->forkexec_idx; | |
2259 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
2260 | ||
2261 | do { | |
2262 | unsigned long load, avg_load; | |
2263 | int local_group; | |
2264 | int i; | |
2265 | ||
da5a5522 BD |
2266 | /* Skip over this group if it has no CPUs allowed */ |
2267 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
3a5c359a | 2268 | continue; |
da5a5522 | 2269 | |
147cbb4b | 2270 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
2271 | |
2272 | /* Tally up the load of all CPUs in the group */ | |
2273 | avg_load = 0; | |
2274 | ||
2275 | for_each_cpu_mask(i, group->cpumask) { | |
2276 | /* Bias balancing toward cpus of our domain */ | |
2277 | if (local_group) | |
2278 | load = source_load(i, load_idx); | |
2279 | else | |
2280 | load = target_load(i, load_idx); | |
2281 | ||
2282 | avg_load += load; | |
2283 | } | |
2284 | ||
2285 | /* Adjust by relative CPU power of the group */ | |
5517d86b ED |
2286 | avg_load = sg_div_cpu_power(group, |
2287 | avg_load * SCHED_LOAD_SCALE); | |
147cbb4b NP |
2288 | |
2289 | if (local_group) { | |
2290 | this_load = avg_load; | |
2291 | this = group; | |
2292 | } else if (avg_load < min_load) { | |
2293 | min_load = avg_load; | |
2294 | idlest = group; | |
2295 | } | |
3a5c359a | 2296 | } while (group = group->next, group != sd->groups); |
147cbb4b NP |
2297 | |
2298 | if (!idlest || 100*this_load < imbalance*min_load) | |
2299 | return NULL; | |
2300 | return idlest; | |
2301 | } | |
2302 | ||
2303 | /* | |
0feaece9 | 2304 | * find_idlest_cpu - find the idlest cpu among the cpus in group. |
147cbb4b | 2305 | */ |
95cdf3b7 | 2306 | static int |
7c16ec58 MT |
2307 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu, |
2308 | cpumask_t *tmp) | |
147cbb4b NP |
2309 | { |
2310 | unsigned long load, min_load = ULONG_MAX; | |
2311 | int idlest = -1; | |
2312 | int i; | |
2313 | ||
da5a5522 | 2314 | /* Traverse only the allowed CPUs */ |
7c16ec58 | 2315 | cpus_and(*tmp, group->cpumask, p->cpus_allowed); |
da5a5522 | 2316 | |
7c16ec58 | 2317 | for_each_cpu_mask(i, *tmp) { |
2dd73a4f | 2318 | load = weighted_cpuload(i); |
147cbb4b NP |
2319 | |
2320 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
2321 | min_load = load; | |
2322 | idlest = i; | |
2323 | } | |
2324 | } | |
2325 | ||
2326 | return idlest; | |
2327 | } | |
2328 | ||
476d139c NP |
2329 | /* |
2330 | * sched_balance_self: balance the current task (running on cpu) in domains | |
2331 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
2332 | * SD_BALANCE_EXEC. | |
2333 | * | |
2334 | * Balance, ie. select the least loaded group. | |
2335 | * | |
2336 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
2337 | * | |
2338 | * preempt must be disabled. | |
2339 | */ | |
2340 | static int sched_balance_self(int cpu, int flag) | |
2341 | { | |
2342 | struct task_struct *t = current; | |
2343 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 2344 | |
c96d145e | 2345 | for_each_domain(cpu, tmp) { |
9761eea8 IM |
2346 | /* |
2347 | * If power savings logic is enabled for a domain, stop there. | |
2348 | */ | |
5c45bf27 SS |
2349 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
2350 | break; | |
476d139c NP |
2351 | if (tmp->flags & flag) |
2352 | sd = tmp; | |
c96d145e | 2353 | } |
476d139c NP |
2354 | |
2355 | while (sd) { | |
7c16ec58 | 2356 | cpumask_t span, tmpmask; |
476d139c | 2357 | struct sched_group *group; |
1a848870 SS |
2358 | int new_cpu, weight; |
2359 | ||
2360 | if (!(sd->flags & flag)) { | |
2361 | sd = sd->child; | |
2362 | continue; | |
2363 | } | |
476d139c NP |
2364 | |
2365 | span = sd->span; | |
2366 | group = find_idlest_group(sd, t, cpu); | |
1a848870 SS |
2367 | if (!group) { |
2368 | sd = sd->child; | |
2369 | continue; | |
2370 | } | |
476d139c | 2371 | |
7c16ec58 | 2372 | new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask); |
1a848870 SS |
2373 | if (new_cpu == -1 || new_cpu == cpu) { |
2374 | /* Now try balancing at a lower domain level of cpu */ | |
2375 | sd = sd->child; | |
2376 | continue; | |
2377 | } | |
476d139c | 2378 | |
1a848870 | 2379 | /* Now try balancing at a lower domain level of new_cpu */ |
476d139c | 2380 | cpu = new_cpu; |
476d139c NP |
2381 | sd = NULL; |
2382 | weight = cpus_weight(span); | |
2383 | for_each_domain(cpu, tmp) { | |
2384 | if (weight <= cpus_weight(tmp->span)) | |
2385 | break; | |
2386 | if (tmp->flags & flag) | |
2387 | sd = tmp; | |
2388 | } | |
2389 | /* while loop will break here if sd == NULL */ | |
2390 | } | |
2391 | ||
2392 | return cpu; | |
2393 | } | |
2394 | ||
2395 | #endif /* CONFIG_SMP */ | |
1da177e4 | 2396 | |
1da177e4 LT |
2397 | /*** |
2398 | * try_to_wake_up - wake up a thread | |
2399 | * @p: the to-be-woken-up thread | |
2400 | * @state: the mask of task states that can be woken | |
2401 | * @sync: do a synchronous wakeup? | |
2402 | * | |
2403 | * Put it on the run-queue if it's not already there. The "current" | |
2404 | * thread is always on the run-queue (except when the actual | |
2405 | * re-schedule is in progress), and as such you're allowed to do | |
2406 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
2407 | * runnable without the overhead of this. | |
2408 | * | |
2409 | * returns failure only if the task is already active. | |
2410 | */ | |
36c8b586 | 2411 | static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync) |
1da177e4 | 2412 | { |
cc367732 | 2413 | int cpu, orig_cpu, this_cpu, success = 0; |
1da177e4 LT |
2414 | unsigned long flags; |
2415 | long old_state; | |
70b97a7f | 2416 | struct rq *rq; |
1da177e4 | 2417 | |
b85d0667 IM |
2418 | if (!sched_feat(SYNC_WAKEUPS)) |
2419 | sync = 0; | |
2420 | ||
04e2f174 | 2421 | smp_wmb(); |
1da177e4 LT |
2422 | rq = task_rq_lock(p, &flags); |
2423 | old_state = p->state; | |
2424 | if (!(old_state & state)) | |
2425 | goto out; | |
2426 | ||
dd41f596 | 2427 | if (p->se.on_rq) |
1da177e4 LT |
2428 | goto out_running; |
2429 | ||
2430 | cpu = task_cpu(p); | |
cc367732 | 2431 | orig_cpu = cpu; |
1da177e4 LT |
2432 | this_cpu = smp_processor_id(); |
2433 | ||
2434 | #ifdef CONFIG_SMP | |
2435 | if (unlikely(task_running(rq, p))) | |
2436 | goto out_activate; | |
2437 | ||
5d2f5a61 DA |
2438 | cpu = p->sched_class->select_task_rq(p, sync); |
2439 | if (cpu != orig_cpu) { | |
2440 | set_task_cpu(p, cpu); | |
1da177e4 LT |
2441 | task_rq_unlock(rq, &flags); |
2442 | /* might preempt at this point */ | |
2443 | rq = task_rq_lock(p, &flags); | |
2444 | old_state = p->state; | |
2445 | if (!(old_state & state)) | |
2446 | goto out; | |
dd41f596 | 2447 | if (p->se.on_rq) |
1da177e4 LT |
2448 | goto out_running; |
2449 | ||
2450 | this_cpu = smp_processor_id(); | |
2451 | cpu = task_cpu(p); | |
2452 | } | |
2453 | ||
e7693a36 GH |
2454 | #ifdef CONFIG_SCHEDSTATS |
2455 | schedstat_inc(rq, ttwu_count); | |
2456 | if (cpu == this_cpu) | |
2457 | schedstat_inc(rq, ttwu_local); | |
2458 | else { | |
2459 | struct sched_domain *sd; | |
2460 | for_each_domain(this_cpu, sd) { | |
2461 | if (cpu_isset(cpu, sd->span)) { | |
2462 | schedstat_inc(sd, ttwu_wake_remote); | |
2463 | break; | |
2464 | } | |
2465 | } | |
2466 | } | |
e7693a36 GH |
2467 | #endif |
2468 | ||
1da177e4 LT |
2469 | out_activate: |
2470 | #endif /* CONFIG_SMP */ | |
bd3bff9e | 2471 | ftrace_wake_up_task(p, rq->curr); |
cc367732 IM |
2472 | schedstat_inc(p, se.nr_wakeups); |
2473 | if (sync) | |
2474 | schedstat_inc(p, se.nr_wakeups_sync); | |
2475 | if (orig_cpu != cpu) | |
2476 | schedstat_inc(p, se.nr_wakeups_migrate); | |
2477 | if (cpu == this_cpu) | |
2478 | schedstat_inc(p, se.nr_wakeups_local); | |
2479 | else | |
2480 | schedstat_inc(p, se.nr_wakeups_remote); | |
2daa3577 | 2481 | update_rq_clock(rq); |
dd41f596 | 2482 | activate_task(rq, p, 1); |
1da177e4 LT |
2483 | success = 1; |
2484 | ||
2485 | out_running: | |
4ae7d5ce IM |
2486 | check_preempt_curr(rq, p); |
2487 | ||
1da177e4 | 2488 | p->state = TASK_RUNNING; |
9a897c5a SR |
2489 | #ifdef CONFIG_SMP |
2490 | if (p->sched_class->task_wake_up) | |
2491 | p->sched_class->task_wake_up(rq, p); | |
2492 | #endif | |
1da177e4 LT |
2493 | out: |
2494 | task_rq_unlock(rq, &flags); | |
2495 | ||
2496 | return success; | |
2497 | } | |
2498 | ||
7ad5b3a5 | 2499 | int wake_up_process(struct task_struct *p) |
1da177e4 | 2500 | { |
d9514f6c | 2501 | return try_to_wake_up(p, TASK_ALL, 0); |
1da177e4 | 2502 | } |
1da177e4 LT |
2503 | EXPORT_SYMBOL(wake_up_process); |
2504 | ||
7ad5b3a5 | 2505 | int wake_up_state(struct task_struct *p, unsigned int state) |
1da177e4 LT |
2506 | { |
2507 | return try_to_wake_up(p, state, 0); | |
2508 | } | |
2509 | ||
1da177e4 LT |
2510 | /* |
2511 | * Perform scheduler related setup for a newly forked process p. | |
2512 | * p is forked by current. | |
dd41f596 IM |
2513 | * |
2514 | * __sched_fork() is basic setup used by init_idle() too: | |
2515 | */ | |
2516 | static void __sched_fork(struct task_struct *p) | |
2517 | { | |
dd41f596 IM |
2518 | p->se.exec_start = 0; |
2519 | p->se.sum_exec_runtime = 0; | |
f6cf891c | 2520 | p->se.prev_sum_exec_runtime = 0; |
4ae7d5ce IM |
2521 | p->se.last_wakeup = 0; |
2522 | p->se.avg_overlap = 0; | |
6cfb0d5d IM |
2523 | |
2524 | #ifdef CONFIG_SCHEDSTATS | |
2525 | p->se.wait_start = 0; | |
dd41f596 IM |
2526 | p->se.sum_sleep_runtime = 0; |
2527 | p->se.sleep_start = 0; | |
dd41f596 IM |
2528 | p->se.block_start = 0; |
2529 | p->se.sleep_max = 0; | |
2530 | p->se.block_max = 0; | |
2531 | p->se.exec_max = 0; | |
eba1ed4b | 2532 | p->se.slice_max = 0; |
dd41f596 | 2533 | p->se.wait_max = 0; |
6cfb0d5d | 2534 | #endif |
476d139c | 2535 | |
fa717060 | 2536 | INIT_LIST_HEAD(&p->rt.run_list); |
dd41f596 | 2537 | p->se.on_rq = 0; |
4a55bd5e | 2538 | INIT_LIST_HEAD(&p->se.group_node); |
476d139c | 2539 | |
e107be36 AK |
2540 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2541 | INIT_HLIST_HEAD(&p->preempt_notifiers); | |
2542 | #endif | |
2543 | ||
1da177e4 LT |
2544 | /* |
2545 | * We mark the process as running here, but have not actually | |
2546 | * inserted it onto the runqueue yet. This guarantees that | |
2547 | * nobody will actually run it, and a signal or other external | |
2548 | * event cannot wake it up and insert it on the runqueue either. | |
2549 | */ | |
2550 | p->state = TASK_RUNNING; | |
dd41f596 IM |
2551 | } |
2552 | ||
2553 | /* | |
2554 | * fork()/clone()-time setup: | |
2555 | */ | |
2556 | void sched_fork(struct task_struct *p, int clone_flags) | |
2557 | { | |
2558 | int cpu = get_cpu(); | |
2559 | ||
2560 | __sched_fork(p); | |
2561 | ||
2562 | #ifdef CONFIG_SMP | |
2563 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
2564 | #endif | |
02e4bac2 | 2565 | set_task_cpu(p, cpu); |
b29739f9 IM |
2566 | |
2567 | /* | |
2568 | * Make sure we do not leak PI boosting priority to the child: | |
2569 | */ | |
2570 | p->prio = current->normal_prio; | |
2ddbf952 HS |
2571 | if (!rt_prio(p->prio)) |
2572 | p->sched_class = &fair_sched_class; | |
b29739f9 | 2573 | |
52f17b6c | 2574 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
dd41f596 | 2575 | if (likely(sched_info_on())) |
52f17b6c | 2576 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
1da177e4 | 2577 | #endif |
d6077cb8 | 2578 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4866cde0 NP |
2579 | p->oncpu = 0; |
2580 | #endif | |
1da177e4 | 2581 | #ifdef CONFIG_PREEMPT |
4866cde0 | 2582 | /* Want to start with kernel preemption disabled. */ |
a1261f54 | 2583 | task_thread_info(p)->preempt_count = 1; |
1da177e4 | 2584 | #endif |
476d139c | 2585 | put_cpu(); |
1da177e4 LT |
2586 | } |
2587 | ||
2588 | /* | |
2589 | * wake_up_new_task - wake up a newly created task for the first time. | |
2590 | * | |
2591 | * This function will do some initial scheduler statistics housekeeping | |
2592 | * that must be done for every newly created context, then puts the task | |
2593 | * on the runqueue and wakes it. | |
2594 | */ | |
7ad5b3a5 | 2595 | void wake_up_new_task(struct task_struct *p, unsigned long clone_flags) |
1da177e4 LT |
2596 | { |
2597 | unsigned long flags; | |
dd41f596 | 2598 | struct rq *rq; |
1da177e4 LT |
2599 | |
2600 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 2601 | BUG_ON(p->state != TASK_RUNNING); |
a8e504d2 | 2602 | update_rq_clock(rq); |
1da177e4 LT |
2603 | |
2604 | p->prio = effective_prio(p); | |
2605 | ||
b9dca1e0 | 2606 | if (!p->sched_class->task_new || !current->se.on_rq) { |
dd41f596 | 2607 | activate_task(rq, p, 0); |
1da177e4 | 2608 | } else { |
1da177e4 | 2609 | /* |
dd41f596 IM |
2610 | * Let the scheduling class do new task startup |
2611 | * management (if any): | |
1da177e4 | 2612 | */ |
ee0827d8 | 2613 | p->sched_class->task_new(rq, p); |
18d95a28 | 2614 | inc_nr_running(rq); |
1da177e4 | 2615 | } |
bd3bff9e | 2616 | ftrace_wake_up_new_task(p, rq->curr); |
dd41f596 | 2617 | check_preempt_curr(rq, p); |
9a897c5a SR |
2618 | #ifdef CONFIG_SMP |
2619 | if (p->sched_class->task_wake_up) | |
2620 | p->sched_class->task_wake_up(rq, p); | |
2621 | #endif | |
dd41f596 | 2622 | task_rq_unlock(rq, &flags); |
1da177e4 LT |
2623 | } |
2624 | ||
e107be36 AK |
2625 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2626 | ||
2627 | /** | |
421cee29 RD |
2628 | * preempt_notifier_register - tell me when current is being being preempted & rescheduled |
2629 | * @notifier: notifier struct to register | |
e107be36 AK |
2630 | */ |
2631 | void preempt_notifier_register(struct preempt_notifier *notifier) | |
2632 | { | |
2633 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); | |
2634 | } | |
2635 | EXPORT_SYMBOL_GPL(preempt_notifier_register); | |
2636 | ||
2637 | /** | |
2638 | * preempt_notifier_unregister - no longer interested in preemption notifications | |
421cee29 | 2639 | * @notifier: notifier struct to unregister |
e107be36 AK |
2640 | * |
2641 | * This is safe to call from within a preemption notifier. | |
2642 | */ | |
2643 | void preempt_notifier_unregister(struct preempt_notifier *notifier) | |
2644 | { | |
2645 | hlist_del(¬ifier->link); | |
2646 | } | |
2647 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); | |
2648 | ||
2649 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2650 | { | |
2651 | struct preempt_notifier *notifier; | |
2652 | struct hlist_node *node; | |
2653 | ||
2654 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2655 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); | |
2656 | } | |
2657 | ||
2658 | static void | |
2659 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2660 | struct task_struct *next) | |
2661 | { | |
2662 | struct preempt_notifier *notifier; | |
2663 | struct hlist_node *node; | |
2664 | ||
2665 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2666 | notifier->ops->sched_out(notifier, next); | |
2667 | } | |
2668 | ||
2669 | #else | |
2670 | ||
2671 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2672 | { | |
2673 | } | |
2674 | ||
2675 | static void | |
2676 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2677 | struct task_struct *next) | |
2678 | { | |
2679 | } | |
2680 | ||
2681 | #endif | |
2682 | ||
4866cde0 NP |
2683 | /** |
2684 | * prepare_task_switch - prepare to switch tasks | |
2685 | * @rq: the runqueue preparing to switch | |
421cee29 | 2686 | * @prev: the current task that is being switched out |
4866cde0 NP |
2687 | * @next: the task we are going to switch to. |
2688 | * | |
2689 | * This is called with the rq lock held and interrupts off. It must | |
2690 | * be paired with a subsequent finish_task_switch after the context | |
2691 | * switch. | |
2692 | * | |
2693 | * prepare_task_switch sets up locking and calls architecture specific | |
2694 | * hooks. | |
2695 | */ | |
e107be36 AK |
2696 | static inline void |
2697 | prepare_task_switch(struct rq *rq, struct task_struct *prev, | |
2698 | struct task_struct *next) | |
4866cde0 | 2699 | { |
e107be36 | 2700 | fire_sched_out_preempt_notifiers(prev, next); |
4866cde0 NP |
2701 | prepare_lock_switch(rq, next); |
2702 | prepare_arch_switch(next); | |
2703 | } | |
2704 | ||
1da177e4 LT |
2705 | /** |
2706 | * finish_task_switch - clean up after a task-switch | |
344babaa | 2707 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
2708 | * @prev: the thread we just switched away from. |
2709 | * | |
4866cde0 NP |
2710 | * finish_task_switch must be called after the context switch, paired |
2711 | * with a prepare_task_switch call before the context switch. | |
2712 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
2713 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
2714 | * |
2715 | * Note that we may have delayed dropping an mm in context_switch(). If | |
41a2d6cf | 2716 | * so, we finish that here outside of the runqueue lock. (Doing it |
1da177e4 LT |
2717 | * with the lock held can cause deadlocks; see schedule() for |
2718 | * details.) | |
2719 | */ | |
a9957449 | 2720 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) |
1da177e4 LT |
2721 | __releases(rq->lock) |
2722 | { | |
1da177e4 | 2723 | struct mm_struct *mm = rq->prev_mm; |
55a101f8 | 2724 | long prev_state; |
1da177e4 LT |
2725 | |
2726 | rq->prev_mm = NULL; | |
2727 | ||
2728 | /* | |
2729 | * A task struct has one reference for the use as "current". | |
c394cc9f | 2730 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
55a101f8 ON |
2731 | * schedule one last time. The schedule call will never return, and |
2732 | * the scheduled task must drop that reference. | |
c394cc9f | 2733 | * The test for TASK_DEAD must occur while the runqueue locks are |
1da177e4 LT |
2734 | * still held, otherwise prev could be scheduled on another cpu, die |
2735 | * there before we look at prev->state, and then the reference would | |
2736 | * be dropped twice. | |
2737 | * Manfred Spraul <manfred@colorfullife.com> | |
2738 | */ | |
55a101f8 | 2739 | prev_state = prev->state; |
4866cde0 NP |
2740 | finish_arch_switch(prev); |
2741 | finish_lock_switch(rq, prev); | |
9a897c5a SR |
2742 | #ifdef CONFIG_SMP |
2743 | if (current->sched_class->post_schedule) | |
2744 | current->sched_class->post_schedule(rq); | |
2745 | #endif | |
e8fa1362 | 2746 | |
e107be36 | 2747 | fire_sched_in_preempt_notifiers(current); |
1da177e4 LT |
2748 | if (mm) |
2749 | mmdrop(mm); | |
c394cc9f | 2750 | if (unlikely(prev_state == TASK_DEAD)) { |
c6fd91f0 | 2751 | /* |
2752 | * Remove function-return probe instances associated with this | |
2753 | * task and put them back on the free list. | |
9761eea8 | 2754 | */ |
c6fd91f0 | 2755 | kprobe_flush_task(prev); |
1da177e4 | 2756 | put_task_struct(prev); |
c6fd91f0 | 2757 | } |
1da177e4 LT |
2758 | } |
2759 | ||
2760 | /** | |
2761 | * schedule_tail - first thing a freshly forked thread must call. | |
2762 | * @prev: the thread we just switched away from. | |
2763 | */ | |
36c8b586 | 2764 | asmlinkage void schedule_tail(struct task_struct *prev) |
1da177e4 LT |
2765 | __releases(rq->lock) |
2766 | { | |
70b97a7f IM |
2767 | struct rq *rq = this_rq(); |
2768 | ||
4866cde0 NP |
2769 | finish_task_switch(rq, prev); |
2770 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
2771 | /* In this case, finish_task_switch does not reenable preemption */ | |
2772 | preempt_enable(); | |
2773 | #endif | |
1da177e4 | 2774 | if (current->set_child_tid) |
b488893a | 2775 | put_user(task_pid_vnr(current), current->set_child_tid); |
1da177e4 LT |
2776 | } |
2777 | ||
2778 | /* | |
2779 | * context_switch - switch to the new MM and the new | |
2780 | * thread's register state. | |
2781 | */ | |
dd41f596 | 2782 | static inline void |
70b97a7f | 2783 | context_switch(struct rq *rq, struct task_struct *prev, |
36c8b586 | 2784 | struct task_struct *next) |
1da177e4 | 2785 | { |
dd41f596 | 2786 | struct mm_struct *mm, *oldmm; |
1da177e4 | 2787 | |
e107be36 | 2788 | prepare_task_switch(rq, prev, next); |
bd3bff9e | 2789 | ftrace_ctx_switch(prev, next); |
dd41f596 IM |
2790 | mm = next->mm; |
2791 | oldmm = prev->active_mm; | |
9226d125 ZA |
2792 | /* |
2793 | * For paravirt, this is coupled with an exit in switch_to to | |
2794 | * combine the page table reload and the switch backend into | |
2795 | * one hypercall. | |
2796 | */ | |
2797 | arch_enter_lazy_cpu_mode(); | |
2798 | ||
dd41f596 | 2799 | if (unlikely(!mm)) { |
1da177e4 LT |
2800 | next->active_mm = oldmm; |
2801 | atomic_inc(&oldmm->mm_count); | |
2802 | enter_lazy_tlb(oldmm, next); | |
2803 | } else | |
2804 | switch_mm(oldmm, mm, next); | |
2805 | ||
dd41f596 | 2806 | if (unlikely(!prev->mm)) { |
1da177e4 | 2807 | prev->active_mm = NULL; |
1da177e4 LT |
2808 | rq->prev_mm = oldmm; |
2809 | } | |
3a5f5e48 IM |
2810 | /* |
2811 | * Since the runqueue lock will be released by the next | |
2812 | * task (which is an invalid locking op but in the case | |
2813 | * of the scheduler it's an obvious special-case), so we | |
2814 | * do an early lockdep release here: | |
2815 | */ | |
2816 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
8a25d5de | 2817 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
3a5f5e48 | 2818 | #endif |
1da177e4 LT |
2819 | |
2820 | /* Here we just switch the register state and the stack. */ | |
2821 | switch_to(prev, next, prev); | |
2822 | ||
dd41f596 IM |
2823 | barrier(); |
2824 | /* | |
2825 | * this_rq must be evaluated again because prev may have moved | |
2826 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
2827 | * frame will be invalid. | |
2828 | */ | |
2829 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
2830 | } |
2831 | ||
2832 | /* | |
2833 | * nr_running, nr_uninterruptible and nr_context_switches: | |
2834 | * | |
2835 | * externally visible scheduler statistics: current number of runnable | |
2836 | * threads, current number of uninterruptible-sleeping threads, total | |
2837 | * number of context switches performed since bootup. | |
2838 | */ | |
2839 | unsigned long nr_running(void) | |
2840 | { | |
2841 | unsigned long i, sum = 0; | |
2842 | ||
2843 | for_each_online_cpu(i) | |
2844 | sum += cpu_rq(i)->nr_running; | |
2845 | ||
2846 | return sum; | |
2847 | } | |
2848 | ||
2849 | unsigned long nr_uninterruptible(void) | |
2850 | { | |
2851 | unsigned long i, sum = 0; | |
2852 | ||
0a945022 | 2853 | for_each_possible_cpu(i) |
1da177e4 LT |
2854 | sum += cpu_rq(i)->nr_uninterruptible; |
2855 | ||
2856 | /* | |
2857 | * Since we read the counters lockless, it might be slightly | |
2858 | * inaccurate. Do not allow it to go below zero though: | |
2859 | */ | |
2860 | if (unlikely((long)sum < 0)) | |
2861 | sum = 0; | |
2862 | ||
2863 | return sum; | |
2864 | } | |
2865 | ||
2866 | unsigned long long nr_context_switches(void) | |
2867 | { | |
cc94abfc SR |
2868 | int i; |
2869 | unsigned long long sum = 0; | |
1da177e4 | 2870 | |
0a945022 | 2871 | for_each_possible_cpu(i) |
1da177e4 LT |
2872 | sum += cpu_rq(i)->nr_switches; |
2873 | ||
2874 | return sum; | |
2875 | } | |
2876 | ||
2877 | unsigned long nr_iowait(void) | |
2878 | { | |
2879 | unsigned long i, sum = 0; | |
2880 | ||
0a945022 | 2881 | for_each_possible_cpu(i) |
1da177e4 LT |
2882 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
2883 | ||
2884 | return sum; | |
2885 | } | |
2886 | ||
db1b1fef JS |
2887 | unsigned long nr_active(void) |
2888 | { | |
2889 | unsigned long i, running = 0, uninterruptible = 0; | |
2890 | ||
2891 | for_each_online_cpu(i) { | |
2892 | running += cpu_rq(i)->nr_running; | |
2893 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
2894 | } | |
2895 | ||
2896 | if (unlikely((long)uninterruptible < 0)) | |
2897 | uninterruptible = 0; | |
2898 | ||
2899 | return running + uninterruptible; | |
2900 | } | |
2901 | ||
48f24c4d | 2902 | /* |
dd41f596 IM |
2903 | * Update rq->cpu_load[] statistics. This function is usually called every |
2904 | * scheduler tick (TICK_NSEC). | |
48f24c4d | 2905 | */ |
dd41f596 | 2906 | static void update_cpu_load(struct rq *this_rq) |
48f24c4d | 2907 | { |
495eca49 | 2908 | unsigned long this_load = this_rq->load.weight; |
dd41f596 IM |
2909 | int i, scale; |
2910 | ||
2911 | this_rq->nr_load_updates++; | |
dd41f596 IM |
2912 | |
2913 | /* Update our load: */ | |
2914 | for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
2915 | unsigned long old_load, new_load; | |
2916 | ||
2917 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
2918 | ||
2919 | old_load = this_rq->cpu_load[i]; | |
2920 | new_load = this_load; | |
a25707f3 IM |
2921 | /* |
2922 | * Round up the averaging division if load is increasing. This | |
2923 | * prevents us from getting stuck on 9 if the load is 10, for | |
2924 | * example. | |
2925 | */ | |
2926 | if (new_load > old_load) | |
2927 | new_load += scale-1; | |
dd41f596 IM |
2928 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i; |
2929 | } | |
48f24c4d IM |
2930 | } |
2931 | ||
dd41f596 IM |
2932 | #ifdef CONFIG_SMP |
2933 | ||
1da177e4 LT |
2934 | /* |
2935 | * double_rq_lock - safely lock two runqueues | |
2936 | * | |
2937 | * Note this does not disable interrupts like task_rq_lock, | |
2938 | * you need to do so manually before calling. | |
2939 | */ | |
70b97a7f | 2940 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
2941 | __acquires(rq1->lock) |
2942 | __acquires(rq2->lock) | |
2943 | { | |
054b9108 | 2944 | BUG_ON(!irqs_disabled()); |
1da177e4 LT |
2945 | if (rq1 == rq2) { |
2946 | spin_lock(&rq1->lock); | |
2947 | __acquire(rq2->lock); /* Fake it out ;) */ | |
2948 | } else { | |
c96d145e | 2949 | if (rq1 < rq2) { |
1da177e4 LT |
2950 | spin_lock(&rq1->lock); |
2951 | spin_lock(&rq2->lock); | |
2952 | } else { | |
2953 | spin_lock(&rq2->lock); | |
2954 | spin_lock(&rq1->lock); | |
2955 | } | |
2956 | } | |
6e82a3be IM |
2957 | update_rq_clock(rq1); |
2958 | update_rq_clock(rq2); | |
1da177e4 LT |
2959 | } |
2960 | ||
2961 | /* | |
2962 | * double_rq_unlock - safely unlock two runqueues | |
2963 | * | |
2964 | * Note this does not restore interrupts like task_rq_unlock, | |
2965 | * you need to do so manually after calling. | |
2966 | */ | |
70b97a7f | 2967 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
2968 | __releases(rq1->lock) |
2969 | __releases(rq2->lock) | |
2970 | { | |
2971 | spin_unlock(&rq1->lock); | |
2972 | if (rq1 != rq2) | |
2973 | spin_unlock(&rq2->lock); | |
2974 | else | |
2975 | __release(rq2->lock); | |
2976 | } | |
2977 | ||
2978 | /* | |
2979 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
2980 | */ | |
e8fa1362 | 2981 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest) |
1da177e4 LT |
2982 | __releases(this_rq->lock) |
2983 | __acquires(busiest->lock) | |
2984 | __acquires(this_rq->lock) | |
2985 | { | |
e8fa1362 SR |
2986 | int ret = 0; |
2987 | ||
054b9108 KK |
2988 | if (unlikely(!irqs_disabled())) { |
2989 | /* printk() doesn't work good under rq->lock */ | |
2990 | spin_unlock(&this_rq->lock); | |
2991 | BUG_ON(1); | |
2992 | } | |
1da177e4 | 2993 | if (unlikely(!spin_trylock(&busiest->lock))) { |
c96d145e | 2994 | if (busiest < this_rq) { |
1da177e4 LT |
2995 | spin_unlock(&this_rq->lock); |
2996 | spin_lock(&busiest->lock); | |
2997 | spin_lock(&this_rq->lock); | |
e8fa1362 | 2998 | ret = 1; |
1da177e4 LT |
2999 | } else |
3000 | spin_lock(&busiest->lock); | |
3001 | } | |
e8fa1362 | 3002 | return ret; |
1da177e4 LT |
3003 | } |
3004 | ||
1da177e4 LT |
3005 | /* |
3006 | * If dest_cpu is allowed for this process, migrate the task to it. | |
3007 | * This is accomplished by forcing the cpu_allowed mask to only | |
41a2d6cf | 3008 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then |
1da177e4 LT |
3009 | * the cpu_allowed mask is restored. |
3010 | */ | |
36c8b586 | 3011 | static void sched_migrate_task(struct task_struct *p, int dest_cpu) |
1da177e4 | 3012 | { |
70b97a7f | 3013 | struct migration_req req; |
1da177e4 | 3014 | unsigned long flags; |
70b97a7f | 3015 | struct rq *rq; |
1da177e4 LT |
3016 | |
3017 | rq = task_rq_lock(p, &flags); | |
3018 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
3019 | || unlikely(cpu_is_offline(dest_cpu))) | |
3020 | goto out; | |
3021 | ||
3022 | /* force the process onto the specified CPU */ | |
3023 | if (migrate_task(p, dest_cpu, &req)) { | |
3024 | /* Need to wait for migration thread (might exit: take ref). */ | |
3025 | struct task_struct *mt = rq->migration_thread; | |
36c8b586 | 3026 | |
1da177e4 LT |
3027 | get_task_struct(mt); |
3028 | task_rq_unlock(rq, &flags); | |
3029 | wake_up_process(mt); | |
3030 | put_task_struct(mt); | |
3031 | wait_for_completion(&req.done); | |
36c8b586 | 3032 | |
1da177e4 LT |
3033 | return; |
3034 | } | |
3035 | out: | |
3036 | task_rq_unlock(rq, &flags); | |
3037 | } | |
3038 | ||
3039 | /* | |
476d139c NP |
3040 | * sched_exec - execve() is a valuable balancing opportunity, because at |
3041 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
3042 | */ |
3043 | void sched_exec(void) | |
3044 | { | |
1da177e4 | 3045 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 3046 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 3047 | put_cpu(); |
476d139c NP |
3048 | if (new_cpu != this_cpu) |
3049 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
3050 | } |
3051 | ||
3052 | /* | |
3053 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
3054 | * Both runqueues must be locked. | |
3055 | */ | |
dd41f596 IM |
3056 | static void pull_task(struct rq *src_rq, struct task_struct *p, |
3057 | struct rq *this_rq, int this_cpu) | |
1da177e4 | 3058 | { |
2e1cb74a | 3059 | deactivate_task(src_rq, p, 0); |
1da177e4 | 3060 | set_task_cpu(p, this_cpu); |
dd41f596 | 3061 | activate_task(this_rq, p, 0); |
1da177e4 LT |
3062 | /* |
3063 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
3064 | * to be always true for them. | |
3065 | */ | |
dd41f596 | 3066 | check_preempt_curr(this_rq, p); |
1da177e4 LT |
3067 | } |
3068 | ||
3069 | /* | |
3070 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3071 | */ | |
858119e1 | 3072 | static |
70b97a7f | 3073 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, |
d15bcfdb | 3074 | struct sched_domain *sd, enum cpu_idle_type idle, |
95cdf3b7 | 3075 | int *all_pinned) |
1da177e4 LT |
3076 | { |
3077 | /* | |
3078 | * We do not migrate tasks that are: | |
3079 | * 1) running (obviously), or | |
3080 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
3081 | * 3) are cache-hot on their current CPU. | |
3082 | */ | |
cc367732 IM |
3083 | if (!cpu_isset(this_cpu, p->cpus_allowed)) { |
3084 | schedstat_inc(p, se.nr_failed_migrations_affine); | |
1da177e4 | 3085 | return 0; |
cc367732 | 3086 | } |
81026794 NP |
3087 | *all_pinned = 0; |
3088 | ||
cc367732 IM |
3089 | if (task_running(rq, p)) { |
3090 | schedstat_inc(p, se.nr_failed_migrations_running); | |
81026794 | 3091 | return 0; |
cc367732 | 3092 | } |
1da177e4 | 3093 | |
da84d961 IM |
3094 | /* |
3095 | * Aggressive migration if: | |
3096 | * 1) task is cache cold, or | |
3097 | * 2) too many balance attempts have failed. | |
3098 | */ | |
3099 | ||
6bc1665b IM |
3100 | if (!task_hot(p, rq->clock, sd) || |
3101 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
da84d961 | 3102 | #ifdef CONFIG_SCHEDSTATS |
cc367732 | 3103 | if (task_hot(p, rq->clock, sd)) { |
da84d961 | 3104 | schedstat_inc(sd, lb_hot_gained[idle]); |
cc367732 IM |
3105 | schedstat_inc(p, se.nr_forced_migrations); |
3106 | } | |
da84d961 IM |
3107 | #endif |
3108 | return 1; | |
3109 | } | |
3110 | ||
cc367732 IM |
3111 | if (task_hot(p, rq->clock, sd)) { |
3112 | schedstat_inc(p, se.nr_failed_migrations_hot); | |
da84d961 | 3113 | return 0; |
cc367732 | 3114 | } |
1da177e4 LT |
3115 | return 1; |
3116 | } | |
3117 | ||
e1d1484f PW |
3118 | static unsigned long |
3119 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3120 | unsigned long max_load_move, struct sched_domain *sd, | |
3121 | enum cpu_idle_type idle, int *all_pinned, | |
3122 | int *this_best_prio, struct rq_iterator *iterator) | |
1da177e4 | 3123 | { |
b82d9fdd | 3124 | int loops = 0, pulled = 0, pinned = 0, skip_for_load; |
dd41f596 IM |
3125 | struct task_struct *p; |
3126 | long rem_load_move = max_load_move; | |
1da177e4 | 3127 | |
e1d1484f | 3128 | if (max_load_move == 0) |
1da177e4 LT |
3129 | goto out; |
3130 | ||
81026794 NP |
3131 | pinned = 1; |
3132 | ||
1da177e4 | 3133 | /* |
dd41f596 | 3134 | * Start the load-balancing iterator: |
1da177e4 | 3135 | */ |
dd41f596 IM |
3136 | p = iterator->start(iterator->arg); |
3137 | next: | |
b82d9fdd | 3138 | if (!p || loops++ > sysctl_sched_nr_migrate) |
1da177e4 | 3139 | goto out; |
50ddd969 | 3140 | /* |
b82d9fdd | 3141 | * To help distribute high priority tasks across CPUs we don't |
50ddd969 PW |
3142 | * skip a task if it will be the highest priority task (i.e. smallest |
3143 | * prio value) on its new queue regardless of its load weight | |
3144 | */ | |
dd41f596 IM |
3145 | skip_for_load = (p->se.load.weight >> 1) > rem_load_move + |
3146 | SCHED_LOAD_SCALE_FUZZ; | |
a4ac01c3 | 3147 | if ((skip_for_load && p->prio >= *this_best_prio) || |
dd41f596 | 3148 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { |
dd41f596 IM |
3149 | p = iterator->next(iterator->arg); |
3150 | goto next; | |
1da177e4 LT |
3151 | } |
3152 | ||
dd41f596 | 3153 | pull_task(busiest, p, this_rq, this_cpu); |
1da177e4 | 3154 | pulled++; |
dd41f596 | 3155 | rem_load_move -= p->se.load.weight; |
1da177e4 | 3156 | |
2dd73a4f | 3157 | /* |
b82d9fdd | 3158 | * We only want to steal up to the prescribed amount of weighted load. |
2dd73a4f | 3159 | */ |
e1d1484f | 3160 | if (rem_load_move > 0) { |
a4ac01c3 PW |
3161 | if (p->prio < *this_best_prio) |
3162 | *this_best_prio = p->prio; | |
dd41f596 IM |
3163 | p = iterator->next(iterator->arg); |
3164 | goto next; | |
1da177e4 LT |
3165 | } |
3166 | out: | |
3167 | /* | |
e1d1484f | 3168 | * Right now, this is one of only two places pull_task() is called, |
1da177e4 LT |
3169 | * so we can safely collect pull_task() stats here rather than |
3170 | * inside pull_task(). | |
3171 | */ | |
3172 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
3173 | |
3174 | if (all_pinned) | |
3175 | *all_pinned = pinned; | |
e1d1484f PW |
3176 | |
3177 | return max_load_move - rem_load_move; | |
1da177e4 LT |
3178 | } |
3179 | ||
dd41f596 | 3180 | /* |
43010659 PW |
3181 | * move_tasks tries to move up to max_load_move weighted load from busiest to |
3182 | * this_rq, as part of a balancing operation within domain "sd". | |
3183 | * Returns 1 if successful and 0 otherwise. | |
dd41f596 IM |
3184 | * |
3185 | * Called with both runqueues locked. | |
3186 | */ | |
3187 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
43010659 | 3188 | unsigned long max_load_move, |
dd41f596 IM |
3189 | struct sched_domain *sd, enum cpu_idle_type idle, |
3190 | int *all_pinned) | |
3191 | { | |
5522d5d5 | 3192 | const struct sched_class *class = sched_class_highest; |
43010659 | 3193 | unsigned long total_load_moved = 0; |
a4ac01c3 | 3194 | int this_best_prio = this_rq->curr->prio; |
dd41f596 IM |
3195 | |
3196 | do { | |
43010659 PW |
3197 | total_load_moved += |
3198 | class->load_balance(this_rq, this_cpu, busiest, | |
e1d1484f | 3199 | max_load_move - total_load_moved, |
a4ac01c3 | 3200 | sd, idle, all_pinned, &this_best_prio); |
dd41f596 | 3201 | class = class->next; |
43010659 | 3202 | } while (class && max_load_move > total_load_moved); |
dd41f596 | 3203 | |
43010659 PW |
3204 | return total_load_moved > 0; |
3205 | } | |
3206 | ||
e1d1484f PW |
3207 | static int |
3208 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3209 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3210 | struct rq_iterator *iterator) | |
3211 | { | |
3212 | struct task_struct *p = iterator->start(iterator->arg); | |
3213 | int pinned = 0; | |
3214 | ||
3215 | while (p) { | |
3216 | if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { | |
3217 | pull_task(busiest, p, this_rq, this_cpu); | |
3218 | /* | |
3219 | * Right now, this is only the second place pull_task() | |
3220 | * is called, so we can safely collect pull_task() | |
3221 | * stats here rather than inside pull_task(). | |
3222 | */ | |
3223 | schedstat_inc(sd, lb_gained[idle]); | |
3224 | ||
3225 | return 1; | |
3226 | } | |
3227 | p = iterator->next(iterator->arg); | |
3228 | } | |
3229 | ||
3230 | return 0; | |
3231 | } | |
3232 | ||
43010659 PW |
3233 | /* |
3234 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3235 | * part of active balancing operations within "domain". | |
3236 | * Returns 1 if successful and 0 otherwise. | |
3237 | * | |
3238 | * Called with both runqueues locked. | |
3239 | */ | |
3240 | static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3241 | struct sched_domain *sd, enum cpu_idle_type idle) | |
3242 | { | |
5522d5d5 | 3243 | const struct sched_class *class; |
43010659 PW |
3244 | |
3245 | for (class = sched_class_highest; class; class = class->next) | |
e1d1484f | 3246 | if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle)) |
43010659 PW |
3247 | return 1; |
3248 | ||
3249 | return 0; | |
dd41f596 IM |
3250 | } |
3251 | ||
1da177e4 LT |
3252 | /* |
3253 | * find_busiest_group finds and returns the busiest CPU group within the | |
48f24c4d IM |
3254 | * domain. It calculates and returns the amount of weighted load which |
3255 | * should be moved to restore balance via the imbalance parameter. | |
1da177e4 LT |
3256 | */ |
3257 | static struct sched_group * | |
3258 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
dd41f596 | 3259 | unsigned long *imbalance, enum cpu_idle_type idle, |
7c16ec58 | 3260 | int *sd_idle, const cpumask_t *cpus, int *balance) |
1da177e4 LT |
3261 | { |
3262 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
3263 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 3264 | unsigned long max_pull; |
2dd73a4f PW |
3265 | unsigned long busiest_load_per_task, busiest_nr_running; |
3266 | unsigned long this_load_per_task, this_nr_running; | |
908a7c1b | 3267 | int load_idx, group_imb = 0; |
5c45bf27 SS |
3268 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
3269 | int power_savings_balance = 1; | |
3270 | unsigned long leader_nr_running = 0, min_load_per_task = 0; | |
3271 | unsigned long min_nr_running = ULONG_MAX; | |
3272 | struct sched_group *group_min = NULL, *group_leader = NULL; | |
3273 | #endif | |
1da177e4 LT |
3274 | |
3275 | max_load = this_load = total_load = total_pwr = 0; | |
2dd73a4f PW |
3276 | busiest_load_per_task = busiest_nr_running = 0; |
3277 | this_load_per_task = this_nr_running = 0; | |
d15bcfdb | 3278 | if (idle == CPU_NOT_IDLE) |
7897986b | 3279 | load_idx = sd->busy_idx; |
d15bcfdb | 3280 | else if (idle == CPU_NEWLY_IDLE) |
7897986b NP |
3281 | load_idx = sd->newidle_idx; |
3282 | else | |
3283 | load_idx = sd->idle_idx; | |
1da177e4 LT |
3284 | |
3285 | do { | |
908a7c1b | 3286 | unsigned long load, group_capacity, max_cpu_load, min_cpu_load; |
1da177e4 LT |
3287 | int local_group; |
3288 | int i; | |
908a7c1b | 3289 | int __group_imb = 0; |
783609c6 | 3290 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
2dd73a4f | 3291 | unsigned long sum_nr_running, sum_weighted_load; |
1da177e4 LT |
3292 | |
3293 | local_group = cpu_isset(this_cpu, group->cpumask); | |
3294 | ||
783609c6 SS |
3295 | if (local_group) |
3296 | balance_cpu = first_cpu(group->cpumask); | |
3297 | ||
1da177e4 | 3298 | /* Tally up the load of all CPUs in the group */ |
2dd73a4f | 3299 | sum_weighted_load = sum_nr_running = avg_load = 0; |
908a7c1b KC |
3300 | max_cpu_load = 0; |
3301 | min_cpu_load = ~0UL; | |
1da177e4 LT |
3302 | |
3303 | for_each_cpu_mask(i, group->cpumask) { | |
0a2966b4 CL |
3304 | struct rq *rq; |
3305 | ||
3306 | if (!cpu_isset(i, *cpus)) | |
3307 | continue; | |
3308 | ||
3309 | rq = cpu_rq(i); | |
2dd73a4f | 3310 | |
9439aab8 | 3311 | if (*sd_idle && rq->nr_running) |
5969fe06 NP |
3312 | *sd_idle = 0; |
3313 | ||
1da177e4 | 3314 | /* Bias balancing toward cpus of our domain */ |
783609c6 SS |
3315 | if (local_group) { |
3316 | if (idle_cpu(i) && !first_idle_cpu) { | |
3317 | first_idle_cpu = 1; | |
3318 | balance_cpu = i; | |
3319 | } | |
3320 | ||
a2000572 | 3321 | load = target_load(i, load_idx); |
908a7c1b | 3322 | } else { |
a2000572 | 3323 | load = source_load(i, load_idx); |
908a7c1b KC |
3324 | if (load > max_cpu_load) |
3325 | max_cpu_load = load; | |
3326 | if (min_cpu_load > load) | |
3327 | min_cpu_load = load; | |
3328 | } | |
1da177e4 LT |
3329 | |
3330 | avg_load += load; | |
2dd73a4f | 3331 | sum_nr_running += rq->nr_running; |
dd41f596 | 3332 | sum_weighted_load += weighted_cpuload(i); |
1da177e4 LT |
3333 | } |
3334 | ||
783609c6 SS |
3335 | /* |
3336 | * First idle cpu or the first cpu(busiest) in this sched group | |
3337 | * is eligible for doing load balancing at this and above | |
9439aab8 SS |
3338 | * domains. In the newly idle case, we will allow all the cpu's |
3339 | * to do the newly idle load balance. | |
783609c6 | 3340 | */ |
9439aab8 SS |
3341 | if (idle != CPU_NEWLY_IDLE && local_group && |
3342 | balance_cpu != this_cpu && balance) { | |
783609c6 SS |
3343 | *balance = 0; |
3344 | goto ret; | |
3345 | } | |
3346 | ||
1da177e4 | 3347 | total_load += avg_load; |
5517d86b | 3348 | total_pwr += group->__cpu_power; |
1da177e4 LT |
3349 | |
3350 | /* Adjust by relative CPU power of the group */ | |
5517d86b ED |
3351 | avg_load = sg_div_cpu_power(group, |
3352 | avg_load * SCHED_LOAD_SCALE); | |
1da177e4 | 3353 | |
908a7c1b KC |
3354 | if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE) |
3355 | __group_imb = 1; | |
3356 | ||
5517d86b | 3357 | group_capacity = group->__cpu_power / SCHED_LOAD_SCALE; |
5c45bf27 | 3358 | |
1da177e4 LT |
3359 | if (local_group) { |
3360 | this_load = avg_load; | |
3361 | this = group; | |
2dd73a4f PW |
3362 | this_nr_running = sum_nr_running; |
3363 | this_load_per_task = sum_weighted_load; | |
3364 | } else if (avg_load > max_load && | |
908a7c1b | 3365 | (sum_nr_running > group_capacity || __group_imb)) { |
1da177e4 LT |
3366 | max_load = avg_load; |
3367 | busiest = group; | |
2dd73a4f PW |
3368 | busiest_nr_running = sum_nr_running; |
3369 | busiest_load_per_task = sum_weighted_load; | |
908a7c1b | 3370 | group_imb = __group_imb; |
1da177e4 | 3371 | } |
5c45bf27 SS |
3372 | |
3373 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3374 | /* | |
3375 | * Busy processors will not participate in power savings | |
3376 | * balance. | |
3377 | */ | |
dd41f596 IM |
3378 | if (idle == CPU_NOT_IDLE || |
3379 | !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
3380 | goto group_next; | |
5c45bf27 SS |
3381 | |
3382 | /* | |
3383 | * If the local group is idle or completely loaded | |
3384 | * no need to do power savings balance at this domain | |
3385 | */ | |
3386 | if (local_group && (this_nr_running >= group_capacity || | |
3387 | !this_nr_running)) | |
3388 | power_savings_balance = 0; | |
3389 | ||
dd41f596 | 3390 | /* |
5c45bf27 SS |
3391 | * If a group is already running at full capacity or idle, |
3392 | * don't include that group in power savings calculations | |
dd41f596 IM |
3393 | */ |
3394 | if (!power_savings_balance || sum_nr_running >= group_capacity | |
5c45bf27 | 3395 | || !sum_nr_running) |
dd41f596 | 3396 | goto group_next; |
5c45bf27 | 3397 | |
dd41f596 | 3398 | /* |
5c45bf27 | 3399 | * Calculate the group which has the least non-idle load. |
dd41f596 IM |
3400 | * This is the group from where we need to pick up the load |
3401 | * for saving power | |
3402 | */ | |
3403 | if ((sum_nr_running < min_nr_running) || | |
3404 | (sum_nr_running == min_nr_running && | |
5c45bf27 SS |
3405 | first_cpu(group->cpumask) < |
3406 | first_cpu(group_min->cpumask))) { | |
dd41f596 IM |
3407 | group_min = group; |
3408 | min_nr_running = sum_nr_running; | |
5c45bf27 SS |
3409 | min_load_per_task = sum_weighted_load / |
3410 | sum_nr_running; | |
dd41f596 | 3411 | } |
5c45bf27 | 3412 | |
dd41f596 | 3413 | /* |
5c45bf27 | 3414 | * Calculate the group which is almost near its |
dd41f596 IM |
3415 | * capacity but still has some space to pick up some load |
3416 | * from other group and save more power | |
3417 | */ | |
3418 | if (sum_nr_running <= group_capacity - 1) { | |
3419 | if (sum_nr_running > leader_nr_running || | |
3420 | (sum_nr_running == leader_nr_running && | |
3421 | first_cpu(group->cpumask) > | |
3422 | first_cpu(group_leader->cpumask))) { | |
3423 | group_leader = group; | |
3424 | leader_nr_running = sum_nr_running; | |
3425 | } | |
48f24c4d | 3426 | } |
5c45bf27 SS |
3427 | group_next: |
3428 | #endif | |
1da177e4 LT |
3429 | group = group->next; |
3430 | } while (group != sd->groups); | |
3431 | ||
2dd73a4f | 3432 | if (!busiest || this_load >= max_load || busiest_nr_running == 0) |
1da177e4 LT |
3433 | goto out_balanced; |
3434 | ||
3435 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
3436 | ||
3437 | if (this_load >= avg_load || | |
3438 | 100*max_load <= sd->imbalance_pct*this_load) | |
3439 | goto out_balanced; | |
3440 | ||
2dd73a4f | 3441 | busiest_load_per_task /= busiest_nr_running; |
908a7c1b KC |
3442 | if (group_imb) |
3443 | busiest_load_per_task = min(busiest_load_per_task, avg_load); | |
3444 | ||
1da177e4 LT |
3445 | /* |
3446 | * We're trying to get all the cpus to the average_load, so we don't | |
3447 | * want to push ourselves above the average load, nor do we wish to | |
3448 | * reduce the max loaded cpu below the average load, as either of these | |
3449 | * actions would just result in more rebalancing later, and ping-pong | |
3450 | * tasks around. Thus we look for the minimum possible imbalance. | |
3451 | * Negative imbalances (*we* are more loaded than anyone else) will | |
3452 | * be counted as no imbalance for these purposes -- we can't fix that | |
41a2d6cf | 3453 | * by pulling tasks to us. Be careful of negative numbers as they'll |
1da177e4 LT |
3454 | * appear as very large values with unsigned longs. |
3455 | */ | |
2dd73a4f PW |
3456 | if (max_load <= busiest_load_per_task) |
3457 | goto out_balanced; | |
3458 | ||
3459 | /* | |
3460 | * In the presence of smp nice balancing, certain scenarios can have | |
3461 | * max load less than avg load(as we skip the groups at or below | |
3462 | * its cpu_power, while calculating max_load..) | |
3463 | */ | |
3464 | if (max_load < avg_load) { | |
3465 | *imbalance = 0; | |
3466 | goto small_imbalance; | |
3467 | } | |
0c117f1b SS |
3468 | |
3469 | /* Don't want to pull so many tasks that a group would go idle */ | |
2dd73a4f | 3470 | max_pull = min(max_load - avg_load, max_load - busiest_load_per_task); |
0c117f1b | 3471 | |
1da177e4 | 3472 | /* How much load to actually move to equalise the imbalance */ |
5517d86b ED |
3473 | *imbalance = min(max_pull * busiest->__cpu_power, |
3474 | (avg_load - this_load) * this->__cpu_power) | |
1da177e4 LT |
3475 | / SCHED_LOAD_SCALE; |
3476 | ||
2dd73a4f PW |
3477 | /* |
3478 | * if *imbalance is less than the average load per runnable task | |
3479 | * there is no gaurantee that any tasks will be moved so we'll have | |
3480 | * a think about bumping its value to force at least one task to be | |
3481 | * moved | |
3482 | */ | |
7fd0d2dd | 3483 | if (*imbalance < busiest_load_per_task) { |
48f24c4d | 3484 | unsigned long tmp, pwr_now, pwr_move; |
2dd73a4f PW |
3485 | unsigned int imbn; |
3486 | ||
3487 | small_imbalance: | |
3488 | pwr_move = pwr_now = 0; | |
3489 | imbn = 2; | |
3490 | if (this_nr_running) { | |
3491 | this_load_per_task /= this_nr_running; | |
3492 | if (busiest_load_per_task > this_load_per_task) | |
3493 | imbn = 1; | |
3494 | } else | |
3495 | this_load_per_task = SCHED_LOAD_SCALE; | |
1da177e4 | 3496 | |
dd41f596 IM |
3497 | if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >= |
3498 | busiest_load_per_task * imbn) { | |
2dd73a4f | 3499 | *imbalance = busiest_load_per_task; |
1da177e4 LT |
3500 | return busiest; |
3501 | } | |
3502 | ||
3503 | /* | |
3504 | * OK, we don't have enough imbalance to justify moving tasks, | |
3505 | * however we may be able to increase total CPU power used by | |
3506 | * moving them. | |
3507 | */ | |
3508 | ||
5517d86b ED |
3509 | pwr_now += busiest->__cpu_power * |
3510 | min(busiest_load_per_task, max_load); | |
3511 | pwr_now += this->__cpu_power * | |
3512 | min(this_load_per_task, this_load); | |
1da177e4 LT |
3513 | pwr_now /= SCHED_LOAD_SCALE; |
3514 | ||
3515 | /* Amount of load we'd subtract */ | |
5517d86b ED |
3516 | tmp = sg_div_cpu_power(busiest, |
3517 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
1da177e4 | 3518 | if (max_load > tmp) |
5517d86b | 3519 | pwr_move += busiest->__cpu_power * |
2dd73a4f | 3520 | min(busiest_load_per_task, max_load - tmp); |
1da177e4 LT |
3521 | |
3522 | /* Amount of load we'd add */ | |
5517d86b | 3523 | if (max_load * busiest->__cpu_power < |
33859f7f | 3524 | busiest_load_per_task * SCHED_LOAD_SCALE) |
5517d86b ED |
3525 | tmp = sg_div_cpu_power(this, |
3526 | max_load * busiest->__cpu_power); | |
1da177e4 | 3527 | else |
5517d86b ED |
3528 | tmp = sg_div_cpu_power(this, |
3529 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
3530 | pwr_move += this->__cpu_power * | |
3531 | min(this_load_per_task, this_load + tmp); | |
1da177e4 LT |
3532 | pwr_move /= SCHED_LOAD_SCALE; |
3533 | ||
3534 | /* Move if we gain throughput */ | |
7fd0d2dd SS |
3535 | if (pwr_move > pwr_now) |
3536 | *imbalance = busiest_load_per_task; | |
1da177e4 LT |
3537 | } |
3538 | ||
1da177e4 LT |
3539 | return busiest; |
3540 | ||
3541 | out_balanced: | |
5c45bf27 | 3542 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
d15bcfdb | 3543 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) |
5c45bf27 | 3544 | goto ret; |
1da177e4 | 3545 | |
5c45bf27 SS |
3546 | if (this == group_leader && group_leader != group_min) { |
3547 | *imbalance = min_load_per_task; | |
3548 | return group_min; | |
3549 | } | |
5c45bf27 | 3550 | #endif |
783609c6 | 3551 | ret: |
1da177e4 LT |
3552 | *imbalance = 0; |
3553 | return NULL; | |
3554 | } | |
3555 | ||
3556 | /* | |
3557 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3558 | */ | |
70b97a7f | 3559 | static struct rq * |
d15bcfdb | 3560 | find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle, |
7c16ec58 | 3561 | unsigned long imbalance, const cpumask_t *cpus) |
1da177e4 | 3562 | { |
70b97a7f | 3563 | struct rq *busiest = NULL, *rq; |
2dd73a4f | 3564 | unsigned long max_load = 0; |
1da177e4 LT |
3565 | int i; |
3566 | ||
3567 | for_each_cpu_mask(i, group->cpumask) { | |
dd41f596 | 3568 | unsigned long wl; |
0a2966b4 CL |
3569 | |
3570 | if (!cpu_isset(i, *cpus)) | |
3571 | continue; | |
3572 | ||
48f24c4d | 3573 | rq = cpu_rq(i); |
dd41f596 | 3574 | wl = weighted_cpuload(i); |
2dd73a4f | 3575 | |
dd41f596 | 3576 | if (rq->nr_running == 1 && wl > imbalance) |
2dd73a4f | 3577 | continue; |
1da177e4 | 3578 | |
dd41f596 IM |
3579 | if (wl > max_load) { |
3580 | max_load = wl; | |
48f24c4d | 3581 | busiest = rq; |
1da177e4 LT |
3582 | } |
3583 | } | |
3584 | ||
3585 | return busiest; | |
3586 | } | |
3587 | ||
77391d71 NP |
3588 | /* |
3589 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3590 | * so long as it is large enough. | |
3591 | */ | |
3592 | #define MAX_PINNED_INTERVAL 512 | |
3593 | ||
1da177e4 LT |
3594 | /* |
3595 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3596 | * tasks if there is an imbalance. | |
1da177e4 | 3597 | */ |
70b97a7f | 3598 | static int load_balance(int this_cpu, struct rq *this_rq, |
d15bcfdb | 3599 | struct sched_domain *sd, enum cpu_idle_type idle, |
7c16ec58 | 3600 | int *balance, cpumask_t *cpus) |
1da177e4 | 3601 | { |
43010659 | 3602 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; |
1da177e4 | 3603 | struct sched_group *group; |
1da177e4 | 3604 | unsigned long imbalance; |
70b97a7f | 3605 | struct rq *busiest; |
fe2eea3f | 3606 | unsigned long flags; |
18d95a28 | 3607 | int unlock_aggregate; |
5969fe06 | 3608 | |
7c16ec58 MT |
3609 | cpus_setall(*cpus); |
3610 | ||
18d95a28 PZ |
3611 | unlock_aggregate = get_aggregate(sd); |
3612 | ||
89c4710e SS |
3613 | /* |
3614 | * When power savings policy is enabled for the parent domain, idle | |
3615 | * sibling can pick up load irrespective of busy siblings. In this case, | |
dd41f596 | 3616 | * let the state of idle sibling percolate up as CPU_IDLE, instead of |
d15bcfdb | 3617 | * portraying it as CPU_NOT_IDLE. |
89c4710e | 3618 | */ |
d15bcfdb | 3619 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3620 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 3621 | sd_idle = 1; |
1da177e4 | 3622 | |
2d72376b | 3623 | schedstat_inc(sd, lb_count[idle]); |
1da177e4 | 3624 | |
0a2966b4 CL |
3625 | redo: |
3626 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
7c16ec58 | 3627 | cpus, balance); |
783609c6 | 3628 | |
06066714 | 3629 | if (*balance == 0) |
783609c6 | 3630 | goto out_balanced; |
783609c6 | 3631 | |
1da177e4 LT |
3632 | if (!group) { |
3633 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3634 | goto out_balanced; | |
3635 | } | |
3636 | ||
7c16ec58 | 3637 | busiest = find_busiest_queue(group, idle, imbalance, cpus); |
1da177e4 LT |
3638 | if (!busiest) { |
3639 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3640 | goto out_balanced; | |
3641 | } | |
3642 | ||
db935dbd | 3643 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
3644 | |
3645 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3646 | ||
43010659 | 3647 | ld_moved = 0; |
1da177e4 LT |
3648 | if (busiest->nr_running > 1) { |
3649 | /* | |
3650 | * Attempt to move tasks. If find_busiest_group has found | |
3651 | * an imbalance but busiest->nr_running <= 1, the group is | |
43010659 | 3652 | * still unbalanced. ld_moved simply stays zero, so it is |
1da177e4 LT |
3653 | * correctly treated as an imbalance. |
3654 | */ | |
fe2eea3f | 3655 | local_irq_save(flags); |
e17224bf | 3656 | double_rq_lock(this_rq, busiest); |
43010659 | 3657 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
48f24c4d | 3658 | imbalance, sd, idle, &all_pinned); |
e17224bf | 3659 | double_rq_unlock(this_rq, busiest); |
fe2eea3f | 3660 | local_irq_restore(flags); |
81026794 | 3661 | |
46cb4b7c SS |
3662 | /* |
3663 | * some other cpu did the load balance for us. | |
3664 | */ | |
43010659 | 3665 | if (ld_moved && this_cpu != smp_processor_id()) |
46cb4b7c SS |
3666 | resched_cpu(this_cpu); |
3667 | ||
81026794 | 3668 | /* All tasks on this runqueue were pinned by CPU affinity */ |
0a2966b4 | 3669 | if (unlikely(all_pinned)) { |
7c16ec58 MT |
3670 | cpu_clear(cpu_of(busiest), *cpus); |
3671 | if (!cpus_empty(*cpus)) | |
0a2966b4 | 3672 | goto redo; |
81026794 | 3673 | goto out_balanced; |
0a2966b4 | 3674 | } |
1da177e4 | 3675 | } |
81026794 | 3676 | |
43010659 | 3677 | if (!ld_moved) { |
1da177e4 LT |
3678 | schedstat_inc(sd, lb_failed[idle]); |
3679 | sd->nr_balance_failed++; | |
3680 | ||
3681 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 | 3682 | |
fe2eea3f | 3683 | spin_lock_irqsave(&busiest->lock, flags); |
fa3b6ddc SS |
3684 | |
3685 | /* don't kick the migration_thread, if the curr | |
3686 | * task on busiest cpu can't be moved to this_cpu | |
3687 | */ | |
3688 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
fe2eea3f | 3689 | spin_unlock_irqrestore(&busiest->lock, flags); |
fa3b6ddc SS |
3690 | all_pinned = 1; |
3691 | goto out_one_pinned; | |
3692 | } | |
3693 | ||
1da177e4 LT |
3694 | if (!busiest->active_balance) { |
3695 | busiest->active_balance = 1; | |
3696 | busiest->push_cpu = this_cpu; | |
81026794 | 3697 | active_balance = 1; |
1da177e4 | 3698 | } |
fe2eea3f | 3699 | spin_unlock_irqrestore(&busiest->lock, flags); |
81026794 | 3700 | if (active_balance) |
1da177e4 LT |
3701 | wake_up_process(busiest->migration_thread); |
3702 | ||
3703 | /* | |
3704 | * We've kicked active balancing, reset the failure | |
3705 | * counter. | |
3706 | */ | |
39507451 | 3707 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 3708 | } |
81026794 | 3709 | } else |
1da177e4 LT |
3710 | sd->nr_balance_failed = 0; |
3711 | ||
81026794 | 3712 | if (likely(!active_balance)) { |
1da177e4 LT |
3713 | /* We were unbalanced, so reset the balancing interval */ |
3714 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
3715 | } else { |
3716 | /* | |
3717 | * If we've begun active balancing, start to back off. This | |
3718 | * case may not be covered by the all_pinned logic if there | |
3719 | * is only 1 task on the busy runqueue (because we don't call | |
3720 | * move_tasks). | |
3721 | */ | |
3722 | if (sd->balance_interval < sd->max_interval) | |
3723 | sd->balance_interval *= 2; | |
1da177e4 LT |
3724 | } |
3725 | ||
43010659 | 3726 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3727 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
18d95a28 PZ |
3728 | ld_moved = -1; |
3729 | ||
3730 | goto out; | |
1da177e4 LT |
3731 | |
3732 | out_balanced: | |
1da177e4 LT |
3733 | schedstat_inc(sd, lb_balanced[idle]); |
3734 | ||
16cfb1c0 | 3735 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
3736 | |
3737 | out_one_pinned: | |
1da177e4 | 3738 | /* tune up the balancing interval */ |
77391d71 NP |
3739 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
3740 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
3741 | sd->balance_interval *= 2; |
3742 | ||
48f24c4d | 3743 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3744 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
18d95a28 PZ |
3745 | ld_moved = -1; |
3746 | else | |
3747 | ld_moved = 0; | |
3748 | out: | |
3749 | if (unlock_aggregate) | |
3750 | put_aggregate(sd); | |
3751 | return ld_moved; | |
1da177e4 LT |
3752 | } |
3753 | ||
3754 | /* | |
3755 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3756 | * tasks if there is an imbalance. | |
3757 | * | |
d15bcfdb | 3758 | * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE). |
1da177e4 LT |
3759 | * this_rq is locked. |
3760 | */ | |
48f24c4d | 3761 | static int |
7c16ec58 MT |
3762 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd, |
3763 | cpumask_t *cpus) | |
1da177e4 LT |
3764 | { |
3765 | struct sched_group *group; | |
70b97a7f | 3766 | struct rq *busiest = NULL; |
1da177e4 | 3767 | unsigned long imbalance; |
43010659 | 3768 | int ld_moved = 0; |
5969fe06 | 3769 | int sd_idle = 0; |
969bb4e4 | 3770 | int all_pinned = 0; |
7c16ec58 MT |
3771 | |
3772 | cpus_setall(*cpus); | |
5969fe06 | 3773 | |
89c4710e SS |
3774 | /* |
3775 | * When power savings policy is enabled for the parent domain, idle | |
3776 | * sibling can pick up load irrespective of busy siblings. In this case, | |
3777 | * let the state of idle sibling percolate up as IDLE, instead of | |
d15bcfdb | 3778 | * portraying it as CPU_NOT_IDLE. |
89c4710e SS |
3779 | */ |
3780 | if (sd->flags & SD_SHARE_CPUPOWER && | |
3781 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 | 3782 | sd_idle = 1; |
1da177e4 | 3783 | |
2d72376b | 3784 | schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]); |
0a2966b4 | 3785 | redo: |
d15bcfdb | 3786 | group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE, |
7c16ec58 | 3787 | &sd_idle, cpus, NULL); |
1da177e4 | 3788 | if (!group) { |
d15bcfdb | 3789 | schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]); |
16cfb1c0 | 3790 | goto out_balanced; |
1da177e4 LT |
3791 | } |
3792 | ||
7c16ec58 | 3793 | busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus); |
db935dbd | 3794 | if (!busiest) { |
d15bcfdb | 3795 | schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]); |
16cfb1c0 | 3796 | goto out_balanced; |
1da177e4 LT |
3797 | } |
3798 | ||
db935dbd NP |
3799 | BUG_ON(busiest == this_rq); |
3800 | ||
d15bcfdb | 3801 | schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance); |
d6d5cfaf | 3802 | |
43010659 | 3803 | ld_moved = 0; |
d6d5cfaf NP |
3804 | if (busiest->nr_running > 1) { |
3805 | /* Attempt to move tasks */ | |
3806 | double_lock_balance(this_rq, busiest); | |
6e82a3be IM |
3807 | /* this_rq->clock is already updated */ |
3808 | update_rq_clock(busiest); | |
43010659 | 3809 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
969bb4e4 SS |
3810 | imbalance, sd, CPU_NEWLY_IDLE, |
3811 | &all_pinned); | |
d6d5cfaf | 3812 | spin_unlock(&busiest->lock); |
0a2966b4 | 3813 | |
969bb4e4 | 3814 | if (unlikely(all_pinned)) { |
7c16ec58 MT |
3815 | cpu_clear(cpu_of(busiest), *cpus); |
3816 | if (!cpus_empty(*cpus)) | |
0a2966b4 CL |
3817 | goto redo; |
3818 | } | |
d6d5cfaf NP |
3819 | } |
3820 | ||
43010659 | 3821 | if (!ld_moved) { |
d15bcfdb | 3822 | schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]); |
89c4710e SS |
3823 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
3824 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 NP |
3825 | return -1; |
3826 | } else | |
16cfb1c0 | 3827 | sd->nr_balance_failed = 0; |
1da177e4 | 3828 | |
43010659 | 3829 | return ld_moved; |
16cfb1c0 NP |
3830 | |
3831 | out_balanced: | |
d15bcfdb | 3832 | schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]); |
48f24c4d | 3833 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3834 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 3835 | return -1; |
16cfb1c0 | 3836 | sd->nr_balance_failed = 0; |
48f24c4d | 3837 | |
16cfb1c0 | 3838 | return 0; |
1da177e4 LT |
3839 | } |
3840 | ||
3841 | /* | |
3842 | * idle_balance is called by schedule() if this_cpu is about to become | |
3843 | * idle. Attempts to pull tasks from other CPUs. | |
3844 | */ | |
70b97a7f | 3845 | static void idle_balance(int this_cpu, struct rq *this_rq) |
1da177e4 LT |
3846 | { |
3847 | struct sched_domain *sd; | |
dd41f596 IM |
3848 | int pulled_task = -1; |
3849 | unsigned long next_balance = jiffies + HZ; | |
7c16ec58 | 3850 | cpumask_t tmpmask; |
1da177e4 LT |
3851 | |
3852 | for_each_domain(this_cpu, sd) { | |
92c4ca5c CL |
3853 | unsigned long interval; |
3854 | ||
3855 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3856 | continue; | |
3857 | ||
3858 | if (sd->flags & SD_BALANCE_NEWIDLE) | |
48f24c4d | 3859 | /* If we've pulled tasks over stop searching: */ |
7c16ec58 MT |
3860 | pulled_task = load_balance_newidle(this_cpu, this_rq, |
3861 | sd, &tmpmask); | |
92c4ca5c CL |
3862 | |
3863 | interval = msecs_to_jiffies(sd->balance_interval); | |
3864 | if (time_after(next_balance, sd->last_balance + interval)) | |
3865 | next_balance = sd->last_balance + interval; | |
3866 | if (pulled_task) | |
3867 | break; | |
1da177e4 | 3868 | } |
dd41f596 | 3869 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
1bd77f2d CL |
3870 | /* |
3871 | * We are going idle. next_balance may be set based on | |
3872 | * a busy processor. So reset next_balance. | |
3873 | */ | |
3874 | this_rq->next_balance = next_balance; | |
dd41f596 | 3875 | } |
1da177e4 LT |
3876 | } |
3877 | ||
3878 | /* | |
3879 | * active_load_balance is run by migration threads. It pushes running tasks | |
3880 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
3881 | * running on each physical CPU where possible, and avoids physical / | |
3882 | * logical imbalances. | |
3883 | * | |
3884 | * Called with busiest_rq locked. | |
3885 | */ | |
70b97a7f | 3886 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) |
1da177e4 | 3887 | { |
39507451 | 3888 | int target_cpu = busiest_rq->push_cpu; |
70b97a7f IM |
3889 | struct sched_domain *sd; |
3890 | struct rq *target_rq; | |
39507451 | 3891 | |
48f24c4d | 3892 | /* Is there any task to move? */ |
39507451 | 3893 | if (busiest_rq->nr_running <= 1) |
39507451 NP |
3894 | return; |
3895 | ||
3896 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
3897 | |
3898 | /* | |
39507451 | 3899 | * This condition is "impossible", if it occurs |
41a2d6cf | 3900 | * we need to fix it. Originally reported by |
39507451 | 3901 | * Bjorn Helgaas on a 128-cpu setup. |
1da177e4 | 3902 | */ |
39507451 | 3903 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 3904 | |
39507451 NP |
3905 | /* move a task from busiest_rq to target_rq */ |
3906 | double_lock_balance(busiest_rq, target_rq); | |
6e82a3be IM |
3907 | update_rq_clock(busiest_rq); |
3908 | update_rq_clock(target_rq); | |
39507451 NP |
3909 | |
3910 | /* Search for an sd spanning us and the target CPU. */ | |
c96d145e | 3911 | for_each_domain(target_cpu, sd) { |
39507451 | 3912 | if ((sd->flags & SD_LOAD_BALANCE) && |
48f24c4d | 3913 | cpu_isset(busiest_cpu, sd->span)) |
39507451 | 3914 | break; |
c96d145e | 3915 | } |
39507451 | 3916 | |
48f24c4d | 3917 | if (likely(sd)) { |
2d72376b | 3918 | schedstat_inc(sd, alb_count); |
39507451 | 3919 | |
43010659 PW |
3920 | if (move_one_task(target_rq, target_cpu, busiest_rq, |
3921 | sd, CPU_IDLE)) | |
48f24c4d IM |
3922 | schedstat_inc(sd, alb_pushed); |
3923 | else | |
3924 | schedstat_inc(sd, alb_failed); | |
3925 | } | |
39507451 | 3926 | spin_unlock(&target_rq->lock); |
1da177e4 LT |
3927 | } |
3928 | ||
46cb4b7c SS |
3929 | #ifdef CONFIG_NO_HZ |
3930 | static struct { | |
3931 | atomic_t load_balancer; | |
41a2d6cf | 3932 | cpumask_t cpu_mask; |
46cb4b7c SS |
3933 | } nohz ____cacheline_aligned = { |
3934 | .load_balancer = ATOMIC_INIT(-1), | |
3935 | .cpu_mask = CPU_MASK_NONE, | |
3936 | }; | |
3937 | ||
7835b98b | 3938 | /* |
46cb4b7c SS |
3939 | * This routine will try to nominate the ilb (idle load balancing) |
3940 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
3941 | * load balancing on behalf of all those cpus. If all the cpus in the system | |
3942 | * go into this tickless mode, then there will be no ilb owner (as there is | |
3943 | * no need for one) and all the cpus will sleep till the next wakeup event | |
3944 | * arrives... | |
3945 | * | |
3946 | * For the ilb owner, tick is not stopped. And this tick will be used | |
3947 | * for idle load balancing. ilb owner will still be part of | |
3948 | * nohz.cpu_mask.. | |
7835b98b | 3949 | * |
46cb4b7c SS |
3950 | * While stopping the tick, this cpu will become the ilb owner if there |
3951 | * is no other owner. And will be the owner till that cpu becomes busy | |
3952 | * or if all cpus in the system stop their ticks at which point | |
3953 | * there is no need for ilb owner. | |
3954 | * | |
3955 | * When the ilb owner becomes busy, it nominates another owner, during the | |
3956 | * next busy scheduler_tick() | |
3957 | */ | |
3958 | int select_nohz_load_balancer(int stop_tick) | |
3959 | { | |
3960 | int cpu = smp_processor_id(); | |
3961 | ||
3962 | if (stop_tick) { | |
3963 | cpu_set(cpu, nohz.cpu_mask); | |
3964 | cpu_rq(cpu)->in_nohz_recently = 1; | |
3965 | ||
3966 | /* | |
3967 | * If we are going offline and still the leader, give up! | |
3968 | */ | |
3969 | if (cpu_is_offline(cpu) && | |
3970 | atomic_read(&nohz.load_balancer) == cpu) { | |
3971 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3972 | BUG(); | |
3973 | return 0; | |
3974 | } | |
3975 | ||
3976 | /* time for ilb owner also to sleep */ | |
3977 | if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
3978 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3979 | atomic_set(&nohz.load_balancer, -1); | |
3980 | return 0; | |
3981 | } | |
3982 | ||
3983 | if (atomic_read(&nohz.load_balancer) == -1) { | |
3984 | /* make me the ilb owner */ | |
3985 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) | |
3986 | return 1; | |
3987 | } else if (atomic_read(&nohz.load_balancer) == cpu) | |
3988 | return 1; | |
3989 | } else { | |
3990 | if (!cpu_isset(cpu, nohz.cpu_mask)) | |
3991 | return 0; | |
3992 | ||
3993 | cpu_clear(cpu, nohz.cpu_mask); | |
3994 | ||
3995 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3996 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3997 | BUG(); | |
3998 | } | |
3999 | return 0; | |
4000 | } | |
4001 | #endif | |
4002 | ||
4003 | static DEFINE_SPINLOCK(balancing); | |
4004 | ||
4005 | /* | |
7835b98b CL |
4006 | * It checks each scheduling domain to see if it is due to be balanced, |
4007 | * and initiates a balancing operation if so. | |
4008 | * | |
4009 | * Balancing parameters are set up in arch_init_sched_domains. | |
4010 | */ | |
a9957449 | 4011 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) |
7835b98b | 4012 | { |
46cb4b7c SS |
4013 | int balance = 1; |
4014 | struct rq *rq = cpu_rq(cpu); | |
7835b98b CL |
4015 | unsigned long interval; |
4016 | struct sched_domain *sd; | |
46cb4b7c | 4017 | /* Earliest time when we have to do rebalance again */ |
c9819f45 | 4018 | unsigned long next_balance = jiffies + 60*HZ; |
f549da84 | 4019 | int update_next_balance = 0; |
7c16ec58 | 4020 | cpumask_t tmp; |
1da177e4 | 4021 | |
46cb4b7c | 4022 | for_each_domain(cpu, sd) { |
1da177e4 LT |
4023 | if (!(sd->flags & SD_LOAD_BALANCE)) |
4024 | continue; | |
4025 | ||
4026 | interval = sd->balance_interval; | |
d15bcfdb | 4027 | if (idle != CPU_IDLE) |
1da177e4 LT |
4028 | interval *= sd->busy_factor; |
4029 | ||
4030 | /* scale ms to jiffies */ | |
4031 | interval = msecs_to_jiffies(interval); | |
4032 | if (unlikely(!interval)) | |
4033 | interval = 1; | |
dd41f596 IM |
4034 | if (interval > HZ*NR_CPUS/10) |
4035 | interval = HZ*NR_CPUS/10; | |
4036 | ||
1da177e4 | 4037 | |
08c183f3 CL |
4038 | if (sd->flags & SD_SERIALIZE) { |
4039 | if (!spin_trylock(&balancing)) | |
4040 | goto out; | |
4041 | } | |
4042 | ||
c9819f45 | 4043 | if (time_after_eq(jiffies, sd->last_balance + interval)) { |
7c16ec58 | 4044 | if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) { |
fa3b6ddc SS |
4045 | /* |
4046 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
4047 | * longer idle, or one of our SMT siblings is |
4048 | * not idle. | |
4049 | */ | |
d15bcfdb | 4050 | idle = CPU_NOT_IDLE; |
1da177e4 | 4051 | } |
1bd77f2d | 4052 | sd->last_balance = jiffies; |
1da177e4 | 4053 | } |
08c183f3 CL |
4054 | if (sd->flags & SD_SERIALIZE) |
4055 | spin_unlock(&balancing); | |
4056 | out: | |
f549da84 | 4057 | if (time_after(next_balance, sd->last_balance + interval)) { |
c9819f45 | 4058 | next_balance = sd->last_balance + interval; |
f549da84 SS |
4059 | update_next_balance = 1; |
4060 | } | |
783609c6 SS |
4061 | |
4062 | /* | |
4063 | * Stop the load balance at this level. There is another | |
4064 | * CPU in our sched group which is doing load balancing more | |
4065 | * actively. | |
4066 | */ | |
4067 | if (!balance) | |
4068 | break; | |
1da177e4 | 4069 | } |
f549da84 SS |
4070 | |
4071 | /* | |
4072 | * next_balance will be updated only when there is a need. | |
4073 | * When the cpu is attached to null domain for ex, it will not be | |
4074 | * updated. | |
4075 | */ | |
4076 | if (likely(update_next_balance)) | |
4077 | rq->next_balance = next_balance; | |
46cb4b7c SS |
4078 | } |
4079 | ||
4080 | /* | |
4081 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
4082 | * In CONFIG_NO_HZ case, the idle load balance owner will do the | |
4083 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
4084 | */ | |
4085 | static void run_rebalance_domains(struct softirq_action *h) | |
4086 | { | |
dd41f596 IM |
4087 | int this_cpu = smp_processor_id(); |
4088 | struct rq *this_rq = cpu_rq(this_cpu); | |
4089 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
4090 | CPU_IDLE : CPU_NOT_IDLE; | |
46cb4b7c | 4091 | |
dd41f596 | 4092 | rebalance_domains(this_cpu, idle); |
46cb4b7c SS |
4093 | |
4094 | #ifdef CONFIG_NO_HZ | |
4095 | /* | |
4096 | * If this cpu is the owner for idle load balancing, then do the | |
4097 | * balancing on behalf of the other idle cpus whose ticks are | |
4098 | * stopped. | |
4099 | */ | |
dd41f596 IM |
4100 | if (this_rq->idle_at_tick && |
4101 | atomic_read(&nohz.load_balancer) == this_cpu) { | |
46cb4b7c SS |
4102 | cpumask_t cpus = nohz.cpu_mask; |
4103 | struct rq *rq; | |
4104 | int balance_cpu; | |
4105 | ||
dd41f596 | 4106 | cpu_clear(this_cpu, cpus); |
46cb4b7c SS |
4107 | for_each_cpu_mask(balance_cpu, cpus) { |
4108 | /* | |
4109 | * If this cpu gets work to do, stop the load balancing | |
4110 | * work being done for other cpus. Next load | |
4111 | * balancing owner will pick it up. | |
4112 | */ | |
4113 | if (need_resched()) | |
4114 | break; | |
4115 | ||
de0cf899 | 4116 | rebalance_domains(balance_cpu, CPU_IDLE); |
46cb4b7c SS |
4117 | |
4118 | rq = cpu_rq(balance_cpu); | |
dd41f596 IM |
4119 | if (time_after(this_rq->next_balance, rq->next_balance)) |
4120 | this_rq->next_balance = rq->next_balance; | |
46cb4b7c SS |
4121 | } |
4122 | } | |
4123 | #endif | |
4124 | } | |
4125 | ||
4126 | /* | |
4127 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
4128 | * | |
4129 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new | |
4130 | * idle load balancing owner or decide to stop the periodic load balancing, | |
4131 | * if the whole system is idle. | |
4132 | */ | |
dd41f596 | 4133 | static inline void trigger_load_balance(struct rq *rq, int cpu) |
46cb4b7c | 4134 | { |
46cb4b7c SS |
4135 | #ifdef CONFIG_NO_HZ |
4136 | /* | |
4137 | * If we were in the nohz mode recently and busy at the current | |
4138 | * scheduler tick, then check if we need to nominate new idle | |
4139 | * load balancer. | |
4140 | */ | |
4141 | if (rq->in_nohz_recently && !rq->idle_at_tick) { | |
4142 | rq->in_nohz_recently = 0; | |
4143 | ||
4144 | if (atomic_read(&nohz.load_balancer) == cpu) { | |
4145 | cpu_clear(cpu, nohz.cpu_mask); | |
4146 | atomic_set(&nohz.load_balancer, -1); | |
4147 | } | |
4148 | ||
4149 | if (atomic_read(&nohz.load_balancer) == -1) { | |
4150 | /* | |
4151 | * simple selection for now: Nominate the | |
4152 | * first cpu in the nohz list to be the next | |
4153 | * ilb owner. | |
4154 | * | |
4155 | * TBD: Traverse the sched domains and nominate | |
4156 | * the nearest cpu in the nohz.cpu_mask. | |
4157 | */ | |
4158 | int ilb = first_cpu(nohz.cpu_mask); | |
4159 | ||
434d53b0 | 4160 | if (ilb < nr_cpu_ids) |
46cb4b7c SS |
4161 | resched_cpu(ilb); |
4162 | } | |
4163 | } | |
4164 | ||
4165 | /* | |
4166 | * If this cpu is idle and doing idle load balancing for all the | |
4167 | * cpus with ticks stopped, is it time for that to stop? | |
4168 | */ | |
4169 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && | |
4170 | cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
4171 | resched_cpu(cpu); | |
4172 | return; | |
4173 | } | |
4174 | ||
4175 | /* | |
4176 | * If this cpu is idle and the idle load balancing is done by | |
4177 | * someone else, then no need raise the SCHED_SOFTIRQ | |
4178 | */ | |
4179 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && | |
4180 | cpu_isset(cpu, nohz.cpu_mask)) | |
4181 | return; | |
4182 | #endif | |
4183 | if (time_after_eq(jiffies, rq->next_balance)) | |
4184 | raise_softirq(SCHED_SOFTIRQ); | |
1da177e4 | 4185 | } |
dd41f596 IM |
4186 | |
4187 | #else /* CONFIG_SMP */ | |
4188 | ||
1da177e4 LT |
4189 | /* |
4190 | * on UP we do not need to balance between CPUs: | |
4191 | */ | |
70b97a7f | 4192 | static inline void idle_balance(int cpu, struct rq *rq) |
1da177e4 LT |
4193 | { |
4194 | } | |
dd41f596 | 4195 | |
1da177e4 LT |
4196 | #endif |
4197 | ||
1da177e4 LT |
4198 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
4199 | ||
4200 | EXPORT_PER_CPU_SYMBOL(kstat); | |
4201 | ||
4202 | /* | |
41b86e9c IM |
4203 | * Return p->sum_exec_runtime plus any more ns on the sched_clock |
4204 | * that have not yet been banked in case the task is currently running. | |
1da177e4 | 4205 | */ |
41b86e9c | 4206 | unsigned long long task_sched_runtime(struct task_struct *p) |
1da177e4 | 4207 | { |
1da177e4 | 4208 | unsigned long flags; |
41b86e9c IM |
4209 | u64 ns, delta_exec; |
4210 | struct rq *rq; | |
48f24c4d | 4211 | |
41b86e9c IM |
4212 | rq = task_rq_lock(p, &flags); |
4213 | ns = p->se.sum_exec_runtime; | |
051a1d1a | 4214 | if (task_current(rq, p)) { |
a8e504d2 IM |
4215 | update_rq_clock(rq); |
4216 | delta_exec = rq->clock - p->se.exec_start; | |
41b86e9c IM |
4217 | if ((s64)delta_exec > 0) |
4218 | ns += delta_exec; | |
4219 | } | |
4220 | task_rq_unlock(rq, &flags); | |
48f24c4d | 4221 | |
1da177e4 LT |
4222 | return ns; |
4223 | } | |
4224 | ||
1da177e4 LT |
4225 | /* |
4226 | * Account user cpu time to a process. | |
4227 | * @p: the process that the cpu time gets accounted to | |
1da177e4 LT |
4228 | * @cputime: the cpu time spent in user space since the last update |
4229 | */ | |
4230 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
4231 | { | |
4232 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4233 | cputime64_t tmp; | |
4234 | ||
4235 | p->utime = cputime_add(p->utime, cputime); | |
4236 | ||
4237 | /* Add user time to cpustat. */ | |
4238 | tmp = cputime_to_cputime64(cputime); | |
4239 | if (TASK_NICE(p) > 0) | |
4240 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
4241 | else | |
4242 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
4243 | } | |
4244 | ||
94886b84 LV |
4245 | /* |
4246 | * Account guest cpu time to a process. | |
4247 | * @p: the process that the cpu time gets accounted to | |
4248 | * @cputime: the cpu time spent in virtual machine since the last update | |
4249 | */ | |
f7402e03 | 4250 | static void account_guest_time(struct task_struct *p, cputime_t cputime) |
94886b84 LV |
4251 | { |
4252 | cputime64_t tmp; | |
4253 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4254 | ||
4255 | tmp = cputime_to_cputime64(cputime); | |
4256 | ||
4257 | p->utime = cputime_add(p->utime, cputime); | |
4258 | p->gtime = cputime_add(p->gtime, cputime); | |
4259 | ||
4260 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
4261 | cpustat->guest = cputime64_add(cpustat->guest, tmp); | |
4262 | } | |
4263 | ||
c66f08be MN |
4264 | /* |
4265 | * Account scaled user cpu time to a process. | |
4266 | * @p: the process that the cpu time gets accounted to | |
4267 | * @cputime: the cpu time spent in user space since the last update | |
4268 | */ | |
4269 | void account_user_time_scaled(struct task_struct *p, cputime_t cputime) | |
4270 | { | |
4271 | p->utimescaled = cputime_add(p->utimescaled, cputime); | |
4272 | } | |
4273 | ||
1da177e4 LT |
4274 | /* |
4275 | * Account system cpu time to a process. | |
4276 | * @p: the process that the cpu time gets accounted to | |
4277 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
4278 | * @cputime: the cpu time spent in kernel space since the last update | |
4279 | */ | |
4280 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
4281 | cputime_t cputime) | |
4282 | { | |
4283 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
70b97a7f | 4284 | struct rq *rq = this_rq(); |
1da177e4 LT |
4285 | cputime64_t tmp; |
4286 | ||
983ed7a6 HH |
4287 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { |
4288 | account_guest_time(p, cputime); | |
4289 | return; | |
4290 | } | |
94886b84 | 4291 | |
1da177e4 LT |
4292 | p->stime = cputime_add(p->stime, cputime); |
4293 | ||
4294 | /* Add system time to cpustat. */ | |
4295 | tmp = cputime_to_cputime64(cputime); | |
4296 | if (hardirq_count() - hardirq_offset) | |
4297 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
4298 | else if (softirq_count()) | |
4299 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
cfb52856 | 4300 | else if (p != rq->idle) |
1da177e4 | 4301 | cpustat->system = cputime64_add(cpustat->system, tmp); |
cfb52856 | 4302 | else if (atomic_read(&rq->nr_iowait) > 0) |
1da177e4 LT |
4303 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
4304 | else | |
4305 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
4306 | /* Account for system time used */ | |
4307 | acct_update_integrals(p); | |
1da177e4 LT |
4308 | } |
4309 | ||
c66f08be MN |
4310 | /* |
4311 | * Account scaled system cpu time to a process. | |
4312 | * @p: the process that the cpu time gets accounted to | |
4313 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
4314 | * @cputime: the cpu time spent in kernel space since the last update | |
4315 | */ | |
4316 | void account_system_time_scaled(struct task_struct *p, cputime_t cputime) | |
4317 | { | |
4318 | p->stimescaled = cputime_add(p->stimescaled, cputime); | |
4319 | } | |
4320 | ||
1da177e4 LT |
4321 | /* |
4322 | * Account for involuntary wait time. | |
4323 | * @p: the process from which the cpu time has been stolen | |
4324 | * @steal: the cpu time spent in involuntary wait | |
4325 | */ | |
4326 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
4327 | { | |
4328 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4329 | cputime64_t tmp = cputime_to_cputime64(steal); | |
70b97a7f | 4330 | struct rq *rq = this_rq(); |
1da177e4 LT |
4331 | |
4332 | if (p == rq->idle) { | |
4333 | p->stime = cputime_add(p->stime, steal); | |
4334 | if (atomic_read(&rq->nr_iowait) > 0) | |
4335 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
4336 | else | |
4337 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
cfb52856 | 4338 | } else |
1da177e4 LT |
4339 | cpustat->steal = cputime64_add(cpustat->steal, tmp); |
4340 | } | |
4341 | ||
7835b98b CL |
4342 | /* |
4343 | * This function gets called by the timer code, with HZ frequency. | |
4344 | * We call it with interrupts disabled. | |
4345 | * | |
4346 | * It also gets called by the fork code, when changing the parent's | |
4347 | * timeslices. | |
4348 | */ | |
4349 | void scheduler_tick(void) | |
4350 | { | |
7835b98b CL |
4351 | int cpu = smp_processor_id(); |
4352 | struct rq *rq = cpu_rq(cpu); | |
dd41f596 | 4353 | struct task_struct *curr = rq->curr; |
3e51f33f PZ |
4354 | |
4355 | sched_clock_tick(); | |
dd41f596 IM |
4356 | |
4357 | spin_lock(&rq->lock); | |
3e51f33f | 4358 | update_rq_clock(rq); |
f1a438d8 | 4359 | update_cpu_load(rq); |
fa85ae24 | 4360 | curr->sched_class->task_tick(rq, curr, 0); |
dd41f596 | 4361 | spin_unlock(&rq->lock); |
7835b98b | 4362 | |
e418e1c2 | 4363 | #ifdef CONFIG_SMP |
dd41f596 IM |
4364 | rq->idle_at_tick = idle_cpu(cpu); |
4365 | trigger_load_balance(rq, cpu); | |
e418e1c2 | 4366 | #endif |
1da177e4 LT |
4367 | } |
4368 | ||
6cd8a4bb SR |
4369 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ |
4370 | defined(CONFIG_PREEMPT_TRACER)) | |
4371 | ||
4372 | static inline unsigned long get_parent_ip(unsigned long addr) | |
4373 | { | |
4374 | if (in_lock_functions(addr)) { | |
4375 | addr = CALLER_ADDR2; | |
4376 | if (in_lock_functions(addr)) | |
4377 | addr = CALLER_ADDR3; | |
4378 | } | |
4379 | return addr; | |
4380 | } | |
1da177e4 | 4381 | |
43627582 | 4382 | void __kprobes add_preempt_count(int val) |
1da177e4 | 4383 | { |
6cd8a4bb | 4384 | #ifdef CONFIG_DEBUG_PREEMPT |
1da177e4 LT |
4385 | /* |
4386 | * Underflow? | |
4387 | */ | |
9a11b49a IM |
4388 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
4389 | return; | |
6cd8a4bb | 4390 | #endif |
1da177e4 | 4391 | preempt_count() += val; |
6cd8a4bb | 4392 | #ifdef CONFIG_DEBUG_PREEMPT |
1da177e4 LT |
4393 | /* |
4394 | * Spinlock count overflowing soon? | |
4395 | */ | |
33859f7f MOS |
4396 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
4397 | PREEMPT_MASK - 10); | |
6cd8a4bb SR |
4398 | #endif |
4399 | if (preempt_count() == val) | |
4400 | trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); | |
1da177e4 LT |
4401 | } |
4402 | EXPORT_SYMBOL(add_preempt_count); | |
4403 | ||
43627582 | 4404 | void __kprobes sub_preempt_count(int val) |
1da177e4 | 4405 | { |
6cd8a4bb | 4406 | #ifdef CONFIG_DEBUG_PREEMPT |
1da177e4 LT |
4407 | /* |
4408 | * Underflow? | |
4409 | */ | |
9a11b49a IM |
4410 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
4411 | return; | |
1da177e4 LT |
4412 | /* |
4413 | * Is the spinlock portion underflowing? | |
4414 | */ | |
9a11b49a IM |
4415 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
4416 | !(preempt_count() & PREEMPT_MASK))) | |
4417 | return; | |
6cd8a4bb | 4418 | #endif |
9a11b49a | 4419 | |
6cd8a4bb SR |
4420 | if (preempt_count() == val) |
4421 | trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); | |
1da177e4 LT |
4422 | preempt_count() -= val; |
4423 | } | |
4424 | EXPORT_SYMBOL(sub_preempt_count); | |
4425 | ||
4426 | #endif | |
4427 | ||
4428 | /* | |
dd41f596 | 4429 | * Print scheduling while atomic bug: |
1da177e4 | 4430 | */ |
dd41f596 | 4431 | static noinline void __schedule_bug(struct task_struct *prev) |
1da177e4 | 4432 | { |
838225b4 SS |
4433 | struct pt_regs *regs = get_irq_regs(); |
4434 | ||
4435 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", | |
4436 | prev->comm, prev->pid, preempt_count()); | |
4437 | ||
dd41f596 IM |
4438 | debug_show_held_locks(prev); |
4439 | if (irqs_disabled()) | |
4440 | print_irqtrace_events(prev); | |
838225b4 SS |
4441 | |
4442 | if (regs) | |
4443 | show_regs(regs); | |
4444 | else | |
4445 | dump_stack(); | |
dd41f596 | 4446 | } |
1da177e4 | 4447 | |
dd41f596 IM |
4448 | /* |
4449 | * Various schedule()-time debugging checks and statistics: | |
4450 | */ | |
4451 | static inline void schedule_debug(struct task_struct *prev) | |
4452 | { | |
1da177e4 | 4453 | /* |
41a2d6cf | 4454 | * Test if we are atomic. Since do_exit() needs to call into |
1da177e4 LT |
4455 | * schedule() atomically, we ignore that path for now. |
4456 | * Otherwise, whine if we are scheduling when we should not be. | |
4457 | */ | |
dd41f596 IM |
4458 | if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state)) |
4459 | __schedule_bug(prev); | |
4460 | ||
1da177e4 LT |
4461 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
4462 | ||
2d72376b | 4463 | schedstat_inc(this_rq(), sched_count); |
b8efb561 IM |
4464 | #ifdef CONFIG_SCHEDSTATS |
4465 | if (unlikely(prev->lock_depth >= 0)) { | |
2d72376b IM |
4466 | schedstat_inc(this_rq(), bkl_count); |
4467 | schedstat_inc(prev, sched_info.bkl_count); | |
b8efb561 IM |
4468 | } |
4469 | #endif | |
dd41f596 IM |
4470 | } |
4471 | ||
4472 | /* | |
4473 | * Pick up the highest-prio task: | |
4474 | */ | |
4475 | static inline struct task_struct * | |
ff95f3df | 4476 | pick_next_task(struct rq *rq, struct task_struct *prev) |
dd41f596 | 4477 | { |
5522d5d5 | 4478 | const struct sched_class *class; |
dd41f596 | 4479 | struct task_struct *p; |
1da177e4 LT |
4480 | |
4481 | /* | |
dd41f596 IM |
4482 | * Optimization: we know that if all tasks are in |
4483 | * the fair class we can call that function directly: | |
1da177e4 | 4484 | */ |
dd41f596 | 4485 | if (likely(rq->nr_running == rq->cfs.nr_running)) { |
fb8d4724 | 4486 | p = fair_sched_class.pick_next_task(rq); |
dd41f596 IM |
4487 | if (likely(p)) |
4488 | return p; | |
1da177e4 LT |
4489 | } |
4490 | ||
dd41f596 IM |
4491 | class = sched_class_highest; |
4492 | for ( ; ; ) { | |
fb8d4724 | 4493 | p = class->pick_next_task(rq); |
dd41f596 IM |
4494 | if (p) |
4495 | return p; | |
4496 | /* | |
4497 | * Will never be NULL as the idle class always | |
4498 | * returns a non-NULL p: | |
4499 | */ | |
4500 | class = class->next; | |
4501 | } | |
4502 | } | |
1da177e4 | 4503 | |
dd41f596 IM |
4504 | /* |
4505 | * schedule() is the main scheduler function. | |
4506 | */ | |
4507 | asmlinkage void __sched schedule(void) | |
4508 | { | |
4509 | struct task_struct *prev, *next; | |
67ca7bde | 4510 | unsigned long *switch_count; |
dd41f596 | 4511 | struct rq *rq; |
dd41f596 IM |
4512 | int cpu; |
4513 | ||
4514 | need_resched: | |
4515 | preempt_disable(); | |
4516 | cpu = smp_processor_id(); | |
4517 | rq = cpu_rq(cpu); | |
4518 | rcu_qsctr_inc(cpu); | |
4519 | prev = rq->curr; | |
4520 | switch_count = &prev->nivcsw; | |
4521 | ||
4522 | release_kernel_lock(prev); | |
4523 | need_resched_nonpreemptible: | |
4524 | ||
4525 | schedule_debug(prev); | |
1da177e4 | 4526 | |
8f4d37ec PZ |
4527 | hrtick_clear(rq); |
4528 | ||
1e819950 IM |
4529 | /* |
4530 | * Do the rq-clock update outside the rq lock: | |
4531 | */ | |
4532 | local_irq_disable(); | |
3e51f33f | 4533 | update_rq_clock(rq); |
1e819950 IM |
4534 | spin_lock(&rq->lock); |
4535 | clear_tsk_need_resched(prev); | |
1da177e4 | 4536 | |
1da177e4 | 4537 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
1da177e4 | 4538 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && |
23e3c3cd | 4539 | signal_pending(prev))) { |
1da177e4 | 4540 | prev->state = TASK_RUNNING; |
dd41f596 | 4541 | } else { |
2e1cb74a | 4542 | deactivate_task(rq, prev, 1); |
1da177e4 | 4543 | } |
dd41f596 | 4544 | switch_count = &prev->nvcsw; |
1da177e4 LT |
4545 | } |
4546 | ||
9a897c5a SR |
4547 | #ifdef CONFIG_SMP |
4548 | if (prev->sched_class->pre_schedule) | |
4549 | prev->sched_class->pre_schedule(rq, prev); | |
4550 | #endif | |
f65eda4f | 4551 | |
dd41f596 | 4552 | if (unlikely(!rq->nr_running)) |
1da177e4 | 4553 | idle_balance(cpu, rq); |
1da177e4 | 4554 | |
31ee529c | 4555 | prev->sched_class->put_prev_task(rq, prev); |
ff95f3df | 4556 | next = pick_next_task(rq, prev); |
1da177e4 | 4557 | |
1da177e4 | 4558 | if (likely(prev != next)) { |
673a90a1 DS |
4559 | sched_info_switch(prev, next); |
4560 | ||
1da177e4 LT |
4561 | rq->nr_switches++; |
4562 | rq->curr = next; | |
4563 | ++*switch_count; | |
4564 | ||
dd41f596 | 4565 | context_switch(rq, prev, next); /* unlocks the rq */ |
8f4d37ec PZ |
4566 | /* |
4567 | * the context switch might have flipped the stack from under | |
4568 | * us, hence refresh the local variables. | |
4569 | */ | |
4570 | cpu = smp_processor_id(); | |
4571 | rq = cpu_rq(cpu); | |
1da177e4 LT |
4572 | } else |
4573 | spin_unlock_irq(&rq->lock); | |
4574 | ||
8f4d37ec PZ |
4575 | hrtick_set(rq); |
4576 | ||
4577 | if (unlikely(reacquire_kernel_lock(current) < 0)) | |
1da177e4 | 4578 | goto need_resched_nonpreemptible; |
8f4d37ec | 4579 | |
1da177e4 LT |
4580 | preempt_enable_no_resched(); |
4581 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
4582 | goto need_resched; | |
4583 | } | |
1da177e4 LT |
4584 | EXPORT_SYMBOL(schedule); |
4585 | ||
4586 | #ifdef CONFIG_PREEMPT | |
4587 | /* | |
2ed6e34f | 4588 | * this is the entry point to schedule() from in-kernel preemption |
41a2d6cf | 4589 | * off of preempt_enable. Kernel preemptions off return from interrupt |
1da177e4 LT |
4590 | * occur there and call schedule directly. |
4591 | */ | |
4592 | asmlinkage void __sched preempt_schedule(void) | |
4593 | { | |
4594 | struct thread_info *ti = current_thread_info(); | |
6478d880 | 4595 | |
1da177e4 LT |
4596 | /* |
4597 | * If there is a non-zero preempt_count or interrupts are disabled, | |
41a2d6cf | 4598 | * we do not want to preempt the current task. Just return.. |
1da177e4 | 4599 | */ |
beed33a8 | 4600 | if (likely(ti->preempt_count || irqs_disabled())) |
1da177e4 LT |
4601 | return; |
4602 | ||
3a5c359a AK |
4603 | do { |
4604 | add_preempt_count(PREEMPT_ACTIVE); | |
3a5c359a | 4605 | schedule(); |
3a5c359a | 4606 | sub_preempt_count(PREEMPT_ACTIVE); |
1da177e4 | 4607 | |
3a5c359a AK |
4608 | /* |
4609 | * Check again in case we missed a preemption opportunity | |
4610 | * between schedule and now. | |
4611 | */ | |
4612 | barrier(); | |
4613 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); | |
1da177e4 | 4614 | } |
1da177e4 LT |
4615 | EXPORT_SYMBOL(preempt_schedule); |
4616 | ||
4617 | /* | |
2ed6e34f | 4618 | * this is the entry point to schedule() from kernel preemption |
1da177e4 LT |
4619 | * off of irq context. |
4620 | * Note, that this is called and return with irqs disabled. This will | |
4621 | * protect us against recursive calling from irq. | |
4622 | */ | |
4623 | asmlinkage void __sched preempt_schedule_irq(void) | |
4624 | { | |
4625 | struct thread_info *ti = current_thread_info(); | |
6478d880 | 4626 | |
2ed6e34f | 4627 | /* Catch callers which need to be fixed */ |
1da177e4 LT |
4628 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
4629 | ||
3a5c359a AK |
4630 | do { |
4631 | add_preempt_count(PREEMPT_ACTIVE); | |
3a5c359a AK |
4632 | local_irq_enable(); |
4633 | schedule(); | |
4634 | local_irq_disable(); | |
3a5c359a | 4635 | sub_preempt_count(PREEMPT_ACTIVE); |
1da177e4 | 4636 | |
3a5c359a AK |
4637 | /* |
4638 | * Check again in case we missed a preemption opportunity | |
4639 | * between schedule and now. | |
4640 | */ | |
4641 | barrier(); | |
4642 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); | |
1da177e4 LT |
4643 | } |
4644 | ||
4645 | #endif /* CONFIG_PREEMPT */ | |
4646 | ||
95cdf3b7 IM |
4647 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
4648 | void *key) | |
1da177e4 | 4649 | { |
48f24c4d | 4650 | return try_to_wake_up(curr->private, mode, sync); |
1da177e4 | 4651 | } |
1da177e4 LT |
4652 | EXPORT_SYMBOL(default_wake_function); |
4653 | ||
4654 | /* | |
41a2d6cf IM |
4655 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
4656 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
1da177e4 LT |
4657 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
4658 | * | |
4659 | * There are circumstances in which we can try to wake a task which has already | |
41a2d6cf | 4660 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
1da177e4 LT |
4661 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
4662 | */ | |
4663 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
4664 | int nr_exclusive, int sync, void *key) | |
4665 | { | |
2e45874c | 4666 | wait_queue_t *curr, *next; |
1da177e4 | 4667 | |
2e45874c | 4668 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { |
48f24c4d IM |
4669 | unsigned flags = curr->flags; |
4670 | ||
1da177e4 | 4671 | if (curr->func(curr, mode, sync, key) && |
48f24c4d | 4672 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) |
1da177e4 LT |
4673 | break; |
4674 | } | |
4675 | } | |
4676 | ||
4677 | /** | |
4678 | * __wake_up - wake up threads blocked on a waitqueue. | |
4679 | * @q: the waitqueue | |
4680 | * @mode: which threads | |
4681 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 4682 | * @key: is directly passed to the wakeup function |
1da177e4 | 4683 | */ |
7ad5b3a5 | 4684 | void __wake_up(wait_queue_head_t *q, unsigned int mode, |
95cdf3b7 | 4685 | int nr_exclusive, void *key) |
1da177e4 LT |
4686 | { |
4687 | unsigned long flags; | |
4688 | ||
4689 | spin_lock_irqsave(&q->lock, flags); | |
4690 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
4691 | spin_unlock_irqrestore(&q->lock, flags); | |
4692 | } | |
1da177e4 LT |
4693 | EXPORT_SYMBOL(__wake_up); |
4694 | ||
4695 | /* | |
4696 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
4697 | */ | |
7ad5b3a5 | 4698 | void __wake_up_locked(wait_queue_head_t *q, unsigned int mode) |
1da177e4 LT |
4699 | { |
4700 | __wake_up_common(q, mode, 1, 0, NULL); | |
4701 | } | |
4702 | ||
4703 | /** | |
67be2dd1 | 4704 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
4705 | * @q: the waitqueue |
4706 | * @mode: which threads | |
4707 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
4708 | * | |
4709 | * The sync wakeup differs that the waker knows that it will schedule | |
4710 | * away soon, so while the target thread will be woken up, it will not | |
4711 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
4712 | * with each other. This can prevent needless bouncing between CPUs. | |
4713 | * | |
4714 | * On UP it can prevent extra preemption. | |
4715 | */ | |
7ad5b3a5 | 4716 | void |
95cdf3b7 | 4717 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
1da177e4 LT |
4718 | { |
4719 | unsigned long flags; | |
4720 | int sync = 1; | |
4721 | ||
4722 | if (unlikely(!q)) | |
4723 | return; | |
4724 | ||
4725 | if (unlikely(!nr_exclusive)) | |
4726 | sync = 0; | |
4727 | ||
4728 | spin_lock_irqsave(&q->lock, flags); | |
4729 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
4730 | spin_unlock_irqrestore(&q->lock, flags); | |
4731 | } | |
4732 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
4733 | ||
b15136e9 | 4734 | void complete(struct completion *x) |
1da177e4 LT |
4735 | { |
4736 | unsigned long flags; | |
4737 | ||
4738 | spin_lock_irqsave(&x->wait.lock, flags); | |
4739 | x->done++; | |
d9514f6c | 4740 | __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); |
1da177e4 LT |
4741 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4742 | } | |
4743 | EXPORT_SYMBOL(complete); | |
4744 | ||
b15136e9 | 4745 | void complete_all(struct completion *x) |
1da177e4 LT |
4746 | { |
4747 | unsigned long flags; | |
4748 | ||
4749 | spin_lock_irqsave(&x->wait.lock, flags); | |
4750 | x->done += UINT_MAX/2; | |
d9514f6c | 4751 | __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); |
1da177e4 LT |
4752 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4753 | } | |
4754 | EXPORT_SYMBOL(complete_all); | |
4755 | ||
8cbbe86d AK |
4756 | static inline long __sched |
4757 | do_wait_for_common(struct completion *x, long timeout, int state) | |
1da177e4 | 4758 | { |
1da177e4 LT |
4759 | if (!x->done) { |
4760 | DECLARE_WAITQUEUE(wait, current); | |
4761 | ||
4762 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
4763 | __add_wait_queue_tail(&x->wait, &wait); | |
4764 | do { | |
009e577e MW |
4765 | if ((state == TASK_INTERRUPTIBLE && |
4766 | signal_pending(current)) || | |
4767 | (state == TASK_KILLABLE && | |
4768 | fatal_signal_pending(current))) { | |
8cbbe86d AK |
4769 | __remove_wait_queue(&x->wait, &wait); |
4770 | return -ERESTARTSYS; | |
4771 | } | |
4772 | __set_current_state(state); | |
1da177e4 LT |
4773 | spin_unlock_irq(&x->wait.lock); |
4774 | timeout = schedule_timeout(timeout); | |
4775 | spin_lock_irq(&x->wait.lock); | |
4776 | if (!timeout) { | |
4777 | __remove_wait_queue(&x->wait, &wait); | |
8cbbe86d | 4778 | return timeout; |
1da177e4 LT |
4779 | } |
4780 | } while (!x->done); | |
4781 | __remove_wait_queue(&x->wait, &wait); | |
4782 | } | |
4783 | x->done--; | |
1da177e4 LT |
4784 | return timeout; |
4785 | } | |
1da177e4 | 4786 | |
8cbbe86d AK |
4787 | static long __sched |
4788 | wait_for_common(struct completion *x, long timeout, int state) | |
1da177e4 | 4789 | { |
1da177e4 LT |
4790 | might_sleep(); |
4791 | ||
4792 | spin_lock_irq(&x->wait.lock); | |
8cbbe86d | 4793 | timeout = do_wait_for_common(x, timeout, state); |
1da177e4 | 4794 | spin_unlock_irq(&x->wait.lock); |
8cbbe86d AK |
4795 | return timeout; |
4796 | } | |
1da177e4 | 4797 | |
b15136e9 | 4798 | void __sched wait_for_completion(struct completion *x) |
8cbbe86d AK |
4799 | { |
4800 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); | |
1da177e4 | 4801 | } |
8cbbe86d | 4802 | EXPORT_SYMBOL(wait_for_completion); |
1da177e4 | 4803 | |
b15136e9 | 4804 | unsigned long __sched |
8cbbe86d | 4805 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
1da177e4 | 4806 | { |
8cbbe86d | 4807 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); |
1da177e4 | 4808 | } |
8cbbe86d | 4809 | EXPORT_SYMBOL(wait_for_completion_timeout); |
1da177e4 | 4810 | |
8cbbe86d | 4811 | int __sched wait_for_completion_interruptible(struct completion *x) |
0fec171c | 4812 | { |
51e97990 AK |
4813 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); |
4814 | if (t == -ERESTARTSYS) | |
4815 | return t; | |
4816 | return 0; | |
0fec171c | 4817 | } |
8cbbe86d | 4818 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
1da177e4 | 4819 | |
b15136e9 | 4820 | unsigned long __sched |
8cbbe86d AK |
4821 | wait_for_completion_interruptible_timeout(struct completion *x, |
4822 | unsigned long timeout) | |
0fec171c | 4823 | { |
8cbbe86d | 4824 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); |
0fec171c | 4825 | } |
8cbbe86d | 4826 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
1da177e4 | 4827 | |
009e577e MW |
4828 | int __sched wait_for_completion_killable(struct completion *x) |
4829 | { | |
4830 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); | |
4831 | if (t == -ERESTARTSYS) | |
4832 | return t; | |
4833 | return 0; | |
4834 | } | |
4835 | EXPORT_SYMBOL(wait_for_completion_killable); | |
4836 | ||
8cbbe86d AK |
4837 | static long __sched |
4838 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) | |
1da177e4 | 4839 | { |
0fec171c IM |
4840 | unsigned long flags; |
4841 | wait_queue_t wait; | |
4842 | ||
4843 | init_waitqueue_entry(&wait, current); | |
1da177e4 | 4844 | |
8cbbe86d | 4845 | __set_current_state(state); |
1da177e4 | 4846 | |
8cbbe86d AK |
4847 | spin_lock_irqsave(&q->lock, flags); |
4848 | __add_wait_queue(q, &wait); | |
4849 | spin_unlock(&q->lock); | |
4850 | timeout = schedule_timeout(timeout); | |
4851 | spin_lock_irq(&q->lock); | |
4852 | __remove_wait_queue(q, &wait); | |
4853 | spin_unlock_irqrestore(&q->lock, flags); | |
4854 | ||
4855 | return timeout; | |
4856 | } | |
4857 | ||
4858 | void __sched interruptible_sleep_on(wait_queue_head_t *q) | |
4859 | { | |
4860 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | |
1da177e4 | 4861 | } |
1da177e4 LT |
4862 | EXPORT_SYMBOL(interruptible_sleep_on); |
4863 | ||
0fec171c | 4864 | long __sched |
95cdf3b7 | 4865 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
1da177e4 | 4866 | { |
8cbbe86d | 4867 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); |
1da177e4 | 4868 | } |
1da177e4 LT |
4869 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
4870 | ||
0fec171c | 4871 | void __sched sleep_on(wait_queue_head_t *q) |
1da177e4 | 4872 | { |
8cbbe86d | 4873 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
1da177e4 | 4874 | } |
1da177e4 LT |
4875 | EXPORT_SYMBOL(sleep_on); |
4876 | ||
0fec171c | 4877 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
1da177e4 | 4878 | { |
8cbbe86d | 4879 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); |
1da177e4 | 4880 | } |
1da177e4 LT |
4881 | EXPORT_SYMBOL(sleep_on_timeout); |
4882 | ||
b29739f9 IM |
4883 | #ifdef CONFIG_RT_MUTEXES |
4884 | ||
4885 | /* | |
4886 | * rt_mutex_setprio - set the current priority of a task | |
4887 | * @p: task | |
4888 | * @prio: prio value (kernel-internal form) | |
4889 | * | |
4890 | * This function changes the 'effective' priority of a task. It does | |
4891 | * not touch ->normal_prio like __setscheduler(). | |
4892 | * | |
4893 | * Used by the rt_mutex code to implement priority inheritance logic. | |
4894 | */ | |
36c8b586 | 4895 | void rt_mutex_setprio(struct task_struct *p, int prio) |
b29739f9 IM |
4896 | { |
4897 | unsigned long flags; | |
83b699ed | 4898 | int oldprio, on_rq, running; |
70b97a7f | 4899 | struct rq *rq; |
cb469845 | 4900 | const struct sched_class *prev_class = p->sched_class; |
b29739f9 IM |
4901 | |
4902 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
4903 | ||
4904 | rq = task_rq_lock(p, &flags); | |
a8e504d2 | 4905 | update_rq_clock(rq); |
b29739f9 | 4906 | |
d5f9f942 | 4907 | oldprio = p->prio; |
dd41f596 | 4908 | on_rq = p->se.on_rq; |
051a1d1a | 4909 | running = task_current(rq, p); |
0e1f3483 | 4910 | if (on_rq) |
69be72c1 | 4911 | dequeue_task(rq, p, 0); |
0e1f3483 HS |
4912 | if (running) |
4913 | p->sched_class->put_prev_task(rq, p); | |
dd41f596 IM |
4914 | |
4915 | if (rt_prio(prio)) | |
4916 | p->sched_class = &rt_sched_class; | |
4917 | else | |
4918 | p->sched_class = &fair_sched_class; | |
4919 | ||
b29739f9 IM |
4920 | p->prio = prio; |
4921 | ||
0e1f3483 HS |
4922 | if (running) |
4923 | p->sched_class->set_curr_task(rq); | |
dd41f596 | 4924 | if (on_rq) { |
8159f87e | 4925 | enqueue_task(rq, p, 0); |
cb469845 SR |
4926 | |
4927 | check_class_changed(rq, p, prev_class, oldprio, running); | |
b29739f9 IM |
4928 | } |
4929 | task_rq_unlock(rq, &flags); | |
4930 | } | |
4931 | ||
4932 | #endif | |
4933 | ||
36c8b586 | 4934 | void set_user_nice(struct task_struct *p, long nice) |
1da177e4 | 4935 | { |
dd41f596 | 4936 | int old_prio, delta, on_rq; |
1da177e4 | 4937 | unsigned long flags; |
70b97a7f | 4938 | struct rq *rq; |
1da177e4 LT |
4939 | |
4940 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
4941 | return; | |
4942 | /* | |
4943 | * We have to be careful, if called from sys_setpriority(), | |
4944 | * the task might be in the middle of scheduling on another CPU. | |
4945 | */ | |
4946 | rq = task_rq_lock(p, &flags); | |
a8e504d2 | 4947 | update_rq_clock(rq); |
1da177e4 LT |
4948 | /* |
4949 | * The RT priorities are set via sched_setscheduler(), but we still | |
4950 | * allow the 'normal' nice value to be set - but as expected | |
4951 | * it wont have any effect on scheduling until the task is | |
dd41f596 | 4952 | * SCHED_FIFO/SCHED_RR: |
1da177e4 | 4953 | */ |
e05606d3 | 4954 | if (task_has_rt_policy(p)) { |
1da177e4 LT |
4955 | p->static_prio = NICE_TO_PRIO(nice); |
4956 | goto out_unlock; | |
4957 | } | |
dd41f596 | 4958 | on_rq = p->se.on_rq; |
18d95a28 | 4959 | if (on_rq) |
69be72c1 | 4960 | dequeue_task(rq, p, 0); |
1da177e4 | 4961 | |
1da177e4 | 4962 | p->static_prio = NICE_TO_PRIO(nice); |
2dd73a4f | 4963 | set_load_weight(p); |
b29739f9 IM |
4964 | old_prio = p->prio; |
4965 | p->prio = effective_prio(p); | |
4966 | delta = p->prio - old_prio; | |
1da177e4 | 4967 | |
dd41f596 | 4968 | if (on_rq) { |
8159f87e | 4969 | enqueue_task(rq, p, 0); |
1da177e4 | 4970 | /* |
d5f9f942 AM |
4971 | * If the task increased its priority or is running and |
4972 | * lowered its priority, then reschedule its CPU: | |
1da177e4 | 4973 | */ |
d5f9f942 | 4974 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
1da177e4 LT |
4975 | resched_task(rq->curr); |
4976 | } | |
4977 | out_unlock: | |
4978 | task_rq_unlock(rq, &flags); | |
4979 | } | |
1da177e4 LT |
4980 | EXPORT_SYMBOL(set_user_nice); |
4981 | ||
e43379f1 MM |
4982 | /* |
4983 | * can_nice - check if a task can reduce its nice value | |
4984 | * @p: task | |
4985 | * @nice: nice value | |
4986 | */ | |
36c8b586 | 4987 | int can_nice(const struct task_struct *p, const int nice) |
e43379f1 | 4988 | { |
024f4747 MM |
4989 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
4990 | int nice_rlim = 20 - nice; | |
48f24c4d | 4991 | |
e43379f1 MM |
4992 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
4993 | capable(CAP_SYS_NICE)); | |
4994 | } | |
4995 | ||
1da177e4 LT |
4996 | #ifdef __ARCH_WANT_SYS_NICE |
4997 | ||
4998 | /* | |
4999 | * sys_nice - change the priority of the current process. | |
5000 | * @increment: priority increment | |
5001 | * | |
5002 | * sys_setpriority is a more generic, but much slower function that | |
5003 | * does similar things. | |
5004 | */ | |
5005 | asmlinkage long sys_nice(int increment) | |
5006 | { | |
48f24c4d | 5007 | long nice, retval; |
1da177e4 LT |
5008 | |
5009 | /* | |
5010 | * Setpriority might change our priority at the same moment. | |
5011 | * We don't have to worry. Conceptually one call occurs first | |
5012 | * and we have a single winner. | |
5013 | */ | |
e43379f1 MM |
5014 | if (increment < -40) |
5015 | increment = -40; | |
1da177e4 LT |
5016 | if (increment > 40) |
5017 | increment = 40; | |
5018 | ||
5019 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
5020 | if (nice < -20) | |
5021 | nice = -20; | |
5022 | if (nice > 19) | |
5023 | nice = 19; | |
5024 | ||
e43379f1 MM |
5025 | if (increment < 0 && !can_nice(current, nice)) |
5026 | return -EPERM; | |
5027 | ||
1da177e4 LT |
5028 | retval = security_task_setnice(current, nice); |
5029 | if (retval) | |
5030 | return retval; | |
5031 | ||
5032 | set_user_nice(current, nice); | |
5033 | return 0; | |
5034 | } | |
5035 | ||
5036 | #endif | |
5037 | ||
5038 | /** | |
5039 | * task_prio - return the priority value of a given task. | |
5040 | * @p: the task in question. | |
5041 | * | |
5042 | * This is the priority value as seen by users in /proc. | |
5043 | * RT tasks are offset by -200. Normal tasks are centered | |
5044 | * around 0, value goes from -16 to +15. | |
5045 | */ | |
36c8b586 | 5046 | int task_prio(const struct task_struct *p) |
1da177e4 LT |
5047 | { |
5048 | return p->prio - MAX_RT_PRIO; | |
5049 | } | |
5050 | ||
5051 | /** | |
5052 | * task_nice - return the nice value of a given task. | |
5053 | * @p: the task in question. | |
5054 | */ | |
36c8b586 | 5055 | int task_nice(const struct task_struct *p) |
1da177e4 LT |
5056 | { |
5057 | return TASK_NICE(p); | |
5058 | } | |
150d8bed | 5059 | EXPORT_SYMBOL(task_nice); |
1da177e4 LT |
5060 | |
5061 | /** | |
5062 | * idle_cpu - is a given cpu idle currently? | |
5063 | * @cpu: the processor in question. | |
5064 | */ | |
5065 | int idle_cpu(int cpu) | |
5066 | { | |
5067 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
5068 | } | |
5069 | ||
1da177e4 LT |
5070 | /** |
5071 | * idle_task - return the idle task for a given cpu. | |
5072 | * @cpu: the processor in question. | |
5073 | */ | |
36c8b586 | 5074 | struct task_struct *idle_task(int cpu) |
1da177e4 LT |
5075 | { |
5076 | return cpu_rq(cpu)->idle; | |
5077 | } | |
5078 | ||
5079 | /** | |
5080 | * find_process_by_pid - find a process with a matching PID value. | |
5081 | * @pid: the pid in question. | |
5082 | */ | |
a9957449 | 5083 | static struct task_struct *find_process_by_pid(pid_t pid) |
1da177e4 | 5084 | { |
228ebcbe | 5085 | return pid ? find_task_by_vpid(pid) : current; |
1da177e4 LT |
5086 | } |
5087 | ||
5088 | /* Actually do priority change: must hold rq lock. */ | |
dd41f596 IM |
5089 | static void |
5090 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) | |
1da177e4 | 5091 | { |
dd41f596 | 5092 | BUG_ON(p->se.on_rq); |
48f24c4d | 5093 | |
1da177e4 | 5094 | p->policy = policy; |
dd41f596 IM |
5095 | switch (p->policy) { |
5096 | case SCHED_NORMAL: | |
5097 | case SCHED_BATCH: | |
5098 | case SCHED_IDLE: | |
5099 | p->sched_class = &fair_sched_class; | |
5100 | break; | |
5101 | case SCHED_FIFO: | |
5102 | case SCHED_RR: | |
5103 | p->sched_class = &rt_sched_class; | |
5104 | break; | |
5105 | } | |
5106 | ||
1da177e4 | 5107 | p->rt_priority = prio; |
b29739f9 IM |
5108 | p->normal_prio = normal_prio(p); |
5109 | /* we are holding p->pi_lock already */ | |
5110 | p->prio = rt_mutex_getprio(p); | |
2dd73a4f | 5111 | set_load_weight(p); |
1da177e4 LT |
5112 | } |
5113 | ||
5114 | /** | |
72fd4a35 | 5115 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. |
1da177e4 LT |
5116 | * @p: the task in question. |
5117 | * @policy: new policy. | |
5118 | * @param: structure containing the new RT priority. | |
5fe1d75f | 5119 | * |
72fd4a35 | 5120 | * NOTE that the task may be already dead. |
1da177e4 | 5121 | */ |
95cdf3b7 IM |
5122 | int sched_setscheduler(struct task_struct *p, int policy, |
5123 | struct sched_param *param) | |
1da177e4 | 5124 | { |
83b699ed | 5125 | int retval, oldprio, oldpolicy = -1, on_rq, running; |
1da177e4 | 5126 | unsigned long flags; |
cb469845 | 5127 | const struct sched_class *prev_class = p->sched_class; |
70b97a7f | 5128 | struct rq *rq; |
1da177e4 | 5129 | |
66e5393a SR |
5130 | /* may grab non-irq protected spin_locks */ |
5131 | BUG_ON(in_interrupt()); | |
1da177e4 LT |
5132 | recheck: |
5133 | /* double check policy once rq lock held */ | |
5134 | if (policy < 0) | |
5135 | policy = oldpolicy = p->policy; | |
5136 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
dd41f596 IM |
5137 | policy != SCHED_NORMAL && policy != SCHED_BATCH && |
5138 | policy != SCHED_IDLE) | |
b0a9499c | 5139 | return -EINVAL; |
1da177e4 LT |
5140 | /* |
5141 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
dd41f596 IM |
5142 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, |
5143 | * SCHED_BATCH and SCHED_IDLE is 0. | |
1da177e4 LT |
5144 | */ |
5145 | if (param->sched_priority < 0 || | |
95cdf3b7 | 5146 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 5147 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 5148 | return -EINVAL; |
e05606d3 | 5149 | if (rt_policy(policy) != (param->sched_priority != 0)) |
1da177e4 LT |
5150 | return -EINVAL; |
5151 | ||
37e4ab3f OC |
5152 | /* |
5153 | * Allow unprivileged RT tasks to decrease priority: | |
5154 | */ | |
5155 | if (!capable(CAP_SYS_NICE)) { | |
e05606d3 | 5156 | if (rt_policy(policy)) { |
8dc3e909 | 5157 | unsigned long rlim_rtprio; |
8dc3e909 ON |
5158 | |
5159 | if (!lock_task_sighand(p, &flags)) | |
5160 | return -ESRCH; | |
5161 | rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; | |
5162 | unlock_task_sighand(p, &flags); | |
5163 | ||
5164 | /* can't set/change the rt policy */ | |
5165 | if (policy != p->policy && !rlim_rtprio) | |
5166 | return -EPERM; | |
5167 | ||
5168 | /* can't increase priority */ | |
5169 | if (param->sched_priority > p->rt_priority && | |
5170 | param->sched_priority > rlim_rtprio) | |
5171 | return -EPERM; | |
5172 | } | |
dd41f596 IM |
5173 | /* |
5174 | * Like positive nice levels, dont allow tasks to | |
5175 | * move out of SCHED_IDLE either: | |
5176 | */ | |
5177 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) | |
5178 | return -EPERM; | |
5fe1d75f | 5179 | |
37e4ab3f OC |
5180 | /* can't change other user's priorities */ |
5181 | if ((current->euid != p->euid) && | |
5182 | (current->euid != p->uid)) | |
5183 | return -EPERM; | |
5184 | } | |
1da177e4 | 5185 | |
b68aa230 PZ |
5186 | #ifdef CONFIG_RT_GROUP_SCHED |
5187 | /* | |
5188 | * Do not allow realtime tasks into groups that have no runtime | |
5189 | * assigned. | |
5190 | */ | |
d0b27fa7 | 5191 | if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0) |
b68aa230 PZ |
5192 | return -EPERM; |
5193 | #endif | |
5194 | ||
1da177e4 LT |
5195 | retval = security_task_setscheduler(p, policy, param); |
5196 | if (retval) | |
5197 | return retval; | |
b29739f9 IM |
5198 | /* |
5199 | * make sure no PI-waiters arrive (or leave) while we are | |
5200 | * changing the priority of the task: | |
5201 | */ | |
5202 | spin_lock_irqsave(&p->pi_lock, flags); | |
1da177e4 LT |
5203 | /* |
5204 | * To be able to change p->policy safely, the apropriate | |
5205 | * runqueue lock must be held. | |
5206 | */ | |
b29739f9 | 5207 | rq = __task_rq_lock(p); |
1da177e4 LT |
5208 | /* recheck policy now with rq lock held */ |
5209 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
5210 | policy = oldpolicy = -1; | |
b29739f9 IM |
5211 | __task_rq_unlock(rq); |
5212 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
5213 | goto recheck; |
5214 | } | |
2daa3577 | 5215 | update_rq_clock(rq); |
dd41f596 | 5216 | on_rq = p->se.on_rq; |
051a1d1a | 5217 | running = task_current(rq, p); |
0e1f3483 | 5218 | if (on_rq) |
2e1cb74a | 5219 | deactivate_task(rq, p, 0); |
0e1f3483 HS |
5220 | if (running) |
5221 | p->sched_class->put_prev_task(rq, p); | |
f6b53205 | 5222 | |
1da177e4 | 5223 | oldprio = p->prio; |
dd41f596 | 5224 | __setscheduler(rq, p, policy, param->sched_priority); |
f6b53205 | 5225 | |
0e1f3483 HS |
5226 | if (running) |
5227 | p->sched_class->set_curr_task(rq); | |
dd41f596 IM |
5228 | if (on_rq) { |
5229 | activate_task(rq, p, 0); | |
cb469845 SR |
5230 | |
5231 | check_class_changed(rq, p, prev_class, oldprio, running); | |
1da177e4 | 5232 | } |
b29739f9 IM |
5233 | __task_rq_unlock(rq); |
5234 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
5235 | ||
95e02ca9 TG |
5236 | rt_mutex_adjust_pi(p); |
5237 | ||
1da177e4 LT |
5238 | return 0; |
5239 | } | |
5240 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
5241 | ||
95cdf3b7 IM |
5242 | static int |
5243 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 | 5244 | { |
1da177e4 LT |
5245 | struct sched_param lparam; |
5246 | struct task_struct *p; | |
36c8b586 | 5247 | int retval; |
1da177e4 LT |
5248 | |
5249 | if (!param || pid < 0) | |
5250 | return -EINVAL; | |
5251 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
5252 | return -EFAULT; | |
5fe1d75f ON |
5253 | |
5254 | rcu_read_lock(); | |
5255 | retval = -ESRCH; | |
1da177e4 | 5256 | p = find_process_by_pid(pid); |
5fe1d75f ON |
5257 | if (p != NULL) |
5258 | retval = sched_setscheduler(p, policy, &lparam); | |
5259 | rcu_read_unlock(); | |
36c8b586 | 5260 | |
1da177e4 LT |
5261 | return retval; |
5262 | } | |
5263 | ||
5264 | /** | |
5265 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
5266 | * @pid: the pid in question. | |
5267 | * @policy: new policy. | |
5268 | * @param: structure containing the new RT priority. | |
5269 | */ | |
41a2d6cf IM |
5270 | asmlinkage long |
5271 | sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 | 5272 | { |
c21761f1 JB |
5273 | /* negative values for policy are not valid */ |
5274 | if (policy < 0) | |
5275 | return -EINVAL; | |
5276 | ||
1da177e4 LT |
5277 | return do_sched_setscheduler(pid, policy, param); |
5278 | } | |
5279 | ||
5280 | /** | |
5281 | * sys_sched_setparam - set/change the RT priority of a thread | |
5282 | * @pid: the pid in question. | |
5283 | * @param: structure containing the new RT priority. | |
5284 | */ | |
5285 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
5286 | { | |
5287 | return do_sched_setscheduler(pid, -1, param); | |
5288 | } | |
5289 | ||
5290 | /** | |
5291 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
5292 | * @pid: the pid in question. | |
5293 | */ | |
5294 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
5295 | { | |
36c8b586 | 5296 | struct task_struct *p; |
3a5c359a | 5297 | int retval; |
1da177e4 LT |
5298 | |
5299 | if (pid < 0) | |
3a5c359a | 5300 | return -EINVAL; |
1da177e4 LT |
5301 | |
5302 | retval = -ESRCH; | |
5303 | read_lock(&tasklist_lock); | |
5304 | p = find_process_by_pid(pid); | |
5305 | if (p) { | |
5306 | retval = security_task_getscheduler(p); | |
5307 | if (!retval) | |
5308 | retval = p->policy; | |
5309 | } | |
5310 | read_unlock(&tasklist_lock); | |
1da177e4 LT |
5311 | return retval; |
5312 | } | |
5313 | ||
5314 | /** | |
5315 | * sys_sched_getscheduler - get the RT priority of a thread | |
5316 | * @pid: the pid in question. | |
5317 | * @param: structure containing the RT priority. | |
5318 | */ | |
5319 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
5320 | { | |
5321 | struct sched_param lp; | |
36c8b586 | 5322 | struct task_struct *p; |
3a5c359a | 5323 | int retval; |
1da177e4 LT |
5324 | |
5325 | if (!param || pid < 0) | |
3a5c359a | 5326 | return -EINVAL; |
1da177e4 LT |
5327 | |
5328 | read_lock(&tasklist_lock); | |
5329 | p = find_process_by_pid(pid); | |
5330 | retval = -ESRCH; | |
5331 | if (!p) | |
5332 | goto out_unlock; | |
5333 | ||
5334 | retval = security_task_getscheduler(p); | |
5335 | if (retval) | |
5336 | goto out_unlock; | |
5337 | ||
5338 | lp.sched_priority = p->rt_priority; | |
5339 | read_unlock(&tasklist_lock); | |
5340 | ||
5341 | /* | |
5342 | * This one might sleep, we cannot do it with a spinlock held ... | |
5343 | */ | |
5344 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
5345 | ||
1da177e4 LT |
5346 | return retval; |
5347 | ||
5348 | out_unlock: | |
5349 | read_unlock(&tasklist_lock); | |
5350 | return retval; | |
5351 | } | |
5352 | ||
b53e921b | 5353 | long sched_setaffinity(pid_t pid, const cpumask_t *in_mask) |
1da177e4 | 5354 | { |
1da177e4 | 5355 | cpumask_t cpus_allowed; |
b53e921b | 5356 | cpumask_t new_mask = *in_mask; |
36c8b586 IM |
5357 | struct task_struct *p; |
5358 | int retval; | |
1da177e4 | 5359 | |
95402b38 | 5360 | get_online_cpus(); |
1da177e4 LT |
5361 | read_lock(&tasklist_lock); |
5362 | ||
5363 | p = find_process_by_pid(pid); | |
5364 | if (!p) { | |
5365 | read_unlock(&tasklist_lock); | |
95402b38 | 5366 | put_online_cpus(); |
1da177e4 LT |
5367 | return -ESRCH; |
5368 | } | |
5369 | ||
5370 | /* | |
5371 | * It is not safe to call set_cpus_allowed with the | |
41a2d6cf | 5372 | * tasklist_lock held. We will bump the task_struct's |
1da177e4 LT |
5373 | * usage count and then drop tasklist_lock. |
5374 | */ | |
5375 | get_task_struct(p); | |
5376 | read_unlock(&tasklist_lock); | |
5377 | ||
5378 | retval = -EPERM; | |
5379 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
5380 | !capable(CAP_SYS_NICE)) | |
5381 | goto out_unlock; | |
5382 | ||
e7834f8f DQ |
5383 | retval = security_task_setscheduler(p, 0, NULL); |
5384 | if (retval) | |
5385 | goto out_unlock; | |
5386 | ||
f9a86fcb | 5387 | cpuset_cpus_allowed(p, &cpus_allowed); |
1da177e4 | 5388 | cpus_and(new_mask, new_mask, cpus_allowed); |
8707d8b8 | 5389 | again: |
7c16ec58 | 5390 | retval = set_cpus_allowed_ptr(p, &new_mask); |
1da177e4 | 5391 | |
8707d8b8 | 5392 | if (!retval) { |
f9a86fcb | 5393 | cpuset_cpus_allowed(p, &cpus_allowed); |
8707d8b8 PM |
5394 | if (!cpus_subset(new_mask, cpus_allowed)) { |
5395 | /* | |
5396 | * We must have raced with a concurrent cpuset | |
5397 | * update. Just reset the cpus_allowed to the | |
5398 | * cpuset's cpus_allowed | |
5399 | */ | |
5400 | new_mask = cpus_allowed; | |
5401 | goto again; | |
5402 | } | |
5403 | } | |
1da177e4 LT |
5404 | out_unlock: |
5405 | put_task_struct(p); | |
95402b38 | 5406 | put_online_cpus(); |
1da177e4 LT |
5407 | return retval; |
5408 | } | |
5409 | ||
5410 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
5411 | cpumask_t *new_mask) | |
5412 | { | |
5413 | if (len < sizeof(cpumask_t)) { | |
5414 | memset(new_mask, 0, sizeof(cpumask_t)); | |
5415 | } else if (len > sizeof(cpumask_t)) { | |
5416 | len = sizeof(cpumask_t); | |
5417 | } | |
5418 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
5419 | } | |
5420 | ||
5421 | /** | |
5422 | * sys_sched_setaffinity - set the cpu affinity of a process | |
5423 | * @pid: pid of the process | |
5424 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5425 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
5426 | */ | |
5427 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
5428 | unsigned long __user *user_mask_ptr) | |
5429 | { | |
5430 | cpumask_t new_mask; | |
5431 | int retval; | |
5432 | ||
5433 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
5434 | if (retval) | |
5435 | return retval; | |
5436 | ||
b53e921b | 5437 | return sched_setaffinity(pid, &new_mask); |
1da177e4 LT |
5438 | } |
5439 | ||
5440 | /* | |
5441 | * Represents all cpu's present in the system | |
5442 | * In systems capable of hotplug, this map could dynamically grow | |
5443 | * as new cpu's are detected in the system via any platform specific | |
5444 | * method, such as ACPI for e.g. | |
5445 | */ | |
5446 | ||
4cef0c61 | 5447 | cpumask_t cpu_present_map __read_mostly; |
1da177e4 LT |
5448 | EXPORT_SYMBOL(cpu_present_map); |
5449 | ||
5450 | #ifndef CONFIG_SMP | |
4cef0c61 | 5451 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 GB |
5452 | EXPORT_SYMBOL(cpu_online_map); |
5453 | ||
4cef0c61 | 5454 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 | 5455 | EXPORT_SYMBOL(cpu_possible_map); |
1da177e4 LT |
5456 | #endif |
5457 | ||
5458 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
5459 | { | |
36c8b586 | 5460 | struct task_struct *p; |
1da177e4 | 5461 | int retval; |
1da177e4 | 5462 | |
95402b38 | 5463 | get_online_cpus(); |
1da177e4 LT |
5464 | read_lock(&tasklist_lock); |
5465 | ||
5466 | retval = -ESRCH; | |
5467 | p = find_process_by_pid(pid); | |
5468 | if (!p) | |
5469 | goto out_unlock; | |
5470 | ||
e7834f8f DQ |
5471 | retval = security_task_getscheduler(p); |
5472 | if (retval) | |
5473 | goto out_unlock; | |
5474 | ||
2f7016d9 | 5475 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
1da177e4 LT |
5476 | |
5477 | out_unlock: | |
5478 | read_unlock(&tasklist_lock); | |
95402b38 | 5479 | put_online_cpus(); |
1da177e4 | 5480 | |
9531b62f | 5481 | return retval; |
1da177e4 LT |
5482 | } |
5483 | ||
5484 | /** | |
5485 | * sys_sched_getaffinity - get the cpu affinity of a process | |
5486 | * @pid: pid of the process | |
5487 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5488 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
5489 | */ | |
5490 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
5491 | unsigned long __user *user_mask_ptr) | |
5492 | { | |
5493 | int ret; | |
5494 | cpumask_t mask; | |
5495 | ||
5496 | if (len < sizeof(cpumask_t)) | |
5497 | return -EINVAL; | |
5498 | ||
5499 | ret = sched_getaffinity(pid, &mask); | |
5500 | if (ret < 0) | |
5501 | return ret; | |
5502 | ||
5503 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
5504 | return -EFAULT; | |
5505 | ||
5506 | return sizeof(cpumask_t); | |
5507 | } | |
5508 | ||
5509 | /** | |
5510 | * sys_sched_yield - yield the current processor to other threads. | |
5511 | * | |
dd41f596 IM |
5512 | * This function yields the current CPU to other tasks. If there are no |
5513 | * other threads running on this CPU then this function will return. | |
1da177e4 LT |
5514 | */ |
5515 | asmlinkage long sys_sched_yield(void) | |
5516 | { | |
70b97a7f | 5517 | struct rq *rq = this_rq_lock(); |
1da177e4 | 5518 | |
2d72376b | 5519 | schedstat_inc(rq, yld_count); |
4530d7ab | 5520 | current->sched_class->yield_task(rq); |
1da177e4 LT |
5521 | |
5522 | /* | |
5523 | * Since we are going to call schedule() anyway, there's | |
5524 | * no need to preempt or enable interrupts: | |
5525 | */ | |
5526 | __release(rq->lock); | |
8a25d5de | 5527 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
1da177e4 LT |
5528 | _raw_spin_unlock(&rq->lock); |
5529 | preempt_enable_no_resched(); | |
5530 | ||
5531 | schedule(); | |
5532 | ||
5533 | return 0; | |
5534 | } | |
5535 | ||
e7b38404 | 5536 | static void __cond_resched(void) |
1da177e4 | 5537 | { |
8e0a43d8 IM |
5538 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
5539 | __might_sleep(__FILE__, __LINE__); | |
5540 | #endif | |
5bbcfd90 IM |
5541 | /* |
5542 | * The BKS might be reacquired before we have dropped | |
5543 | * PREEMPT_ACTIVE, which could trigger a second | |
5544 | * cond_resched() call. | |
5545 | */ | |
1da177e4 LT |
5546 | do { |
5547 | add_preempt_count(PREEMPT_ACTIVE); | |
5548 | schedule(); | |
5549 | sub_preempt_count(PREEMPT_ACTIVE); | |
5550 | } while (need_resched()); | |
5551 | } | |
5552 | ||
02b67cc3 | 5553 | int __sched _cond_resched(void) |
1da177e4 | 5554 | { |
9414232f IM |
5555 | if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) && |
5556 | system_state == SYSTEM_RUNNING) { | |
1da177e4 LT |
5557 | __cond_resched(); |
5558 | return 1; | |
5559 | } | |
5560 | return 0; | |
5561 | } | |
02b67cc3 | 5562 | EXPORT_SYMBOL(_cond_resched); |
1da177e4 LT |
5563 | |
5564 | /* | |
5565 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
5566 | * call schedule, and on return reacquire the lock. | |
5567 | * | |
41a2d6cf | 5568 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
1da177e4 LT |
5569 | * operations here to prevent schedule() from being called twice (once via |
5570 | * spin_unlock(), once by hand). | |
5571 | */ | |
95cdf3b7 | 5572 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 5573 | { |
95c354fe | 5574 | int resched = need_resched() && system_state == SYSTEM_RUNNING; |
6df3cecb JK |
5575 | int ret = 0; |
5576 | ||
95c354fe | 5577 | if (spin_needbreak(lock) || resched) { |
1da177e4 | 5578 | spin_unlock(lock); |
95c354fe NP |
5579 | if (resched && need_resched()) |
5580 | __cond_resched(); | |
5581 | else | |
5582 | cpu_relax(); | |
6df3cecb | 5583 | ret = 1; |
1da177e4 | 5584 | spin_lock(lock); |
1da177e4 | 5585 | } |
6df3cecb | 5586 | return ret; |
1da177e4 | 5587 | } |
1da177e4 LT |
5588 | EXPORT_SYMBOL(cond_resched_lock); |
5589 | ||
5590 | int __sched cond_resched_softirq(void) | |
5591 | { | |
5592 | BUG_ON(!in_softirq()); | |
5593 | ||
9414232f | 5594 | if (need_resched() && system_state == SYSTEM_RUNNING) { |
98d82567 | 5595 | local_bh_enable(); |
1da177e4 LT |
5596 | __cond_resched(); |
5597 | local_bh_disable(); | |
5598 | return 1; | |
5599 | } | |
5600 | return 0; | |
5601 | } | |
1da177e4 LT |
5602 | EXPORT_SYMBOL(cond_resched_softirq); |
5603 | ||
1da177e4 LT |
5604 | /** |
5605 | * yield - yield the current processor to other threads. | |
5606 | * | |
72fd4a35 | 5607 | * This is a shortcut for kernel-space yielding - it marks the |
1da177e4 LT |
5608 | * thread runnable and calls sys_sched_yield(). |
5609 | */ | |
5610 | void __sched yield(void) | |
5611 | { | |
5612 | set_current_state(TASK_RUNNING); | |
5613 | sys_sched_yield(); | |
5614 | } | |
1da177e4 LT |
5615 | EXPORT_SYMBOL(yield); |
5616 | ||
5617 | /* | |
41a2d6cf | 5618 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
1da177e4 LT |
5619 | * that process accounting knows that this is a task in IO wait state. |
5620 | * | |
5621 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
5622 | * has set its backing_dev_info: the queue against which it should throttle) | |
5623 | */ | |
5624 | void __sched io_schedule(void) | |
5625 | { | |
70b97a7f | 5626 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 | 5627 | |
0ff92245 | 5628 | delayacct_blkio_start(); |
1da177e4 LT |
5629 | atomic_inc(&rq->nr_iowait); |
5630 | schedule(); | |
5631 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 5632 | delayacct_blkio_end(); |
1da177e4 | 5633 | } |
1da177e4 LT |
5634 | EXPORT_SYMBOL(io_schedule); |
5635 | ||
5636 | long __sched io_schedule_timeout(long timeout) | |
5637 | { | |
70b97a7f | 5638 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
5639 | long ret; |
5640 | ||
0ff92245 | 5641 | delayacct_blkio_start(); |
1da177e4 LT |
5642 | atomic_inc(&rq->nr_iowait); |
5643 | ret = schedule_timeout(timeout); | |
5644 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 5645 | delayacct_blkio_end(); |
1da177e4 LT |
5646 | return ret; |
5647 | } | |
5648 | ||
5649 | /** | |
5650 | * sys_sched_get_priority_max - return maximum RT priority. | |
5651 | * @policy: scheduling class. | |
5652 | * | |
5653 | * this syscall returns the maximum rt_priority that can be used | |
5654 | * by a given scheduling class. | |
5655 | */ | |
5656 | asmlinkage long sys_sched_get_priority_max(int policy) | |
5657 | { | |
5658 | int ret = -EINVAL; | |
5659 | ||
5660 | switch (policy) { | |
5661 | case SCHED_FIFO: | |
5662 | case SCHED_RR: | |
5663 | ret = MAX_USER_RT_PRIO-1; | |
5664 | break; | |
5665 | case SCHED_NORMAL: | |
b0a9499c | 5666 | case SCHED_BATCH: |
dd41f596 | 5667 | case SCHED_IDLE: |
1da177e4 LT |
5668 | ret = 0; |
5669 | break; | |
5670 | } | |
5671 | return ret; | |
5672 | } | |
5673 | ||
5674 | /** | |
5675 | * sys_sched_get_priority_min - return minimum RT priority. | |
5676 | * @policy: scheduling class. | |
5677 | * | |
5678 | * this syscall returns the minimum rt_priority that can be used | |
5679 | * by a given scheduling class. | |
5680 | */ | |
5681 | asmlinkage long sys_sched_get_priority_min(int policy) | |
5682 | { | |
5683 | int ret = -EINVAL; | |
5684 | ||
5685 | switch (policy) { | |
5686 | case SCHED_FIFO: | |
5687 | case SCHED_RR: | |
5688 | ret = 1; | |
5689 | break; | |
5690 | case SCHED_NORMAL: | |
b0a9499c | 5691 | case SCHED_BATCH: |
dd41f596 | 5692 | case SCHED_IDLE: |
1da177e4 LT |
5693 | ret = 0; |
5694 | } | |
5695 | return ret; | |
5696 | } | |
5697 | ||
5698 | /** | |
5699 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
5700 | * @pid: pid of the process. | |
5701 | * @interval: userspace pointer to the timeslice value. | |
5702 | * | |
5703 | * this syscall writes the default timeslice value of a given process | |
5704 | * into the user-space timespec buffer. A value of '0' means infinity. | |
5705 | */ | |
5706 | asmlinkage | |
5707 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
5708 | { | |
36c8b586 | 5709 | struct task_struct *p; |
a4ec24b4 | 5710 | unsigned int time_slice; |
3a5c359a | 5711 | int retval; |
1da177e4 | 5712 | struct timespec t; |
1da177e4 LT |
5713 | |
5714 | if (pid < 0) | |
3a5c359a | 5715 | return -EINVAL; |
1da177e4 LT |
5716 | |
5717 | retval = -ESRCH; | |
5718 | read_lock(&tasklist_lock); | |
5719 | p = find_process_by_pid(pid); | |
5720 | if (!p) | |
5721 | goto out_unlock; | |
5722 | ||
5723 | retval = security_task_getscheduler(p); | |
5724 | if (retval) | |
5725 | goto out_unlock; | |
5726 | ||
77034937 IM |
5727 | /* |
5728 | * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER | |
5729 | * tasks that are on an otherwise idle runqueue: | |
5730 | */ | |
5731 | time_slice = 0; | |
5732 | if (p->policy == SCHED_RR) { | |
a4ec24b4 | 5733 | time_slice = DEF_TIMESLICE; |
1868f958 | 5734 | } else if (p->policy != SCHED_FIFO) { |
a4ec24b4 DA |
5735 | struct sched_entity *se = &p->se; |
5736 | unsigned long flags; | |
5737 | struct rq *rq; | |
5738 | ||
5739 | rq = task_rq_lock(p, &flags); | |
77034937 IM |
5740 | if (rq->cfs.load.weight) |
5741 | time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
a4ec24b4 DA |
5742 | task_rq_unlock(rq, &flags); |
5743 | } | |
1da177e4 | 5744 | read_unlock(&tasklist_lock); |
a4ec24b4 | 5745 | jiffies_to_timespec(time_slice, &t); |
1da177e4 | 5746 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
1da177e4 | 5747 | return retval; |
3a5c359a | 5748 | |
1da177e4 LT |
5749 | out_unlock: |
5750 | read_unlock(&tasklist_lock); | |
5751 | return retval; | |
5752 | } | |
5753 | ||
7c731e0a | 5754 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; |
36c8b586 | 5755 | |
82a1fcb9 | 5756 | void sched_show_task(struct task_struct *p) |
1da177e4 | 5757 | { |
1da177e4 | 5758 | unsigned long free = 0; |
36c8b586 | 5759 | unsigned state; |
1da177e4 | 5760 | |
1da177e4 | 5761 | state = p->state ? __ffs(p->state) + 1 : 0; |
cc4ea795 | 5762 | printk(KERN_INFO "%-13.13s %c", p->comm, |
2ed6e34f | 5763 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); |
4bd77321 | 5764 | #if BITS_PER_LONG == 32 |
1da177e4 | 5765 | if (state == TASK_RUNNING) |
cc4ea795 | 5766 | printk(KERN_CONT " running "); |
1da177e4 | 5767 | else |
cc4ea795 | 5768 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); |
1da177e4 LT |
5769 | #else |
5770 | if (state == TASK_RUNNING) | |
cc4ea795 | 5771 | printk(KERN_CONT " running task "); |
1da177e4 | 5772 | else |
cc4ea795 | 5773 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); |
1da177e4 LT |
5774 | #endif |
5775 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
5776 | { | |
10ebffde | 5777 | unsigned long *n = end_of_stack(p); |
1da177e4 LT |
5778 | while (!*n) |
5779 | n++; | |
10ebffde | 5780 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
1da177e4 LT |
5781 | } |
5782 | #endif | |
ba25f9dc | 5783 | printk(KERN_CONT "%5lu %5d %6d\n", free, |
fcfd50af | 5784 | task_pid_nr(p), task_pid_nr(p->real_parent)); |
1da177e4 | 5785 | |
5fb5e6de | 5786 | show_stack(p, NULL); |
1da177e4 LT |
5787 | } |
5788 | ||
e59e2ae2 | 5789 | void show_state_filter(unsigned long state_filter) |
1da177e4 | 5790 | { |
36c8b586 | 5791 | struct task_struct *g, *p; |
1da177e4 | 5792 | |
4bd77321 IM |
5793 | #if BITS_PER_LONG == 32 |
5794 | printk(KERN_INFO | |
5795 | " task PC stack pid father\n"); | |
1da177e4 | 5796 | #else |
4bd77321 IM |
5797 | printk(KERN_INFO |
5798 | " task PC stack pid father\n"); | |
1da177e4 LT |
5799 | #endif |
5800 | read_lock(&tasklist_lock); | |
5801 | do_each_thread(g, p) { | |
5802 | /* | |
5803 | * reset the NMI-timeout, listing all files on a slow | |
5804 | * console might take alot of time: | |
5805 | */ | |
5806 | touch_nmi_watchdog(); | |
39bc89fd | 5807 | if (!state_filter || (p->state & state_filter)) |
82a1fcb9 | 5808 | sched_show_task(p); |
1da177e4 LT |
5809 | } while_each_thread(g, p); |
5810 | ||
04c9167f JF |
5811 | touch_all_softlockup_watchdogs(); |
5812 | ||
dd41f596 IM |
5813 | #ifdef CONFIG_SCHED_DEBUG |
5814 | sysrq_sched_debug_show(); | |
5815 | #endif | |
1da177e4 | 5816 | read_unlock(&tasklist_lock); |
e59e2ae2 IM |
5817 | /* |
5818 | * Only show locks if all tasks are dumped: | |
5819 | */ | |
5820 | if (state_filter == -1) | |
5821 | debug_show_all_locks(); | |
1da177e4 LT |
5822 | } |
5823 | ||
1df21055 IM |
5824 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) |
5825 | { | |
dd41f596 | 5826 | idle->sched_class = &idle_sched_class; |
1df21055 IM |
5827 | } |
5828 | ||
f340c0d1 IM |
5829 | /** |
5830 | * init_idle - set up an idle thread for a given CPU | |
5831 | * @idle: task in question | |
5832 | * @cpu: cpu the idle task belongs to | |
5833 | * | |
5834 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
5835 | * flag, to make booting more robust. | |
5836 | */ | |
5c1e1767 | 5837 | void __cpuinit init_idle(struct task_struct *idle, int cpu) |
1da177e4 | 5838 | { |
70b97a7f | 5839 | struct rq *rq = cpu_rq(cpu); |
1da177e4 LT |
5840 | unsigned long flags; |
5841 | ||
dd41f596 IM |
5842 | __sched_fork(idle); |
5843 | idle->se.exec_start = sched_clock(); | |
5844 | ||
b29739f9 | 5845 | idle->prio = idle->normal_prio = MAX_PRIO; |
1da177e4 | 5846 | idle->cpus_allowed = cpumask_of_cpu(cpu); |
dd41f596 | 5847 | __set_task_cpu(idle, cpu); |
1da177e4 LT |
5848 | |
5849 | spin_lock_irqsave(&rq->lock, flags); | |
5850 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
5851 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
5852 | idle->oncpu = 1; | |
5853 | #endif | |
1da177e4 LT |
5854 | spin_unlock_irqrestore(&rq->lock, flags); |
5855 | ||
5856 | /* Set the preempt count _outside_ the spinlocks! */ | |
8e3e076c LT |
5857 | #if defined(CONFIG_PREEMPT) |
5858 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); | |
5859 | #else | |
a1261f54 | 5860 | task_thread_info(idle)->preempt_count = 0; |
8e3e076c | 5861 | #endif |
dd41f596 IM |
5862 | /* |
5863 | * The idle tasks have their own, simple scheduling class: | |
5864 | */ | |
5865 | idle->sched_class = &idle_sched_class; | |
1da177e4 LT |
5866 | } |
5867 | ||
5868 | /* | |
5869 | * In a system that switches off the HZ timer nohz_cpu_mask | |
5870 | * indicates which cpus entered this state. This is used | |
5871 | * in the rcu update to wait only for active cpus. For system | |
5872 | * which do not switch off the HZ timer nohz_cpu_mask should | |
5873 | * always be CPU_MASK_NONE. | |
5874 | */ | |
5875 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
5876 | ||
19978ca6 IM |
5877 | /* |
5878 | * Increase the granularity value when there are more CPUs, | |
5879 | * because with more CPUs the 'effective latency' as visible | |
5880 | * to users decreases. But the relationship is not linear, | |
5881 | * so pick a second-best guess by going with the log2 of the | |
5882 | * number of CPUs. | |
5883 | * | |
5884 | * This idea comes from the SD scheduler of Con Kolivas: | |
5885 | */ | |
5886 | static inline void sched_init_granularity(void) | |
5887 | { | |
5888 | unsigned int factor = 1 + ilog2(num_online_cpus()); | |
5889 | const unsigned long limit = 200000000; | |
5890 | ||
5891 | sysctl_sched_min_granularity *= factor; | |
5892 | if (sysctl_sched_min_granularity > limit) | |
5893 | sysctl_sched_min_granularity = limit; | |
5894 | ||
5895 | sysctl_sched_latency *= factor; | |
5896 | if (sysctl_sched_latency > limit) | |
5897 | sysctl_sched_latency = limit; | |
5898 | ||
5899 | sysctl_sched_wakeup_granularity *= factor; | |
19978ca6 IM |
5900 | } |
5901 | ||
1da177e4 LT |
5902 | #ifdef CONFIG_SMP |
5903 | /* | |
5904 | * This is how migration works: | |
5905 | * | |
70b97a7f | 5906 | * 1) we queue a struct migration_req structure in the source CPU's |
1da177e4 LT |
5907 | * runqueue and wake up that CPU's migration thread. |
5908 | * 2) we down() the locked semaphore => thread blocks. | |
5909 | * 3) migration thread wakes up (implicitly it forces the migrated | |
5910 | * thread off the CPU) | |
5911 | * 4) it gets the migration request and checks whether the migrated | |
5912 | * task is still in the wrong runqueue. | |
5913 | * 5) if it's in the wrong runqueue then the migration thread removes | |
5914 | * it and puts it into the right queue. | |
5915 | * 6) migration thread up()s the semaphore. | |
5916 | * 7) we wake up and the migration is done. | |
5917 | */ | |
5918 | ||
5919 | /* | |
5920 | * Change a given task's CPU affinity. Migrate the thread to a | |
5921 | * proper CPU and schedule it away if the CPU it's executing on | |
5922 | * is removed from the allowed bitmask. | |
5923 | * | |
5924 | * NOTE: the caller must have a valid reference to the task, the | |
41a2d6cf | 5925 | * task must not exit() & deallocate itself prematurely. The |
1da177e4 LT |
5926 | * call is not atomic; no spinlocks may be held. |
5927 | */ | |
cd8ba7cd | 5928 | int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask) |
1da177e4 | 5929 | { |
70b97a7f | 5930 | struct migration_req req; |
1da177e4 | 5931 | unsigned long flags; |
70b97a7f | 5932 | struct rq *rq; |
48f24c4d | 5933 | int ret = 0; |
1da177e4 LT |
5934 | |
5935 | rq = task_rq_lock(p, &flags); | |
cd8ba7cd | 5936 | if (!cpus_intersects(*new_mask, cpu_online_map)) { |
1da177e4 LT |
5937 | ret = -EINVAL; |
5938 | goto out; | |
5939 | } | |
5940 | ||
73fe6aae | 5941 | if (p->sched_class->set_cpus_allowed) |
cd8ba7cd | 5942 | p->sched_class->set_cpus_allowed(p, new_mask); |
73fe6aae | 5943 | else { |
cd8ba7cd MT |
5944 | p->cpus_allowed = *new_mask; |
5945 | p->rt.nr_cpus_allowed = cpus_weight(*new_mask); | |
73fe6aae GH |
5946 | } |
5947 | ||
1da177e4 | 5948 | /* Can the task run on the task's current CPU? If so, we're done */ |
cd8ba7cd | 5949 | if (cpu_isset(task_cpu(p), *new_mask)) |
1da177e4 LT |
5950 | goto out; |
5951 | ||
cd8ba7cd | 5952 | if (migrate_task(p, any_online_cpu(*new_mask), &req)) { |
1da177e4 LT |
5953 | /* Need help from migration thread: drop lock and wait. */ |
5954 | task_rq_unlock(rq, &flags); | |
5955 | wake_up_process(rq->migration_thread); | |
5956 | wait_for_completion(&req.done); | |
5957 | tlb_migrate_finish(p->mm); | |
5958 | return 0; | |
5959 | } | |
5960 | out: | |
5961 | task_rq_unlock(rq, &flags); | |
48f24c4d | 5962 | |
1da177e4 LT |
5963 | return ret; |
5964 | } | |
cd8ba7cd | 5965 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); |
1da177e4 LT |
5966 | |
5967 | /* | |
41a2d6cf | 5968 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
1da177e4 LT |
5969 | * this because either it can't run here any more (set_cpus_allowed() |
5970 | * away from this CPU, or CPU going down), or because we're | |
5971 | * attempting to rebalance this task on exec (sched_exec). | |
5972 | * | |
5973 | * So we race with normal scheduler movements, but that's OK, as long | |
5974 | * as the task is no longer on this CPU. | |
efc30814 KK |
5975 | * |
5976 | * Returns non-zero if task was successfully migrated. | |
1da177e4 | 5977 | */ |
efc30814 | 5978 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
1da177e4 | 5979 | { |
70b97a7f | 5980 | struct rq *rq_dest, *rq_src; |
dd41f596 | 5981 | int ret = 0, on_rq; |
1da177e4 LT |
5982 | |
5983 | if (unlikely(cpu_is_offline(dest_cpu))) | |
efc30814 | 5984 | return ret; |
1da177e4 LT |
5985 | |
5986 | rq_src = cpu_rq(src_cpu); | |
5987 | rq_dest = cpu_rq(dest_cpu); | |
5988 | ||
5989 | double_rq_lock(rq_src, rq_dest); | |
5990 | /* Already moved. */ | |
5991 | if (task_cpu(p) != src_cpu) | |
5992 | goto out; | |
5993 | /* Affinity changed (again). */ | |
5994 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
5995 | goto out; | |
5996 | ||
dd41f596 | 5997 | on_rq = p->se.on_rq; |
6e82a3be | 5998 | if (on_rq) |
2e1cb74a | 5999 | deactivate_task(rq_src, p, 0); |
6e82a3be | 6000 | |
1da177e4 | 6001 | set_task_cpu(p, dest_cpu); |
dd41f596 IM |
6002 | if (on_rq) { |
6003 | activate_task(rq_dest, p, 0); | |
6004 | check_preempt_curr(rq_dest, p); | |
1da177e4 | 6005 | } |
efc30814 | 6006 | ret = 1; |
1da177e4 LT |
6007 | out: |
6008 | double_rq_unlock(rq_src, rq_dest); | |
efc30814 | 6009 | return ret; |
1da177e4 LT |
6010 | } |
6011 | ||
6012 | /* | |
6013 | * migration_thread - this is a highprio system thread that performs | |
6014 | * thread migration by bumping thread off CPU then 'pushing' onto | |
6015 | * another runqueue. | |
6016 | */ | |
95cdf3b7 | 6017 | static int migration_thread(void *data) |
1da177e4 | 6018 | { |
1da177e4 | 6019 | int cpu = (long)data; |
70b97a7f | 6020 | struct rq *rq; |
1da177e4 LT |
6021 | |
6022 | rq = cpu_rq(cpu); | |
6023 | BUG_ON(rq->migration_thread != current); | |
6024 | ||
6025 | set_current_state(TASK_INTERRUPTIBLE); | |
6026 | while (!kthread_should_stop()) { | |
70b97a7f | 6027 | struct migration_req *req; |
1da177e4 | 6028 | struct list_head *head; |
1da177e4 | 6029 | |
1da177e4 LT |
6030 | spin_lock_irq(&rq->lock); |
6031 | ||
6032 | if (cpu_is_offline(cpu)) { | |
6033 | spin_unlock_irq(&rq->lock); | |
6034 | goto wait_to_die; | |
6035 | } | |
6036 | ||
6037 | if (rq->active_balance) { | |
6038 | active_load_balance(rq, cpu); | |
6039 | rq->active_balance = 0; | |
6040 | } | |
6041 | ||
6042 | head = &rq->migration_queue; | |
6043 | ||
6044 | if (list_empty(head)) { | |
6045 | spin_unlock_irq(&rq->lock); | |
6046 | schedule(); | |
6047 | set_current_state(TASK_INTERRUPTIBLE); | |
6048 | continue; | |
6049 | } | |
70b97a7f | 6050 | req = list_entry(head->next, struct migration_req, list); |
1da177e4 LT |
6051 | list_del_init(head->next); |
6052 | ||
674311d5 NP |
6053 | spin_unlock(&rq->lock); |
6054 | __migrate_task(req->task, cpu, req->dest_cpu); | |
6055 | local_irq_enable(); | |
1da177e4 LT |
6056 | |
6057 | complete(&req->done); | |
6058 | } | |
6059 | __set_current_state(TASK_RUNNING); | |
6060 | return 0; | |
6061 | ||
6062 | wait_to_die: | |
6063 | /* Wait for kthread_stop */ | |
6064 | set_current_state(TASK_INTERRUPTIBLE); | |
6065 | while (!kthread_should_stop()) { | |
6066 | schedule(); | |
6067 | set_current_state(TASK_INTERRUPTIBLE); | |
6068 | } | |
6069 | __set_current_state(TASK_RUNNING); | |
6070 | return 0; | |
6071 | } | |
6072 | ||
6073 | #ifdef CONFIG_HOTPLUG_CPU | |
f7b4cddc ON |
6074 | |
6075 | static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu) | |
6076 | { | |
6077 | int ret; | |
6078 | ||
6079 | local_irq_disable(); | |
6080 | ret = __migrate_task(p, src_cpu, dest_cpu); | |
6081 | local_irq_enable(); | |
6082 | return ret; | |
6083 | } | |
6084 | ||
054b9108 | 6085 | /* |
3a4fa0a2 | 6086 | * Figure out where task on dead CPU should go, use force if necessary. |
054b9108 KK |
6087 | * NOTE: interrupts should be disabled by the caller |
6088 | */ | |
48f24c4d | 6089 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) |
1da177e4 | 6090 | { |
efc30814 | 6091 | unsigned long flags; |
1da177e4 | 6092 | cpumask_t mask; |
70b97a7f IM |
6093 | struct rq *rq; |
6094 | int dest_cpu; | |
1da177e4 | 6095 | |
3a5c359a AK |
6096 | do { |
6097 | /* On same node? */ | |
6098 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
6099 | cpus_and(mask, mask, p->cpus_allowed); | |
6100 | dest_cpu = any_online_cpu(mask); | |
6101 | ||
6102 | /* On any allowed CPU? */ | |
434d53b0 | 6103 | if (dest_cpu >= nr_cpu_ids) |
3a5c359a AK |
6104 | dest_cpu = any_online_cpu(p->cpus_allowed); |
6105 | ||
6106 | /* No more Mr. Nice Guy. */ | |
434d53b0 | 6107 | if (dest_cpu >= nr_cpu_ids) { |
f9a86fcb MT |
6108 | cpumask_t cpus_allowed; |
6109 | ||
6110 | cpuset_cpus_allowed_locked(p, &cpus_allowed); | |
470fd646 CW |
6111 | /* |
6112 | * Try to stay on the same cpuset, where the | |
6113 | * current cpuset may be a subset of all cpus. | |
6114 | * The cpuset_cpus_allowed_locked() variant of | |
41a2d6cf | 6115 | * cpuset_cpus_allowed() will not block. It must be |
470fd646 CW |
6116 | * called within calls to cpuset_lock/cpuset_unlock. |
6117 | */ | |
3a5c359a | 6118 | rq = task_rq_lock(p, &flags); |
470fd646 | 6119 | p->cpus_allowed = cpus_allowed; |
3a5c359a AK |
6120 | dest_cpu = any_online_cpu(p->cpus_allowed); |
6121 | task_rq_unlock(rq, &flags); | |
1da177e4 | 6122 | |
3a5c359a AK |
6123 | /* |
6124 | * Don't tell them about moving exiting tasks or | |
6125 | * kernel threads (both mm NULL), since they never | |
6126 | * leave kernel. | |
6127 | */ | |
41a2d6cf | 6128 | if (p->mm && printk_ratelimit()) { |
3a5c359a AK |
6129 | printk(KERN_INFO "process %d (%s) no " |
6130 | "longer affine to cpu%d\n", | |
41a2d6cf IM |
6131 | task_pid_nr(p), p->comm, dead_cpu); |
6132 | } | |
3a5c359a | 6133 | } |
f7b4cddc | 6134 | } while (!__migrate_task_irq(p, dead_cpu, dest_cpu)); |
1da177e4 LT |
6135 | } |
6136 | ||
6137 | /* | |
6138 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
6139 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
6140 | * for performance reasons the counter is not stricly tracking tasks to | |
6141 | * their home CPUs. So we just add the counter to another CPU's counter, | |
6142 | * to keep the global sum constant after CPU-down: | |
6143 | */ | |
70b97a7f | 6144 | static void migrate_nr_uninterruptible(struct rq *rq_src) |
1da177e4 | 6145 | { |
7c16ec58 | 6146 | struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR)); |
1da177e4 LT |
6147 | unsigned long flags; |
6148 | ||
6149 | local_irq_save(flags); | |
6150 | double_rq_lock(rq_src, rq_dest); | |
6151 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
6152 | rq_src->nr_uninterruptible = 0; | |
6153 | double_rq_unlock(rq_src, rq_dest); | |
6154 | local_irq_restore(flags); | |
6155 | } | |
6156 | ||
6157 | /* Run through task list and migrate tasks from the dead cpu. */ | |
6158 | static void migrate_live_tasks(int src_cpu) | |
6159 | { | |
48f24c4d | 6160 | struct task_struct *p, *t; |
1da177e4 | 6161 | |
f7b4cddc | 6162 | read_lock(&tasklist_lock); |
1da177e4 | 6163 | |
48f24c4d IM |
6164 | do_each_thread(t, p) { |
6165 | if (p == current) | |
1da177e4 LT |
6166 | continue; |
6167 | ||
48f24c4d IM |
6168 | if (task_cpu(p) == src_cpu) |
6169 | move_task_off_dead_cpu(src_cpu, p); | |
6170 | } while_each_thread(t, p); | |
1da177e4 | 6171 | |
f7b4cddc | 6172 | read_unlock(&tasklist_lock); |
1da177e4 LT |
6173 | } |
6174 | ||
dd41f596 IM |
6175 | /* |
6176 | * Schedules idle task to be the next runnable task on current CPU. | |
94bc9a7b DA |
6177 | * It does so by boosting its priority to highest possible. |
6178 | * Used by CPU offline code. | |
1da177e4 LT |
6179 | */ |
6180 | void sched_idle_next(void) | |
6181 | { | |
48f24c4d | 6182 | int this_cpu = smp_processor_id(); |
70b97a7f | 6183 | struct rq *rq = cpu_rq(this_cpu); |
1da177e4 LT |
6184 | struct task_struct *p = rq->idle; |
6185 | unsigned long flags; | |
6186 | ||
6187 | /* cpu has to be offline */ | |
48f24c4d | 6188 | BUG_ON(cpu_online(this_cpu)); |
1da177e4 | 6189 | |
48f24c4d IM |
6190 | /* |
6191 | * Strictly not necessary since rest of the CPUs are stopped by now | |
6192 | * and interrupts disabled on the current cpu. | |
1da177e4 LT |
6193 | */ |
6194 | spin_lock_irqsave(&rq->lock, flags); | |
6195 | ||
dd41f596 | 6196 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
48f24c4d | 6197 | |
94bc9a7b DA |
6198 | update_rq_clock(rq); |
6199 | activate_task(rq, p, 0); | |
1da177e4 LT |
6200 | |
6201 | spin_unlock_irqrestore(&rq->lock, flags); | |
6202 | } | |
6203 | ||
48f24c4d IM |
6204 | /* |
6205 | * Ensures that the idle task is using init_mm right before its cpu goes | |
1da177e4 LT |
6206 | * offline. |
6207 | */ | |
6208 | void idle_task_exit(void) | |
6209 | { | |
6210 | struct mm_struct *mm = current->active_mm; | |
6211 | ||
6212 | BUG_ON(cpu_online(smp_processor_id())); | |
6213 | ||
6214 | if (mm != &init_mm) | |
6215 | switch_mm(mm, &init_mm, current); | |
6216 | mmdrop(mm); | |
6217 | } | |
6218 | ||
054b9108 | 6219 | /* called under rq->lock with disabled interrupts */ |
36c8b586 | 6220 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) |
1da177e4 | 6221 | { |
70b97a7f | 6222 | struct rq *rq = cpu_rq(dead_cpu); |
1da177e4 LT |
6223 | |
6224 | /* Must be exiting, otherwise would be on tasklist. */ | |
270f722d | 6225 | BUG_ON(!p->exit_state); |
1da177e4 LT |
6226 | |
6227 | /* Cannot have done final schedule yet: would have vanished. */ | |
c394cc9f | 6228 | BUG_ON(p->state == TASK_DEAD); |
1da177e4 | 6229 | |
48f24c4d | 6230 | get_task_struct(p); |
1da177e4 LT |
6231 | |
6232 | /* | |
6233 | * Drop lock around migration; if someone else moves it, | |
41a2d6cf | 6234 | * that's OK. No task can be added to this CPU, so iteration is |
1da177e4 LT |
6235 | * fine. |
6236 | */ | |
f7b4cddc | 6237 | spin_unlock_irq(&rq->lock); |
48f24c4d | 6238 | move_task_off_dead_cpu(dead_cpu, p); |
f7b4cddc | 6239 | spin_lock_irq(&rq->lock); |
1da177e4 | 6240 | |
48f24c4d | 6241 | put_task_struct(p); |
1da177e4 LT |
6242 | } |
6243 | ||
6244 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
6245 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
6246 | { | |
70b97a7f | 6247 | struct rq *rq = cpu_rq(dead_cpu); |
dd41f596 | 6248 | struct task_struct *next; |
48f24c4d | 6249 | |
dd41f596 IM |
6250 | for ( ; ; ) { |
6251 | if (!rq->nr_running) | |
6252 | break; | |
a8e504d2 | 6253 | update_rq_clock(rq); |
ff95f3df | 6254 | next = pick_next_task(rq, rq->curr); |
dd41f596 IM |
6255 | if (!next) |
6256 | break; | |
6257 | migrate_dead(dead_cpu, next); | |
e692ab53 | 6258 | |
1da177e4 LT |
6259 | } |
6260 | } | |
6261 | #endif /* CONFIG_HOTPLUG_CPU */ | |
6262 | ||
e692ab53 NP |
6263 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) |
6264 | ||
6265 | static struct ctl_table sd_ctl_dir[] = { | |
e0361851 AD |
6266 | { |
6267 | .procname = "sched_domain", | |
c57baf1e | 6268 | .mode = 0555, |
e0361851 | 6269 | }, |
38605cae | 6270 | {0, }, |
e692ab53 NP |
6271 | }; |
6272 | ||
6273 | static struct ctl_table sd_ctl_root[] = { | |
e0361851 | 6274 | { |
c57baf1e | 6275 | .ctl_name = CTL_KERN, |
e0361851 | 6276 | .procname = "kernel", |
c57baf1e | 6277 | .mode = 0555, |
e0361851 AD |
6278 | .child = sd_ctl_dir, |
6279 | }, | |
38605cae | 6280 | {0, }, |
e692ab53 NP |
6281 | }; |
6282 | ||
6283 | static struct ctl_table *sd_alloc_ctl_entry(int n) | |
6284 | { | |
6285 | struct ctl_table *entry = | |
5cf9f062 | 6286 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); |
e692ab53 | 6287 | |
e692ab53 NP |
6288 | return entry; |
6289 | } | |
6290 | ||
6382bc90 MM |
6291 | static void sd_free_ctl_entry(struct ctl_table **tablep) |
6292 | { | |
cd790076 | 6293 | struct ctl_table *entry; |
6382bc90 | 6294 | |
cd790076 MM |
6295 | /* |
6296 | * In the intermediate directories, both the child directory and | |
6297 | * procname are dynamically allocated and could fail but the mode | |
41a2d6cf | 6298 | * will always be set. In the lowest directory the names are |
cd790076 MM |
6299 | * static strings and all have proc handlers. |
6300 | */ | |
6301 | for (entry = *tablep; entry->mode; entry++) { | |
6382bc90 MM |
6302 | if (entry->child) |
6303 | sd_free_ctl_entry(&entry->child); | |
cd790076 MM |
6304 | if (entry->proc_handler == NULL) |
6305 | kfree(entry->procname); | |
6306 | } | |
6382bc90 MM |
6307 | |
6308 | kfree(*tablep); | |
6309 | *tablep = NULL; | |
6310 | } | |
6311 | ||
e692ab53 | 6312 | static void |
e0361851 | 6313 | set_table_entry(struct ctl_table *entry, |
e692ab53 NP |
6314 | const char *procname, void *data, int maxlen, |
6315 | mode_t mode, proc_handler *proc_handler) | |
6316 | { | |
e692ab53 NP |
6317 | entry->procname = procname; |
6318 | entry->data = data; | |
6319 | entry->maxlen = maxlen; | |
6320 | entry->mode = mode; | |
6321 | entry->proc_handler = proc_handler; | |
6322 | } | |
6323 | ||
6324 | static struct ctl_table * | |
6325 | sd_alloc_ctl_domain_table(struct sched_domain *sd) | |
6326 | { | |
ace8b3d6 | 6327 | struct ctl_table *table = sd_alloc_ctl_entry(12); |
e692ab53 | 6328 | |
ad1cdc1d MM |
6329 | if (table == NULL) |
6330 | return NULL; | |
6331 | ||
e0361851 | 6332 | set_table_entry(&table[0], "min_interval", &sd->min_interval, |
e692ab53 | 6333 | sizeof(long), 0644, proc_doulongvec_minmax); |
e0361851 | 6334 | set_table_entry(&table[1], "max_interval", &sd->max_interval, |
e692ab53 | 6335 | sizeof(long), 0644, proc_doulongvec_minmax); |
e0361851 | 6336 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, |
e692ab53 | 6337 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6338 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, |
e692ab53 | 6339 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6340 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, |
e692ab53 | 6341 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6342 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, |
e692ab53 | 6343 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6344 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, |
e692ab53 | 6345 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6346 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, |
e692ab53 | 6347 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6348 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, |
e692ab53 | 6349 | sizeof(int), 0644, proc_dointvec_minmax); |
ace8b3d6 | 6350 | set_table_entry(&table[9], "cache_nice_tries", |
e692ab53 NP |
6351 | &sd->cache_nice_tries, |
6352 | sizeof(int), 0644, proc_dointvec_minmax); | |
ace8b3d6 | 6353 | set_table_entry(&table[10], "flags", &sd->flags, |
e692ab53 | 6354 | sizeof(int), 0644, proc_dointvec_minmax); |
6323469f | 6355 | /* &table[11] is terminator */ |
e692ab53 NP |
6356 | |
6357 | return table; | |
6358 | } | |
6359 | ||
9a4e7159 | 6360 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) |
e692ab53 NP |
6361 | { |
6362 | struct ctl_table *entry, *table; | |
6363 | struct sched_domain *sd; | |
6364 | int domain_num = 0, i; | |
6365 | char buf[32]; | |
6366 | ||
6367 | for_each_domain(cpu, sd) | |
6368 | domain_num++; | |
6369 | entry = table = sd_alloc_ctl_entry(domain_num + 1); | |
ad1cdc1d MM |
6370 | if (table == NULL) |
6371 | return NULL; | |
e692ab53 NP |
6372 | |
6373 | i = 0; | |
6374 | for_each_domain(cpu, sd) { | |
6375 | snprintf(buf, 32, "domain%d", i); | |
e692ab53 | 6376 | entry->procname = kstrdup(buf, GFP_KERNEL); |
c57baf1e | 6377 | entry->mode = 0555; |
e692ab53 NP |
6378 | entry->child = sd_alloc_ctl_domain_table(sd); |
6379 | entry++; | |
6380 | i++; | |
6381 | } | |
6382 | return table; | |
6383 | } | |
6384 | ||
6385 | static struct ctl_table_header *sd_sysctl_header; | |
6382bc90 | 6386 | static void register_sched_domain_sysctl(void) |
e692ab53 NP |
6387 | { |
6388 | int i, cpu_num = num_online_cpus(); | |
6389 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); | |
6390 | char buf[32]; | |
6391 | ||
7378547f MM |
6392 | WARN_ON(sd_ctl_dir[0].child); |
6393 | sd_ctl_dir[0].child = entry; | |
6394 | ||
ad1cdc1d MM |
6395 | if (entry == NULL) |
6396 | return; | |
6397 | ||
97b6ea7b | 6398 | for_each_online_cpu(i) { |
e692ab53 | 6399 | snprintf(buf, 32, "cpu%d", i); |
e692ab53 | 6400 | entry->procname = kstrdup(buf, GFP_KERNEL); |
c57baf1e | 6401 | entry->mode = 0555; |
e692ab53 | 6402 | entry->child = sd_alloc_ctl_cpu_table(i); |
97b6ea7b | 6403 | entry++; |
e692ab53 | 6404 | } |
7378547f MM |
6405 | |
6406 | WARN_ON(sd_sysctl_header); | |
e692ab53 NP |
6407 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); |
6408 | } | |
6382bc90 | 6409 | |
7378547f | 6410 | /* may be called multiple times per register */ |
6382bc90 MM |
6411 | static void unregister_sched_domain_sysctl(void) |
6412 | { | |
7378547f MM |
6413 | if (sd_sysctl_header) |
6414 | unregister_sysctl_table(sd_sysctl_header); | |
6382bc90 | 6415 | sd_sysctl_header = NULL; |
7378547f MM |
6416 | if (sd_ctl_dir[0].child) |
6417 | sd_free_ctl_entry(&sd_ctl_dir[0].child); | |
6382bc90 | 6418 | } |
e692ab53 | 6419 | #else |
6382bc90 MM |
6420 | static void register_sched_domain_sysctl(void) |
6421 | { | |
6422 | } | |
6423 | static void unregister_sched_domain_sysctl(void) | |
e692ab53 NP |
6424 | { |
6425 | } | |
6426 | #endif | |
6427 | ||
1da177e4 LT |
6428 | /* |
6429 | * migration_call - callback that gets triggered when a CPU is added. | |
6430 | * Here we can start up the necessary migration thread for the new CPU. | |
6431 | */ | |
48f24c4d IM |
6432 | static int __cpuinit |
6433 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
1da177e4 | 6434 | { |
1da177e4 | 6435 | struct task_struct *p; |
48f24c4d | 6436 | int cpu = (long)hcpu; |
1da177e4 | 6437 | unsigned long flags; |
70b97a7f | 6438 | struct rq *rq; |
1da177e4 LT |
6439 | |
6440 | switch (action) { | |
5be9361c | 6441 | |
1da177e4 | 6442 | case CPU_UP_PREPARE: |
8bb78442 | 6443 | case CPU_UP_PREPARE_FROZEN: |
dd41f596 | 6444 | p = kthread_create(migration_thread, hcpu, "migration/%d", cpu); |
1da177e4 LT |
6445 | if (IS_ERR(p)) |
6446 | return NOTIFY_BAD; | |
1da177e4 LT |
6447 | kthread_bind(p, cpu); |
6448 | /* Must be high prio: stop_machine expects to yield to it. */ | |
6449 | rq = task_rq_lock(p, &flags); | |
dd41f596 | 6450 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
1da177e4 LT |
6451 | task_rq_unlock(rq, &flags); |
6452 | cpu_rq(cpu)->migration_thread = p; | |
6453 | break; | |
48f24c4d | 6454 | |
1da177e4 | 6455 | case CPU_ONLINE: |
8bb78442 | 6456 | case CPU_ONLINE_FROZEN: |
3a4fa0a2 | 6457 | /* Strictly unnecessary, as first user will wake it. */ |
1da177e4 | 6458 | wake_up_process(cpu_rq(cpu)->migration_thread); |
1f94ef59 GH |
6459 | |
6460 | /* Update our root-domain */ | |
6461 | rq = cpu_rq(cpu); | |
6462 | spin_lock_irqsave(&rq->lock, flags); | |
6463 | if (rq->rd) { | |
6464 | BUG_ON(!cpu_isset(cpu, rq->rd->span)); | |
6465 | cpu_set(cpu, rq->rd->online); | |
6466 | } | |
6467 | spin_unlock_irqrestore(&rq->lock, flags); | |
1da177e4 | 6468 | break; |
48f24c4d | 6469 | |
1da177e4 LT |
6470 | #ifdef CONFIG_HOTPLUG_CPU |
6471 | case CPU_UP_CANCELED: | |
8bb78442 | 6472 | case CPU_UP_CANCELED_FROZEN: |
fc75cdfa HC |
6473 | if (!cpu_rq(cpu)->migration_thread) |
6474 | break; | |
41a2d6cf | 6475 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
a4c4af7c HC |
6476 | kthread_bind(cpu_rq(cpu)->migration_thread, |
6477 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
6478 | kthread_stop(cpu_rq(cpu)->migration_thread); |
6479 | cpu_rq(cpu)->migration_thread = NULL; | |
6480 | break; | |
48f24c4d | 6481 | |
1da177e4 | 6482 | case CPU_DEAD: |
8bb78442 | 6483 | case CPU_DEAD_FROZEN: |
470fd646 | 6484 | cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */ |
1da177e4 LT |
6485 | migrate_live_tasks(cpu); |
6486 | rq = cpu_rq(cpu); | |
6487 | kthread_stop(rq->migration_thread); | |
6488 | rq->migration_thread = NULL; | |
6489 | /* Idle task back to normal (off runqueue, low prio) */ | |
d2da272a | 6490 | spin_lock_irq(&rq->lock); |
a8e504d2 | 6491 | update_rq_clock(rq); |
2e1cb74a | 6492 | deactivate_task(rq, rq->idle, 0); |
1da177e4 | 6493 | rq->idle->static_prio = MAX_PRIO; |
dd41f596 IM |
6494 | __setscheduler(rq, rq->idle, SCHED_NORMAL, 0); |
6495 | rq->idle->sched_class = &idle_sched_class; | |
1da177e4 | 6496 | migrate_dead_tasks(cpu); |
d2da272a | 6497 | spin_unlock_irq(&rq->lock); |
470fd646 | 6498 | cpuset_unlock(); |
1da177e4 LT |
6499 | migrate_nr_uninterruptible(rq); |
6500 | BUG_ON(rq->nr_running != 0); | |
6501 | ||
41a2d6cf IM |
6502 | /* |
6503 | * No need to migrate the tasks: it was best-effort if | |
6504 | * they didn't take sched_hotcpu_mutex. Just wake up | |
6505 | * the requestors. | |
6506 | */ | |
1da177e4 LT |
6507 | spin_lock_irq(&rq->lock); |
6508 | while (!list_empty(&rq->migration_queue)) { | |
70b97a7f IM |
6509 | struct migration_req *req; |
6510 | ||
1da177e4 | 6511 | req = list_entry(rq->migration_queue.next, |
70b97a7f | 6512 | struct migration_req, list); |
1da177e4 LT |
6513 | list_del_init(&req->list); |
6514 | complete(&req->done); | |
6515 | } | |
6516 | spin_unlock_irq(&rq->lock); | |
6517 | break; | |
57d885fe | 6518 | |
08f503b0 GH |
6519 | case CPU_DYING: |
6520 | case CPU_DYING_FROZEN: | |
57d885fe GH |
6521 | /* Update our root-domain */ |
6522 | rq = cpu_rq(cpu); | |
6523 | spin_lock_irqsave(&rq->lock, flags); | |
6524 | if (rq->rd) { | |
6525 | BUG_ON(!cpu_isset(cpu, rq->rd->span)); | |
6526 | cpu_clear(cpu, rq->rd->online); | |
6527 | } | |
6528 | spin_unlock_irqrestore(&rq->lock, flags); | |
6529 | break; | |
1da177e4 LT |
6530 | #endif |
6531 | } | |
6532 | return NOTIFY_OK; | |
6533 | } | |
6534 | ||
6535 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
6536 | * happens before everything else. | |
6537 | */ | |
26c2143b | 6538 | static struct notifier_block __cpuinitdata migration_notifier = { |
1da177e4 LT |
6539 | .notifier_call = migration_call, |
6540 | .priority = 10 | |
6541 | }; | |
6542 | ||
e6fe6649 | 6543 | void __init migration_init(void) |
1da177e4 LT |
6544 | { |
6545 | void *cpu = (void *)(long)smp_processor_id(); | |
07dccf33 | 6546 | int err; |
48f24c4d IM |
6547 | |
6548 | /* Start one for the boot CPU: */ | |
07dccf33 AM |
6549 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
6550 | BUG_ON(err == NOTIFY_BAD); | |
1da177e4 LT |
6551 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
6552 | register_cpu_notifier(&migration_notifier); | |
1da177e4 LT |
6553 | } |
6554 | #endif | |
6555 | ||
6556 | #ifdef CONFIG_SMP | |
476f3534 | 6557 | |
3e9830dc | 6558 | #ifdef CONFIG_SCHED_DEBUG |
4dcf6aff | 6559 | |
7c16ec58 MT |
6560 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, |
6561 | cpumask_t *groupmask) | |
1da177e4 | 6562 | { |
4dcf6aff | 6563 | struct sched_group *group = sd->groups; |
434d53b0 | 6564 | char str[256]; |
1da177e4 | 6565 | |
434d53b0 | 6566 | cpulist_scnprintf(str, sizeof(str), sd->span); |
7c16ec58 | 6567 | cpus_clear(*groupmask); |
4dcf6aff IM |
6568 | |
6569 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); | |
6570 | ||
6571 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
6572 | printk("does not load-balance\n"); | |
6573 | if (sd->parent) | |
6574 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" | |
6575 | " has parent"); | |
6576 | return -1; | |
41c7ce9a NP |
6577 | } |
6578 | ||
4dcf6aff IM |
6579 | printk(KERN_CONT "span %s\n", str); |
6580 | ||
6581 | if (!cpu_isset(cpu, sd->span)) { | |
6582 | printk(KERN_ERR "ERROR: domain->span does not contain " | |
6583 | "CPU%d\n", cpu); | |
6584 | } | |
6585 | if (!cpu_isset(cpu, group->cpumask)) { | |
6586 | printk(KERN_ERR "ERROR: domain->groups does not contain" | |
6587 | " CPU%d\n", cpu); | |
6588 | } | |
1da177e4 | 6589 | |
4dcf6aff | 6590 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); |
1da177e4 | 6591 | do { |
4dcf6aff IM |
6592 | if (!group) { |
6593 | printk("\n"); | |
6594 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
1da177e4 LT |
6595 | break; |
6596 | } | |
6597 | ||
4dcf6aff IM |
6598 | if (!group->__cpu_power) { |
6599 | printk(KERN_CONT "\n"); | |
6600 | printk(KERN_ERR "ERROR: domain->cpu_power not " | |
6601 | "set\n"); | |
6602 | break; | |
6603 | } | |
1da177e4 | 6604 | |
4dcf6aff IM |
6605 | if (!cpus_weight(group->cpumask)) { |
6606 | printk(KERN_CONT "\n"); | |
6607 | printk(KERN_ERR "ERROR: empty group\n"); | |
6608 | break; | |
6609 | } | |
1da177e4 | 6610 | |
7c16ec58 | 6611 | if (cpus_intersects(*groupmask, group->cpumask)) { |
4dcf6aff IM |
6612 | printk(KERN_CONT "\n"); |
6613 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
6614 | break; | |
6615 | } | |
1da177e4 | 6616 | |
7c16ec58 | 6617 | cpus_or(*groupmask, *groupmask, group->cpumask); |
1da177e4 | 6618 | |
434d53b0 | 6619 | cpulist_scnprintf(str, sizeof(str), group->cpumask); |
4dcf6aff | 6620 | printk(KERN_CONT " %s", str); |
1da177e4 | 6621 | |
4dcf6aff IM |
6622 | group = group->next; |
6623 | } while (group != sd->groups); | |
6624 | printk(KERN_CONT "\n"); | |
1da177e4 | 6625 | |
7c16ec58 | 6626 | if (!cpus_equal(sd->span, *groupmask)) |
4dcf6aff | 6627 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
1da177e4 | 6628 | |
7c16ec58 | 6629 | if (sd->parent && !cpus_subset(*groupmask, sd->parent->span)) |
4dcf6aff IM |
6630 | printk(KERN_ERR "ERROR: parent span is not a superset " |
6631 | "of domain->span\n"); | |
6632 | return 0; | |
6633 | } | |
1da177e4 | 6634 | |
4dcf6aff IM |
6635 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
6636 | { | |
7c16ec58 | 6637 | cpumask_t *groupmask; |
4dcf6aff | 6638 | int level = 0; |
1da177e4 | 6639 | |
4dcf6aff IM |
6640 | if (!sd) { |
6641 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
6642 | return; | |
6643 | } | |
1da177e4 | 6644 | |
4dcf6aff IM |
6645 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
6646 | ||
7c16ec58 MT |
6647 | groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL); |
6648 | if (!groupmask) { | |
6649 | printk(KERN_DEBUG "Cannot load-balance (out of memory)\n"); | |
6650 | return; | |
6651 | } | |
6652 | ||
4dcf6aff | 6653 | for (;;) { |
7c16ec58 | 6654 | if (sched_domain_debug_one(sd, cpu, level, groupmask)) |
4dcf6aff | 6655 | break; |
1da177e4 LT |
6656 | level++; |
6657 | sd = sd->parent; | |
33859f7f | 6658 | if (!sd) |
4dcf6aff IM |
6659 | break; |
6660 | } | |
7c16ec58 | 6661 | kfree(groupmask); |
1da177e4 LT |
6662 | } |
6663 | #else | |
48f24c4d | 6664 | # define sched_domain_debug(sd, cpu) do { } while (0) |
1da177e4 LT |
6665 | #endif |
6666 | ||
1a20ff27 | 6667 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
6668 | { |
6669 | if (cpus_weight(sd->span) == 1) | |
6670 | return 1; | |
6671 | ||
6672 | /* Following flags need at least 2 groups */ | |
6673 | if (sd->flags & (SD_LOAD_BALANCE | | |
6674 | SD_BALANCE_NEWIDLE | | |
6675 | SD_BALANCE_FORK | | |
89c4710e SS |
6676 | SD_BALANCE_EXEC | |
6677 | SD_SHARE_CPUPOWER | | |
6678 | SD_SHARE_PKG_RESOURCES)) { | |
245af2c7 SS |
6679 | if (sd->groups != sd->groups->next) |
6680 | return 0; | |
6681 | } | |
6682 | ||
6683 | /* Following flags don't use groups */ | |
6684 | if (sd->flags & (SD_WAKE_IDLE | | |
6685 | SD_WAKE_AFFINE | | |
6686 | SD_WAKE_BALANCE)) | |
6687 | return 0; | |
6688 | ||
6689 | return 1; | |
6690 | } | |
6691 | ||
48f24c4d IM |
6692 | static int |
6693 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
245af2c7 SS |
6694 | { |
6695 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
6696 | ||
6697 | if (sd_degenerate(parent)) | |
6698 | return 1; | |
6699 | ||
6700 | if (!cpus_equal(sd->span, parent->span)) | |
6701 | return 0; | |
6702 | ||
6703 | /* Does parent contain flags not in child? */ | |
6704 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
6705 | if (cflags & SD_WAKE_AFFINE) | |
6706 | pflags &= ~SD_WAKE_BALANCE; | |
6707 | /* Flags needing groups don't count if only 1 group in parent */ | |
6708 | if (parent->groups == parent->groups->next) { | |
6709 | pflags &= ~(SD_LOAD_BALANCE | | |
6710 | SD_BALANCE_NEWIDLE | | |
6711 | SD_BALANCE_FORK | | |
89c4710e SS |
6712 | SD_BALANCE_EXEC | |
6713 | SD_SHARE_CPUPOWER | | |
6714 | SD_SHARE_PKG_RESOURCES); | |
245af2c7 SS |
6715 | } |
6716 | if (~cflags & pflags) | |
6717 | return 0; | |
6718 | ||
6719 | return 1; | |
6720 | } | |
6721 | ||
57d885fe GH |
6722 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) |
6723 | { | |
6724 | unsigned long flags; | |
6725 | const struct sched_class *class; | |
6726 | ||
6727 | spin_lock_irqsave(&rq->lock, flags); | |
6728 | ||
6729 | if (rq->rd) { | |
6730 | struct root_domain *old_rd = rq->rd; | |
6731 | ||
0eab9146 | 6732 | for (class = sched_class_highest; class; class = class->next) { |
57d885fe GH |
6733 | if (class->leave_domain) |
6734 | class->leave_domain(rq); | |
0eab9146 | 6735 | } |
57d885fe | 6736 | |
dc938520 GH |
6737 | cpu_clear(rq->cpu, old_rd->span); |
6738 | cpu_clear(rq->cpu, old_rd->online); | |
6739 | ||
57d885fe GH |
6740 | if (atomic_dec_and_test(&old_rd->refcount)) |
6741 | kfree(old_rd); | |
6742 | } | |
6743 | ||
6744 | atomic_inc(&rd->refcount); | |
6745 | rq->rd = rd; | |
6746 | ||
dc938520 | 6747 | cpu_set(rq->cpu, rd->span); |
1f94ef59 GH |
6748 | if (cpu_isset(rq->cpu, cpu_online_map)) |
6749 | cpu_set(rq->cpu, rd->online); | |
dc938520 | 6750 | |
0eab9146 | 6751 | for (class = sched_class_highest; class; class = class->next) { |
57d885fe GH |
6752 | if (class->join_domain) |
6753 | class->join_domain(rq); | |
0eab9146 | 6754 | } |
57d885fe GH |
6755 | |
6756 | spin_unlock_irqrestore(&rq->lock, flags); | |
6757 | } | |
6758 | ||
dc938520 | 6759 | static void init_rootdomain(struct root_domain *rd) |
57d885fe GH |
6760 | { |
6761 | memset(rd, 0, sizeof(*rd)); | |
6762 | ||
dc938520 GH |
6763 | cpus_clear(rd->span); |
6764 | cpus_clear(rd->online); | |
57d885fe GH |
6765 | } |
6766 | ||
6767 | static void init_defrootdomain(void) | |
6768 | { | |
dc938520 | 6769 | init_rootdomain(&def_root_domain); |
57d885fe GH |
6770 | atomic_set(&def_root_domain.refcount, 1); |
6771 | } | |
6772 | ||
dc938520 | 6773 | static struct root_domain *alloc_rootdomain(void) |
57d885fe GH |
6774 | { |
6775 | struct root_domain *rd; | |
6776 | ||
6777 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); | |
6778 | if (!rd) | |
6779 | return NULL; | |
6780 | ||
dc938520 | 6781 | init_rootdomain(rd); |
57d885fe GH |
6782 | |
6783 | return rd; | |
6784 | } | |
6785 | ||
1da177e4 | 6786 | /* |
0eab9146 | 6787 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
1da177e4 LT |
6788 | * hold the hotplug lock. |
6789 | */ | |
0eab9146 IM |
6790 | static void |
6791 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) | |
1da177e4 | 6792 | { |
70b97a7f | 6793 | struct rq *rq = cpu_rq(cpu); |
245af2c7 SS |
6794 | struct sched_domain *tmp; |
6795 | ||
6796 | /* Remove the sched domains which do not contribute to scheduling. */ | |
6797 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
6798 | struct sched_domain *parent = tmp->parent; | |
6799 | if (!parent) | |
6800 | break; | |
1a848870 | 6801 | if (sd_parent_degenerate(tmp, parent)) { |
245af2c7 | 6802 | tmp->parent = parent->parent; |
1a848870 SS |
6803 | if (parent->parent) |
6804 | parent->parent->child = tmp; | |
6805 | } | |
245af2c7 SS |
6806 | } |
6807 | ||
1a848870 | 6808 | if (sd && sd_degenerate(sd)) { |
245af2c7 | 6809 | sd = sd->parent; |
1a848870 SS |
6810 | if (sd) |
6811 | sd->child = NULL; | |
6812 | } | |
1da177e4 LT |
6813 | |
6814 | sched_domain_debug(sd, cpu); | |
6815 | ||
57d885fe | 6816 | rq_attach_root(rq, rd); |
674311d5 | 6817 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
6818 | } |
6819 | ||
6820 | /* cpus with isolated domains */ | |
67af63a6 | 6821 | static cpumask_t cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
6822 | |
6823 | /* Setup the mask of cpus configured for isolated domains */ | |
6824 | static int __init isolated_cpu_setup(char *str) | |
6825 | { | |
6826 | int ints[NR_CPUS], i; | |
6827 | ||
6828 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
6829 | cpus_clear(cpu_isolated_map); | |
6830 | for (i = 1; i <= ints[0]; i++) | |
6831 | if (ints[i] < NR_CPUS) | |
6832 | cpu_set(ints[i], cpu_isolated_map); | |
6833 | return 1; | |
6834 | } | |
6835 | ||
8927f494 | 6836 | __setup("isolcpus=", isolated_cpu_setup); |
1da177e4 LT |
6837 | |
6838 | /* | |
6711cab4 SS |
6839 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer |
6840 | * to a function which identifies what group(along with sched group) a CPU | |
6841 | * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS | |
6842 | * (due to the fact that we keep track of groups covered with a cpumask_t). | |
1da177e4 LT |
6843 | * |
6844 | * init_sched_build_groups will build a circular linked list of the groups | |
6845 | * covered by the given span, and will set each group's ->cpumask correctly, | |
6846 | * and ->cpu_power to 0. | |
6847 | */ | |
a616058b | 6848 | static void |
7c16ec58 | 6849 | init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map, |
6711cab4 | 6850 | int (*group_fn)(int cpu, const cpumask_t *cpu_map, |
7c16ec58 MT |
6851 | struct sched_group **sg, |
6852 | cpumask_t *tmpmask), | |
6853 | cpumask_t *covered, cpumask_t *tmpmask) | |
1da177e4 LT |
6854 | { |
6855 | struct sched_group *first = NULL, *last = NULL; | |
1da177e4 LT |
6856 | int i; |
6857 | ||
7c16ec58 MT |
6858 | cpus_clear(*covered); |
6859 | ||
6860 | for_each_cpu_mask(i, *span) { | |
6711cab4 | 6861 | struct sched_group *sg; |
7c16ec58 | 6862 | int group = group_fn(i, cpu_map, &sg, tmpmask); |
1da177e4 LT |
6863 | int j; |
6864 | ||
7c16ec58 | 6865 | if (cpu_isset(i, *covered)) |
1da177e4 LT |
6866 | continue; |
6867 | ||
7c16ec58 | 6868 | cpus_clear(sg->cpumask); |
5517d86b | 6869 | sg->__cpu_power = 0; |
1da177e4 | 6870 | |
7c16ec58 MT |
6871 | for_each_cpu_mask(j, *span) { |
6872 | if (group_fn(j, cpu_map, NULL, tmpmask) != group) | |
1da177e4 LT |
6873 | continue; |
6874 | ||
7c16ec58 | 6875 | cpu_set(j, *covered); |
1da177e4 LT |
6876 | cpu_set(j, sg->cpumask); |
6877 | } | |
6878 | if (!first) | |
6879 | first = sg; | |
6880 | if (last) | |
6881 | last->next = sg; | |
6882 | last = sg; | |
6883 | } | |
6884 | last->next = first; | |
6885 | } | |
6886 | ||
9c1cfda2 | 6887 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 6888 | |
9c1cfda2 | 6889 | #ifdef CONFIG_NUMA |
198e2f18 | 6890 | |
9c1cfda2 JH |
6891 | /** |
6892 | * find_next_best_node - find the next node to include in a sched_domain | |
6893 | * @node: node whose sched_domain we're building | |
6894 | * @used_nodes: nodes already in the sched_domain | |
6895 | * | |
41a2d6cf | 6896 | * Find the next node to include in a given scheduling domain. Simply |
9c1cfda2 JH |
6897 | * finds the closest node not already in the @used_nodes map. |
6898 | * | |
6899 | * Should use nodemask_t. | |
6900 | */ | |
c5f59f08 | 6901 | static int find_next_best_node(int node, nodemask_t *used_nodes) |
9c1cfda2 JH |
6902 | { |
6903 | int i, n, val, min_val, best_node = 0; | |
6904 | ||
6905 | min_val = INT_MAX; | |
6906 | ||
6907 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6908 | /* Start at @node */ | |
6909 | n = (node + i) % MAX_NUMNODES; | |
6910 | ||
6911 | if (!nr_cpus_node(n)) | |
6912 | continue; | |
6913 | ||
6914 | /* Skip already used nodes */ | |
c5f59f08 | 6915 | if (node_isset(n, *used_nodes)) |
9c1cfda2 JH |
6916 | continue; |
6917 | ||
6918 | /* Simple min distance search */ | |
6919 | val = node_distance(node, n); | |
6920 | ||
6921 | if (val < min_val) { | |
6922 | min_val = val; | |
6923 | best_node = n; | |
6924 | } | |
6925 | } | |
6926 | ||
c5f59f08 | 6927 | node_set(best_node, *used_nodes); |
9c1cfda2 JH |
6928 | return best_node; |
6929 | } | |
6930 | ||
6931 | /** | |
6932 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
6933 | * @node: node whose cpumask we're constructing | |
73486722 | 6934 | * @span: resulting cpumask |
9c1cfda2 | 6935 | * |
41a2d6cf | 6936 | * Given a node, construct a good cpumask for its sched_domain to span. It |
9c1cfda2 JH |
6937 | * should be one that prevents unnecessary balancing, but also spreads tasks |
6938 | * out optimally. | |
6939 | */ | |
4bdbaad3 | 6940 | static void sched_domain_node_span(int node, cpumask_t *span) |
9c1cfda2 | 6941 | { |
c5f59f08 | 6942 | nodemask_t used_nodes; |
c5f59f08 | 6943 | node_to_cpumask_ptr(nodemask, node); |
48f24c4d | 6944 | int i; |
9c1cfda2 | 6945 | |
4bdbaad3 | 6946 | cpus_clear(*span); |
c5f59f08 | 6947 | nodes_clear(used_nodes); |
9c1cfda2 | 6948 | |
4bdbaad3 | 6949 | cpus_or(*span, *span, *nodemask); |
c5f59f08 | 6950 | node_set(node, used_nodes); |
9c1cfda2 JH |
6951 | |
6952 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
c5f59f08 | 6953 | int next_node = find_next_best_node(node, &used_nodes); |
48f24c4d | 6954 | |
c5f59f08 | 6955 | node_to_cpumask_ptr_next(nodemask, next_node); |
4bdbaad3 | 6956 | cpus_or(*span, *span, *nodemask); |
9c1cfda2 | 6957 | } |
9c1cfda2 JH |
6958 | } |
6959 | #endif | |
6960 | ||
5c45bf27 | 6961 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
48f24c4d | 6962 | |
9c1cfda2 | 6963 | /* |
48f24c4d | 6964 | * SMT sched-domains: |
9c1cfda2 | 6965 | */ |
1da177e4 LT |
6966 | #ifdef CONFIG_SCHED_SMT |
6967 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
6711cab4 | 6968 | static DEFINE_PER_CPU(struct sched_group, sched_group_cpus); |
48f24c4d | 6969 | |
41a2d6cf | 6970 | static int |
7c16ec58 MT |
6971 | cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
6972 | cpumask_t *unused) | |
1da177e4 | 6973 | { |
6711cab4 SS |
6974 | if (sg) |
6975 | *sg = &per_cpu(sched_group_cpus, cpu); | |
1da177e4 LT |
6976 | return cpu; |
6977 | } | |
6978 | #endif | |
6979 | ||
48f24c4d IM |
6980 | /* |
6981 | * multi-core sched-domains: | |
6982 | */ | |
1e9f28fa SS |
6983 | #ifdef CONFIG_SCHED_MC |
6984 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
6711cab4 | 6985 | static DEFINE_PER_CPU(struct sched_group, sched_group_core); |
1e9f28fa SS |
6986 | #endif |
6987 | ||
6988 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
41a2d6cf | 6989 | static int |
7c16ec58 MT |
6990 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
6991 | cpumask_t *mask) | |
1e9f28fa | 6992 | { |
6711cab4 | 6993 | int group; |
7c16ec58 MT |
6994 | |
6995 | *mask = per_cpu(cpu_sibling_map, cpu); | |
6996 | cpus_and(*mask, *mask, *cpu_map); | |
6997 | group = first_cpu(*mask); | |
6711cab4 SS |
6998 | if (sg) |
6999 | *sg = &per_cpu(sched_group_core, group); | |
7000 | return group; | |
1e9f28fa SS |
7001 | } |
7002 | #elif defined(CONFIG_SCHED_MC) | |
41a2d6cf | 7003 | static int |
7c16ec58 MT |
7004 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
7005 | cpumask_t *unused) | |
1e9f28fa | 7006 | { |
6711cab4 SS |
7007 | if (sg) |
7008 | *sg = &per_cpu(sched_group_core, cpu); | |
1e9f28fa SS |
7009 | return cpu; |
7010 | } | |
7011 | #endif | |
7012 | ||
1da177e4 | 7013 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
6711cab4 | 7014 | static DEFINE_PER_CPU(struct sched_group, sched_group_phys); |
48f24c4d | 7015 | |
41a2d6cf | 7016 | static int |
7c16ec58 MT |
7017 | cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
7018 | cpumask_t *mask) | |
1da177e4 | 7019 | { |
6711cab4 | 7020 | int group; |
48f24c4d | 7021 | #ifdef CONFIG_SCHED_MC |
7c16ec58 MT |
7022 | *mask = cpu_coregroup_map(cpu); |
7023 | cpus_and(*mask, *mask, *cpu_map); | |
7024 | group = first_cpu(*mask); | |
1e9f28fa | 7025 | #elif defined(CONFIG_SCHED_SMT) |
7c16ec58 MT |
7026 | *mask = per_cpu(cpu_sibling_map, cpu); |
7027 | cpus_and(*mask, *mask, *cpu_map); | |
7028 | group = first_cpu(*mask); | |
1da177e4 | 7029 | #else |
6711cab4 | 7030 | group = cpu; |
1da177e4 | 7031 | #endif |
6711cab4 SS |
7032 | if (sg) |
7033 | *sg = &per_cpu(sched_group_phys, group); | |
7034 | return group; | |
1da177e4 LT |
7035 | } |
7036 | ||
7037 | #ifdef CONFIG_NUMA | |
1da177e4 | 7038 | /* |
9c1cfda2 JH |
7039 | * The init_sched_build_groups can't handle what we want to do with node |
7040 | * groups, so roll our own. Now each node has its own list of groups which | |
7041 | * gets dynamically allocated. | |
1da177e4 | 7042 | */ |
9c1cfda2 | 7043 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
434d53b0 | 7044 | static struct sched_group ***sched_group_nodes_bycpu; |
1da177e4 | 7045 | |
9c1cfda2 | 7046 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
6711cab4 | 7047 | static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes); |
9c1cfda2 | 7048 | |
6711cab4 | 7049 | static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map, |
7c16ec58 | 7050 | struct sched_group **sg, cpumask_t *nodemask) |
9c1cfda2 | 7051 | { |
6711cab4 SS |
7052 | int group; |
7053 | ||
7c16ec58 MT |
7054 | *nodemask = node_to_cpumask(cpu_to_node(cpu)); |
7055 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7056 | group = first_cpu(*nodemask); | |
6711cab4 SS |
7057 | |
7058 | if (sg) | |
7059 | *sg = &per_cpu(sched_group_allnodes, group); | |
7060 | return group; | |
1da177e4 | 7061 | } |
6711cab4 | 7062 | |
08069033 SS |
7063 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
7064 | { | |
7065 | struct sched_group *sg = group_head; | |
7066 | int j; | |
7067 | ||
7068 | if (!sg) | |
7069 | return; | |
3a5c359a AK |
7070 | do { |
7071 | for_each_cpu_mask(j, sg->cpumask) { | |
7072 | struct sched_domain *sd; | |
08069033 | 7073 | |
3a5c359a AK |
7074 | sd = &per_cpu(phys_domains, j); |
7075 | if (j != first_cpu(sd->groups->cpumask)) { | |
7076 | /* | |
7077 | * Only add "power" once for each | |
7078 | * physical package. | |
7079 | */ | |
7080 | continue; | |
7081 | } | |
08069033 | 7082 | |
3a5c359a AK |
7083 | sg_inc_cpu_power(sg, sd->groups->__cpu_power); |
7084 | } | |
7085 | sg = sg->next; | |
7086 | } while (sg != group_head); | |
08069033 | 7087 | } |
1da177e4 LT |
7088 | #endif |
7089 | ||
a616058b | 7090 | #ifdef CONFIG_NUMA |
51888ca2 | 7091 | /* Free memory allocated for various sched_group structures */ |
7c16ec58 | 7092 | static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask) |
51888ca2 | 7093 | { |
a616058b | 7094 | int cpu, i; |
51888ca2 SV |
7095 | |
7096 | for_each_cpu_mask(cpu, *cpu_map) { | |
51888ca2 SV |
7097 | struct sched_group **sched_group_nodes |
7098 | = sched_group_nodes_bycpu[cpu]; | |
7099 | ||
51888ca2 SV |
7100 | if (!sched_group_nodes) |
7101 | continue; | |
7102 | ||
7103 | for (i = 0; i < MAX_NUMNODES; i++) { | |
51888ca2 SV |
7104 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; |
7105 | ||
7c16ec58 MT |
7106 | *nodemask = node_to_cpumask(i); |
7107 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7108 | if (cpus_empty(*nodemask)) | |
51888ca2 SV |
7109 | continue; |
7110 | ||
7111 | if (sg == NULL) | |
7112 | continue; | |
7113 | sg = sg->next; | |
7114 | next_sg: | |
7115 | oldsg = sg; | |
7116 | sg = sg->next; | |
7117 | kfree(oldsg); | |
7118 | if (oldsg != sched_group_nodes[i]) | |
7119 | goto next_sg; | |
7120 | } | |
7121 | kfree(sched_group_nodes); | |
7122 | sched_group_nodes_bycpu[cpu] = NULL; | |
7123 | } | |
51888ca2 | 7124 | } |
a616058b | 7125 | #else |
7c16ec58 | 7126 | static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask) |
a616058b SS |
7127 | { |
7128 | } | |
7129 | #endif | |
51888ca2 | 7130 | |
89c4710e SS |
7131 | /* |
7132 | * Initialize sched groups cpu_power. | |
7133 | * | |
7134 | * cpu_power indicates the capacity of sched group, which is used while | |
7135 | * distributing the load between different sched groups in a sched domain. | |
7136 | * Typically cpu_power for all the groups in a sched domain will be same unless | |
7137 | * there are asymmetries in the topology. If there are asymmetries, group | |
7138 | * having more cpu_power will pickup more load compared to the group having | |
7139 | * less cpu_power. | |
7140 | * | |
7141 | * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents | |
7142 | * the maximum number of tasks a group can handle in the presence of other idle | |
7143 | * or lightly loaded groups in the same sched domain. | |
7144 | */ | |
7145 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) | |
7146 | { | |
7147 | struct sched_domain *child; | |
7148 | struct sched_group *group; | |
7149 | ||
7150 | WARN_ON(!sd || !sd->groups); | |
7151 | ||
7152 | if (cpu != first_cpu(sd->groups->cpumask)) | |
7153 | return; | |
7154 | ||
7155 | child = sd->child; | |
7156 | ||
5517d86b ED |
7157 | sd->groups->__cpu_power = 0; |
7158 | ||
89c4710e SS |
7159 | /* |
7160 | * For perf policy, if the groups in child domain share resources | |
7161 | * (for example cores sharing some portions of the cache hierarchy | |
7162 | * or SMT), then set this domain groups cpu_power such that each group | |
7163 | * can handle only one task, when there are other idle groups in the | |
7164 | * same sched domain. | |
7165 | */ | |
7166 | if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) && | |
7167 | (child->flags & | |
7168 | (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) { | |
5517d86b | 7169 | sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE); |
89c4710e SS |
7170 | return; |
7171 | } | |
7172 | ||
89c4710e SS |
7173 | /* |
7174 | * add cpu_power of each child group to this groups cpu_power | |
7175 | */ | |
7176 | group = child->groups; | |
7177 | do { | |
5517d86b | 7178 | sg_inc_cpu_power(sd->groups, group->__cpu_power); |
89c4710e SS |
7179 | group = group->next; |
7180 | } while (group != child->groups); | |
7181 | } | |
7182 | ||
7c16ec58 MT |
7183 | /* |
7184 | * Initializers for schedule domains | |
7185 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() | |
7186 | */ | |
7187 | ||
7188 | #define SD_INIT(sd, type) sd_init_##type(sd) | |
7189 | #define SD_INIT_FUNC(type) \ | |
7190 | static noinline void sd_init_##type(struct sched_domain *sd) \ | |
7191 | { \ | |
7192 | memset(sd, 0, sizeof(*sd)); \ | |
7193 | *sd = SD_##type##_INIT; \ | |
1d3504fc | 7194 | sd->level = SD_LV_##type; \ |
7c16ec58 MT |
7195 | } |
7196 | ||
7197 | SD_INIT_FUNC(CPU) | |
7198 | #ifdef CONFIG_NUMA | |
7199 | SD_INIT_FUNC(ALLNODES) | |
7200 | SD_INIT_FUNC(NODE) | |
7201 | #endif | |
7202 | #ifdef CONFIG_SCHED_SMT | |
7203 | SD_INIT_FUNC(SIBLING) | |
7204 | #endif | |
7205 | #ifdef CONFIG_SCHED_MC | |
7206 | SD_INIT_FUNC(MC) | |
7207 | #endif | |
7208 | ||
7209 | /* | |
7210 | * To minimize stack usage kmalloc room for cpumasks and share the | |
7211 | * space as the usage in build_sched_domains() dictates. Used only | |
7212 | * if the amount of space is significant. | |
7213 | */ | |
7214 | struct allmasks { | |
7215 | cpumask_t tmpmask; /* make this one first */ | |
7216 | union { | |
7217 | cpumask_t nodemask; | |
7218 | cpumask_t this_sibling_map; | |
7219 | cpumask_t this_core_map; | |
7220 | }; | |
7221 | cpumask_t send_covered; | |
7222 | ||
7223 | #ifdef CONFIG_NUMA | |
7224 | cpumask_t domainspan; | |
7225 | cpumask_t covered; | |
7226 | cpumask_t notcovered; | |
7227 | #endif | |
7228 | }; | |
7229 | ||
7230 | #if NR_CPUS > 128 | |
7231 | #define SCHED_CPUMASK_ALLOC 1 | |
7232 | #define SCHED_CPUMASK_FREE(v) kfree(v) | |
7233 | #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v | |
7234 | #else | |
7235 | #define SCHED_CPUMASK_ALLOC 0 | |
7236 | #define SCHED_CPUMASK_FREE(v) | |
7237 | #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v | |
7238 | #endif | |
7239 | ||
7240 | #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \ | |
7241 | ((unsigned long)(a) + offsetof(struct allmasks, v)) | |
7242 | ||
1d3504fc HS |
7243 | static int default_relax_domain_level = -1; |
7244 | ||
7245 | static int __init setup_relax_domain_level(char *str) | |
7246 | { | |
7247 | default_relax_domain_level = simple_strtoul(str, NULL, 0); | |
7248 | return 1; | |
7249 | } | |
7250 | __setup("relax_domain_level=", setup_relax_domain_level); | |
7251 | ||
7252 | static void set_domain_attribute(struct sched_domain *sd, | |
7253 | struct sched_domain_attr *attr) | |
7254 | { | |
7255 | int request; | |
7256 | ||
7257 | if (!attr || attr->relax_domain_level < 0) { | |
7258 | if (default_relax_domain_level < 0) | |
7259 | return; | |
7260 | else | |
7261 | request = default_relax_domain_level; | |
7262 | } else | |
7263 | request = attr->relax_domain_level; | |
7264 | if (request < sd->level) { | |
7265 | /* turn off idle balance on this domain */ | |
7266 | sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE); | |
7267 | } else { | |
7268 | /* turn on idle balance on this domain */ | |
7269 | sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE); | |
7270 | } | |
7271 | } | |
7272 | ||
1da177e4 | 7273 | /* |
1a20ff27 DG |
7274 | * Build sched domains for a given set of cpus and attach the sched domains |
7275 | * to the individual cpus | |
1da177e4 | 7276 | */ |
1d3504fc HS |
7277 | static int __build_sched_domains(const cpumask_t *cpu_map, |
7278 | struct sched_domain_attr *attr) | |
1da177e4 LT |
7279 | { |
7280 | int i; | |
57d885fe | 7281 | struct root_domain *rd; |
7c16ec58 MT |
7282 | SCHED_CPUMASK_DECLARE(allmasks); |
7283 | cpumask_t *tmpmask; | |
d1b55138 JH |
7284 | #ifdef CONFIG_NUMA |
7285 | struct sched_group **sched_group_nodes = NULL; | |
6711cab4 | 7286 | int sd_allnodes = 0; |
d1b55138 JH |
7287 | |
7288 | /* | |
7289 | * Allocate the per-node list of sched groups | |
7290 | */ | |
5cf9f062 | 7291 | sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *), |
41a2d6cf | 7292 | GFP_KERNEL); |
d1b55138 JH |
7293 | if (!sched_group_nodes) { |
7294 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
51888ca2 | 7295 | return -ENOMEM; |
d1b55138 | 7296 | } |
d1b55138 | 7297 | #endif |
1da177e4 | 7298 | |
dc938520 | 7299 | rd = alloc_rootdomain(); |
57d885fe GH |
7300 | if (!rd) { |
7301 | printk(KERN_WARNING "Cannot alloc root domain\n"); | |
7c16ec58 MT |
7302 | #ifdef CONFIG_NUMA |
7303 | kfree(sched_group_nodes); | |
7304 | #endif | |
57d885fe GH |
7305 | return -ENOMEM; |
7306 | } | |
7307 | ||
7c16ec58 MT |
7308 | #if SCHED_CPUMASK_ALLOC |
7309 | /* get space for all scratch cpumask variables */ | |
7310 | allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL); | |
7311 | if (!allmasks) { | |
7312 | printk(KERN_WARNING "Cannot alloc cpumask array\n"); | |
7313 | kfree(rd); | |
7314 | #ifdef CONFIG_NUMA | |
7315 | kfree(sched_group_nodes); | |
7316 | #endif | |
7317 | return -ENOMEM; | |
7318 | } | |
7319 | #endif | |
7320 | tmpmask = (cpumask_t *)allmasks; | |
7321 | ||
7322 | ||
7323 | #ifdef CONFIG_NUMA | |
7324 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
7325 | #endif | |
7326 | ||
1da177e4 | 7327 | /* |
1a20ff27 | 7328 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 7329 | */ |
1a20ff27 | 7330 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 7331 | struct sched_domain *sd = NULL, *p; |
7c16ec58 | 7332 | SCHED_CPUMASK_VAR(nodemask, allmasks); |
1da177e4 | 7333 | |
7c16ec58 MT |
7334 | *nodemask = node_to_cpumask(cpu_to_node(i)); |
7335 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
1da177e4 LT |
7336 | |
7337 | #ifdef CONFIG_NUMA | |
dd41f596 | 7338 | if (cpus_weight(*cpu_map) > |
7c16ec58 | 7339 | SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) { |
9c1cfda2 | 7340 | sd = &per_cpu(allnodes_domains, i); |
7c16ec58 | 7341 | SD_INIT(sd, ALLNODES); |
1d3504fc | 7342 | set_domain_attribute(sd, attr); |
9c1cfda2 | 7343 | sd->span = *cpu_map; |
18d95a28 | 7344 | sd->first_cpu = first_cpu(sd->span); |
7c16ec58 | 7345 | cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask); |
9c1cfda2 | 7346 | p = sd; |
6711cab4 | 7347 | sd_allnodes = 1; |
9c1cfda2 JH |
7348 | } else |
7349 | p = NULL; | |
7350 | ||
1da177e4 | 7351 | sd = &per_cpu(node_domains, i); |
7c16ec58 | 7352 | SD_INIT(sd, NODE); |
1d3504fc | 7353 | set_domain_attribute(sd, attr); |
4bdbaad3 | 7354 | sched_domain_node_span(cpu_to_node(i), &sd->span); |
18d95a28 | 7355 | sd->first_cpu = first_cpu(sd->span); |
9c1cfda2 | 7356 | sd->parent = p; |
1a848870 SS |
7357 | if (p) |
7358 | p->child = sd; | |
9c1cfda2 | 7359 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
7360 | #endif |
7361 | ||
7362 | p = sd; | |
7363 | sd = &per_cpu(phys_domains, i); | |
7c16ec58 | 7364 | SD_INIT(sd, CPU); |
1d3504fc | 7365 | set_domain_attribute(sd, attr); |
7c16ec58 | 7366 | sd->span = *nodemask; |
18d95a28 | 7367 | sd->first_cpu = first_cpu(sd->span); |
1da177e4 | 7368 | sd->parent = p; |
1a848870 SS |
7369 | if (p) |
7370 | p->child = sd; | |
7c16ec58 | 7371 | cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask); |
1da177e4 | 7372 | |
1e9f28fa SS |
7373 | #ifdef CONFIG_SCHED_MC |
7374 | p = sd; | |
7375 | sd = &per_cpu(core_domains, i); | |
7c16ec58 | 7376 | SD_INIT(sd, MC); |
1d3504fc | 7377 | set_domain_attribute(sd, attr); |
1e9f28fa | 7378 | sd->span = cpu_coregroup_map(i); |
18d95a28 | 7379 | sd->first_cpu = first_cpu(sd->span); |
1e9f28fa SS |
7380 | cpus_and(sd->span, sd->span, *cpu_map); |
7381 | sd->parent = p; | |
1a848870 | 7382 | p->child = sd; |
7c16ec58 | 7383 | cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask); |
1e9f28fa SS |
7384 | #endif |
7385 | ||
1da177e4 LT |
7386 | #ifdef CONFIG_SCHED_SMT |
7387 | p = sd; | |
7388 | sd = &per_cpu(cpu_domains, i); | |
7c16ec58 | 7389 | SD_INIT(sd, SIBLING); |
1d3504fc | 7390 | set_domain_attribute(sd, attr); |
d5a7430d | 7391 | sd->span = per_cpu(cpu_sibling_map, i); |
18d95a28 | 7392 | sd->first_cpu = first_cpu(sd->span); |
1a20ff27 | 7393 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 | 7394 | sd->parent = p; |
1a848870 | 7395 | p->child = sd; |
7c16ec58 | 7396 | cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask); |
1da177e4 LT |
7397 | #endif |
7398 | } | |
7399 | ||
7400 | #ifdef CONFIG_SCHED_SMT | |
7401 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 7402 | for_each_cpu_mask(i, *cpu_map) { |
7c16ec58 MT |
7403 | SCHED_CPUMASK_VAR(this_sibling_map, allmasks); |
7404 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7405 | ||
7406 | *this_sibling_map = per_cpu(cpu_sibling_map, i); | |
7407 | cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map); | |
7408 | if (i != first_cpu(*this_sibling_map)) | |
1da177e4 LT |
7409 | continue; |
7410 | ||
dd41f596 | 7411 | init_sched_build_groups(this_sibling_map, cpu_map, |
7c16ec58 MT |
7412 | &cpu_to_cpu_group, |
7413 | send_covered, tmpmask); | |
1da177e4 LT |
7414 | } |
7415 | #endif | |
7416 | ||
1e9f28fa SS |
7417 | #ifdef CONFIG_SCHED_MC |
7418 | /* Set up multi-core groups */ | |
7419 | for_each_cpu_mask(i, *cpu_map) { | |
7c16ec58 MT |
7420 | SCHED_CPUMASK_VAR(this_core_map, allmasks); |
7421 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7422 | ||
7423 | *this_core_map = cpu_coregroup_map(i); | |
7424 | cpus_and(*this_core_map, *this_core_map, *cpu_map); | |
7425 | if (i != first_cpu(*this_core_map)) | |
1e9f28fa | 7426 | continue; |
7c16ec58 | 7427 | |
dd41f596 | 7428 | init_sched_build_groups(this_core_map, cpu_map, |
7c16ec58 MT |
7429 | &cpu_to_core_group, |
7430 | send_covered, tmpmask); | |
1e9f28fa SS |
7431 | } |
7432 | #endif | |
7433 | ||
1da177e4 LT |
7434 | /* Set up physical groups */ |
7435 | for (i = 0; i < MAX_NUMNODES; i++) { | |
7c16ec58 MT |
7436 | SCHED_CPUMASK_VAR(nodemask, allmasks); |
7437 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
1da177e4 | 7438 | |
7c16ec58 MT |
7439 | *nodemask = node_to_cpumask(i); |
7440 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7441 | if (cpus_empty(*nodemask)) | |
1da177e4 LT |
7442 | continue; |
7443 | ||
7c16ec58 MT |
7444 | init_sched_build_groups(nodemask, cpu_map, |
7445 | &cpu_to_phys_group, | |
7446 | send_covered, tmpmask); | |
1da177e4 LT |
7447 | } |
7448 | ||
7449 | #ifdef CONFIG_NUMA | |
7450 | /* Set up node groups */ | |
7c16ec58 MT |
7451 | if (sd_allnodes) { |
7452 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7453 | ||
7454 | init_sched_build_groups(cpu_map, cpu_map, | |
7455 | &cpu_to_allnodes_group, | |
7456 | send_covered, tmpmask); | |
7457 | } | |
9c1cfda2 JH |
7458 | |
7459 | for (i = 0; i < MAX_NUMNODES; i++) { | |
7460 | /* Set up node groups */ | |
7461 | struct sched_group *sg, *prev; | |
7c16ec58 MT |
7462 | SCHED_CPUMASK_VAR(nodemask, allmasks); |
7463 | SCHED_CPUMASK_VAR(domainspan, allmasks); | |
7464 | SCHED_CPUMASK_VAR(covered, allmasks); | |
9c1cfda2 JH |
7465 | int j; |
7466 | ||
7c16ec58 MT |
7467 | *nodemask = node_to_cpumask(i); |
7468 | cpus_clear(*covered); | |
7469 | ||
7470 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7471 | if (cpus_empty(*nodemask)) { | |
d1b55138 | 7472 | sched_group_nodes[i] = NULL; |
9c1cfda2 | 7473 | continue; |
d1b55138 | 7474 | } |
9c1cfda2 | 7475 | |
4bdbaad3 | 7476 | sched_domain_node_span(i, domainspan); |
7c16ec58 | 7477 | cpus_and(*domainspan, *domainspan, *cpu_map); |
9c1cfda2 | 7478 | |
15f0b676 | 7479 | sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i); |
51888ca2 SV |
7480 | if (!sg) { |
7481 | printk(KERN_WARNING "Can not alloc domain group for " | |
7482 | "node %d\n", i); | |
7483 | goto error; | |
7484 | } | |
9c1cfda2 | 7485 | sched_group_nodes[i] = sg; |
7c16ec58 | 7486 | for_each_cpu_mask(j, *nodemask) { |
9c1cfda2 | 7487 | struct sched_domain *sd; |
9761eea8 | 7488 | |
9c1cfda2 JH |
7489 | sd = &per_cpu(node_domains, j); |
7490 | sd->groups = sg; | |
9c1cfda2 | 7491 | } |
5517d86b | 7492 | sg->__cpu_power = 0; |
7c16ec58 | 7493 | sg->cpumask = *nodemask; |
51888ca2 | 7494 | sg->next = sg; |
7c16ec58 | 7495 | cpus_or(*covered, *covered, *nodemask); |
9c1cfda2 JH |
7496 | prev = sg; |
7497 | ||
7498 | for (j = 0; j < MAX_NUMNODES; j++) { | |
7c16ec58 | 7499 | SCHED_CPUMASK_VAR(notcovered, allmasks); |
9c1cfda2 | 7500 | int n = (i + j) % MAX_NUMNODES; |
c5f59f08 | 7501 | node_to_cpumask_ptr(pnodemask, n); |
9c1cfda2 | 7502 | |
7c16ec58 MT |
7503 | cpus_complement(*notcovered, *covered); |
7504 | cpus_and(*tmpmask, *notcovered, *cpu_map); | |
7505 | cpus_and(*tmpmask, *tmpmask, *domainspan); | |
7506 | if (cpus_empty(*tmpmask)) | |
9c1cfda2 JH |
7507 | break; |
7508 | ||
7c16ec58 MT |
7509 | cpus_and(*tmpmask, *tmpmask, *pnodemask); |
7510 | if (cpus_empty(*tmpmask)) | |
9c1cfda2 JH |
7511 | continue; |
7512 | ||
15f0b676 SV |
7513 | sg = kmalloc_node(sizeof(struct sched_group), |
7514 | GFP_KERNEL, i); | |
9c1cfda2 JH |
7515 | if (!sg) { |
7516 | printk(KERN_WARNING | |
7517 | "Can not alloc domain group for node %d\n", j); | |
51888ca2 | 7518 | goto error; |
9c1cfda2 | 7519 | } |
5517d86b | 7520 | sg->__cpu_power = 0; |
7c16ec58 | 7521 | sg->cpumask = *tmpmask; |
51888ca2 | 7522 | sg->next = prev->next; |
7c16ec58 | 7523 | cpus_or(*covered, *covered, *tmpmask); |
9c1cfda2 JH |
7524 | prev->next = sg; |
7525 | prev = sg; | |
7526 | } | |
9c1cfda2 | 7527 | } |
1da177e4 LT |
7528 | #endif |
7529 | ||
7530 | /* Calculate CPU power for physical packages and nodes */ | |
5c45bf27 | 7531 | #ifdef CONFIG_SCHED_SMT |
1a20ff27 | 7532 | for_each_cpu_mask(i, *cpu_map) { |
dd41f596 IM |
7533 | struct sched_domain *sd = &per_cpu(cpu_domains, i); |
7534 | ||
89c4710e | 7535 | init_sched_groups_power(i, sd); |
5c45bf27 | 7536 | } |
1da177e4 | 7537 | #endif |
1e9f28fa | 7538 | #ifdef CONFIG_SCHED_MC |
5c45bf27 | 7539 | for_each_cpu_mask(i, *cpu_map) { |
dd41f596 IM |
7540 | struct sched_domain *sd = &per_cpu(core_domains, i); |
7541 | ||
89c4710e | 7542 | init_sched_groups_power(i, sd); |
5c45bf27 SS |
7543 | } |
7544 | #endif | |
1e9f28fa | 7545 | |
5c45bf27 | 7546 | for_each_cpu_mask(i, *cpu_map) { |
dd41f596 IM |
7547 | struct sched_domain *sd = &per_cpu(phys_domains, i); |
7548 | ||
89c4710e | 7549 | init_sched_groups_power(i, sd); |
1da177e4 LT |
7550 | } |
7551 | ||
9c1cfda2 | 7552 | #ifdef CONFIG_NUMA |
08069033 SS |
7553 | for (i = 0; i < MAX_NUMNODES; i++) |
7554 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
9c1cfda2 | 7555 | |
6711cab4 SS |
7556 | if (sd_allnodes) { |
7557 | struct sched_group *sg; | |
f712c0c7 | 7558 | |
7c16ec58 MT |
7559 | cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg, |
7560 | tmpmask); | |
f712c0c7 SS |
7561 | init_numa_sched_groups_power(sg); |
7562 | } | |
9c1cfda2 JH |
7563 | #endif |
7564 | ||
1da177e4 | 7565 | /* Attach the domains */ |
1a20ff27 | 7566 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
7567 | struct sched_domain *sd; |
7568 | #ifdef CONFIG_SCHED_SMT | |
7569 | sd = &per_cpu(cpu_domains, i); | |
1e9f28fa SS |
7570 | #elif defined(CONFIG_SCHED_MC) |
7571 | sd = &per_cpu(core_domains, i); | |
1da177e4 LT |
7572 | #else |
7573 | sd = &per_cpu(phys_domains, i); | |
7574 | #endif | |
57d885fe | 7575 | cpu_attach_domain(sd, rd, i); |
1da177e4 | 7576 | } |
51888ca2 | 7577 | |
7c16ec58 | 7578 | SCHED_CPUMASK_FREE((void *)allmasks); |
51888ca2 SV |
7579 | return 0; |
7580 | ||
a616058b | 7581 | #ifdef CONFIG_NUMA |
51888ca2 | 7582 | error: |
7c16ec58 MT |
7583 | free_sched_groups(cpu_map, tmpmask); |
7584 | SCHED_CPUMASK_FREE((void *)allmasks); | |
51888ca2 | 7585 | return -ENOMEM; |
a616058b | 7586 | #endif |
1da177e4 | 7587 | } |
029190c5 | 7588 | |
1d3504fc HS |
7589 | static int build_sched_domains(const cpumask_t *cpu_map) |
7590 | { | |
7591 | return __build_sched_domains(cpu_map, NULL); | |
7592 | } | |
7593 | ||
029190c5 PJ |
7594 | static cpumask_t *doms_cur; /* current sched domains */ |
7595 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ | |
1d3504fc HS |
7596 | static struct sched_domain_attr *dattr_cur; /* attribues of custom domains |
7597 | in 'doms_cur' */ | |
029190c5 PJ |
7598 | |
7599 | /* | |
7600 | * Special case: If a kmalloc of a doms_cur partition (array of | |
7601 | * cpumask_t) fails, then fallback to a single sched domain, | |
7602 | * as determined by the single cpumask_t fallback_doms. | |
7603 | */ | |
7604 | static cpumask_t fallback_doms; | |
7605 | ||
22e52b07 HC |
7606 | void __attribute__((weak)) arch_update_cpu_topology(void) |
7607 | { | |
7608 | } | |
7609 | ||
1a20ff27 | 7610 | /* |
41a2d6cf | 7611 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
029190c5 PJ |
7612 | * For now this just excludes isolated cpus, but could be used to |
7613 | * exclude other special cases in the future. | |
1a20ff27 | 7614 | */ |
51888ca2 | 7615 | static int arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 | 7616 | { |
7378547f MM |
7617 | int err; |
7618 | ||
22e52b07 | 7619 | arch_update_cpu_topology(); |
029190c5 PJ |
7620 | ndoms_cur = 1; |
7621 | doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL); | |
7622 | if (!doms_cur) | |
7623 | doms_cur = &fallback_doms; | |
7624 | cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map); | |
1d3504fc | 7625 | dattr_cur = NULL; |
7378547f | 7626 | err = build_sched_domains(doms_cur); |
6382bc90 | 7627 | register_sched_domain_sysctl(); |
7378547f MM |
7628 | |
7629 | return err; | |
1a20ff27 DG |
7630 | } |
7631 | ||
7c16ec58 MT |
7632 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map, |
7633 | cpumask_t *tmpmask) | |
1da177e4 | 7634 | { |
7c16ec58 | 7635 | free_sched_groups(cpu_map, tmpmask); |
9c1cfda2 | 7636 | } |
1da177e4 | 7637 | |
1a20ff27 DG |
7638 | /* |
7639 | * Detach sched domains from a group of cpus specified in cpu_map | |
7640 | * These cpus will now be attached to the NULL domain | |
7641 | */ | |
858119e1 | 7642 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
1a20ff27 | 7643 | { |
7c16ec58 | 7644 | cpumask_t tmpmask; |
1a20ff27 DG |
7645 | int i; |
7646 | ||
6382bc90 MM |
7647 | unregister_sched_domain_sysctl(); |
7648 | ||
1a20ff27 | 7649 | for_each_cpu_mask(i, *cpu_map) |
57d885fe | 7650 | cpu_attach_domain(NULL, &def_root_domain, i); |
1a20ff27 | 7651 | synchronize_sched(); |
7c16ec58 | 7652 | arch_destroy_sched_domains(cpu_map, &tmpmask); |
1a20ff27 DG |
7653 | } |
7654 | ||
1d3504fc HS |
7655 | /* handle null as "default" */ |
7656 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, | |
7657 | struct sched_domain_attr *new, int idx_new) | |
7658 | { | |
7659 | struct sched_domain_attr tmp; | |
7660 | ||
7661 | /* fast path */ | |
7662 | if (!new && !cur) | |
7663 | return 1; | |
7664 | ||
7665 | tmp = SD_ATTR_INIT; | |
7666 | return !memcmp(cur ? (cur + idx_cur) : &tmp, | |
7667 | new ? (new + idx_new) : &tmp, | |
7668 | sizeof(struct sched_domain_attr)); | |
7669 | } | |
7670 | ||
029190c5 PJ |
7671 | /* |
7672 | * Partition sched domains as specified by the 'ndoms_new' | |
41a2d6cf | 7673 | * cpumasks in the array doms_new[] of cpumasks. This compares |
029190c5 PJ |
7674 | * doms_new[] to the current sched domain partitioning, doms_cur[]. |
7675 | * It destroys each deleted domain and builds each new domain. | |
7676 | * | |
7677 | * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'. | |
41a2d6cf IM |
7678 | * The masks don't intersect (don't overlap.) We should setup one |
7679 | * sched domain for each mask. CPUs not in any of the cpumasks will | |
7680 | * not be load balanced. If the same cpumask appears both in the | |
029190c5 PJ |
7681 | * current 'doms_cur' domains and in the new 'doms_new', we can leave |
7682 | * it as it is. | |
7683 | * | |
41a2d6cf IM |
7684 | * The passed in 'doms_new' should be kmalloc'd. This routine takes |
7685 | * ownership of it and will kfree it when done with it. If the caller | |
029190c5 PJ |
7686 | * failed the kmalloc call, then it can pass in doms_new == NULL, |
7687 | * and partition_sched_domains() will fallback to the single partition | |
7688 | * 'fallback_doms'. | |
7689 | * | |
7690 | * Call with hotplug lock held | |
7691 | */ | |
1d3504fc HS |
7692 | void partition_sched_domains(int ndoms_new, cpumask_t *doms_new, |
7693 | struct sched_domain_attr *dattr_new) | |
029190c5 PJ |
7694 | { |
7695 | int i, j; | |
7696 | ||
712555ee | 7697 | mutex_lock(&sched_domains_mutex); |
a1835615 | 7698 | |
7378547f MM |
7699 | /* always unregister in case we don't destroy any domains */ |
7700 | unregister_sched_domain_sysctl(); | |
7701 | ||
029190c5 PJ |
7702 | if (doms_new == NULL) { |
7703 | ndoms_new = 1; | |
7704 | doms_new = &fallback_doms; | |
7705 | cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map); | |
1d3504fc | 7706 | dattr_new = NULL; |
029190c5 PJ |
7707 | } |
7708 | ||
7709 | /* Destroy deleted domains */ | |
7710 | for (i = 0; i < ndoms_cur; i++) { | |
7711 | for (j = 0; j < ndoms_new; j++) { | |
1d3504fc HS |
7712 | if (cpus_equal(doms_cur[i], doms_new[j]) |
7713 | && dattrs_equal(dattr_cur, i, dattr_new, j)) | |
029190c5 PJ |
7714 | goto match1; |
7715 | } | |
7716 | /* no match - a current sched domain not in new doms_new[] */ | |
7717 | detach_destroy_domains(doms_cur + i); | |
7718 | match1: | |
7719 | ; | |
7720 | } | |
7721 | ||
7722 | /* Build new domains */ | |
7723 | for (i = 0; i < ndoms_new; i++) { | |
7724 | for (j = 0; j < ndoms_cur; j++) { | |
1d3504fc HS |
7725 | if (cpus_equal(doms_new[i], doms_cur[j]) |
7726 | && dattrs_equal(dattr_new, i, dattr_cur, j)) | |
029190c5 PJ |
7727 | goto match2; |
7728 | } | |
7729 | /* no match - add a new doms_new */ | |
1d3504fc HS |
7730 | __build_sched_domains(doms_new + i, |
7731 | dattr_new ? dattr_new + i : NULL); | |
029190c5 PJ |
7732 | match2: |
7733 | ; | |
7734 | } | |
7735 | ||
7736 | /* Remember the new sched domains */ | |
7737 | if (doms_cur != &fallback_doms) | |
7738 | kfree(doms_cur); | |
1d3504fc | 7739 | kfree(dattr_cur); /* kfree(NULL) is safe */ |
029190c5 | 7740 | doms_cur = doms_new; |
1d3504fc | 7741 | dattr_cur = dattr_new; |
029190c5 | 7742 | ndoms_cur = ndoms_new; |
7378547f MM |
7743 | |
7744 | register_sched_domain_sysctl(); | |
a1835615 | 7745 | |
712555ee | 7746 | mutex_unlock(&sched_domains_mutex); |
029190c5 PJ |
7747 | } |
7748 | ||
5c45bf27 | 7749 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
9aefd0ab | 7750 | int arch_reinit_sched_domains(void) |
5c45bf27 SS |
7751 | { |
7752 | int err; | |
7753 | ||
95402b38 | 7754 | get_online_cpus(); |
712555ee | 7755 | mutex_lock(&sched_domains_mutex); |
5c45bf27 SS |
7756 | detach_destroy_domains(&cpu_online_map); |
7757 | err = arch_init_sched_domains(&cpu_online_map); | |
712555ee | 7758 | mutex_unlock(&sched_domains_mutex); |
95402b38 | 7759 | put_online_cpus(); |
5c45bf27 SS |
7760 | |
7761 | return err; | |
7762 | } | |
7763 | ||
7764 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
7765 | { | |
7766 | int ret; | |
7767 | ||
7768 | if (buf[0] != '0' && buf[0] != '1') | |
7769 | return -EINVAL; | |
7770 | ||
7771 | if (smt) | |
7772 | sched_smt_power_savings = (buf[0] == '1'); | |
7773 | else | |
7774 | sched_mc_power_savings = (buf[0] == '1'); | |
7775 | ||
7776 | ret = arch_reinit_sched_domains(); | |
7777 | ||
7778 | return ret ? ret : count; | |
7779 | } | |
7780 | ||
5c45bf27 SS |
7781 | #ifdef CONFIG_SCHED_MC |
7782 | static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page) | |
7783 | { | |
7784 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
7785 | } | |
48f24c4d IM |
7786 | static ssize_t sched_mc_power_savings_store(struct sys_device *dev, |
7787 | const char *buf, size_t count) | |
5c45bf27 SS |
7788 | { |
7789 | return sched_power_savings_store(buf, count, 0); | |
7790 | } | |
6707de00 AB |
7791 | static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show, |
7792 | sched_mc_power_savings_store); | |
5c45bf27 SS |
7793 | #endif |
7794 | ||
7795 | #ifdef CONFIG_SCHED_SMT | |
7796 | static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page) | |
7797 | { | |
7798 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
7799 | } | |
48f24c4d IM |
7800 | static ssize_t sched_smt_power_savings_store(struct sys_device *dev, |
7801 | const char *buf, size_t count) | |
5c45bf27 SS |
7802 | { |
7803 | return sched_power_savings_store(buf, count, 1); | |
7804 | } | |
6707de00 AB |
7805 | static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show, |
7806 | sched_smt_power_savings_store); | |
7807 | #endif | |
7808 | ||
7809 | int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
7810 | { | |
7811 | int err = 0; | |
7812 | ||
7813 | #ifdef CONFIG_SCHED_SMT | |
7814 | if (smt_capable()) | |
7815 | err = sysfs_create_file(&cls->kset.kobj, | |
7816 | &attr_sched_smt_power_savings.attr); | |
7817 | #endif | |
7818 | #ifdef CONFIG_SCHED_MC | |
7819 | if (!err && mc_capable()) | |
7820 | err = sysfs_create_file(&cls->kset.kobj, | |
7821 | &attr_sched_mc_power_savings.attr); | |
7822 | #endif | |
7823 | return err; | |
7824 | } | |
5c45bf27 SS |
7825 | #endif |
7826 | ||
1da177e4 | 7827 | /* |
41a2d6cf | 7828 | * Force a reinitialization of the sched domains hierarchy. The domains |
1da177e4 | 7829 | * and groups cannot be updated in place without racing with the balancing |
41c7ce9a | 7830 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
7831 | * which will prevent rebalancing while the sched domains are recalculated. |
7832 | */ | |
7833 | static int update_sched_domains(struct notifier_block *nfb, | |
7834 | unsigned long action, void *hcpu) | |
7835 | { | |
1da177e4 LT |
7836 | switch (action) { |
7837 | case CPU_UP_PREPARE: | |
8bb78442 | 7838 | case CPU_UP_PREPARE_FROZEN: |
1da177e4 | 7839 | case CPU_DOWN_PREPARE: |
8bb78442 | 7840 | case CPU_DOWN_PREPARE_FROZEN: |
1a20ff27 | 7841 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
7842 | return NOTIFY_OK; |
7843 | ||
7844 | case CPU_UP_CANCELED: | |
8bb78442 | 7845 | case CPU_UP_CANCELED_FROZEN: |
1da177e4 | 7846 | case CPU_DOWN_FAILED: |
8bb78442 | 7847 | case CPU_DOWN_FAILED_FROZEN: |
1da177e4 | 7848 | case CPU_ONLINE: |
8bb78442 | 7849 | case CPU_ONLINE_FROZEN: |
1da177e4 | 7850 | case CPU_DEAD: |
8bb78442 | 7851 | case CPU_DEAD_FROZEN: |
1da177e4 LT |
7852 | /* |
7853 | * Fall through and re-initialise the domains. | |
7854 | */ | |
7855 | break; | |
7856 | default: | |
7857 | return NOTIFY_DONE; | |
7858 | } | |
7859 | ||
7860 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 7861 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
7862 | |
7863 | return NOTIFY_OK; | |
7864 | } | |
1da177e4 LT |
7865 | |
7866 | void __init sched_init_smp(void) | |
7867 | { | |
5c1e1767 NP |
7868 | cpumask_t non_isolated_cpus; |
7869 | ||
434d53b0 MT |
7870 | #if defined(CONFIG_NUMA) |
7871 | sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **), | |
7872 | GFP_KERNEL); | |
7873 | BUG_ON(sched_group_nodes_bycpu == NULL); | |
7874 | #endif | |
95402b38 | 7875 | get_online_cpus(); |
712555ee | 7876 | mutex_lock(&sched_domains_mutex); |
1a20ff27 | 7877 | arch_init_sched_domains(&cpu_online_map); |
e5e5673f | 7878 | cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map); |
5c1e1767 NP |
7879 | if (cpus_empty(non_isolated_cpus)) |
7880 | cpu_set(smp_processor_id(), non_isolated_cpus); | |
712555ee | 7881 | mutex_unlock(&sched_domains_mutex); |
95402b38 | 7882 | put_online_cpus(); |
1da177e4 LT |
7883 | /* XXX: Theoretical race here - CPU may be hotplugged now */ |
7884 | hotcpu_notifier(update_sched_domains, 0); | |
b328ca18 | 7885 | init_hrtick(); |
5c1e1767 NP |
7886 | |
7887 | /* Move init over to a non-isolated CPU */ | |
7c16ec58 | 7888 | if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0) |
5c1e1767 | 7889 | BUG(); |
19978ca6 | 7890 | sched_init_granularity(); |
1da177e4 LT |
7891 | } |
7892 | #else | |
7893 | void __init sched_init_smp(void) | |
7894 | { | |
19978ca6 | 7895 | sched_init_granularity(); |
1da177e4 LT |
7896 | } |
7897 | #endif /* CONFIG_SMP */ | |
7898 | ||
7899 | int in_sched_functions(unsigned long addr) | |
7900 | { | |
1da177e4 LT |
7901 | return in_lock_functions(addr) || |
7902 | (addr >= (unsigned long)__sched_text_start | |
7903 | && addr < (unsigned long)__sched_text_end); | |
7904 | } | |
7905 | ||
a9957449 | 7906 | static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq) |
dd41f596 IM |
7907 | { |
7908 | cfs_rq->tasks_timeline = RB_ROOT; | |
4a55bd5e | 7909 | INIT_LIST_HEAD(&cfs_rq->tasks); |
dd41f596 IM |
7910 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7911 | cfs_rq->rq = rq; | |
7912 | #endif | |
67e9fb2a | 7913 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
dd41f596 IM |
7914 | } |
7915 | ||
fa85ae24 PZ |
7916 | static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) |
7917 | { | |
7918 | struct rt_prio_array *array; | |
7919 | int i; | |
7920 | ||
7921 | array = &rt_rq->active; | |
7922 | for (i = 0; i < MAX_RT_PRIO; i++) { | |
7923 | INIT_LIST_HEAD(array->queue + i); | |
7924 | __clear_bit(i, array->bitmap); | |
7925 | } | |
7926 | /* delimiter for bitsearch: */ | |
7927 | __set_bit(MAX_RT_PRIO, array->bitmap); | |
7928 | ||
052f1dc7 | 7929 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
48d5e258 PZ |
7930 | rt_rq->highest_prio = MAX_RT_PRIO; |
7931 | #endif | |
fa85ae24 PZ |
7932 | #ifdef CONFIG_SMP |
7933 | rt_rq->rt_nr_migratory = 0; | |
fa85ae24 PZ |
7934 | rt_rq->overloaded = 0; |
7935 | #endif | |
7936 | ||
7937 | rt_rq->rt_time = 0; | |
7938 | rt_rq->rt_throttled = 0; | |
ac086bc2 PZ |
7939 | rt_rq->rt_runtime = 0; |
7940 | spin_lock_init(&rt_rq->rt_runtime_lock); | |
6f505b16 | 7941 | |
052f1dc7 | 7942 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc | 7943 | rt_rq->rt_nr_boosted = 0; |
6f505b16 PZ |
7944 | rt_rq->rq = rq; |
7945 | #endif | |
fa85ae24 PZ |
7946 | } |
7947 | ||
6f505b16 | 7948 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ec7dc8ac DG |
7949 | static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, |
7950 | struct sched_entity *se, int cpu, int add, | |
7951 | struct sched_entity *parent) | |
6f505b16 | 7952 | { |
ec7dc8ac | 7953 | struct rq *rq = cpu_rq(cpu); |
6f505b16 PZ |
7954 | tg->cfs_rq[cpu] = cfs_rq; |
7955 | init_cfs_rq(cfs_rq, rq); | |
7956 | cfs_rq->tg = tg; | |
7957 | if (add) | |
7958 | list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list); | |
7959 | ||
7960 | tg->se[cpu] = se; | |
354d60c2 DG |
7961 | /* se could be NULL for init_task_group */ |
7962 | if (!se) | |
7963 | return; | |
7964 | ||
ec7dc8ac DG |
7965 | if (!parent) |
7966 | se->cfs_rq = &rq->cfs; | |
7967 | else | |
7968 | se->cfs_rq = parent->my_q; | |
7969 | ||
6f505b16 PZ |
7970 | se->my_q = cfs_rq; |
7971 | se->load.weight = tg->shares; | |
e05510d0 | 7972 | se->load.inv_weight = 0; |
ec7dc8ac | 7973 | se->parent = parent; |
6f505b16 | 7974 | } |
052f1dc7 | 7975 | #endif |
6f505b16 | 7976 | |
052f1dc7 | 7977 | #ifdef CONFIG_RT_GROUP_SCHED |
ec7dc8ac DG |
7978 | static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, |
7979 | struct sched_rt_entity *rt_se, int cpu, int add, | |
7980 | struct sched_rt_entity *parent) | |
6f505b16 | 7981 | { |
ec7dc8ac DG |
7982 | struct rq *rq = cpu_rq(cpu); |
7983 | ||
6f505b16 PZ |
7984 | tg->rt_rq[cpu] = rt_rq; |
7985 | init_rt_rq(rt_rq, rq); | |
7986 | rt_rq->tg = tg; | |
7987 | rt_rq->rt_se = rt_se; | |
ac086bc2 | 7988 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; |
6f505b16 PZ |
7989 | if (add) |
7990 | list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list); | |
7991 | ||
7992 | tg->rt_se[cpu] = rt_se; | |
354d60c2 DG |
7993 | if (!rt_se) |
7994 | return; | |
7995 | ||
ec7dc8ac DG |
7996 | if (!parent) |
7997 | rt_se->rt_rq = &rq->rt; | |
7998 | else | |
7999 | rt_se->rt_rq = parent->my_q; | |
8000 | ||
6f505b16 PZ |
8001 | rt_se->rt_rq = &rq->rt; |
8002 | rt_se->my_q = rt_rq; | |
ec7dc8ac | 8003 | rt_se->parent = parent; |
6f505b16 PZ |
8004 | INIT_LIST_HEAD(&rt_se->run_list); |
8005 | } | |
8006 | #endif | |
8007 | ||
1da177e4 LT |
8008 | void __init sched_init(void) |
8009 | { | |
dd41f596 | 8010 | int i, j; |
434d53b0 MT |
8011 | unsigned long alloc_size = 0, ptr; |
8012 | ||
8013 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8014 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
8015 | #endif | |
8016 | #ifdef CONFIG_RT_GROUP_SCHED | |
8017 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
eff766a6 PZ |
8018 | #endif |
8019 | #ifdef CONFIG_USER_SCHED | |
8020 | alloc_size *= 2; | |
434d53b0 MT |
8021 | #endif |
8022 | /* | |
8023 | * As sched_init() is called before page_alloc is setup, | |
8024 | * we use alloc_bootmem(). | |
8025 | */ | |
8026 | if (alloc_size) { | |
5a9d3225 | 8027 | ptr = (unsigned long)alloc_bootmem(alloc_size); |
434d53b0 MT |
8028 | |
8029 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8030 | init_task_group.se = (struct sched_entity **)ptr; | |
8031 | ptr += nr_cpu_ids * sizeof(void **); | |
8032 | ||
8033 | init_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
8034 | ptr += nr_cpu_ids * sizeof(void **); | |
eff766a6 PZ |
8035 | |
8036 | #ifdef CONFIG_USER_SCHED | |
8037 | root_task_group.se = (struct sched_entity **)ptr; | |
8038 | ptr += nr_cpu_ids * sizeof(void **); | |
8039 | ||
8040 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
8041 | ptr += nr_cpu_ids * sizeof(void **); | |
8042 | #endif | |
434d53b0 MT |
8043 | #endif |
8044 | #ifdef CONFIG_RT_GROUP_SCHED | |
8045 | init_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
8046 | ptr += nr_cpu_ids * sizeof(void **); | |
8047 | ||
8048 | init_task_group.rt_rq = (struct rt_rq **)ptr; | |
eff766a6 PZ |
8049 | ptr += nr_cpu_ids * sizeof(void **); |
8050 | ||
8051 | #ifdef CONFIG_USER_SCHED | |
8052 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
8053 | ptr += nr_cpu_ids * sizeof(void **); | |
8054 | ||
8055 | root_task_group.rt_rq = (struct rt_rq **)ptr; | |
8056 | ptr += nr_cpu_ids * sizeof(void **); | |
8057 | #endif | |
434d53b0 MT |
8058 | #endif |
8059 | } | |
dd41f596 | 8060 | |
57d885fe | 8061 | #ifdef CONFIG_SMP |
18d95a28 | 8062 | init_aggregate(); |
57d885fe GH |
8063 | init_defrootdomain(); |
8064 | #endif | |
8065 | ||
d0b27fa7 PZ |
8066 | init_rt_bandwidth(&def_rt_bandwidth, |
8067 | global_rt_period(), global_rt_runtime()); | |
8068 | ||
8069 | #ifdef CONFIG_RT_GROUP_SCHED | |
8070 | init_rt_bandwidth(&init_task_group.rt_bandwidth, | |
8071 | global_rt_period(), global_rt_runtime()); | |
eff766a6 PZ |
8072 | #ifdef CONFIG_USER_SCHED |
8073 | init_rt_bandwidth(&root_task_group.rt_bandwidth, | |
8074 | global_rt_period(), RUNTIME_INF); | |
8075 | #endif | |
d0b27fa7 PZ |
8076 | #endif |
8077 | ||
052f1dc7 | 8078 | #ifdef CONFIG_GROUP_SCHED |
6f505b16 | 8079 | list_add(&init_task_group.list, &task_groups); |
f473aa5e PZ |
8080 | INIT_LIST_HEAD(&init_task_group.children); |
8081 | ||
8082 | #ifdef CONFIG_USER_SCHED | |
8083 | INIT_LIST_HEAD(&root_task_group.children); | |
8084 | init_task_group.parent = &root_task_group; | |
8085 | list_add(&init_task_group.siblings, &root_task_group.children); | |
8086 | #endif | |
6f505b16 PZ |
8087 | #endif |
8088 | ||
0a945022 | 8089 | for_each_possible_cpu(i) { |
70b97a7f | 8090 | struct rq *rq; |
1da177e4 LT |
8091 | |
8092 | rq = cpu_rq(i); | |
8093 | spin_lock_init(&rq->lock); | |
fcb99371 | 8094 | lockdep_set_class(&rq->lock, &rq->rq_lock_key); |
7897986b | 8095 | rq->nr_running = 0; |
dd41f596 | 8096 | init_cfs_rq(&rq->cfs, rq); |
6f505b16 | 8097 | init_rt_rq(&rq->rt, rq); |
dd41f596 | 8098 | #ifdef CONFIG_FAIR_GROUP_SCHED |
4cf86d77 | 8099 | init_task_group.shares = init_task_group_load; |
6f505b16 | 8100 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
354d60c2 DG |
8101 | #ifdef CONFIG_CGROUP_SCHED |
8102 | /* | |
8103 | * How much cpu bandwidth does init_task_group get? | |
8104 | * | |
8105 | * In case of task-groups formed thr' the cgroup filesystem, it | |
8106 | * gets 100% of the cpu resources in the system. This overall | |
8107 | * system cpu resource is divided among the tasks of | |
8108 | * init_task_group and its child task-groups in a fair manner, | |
8109 | * based on each entity's (task or task-group's) weight | |
8110 | * (se->load.weight). | |
8111 | * | |
8112 | * In other words, if init_task_group has 10 tasks of weight | |
8113 | * 1024) and two child groups A0 and A1 (of weight 1024 each), | |
8114 | * then A0's share of the cpu resource is: | |
8115 | * | |
8116 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% | |
8117 | * | |
8118 | * We achieve this by letting init_task_group's tasks sit | |
8119 | * directly in rq->cfs (i.e init_task_group->se[] = NULL). | |
8120 | */ | |
ec7dc8ac | 8121 | init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL); |
354d60c2 | 8122 | #elif defined CONFIG_USER_SCHED |
eff766a6 PZ |
8123 | root_task_group.shares = NICE_0_LOAD; |
8124 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL); | |
354d60c2 DG |
8125 | /* |
8126 | * In case of task-groups formed thr' the user id of tasks, | |
8127 | * init_task_group represents tasks belonging to root user. | |
8128 | * Hence it forms a sibling of all subsequent groups formed. | |
8129 | * In this case, init_task_group gets only a fraction of overall | |
8130 | * system cpu resource, based on the weight assigned to root | |
8131 | * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished | |
8132 | * by letting tasks of init_task_group sit in a separate cfs_rq | |
8133 | * (init_cfs_rq) and having one entity represent this group of | |
8134 | * tasks in rq->cfs (i.e init_task_group->se[] != NULL). | |
8135 | */ | |
ec7dc8ac | 8136 | init_tg_cfs_entry(&init_task_group, |
6f505b16 | 8137 | &per_cpu(init_cfs_rq, i), |
eff766a6 PZ |
8138 | &per_cpu(init_sched_entity, i), i, 1, |
8139 | root_task_group.se[i]); | |
6f505b16 | 8140 | |
052f1dc7 | 8141 | #endif |
354d60c2 DG |
8142 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
8143 | ||
8144 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; | |
052f1dc7 | 8145 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 8146 | INIT_LIST_HEAD(&rq->leaf_rt_rq_list); |
354d60c2 | 8147 | #ifdef CONFIG_CGROUP_SCHED |
ec7dc8ac | 8148 | init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL); |
354d60c2 | 8149 | #elif defined CONFIG_USER_SCHED |
eff766a6 | 8150 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL); |
ec7dc8ac | 8151 | init_tg_rt_entry(&init_task_group, |
6f505b16 | 8152 | &per_cpu(init_rt_rq, i), |
eff766a6 PZ |
8153 | &per_cpu(init_sched_rt_entity, i), i, 1, |
8154 | root_task_group.rt_se[i]); | |
354d60c2 | 8155 | #endif |
dd41f596 | 8156 | #endif |
1da177e4 | 8157 | |
dd41f596 IM |
8158 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) |
8159 | rq->cpu_load[j] = 0; | |
1da177e4 | 8160 | #ifdef CONFIG_SMP |
41c7ce9a | 8161 | rq->sd = NULL; |
57d885fe | 8162 | rq->rd = NULL; |
1da177e4 | 8163 | rq->active_balance = 0; |
dd41f596 | 8164 | rq->next_balance = jiffies; |
1da177e4 | 8165 | rq->push_cpu = 0; |
0a2966b4 | 8166 | rq->cpu = i; |
1da177e4 LT |
8167 | rq->migration_thread = NULL; |
8168 | INIT_LIST_HEAD(&rq->migration_queue); | |
dc938520 | 8169 | rq_attach_root(rq, &def_root_domain); |
1da177e4 | 8170 | #endif |
8f4d37ec | 8171 | init_rq_hrtick(rq); |
1da177e4 | 8172 | atomic_set(&rq->nr_iowait, 0); |
1da177e4 LT |
8173 | } |
8174 | ||
2dd73a4f | 8175 | set_load_weight(&init_task); |
b50f60ce | 8176 | |
e107be36 AK |
8177 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
8178 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); | |
8179 | #endif | |
8180 | ||
c9819f45 CL |
8181 | #ifdef CONFIG_SMP |
8182 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL); | |
8183 | #endif | |
8184 | ||
b50f60ce HC |
8185 | #ifdef CONFIG_RT_MUTEXES |
8186 | plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); | |
8187 | #endif | |
8188 | ||
1da177e4 LT |
8189 | /* |
8190 | * The boot idle thread does lazy MMU switching as well: | |
8191 | */ | |
8192 | atomic_inc(&init_mm.mm_count); | |
8193 | enter_lazy_tlb(&init_mm, current); | |
8194 | ||
8195 | /* | |
8196 | * Make us the idle thread. Technically, schedule() should not be | |
8197 | * called from this thread, however somewhere below it might be, | |
8198 | * but because we are the idle thread, we just pick up running again | |
8199 | * when this runqueue becomes "idle". | |
8200 | */ | |
8201 | init_idle(current, smp_processor_id()); | |
dd41f596 IM |
8202 | /* |
8203 | * During early bootup we pretend to be a normal task: | |
8204 | */ | |
8205 | current->sched_class = &fair_sched_class; | |
6892b75e IM |
8206 | |
8207 | scheduler_running = 1; | |
1da177e4 LT |
8208 | } |
8209 | ||
8210 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
8211 | void __might_sleep(char *file, int line) | |
8212 | { | |
48f24c4d | 8213 | #ifdef in_atomic |
1da177e4 LT |
8214 | static unsigned long prev_jiffy; /* ratelimiting */ |
8215 | ||
8216 | if ((in_atomic() || irqs_disabled()) && | |
8217 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
8218 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
8219 | return; | |
8220 | prev_jiffy = jiffies; | |
91368d73 | 8221 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
1da177e4 LT |
8222 | " context at %s:%d\n", file, line); |
8223 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
8224 | in_atomic(), irqs_disabled()); | |
a4c410f0 | 8225 | debug_show_held_locks(current); |
3117df04 IM |
8226 | if (irqs_disabled()) |
8227 | print_irqtrace_events(current); | |
1da177e4 LT |
8228 | dump_stack(); |
8229 | } | |
8230 | #endif | |
8231 | } | |
8232 | EXPORT_SYMBOL(__might_sleep); | |
8233 | #endif | |
8234 | ||
8235 | #ifdef CONFIG_MAGIC_SYSRQ | |
3a5e4dc1 AK |
8236 | static void normalize_task(struct rq *rq, struct task_struct *p) |
8237 | { | |
8238 | int on_rq; | |
3e51f33f | 8239 | |
3a5e4dc1 AK |
8240 | update_rq_clock(rq); |
8241 | on_rq = p->se.on_rq; | |
8242 | if (on_rq) | |
8243 | deactivate_task(rq, p, 0); | |
8244 | __setscheduler(rq, p, SCHED_NORMAL, 0); | |
8245 | if (on_rq) { | |
8246 | activate_task(rq, p, 0); | |
8247 | resched_task(rq->curr); | |
8248 | } | |
8249 | } | |
8250 | ||
1da177e4 LT |
8251 | void normalize_rt_tasks(void) |
8252 | { | |
a0f98a1c | 8253 | struct task_struct *g, *p; |
1da177e4 | 8254 | unsigned long flags; |
70b97a7f | 8255 | struct rq *rq; |
1da177e4 | 8256 | |
4cf5d77a | 8257 | read_lock_irqsave(&tasklist_lock, flags); |
a0f98a1c | 8258 | do_each_thread(g, p) { |
178be793 IM |
8259 | /* |
8260 | * Only normalize user tasks: | |
8261 | */ | |
8262 | if (!p->mm) | |
8263 | continue; | |
8264 | ||
6cfb0d5d | 8265 | p->se.exec_start = 0; |
6cfb0d5d | 8266 | #ifdef CONFIG_SCHEDSTATS |
dd41f596 | 8267 | p->se.wait_start = 0; |
dd41f596 | 8268 | p->se.sleep_start = 0; |
dd41f596 | 8269 | p->se.block_start = 0; |
6cfb0d5d | 8270 | #endif |
dd41f596 IM |
8271 | |
8272 | if (!rt_task(p)) { | |
8273 | /* | |
8274 | * Renice negative nice level userspace | |
8275 | * tasks back to 0: | |
8276 | */ | |
8277 | if (TASK_NICE(p) < 0 && p->mm) | |
8278 | set_user_nice(p, 0); | |
1da177e4 | 8279 | continue; |
dd41f596 | 8280 | } |
1da177e4 | 8281 | |
4cf5d77a | 8282 | spin_lock(&p->pi_lock); |
b29739f9 | 8283 | rq = __task_rq_lock(p); |
1da177e4 | 8284 | |
178be793 | 8285 | normalize_task(rq, p); |
3a5e4dc1 | 8286 | |
b29739f9 | 8287 | __task_rq_unlock(rq); |
4cf5d77a | 8288 | spin_unlock(&p->pi_lock); |
a0f98a1c IM |
8289 | } while_each_thread(g, p); |
8290 | ||
4cf5d77a | 8291 | read_unlock_irqrestore(&tasklist_lock, flags); |
1da177e4 LT |
8292 | } |
8293 | ||
8294 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
8295 | |
8296 | #ifdef CONFIG_IA64 | |
8297 | /* | |
8298 | * These functions are only useful for the IA64 MCA handling. | |
8299 | * | |
8300 | * They can only be called when the whole system has been | |
8301 | * stopped - every CPU needs to be quiescent, and no scheduling | |
8302 | * activity can take place. Using them for anything else would | |
8303 | * be a serious bug, and as a result, they aren't even visible | |
8304 | * under any other configuration. | |
8305 | */ | |
8306 | ||
8307 | /** | |
8308 | * curr_task - return the current task for a given cpu. | |
8309 | * @cpu: the processor in question. | |
8310 | * | |
8311 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8312 | */ | |
36c8b586 | 8313 | struct task_struct *curr_task(int cpu) |
1df5c10a LT |
8314 | { |
8315 | return cpu_curr(cpu); | |
8316 | } | |
8317 | ||
8318 | /** | |
8319 | * set_curr_task - set the current task for a given cpu. | |
8320 | * @cpu: the processor in question. | |
8321 | * @p: the task pointer to set. | |
8322 | * | |
8323 | * Description: This function must only be used when non-maskable interrupts | |
41a2d6cf IM |
8324 | * are serviced on a separate stack. It allows the architecture to switch the |
8325 | * notion of the current task on a cpu in a non-blocking manner. This function | |
1df5c10a LT |
8326 | * must be called with all CPU's synchronized, and interrupts disabled, the |
8327 | * and caller must save the original value of the current task (see | |
8328 | * curr_task() above) and restore that value before reenabling interrupts and | |
8329 | * re-starting the system. | |
8330 | * | |
8331 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8332 | */ | |
36c8b586 | 8333 | void set_curr_task(int cpu, struct task_struct *p) |
1df5c10a LT |
8334 | { |
8335 | cpu_curr(cpu) = p; | |
8336 | } | |
8337 | ||
8338 | #endif | |
29f59db3 | 8339 | |
bccbe08a PZ |
8340 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8341 | static void free_fair_sched_group(struct task_group *tg) | |
6f505b16 PZ |
8342 | { |
8343 | int i; | |
8344 | ||
8345 | for_each_possible_cpu(i) { | |
8346 | if (tg->cfs_rq) | |
8347 | kfree(tg->cfs_rq[i]); | |
8348 | if (tg->se) | |
8349 | kfree(tg->se[i]); | |
6f505b16 PZ |
8350 | } |
8351 | ||
8352 | kfree(tg->cfs_rq); | |
8353 | kfree(tg->se); | |
6f505b16 PZ |
8354 | } |
8355 | ||
ec7dc8ac DG |
8356 | static |
8357 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
29f59db3 | 8358 | { |
29f59db3 | 8359 | struct cfs_rq *cfs_rq; |
ec7dc8ac | 8360 | struct sched_entity *se, *parent_se; |
9b5b7751 | 8361 | struct rq *rq; |
29f59db3 SV |
8362 | int i; |
8363 | ||
434d53b0 | 8364 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); |
29f59db3 SV |
8365 | if (!tg->cfs_rq) |
8366 | goto err; | |
434d53b0 | 8367 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); |
29f59db3 SV |
8368 | if (!tg->se) |
8369 | goto err; | |
052f1dc7 PZ |
8370 | |
8371 | tg->shares = NICE_0_LOAD; | |
29f59db3 SV |
8372 | |
8373 | for_each_possible_cpu(i) { | |
9b5b7751 | 8374 | rq = cpu_rq(i); |
29f59db3 | 8375 | |
6f505b16 PZ |
8376 | cfs_rq = kmalloc_node(sizeof(struct cfs_rq), |
8377 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
29f59db3 SV |
8378 | if (!cfs_rq) |
8379 | goto err; | |
8380 | ||
6f505b16 PZ |
8381 | se = kmalloc_node(sizeof(struct sched_entity), |
8382 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
29f59db3 SV |
8383 | if (!se) |
8384 | goto err; | |
8385 | ||
ec7dc8ac DG |
8386 | parent_se = parent ? parent->se[i] : NULL; |
8387 | init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se); | |
bccbe08a PZ |
8388 | } |
8389 | ||
8390 | return 1; | |
8391 | ||
8392 | err: | |
8393 | return 0; | |
8394 | } | |
8395 | ||
8396 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8397 | { | |
8398 | list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list, | |
8399 | &cpu_rq(cpu)->leaf_cfs_rq_list); | |
8400 | } | |
8401 | ||
8402 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8403 | { | |
8404 | list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list); | |
8405 | } | |
8406 | #else | |
8407 | static inline void free_fair_sched_group(struct task_group *tg) | |
8408 | { | |
8409 | } | |
8410 | ||
ec7dc8ac DG |
8411 | static inline |
8412 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
bccbe08a PZ |
8413 | { |
8414 | return 1; | |
8415 | } | |
8416 | ||
8417 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8418 | { | |
8419 | } | |
8420 | ||
8421 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8422 | { | |
8423 | } | |
052f1dc7 PZ |
8424 | #endif |
8425 | ||
8426 | #ifdef CONFIG_RT_GROUP_SCHED | |
bccbe08a PZ |
8427 | static void free_rt_sched_group(struct task_group *tg) |
8428 | { | |
8429 | int i; | |
8430 | ||
d0b27fa7 PZ |
8431 | destroy_rt_bandwidth(&tg->rt_bandwidth); |
8432 | ||
bccbe08a PZ |
8433 | for_each_possible_cpu(i) { |
8434 | if (tg->rt_rq) | |
8435 | kfree(tg->rt_rq[i]); | |
8436 | if (tg->rt_se) | |
8437 | kfree(tg->rt_se[i]); | |
8438 | } | |
8439 | ||
8440 | kfree(tg->rt_rq); | |
8441 | kfree(tg->rt_se); | |
8442 | } | |
8443 | ||
ec7dc8ac DG |
8444 | static |
8445 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
bccbe08a PZ |
8446 | { |
8447 | struct rt_rq *rt_rq; | |
ec7dc8ac | 8448 | struct sched_rt_entity *rt_se, *parent_se; |
bccbe08a PZ |
8449 | struct rq *rq; |
8450 | int i; | |
8451 | ||
434d53b0 | 8452 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); |
bccbe08a PZ |
8453 | if (!tg->rt_rq) |
8454 | goto err; | |
434d53b0 | 8455 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); |
bccbe08a PZ |
8456 | if (!tg->rt_se) |
8457 | goto err; | |
8458 | ||
d0b27fa7 PZ |
8459 | init_rt_bandwidth(&tg->rt_bandwidth, |
8460 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); | |
bccbe08a PZ |
8461 | |
8462 | for_each_possible_cpu(i) { | |
8463 | rq = cpu_rq(i); | |
8464 | ||
6f505b16 PZ |
8465 | rt_rq = kmalloc_node(sizeof(struct rt_rq), |
8466 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8467 | if (!rt_rq) | |
8468 | goto err; | |
29f59db3 | 8469 | |
6f505b16 PZ |
8470 | rt_se = kmalloc_node(sizeof(struct sched_rt_entity), |
8471 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8472 | if (!rt_se) | |
8473 | goto err; | |
29f59db3 | 8474 | |
ec7dc8ac DG |
8475 | parent_se = parent ? parent->rt_se[i] : NULL; |
8476 | init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se); | |
29f59db3 SV |
8477 | } |
8478 | ||
bccbe08a PZ |
8479 | return 1; |
8480 | ||
8481 | err: | |
8482 | return 0; | |
8483 | } | |
8484 | ||
8485 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8486 | { | |
8487 | list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list, | |
8488 | &cpu_rq(cpu)->leaf_rt_rq_list); | |
8489 | } | |
8490 | ||
8491 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8492 | { | |
8493 | list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list); | |
8494 | } | |
8495 | #else | |
8496 | static inline void free_rt_sched_group(struct task_group *tg) | |
8497 | { | |
8498 | } | |
8499 | ||
ec7dc8ac DG |
8500 | static inline |
8501 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
bccbe08a PZ |
8502 | { |
8503 | return 1; | |
8504 | } | |
8505 | ||
8506 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8507 | { | |
8508 | } | |
8509 | ||
8510 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8511 | { | |
8512 | } | |
8513 | #endif | |
8514 | ||
d0b27fa7 | 8515 | #ifdef CONFIG_GROUP_SCHED |
bccbe08a PZ |
8516 | static void free_sched_group(struct task_group *tg) |
8517 | { | |
8518 | free_fair_sched_group(tg); | |
8519 | free_rt_sched_group(tg); | |
8520 | kfree(tg); | |
8521 | } | |
8522 | ||
8523 | /* allocate runqueue etc for a new task group */ | |
ec7dc8ac | 8524 | struct task_group *sched_create_group(struct task_group *parent) |
bccbe08a PZ |
8525 | { |
8526 | struct task_group *tg; | |
8527 | unsigned long flags; | |
8528 | int i; | |
8529 | ||
8530 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); | |
8531 | if (!tg) | |
8532 | return ERR_PTR(-ENOMEM); | |
8533 | ||
ec7dc8ac | 8534 | if (!alloc_fair_sched_group(tg, parent)) |
bccbe08a PZ |
8535 | goto err; |
8536 | ||
ec7dc8ac | 8537 | if (!alloc_rt_sched_group(tg, parent)) |
bccbe08a PZ |
8538 | goto err; |
8539 | ||
8ed36996 | 8540 | spin_lock_irqsave(&task_group_lock, flags); |
9b5b7751 | 8541 | for_each_possible_cpu(i) { |
bccbe08a PZ |
8542 | register_fair_sched_group(tg, i); |
8543 | register_rt_sched_group(tg, i); | |
9b5b7751 | 8544 | } |
6f505b16 | 8545 | list_add_rcu(&tg->list, &task_groups); |
f473aa5e PZ |
8546 | |
8547 | WARN_ON(!parent); /* root should already exist */ | |
8548 | ||
8549 | tg->parent = parent; | |
8550 | list_add_rcu(&tg->siblings, &parent->children); | |
8551 | INIT_LIST_HEAD(&tg->children); | |
8ed36996 | 8552 | spin_unlock_irqrestore(&task_group_lock, flags); |
29f59db3 | 8553 | |
9b5b7751 | 8554 | return tg; |
29f59db3 SV |
8555 | |
8556 | err: | |
6f505b16 | 8557 | free_sched_group(tg); |
29f59db3 SV |
8558 | return ERR_PTR(-ENOMEM); |
8559 | } | |
8560 | ||
9b5b7751 | 8561 | /* rcu callback to free various structures associated with a task group */ |
6f505b16 | 8562 | static void free_sched_group_rcu(struct rcu_head *rhp) |
29f59db3 | 8563 | { |
29f59db3 | 8564 | /* now it should be safe to free those cfs_rqs */ |
6f505b16 | 8565 | free_sched_group(container_of(rhp, struct task_group, rcu)); |
29f59db3 SV |
8566 | } |
8567 | ||
9b5b7751 | 8568 | /* Destroy runqueue etc associated with a task group */ |
4cf86d77 | 8569 | void sched_destroy_group(struct task_group *tg) |
29f59db3 | 8570 | { |
8ed36996 | 8571 | unsigned long flags; |
9b5b7751 | 8572 | int i; |
29f59db3 | 8573 | |
8ed36996 | 8574 | spin_lock_irqsave(&task_group_lock, flags); |
9b5b7751 | 8575 | for_each_possible_cpu(i) { |
bccbe08a PZ |
8576 | unregister_fair_sched_group(tg, i); |
8577 | unregister_rt_sched_group(tg, i); | |
9b5b7751 | 8578 | } |
6f505b16 | 8579 | list_del_rcu(&tg->list); |
f473aa5e | 8580 | list_del_rcu(&tg->siblings); |
8ed36996 | 8581 | spin_unlock_irqrestore(&task_group_lock, flags); |
9b5b7751 | 8582 | |
9b5b7751 | 8583 | /* wait for possible concurrent references to cfs_rqs complete */ |
6f505b16 | 8584 | call_rcu(&tg->rcu, free_sched_group_rcu); |
29f59db3 SV |
8585 | } |
8586 | ||
9b5b7751 | 8587 | /* change task's runqueue when it moves between groups. |
3a252015 IM |
8588 | * The caller of this function should have put the task in its new group |
8589 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to | |
8590 | * reflect its new group. | |
9b5b7751 SV |
8591 | */ |
8592 | void sched_move_task(struct task_struct *tsk) | |
29f59db3 SV |
8593 | { |
8594 | int on_rq, running; | |
8595 | unsigned long flags; | |
8596 | struct rq *rq; | |
8597 | ||
8598 | rq = task_rq_lock(tsk, &flags); | |
8599 | ||
29f59db3 SV |
8600 | update_rq_clock(rq); |
8601 | ||
051a1d1a | 8602 | running = task_current(rq, tsk); |
29f59db3 SV |
8603 | on_rq = tsk->se.on_rq; |
8604 | ||
0e1f3483 | 8605 | if (on_rq) |
29f59db3 | 8606 | dequeue_task(rq, tsk, 0); |
0e1f3483 HS |
8607 | if (unlikely(running)) |
8608 | tsk->sched_class->put_prev_task(rq, tsk); | |
29f59db3 | 8609 | |
6f505b16 | 8610 | set_task_rq(tsk, task_cpu(tsk)); |
29f59db3 | 8611 | |
810b3817 PZ |
8612 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8613 | if (tsk->sched_class->moved_group) | |
8614 | tsk->sched_class->moved_group(tsk); | |
8615 | #endif | |
8616 | ||
0e1f3483 HS |
8617 | if (unlikely(running)) |
8618 | tsk->sched_class->set_curr_task(rq); | |
8619 | if (on_rq) | |
7074badb | 8620 | enqueue_task(rq, tsk, 0); |
29f59db3 | 8621 | |
29f59db3 SV |
8622 | task_rq_unlock(rq, &flags); |
8623 | } | |
d0b27fa7 | 8624 | #endif |
29f59db3 | 8625 | |
052f1dc7 | 8626 | #ifdef CONFIG_FAIR_GROUP_SCHED |
18d95a28 | 8627 | static void __set_se_shares(struct sched_entity *se, unsigned long shares) |
29f59db3 SV |
8628 | { |
8629 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
29f59db3 SV |
8630 | int on_rq; |
8631 | ||
29f59db3 | 8632 | on_rq = se->on_rq; |
62fb1851 | 8633 | if (on_rq) |
29f59db3 SV |
8634 | dequeue_entity(cfs_rq, se, 0); |
8635 | ||
8636 | se->load.weight = shares; | |
e05510d0 | 8637 | se->load.inv_weight = 0; |
29f59db3 | 8638 | |
62fb1851 | 8639 | if (on_rq) |
29f59db3 | 8640 | enqueue_entity(cfs_rq, se, 0); |
18d95a28 | 8641 | } |
62fb1851 | 8642 | |
18d95a28 PZ |
8643 | static void set_se_shares(struct sched_entity *se, unsigned long shares) |
8644 | { | |
8645 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
8646 | struct rq *rq = cfs_rq->rq; | |
8647 | unsigned long flags; | |
8648 | ||
8649 | spin_lock_irqsave(&rq->lock, flags); | |
8650 | __set_se_shares(se, shares); | |
8651 | spin_unlock_irqrestore(&rq->lock, flags); | |
29f59db3 SV |
8652 | } |
8653 | ||
8ed36996 PZ |
8654 | static DEFINE_MUTEX(shares_mutex); |
8655 | ||
4cf86d77 | 8656 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) |
29f59db3 SV |
8657 | { |
8658 | int i; | |
8ed36996 | 8659 | unsigned long flags; |
c61935fd | 8660 | |
ec7dc8ac DG |
8661 | /* |
8662 | * We can't change the weight of the root cgroup. | |
8663 | */ | |
8664 | if (!tg->se[0]) | |
8665 | return -EINVAL; | |
8666 | ||
18d95a28 PZ |
8667 | if (shares < MIN_SHARES) |
8668 | shares = MIN_SHARES; | |
cb4ad1ff MX |
8669 | else if (shares > MAX_SHARES) |
8670 | shares = MAX_SHARES; | |
62fb1851 | 8671 | |
8ed36996 | 8672 | mutex_lock(&shares_mutex); |
9b5b7751 | 8673 | if (tg->shares == shares) |
5cb350ba | 8674 | goto done; |
29f59db3 | 8675 | |
8ed36996 | 8676 | spin_lock_irqsave(&task_group_lock, flags); |
bccbe08a PZ |
8677 | for_each_possible_cpu(i) |
8678 | unregister_fair_sched_group(tg, i); | |
f473aa5e | 8679 | list_del_rcu(&tg->siblings); |
8ed36996 | 8680 | spin_unlock_irqrestore(&task_group_lock, flags); |
6b2d7700 SV |
8681 | |
8682 | /* wait for any ongoing reference to this group to finish */ | |
8683 | synchronize_sched(); | |
8684 | ||
8685 | /* | |
8686 | * Now we are free to modify the group's share on each cpu | |
8687 | * w/o tripping rebalance_share or load_balance_fair. | |
8688 | */ | |
9b5b7751 | 8689 | tg->shares = shares; |
18d95a28 PZ |
8690 | for_each_possible_cpu(i) { |
8691 | /* | |
8692 | * force a rebalance | |
8693 | */ | |
8694 | cfs_rq_set_shares(tg->cfs_rq[i], 0); | |
cb4ad1ff | 8695 | set_se_shares(tg->se[i], shares); |
18d95a28 | 8696 | } |
29f59db3 | 8697 | |
6b2d7700 SV |
8698 | /* |
8699 | * Enable load balance activity on this group, by inserting it back on | |
8700 | * each cpu's rq->leaf_cfs_rq_list. | |
8701 | */ | |
8ed36996 | 8702 | spin_lock_irqsave(&task_group_lock, flags); |
bccbe08a PZ |
8703 | for_each_possible_cpu(i) |
8704 | register_fair_sched_group(tg, i); | |
f473aa5e | 8705 | list_add_rcu(&tg->siblings, &tg->parent->children); |
8ed36996 | 8706 | spin_unlock_irqrestore(&task_group_lock, flags); |
5cb350ba | 8707 | done: |
8ed36996 | 8708 | mutex_unlock(&shares_mutex); |
9b5b7751 | 8709 | return 0; |
29f59db3 SV |
8710 | } |
8711 | ||
5cb350ba DG |
8712 | unsigned long sched_group_shares(struct task_group *tg) |
8713 | { | |
8714 | return tg->shares; | |
8715 | } | |
052f1dc7 | 8716 | #endif |
5cb350ba | 8717 | |
052f1dc7 | 8718 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 8719 | /* |
9f0c1e56 | 8720 | * Ensure that the real time constraints are schedulable. |
6f505b16 | 8721 | */ |
9f0c1e56 PZ |
8722 | static DEFINE_MUTEX(rt_constraints_mutex); |
8723 | ||
8724 | static unsigned long to_ratio(u64 period, u64 runtime) | |
8725 | { | |
8726 | if (runtime == RUNTIME_INF) | |
8727 | return 1ULL << 16; | |
8728 | ||
6f6d6a1a | 8729 | return div64_u64(runtime << 16, period); |
9f0c1e56 PZ |
8730 | } |
8731 | ||
b40b2e8e PZ |
8732 | #ifdef CONFIG_CGROUP_SCHED |
8733 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
8734 | { | |
8735 | struct task_group *tgi, *parent = tg->parent; | |
8736 | unsigned long total = 0; | |
8737 | ||
8738 | if (!parent) { | |
8739 | if (global_rt_period() < period) | |
8740 | return 0; | |
8741 | ||
8742 | return to_ratio(period, runtime) < | |
8743 | to_ratio(global_rt_period(), global_rt_runtime()); | |
8744 | } | |
8745 | ||
8746 | if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period) | |
8747 | return 0; | |
8748 | ||
8749 | rcu_read_lock(); | |
8750 | list_for_each_entry_rcu(tgi, &parent->children, siblings) { | |
8751 | if (tgi == tg) | |
8752 | continue; | |
8753 | ||
8754 | total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period), | |
8755 | tgi->rt_bandwidth.rt_runtime); | |
8756 | } | |
8757 | rcu_read_unlock(); | |
8758 | ||
8759 | return total + to_ratio(period, runtime) < | |
8760 | to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period), | |
8761 | parent->rt_bandwidth.rt_runtime); | |
8762 | } | |
8763 | #elif defined CONFIG_USER_SCHED | |
9f0c1e56 | 8764 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) |
6f505b16 PZ |
8765 | { |
8766 | struct task_group *tgi; | |
8767 | unsigned long total = 0; | |
9f0c1e56 | 8768 | unsigned long global_ratio = |
d0b27fa7 | 8769 | to_ratio(global_rt_period(), global_rt_runtime()); |
6f505b16 PZ |
8770 | |
8771 | rcu_read_lock(); | |
9f0c1e56 PZ |
8772 | list_for_each_entry_rcu(tgi, &task_groups, list) { |
8773 | if (tgi == tg) | |
8774 | continue; | |
6f505b16 | 8775 | |
d0b27fa7 PZ |
8776 | total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period), |
8777 | tgi->rt_bandwidth.rt_runtime); | |
9f0c1e56 PZ |
8778 | } |
8779 | rcu_read_unlock(); | |
6f505b16 | 8780 | |
9f0c1e56 | 8781 | return total + to_ratio(period, runtime) < global_ratio; |
6f505b16 | 8782 | } |
b40b2e8e | 8783 | #endif |
6f505b16 | 8784 | |
521f1a24 DG |
8785 | /* Must be called with tasklist_lock held */ |
8786 | static inline int tg_has_rt_tasks(struct task_group *tg) | |
8787 | { | |
8788 | struct task_struct *g, *p; | |
8789 | do_each_thread(g, p) { | |
8790 | if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg) | |
8791 | return 1; | |
8792 | } while_each_thread(g, p); | |
8793 | return 0; | |
8794 | } | |
8795 | ||
d0b27fa7 PZ |
8796 | static int tg_set_bandwidth(struct task_group *tg, |
8797 | u64 rt_period, u64 rt_runtime) | |
6f505b16 | 8798 | { |
ac086bc2 | 8799 | int i, err = 0; |
9f0c1e56 | 8800 | |
9f0c1e56 | 8801 | mutex_lock(&rt_constraints_mutex); |
521f1a24 | 8802 | read_lock(&tasklist_lock); |
ac086bc2 | 8803 | if (rt_runtime == 0 && tg_has_rt_tasks(tg)) { |
521f1a24 DG |
8804 | err = -EBUSY; |
8805 | goto unlock; | |
8806 | } | |
9f0c1e56 PZ |
8807 | if (!__rt_schedulable(tg, rt_period, rt_runtime)) { |
8808 | err = -EINVAL; | |
8809 | goto unlock; | |
8810 | } | |
ac086bc2 PZ |
8811 | |
8812 | spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
d0b27fa7 PZ |
8813 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); |
8814 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
ac086bc2 PZ |
8815 | |
8816 | for_each_possible_cpu(i) { | |
8817 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
8818 | ||
8819 | spin_lock(&rt_rq->rt_runtime_lock); | |
8820 | rt_rq->rt_runtime = rt_runtime; | |
8821 | spin_unlock(&rt_rq->rt_runtime_lock); | |
8822 | } | |
8823 | spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
9f0c1e56 | 8824 | unlock: |
521f1a24 | 8825 | read_unlock(&tasklist_lock); |
9f0c1e56 PZ |
8826 | mutex_unlock(&rt_constraints_mutex); |
8827 | ||
8828 | return err; | |
6f505b16 PZ |
8829 | } |
8830 | ||
d0b27fa7 PZ |
8831 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) |
8832 | { | |
8833 | u64 rt_runtime, rt_period; | |
8834 | ||
8835 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8836 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
8837 | if (rt_runtime_us < 0) | |
8838 | rt_runtime = RUNTIME_INF; | |
8839 | ||
8840 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8841 | } | |
8842 | ||
9f0c1e56 PZ |
8843 | long sched_group_rt_runtime(struct task_group *tg) |
8844 | { | |
8845 | u64 rt_runtime_us; | |
8846 | ||
d0b27fa7 | 8847 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) |
9f0c1e56 PZ |
8848 | return -1; |
8849 | ||
d0b27fa7 | 8850 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; |
9f0c1e56 PZ |
8851 | do_div(rt_runtime_us, NSEC_PER_USEC); |
8852 | return rt_runtime_us; | |
8853 | } | |
d0b27fa7 PZ |
8854 | |
8855 | int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) | |
8856 | { | |
8857 | u64 rt_runtime, rt_period; | |
8858 | ||
8859 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; | |
8860 | rt_runtime = tg->rt_bandwidth.rt_runtime; | |
8861 | ||
8862 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8863 | } | |
8864 | ||
8865 | long sched_group_rt_period(struct task_group *tg) | |
8866 | { | |
8867 | u64 rt_period_us; | |
8868 | ||
8869 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8870 | do_div(rt_period_us, NSEC_PER_USEC); | |
8871 | return rt_period_us; | |
8872 | } | |
8873 | ||
8874 | static int sched_rt_global_constraints(void) | |
8875 | { | |
8876 | int ret = 0; | |
8877 | ||
8878 | mutex_lock(&rt_constraints_mutex); | |
8879 | if (!__rt_schedulable(NULL, 1, 0)) | |
8880 | ret = -EINVAL; | |
8881 | mutex_unlock(&rt_constraints_mutex); | |
8882 | ||
8883 | return ret; | |
8884 | } | |
8885 | #else | |
8886 | static int sched_rt_global_constraints(void) | |
8887 | { | |
ac086bc2 PZ |
8888 | unsigned long flags; |
8889 | int i; | |
8890 | ||
8891 | spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8892 | for_each_possible_cpu(i) { | |
8893 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
8894 | ||
8895 | spin_lock(&rt_rq->rt_runtime_lock); | |
8896 | rt_rq->rt_runtime = global_rt_runtime(); | |
8897 | spin_unlock(&rt_rq->rt_runtime_lock); | |
8898 | } | |
8899 | spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8900 | ||
d0b27fa7 PZ |
8901 | return 0; |
8902 | } | |
052f1dc7 | 8903 | #endif |
d0b27fa7 PZ |
8904 | |
8905 | int sched_rt_handler(struct ctl_table *table, int write, | |
8906 | struct file *filp, void __user *buffer, size_t *lenp, | |
8907 | loff_t *ppos) | |
8908 | { | |
8909 | int ret; | |
8910 | int old_period, old_runtime; | |
8911 | static DEFINE_MUTEX(mutex); | |
8912 | ||
8913 | mutex_lock(&mutex); | |
8914 | old_period = sysctl_sched_rt_period; | |
8915 | old_runtime = sysctl_sched_rt_runtime; | |
8916 | ||
8917 | ret = proc_dointvec(table, write, filp, buffer, lenp, ppos); | |
8918 | ||
8919 | if (!ret && write) { | |
8920 | ret = sched_rt_global_constraints(); | |
8921 | if (ret) { | |
8922 | sysctl_sched_rt_period = old_period; | |
8923 | sysctl_sched_rt_runtime = old_runtime; | |
8924 | } else { | |
8925 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
8926 | def_rt_bandwidth.rt_period = | |
8927 | ns_to_ktime(global_rt_period()); | |
8928 | } | |
8929 | } | |
8930 | mutex_unlock(&mutex); | |
8931 | ||
8932 | return ret; | |
8933 | } | |
68318b8e | 8934 | |
052f1dc7 | 8935 | #ifdef CONFIG_CGROUP_SCHED |
68318b8e SV |
8936 | |
8937 | /* return corresponding task_group object of a cgroup */ | |
2b01dfe3 | 8938 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) |
68318b8e | 8939 | { |
2b01dfe3 PM |
8940 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), |
8941 | struct task_group, css); | |
68318b8e SV |
8942 | } |
8943 | ||
8944 | static struct cgroup_subsys_state * | |
2b01dfe3 | 8945 | cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) |
68318b8e | 8946 | { |
ec7dc8ac | 8947 | struct task_group *tg, *parent; |
68318b8e | 8948 | |
2b01dfe3 | 8949 | if (!cgrp->parent) { |
68318b8e | 8950 | /* This is early initialization for the top cgroup */ |
2b01dfe3 | 8951 | init_task_group.css.cgroup = cgrp; |
68318b8e SV |
8952 | return &init_task_group.css; |
8953 | } | |
8954 | ||
ec7dc8ac DG |
8955 | parent = cgroup_tg(cgrp->parent); |
8956 | tg = sched_create_group(parent); | |
68318b8e SV |
8957 | if (IS_ERR(tg)) |
8958 | return ERR_PTR(-ENOMEM); | |
8959 | ||
8960 | /* Bind the cgroup to task_group object we just created */ | |
2b01dfe3 | 8961 | tg->css.cgroup = cgrp; |
68318b8e SV |
8962 | |
8963 | return &tg->css; | |
8964 | } | |
8965 | ||
41a2d6cf IM |
8966 | static void |
8967 | cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
68318b8e | 8968 | { |
2b01dfe3 | 8969 | struct task_group *tg = cgroup_tg(cgrp); |
68318b8e SV |
8970 | |
8971 | sched_destroy_group(tg); | |
8972 | } | |
8973 | ||
41a2d6cf IM |
8974 | static int |
8975 | cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, | |
8976 | struct task_struct *tsk) | |
68318b8e | 8977 | { |
b68aa230 PZ |
8978 | #ifdef CONFIG_RT_GROUP_SCHED |
8979 | /* Don't accept realtime tasks when there is no way for them to run */ | |
d0b27fa7 | 8980 | if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0) |
b68aa230 PZ |
8981 | return -EINVAL; |
8982 | #else | |
68318b8e SV |
8983 | /* We don't support RT-tasks being in separate groups */ |
8984 | if (tsk->sched_class != &fair_sched_class) | |
8985 | return -EINVAL; | |
b68aa230 | 8986 | #endif |
68318b8e SV |
8987 | |
8988 | return 0; | |
8989 | } | |
8990 | ||
8991 | static void | |
2b01dfe3 | 8992 | cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
68318b8e SV |
8993 | struct cgroup *old_cont, struct task_struct *tsk) |
8994 | { | |
8995 | sched_move_task(tsk); | |
8996 | } | |
8997 | ||
052f1dc7 | 8998 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f4c753b7 | 8999 | static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, |
2b01dfe3 | 9000 | u64 shareval) |
68318b8e | 9001 | { |
2b01dfe3 | 9002 | return sched_group_set_shares(cgroup_tg(cgrp), shareval); |
68318b8e SV |
9003 | } |
9004 | ||
f4c753b7 | 9005 | static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) |
68318b8e | 9006 | { |
2b01dfe3 | 9007 | struct task_group *tg = cgroup_tg(cgrp); |
68318b8e SV |
9008 | |
9009 | return (u64) tg->shares; | |
9010 | } | |
052f1dc7 | 9011 | #endif |
68318b8e | 9012 | |
052f1dc7 | 9013 | #ifdef CONFIG_RT_GROUP_SCHED |
0c70814c | 9014 | static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, |
06ecb27c | 9015 | s64 val) |
6f505b16 | 9016 | { |
06ecb27c | 9017 | return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); |
6f505b16 PZ |
9018 | } |
9019 | ||
06ecb27c | 9020 | static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) |
6f505b16 | 9021 | { |
06ecb27c | 9022 | return sched_group_rt_runtime(cgroup_tg(cgrp)); |
6f505b16 | 9023 | } |
d0b27fa7 PZ |
9024 | |
9025 | static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, | |
9026 | u64 rt_period_us) | |
9027 | { | |
9028 | return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); | |
9029 | } | |
9030 | ||
9031 | static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) | |
9032 | { | |
9033 | return sched_group_rt_period(cgroup_tg(cgrp)); | |
9034 | } | |
052f1dc7 | 9035 | #endif |
6f505b16 | 9036 | |
fe5c7cc2 | 9037 | static struct cftype cpu_files[] = { |
052f1dc7 | 9038 | #ifdef CONFIG_FAIR_GROUP_SCHED |
fe5c7cc2 PM |
9039 | { |
9040 | .name = "shares", | |
f4c753b7 PM |
9041 | .read_u64 = cpu_shares_read_u64, |
9042 | .write_u64 = cpu_shares_write_u64, | |
fe5c7cc2 | 9043 | }, |
052f1dc7 PZ |
9044 | #endif |
9045 | #ifdef CONFIG_RT_GROUP_SCHED | |
6f505b16 | 9046 | { |
9f0c1e56 | 9047 | .name = "rt_runtime_us", |
06ecb27c PM |
9048 | .read_s64 = cpu_rt_runtime_read, |
9049 | .write_s64 = cpu_rt_runtime_write, | |
6f505b16 | 9050 | }, |
d0b27fa7 PZ |
9051 | { |
9052 | .name = "rt_period_us", | |
f4c753b7 PM |
9053 | .read_u64 = cpu_rt_period_read_uint, |
9054 | .write_u64 = cpu_rt_period_write_uint, | |
d0b27fa7 | 9055 | }, |
052f1dc7 | 9056 | #endif |
68318b8e SV |
9057 | }; |
9058 | ||
9059 | static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) | |
9060 | { | |
fe5c7cc2 | 9061 | return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); |
68318b8e SV |
9062 | } |
9063 | ||
9064 | struct cgroup_subsys cpu_cgroup_subsys = { | |
38605cae IM |
9065 | .name = "cpu", |
9066 | .create = cpu_cgroup_create, | |
9067 | .destroy = cpu_cgroup_destroy, | |
9068 | .can_attach = cpu_cgroup_can_attach, | |
9069 | .attach = cpu_cgroup_attach, | |
9070 | .populate = cpu_cgroup_populate, | |
9071 | .subsys_id = cpu_cgroup_subsys_id, | |
68318b8e SV |
9072 | .early_init = 1, |
9073 | }; | |
9074 | ||
052f1dc7 | 9075 | #endif /* CONFIG_CGROUP_SCHED */ |
d842de87 SV |
9076 | |
9077 | #ifdef CONFIG_CGROUP_CPUACCT | |
9078 | ||
9079 | /* | |
9080 | * CPU accounting code for task groups. | |
9081 | * | |
9082 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh | |
9083 | * (balbir@in.ibm.com). | |
9084 | */ | |
9085 | ||
9086 | /* track cpu usage of a group of tasks */ | |
9087 | struct cpuacct { | |
9088 | struct cgroup_subsys_state css; | |
9089 | /* cpuusage holds pointer to a u64-type object on every cpu */ | |
9090 | u64 *cpuusage; | |
9091 | }; | |
9092 | ||
9093 | struct cgroup_subsys cpuacct_subsys; | |
9094 | ||
9095 | /* return cpu accounting group corresponding to this container */ | |
32cd756a | 9096 | static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) |
d842de87 | 9097 | { |
32cd756a | 9098 | return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), |
d842de87 SV |
9099 | struct cpuacct, css); |
9100 | } | |
9101 | ||
9102 | /* return cpu accounting group to which this task belongs */ | |
9103 | static inline struct cpuacct *task_ca(struct task_struct *tsk) | |
9104 | { | |
9105 | return container_of(task_subsys_state(tsk, cpuacct_subsys_id), | |
9106 | struct cpuacct, css); | |
9107 | } | |
9108 | ||
9109 | /* create a new cpu accounting group */ | |
9110 | static struct cgroup_subsys_state *cpuacct_create( | |
32cd756a | 9111 | struct cgroup_subsys *ss, struct cgroup *cgrp) |
d842de87 SV |
9112 | { |
9113 | struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); | |
9114 | ||
9115 | if (!ca) | |
9116 | return ERR_PTR(-ENOMEM); | |
9117 | ||
9118 | ca->cpuusage = alloc_percpu(u64); | |
9119 | if (!ca->cpuusage) { | |
9120 | kfree(ca); | |
9121 | return ERR_PTR(-ENOMEM); | |
9122 | } | |
9123 | ||
9124 | return &ca->css; | |
9125 | } | |
9126 | ||
9127 | /* destroy an existing cpu accounting group */ | |
41a2d6cf | 9128 | static void |
32cd756a | 9129 | cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) |
d842de87 | 9130 | { |
32cd756a | 9131 | struct cpuacct *ca = cgroup_ca(cgrp); |
d842de87 SV |
9132 | |
9133 | free_percpu(ca->cpuusage); | |
9134 | kfree(ca); | |
9135 | } | |
9136 | ||
9137 | /* return total cpu usage (in nanoseconds) of a group */ | |
32cd756a | 9138 | static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) |
d842de87 | 9139 | { |
32cd756a | 9140 | struct cpuacct *ca = cgroup_ca(cgrp); |
d842de87 SV |
9141 | u64 totalcpuusage = 0; |
9142 | int i; | |
9143 | ||
9144 | for_each_possible_cpu(i) { | |
9145 | u64 *cpuusage = percpu_ptr(ca->cpuusage, i); | |
9146 | ||
9147 | /* | |
9148 | * Take rq->lock to make 64-bit addition safe on 32-bit | |
9149 | * platforms. | |
9150 | */ | |
9151 | spin_lock_irq(&cpu_rq(i)->lock); | |
9152 | totalcpuusage += *cpuusage; | |
9153 | spin_unlock_irq(&cpu_rq(i)->lock); | |
9154 | } | |
9155 | ||
9156 | return totalcpuusage; | |
9157 | } | |
9158 | ||
0297b803 DG |
9159 | static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, |
9160 | u64 reset) | |
9161 | { | |
9162 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9163 | int err = 0; | |
9164 | int i; | |
9165 | ||
9166 | if (reset) { | |
9167 | err = -EINVAL; | |
9168 | goto out; | |
9169 | } | |
9170 | ||
9171 | for_each_possible_cpu(i) { | |
9172 | u64 *cpuusage = percpu_ptr(ca->cpuusage, i); | |
9173 | ||
9174 | spin_lock_irq(&cpu_rq(i)->lock); | |
9175 | *cpuusage = 0; | |
9176 | spin_unlock_irq(&cpu_rq(i)->lock); | |
9177 | } | |
9178 | out: | |
9179 | return err; | |
9180 | } | |
9181 | ||
d842de87 SV |
9182 | static struct cftype files[] = { |
9183 | { | |
9184 | .name = "usage", | |
f4c753b7 PM |
9185 | .read_u64 = cpuusage_read, |
9186 | .write_u64 = cpuusage_write, | |
d842de87 SV |
9187 | }, |
9188 | }; | |
9189 | ||
32cd756a | 9190 | static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) |
d842de87 | 9191 | { |
32cd756a | 9192 | return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); |
d842de87 SV |
9193 | } |
9194 | ||
9195 | /* | |
9196 | * charge this task's execution time to its accounting group. | |
9197 | * | |
9198 | * called with rq->lock held. | |
9199 | */ | |
9200 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime) | |
9201 | { | |
9202 | struct cpuacct *ca; | |
9203 | ||
9204 | if (!cpuacct_subsys.active) | |
9205 | return; | |
9206 | ||
9207 | ca = task_ca(tsk); | |
9208 | if (ca) { | |
9209 | u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk)); | |
9210 | ||
9211 | *cpuusage += cputime; | |
9212 | } | |
9213 | } | |
9214 | ||
9215 | struct cgroup_subsys cpuacct_subsys = { | |
9216 | .name = "cpuacct", | |
9217 | .create = cpuacct_create, | |
9218 | .destroy = cpuacct_destroy, | |
9219 | .populate = cpuacct_populate, | |
9220 | .subsys_id = cpuacct_subsys_id, | |
9221 | }; | |
9222 | #endif /* CONFIG_CGROUP_CPUACCT */ |